diff --git a/parse/train/B1lG42C9Km/B1lG42C9Km.md b/parse/train/B1lG42C9Km/B1lG42C9Km.md new file mode 100644 index 0000000000000000000000000000000000000000..b956f92b012a5b214faf661ec0dbbb0bad50b7ac --- /dev/null +++ b/parse/train/B1lG42C9Km/B1lG42C9Km.md @@ -0,0 +1,392 @@ +# INTRINSIC SOCIAL MOTIVATION VIA CAUSAL INFLUENCE IN MULTI-AGENT RL + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +We derive a new intrinsic social motivation for multi-agent reinforcement learning (MARL), in which agents are rewarded for having causal influence over another agent’s actions. Causal influence is assessed using counterfactual reasoning. The reward does not depend on observing another agent’s reward function, and is thus a more realistic approach to MARL than taken in previous work. We show that the causal influence reward is related to maximizing the mutual information between agents’ actions. We test the approach in challenging social dilemma environments, where it consistently leads to enhanced cooperation between agents and higher collective reward. Moreover, we find that rewarding influence can lead agents to develop emergent communication protocols. We therefore employ influence to train agents to use an explicit communication channel, and find that it leads to more effective communication and higher collective reward. Finally, we show that influence can be computed by equipping each agent with an internal model that predicts the actions of other agents. This allows the social influence reward to be computed without the use of a centralised controller, and as such represents a significantly more general and scalable inductive bias for MARL with independent agents. + +# 1 INTRODUCTION + +Deep reinforcement learning (RL) has made impressive progress on specific tasks with well-defined reward functions, but is still difficult to learn intelligent behavior that generalizes across multiple domains. Intrinsic motivation is a technique for solving this problem by developing general reward functions that encourage an agent to learn across a variety of tasks (Singh et al., 2004). Previous approaches to intrinsic motivation have broadly fallen into two categories: (1) curiosity, or a drive for novelty (e.g. Pathak et al. (2017); Schmidhuber (2010)), and (2) empowerment, or a drive to be able to manipulate the environment (Klyubin et al., 2005). + +We posit that this body of work has largely overlooked an important intrinsic motivation that is key to human learning: social interaction. Humans have remarkable social learning abilities; some authors suggest that it is social learning that has given rise to cultural evolution, and allowed us to achieve unprecedented progress and coordination on a massive scale (van Schaik & Burkart, 2011; Herrmann et al., 2007). Others emphasize that our impressive capacity to learn from others far surpasses that of other animals, apes, and even other proto-human species (Henrich, 2015; Harari, 2014; Laland, 2017). + +Therefore, we propose an intrinsic reward function designed for multi-agent RL (MARL), which awards agents for having a causal influence on other agents’ actions. Causal influence is assessed using counterfactual reasoning; at each timestep, an agent simulates alternate, counterfactual actions that it could have taken, and assesses their effect on another agent’s behavior. Actions that lead to relatively higher change in the other agent are considered to be highly influential and are rewarded. We show how this reward is related to maximizing the mutual information between agents’ actions, and is thus a form of social empowerment. We hypothesize that rewarding influence may therefore encourage cooperation between agents. We also take inspiration from experiments in human cognition, showing that newborn infants are sensitive to correspondences between their own actions and the actions of other people, and use this to coordinate their behavior with others (Tomasello, 2009; Melis & Semmann, 2010). + +To study our proposed social influence reward in the MARL setting, we adopt the Sequential Social Dilemmas (SSDs) of Leibo et al. (2017). These are challenging MA environments with a game-theoretic reward structure, similar to Prisoner’s Dilemma. For each individual agent, ‘defecting’ (non-cooperative behavior) has the highest payoff. However, the collective reward will be better if all agents choose to cooperate. The paradoxical payoff structure of these tasks make achieving cooperative social dynamics extremely challenging for typical RL agents. We show that social influence allows agents to learn to cooperate in these environments, and make the following contributions: + +• We demonstrate that deep RL agents trained with the proposed social influence reward cooperate to attain higher collective reward than baseline deep RL agents (Mnih et al., 2016). In some cases, this cooperation is attained because influencer agents learn to use their actions as an emergent communication protocol, analogous to behavior seen in animals (von Frisch, 1969). Motivated by the previous point, we apply the influence reward to training deep RL agents to use an explicit communication channel, as in Foerster et al. (2016). We demonstrate that the communication protocols trained with the influence reward meaningfully relate to agents’ actions, and that once again, agents trained with the influence reward achieve better collective outcomes. We demonstrate that there is a significant correlation between being influenced through communication messages and obtaining higher individual return, suggesting that influential communication is beneficial to the agents that receive it. Finally, rather than computing social influence using a centralised training framework as in prior work (e.g. Foerster et al. (2017; 2016)), we extend the approach by attaching an internal Model of Other Agents (MOA) network to each agent and training it to predict the actions of every other agent. The agent can then simulate counterfactual actions and use its own internal MOA to predict how these will affect other agents, thus computing its own intrinsic influence reward. + +Using a MOA to predict and reward influence allows us to compute an intrinsic social reward by observing other agents’ past actions, without a centralised controller, and without requiring access to another agent’s reward function. We believe this is an important innovation over prior work (e.g. (Hughes et al., 2018; Foerster et al., 2017; 2016)). When we consider likely future applications of MARL, such as autonomous driving, it becomes apparent that centralised training or the sharing of reward functions are unrealistic assumptions, since autonomous vehicles are likely to be produced by a wide variety of organizations and institutions with mixed motivations. Rather, a social reward function which only depends on observing the behavior of agents acting in the environment, and which can give rise to coordinated, cooperative behavior, represents a more promising approach. + +# 2 METHODS + +We consider a MARL Markov game defined by the tuple $\langle S , T , A , r \rangle$ , in which multiple agents which do not share weights are trained to independently maximize their own individual reward. The environment state is given by $s \in S$ . At each timestep $t$ , each agent $k$ chooses an action $a _ { t } ^ { k } \in A$ . The actions of all $N$ agents are combined to form a joint action $\pmb { a } _ { t } = [ a _ { t } ^ { 0 } , . . . a _ { t } ^ { N } ]$ , which produces a transition in the environment $T _ { \cdot } ( s _ { t + 1 } | \mathbf { a } _ { t } , s _ { t } )$ , according to the state transition function $T$ . Each agent then receives its own reward $r ^ { k } ( a _ { t } , s _ { t } )$ , which may depend on the actions of other agents. A history of these variables over time is termed a trajectory, ${ \boldsymbol { \tau } } = \left\{ { { s _ { t } } , { a _ { t } } , { r _ { t } } } \right\} _ { t = 0 } ^ { T }$ . We consider a partially observable setting in which each agent $k$ can only vieture reward, true sta, where , $s _ { t } ^ { k }$ . Each agent seeks to maximize its own total discount factor. A distributed asynchronous a ectedntage $\begin{array} { r } { R ^ { k } = \sum _ { i = 0 } ^ { \infty } \gamma ^ { i } r _ { t + i } ^ { k } } \end{array}$ $\gamma$ +$\pi ^ { k }$ policy is learned via REINFORCE with baseline (Williams, 1992). Architecturally, our agents consist of a convolutional layer, fully connected layers, a Long Short Term Memory (LSTM) network (Gers et al., 1999), and linear layers which output $\pi ^ { \dot { k } }$ and the value function $V ^ { \pi _ { k } } ( s )$ . We will refer to the internal LSTM state of agent $k$ at timestep $t$ as $\dot { u } _ { t } ^ { k }$ . + +# 2.1 INTRINSIC SOCIAL MOTIVATION VIA CAUSAL INFLUENCE + +Social influence intrinsic motivation modifies an agent’s reward function so that it becomes $R ^ { k } = \alpha E ^ { k } + \beta I ^ { k }$ , where $E ^ { k }$ is the extrinsic or environmental reward, and $I ^ { k }$ is the causal influence reward. We compute $I ^ { k }$ by generating counterfactual actions that the agent could have taken at each timestep, and assessing how taking these would have affected other agents’ behavior. A counterfactual is the estimated probability that $^ { 6 6 } Y$ would be $y$ had $X$ been $x$ , in situation $Z = z ^ { \prime }$ , where $X , Y ,$ , and $Z$ are random variables, and $x , y$ and $z$ are their values (Pearl et al., 2016). Importantly, it is a counterfactual because we condition on a set of evidence $z$ , and because the assignment $X = x$ is counter to what we actually observed; in reality, $X$ took on some other value. + +![](images/d4c9d865d0914f582cec2f90cb9102528669a3d0feda5b6e015295234f191f06.jpg) +Figure 1: Causal diagram of agent $A$ ’s effect on $B$ ’s action. We condition on each agent’s view of the environment and LSTM state $u$ (shaded nodes), and intervene on $a _ { t } ^ { \dot { A } }$ (blue). + +To see how we can compute the causal effect of one agent on another, suppose there are two agents, $A$ and $B$ , and that agent $B$ receives $A$ ’s action at time $t$ taction, know a , as input1. Agent $p \big ( a _ { t } ^ { B } | a _ { t } ^ { A } , \overline { { s _ { t } ^ { B } } } , u _ { t } ^ { B } \big )$ $B$ then uses this to compute a distribution over its own ecause we have built the model , and its own internal LSTM state, agent , as sh $B$ , wen in $\bar { a } _ { t } ^ { A } , s _ { t } ^ { B }$ $u _ { t } ^ { B }$ Figure 1. This allows us to exactly isolate the causal effect of $A$ ’s action on $B$ by conditioning on the values we observed for the other inputs at this timestep (note that by conditioning on these variables, including the LSTM state $u _ { t } ^ { B }$ , we remove any dependency on previous timesteps in the trajectory). We can then intervene on $a _ { t } ^ { A }$ by replacing it with a counterfactual action, $d o \dot { ( a _ { t } ^ { A } ) }$ . This counterfactual action is used to compute a new estimate of $p ( a _ { t } ^ { B } | d o ( \tilde { a } _ { t } ^ { A } ) , s _ { t } ^ { B } , u _ { t } ^ { B } )$ . Essentially, the agent asks a retrospective question: “How would $B$ ’s action change if I had acted differently in this situation?”. + +To simplify notation, let $\boldsymbol { z } _ { t } = \langle \boldsymbol { u } _ { t } ^ { B } , \boldsymbol { s } _ { t } ^ { B } \rangle$ , so that conditioning on $z _ { t }$ is equivalent to conditioning on all relevant background variables (all shaded variables in Figure 1). We can also forego the $d o$ operator, noting that $p ( \dot { a } _ { t } ^ { B } | d o ( \tilde { a } _ { t } ^ { A } ) , z _ { t } ) \equiv p ( a _ { t } ^ { B } | \tilde { a } _ { t } ^ { \tilde { A } } , z _ { t } )$ in this case, because $z$ satisfies the back-door criterion (Pearl & Mackenzie, 2018). Now, consider averaging over several counterfactuals $\tilde { a } _ { t } ^ { A }$ . This gives us the marginal policy of $B$ , $\begin{array} { r } { p ( a _ { t } ^ { B } | \tilde { z _ { t } } ) = \sum _ { \tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } | \tilde { a } _ { t } ^ { A } , z _ { t } ) p ( \tilde { a } _ { t } ^ { A } | z _ { t } ) } \end{array}$ —in other words, $B$ ’s policy if $A$ were not considered. The discrepancy between the marginal policy of $B$ and the conditional policy of $B$ given $A$ ’s action is a measure of the causal influence of + +$A$ on $B$ ; it gives the degree to which $B$ changes its planned action distribution because of $A$ ’s behavior. Thus, the causal influence intrinsic reward for agent $A$ is + +$$ +{ I _ { t } ^ { A } } = D _ { K L } \left[ p ( a _ { t } ^ { B } \mid a _ { t } ^ { A } , z _ { t } ) \right| \left| \sum _ { \tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } \mid z _ { t } , \tilde { a } _ { t } ^ { A } ) p ( \tilde { a } _ { t } ^ { A } \mid z _ { t } ) \right] = D _ { K L } \left[ p ( a _ { t } ^ { B } \mid a _ { t } ^ { A } , z _ { t } ) \right| \left| p ( a _ { t } ^ { B } \mid z _ { t } ) \right] . +$$ + +# 2.2 RELATIONSHIP TO MUTUAL INFORMATION AND EMPOWERMENT + +The causal influence reward in Eq. 1 is related to the mutual information (MI) between the actions of agents $A$ and $B$ , which is given by + +$$ +I ( A ^ { B } ; A ^ { A } | z ) = \sum _ { a ^ { A } , a ^ { B } } p ( a ^ { B } , a ^ { A } | z ) \mathrm { l o g } \frac { p ( a ^ { B } , a ^ { A } | z ) } { p ( a ^ { B } | z ) p ( a ^ { A } | z ) } { = \sum _ { a ^ { A } } p ( a ^ { A } | z ) D _ { \mathrm { K L } } \left[ p ( a ^ { B } | a ^ { A } , z ) \right] \left| p ( a ^ { B } | z ) \right] } , +$$ + +where we see that the $D _ { K L }$ factor in Eq. 2 is the causal influence reward given in Eq. 1. The connection to mutual information is interesting, because a frequently used intrinsic motivation for single agent RL is empowerment, which rewards the agent for having high mutual information between its actions and the future state of the environment (e.g. Klyubin et al. (2005); Capdepuy et al. (2007)). To the extent that the social influence reward defined in Eq. 1 is an approximation of the MI, $A$ is rewarded for having empowerment over $B ^ { \prime } s$ actions. + +By sampling $N$ independent trajectories $\tau _ { n }$ from the environment, where A’s actions $a _ { n } ^ { A }$ are drawn according to $\overset { \cdot } { p } ( a ^ { A } | z )$ , we perform a Monte-Carlo approximation of the MI (see e.g. Strouse et al. (2018)), + +$$ +I ( A ^ { A } ; A ^ { B } | z ) = \mathbb { E } _ { \tau } \Big [ D _ { \mathrm { K L } } \big [ p ( A ^ { B } | A ^ { A } , z ) \big \| p ( A ^ { B } | z ) \big ] \Big | z \Big ] \approx \frac { 1 } { N } \sum _ { n } D _ { \mathrm { K L } } \big [ p ( A ^ { B } | a _ { n } ^ { A } , z ) \big \| p ( A ^ { B } | z ) \big ] . +$$ + +Thus, in expectation, the social influence reward is the MI between agents’ actions. + +Whether the policy trained with Eq. 1 actually learns to approximate the MI depends on the learning dynamics. We calculate the intrinsic social influence reward using Eq. 1, because unlike Eq. 2, which gives an estimate of the symmetric bandwidth between $A$ and $B$ , Eq. 1 gives the directed causal effect of the specific action taken by agent $A$ , $a _ { t } ^ { A }$ . We believe this will result in an easier reward to learn, since it allows for better credit assignment; agent $A$ can more easily learn which of its actions lead to high influence. We also experiment with replacing the KL-divergence with several other measures, including the Jensen-Shannon Divergence (JSD), and find that the influence reward is robust to the choice of measure. + +# 2.3 INFLUENCE THROUGH COMMUNICATION + +According to Melis & Semmann (2010), human children rapidly learn to use communication to influence the behavior of others when engaging in cooperative activities. They explain that “this ability to influence the partner via communication has been interpreted as evidence for a capacity to form shared goals with others”, and that this capacity may be “what allows humans to engage in a wide range of cooperative activities”. Therefore, we investigate a second use of the social influence reward: learning inter-agent communication protocols. Using a similar approach to Reinforced Inter-Agent Learning (RIAL) (Foerster et al., 2016), we equip the agents with an explicit communication channel. At each timestep, each agent $k$ chooses a discrete communication symbol $\Dot { m } _ { t } ^ { k }$ ; these symbols are concatenated into a combined message vector $\pmb { m } _ { t } = [ m _ { t } ^ { 0 } , m _ { t } ^ { 1 } . . . m _ { t } ^ { N } ] ,$ , for $N$ agents. This message vector $\mathbf { \nabla } m _ { t }$ is then shown to every other agent in the next timestep, as in Figure 2. To train the agents to communicate, we augment our initial network with an additional A3C output head, that learns a communication policy $\pi _ { c }$ over which symbol to emit, and a communication value function $V _ { c }$ (this is separate from the normal policy and value function used for acting in the environment, $\pi _ { e }$ and $V _ { e }$ , which are trained only with environmental reward $E$ ). + +The influence reward is used, in addition to environmental reward, to train the communication policy $\pi _ { c }$ Counterfactuals are employed to assess how muon another agent’s action in the next timestep, luence an agent’s communication message, . Importantly, we hypothesize that comm $m _ { t } ^ { A }$ , hastion $a _ { t + 1 } ^ { B }$ can only be influential if it is useful to another agent. There is nothing that compels agent to act based on agent $A$ ’s communication message; if it does not contain valuable information, $B$ is free to ignore it. In fact, previous work has shown that selfish agents do not learn to use this type of ungrounded, cheap talk communication channel effectively (Cao et al., 2018). In contrast, for $A$ to gain influence via communication, $m _ { t } ^ { A }$ must contain valuable information that informs $B$ about how best to maximize its own reward, so much so that it actually causes $B$ to change its intended action. + +![](images/caa2b9dc3c9fe0bdc22237d63bfba2b09073071b89e2a88a719924f5aa8e5a38.jpg) +Figure 2: The communication model has two A2C heads, which learn a normal policy, $\pi _ { e }$ , and a policy for emitting communication symbols, $\pi _ { c }$ . Other agents’ communication messages $\mathbf { \nabla } m _ { t }$ are input to the LSTM. + +![](images/64e00c7a40f5d8bfca22499a480991f9780f7757dc5984363b283b0279a25fc0.jpg) +Figure 3: The Model of Other Agents (MOA) architecture learns both an RL policy $\pi _ { e }$ , and a supervised model that predicts the actions of other agents, $\mathbf { \pmb { a } } _ { t + 1 }$ The predictions of the supervised model are used for computing the influence reward. + +# 2.4 INFLUENCE VIA MODELING OTHER AGENTS + +Computing the causal influence reward as introduced in Section 2.1 requires knowing the probability of $B$ ’s next action given a counterfactual, $p ( a _ { t } ^ { B } | \tilde { a } _ { t } ^ { A } , s _ { t } ^ { B } )$ , which we previously solved by using a centralised controller that could access other agent’s policy networks. While using a centralised training framework is common in MARL (e.g. Foerster et al. (2017; 2016)), it is less realistic than a scenario in which each agent is trained independently. We can relax this assumption and achieve independent training by equipping each agent with its own internal Model of Other Agents (MOA). The MOA consists of a second set of fully-connected and LSTM layers connected to the agent’s convolutional layer (see Figure 3), and is trained to predict all other agents’ next actions given their current action, and the agent’s egocentric view of the state: $p ( { \pmb a } _ { t + 1 } | { \pmb a } _ { t } , { \pmb s } _ { t } ^ { A } )$ . The MOA is trained using cross-entropy loss over observed action trajectories. + +A trained MOA can be used to compute the social influence reward in the following way. Each agent can “imagine” counterfactual actions that it could have taken at each timestep, and use its internal MOA to predict the effect on other agents. It can then give itself reward for taking actions that it estimates were the most influential. This has an intuitive appeal, because it resembles how humans reason about their effect on others (Ferguson et al., 2010). We may often find ourselves asking counterfactual questions of the form, “How would she have reacted if I had said or done something else in that situation?”, which we can only answer using our internal model of others. + +Both the MOA and communication approaches are an important improvement over the original model shown in Figure 1, which computed influence within a given timestep and required that agent $A$ choose its action $a _ { t } ^ { A }$ first, and this action be transmitted to agent $B$ as input. This meant that only some agents (those acting first) could be influencers. In contrast, using influence for communication or with a MOA are general approaches that can be implemented in any agent, and allow all agents to mutually influence each other. + +We now seek to estimate influence in the next timestep, meaning the influence of $a _ { t } ^ { A }$ on $a _ { t + 1 } ^ { B }$ , which requires modeling $p ( a _ { t + 1 } ^ { B } | \underline { { a } } _ { t } ^ { A } , s _ { t } ^ { A } )$ . The corresponding causal diagrof $a _ { t } ^ { A }$ is son $a _ { t + 1 } ^ { B }$ in Figure 4. We can infer the causal effectby conditioning on the shaded variables (so that there are no back-door paths) (Pearl et al., 2016). Learning a model of $p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { \bar { A } } , s _ { t } ^ { A } )$ requires implicitly modeling both the environment transition function $T$ (to predict $s _ { t + 1 } )$ , as well as relevant aspects of the internal LSTM state of the other agent, $u _ { t + 1 } ^ { B ^ { * } }$ , as highlighted in Figure 4. + +We enable agents to condition their policy on the actions of other agents in the previous timestep (actions are visible), and only give the social influence reward to an agent when the agent it is attempting to influence is within its field-of-view, because the estimates of $p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { A } )$ are likely to be more accurate when $B$ is visible to $A ^ { 2 }$ . The latter constraint could have the sideeffect of encouraging agents to stay in closer proximity. + +![](images/c636b44cc733a0638da4fbce914888df05874d72cc520f1bcf72e309b465e051.jpg) +Figure 4: Causal diagram in the MOA case. Shaded nodes are conditioned on, and we intervene on $a _ { t } ^ { A }$ (blue node) by replacing it with counterfactuals. Nodes with a green background must be modeled using the MOA module. Note that there is no backdoor path between $a _ { t } ^ { A }$ and $s _ { t }$ since it would require traversing a collider that is not in the conditioning set. + +However, an intrinsic social reward based on proximity is also a reasonable approach to approximating human social motivation. Humans seek affiliation and to spend time near other people (Tomasello, 2009). + +# 2.5 SEQUENTIAL SOCIAL DILEMMAS + +First proposed by Leibo et al. (2017), Sequential Social Dilemmas (SSDs) are spatially and temporally extended multi-agent games that have a payoff structure similar to that of Prisoner’s Dilemma (PD). That is, an individual agent can obtain higher reward by engaging in defecting, non-cooperative behavior (and thus is rationally motivated to defect), but the average payoff per agent will be higher if all agents cooperate (see Figure 9 of the Appendix). The paradoxical reward structure makes it extremely difficult for traditional RL agents to learn to coordinate to solve the tasks (Hughes et al., 2018). We experiment with two SSDs in this work, a public goods game Cleanup, and a tragedy-of-the-commons game Harvest (see Figure 5). In both games apples (green tiles) provide the rewards, and agents also have the ability to punish each other with a fining beam. Further details are available in Appendix Section 6.1. + +![](images/20206990cd332b2c23f36bfab8c7c99fd397bf6323218026ab67daf49c621e5f.jpg) +Figure 5: The two SSD environments, Cleanup (left) and Harvest (right). Agents can exploit other agents for immediate payoff, but at the expense of the long-term collective reward of the group. + +# 3 RELATED WORK + +Several attempts have been made to develop intrinsic social motivation rewards3. Sequeira et al. (2011) developed hand-crafted rewards specific to a foraging environment, in which agents were punished for eating more than their fair share of food. Another approach gave agents an emotional intrinsic reward based on their perception of their neighbours’ cooperativeness in a networked version of the iterated prisoner’s dilemma (Yu et al., 2013). This approach is limited to scenarios in which it is possible to directly classify each action as cooperative or non-cooperative, which is untenable in complex settings with long-term strategies, such as the SSDs under investigation here. Hughes et al. (2018) introduced an inequity aversion motivation, which penalized agents if their rewards differed too much from those of the group. Another approach used prosocial reward shaping to show that if even a single agent is trained to optimize for the rewards of other agents, it can help the group obtain better collective outcomes (Peysakhovich & Lerer, 2018). However, these both require the ability to observe other agent’s rewards, which may be an unrealistic assumption, depending on the application. + +Another body of work has focused on training agents to learn emergent communication protocols (Foerster et al., 2016; Cao et al., 2018; Choi et al., 2018; Lazaridou et al., 2018; Bogin et al., 2018), with many authors finding that selfish agents do not learn to use an ungrounded, cheap talk communication channel effectively. Crawford & Sobel (1982) find that in theory, the information revealed in communication (in equilibrium) is proportional to amount of common interest; thus, as agents’ interests diverge, no communication is to be expected. And while communication can emerge when agents are prosocial (Foerster et al., 2016; Lazaridou et al., 2018) or hand-crafted (Crandall et al., 2017), self-interested agents do not to learn to communicate (Cao et al., 2018). We test whether the social influence reward can encourage agents to learn to communicate more effectively in complex environments with challenging social dilemma dynamics. + +Interestingly, Oudeyer & Kaplan (2006) show that a robot trained with a curiosity-based intrinsic motivation to maximize learning progress learns to prefer vocalizing sounds imitated by another robot over interaction with other objects in the environment. Follow-up papers suggest that curiosity may be a sufficient motivation to encourage agents, or even children, to learn to communicate with others (Oudeyer & Smith, 2016; Forestier & Oudeyer, 2017). + +Our MOA network is related to work on machine theory of mind (Rabinowitz et al., 2018), which demonstrated that a model trained to predict agents’ actions is able to model false beliefs. With LOLA, Foerster et al. (2018) train agents that model the impact of their policy on the parameter updates of other agents, and directly incorporate this into the agent’s own learning rule. + +Barton et al. (2018) propose causal influence as a way to measure coordination between agents, specifically using Convergence Cross Mapping (CCM) to analyze the degree of dependence between two agents’ policies. The limitation of this approach is that CCM estimates of causality are known to degrade in the presence of stochastic effects (Tajima et al., 2015). Counterfactual reasoning has also been used in a multi-agent setting, to marginalize out the effect of one agent on a predicted global value function estimating collective reward, and thus obtain an improved baseline for computing each agent’s advantage function (Foerster et al., 2017). A similar paper shows that counterfactuals can be used with potential-based reward shaping to improve credit assignment for training a joint policy in multi-agent RL Devlin et al. (2014). However, once again these approaches rely on a centralised controller. + +Following in the tradition of the empowerment literature, authors have investigated mutual information (MI) as a powerful tool for designing social rewards. Strouse et al. (2018) train agents to maximize or minimize the MI between their actions and a categorical goal, and show how this can be used to signal or hide the agent’s intentions. However, this approach depends on agents pursuing a known, categorical goal. Guckelsberger et al. (2018), in pursuit of the ultimate video game adversary, develop an agent that maximizes its empowerment over its own states, minimizes the player’s empowerment over their states, and maximizes its empowerment over the player’s next state. This third goal, termed transfer empowerment, is obtained by maximizing the MI between the agent’s actions and the player’s future state. While similar to our approach, the authors find that agents trained with transfer empowerment simply tend to stay near the player. Further, the agents are not trained with RL, but rather analytically compute these measures in simple grid-world environments. As such, the agent cannot learn to model other agents or the environment. + +# 4 EXPERIMENTS + +The following sections present the results of training agents with the social influence reward in three settings: (1) using a centralised controller, (2) using an explicit communication channel, and (3) using a learned model of other agents (MOA). In each case we compare against a standard A3C agent, and an ablated version of the model which is architecturally identical, but does not receive the influence reward. We measure the total collective reward obtained using the best hyperparameter setting tested with 5 random seeds. It is worth noting that we use a curriculum learning approach which gradually increases the weight of the social influence reward over $C$ steps $( C \in [ 0 . 2 \mathrm { - } 3 . 5 ] \times \bar { 1 } 0 ^ { 8 } )$ ; this can lead to a slight delay before the influence models’ performance improves. + +We also provide the results of an additional experiment Section 6.2 of the Appendix, which tests the social influence reward in a simplified environment where the effects of influence are clear. We encourage the reader to examine that section to gain a better intuition for how social influence can foster cooperative behavior in an otherwise selfish agent. + +# 4.1 CENTRALISED CONTROLLER + +Figures 6(a) and 6(d) show the results of training influence with a centralised controller as described in Section 2.1. With this method, the influencer agents transmit their intended action to the influenced agents at each timestep. Therefore, we benchmark against an ablated version of the influence model with visible actions but no influence reward. As is evident in Figures 6(a) and 6(d), introducing an awareness of other agents’ actions helps, but having the social influence reward eventually leads to significantly higher collective reward in both games. + +![](images/3b9225cc90c9a95e47d5cf85514dba567067d0da375c94be1d1955429ea1bdd4.jpg) +Figure 6: Total collective reward obtained in all experiments. Error bars show a $9 9 . 5 \%$ confidence interval (CI) over 5 random seeds, computed within a sliding window of 200 agent steps. The models trained with influence reward (red) significantly outperform the baseline and ablated models. + +While these aggregated results demonstrate the success of our models, they are not sufficient to understand the mechanism through which social influence is helping the agents achieve cooperative behavior. Therefore, we investigated the trajectories produced by high scoring models in both Cleanup and Harvest; the analysis revealed interesting behavior. As an example, in the Cleanup video available here: https: //youtu.be/iH_V5WKQxmo a single agent (shown in purple) was trained with the social influence reward. We see that unlike the other agents, which continue to randomly move and explore while waiting for apples to spawn, the influencer has a strange economy of motion; it only moves on the map when it is pursuing an apple, then stops. Interestingly, examining the trajectory reveals that the influencer uses only two moves to explore the map: turn left, which turns the agent in place without traversing the map, and move right, which moves the agent one square to the right on the map. + +![](images/15faf4dfa88bd78c4d0e9ddf505f7cecc97572565601a0318358c9465b4a4695.jpg) +Figure 7: A moment of high influence when the purple influencer signals the presence of an apple outside the yellow influencee’s field-of-view (yellow outlined box). + +Why did the influencer learn to use only these two moves? We can see that the influencer agent only chooses to move right (i.e. traverse the map) when it is pursuing an apple which is present. The rest of the time it simply turns left on the spot. At $t = 4 9$ , there is a moment of high influence between the influencer and the yellow influencee, which is shown in Figure 7. The influencer has chosen to move right towards an apple that is outside of the ego-centric field-of-view of the yellow agent. Because the purple agent only moves when apples are available, this signals to the yellow agent that an apple must be present above it which it cannot see. This changes the yellow agent’s distribution over its planned action, $p ( a _ { t } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { B } )$ and allows the purple agent to gain influence. A similar moment occurs when the influencer signals to an agent that has been cleaning the river that no apples have appeared by continuing to turn left (see Figure 12 in the Appendix). + +In this example, the influencer agent learned to use its own actions as a sort of binary code, which signals the presence or absence of apples in the environment. We also observe this effect in the influence agents in the Harvest task. This type of action-based communication could be likened to the bee waggle dance discovered by von Frisch (1969). Thus, rewarding agents for increasing the mutual information between their actions gave rise not only to cooperative behavior, but in this case, to emergent communication. These results further support the idea of using influence as a reward for training agents to communicate. + +# 4.2 INFLUENCE THROUGH COMMUNICATION + +Figures 6(b) and 6(e) show the results of training the agents to use an explicit communication channel, and its effect on their collective reward. In this case, the ablated baseline is a model that has the same structure as in Figure 2, but in which the communication policy $\pi _ { c }$ is trained only with environmental reward. We observe that the agents which are trained to use the communication channel with additional social influence reward achieve significantly higher collective reward in both games. In fact, in the case of Cleanup, we found that $\alpha = 0$ in the optimal hyperparameter settings, meaning that it was most effective to train the communication head with zero extrinsic or environmental reward (see Table 2 in the Appendix). This suggests that influence alone can be a sufficient mechanism for training an effective communication policy. + +To analyze the communication behaviour learned by the agents, we introduce three metrics. Speaker consistency, is a normalized score $\in [ 0 , 1 ]$ which assesses the entropy of $p ( a ^ { k } | m ^ { k } )$ and $p ( m ^ { k } | a ^ { k } )$ to determine how consistently a speaker agent emits a particular symbol when it takes a particular action, and vice versa (the formula is given in Appendix Section 6.3.4). We expect this measure to be high if, for example, the speaker always emits the same symbol when it is cleaning the river. We also introduce two measures of i(1) symbol/action $\mathbf { I C } = I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )$ dination (IC), which are both measures of mutual information (MI): measures the MI between the influencer/speaker’s symbol and the influencee/listener’s next action, and (2) action/action $\begin{array} { r } { \mathbf { I C } = I ( a _ { t } ^ { A } ; a _ { t + 1 } ^ { B } ) } \end{array}$ measures the MI between the influencer’s action and the influencee’s action in the next timestep. To compute these measures we first average over all trajectory steps, then take the maximum value between any two agents, to determine if any pair of agents are coordinating. Note that these measures are all instantaneous, as they consider only short-term dependencies across two consecutive timesteps, and cannot capture if an agent communicates influential compositional messages, i.e. information that requires several consecutive symbols to transmit and only then affects the other agents behavior. + +![](images/cfabcddf1dd58ea122c746b3e02eae517a3782b403a73cf07bb8bc425f0202f7.jpg) +Figure 8 presents the results. The speaker consistencies metric reveals that agents trained with the influence reward communicate less ambiguously about their own actions, indicating that the emergent communication is more meaningful. The instantaneous coordination metrics demonstrate +Figure 8: Metrics describing the quality of learned communication protocols. The models trained with influence reward exhibit more consistent communication and more coordination, especially in moments where influence is high. + +that the baseline agents trained without influence reward show almost no signs of co-ordinating behavior with communication, i.e. speakers saying A and listeners doing B consistently. This result is aligned with both theoretical results in cheap-talk literature (Crawford & Sobel, 1982), and recent empirical results in MARL (e.g. (Foerster et al., 2016; Lazaridou et al., 2018; Cao et al., 2018)). + +In contrast, we do see highly coordinated behavior between influence agents, but only when we limit the analysis to timesteps on which influence was high (cf. influential moments in Figure 8). If we inspect the results for agents trained with influence on the two tasks, a common pattern emerges: influence is sparse in time. An agent’s influence is only greater than its mean influence in less than $10 \%$ of timesteps. Because the listener agent is not compelled to listen to any given speaker, listeners selectively listen to a speaker only when it is beneficial, and influence cannot occur all the time. Only when the listener decides to change its action based on the speaker’s message does influence occur, and in these moments we observe high $I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )$ ; an effect that is lost when averaging over the entire trajectory. It appears the influencers have learned a strategy of communicating meaningful information about their own actions, and gaining influence when this becomes relevant enough for the listener to act upon it. + +Examining the relationship between the reward obtained by individual agents and the degree to which they were influenced by other agents gives a compelling result: agents that are the most influenced also achieve higher individual environmental reward, $\dot { E } ^ { k }$ . We sampled 100 different experimental conditions (i.e., hyper-parameters and random seeds) for both games, collected the influence and individual rewards, normalized them across the 5 agents in each condition, and correlated the resulting list of values. We found that agents who are more often influenced tend to achieve higher task reward in both Cleanup, $\rho = . 6 7$ , $p { < } 0 . 0 0 1$ , and Harvest, $\rho = . 3 4$ , $p { < } 0 . 0 0 1$ . This supports the hypothesis stated in Section 2.3: in order to gain influence from another agent by communicating with it, the communication message should contain information that helps the listener maximize its own environmental reward. Since better listeners/influencees are more successful in terms of task reward, we have evidence that useful information was transmitted to them. + +# 4.3 INFLUENCE VIA MODELING OTHER AGENTS + +Finally, we investigate whether the influence reward is still effective when computed without a centralised controller, but rather through each agent’s own internal Model of Other Agents (MOA) network. In this case, we extend the training period from $3 \cdot 1 0 ^ { 8 }$ steps to $5 \cdot 1 0 ^ { 8 }$ , in order to give the MOA model time to train. We also allow the policy LSTM to condition on the actions of other agents in the last timestep. We compare against an ablated version of this architecture (shown in Figure 3), which does not use the output of the MOA module to compute a reward; rather, the MOA module can be thought of as an unsupervised auxiliary task that may help the model to learn a better shared embedding layer, encouraging it to encode information relevant to predicting other agents’ behavior. + +Figures 6(c) and 6(f) show the collective reward obtained for agents trained with a MOA module. While we see that the auxiliary task does help to improve reward over the A3C baseline, the influence agent gets consistently higher collective reward. Impressively, for Cleanup, the MOA model scores higher than the original influence agents computed using the centralised controller (CC). As shown in Figure 6(c), the MOA baseline also achieves high collective reward, suggesting that the auxiliary task of modeling other agents helps the MOA agents cooperate more effectively in Cleanup. Further, the independent design of the MOA method allows each agent to influence every other agent, thus generating more reward signal and a greater chance to develop two-way cooperative behavior. + +Table 4 of the Appendix gives the final collective reward obtained by each model for all three experiments. Interestingly, several influence models are able to achieve higher collective reward than the previous state-of-the-art scores for these environments (275 for Cleanup and 750 for Harvest) (Hughes et al., 2018). This is compelling, given that previous work relied on the assumption that agents could view one another’s rewards; we make no such assumption, instead relying only on agents viewing each other’s actions. + +# 5 DISCUSSION AND CONCLUSIONS + +The experiments above have demonstrated that an intrinsic social reward based on having causal influence on the actions of other agents consistently improves cooperation and leads to higher collective return in the MA social dilemmas under investigation. In some cases, the influence reward drove agents to learn an emergent communication protocol via their actions. This is compelling, and confirms the connection between maximizing influence and maximizing the mutual information between agents’ actions. + +However, it is important to consider the limitations of the influence reward. Whether it will always give rise to cooperative behavior may depend on the specifics of the environment, task, and the trade-off between environmental and influence reward. Although influence is arguably necessary for cooperation (e.g. two agents cooperating to lift a box would have a high degree of influence between their actions), it may not be sufficient, in that it may be possible to influence another agent without helping it. For example, it is possible that agents could have gained influence in the tasks studied here by threatening to attack other agents with their fining beam. We believe this type of behavior did not emerge because communicating information represents the cheapest and most effective way to gain influence. Influencers do not have to sacrifice much in terms of their own environmental reward in order to communicate to other agents. + +Rewarding influence over an explicit communication channel may not be subject to this limitation, because influential communication may be inherently beneficial to the listener (at least in the case where listeners and speakers interact repeatedly). Since listeners can easily ignore communication messages if they do not help to obtain environmental reward, a speaker must transmit valuable information in order to gain influence through communication. There is no advantage to the speaker for communicating unreliably, because it would lose influence with the listener over time (although this is no longer guaranteed in one-shot interactions). Indeed, our results reveal that agents benefit from being influenced by (listening to) communication messages by obtaining higher individual reward, suggesting that the messages contain valuable information. Further, we found that the communication protocols learned via influence reward were more meaningful, and that the influence reward allowed agents to obtain higher collective return. Therefore, we suggest that influence could be a promising way to train emergent communication protocols in various settings. + +Finally, we have shown that influence can be computed by augmenting agents with an internal model that predicts the actions of other agents, and using this MOA model to simulate the effect of an agent’s actions on others. This represents an important step forward in multi-agent intrinsic social motivation, because it implies that the influence reward can be computed without having access to another agent’s reward function, or requiring a centralised controller. + +# 5.1 FUTURE WORK + +Using counterfactuals to allow agents to understand the effects of their actions on other agents could be a promising approach with a number of extensions. Perhaps agents could use counterfactuals to develop a form of ‘empathy‘, by simulating how their actions affect another agent’s value function. Or, social influence could be used to drive coordinated behavior in robots attempting to do cooperative manipulation and control tasks. Finally, if we view multi-agent networks as a single agent, influence could be used as a regularizer to encourage different modules of the network to integrate information from other networks; for example, perhaps it could prevent collapse in hierarchical RL. + +# REFERENCES + +Sean L Barton, Nicholas R Waytowich, Erin Zaroukian, and Derrik E Asher. Measuring collaborative emergent behavior in multi-agent reinforcement learning. arXiv preprint arXiv:1807.08663, 2018. + +Ben Bogin, Mor Geva, and Jonathan Berant. Emergence of communication in an interactive world with consistent speakers. arXiv preprint arXiv:1809.00549, 2018. + +Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent communication through negotiation. arXiv preprint arXiv:1804.03980, 2018. + +Philippe Capdepuy, Daniel Polani, and Chrystopher L Nehaniv. Maximization of potential information flow as a universal utility for collective behaviour. In Artificial Life, 2007. ALIFE’07. IEEE Symposium on, pp. 207–213. Ieee, 2007. + +Edward Choi, Angeliki Lazaridou, and Nando de Freitas. Compositional obverter communication learning from raw visual input. arXiv preprint arXiv:1804.02341, 2018. + +Jacob W. Crandall, Mayada Oudah, Tennom Chenlinangjia, Fatimah Ishowo-Oloko, Sherief Abdallah, Jean-Franc¸ois Bonnefon, Manuel Cebrian, Azim Shariff, Michael A. Goodrich, and Iyad Rahwan. ´ Cooperating with machines. CoRR, abs/1703.06207, 2017. URL http://arxiv.org/abs/ 1703.06207. + +Vincent P Crawford and Joel Sobel. Strategic information transmission. Econometrica: Journal of the Econometric Society, pp. 1431–1451, 1982. + +Sam Devlin, Logan Yliniemi, Daniel Kudenko, and Kagan Tumer. Potential-based difference rewards for multiagent reinforcement learning. In Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems, pp. 165–172. International Foundation for Autonomous Agents and Multiagent Systems, 2014. + +Heather J. Ferguson, Christoph Scheepers, and Anthony J. Sanford. Expectations in counterfactual and theory of mind reasoning. Language and Cognitive Processes, 25(3):297–346, 2010. doi: 10.1080/ 01690960903041174. URL https://doi.org/10.1080/01690960903041174. + +Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pp. 2137–2145, 2016. + +Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. arXiv preprint arXiv:1705.08926, 2017. + +Jakob Foerster, Richard Y Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch. Learning with opponent-learning awareness. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp. 122–130. International Foundation for Autonomous Agents and Multiagent Systems, 2018. + +Sebastien Forestier and Pierre-Yves Oudeyer. A unified model of speech and tool use early development.´ In 39th Annual Conference of the Cognitive Science Society (CogSci 2017), 2017. + +Felix A Gers, Jurgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm. ¨ 1999. + +Christian Guckelsberger, Christoph Salge, and Julian Togelius. New and surprising ways to be mean. adversarial npcs with coupled empowerment minimisation. arXiv preprint arXiv:1806.01387, 2018. + +Yuval Noah Harari. Sapiens: A brief history of humankind. Random House, 2014. + +Joseph Henrich. The Secret of Our Success: How culture is driving human evolution, domesticating our species, and making us smart. Princeton University Press, Princeton, NJ, 2015. URL http: //press.princeton.edu/titles/10543.html. + +Esther Herrmann, Josep Call, Mar´ıa Victoria Hernandez-Lloreda, Brian Hare, and Michael Tomasello. \` Humans have evolved specialized skills of social cognition: The cultural intelligence hypothesis. Science, 317(5843):1360–1366, 2007. ISSN 0036-8075. doi: 10.1126/science.1146282. URL http: //science.sciencemag.org/content/317/5843/1360. + +Edward Hughes, Joel Z Leibo, Matthew G Phillips, Karl Tuyls, Edgar A Due´nez-Guzm ˜ an, Antonio Garc ´ ´ıa Castaneda, Iain Dunning, Tina Zhu, Kevin R McKee, Raphael Koster, et al. Inequity aversion improves ˜ cooperation in intertemporal social dilemmas. In Advances in neural information processing systems (NIPS), Montreal, Canada, 2018. + +Alexander S Klyubin, Daniel Polani, and Chrystopher L Nehaniv. Empowerment: A universal agent-centric measure of control. In Evolutionary Computation, 2005. The 2005 IEEE Congress on, volume 1, pp. 128–135. IEEE, 2005. + +Kevin N. Laland. Darwin’s unfinished symphony $\therefore$ how culture made the human mind / Kevin N. Laland. Princeton University Press Princeton, 2017. ISBN 9781400884872 140088487. + +Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic communication from referential games with symbolic and pixel input. arXiv preprint arXiv:1804.03984, 2018. + +Joel Z Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. Multi-agent reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, pp. 464–473. International Foundation for Autonomous Agents and Multiagent Systems, 2017. + +Joseph T Lizier and Mikhail Prokopenko. Differentiating information transfer and causal effect. The European Physical Journal B, 73(4):605–615, 2010. + +Alicia P Melis and Dirk Semmann. How is human cooperation different? Philosophical Transactions of the Royal Society of London B: Biological Sciences, 365(1553):2663–2674, 2010. + +Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pp. 1928–1937, 2016. + +Pierre-Yves Oudeyer and Fred´ eric Kaplan. Discovering communication. ´ Connection Science, 18(2): 189–206, 2006. + +Pierre-Yves Oudeyer and Linda B Smith. How evolution may work through curiosity-driven developmental process. Topics in Cognitive Science, 8(2):492–502, 2016. + +Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning (ICML), volume 2017, 2017. + +Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018. + +Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: a primer. John Wiley & Sons, 2016. + +Julien Perolat, Joel Z Leibo, Vinicius Zambaldi, Charles Beattie, Karl Tuyls, and Thore Graepel. A multi-agent reinforcement learning model of common-pool resource appropriation. In Advances in Neural Information Processing Systems, pp. 3643–3652, 2017. + +Alexander Peysakhovich and Adam Lerer. Prosocial learning agents solve generalized stag hunts better than selfish ones. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp. 2043–2044. International Foundation for Autonomous Agents and Multiagent Systems, 2018. + +Neil C Rabinowitz, Frank Perbet, H Francis Song, Chiyuan Zhang, SM Eslami, and Matthew Botvinick. Machine theory of mind. arXiv preprint arXiv:1802.07740, 2018. + +Thomas C Schelling. Hockey helmets, concealed weapons, and daylight saving: A study of binary choices with externalities. Journal of Conflict resolution, 17(3):381–428, 1973. + +Jurgen Schmidhuber. Formal theory of creativity, fun, and intrinsic motivation (1990–2010). ¨ IEEE Transactions on Autonomous Mental Development, 2(3):230–247, 2010. + +Pedro Sequeira, Francisco S Melo, Rui Prada, and Ana Paiva. Emerging social awareness: Exploring intrinsic motivation in multiagent learning. In Development and Learning (ICDL), 2011 IEEE International Conference on, volume 2, pp. 1–6. IEEE, 2011. + +Satinder P. Singh, Andrew G. Barto, and Nuttapong Chentanez. Intrinsically motivated reinforcement learning. In Advances in Neural Information Processing Systems 17 [Neural Information Processing Systems, NIPS 2004, December 13-18, 2004, Vancouver, British Columbia, Canada], pp. 1281–1288, 2004. URL http://papers.nips.cc/paper/ 2552-intrinsically-motivated-reinforcement-learning. + +Katherine KM Stavropoulos and Leslie J Carver. Research review: social motivation and oxytocin in autism–implications for joint attention development and intervention. Journal of Child Psychology and Psychiatry, 54(6):603–618, 2013. + +DJ Strouse, Max Kleiman-Weiner, Josh Tenenbaum, Matt Botvinick, and David Schwab. Learning to share and hide intentions using information regularization. arXiv preprint arXiv:1808.02093, 2018. + +Satohiro Tajima, Toru Yanagawa, Naotaka Fujii, and Taro Toyoizumi. Untangling brain-wide dynamics in consciousness by cross-embedding. PLoS computational biology, 11(11):e1004537, 2015. + +Michael Tomasello. Why we cooperate. MIT press, 2009. + +Carel P van Schaik and Judith M Burkart. Social learning and evolution: the cultural intelligence hypothesis. Philosophical Transactions of the Royal Society B: Biological Sciences, 366(1567):1008–1016, 2011. + +Karl von Frisch. The dance language and orientation of bees. 5, 06 1969. + +Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256, 1992. + +Chao Yu, Minjie Zhang, and Fenghui Ren. Emotional multiagent reinforcement learning in social dilemmas. In International Conference on Principles and Practice of Multi-Agent Systems, pp. 372–387. Springer, 2013. + +# 6 APPENDIX + +# 6.1 SEQUENTIAL SOCIAL DILEMMAS + +In each of the sequential social dilemma (SSD) games studied above, an agent is rewarded $+ 1$ for every apple it collects, but the apples are a limited resource. In Harvest (a tragedy of the commons game), apples regenerate more slowly the faster they are harvested, and if an exploiting agent consumes all of the apples, they will not grow back; agents must cooperate to harvest sustainably. In Cleanup (a public goods game), apples are generated based on the amount of waste in a nearby river. Agents can use a cleaning beam action to clean the river when they are positioned in it; or they can simply consume the apples the other agent produces. Agents also have a fining beam action which they can use to fine nearby agents $- 5 0$ reward. + +Figure 9 gives the Schelling diagram for both SSD tasks under investigation. A Schelling diagram (Schelling, 1973; Perolat et al., 2017) shows the relative payoffs for a single agent’s strategy given a fixed number of other agents who are cooperative. Schelling diagrams generalize payoff matrices to multi-agent settings, and make it easy to visually recognize game-theoretic properties like Nash equilibria (see Schelling (1973) for more details). + +![](images/cd8620142473269dee9b7be7e5b1fb98864ebd4ee20a85b56ccae63f16b46cc2.jpg) +Figure 9: Schelling diagrams for the two social dilemma tasks show that an individual is almost always motivated to defect, even though the group will get higher reward if there are more cooperators. + +# 6.2 ADDITIONAL EXPERIMENT - BOX TRAPPED + +As a proof-of-concept experiment to test whether the influence reward works as expected, we constructed a special environment, shown in Figure 10. In this environment, one agent (teal) is trapped in a box. The other agent (purple) has a special action it can use to open the box... or it can simply choose to consume apples, which exist outside the box and are inexhaustible in this environment. + +As expected, a vanilla A3C agent learns to act selfishly; the purple agent will simply consume apples, and chooses the open box action in $0 \%$ of trajectories once the policy has converged. A video of A3C agents trained in this environment is available at: https://youtu.be/C8SE9_ $\Upsilon \mathrm { K z x I }$ , which shows that the purple agent leaves its compatriot trapped in the box throughout the trajectory. + +![](images/3f18791740d5f30d7f41c7fce3b78ff4b93cfac39e1f81dcfeb16188b74db18f.jpg) +Figure 10: The Box trapped environment in which the teal agent is trapped, and the purple agent can release it with a special open box action. + +In contrast, an agent trained with the social influ +ence reward chooses the open box action in $88 \%$ of trajectories, releasing its fellow agent so that they are both able to consume apples. A video of this behavior is shown at: https://youtu.be/Gfo248-qt3c. Further, as Figure 11(a) reveals, the purple influencer agent usually chooses to open the box within the first few steps of the trajetory, giving its fellow agent more time to collect reward. + +Most importantly though, Figure 11(b) shows the influence reward over the course of a trajectory in the Box trapped environment. The agent chooses the open box action in the second timestep; at this point, we see a corresponding spike in the influence reward. This reveals that the influence reward works as expected, incentivizing an action which has a strong — and in this case, prosocial — effect on the other agent’s behavior. + +![](images/87f2f4c8939684aaa1a065bb1bd2a40dcabf6619e1b311cd76ab190958323091.jpg) + +(a) Number of times the open box action occurs at each trajectory step over 100 trajectories. + +![](images/aa590ebca444f816660ca98874365417e4da0912c85a637d878b313133db889b.jpg) +(b) Influence reward over a trajectory in Box trapped +Figure 11: The Box trapped proof-of-concept experiment reveals that an agent gets high influence for letting another agent out of a box in which it is trapped. + +# 6.3 IMPLEMENTATION DETAILS + +All models are trained with a single convolutional layer with a kernel of size 3, stride of size 1, and 6 output channels. This is connected to two fully connected layers of size 32 each, and an LSTM with 128 cells. We use a discount factor $\gamma = . 9 9$ . The number of agents $N$ is fixed to 5. + +As mentioned in Section 2.2, the social influence reward can be computed using a number of divergence measures, including JSD. We also experiment with training the agents using the pointwise mutual information (the innermost term of Eq. 3), which is given by: + +$$ +p m i ( a ^ { A } ; a ^ { B } | Z { = } z ) { = } \log \frac { p ( a ^ { B } | a ^ { A } , z ) } { p ( a ^ { B } | z ) } { = } \mathrm { l o g } \frac { p ( a ^ { A } , a ^ { B } | z ) } { p ( a ^ { A } | z ) p ( a ^ { B } | z ) } . +$$ + +This PMI term is precisely the local information flow proposed by Lizier & Prokopenko (2010) as a measure of direct causal effect; the expectation of the PMI over $p \dot { ( } a ^ { B } , a ^ { A } | z )$ is the MI. and gives us a measure of influence of a single action of $A$ on the single action taken by $B$ . + +In addition to the comparison function used to compute influence, there are many other hyperparameters that can be tuned for each model. We use a random search over hyperparameters, ensuring a fair comparison with the search size over the baseline parameters that are shared with the influence models. For all models we search for the optimal entropy reward and learning rate, where we anneal the learning rate from an initial value lr init to lr final. The below sections give the parameters found to be most effective for each of the three experiments. + +# 6.3.1 CENTRALISED CONTROLLER HYPERPARAMETERS + +In this setting we vary the number of influencers from 1−4, the influence reward weight $\beta$ , and the number of curriculum steps over which the weight of the influence reward is linearly increased $C$ . In this setting, since we have a centralised controller, we also experiment with giving the influence reward to the agent being influenced as well, and find that this sometimes helps. This ‘influencee’ reward is not used in the other two experiments, since it precludes independent training. The hyperparameters found to give the best performance for each model are shown in Table 1. + +# 6.3.2 COMMUNICATION HYPERPARAMETERS + +Because the communication models have an extra A2C output head for the communication policy, we use an additional entropy regularization term just for this head, and apply a weight to the communication loss in the loss function. We also vary the number of communication symbols that the agents can emit, and the size of the linear layer that connects the LSTM to the communication policy layer, which we term the communication embedding size. Finally, in the communication regime, we experiment to setting the weight on the extrinsic reward E, $\alpha$ , to zero. The best hyperparameters for each of the communication models are shown in Table 2. + +Table 1: Optimal hyperparameter settings for the models in the centralised controller experiment. + +
HyperparameterCleanupHarvest
A3CVisibleactionsInfluenceA3C baselineVisibleactionsInfluence
Entropy reg.baseline .00176baseline .00176.000248.000687baseline .00184.00025
lr_init.00126.00126.00107.00136.00215.00107
lr_end.000012.000012.000042.000028.000013.000042
Number of influencers3133
Influence weight β0.1460.224
Curriculum C11401140
Policy comparisonJSDPMI
Influencee reward10
+ +Table 2: Optimal hyperparameter settings for the models in the communication experiment. + +
HyperparameterCleanupHarvest
A3C baselineComm. baselineInfluence comm.A3C Comm. baseline baselineInfluence comm.
Entropy reg..00176.000249.00305.000687.000174.00220
lr_init.00126.00223.00249.00136.00137.000413
lr_end.000012.000022.0000127.000028.0000127.000049
Influence weight β=02.752=04.825
Extrinsic reward01.0
weight α
Curriculum C1 KL8
Policy comparison.000789KL
Comm. entropy reg..0758.00208
Comm. loss weight9.0709
Symbol vocab size327
Comm. embedding16
+ +Table 3: Optimal hyperparameter settings for the models in the model of other agents (MOA) experiment. + +
HyperparameterCleanupHarvest
A3C baselineMOA baselineInfluence MOAA3C baselineMOA baselineInfluence MOA
Entropy reg..00176.00176.00176.000687.00495.00223
lr_init.00126.00123.00123.00136.00206.00120
lr_end.000012.000012.000012.000028.000022.000044
Influence weight β0.620=02.521
MOA loss weight1.31215.007=1.71110.911
Curriculum C-40=1226
Policy comparison=KL==KL
Train MOA onlyFalseTrueFalseTrue
when visible
+ +# 6.3.3 MODEL OF OTHER AGENTS (MOA) HYPERPARAMETERS + +The MOA hyperparameters include whether to only train the MOA with cross-entropy loss on the actions of agents that are visible, and how much to weight the supervised loss in the overall loss of the model. The best hyperparameters are shown in Table 3. + +
CleanupHarvest
A3Cbaseline89485
Inequity aversion (Hughes et al., 2018)275750
Influence - Basic1901073
Influence - Communication166951
Influence -Model of other agents392588
+ +Table 4: Final collective reward over the last 50 agent steps for each of the models considered. Bolded entries represent experiments in which the influence models significantly outperformed the scores reported in previous work on inequity aversion(Hughes et al., 2018). This is impressive, considering the inequity averse agents are able to view all other agents’ rewards. We make no such assumption, and yet are able to achieve similar or superior performance. + +# 6.3.4 COMMUNICATION ANALYSIS + +The speaker consistency metric is calculated as: + +$$ +\sum _ { k = 1 } ^ { N } 0 . 5 [ \sum _ { c } 1 - \frac { H ( p ( a ^ { k } | m ^ { k } = c ) ) } { H _ { m a x } } + \sum _ { a } 1 - \frac { H ( p ( m ^ { k } | a ^ { k } = a ) ) } { H _ { m a x } } ] , +$$ + +where $H$ is the entropy function and $H _ { m a x }$ is the maximum entropy based on the number of discrete symbols or actions. The goal of the metric is to measure how much of a 1:1 correspondence exists between a speaker’s action and the speaker’s communication message. + +# 6.4 ADDITIONAL RESULTS + +Figure 12 shows an additional moment of high influence in the Cleanup game. The purple influencer agent can see the area within the white box, and therefore all of the apple patch. The field-of-view of the magenta influencee is outlined with the magenta box; it cannot see if apples have appeared, even though it has been cleaning the river, which is the action required to cause apples to appear. When the purple influencer turns left and does not move towards the apple patch, this signals to the magenta agent that no apples have appeared, since otherwise the influence would move right. + +![](images/fda9989ec537ce17189a8d0151565ce65db1893467693e2932bcd54bd93d57e3.jpg) +Figure 12: A moment of high influence between the purple influencer and magenta influencee. + +Table 4 presents the final collective reward obtained by each of the models tested in the experiments presented in Section 4. We see that in several cases, the influence agents are even able to out-perform the state-of-the-art results on these tasks reported by Hughes et al. (2018), despite the fact that the solution proposed by Hughes et al. (2018) requires that agents can view other agents’ rewards, whereas we do not make this assumption, and instead only require that agents can view each others’ actions. + +It is important to note that collective reward is not always the perfect metric of cooperative behavior, a finding that was also discovered by Barton et al. (2018) and emphasized by Leibo et al. (2017). In the case, we find that there is a spurious solution to the Harvest game, in which one agent fails to learn and fails to collect any apples. This leads to very high collective reward, since it means there is one fewer agent that can exploit the others, and makes sustainable harvesting easier to achieve. Therefore, for the results shown in the paper, we eliminate any random seed in Harvest for which one of the agents has failed to learn to collect apples, as in previous work (Hughes et al., 2018). + +However, here we also present an alternative strategy for assessing the overall collective outcomes: weighting the total collective reward by an index of equality of the individual returns. Specifically, we compute the Gini coefficient over the $N$ agents’ individual returns: + +$$ +G = \frac { \sum _ { i = 1 } ^ { N } \sum _ { j = 1 } ^ { N } \left| { r ^ { i } - r ^ { j } } \right| } { 2 N \sum _ { i = 1 } ^ { N } r ^ { i } } , +$$ + +which gives us a measure of the inequality of the returns, where $G \in [ 0 , 1 ]$ , with $G = 0$ indicating perfect equality. Thus, $1 - G$ is a measure of equality; we use this to weight the collective reward for each experiment, and plot the results in Figure 13. Once again, we see that the influence models give the highest final performance, even with this new metric. + +![](images/c923ae01cca519ebc26562fb33c66fec40b0011ec5f33643088a3be15814576d.jpg) +Figure 13: Total collective reward times equality, $R * ( 1 - G )$ , obtained in all experiments. Error bars show a $9 9 . 5 \%$ confidence interval (CI) over 5 random seeds, computed within a sliding window of 200 agent steps. Once again, the models trained with influence reward (red) significantly outperform the baseline and ablated models. + +Finally, we would like to show that the influence reward is robust to the choice of hyperparameter settings. Therefore, in Figure 14, we plot the collective reward of the top 5 best hyperparameter settings for each experiment, over 5 random seeds each. Once again, the influence models result in higher collective reward, which provides evidence that the model is robust to the choice of hyperparameters. + +![](images/1408fa6f4ede73df1fdda59c904bd76d77d6315c5152307ac31700974d00769d.jpg) +Figure 14: Total collective reward over the top 5 hyperparameter settings, with 5 random seeds each, for all experiments. Error bars show a $9 9 . 5 \%$ confidence interval (CI) computed within a sliding window of 200 agent steps. The influence models still maintain an advantage over the baselines and ablated models, suggesting the technique is robust to the hyperparameter settings. + +# 6.4.1 OPTIMIZING FOR COLLECTIVE REWARD + +In this section we include the results of training explicitly prosocial agents, which directly optimize for the collective reward of all agents. Previous work (e.g. Peysakhovich & Lerer (2018)) has shown that training agents to optimize for the rewards of other agents can help the group to obtain better collective outcomes. Following a similar prown individual reward $E ^ { k }$ ple, we implemented agents that optimize foand the collective reward of all other agents, $\textstyle \sum _ { i = 1 , i \neq k } ^ { N } E ^ { i }$ mbination of their. Thus, the reward function for agent $k$ is $\begin{array} { r } { { R ^ { k } } = E ^ { k } + \eta \sum _ { i = 1 , i \neq k } ^ { N } E ^ { i } } \end{array}$ . We conducted the same hyperparameter search over the parameters mentioned in Section 6.3.1 varying the weight placed on the collective reward, $\eta \in [ 0 , 2 ]$ . + +As expected, we find that agents trained to optimize for collective reward attain higher collective reward in both Cleanup and Harvest, as is shown in Figure 15. In both games, the optimal value for $\eta = 0 . 8 5$ . + +![](images/fe3d4f89e8a19d1b5fe2a4f19403d717b5760772f2debd3fb1d1b879cb6bc09c.jpg) +Figure 15: Total collective reward obtained by agents trained to optimize for the collective reward, for the 5 best hyperparameter settings with 5 random seeds each. Error bars show a $9 9 . 5 \%$ confidence interval (CI) computed within a sliding window of 200 agent steps. + +Interestingly, however, the equality in the individual returns for these agents is extremely low. Across the hyperparameter sweep, no solution to the Cleanup game which scored more than 20 points in terms of collective return was found in which all agents scored an individual return above 0. It seems that in Cleanup, when agents are trained to optimize for collective return, they converge on a solution in which some agents never receive any reward. + +Note that training agents to optimize for collective reward requires that each agent can view the rewards obtained by other agents. As discussed previously, the social influence reward is a novel way to obtain cooperative behavior, that does not require making this assumption. \ No newline at end of file diff --git a/parse/train/B1lG42C9Km/B1lG42C9Km_content_list.json b/parse/train/B1lG42C9Km/B1lG42C9Km_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a1dfb7993eb4f2cedb3e34a7ce6f9f75001ffe2 --- /dev/null +++ b/parse/train/B1lG42C9Km/B1lG42C9Km_content_list.json @@ -0,0 +1,2147 @@ +[ + { + "type": "text", + "text": "INTRINSIC SOCIAL MOTIVATION VIA CAUSAL INFLUENCE IN MULTI-AGENT RL ", + "text_level": 1, + "bbox": [ + 174, + 98, + 697, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 184, + 171, + 387, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 455, + 234, + 542, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We derive a new intrinsic social motivation for multi-agent reinforcement learning (MARL), in which agents are rewarded for having causal influence over another agent’s actions. Causal influence is assessed using counterfactual reasoning. The reward does not depend on observing another agent’s reward function, and is thus a more realistic approach to MARL than taken in previous work. We show that the causal influence reward is related to maximizing the mutual information between agents’ actions. We test the approach in challenging social dilemma environments, where it consistently leads to enhanced cooperation between agents and higher collective reward. Moreover, we find that rewarding influence can lead agents to develop emergent communication protocols. We therefore employ influence to train agents to use an explicit communication channel, and find that it leads to more effective communication and higher collective reward. Finally, we show that influence can be computed by equipping each agent with an internal model that predicts the actions of other agents. This allows the social influence reward to be computed without the use of a centralised controller, and as such represents a significantly more general and scalable inductive bias for MARL with independent agents. ", + "bbox": [ + 233, + 267, + 766, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 502, + 330, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep reinforcement learning (RL) has made impressive progress on specific tasks with well-defined reward functions, but is still difficult to learn intelligent behavior that generalizes across multiple domains. Intrinsic motivation is a technique for solving this problem by developing general reward functions that encourage an agent to learn across a variety of tasks (Singh et al., 2004). Previous approaches to intrinsic motivation have broadly fallen into two categories: (1) curiosity, or a drive for novelty (e.g. Pathak et al. (2017); Schmidhuber (2010)), and (2) empowerment, or a drive to be able to manipulate the environment (Klyubin et al., 2005). ", + "bbox": [ + 174, + 526, + 825, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We posit that this body of work has largely overlooked an important intrinsic motivation that is key to human learning: social interaction. Humans have remarkable social learning abilities; some authors suggest that it is social learning that has given rise to cultural evolution, and allowed us to achieve unprecedented progress and coordination on a massive scale (van Schaik & Burkart, 2011; Herrmann et al., 2007). Others emphasize that our impressive capacity to learn from others far surpasses that of other animals, apes, and even other proto-human species (Henrich, 2015; Harari, 2014; Laland, 2017). ", + "bbox": [ + 174, + 617, + 825, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, we propose an intrinsic reward function designed for multi-agent RL (MARL), which awards agents for having a causal influence on other agents’ actions. Causal influence is assessed using counterfactual reasoning; at each timestep, an agent simulates alternate, counterfactual actions that it could have taken, and assesses their effect on another agent’s behavior. Actions that lead to relatively higher change in the other agent are considered to be highly influential and are rewarded. We show how this reward is related to maximizing the mutual information between agents’ actions, and is thus a form of social empowerment. We hypothesize that rewarding influence may therefore encourage cooperation between agents. We also take inspiration from experiments in human cognition, showing that newborn infants are sensitive to correspondences between their own actions and the actions of other people, and use this to coordinate their behavior with others (Tomasello, 2009; Melis & Semmann, 2010). ", + "bbox": [ + 174, + 708, + 825, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To study our proposed social influence reward in the MARL setting, we adopt the Sequential Social Dilemmas (SSDs) of Leibo et al. (2017). These are challenging MA environments with a game-theoretic reward structure, similar to Prisoner’s Dilemma. For each individual agent, ‘defecting’ (non-cooperative behavior) has the highest payoff. However, the collective reward will be better if all agents choose to cooperate. The paradoxical payoff structure of these tasks make achieving cooperative social dynamics extremely challenging for typical RL agents. We show that social influence allows agents to learn to cooperate in these environments, and make the following contributions: ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We demonstrate that deep RL agents trained with the proposed social influence reward cooperate to attain higher collective reward than baseline deep RL agents (Mnih et al., 2016). In some cases, this cooperation is attained because influencer agents learn to use their actions as an emergent communication protocol, analogous to behavior seen in animals (von Frisch, 1969). Motivated by the previous point, we apply the influence reward to training deep RL agents to use an explicit communication channel, as in Foerster et al. (2016). We demonstrate that the communication protocols trained with the influence reward meaningfully relate to agents’ actions, and that once again, agents trained with the influence reward achieve better collective outcomes. We demonstrate that there is a significant correlation between being influenced through communication messages and obtaining higher individual return, suggesting that influential communication is beneficial to the agents that receive it. Finally, rather than computing social influence using a centralised training framework as in prior work (e.g. Foerster et al. (2017; 2016)), we extend the approach by attaching an internal Model of Other Agents (MOA) network to each agent and training it to predict the actions of every other agent. The agent can then simulate counterfactual actions and use its own internal MOA to predict how these will affect other agents, thus computing its own intrinsic influence reward. ", + "bbox": [ + 215, + 143, + 826, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using a MOA to predict and reward influence allows us to compute an intrinsic social reward by observing other agents’ past actions, without a centralised controller, and without requiring access to another agent’s reward function. We believe this is an important innovation over prior work (e.g. (Hughes et al., 2018; Foerster et al., 2017; 2016)). When we consider likely future applications of MARL, such as autonomous driving, it becomes apparent that centralised training or the sharing of reward functions are unrealistic assumptions, since autonomous vehicles are likely to be produced by a wide variety of organizations and institutions with mixed motivations. Rather, a social reward function which only depends on observing the behavior of agents acting in the environment, and which can give rise to coordinated, cooperative behavior, represents a more promising approach. ", + "bbox": [ + 174, + 397, + 825, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 METHODS ", + "text_level": 1, + "bbox": [ + 174, + 541, + 287, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider a MARL Markov game defined by the tuple $\\langle S , T , A , r \\rangle$ , in which multiple agents which do not share weights are trained to independently maximize their own individual reward. The environment state is given by $s \\in S$ . At each timestep $t$ , each agent $k$ chooses an action $a _ { t } ^ { k } \\in A$ . The actions of all $N$ agents are combined to form a joint action $\\pmb { a } _ { t } = [ a _ { t } ^ { 0 } , . . . a _ { t } ^ { N } ]$ , which produces a transition in the environment $T _ { \\cdot } ( s _ { t + 1 } | \\mathbf { a } _ { t } , s _ { t } )$ , according to the state transition function $T$ . Each agent then receives its own reward $r ^ { k } ( a _ { t } , s _ { t } )$ , which may depend on the actions of other agents. A history of these variables over time is termed a trajectory, ${ \\boldsymbol { \\tau } } = \\left\\{ { { s _ { t } } , { a _ { t } } , { r _ { t } } } \\right\\} _ { t = 0 } ^ { T }$ . We consider a partially observable setting in which each agent $k$ can only vieture reward, true sta, where , $s _ { t } ^ { k }$ . Each agent seeks to maximize its own total discount factor. A distributed asynchronous a ectedntage $\\begin{array} { r } { R ^ { k } = \\sum _ { i = 0 } ^ { \\infty } \\gamma ^ { i } r _ { t + i } ^ { k } } \\end{array}$ $\\gamma$ \n$\\pi ^ { k }$ policy is learned via REINFORCE with baseline (Williams, 1992). Architecturally, our agents consist of a convolutional layer, fully connected layers, a Long Short Term Memory (LSTM) network (Gers et al., 1999), and linear layers which output $\\pi ^ { \\dot { k } }$ and the value function $V ^ { \\pi _ { k } } ( s )$ . We will refer to the internal LSTM state of agent $k$ at timestep $t$ as $\\dot { u } _ { t } ^ { k }$ . ", + "bbox": [ + 173, + 569, + 825, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 INTRINSIC SOCIAL MOTIVATION VIA CAUSAL INFLUENCE ", + "text_level": 1, + "bbox": [ + 174, + 787, + 589, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Social influence intrinsic motivation modifies an agent’s reward function so that it becomes $R ^ { k } = \\alpha E ^ { k } + \\beta I ^ { k }$ , where $E ^ { k }$ is the extrinsic or environmental reward, and $I ^ { k }$ is the causal influence reward. We compute $I ^ { k }$ by generating counterfactual actions that the agent could have taken at each timestep, and assessing how taking these would have affected other agents’ behavior. A counterfactual is the estimated probability that $^ { 6 6 } Y$ would be $y$ had $X$ been $x$ , in situation $Z = z ^ { \\prime }$ , where $X , Y ,$ , and $Z$ are random variables, and $x , y$ and $z$ are their values (Pearl et al., 2016). Importantly, it is a counterfactual because we condition on a set of evidence $z$ , and because the assignment $X = x$ is counter to what we actually observed; in reality, $X$ took on some other value. ", + "bbox": [ + 174, + 809, + 825, + 921 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d4c9d865d0914f582cec2f90cb9102528669a3d0feda5b6e015295234f191f06.jpg", + "image_caption": [ + "Figure 1: Causal diagram of agent $A$ ’s effect on $B$ ’s action. We condition on each agent’s view of the environment and LSTM state $u$ (shaded nodes), and intervene on $a _ { t } ^ { \\dot { A } }$ (blue). " + ], + "image_footnote": [], + "bbox": [ + 176, + 103, + 303, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To see how we can compute the causal effect of one agent on another, suppose there are two agents, $A$ and $B$ , and that agent $B$ receives $A$ ’s action at time $t$ taction, know a , as input1. Agent $p \\big ( a _ { t } ^ { B } | a _ { t } ^ { A } , \\overline { { s _ { t } ^ { B } } } , u _ { t } ^ { B } \\big )$ $B$ then uses this to compute a distribution over its own ecause we have built the model , and its own internal LSTM state, agent , as sh $B$ , wen in $\\bar { a } _ { t } ^ { A } , s _ { t } ^ { B }$ $u _ { t } ^ { B }$ Figure 1. This allows us to exactly isolate the causal effect of $A$ ’s action on $B$ by conditioning on the values we observed for the other inputs at this timestep (note that by conditioning on these variables, including the LSTM state $u _ { t } ^ { B }$ , we remove any dependency on previous timesteps in the trajectory). We can then intervene on $a _ { t } ^ { A }$ by replacing it with a counterfactual action, $d o \\dot { ( a _ { t } ^ { A } ) }$ . This counterfactual action is used to compute a new estimate of $p ( a _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , s _ { t } ^ { B } , u _ { t } ^ { B } )$ . Essentially, the agent asks a retrospective question: “How would $B$ ’s action change if I had acted differently in this situation?”. ", + "bbox": [ + 318, + 103, + 825, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To simplify notation, let $\\boldsymbol { z } _ { t } = \\langle \\boldsymbol { u } _ { t } ^ { B } , \\boldsymbol { s } _ { t } ^ { B } \\rangle$ , so that conditioning on $z _ { t }$ is equivalent to conditioning on all relevant background variables (all shaded variables in Figure 1). We can also forego the $d o$ operator, noting that $p ( \\dot { a } _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , z _ { t } ) \\equiv p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { \\tilde { A } } , z _ { t } )$ in this case, because $z$ satisfies the back-door criterion (Pearl & Mackenzie, 2018). Now, consider averaging over several counterfactuals $\\tilde { a } _ { t } ^ { A }$ . This gives us the marginal policy of $B$ , $\\begin{array} { r } { p ( a _ { t } ^ { B } | \\tilde { z _ { t } } ) = \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , z _ { t } ) p ( \\tilde { a } _ { t } ^ { A } | z _ { t } ) } \\end{array}$ —in other words, $B$ ’s policy if $A$ were not considered. The discrepancy between the marginal policy of $B$ and the conditional policy of $B$ given $A$ ’s action is a measure of the causal influence of ", + "bbox": [ + 318, + 290, + 825, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$A$ on $B$ ; it gives the degree to which $B$ changes its planned action distribution because of $A$ ’s behavior. Thus, the causal influence intrinsic reward for agent $A$ is ", + "bbox": [ + 171, + 405, + 823, + 433 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/4c598b57c8ef06c70864098ccc7424fb311f9df5bf10e8fc304f1cd45bcbf3e0.jpg", + "text": "$$\n{ I _ { t } ^ { A } } = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } \\mid z _ { t } , \\tilde { a } _ { t } ^ { A } ) p ( \\tilde { a } _ { t } ^ { A } \\mid z _ { t } ) \\right] = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| p ( a _ { t } ^ { B } \\mid z _ { t } ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 218, + 438, + 779, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 RELATIONSHIP TO MUTUAL INFORMATION AND EMPOWERMENT ", + "text_level": 1, + "bbox": [ + 173, + 491, + 643, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The causal influence reward in Eq. 1 is related to the mutual information (MI) between the actions of agents $A$ and $B$ , which is given by ", + "bbox": [ + 176, + 512, + 825, + 541 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/fc2dfdef03318cc35a94ea138af5315729d92b04c7b2a99e379f024a438c3e59.jpg", + "text": "$$\nI ( A ^ { B } ; A ^ { A } | z ) = \\sum _ { a ^ { A } , a ^ { B } } p ( a ^ { B } , a ^ { A } | z ) \\mathrm { l o g } \\frac { p ( a ^ { B } , a ^ { A } | z ) } { p ( a ^ { B } | z ) p ( a ^ { A } | z ) } { = \\sum _ { a ^ { A } } p ( a ^ { A } | z ) D _ { \\mathrm { K L } } \\left[ p ( a ^ { B } | a ^ { A } , z ) \\right] \\left| p ( a ^ { B } | z ) \\right] } ,\n$$", + "text_format": "latex", + "bbox": [ + 187, + 546, + 790, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we see that the $D _ { K L }$ factor in Eq. 2 is the causal influence reward given in Eq. 1. The connection to mutual information is interesting, because a frequently used intrinsic motivation for single agent RL is empowerment, which rewards the agent for having high mutual information between its actions and the future state of the environment (e.g. Klyubin et al. (2005); Capdepuy et al. (2007)). To the extent that the social influence reward defined in Eq. 1 is an approximation of the MI, $A$ is rewarded for having empowerment over $B ^ { \\prime } s$ actions. ", + "bbox": [ + 174, + 593, + 825, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By sampling $N$ independent trajectories $\\tau _ { n }$ from the environment, where A’s actions $a _ { n } ^ { A }$ are drawn according to $\\overset { \\cdot } { p } ( a ^ { A } | z )$ , we perform a Monte-Carlo approximation of the MI (see e.g. Strouse et al. (2018)), ", + "bbox": [ + 171, + 684, + 825, + 713 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/fc5ac4b9ac6bc37a9d7073bded53bdb90fc53b4246ba47725aac71ca7c4aa90e.jpg", + "text": "$$\nI ( A ^ { A } ; A ^ { B } | z ) = \\mathbb { E } _ { \\tau } \\Big [ D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | A ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] \\Big | z \\Big ] \\approx \\frac { 1 } { N } \\sum _ { n } D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | a _ { n } ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] .\n$$", + "text_format": "latex", + "bbox": [ + 210, + 719, + 785, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, in expectation, the social influence reward is the MI between agents’ actions. ", + "bbox": [ + 176, + 761, + 687, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Whether the policy trained with Eq. 1 actually learns to approximate the MI depends on the learning dynamics. We calculate the intrinsic social influence reward using Eq. 1, because unlike Eq. 2, which gives an estimate of the symmetric bandwidth between $A$ and $B$ , Eq. 1 gives the directed causal effect of the specific action taken by agent $A$ , $a _ { t } ^ { A }$ . We believe this will result in an easier reward to learn, since it allows for better credit assignment; agent $A$ can more easily learn which of its actions lead to high influence. We also experiment with replacing the KL-divergence with several other measures, including the Jensen-Shannon Divergence (JSD), and find that the influence reward is robust to the choice of measure. ", + "bbox": [ + 174, + 781, + 825, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 INFLUENCE THROUGH COMMUNICATION ", + "text_level": 1, + "bbox": [ + 176, + 174, + 480, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Melis & Semmann (2010), human children rapidly learn to use communication to influence the behavior of others when engaging in cooperative activities. They explain that “this ability to influence the partner via communication has been interpreted as evidence for a capacity to form shared goals with others”, and that this capacity may be “what allows humans to engage in a wide range of cooperative activities”. Therefore, we investigate a second use of the social influence reward: learning inter-agent communication protocols. Using a similar approach to Reinforced Inter-Agent Learning (RIAL) (Foerster et al., 2016), we equip the agents with an explicit communication channel. At each timestep, each agent $k$ chooses a discrete communication symbol $\\Dot { m } _ { t } ^ { k }$ ; these symbols are concatenated into a combined message vector $\\pmb { m } _ { t } = [ m _ { t } ^ { 0 } , m _ { t } ^ { 1 } . . . m _ { t } ^ { N } ] ,$ , for $N$ agents. This message vector $\\mathbf { \\nabla } m _ { t }$ is then shown to every other agent in the next timestep, as in Figure 2. To train the agents to communicate, we augment our initial network with an additional A3C output head, that learns a communication policy $\\pi _ { c }$ over which symbol to emit, and a communication value function $V _ { c }$ (this is separate from the normal policy and value function used for acting in the environment, $\\pi _ { e }$ and $V _ { e }$ , which are trained only with environmental reward $E$ ). ", + "bbox": [ + 173, + 193, + 825, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The influence reward is used, in addition to environmental reward, to train the communication policy $\\pi _ { c }$ Counterfactuals are employed to assess how muon another agent’s action in the next timestep, luence an agent’s communication message, . Importantly, we hypothesize that comm $m _ { t } ^ { A }$ , hastion $a _ { t + 1 } ^ { B }$ can only be influential if it is useful to another agent. There is nothing that compels agent to act based on agent $A$ ’s communication message; if it does not contain valuable information, $B$ is free to ignore it. In fact, previous work has shown that selfish agents do not learn to use this type of ungrounded, cheap talk communication channel effectively (Cao et al., 2018). In contrast, for $A$ to gain influence via communication, $m _ { t } ^ { A }$ must contain valuable information that informs $B$ about how best to maximize its own reward, so much so that it actually causes $B$ to change its intended action. ", + "bbox": [ + 173, + 380, + 825, + 506 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/caa2b9dc3c9fe0bdc22237d63bfba2b09073071b89e2a88a719924f5aa8e5a38.jpg", + "image_caption": [ + "Figure 2: The communication model has two A2C heads, which learn a normal policy, $\\pi _ { e }$ , and a policy for emitting communication symbols, $\\pi _ { c }$ . Other agents’ communication messages $\\mathbf { \\nabla } m _ { t }$ are input to the LSTM. " + ], + "image_footnote": [], + "bbox": [ + 174, + 522, + 468, + 651 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64e00c7a40f5d8bfca22499a480991f9780f7757dc5984363b283b0279a25fc0.jpg", + "image_caption": [ + "Figure 3: The Model of Other Agents (MOA) architecture learns both an RL policy $\\pi _ { e }$ , and a supervised model that predicts the actions of other agents, $\\mathbf { \\pmb { a } } _ { t + 1 }$ The predictions of the supervised model are used for computing the influence reward. " + ], + "image_footnote": [], + "bbox": [ + 498, + 520, + 816, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 INFLUENCE VIA MODELING OTHER AGENTS ", + "text_level": 1, + "bbox": [ + 178, + 744, + 501, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computing the causal influence reward as introduced in Section 2.1 requires knowing the probability of $B$ ’s next action given a counterfactual, $p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , s _ { t } ^ { B } )$ , which we previously solved by using a centralised controller that could access other agent’s policy networks. While using a centralised training framework is common in MARL (e.g. Foerster et al. (2017; 2016)), it is less realistic than a scenario in which each agent is trained independently. We can relax this assumption and achieve independent training by equipping each agent with its own internal Model of Other Agents (MOA). The MOA consists of a second set of fully-connected and LSTM layers connected to the agent’s convolutional layer (see Figure 3), and is trained to predict all other agents’ next actions given their current action, and the agent’s egocentric view of the state: $p ( { \\pmb a } _ { t + 1 } | { \\pmb a } _ { t } , { \\pmb s } _ { t } ^ { A } )$ . The MOA is trained using cross-entropy loss over observed action trajectories. ", + "bbox": [ + 174, + 762, + 825, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A trained MOA can be used to compute the social influence reward in the following way. Each agent can “imagine” counterfactual actions that it could have taken at each timestep, and use its internal MOA to predict the effect on other agents. It can then give itself reward for taking actions that it estimates were the most influential. This has an intuitive appeal, because it resembles how humans reason about their effect on others (Ferguson et al., 2010). We may often find ourselves asking counterfactual questions of the form, “How would she have reacted if I had said or done something else in that situation?”, which we can only answer using our internal model of others. ", + "bbox": [ + 171, + 895, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both the MOA and communication approaches are an important improvement over the original model shown in Figure 1, which computed influence within a given timestep and required that agent $A$ choose its action $a _ { t } ^ { A }$ first, and this action be transmitted to agent $B$ as input. This meant that only some agents (those acting first) could be influencers. In contrast, using influence for communication or with a MOA are general approaches that can be implemented in any agent, and allow all agents to mutually influence each other. ", + "bbox": [ + 174, + 180, + 825, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now seek to estimate influence in the next timestep, meaning the influence of $a _ { t } ^ { A }$ on $a _ { t + 1 } ^ { B }$ , which requires modeling $p ( a _ { t + 1 } ^ { B } | \\underline { { a } } _ { t } ^ { A } , s _ { t } ^ { A } )$ . The corresponding causal diagrof $a _ { t } ^ { A }$ is son $a _ { t + 1 } ^ { B }$ in Figure 4. We can infer the causal effectby conditioning on the shaded variables (so that there are no back-door paths) (Pearl et al., 2016). Learning a model of $p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { \\bar { A } } , s _ { t } ^ { A } )$ requires implicitly modeling both the environment transition function $T$ (to predict $s _ { t + 1 } )$ , as well as relevant aspects of the internal LSTM state of the other agent, $u _ { t + 1 } ^ { B ^ { * } }$ , as highlighted in Figure 4. ", + "bbox": [ + 174, + 271, + 517, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We enable agents to condition their policy on the actions of other agents in the previous timestep (actions are visible), and only give the social influence reward to an agent when the agent it is attempting to influence is within its field-of-view, because the estimates of $p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { A } )$ are likely to be more accurate when $B$ is visible to $A ^ { 2 }$ . The latter constraint could have the sideeffect of encouraging agents to stay in closer proximity. ", + "bbox": [ + 174, + 433, + 516, + 546 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c636b44cc733a0638da4fbce914888df05874d72cc520f1bcf72e309b465e051.jpg", + "image_caption": [ + "Figure 4: Causal diagram in the MOA case. Shaded nodes are conditioned on, and we intervene on $a _ { t } ^ { A }$ (blue node) by replacing it with counterfactuals. Nodes with a green background must be modeled using the MOA module. Note that there is no backdoor path between $a _ { t } ^ { A }$ and $s _ { t }$ since it would require traversing a collider that is not in the conditioning set. " + ], + "image_footnote": [], + "bbox": [ + 534, + 270, + 789, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, an intrinsic social reward based on proximity is also a reasonable approach to approximating human social motivation. Humans seek affiliation and to spend time near other people (Tomasello, 2009). ", + "bbox": [ + 174, + 546, + 825, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.5 SEQUENTIAL SOCIAL DILEMMAS ", + "text_level": 1, + "bbox": [ + 176, + 597, + 429, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First proposed by Leibo et al. (2017), Sequential Social Dilemmas (SSDs) are spatially and temporally extended multi-agent games that have a payoff structure similar to that of Prisoner’s Dilemma (PD). That is, an individual agent can obtain higher reward by engaging in defecting, non-cooperative behavior (and thus is rationally motivated to defect), but the average payoff per agent will be higher if all agents cooperate (see Figure 9 of the Appendix). The paradoxical reward structure makes it extremely difficult for traditional RL agents to learn to coordinate to solve the tasks (Hughes et al., 2018). We experiment with two SSDs in this work, a public goods game Cleanup, and a tragedy-of-the-commons game Harvest (see Figure 5). In both games apples (green tiles) provide the rewards, and agents also have the ability to punish each other with a fining beam. Further details are available in Appendix Section 6.1. ", + "bbox": [ + 173, + 621, + 825, + 747 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20206990cd332b2c23f36bfab8c7c99fd397bf6323218026ab67daf49c621e5f.jpg", + "image_caption": [ + "Figure 5: The two SSD environments, Cleanup (left) and Harvest (right). Agents can exploit other agents for immediate payoff, but at the expense of the long-term collective reward of the group. " + ], + "image_footnote": [], + "bbox": [ + 183, + 763, + 818, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 334, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several attempts have been made to develop intrinsic social motivation rewards3. Sequeira et al. (2011) developed hand-crafted rewards specific to a foraging environment, in which agents were punished for eating more than their fair share of food. Another approach gave agents an emotional intrinsic reward based on their perception of their neighbours’ cooperativeness in a networked version of the iterated prisoner’s dilemma (Yu et al., 2013). This approach is limited to scenarios in which it is possible to directly classify each action as cooperative or non-cooperative, which is untenable in complex settings with long-term strategies, such as the SSDs under investigation here. Hughes et al. (2018) introduced an inequity aversion motivation, which penalized agents if their rewards differed too much from those of the group. Another approach used prosocial reward shaping to show that if even a single agent is trained to optimize for the rewards of other agents, it can help the group obtain better collective outcomes (Peysakhovich & Lerer, 2018). However, these both require the ability to observe other agent’s rewards, which may be an unrealistic assumption, depending on the application. ", + "bbox": [ + 174, + 137, + 825, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another body of work has focused on training agents to learn emergent communication protocols (Foerster et al., 2016; Cao et al., 2018; Choi et al., 2018; Lazaridou et al., 2018; Bogin et al., 2018), with many authors finding that selfish agents do not learn to use an ungrounded, cheap talk communication channel effectively. Crawford & Sobel (1982) find that in theory, the information revealed in communication (in equilibrium) is proportional to amount of common interest; thus, as agents’ interests diverge, no communication is to be expected. And while communication can emerge when agents are prosocial (Foerster et al., 2016; Lazaridou et al., 2018) or hand-crafted (Crandall et al., 2017), self-interested agents do not to learn to communicate (Cao et al., 2018). We test whether the social influence reward can encourage agents to learn to communicate more effectively in complex environments with challenging social dilemma dynamics. ", + "bbox": [ + 174, + 311, + 825, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, Oudeyer & Kaplan (2006) show that a robot trained with a curiosity-based intrinsic motivation to maximize learning progress learns to prefer vocalizing sounds imitated by another robot over interaction with other objects in the environment. Follow-up papers suggest that curiosity may be a sufficient motivation to encourage agents, or even children, to learn to communicate with others (Oudeyer & Smith, 2016; Forestier & Oudeyer, 2017). ", + "bbox": [ + 174, + 444, + 825, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our MOA network is related to work on machine theory of mind (Rabinowitz et al., 2018), which demonstrated that a model trained to predict agents’ actions is able to model false beliefs. With LOLA, Foerster et al. (2018) train agents that model the impact of their policy on the parameter updates of other agents, and directly incorporate this into the agent’s own learning rule. ", + "bbox": [ + 174, + 520, + 825, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Barton et al. (2018) propose causal influence as a way to measure coordination between agents, specifically using Convergence Cross Mapping (CCM) to analyze the degree of dependence between two agents’ policies. The limitation of this approach is that CCM estimates of causality are known to degrade in the presence of stochastic effects (Tajima et al., 2015). Counterfactual reasoning has also been used in a multi-agent setting, to marginalize out the effect of one agent on a predicted global value function estimating collective reward, and thus obtain an improved baseline for computing each agent’s advantage function (Foerster et al., 2017). A similar paper shows that counterfactuals can be used with potential-based reward shaping to improve credit assignment for training a joint policy in multi-agent RL Devlin et al. (2014). However, once again these approaches rely on a centralised controller. ", + "bbox": [ + 174, + 583, + 825, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following in the tradition of the empowerment literature, authors have investigated mutual information (MI) as a powerful tool for designing social rewards. Strouse et al. (2018) train agents to maximize or minimize the MI between their actions and a categorical goal, and show how this can be used to signal or hide the agent’s intentions. However, this approach depends on agents pursuing a known, categorical goal. Guckelsberger et al. (2018), in pursuit of the ultimate video game adversary, develop an agent that maximizes its empowerment over its own states, minimizes the player’s empowerment over their states, and maximizes its empowerment over the player’s next state. This third goal, termed transfer empowerment, is obtained by maximizing the MI between the agent’s actions and the player’s future state. While similar to our approach, the authors find that agents trained with transfer empowerment simply tend to stay near the player. Further, the agents are not trained with RL, but rather analytically compute these measures in simple grid-world environments. As such, the agent cannot learn to model other agents or the environment. ", + "bbox": [ + 174, + 715, + 825, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 320, + 117 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The following sections present the results of training agents with the social influence reward in three settings: (1) using a centralised controller, (2) using an explicit communication channel, and (3) using a learned model of other agents (MOA). In each case we compare against a standard A3C agent, and an ablated version of the model which is architecturally identical, but does not receive the influence reward. We measure the total collective reward obtained using the best hyperparameter setting tested with 5 random seeds. It is worth noting that we use a curriculum learning approach which gradually increases the weight of the social influence reward over $C$ steps $( C \\in [ 0 . 2 \\mathrm { - } 3 . 5 ] \\times \\bar { 1 } 0 ^ { 8 } )$ ; this can lead to a slight delay before the influence models’ performance improves. ", + "bbox": [ + 173, + 131, + 825, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also provide the results of an additional experiment Section 6.2 of the Appendix, which tests the social influence reward in a simplified environment where the effects of influence are clear. We encourage the reader to examine that section to gain a better intuition for how social influence can foster cooperative behavior in an otherwise selfish agent. ", + "bbox": [ + 174, + 250, + 825, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 CENTRALISED CONTROLLER ", + "text_level": 1, + "bbox": [ + 176, + 329, + 401, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figures 6(a) and 6(d) show the results of training influence with a centralised controller as described in Section 2.1. With this method, the influencer agents transmit their intended action to the influenced agents at each timestep. Therefore, we benchmark against an ablated version of the influence model with visible actions but no influence reward. As is evident in Figures 6(a) and 6(d), introducing an awareness of other agents’ actions helps, but having the social influence reward eventually leads to significantly higher collective reward in both games. ", + "bbox": [ + 173, + 353, + 825, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b9225cc90c9a95e47d5cf85514dba567067d0da375c94be1d1955429ea1bdd4.jpg", + "image_caption": [ + "Figure 6: Total collective reward obtained in all experiments. Error bars show a $9 9 . 5 \\%$ confidence interval (CI) over 5 random seeds, computed within a sliding window of 200 agent steps. The models trained with influence reward (red) significantly outperform the baseline and ablated models. " + ], + "image_footnote": [], + "bbox": [ + 178, + 457, + 813, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While these aggregated results demonstrate the success of our models, they are not sufficient to understand the mechanism through which social influence is helping the agents achieve cooperative behavior. Therefore, we investigated the trajectories produced by high scoring models in both Cleanup and Harvest; the analysis revealed interesting behavior. As an example, in the Cleanup video available here: https: //youtu.be/iH_V5WKQxmo a single agent (shown in purple) was trained with the social influence reward. We see that unlike the other agents, which continue to randomly move and explore while waiting for apples to spawn, the influencer has a strange economy of motion; it only moves on the map when it is pursuing an apple, then stops. Interestingly, examining the trajectory reveals that the influencer uses only two moves to explore the map: turn left, which turns the agent in place without traversing the map, and move right, which moves the agent one square to the right on the map. ", + "bbox": [ + 174, + 781, + 825, + 921 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15faf4dfa88bd78c4d0e9ddf505f7cecc97572565601a0318358c9465b4a4695.jpg", + "image_caption": [ + "Figure 7: A moment of high influence when the purple influencer signals the presence of an apple outside the yellow influencee’s field-of-view (yellow outlined box). " + ], + "image_footnote": [], + "bbox": [ + 176, + 103, + 302, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why did the influencer learn to use only these two moves? We can see that the influencer agent only chooses to move right (i.e. traverse the map) when it is pursuing an apple which is present. The rest of the time it simply turns left on the spot. At $t = 4 9$ , there is a moment of high influence between the influencer and the yellow influencee, which is shown in Figure 7. The influencer has chosen to move right towards an apple that is outside of the ego-centric field-of-view of the yellow agent. Because the purple agent only moves when apples are available, this signals to the yellow agent that an apple must be present above it which it cannot see. This changes the yellow agent’s distribution over its planned action, $p ( a _ { t } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { B } )$ and allows the purple agent to gain influence. A similar moment occurs when the influencer signals to an agent that has been cleaning the river that no apples have appeared by continuing to turn left (see Figure 12 in the Appendix). ", + "bbox": [ + 320, + 103, + 825, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this example, the influencer agent learned to use its own actions as a sort of binary code, which signals the presence or absence of apples in the environment. We also observe this effect in the influence agents in the Harvest task. This type of action-based communication could be likened to the bee waggle dance discovered by von Frisch (1969). Thus, rewarding agents for increasing the mutual information between their actions gave rise not only to cooperative behavior, but in this case, to emergent communication. These results further support the idea of using influence as a reward for training agents to communicate. ", + "bbox": [ + 320, + 277, + 825, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 INFLUENCE THROUGH COMMUNICATION ", + "text_level": 1, + "bbox": [ + 176, + 402, + 482, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figures 6(b) and 6(e) show the results of training the agents to use an explicit communication channel, and its effect on their collective reward. In this case, the ablated baseline is a model that has the same structure as in Figure 2, but in which the communication policy $\\pi _ { c }$ is trained only with environmental reward. We observe that the agents which are trained to use the communication channel with additional social influence reward achieve significantly higher collective reward in both games. In fact, in the case of Cleanup, we found that $\\alpha = 0$ in the optimal hyperparameter settings, meaning that it was most effective to train the communication head with zero extrinsic or environmental reward (see Table 2 in the Appendix). This suggests that influence alone can be a sufficient mechanism for training an effective communication policy. ", + "bbox": [ + 173, + 421, + 825, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To analyze the communication behaviour learned by the agents, we introduce three metrics. Speaker consistency, is a normalized score $\\in [ 0 , 1 ]$ which assesses the entropy of $p ( a ^ { k } | m ^ { k } )$ and $p ( m ^ { k } | a ^ { k } )$ to determine how consistently a speaker agent emits a particular symbol when it takes a particular action, and vice versa (the formula is given in Appendix Section 6.3.4). We expect this measure to be high if, for example, the speaker always emits the same symbol when it is cleaning the river. We also introduce two measures of i(1) symbol/action $\\mathbf { I C } = I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )$ dination (IC), which are both measures of mutual information (MI): measures the MI between the influencer/speaker’s symbol and the influencee/listener’s next action, and (2) action/action $\\begin{array} { r } { \\mathbf { I C } = I ( a _ { t } ^ { A } ; a _ { t + 1 } ^ { B } ) } \\end{array}$ measures the MI between the influencer’s action and the influencee’s action in the next timestep. To compute these measures we first average over all trajectory steps, then take the maximum value between any two agents, to determine if any pair of agents are coordinating. Note that these measures are all instantaneous, as they consider only short-term dependencies across two consecutive timesteps, and cannot capture if an agent communicates influential compositional messages, i.e. information that requires several consecutive symbols to transmit and only then affects the other agents behavior. ", + "bbox": [ + 173, + 540, + 826, + 736 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfabcddf1dd58ea122c746b3e02eae517a3782b403a73cf07bb8bc425f0202f7.jpg", + "image_caption": [ + "Figure 8 presents the results. The speaker consistencies metric reveals that agents trained with the influence reward communicate less ambiguously about their own actions, indicating that the emergent communication is more meaningful. The instantaneous coordination metrics demonstrate ", + "Figure 8: Metrics describing the quality of learned communication protocols. The models trained with influence reward exhibit more consistent communication and more coordination, especially in moments where influence is high. " + ], + "image_footnote": [], + "bbox": [ + 338, + 743, + 820, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the baseline agents trained without influence reward show almost no signs of co-ordinating behavior with communication, i.e. speakers saying A and listeners doing B consistently. This result is aligned with both theoretical results in cheap-talk literature (Crawford & Sobel, 1982), and recent empirical results in MARL (e.g. (Foerster et al., 2016; Lazaridou et al., 2018; Cao et al., 2018)). ", + "bbox": [ + 174, + 103, + 823, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In contrast, we do see highly coordinated behavior between influence agents, but only when we limit the analysis to timesteps on which influence was high (cf. influential moments in Figure 8). If we inspect the results for agents trained with influence on the two tasks, a common pattern emerges: influence is sparse in time. An agent’s influence is only greater than its mean influence in less than $10 \\%$ of timesteps. Because the listener agent is not compelled to listen to any given speaker, listeners selectively listen to a speaker only when it is beneficial, and influence cannot occur all the time. Only when the listener decides to change its action based on the speaker’s message does influence occur, and in these moments we observe high $I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )$ ; an effect that is lost when averaging over the entire trajectory. It appears the influencers have learned a strategy of communicating meaningful information about their own actions, and gaining influence when this becomes relevant enough for the listener to act upon it. ", + "bbox": [ + 174, + 166, + 825, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Examining the relationship between the reward obtained by individual agents and the degree to which they were influenced by other agents gives a compelling result: agents that are the most influenced also achieve higher individual environmental reward, $\\dot { E } ^ { k }$ . We sampled 100 different experimental conditions (i.e., hyper-parameters and random seeds) for both games, collected the influence and individual rewards, normalized them across the 5 agents in each condition, and correlated the resulting list of values. We found that agents who are more often influenced tend to achieve higher task reward in both Cleanup, $\\rho = . 6 7$ , $p { < } 0 . 0 0 1$ , and Harvest, $\\rho = . 3 4$ , $p { < } 0 . 0 0 1$ . This supports the hypothesis stated in Section 2.3: in order to gain influence from another agent by communicating with it, the communication message should contain information that helps the listener maximize its own environmental reward. Since better listeners/influencees are more successful in terms of task reward, we have evidence that useful information was transmitted to them. ", + "bbox": [ + 174, + 311, + 825, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 INFLUENCE VIA MODELING OTHER AGENTS ", + "text_level": 1, + "bbox": [ + 173, + 483, + 501, + 496 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we investigate whether the influence reward is still effective when computed without a centralised controller, but rather through each agent’s own internal Model of Other Agents (MOA) network. In this case, we extend the training period from $3 \\cdot 1 0 ^ { 8 }$ steps to $5 \\cdot 1 0 ^ { 8 }$ , in order to give the MOA model time to train. We also allow the policy LSTM to condition on the actions of other agents in the last timestep. We compare against an ablated version of this architecture (shown in Figure 3), which does not use the output of the MOA module to compute a reward; rather, the MOA module can be thought of as an unsupervised auxiliary task that may help the model to learn a better shared embedding layer, encouraging it to encode information relevant to predicting other agents’ behavior. ", + "bbox": [ + 174, + 503, + 825, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figures 6(c) and 6(f) show the collective reward obtained for agents trained with a MOA module. While we see that the auxiliary task does help to improve reward over the A3C baseline, the influence agent gets consistently higher collective reward. Impressively, for Cleanup, the MOA model scores higher than the original influence agents computed using the centralised controller (CC). As shown in Figure 6(c), the MOA baseline also achieves high collective reward, suggesting that the auxiliary task of modeling other agents helps the MOA agents cooperate more effectively in Cleanup. Further, the independent design of the MOA method allows each agent to influence every other agent, thus generating more reward signal and a greater chance to develop two-way cooperative behavior. ", + "bbox": [ + 174, + 622, + 825, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4 of the Appendix gives the final collective reward obtained by each model for all three experiments. Interestingly, several influence models are able to achieve higher collective reward than the previous state-of-the-art scores for these environments (275 for Cleanup and 750 for Harvest) (Hughes et al., 2018). This is compelling, given that previous work relied on the assumption that agents could view one another’s rewards; we make no such assumption, instead relying only on agents viewing each other’s actions. ", + "bbox": [ + 174, + 741, + 825, + 810 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 DISCUSSION AND CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 178, + 829, + 464, + 844 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The experiments above have demonstrated that an intrinsic social reward based on having causal influence on the actions of other agents consistently improves cooperation and leads to higher collective return in the MA social dilemmas under investigation. In some cases, the influence reward drove agents to learn an emergent communication protocol via their actions. This is compelling, and confirms the connection between maximizing influence and maximizing the mutual information between agents’ actions. ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "However, it is important to consider the limitations of the influence reward. Whether it will always give rise to cooperative behavior may depend on the specifics of the environment, task, and the trade-off between environmental and influence reward. Although influence is arguably necessary for cooperation (e.g. two agents cooperating to lift a box would have a high degree of influence between their actions), it may not be sufficient, in that it may be possible to influence another agent without helping it. For example, it is possible that agents could have gained influence in the tasks studied here by threatening to attack other agents with their fining beam. We believe this type of behavior did not emerge because communicating information represents the cheapest and most effective way to gain influence. Influencers do not have to sacrifice much in terms of their own environmental reward in order to communicate to other agents. ", + "bbox": [ + 174, + 103, + 823, + 228 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rewarding influence over an explicit communication channel may not be subject to this limitation, because influential communication may be inherently beneficial to the listener (at least in the case where listeners and speakers interact repeatedly). Since listeners can easily ignore communication messages if they do not help to obtain environmental reward, a speaker must transmit valuable information in order to gain influence through communication. There is no advantage to the speaker for communicating unreliably, because it would lose influence with the listener over time (although this is no longer guaranteed in one-shot interactions). Indeed, our results reveal that agents benefit from being influenced by (listening to) communication messages by obtaining higher individual reward, suggesting that the messages contain valuable information. Further, we found that the communication protocols learned via influence reward were more meaningful, and that the influence reward allowed agents to obtain higher collective return. Therefore, we suggest that influence could be a promising way to train emergent communication protocols in various settings. ", + "bbox": [ + 174, + 236, + 825, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we have shown that influence can be computed by augmenting agents with an internal model that predicts the actions of other agents, and using this MOA model to simulate the effect of an agent’s actions on others. This represents an important step forward in multi-agent intrinsic social motivation, because it implies that the influence reward can be computed without having access to another agent’s reward function, or requiring a centralised controller. ", + "bbox": [ + 174, + 410, + 825, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 FUTURE WORK ", + "text_level": 1, + "bbox": [ + 174, + 494, + 316, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Using counterfactuals to allow agents to understand the effects of their actions on other agents could be a promising approach with a number of extensions. Perhaps agents could use counterfactuals to develop a form of ‘empathy‘, by simulating how their actions affect another agent’s value function. Or, social influence could be used to drive coordinated behavior in robots attempting to do cooperative manipulation and control tasks. Finally, if we view multi-agent networks as a single agent, influence could be used as a regularizer to encourage different modules of the network to integrate information from other networks; for example, perhaps it could prevent collapse in hierarchical RL. ", + "bbox": [ + 174, + 517, + 825, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 633, + 281, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sean L Barton, Nicholas R Waytowich, Erin Zaroukian, and Derrik E Asher. Measuring collaborative emergent behavior in multi-agent reinforcement learning. arXiv preprint arXiv:1807.08663, 2018. ", + "bbox": [ + 174, + 657, + 823, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ben Bogin, Mor Geva, and Jonathan Berant. Emergence of communication in an interactive world with consistent speakers. arXiv preprint arXiv:1809.00549, 2018. ", + "bbox": [ + 173, + 696, + 821, + 726 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent communication through negotiation. arXiv preprint arXiv:1804.03980, 2018. ", + "bbox": [ + 173, + 736, + 823, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Philippe Capdepuy, Daniel Polani, and Chrystopher L Nehaniv. Maximization of potential information flow as a universal utility for collective behaviour. In Artificial Life, 2007. ALIFE’07. IEEE Symposium on, pp. 207–213. Ieee, 2007. ", + "bbox": [ + 173, + 775, + 821, + 818 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Edward Choi, Angeliki Lazaridou, and Nando de Freitas. Compositional obverter communication learning from raw visual input. arXiv preprint arXiv:1804.02341, 2018. ", + "bbox": [ + 169, + 828, + 825, + 857 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jacob W. Crandall, Mayada Oudah, Tennom Chenlinangjia, Fatimah Ishowo-Oloko, Sherief Abdallah, Jean-Franc¸ois Bonnefon, Manuel Cebrian, Azim Shariff, Michael A. Goodrich, and Iyad Rahwan. ´ Cooperating with machines. CoRR, abs/1703.06207, 2017. URL http://arxiv.org/abs/ 1703.06207. ", + "bbox": [ + 174, + 867, + 826, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Vincent P Crawford and Joel Sobel. Strategic information transmission. Econometrica: Journal of the Econometric Society, pp. 1431–1451, 1982. ", + "bbox": [ + 169, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sam Devlin, Logan Yliniemi, Daniel Kudenko, and Kagan Tumer. Potential-based difference rewards for multiagent reinforcement learning. In Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems, pp. 165–172. International Foundation for Autonomous Agents and Multiagent Systems, 2014. ", + "bbox": [ + 174, + 140, + 825, + 196 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Heather J. Ferguson, Christoph Scheepers, and Anthony J. Sanford. Expectations in counterfactual and theory of mind reasoning. Language and Cognitive Processes, 25(3):297–346, 2010. doi: 10.1080/ 01690960903041174. URL https://doi.org/10.1080/01690960903041174. ", + "bbox": [ + 174, + 205, + 825, + 248 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pp. 2137–2145, 2016. ", + "bbox": [ + 173, + 257, + 826, + 300 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. arXiv preprint arXiv:1705.08926, 2017. ", + "bbox": [ + 173, + 308, + 823, + 337 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jakob Foerster, Richard Y Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch. Learning with opponent-learning awareness. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp. 122–130. International Foundation for Autonomous Agents and Multiagent Systems, 2018. ", + "bbox": [ + 173, + 344, + 826, + 402 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sebastien Forestier and Pierre-Yves Oudeyer. A unified model of speech and tool use early development.´ In 39th Annual Conference of the Cognitive Science Society (CogSci 2017), 2017. ", + "bbox": [ + 173, + 410, + 823, + 440 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Felix A Gers, Jurgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm. ¨ 1999. ", + "bbox": [ + 173, + 446, + 823, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Christian Guckelsberger, Christoph Salge, and Julian Togelius. New and surprising ways to be mean. adversarial npcs with coupled empowerment minimisation. arXiv preprint arXiv:1806.01387, 2018. ", + "bbox": [ + 173, + 484, + 823, + 515 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yuval Noah Harari. Sapiens: A brief history of humankind. Random House, 2014. ", + "bbox": [ + 173, + 522, + 689, + 539 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Joseph Henrich. The Secret of Our Success: How culture is driving human evolution, domesticating our species, and making us smart. Princeton University Press, Princeton, NJ, 2015. URL http: //press.princeton.edu/titles/10543.html. ", + "bbox": [ + 173, + 546, + 826, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Esther Herrmann, Josep Call, Mar´ıa Victoria Hernandez-Lloreda, Brian Hare, and Michael Tomasello. \\` Humans have evolved specialized skills of social cognition: The cultural intelligence hypothesis. Science, 317(5843):1360–1366, 2007. ISSN 0036-8075. doi: 10.1126/science.1146282. URL http: //science.sciencemag.org/content/317/5843/1360. ", + "bbox": [ + 173, + 597, + 826, + 655 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Edward Hughes, Joel Z Leibo, Matthew G Phillips, Karl Tuyls, Edgar A Due´nez-Guzm ˜ an, Antonio Garc ´ ´ıa Castaneda, Iain Dunning, Tina Zhu, Kevin R McKee, Raphael Koster, et al. Inequity aversion improves ˜ cooperation in intertemporal social dilemmas. In Advances in neural information processing systems (NIPS), Montreal, Canada, 2018. ", + "bbox": [ + 173, + 662, + 826, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexander S Klyubin, Daniel Polani, and Chrystopher L Nehaniv. Empowerment: A universal agent-centric measure of control. In Evolutionary Computation, 2005. The 2005 IEEE Congress on, volume 1, pp. 128–135. IEEE, 2005. ", + "bbox": [ + 173, + 727, + 823, + 770 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kevin N. Laland. Darwin’s unfinished symphony $\\therefore$ how culture made the human mind / Kevin N. Laland. Princeton University Press Princeton, 2017. ISBN 9781400884872 140088487. ", + "bbox": [ + 169, + 779, + 825, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic communication from referential games with symbolic and pixel input. arXiv preprint arXiv:1804.03984, 2018. ", + "bbox": [ + 173, + 815, + 825, + 858 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Joel Z Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. Multi-agent reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, pp. 464–473. International Foundation for Autonomous Agents and Multiagent Systems, 2017. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Joseph T Lizier and Mikhail Prokopenko. Differentiating information transfer and causal effect. The European Physical Journal B, 73(4):605–615, 2010. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alicia P Melis and Dirk Semmann. How is human cooperation different? Philosophical Transactions of the Royal Society of London B: Biological Sciences, 365(1553):2663–2674, 2010. ", + "bbox": [ + 171, + 141, + 825, + 171 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pp. 1928–1937, 2016. ", + "bbox": [ + 176, + 179, + 823, + 223 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pierre-Yves Oudeyer and Fred´ eric Kaplan. Discovering communication. ´ Connection Science, 18(2): 189–206, 2006. ", + "bbox": [ + 173, + 232, + 823, + 261 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pierre-Yves Oudeyer and Linda B Smith. How evolution may work through curiosity-driven developmental process. Topics in Cognitive Science, 8(2):492–502, 2016. ", + "bbox": [ + 173, + 270, + 823, + 300 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning (ICML), volume 2017, 2017. ", + "bbox": [ + 173, + 309, + 826, + 352 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018. ", + "bbox": [ + 173, + 361, + 825, + 390 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: a primer. John Wiley & Sons, 2016. ", + "bbox": [ + 173, + 400, + 823, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Julien Perolat, Joel Z Leibo, Vinicius Zambaldi, Charles Beattie, Karl Tuyls, and Thore Graepel. A multi-agent reinforcement learning model of common-pool resource appropriation. In Advances in Neural Information Processing Systems, pp. 3643–3652, 2017. ", + "bbox": [ + 173, + 438, + 823, + 482 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alexander Peysakhovich and Adam Lerer. Prosocial learning agents solve generalized stag hunts better than selfish ones. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp. 2043–2044. International Foundation for Autonomous Agents and Multiagent Systems, 2018. ", + "bbox": [ + 173, + 489, + 825, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Neil C Rabinowitz, Frank Perbet, H Francis Song, Chiyuan Zhang, SM Eslami, and Matthew Botvinick. Machine theory of mind. arXiv preprint arXiv:1802.07740, 2018. ", + "bbox": [ + 173, + 556, + 823, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Thomas C Schelling. Hockey helmets, concealed weapons, and daylight saving: A study of binary choices with externalities. Journal of Conflict resolution, 17(3):381–428, 1973. ", + "bbox": [ + 173, + 594, + 823, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jurgen Schmidhuber. Formal theory of creativity, fun, and intrinsic motivation (1990–2010). ¨ IEEE Transactions on Autonomous Mental Development, 2(3):230–247, 2010. ", + "bbox": [ + 173, + 633, + 823, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pedro Sequeira, Francisco S Melo, Rui Prada, and Ana Paiva. Emerging social awareness: Exploring intrinsic motivation in multiagent learning. In Development and Learning (ICDL), 2011 IEEE International Conference on, volume 2, pp. 1–6. IEEE, 2011. ", + "bbox": [ + 173, + 671, + 826, + 714 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Satinder P. Singh, Andrew G. Barto, and Nuttapong Chentanez. Intrinsically motivated reinforcement learning. In Advances in Neural Information Processing Systems 17 [Neural Information Processing Systems, NIPS 2004, December 13-18, 2004, Vancouver, British Columbia, Canada], pp. 1281–1288, 2004. URL http://papers.nips.cc/paper/ 2552-intrinsically-motivated-reinforcement-learning. ", + "bbox": [ + 173, + 724, + 825, + 795 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Katherine KM Stavropoulos and Leslie J Carver. Research review: social motivation and oxytocin in autism–implications for joint attention development and intervention. Journal of Child Psychology and Psychiatry, 54(6):603–618, 2013. ", + "bbox": [ + 176, + 804, + 823, + 847 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "DJ Strouse, Max Kleiman-Weiner, Josh Tenenbaum, Matt Botvinick, and David Schwab. Learning to share and hide intentions using information regularization. arXiv preprint arXiv:1808.02093, 2018. ", + "bbox": [ + 174, + 856, + 821, + 886 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Satohiro Tajima, Toru Yanagawa, Naotaka Fujii, and Taro Toyoizumi. Untangling brain-wide dynamics in consciousness by cross-embedding. PLoS computational biology, 11(11):e1004537, 2015. ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Michael Tomasello. Why we cooperate. MIT press, 2009. ", + "bbox": [ + 173, + 103, + 537, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Carel P van Schaik and Judith M Burkart. Social learning and evolution: the cultural intelligence hypothesis. Philosophical Transactions of the Royal Society B: Biological Sciences, 366(1567):1008–1016, 2011. ", + "bbox": [ + 173, + 127, + 823, + 156 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Karl von Frisch. The dance language and orientation of bees. 5, 06 1969. ", + "bbox": [ + 173, + 165, + 629, + 180 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256, 1992. ", + "bbox": [ + 173, + 189, + 823, + 218 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chao Yu, Minjie Zhang, and Fenghui Ren. Emotional multiagent reinforcement learning in social dilemmas. In International Conference on Principles and Practice of Multi-Agent Systems, pp. 372–387. Springer, 2013. ", + "bbox": [ + 176, + 227, + 825, + 268 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 APPENDIX ", + "text_level": 1, + "bbox": [ + 174, + 102, + 289, + 117 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.1 SEQUENTIAL SOCIAL DILEMMAS ", + "text_level": 1, + "bbox": [ + 176, + 126, + 434, + 140 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In each of the sequential social dilemma (SSD) games studied above, an agent is rewarded $+ 1$ for every apple it collects, but the apples are a limited resource. In Harvest (a tragedy of the commons game), apples regenerate more slowly the faster they are harvested, and if an exploiting agent consumes all of the apples, they will not grow back; agents must cooperate to harvest sustainably. In Cleanup (a public goods game), apples are generated based on the amount of waste in a nearby river. Agents can use a cleaning beam action to clean the river when they are positioned in it; or they can simply consume the apples the other agent produces. Agents also have a fining beam action which they can use to fine nearby agents $- 5 0$ reward. ", + "bbox": [ + 173, + 145, + 825, + 243 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Figure 9 gives the Schelling diagram for both SSD tasks under investigation. A Schelling diagram (Schelling, 1973; Perolat et al., 2017) shows the relative payoffs for a single agent’s strategy given a fixed number of other agents who are cooperative. Schelling diagrams generalize payoff matrices to multi-agent settings, and make it easy to visually recognize game-theoretic properties like Nash equilibria (see Schelling (1973) for more details). ", + "bbox": [ + 173, + 250, + 825, + 319 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cd8620142473269dee9b7be7e5b1fb98864ebd4ee20a85b56ccae63f16b46cc2.jpg", + "image_caption": [ + "Figure 9: Schelling diagrams for the two social dilemma tasks show that an individual is almost always motivated to defect, even though the group will get higher reward if there are more cooperators. " + ], + "image_footnote": [], + "bbox": [ + 184, + 330, + 818, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.2 ADDITIONAL EXPERIMENT - BOX TRAPPED ", + "text_level": 1, + "bbox": [ + 173, + 584, + 503, + 598 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As a proof-of-concept experiment to test whether the influence reward works as expected, we constructed a special environment, shown in Figure 10. In this environment, one agent (teal) is trapped in a box. The other agent (purple) has a special action it can use to open the box... or it can simply choose to consume apples, which exist outside the box and are inexhaustible in this environment. ", + "bbox": [ + 174, + 604, + 485, + 714 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As expected, a vanilla A3C agent learns to act selfishly; the purple agent will simply consume apples, and chooses the open box action in $0 \\%$ of trajectories once the policy has converged. A video of A3C agents trained in this environment is available at: https://youtu.be/C8SE9_ $\\Upsilon \\mathrm { K z x I }$ , which shows that the purple agent leaves its compatriot trapped in the box throughout the trajectory. ", + "bbox": [ + 174, + 722, + 483, + 847 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3f18791740d5f30d7f41c7fce3b78ff4b93cfac39e1f81dcfeb16188b74db18f.jpg", + "image_caption": [ + "Figure 10: The Box trapped environment in which the teal agent is trapped, and the purple agent can release it with a special open box action. " + ], + "image_footnote": [], + "bbox": [ + 503, + 616, + 813, + 785 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In contrast, an agent trained with the social influ \nence reward chooses the open box action in $88 \\%$ of trajectories, releasing its fellow agent so that they are both able to consume apples. A video of this behavior is shown at: https://youtu.be/Gfo248-qt3c. Further, as Figure 11(a) reveals, the purple influencer agent usually chooses to open the box within the first few steps of the trajetory, giving its fellow agent more time to collect reward. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Most importantly though, Figure 11(b) shows the influence reward over the course of a trajectory in the Box trapped environment. The agent chooses the open box action in the second timestep; at this point, we see a corresponding spike in the influence reward. This reveals that the influence reward works as expected, incentivizing an action which has a strong — and in this case, prosocial — effect on the other agent’s behavior. ", + "bbox": [ + 174, + 103, + 825, + 172 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/87f2f4c8939684aaa1a065bb1bd2a40dcabf6619e1b311cd76ab190958323091.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 186, + 483, + 348 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "(a) Number of times the open box action occurs at each trajectory step over 100 trajectories. ", + "bbox": [ + 173, + 357, + 501, + 382 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/aa590ebca444f816660ca98874365417e4da0912c85a637d878b313133db889b.jpg", + "image_caption": [ + "(b) Influence reward over a trajectory in Box trapped ", + "Figure 11: The Box trapped proof-of-concept experiment reveals that an agent gets high influence for letting another agent out of a box in which it is trapped. " + ], + "image_footnote": [], + "bbox": [ + 516, + 193, + 808, + 353 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "6.3 IMPLEMENTATION DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 446, + 393, + 460 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "All models are trained with a single convolutional layer with a kernel of size 3, stride of size 1, and 6 output channels. This is connected to two fully connected layers of size 32 each, and an LSTM with 128 cells. We use a discount factor $\\gamma = . 9 9$ . The number of agents $N$ is fixed to 5. ", + "bbox": [ + 174, + 467, + 825, + 508 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "As mentioned in Section 2.2, the social influence reward can be computed using a number of divergence measures, including JSD. We also experiment with training the agents using the pointwise mutual information (the innermost term of Eq. 3), which is given by: ", + "bbox": [ + 176, + 515, + 823, + 559 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/3711246b82494f6bba66f95bcb17310950d9a545ef4a431d8f248a34b3326879.jpg", + "text": "$$\np m i ( a ^ { A } ; a ^ { B } | Z { = } z ) { = } \\log \\frac { p ( a ^ { B } | a ^ { A } , z ) } { p ( a ^ { B } | z ) } { = } \\mathrm { l o g } \\frac { p ( a ^ { A } , a ^ { B } | z ) } { p ( a ^ { A } | z ) p ( a ^ { B } | z ) } .\n$$", + "text_format": "latex", + "bbox": [ + 300, + 563, + 696, + 598 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This PMI term is precisely the local information flow proposed by Lizier & Prokopenko (2010) as a measure of direct causal effect; the expectation of the PMI over $p \\dot { ( } a ^ { B } , a ^ { A } | z )$ is the MI. and gives us a measure of influence of a single action of $A$ on the single action taken by $B$ . ", + "bbox": [ + 176, + 602, + 821, + 643 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In addition to the comparison function used to compute influence, there are many other hyperparameters that can be tuned for each model. We use a random search over hyperparameters, ensuring a fair comparison with the search size over the baseline parameters that are shared with the influence models. For all models we search for the optimal entropy reward and learning rate, where we anneal the learning rate from an initial value lr init to lr final. The below sections give the parameters found to be most effective for each of the three experiments. ", + "bbox": [ + 173, + 650, + 825, + 734 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "6.3.1 CENTRALISED CONTROLLER HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 173, + 750, + 553, + 763 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In this setting we vary the number of influencers from 1−4, the influence reward weight $\\beta$ , and the number of curriculum steps over which the weight of the influence reward is linearly increased $C$ . In this setting, since we have a centralised controller, we also experiment with giving the influence reward to the agent being influenced as well, and find that this sometimes helps. This ‘influencee’ reward is not used in the other two experiments, since it precludes independent training. The hyperparameters found to give the best performance for each model are shown in Table 1. ", + "bbox": [ + 174, + 772, + 825, + 857 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "6.3.2 COMMUNICATION HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 176, + 871, + 483, + 886 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Because the communication models have an extra A2C output head for the communication policy, we use an additional entropy regularization term just for this head, and apply a weight to the communication loss in the loss function. We also vary the number of communication symbols that the agents can emit, and the size of the linear layer that connects the LSTM to the communication policy layer, which we term the communication embedding size. Finally, in the communication regime, we experiment to setting the weight on the extrinsic reward E, $\\alpha$ , to zero. The best hyperparameters for each of the communication models are shown in Table 2. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/f4c2027e5c945541b51744fd4ee3d446041a4b1d05ad3603be18149e0ee814e4.jpg", + "table_caption": [ + "Table 1: Optimal hyperparameter settings for the models in the centralised controller experiment. " + ], + "table_footnote": [], + "table_body": "
HyperparameterCleanupHarvest
A3CVisibleactionsInfluenceA3C baselineVisibleactionsInfluence
Entropy reg.baseline .00176baseline .00176.000248.000687baseline .00184.00025
lr_init.00126.00126.00107.00136.00215.00107
lr_end.000012.000012.000042.000028.000013.000042
Number of influencers3133
Influence weight β0.1460.224
Curriculum C11401140
Policy comparisonJSDPMI
Influencee reward10
", + "bbox": [ + 192, + 99, + 812, + 243 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/a539e1e377f59920fd6287d7d14b0f7b1c52e3c2927f2e15dece92f3ff011d12.jpg", + "table_caption": [ + "Table 2: Optimal hyperparameter settings for the models in the communication experiment. " + ], + "table_footnote": [], + "table_body": "
HyperparameterCleanupHarvest
A3C baselineComm. baselineInfluence comm.A3C Comm. baseline baselineInfluence comm.
Entropy reg..00176.000249.00305.000687.000174.00220
lr_init.00126.00223.00249.00136.00137.000413
lr_end.000012.000022.0000127.000028.0000127.000049
Influence weight β=02.752=04.825
Extrinsic reward01.0
weight α
Curriculum C1 KL8
Policy comparison.000789KL
Comm. entropy reg..0758.00208
Comm. loss weight9.0709
Symbol vocab size327
Comm. embedding16
", + "bbox": [ + 223, + 292, + 772, + 486 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/d4fbff5faab5fec4aac6c7522438316d93a7beef9c4c8cf0187176e97538ea27.jpg", + "table_caption": [ + "Table 3: Optimal hyperparameter settings for the models in the model of other agents (MOA) experiment. " + ], + "table_footnote": [], + "table_body": "
HyperparameterCleanupHarvest
A3C baselineMOA baselineInfluence MOAA3C baselineMOA baselineInfluence MOA
Entropy reg..00176.00176.00176.000687.00495.00223
lr_init.00126.00123.00123.00136.00206.00120
lr_end.000012.000012.000012.000028.000022.000044
Influence weight β0.620=02.521
MOA loss weight1.31215.007=1.71110.911
Curriculum C-40=1226
Policy comparison=KL==KL
Train MOA onlyFalseTrueFalseTrue
when visible
", + "bbox": [ + 233, + 536, + 766, + 690 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 753, + 825, + 824 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "6.3.3 MODEL OF OTHER AGENTS (MOA) HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 174, + 852, + 598, + 867 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The MOA hyperparameters include whether to only train the MOA with cross-entropy loss on the actions of agents that are visible, and how much to weight the supervised loss in the overall loss of the model. The best hyperparameters are shown in Table 3. ", + "bbox": [ + 174, + 882, + 823, + 922 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/4ecc39bab522686cadd4d604d2910d079ffcab346895d226ba7b33dec2aefc19.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CleanupHarvest
A3Cbaseline89485
Inequity aversion (Hughes et al., 2018)275750
Influence - Basic1901073
Influence - Communication166951
Influence -Model of other agents392588
", + "bbox": [ + 300, + 101, + 699, + 186 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Table 4: Final collective reward over the last 50 agent steps for each of the models considered. Bolded entries represent experiments in which the influence models significantly outperformed the scores reported in previous work on inequity aversion(Hughes et al., 2018). This is impressive, considering the inequity averse agents are able to view all other agents’ rewards. We make no such assumption, and yet are able to achieve similar or superior performance. ", + "bbox": [ + 173, + 196, + 826, + 267 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "6.3.4 COMMUNICATION ANALYSIS ", + "text_level": 1, + "bbox": [ + 174, + 292, + 415, + 306 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The speaker consistency metric is calculated as: ", + "bbox": [ + 176, + 316, + 470, + 332 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/3426eefe40c38e35aa09a3955c858cd87f99b15cce7a8836f10826943feba7cc.jpg", + "text": "$$\n\\sum _ { k = 1 } ^ { N } 0 . 5 [ \\sum _ { c } 1 - \\frac { H ( p ( a ^ { k } | m ^ { k } = c ) ) } { H _ { m a x } } + \\sum _ { a } 1 - \\frac { H ( p ( m ^ { k } | a ^ { k } = a ) ) } { H _ { m a x } } ] ,\n$$", + "text_format": "latex", + "bbox": [ + 300, + 335, + 694, + 380 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where $H$ is the entropy function and $H _ { m a x }$ is the maximum entropy based on the number of discrete symbols or actions. The goal of the metric is to measure how much of a 1:1 correspondence exists between a speaker’s action and the speaker’s communication message. ", + "bbox": [ + 176, + 383, + 825, + 426 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "6.4 ADDITIONAL RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 440, + 362, + 454 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Figure 12 shows an additional moment of high influence in the Cleanup game. The purple influencer agent can see the area within the white box, and therefore all of the apple patch. The field-of-view of the magenta influencee is outlined with the magenta box; it cannot see if apples have appeared, even though it has been cleaning the river, which is the action required to cause apples to appear. When the purple influencer turns left and does not move towards the apple patch, this signals to the magenta agent that no apples have appeared, since otherwise the influence would move right. ", + "bbox": [ + 352, + 460, + 825, + 573 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/fda9989ec537ce17189a8d0151565ce65db1893467693e2932bcd54bd93d57e3.jpg", + "image_caption": [ + "Figure 12: A moment of high influence between the purple influencer and magenta influencee. " + ], + "image_footnote": [], + "bbox": [ + 176, + 464, + 330, + 640 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Table 4 presents the final collective reward obtained by each of the models tested in the experiments presented in Section 4. We see that in several cases, the influence agents are even able to out-perform the state-of-the-art results on these tasks reported by Hughes et al. (2018), despite the fact that the solution proposed by Hughes et al. (2018) requires that agents can view other agents’ rewards, whereas we do not make this assumption, and instead only require that agents can view each others’ actions. ", + "bbox": [ + 352, + 579, + 825, + 676 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "It is important to note that collective reward is not always the perfect metric of cooperative behavior, a finding that was also discovered by Barton et al. (2018) and emphasized by Leibo et al. (2017). In the case, we find that there is a spurious solution to the Harvest game, in which one agent fails to learn and fails to collect any apples. This leads to very high collective reward, since it means there is one fewer agent that can exploit the others, and makes sustainable harvesting easier to achieve. Therefore, for the results shown in the paper, we eliminate any random seed in Harvest for which one of the agents has failed to learn to collect apples, as in previous work (Hughes et al., 2018). ", + "bbox": [ + 351, + 683, + 825, + 726 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 726, + 825, + 795 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "However, here we also present an alternative strategy for assessing the overall collective outcomes: weighting the total collective reward by an index of equality of the individual returns. Specifically, we compute the Gini coefficient over the $N$ agents’ individual returns: ", + "bbox": [ + 174, + 801, + 823, + 843 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/84ca2cb9245eceaa6294138a836191940ef7378fa0c97c845f4e2a3c0ef3313a.jpg", + "text": "$$\nG = \\frac { \\sum _ { i = 1 } ^ { N } \\sum _ { j = 1 } ^ { N } \\left| { r ^ { i } - r ^ { j } } \\right| } { 2 N \\sum _ { i = 1 } ^ { N } r ^ { i } } ,\n$$", + "text_format": "latex", + "bbox": [ + 415, + 848, + 581, + 891 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "which gives us a measure of the inequality of the returns, where $G \\in [ 0 , 1 ]$ , with $G = 0$ indicating perfect equality. Thus, $1 - G$ is a measure of equality; we use this to weight the collective reward for each experiment, and plot the results in Figure 13. Once again, we see that the influence models give the highest final performance, even with this new metric. ", + "bbox": [ + 174, + 895, + 825, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 132 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/c923ae01cca519ebc26562fb33c66fec40b0011ec5f33643088a3be15814576d.jpg", + "image_caption": [ + "Figure 13: Total collective reward times equality, $R * ( 1 - G )$ , obtained in all experiments. Error bars show a $9 9 . 5 \\%$ confidence interval (CI) over 5 random seeds, computed within a sliding window of 200 agent steps. Once again, the models trained with influence reward (red) significantly outperform the baseline and ablated models. " + ], + "image_footnote": [], + "bbox": [ + 187, + 180, + 818, + 744 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Finally, we would like to show that the influence reward is robust to the choice of hyperparameter settings. Therefore, in Figure 14, we plot the collective reward of the top 5 best hyperparameter settings for each experiment, over 5 random seeds each. Once again, the influence models result in higher collective reward, which provides evidence that the model is robust to the choice of hyperparameters. ", + "bbox": [ + 174, + 867, + 826, + 924 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/1408fa6f4ede73df1fdda59c904bd76d77d6315c5152307ac31700974d00769d.jpg", + "image_caption": [ + "Figure 14: Total collective reward over the top 5 hyperparameter settings, with 5 random seeds each, for all experiments. Error bars show a $9 9 . 5 \\%$ confidence interval (CI) computed within a sliding window of 200 agent steps. The influence models still maintain an advantage over the baselines and ablated models, suggesting the technique is robust to the hyperparameter settings. " + ], + "image_footnote": [], + "bbox": [ + 179, + 98, + 818, + 662 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "6.4.1 OPTIMIZING FOR COLLECTIVE REWARD ", + "text_level": 1, + "bbox": [ + 176, + 756, + 488, + 770 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "In this section we include the results of training explicitly prosocial agents, which directly optimize for the collective reward of all agents. Previous work (e.g. Peysakhovich & Lerer (2018)) has shown that training agents to optimize for the rewards of other agents can help the group to obtain better collective outcomes. Following a similar prown individual reward $E ^ { k }$ ple, we implemented agents that optimize foand the collective reward of all other agents, $\\textstyle \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i }$ mbination of their. Thus, the reward function for agent $k$ is $\\begin{array} { r } { { R ^ { k } } = E ^ { k } + \\eta \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i } } \\end{array}$ . We conducted the same hyperparameter search over the parameters mentioned in Section 6.3.1 varying the weight placed on the collective reward, $\\eta \\in [ 0 , 2 ]$ . ", + "bbox": [ + 173, + 780, + 826, + 890 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "As expected, we find that agents trained to optimize for collective reward attain higher collective reward in both Cleanup and Harvest, as is shown in Figure 15. In both games, the optimal value for $\\eta = 0 . 8 5$ . ", + "bbox": [ + 173, + 895, + 825, + 924 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/fe3d4f89e8a19d1b5fe2a4f19403d717b5760772f2debd3fb1d1b879cb6bc09c.jpg", + "image_caption": [ + "Figure 15: Total collective reward obtained by agents trained to optimize for the collective reward, for the 5 best hyperparameter settings with 5 random seeds each. Error bars show a $9 9 . 5 \\%$ confidence interval (CI) computed within a sliding window of 200 agent steps. " + ], + "image_footnote": [], + "bbox": [ + 191, + 97, + 807, + 270 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Interestingly, however, the equality in the individual returns for these agents is extremely low. Across the hyperparameter sweep, no solution to the Cleanup game which scored more than 20 points in terms of collective return was found in which all agents scored an individual return above 0. It seems that in Cleanup, when agents are trained to optimize for collective return, they converge on a solution in which some agents never receive any reward. ", + "bbox": [ + 174, + 351, + 825, + 421 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Note that training agents to optimize for collective reward requires that each agent can view the rewards obtained by other agents. As discussed previously, the social influence reward is a novel way to obtain cooperative behavior, that does not require making this assumption. ", + "bbox": [ + 174, + 428, + 823, + 470 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/parse/train/B1lG42C9Km/B1lG42C9Km_middle.json b/parse/train/B1lG42C9Km/B1lG42C9Km_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..2524a29122eb2e267a9a1f35a78106b835e5a999 --- /dev/null +++ b/parse/train/B1lG42C9Km/B1lG42C9Km_middle.json @@ -0,0 +1,52959 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 427, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 430, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 430, + 98 + ], + "score": 1.0, + "content": "INTRINSIC SOCIAL MOTIVATION VIA CAUSAL", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 345, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 345, + 117 + ], + "score": 1.0, + "content": "INFLUENCE IN MULTI-AGENT RL", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 136, + 237, + 157 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 197, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 197, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 238, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 238, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 279, + 186, + 332, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 334, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 334, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 469, + 377 + ], + "lines": [ + { + "bbox": [ + 141, + 211, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 211, + 469, + 225 + ], + "score": 1.0, + "content": "We derive a new intrinsic social motivation for multi-agent reinforcement learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "score": 1.0, + "content": "(MARL), in which agents are rewarded for having causal influence over another agent’s", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "actions. Causal influence is assessed using counterfactual reasoning. The reward does", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 258 + ], + "score": 1.0, + "content": "not depend on observing another agent’s reward function, and is thus a more realistic", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "score": 1.0, + "content": "approach to MARL than taken in previous work. We show that the causal influence", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "score": 1.0, + "content": "reward is related to maximizing the mutual information between agents’ actions. We test", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "score": 1.0, + "content": "the approach in challenging social dilemma environments, where it consistently leads to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 290, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 470, + 301 + ], + "score": 1.0, + "content": "enhanced cooperation between agents and higher collective reward. Moreover, we find", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 471, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 471, + 312 + ], + "score": 1.0, + "content": "that rewarding influence can lead agents to develop emergent communication protocols.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 310, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 324 + ], + "score": 1.0, + "content": "We therefore employ influence to train agents to use an explicit communication channel,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "score": 1.0, + "content": "and find that it leads to more effective communication and higher collective reward. Fi-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "nally, we show that influence can be computed by equipping each agent with an internal", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 344, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 142, + 344, + 469, + 356 + ], + "score": 1.0, + "content": "model that predicts the actions of other agents. This allows the social influence reward", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 353, + 471, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 471, + 368 + ], + "score": 1.0, + "content": "to be computed without the use of a centralised controller, and as such represents a sig-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 140, + 364, + 471, + 380 + ], + "spans": [ + { + "bbox": [ + 140, + 364, + 471, + 380 + ], + "score": 1.0, + "content": "nificantly more general and scalable inductive bias for MARL with independent agents.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 398, + 202, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 204, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 204, + 412 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) has made impressive progress on specific tasks with well-defined reward", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "functions, but is still difficult to learn intelligent behavior that generalizes across multiple domains. Intrinsic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "motivation is a technique for solving this problem by developing general reward functions that encourage an", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "agent to learn across a variety of tasks (Singh et al., 2004). Previous approaches to intrinsic motivation have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "broadly fallen into two categories: (1) curiosity, or a drive for novelty (e.g. Pathak et al. (2017); Schmidhuber", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "score": 1.0, + "content": "(2010)), and (2) empowerment, or a drive to be able to manipulate the environment (Klyubin et al., 2005).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "We posit that this body of work has largely overlooked an important intrinsic motivation that is key to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "human learning: social interaction. Humans have remarkable social learning abilities; some authors suggest", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "that it is social learning that has given rise to cultural evolution, and allowed us to achieve unprecedented", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "progress and coordination on a massive scale (van Schaik & Burkart, 2011; Herrmann et al., 2007). Others", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 545 + ], + "score": 1.0, + "content": "emphasize that our impressive capacity to learn from others far surpasses that of other animals, apes, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 400, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 400, + 556 + ], + "score": 1.0, + "content": "even other proto-human species (Henrich, 2015; Harari, 2014; Laland, 2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Therefore, we propose an intrinsic reward function designed for multi-agent RL (MARL), which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 586 + ], + "score": 1.0, + "content": "awards agents for having a causal influence on other agents’ actions. Causal influence is assessed using", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "counterfactual reasoning; at each timestep, an agent simulates alternate, counterfactual actions that it could", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "have taken, and assesses their effect on another agent’s behavior. Actions that lead to relatively higher", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "change in the other agent are considered to be highly influential and are rewarded. We show how this", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "reward is related to maximizing the mutual information between agents’ actions, and is thus a form of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "social empowerment. We hypothesize that rewarding influence may therefore encourage cooperation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "between agents. We also take inspiration from experiments in human cognition, showing that newborn", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "infants are sensitive to correspondences between their own actions and the actions of other people, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 660, + 462, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 462, + 672 + ], + "score": 1.0, + "content": "use this to coordinate their behavior with others (Tomasello, 2009; Melis & Semmann, 2010).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "To study our proposed social influence reward in the MARL setting, we adopt the Sequential Social", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Dilemmas (SSDs) of Leibo et al. (2017). These are challenging MA environments with a game-theoretic", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "reward structure, similar to Prisoner’s Dilemma. For each individual agent, ‘defecting’ (non-cooperative", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "behavior) has the highest payoff. However, the collective reward will be better if all agents choose to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "cooperate. The paradoxical payoff structure of these tasks make achieving cooperative social dynamics", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 298, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 427, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 430, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 430, + 98 + ], + "score": 1.0, + "content": "INTRINSIC SOCIAL MOTIVATION VIA CAUSAL", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 345, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 345, + 117 + ], + "score": 1.0, + "content": "INFLUENCE IN MULTI-AGENT RL", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 136, + 237, + 157 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 197, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 197, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 238, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 238, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 136, + 238, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 279, + 186, + 332, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 334, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 334, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 469, + 377 + ], + "lines": [ + { + "bbox": [ + 141, + 211, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 211, + 469, + 225 + ], + "score": 1.0, + "content": "We derive a new intrinsic social motivation for multi-agent reinforcement learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "score": 1.0, + "content": "(MARL), in which agents are rewarded for having causal influence over another agent’s", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "actions. Causal influence is assessed using counterfactual reasoning. The reward does", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 258 + ], + "score": 1.0, + "content": "not depend on observing another agent’s reward function, and is thus a more realistic", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 469, + 268 + ], + "score": 1.0, + "content": "approach to MARL than taken in previous work. We show that the causal influence", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "score": 1.0, + "content": "reward is related to maximizing the mutual information between agents’ actions. We test", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 290 + ], + "score": 1.0, + "content": "the approach in challenging social dilemma environments, where it consistently leads to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 290, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 470, + 301 + ], + "score": 1.0, + "content": "enhanced cooperation between agents and higher collective reward. Moreover, we find", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 471, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 471, + 312 + ], + "score": 1.0, + "content": "that rewarding influence can lead agents to develop emergent communication protocols.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 310, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 324 + ], + "score": 1.0, + "content": "We therefore employ influence to train agents to use an explicit communication channel,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 333 + ], + "score": 1.0, + "content": "and find that it leads to more effective communication and higher collective reward. Fi-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "nally, we show that influence can be computed by equipping each agent with an internal", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 344, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 142, + 344, + 469, + 356 + ], + "score": 1.0, + "content": "model that predicts the actions of other agents. This allows the social influence reward", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 353, + 471, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 471, + 368 + ], + "score": 1.0, + "content": "to be computed without the use of a centralised controller, and as such represents a sig-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 140, + 364, + 471, + 380 + ], + "spans": [ + { + "bbox": [ + 140, + 364, + 471, + 380 + ], + "score": 1.0, + "content": "nificantly more general and scalable inductive bias for MARL with independent agents.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12, + "bbox_fs": [ + 140, + 211, + 471, + 380 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 398, + 202, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 204, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 204, + 412 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) has made impressive progress on specific tasks with well-defined reward", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "functions, but is still difficult to learn intelligent behavior that generalizes across multiple domains. Intrinsic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "motivation is a technique for solving this problem by developing general reward functions that encourage an", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "agent to learn across a variety of tasks (Singh et al., 2004). Previous approaches to intrinsic motivation have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "broadly fallen into two categories: (1) curiosity, or a drive for novelty (e.g. Pathak et al. (2017); Schmidhuber", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "score": 1.0, + "content": "(2010)), and (2) empowerment, or a drive to be able to manipulate the environment (Klyubin et al., 2005).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 417, + 506, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "We posit that this body of work has largely overlooked an important intrinsic motivation that is key to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "human learning: social interaction. Humans have remarkable social learning abilities; some authors suggest", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "that it is social learning that has given rise to cultural evolution, and allowed us to achieve unprecedented", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "progress and coordination on a massive scale (van Schaik & Burkart, 2011; Herrmann et al., 2007). Others", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 545 + ], + "score": 1.0, + "content": "emphasize that our impressive capacity to learn from others far surpasses that of other animals, apes, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 400, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 400, + 556 + ], + "score": 1.0, + "content": "even other proto-human species (Henrich, 2015; Harari, 2014; Laland, 2017).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 488, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Therefore, we propose an intrinsic reward function designed for multi-agent RL (MARL), which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 586 + ], + "score": 1.0, + "content": "awards agents for having a causal influence on other agents’ actions. Causal influence is assessed using", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "counterfactual reasoning; at each timestep, an agent simulates alternate, counterfactual actions that it could", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "have taken, and assesses their effect on another agent’s behavior. Actions that lead to relatively higher", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "change in the other agent are considered to be highly influential and are rewarded. We show how this", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "reward is related to maximizing the mutual information between agents’ actions, and is thus a form of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "social empowerment. We hypothesize that rewarding influence may therefore encourage cooperation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "between agents. We also take inspiration from experiments in human cognition, showing that newborn", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "infants are sensitive to correspondences between their own actions and the actions of other people, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 660, + 462, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 462, + 672 + ], + "score": 1.0, + "content": "use this to coordinate their behavior with others (Tomasello, 2009; Melis & Semmann, 2010).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 561, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "To study our proposed social influence reward in the MARL setting, we adopt the Sequential Social", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Dilemmas (SSDs) of Leibo et al. (2017). These are challenging MA environments with a game-theoretic", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "reward structure, similar to Prisoner’s Dilemma. For each individual agent, ‘defecting’ (non-cooperative", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "behavior) has the highest payoff. However, the collective reward will be better if all agents choose to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "cooperate. The paradoxical payoff structure of these tasks make achieving cooperative social dynamics", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "extremely challenging for typical RL agents. We show that social influence allows agents to learn to", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 378, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 378, + 106 + ], + "score": 1.0, + "content": "cooperate in these environments, and make the following contributions:", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 676, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "extremely challenging for typical RL agents. We show that social influence allows agents to learn to", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 378, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 378, + 106 + ], + "score": 1.0, + "content": "cooperate in these environments, and make the following contributions:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 132, + 114, + 506, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "• We demonstrate that deep RL agents trained with the proposed social influence reward cooperate", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 141, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "to attain higher collective reward than baseline deep RL agents (Mnih et al., 2016). In some cases,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 142, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "this cooperation is attained because influencer agents learn to use their actions as an emergent", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 148, + 459, + 160 + ], + "spans": [ + { + "bbox": [ + 141, + 148, + 459, + 160 + ], + "score": 1.0, + "content": "communication protocol, analogous to behavior seen in animals (von Frisch, 1969).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 165, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 140, + 165, + 505, + 176 + ], + "score": 1.0, + "content": "Motivated by the previous point, we apply the influence reward to training deep RL agents to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 141, + 175, + 505, + 187 + ], + "score": 1.0, + "content": "use an explicit communication channel, as in Foerster et al. (2016). We demonstrate that the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 141, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "communication protocols trained with the influence reward meaningfully relate to agents’ actions,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 197, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 142, + 197, + 506, + 209 + ], + "score": 1.0, + "content": "and that once again, agents trained with the influence reward achieve better collective outcomes.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 213, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 213, + 504, + 225 + ], + "score": 1.0, + "content": "We demonstrate that there is a significant correlation between being influenced through", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 225, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 142, + 225, + 505, + 236 + ], + "score": 1.0, + "content": "communication messages and obtaining higher individual return, suggesting that influential", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 235, + 357, + 246 + ], + "spans": [ + { + "bbox": [ + 142, + 235, + 357, + 246 + ], + "score": 1.0, + "content": "communication is beneficial to the agents that receive it.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 251, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 137, + 251, + 505, + 262 + ], + "score": 1.0, + "content": "Finally, rather than computing social influence using a centralised training framework as in prior", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "work (e.g. Foerster et al. (2017; 2016)), we extend the approach by attaching an internal Model", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "of Other Agents (MOA) network to each agent and training it to predict the actions of every", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 284, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 142, + 284, + 505, + 295 + ], + "score": 1.0, + "content": "other agent. The agent can then simulate counterfactual actions and use its own internal MOA", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 294, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 504, + 306 + ], + "score": 1.0, + "content": "to predict how these will affect other agents, thus computing its own intrinsic influence reward.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "Using a MOA to predict and reward influence allows us to compute an intrinsic social reward by observing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "other agents’ past actions, without a centralised controller, and without requiring access to another agent’s", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "reward function. We believe this is an important innovation over prior work (e.g. (Hughes et al., 2018;", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "Foerster et al., 2017; 2016)). When we consider likely future applications of MARL, such as autonomous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "driving, it becomes apparent that centralised training or the sharing of reward functions are unrealistic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "assumptions, since autonomous vehicles are likely to be produced by a wide variety of organizations and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "score": 1.0, + "content": "institutions with mixed motivations. Rather, a social reward function which only depends on observing the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 507, + 407 + ], + "score": 1.0, + "content": "behavior of agents acting in the environment, and which can give rise to coordinated, cooperative behavior,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 404, + 255, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 255, + 416 + ], + "score": 1.0, + "content": "represents a more promising approach.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 107, + 429, + 176, + 442 + ], + "lines": [ + { + "bbox": [ + 104, + 428, + 178, + 444 + ], + "spans": [ + { + "bbox": [ + 104, + 428, + 178, + 444 + ], + "score": 1.0, + "content": "2 METHODS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 334, + 464 + ], + "score": 1.0, + "content": "We consider a MARL Markov game defined by the tuple", + "type": "text" + }, + { + "bbox": [ + 334, + 451, + 375, + 463 + ], + "score": 0.87, + "content": "\\langle S , T , A , r \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 450, + 505, + 464 + ], + "score": 1.0, + "content": ", in which multiple agents which", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "do not share weights are trained to independently maximize their own individual reward. The environment", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 472, + 504, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 171, + 484 + ], + "score": 1.0, + "content": "state is given by", + "type": "text" + }, + { + "bbox": [ + 171, + 473, + 192, + 483 + ], + "score": 0.89, + "content": "s \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 473, + 263, + 484 + ], + "score": 1.0, + "content": ". At each timestep", + "type": "text" + }, + { + "bbox": [ + 263, + 474, + 268, + 483 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 473, + 315, + 484 + ], + "score": 1.0, + "content": ", each agent", + "type": "text" + }, + { + "bbox": [ + 315, + 473, + 321, + 482 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 473, + 393, + 484 + ], + "score": 1.0, + "content": "chooses an action", + "type": "text" + }, + { + "bbox": [ + 393, + 472, + 420, + 484 + ], + "score": 0.92, + "content": "a _ { t } ^ { k } \\in A", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 473, + 493, + 484 + ], + "score": 1.0, + "content": ". The actions of all", + "type": "text" + }, + { + "bbox": [ + 494, + 473, + 504, + 483 + ], + "score": 0.77, + "content": "N", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 482, + 507, + 498 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 267, + 498 + ], + "score": 1.0, + "content": "agents are combined to form a joint action", + "type": "text" + }, + { + "bbox": [ + 267, + 483, + 324, + 496 + ], + "score": 0.92, + "content": "\\pmb { a } _ { t } = [ a _ { t } ^ { 0 } , . . . a _ { t } ^ { N } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 482, + 507, + 498 + ], + "score": 1.0, + "content": ", which produces a transition in the environment", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 493, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 107, + 495, + 162, + 506 + ], + "score": 0.89, + "content": "T _ { \\cdot } ( s _ { t + 1 } | \\mathbf { a } _ { t } , s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 493, + 328, + 508 + ], + "score": 1.0, + "content": ", according to the state transition function", + "type": "text" + }, + { + "bbox": [ + 329, + 495, + 337, + 505 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 493, + 506, + 508 + ], + "score": 1.0, + "content": ". Each agent then receives its own reward", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 504, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 144, + 518 + ], + "score": 0.89, + "content": "r ^ { k } ( a _ { t } , s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 504, + 507, + 519 + ], + "score": 1.0, + "content": ", which may depend on the actions of other agents. A history of these variables over time is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 101, + 512, + 509, + 537 + ], + "spans": [ + { + "bbox": [ + 101, + 512, + 184, + 537 + ], + "score": 1.0, + "content": "termed a trajectory,", + "type": "text" + }, + { + "bbox": [ + 185, + 516, + 254, + 531 + ], + "score": 0.93, + "content": "{ \\boldsymbol { \\tau } } = \\left\\{ { { s _ { t } } , { a _ { t } } , { r _ { t } } } \\right\\} _ { t = 0 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 512, + 509, + 537 + ], + "score": 1.0, + "content": ". We consider a partially observable setting in which each agent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 529, + 509, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 114, + 541 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 529, + 164, + 568 + ], + "score": 1.0, + "content": "can only vieture reward,", + "type": "text" + }, + { + "bbox": [ + 237, + 529, + 266, + 568 + ], + "score": 1.0, + "content": "true sta, where", + "type": "text" + }, + { + "bbox": [ + 275, + 529, + 278, + 568 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 278, + 530, + 288, + 543 + ], + "score": 0.89, + "content": "s _ { t } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 529, + 471, + 568 + ], + "score": 1.0, + "content": ". Each agent seeks to maximize its own total discount factor. A distributed asynchronous a", + "type": "text" + }, + { + "bbox": [ + 483, + 529, + 509, + 568 + ], + "score": 1.0, + "content": "ectedntage", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 164, + 541, + 274, + 555 + ], + "spans": [ + { + "bbox": [ + 164, + 541, + 236, + 555 + ], + "score": 0.94, + "content": "\\begin{array} { r } { R ^ { k } = \\sum _ { i = 0 } ^ { \\infty } \\gamma ^ { i } r _ { t + i } ^ { k } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 544, + 274, + 554 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 471, + 554, + 483, + 565 + ], + "spans": [ + { + "bbox": [ + 471, + 554, + 483, + 565 + ], + "score": 0.86, + "content": "\\pi ^ { k }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "policy is learned via REINFORCE with baseline (Williams, 1992). Architecturally, our agents consist", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "of a convolutional layer, fully connected layers, a Long Short Term Memory (LSTM) network (Gers", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 269, + 601 + ], + "score": 1.0, + "content": "et al., 1999), and linear layers which output", + "type": "text" + }, + { + "bbox": [ + 269, + 587, + 281, + 598 + ], + "score": 0.88, + "content": "\\pi ^ { \\dot { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 587, + 367, + 601 + ], + "score": 1.0, + "content": "and the value function", + "type": "text" + }, + { + "bbox": [ + 367, + 588, + 396, + 600 + ], + "score": 0.92, + "content": "V ^ { \\pi _ { k } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 587, + 506, + 601 + ], + "score": 1.0, + "content": ". We will refer to the internal", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 103, + 595, + 271, + 614 + ], + "spans": [ + { + "bbox": [ + 103, + 595, + 187, + 614 + ], + "score": 1.0, + "content": "LSTM state of agent", + "type": "text" + }, + { + "bbox": [ + 187, + 599, + 194, + 609 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 595, + 237, + 614 + ], + "score": 1.0, + "content": "at timestep", + "type": "text" + }, + { + "bbox": [ + 238, + 600, + 243, + 609 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 595, + 254, + 614 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 254, + 598, + 265, + 611 + ], + "score": 0.9, + "content": "\\dot { u } _ { t } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 595, + 271, + 614 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 107, + 624, + 361, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 362, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 362, + 636 + ], + "score": 1.0, + "content": "2.1 INTRINSIC SOCIAL MOTIVATION VIA CAUSAL INFLUENCE", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 641, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "Social influence intrinsic motivation modifies an agent’s reward function so that it becomes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 650, + 507, + 666 + ], + "spans": [ + { + "bbox": [ + 107, + 652, + 172, + 664 + ], + "score": 0.9, + "content": "R ^ { k } = \\alpha E ^ { k } + \\beta I ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 650, + 200, + 666 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 201, + 652, + 214, + 663 + ], + "score": 0.89, + "content": "E ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 650, + 381, + 666 + ], + "score": 1.0, + "content": "is the extrinsic or environmental reward, and", + "type": "text" + }, + { + "bbox": [ + 381, + 652, + 392, + 663 + ], + "score": 0.88, + "content": "I ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 650, + 507, + 666 + ], + "score": 1.0, + "content": "is the causal influence reward.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 663, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 155, + 676 + ], + "score": 1.0, + "content": "We compute", + "type": "text" + }, + { + "bbox": [ + 156, + 663, + 167, + 674 + ], + "score": 0.86, + "content": "I ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 663, + 506, + 676 + ], + "score": 1.0, + "content": "by generating counterfactual actions that the agent could have taken at each timestep, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "score": 1.0, + "content": "assessing how taking these would have affected other agents’ behavior. A counterfactual is the estimated", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 165, + 698 + ], + "score": 1.0, + "content": "probability that", + "type": "text" + }, + { + "bbox": [ + 165, + 686, + 178, + 695 + ], + "score": 0.63, + "content": "^ { 6 6 } Y", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 685, + 215, + 698 + ], + "score": 1.0, + "content": "would be", + "type": "text" + }, + { + "bbox": [ + 216, + 687, + 222, + 697 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 685, + 239, + 698 + ], + "score": 1.0, + "content": "had", + "type": "text" + }, + { + "bbox": [ + 239, + 687, + 248, + 695 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 685, + 269, + 698 + ], + "score": 1.0, + "content": "been", + "type": "text" + }, + { + "bbox": [ + 270, + 687, + 276, + 695 + ], + "score": 0.74, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 685, + 322, + 698 + ], + "score": 1.0, + "content": ", in situation", + "type": "text" + }, + { + "bbox": [ + 322, + 686, + 349, + 695 + ], + "score": 0.89, + "content": "Z = z ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 685, + 378, + 698 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 378, + 686, + 397, + 696 + ], + "score": 0.68, + "content": "X , Y ,", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 685, + 415, + 698 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 415, + 686, + 424, + 695 + ], + "score": 0.79, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 685, + 506, + 698 + ], + "score": 1.0, + "content": "are random variables,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 697, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 122, + 709 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 698, + 137, + 708 + ], + "score": 0.8, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 697, + 154, + 709 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 699, + 160, + 707 + ], + "score": 0.76, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "are their values (Pearl et al., 2016). Importantly, it is a counterfactual because we condition", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 708, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 186, + 720 + ], + "score": 1.0, + "content": "on a set of evidence", + "type": "text" + }, + { + "bbox": [ + 187, + 710, + 193, + 717 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 708, + 307, + 720 + ], + "score": 1.0, + "content": ", and because the assignment", + "type": "text" + }, + { + "bbox": [ + 308, + 708, + 334, + 717 + ], + "score": 0.91, + "content": "X = x", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 708, + 505, + 720 + ], + "score": 1.0, + "content": "is counter to what we actually observed; in", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 718, + 246, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 134, + 730 + ], + "score": 1.0, + "content": "reality,", + "type": "text" + }, + { + "bbox": [ + 135, + 719, + 144, + 728 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 718, + 246, + 730 + ], + "score": 1.0, + "content": "took on some other value.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 132, + 114, + 506, + 306 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "• We demonstrate that deep RL agents trained with the proposed social influence reward cooperate", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 141, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "to attain higher collective reward than baseline deep RL agents (Mnih et al., 2016). In some cases,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 142, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "this cooperation is attained because influencer agents learn to use their actions as an emergent", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 148, + 459, + 160 + ], + "spans": [ + { + "bbox": [ + 141, + 148, + 459, + 160 + ], + "score": 1.0, + "content": "communication protocol, analogous to behavior seen in animals (von Frisch, 1969).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 165, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 140, + 165, + 505, + 176 + ], + "score": 1.0, + "content": "Motivated by the previous point, we apply the influence reward to training deep RL agents to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 141, + 175, + 505, + 187 + ], + "score": 1.0, + "content": "use an explicit communication channel, as in Foerster et al. (2016). We demonstrate that the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 141, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "communication protocols trained with the influence reward meaningfully relate to agents’ actions,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 197, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 142, + 197, + 506, + 209 + ], + "score": 1.0, + "content": "and that once again, agents trained with the influence reward achieve better collective outcomes.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 213, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 213, + 504, + 225 + ], + "score": 1.0, + "content": "We demonstrate that there is a significant correlation between being influenced through", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 225, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 142, + 225, + 505, + 236 + ], + "score": 1.0, + "content": "communication messages and obtaining higher individual return, suggesting that influential", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 235, + 357, + 246 + ], + "spans": [ + { + "bbox": [ + 142, + 235, + 357, + 246 + ], + "score": 1.0, + "content": "communication is beneficial to the agents that receive it.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 251, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 137, + 251, + 505, + 262 + ], + "score": 1.0, + "content": "Finally, rather than computing social influence using a centralised training framework as in prior", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "work (e.g. Foerster et al. (2017; 2016)), we extend the approach by attaching an internal Model", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "of Other Agents (MOA) network to each agent and training it to predict the actions of every", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 284, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 142, + 284, + 505, + 295 + ], + "score": 1.0, + "content": "other agent. The agent can then simulate counterfactual actions and use its own internal MOA", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 294, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 504, + 306 + ], + "score": 1.0, + "content": "to predict how these will affect other agents, thus computing its own intrinsic influence reward.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 9.5, + "bbox_fs": [ + 132, + 115, + 506, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "Using a MOA to predict and reward influence allows us to compute an intrinsic social reward by observing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "other agents’ past actions, without a centralised controller, and without requiring access to another agent’s", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "reward function. We believe this is an important innovation over prior work (e.g. (Hughes et al., 2018;", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "Foerster et al., 2017; 2016)). When we consider likely future applications of MARL, such as autonomous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "driving, it becomes apparent that centralised training or the sharing of reward functions are unrealistic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "assumptions, since autonomous vehicles are likely to be produced by a wide variety of organizations and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "score": 1.0, + "content": "institutions with mixed motivations. Rather, a social reward function which only depends on observing the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 507, + 407 + ], + "score": 1.0, + "content": "behavior of agents acting in the environment, and which can give rise to coordinated, cooperative behavior,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 404, + 255, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 255, + 416 + ], + "score": 1.0, + "content": "represents a more promising approach.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 315, + 507, + 416 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 429, + 176, + 442 + ], + "lines": [ + { + "bbox": [ + 104, + 428, + 178, + 444 + ], + "spans": [ + { + "bbox": [ + 104, + 428, + 178, + 444 + ], + "score": 1.0, + "content": "2 METHODS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "list", + "bbox": [ + 106, + 451, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 334, + 464 + ], + "score": 1.0, + "content": "We consider a MARL Markov game defined by the tuple", + "type": "text" + }, + { + "bbox": [ + 334, + 451, + 375, + 463 + ], + "score": 0.87, + "content": "\\langle S , T , A , r \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 450, + 505, + 464 + ], + "score": 1.0, + "content": ", in which multiple agents which", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "do not share weights are trained to independently maximize their own individual reward. The environment", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 472, + 504, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 171, + 484 + ], + "score": 1.0, + "content": "state is given by", + "type": "text" + }, + { + "bbox": [ + 171, + 473, + 192, + 483 + ], + "score": 0.89, + "content": "s \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 473, + 263, + 484 + ], + "score": 1.0, + "content": ". At each timestep", + "type": "text" + }, + { + "bbox": [ + 263, + 474, + 268, + 483 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 473, + 315, + 484 + ], + "score": 1.0, + "content": ", each agent", + "type": "text" + }, + { + "bbox": [ + 315, + 473, + 321, + 482 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 473, + 393, + 484 + ], + "score": 1.0, + "content": "chooses an action", + "type": "text" + }, + { + "bbox": [ + 393, + 472, + 420, + 484 + ], + "score": 0.92, + "content": "a _ { t } ^ { k } \\in A", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 473, + 493, + 484 + ], + "score": 1.0, + "content": ". The actions of all", + "type": "text" + }, + { + "bbox": [ + 494, + 473, + 504, + 483 + ], + "score": 0.77, + "content": "N", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 482, + 507, + 498 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 267, + 498 + ], + "score": 1.0, + "content": "agents are combined to form a joint action", + "type": "text" + }, + { + "bbox": [ + 267, + 483, + 324, + 496 + ], + "score": 0.92, + "content": "\\pmb { a } _ { t } = [ a _ { t } ^ { 0 } , . . . a _ { t } ^ { N } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 482, + 507, + 498 + ], + "score": 1.0, + "content": ", which produces a transition in the environment", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 493, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 107, + 495, + 162, + 506 + ], + "score": 0.89, + "content": "T _ { \\cdot } ( s _ { t + 1 } | \\mathbf { a } _ { t } , s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 493, + 328, + 508 + ], + "score": 1.0, + "content": ", according to the state transition function", + "type": "text" + }, + { + "bbox": [ + 329, + 495, + 337, + 505 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 493, + 506, + 508 + ], + "score": 1.0, + "content": ". Each agent then receives its own reward", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 504, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 144, + 518 + ], + "score": 0.89, + "content": "r ^ { k } ( a _ { t } , s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 504, + 507, + 519 + ], + "score": 1.0, + "content": ", which may depend on the actions of other agents. A history of these variables over time is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 101, + 512, + 509, + 537 + ], + "spans": [ + { + "bbox": [ + 101, + 512, + 184, + 537 + ], + "score": 1.0, + "content": "termed a trajectory,", + "type": "text" + }, + { + "bbox": [ + 185, + 516, + 254, + 531 + ], + "score": 0.93, + "content": "{ \\boldsymbol { \\tau } } = \\left\\{ { { s _ { t } } , { a _ { t } } , { r _ { t } } } \\right\\} _ { t = 0 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 512, + 509, + 537 + ], + "score": 1.0, + "content": ". We consider a partially observable setting in which each agent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 529, + 509, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 114, + 541 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 529, + 164, + 568 + ], + "score": 1.0, + "content": "can only vieture reward,", + "type": "text" + }, + { + "bbox": [ + 237, + 529, + 266, + 568 + ], + "score": 1.0, + "content": "true sta, where", + "type": "text" + }, + { + "bbox": [ + 275, + 529, + 278, + 568 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 278, + 530, + 288, + 543 + ], + "score": 0.89, + "content": "s _ { t } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 529, + 471, + 568 + ], + "score": 1.0, + "content": ". Each agent seeks to maximize its own total discount factor. A distributed asynchronous a", + "type": "text" + }, + { + "bbox": [ + 483, + 529, + 509, + 568 + ], + "score": 1.0, + "content": "ectedntage", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 164, + 541, + 274, + 555 + ], + "spans": [ + { + "bbox": [ + 164, + 541, + 236, + 555 + ], + "score": 0.94, + "content": "\\begin{array} { r } { R ^ { k } = \\sum _ { i = 0 } ^ { \\infty } \\gamma ^ { i } r _ { t + i } ^ { k } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 544, + 274, + 554 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 471, + 554, + 483, + 565 + ], + "spans": [ + { + "bbox": [ + 471, + 554, + 483, + 565 + ], + "score": 0.86, + "content": "\\pi ^ { k }", + "type": "inline_equation" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "policy is learned via REINFORCE with baseline (Williams, 1992). Architecturally, our agents consist", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "of a convolutional layer, fully connected layers, a Long Short Term Memory (LSTM) network (Gers", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 269, + 601 + ], + "score": 1.0, + "content": "et al., 1999), and linear layers which output", + "type": "text" + }, + { + "bbox": [ + 269, + 587, + 281, + 598 + ], + "score": 0.88, + "content": "\\pi ^ { \\dot { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 587, + 367, + 601 + ], + "score": 1.0, + "content": "and the value function", + "type": "text" + }, + { + "bbox": [ + 367, + 588, + 396, + 600 + ], + "score": 0.92, + "content": "V ^ { \\pi _ { k } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 587, + 506, + 601 + ], + "score": 1.0, + "content": ". We will refer to the internal", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 103, + 595, + 271, + 614 + ], + "spans": [ + { + "bbox": [ + 103, + 595, + 187, + 614 + ], + "score": 1.0, + "content": "LSTM state of agent", + "type": "text" + }, + { + "bbox": [ + 187, + 599, + 194, + 609 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 595, + 237, + 614 + ], + "score": 1.0, + "content": "at timestep", + "type": "text" + }, + { + "bbox": [ + 238, + 600, + 243, + 609 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 595, + 254, + 614 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 254, + 598, + 265, + 611 + ], + "score": 0.9, + "content": "\\dot { u } _ { t } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 595, + 271, + 614 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + } + ], + "index": 34.5, + "bbox_fs": [ + 101, + 450, + 509, + 614 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 624, + 361, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 362, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 362, + 636 + ], + "score": 1.0, + "content": "2.1 INTRINSIC SOCIAL MOTIVATION VIA CAUSAL INFLUENCE", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 641, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "Social influence intrinsic motivation modifies an agent’s reward function so that it becomes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 650, + 507, + 666 + ], + "spans": [ + { + "bbox": [ + 107, + 652, + 172, + 664 + ], + "score": 0.9, + "content": "R ^ { k } = \\alpha E ^ { k } + \\beta I ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 650, + 200, + 666 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 201, + 652, + 214, + 663 + ], + "score": 0.89, + "content": "E ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 650, + 381, + 666 + ], + "score": 1.0, + "content": "is the extrinsic or environmental reward, and", + "type": "text" + }, + { + "bbox": [ + 381, + 652, + 392, + 663 + ], + "score": 0.88, + "content": "I ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 650, + 507, + 666 + ], + "score": 1.0, + "content": "is the causal influence reward.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 663, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 155, + 676 + ], + "score": 1.0, + "content": "We compute", + "type": "text" + }, + { + "bbox": [ + 156, + 663, + 167, + 674 + ], + "score": 0.86, + "content": "I ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 663, + 506, + 676 + ], + "score": 1.0, + "content": "by generating counterfactual actions that the agent could have taken at each timestep, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "score": 1.0, + "content": "assessing how taking these would have affected other agents’ behavior. A counterfactual is the estimated", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 165, + 698 + ], + "score": 1.0, + "content": "probability that", + "type": "text" + }, + { + "bbox": [ + 165, + 686, + 178, + 695 + ], + "score": 0.63, + "content": "^ { 6 6 } Y", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 685, + 215, + 698 + ], + "score": 1.0, + "content": "would be", + "type": "text" + }, + { + "bbox": [ + 216, + 687, + 222, + 697 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 685, + 239, + 698 + ], + "score": 1.0, + "content": "had", + "type": "text" + }, + { + "bbox": [ + 239, + 687, + 248, + 695 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 685, + 269, + 698 + ], + "score": 1.0, + "content": "been", + "type": "text" + }, + { + "bbox": [ + 270, + 687, + 276, + 695 + ], + "score": 0.74, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 685, + 322, + 698 + ], + "score": 1.0, + "content": ", in situation", + "type": "text" + }, + { + "bbox": [ + 322, + 686, + 349, + 695 + ], + "score": 0.89, + "content": "Z = z ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 685, + 378, + 698 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 378, + 686, + 397, + 696 + ], + "score": 0.68, + "content": "X , Y ,", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 685, + 415, + 698 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 415, + 686, + 424, + 695 + ], + "score": 0.79, + "content": "Z", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 685, + 506, + 698 + ], + "score": 1.0, + "content": "are random variables,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 697, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 122, + 709 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 698, + 137, + 708 + ], + "score": 0.8, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 697, + 154, + 709 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 699, + 160, + 707 + ], + "score": 0.76, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "are their values (Pearl et al., 2016). Importantly, it is a counterfactual because we condition", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 708, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 186, + 720 + ], + "score": 1.0, + "content": "on a set of evidence", + "type": "text" + }, + { + "bbox": [ + 187, + 710, + 193, + 717 + ], + "score": 0.72, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 708, + 307, + 720 + ], + "score": 1.0, + "content": ", and because the assignment", + "type": "text" + }, + { + "bbox": [ + 308, + 708, + 334, + 717 + ], + "score": 0.91, + "content": "X = x", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 708, + 505, + 720 + ], + "score": 1.0, + "content": "is counter to what we actually observed; in", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 718, + 246, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 134, + 730 + ], + "score": 1.0, + "content": "reality,", + "type": "text" + }, + { + "bbox": [ + 135, + 719, + 144, + 728 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 718, + 246, + 730 + ], + "score": 1.0, + "content": "took on some other value.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 641, + 507, + 730 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 186, + 191 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 186, + 191 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 186, + 191 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 186, + 191 + ], + "score": 0.946, + "type": "image", + "image_path": "d4c9d865d0914f582cec2f90cb9102528669a3d0feda5b6e015295234f191f06.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 186, + 191 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 199, + 187, + 309 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 198, + 189, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 189, + 210 + ], + "score": 1.0, + "content": "Figure 1: Causal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 209, + 189, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 189, + 222 + ], + "score": 1.0, + "content": "diagram of agent", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 220, + 189, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 221, + 115, + 231 + ], + "score": 0.51, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 220, + 171, + 232 + ], + "score": 1.0, + "content": "’s effect on", + "type": "text" + }, + { + "bbox": [ + 171, + 221, + 181, + 231 + ], + "score": 0.62, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 220, + 189, + 232 + ], + "score": 1.0, + "content": "’s", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 231, + 189, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 189, + 243 + ], + "score": 1.0, + "content": "action. We condition", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 243, + 189, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 189, + 254 + ], + "score": 1.0, + "content": "on each agent’s view", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 253, + 189, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 189, + 265 + ], + "score": 1.0, + "content": "of the environment", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 264, + 189, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 189, + 276 + ], + "score": 1.0, + "content": "and LSTM state", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 275, + 190, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 114, + 286 + ], + "score": 0.65, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 275, + 190, + 288 + ], + "score": 1.0, + "content": "(shaded nodes),", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 284, + 187, + 300 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 174, + 300 + ], + "score": 1.0, + "content": "and intervene on", + "type": "text" + }, + { + "bbox": [ + 174, + 286, + 187, + 298 + ], + "score": 0.89, + "content": "a _ { t } ^ { \\dot { A } }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 296, + 135, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 135, + 311 + ], + "score": 1.0, + "content": "(blue).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 195, + 82, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 194, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 194, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "To see how we can compute the causal effect of one agent on another, suppose", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 195, + 94, + 503, + 106 + ], + "spans": [ + { + "bbox": [ + 195, + 94, + 280, + 106 + ], + "score": 1.0, + "content": "there are two agents,", + "type": "text" + }, + { + "bbox": [ + 281, + 94, + 289, + 104 + ], + "score": 0.79, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 94, + 308, + 106 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 94, + 317, + 104 + ], + "score": 0.77, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 94, + 379, + 106 + ], + "score": 1.0, + "content": ", and that agent", + "type": "text" + }, + { + "bbox": [ + 380, + 94, + 389, + 104 + ], + "score": 0.78, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 94, + 424, + 106 + ], + "score": 1.0, + "content": "receives", + "type": "text" + }, + { + "bbox": [ + 425, + 94, + 433, + 104 + ], + "score": 0.67, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 94, + 497, + 106 + ], + "score": 1.0, + "content": "’s action at time", + "type": "text" + }, + { + "bbox": [ + 498, + 95, + 503, + 104 + ], + "score": 0.7, + "content": "t", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 191, + 101, + 509, + 143 + ], + "spans": [ + { + "bbox": [ + 191, + 108, + 225, + 143 + ], + "score": 1.0, + "content": "taction, know a", + "type": "text" + }, + { + "bbox": [ + 208, + 101, + 283, + 118 + ], + "score": 1.0, + "content": ", as input1. Agent", + "type": "text" + }, + { + "bbox": [ + 226, + 114, + 294, + 127 + ], + "score": 0.92, + "content": "p \\big ( a _ { t } ^ { B } | a _ { t } ^ { A } , \\overline { { s _ { t } ^ { B } } } , u _ { t } ^ { B } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 105, + 292, + 114 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 101, + 508, + 118 + ], + "score": 1.0, + "content": "then uses this to compute a distribution over its own", + "type": "text" + }, + { + "bbox": [ + 310, + 108, + 440, + 143 + ], + "score": 1.0, + "content": "ecause we have built the model , and its own internal LSTM state,", + "type": "text" + }, + { + "bbox": [ + 454, + 108, + 477, + 143 + ], + "score": 1.0, + "content": "agent , as sh", + "type": "text" + }, + { + "bbox": [ + 477, + 116, + 486, + 125 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 108, + 509, + 143 + ], + "score": 1.0, + "content": ", wen in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 281, + 126, + 454, + 138 + ], + "spans": [ + { + "bbox": [ + 281, + 126, + 309, + 138 + ], + "score": 0.43, + "content": "\\bar { a } _ { t } ^ { A } , s _ { t } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 126, + 454, + 138 + ], + "score": 0.91, + "content": "u _ { t } ^ { B }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 194, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 194, + 137, + 430, + 150 + ], + "score": 1.0, + "content": "Figure 1. This allows us to exactly isolate the causal effect of", + "type": "text" + }, + { + "bbox": [ + 430, + 138, + 438, + 147 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 137, + 483, + 150 + ], + "score": 1.0, + "content": "’s action on", + "type": "text" + }, + { + "bbox": [ + 483, + 138, + 492, + 147 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "conditioning on the values we observed for the other inputs at this timestep (note", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 194, + 157, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 194, + 157, + 444, + 173 + ], + "score": 1.0, + "content": "that by conditioning on these variables, including the LSTM state", + "type": "text" + }, + { + "bbox": [ + 444, + 159, + 457, + 171 + ], + "score": 0.91, + "content": "u _ { t } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 157, + 506, + 173 + ], + "score": 1.0, + "content": ", we remove", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 194, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 194, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "any dependency on previous timesteps in the trajectory). We can then intervene", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 194, + 180, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 194, + 180, + 208, + 193 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 208, + 181, + 221, + 193 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 180, + 393, + 193 + ], + "score": 1.0, + "content": "by replacing it with a counterfactual action,", + "type": "text" + }, + { + "bbox": [ + 394, + 181, + 422, + 192 + ], + "score": 0.89, + "content": "d o \\dot { ( a _ { t } ^ { A } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 180, + 506, + 193 + ], + "score": 1.0, + "content": ". This counterfactual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 193, + 189, + 507, + 207 + ], + "spans": [ + { + "bbox": [ + 193, + 189, + 359, + 207 + ], + "score": 1.0, + "content": "action is used to compute a new estimate of", + "type": "text" + }, + { + "bbox": [ + 360, + 191, + 443, + 204 + ], + "score": 0.94, + "content": "p ( a _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , s _ { t } ^ { B } , u _ { t } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 189, + 507, + 207 + ], + "score": 1.0, + "content": ". Essentially, the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 194, + 203, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 194, + 203, + 382, + 215 + ], + "score": 1.0, + "content": "agent asks a retrospective question: “How would", + "type": "text" + }, + { + "bbox": [ + 383, + 204, + 392, + 213 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 203, + 505, + 215 + ], + "score": 1.0, + "content": "’s action change if I had acted", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 195, + 213, + 307, + 226 + ], + "spans": [ + { + "bbox": [ + 195, + 213, + 307, + 226 + ], + "score": 1.0, + "content": "differently in this situation?”.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 195, + 230, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 193, + 227, + 507, + 246 + ], + "spans": [ + { + "bbox": [ + 193, + 227, + 289, + 246 + ], + "score": 1.0, + "content": "To simplify notation, let", + "type": "text" + }, + { + "bbox": [ + 290, + 230, + 341, + 243 + ], + "score": 0.93, + "content": "\\boldsymbol { z } _ { t } = \\langle \\boldsymbol { u } _ { t } ^ { B } , \\boldsymbol { s } _ { t } ^ { B } \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 227, + 434, + 246 + ], + "score": 1.0, + "content": ", so that conditioning on", + "type": "text" + }, + { + "bbox": [ + 434, + 233, + 444, + 242 + ], + "score": 0.85, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 227, + 507, + 246 + ], + "score": 1.0, + "content": "is equivalent to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 195, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 195, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "conditioning on all relevant background variables (all shaded variables in Figure 1).", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 194, + 252, + 507, + 266 + ], + "spans": [ + { + "bbox": [ + 194, + 252, + 282, + 266 + ], + "score": 1.0, + "content": "We can also forego the", + "type": "text" + }, + { + "bbox": [ + 282, + 253, + 293, + 263 + ], + "score": 0.59, + "content": "d o", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 252, + 371, + 266 + ], + "score": 1.0, + "content": "operator, noting that", + "type": "text" + }, + { + "bbox": [ + 371, + 252, + 494, + 265 + ], + "score": 0.92, + "content": "p ( \\dot { a } _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , z _ { t } ) \\equiv p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { \\tilde { A } } , z _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 252, + 507, + 266 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 194, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 194, + 263, + 266, + 276 + ], + "score": 1.0, + "content": "this case, because", + "type": "text" + }, + { + "bbox": [ + 266, + 266, + 273, + 274 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "satisfies the back-door criterion (Pearl & Mackenzie, 2018).", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 194, + 273, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 194, + 273, + 391, + 287 + ], + "score": 1.0, + "content": "Now, consider averaging over several counterfactuals", + "type": "text" + }, + { + "bbox": [ + 392, + 275, + 404, + 286 + ], + "score": 0.87, + "content": "\\tilde { a } _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 273, + 506, + 287 + ], + "score": 1.0, + "content": ". This gives us the marginal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 194, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 194, + 285, + 232, + 299 + ], + "score": 1.0, + "content": "policy of", + "type": "text" + }, + { + "bbox": [ + 232, + 286, + 241, + 296 + ], + "score": 0.67, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 285, + 245, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 245, + 285, + 392, + 300 + ], + "score": 0.89, + "content": "\\begin{array} { r } { p ( a _ { t } ^ { B } | \\tilde { z _ { t } } ) = \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , z _ { t } ) p ( \\tilde { a } _ { t } ^ { A } | z _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 285, + 462, + 299 + ], + "score": 1.0, + "content": "—in other words,", + "type": "text" + }, + { + "bbox": [ + 462, + 287, + 471, + 296 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "’s policy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 194, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 194, + 297, + 204, + 311 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 204, + 299, + 212, + 308 + ], + "score": 0.74, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 297, + 478, + 311 + ], + "score": 1.0, + "content": "were not considered. The discrepancy between the marginal policy of", + "type": "text" + }, + { + "bbox": [ + 479, + 300, + 487, + 308 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 195, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 195, + 309, + 288, + 322 + ], + "score": 1.0, + "content": "the conditional policy of", + "type": "text" + }, + { + "bbox": [ + 289, + 310, + 298, + 320 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 309, + 321, + 322 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 321, + 310, + 329, + 320 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "’s action is a measure of the causal influence of", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 105, + 321, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 107, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 107, + 321, + 115, + 330 + ], + "score": 0.75, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 320, + 128, + 333 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 128, + 321, + 137, + 330 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 320, + 249, + 333 + ], + "score": 1.0, + "content": "; it gives the degree to which", + "type": "text" + }, + { + "bbox": [ + 250, + 321, + 259, + 330 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 320, + 453, + 333 + ], + "score": 1.0, + "content": "changes its planned action distribution because of", + "type": "text" + }, + { + "bbox": [ + 453, + 321, + 461, + 330 + ], + "score": 0.83, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "’s behavior.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 331, + 323, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 304, + 344 + ], + "score": 1.0, + "content": "Thus, the causal influence intrinsic reward for agent", + "type": "text" + }, + { + "bbox": [ + 304, + 332, + 312, + 342 + ], + "score": 0.81, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 331, + 323, + 344 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "interline_equation", + "bbox": [ + 134, + 347, + 477, + 378 + ], + "lines": [ + { + "bbox": [ + 134, + 347, + 477, + 378 + ], + "spans": [ + { + "bbox": [ + 134, + 347, + 477, + 378 + ], + "score": 0.92, + "content": "{ I _ { t } ^ { A } } = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } \\mid z _ { t } , \\tilde { a } _ { t } ^ { A } ) p ( \\tilde { a } _ { t } ^ { A } \\mid z _ { t } ) \\right] = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| p ( a _ { t } ^ { B } \\mid z _ { t } ) \\right] .", + "type": "interline_equation", + "image_path": "4c598b57c8ef06c70864098ccc7424fb311f9df5bf10e8fc304f1cd45bcbf3e0.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 134, + 347, + 477, + 357.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 134, + 357.3333333333333, + 477, + 367.66666666666663 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 134, + 367.66666666666663, + 477, + 377.99999999999994 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 389, + 394, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 396, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 396, + 402 + ], + "score": 1.0, + "content": "2.2 RELATIONSHIP TO MUTUAL INFORMATION AND EMPOWERMENT", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 406, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "The causal influence reward in Eq. 1 is related to the mutual information (MI) between the actions of agents", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 415, + 213, + 430 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 115, + 426 + ], + "score": 0.82, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 415, + 132, + 430 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 132, + 417, + 141, + 426 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 415, + 213, + 430 + ], + "score": 1.0, + "content": ", which is given by", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 433, + 484, + 466 + ], + "lines": [ + { + "bbox": [ + 115, + 433, + 484, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 484, + 466 + ], + "score": 0.93, + "content": "I ( A ^ { B } ; A ^ { A } | z ) = \\sum _ { a ^ { A } , a ^ { B } } p ( a ^ { B } , a ^ { A } | z ) \\mathrm { l o g } \\frac { p ( a ^ { B } , a ^ { A } | z ) } { p ( a ^ { B } | z ) p ( a ^ { A } | z ) } { = \\sum _ { a ^ { A } } p ( a ^ { A } | z ) D _ { \\mathrm { K L } } \\left[ p ( a ^ { B } | a ^ { A } , z ) \\right] \\left| p ( a ^ { B } | z ) \\right] } ,", + "type": "interline_equation", + "image_path": "fc2dfdef03318cc35a94ea138af5315729d92b04c7b2a99e379f024a438c3e59.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 115, + 433, + 484, + 444.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 115, + 444.0, + 484, + 455.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 115, + 455.0, + 484, + 466.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 190, + 483 + ], + "score": 1.0, + "content": "where we see that the", + "type": "text" + }, + { + "bbox": [ + 191, + 471, + 212, + 482 + ], + "score": 0.91, + "content": "D _ { K L }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "factor in Eq. 2 is the causal influence reward given in Eq. 1. The connection", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "to mutual information is interesting, because a frequently used intrinsic motivation for single agent RL", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 494, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 504 + ], + "score": 1.0, + "content": "is empowerment, which rewards the agent for having high mutual information between its actions and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "the future state of the environment (e.g. Klyubin et al. (2005); Capdepuy et al. (2007)). To the extent", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 513, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 408, + 528 + ], + "score": 1.0, + "content": "that the social influence reward defined in Eq. 1 is an approximation of the MI,", + "type": "text" + }, + { + "bbox": [ + 408, + 515, + 416, + 524 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 513, + 506, + 528 + ], + "score": 1.0, + "content": "is rewarded for having", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 526, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 182, + 538 + ], + "score": 1.0, + "content": "empowerment over", + "type": "text" + }, + { + "bbox": [ + 182, + 526, + 198, + 536 + ], + "score": 0.88, + "content": "B ^ { \\prime } s", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 526, + 230, + 538 + ], + "score": 1.0, + "content": "actions.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 105, + 542, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 160, + 556 + ], + "score": 1.0, + "content": "By sampling", + "type": "text" + }, + { + "bbox": [ + 161, + 543, + 171, + 552 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 541, + 269, + 556 + ], + "score": 1.0, + "content": "independent trajectories", + "type": "text" + }, + { + "bbox": [ + 269, + 545, + 280, + 554 + ], + "score": 0.85, + "content": "\\tau _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 541, + 448, + 556 + ], + "score": 1.0, + "content": "from the environment, where A’s actions", + "type": "text" + }, + { + "bbox": [ + 448, + 542, + 461, + 555 + ], + "score": 0.9, + "content": "a _ { n } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 541, + 506, + 556 + ], + "score": 1.0, + "content": "are drawn", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 155, + 566 + ], + "score": 1.0, + "content": "according to", + "type": "text" + }, + { + "bbox": [ + 156, + 553, + 187, + 565 + ], + "score": 0.92, + "content": "\\overset { \\cdot } { p } ( a ^ { A } | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 552, + 506, + 566 + ], + "score": 1.0, + "content": ", we perform a Monte-Carlo approximation of the MI (see e.g. Strouse et al. (2018)),", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + }, + { + "type": "interline_equation", + "bbox": [ + 129, + 570, + 481, + 599 + ], + "lines": [ + { + "bbox": [ + 129, + 570, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 129, + 570, + 481, + 599 + ], + "score": 0.94, + "content": "I ( A ^ { A } ; A ^ { B } | z ) = \\mathbb { E } _ { \\tau } \\Big [ D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | A ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] \\Big | z \\Big ] \\approx \\frac { 1 } { N } \\sum _ { n } D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | a _ { n } ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] .", + "type": "interline_equation", + "image_path": "fc5ac4b9ac6bc37a9d7073bded53bdb90fc53b4246ba47725aac71ca7c4aa90e.jpg" + } + ] + } + ], + "index": 51, + "virtual_lines": [ + { + "bbox": [ + 129, + 570, + 481, + 579.6666666666666 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 129, + 579.6666666666666, + 481, + 589.3333333333333 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 129, + 589.3333333333333, + 481, + 598.9999999999999 + ], + "spans": [], + "index": 52 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 603, + 421, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 602, + 421, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 421, + 616 + ], + "score": 1.0, + "content": "Thus, in expectation, the social influence reward is the MI between agents’ actions.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "Whether the policy trained with Eq. 1 actually learns to approximate the MI depends on the learning", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "dynamics. We calculate the intrinsic social influence reward using Eq. 1, because unlike Eq. 2, which", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 642, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 321, + 654 + ], + "score": 1.0, + "content": "gives an estimate of the symmetric bandwidth between", + "type": "text" + }, + { + "bbox": [ + 322, + 642, + 330, + 652 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 642, + 348, + 654 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 348, + 642, + 357, + 652 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 642, + 506, + 654 + ], + "score": 1.0, + "content": ", Eq. 1 gives the directed causal effect", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 55 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 661, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 660, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 119, + 660, + 237, + 674 + ], + "score": 1.0, + "content": "1Note that this requires that agent", + "type": "text" + }, + { + "bbox": [ + 238, + 663, + 245, + 671 + ], + "score": 0.72, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 660, + 330, + 674 + ], + "score": 1.0, + "content": "choose its action before", + "type": "text" + }, + { + "bbox": [ + 331, + 663, + 339, + 670 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 660, + 390, + 674 + ], + "score": 1.0, + "content": ", and therefore", + "type": "text" + }, + { + "bbox": [ + 390, + 662, + 398, + 670 + ], + "score": 0.76, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 660, + 447, + 674 + ], + "score": 1.0, + "content": "can influence", + "type": "text" + }, + { + "bbox": [ + 448, + 663, + 456, + 670 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 660, + 470, + 674 + ], + "score": 1.0, + "content": "but", + "type": "text" + }, + { + "bbox": [ + 470, + 663, + 478, + 670 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 660, + 506, + 674 + ], + "score": 1.0, + "content": "cannot", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 672, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 141, + 684 + ], + "score": 1.0, + "content": "influence", + "type": "text" + }, + { + "bbox": [ + 141, + 672, + 149, + 681 + ], + "score": 0.75, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 672, + 505, + 684 + ], + "score": 1.0, + "content": "; in other words, we must impose a sequential ordering on agents’ actions, and there cannot be mutual", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "influence. We improve upon this approach in Section 2.4. For now, we allow only a fixed number of agents", + "type": "text" + } + ] + }, + { + "bbox": [ + 107, + 691, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 691, + 151, + 702 + ], + "score": 0.9, + "content": "( \\in [ 1 , N - 1 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 691, + 506, + 704 + ], + "score": 1.0, + "content": "to be influencers, and the rest are influencees. Only an influencer gets the causal influence reward, and", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "only an influencee can be influenced. At each timestep, the influencers choose their actions first, and these actions are", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 367, + 723 + ], + "score": 1.0, + "content": "then given as input to the influencees. If agent A and B are influencers, and", + "type": "text" + }, + { + "bbox": [ + 367, + 714, + 374, + 720 + ], + "score": 0.68, + "content": "\\mathrm { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "is an influencee, then C receives both", + "type": "text" + } + ] + }, + { + "bbox": [ + 107, + 720, + 483, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 118, + 732 + ], + "score": 0.89, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 720, + 133, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 134, + 721, + 146, + 732 + ], + "score": 0.9, + "content": "a _ { t } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 720, + 389, + 734 + ], + "score": 1.0, + "content": "as input. When computing the causal influence of A on C, we also add", + "type": "text" + }, + { + "bbox": [ + 390, + 722, + 401, + 732 + ], + "score": 0.91, + "content": "a _ { t } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 720, + 483, + 734 + ], + "score": 1.0, + "content": "to the conditioning set.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 186, + 191 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 186, + 191 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 186, + 191 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 186, + 191 + ], + "score": 0.946, + "type": "image", + "image_path": "d4c9d865d0914f582cec2f90cb9102528669a3d0feda5b6e015295234f191f06.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 186, + 191 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 199, + 187, + 309 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 198, + 189, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 189, + 210 + ], + "score": 1.0, + "content": "Figure 1: Causal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 209, + 189, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 189, + 222 + ], + "score": 1.0, + "content": "diagram of agent", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 220, + 189, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 221, + 115, + 231 + ], + "score": 0.51, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 220, + 171, + 232 + ], + "score": 1.0, + "content": "’s effect on", + "type": "text" + }, + { + "bbox": [ + 171, + 221, + 181, + 231 + ], + "score": 0.62, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 220, + 189, + 232 + ], + "score": 1.0, + "content": "’s", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 231, + 189, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 189, + 243 + ], + "score": 1.0, + "content": "action. We condition", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 243, + 189, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 189, + 254 + ], + "score": 1.0, + "content": "on each agent’s view", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 253, + 189, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 189, + 265 + ], + "score": 1.0, + "content": "of the environment", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 264, + 189, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 189, + 276 + ], + "score": 1.0, + "content": "and LSTM state", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 275, + 190, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 114, + 286 + ], + "score": 0.65, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 275, + 190, + 288 + ], + "score": 1.0, + "content": "(shaded nodes),", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 284, + 187, + 300 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 174, + 300 + ], + "score": 1.0, + "content": "and intervene on", + "type": "text" + }, + { + "bbox": [ + 174, + 286, + 187, + 298 + ], + "score": 0.89, + "content": "a _ { t } ^ { \\dot { A } }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 296, + 135, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 135, + 311 + ], + "score": 1.0, + "content": "(blue).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 195, + 82, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 194, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 194, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "To see how we can compute the causal effect of one agent on another, suppose", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 195, + 94, + 503, + 106 + ], + "spans": [ + { + "bbox": [ + 195, + 94, + 280, + 106 + ], + "score": 1.0, + "content": "there are two agents,", + "type": "text" + }, + { + "bbox": [ + 281, + 94, + 289, + 104 + ], + "score": 0.79, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 94, + 308, + 106 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 94, + 317, + 104 + ], + "score": 0.77, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 94, + 379, + 106 + ], + "score": 1.0, + "content": ", and that agent", + "type": "text" + }, + { + "bbox": [ + 380, + 94, + 389, + 104 + ], + "score": 0.78, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 94, + 424, + 106 + ], + "score": 1.0, + "content": "receives", + "type": "text" + }, + { + "bbox": [ + 425, + 94, + 433, + 104 + ], + "score": 0.67, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 94, + 497, + 106 + ], + "score": 1.0, + "content": "’s action at time", + "type": "text" + }, + { + "bbox": [ + 498, + 95, + 503, + 104 + ], + "score": 0.7, + "content": "t", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 191, + 101, + 509, + 143 + ], + "spans": [ + { + "bbox": [ + 191, + 108, + 225, + 143 + ], + "score": 1.0, + "content": "taction, know a", + "type": "text" + }, + { + "bbox": [ + 208, + 101, + 283, + 118 + ], + "score": 1.0, + "content": ", as input1. Agent", + "type": "text" + }, + { + "bbox": [ + 226, + 114, + 294, + 127 + ], + "score": 0.92, + "content": "p \\big ( a _ { t } ^ { B } | a _ { t } ^ { A } , \\overline { { s _ { t } ^ { B } } } , u _ { t } ^ { B } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 105, + 292, + 114 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 101, + 508, + 118 + ], + "score": 1.0, + "content": "then uses this to compute a distribution over its own", + "type": "text" + }, + { + "bbox": [ + 310, + 108, + 440, + 143 + ], + "score": 1.0, + "content": "ecause we have built the model , and its own internal LSTM state,", + "type": "text" + }, + { + "bbox": [ + 454, + 108, + 477, + 143 + ], + "score": 1.0, + "content": "agent , as sh", + "type": "text" + }, + { + "bbox": [ + 477, + 116, + 486, + 125 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 108, + 509, + 143 + ], + "score": 1.0, + "content": ", wen in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 281, + 126, + 454, + 138 + ], + "spans": [ + { + "bbox": [ + 281, + 126, + 309, + 138 + ], + "score": 0.43, + "content": "\\bar { a } _ { t } ^ { A } , s _ { t } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 126, + 454, + 138 + ], + "score": 0.91, + "content": "u _ { t } ^ { B }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 194, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 194, + 137, + 430, + 150 + ], + "score": 1.0, + "content": "Figure 1. This allows us to exactly isolate the causal effect of", + "type": "text" + }, + { + "bbox": [ + 430, + 138, + 438, + 147 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 137, + 483, + 150 + ], + "score": 1.0, + "content": "’s action on", + "type": "text" + }, + { + "bbox": [ + 483, + 138, + 492, + 147 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "conditioning on the values we observed for the other inputs at this timestep (note", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 194, + 157, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 194, + 157, + 444, + 173 + ], + "score": 1.0, + "content": "that by conditioning on these variables, including the LSTM state", + "type": "text" + }, + { + "bbox": [ + 444, + 159, + 457, + 171 + ], + "score": 0.91, + "content": "u _ { t } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 157, + 506, + 173 + ], + "score": 1.0, + "content": ", we remove", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 194, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 194, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "any dependency on previous timesteps in the trajectory). We can then intervene", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 194, + 180, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 194, + 180, + 208, + 193 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 208, + 181, + 221, + 193 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 180, + 393, + 193 + ], + "score": 1.0, + "content": "by replacing it with a counterfactual action,", + "type": "text" + }, + { + "bbox": [ + 394, + 181, + 422, + 192 + ], + "score": 0.89, + "content": "d o \\dot { ( a _ { t } ^ { A } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 180, + 506, + 193 + ], + "score": 1.0, + "content": ". This counterfactual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 193, + 189, + 507, + 207 + ], + "spans": [ + { + "bbox": [ + 193, + 189, + 359, + 207 + ], + "score": 1.0, + "content": "action is used to compute a new estimate of", + "type": "text" + }, + { + "bbox": [ + 360, + 191, + 443, + 204 + ], + "score": 0.94, + "content": "p ( a _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , s _ { t } ^ { B } , u _ { t } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 189, + 507, + 207 + ], + "score": 1.0, + "content": ". Essentially, the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 194, + 203, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 194, + 203, + 382, + 215 + ], + "score": 1.0, + "content": "agent asks a retrospective question: “How would", + "type": "text" + }, + { + "bbox": [ + 383, + 204, + 392, + 213 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 203, + 505, + 215 + ], + "score": 1.0, + "content": "’s action change if I had acted", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 195, + 213, + 307, + 226 + ], + "spans": [ + { + "bbox": [ + 195, + 213, + 307, + 226 + ], + "score": 1.0, + "content": "differently in this situation?”.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 16.5, + "bbox_fs": [ + 191, + 81, + 509, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 195, + 230, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 193, + 227, + 507, + 246 + ], + "spans": [ + { + "bbox": [ + 193, + 227, + 289, + 246 + ], + "score": 1.0, + "content": "To simplify notation, let", + "type": "text" + }, + { + "bbox": [ + 290, + 230, + 341, + 243 + ], + "score": 0.93, + "content": "\\boldsymbol { z } _ { t } = \\langle \\boldsymbol { u } _ { t } ^ { B } , \\boldsymbol { s } _ { t } ^ { B } \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 227, + 434, + 246 + ], + "score": 1.0, + "content": ", so that conditioning on", + "type": "text" + }, + { + "bbox": [ + 434, + 233, + 444, + 242 + ], + "score": 0.85, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 227, + 507, + 246 + ], + "score": 1.0, + "content": "is equivalent to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 195, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 195, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "conditioning on all relevant background variables (all shaded variables in Figure 1).", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 194, + 252, + 507, + 266 + ], + "spans": [ + { + "bbox": [ + 194, + 252, + 282, + 266 + ], + "score": 1.0, + "content": "We can also forego the", + "type": "text" + }, + { + "bbox": [ + 282, + 253, + 293, + 263 + ], + "score": 0.59, + "content": "d o", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 252, + 371, + 266 + ], + "score": 1.0, + "content": "operator, noting that", + "type": "text" + }, + { + "bbox": [ + 371, + 252, + 494, + 265 + ], + "score": 0.92, + "content": "p ( \\dot { a } _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , z _ { t } ) \\equiv p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { \\tilde { A } } , z _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 252, + 507, + 266 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 194, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 194, + 263, + 266, + 276 + ], + "score": 1.0, + "content": "this case, because", + "type": "text" + }, + { + "bbox": [ + 266, + 266, + 273, + 274 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "satisfies the back-door criterion (Pearl & Mackenzie, 2018).", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 194, + 273, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 194, + 273, + 391, + 287 + ], + "score": 1.0, + "content": "Now, consider averaging over several counterfactuals", + "type": "text" + }, + { + "bbox": [ + 392, + 275, + 404, + 286 + ], + "score": 0.87, + "content": "\\tilde { a } _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 273, + 506, + 287 + ], + "score": 1.0, + "content": ". This gives us the marginal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 194, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 194, + 285, + 232, + 299 + ], + "score": 1.0, + "content": "policy of", + "type": "text" + }, + { + "bbox": [ + 232, + 286, + 241, + 296 + ], + "score": 0.67, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 285, + 245, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 245, + 285, + 392, + 300 + ], + "score": 0.89, + "content": "\\begin{array} { r } { p ( a _ { t } ^ { B } | \\tilde { z _ { t } } ) = \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , z _ { t } ) p ( \\tilde { a } _ { t } ^ { A } | z _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 285, + 462, + 299 + ], + "score": 1.0, + "content": "—in other words,", + "type": "text" + }, + { + "bbox": [ + 462, + 287, + 471, + 296 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "’s policy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 194, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 194, + 297, + 204, + 311 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 204, + 299, + 212, + 308 + ], + "score": 0.74, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 297, + 478, + 311 + ], + "score": 1.0, + "content": "were not considered. The discrepancy between the marginal policy of", + "type": "text" + }, + { + "bbox": [ + 479, + 300, + 487, + 308 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 195, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 195, + 309, + 288, + 322 + ], + "score": 1.0, + "content": "the conditional policy of", + "type": "text" + }, + { + "bbox": [ + 289, + 310, + 298, + 320 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 309, + 321, + 322 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 321, + 310, + 329, + 320 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "’s action is a measure of the causal influence of", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 193, + 227, + 507, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 321, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 107, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 107, + 321, + 115, + 330 + ], + "score": 0.75, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 320, + 128, + 333 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 128, + 321, + 137, + 330 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 320, + 249, + 333 + ], + "score": 1.0, + "content": "; it gives the degree to which", + "type": "text" + }, + { + "bbox": [ + 250, + 321, + 259, + 330 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 320, + 453, + 333 + ], + "score": 1.0, + "content": "changes its planned action distribution because of", + "type": "text" + }, + { + "bbox": [ + 453, + 321, + 461, + 330 + ], + "score": 0.83, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "’s behavior.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 331, + 323, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 304, + 344 + ], + "score": 1.0, + "content": "Thus, the causal influence intrinsic reward for agent", + "type": "text" + }, + { + "bbox": [ + 304, + 332, + 312, + 342 + ], + "score": 0.81, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 331, + 323, + 344 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 320, + 506, + 344 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 134, + 347, + 477, + 378 + ], + "lines": [ + { + "bbox": [ + 134, + 347, + 477, + 378 + ], + "spans": [ + { + "bbox": [ + 134, + 347, + 477, + 378 + ], + "score": 0.92, + "content": "{ I _ { t } ^ { A } } = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } \\mid z _ { t } , \\tilde { a } _ { t } ^ { A } ) p ( \\tilde { a } _ { t } ^ { A } \\mid z _ { t } ) \\right] = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| p ( a _ { t } ^ { B } \\mid z _ { t } ) \\right] .", + "type": "interline_equation", + "image_path": "4c598b57c8ef06c70864098ccc7424fb311f9df5bf10e8fc304f1cd45bcbf3e0.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 134, + 347, + 477, + 357.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 134, + 357.3333333333333, + 477, + 367.66666666666663 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 134, + 367.66666666666663, + 477, + 377.99999999999994 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 389, + 394, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 396, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 396, + 402 + ], + "score": 1.0, + "content": "2.2 RELATIONSHIP TO MUTUAL INFORMATION AND EMPOWERMENT", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 406, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "The causal influence reward in Eq. 1 is related to the mutual information (MI) between the actions of agents", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 415, + 213, + 430 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 115, + 426 + ], + "score": 0.82, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 415, + 132, + 430 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 132, + 417, + 141, + 426 + ], + "score": 0.84, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 415, + 213, + 430 + ], + "score": 1.0, + "content": ", which is given by", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 405, + 505, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 433, + 484, + 466 + ], + "lines": [ + { + "bbox": [ + 115, + 433, + 484, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 484, + 466 + ], + "score": 0.93, + "content": "I ( A ^ { B } ; A ^ { A } | z ) = \\sum _ { a ^ { A } , a ^ { B } } p ( a ^ { B } , a ^ { A } | z ) \\mathrm { l o g } \\frac { p ( a ^ { B } , a ^ { A } | z ) } { p ( a ^ { B } | z ) p ( a ^ { A } | z ) } { = \\sum _ { a ^ { A } } p ( a ^ { A } | z ) D _ { \\mathrm { K L } } \\left[ p ( a ^ { B } | a ^ { A } , z ) \\right] \\left| p ( a ^ { B } | z ) \\right] } ,", + "type": "interline_equation", + "image_path": "fc2dfdef03318cc35a94ea138af5315729d92b04c7b2a99e379f024a438c3e59.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 115, + 433, + 484, + 444.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 115, + 444.0, + 484, + 455.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 115, + 455.0, + 484, + 466.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 190, + 483 + ], + "score": 1.0, + "content": "where we see that the", + "type": "text" + }, + { + "bbox": [ + 191, + 471, + 212, + 482 + ], + "score": 0.91, + "content": "D _ { K L }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "factor in Eq. 2 is the causal influence reward given in Eq. 1. The connection", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "to mutual information is interesting, because a frequently used intrinsic motivation for single agent RL", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 494, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 504 + ], + "score": 1.0, + "content": "is empowerment, which rewards the agent for having high mutual information between its actions and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "the future state of the environment (e.g. Klyubin et al. (2005); Capdepuy et al. (2007)). To the extent", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 513, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 408, + 528 + ], + "score": 1.0, + "content": "that the social influence reward defined in Eq. 1 is an approximation of the MI,", + "type": "text" + }, + { + "bbox": [ + 408, + 515, + 416, + 524 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 513, + 506, + 528 + ], + "score": 1.0, + "content": "is rewarded for having", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 526, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 182, + 538 + ], + "score": 1.0, + "content": "empowerment over", + "type": "text" + }, + { + "bbox": [ + 182, + 526, + 198, + 536 + ], + "score": 0.88, + "content": "B ^ { \\prime } s", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 526, + 230, + 538 + ], + "score": 1.0, + "content": "actions.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 471, + 506, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 542, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 160, + 556 + ], + "score": 1.0, + "content": "By sampling", + "type": "text" + }, + { + "bbox": [ + 161, + 543, + 171, + 552 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 541, + 269, + 556 + ], + "score": 1.0, + "content": "independent trajectories", + "type": "text" + }, + { + "bbox": [ + 269, + 545, + 280, + 554 + ], + "score": 0.85, + "content": "\\tau _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 541, + 448, + 556 + ], + "score": 1.0, + "content": "from the environment, where A’s actions", + "type": "text" + }, + { + "bbox": [ + 448, + 542, + 461, + 555 + ], + "score": 0.9, + "content": "a _ { n } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 541, + 506, + 556 + ], + "score": 1.0, + "content": "are drawn", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 155, + 566 + ], + "score": 1.0, + "content": "according to", + "type": "text" + }, + { + "bbox": [ + 156, + 553, + 187, + 565 + ], + "score": 0.92, + "content": "\\overset { \\cdot } { p } ( a ^ { A } | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 552, + 506, + 566 + ], + "score": 1.0, + "content": ", we perform a Monte-Carlo approximation of the MI (see e.g. Strouse et al. (2018)),", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 541, + 506, + 566 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 129, + 570, + 481, + 599 + ], + "lines": [ + { + "bbox": [ + 129, + 570, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 129, + 570, + 481, + 599 + ], + "score": 0.94, + "content": "I ( A ^ { A } ; A ^ { B } | z ) = \\mathbb { E } _ { \\tau } \\Big [ D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | A ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] \\Big | z \\Big ] \\approx \\frac { 1 } { N } \\sum _ { n } D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | a _ { n } ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] .", + "type": "interline_equation", + "image_path": "fc5ac4b9ac6bc37a9d7073bded53bdb90fc53b4246ba47725aac71ca7c4aa90e.jpg" + } + ] + } + ], + "index": 51, + "virtual_lines": [ + { + "bbox": [ + 129, + 570, + 481, + 579.6666666666666 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 129, + 579.6666666666666, + 481, + 589.3333333333333 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 129, + 589.3333333333333, + 481, + 598.9999999999999 + ], + "spans": [], + "index": 52 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 603, + 421, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 602, + 421, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 421, + 616 + ], + "score": 1.0, + "content": "Thus, in expectation, the social influence reward is the MI between agents’ actions.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53, + "bbox_fs": [ + 106, + 602, + 421, + 616 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "Whether the policy trained with Eq. 1 actually learns to approximate the MI depends on the learning", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "dynamics. We calculate the intrinsic social influence reward using Eq. 1, because unlike Eq. 2, which", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 642, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 321, + 654 + ], + "score": 1.0, + "content": "gives an estimate of the symmetric bandwidth between", + "type": "text" + }, + { + "bbox": [ + 322, + 642, + 330, + 652 + ], + "score": 0.8, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 642, + 348, + 654 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 348, + 642, + 357, + 652 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 642, + 506, + 654 + ], + "score": 1.0, + "content": ", Eq. 1 gives the directed causal effect", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 246, + 96 + ], + "score": 1.0, + "content": "of the specific action taken by agent", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 246, + 83, + 255, + 93 + ], + "score": 0.61, + "content": "A", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 255, + 81, + 258, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 258, + 82, + 270, + 95 + ], + "score": 0.84, + "content": "a _ { t } ^ { A }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 271, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ". We believe this will result in an easier reward to learn, since", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 279, + 105 + ], + "score": 1.0, + "content": "it allows for better credit assignment; agent", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 279, + 94, + 288, + 104 + ], + "score": 0.69, + "content": "A", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 288, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "can more easily learn which of its actions lead to high", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "influence. We also experiment with replacing the KL-divergence with several other measures, including the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 502, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 502, + 127 + ], + "score": 1.0, + "content": "Jensen-Shannon Divergence (JSD), and find that the influence reward is robust to the choice of measure.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 55, + "bbox_fs": [ + 105, + 619, + 506, + 654 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 246, + 96 + ], + "score": 1.0, + "content": "of the specific action taken by agent", + "type": "text" + }, + { + "bbox": [ + 246, + 83, + 255, + 93 + ], + "score": 0.61, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 81, + 258, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 258, + 82, + 270, + 95 + ], + "score": 0.84, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ". We believe this will result in an easier reward to learn, since", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 279, + 105 + ], + "score": 1.0, + "content": "it allows for better credit assignment; agent", + "type": "text" + }, + { + "bbox": [ + 279, + 94, + 288, + 104 + ], + "score": 0.69, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "can more easily learn which of its actions lead to high", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "influence. We also experiment with replacing the KL-divergence with several other measures, including the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 502, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 502, + 127 + ], + "score": 1.0, + "content": "Jensen-Shannon Divergence (JSD), and find that the influence reward is robust to the choice of measure.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 108, + 138, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 137, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 297, + 150 + ], + "score": 1.0, + "content": "2.3 INFLUENCE THROUGH COMMUNICATION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 165 + ], + "score": 1.0, + "content": "According to Melis & Semmann (2010), human children rapidly learn to use communication to influence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "the behavior of others when engaging in cooperative activities. They explain that “this ability to influence", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "the partner via communication has been interpreted as evidence for a capacity to form shared goals with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "others”, and that this capacity may be “what allows humans to engage in a wide range of cooperative", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 211 + ], + "score": 1.0, + "content": "activities”. Therefore, we investigate a second use of the social influence reward: learning inter-agent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "score": 1.0, + "content": "communication protocols. Using a similar approach to Reinforced Inter-Agent Learning (RIAL) (Foerster", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 220, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 497, + 232 + ], + "score": 1.0, + "content": "et al., 2016), we equip the agents with an explicit communication channel. At each timestep, each agent", + "type": "text" + }, + { + "bbox": [ + 497, + 220, + 504, + 229 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 228, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 268, + 244 + ], + "score": 1.0, + "content": "chooses a discrete communication symbol", + "type": "text" + }, + { + "bbox": [ + 268, + 230, + 282, + 242 + ], + "score": 0.9, + "content": "\\Dot { m } _ { t } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 228, + 505, + 244 + ], + "score": 1.0, + "content": "; these symbols are concatenated into a combined message", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 132, + 255 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 132, + 240, + 213, + 253 + ], + "score": 0.93, + "content": "\\pmb { m } _ { t } = [ m _ { t } ^ { 0 } , m _ { t } ^ { 1 } . . . m _ { t } ^ { N } ] ,", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 241, + 229, + 255 + ], + "score": 1.0, + "content": ", for", + "type": "text" + }, + { + "bbox": [ + 230, + 241, + 240, + 251 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 241, + 348, + 255 + ], + "score": 1.0, + "content": "agents. This message vector", + "type": "text" + }, + { + "bbox": [ + 349, + 243, + 363, + 252 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } m _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "is then shown to every other agent in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "the next timestep, as in Figure 2. To train the agents to communicate, we augment our initial network with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 262, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 367, + 277 + ], + "score": 1.0, + "content": "an additional A3C output head, that learns a communication policy", + "type": "text" + }, + { + "bbox": [ + 368, + 265, + 379, + 274 + ], + "score": 0.84, + "content": "\\pi _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 262, + 506, + 277 + ], + "score": 1.0, + "content": "over which symbol to emit, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 232, + 287 + ], + "score": 1.0, + "content": "a communication value function", + "type": "text" + }, + { + "bbox": [ + 233, + 274, + 244, + 285 + ], + "score": 0.86, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "(this is separate from the normal policy and value function used for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 284, + 455, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 207, + 297 + ], + "score": 1.0, + "content": "acting in the environment,", + "type": "text" + }, + { + "bbox": [ + 207, + 286, + 218, + 296 + ], + "score": 0.86, + "content": "\\pi _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 284, + 235, + 297 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 235, + 286, + 246, + 296 + ], + "score": 0.88, + "content": "V _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 284, + 440, + 297 + ], + "score": 1.0, + "content": ", which are trained only with environmental reward", + "type": "text" + }, + { + "bbox": [ + 441, + 285, + 449, + 295 + ], + "score": 0.77, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 284, + 455, + 297 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 503, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 492, + 316 + ], + "score": 1.0, + "content": "The influence reward is used, in addition to environmental reward, to train the communication policy", + "type": "text" + }, + { + "bbox": [ + 493, + 304, + 503, + 313 + ], + "score": 0.81, + "content": "\\pi _ { c }", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 102, + 311, + 509, + 342 + ], + "spans": [ + { + "bbox": [ + 102, + 311, + 288, + 342 + ], + "score": 1.0, + "content": "Counterfactuals are employed to assess how muon another agent’s action in the next timestep,", + "type": "text" + }, + { + "bbox": [ + 308, + 311, + 470, + 342 + ], + "score": 1.0, + "content": "luence an agent’s communication message, . Importantly, we hypothesize that comm", + "type": "text" + }, + { + "bbox": [ + 471, + 313, + 486, + 325 + ], + "score": 0.91, + "content": "m _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 311, + 509, + 342 + ], + "score": 1.0, + "content": ", hastion", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 288, + 323, + 307, + 336 + ], + "spans": [ + { + "bbox": [ + 288, + 323, + 307, + 336 + ], + "score": 0.92, + "content": "a _ { t + 1 } ^ { B }", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 447, + 348 + ], + "score": 1.0, + "content": "can only be influential if it is useful to another agent. There is nothing that compels agent", + "type": "text" + }, + { + "bbox": [ + 457, + 334, + 506, + 348 + ], + "score": 1.0, + "content": "to act based", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 143, + 359 + ], + "score": 1.0, + "content": "on agent", + "type": "text" + }, + { + "bbox": [ + 143, + 346, + 151, + 356 + ], + "score": 0.66, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 345, + 429, + 359 + ], + "score": 1.0, + "content": "’s communication message; if it does not contain valuable information,", + "type": "text" + }, + { + "bbox": [ + 430, + 346, + 439, + 356 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "is free to ignore", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "it. In fact, previous work has shown that selfish agents do not learn to use this type of ungrounded,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 368, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 415, + 380 + ], + "score": 1.0, + "content": "cheap talk communication channel effectively (Cao et al., 2018). In contrast, for", + "type": "text" + }, + { + "bbox": [ + 415, + 368, + 424, + 378 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 368, + 506, + 380 + ], + "score": 1.0, + "content": "to gain influence via", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 171, + 392 + ], + "score": 1.0, + "content": "communication,", + "type": "text" + }, + { + "bbox": [ + 171, + 378, + 187, + 390 + ], + "score": 0.92, + "content": "m _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 379, + 372, + 392 + ], + "score": 1.0, + "content": "must contain valuable information that informs", + "type": "text" + }, + { + "bbox": [ + 372, + 379, + 381, + 389 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "about how best to maximize its", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 390, + 406, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 283, + 402 + ], + "score": 1.0, + "content": "own reward, so much so that it actually causes", + "type": "text" + }, + { + "bbox": [ + 284, + 390, + 293, + 399 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 390, + 406, + 402 + ], + "score": 1.0, + "content": "to change its intended action.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 107, + 414, + 287, + 516 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 414, + 287, + 516 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 414, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 107, + 414, + 287, + 516 + ], + "score": 0.962, + "type": "image", + "image_path": "caa2b9dc3c9fe0bdc22237d63bfba2b09073071b89e2a88a719924f5aa8e5a38.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 107, + 414, + 287, + 428.57142857142856 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 107, + 428.57142857142856, + 287, + 443.1428571428571 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 107, + 443.1428571428571, + 287, + 457.71428571428567 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 107, + 457.71428571428567, + 287, + 472.2857142857142 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 107, + 472.2857142857142, + 287, + 486.8571428571428 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 107, + 486.8571428571428, + 287, + 501.42857142857133 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 107, + 501.42857142857133, + 287, + 515.9999999999999 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 525, + 288, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 525, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 287, + 537 + ], + "score": 1.0, + "content": "Figure 2: The communication model has two", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 536, + 288, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 257, + 549 + ], + "score": 1.0, + "content": "A2C heads, which learn a normal policy,", + "type": "text" + }, + { + "bbox": [ + 257, + 538, + 268, + 547 + ], + "score": 0.85, + "content": "\\pi _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 536, + 288, + 549 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 547, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 288, + 559 + ], + "score": 1.0, + "content": "a policy for emitting communication symbols,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 557, + 286, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 117, + 569 + ], + "score": 0.84, + "content": "\\pi _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 557, + 271, + 572 + ], + "score": 1.0, + "content": ". Other agents’ communication messages", + "type": "text" + }, + { + "bbox": [ + 271, + 560, + 286, + 569 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } m _ { t }", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 570, + 196, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 196, + 581 + ], + "score": 1.0, + "content": "are input to the LSTM.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + } + ], + "index": 36.5 + }, + { + "type": "image", + "bbox": [ + 305, + 412, + 500, + 514 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 305, + 412, + 500, + 514 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 305, + 412, + 500, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 500, + 514 + ], + "score": 0.964, + "type": "image", + "image_path": "64e00c7a40f5d8bfca22499a480991f9780f7757dc5984363b283b0279a25fc0.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 305, + 412, + 500, + 426.57142857142856 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 305, + 426.57142857142856, + 500, + 441.1428571428571 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 305, + 441.1428571428571, + 500, + 455.71428571428567 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 305, + 455.71428571428567, + 500, + 470.2857142857142 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 305, + 470.2857142857142, + 500, + 484.8571428571428 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 305, + 484.8571428571428, + 500, + 499.42857142857133 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 305, + 499.42857142857133, + 500, + 513.9999999999999 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 301, + 525, + 503, + 580 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 302, + 525, + 503, + 537 + ], + "spans": [ + { + "bbox": [ + 302, + 525, + 503, + 537 + ], + "score": 1.0, + "content": "Figure 3: The Model of Other Agents (MOA) archi-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 301, + 537, + 503, + 549 + ], + "spans": [ + { + "bbox": [ + 301, + 537, + 423, + 549 + ], + "score": 1.0, + "content": "tecture learns both an RL policy", + "type": "text" + }, + { + "bbox": [ + 424, + 538, + 434, + 547 + ], + "score": 0.84, + "content": "\\pi _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 537, + 503, + 549 + ], + "score": 1.0, + "content": ", and a supervised", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 300, + 546, + 500, + 561 + ], + "spans": [ + { + "bbox": [ + 300, + 546, + 480, + 561 + ], + "score": 1.0, + "content": "model that predicts the actions of other agents,", + "type": "text" + }, + { + "bbox": [ + 480, + 549, + 500, + 559 + ], + "score": 0.88, + "content": "\\mathbf { \\pmb { a } } _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 302, + 559, + 503, + 570 + ], + "spans": [ + { + "bbox": [ + 302, + 559, + 503, + 570 + ], + "score": 1.0, + "content": "The predictions of the supervised model are used for", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 302, + 570, + 426, + 581 + ], + "spans": [ + { + "bbox": [ + 302, + 570, + 426, + 581 + ], + "score": 1.0, + "content": "computing the influence reward.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "index": 42.5 + }, + { + "type": "title", + "bbox": [ + 109, + 590, + 307, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 309, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 309, + 601 + ], + "score": 1.0, + "content": "2.4 INFLUENCE VIA MODELING OTHER AGENTS", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "Computing the causal influence reward as introduced in Section 2.1 requires knowing the probability of", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 107, + 612, + 508, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 617, + 116, + 626 + ], + "score": 0.71, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 612, + 255, + 633 + ], + "score": 1.0, + "content": "’s next action given a counterfactual,", + "type": "text" + }, + { + "bbox": [ + 256, + 615, + 308, + 628 + ], + "score": 0.93, + "content": "p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , s _ { t } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 612, + 508, + 633 + ], + "score": 1.0, + "content": ", which we previously solved by using a centralised", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "controller that could access other agent’s policy networks. While using a centralised training framework is", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "common in MARL (e.g. Foerster et al. (2017; 2016)), it is less realistic than a scenario in which each agent", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 104, + 647, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 647, + 505, + 663 + ], + "score": 1.0, + "content": "is trained independently. We can relax this assumption and achieve independent training by equipping", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "each agent with its own internal Model of Other Agents (MOA). The MOA consists of a second set of", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "fully-connected and LSTM layers connected to the agent’s convolutional layer (see Figure 3), and is trained", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "score": 1.0, + "content": "to predict all other agents’ next actions given their current action, and the agent’s egocentric view of the", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 692, + 489, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 129, + 706 + ], + "score": 1.0, + "content": "state:", + "type": "text" + }, + { + "bbox": [ + 129, + 692, + 186, + 705 + ], + "score": 0.93, + "content": "p ( { \\pmb a } _ { t + 1 } | { \\pmb a } _ { t } , { \\pmb s } _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 693, + 489, + 706 + ], + "score": 1.0, + "content": ". The MOA is trained using cross-entropy loss over observed action trajectories.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 56 + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "A trained MOA can be used to compute the social influence reward in the following way. Each agent can", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "“imagine” counterfactual actions that it could have taken at each timestep, and use its internal MOA to", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 61.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 126 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 81, + 506, + 127 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 138, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 137, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 297, + 150 + ], + "score": 1.0, + "content": "2.3 INFLUENCE THROUGH COMMUNICATION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 165 + ], + "score": 1.0, + "content": "According to Melis & Semmann (2010), human children rapidly learn to use communication to influence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "the behavior of others when engaging in cooperative activities. They explain that “this ability to influence", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "the partner via communication has been interpreted as evidence for a capacity to form shared goals with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "others”, and that this capacity may be “what allows humans to engage in a wide range of cooperative", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 211 + ], + "score": 1.0, + "content": "activities”. Therefore, we investigate a second use of the social influence reward: learning inter-agent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "score": 1.0, + "content": "communication protocols. Using a similar approach to Reinforced Inter-Agent Learning (RIAL) (Foerster", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 220, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 497, + 232 + ], + "score": 1.0, + "content": "et al., 2016), we equip the agents with an explicit communication channel. At each timestep, each agent", + "type": "text" + }, + { + "bbox": [ + 497, + 220, + 504, + 229 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 228, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 268, + 244 + ], + "score": 1.0, + "content": "chooses a discrete communication symbol", + "type": "text" + }, + { + "bbox": [ + 268, + 230, + 282, + 242 + ], + "score": 0.9, + "content": "\\Dot { m } _ { t } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 228, + 505, + 244 + ], + "score": 1.0, + "content": "; these symbols are concatenated into a combined message", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 132, + 255 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 132, + 240, + 213, + 253 + ], + "score": 0.93, + "content": "\\pmb { m } _ { t } = [ m _ { t } ^ { 0 } , m _ { t } ^ { 1 } . . . m _ { t } ^ { N } ] ,", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 241, + 229, + 255 + ], + "score": 1.0, + "content": ", for", + "type": "text" + }, + { + "bbox": [ + 230, + 241, + 240, + 251 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 241, + 348, + 255 + ], + "score": 1.0, + "content": "agents. This message vector", + "type": "text" + }, + { + "bbox": [ + 349, + 243, + 363, + 252 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } m _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "is then shown to every other agent in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "the next timestep, as in Figure 2. To train the agents to communicate, we augment our initial network with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 262, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 367, + 277 + ], + "score": 1.0, + "content": "an additional A3C output head, that learns a communication policy", + "type": "text" + }, + { + "bbox": [ + 368, + 265, + 379, + 274 + ], + "score": 0.84, + "content": "\\pi _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 262, + 506, + 277 + ], + "score": 1.0, + "content": "over which symbol to emit, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 232, + 287 + ], + "score": 1.0, + "content": "a communication value function", + "type": "text" + }, + { + "bbox": [ + 233, + 274, + 244, + 285 + ], + "score": 0.86, + "content": "V _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "(this is separate from the normal policy and value function used for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 284, + 455, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 207, + 297 + ], + "score": 1.0, + "content": "acting in the environment,", + "type": "text" + }, + { + "bbox": [ + 207, + 286, + 218, + 296 + ], + "score": 0.86, + "content": "\\pi _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 284, + 235, + 297 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 235, + 286, + 246, + 296 + ], + "score": 0.88, + "content": "V _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 284, + 440, + 297 + ], + "score": 1.0, + "content": ", which are trained only with environmental reward", + "type": "text" + }, + { + "bbox": [ + 441, + 285, + 449, + 295 + ], + "score": 0.77, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 284, + 455, + 297 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 154, + 506, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 503, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 492, + 316 + ], + "score": 1.0, + "content": "The influence reward is used, in addition to environmental reward, to train the communication policy", + "type": "text" + }, + { + "bbox": [ + 493, + 304, + 503, + 313 + ], + "score": 0.81, + "content": "\\pi _ { c }", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 102, + 311, + 509, + 342 + ], + "spans": [ + { + "bbox": [ + 102, + 311, + 288, + 342 + ], + "score": 1.0, + "content": "Counterfactuals are employed to assess how muon another agent’s action in the next timestep,", + "type": "text" + }, + { + "bbox": [ + 308, + 311, + 470, + 342 + ], + "score": 1.0, + "content": "luence an agent’s communication message, . Importantly, we hypothesize that comm", + "type": "text" + }, + { + "bbox": [ + 471, + 313, + 486, + 325 + ], + "score": 0.91, + "content": "m _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 311, + 509, + 342 + ], + "score": 1.0, + "content": ", hastion", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 288, + 323, + 307, + 336 + ], + "spans": [ + { + "bbox": [ + 288, + 323, + 307, + 336 + ], + "score": 0.92, + "content": "a _ { t + 1 } ^ { B }", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 447, + 348 + ], + "score": 1.0, + "content": "can only be influential if it is useful to another agent. There is nothing that compels agent", + "type": "text" + }, + { + "bbox": [ + 457, + 334, + 506, + 348 + ], + "score": 1.0, + "content": "to act based", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 143, + 359 + ], + "score": 1.0, + "content": "on agent", + "type": "text" + }, + { + "bbox": [ + 143, + 346, + 151, + 356 + ], + "score": 0.66, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 345, + 429, + 359 + ], + "score": 1.0, + "content": "’s communication message; if it does not contain valuable information,", + "type": "text" + }, + { + "bbox": [ + 430, + 346, + 439, + 356 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "is free to ignore", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "it. In fact, previous work has shown that selfish agents do not learn to use this type of ungrounded,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 368, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 415, + 380 + ], + "score": 1.0, + "content": "cheap talk communication channel effectively (Cao et al., 2018). In contrast, for", + "type": "text" + }, + { + "bbox": [ + 415, + 368, + 424, + 378 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 368, + 506, + 380 + ], + "score": 1.0, + "content": "to gain influence via", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 171, + 392 + ], + "score": 1.0, + "content": "communication,", + "type": "text" + }, + { + "bbox": [ + 171, + 378, + 187, + 390 + ], + "score": 0.92, + "content": "m _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 379, + 372, + 392 + ], + "score": 1.0, + "content": "must contain valuable information that informs", + "type": "text" + }, + { + "bbox": [ + 372, + 379, + 381, + 389 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "about how best to maximize its", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 390, + 406, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 283, + 402 + ], + "score": 1.0, + "content": "own reward, so much so that it actually causes", + "type": "text" + }, + { + "bbox": [ + 284, + 390, + 293, + 399 + ], + "score": 0.82, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 390, + 406, + 402 + ], + "score": 1.0, + "content": "to change its intended action.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 102, + 299, + 509, + 402 + ] + }, + { + "type": "image", + "bbox": [ + 107, + 414, + 287, + 516 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 414, + 287, + 516 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 414, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 107, + 414, + 287, + 516 + ], + "score": 0.962, + "type": "image", + "image_path": "caa2b9dc3c9fe0bdc22237d63bfba2b09073071b89e2a88a719924f5aa8e5a38.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 107, + 414, + 287, + 428.57142857142856 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 107, + 428.57142857142856, + 287, + 443.1428571428571 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 107, + 443.1428571428571, + 287, + 457.71428571428567 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 107, + 457.71428571428567, + 287, + 472.2857142857142 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 107, + 472.2857142857142, + 287, + 486.8571428571428 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 107, + 486.8571428571428, + 287, + 501.42857142857133 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 107, + 501.42857142857133, + 287, + 515.9999999999999 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 525, + 288, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 525, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 287, + 537 + ], + "score": 1.0, + "content": "Figure 2: The communication model has two", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 536, + 288, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 257, + 549 + ], + "score": 1.0, + "content": "A2C heads, which learn a normal policy,", + "type": "text" + }, + { + "bbox": [ + 257, + 538, + 268, + 547 + ], + "score": 0.85, + "content": "\\pi _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 536, + 288, + 549 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 547, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 288, + 559 + ], + "score": 1.0, + "content": "a policy for emitting communication symbols,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 557, + 286, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 117, + 569 + ], + "score": 0.84, + "content": "\\pi _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 557, + 271, + 572 + ], + "score": 1.0, + "content": ". Other agents’ communication messages", + "type": "text" + }, + { + "bbox": [ + 271, + 560, + 286, + 569 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } m _ { t }", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 570, + 196, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 196, + 581 + ], + "score": 1.0, + "content": "are input to the LSTM.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + } + ], + "index": 36.5 + }, + { + "type": "image", + "bbox": [ + 305, + 412, + 500, + 514 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 305, + 412, + 500, + 514 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 305, + 412, + 500, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 500, + 514 + ], + "score": 0.964, + "type": "image", + "image_path": "64e00c7a40f5d8bfca22499a480991f9780f7757dc5984363b283b0279a25fc0.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 305, + 412, + 500, + 426.57142857142856 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 305, + 426.57142857142856, + 500, + 441.1428571428571 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 305, + 441.1428571428571, + 500, + 455.71428571428567 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 305, + 455.71428571428567, + 500, + 470.2857142857142 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 305, + 470.2857142857142, + 500, + 484.8571428571428 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 305, + 484.8571428571428, + 500, + 499.42857142857133 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 305, + 499.42857142857133, + 500, + 513.9999999999999 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 301, + 525, + 503, + 580 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 302, + 525, + 503, + 537 + ], + "spans": [ + { + "bbox": [ + 302, + 525, + 503, + 537 + ], + "score": 1.0, + "content": "Figure 3: The Model of Other Agents (MOA) archi-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 301, + 537, + 503, + 549 + ], + "spans": [ + { + "bbox": [ + 301, + 537, + 423, + 549 + ], + "score": 1.0, + "content": "tecture learns both an RL policy", + "type": "text" + }, + { + "bbox": [ + 424, + 538, + 434, + 547 + ], + "score": 0.84, + "content": "\\pi _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 537, + 503, + 549 + ], + "score": 1.0, + "content": ", and a supervised", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 300, + 546, + 500, + 561 + ], + "spans": [ + { + "bbox": [ + 300, + 546, + 480, + 561 + ], + "score": 1.0, + "content": "model that predicts the actions of other agents,", + "type": "text" + }, + { + "bbox": [ + 480, + 549, + 500, + 559 + ], + "score": 0.88, + "content": "\\mathbf { \\pmb { a } } _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 302, + 559, + 503, + 570 + ], + "spans": [ + { + "bbox": [ + 302, + 559, + 503, + 570 + ], + "score": 1.0, + "content": "The predictions of the supervised model are used for", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 302, + 570, + 426, + 581 + ], + "spans": [ + { + "bbox": [ + 302, + 570, + 426, + 581 + ], + "score": 1.0, + "content": "computing the influence reward.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "index": 42.5 + }, + { + "type": "title", + "bbox": [ + 109, + 590, + 307, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 309, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 309, + 601 + ], + "score": 1.0, + "content": "2.4 INFLUENCE VIA MODELING OTHER AGENTS", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "Computing the causal influence reward as introduced in Section 2.1 requires knowing the probability of", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 107, + 612, + 508, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 617, + 116, + 626 + ], + "score": 0.71, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 612, + 255, + 633 + ], + "score": 1.0, + "content": "’s next action given a counterfactual,", + "type": "text" + }, + { + "bbox": [ + 256, + 615, + 308, + 628 + ], + "score": 0.93, + "content": "p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , s _ { t } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 612, + 508, + 633 + ], + "score": 1.0, + "content": ", which we previously solved by using a centralised", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "controller that could access other agent’s policy networks. While using a centralised training framework is", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "common in MARL (e.g. Foerster et al. (2017; 2016)), it is less realistic than a scenario in which each agent", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 104, + 647, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 647, + 505, + 663 + ], + "score": 1.0, + "content": "is trained independently. We can relax this assumption and achieve independent training by equipping", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "each agent with its own internal Model of Other Agents (MOA). The MOA consists of a second set of", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "fully-connected and LSTM layers connected to the agent’s convolutional layer (see Figure 3), and is trained", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "score": 1.0, + "content": "to predict all other agents’ next actions given their current action, and the agent’s egocentric view of the", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 692, + 489, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 129, + 706 + ], + "score": 1.0, + "content": "state:", + "type": "text" + }, + { + "bbox": [ + 129, + 692, + 186, + 705 + ], + "score": 0.93, + "content": "p ( { \\pmb a } _ { t + 1 } | { \\pmb a } _ { t } , { \\pmb s } _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 693, + 489, + 706 + ], + "score": 1.0, + "content": ". The MOA is trained using cross-entropy loss over observed action trajectories.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 56, + "bbox_fs": [ + 104, + 604, + 508, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "A trained MOA can be used to compute the social influence reward in the following way. Each agent can", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "“imagine” counterfactual actions that it could have taken at each timestep, and use its internal MOA to", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "predict the effect on other agents. It can then give itself reward for taking actions that it estimates were the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "most influential. This has an intuitive appeal, because it resembles how humans reason about their effect", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "on others (Ferguson et al., 2010). We may often find ourselves asking counterfactual questions of the form,", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "“How would she have reacted if I had said or done something else in that situation?”, which we can only", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 268, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 268, + 139 + ], + "score": 1.0, + "content": "answer using our internal model of others.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 61.5, + "bbox_fs": [ + 105, + 708, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "predict the effect on other agents. It can then give itself reward for taking actions that it estimates were the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "most influential. This has an intuitive appeal, because it resembles how humans reason about their effect", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "on others (Ferguson et al., 2010). We may often find ourselves asking counterfactual questions of the form,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "“How would she have reacted if I had said or done something else in that situation?”, which we can only", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 268, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 268, + 139 + ], + "score": 1.0, + "content": "answer using our internal model of others.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "Both the MOA and communication approaches are an important improvement over the original model", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 456, + 166 + ], + "score": 1.0, + "content": "shown in Figure 1, which computed influence within a given timestep and required that agent", + "type": "text" + }, + { + "bbox": [ + 456, + 155, + 465, + 164 + ], + "score": 0.66, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "choose its", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 131, + 178 + ], + "score": 1.0, + "content": "action", + "type": "text" + }, + { + "bbox": [ + 132, + 165, + 144, + 177 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 165, + 308, + 178 + ], + "score": 1.0, + "content": "first, and this action be transmitted to agent", + "type": "text" + }, + { + "bbox": [ + 308, + 166, + 317, + 175 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "as input. This meant that only some agents (those", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "acting first) could be influencers. In contrast, using influence for communication or with a MOA are general", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "approaches that can be implemented in any agent, and allow all agents to mutually influence each other.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 317, + 338 + ], + "lines": [ + { + "bbox": [ + 105, + 213, + 318, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 318, + 228 + ], + "score": 1.0, + "content": "We now seek to estimate influence in the next timestep,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 102, + 220, + 321, + 243 + ], + "spans": [ + { + "bbox": [ + 102, + 220, + 207, + 243 + ], + "score": 1.0, + "content": "meaning the influence of", + "type": "text" + }, + { + "bbox": [ + 207, + 225, + 219, + 238 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 220, + 233, + 243 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 234, + 225, + 253, + 239 + ], + "score": 0.92, + "content": "a _ { t + 1 } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 220, + 321, + 243 + ], + "score": 1.0, + "content": ", which requires", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 103, + 234, + 321, + 255 + ], + "spans": [ + { + "bbox": [ + 103, + 234, + 144, + 255 + ], + "score": 1.0, + "content": "modeling", + "type": "text" + }, + { + "bbox": [ + 144, + 237, + 201, + 252 + ], + "score": 0.92, + "content": "p ( a _ { t + 1 } ^ { B } | \\underline { { a } } _ { t } ^ { A } , s _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 234, + 321, + 255 + ], + "score": 1.0, + "content": ". The corresponding causal dia-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 250, + 318, + 278 + ], + "spans": [ + { + "bbox": [ + 102, + 250, + 117, + 278 + ], + "score": 1.0, + "content": "grof", + "type": "text" + }, + { + "bbox": [ + 117, + 260, + 129, + 273 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 250, + 142, + 278 + ], + "score": 1.0, + "content": "is son", + "type": "text" + }, + { + "bbox": [ + 143, + 260, + 162, + 273 + ], + "score": 0.92, + "content": "a _ { t + 1 } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 250, + 318, + 278 + ], + "score": 1.0, + "content": "in Figure 4. We can infer the causal effectby conditioning on the shaded variables", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 318, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 318, + 284 + ], + "score": 1.0, + "content": "(so that there are no back-door paths) (Pearl et al., 2016).", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 103, + 279, + 320, + 298 + ], + "spans": [ + { + "bbox": [ + 103, + 279, + 186, + 298 + ], + "score": 1.0, + "content": "Learning a model of", + "type": "text" + }, + { + "bbox": [ + 186, + 282, + 244, + 295 + ], + "score": 0.93, + "content": "p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { \\bar { A } } , s _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 279, + 320, + 298 + ], + "score": 1.0, + "content": "requires implicitly", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 317, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 294, + 304 + ], + "score": 1.0, + "content": "modeling both the environment transition function", + "type": "text" + }, + { + "bbox": [ + 295, + 294, + 303, + 304 + ], + "score": 0.77, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 294, + 317, + 304 + ], + "score": 1.0, + "content": "(to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 317, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 135, + 316 + ], + "score": 1.0, + "content": "predict", + "type": "text" + }, + { + "bbox": [ + 135, + 305, + 155, + 316 + ], + "score": 0.89, + "content": "s _ { t + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 304, + 317, + 316 + ], + "score": 1.0, + "content": ", as well as relevant aspects of the internal", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 102, + 310, + 320, + 331 + ], + "spans": [ + { + "bbox": [ + 102, + 310, + 227, + 331 + ], + "score": 1.0, + "content": "LSTM state of the other agent,", + "type": "text" + }, + { + "bbox": [ + 227, + 315, + 246, + 328 + ], + "score": 0.92, + "content": "u _ { t + 1 } ^ { B ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 310, + 320, + 331 + ], + "score": 1.0, + "content": ", as highlighted in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 144, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 144, + 339 + ], + "score": 1.0, + "content": "Figure 4.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 316, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 318, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 318, + 356 + ], + "score": 1.0, + "content": "We enable agents to condition their policy on the ac-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 355, + 317, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 317, + 366 + ], + "score": 1.0, + "content": "tions of other agents in the previous timestep (actions", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 365, + 317, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 317, + 376 + ], + "score": 1.0, + "content": "are visible), and only give the social influence reward", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 377, + 317, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 317, + 387 + ], + "score": 1.0, + "content": "to an agent when the agent it is attempting to influ-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 387, + 317, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 317, + 398 + ], + "score": 1.0, + "content": "ence is within its field-of-view, because the estimates of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 396, + 316, + 412 + ], + "spans": [ + { + "bbox": [ + 107, + 398, + 165, + 411 + ], + "score": 0.92, + "content": "p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 396, + 306, + 412 + ], + "score": 1.0, + "content": "are likely to be more accurate when", + "type": "text" + }, + { + "bbox": [ + 306, + 398, + 316, + 408 + ], + "score": 0.72, + "content": "B", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 410, + 318, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 151, + 422 + ], + "score": 1.0, + "content": "is visible to", + "type": "text" + }, + { + "bbox": [ + 151, + 411, + 163, + 421 + ], + "score": 0.86, + "content": "A ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 410, + 318, + 422 + ], + "score": 1.0, + "content": ". The latter constraint could have the side-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 421, + 318, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 318, + 434 + ], + "score": 1.0, + "content": "effect of encouraging agents to stay in closer proximity.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 327, + 214, + 483, + 323 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 327, + 214, + 483, + 323 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 327, + 214, + 483, + 323 + ], + "spans": [ + { + "bbox": [ + 327, + 214, + 483, + 323 + ], + "score": 0.968, + "type": "image", + "image_path": "c636b44cc733a0638da4fbce914888df05874d72cc520f1bcf72e309b465e051.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 327, + 214, + 483, + 227.625 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 327, + 227.625, + 483, + 241.25 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 327, + 241.25, + 483, + 254.875 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 327, + 254.875, + 483, + 268.5 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 327, + 268.5, + 483, + 282.125 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 327, + 282.125, + 483, + 295.75 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 327, + 295.75, + 483, + 309.375 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 327, + 309.375, + 483, + 323.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 332, + 505, + 420 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 323, + 331, + 507, + 345 + ], + "spans": [ + { + "bbox": [ + 323, + 331, + 507, + 345 + ], + "score": 1.0, + "content": "Figure 4: Causal diagram in the MOA case.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 324, + 343, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 324, + 343, + 506, + 355 + ], + "score": 1.0, + "content": "Shaded nodes are conditioned on, and we in-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 324, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 324, + 354, + 367, + 366 + ], + "score": 1.0, + "content": "tervene on", + "type": "text" + }, + { + "bbox": [ + 367, + 354, + 380, + 366 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "(blue node) by replacing it with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 324, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 324, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "counterfactuals. Nodes with a green background", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 324, + 376, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 324, + 376, + 504, + 387 + ], + "score": 1.0, + "content": "must be modeled using the MOA module. Note", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 324, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 324, + 387, + 474, + 399 + ], + "score": 1.0, + "content": "that there is no backdoor path between", + "type": "text" + }, + { + "bbox": [ + 475, + 387, + 487, + 399 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 324, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 324, + 399, + 334, + 409 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "since it would require traversing a collider", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 323, + 408, + 452, + 421 + ], + "spans": [ + { + "bbox": [ + 323, + 408, + 452, + 421 + ], + "score": 1.0, + "content": "that is not in the conditioning set.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "However, an intrinsic social reward based on proximity is also a reasonable approach to approximating", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 444, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 455 + ], + "score": 1.0, + "content": "human social motivation. Humans seek affiliation and to spend time near other people (Tomasello, 2009).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "title", + "bbox": [ + 108, + 473, + 263, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 265, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 265, + 486 + ], + "score": 1.0, + "content": "2.5 SEQUENTIAL SOCIAL DILEMMAS", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 106, + 492, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 493, + 504, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 504, + 505 + ], + "score": 1.0, + "content": "First proposed by Leibo et al. (2017), Sequential Social Dilemmas (SSDs) are spatially and temporally", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "score": 1.0, + "content": "extended multi-agent games that have a payoff structure similar to that of Prisoner’s Dilemma (PD). That is,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "an individual agent can obtain higher reward by engaging in defecting, non-cooperative behavior (and thus", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 524, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 524, + 506, + 540 + ], + "score": 1.0, + "content": "is rationally motivated to defect), but the average payoff per agent will be higher if all agents cooperate (see", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "Figure 9 of the Appendix). The paradoxical reward structure makes it extremely difficult for traditional RL", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 104, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 104, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "agents to learn to coordinate to solve the tasks (Hughes et al., 2018). We experiment with two SSDs in this", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "work, a public goods game Cleanup, and a tragedy-of-the-commons game Harvest (see Figure 5). In both", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "games apples (green tiles) provide the rewards, and agents also have the ability to punish each other with a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 581, + 360, + 594 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 360, + 594 + ], + "score": 1.0, + "content": "fining beam. Further details are available in Appendix Section 6.1.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 112, + 605, + 501, + 674 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 605, + 501, + 674 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 605, + 501, + 674 + ], + "spans": [ + { + "bbox": [ + 112, + 605, + 501, + 674 + ], + "score": 0.961, + "type": "image", + "image_path": "20206990cd332b2c23f36bfab8c7c99fd397bf6323218026ab67daf49c621e5f.jpg" + } + ] + } + ], + "index": 57, + "virtual_lines": [ + { + "bbox": [ + 112, + 605, + 501, + 628.0 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 112, + 628.0, + 501, + 651.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 112, + 651.0, + 501, + 674.0 + ], + "spans": [], + "index": 58 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 682, + 503, + 705 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 682, + 504, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 504, + 695 + ], + "score": 1.0, + "content": "Figure 5: The two SSD environments, Cleanup (left) and Harvest (right). Agents can exploit other agents", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 692, + 442, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 442, + 707 + ], + "score": 1.0, + "content": "for immediate payoff, but at the expense of the long-term collective reward of the group.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 59.5 + } + ], + "index": 58.25 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 113, + 721, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 718, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 718, + 505, + 734 + ], + "score": 1.0, + "content": "2This contrasts with our previous models in which the influence reward was obtained even from non-visible agents.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 104, + 82, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "Both the MOA and communication approaches are an important improvement over the original model", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 456, + 166 + ], + "score": 1.0, + "content": "shown in Figure 1, which computed influence within a given timestep and required that agent", + "type": "text" + }, + { + "bbox": [ + 456, + 155, + 465, + 164 + ], + "score": 0.66, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "choose its", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 131, + 178 + ], + "score": 1.0, + "content": "action", + "type": "text" + }, + { + "bbox": [ + 132, + 165, + 144, + 177 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 165, + 308, + 178 + ], + "score": 1.0, + "content": "first, and this action be transmitted to agent", + "type": "text" + }, + { + "bbox": [ + 308, + 166, + 317, + 175 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "as input. This meant that only some agents (those", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "acting first) could be influencers. In contrast, using influence for communication or with a MOA are general", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "approaches that can be implemented in any agent, and allow all agents to mutually influence each other.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 143, + 506, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 317, + 338 + ], + "lines": [ + { + "bbox": [ + 105, + 213, + 318, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 318, + 228 + ], + "score": 1.0, + "content": "We now seek to estimate influence in the next timestep,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 102, + 220, + 321, + 243 + ], + "spans": [ + { + "bbox": [ + 102, + 220, + 207, + 243 + ], + "score": 1.0, + "content": "meaning the influence of", + "type": "text" + }, + { + "bbox": [ + 207, + 225, + 219, + 238 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 220, + 233, + 243 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 234, + 225, + 253, + 239 + ], + "score": 0.92, + "content": "a _ { t + 1 } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 220, + 321, + 243 + ], + "score": 1.0, + "content": ", which requires", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 103, + 234, + 321, + 255 + ], + "spans": [ + { + "bbox": [ + 103, + 234, + 144, + 255 + ], + "score": 1.0, + "content": "modeling", + "type": "text" + }, + { + "bbox": [ + 144, + 237, + 201, + 252 + ], + "score": 0.92, + "content": "p ( a _ { t + 1 } ^ { B } | \\underline { { a } } _ { t } ^ { A } , s _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 234, + 321, + 255 + ], + "score": 1.0, + "content": ". The corresponding causal dia-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 250, + 318, + 278 + ], + "spans": [ + { + "bbox": [ + 102, + 250, + 117, + 278 + ], + "score": 1.0, + "content": "grof", + "type": "text" + }, + { + "bbox": [ + 117, + 260, + 129, + 273 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 250, + 142, + 278 + ], + "score": 1.0, + "content": "is son", + "type": "text" + }, + { + "bbox": [ + 143, + 260, + 162, + 273 + ], + "score": 0.92, + "content": "a _ { t + 1 } ^ { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 250, + 318, + 278 + ], + "score": 1.0, + "content": "in Figure 4. We can infer the causal effectby conditioning on the shaded variables", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 318, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 318, + 284 + ], + "score": 1.0, + "content": "(so that there are no back-door paths) (Pearl et al., 2016).", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 103, + 279, + 320, + 298 + ], + "spans": [ + { + "bbox": [ + 103, + 279, + 186, + 298 + ], + "score": 1.0, + "content": "Learning a model of", + "type": "text" + }, + { + "bbox": [ + 186, + 282, + 244, + 295 + ], + "score": 0.93, + "content": "p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { \\bar { A } } , s _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 279, + 320, + 298 + ], + "score": 1.0, + "content": "requires implicitly", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 317, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 294, + 304 + ], + "score": 1.0, + "content": "modeling both the environment transition function", + "type": "text" + }, + { + "bbox": [ + 295, + 294, + 303, + 304 + ], + "score": 0.77, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 294, + 317, + 304 + ], + "score": 1.0, + "content": "(to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 317, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 135, + 316 + ], + "score": 1.0, + "content": "predict", + "type": "text" + }, + { + "bbox": [ + 135, + 305, + 155, + 316 + ], + "score": 0.89, + "content": "s _ { t + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 304, + 317, + 316 + ], + "score": 1.0, + "content": ", as well as relevant aspects of the internal", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 102, + 310, + 320, + 331 + ], + "spans": [ + { + "bbox": [ + 102, + 310, + 227, + 331 + ], + "score": 1.0, + "content": "LSTM state of the other agent,", + "type": "text" + }, + { + "bbox": [ + 227, + 315, + 246, + 328 + ], + "score": 0.92, + "content": "u _ { t + 1 } ^ { B ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 310, + 320, + 331 + ], + "score": 1.0, + "content": ", as highlighted in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 144, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 144, + 339 + ], + "score": 1.0, + "content": "Figure 4.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5, + "bbox_fs": [ + 102, + 213, + 321, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 316, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 318, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 318, + 356 + ], + "score": 1.0, + "content": "We enable agents to condition their policy on the ac-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 355, + 317, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 317, + 366 + ], + "score": 1.0, + "content": "tions of other agents in the previous timestep (actions", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 365, + 317, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 317, + 376 + ], + "score": 1.0, + "content": "are visible), and only give the social influence reward", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 377, + 317, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 317, + 387 + ], + "score": 1.0, + "content": "to an agent when the agent it is attempting to influ-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 387, + 317, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 317, + 398 + ], + "score": 1.0, + "content": "ence is within its field-of-view, because the estimates of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 396, + 316, + 412 + ], + "spans": [ + { + "bbox": [ + 107, + 398, + 165, + 411 + ], + "score": 0.92, + "content": "p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 396, + 306, + 412 + ], + "score": 1.0, + "content": "are likely to be more accurate when", + "type": "text" + }, + { + "bbox": [ + 306, + 398, + 316, + 408 + ], + "score": 0.72, + "content": "B", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 410, + 318, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 151, + 422 + ], + "score": 1.0, + "content": "is visible to", + "type": "text" + }, + { + "bbox": [ + 151, + 411, + 163, + 421 + ], + "score": 0.86, + "content": "A ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 410, + 318, + 422 + ], + "score": 1.0, + "content": ". The latter constraint could have the side-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 421, + 318, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 318, + 434 + ], + "score": 1.0, + "content": "effect of encouraging agents to stay in closer proximity.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 342, + 318, + 434 + ] + }, + { + "type": "image", + "bbox": [ + 327, + 214, + 483, + 323 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 327, + 214, + 483, + 323 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 327, + 214, + 483, + 323 + ], + "spans": [ + { + "bbox": [ + 327, + 214, + 483, + 323 + ], + "score": 0.968, + "type": "image", + "image_path": "c636b44cc733a0638da4fbce914888df05874d72cc520f1bcf72e309b465e051.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 327, + 214, + 483, + 227.625 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 327, + 227.625, + 483, + 241.25 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 327, + 241.25, + 483, + 254.875 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 327, + 254.875, + 483, + 268.5 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 327, + 268.5, + 483, + 282.125 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 327, + 282.125, + 483, + 295.75 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 327, + 295.75, + 483, + 309.375 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 327, + 309.375, + 483, + 323.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 332, + 505, + 420 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 323, + 331, + 507, + 345 + ], + "spans": [ + { + "bbox": [ + 323, + 331, + 507, + 345 + ], + "score": 1.0, + "content": "Figure 4: Causal diagram in the MOA case.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 324, + 343, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 324, + 343, + 506, + 355 + ], + "score": 1.0, + "content": "Shaded nodes are conditioned on, and we in-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 324, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 324, + 354, + 367, + 366 + ], + "score": 1.0, + "content": "tervene on", + "type": "text" + }, + { + "bbox": [ + 367, + 354, + 380, + 366 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "(blue node) by replacing it with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 324, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 324, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "counterfactuals. Nodes with a green background", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 324, + 376, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 324, + 376, + 504, + 387 + ], + "score": 1.0, + "content": "must be modeled using the MOA module. Note", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 324, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 324, + 387, + 474, + 399 + ], + "score": 1.0, + "content": "that there is no backdoor path between", + "type": "text" + }, + { + "bbox": [ + 475, + 387, + 487, + 399 + ], + "score": 0.9, + "content": "a _ { t } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 324, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 324, + 399, + 334, + 409 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "since it would require traversing a collider", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 323, + 408, + 452, + 421 + ], + "spans": [ + { + "bbox": [ + 323, + 408, + 452, + 421 + ], + "score": 1.0, + "content": "that is not in the conditioning set.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "However, an intrinsic social reward based on proximity is also a reasonable approach to approximating", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 444, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 455 + ], + "score": 1.0, + "content": "human social motivation. Humans seek affiliation and to spend time near other people (Tomasello, 2009).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 432, + 506, + 455 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 473, + 263, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 265, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 265, + 486 + ], + "score": 1.0, + "content": "2.5 SEQUENTIAL SOCIAL DILEMMAS", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 106, + 492, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 493, + 504, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 504, + 505 + ], + "score": 1.0, + "content": "First proposed by Leibo et al. (2017), Sequential Social Dilemmas (SSDs) are spatially and temporally", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "score": 1.0, + "content": "extended multi-agent games that have a payoff structure similar to that of Prisoner’s Dilemma (PD). That is,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "an individual agent can obtain higher reward by engaging in defecting, non-cooperative behavior (and thus", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 524, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 524, + 506, + 540 + ], + "score": 1.0, + "content": "is rationally motivated to defect), but the average payoff per agent will be higher if all agents cooperate (see", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "Figure 9 of the Appendix). The paradoxical reward structure makes it extremely difficult for traditional RL", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 104, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 104, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "agents to learn to coordinate to solve the tasks (Hughes et al., 2018). We experiment with two SSDs in this", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "work, a public goods game Cleanup, and a tragedy-of-the-commons game Harvest (see Figure 5). In both", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "games apples (green tiles) provide the rewards, and agents also have the ability to punish each other with a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 581, + 360, + 594 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 360, + 594 + ], + "score": 1.0, + "content": "fining beam. Further details are available in Appendix Section 6.1.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 51, + "bbox_fs": [ + 104, + 493, + 506, + 594 + ] + }, + { + "type": "image", + "bbox": [ + 112, + 605, + 501, + 674 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 605, + 501, + 674 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 605, + 501, + 674 + ], + "spans": [ + { + "bbox": [ + 112, + 605, + 501, + 674 + ], + "score": 0.961, + "type": "image", + "image_path": "20206990cd332b2c23f36bfab8c7c99fd397bf6323218026ab67daf49c621e5f.jpg" + } + ] + } + ], + "index": 57, + "virtual_lines": [ + { + "bbox": [ + 112, + 605, + 501, + 628.0 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 112, + 628.0, + 501, + 651.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 112, + 651.0, + 501, + 674.0 + ], + "spans": [], + "index": 58 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 682, + 503, + 705 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 682, + 504, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 504, + 695 + ], + "score": 1.0, + "content": "Figure 5: The two SSD environments, Cleanup (left) and Harvest (right). Agents can exploit other agents", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 692, + 442, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 442, + 707 + ], + "score": 1.0, + "content": "for immediate payoff, but at the expense of the long-term collective reward of the group.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 59.5 + } + ], + "index": 58.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 205, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 207, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 207, + 96 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 109, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 105, + 108, + 507, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 507, + 123 + ], + "score": 1.0, + "content": "Several attempts have been made to develop intrinsic social motivation rewards3. Sequeira et al. (2011)", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 120, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 133 + ], + "score": 1.0, + "content": "developed hand-crafted rewards specific to a foraging environment, in which agents were punished for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "eating more than their fair share of food. Another approach gave agents an emotional intrinsic reward based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 144, + 504, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 504, + 154 + ], + "score": 1.0, + "content": "on their perception of their neighbours’ cooperativeness in a networked version of the iterated prisoner’s", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "score": 1.0, + "content": "dilemma (Yu et al., 2013). This approach is limited to scenarios in which it is possible to directly classify", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "each action as cooperative or non-cooperative, which is untenable in complex settings with long-term", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "strategies, such as the SSDs under investigation here. Hughes et al. (2018) introduced an inequity aversion", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "motivation, which penalized agents if their rewards differed too much from those of the group. Another", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "approach used prosocial reward shaping to show that if even a single agent is trained to optimize for the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 506, + 221 + ], + "score": 1.0, + "content": "rewards of other agents, it can help the group obtain better collective outcomes (Peysakhovich & Lerer,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "2018). However, these both require the ability to observe other agent’s rewards, which may be an unrealistic", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 231, + 267, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 267, + 243 + ], + "score": 1.0, + "content": "assumption, depending on the application.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 260 + ], + "score": 1.0, + "content": "Another body of work has focused on training agents to learn emergent communication protocols (Foerster", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "et al., 2016; Cao et al., 2018; Choi et al., 2018; Lazaridou et al., 2018; Bogin et al., 2018), with many authors", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "finding that selfish agents do not learn to use an ungrounded, cheap talk communication channel effectively.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "score": 1.0, + "content": "Crawford & Sobel (1982) find that in theory, the information revealed in communication (in equilibrium)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "is proportional to amount of common interest; thus, as agents’ interests diverge, no communication is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "to be expected. And while communication can emerge when agents are prosocial (Foerster et al., 2016;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "score": 1.0, + "content": "Lazaridou et al., 2018) or hand-crafted (Crandall et al., 2017), self-interested agents do not to learn to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 324, + 504, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 504, + 336 + ], + "score": 1.0, + "content": "communicate (Cao et al., 2018). We test whether the social influence reward can encourage agents to learn", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 335, + 497, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 497, + 347 + ], + "score": 1.0, + "content": "to communicate more effectively in complex environments with challenging social dilemma dynamics.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "Interestingly, Oudeyer & Kaplan (2006) show that a robot trained with a curiosity-based intrinsic motivation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "to maximize learning progress learns to prefer vocalizing sounds imitated by another robot over interaction", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "with other objects in the environment. Follow-up papers suggest that curiosity may be a sufficient motivation", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "to encourage agents, or even children, to learn to communicate with others (Oudeyer & Smith, 2016;", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 396, + 216, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 216, + 408 + ], + "score": 1.0, + "content": "Forestier & Oudeyer, 2017).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "Our MOA network is related to work on machine theory of mind (Rabinowitz et al., 2018), which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 424, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 435 + ], + "score": 1.0, + "content": "demonstrated that a model trained to predict agents’ actions is able to model false beliefs. With LOLA,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "Foerster et al. (2018) train agents that model the impact of their policy on the parameter updates of other", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 374, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 374, + 458 + ], + "score": 1.0, + "content": "agents, and directly incorporate this into the agent’s own learning rule.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "score": 1.0, + "content": "Barton et al. (2018) propose causal influence as a way to measure coordination between agents, specifically", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 504, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 504, + 485 + ], + "score": 1.0, + "content": "using Convergence Cross Mapping (CCM) to analyze the degree of dependence between two agents’", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "policies. The limitation of this approach is that CCM estimates of causality are known to degrade in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 508 + ], + "score": 1.0, + "content": "presence of stochastic effects (Tajima et al., 2015). Counterfactual reasoning has also been used in a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "multi-agent setting, to marginalize out the effect of one agent on a predicted global value function estimating", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "collective reward, and thus obtain an improved baseline for computing each agent’s advantage function", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "(Foerster et al., 2017). A similar paper shows that counterfactuals can be used with potential-based reward", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "shaping to improve credit assignment for training a joint policy in multi-agent RL Devlin et al. (2014).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 551, + 373, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 373, + 562 + ], + "score": 1.0, + "content": "However, once again these approaches rely on a centralised controller.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "Following in the tradition of the empowerment literature, authors have investigated mutual information", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "(MI) as a powerful tool for designing social rewards. Strouse et al. (2018) train agents to maximize or", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "minimize the MI between their actions and a categorical goal, and show how this can be used to signal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "or hide the agent’s intentions. However, this approach depends on agents pursuing a known, categorical", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 610, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 507, + 624 + ], + "score": 1.0, + "content": "goal. Guckelsberger et al. (2018), in pursuit of the ultimate video game adversary, develop an agent that", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "maximizes its empowerment over its own states, minimizes the player’s empowerment over their states, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "maximizes its empowerment over the player’s next state. This third goal, termed transfer empowerment, is", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "obtained by maximizing the MI between the agent’s actions and the player’s future state. While similar", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "to our approach, the authors find that agents trained with transfer empowerment simply tend to stay near", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "the player. Further, the agents are not trained with RL, but rather analytically compute these measures in", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "simple grid-world environments. As such, the agent cannot learn to model other agents or the environment.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 45 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 711, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 708, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 505, + 725 + ], + "score": 1.0, + "content": "3Note that intrinsic is not a synonym of internal; it is possible to be intrinsically motivated by other people", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "score": 1.0, + "content": "(Stavropoulos & Carver, 2013).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 298, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 205, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 207, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 207, + 96 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 109, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 105, + 108, + 507, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 507, + 123 + ], + "score": 1.0, + "content": "Several attempts have been made to develop intrinsic social motivation rewards3. Sequeira et al. (2011)", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 120, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 133 + ], + "score": 1.0, + "content": "developed hand-crafted rewards specific to a foraging environment, in which agents were punished for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "eating more than their fair share of food. Another approach gave agents an emotional intrinsic reward based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 144, + 504, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 504, + 154 + ], + "score": 1.0, + "content": "on their perception of their neighbours’ cooperativeness in a networked version of the iterated prisoner’s", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "score": 1.0, + "content": "dilemma (Yu et al., 2013). This approach is limited to scenarios in which it is possible to directly classify", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "each action as cooperative or non-cooperative, which is untenable in complex settings with long-term", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "strategies, such as the SSDs under investigation here. Hughes et al. (2018) introduced an inequity aversion", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "motivation, which penalized agents if their rewards differed too much from those of the group. Another", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "approach used prosocial reward shaping to show that if even a single agent is trained to optimize for the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 506, + 221 + ], + "score": 1.0, + "content": "rewards of other agents, it can help the group obtain better collective outcomes (Peysakhovich & Lerer,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "2018). However, these both require the ability to observe other agent’s rewards, which may be an unrealistic", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 231, + 267, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 267, + 243 + ], + "score": 1.0, + "content": "assumption, depending on the application.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 108, + 507, + 243 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 260 + ], + "score": 1.0, + "content": "Another body of work has focused on training agents to learn emergent communication protocols (Foerster", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "et al., 2016; Cao et al., 2018; Choi et al., 2018; Lazaridou et al., 2018; Bogin et al., 2018), with many authors", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "finding that selfish agents do not learn to use an ungrounded, cheap talk communication channel effectively.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 293 + ], + "score": 1.0, + "content": "Crawford & Sobel (1982) find that in theory, the information revealed in communication (in equilibrium)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "is proportional to amount of common interest; thus, as agents’ interests diverge, no communication is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "to be expected. And while communication can emerge when agents are prosocial (Foerster et al., 2016;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 506, + 326 + ], + "score": 1.0, + "content": "Lazaridou et al., 2018) or hand-crafted (Crandall et al., 2017), self-interested agents do not to learn to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 324, + 504, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 504, + 336 + ], + "score": 1.0, + "content": "communicate (Cao et al., 2018). We test whether the social influence reward can encourage agents to learn", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 335, + 497, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 497, + 347 + ], + "score": 1.0, + "content": "to communicate more effectively in complex environments with challenging social dilemma dynamics.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 246, + 506, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "Interestingly, Oudeyer & Kaplan (2006) show that a robot trained with a curiosity-based intrinsic motivation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "to maximize learning progress learns to prefer vocalizing sounds imitated by another robot over interaction", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "with other objects in the environment. Follow-up papers suggest that curiosity may be a sufficient motivation", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "to encourage agents, or even children, to learn to communicate with others (Oudeyer & Smith, 2016;", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 396, + 216, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 216, + 408 + ], + "score": 1.0, + "content": "Forestier & Oudeyer, 2017).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 352, + 506, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "Our MOA network is related to work on machine theory of mind (Rabinowitz et al., 2018), which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 424, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 435 + ], + "score": 1.0, + "content": "demonstrated that a model trained to predict agents’ actions is able to model false beliefs. With LOLA,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "Foerster et al. (2018) train agents that model the impact of their policy on the parameter updates of other", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 374, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 374, + 458 + ], + "score": 1.0, + "content": "agents, and directly incorporate this into the agent’s own learning rule.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 412, + 506, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "score": 1.0, + "content": "Barton et al. (2018) propose causal influence as a way to measure coordination between agents, specifically", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 504, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 504, + 485 + ], + "score": 1.0, + "content": "using Convergence Cross Mapping (CCM) to analyze the degree of dependence between two agents’", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "policies. The limitation of this approach is that CCM estimates of causality are known to degrade in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 508 + ], + "score": 1.0, + "content": "presence of stochastic effects (Tajima et al., 2015). Counterfactual reasoning has also been used in a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "multi-agent setting, to marginalize out the effect of one agent on a predicted global value function estimating", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "collective reward, and thus obtain an improved baseline for computing each agent’s advantage function", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "(Foerster et al., 2017). A similar paper shows that counterfactuals can be used with potential-based reward", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "shaping to improve credit assignment for training a joint policy in multi-agent RL Devlin et al. (2014).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 551, + 373, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 373, + 562 + ], + "score": 1.0, + "content": "However, once again these approaches rely on a centralised controller.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 463, + 506, + 562 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "Following in the tradition of the empowerment literature, authors have investigated mutual information", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "(MI) as a powerful tool for designing social rewards. Strouse et al. (2018) train agents to maximize or", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "minimize the MI between their actions and a categorical goal, and show how this can be used to signal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "or hide the agent’s intentions. However, this approach depends on agents pursuing a known, categorical", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 610, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 507, + 624 + ], + "score": 1.0, + "content": "goal. Guckelsberger et al. (2018), in pursuit of the ultimate video game adversary, develop an agent that", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "maximizes its empowerment over its own states, minimizes the player’s empowerment over their states, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "maximizes its empowerment over the player’s next state. This third goal, termed transfer empowerment, is", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "obtained by maximizing the MI between the agent’s actions and the player’s future state. While similar", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "to our approach, the authors find that agents trained with transfer empowerment simply tend to stay near", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "the player. Further, the agents are not trained with RL, but rather analytically compute these measures in", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "simple grid-world environments. As such, the agent cannot learn to model other agents or the environment.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 566, + 507, + 689 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 196, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 198, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 198, + 96 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "The following sections present the results of training agents with the social influence reward in three", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "settings: (1) using a centralised controller, (2) using an explicit communication channel, and (3) using a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "learned model of other agents (MOA). In each case we compare against a standard A3C agent, and an", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 149 + ], + "score": 1.0, + "content": "ablated version of the model which is architecturally identical, but does not receive the influence reward.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "We measure the total collective reward obtained using the best hyperparameter setting tested with 5 random", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "seeds. It is worth noting that we use a curriculum learning approach which gradually increases the weight", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 237, + 182 + ], + "score": 1.0, + "content": "of the social influence reward over", + "type": "text" + }, + { + "bbox": [ + 237, + 170, + 246, + 180 + ], + "score": 0.79, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 169, + 269, + 182 + ], + "score": 1.0, + "content": "steps", + "type": "text" + }, + { + "bbox": [ + 269, + 169, + 351, + 182 + ], + "score": 0.9, + "content": "( C \\in [ 0 . 2 \\mathrm { - } 3 . 5 ] \\times \\bar { 1 } 0 ^ { 8 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "; this can lead to a slight delay before the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 180, + 266, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 266, + 193 + ], + "score": 1.0, + "content": "influence models’ performance improves.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "We also provide the results of an additional experiment Section 6.2 of the Appendix, which tests the social", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "influence reward in a simplified environment where the effects of influence are clear. We encourage the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "reader to examine that section to gain a better intuition for how social influence can foster cooperative", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 230, + 253, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 253, + 243 + ], + "score": 1.0, + "content": "behavior in an otherwise selfish agent.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 261, + 246, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 248, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 248, + 272 + ], + "score": 1.0, + "content": "4.1 CENTRALISED CONTROLLER", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 280, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "Figures 6(a) and 6(d) show the results of training influence with a centralised controller as described", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "in Section 2.1. With this method, the influencer agents transmit their intended action to the influenced", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "score": 1.0, + "content": "agents at each timestep. Therefore, we benchmark against an ablated version of the influence model with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 315, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 326 + ], + "score": 1.0, + "content": "visible actions but no influence reward. As is evident in Figures 6(a) and 6(d), introducing an awareness of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "other agents’ actions helps, but having the social influence reward eventually leads to significantly higher", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 231, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 231, + 348 + ], + "score": 1.0, + "content": "collective reward in both games.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "image", + "bbox": [ + 109, + 362, + 498, + 569 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 362, + 498, + 569 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 362, + 498, + 569 + ], + "spans": [ + { + "bbox": [ + 109, + 362, + 498, + 569 + ], + "score": 0.972, + "type": "image", + "image_path": "3b9225cc90c9a95e47d5cf85514dba567067d0da375c94be1d1955429ea1bdd4.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 109, + 362, + 498, + 431.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 109, + 431.0, + 498, + 500.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 109, + 500.0, + 498, + 569.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 579, + 503, + 612 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 404, + 591 + ], + "score": 1.0, + "content": "Figure 6: Total collective reward obtained in all experiments. Error bars show a", + "type": "text" + }, + { + "bbox": [ + 405, + 579, + 430, + 589 + ], + "score": 0.88, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "confidence interval", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "(CI) over 5 random seeds, computed within a sliding window of 200 agent steps. The models trained with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 600, + 411, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 411, + 613 + ], + "score": 1.0, + "content": "influence reward (red) significantly outperform the baseline and ablated models.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 631 + ], + "score": 1.0, + "content": "While these aggregated results demonstrate the success of our models, they are not sufficient to understand", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "score": 1.0, + "content": "the mechanism through which social influence is helping the agents achieve cooperative behavior. Therefore,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "we investigated the trajectories produced by high scoring models in both Cleanup and Harvest; the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 653, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 506, + 666 + ], + "score": 1.0, + "content": "analysis revealed interesting behavior. As an example, in the Cleanup video available here: https:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "//youtu.be/iH_V5WKQxmo a single agent (shown in purple) was trained with the social influence", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 673, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 104, + 673, + 506, + 689 + ], + "score": 1.0, + "content": "reward. We see that unlike the other agents, which continue to randomly move and explore while waiting", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 686, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 505, + 699 + ], + "score": 1.0, + "content": "for apples to spawn, the influencer has a strange economy of motion; it only moves on the map when it is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 697, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "pursuing an apple, then stops. Interestingly, examining the trajectory reveals that the influencer uses only", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "score": 1.0, + "content": "two moves to explore the map: turn left, which turns the agent in place without traversing the map, and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 718, + 374, + 731 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 374, + 731 + ], + "score": 1.0, + "content": "move right, which moves the agent one square to the right on the map.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 7 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 196, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 198, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 198, + 96 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "The following sections present the results of training agents with the social influence reward in three", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "settings: (1) using a centralised controller, (2) using an explicit communication channel, and (3) using a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "learned model of other agents (MOA). In each case we compare against a standard A3C agent, and an", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 149 + ], + "score": 1.0, + "content": "ablated version of the model which is architecturally identical, but does not receive the influence reward.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "We measure the total collective reward obtained using the best hyperparameter setting tested with 5 random", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "seeds. It is worth noting that we use a curriculum learning approach which gradually increases the weight", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 237, + 182 + ], + "score": 1.0, + "content": "of the social influence reward over", + "type": "text" + }, + { + "bbox": [ + 237, + 170, + 246, + 180 + ], + "score": 0.79, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 169, + 269, + 182 + ], + "score": 1.0, + "content": "steps", + "type": "text" + }, + { + "bbox": [ + 269, + 169, + 351, + 182 + ], + "score": 0.9, + "content": "( C \\in [ 0 . 2 \\mathrm { - } 3 . 5 ] \\times \\bar { 1 } 0 ^ { 8 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 169, + 505, + 182 + ], + "score": 1.0, + "content": "; this can lead to a slight delay before the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 180, + 266, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 266, + 193 + ], + "score": 1.0, + "content": "influence models’ performance improves.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 104, + 506, + 193 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "We also provide the results of an additional experiment Section 6.2 of the Appendix, which tests the social", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "influence reward in a simplified environment where the effects of influence are clear. We encourage the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "reader to examine that section to gain a better intuition for how social influence can foster cooperative", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 230, + 253, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 253, + 243 + ], + "score": 1.0, + "content": "behavior in an otherwise selfish agent.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 197, + 505, + 243 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 261, + 246, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 248, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 248, + 272 + ], + "score": 1.0, + "content": "4.1 CENTRALISED CONTROLLER", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 280, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "Figures 6(a) and 6(d) show the results of training influence with a centralised controller as described", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "in Section 2.1. With this method, the influencer agents transmit their intended action to the influenced", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "score": 1.0, + "content": "agents at each timestep. Therefore, we benchmark against an ablated version of the influence model with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 315, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 326 + ], + "score": 1.0, + "content": "visible actions but no influence reward. As is evident in Figures 6(a) and 6(d), introducing an awareness of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "other agents’ actions helps, but having the social influence reward eventually leads to significantly higher", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 231, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 231, + 348 + ], + "score": 1.0, + "content": "collective reward in both games.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 281, + 506, + 348 + ] + }, + { + "type": "image", + "bbox": [ + 109, + 362, + 498, + 569 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 362, + 498, + 569 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 362, + 498, + 569 + ], + "spans": [ + { + "bbox": [ + 109, + 362, + 498, + 569 + ], + "score": 0.972, + "type": "image", + "image_path": "3b9225cc90c9a95e47d5cf85514dba567067d0da375c94be1d1955429ea1bdd4.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 109, + 362, + 498, + 431.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 109, + 431.0, + 498, + 500.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 109, + 500.0, + 498, + 569.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 579, + 503, + 612 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 404, + 591 + ], + "score": 1.0, + "content": "Figure 6: Total collective reward obtained in all experiments. Error bars show a", + "type": "text" + }, + { + "bbox": [ + 405, + 579, + 430, + 589 + ], + "score": 0.88, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "confidence interval", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "(CI) over 5 random seeds, computed within a sliding window of 200 agent steps. The models trained with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 600, + 411, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 411, + 613 + ], + "score": 1.0, + "content": "influence reward (red) significantly outperform the baseline and ablated models.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 620, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 631 + ], + "score": 1.0, + "content": "While these aggregated results demonstrate the success of our models, they are not sufficient to understand", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "score": 1.0, + "content": "the mechanism through which social influence is helping the agents achieve cooperative behavior. Therefore,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "we investigated the trajectories produced by high scoring models in both Cleanup and Harvest; the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 653, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 506, + 666 + ], + "score": 1.0, + "content": "analysis revealed interesting behavior. As an example, in the Cleanup video available here: https:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "//youtu.be/iH_V5WKQxmo a single agent (shown in purple) was trained with the social influence", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 673, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 104, + 673, + 506, + 689 + ], + "score": 1.0, + "content": "reward. We see that unlike the other agents, which continue to randomly move and explore while waiting", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 686, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 505, + 699 + ], + "score": 1.0, + "content": "for apples to spawn, the influencer has a strange economy of motion; it only moves on the map when it is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 697, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "pursuing an apple, then stops. Interestingly, examining the trajectory reveals that the influencer uses only", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "score": 1.0, + "content": "two moves to explore the map: turn left, which turns the agent in place without traversing the map, and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 718, + 374, + 731 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 374, + 731 + ], + "score": 1.0, + "content": "move right, which moves the agent one square to the right on the map.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 620, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 185, + 191 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 185, + 191 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 185, + 191 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 185, + 191 + ], + "score": 0.963, + "type": "image", + "image_path": "15faf4dfa88bd78c4d0e9ddf505f7cecc97572565601a0318358c9465b4a4695.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 185, + 191 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 201, + 188, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 200, + 191, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 191, + 213 + ], + "score": 1.0, + "content": "Figure 7: A mo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 211, + 190, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 190, + 223 + ], + "score": 1.0, + "content": "ment of high influ-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 222, + 190, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 190, + 235 + ], + "score": 1.0, + "content": "ence when the pur-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 232, + 190, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 190, + 247 + ], + "score": 1.0, + "content": "ple influencer sig-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 244, + 190, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 190, + 257 + ], + "score": 1.0, + "content": "nals the presence of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 255, + 189, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 189, + 267 + ], + "score": 1.0, + "content": "an apple outside the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 266, + 189, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 189, + 278 + ], + "score": 1.0, + "content": "yellow influencee’s", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 276, + 189, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 189, + 290 + ], + "score": 1.0, + "content": "field-of-view (yellow", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 288, + 162, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 162, + 301 + ], + "score": 1.0, + "content": "outlined box).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 196, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 195, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 195, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Why did the influencer learn to use only these two moves? We can see that the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 195, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 195, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "influencer agent only chooses to move right (i.e. traverse the map) when it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 194, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "pursuing an apple which is present. The rest of the time it simply turns left on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 195, + 115, + 227, + 127 + ], + "score": 1.0, + "content": "spot. At", + "type": "text" + }, + { + "bbox": [ + 228, + 116, + 252, + 126 + ], + "score": 0.89, + "content": "t = 4 9", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 115, + 505, + 127 + ], + "score": 1.0, + "content": ", there is a moment of high influence between the influencer and the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 195, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "yellow influencee, which is shown in Figure 7. The influencer has chosen to move", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 195, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 195, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "right towards an apple that is outside of the ego-centric field-of-view of the yellow", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "agent. Because the purple agent only moves when apples are available, this signals", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 195, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 195, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "to the yellow agent that an apple must be present above it which it cannot see.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 192, + 166, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 192, + 166, + 451, + 186 + ], + "score": 1.0, + "content": "This changes the yellow agent’s distribution over its planned action,", + "type": "text" + }, + { + "bbox": [ + 451, + 171, + 502, + 182 + ], + "score": 0.93, + "content": "p ( a _ { t } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { B } )", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 195, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 195, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "and allows the purple agent to gain influence. A similar moment occurs when the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 195, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 195, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "influencer signals to an agent that has been cleaning the river that no apples have", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 195, + 203, + 454, + 216 + ], + "spans": [ + { + "bbox": [ + 195, + 203, + 454, + 216 + ], + "score": 1.0, + "content": "appeared by continuing to turn left (see Figure 12 in the Appendix).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 196, + 220, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 195, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 195, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "In this example, the influencer agent learned to use its own actions as a sort of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 195, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 195, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "binary code, which signals the presence or absence of apples in the environment.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 195, + 242, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 195, + 242, + 505, + 253 + ], + "score": 1.0, + "content": "We also observe this effect in the influence agents in the Harvest task. This type of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 195, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 195, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "action-based communication could be likened to the bee waggle dance discovered", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 195, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 195, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "by von Frisch (1969). Thus, rewarding agents for increasing the mutual information", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 195, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 195, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "between their actions gave rise not only to cooperative behavior, but in this case, to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 195, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 195, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "emergent communication. These results further support the idea of using influence", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 195, + 297, + 377, + 309 + ], + "spans": [ + { + "bbox": [ + 195, + 297, + 377, + 309 + ], + "score": 1.0, + "content": "as a reward for training agents to communicate.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 319, + 295, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 297, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 297, + 332 + ], + "score": 1.0, + "content": "4.2 INFLUENCE THROUGH COMMUNICATION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "Figures 6(b) and 6(e) show the results of training the agents to use an explicit communication channel, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 345, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 506, + 357 + ], + "score": 1.0, + "content": "its effect on their collective reward. In this case, the ablated baseline is a model that has the same structure", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 314, + 368 + ], + "score": 1.0, + "content": "as in Figure 2, but in which the communication policy", + "type": "text" + }, + { + "bbox": [ + 314, + 358, + 325, + 367 + ], + "score": 0.85, + "content": "\\pi _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "is trained only with environmental reward. We", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "observe that the agents which are trained to use the communication channel with additional social influence", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "reward achieve significantly higher collective reward in both games. In fact, in the case of Cleanup, we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 148, + 402 + ], + "score": 1.0, + "content": "found that", + "type": "text" + }, + { + "bbox": [ + 149, + 390, + 172, + 399 + ], + "score": 0.89, + "content": "\\alpha = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "in the optimal hyperparameter settings, meaning that it was most effective to train the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 400, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 412 + ], + "score": 1.0, + "content": "communication head with zero extrinsic or environmental reward (see Table 2 in the Appendix). This", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 410, + 507, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 507, + 425 + ], + "score": 1.0, + "content": "suggests that influence alone can be a sufficient mechanism for training an effective communication policy.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 428, + 506, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "To analyze the communication behaviour learned by the agents, we introduce three metrics. Speaker", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 437, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 244, + 453 + ], + "score": 1.0, + "content": "consistency, is a normalized score", + "type": "text" + }, + { + "bbox": [ + 244, + 439, + 273, + 451 + ], + "score": 0.88, + "content": "\\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 437, + 396, + 453 + ], + "score": 1.0, + "content": "which assesses the entropy of", + "type": "text" + }, + { + "bbox": [ + 396, + 438, + 434, + 451 + ], + "score": 0.94, + "content": "p ( a ^ { k } | m ^ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 437, + 453, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 454, + 439, + 493, + 451 + ], + "score": 0.94, + "content": "p ( m ^ { k } | a ^ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 437, + 506, + 453 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "determine how consistently a speaker agent emits a particular symbol when it takes a particular action,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "and vice versa (the formula is given in Appendix Section 6.3.4). We expect this measure to be high if,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "for example, the speaker always emits the same symbol when it is cleaning the river. We also introduce", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 102, + 483, + 509, + 512 + ], + "spans": [ + { + "bbox": [ + 102, + 483, + 177, + 512 + ], + "score": 1.0, + "content": "two measures of i(1) symbol/action", + "type": "text" + }, + { + "bbox": [ + 177, + 493, + 246, + 506 + ], + "score": 0.92, + "content": "\\mathbf { I C } = I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 483, + 509, + 512 + ], + "score": 1.0, + "content": "dination (IC), which are both measures of mutual information (MI): measures the MI between the influencer/speaker’s symbol and the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 504, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 317, + 522 + ], + "score": 1.0, + "content": "influencee/listener’s next action, and (2) action/action", + "type": "text" + }, + { + "bbox": [ + 317, + 505, + 386, + 519 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathbf { I C } = I ( a _ { t } ^ { A } ; a _ { t + 1 } ^ { B } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 504, + 506, + 522 + ], + "score": 1.0, + "content": "measures the MI between the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 516, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 530 + ], + "score": 1.0, + "content": "influencer’s action and the influencee’s action in the next timestep. To compute these measures we first", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "score": 1.0, + "content": "average over all trajectory steps, then take the maximum value between any two agents, to determine if", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "any pair of agents are coordinating. Note that these measures are all instantaneous, as they consider only", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "short-term dependencies across two consecutive timesteps, and cannot capture if an agent communicates", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "influential compositional messages, i.e. information that requires several consecutive symbols to transmit", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 573, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 286, + 584 + ], + "score": 1.0, + "content": "and only then affects the other agents behavior.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 207, + 589, + 502, + 676 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 589, + 198, + 731 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 588, + 198, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 198, + 601 + ], + "score": 1.0, + "content": "Figure 8 presents the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 599, + 199, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 199, + 612 + ], + "score": 1.0, + "content": "results. The speaker", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 611, + 199, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 199, + 622 + ], + "score": 1.0, + "content": "consistencies metric re-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 622, + 198, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 198, + 633 + ], + "score": 1.0, + "content": "veals that agents trained", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 632, + 199, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 199, + 644 + ], + "score": 1.0, + "content": "with the influence", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 644, + 198, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 198, + 654 + ], + "score": 1.0, + "content": "reward communicate", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 655, + 198, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 198, + 666 + ], + "score": 1.0, + "content": "less ambiguously about", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 666, + 199, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 199, + 677 + ], + "score": 1.0, + "content": "their own actions, indi-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 677, + 199, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 199, + 688 + ], + "score": 1.0, + "content": "cating that the emergent", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 688, + 198, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 198, + 699 + ], + "score": 1.0, + "content": "communication is more", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 699, + 199, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 199, + 710 + ], + "score": 1.0, + "content": "meaningful. The instan-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 710, + 198, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 198, + 721 + ], + "score": 1.0, + "content": "taneous coordination", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 720, + 198, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 198, + 732 + ], + "score": 1.0, + "content": "metrics demonstrate", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 58 + }, + { + "type": "image_body", + "bbox": [ + 207, + 589, + 502, + 676 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 207, + 589, + 502, + 676 + ], + "spans": [ + { + "bbox": [ + 207, + 589, + 502, + 676 + ], + "score": 0.966, + "type": "image", + "image_path": "cfabcddf1dd58ea122c746b3e02eae517a3782b403a73cf07bb8bc425f0202f7.jpg" + } + ] + } + ], + "index": 66, + "virtual_lines": [ + { + "bbox": [ + 207, + 589, + 502, + 618.0 + ], + "spans": [], + "index": 65 + }, + { + "bbox": [ + 207, + 618.0, + 502, + 647.0 + ], + "spans": [], + "index": 66 + }, + { + "bbox": [ + 207, + 647.0, + 502, + 676.0 + ], + "spans": [], + "index": 67 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 205, + 685, + 506, + 718 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 205, + 684, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 205, + 684, + 506, + 697 + ], + "score": 1.0, + "content": "Figure 8: Metrics describing the quality of learned communication protocols.", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 205, + 695, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 205, + 695, + 506, + 707 + ], + "score": 1.0, + "content": "The models trained with influence reward exhibit more consistent communica-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 205, + 707, + 492, + 719 + ], + "spans": [ + { + "bbox": [ + 205, + 707, + 492, + 719 + ], + "score": 1.0, + "content": "tion and more coordination, especially in moments where influence is high.", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 69 + } + ], + "index": 66 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 298, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 298, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 185, + 191 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 185, + 191 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 185, + 191 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 185, + 191 + ], + "score": 0.963, + "type": "image", + "image_path": "15faf4dfa88bd78c4d0e9ddf505f7cecc97572565601a0318358c9465b4a4695.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 185, + 191 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 201, + 188, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 200, + 191, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 191, + 213 + ], + "score": 1.0, + "content": "Figure 7: A mo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 211, + 190, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 190, + 223 + ], + "score": 1.0, + "content": "ment of high influ-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 222, + 190, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 190, + 235 + ], + "score": 1.0, + "content": "ence when the pur-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 232, + 190, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 190, + 247 + ], + "score": 1.0, + "content": "ple influencer sig-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 244, + 190, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 190, + 257 + ], + "score": 1.0, + "content": "nals the presence of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 255, + 189, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 189, + 267 + ], + "score": 1.0, + "content": "an apple outside the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 266, + 189, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 189, + 278 + ], + "score": 1.0, + "content": "yellow influencee’s", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 276, + 189, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 189, + 290 + ], + "score": 1.0, + "content": "field-of-view (yellow", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 288, + 162, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 162, + 301 + ], + "score": 1.0, + "content": "outlined box).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 196, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 195, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 195, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Why did the influencer learn to use only these two moves? We can see that the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 195, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 195, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "influencer agent only chooses to move right (i.e. traverse the map) when it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 194, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "pursuing an apple which is present. The rest of the time it simply turns left on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 195, + 115, + 227, + 127 + ], + "score": 1.0, + "content": "spot. At", + "type": "text" + }, + { + "bbox": [ + 228, + 116, + 252, + 126 + ], + "score": 0.89, + "content": "t = 4 9", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 115, + 505, + 127 + ], + "score": 1.0, + "content": ", there is a moment of high influence between the influencer and the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 195, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "yellow influencee, which is shown in Figure 7. The influencer has chosen to move", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 195, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 195, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "right towards an apple that is outside of the ego-centric field-of-view of the yellow", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 195, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "agent. Because the purple agent only moves when apples are available, this signals", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 195, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 195, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "to the yellow agent that an apple must be present above it which it cannot see.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 192, + 166, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 192, + 166, + 451, + 186 + ], + "score": 1.0, + "content": "This changes the yellow agent’s distribution over its planned action,", + "type": "text" + }, + { + "bbox": [ + 451, + 171, + 502, + 182 + ], + "score": 0.93, + "content": "p ( a _ { t } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { B } )", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 195, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 195, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "and allows the purple agent to gain influence. A similar moment occurs when the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 195, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 195, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "influencer signals to an agent that has been cleaning the river that no apples have", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 195, + 203, + 454, + 216 + ], + "spans": [ + { + "bbox": [ + 195, + 203, + 454, + 216 + ], + "score": 1.0, + "content": "appeared by continuing to turn left (see Figure 12 in the Appendix).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15.5, + "bbox_fs": [ + 192, + 82, + 506, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 196, + 220, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 195, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 195, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "In this example, the influencer agent learned to use its own actions as a sort of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 195, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 195, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "binary code, which signals the presence or absence of apples in the environment.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 195, + 242, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 195, + 242, + 505, + 253 + ], + "score": 1.0, + "content": "We also observe this effect in the influence agents in the Harvest task. This type of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 195, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 195, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "action-based communication could be likened to the bee waggle dance discovered", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 195, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 195, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "by von Frisch (1969). Thus, rewarding agents for increasing the mutual information", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 195, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 195, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "between their actions gave rise not only to cooperative behavior, but in this case, to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 195, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 195, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "emergent communication. These results further support the idea of using influence", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 195, + 297, + 377, + 309 + ], + "spans": [ + { + "bbox": [ + 195, + 297, + 377, + 309 + ], + "score": 1.0, + "content": "as a reward for training agents to communicate.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 195, + 220, + 506, + 309 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 319, + 295, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 297, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 297, + 332 + ], + "score": 1.0, + "content": "4.2 INFLUENCE THROUGH COMMUNICATION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "Figures 6(b) and 6(e) show the results of training the agents to use an explicit communication channel, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 345, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 506, + 357 + ], + "score": 1.0, + "content": "its effect on their collective reward. In this case, the ablated baseline is a model that has the same structure", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 314, + 368 + ], + "score": 1.0, + "content": "as in Figure 2, but in which the communication policy", + "type": "text" + }, + { + "bbox": [ + 314, + 358, + 325, + 367 + ], + "score": 0.85, + "content": "\\pi _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "is trained only with environmental reward. We", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "observe that the agents which are trained to use the communication channel with additional social influence", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "reward achieve significantly higher collective reward in both games. In fact, in the case of Cleanup, we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 148, + 402 + ], + "score": 1.0, + "content": "found that", + "type": "text" + }, + { + "bbox": [ + 149, + 390, + 172, + 399 + ], + "score": 0.89, + "content": "\\alpha = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "in the optimal hyperparameter settings, meaning that it was most effective to train the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 400, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 412 + ], + "score": 1.0, + "content": "communication head with zero extrinsic or environmental reward (see Table 2 in the Appendix). This", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 410, + 507, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 507, + 425 + ], + "score": 1.0, + "content": "suggests that influence alone can be a sufficient mechanism for training an effective communication policy.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 334, + 507, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 428, + 506, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "To analyze the communication behaviour learned by the agents, we introduce three metrics. Speaker", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 437, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 244, + 453 + ], + "score": 1.0, + "content": "consistency, is a normalized score", + "type": "text" + }, + { + "bbox": [ + 244, + 439, + 273, + 451 + ], + "score": 0.88, + "content": "\\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 437, + 396, + 453 + ], + "score": 1.0, + "content": "which assesses the entropy of", + "type": "text" + }, + { + "bbox": [ + 396, + 438, + 434, + 451 + ], + "score": 0.94, + "content": "p ( a ^ { k } | m ^ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 437, + 453, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 454, + 439, + 493, + 451 + ], + "score": 0.94, + "content": "p ( m ^ { k } | a ^ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 437, + 506, + 453 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 450, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 506, + 463 + ], + "score": 1.0, + "content": "determine how consistently a speaker agent emits a particular symbol when it takes a particular action,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "and vice versa (the formula is given in Appendix Section 6.3.4). We expect this measure to be high if,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "for example, the speaker always emits the same symbol when it is cleaning the river. We also introduce", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 102, + 483, + 509, + 512 + ], + "spans": [ + { + "bbox": [ + 102, + 483, + 177, + 512 + ], + "score": 1.0, + "content": "two measures of i(1) symbol/action", + "type": "text" + }, + { + "bbox": [ + 177, + 493, + 246, + 506 + ], + "score": 0.92, + "content": "\\mathbf { I C } = I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 483, + 509, + 512 + ], + "score": 1.0, + "content": "dination (IC), which are both measures of mutual information (MI): measures the MI between the influencer/speaker’s symbol and the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 504, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 317, + 522 + ], + "score": 1.0, + "content": "influencee/listener’s next action, and (2) action/action", + "type": "text" + }, + { + "bbox": [ + 317, + 505, + 386, + 519 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathbf { I C } = I ( a _ { t } ^ { A } ; a _ { t + 1 } ^ { B } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 504, + 506, + 522 + ], + "score": 1.0, + "content": "measures the MI between the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 516, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 530 + ], + "score": 1.0, + "content": "influencer’s action and the influencee’s action in the next timestep. To compute these measures we first", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "score": 1.0, + "content": "average over all trajectory steps, then take the maximum value between any two agents, to determine if", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "any pair of agents are coordinating. Note that these measures are all instantaneous, as they consider only", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "short-term dependencies across two consecutive timesteps, and cannot capture if an agent communicates", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "influential compositional messages, i.e. information that requires several consecutive symbols to transmit", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 573, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 286, + 584 + ], + "score": 1.0, + "content": "and only then affects the other agents behavior.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45, + "bbox_fs": [ + 102, + 428, + 509, + 584 + ] + }, + { + "type": "image", + "bbox": [ + 207, + 589, + 502, + 676 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 106, + 589, + 198, + 731 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 588, + 198, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 198, + 601 + ], + "score": 1.0, + "content": "Figure 8 presents the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 599, + 199, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 199, + 612 + ], + "score": 1.0, + "content": "results. The speaker", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 611, + 199, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 199, + 622 + ], + "score": 1.0, + "content": "consistencies metric re-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 622, + 198, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 198, + 633 + ], + "score": 1.0, + "content": "veals that agents trained", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 632, + 199, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 199, + 644 + ], + "score": 1.0, + "content": "with the influence", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 644, + 198, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 198, + 654 + ], + "score": 1.0, + "content": "reward communicate", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 655, + 198, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 198, + 666 + ], + "score": 1.0, + "content": "less ambiguously about", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 666, + 199, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 199, + 677 + ], + "score": 1.0, + "content": "their own actions, indi-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 677, + 199, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 199, + 688 + ], + "score": 1.0, + "content": "cating that the emergent", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 688, + 198, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 198, + 699 + ], + "score": 1.0, + "content": "communication is more", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 699, + 199, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 199, + 710 + ], + "score": 1.0, + "content": "meaningful. The instan-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 710, + 198, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 198, + 721 + ], + "score": 1.0, + "content": "taneous coordination", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 720, + 198, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 198, + 732 + ], + "score": 1.0, + "content": "metrics demonstrate", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 58 + }, + { + "type": "image_body", + "bbox": [ + 207, + 589, + 502, + 676 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 207, + 589, + 502, + 676 + ], + "spans": [ + { + "bbox": [ + 207, + 589, + 502, + 676 + ], + "score": 0.966, + "type": "image", + "image_path": "cfabcddf1dd58ea122c746b3e02eae517a3782b403a73cf07bb8bc425f0202f7.jpg" + } + ] + } + ], + "index": 66, + "virtual_lines": [ + { + "bbox": [ + 207, + 589, + 502, + 618.0 + ], + "spans": [], + "index": 65 + }, + { + "bbox": [ + 207, + 618.0, + 502, + 647.0 + ], + "spans": [], + "index": 66 + }, + { + "bbox": [ + 207, + 647.0, + 502, + 676.0 + ], + "spans": [], + "index": 67 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 205, + 685, + 506, + 718 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 205, + 684, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 205, + 684, + 506, + 697 + ], + "score": 1.0, + "content": "Figure 8: Metrics describing the quality of learned communication protocols.", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 205, + 695, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 205, + 695, + 506, + 707 + ], + "score": 1.0, + "content": "The models trained with influence reward exhibit more consistent communica-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 205, + 707, + 492, + 719 + ], + "spans": [ + { + "bbox": [ + 205, + 707, + 492, + 719 + ], + "score": 1.0, + "content": "tion and more coordination, especially in moments where influence is high.", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 69 + } + ], + "index": 66 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "that the baseline agents trained without influence reward show almost no signs of co-ordinating behavior", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "with communication, i.e. speakers saying A and listeners doing B consistently. This result is aligned with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "both theoretical results in cheap-talk literature (Crawford & Sobel, 1982), and recent empirical results in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 397, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 397, + 128 + ], + "score": 1.0, + "content": "MARL (e.g. (Foerster et al., 2016; Lazaridou et al., 2018; Cao et al., 2018)).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 504, + 144 + ], + "score": 1.0, + "content": "In contrast, we do see highly coordinated behavior between influence agents, but only when we limit the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "analysis to timesteps on which influence was high (cf. influential moments in Figure 8). If we inspect the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "results for agents trained with influence on the two tasks, a common pattern emerges: influence is sparse in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 400, + 178 + ], + "score": 1.0, + "content": "time. An agent’s influence is only greater than its mean influence in less than", + "type": "text" + }, + { + "bbox": [ + 400, + 165, + 419, + 176 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "of timesteps. Because", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "the listener agent is not compelled to listen to any given speaker, listeners selectively listen to a speaker only", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "when it is beneficial, and influence cannot occur all the time. Only when the listener decides to change", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "its action based on the speaker’s message does influence occur, and in these moments we observe high", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 206, + 507, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 208, + 155, + 222 + ], + "score": 0.92, + "content": "I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 206, + 507, + 226 + ], + "score": 1.0, + "content": "; an effect that is lost when averaging over the entire trajectory. It appears the influencers", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "score": 1.0, + "content": "have learned a strategy of communicating meaningful information about their own actions, and gaining", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 392, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 392, + 244 + ], + "score": 1.0, + "content": "influence when this becomes relevant enough for the listener to act upon it.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "Examining the relationship between the reward obtained by individual agents and the degree to which", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "they were influenced by other agents gives a compelling result: agents that are the most influenced also", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 292, + 282 + ], + "score": 1.0, + "content": "achieve higher individual environmental reward,", + "type": "text" + }, + { + "bbox": [ + 293, + 269, + 306, + 280 + ], + "score": 0.87, + "content": "\\dot { E } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 268, + 506, + 282 + ], + "score": 1.0, + "content": ". We sampled 100 different experimental conditions", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "(i.e., hyper-parameters and random seeds) for both games, collected the influence and individual rewards,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 303 + ], + "score": 1.0, + "content": "normalized them across the 5 agents in each condition, and correlated the resulting list of values. We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 507, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 507, + 316 + ], + "score": 1.0, + "content": "found that agents who are more often influenced tend to achieve higher task reward in both Cleanup,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 313, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 137, + 325 + ], + "score": 0.87, + "content": "\\rho = . 6 7", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 313, + 140, + 326 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 141, + 314, + 176, + 325 + ], + "score": 0.84, + "content": "p { < } 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 313, + 232, + 326 + ], + "score": 1.0, + "content": ", and Harvest,", + "type": "text" + }, + { + "bbox": [ + 232, + 314, + 263, + 325 + ], + "score": 0.87, + "content": "\\rho = . 3 4", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 313, + 266, + 326 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 267, + 314, + 302, + 325 + ], + "score": 0.81, + "content": "p { < } 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 313, + 507, + 326 + ], + "score": 1.0, + "content": ". This supports the hypothesis stated in Section 2.3:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "in order to gain influence from another agent by communicating with it, the communication message", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 336, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 506, + 347 + ], + "score": 1.0, + "content": "should contain information that helps the listener maximize its own environmental reward. Since better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "score": 1.0, + "content": "listeners/influencees are more successful in terms of task reward, we have evidence that useful information", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 358, + 201, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 201, + 369 + ], + "score": 1.0, + "content": "was transmitted to them.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 106, + 383, + 307, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 308, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 308, + 394 + ], + "score": 1.0, + "content": "4.3 INFLUENCE VIA MODELING OTHER AGENTS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 399, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 504, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 504, + 411 + ], + "score": 1.0, + "content": "Finally, we investigate whether the influence reward is still effective when computed without a centralised", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "controller, but rather through each agent’s own internal Model of Other Agents (MOA) network. In this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 421, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 264, + 434 + ], + "score": 1.0, + "content": "case, we extend the training period from", + "type": "text" + }, + { + "bbox": [ + 264, + 421, + 288, + 432 + ], + "score": 0.88, + "content": "3 \\cdot 1 0 ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 421, + 321, + 434 + ], + "score": 1.0, + "content": "steps to", + "type": "text" + }, + { + "bbox": [ + 321, + 421, + 344, + 432 + ], + "score": 0.88, + "content": "5 \\cdot 1 0 ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 421, + 506, + 434 + ], + "score": 1.0, + "content": ", in order to give the MOA model time to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 433, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 445 + ], + "score": 1.0, + "content": "train. We also allow the policy LSTM to condition on the actions of other agents in the last timestep. We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "compare against an ablated version of this architecture (shown in Figure 3), which does not use the output", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "of the MOA module to compute a reward; rather, the MOA module can be thought of as an unsupervised", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "auxiliary task that may help the model to learn a better shared embedding layer, encouraging it to encode", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 476, + 324, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 324, + 489 + ], + "score": 1.0, + "content": "information relevant to predicting other agents’ behavior.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "Figures 6(c) and 6(f) show the collective reward obtained for agents trained with a MOA module. While", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "we see that the auxiliary task does help to improve reward over the A3C baseline, the influence agent gets", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "consistently higher collective reward. Impressively, for Cleanup, the MOA model scores higher than the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 504, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 504, + 538 + ], + "score": 1.0, + "content": "original influence agents computed using the centralised controller (CC). As shown in Figure 6(c), the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "MOA baseline also achieves high collective reward, suggesting that the auxiliary task of modeling other", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 548, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 506, + 561 + ], + "score": 1.0, + "content": "agents helps the MOA agents cooperate more effectively in Cleanup. Further, the independent design of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "the MOA method allows each agent to influence every other agent, thus generating more reward signal and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 570, + 330, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 330, + 583 + ], + "score": 1.0, + "content": "a greater chance to develop two-way cooperative behavior.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 507, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 507, + 600 + ], + "score": 1.0, + "content": "Table 4 of the Appendix gives the final collective reward obtained by each model for all three experiments.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Interestingly, several influence models are able to achieve higher collective reward than the previous", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 609, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 621 + ], + "score": 1.0, + "content": "state-of-the-art scores for these environments (275 for Cleanup and 750 for Harvest) (Hughes et al., 2018).", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "This is compelling, given that previous work relied on the assumption that agents could view one another’s", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 631, + 483, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 483, + 644 + ], + "score": 1.0, + "content": "rewards; we make no such assumption, instead relying only on agents viewing each other’s actions.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + }, + { + "type": "title", + "bbox": [ + 109, + 657, + 284, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 286, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 286, + 672 + ], + "score": 1.0, + "content": "5 DISCUSSION AND CONCLUSIONS", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "The experiments above have demonstrated that an intrinsic social reward based on having causal influence", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "on the actions of other agents consistently improves cooperation and leads to higher collective return in", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "the MA social dilemmas under investigation. In some cases, the influence reward drove agents to learn", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "an emergent communication protocol via their actions. This is compelling, and confirms the connection", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 472, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 472, + 733 + ], + "score": 1.0, + "content": "between maximizing influence and maximizing the mutual information between agents’ actions.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 126 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "that the baseline agents trained without influence reward show almost no signs of co-ordinating behavior", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "with communication, i.e. speakers saying A and listeners doing B consistently. This result is aligned with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "both theoretical results in cheap-talk literature (Crawford & Sobel, 1982), and recent empirical results in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 397, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 397, + 128 + ], + "score": 1.0, + "content": "MARL (e.g. (Foerster et al., 2016; Lazaridou et al., 2018; Cao et al., 2018)).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 83, + 506, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 504, + 144 + ], + "score": 1.0, + "content": "In contrast, we do see highly coordinated behavior between influence agents, but only when we limit the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "analysis to timesteps on which influence was high (cf. influential moments in Figure 8). If we inspect the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "results for agents trained with influence on the two tasks, a common pattern emerges: influence is sparse in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 400, + 178 + ], + "score": 1.0, + "content": "time. An agent’s influence is only greater than its mean influence in less than", + "type": "text" + }, + { + "bbox": [ + 400, + 165, + 419, + 176 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "of timesteps. Because", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "the listener agent is not compelled to listen to any given speaker, listeners selectively listen to a speaker only", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "when it is beneficial, and influence cannot occur all the time. Only when the listener decides to change", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "its action based on the speaker’s message does influence occur, and in these moments we observe high", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 206, + 507, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 208, + 155, + 222 + ], + "score": 0.92, + "content": "I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 206, + 507, + 226 + ], + "score": 1.0, + "content": "; an effect that is lost when averaging over the entire trajectory. It appears the influencers", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 234 + ], + "score": 1.0, + "content": "have learned a strategy of communicating meaningful information about their own actions, and gaining", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 392, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 392, + 244 + ], + "score": 1.0, + "content": "influence when this becomes relevant enough for the listener to act upon it.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 132, + 507, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "Examining the relationship between the reward obtained by individual agents and the degree to which", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "they were influenced by other agents gives a compelling result: agents that are the most influenced also", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 292, + 282 + ], + "score": 1.0, + "content": "achieve higher individual environmental reward,", + "type": "text" + }, + { + "bbox": [ + 293, + 269, + 306, + 280 + ], + "score": 0.87, + "content": "\\dot { E } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 268, + 506, + 282 + ], + "score": 1.0, + "content": ". We sampled 100 different experimental conditions", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "(i.e., hyper-parameters and random seeds) for both games, collected the influence and individual rewards,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 303 + ], + "score": 1.0, + "content": "normalized them across the 5 agents in each condition, and correlated the resulting list of values. We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 507, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 507, + 316 + ], + "score": 1.0, + "content": "found that agents who are more often influenced tend to achieve higher task reward in both Cleanup,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 313, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 137, + 325 + ], + "score": 0.87, + "content": "\\rho = . 6 7", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 313, + 140, + 326 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 141, + 314, + 176, + 325 + ], + "score": 0.84, + "content": "p { < } 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 313, + 232, + 326 + ], + "score": 1.0, + "content": ", and Harvest,", + "type": "text" + }, + { + "bbox": [ + 232, + 314, + 263, + 325 + ], + "score": 0.87, + "content": "\\rho = . 3 4", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 313, + 266, + 326 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 267, + 314, + 302, + 325 + ], + "score": 0.81, + "content": "p { < } 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 313, + 507, + 326 + ], + "score": 1.0, + "content": ". This supports the hypothesis stated in Section 2.3:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "in order to gain influence from another agent by communicating with it, the communication message", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 336, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 506, + 347 + ], + "score": 1.0, + "content": "should contain information that helps the listener maximize its own environmental reward. Since better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "score": 1.0, + "content": "listeners/influencees are more successful in terms of task reward, we have evidence that useful information", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 358, + 201, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 201, + 369 + ], + "score": 1.0, + "content": "was transmitted to them.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 248, + 507, + 369 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 383, + 307, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 308, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 308, + 394 + ], + "score": 1.0, + "content": "4.3 INFLUENCE VIA MODELING OTHER AGENTS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 399, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 504, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 504, + 411 + ], + "score": 1.0, + "content": "Finally, we investigate whether the influence reward is still effective when computed without a centralised", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "controller, but rather through each agent’s own internal Model of Other Agents (MOA) network. In this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 421, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 264, + 434 + ], + "score": 1.0, + "content": "case, we extend the training period from", + "type": "text" + }, + { + "bbox": [ + 264, + 421, + 288, + 432 + ], + "score": 0.88, + "content": "3 \\cdot 1 0 ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 421, + 321, + 434 + ], + "score": 1.0, + "content": "steps to", + "type": "text" + }, + { + "bbox": [ + 321, + 421, + 344, + 432 + ], + "score": 0.88, + "content": "5 \\cdot 1 0 ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 421, + 506, + 434 + ], + "score": 1.0, + "content": ", in order to give the MOA model time to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 433, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 445 + ], + "score": 1.0, + "content": "train. We also allow the policy LSTM to condition on the actions of other agents in the last timestep. We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "compare against an ablated version of this architecture (shown in Figure 3), which does not use the output", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "of the MOA module to compute a reward; rather, the MOA module can be thought of as an unsupervised", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "auxiliary task that may help the model to learn a better shared embedding layer, encouraging it to encode", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 476, + 324, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 324, + 489 + ], + "score": 1.0, + "content": "information relevant to predicting other agents’ behavior.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 400, + 506, + 489 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "Figures 6(c) and 6(f) show the collective reward obtained for agents trained with a MOA module. While", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "we see that the auxiliary task does help to improve reward over the A3C baseline, the influence agent gets", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "consistently higher collective reward. Impressively, for Cleanup, the MOA model scores higher than the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 526, + 504, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 504, + 538 + ], + "score": 1.0, + "content": "original influence agents computed using the centralised controller (CC). As shown in Figure 6(c), the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "MOA baseline also achieves high collective reward, suggesting that the auxiliary task of modeling other", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 548, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 506, + 561 + ], + "score": 1.0, + "content": "agents helps the MOA agents cooperate more effectively in Cleanup. Further, the independent design of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "the MOA method allows each agent to influence every other agent, thus generating more reward signal and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 570, + 330, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 330, + 583 + ], + "score": 1.0, + "content": "a greater chance to develop two-way cooperative behavior.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 493, + 506, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 507, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 507, + 600 + ], + "score": 1.0, + "content": "Table 4 of the Appendix gives the final collective reward obtained by each model for all three experiments.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Interestingly, several influence models are able to achieve higher collective reward than the previous", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 609, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 621 + ], + "score": 1.0, + "content": "state-of-the-art scores for these environments (275 for Cleanup and 750 for Harvest) (Hughes et al., 2018).", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "This is compelling, given that previous work relied on the assumption that agents could view one another’s", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 631, + 483, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 483, + 644 + ], + "score": 1.0, + "content": "rewards; we make no such assumption, instead relying only on agents viewing each other’s actions.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 586, + 507, + 644 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 657, + 284, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 286, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 286, + 672 + ], + "score": 1.0, + "content": "5 DISCUSSION AND CONCLUSIONS", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "The experiments above have demonstrated that an intrinsic social reward based on having causal influence", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "on the actions of other agents consistently improves cooperation and leads to higher collective return in", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "the MA social dilemmas under investigation. In some cases, the influence reward drove agents to learn", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "an emergent communication protocol via their actions. This is compelling, and confirms the connection", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 472, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 472, + 733 + ], + "score": 1.0, + "content": "between maximizing influence and maximizing the mutual information between agents’ actions.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 677, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "However, it is important to consider the limitations of the influence reward. Whether it will always give rise", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "to cooperative behavior may depend on the specifics of the environment, task, and the trade-off between", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "environmental and influence reward. Although influence is arguably necessary for cooperation (e.g. two", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "score": 1.0, + "content": "agents cooperating to lift a box would have a high degree of influence between their actions), it may not", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "be sufficient, in that it may be possible to influence another agent without helping it. For example, it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "possible that agents could have gained influence in the tasks studied here by threatening to attack other", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "agents with their fining beam. We believe this type of behavior did not emerge because communicating", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "information represents the cheapest and most effective way to gain influence. Influencers do not have to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 484, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 484, + 183 + ], + "score": 1.0, + "content": "sacrifice much in terms of their own environmental reward in order to communicate to other agents.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Rewarding influence over an explicit communication channel may not be subject to this limitation, because", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "influential communication may be inherently beneficial to the listener (at least in the case where listeners", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "and speakers interact repeatedly). Since listeners can easily ignore communication messages if they do", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "not help to obtain environmental reward, a speaker must transmit valuable information in order to gain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "influence through communication. There is no advantage to the speaker for communicating unreliably,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "because it would lose influence with the listener over time (although this is no longer guaranteed in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 251, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 267 + ], + "score": 1.0, + "content": "one-shot interactions). Indeed, our results reveal that agents benefit from being influenced by (listening", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "to) communication messages by obtaining higher individual reward, suggesting that the messages contain", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "valuable information. Further, we found that the communication protocols learned via influence reward", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "were more meaningful, and that the influence reward allowed agents to obtain higher collective return.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 310 + ], + "score": 1.0, + "content": "Therefore, we suggest that influence could be a promising way to train emergent communication protocols", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 179, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 179, + 321 + ], + "score": 1.0, + "content": "in various settings.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "Finally, we have shown that influence can be computed by augmenting agents with an internal model that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "predicts the actions of other agents, and using this MOA model to simulate the effect of an agent’s actions", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "on others. This represents an important step forward in multi-agent intrinsic social motivation, because", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "it implies that the influence reward can be computed without having access to another agent’s reward", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 369, + 279, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 279, + 382 + ], + "score": 1.0, + "content": "function, or requiring a centralised controller.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 392, + 194, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 196, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 196, + 405 + ], + "score": 1.0, + "content": "5.1 FUTURE WORK", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 410, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 506, + 423 + ], + "score": 1.0, + "content": "Using counterfactuals to allow agents to understand the effects of their actions on other agents could be a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "promising approach with a number of extensions. Perhaps agents could use counterfactuals to develop", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "a form of ‘empathy‘, by simulating how their actions affect another agent’s value function. Or, social", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "influence could be used to drive coordinated behavior in robots attempting to do cooperative manipulation", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "score": 1.0, + "content": "and control tasks. Finally, if we view multi-agent networks as a single agent, influence could be used as a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "regularizer to encourage different modules of the network to integrate information from other networks; for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 477, + 343, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 343, + 488 + ], + "score": 1.0, + "content": "example, perhaps it could prevent collapse in hierarchical RL.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 502, + 172, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 173, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 173, + 516 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Sean L Barton, Nicholas R Waytowich, Erin Zaroukian, and Derrik E Asher. Measuring collaborative", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 533, + 490, + 545 + ], + "spans": [ + { + "bbox": [ + 116, + 533, + 490, + 545 + ], + "score": 1.0, + "content": "emergent behavior in multi-agent reinforcement learning. arXiv preprint arXiv:1807.08663, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 552, + 503, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "Ben Bogin, Mor Geva, and Jonathan Berant. Emergence of communication in an interactive world with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 564, + 349, + 576 + ], + "spans": [ + { + "bbox": [ + 116, + 564, + 349, + 576 + ], + "score": 1.0, + "content": "consistent speakers. arXiv preprint arXiv:1809.00549, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 583, + 504, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 595, + 412, + 606 + ], + "spans": [ + { + "bbox": [ + 115, + 595, + 412, + 606 + ], + "score": 1.0, + "content": "communication through negotiation. arXiv preprint arXiv:1804.03980, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 503, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "score": 1.0, + "content": "Philippe Capdepuy, Daniel Polani, and Chrystopher L Nehaniv. Maximization of potential information", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 626, + 504, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 504, + 638 + ], + "score": 1.0, + "content": "flow as a universal utility for collective behaviour. In Artificial Life, 2007. ALIFE’07. IEEE Symposium", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 637, + 227, + 648 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 227, + 648 + ], + "score": 1.0, + "content": "on, pp. 207–213. Ieee, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 104, + 656, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 104, + 655, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 104, + 655, + 506, + 671 + ], + "score": 1.0, + "content": "Edward Choi, Angeliki Lazaridou, and Nando de Freitas. Compositional obverter communication learning", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 667, + 358, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 358, + 680 + ], + "score": 1.0, + "content": "from raw visual input. arXiv preprint arXiv:1804.02341, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 699 + ], + "score": 1.0, + "content": "Jacob W. Crandall, Mayada Oudah, Tennom Chenlinangjia, Fatimah Ishowo-Oloko, Sherief Abdallah,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Jean-Franc¸ois Bonnefon, Manuel Cebrian, Azim Shariff, Michael A. Goodrich, and Iyad Rahwan. ´", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "Cooperating with machines. CoRR, abs/1703.06207, 2017. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 721, + 178, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 178, + 731 + ], + "score": 1.0, + "content": "1703.06207.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "However, it is important to consider the limitations of the influence reward. Whether it will always give rise", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "to cooperative behavior may depend on the specifics of the environment, task, and the trade-off between", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "environmental and influence reward. Although influence is arguably necessary for cooperation (e.g. two", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "score": 1.0, + "content": "agents cooperating to lift a box would have a high degree of influence between their actions), it may not", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "be sufficient, in that it may be possible to influence another agent without helping it. For example, it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "possible that agents could have gained influence in the tasks studied here by threatening to attack other", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "agents with their fining beam. We believe this type of behavior did not emerge because communicating", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 172 + ], + "score": 1.0, + "content": "information represents the cheapest and most effective way to gain influence. Influencers do not have to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 484, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 484, + 183 + ], + "score": 1.0, + "content": "sacrifice much in terms of their own environmental reward in order to communicate to other agents.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 83, + 506, + 183 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Rewarding influence over an explicit communication channel may not be subject to this limitation, because", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "influential communication may be inherently beneficial to the listener (at least in the case where listeners", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "and speakers interact repeatedly). Since listeners can easily ignore communication messages if they do", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "not help to obtain environmental reward, a speaker must transmit valuable information in order to gain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "influence through communication. There is no advantage to the speaker for communicating unreliably,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "because it would lose influence with the listener over time (although this is no longer guaranteed in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 251, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 267 + ], + "score": 1.0, + "content": "one-shot interactions). Indeed, our results reveal that agents benefit from being influenced by (listening", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "to) communication messages by obtaining higher individual reward, suggesting that the messages contain", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "valuable information. Further, we found that the communication protocols learned via influence reward", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "were more meaningful, and that the influence reward allowed agents to obtain higher collective return.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 310 + ], + "score": 1.0, + "content": "Therefore, we suggest that influence could be a promising way to train emergent communication protocols", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 179, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 179, + 321 + ], + "score": 1.0, + "content": "in various settings.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 187, + 506, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "Finally, we have shown that influence can be computed by augmenting agents with an internal model that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "predicts the actions of other agents, and using this MOA model to simulate the effect of an agent’s actions", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "on others. This represents an important step forward in multi-agent intrinsic social motivation, because", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "it implies that the influence reward can be computed without having access to another agent’s reward", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 369, + 279, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 279, + 382 + ], + "score": 1.0, + "content": "function, or requiring a centralised controller.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 324, + 505, + 382 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 392, + 194, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 196, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 196, + 405 + ], + "score": 1.0, + "content": "5.1 FUTURE WORK", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 410, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 506, + 423 + ], + "score": 1.0, + "content": "Using counterfactuals to allow agents to understand the effects of their actions on other agents could be a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "promising approach with a number of extensions. Perhaps agents could use counterfactuals to develop", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "a form of ‘empathy‘, by simulating how their actions affect another agent’s value function. Or, social", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "influence could be used to drive coordinated behavior in robots attempting to do cooperative manipulation", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "score": 1.0, + "content": "and control tasks. Finally, if we view multi-agent networks as a single agent, influence could be used as a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "regularizer to encourage different modules of the network to integrate information from other networks; for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 477, + 343, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 343, + 488 + ], + "score": 1.0, + "content": "example, perhaps it could prevent collapse in hierarchical RL.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 409, + 506, + 488 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 502, + 172, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 173, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 173, + 516 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Sean L Barton, Nicholas R Waytowich, Erin Zaroukian, and Derrik E Asher. Measuring collaborative", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 533, + 490, + 545 + ], + "spans": [ + { + "bbox": [ + 116, + 533, + 490, + 545 + ], + "score": 1.0, + "content": "emergent behavior in multi-agent reinforcement learning. arXiv preprint arXiv:1807.08663, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 520, + 505, + 545 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 552, + 503, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "Ben Bogin, Mor Geva, and Jonathan Berant. Emergence of communication in an interactive world with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 564, + 349, + 576 + ], + "spans": [ + { + "bbox": [ + 116, + 564, + 349, + 576 + ], + "score": 1.0, + "content": "consistent speakers. arXiv preprint arXiv:1809.00549, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 552, + 505, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 583, + 504, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 595, + 412, + 606 + ], + "spans": [ + { + "bbox": [ + 115, + 595, + 412, + 606 + ], + "score": 1.0, + "content": "communication through negotiation. arXiv preprint arXiv:1804.03980, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 583, + 505, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 503, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "score": 1.0, + "content": "Philippe Capdepuy, Daniel Polani, and Chrystopher L Nehaniv. Maximization of potential information", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 626, + 504, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 504, + 638 + ], + "score": 1.0, + "content": "flow as a universal utility for collective behaviour. In Artificial Life, 2007. ALIFE’07. IEEE Symposium", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 637, + 227, + 648 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 227, + 648 + ], + "score": 1.0, + "content": "on, pp. 207–213. Ieee, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 614, + 505, + 648 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 656, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 104, + 655, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 104, + 655, + 506, + 671 + ], + "score": 1.0, + "content": "Edward Choi, Angeliki Lazaridou, and Nando de Freitas. Compositional obverter communication learning", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 667, + 358, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 358, + 680 + ], + "score": 1.0, + "content": "from raw visual input. arXiv preprint arXiv:1804.02341, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 104, + 655, + 506, + 680 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 699 + ], + "score": 1.0, + "content": "Jacob W. Crandall, Mayada Oudah, Tennom Chenlinangjia, Fatimah Ishowo-Oloko, Sherief Abdallah,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Jean-Franc¸ois Bonnefon, Manuel Cebrian, Azim Shariff, Michael A. Goodrich, and Iyad Rahwan. ´", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "Cooperating with machines. CoRR, abs/1703.06207, 2017. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 721, + 178, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 178, + 731 + ], + "score": 1.0, + "content": "1703.06207.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 688, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Vincent P Crawford and Joel Sobel. Strategic information transmission. Econometrica: Journal of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 284, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 284, + 105 + ], + "score": 1.0, + "content": "Econometric Society, pp. 1431–1451, 1982.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Sam Devlin, Logan Yliniemi, Daniel Kudenko, and Kagan Tumer. Potential-based difference rewards for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "multiagent reinforcement learning. In Proceedings of the 2014 international conference on Autonomous", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 146 + ], + "score": 1.0, + "content": "agents and multi-agent systems, pp. 165–172. International Foundation for Autonomous Agents and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 221, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 221, + 158 + ], + "score": 1.0, + "content": "Multiagent Systems, 2014.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 163, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "Heather J. Ferguson, Christoph Scheepers, and Anthony J. Sanford. Expectations in counterfactual and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 506, + 187 + ], + "score": 1.0, + "content": "theory of mind reasoning. Language and Cognitive Processes, 25(3):297–346, 2010. doi: 10.1080/", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 465, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 465, + 198 + ], + "score": 1.0, + "content": "01690960903041174. URL https://doi.org/10.1080/01690960903041174.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 506, + 238 + ], + "lines": [ + { + "bbox": [ + 104, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "communicate with deep multi-agent reinforcement learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 226, + 280, + 238 + ], + "spans": [ + { + "bbox": [ + 116, + 226, + 280, + 238 + ], + "score": 1.0, + "content": "Processing Systems, pp. 2137–2145, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 504, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "score": 1.0, + "content": "Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counter-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 255, + 411, + 268 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 411, + 268 + ], + "score": 1.0, + "content": "factual multi-agent policy gradients. arXiv preprint arXiv:1705.08926, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 506, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "Jakob Foerster, Richard Y Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "Learning with opponent-learning awareness. In Proceedings of the 17th International Conference on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Autonomous Agents and MultiAgent Systems, pp. 122–130. International Foundation for Autonomous", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 308, + 264, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 264, + 320 + ], + "score": 1.0, + "content": "Agents and Multiagent Systems, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 504, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "score": 1.0, + "content": "Sebastien Forestier and Pierre-Yves Oudeyer. A unified model of speech and tool use early development.´", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 337, + 428, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 428, + 350 + ], + "score": 1.0, + "content": "In 39th Annual Conference of the Cognitive Science Society (CogSci 2017), 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 504, + 378 + ], + "lines": [ + { + "bbox": [ + 104, + 353, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 104, + 353, + 506, + 368 + ], + "score": 1.0, + "content": "Felix A Gers, Jurgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm. ¨", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 365, + 141, + 379 + ], + "spans": [ + { + "bbox": [ + 114, + 365, + 141, + 379 + ], + "score": 1.0, + "content": "1999.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 398 + ], + "score": 1.0, + "content": "Christian Guckelsberger, Christoph Salge, and Julian Togelius. New and surprising ways to be mean.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 396, + 497, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 497, + 408 + ], + "score": 1.0, + "content": "adversarial npcs with coupled empowerment minimisation. arXiv preprint arXiv:1806.01387, 2018.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 422, + 427 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 421, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 421, + 428 + ], + "score": 1.0, + "content": "Yuval Noah Harari. Sapiens: A brief history of humankind. Random House, 2014.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 506, + 466 + ], + "lines": [ + { + "bbox": [ + 104, + 432, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 506, + 447 + ], + "score": 1.0, + "content": "Joseph Henrich. The Secret of Our Success: How culture is driving human evolution, domesticating", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 445, + 507, + 457 + ], + "spans": [ + { + "bbox": [ + 116, + 445, + 507, + 457 + ], + "score": 1.0, + "content": "our species, and making us smart. Princeton University Press, Princeton, NJ, 2015. URL http:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 455, + 344, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 344, + 468 + ], + "score": 1.0, + "content": "//press.princeton.edu/titles/10543.html.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 506, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 507, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 507, + 486 + ], + "score": 1.0, + "content": "Esther Herrmann, Josep Call, Mar´ıa Victoria Hernandez-Lloreda, Brian Hare, and Michael Tomasello. `", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 486, + 507, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 486, + 507, + 497 + ], + "score": 1.0, + "content": "Humans have evolved specialized skills of social cognition: The cultural intelligence hypothesis.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "Science, 317(5843):1360–1366, 2007. ISSN 0036-8075. doi: 10.1126/science.1146282. URL http:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 507, + 384, + 519 + ], + "spans": [ + { + "bbox": [ + 116, + 507, + 384, + 519 + ], + "score": 1.0, + "content": "//science.sciencemag.org/content/317/5843/1360.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 506, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "Edward Hughes, Joel Z Leibo, Matthew G Phillips, Karl Tuyls, Edgar A Due´nez-Guzm ˜ an, Antonio Garc ´ ´ıa", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 116, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Castaneda, Iain Dunning, Tina Zhu, Kevin R McKee, Raphael Koster, et al. Inequity aversion improves ˜", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "cooperation in intertemporal social dilemmas. In Advances in neural information processing systems", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 559, + 243, + 570 + ], + "spans": [ + { + "bbox": [ + 116, + 559, + 243, + 570 + ], + "score": 1.0, + "content": "(NIPS), Montreal, Canada, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "Alexander S Klyubin, Daniel Polani, and Chrystopher L Nehaniv. Empowerment: A universal agent-centric", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 586, + 507, + 603 + ], + "spans": [ + { + "bbox": [ + 114, + 586, + 507, + 603 + ], + "score": 1.0, + "content": "measure of control. In Evolutionary Computation, 2005. The 2005 IEEE Congress on, volume 1, pp.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 598, + 204, + 610 + ], + "spans": [ + { + "bbox": [ + 116, + 598, + 204, + 610 + ], + "score": 1.0, + "content": "128–135. IEEE, 2005.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 104, + 617, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 295, + 630 + ], + "score": 1.0, + "content": "Kevin N. Laland. Darwin’s unfinished symphony", + "type": "text" + }, + { + "bbox": [ + 295, + 620, + 301, + 627 + ], + "score": 0.33, + "content": "\\therefore", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 617, + 507, + 630 + ], + "score": 1.0, + "content": "how culture made the human mind / Kevin N. Laland.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 629, + 420, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 629, + 420, + 640 + ], + "score": 1.0, + "content": "Princeton University Press Princeton, 2017. ISBN 9781400884872 140088487.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 658, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 506, + 670 + ], + "score": 1.0, + "content": "communication from referential games with symbolic and pixel input. arXiv preprint arXiv:1804.03984,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 668, + 141, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 141, + 680 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "score": 1.0, + "content": "Joel Z Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. Multi-agent", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Autonomous Agents and MultiAgent Systems, pp. 464–473. International Foundation for Autonomous", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 265, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 265, + 732 + ], + "score": 1.0, + "content": "Agents and Multiagent Systems, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 298, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Vincent P Crawford and Joel Sobel. Strategic information transmission. Econometrica: Journal of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 284, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 284, + 105 + ], + "score": 1.0, + "content": "Econometric Society, pp. 1431–1451, 1982.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 107, + 82, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Sam Devlin, Logan Yliniemi, Daniel Kudenko, and Kagan Tumer. Potential-based difference rewards for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "multiagent reinforcement learning. In Proceedings of the 2014 international conference on Autonomous", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 146 + ], + "score": 1.0, + "content": "agents and multi-agent systems, pp. 165–172. International Foundation for Autonomous Agents and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 221, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 221, + 158 + ], + "score": 1.0, + "content": "Multiagent Systems, 2014.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 112, + 506, + 158 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 163, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "Heather J. Ferguson, Christoph Scheepers, and Anthony J. Sanford. Expectations in counterfactual and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 506, + 187 + ], + "score": 1.0, + "content": "theory of mind reasoning. Language and Cognitive Processes, 25(3):297–346, 2010. doi: 10.1080/", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 465, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 465, + 198 + ], + "score": 1.0, + "content": "01690960903041174. URL https://doi.org/10.1080/01690960903041174.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 163, + 506, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 506, + 238 + ], + "lines": [ + { + "bbox": [ + 104, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "communicate with deep multi-agent reinforcement learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 226, + 280, + 238 + ], + "spans": [ + { + "bbox": [ + 116, + 226, + 280, + 238 + ], + "score": 1.0, + "content": "Processing Systems, pp. 2137–2145, 2016.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 203, + 506, + 238 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 504, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "score": 1.0, + "content": "Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counter-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 255, + 411, + 268 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 411, + 268 + ], + "score": 1.0, + "content": "factual multi-agent policy gradients. arXiv preprint arXiv:1705.08926, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 243, + 506, + 268 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 506, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "Jakob Foerster, Richard Y Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "Learning with opponent-learning awareness. In Proceedings of the 17th International Conference on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Autonomous Agents and MultiAgent Systems, pp. 122–130. International Foundation for Autonomous", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 308, + 264, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 264, + 320 + ], + "score": 1.0, + "content": "Agents and Multiagent Systems, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 274, + 506, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 504, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "score": 1.0, + "content": "Sebastien Forestier and Pierre-Yves Oudeyer. A unified model of speech and tool use early development.´", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 337, + 428, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 428, + 350 + ], + "score": 1.0, + "content": "In 39th Annual Conference of the Cognitive Science Society (CogSci 2017), 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 325, + 505, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 504, + 378 + ], + "lines": [ + { + "bbox": [ + 104, + 353, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 104, + 353, + 506, + 368 + ], + "score": 1.0, + "content": "Felix A Gers, Jurgen Schmidhuber, and Fred Cummins. Learning to forget: Continual prediction with lstm. ¨", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 365, + 141, + 379 + ], + "spans": [ + { + "bbox": [ + 114, + 365, + 141, + 379 + ], + "score": 1.0, + "content": "1999.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 353, + 506, + 379 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 398 + ], + "score": 1.0, + "content": "Christian Guckelsberger, Christoph Salge, and Julian Togelius. New and surprising ways to be mean.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 396, + 497, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 497, + 408 + ], + "score": 1.0, + "content": "adversarial npcs with coupled empowerment minimisation. arXiv preprint arXiv:1806.01387, 2018.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 384, + 505, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 422, + 427 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 421, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 421, + 428 + ], + "score": 1.0, + "content": "Yuval Noah Harari. Sapiens: A brief history of humankind. Random House, 2014.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 414, + 421, + 428 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 506, + 466 + ], + "lines": [ + { + "bbox": [ + 104, + 432, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 506, + 447 + ], + "score": 1.0, + "content": "Joseph Henrich. The Secret of Our Success: How culture is driving human evolution, domesticating", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 445, + 507, + 457 + ], + "spans": [ + { + "bbox": [ + 116, + 445, + 507, + 457 + ], + "score": 1.0, + "content": "our species, and making us smart. Princeton University Press, Princeton, NJ, 2015. URL http:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 455, + 344, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 344, + 468 + ], + "score": 1.0, + "content": "//press.princeton.edu/titles/10543.html.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 432, + 507, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 506, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 507, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 507, + 486 + ], + "score": 1.0, + "content": "Esther Herrmann, Josep Call, Mar´ıa Victoria Hernandez-Lloreda, Brian Hare, and Michael Tomasello. `", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 486, + 507, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 486, + 507, + 497 + ], + "score": 1.0, + "content": "Humans have evolved specialized skills of social cognition: The cultural intelligence hypothesis.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "Science, 317(5843):1360–1366, 2007. ISSN 0036-8075. doi: 10.1126/science.1146282. URL http:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 507, + 384, + 519 + ], + "spans": [ + { + "bbox": [ + 116, + 507, + 384, + 519 + ], + "score": 1.0, + "content": "//science.sciencemag.org/content/317/5843/1360.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 473, + 507, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 506, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "Edward Hughes, Joel Z Leibo, Matthew G Phillips, Karl Tuyls, Edgar A Due´nez-Guzm ˜ an, Antonio Garc ´ ´ıa", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 116, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Castaneda, Iain Dunning, Tina Zhu, Kevin R McKee, Raphael Koster, et al. Inequity aversion improves ˜", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "cooperation in intertemporal social dilemmas. In Advances in neural information processing systems", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 559, + 243, + 570 + ], + "spans": [ + { + "bbox": [ + 116, + 559, + 243, + 570 + ], + "score": 1.0, + "content": "(NIPS), Montreal, Canada, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 525, + 506, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "Alexander S Klyubin, Daniel Polani, and Chrystopher L Nehaniv. Empowerment: A universal agent-centric", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 586, + 507, + 603 + ], + "spans": [ + { + "bbox": [ + 114, + 586, + 507, + 603 + ], + "score": 1.0, + "content": "measure of control. In Evolutionary Computation, 2005. The 2005 IEEE Congress on, volume 1, pp.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 598, + 204, + 610 + ], + "spans": [ + { + "bbox": [ + 116, + 598, + 204, + 610 + ], + "score": 1.0, + "content": "128–135. IEEE, 2005.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 577, + 507, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 617, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 295, + 630 + ], + "score": 1.0, + "content": "Kevin N. Laland. Darwin’s unfinished symphony", + "type": "text" + }, + { + "bbox": [ + 295, + 620, + 301, + 627 + ], + "score": 0.33, + "content": "\\therefore", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 617, + 507, + 630 + ], + "score": 1.0, + "content": "how culture made the human mind / Kevin N. Laland.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 629, + 420, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 629, + 420, + 640 + ], + "score": 1.0, + "content": "Princeton University Press Princeton, 2017. ISBN 9781400884872 140088487.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 617, + 507, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 658, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 506, + 670 + ], + "score": 1.0, + "content": "communication from referential games with symbolic and pixel input. arXiv preprint arXiv:1804.03984,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 668, + 141, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 141, + 680 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 646, + 506, + 680 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "score": 1.0, + "content": "Joel Z Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. Multi-agent", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Autonomous Agents and MultiAgent Systems, pp. 464–473. International Foundation for Autonomous", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 265, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 265, + 732 + ], + "score": 1.0, + "content": "Agents and Multiagent Systems, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 686, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Joseph T Lizier and Mikhail Prokopenko. Differentiating information transfer and causal effect. The", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 317, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 317, + 106 + ], + "score": 1.0, + "content": "European Physical Journal B, 73(4):605–615, 2010.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 505, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "score": 1.0, + "content": "Alicia P Melis and Dirk Semmann. How is human cooperation different? Philosophical Transactions of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 428, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 428, + 136 + ], + "score": 1.0, + "content": "the Royal Society of London B: Biological Sciences, 365(1553):2663–2674, 2010.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 142, + 504, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 157 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 165, + 379, + 178 + ], + "spans": [ + { + "bbox": [ + 117, + 165, + 379, + 178 + ], + "score": 1.0, + "content": "International conference on machine learning, pp. 1928–1937, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 184, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 198 + ], + "score": 1.0, + "content": "Pierre-Yves Oudeyer and Fred´ eric Kaplan. Discovering communication. ´ Connection Science, 18(2):", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 178, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 178, + 208 + ], + "score": 1.0, + "content": "189–206, 2006.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 504, + 238 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 504, + 228 + ], + "score": 1.0, + "content": "Pierre-Yves Oudeyer and Linda B Smith. How evolution may work through curiosity-driven developmental", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 227, + 340, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 227, + 340, + 238 + ], + "score": 1.0, + "content": "process. Topics in Cognitive Science, 8(2):492–502, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 506, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 259 + ], + "score": 1.0, + "content": "Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 255, + 507, + 270 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 507, + 270 + ], + "score": 1.0, + "content": "self-supervised prediction. In International Conference on Machine Learning (ICML), volume 2017,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 266, + 144, + 280 + ], + "spans": [ + { + "bbox": [ + 114, + 266, + 144, + 280 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic Books,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 296, + 142, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 142, + 311 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: a primer. John", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 327, + 198, + 341 + ], + "spans": [ + { + "bbox": [ + 116, + 327, + 198, + 341 + ], + "score": 1.0, + "content": "Wiley & Sons, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 504, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "score": 1.0, + "content": "Julien Perolat, Joel Z Leibo, Vinicius Zambaldi, Charles Beattie, Karl Tuyls, and Thore Graepel. A", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 116, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "multi-agent reinforcement learning model of common-pool resource appropriation. In Advances in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 370, + 356, + 382 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 356, + 382 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 3643–3652, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "Alexander Peysakhovich and Adam Lerer. Prosocial learning agents solve generalized stag hunts better", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 115, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "than selfish ones. In Proceedings of the 17th International Conference on Autonomous Agents and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 506, + 424 + ], + "score": 1.0, + "content": "MultiAgent Systems, pp. 2043–2044. International Foundation for Autonomous Agents and Multiagent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 423, + 177, + 435 + ], + "spans": [ + { + "bbox": [ + 116, + 423, + 177, + 435 + ], + "score": 1.0, + "content": "Systems, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "Neil C Rabinowitz, Frank Perbet, H Francis Song, Chiyuan Zhang, SM Eslami, and Matthew Botvinick.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 452, + 369, + 465 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 369, + 465 + ], + "score": 1.0, + "content": "Machine theory of mind. arXiv preprint arXiv:1802.07740, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 504, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 504, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 504, + 485 + ], + "score": 1.0, + "content": "Thomas C Schelling. Hockey helmets, concealed weapons, and daylight saving: A study of binary choices", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 483, + 389, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 389, + 495 + ], + "score": 1.0, + "content": "with externalities. Journal of Conflict resolution, 17(3):381–428, 1973.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 504, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "Jurgen Schmidhuber. Formal theory of creativity, fun, and intrinsic motivation (1990–2010). ¨ IEEE", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 513, + 391, + 525 + ], + "spans": [ + { + "bbox": [ + 117, + 513, + 391, + 525 + ], + "score": 1.0, + "content": "Transactions on Autonomous Mental Development, 2(3):230–247, 2010.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 506, + 566 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "Pedro Sequeira, Francisco S Melo, Rui Prada, and Ana Paiva. Emerging social awareness: Exploring intrin-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 544, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 544, + 506, + 556 + ], + "score": 1.0, + "content": "sic motivation in multiagent learning. In Development and Learning (ICDL), 2011 IEEE International", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 555, + 299, + 567 + ], + "spans": [ + { + "bbox": [ + 117, + 555, + 299, + 567 + ], + "score": 1.0, + "content": "Conference on, volume 2, pp. 1–6. IEEE, 2011.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "Satinder P. Singh, Andrew G. Barto, and Nuttapong Chentanez. Intrinsically motivated re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 116, + 584, + 206, + 598 + ], + "score": 1.0, + "content": "inforcement learning.", + "type": "text" + }, + { + "bbox": [ + 218, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems 17 [Neu-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 115, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "ral Information Processing Systems, NIPS 2004, December 13-18, 2004, Vancouver, British", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 116, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "Columbia, Canada], pp. 1281–1288, 2004. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 617, + 416, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 416, + 632 + ], + "score": 1.0, + "content": "2552-intrinsically-motivated-reinforcement-learning.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 637, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "Katherine KM Stavropoulos and Leslie J Carver. Research review: social motivation and oxytocin in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "autism–implications for joint attention development and intervention. Journal of Child Psychology and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 660, + 246, + 672 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 246, + 672 + ], + "score": 1.0, + "content": "Psychiatry, 54(6):603–618, 2013.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "DJ Strouse, Max Kleiman-Weiner, Josh Tenenbaum, Matt Botvinick, and David Schwab. Learning to share", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 690, + 472, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 472, + 702 + ], + "score": 1.0, + "content": "and hide intentions using information regularization. arXiv preprint arXiv:1808.02093, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Satohiro Tajima, Toru Yanagawa, Naotaka Fujii, and Taro Toyoizumi. Untangling brain-wide dynamics in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 721, + 461, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 461, + 732 + ], + "score": 1.0, + "content": "consciousness by cross-embedding. PLoS computational biology, 11(11):e1004537, 2015.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 298, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Joseph T Lizier and Mikhail Prokopenko. Differentiating information transfer and causal effect. The", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 317, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 317, + 106 + ], + "score": 1.0, + "content": "European Physical Journal B, 73(4):605–615, 2010.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 505, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 127 + ], + "score": 1.0, + "content": "Alicia P Melis and Dirk Semmann. How is human cooperation different? Philosophical Transactions of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 428, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 428, + 136 + ], + "score": 1.0, + "content": "the Royal Society of London B: Biological Sciences, 365(1553):2663–2674, 2010.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 506, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 142, + 504, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 157 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 165, + 379, + 178 + ], + "spans": [ + { + "bbox": [ + 117, + 165, + 379, + 178 + ], + "score": 1.0, + "content": "International conference on machine learning, pp. 1928–1937, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 142, + 506, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 184, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 198 + ], + "score": 1.0, + "content": "Pierre-Yves Oudeyer and Fred´ eric Kaplan. Discovering communication. ´ Connection Science, 18(2):", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 178, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 178, + 208 + ], + "score": 1.0, + "content": "189–206, 2006.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 182, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 504, + 238 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 504, + 228 + ], + "score": 1.0, + "content": "Pierre-Yves Oudeyer and Linda B Smith. How evolution may work through curiosity-driven developmental", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 227, + 340, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 227, + 340, + 238 + ], + "score": 1.0, + "content": "process. Topics in Cognitive Science, 8(2):492–502, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 215, + 504, + 238 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 506, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 259 + ], + "score": 1.0, + "content": "Deepak Pathak, Pulkit Agrawal, Alexei A Efros, and Trevor Darrell. Curiosity-driven exploration by", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 255, + 507, + 270 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 507, + 270 + ], + "score": 1.0, + "content": "self-supervised prediction. In International Conference on Machine Learning (ICML), volume 2017,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 266, + 144, + 280 + ], + "spans": [ + { + "bbox": [ + 114, + 266, + 144, + 280 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 243, + 507, + 280 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic Books,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 296, + 142, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 142, + 311 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 286, + 506, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: a primer. John", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 327, + 198, + 341 + ], + "spans": [ + { + "bbox": [ + 116, + 327, + 198, + 341 + ], + "score": 1.0, + "content": "Wiley & Sons, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 317, + 505, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 504, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "score": 1.0, + "content": "Julien Perolat, Joel Z Leibo, Vinicius Zambaldi, Charles Beattie, Karl Tuyls, and Thore Graepel. A", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 116, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "multi-agent reinforcement learning model of common-pool resource appropriation. In Advances in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 370, + 356, + 382 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 356, + 382 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 3643–3652, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 346, + 506, + 382 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "Alexander Peysakhovich and Adam Lerer. Prosocial learning agents solve generalized stag hunts better", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 115, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "than selfish ones. In Proceedings of the 17th International Conference on Autonomous Agents and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 506, + 424 + ], + "score": 1.0, + "content": "MultiAgent Systems, pp. 2043–2044. International Foundation for Autonomous Agents and Multiagent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 423, + 177, + 435 + ], + "spans": [ + { + "bbox": [ + 116, + 423, + 177, + 435 + ], + "score": 1.0, + "content": "Systems, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 388, + 506, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "Neil C Rabinowitz, Frank Perbet, H Francis Song, Chiyuan Zhang, SM Eslami, and Matthew Botvinick.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 452, + 369, + 465 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 369, + 465 + ], + "score": 1.0, + "content": "Machine theory of mind. arXiv preprint arXiv:1802.07740, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 442, + 505, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 504, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 504, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 504, + 485 + ], + "score": 1.0, + "content": "Thomas C Schelling. Hockey helmets, concealed weapons, and daylight saving: A study of binary choices", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 483, + 389, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 389, + 495 + ], + "score": 1.0, + "content": "with externalities. Journal of Conflict resolution, 17(3):381–428, 1973.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 471, + 504, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 504, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "Jurgen Schmidhuber. Formal theory of creativity, fun, and intrinsic motivation (1990–2010). ¨ IEEE", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 513, + 391, + 525 + ], + "spans": [ + { + "bbox": [ + 117, + 513, + 391, + 525 + ], + "score": 1.0, + "content": "Transactions on Autonomous Mental Development, 2(3):230–247, 2010.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 502, + 506, + 525 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 506, + 566 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "Pedro Sequeira, Francisco S Melo, Rui Prada, and Ana Paiva. Emerging social awareness: Exploring intrin-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 544, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 544, + 506, + 556 + ], + "score": 1.0, + "content": "sic motivation in multiagent learning. In Development and Learning (ICDL), 2011 IEEE International", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 555, + 299, + 567 + ], + "spans": [ + { + "bbox": [ + 117, + 555, + 299, + 567 + ], + "score": 1.0, + "content": "Conference on, volume 2, pp. 1–6. IEEE, 2011.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 533, + 506, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "Satinder P. Singh, Andrew G. Barto, and Nuttapong Chentanez. Intrinsically motivated re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 116, + 584, + 206, + 598 + ], + "score": 1.0, + "content": "inforcement learning.", + "type": "text" + }, + { + "bbox": [ + 218, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems 17 [Neu-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 115, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "ral Information Processing Systems, NIPS 2004, December 13-18, 2004, Vancouver, British", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 116, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "Columbia, Canada], pp. 1281–1288, 2004. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 617, + 416, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 416, + 632 + ], + "score": 1.0, + "content": "2552-intrinsically-motivated-reinforcement-learning.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 573, + 506, + 632 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 637, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "Katherine KM Stavropoulos and Leslie J Carver. Research review: social motivation and oxytocin in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "autism–implications for joint attention development and intervention. Journal of Child Psychology and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 660, + 246, + 672 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 246, + 672 + ], + "score": 1.0, + "content": "Psychiatry, 54(6):603–618, 2013.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 637, + 505, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "DJ Strouse, Max Kleiman-Weiner, Josh Tenenbaum, Matt Botvinick, and David Schwab. Learning to share", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 690, + 472, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 472, + 702 + ], + "score": 1.0, + "content": "and hide intentions using information regularization. arXiv preprint arXiv:1808.02093, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 678, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Satohiro Tajima, Toru Yanagawa, Naotaka Fujii, and Taro Toyoizumi. Untangling brain-wide dynamics in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 721, + 461, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 461, + 732 + ], + "score": 1.0, + "content": "consciousness by cross-embedding. PLoS computational biology, 11(11):e1004537, 2015.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 329, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 329, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 329, + 96 + ], + "score": 1.0, + "content": "Michael Tomasello. Why we cooperate. MIT press, 2009.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 101, + 504, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 506, + 114 + ], + "score": 1.0, + "content": "Carel P van Schaik and Judith M Burkart. Social learning and evolution: the cultural intelligence hypothesis.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 112, + 502, + 125 + ], + "spans": [ + { + "bbox": [ + 116, + 112, + 502, + 125 + ], + "score": 1.0, + "content": "Philosophical Transactions of the Royal Society B: Biological Sciences, 366(1567):1008–1016, 2011.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 385, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 385, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 385, + 144 + ], + "score": 1.0, + "content": "Karl von Frisch. The dance language and orientation of bees. 5, 06 1969.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 504, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 161, + 313, + 174 + ], + "spans": [ + { + "bbox": [ + 116, + 161, + 313, + 174 + ], + "score": 1.0, + "content": "learning. Machine learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 180, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 506, + 193 + ], + "score": 1.0, + "content": "Chao Yu, Minjie Zhang, and Fenghui Ren. Emotional multiagent reinforcement learning in social dilemmas.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 190, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 190, + 506, + 205 + ], + "score": 1.0, + "content": "In International Conference on Principles and Practice of Multi-Agent Systems, pp. 372–387. Springer,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 200, + 142, + 215 + ], + "spans": [ + { + "bbox": [ + 115, + 200, + 142, + 215 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 329, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 329, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 329, + 96 + ], + "score": 1.0, + "content": "Michael Tomasello. Why we cooperate. MIT press, 2009.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 329, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 101, + 504, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 506, + 114 + ], + "score": 1.0, + "content": "Carel P van Schaik and Judith M Burkart. Social learning and evolution: the cultural intelligence hypothesis.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 112, + 502, + 125 + ], + "spans": [ + { + "bbox": [ + 116, + 112, + 502, + 125 + ], + "score": 1.0, + "content": "Philosophical Transactions of the Royal Society B: Biological Sciences, 366(1567):1008–1016, 2011.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 100, + 506, + 125 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 385, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 385, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 385, + 144 + ], + "score": 1.0, + "content": "Karl von Frisch. The dance language and orientation of bees. 5, 06 1969.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 131, + 385, + 144 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 504, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 161, + 313, + 174 + ], + "spans": [ + { + "bbox": [ + 116, + 161, + 313, + 174 + ], + "score": 1.0, + "content": "learning. Machine learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 150, + 505, + 174 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 180, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 506, + 193 + ], + "score": 1.0, + "content": "Chao Yu, Minjie Zhang, and Fenghui Ren. Emotional multiagent reinforcement learning in social dilemmas.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 190, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 190, + 506, + 205 + ], + "score": 1.0, + "content": "In International Conference on Principles and Practice of Multi-Agent Systems, pp. 372–387. Springer,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 200, + 142, + 215 + ], + "spans": [ + { + "bbox": [ + 115, + 200, + 142, + 215 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 179, + 506, + 215 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 177, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 180, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 180, + 96 + ], + "score": 1.0, + "content": "6 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 100, + 266, + 111 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 267, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 267, + 113 + ], + "score": 1.0, + "content": "6.1 SEQUENTIAL SOCIAL DILEMMAS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 115, + 505, + 193 + ], + "lines": [ + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 453, + 129 + ], + "score": 1.0, + "content": "In each of the sequential social dilemma (SSD) games studied above, an agent is rewarded", + "type": "text" + }, + { + "bbox": [ + 454, + 116, + 468, + 126 + ], + "score": 0.84, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "for every", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "apple it collects, but the apples are a limited resource. In Harvest (a tragedy of the commons game), apples", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "regenerate more slowly the faster they are harvested, and if an exploiting agent consumes all of the apples,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "they will not grow back; agents must cooperate to harvest sustainably. In Cleanup (a public goods game),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "apples are generated based on the amount of waste in a nearby river. Agents can use a cleaning beam action", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "to clean the river when they are positioned in it; or they can simply consume the apples the other agent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 497, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 447, + 194 + ], + "score": 1.0, + "content": "produces. Agents also have a fining beam action which they can use to fine nearby agents", + "type": "text" + }, + { + "bbox": [ + 447, + 182, + 466, + 192 + ], + "score": 0.7, + "content": "- 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 181, + 497, + 194 + ], + "score": 1.0, + "content": "reward.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "Figure 9 gives the Schelling diagram for both SSD tasks under investigation. A Schelling diagram", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "(Schelling, 1973; Perolat et al., 2017) shows the relative payoffs for a single agent’s strategy given a fixed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "number of other agents who are cooperative. Schelling diagrams generalize payoff matrices to multi-agent", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 229, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 104, + 229, + 505, + 245 + ], + "score": 1.0, + "content": "settings, and make it easy to visually recognize game-theoretic properties like Nash equilibria (see Schelling", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 201, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 201, + 255 + ], + "score": 1.0, + "content": "(1973) for more details).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 113, + 262, + 501, + 415 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 262, + 501, + 415 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 262, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 113, + 262, + 501, + 415 + ], + "score": 0.973, + "type": "image", + "image_path": "cd8620142473269dee9b7be7e5b1fb98864ebd4ee20a85b56ccae63f16b46cc2.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 113, + 262, + 501, + 313.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 113, + 313.0, + 501, + 364.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 113, + 364.0, + 501, + 415.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 424, + 506, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "Figure 9: Schelling diagrams for the two social dilemma tasks show that an individual is almost always", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 435, + 469, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 469, + 448 + ], + "score": 1.0, + "content": "motivated to defect, even though the group will get higher reward if there are more cooperators.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 16.25 + }, + { + "type": "title", + "bbox": [ + 106, + 463, + 308, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 308, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 308, + 475 + ], + "score": 1.0, + "content": "6.2 ADDITIONAL EXPERIMENT - BOX TRAPPED", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 297, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 298, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 298, + 491 + ], + "score": 1.0, + "content": "As a proof-of-concept experiment to test whether", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 489, + 298, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 298, + 501 + ], + "score": 1.0, + "content": "the influence reward works as expected, we con-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 500, + 298, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 298, + 512 + ], + "score": 1.0, + "content": "structed a special environment, shown in Figure 10.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 510, + 298, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 298, + 524 + ], + "score": 1.0, + "content": "In this environment, one agent (teal) is trapped in a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 522, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 297, + 534 + ], + "score": 1.0, + "content": "box. The other agent (purple) has a special action", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 533, + 297, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 297, + 545 + ], + "score": 1.0, + "content": "it can use to open the box... or it can simply choose", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 545, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 297, + 555 + ], + "score": 1.0, + "content": "to consume apples, which exist outside the box and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 555, + 249, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 249, + 566 + ], + "score": 1.0, + "content": "are inexhaustible in this environment.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 296, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 297, + 584 + ], + "score": 1.0, + "content": "As expected, a vanilla A3C agent learns to act", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 298, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 298, + 595 + ], + "score": 1.0, + "content": "selfishly; the purple agent will simply consume", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 594, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 281, + 606 + ], + "score": 1.0, + "content": "apples, and chooses the open box action in", + "type": "text" + }, + { + "bbox": [ + 281, + 594, + 296, + 605 + ], + "score": 0.83, + "content": "0 \\%", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 605, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 297, + 617 + ], + "score": 1.0, + "content": "of trajectories once the policy has converged. A", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 616, + 297, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 297, + 627 + ], + "score": 1.0, + "content": "video of A3C agents trained in this environment", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 627, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 296, + 638 + ], + "score": 1.0, + "content": "is available at: https://youtu.be/C8SE9_", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 637, + 297, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 136, + 649 + ], + "score": 0.33, + "content": "\\Upsilon \\mathrm { K z x I }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 637, + 297, + 650 + ], + "score": 1.0, + "content": ", which shows that the purple agent leaves", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 649, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 297, + 661 + ], + "score": 1.0, + "content": "its compatriot trapped in the box throughout the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 659, + 147, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 147, + 674 + ], + "score": 1.0, + "content": "trajectory.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 308, + 488, + 498, + 622 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 488, + 498, + 622 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 308, + 488, + 498, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 498, + 622 + ], + "score": 0.908, + "type": "image", + "image_path": "3f18791740d5f30d7f41c7fce3b78ff4b93cfac39e1f81dcfeb16188b74db18f.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 308, + 488, + 498, + 501.4 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 308, + 501.4, + 498, + 514.8 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 308, + 514.8, + 498, + 528.1999999999999 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 308, + 528.1999999999999, + 498, + 541.5999999999999 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 308, + 541.5999999999999, + 498, + 554.9999999999999 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 308, + 554.9999999999999, + 498, + 568.3999999999999 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 308, + 568.3999999999999, + 498, + 581.7999999999998 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 308, + 581.7999999999998, + 498, + 595.1999999999998 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 308, + 595.1999999999998, + 498, + 608.5999999999998 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 308, + 608.5999999999998, + 498, + 621.9999999999998 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 630, + 504, + 664 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 304, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "Figure 10: The Box trapped environment in which", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "the teal agent is trapped, and the purple agent can", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 653, + 459, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 459, + 665 + ], + "score": 1.0, + "content": "release it with a special open box action.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48 + } + ], + "index": 40.25 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 298, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 298, + 688 + ], + "score": 1.0, + "content": "In contrast, an agent trained with the social influ-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 264, + 700 + ], + "score": 1.0, + "content": "ence reward chooses the open box action in", + "type": "text" + }, + { + "bbox": [ + 265, + 688, + 283, + 698 + ], + "score": 0.86, + "content": "88 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "of trajectories, releasing its fellow agent so that they are both", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "able to consume apples. A video of this behavior is shown at: https://youtu.be/Gfo248-qt3c.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "Further, as Figure 11(a) reveals, the purple influencer agent usually chooses to open the box within the first", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 398, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 398, + 734 + ], + "score": 1.0, + "content": "few steps of the trajetory, giving its fellow agent more time to collect reward.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 752, + 310, + 760 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 177, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 180, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 180, + 96 + ], + "score": 1.0, + "content": "6 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 100, + 266, + 111 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 267, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 267, + 113 + ], + "score": 1.0, + "content": "6.1 SEQUENTIAL SOCIAL DILEMMAS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 115, + 505, + 193 + ], + "lines": [ + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 453, + 129 + ], + "score": 1.0, + "content": "In each of the sequential social dilemma (SSD) games studied above, an agent is rewarded", + "type": "text" + }, + { + "bbox": [ + 454, + 116, + 468, + 126 + ], + "score": 0.84, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "for every", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "apple it collects, but the apples are a limited resource. In Harvest (a tragedy of the commons game), apples", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "regenerate more slowly the faster they are harvested, and if an exploiting agent consumes all of the apples,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "they will not grow back; agents must cooperate to harvest sustainably. In Cleanup (a public goods game),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "apples are generated based on the amount of waste in a nearby river. Agents can use a cleaning beam action", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "to clean the river when they are positioned in it; or they can simply consume the apples the other agent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 497, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 447, + 194 + ], + "score": 1.0, + "content": "produces. Agents also have a fining beam action which they can use to fine nearby agents", + "type": "text" + }, + { + "bbox": [ + 447, + 182, + 466, + 192 + ], + "score": 0.7, + "content": "- 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 181, + 497, + 194 + ], + "score": 1.0, + "content": "reward.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 114, + 506, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "Figure 9 gives the Schelling diagram for both SSD tasks under investigation. A Schelling diagram", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "(Schelling, 1973; Perolat et al., 2017) shows the relative payoffs for a single agent’s strategy given a fixed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "number of other agents who are cooperative. Schelling diagrams generalize payoff matrices to multi-agent", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 229, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 104, + 229, + 505, + 245 + ], + "score": 1.0, + "content": "settings, and make it easy to visually recognize game-theoretic properties like Nash equilibria (see Schelling", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 201, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 201, + 255 + ], + "score": 1.0, + "content": "(1973) for more details).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 198, + 505, + 255 + ] + }, + { + "type": "image", + "bbox": [ + 113, + 262, + 501, + 415 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 262, + 501, + 415 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 262, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 113, + 262, + 501, + 415 + ], + "score": 0.973, + "type": "image", + "image_path": "cd8620142473269dee9b7be7e5b1fb98864ebd4ee20a85b56ccae63f16b46cc2.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 113, + 262, + 501, + 313.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 113, + 313.0, + 501, + 364.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 113, + 364.0, + 501, + 415.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 424, + 506, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "Figure 9: Schelling diagrams for the two social dilemma tasks show that an individual is almost always", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 435, + 469, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 469, + 448 + ], + "score": 1.0, + "content": "motivated to defect, even though the group will get higher reward if there are more cooperators.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 16.25 + }, + { + "type": "title", + "bbox": [ + 106, + 463, + 308, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 308, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 308, + 475 + ], + "score": 1.0, + "content": "6.2 ADDITIONAL EXPERIMENT - BOX TRAPPED", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 297, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 298, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 298, + 491 + ], + "score": 1.0, + "content": "As a proof-of-concept experiment to test whether", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 489, + 298, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 298, + 501 + ], + "score": 1.0, + "content": "the influence reward works as expected, we con-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 500, + 298, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 298, + 512 + ], + "score": 1.0, + "content": "structed a special environment, shown in Figure 10.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 510, + 298, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 298, + 524 + ], + "score": 1.0, + "content": "In this environment, one agent (teal) is trapped in a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 522, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 297, + 534 + ], + "score": 1.0, + "content": "box. The other agent (purple) has a special action", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 533, + 297, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 297, + 545 + ], + "score": 1.0, + "content": "it can use to open the box... or it can simply choose", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 545, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 297, + 555 + ], + "score": 1.0, + "content": "to consume apples, which exist outside the box and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 555, + 249, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 249, + 566 + ], + "score": 1.0, + "content": "are inexhaustible in this environment.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 479, + 298, + 566 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 296, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 297, + 584 + ], + "score": 1.0, + "content": "As expected, a vanilla A3C agent learns to act", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 298, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 298, + 595 + ], + "score": 1.0, + "content": "selfishly; the purple agent will simply consume", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 594, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 281, + 606 + ], + "score": 1.0, + "content": "apples, and chooses the open box action in", + "type": "text" + }, + { + "bbox": [ + 281, + 594, + 296, + 605 + ], + "score": 0.83, + "content": "0 \\%", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 605, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 297, + 617 + ], + "score": 1.0, + "content": "of trajectories once the policy has converged. A", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 616, + 297, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 297, + 627 + ], + "score": 1.0, + "content": "video of A3C agents trained in this environment", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 627, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 296, + 638 + ], + "score": 1.0, + "content": "is available at: https://youtu.be/C8SE9_", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 637, + 297, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 136, + 649 + ], + "score": 0.33, + "content": "\\Upsilon \\mathrm { K z x I }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 637, + 297, + 650 + ], + "score": 1.0, + "content": ", which shows that the purple agent leaves", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 649, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 297, + 661 + ], + "score": 1.0, + "content": "its compatriot trapped in the box throughout the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 659, + 147, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 147, + 674 + ], + "score": 1.0, + "content": "trajectory.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 572, + 298, + 674 + ] + }, + { + "type": "image", + "bbox": [ + 308, + 488, + 498, + 622 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 488, + 498, + 622 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 308, + 488, + 498, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 498, + 622 + ], + "score": 0.908, + "type": "image", + "image_path": "3f18791740d5f30d7f41c7fce3b78ff4b93cfac39e1f81dcfeb16188b74db18f.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 308, + 488, + 498, + 501.4 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 308, + 501.4, + 498, + 514.8 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 308, + 514.8, + 498, + 528.1999999999999 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 308, + 528.1999999999999, + 498, + 541.5999999999999 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 308, + 541.5999999999999, + 498, + 554.9999999999999 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 308, + 554.9999999999999, + 498, + 568.3999999999999 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 308, + 568.3999999999999, + 498, + 581.7999999999998 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 308, + 581.7999999999998, + 498, + 595.1999999999998 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 308, + 595.1999999999998, + 498, + 608.5999999999998 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 308, + 608.5999999999998, + 498, + 621.9999999999998 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 630, + 504, + 664 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 304, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "Figure 10: The Box trapped environment in which", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "the teal agent is trapped, and the purple agent can", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 653, + 459, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 459, + 665 + ], + "score": 1.0, + "content": "release it with a special open box action.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48 + } + ], + "index": 40.25 + }, + { + "type": "list", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 298, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 298, + 688 + ], + "score": 1.0, + "content": "In contrast, an agent trained with the social influ-", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 264, + 700 + ], + "score": 1.0, + "content": "ence reward chooses the open box action in", + "type": "text" + }, + { + "bbox": [ + 265, + 688, + 283, + 698 + ], + "score": 0.86, + "content": "88 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "of trajectories, releasing its fellow agent so that they are both", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "able to consume apples. A video of this behavior is shown at: https://youtu.be/Gfo248-qt3c.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "Further, as Figure 11(a) reveals, the purple influencer agent usually chooses to open the box within the first", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 398, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 398, + 734 + ], + "score": 1.0, + "content": "few steps of the trajetory, giving its fellow agent more time to collect reward.", + "type": "text" + } + ], + "index": 54, + "is_list_end_line": true + } + ], + "index": 52, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Most importantly though, Figure 11(b) shows the influence reward over the course of a trajectory in the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Box trapped environment. The agent chooses the open box action in the second timestep; at this point, we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "see a corresponding spike in the influence reward. This reveals that the influence reward works as expected,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "incentivizing an action which has a strong — and in this case, prosocial — effect on the other agent’s", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 144, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 144, + 139 + ], + "score": 1.0, + "content": "behavior.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 148, + 296, + 276 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 148, + 296, + 276 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 118, + 148, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 118, + 148, + 296, + 276 + ], + "score": 0.964, + "type": "image", + "image_path": "87f2f4c8939684aaa1a065bb1bd2a40dcabf6619e1b311cd76ab190958323091.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 118, + 148, + 296, + 162.22222222222223 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 118, + 162.22222222222223, + 296, + 176.44444444444446 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 118, + 176.44444444444446, + 296, + 190.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 118, + 190.66666666666669, + 296, + 204.8888888888889 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 118, + 204.8888888888889, + 296, + 219.11111111111114 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 118, + 219.11111111111114, + 296, + 233.33333333333337 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 118, + 233.33333333333337, + 296, + 247.5555555555556 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 118, + 247.5555555555556, + 296, + 261.7777777777778 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 118, + 261.7777777777778, + 296, + 276.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 307, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 307, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 307, + 294 + ], + "score": 1.0, + "content": "(a) Number of times the open box action occurs at each", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 293, + 230, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 230, + 304 + ], + "score": 1.0, + "content": "trajectory step over 100 trajectories.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 316, + 153, + 495, + 280 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 316, + 153, + 495, + 280 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 495, + 280 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 495, + 280 + ], + "score": 0.957, + "type": "image", + "image_path": "aa590ebca444f816660ca98874365417e4da0912c85a637d878b313133db889b.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 316, + 153, + 495, + 167.11111111111111 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 316, + 167.11111111111111, + 495, + 181.22222222222223 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 316, + 181.22222222222223, + 495, + 195.33333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 316, + 195.33333333333334, + 495, + 209.44444444444446 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 316, + 209.44444444444446, + 495, + 223.55555555555557 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 316, + 223.55555555555557, + 495, + 237.66666666666669 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 316, + 237.66666666666669, + 495, + 251.7777777777778 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 316, + 251.7777777777778, + 495, + 265.8888888888889 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 316, + 265.8888888888889, + 495, + 280.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 315, + 286, + 495, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 312, + 285, + 496, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 285, + 496, + 299 + ], + "score": 1.0, + "content": "(b) Influence reward over a trajectory in Box trapped", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "image_caption", + "bbox": [ + 105, + 312, + 504, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "Figure 11: The Box trapped proof-of-concept experiment reveals that an agent gets high influence for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 323, + 318, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 318, + 336 + ], + "score": 1.0, + "content": "letting another agent out of a box in which it is trapped.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 354, + 241, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 243, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 243, + 366 + ], + "score": 1.0, + "content": "6.3 IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "All models are trained with a single convolutional layer with a kernel of size 3, stride of size 1, and 6 output", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "channels. This is connected to two fully connected layers of size 32 each, and an LSTM with 128 cells.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 391, + 380, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 200, + 404 + ], + "score": 1.0, + "content": "We use a discount factor", + "type": "text" + }, + { + "bbox": [ + 201, + 393, + 229, + 403 + ], + "score": 0.9, + "content": "\\gamma = . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 391, + 318, + 404 + ], + "score": 1.0, + "content": ". The number of agents", + "type": "text" + }, + { + "bbox": [ + 319, + 393, + 329, + 402 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 391, + 380, + 404 + ], + "score": 1.0, + "content": "is fixed to 5.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 408, + 504, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "As mentioned in Section 2.2, the social influence reward can be computed using a number of diver-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "score": 1.0, + "content": "gence measures, including JSD. We also experiment with training the agents using the pointwise mutual", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 430, + 340, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 340, + 444 + ], + "score": 1.0, + "content": "information (the innermost term of Eq. 3), which is given by:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 446, + 426, + 474 + ], + "lines": [ + { + "bbox": [ + 184, + 446, + 426, + 474 + ], + "spans": [ + { + "bbox": [ + 184, + 446, + 426, + 474 + ], + "score": 0.93, + "content": "p m i ( a ^ { A } ; a ^ { B } | Z { = } z ) { = } \\log \\frac { p ( a ^ { B } | a ^ { A } , z ) } { p ( a ^ { B } | z ) } { = } \\mathrm { l o g } \\frac { p ( a ^ { A } , a ^ { B } | z ) } { p ( a ^ { A } | z ) p ( a ^ { B } | z ) } .", + "type": "interline_equation", + "image_path": "3711246b82494f6bba66f95bcb17310950d9a545ef4a431d8f248a34b3326879.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 446, + 426, + 460.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 184, + 460.0, + 426, + 474.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 477, + 503, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "This PMI term is precisely the local information flow proposed by Lizier & Prokopenko (2010) as a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 358, + 501 + ], + "score": 1.0, + "content": "measure of direct causal effect; the expectation of the PMI over", + "type": "text" + }, + { + "bbox": [ + 358, + 487, + 405, + 500 + ], + "score": 0.91, + "content": "p \\dot { ( } a ^ { B } , a ^ { A } | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "is the MI. and gives us a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 499, + 398, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 265, + 511 + ], + "score": 1.0, + "content": "measure of influence of a single action of", + "type": "text" + }, + { + "bbox": [ + 265, + 499, + 274, + 509 + ], + "score": 0.81, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 499, + 385, + 511 + ], + "score": 1.0, + "content": "on the single action taken by", + "type": "text" + }, + { + "bbox": [ + 385, + 500, + 394, + 509 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 499, + 398, + 511 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "In addition to the comparison function used to compute influence, there are many other hyperparameters", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "that can be tuned for each model. We use a random search over hyperparameters, ensuring a fair comparison", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "with the search size over the baseline parameters that are shared with the influence models. For all models", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 549, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 560 + ], + "score": 1.0, + "content": "we search for the optimal entropy reward and learning rate, where we anneal the learning rate from an initial", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "value lr init to lr final. The below sections give the parameters found to be most effective for", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 571, + 222, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 222, + 582 + ], + "score": 1.0, + "content": "each of the three experiments.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + }, + { + "type": "title", + "bbox": [ + 106, + 594, + 339, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 340, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 340, + 606 + ], + "score": 1.0, + "content": "6.3.1 CENTRALISED CONTROLLER HYPERPARAMETERS", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 434, + 624 + ], + "score": 1.0, + "content": "In this setting we vary the number of influencers from 1−4, the influence reward weight", + "type": "text" + }, + { + "bbox": [ + 435, + 614, + 442, + 624 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 612, + 506, + 624 + ], + "score": 1.0, + "content": ", and the number", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 622, + 507, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 437, + 637 + ], + "score": 1.0, + "content": "of curriculum steps over which the weight of the influence reward is linearly increased", + "type": "text" + }, + { + "bbox": [ + 437, + 624, + 446, + 633 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 622, + 507, + 637 + ], + "score": 1.0, + "content": ". In this setting,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "since we have a centralised controller, we also experiment with giving the influence reward to the agent", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "being influenced as well, and find that this sometimes helps. This ‘influencee’ reward is not used in the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "other two experiments, since it precludes independent training. The hyperparameters found to give the best", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 668, + 299, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 299, + 679 + ], + "score": 1.0, + "content": "performance for each model are shown in Table 1.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5 + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 296, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 298, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 298, + 703 + ], + "score": 1.0, + "content": "6.3.2 COMMUNICATION HYPERPARAMETERS", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Because the communication models have an extra A2C output head for the communication policy, we use", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "an additional entropy regularization term just for this head, and apply a weight to the communication loss", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Most importantly though, Figure 11(b) shows the influence reward over the course of a trajectory in the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Box trapped environment. The agent chooses the open box action in the second timestep; at this point, we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "see a corresponding spike in the influence reward. This reveals that the influence reward works as expected,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "incentivizing an action which has a strong — and in this case, prosocial — effect on the other agent’s", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 144, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 144, + 139 + ], + "score": 1.0, + "content": "behavior.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 83, + 506, + 139 + ] + }, + { + "type": "image", + "bbox": [ + 118, + 148, + 296, + 276 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 148, + 296, + 276 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 118, + 148, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 118, + 148, + 296, + 276 + ], + "score": 0.964, + "type": "image", + "image_path": "87f2f4c8939684aaa1a065bb1bd2a40dcabf6619e1b311cd76ab190958323091.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 118, + 148, + 296, + 162.22222222222223 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 118, + 162.22222222222223, + 296, + 176.44444444444446 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 118, + 176.44444444444446, + 296, + 190.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 118, + 190.66666666666669, + 296, + 204.8888888888889 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 118, + 204.8888888888889, + 296, + 219.11111111111114 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 118, + 219.11111111111114, + 296, + 233.33333333333337 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 118, + 233.33333333333337, + 296, + 247.5555555555556 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 118, + 247.5555555555556, + 296, + 261.7777777777778 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 118, + 261.7777777777778, + 296, + 276.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 307, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 307, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 307, + 294 + ], + "score": 1.0, + "content": "(a) Number of times the open box action occurs at each", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 293, + 230, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 230, + 304 + ], + "score": 1.0, + "content": "trajectory step over 100 trajectories.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 282, + 307, + 304 + ] + }, + { + "type": "image", + "bbox": [ + 316, + 153, + 495, + 280 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 316, + 153, + 495, + 280 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 495, + 280 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 495, + 280 + ], + "score": 0.957, + "type": "image", + "image_path": "aa590ebca444f816660ca98874365417e4da0912c85a637d878b313133db889b.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 316, + 153, + 495, + 167.11111111111111 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 316, + 167.11111111111111, + 495, + 181.22222222222223 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 316, + 181.22222222222223, + 495, + 195.33333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 316, + 195.33333333333334, + 495, + 209.44444444444446 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 316, + 209.44444444444446, + 495, + 223.55555555555557 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 316, + 223.55555555555557, + 495, + 237.66666666666669 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 316, + 237.66666666666669, + 495, + 251.7777777777778 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 316, + 251.7777777777778, + 495, + 265.8888888888889 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 316, + 265.8888888888889, + 495, + 280.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 315, + 286, + 495, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 312, + 285, + 496, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 285, + 496, + 299 + ], + "score": 1.0, + "content": "(b) Influence reward over a trajectory in Box trapped", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "image_caption", + "bbox": [ + 105, + 312, + 504, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "Figure 11: The Box trapped proof-of-concept experiment reveals that an agent gets high influence for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 323, + 318, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 318, + 336 + ], + "score": 1.0, + "content": "letting another agent out of a box in which it is trapped.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 354, + 241, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 243, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 243, + 366 + ], + "score": 1.0, + "content": "6.3 IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "All models are trained with a single convolutional layer with a kernel of size 3, stride of size 1, and 6 output", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "channels. This is connected to two fully connected layers of size 32 each, and an LSTM with 128 cells.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 391, + 380, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 200, + 404 + ], + "score": 1.0, + "content": "We use a discount factor", + "type": "text" + }, + { + "bbox": [ + 201, + 393, + 229, + 403 + ], + "score": 0.9, + "content": "\\gamma = . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 391, + 318, + 404 + ], + "score": 1.0, + "content": ". The number of agents", + "type": "text" + }, + { + "bbox": [ + 319, + 393, + 329, + 402 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 391, + 380, + 404 + ], + "score": 1.0, + "content": "is fixed to 5.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 370, + 506, + 404 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 408, + 504, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "As mentioned in Section 2.2, the social influence reward can be computed using a number of diver-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "score": 1.0, + "content": "gence measures, including JSD. We also experiment with training the agents using the pointwise mutual", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 430, + 340, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 340, + 444 + ], + "score": 1.0, + "content": "information (the innermost term of Eq. 3), which is given by:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 408, + 506, + 444 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 446, + 426, + 474 + ], + "lines": [ + { + "bbox": [ + 184, + 446, + 426, + 474 + ], + "spans": [ + { + "bbox": [ + 184, + 446, + 426, + 474 + ], + "score": 0.93, + "content": "p m i ( a ^ { A } ; a ^ { B } | Z { = } z ) { = } \\log \\frac { p ( a ^ { B } | a ^ { A } , z ) } { p ( a ^ { B } | z ) } { = } \\mathrm { l o g } \\frac { p ( a ^ { A } , a ^ { B } | z ) } { p ( a ^ { A } | z ) p ( a ^ { B } | z ) } .", + "type": "interline_equation", + "image_path": "3711246b82494f6bba66f95bcb17310950d9a545ef4a431d8f248a34b3326879.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 446, + 426, + 460.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 184, + 460.0, + 426, + 474.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 477, + 503, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "This PMI term is precisely the local information flow proposed by Lizier & Prokopenko (2010) as a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 358, + 501 + ], + "score": 1.0, + "content": "measure of direct causal effect; the expectation of the PMI over", + "type": "text" + }, + { + "bbox": [ + 358, + 487, + 405, + 500 + ], + "score": 0.91, + "content": "p \\dot { ( } a ^ { B } , a ^ { A } | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "is the MI. and gives us a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 499, + 398, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 265, + 511 + ], + "score": 1.0, + "content": "measure of influence of a single action of", + "type": "text" + }, + { + "bbox": [ + 265, + 499, + 274, + 509 + ], + "score": 0.81, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 499, + 385, + 511 + ], + "score": 1.0, + "content": "on the single action taken by", + "type": "text" + }, + { + "bbox": [ + 385, + 500, + 394, + 509 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 499, + 398, + 511 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 476, + 506, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "In addition to the comparison function used to compute influence, there are many other hyperparameters", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "that can be tuned for each model. We use a random search over hyperparameters, ensuring a fair comparison", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "with the search size over the baseline parameters that are shared with the influence models. For all models", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 549, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 560 + ], + "score": 1.0, + "content": "we search for the optimal entropy reward and learning rate, where we anneal the learning rate from an initial", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "value lr init to lr final. The below sections give the parameters found to be most effective for", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 571, + 222, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 222, + 582 + ], + "score": 1.0, + "content": "each of the three experiments.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 515, + 506, + 582 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 594, + 339, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 340, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 340, + 606 + ], + "score": 1.0, + "content": "6.3.1 CENTRALISED CONTROLLER HYPERPARAMETERS", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 434, + 624 + ], + "score": 1.0, + "content": "In this setting we vary the number of influencers from 1−4, the influence reward weight", + "type": "text" + }, + { + "bbox": [ + 435, + 614, + 442, + 624 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 612, + 506, + 624 + ], + "score": 1.0, + "content": ", and the number", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 622, + 507, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 437, + 637 + ], + "score": 1.0, + "content": "of curriculum steps over which the weight of the influence reward is linearly increased", + "type": "text" + }, + { + "bbox": [ + 437, + 624, + 446, + 633 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 622, + 507, + 637 + ], + "score": 1.0, + "content": ". In this setting,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "since we have a centralised controller, we also experiment with giving the influence reward to the agent", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "being influenced as well, and find that this sometimes helps. This ‘influencee’ reward is not used in the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "other two experiments, since it precludes independent training. The hyperparameters found to give the best", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 668, + 299, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 299, + 679 + ], + "score": 1.0, + "content": "performance for each model are shown in Table 1.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 612, + 507, + 679 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 296, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 298, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 298, + 703 + ], + "score": 1.0, + "content": "6.3.2 COMMUNICATION HYPERPARAMETERS", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Because the communication models have an extra A2C output head for the communication policy, we use", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "an additional entropy regularization term just for this head, and apply a weight to the communication loss", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "in the loss function. We also vary the number of communication symbols that the agents can emit, and", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "the size of the linear layer that connects the LSTM to the communication policy layer, which we term", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "the communication embedding size. Finally, in the communication regime, we experiment to setting the", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 236, + 644 + ], + "score": 1.0, + "content": "weight on the extrinsic reward E,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 237, + 633, + 244, + 641 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 245, + 630, + 505, + 644 + ], + "score": 1.0, + "content": ", to zero. The best hyperparameters for each of the communication", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 642, + 219, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 219, + 653 + ], + "score": 1.0, + "content": "models are shown in Table 2.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 54.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 79, + 497, + 193 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 118, + 79, + 497, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 79, + 497, + 193 + ], + "spans": [ + { + "bbox": [ + 118, + 79, + 497, + 193 + ], + "score": 0.979, + "html": "
HyperparameterCleanupHarvest
A3CVisibleactionsInfluenceA3C baselineVisibleactionsInfluence
Entropy reg.baseline .00176baseline .00176.000248.000687baseline .00184.00025
lr_init.00126.00126.00107.00136.00215.00107
lr_end.000012.000012.000042.000028.000013.000042
Number of influencers3133
Influence weight β0.1460.224
Curriculum C11401140
Policy comparisonJSDPMI
Influencee reward10
", + "type": "table", + "image_path": "f4c2027e5c945541b51744fd4ee3d446041a4b1d05ad3603be18149e0ee814e4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 79, + 497, + 117.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 117.0, + 497, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 155.0, + 497, + 193.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 121, + 200, + 488, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 199, + 490, + 214 + ], + "spans": [ + { + "bbox": [ + 120, + 199, + 490, + 214 + ], + "score": 1.0, + "content": "Table 1: Optimal hyperparameter settings for the models in the centralised controller experiment.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 137, + 232, + 473, + 385 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 137, + 232, + 473, + 385 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 137, + 232, + 473, + 385 + ], + "spans": [ + { + "bbox": [ + 137, + 232, + 473, + 385 + ], + "score": 0.98, + "html": "
HyperparameterCleanupHarvest
A3C baselineComm. baselineInfluence comm.A3C Comm. baseline baselineInfluence comm.
Entropy reg..00176.000249.00305.000687.000174.00220
lr_init.00126.00223.00249.00136.00137.000413
lr_end.000012.000022.0000127.000028.0000127.000049
Influence weight β=02.752=04.825
Extrinsic reward01.0
weight α
Curriculum C1 KL8
Policy comparison.000789KL
Comm. entropy reg..0758.00208
Comm. loss weight9.0709
Symbol vocab size327
Comm. embedding16
", + "type": "table", + "image_path": "a539e1e377f59920fd6287d7d14b0f7b1c52e3c2927f2e15dece92f3ff011d12.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 137, + 232, + 473, + 283.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 137, + 283.0, + 473, + 334.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 137, + 334.0, + 473, + 385.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 132, + 392, + 477, + 404 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 130, + 391, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 479, + 406 + ], + "score": 1.0, + "content": "Table 2: Optimal hyperparameter settings for the models in the communication experiment.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "table", + "bbox": [ + 143, + 425, + 469, + 547 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 143, + 425, + 469, + 547 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 143, + 425, + 469, + 547 + ], + "spans": [ + { + "bbox": [ + 143, + 425, + 469, + 547 + ], + "score": 0.979, + "html": "
HyperparameterCleanupHarvest
A3C baselineMOA baselineInfluence MOAA3C baselineMOA baselineInfluence MOA
Entropy reg..00176.00176.00176.000687.00495.00223
lr_init.00126.00123.00123.00136.00206.00120
lr_end.000012.000012.000012.000028.000022.000044
Influence weight β0.620=02.521
MOA loss weight1.31215.007=1.71110.911
Curriculum C-40=1226
Policy comparison=KL==KL
Train MOA onlyFalseTrueFalseTrue
when visible
", + "type": "table", + "image_path": "d4fbff5faab5fec4aac6c7522438316d93a7beef9c4c8cf0187176e97538ea27.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 143, + 425, + 469, + 465.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 143, + 465.6666666666667, + 469, + 506.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 143, + 506.33333333333337, + 469, + 547.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 555, + 504, + 567 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 569 + ], + "score": 1.0, + "content": "Table 3: Optimal hyperparameter settings for the models in the model of other agents (MOA) experiment.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "in the loss function. We also vary the number of communication symbols that the agents can emit, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "the size of the linear layer that connects the LSTM to the communication policy layer, which we term", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "the communication embedding size. Finally, in the communication regime, we experiment to setting the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 236, + 644 + ], + "score": 1.0, + "content": "weight on the extrinsic reward E,", + "type": "text" + }, + { + "bbox": [ + 237, + 633, + 244, + 641 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 630, + 505, + 644 + ], + "score": 1.0, + "content": ", to zero. The best hyperparameters for each of the communication", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 642, + 219, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 219, + 653 + ], + "score": 1.0, + "content": "models are shown in Table 2.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 675, + 366, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 675, + 367, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 367, + 688 + ], + "score": 1.0, + "content": "6.3.3 MODEL OF OTHER AGENTS (MOA) HYPERPARAMETERS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "The MOA hyperparameters include whether to only train the MOA with cross-entropy loss on the actions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "of agents that are visible, and how much to weight the supervised loss in the overall loss of the model. The", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 720, + 273, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 273, + 732 + ], + "score": 1.0, + "content": "best hyperparameters are shown in Table 3.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 298, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 299, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 79, + 497, + 193 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 118, + 79, + 497, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 79, + 497, + 193 + ], + "spans": [ + { + "bbox": [ + 118, + 79, + 497, + 193 + ], + "score": 0.979, + "html": "
HyperparameterCleanupHarvest
A3CVisibleactionsInfluenceA3C baselineVisibleactionsInfluence
Entropy reg.baseline .00176baseline .00176.000248.000687baseline .00184.00025
lr_init.00126.00126.00107.00136.00215.00107
lr_end.000012.000012.000042.000028.000013.000042
Number of influencers3133
Influence weight β0.1460.224
Curriculum C11401140
Policy comparisonJSDPMI
Influencee reward10
", + "type": "table", + "image_path": "f4c2027e5c945541b51744fd4ee3d446041a4b1d05ad3603be18149e0ee814e4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 79, + 497, + 117.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 117.0, + 497, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 155.0, + 497, + 193.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 121, + 200, + 488, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 199, + 490, + 214 + ], + "spans": [ + { + "bbox": [ + 120, + 199, + 490, + 214 + ], + "score": 1.0, + "content": "Table 1: Optimal hyperparameter settings for the models in the centralised controller experiment.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 137, + 232, + 473, + 385 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 137, + 232, + 473, + 385 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 137, + 232, + 473, + 385 + ], + "spans": [ + { + "bbox": [ + 137, + 232, + 473, + 385 + ], + "score": 0.98, + "html": "
HyperparameterCleanupHarvest
A3C baselineComm. baselineInfluence comm.A3C Comm. baseline baselineInfluence comm.
Entropy reg..00176.000249.00305.000687.000174.00220
lr_init.00126.00223.00249.00136.00137.000413
lr_end.000012.000022.0000127.000028.0000127.000049
Influence weight β=02.752=04.825
Extrinsic reward01.0
weight α
Curriculum C1 KL8
Policy comparison.000789KL
Comm. entropy reg..0758.00208
Comm. loss weight9.0709
Symbol vocab size327
Comm. embedding16
", + "type": "table", + "image_path": "a539e1e377f59920fd6287d7d14b0f7b1c52e3c2927f2e15dece92f3ff011d12.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 137, + 232, + 473, + 283.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 137, + 283.0, + 473, + 334.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 137, + 334.0, + 473, + 385.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 132, + 392, + 477, + 404 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 130, + 391, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 479, + 406 + ], + "score": 1.0, + "content": "Table 2: Optimal hyperparameter settings for the models in the communication experiment.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "table", + "bbox": [ + 143, + 425, + 469, + 547 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 143, + 425, + 469, + 547 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 143, + 425, + 469, + 547 + ], + "spans": [ + { + "bbox": [ + 143, + 425, + 469, + 547 + ], + "score": 0.979, + "html": "
HyperparameterCleanupHarvest
A3C baselineMOA baselineInfluence MOAA3C baselineMOA baselineInfluence MOA
Entropy reg..00176.00176.00176.000687.00495.00223
lr_init.00126.00123.00123.00136.00206.00120
lr_end.000012.000012.000012.000028.000022.000044
Influence weight β0.620=02.521
MOA loss weight1.31215.007=1.71110.911
Curriculum C-40=1226
Policy comparison=KL==KL
Train MOA onlyFalseTrueFalseTrue
when visible
", + "type": "table", + "image_path": "d4fbff5faab5fec4aac6c7522438316d93a7beef9c4c8cf0187176e97538ea27.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 143, + 425, + 469, + 465.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 143, + 465.6666666666667, + 469, + 506.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 143, + 506.33333333333337, + 469, + 547.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 555, + 504, + 567 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 569 + ], + "score": 1.0, + "content": "Table 3: Optimal hyperparameter settings for the models in the model of other agents (MOA) experiment.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 653 + ], + "lines": [], + "index": 14, + "bbox_fs": [ + 105, + 598, + 505, + 653 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 675, + 366, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 675, + 367, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 367, + 688 + ], + "score": 1.0, + "content": "6.3.3 MODEL OF OTHER AGENTS (MOA) HYPERPARAMETERS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "The MOA hyperparameters include whether to only train the MOA with cross-entropy loss on the actions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "of agents that are visible, and how much to weight the supervised loss in the overall loss of the model. The", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 720, + 273, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 273, + 732 + ], + "score": 1.0, + "content": "best hyperparameters are shown in Table 3.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 699, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 80, + 428, + 148 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 184, + 80, + 428, + 148 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 80, + 428, + 148 + ], + "spans": [ + { + "bbox": [ + 184, + 80, + 428, + 148 + ], + "score": 0.969, + "html": "
CleanupHarvest
A3Cbaseline89485
Inequity aversion (Hughes et al., 2018)275750
Influence - Basic1901073
Influence - Communication166951
Influence -Model of other agents392588
", + "type": "table", + "image_path": "4ecc39bab522686cadd4d604d2910d079ffcab346895d226ba7b33dec2aefc19.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 184, + 80, + 428, + 93.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 184, + 93.6, + 428, + 107.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 184, + 107.19999999999999, + 428, + 120.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 184, + 120.79999999999998, + 428, + 134.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 184, + 134.39999999999998, + 428, + 147.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 156, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "Table 4: Final collective reward over the last 50 agent steps for each of the models considered. Bolded", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "score": 1.0, + "content": "entries represent experiments in which the influence models significantly outperformed the scores reported", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 177, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 192 + ], + "score": 1.0, + "content": "in previous work on inequity aversion(Hughes et al., 2018). This is impressive, considering the inequity", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "score": 1.0, + "content": "averse agents are able to view all other agents’ rewards. We make no such assumption, and yet are able to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 200, + 261, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 261, + 212 + ], + "score": 1.0, + "content": "achieve similar or superior performance.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 232, + 254, + 243 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 256, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 256, + 245 + ], + "score": 1.0, + "content": "6.3.4 COMMUNICATION ANALYSIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 251, + 288, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 289, + 264 + ], + "score": 1.0, + "content": "The speaker consistency metric is calculated as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 266, + 425, + 301 + ], + "lines": [ + { + "bbox": [ + 184, + 266, + 425, + 301 + ], + "spans": [ + { + "bbox": [ + 184, + 266, + 425, + 301 + ], + "score": 0.94, + "content": "\\sum _ { k = 1 } ^ { N } 0 . 5 [ \\sum _ { c } 1 - \\frac { H ( p ( a ^ { k } | m ^ { k } = c ) ) } { H _ { m a x } } + \\sum _ { a } 1 - \\frac { H ( p ( m ^ { k } | a ^ { k } = a ) ) } { H _ { m a x } } ] ,", + "type": "interline_equation", + "image_path": "3426eefe40c38e35aa09a3955c858cd87f99b15cce7a8836f10826943feba7cc.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 266, + 425, + 283.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 184, + 283.5, + 425, + 301.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 304, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 132, + 317 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 305, + 143, + 315 + ], + "score": 0.82, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 304, + 252, + 317 + ], + "score": 1.0, + "content": "is the entropy function and", + "type": "text" + }, + { + "bbox": [ + 252, + 305, + 277, + 316 + ], + "score": 0.92, + "content": "H _ { m a x }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "is the maximum entropy based on the number of discrete", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 315, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 329 + ], + "score": 1.0, + "content": "symbols or actions. The goal of the metric is to measure how much of a 1:1 correspondence exists between", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 326, + 341, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 341, + 340 + ], + "score": 1.0, + "content": "a speaker’s action and the speaker’s communication message.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 107, + 349, + 222, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 223, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 223, + 361 + ], + "score": 1.0, + "content": "6.4 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 216, + 365, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 214, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 214, + 363, + 505, + 378 + ], + "score": 1.0, + "content": "Figure 12 shows an additional moment of high influence in the Cleanup", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 214, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 214, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "game. The purple influencer agent can see the area within the white box, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 215, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 215, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "therefore all of the apple patch. The field-of-view of the magenta influencee", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 214, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 214, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "is outlined with the magenta box; it cannot see if apples have appeared, even", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 214, + 409, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 214, + 409, + 505, + 421 + ], + "score": 1.0, + "content": "though it has been cleaning the river, which is the action required to cause", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 214, + 420, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 214, + 420, + 506, + 432 + ], + "score": 1.0, + "content": "apples to appear. When the purple influencer turns left and does not move", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 215, + 432, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 215, + 432, + 505, + 443 + ], + "score": 1.0, + "content": "towards the apple patch, this signals to the magenta agent that no apples have", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 214, + 442, + 437, + 455 + ], + "spans": [ + { + "bbox": [ + 214, + 442, + 437, + 455 + ], + "score": 1.0, + "content": "appeared, since otherwise the influence would move right.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "image", + "bbox": [ + 108, + 368, + 202, + 507 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 368, + 202, + 507 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 368, + 202, + 507 + ], + "spans": [ + { + "bbox": [ + 108, + 368, + 202, + 507 + ], + "score": 0.91, + "type": "image", + "image_path": "fda9989ec537ce17189a8d0151565ce65db1893467693e2932bcd54bd93d57e3.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 108, + 368, + 202, + 507 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 517, + 208, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 516, + 208, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 208, + 528 + ], + "score": 1.0, + "content": "Figure 12: A moment of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 527, + 208, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 208, + 539 + ], + "score": 1.0, + "content": "high influence between the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 209, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 209, + 551 + ], + "score": 1.0, + "content": "purple influencer and ma-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 550, + 173, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 173, + 561 + ], + "score": 1.0, + "content": "genta influencee.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + } + ], + "index": 26.25 + }, + { + "type": "text", + "bbox": [ + 216, + 459, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 214, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 214, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "Table 4 presents the final collective reward obtained by each of the models", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 214, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 214, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "tested in the experiments presented in Section 4. We see that in several cases,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 215, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 215, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "the influence agents are even able to out-perform the state-of-the-art results on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 215, + 492, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 215, + 492, + 505, + 503 + ], + "score": 1.0, + "content": "these tasks reported by Hughes et al. (2018), despite the fact that the solution", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 214, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 214, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "proposed by Hughes et al. (2018) requires that agents can view other agents’", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 214, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 214, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "rewards, whereas we do not make this assumption, and instead only require", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 214, + 525, + 374, + 536 + ], + "spans": [ + { + "bbox": [ + 214, + 525, + 374, + 536 + ], + "score": 1.0, + "content": "that agents can view each others’ actions.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 215, + 541, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 214, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 214, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "It is important to note that collective reward is not always the perfect metric", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 214, + 552, + 507, + 565 + ], + "spans": [ + { + "bbox": [ + 214, + 552, + 507, + 565 + ], + "score": 1.0, + "content": "of cooperative behavior, a finding that was also discovered by Barton et al.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 214, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 214, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "(2018) and emphasized by Leibo et al. (2017). In the case, we find that there", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "score": 1.0, + "content": "is a spurious solution to the Harvest game, in which one agent fails to learn and fails to collect any apples.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "This leads to very high collective reward, since it means there is one fewer agent that can exploit the others,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "and makes sustainable harvesting easier to achieve. Therefore, for the results shown in the paper, we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "eliminate any random seed in Harvest for which one of the agents has failed to learn to collect apples, as in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 618, + 246, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 246, + 630 + ], + "score": 1.0, + "content": "previous work (Hughes et al., 2018).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 635, + 504, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "However, here we also present an alternative strategy for assessing the overall collective outcomes:", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "score": 1.0, + "content": "weighting the total collective reward by an index of equality of the individual returns. Specifically, we", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 657, + 362, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 249, + 670 + ], + "score": 1.0, + "content": "compute the Gini coefficient over the", + "type": "text" + }, + { + "bbox": [ + 250, + 658, + 260, + 667 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 657, + 362, + 670 + ], + "score": 1.0, + "content": "agents’ individual returns:", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + }, + { + "type": "interline_equation", + "bbox": [ + 254, + 672, + 356, + 706 + ], + "lines": [ + { + "bbox": [ + 254, + 672, + 356, + 706 + ], + "spans": [ + { + "bbox": [ + 254, + 672, + 356, + 706 + ], + "score": 0.95, + "content": "G = \\frac { \\sum _ { i = 1 } ^ { N } \\sum _ { j = 1 } ^ { N } \\left| { r ^ { i } - r ^ { j } } \\right| } { 2 N \\sum _ { i = 1 } ^ { N } r ^ { i } } ,", + "type": "interline_equation", + "image_path": "84ca2cb9245eceaa6294138a836191940ef7378fa0c97c845f4e2a3c0ef3313a.jpg" + } + ] + } + ], + "index": 49.5, + "virtual_lines": [ + { + "bbox": [ + 254, + 672, + 356, + 689.0 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 254, + 689.0, + 356, + 706.0 + ], + "spans": [], + "index": 50 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 353, + 722 + ], + "score": 1.0, + "content": "which gives us a measure of the inequality of the returns, where", + "type": "text" + }, + { + "bbox": [ + 353, + 709, + 388, + 721 + ], + "score": 0.88, + "content": "G \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 709, + 411, + 722 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 411, + 710, + 434, + 720 + ], + "score": 0.89, + "content": "G = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "indicating perfect", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 170, + 733 + ], + "score": 1.0, + "content": "equality. Thus,", + "type": "text" + }, + { + "bbox": [ + 170, + 721, + 195, + 731 + ], + "score": 0.87, + "content": "1 - G", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is a measure of equality; we use this to weight the collective reward for each", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 298, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 80, + 428, + 148 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 184, + 80, + 428, + 148 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 80, + 428, + 148 + ], + "spans": [ + { + "bbox": [ + 184, + 80, + 428, + 148 + ], + "score": 0.969, + "html": "
CleanupHarvest
A3Cbaseline89485
Inequity aversion (Hughes et al., 2018)275750
Influence - Basic1901073
Influence - Communication166951
Influence -Model of other agents392588
", + "type": "table", + "image_path": "4ecc39bab522686cadd4d604d2910d079ffcab346895d226ba7b33dec2aefc19.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 184, + 80, + 428, + 93.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 184, + 93.6, + 428, + 107.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 184, + 107.19999999999999, + 428, + 120.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 184, + 120.79999999999998, + 428, + 134.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 184, + 134.39999999999998, + 428, + 147.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 156, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "Table 4: Final collective reward over the last 50 agent steps for each of the models considered. Bolded", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "score": 1.0, + "content": "entries represent experiments in which the influence models significantly outperformed the scores reported", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 177, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 192 + ], + "score": 1.0, + "content": "in previous work on inequity aversion(Hughes et al., 2018). This is impressive, considering the inequity", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "score": 1.0, + "content": "averse agents are able to view all other agents’ rewards. We make no such assumption, and yet are able to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 200, + 261, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 261, + 212 + ], + "score": 1.0, + "content": "achieve similar or superior performance.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 155, + 506, + 212 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 232, + 254, + 243 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 256, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 256, + 245 + ], + "score": 1.0, + "content": "6.3.4 COMMUNICATION ANALYSIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 251, + 288, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 250, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 289, + 264 + ], + "score": 1.0, + "content": "The speaker consistency metric is calculated as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 250, + 289, + 264 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 266, + 425, + 301 + ], + "lines": [ + { + "bbox": [ + 184, + 266, + 425, + 301 + ], + "spans": [ + { + "bbox": [ + 184, + 266, + 425, + 301 + ], + "score": 0.94, + "content": "\\sum _ { k = 1 } ^ { N } 0 . 5 [ \\sum _ { c } 1 - \\frac { H ( p ( a ^ { k } | m ^ { k } = c ) ) } { H _ { m a x } } + \\sum _ { a } 1 - \\frac { H ( p ( m ^ { k } | a ^ { k } = a ) ) } { H _ { m a x } } ] ,", + "type": "interline_equation", + "image_path": "3426eefe40c38e35aa09a3955c858cd87f99b15cce7a8836f10826943feba7cc.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 266, + 425, + 283.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 184, + 283.5, + 425, + 301.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 304, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 132, + 317 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 305, + 143, + 315 + ], + "score": 0.82, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 304, + 252, + 317 + ], + "score": 1.0, + "content": "is the entropy function and", + "type": "text" + }, + { + "bbox": [ + 252, + 305, + 277, + 316 + ], + "score": 0.92, + "content": "H _ { m a x }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "is the maximum entropy based on the number of discrete", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 315, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 329 + ], + "score": 1.0, + "content": "symbols or actions. The goal of the metric is to measure how much of a 1:1 correspondence exists between", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 326, + 341, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 341, + 340 + ], + "score": 1.0, + "content": "a speaker’s action and the speaker’s communication message.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 304, + 506, + 340 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 349, + 222, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 223, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 223, + 361 + ], + "score": 1.0, + "content": "6.4 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 216, + 365, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 214, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 214, + 363, + 505, + 378 + ], + "score": 1.0, + "content": "Figure 12 shows an additional moment of high influence in the Cleanup", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 214, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 214, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "game. The purple influencer agent can see the area within the white box, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 215, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 215, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "therefore all of the apple patch. The field-of-view of the magenta influencee", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 214, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 214, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "is outlined with the magenta box; it cannot see if apples have appeared, even", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 214, + 409, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 214, + 409, + 505, + 421 + ], + "score": 1.0, + "content": "though it has been cleaning the river, which is the action required to cause", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 214, + 420, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 214, + 420, + 506, + 432 + ], + "score": 1.0, + "content": "apples to appear. When the purple influencer turns left and does not move", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 215, + 432, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 215, + 432, + 505, + 443 + ], + "score": 1.0, + "content": "towards the apple patch, this signals to the magenta agent that no apples have", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 214, + 442, + 437, + 455 + ], + "spans": [ + { + "bbox": [ + 214, + 442, + 437, + 455 + ], + "score": 1.0, + "content": "appeared, since otherwise the influence would move right.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 214, + 363, + 506, + 455 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 368, + 202, + 507 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 368, + 202, + 507 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 368, + 202, + 507 + ], + "spans": [ + { + "bbox": [ + 108, + 368, + 202, + 507 + ], + "score": 0.91, + "type": "image", + "image_path": "fda9989ec537ce17189a8d0151565ce65db1893467693e2932bcd54bd93d57e3.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 108, + 368, + 202, + 507 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 517, + 208, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 516, + 208, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 208, + 528 + ], + "score": 1.0, + "content": "Figure 12: A moment of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 527, + 208, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 208, + 539 + ], + "score": 1.0, + "content": "high influence between the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 209, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 209, + 551 + ], + "score": 1.0, + "content": "purple influencer and ma-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 550, + 173, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 173, + 561 + ], + "score": 1.0, + "content": "genta influencee.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + } + ], + "index": 26.25 + }, + { + "type": "text", + "bbox": [ + 216, + 459, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 214, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 214, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "Table 4 presents the final collective reward obtained by each of the models", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 214, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 214, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "tested in the experiments presented in Section 4. We see that in several cases,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 215, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 215, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "the influence agents are even able to out-perform the state-of-the-art results on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 215, + 492, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 215, + 492, + 505, + 503 + ], + "score": 1.0, + "content": "these tasks reported by Hughes et al. (2018), despite the fact that the solution", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 214, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 214, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "proposed by Hughes et al. (2018) requires that agents can view other agents’", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 214, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 214, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "rewards, whereas we do not make this assumption, and instead only require", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 214, + 525, + 374, + 536 + ], + "spans": [ + { + "bbox": [ + 214, + 525, + 374, + 536 + ], + "score": 1.0, + "content": "that agents can view each others’ actions.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 214, + 459, + 506, + 536 + ] + }, + { + "type": "text", + "bbox": [ + 215, + 541, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 214, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 214, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "It is important to note that collective reward is not always the perfect metric", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 214, + 552, + 507, + 565 + ], + "spans": [ + { + "bbox": [ + 214, + 552, + 507, + 565 + ], + "score": 1.0, + "content": "of cooperative behavior, a finding that was also discovered by Barton et al.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 214, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 214, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "(2018) and emphasized by Leibo et al. (2017). In the case, we find that there", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "score": 1.0, + "content": "is a spurious solution to the Harvest game, in which one agent fails to learn and fails to collect any apples.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "This leads to very high collective reward, since it means there is one fewer agent that can exploit the others,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "and makes sustainable harvesting easier to achieve. Therefore, for the results shown in the paper, we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "eliminate any random seed in Harvest for which one of the agents has failed to learn to collect apples, as in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 618, + 246, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 246, + 630 + ], + "score": 1.0, + "content": "previous work (Hughes et al., 2018).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39, + "bbox_fs": [ + 214, + 541, + 507, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 630 + ], + "lines": [], + "index": 43, + "bbox_fs": [ + 105, + 575, + 506, + 630 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 635, + 504, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "However, here we also present an alternative strategy for assessing the overall collective outcomes:", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "score": 1.0, + "content": "weighting the total collective reward by an index of equality of the individual returns. Specifically, we", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 657, + 362, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 249, + 670 + ], + "score": 1.0, + "content": "compute the Gini coefficient over the", + "type": "text" + }, + { + "bbox": [ + 250, + 658, + 260, + 667 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 657, + 362, + 670 + ], + "score": 1.0, + "content": "agents’ individual returns:", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 635, + 506, + 670 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 254, + 672, + 356, + 706 + ], + "lines": [ + { + "bbox": [ + 254, + 672, + 356, + 706 + ], + "spans": [ + { + "bbox": [ + 254, + 672, + 356, + 706 + ], + "score": 0.95, + "content": "G = \\frac { \\sum _ { i = 1 } ^ { N } \\sum _ { j = 1 } ^ { N } \\left| { r ^ { i } - r ^ { j } } \\right| } { 2 N \\sum _ { i = 1 } ^ { N } r ^ { i } } ,", + "type": "interline_equation", + "image_path": "84ca2cb9245eceaa6294138a836191940ef7378fa0c97c845f4e2a3c0ef3313a.jpg" + } + ] + } + ], + "index": 49.5, + "virtual_lines": [ + { + "bbox": [ + 254, + 672, + 356, + 689.0 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 254, + 689.0, + 356, + 706.0 + ], + "spans": [], + "index": 50 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 353, + 722 + ], + "score": 1.0, + "content": "which gives us a measure of the inequality of the returns, where", + "type": "text" + }, + { + "bbox": [ + 353, + 709, + 388, + 721 + ], + "score": 0.88, + "content": "G \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 709, + 411, + 722 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 411, + 710, + 434, + 720 + ], + "score": 0.89, + "content": "G = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "indicating perfect", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 170, + 733 + ], + "score": 1.0, + "content": "equality. Thus,", + "type": "text" + }, + { + "bbox": [ + 170, + 721, + 195, + 731 + ], + "score": 0.87, + "content": "1 - G", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is a measure of equality; we use this to weight the collective reward for each", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "experiment, and plot the results in Figure 13. Once again, we see that the influence models give the highest", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 280, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 280, + 105 + ], + "score": 1.0, + "content": "final performance, even with this new metric.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "experiment, and plot the results in Figure 13. Once again, we see that the influence models give the highest", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 280, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 280, + 105 + ], + "score": 1.0, + "content": "final performance, even with this new metric.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "image", + "bbox": [ + 115, + 143, + 501, + 590 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 143, + 501, + 590 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 143, + 501, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 501, + 590 + ], + "score": 0.977, + "type": "image", + "image_path": "c923ae01cca519ebc26562fb33c66fec40b0011ec5f33643088a3be15814576d.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 115, + 143, + 501, + 292.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 115, + 292.0, + 501, + 441.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 115, + 441.0, + 501, + 590.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 599, + 505, + 642 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 291, + 611 + ], + "score": 1.0, + "content": "Figure 13: Total collective reward times equality,", + "type": "text" + }, + { + "bbox": [ + 291, + 599, + 334, + 611 + ], + "score": 0.93, + "content": "R * ( 1 - G )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 599, + 505, + 611 + ], + "score": 1.0, + "content": ", obtained in all experiments. Error bars show", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 113, + 622 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 113, + 610, + 140, + 621 + ], + "score": 0.91, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "confidence interval (CI) over 5 random seeds, computed within a sliding window of 200 agent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "steps. Once again, the models trained with influence reward (red) significantly outperform the baseline and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 632, + 168, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 168, + 643 + ], + "score": 1.0, + "content": "ablated models.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Finally, we would like to show that the influence reward is robust to the choice of hyperparameter settings.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Therefore, in Figure 14, we plot the collective reward of the top 5 best hyperparameter settings for each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "experiment, over 5 random seeds each. Once again, the influence models result in higher collective reward,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 720, + 420, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 420, + 732 + ], + "score": 1.0, + "content": "which provides evidence that the model is robust to the choice of hyperparameters.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 105 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 115, + 143, + 501, + 590 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 143, + 501, + 590 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 143, + 501, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 501, + 590 + ], + "score": 0.977, + "type": "image", + "image_path": "c923ae01cca519ebc26562fb33c66fec40b0011ec5f33643088a3be15814576d.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 115, + 143, + 501, + 292.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 115, + 292.0, + 501, + 441.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 115, + 441.0, + 501, + 590.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 599, + 505, + 642 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 291, + 611 + ], + "score": 1.0, + "content": "Figure 13: Total collective reward times equality,", + "type": "text" + }, + { + "bbox": [ + 291, + 599, + 334, + 611 + ], + "score": 0.93, + "content": "R * ( 1 - G )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 599, + 505, + 611 + ], + "score": 1.0, + "content": ", obtained in all experiments. Error bars show", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 113, + 622 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 113, + 610, + 140, + 621 + ], + "score": 0.91, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "confidence interval (CI) over 5 random seeds, computed within a sliding window of 200 agent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "steps. Once again, the models trained with influence reward (red) significantly outperform the baseline and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 632, + 168, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 168, + 643 + ], + "score": 1.0, + "content": "ablated models.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Finally, we would like to show that the influence reward is robust to the choice of hyperparameter settings.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Therefore, in Figure 14, we plot the collective reward of the top 5 best hyperparameter settings for each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "experiment, over 5 random seeds each. Once again, the influence models result in higher collective reward,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 720, + 420, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 420, + 732 + ], + "score": 1.0, + "content": "which provides evidence that the model is robust to the choice of hyperparameters.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 78, + 501, + 525 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 78, + 501, + 525 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 78, + 501, + 525 + ], + "spans": [ + { + "bbox": [ + 110, + 78, + 501, + 525 + ], + "score": 0.976, + "type": "image", + "image_path": "1408fa6f4ede73df1fdda59c904bd76d77d6315c5152307ac31700974d00769d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 78, + 501, + 227.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 227.0, + 501, + 376.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 376.0, + 501, + 525.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 532, + 506, + 577 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "Figure 14: Total collective reward over the top 5 hyperparameter settings, with 5 random seeds each, for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 543, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 239, + 555 + ], + "score": 1.0, + "content": "all experiments. Error bars show a", + "type": "text" + }, + { + "bbox": [ + 240, + 543, + 266, + 554 + ], + "score": 0.9, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 543, + 506, + 555 + ], + "score": 1.0, + "content": "confidence interval (CI) computed within a sliding window of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 554, + 507, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 507, + 567 + ], + "score": 1.0, + "content": "200 agent steps. The influence models still maintain an advantage over the baselines and ablated models,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 565, + 354, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 354, + 578 + ], + "score": 1.0, + "content": "suggesting the technique is robust to the hyperparameter settings.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "title", + "bbox": [ + 108, + 599, + 299, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 301, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 301, + 612 + ], + "score": 1.0, + "content": "6.4.1 OPTIMIZING FOR COLLECTIVE REWARD", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 618, + 506, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "In this section we include the results of training explicitly prosocial agents, which directly optimize for the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "collective reward of all agents. Previous work (e.g. Peysakhovich & Lerer (2018)) has shown that training", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "agents to optimize for the rewards of other agents can help the group to obtain better collective outcomes.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 649, + 506, + 686 + ], + "spans": [ + { + "bbox": [ + 100, + 649, + 195, + 686 + ], + "score": 1.0, + "content": "Following a similar prown individual reward", + "type": "text" + }, + { + "bbox": [ + 196, + 664, + 209, + 675 + ], + "score": 0.89, + "content": "E ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 649, + 383, + 686 + ], + "score": 1.0, + "content": "ple, we implemented agents that optimize foand the collective reward of all other agents,", + "type": "text" + }, + { + "bbox": [ + 383, + 663, + 433, + 679 + ], + "score": 0.94, + "content": "\\textstyle \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 649, + 506, + 686 + ], + "score": 1.0, + "content": "mbination of their. Thus, the reward", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 103, + 676, + 507, + 698 + ], + "spans": [ + { + "bbox": [ + 103, + 676, + 176, + 698 + ], + "score": 1.0, + "content": "function for agent", + "type": "text" + }, + { + "bbox": [ + 176, + 680, + 182, + 690 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 676, + 192, + 698 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 192, + 677, + 290, + 694 + ], + "score": 0.92, + "content": "\\begin{array} { r } { { R ^ { k } } = E ^ { k } + \\eta \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 676, + 507, + 698 + ], + "score": 1.0, + "content": ". We conducted the same hyperparameter search over the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 692, + 484, + 706 + ], + "spans": [ + { + "bbox": [ + 104, + 692, + 448, + 706 + ], + "score": 1.0, + "content": "parameters mentioned in Section 6.3.1 varying the weight placed on the collective reward,", + "type": "text" + }, + { + "bbox": [ + 448, + 692, + 480, + 705 + ], + "score": 0.69, + "content": "\\eta \\in [ 0 , 2 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 692, + 484, + 706 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "As expected, we find that agents trained to optimize for collective reward attain higher collective reward", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 469, + 733 + ], + "score": 1.0, + "content": "in both Cleanup and Harvest, as is shown in Figure 15. In both games, the optimal value for", + "type": "text" + }, + { + "bbox": [ + 469, + 721, + 503, + 732 + ], + "score": 0.9, + "content": "\\eta = 0 . 8 5", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 720, + 506, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 298, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 298, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 78, + 501, + 525 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 78, + 501, + 525 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 78, + 501, + 525 + ], + "spans": [ + { + "bbox": [ + 110, + 78, + 501, + 525 + ], + "score": 0.976, + "type": "image", + "image_path": "1408fa6f4ede73df1fdda59c904bd76d77d6315c5152307ac31700974d00769d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 78, + 501, + 227.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 227.0, + 501, + 376.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 376.0, + 501, + 525.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 532, + 506, + 577 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "Figure 14: Total collective reward over the top 5 hyperparameter settings, with 5 random seeds each, for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 543, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 239, + 555 + ], + "score": 1.0, + "content": "all experiments. Error bars show a", + "type": "text" + }, + { + "bbox": [ + 240, + 543, + 266, + 554 + ], + "score": 0.9, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 543, + 506, + 555 + ], + "score": 1.0, + "content": "confidence interval (CI) computed within a sliding window of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 554, + 507, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 507, + 567 + ], + "score": 1.0, + "content": "200 agent steps. The influence models still maintain an advantage over the baselines and ablated models,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 565, + 354, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 354, + 578 + ], + "score": 1.0, + "content": "suggesting the technique is robust to the hyperparameter settings.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "title", + "bbox": [ + 108, + 599, + 299, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 301, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 301, + 612 + ], + "score": 1.0, + "content": "6.4.1 OPTIMIZING FOR COLLECTIVE REWARD", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 618, + 506, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "In this section we include the results of training explicitly prosocial agents, which directly optimize for the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "collective reward of all agents. Previous work (e.g. Peysakhovich & Lerer (2018)) has shown that training", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "agents to optimize for the rewards of other agents can help the group to obtain better collective outcomes.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 649, + 506, + 686 + ], + "spans": [ + { + "bbox": [ + 100, + 649, + 195, + 686 + ], + "score": 1.0, + "content": "Following a similar prown individual reward", + "type": "text" + }, + { + "bbox": [ + 196, + 664, + 209, + 675 + ], + "score": 0.89, + "content": "E ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 649, + 383, + 686 + ], + "score": 1.0, + "content": "ple, we implemented agents that optimize foand the collective reward of all other agents,", + "type": "text" + }, + { + "bbox": [ + 383, + 663, + 433, + 679 + ], + "score": 0.94, + "content": "\\textstyle \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 649, + 506, + 686 + ], + "score": 1.0, + "content": "mbination of their. Thus, the reward", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 103, + 676, + 507, + 698 + ], + "spans": [ + { + "bbox": [ + 103, + 676, + 176, + 698 + ], + "score": 1.0, + "content": "function for agent", + "type": "text" + }, + { + "bbox": [ + 176, + 680, + 182, + 690 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 676, + 192, + 698 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 192, + 677, + 290, + 694 + ], + "score": 0.92, + "content": "\\begin{array} { r } { { R ^ { k } } = E ^ { k } + \\eta \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 676, + 507, + 698 + ], + "score": 1.0, + "content": ". We conducted the same hyperparameter search over the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 692, + 484, + 706 + ], + "spans": [ + { + "bbox": [ + 104, + 692, + 448, + 706 + ], + "score": 1.0, + "content": "parameters mentioned in Section 6.3.1 varying the weight placed on the collective reward,", + "type": "text" + }, + { + "bbox": [ + 448, + 692, + 480, + 705 + ], + "score": 0.69, + "content": "\\eta \\in [ 0 , 2 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 692, + 484, + 706 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 100, + 619, + 507, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "As expected, we find that agents trained to optimize for collective reward attain higher collective reward", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 469, + 733 + ], + "score": 1.0, + "content": "in both Cleanup and Harvest, as is shown in Figure 15. In both games, the optimal value for", + "type": "text" + }, + { + "bbox": [ + 469, + 721, + 503, + 732 + ], + "score": 0.9, + "content": "\\eta = 0 . 8 5", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 720, + 506, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 709, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 77, + 494, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 77, + 494, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 77, + 494, + 214 + ], + "spans": [ + { + "bbox": [ + 117, + 77, + 494, + 214 + ], + "score": 0.959, + "type": "image", + "image_path": "fe3d4f89e8a19d1b5fe2a4f19403d717b5760772f2debd3fb1d1b879cb6bc09c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 77, + 494, + 122.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 122.66666666666666, + 494, + 168.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 168.33333333333331, + 494, + 213.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 223, + 505, + 257 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "Figure 15: Total collective reward obtained by agents trained to optimize for the collective reward, for the 5", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 387, + 247 + ], + "score": 1.0, + "content": "best hyperparameter settings with 5 random seeds each. Error bars show a", + "type": "text" + }, + { + "bbox": [ + 387, + 235, + 413, + 245 + ], + "score": 0.87, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "confidence interval (CI)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 245, + 313, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 313, + 259 + ], + "score": 1.0, + "content": "computed within a sliding window of 200 agent steps.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Interestingly, however, the equality in the individual returns for these agents is extremely low. Across", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "the hyperparameter sweep, no solution to the Cleanup game which scored more than 20 points in terms", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "of collective return was found in which all agents scored an individual return above 0. It seems that in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "Cleanup, when agents are trained to optimize for collective return, they converge on a solution in which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 324, + 253, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 253, + 334 + ], + "score": 1.0, + "content": "some agents never receive any reward.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "Note that training agents to optimize for collective reward requires that each agent can view the rewards", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 350, + 504, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 504, + 362 + ], + "score": 1.0, + "content": "obtained by other agents. As discussed previously, the social influence reward is a novel way to obtain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 362, + 362, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 362, + 374 + ], + "score": 1.0, + "content": "cooperative behavior, that does not require making this assumption.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 297, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 299, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 299, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "20", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 77, + 494, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 77, + 494, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 77, + 494, + 214 + ], + "spans": [ + { + "bbox": [ + 117, + 77, + 494, + 214 + ], + "score": 0.959, + "type": "image", + "image_path": "fe3d4f89e8a19d1b5fe2a4f19403d717b5760772f2debd3fb1d1b879cb6bc09c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 77, + 494, + 122.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 122.66666666666666, + 494, + 168.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 168.33333333333331, + 494, + 213.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 223, + 505, + 257 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "Figure 15: Total collective reward obtained by agents trained to optimize for the collective reward, for the 5", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 235, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 387, + 247 + ], + "score": 1.0, + "content": "best hyperparameter settings with 5 random seeds each. Error bars show a", + "type": "text" + }, + { + "bbox": [ + 387, + 235, + 413, + 245 + ], + "score": 0.87, + "content": "9 9 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 235, + 505, + 247 + ], + "score": 1.0, + "content": "confidence interval (CI)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 245, + 313, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 313, + 259 + ], + "score": 1.0, + "content": "computed within a sliding window of 200 agent steps.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Interestingly, however, the equality in the individual returns for these agents is extremely low. Across", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "the hyperparameter sweep, no solution to the Cleanup game which scored more than 20 points in terms", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "of collective return was found in which all agents scored an individual return above 0. It seems that in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "Cleanup, when agents are trained to optimize for collective return, they converge on a solution in which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 324, + 253, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 253, + 334 + ], + "score": 1.0, + "content": "some agents never receive any reward.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 278, + 505, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "Note that training agents to optimize for collective reward requires that each agent can view the rewards", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 350, + 504, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 504, + 362 + ], + "score": 1.0, + "content": "obtained by other agents. As discussed previously, the social influence reward is a novel way to obtain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 362, + 362, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 362, + 374 + ], + "score": 1.0, + "content": "cooperative behavior, that does not require making this assumption.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 340, + 505, + 374 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/B1lG42C9Km/B1lG42C9Km_model.json b/parse/train/B1lG42C9Km/B1lG42C9Km_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ab988b166a3348e5432aeadc1c82a04e81bf2c1 --- /dev/null +++ b/parse/train/B1lG42C9Km/B1lG42C9Km_model.json @@ -0,0 +1,31159 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 590, + 1304, + 590, + 1304, + 1048, + 398, + 1048 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1560, + 1403, + 1560, + 1403, + 1865, + 298, + 1865 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1360, + 1403, + 1360, + 1403, + 1544, + 298, + 1544 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1161, + 1403, + 1161, + 1403, + 1345, + 298, + 1345 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1401, + 1881, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 298, + 219, + 1188, + 219, + 1188, + 324, + 298, + 324 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 315, + 378, + 661, + 378, + 661, + 438, + 315, + 438 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 302, + 1106, + 562, + 1106, + 562, + 1140, + 302, + 1140 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 828, + 75, + 828, + 104, + 300, + 104 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 776, + 519, + 924, + 519, + 924, + 553, + 776, + 553 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.701 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 216.0, + 1195.0, + 216.0, + 1195.0, + 273.0, + 292.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 275.0, + 960.0, + 275.0, + 960.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1103.0, + 567.0, + 1103.0, + 567.0, + 1146.0, + 293.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 831.0, + 72.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 518.0, + 928.0, + 518.0, + 928.0, + 556.0, + 772.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 587.0, + 1305.0, + 587.0, + 1305.0, + 626.0, + 394.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 621.0, + 1305.0, + 621.0, + 1305.0, + 656.0, + 394.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 652.0, + 1304.0, + 652.0, + 1304.0, + 684.0, + 394.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 683.0, + 1305.0, + 683.0, + 1305.0, + 717.0, + 394.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 713.0, + 1305.0, + 713.0, + 1305.0, + 746.0, + 393.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 743.0, + 1306.0, + 743.0, + 1306.0, + 778.0, + 392.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 775.0, + 1305.0, + 775.0, + 1305.0, + 807.0, + 394.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 806.0, + 1306.0, + 806.0, + 1306.0, + 837.0, + 395.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 834.0, + 1309.0, + 834.0, + 1309.0, + 869.0, + 392.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 862.0, + 1308.0, + 862.0, + 1308.0, + 900.0, + 393.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 896.0, + 1308.0, + 896.0, + 1308.0, + 927.0, + 394.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 926.0, + 1306.0, + 926.0, + 1306.0, + 961.0, + 394.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 957.0, + 1305.0, + 957.0, + 1305.0, + 989.0, + 395.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 983.0, + 1309.0, + 983.0, + 1309.0, + 1024.0, + 392.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1013.0, + 1311.0, + 1013.0, + 1311.0, + 1057.0, + 391.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1595.0, + 296.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1587.0, + 1407.0, + 1587.0, + 1407.0, + 1628.0, + 292.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1407.0, + 1681.0, + 1407.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1714.0, + 1405.0, + 1714.0, + 1405.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 1404.0, + 1744.0, + 1404.0, + 1776.0, + 296.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1837.0, + 293.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1835.0, + 1285.0, + 1835.0, + 1285.0, + 1867.0, + 296.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1407.0, + 1358.0, + 1407.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 1407.0, + 1390.0, + 1407.0, + 1427.0, + 293.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1456.0, + 293.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1485.0, + 293.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1484.0, + 1404.0, + 1484.0, + 1404.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 1113.0, + 1514.0, + 1113.0, + 1546.0, + 294.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1159.0, + 1407.0, + 1159.0, + 1407.0, + 1198.0, + 293.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1191.0, + 1407.0, + 1191.0, + 1407.0, + 1228.0, + 293.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1223.0, + 1405.0, + 1223.0, + 1405.0, + 1258.0, + 293.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1253.0, + 1405.0, + 1253.0, + 1405.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1280.0, + 1408.0, + 1280.0, + 1408.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1314.0, + 1408.0, + 1314.0, + 1408.0, + 1349.0, + 296.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1403.0, + 1909.0, + 1403.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 1405.0, + 1968.0, + 1405.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 378.0, + 549.0, + 378.0, + 549.0, + 410.0, + 313.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 407.0, + 663.0, + 407.0, + 663.0, + 440.0, + 310.0, + 440.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1253, + 1404, + 1253, + 1404, + 1697, + 297, + 1697 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 877, + 1405, + 877, + 1405, + 1154, + 298, + 1154 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1783, + 1405, + 1783, + 1405, + 2028, + 298, + 2028 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1400, + 229, + 1400, + 292, + 296, + 292 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 368, + 318, + 1407, + 318, + 1407, + 850, + 368, + 850 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 298, + 1736, + 1005, + 1736, + 1005, + 1765, + 298, + 1765 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 298, + 1194, + 491, + 1194, + 491, + 1229, + 298, + 1229 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 301, + 76, + 827, + 76, + 827, + 104, + 301, + 104 + ], + "score": 0.75 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.712 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 827, + 76, + 827, + 104, + 300, + 104 + ], + "score": 0.446 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.108 + }, + { + "category_id": 13, + "poly": [ + 457, + 1505, + 658, + 1505, + 658, + 1543, + 457, + 1543 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { R ^ { k } = \\sum _ { i = 0 } ^ { \\infty } \\gamma ^ { i } r _ { t + i } ^ { k } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 514, + 1436, + 708, + 1436, + 708, + 1477, + 514, + 1477 + ], + "score": 0.93, + "latex": "{ \\boldsymbol { \\tau } } = \\left\\{ { { s _ { t } } , { a _ { t } } , { r _ { t } } } \\right\\} _ { t = 0 } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 744, + 1343, + 901, + 1343, + 901, + 1378, + 744, + 1378 + ], + "score": 0.92, + "latex": "\\pmb { a } _ { t } = [ a _ { t } ^ { 0 } , . . . a _ { t } ^ { N } ]" + }, + { + "category_id": 13, + "poly": [ + 1021, + 1634, + 1101, + 1634, + 1101, + 1668, + 1021, + 1668 + ], + "score": 0.92, + "latex": "V ^ { \\pi _ { k } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 1093, + 1313, + 1167, + 1313, + 1167, + 1347, + 1093, + 1347 + ], + "score": 0.92, + "latex": "a _ { t } ^ { k } \\in A" + }, + { + "category_id": 13, + "poly": [ + 856, + 1968, + 929, + 1968, + 929, + 1994, + 856, + 1994 + ], + "score": 0.91, + "latex": "X = x" + }, + { + "category_id": 13, + "poly": [ + 299, + 1812, + 479, + 1812, + 479, + 1845, + 299, + 1845 + ], + "score": 0.9, + "latex": "R ^ { k } = \\alpha E ^ { k } + \\beta I ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1663, + 738, + 1663, + 738, + 1698, + 707, + 1698 + ], + "score": 0.9, + "latex": "\\dot { u } _ { t } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1375, + 451, + 1375, + 451, + 1408, + 299, + 1408 + ], + "score": 0.89, + "latex": "T _ { \\cdot } ( s _ { t + 1 } | \\mathbf { a } _ { t } , s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 774, + 1474, + 802, + 1474, + 802, + 1510, + 774, + 1510 + ], + "score": 0.89, + "latex": "s _ { t } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 559, + 1813, + 595, + 1813, + 595, + 1842, + 559, + 1842 + ], + "score": 0.89, + "latex": "E ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 897, + 1906, + 971, + 1906, + 971, + 1933, + 897, + 1933 + ], + "score": 0.89, + "latex": "Z = z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1406, + 402, + 1406, + 402, + 1439, + 298, + 1439 + ], + "score": 0.89, + "latex": "r ^ { k } ( a _ { t } , s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 476, + 1316, + 534, + 1316, + 534, + 1343, + 476, + 1343 + ], + "score": 0.89, + "latex": "s \\in S" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1813, + 1090, + 1813, + 1090, + 1842, + 1060, + 1842 + ], + "score": 0.88, + "latex": "I ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 749, + 1633, + 782, + 1633, + 782, + 1662, + 749, + 1662 + ], + "score": 0.88, + "latex": "\\pi ^ { \\dot { k } }" + }, + { + "category_id": 13, + "poly": [ + 930, + 1253, + 1044, + 1253, + 1044, + 1287, + 930, + 1287 + ], + "score": 0.87, + "latex": "\\langle S , T , A , r \\rangle" + }, + { + "category_id": 13, + "poly": [ + 1311, + 1541, + 1343, + 1541, + 1343, + 1572, + 1311, + 1572 + ], + "score": 0.86, + "latex": "\\pi ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 434, + 1844, + 464, + 1844, + 464, + 1873, + 434, + 1873 + ], + "score": 0.86, + "latex": "I ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 665, + 1909, + 691, + 1909, + 691, + 1932, + 665, + 1932 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 375, + 1998, + 402, + 1998, + 402, + 2024, + 375, + 2024 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 521, + 1666, + 540, + 1666, + 540, + 1692, + 521, + 1692 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 742, + 1513, + 763, + 1513, + 763, + 1540, + 742, + 1540 + ], + "score": 0.82, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 600, + 1910, + 618, + 1910, + 618, + 1937, + 600, + 1937 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 342, + 1941, + 382, + 1941, + 382, + 1968, + 342, + 1968 + ], + "score": 0.8, + "latex": "x , y" + }, + { + "category_id": 13, + "poly": [ + 876, + 1316, + 894, + 1316, + 894, + 1341, + 876, + 1341 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1154, + 1907, + 1178, + 1907, + 1178, + 1933, + 1154, + 1933 + ], + "score": 0.79, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 297, + 1478, + 317, + 1478, + 317, + 1504, + 297, + 1504 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 662, + 1667, + 676, + 1667, + 676, + 1692, + 662, + 1692 + ], + "score": 0.77, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1316, + 1401, + 1316, + 1401, + 1342, + 1373, + 1342 + ], + "score": 0.77, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 914, + 1377, + 937, + 1377, + 937, + 1403, + 914, + 1403 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 429, + 1943, + 447, + 1943, + 447, + 1964, + 429, + 1964 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 750, + 1911, + 767, + 1911, + 767, + 1933, + 750, + 1933 + ], + "score": 0.74, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 520, + 1973, + 537, + 1973, + 537, + 1994, + 520, + 1994 + ], + "score": 0.72, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 733, + 1318, + 746, + 1318, + 746, + 1342, + 733, + 1342 + ], + "score": 0.72, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1906, + 1103, + 1906, + 1103, + 1936, + 1051, + 1936 + ], + "score": 0.68, + "latex": "X , Y ," + }, + { + "category_id": 13, + "poly": [ + 461, + 1907, + 496, + 1907, + 496, + 1933, + 461, + 1933 + ], + "score": 0.63, + "latex": "^ { 6 6 } Y" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1733.0, + 1008.0, + 1733.0, + 1008.0, + 1769.0, + 294.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1189.0, + 496.0, + 1189.0, + 496.0, + 1236.0, + 290.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 297.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 296.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1252.0, + 929.0, + 1252.0, + 929.0, + 1289.0, + 294.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1252.0, + 1405.0, + 1252.0, + 1405.0, + 1289.0, + 1045.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1283.0, + 1406.0, + 1283.0, + 1406.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1315.0, + 475.0, + 1315.0, + 475.0, + 1346.0, + 293.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1315.0, + 732.0, + 1315.0, + 732.0, + 1346.0, + 535.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1315.0, + 875.0, + 1315.0, + 875.0, + 1346.0, + 747.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1315.0, + 1092.0, + 1315.0, + 1092.0, + 1346.0, + 895.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1315.0, + 1372.0, + 1315.0, + 1372.0, + 1346.0, + 1168.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1315.0, + 1405.0, + 1315.0, + 1405.0, + 1346.0, + 1402.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1339.0, + 743.0, + 1339.0, + 743.0, + 1385.0, + 291.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1339.0, + 1410.0, + 1339.0, + 1410.0, + 1385.0, + 902.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1371.0, + 298.0, + 1371.0, + 298.0, + 1413.0, + 291.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1371.0, + 913.0, + 1371.0, + 913.0, + 1413.0, + 452.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1371.0, + 1407.0, + 1371.0, + 1407.0, + 1413.0, + 938.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1400.0, + 297.0, + 1400.0, + 297.0, + 1443.0, + 291.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1400.0, + 1409.0, + 1400.0, + 1409.0, + 1443.0, + 403.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1423.0, + 513.0, + 1423.0, + 513.0, + 1492.0, + 283.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1423.0, + 1416.0, + 1423.0, + 1416.0, + 1492.0, + 709.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1471.0, + 296.0, + 1471.0, + 296.0, + 1578.0, + 283.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1471.0, + 456.0, + 1471.0, + 456.0, + 1578.0, + 318.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1471.0, + 741.0, + 1471.0, + 741.0, + 1578.0, + 659.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1471.0, + 773.0, + 1471.0, + 773.0, + 1578.0, + 764.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 1471.0, + 1310.0, + 1471.0, + 1310.0, + 1578.0, + 803.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 1471.0, + 1415.0, + 1471.0, + 1415.0, + 1578.0, + 1344.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1572.0, + 1407.0, + 1572.0, + 1407.0, + 1609.0, + 292.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1604.0, + 1404.0, + 1604.0, + 1404.0, + 1639.0, + 294.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1633.0, + 748.0, + 1633.0, + 748.0, + 1671.0, + 292.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1633.0, + 1020.0, + 1633.0, + 1020.0, + 1671.0, + 783.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1633.0, + 1406.0, + 1633.0, + 1406.0, + 1671.0, + 1102.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1655.0, + 520.0, + 1655.0, + 520.0, + 1708.0, + 287.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1655.0, + 661.0, + 1655.0, + 661.0, + 1708.0, + 541.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1655.0, + 706.0, + 1655.0, + 706.0, + 1708.0, + 677.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1655.0, + 755.0, + 1655.0, + 755.0, + 1708.0, + 739.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 915.0, + 295.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 909.0, + 1405.0, + 909.0, + 1405.0, + 942.0, + 297.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 938.0, + 1407.0, + 938.0, + 1407.0, + 975.0, + 295.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 968.0, + 1406.0, + 968.0, + 1406.0, + 1005.0, + 295.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1000.0, + 1405.0, + 1000.0, + 1405.0, + 1036.0, + 295.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1031.0, + 1403.0, + 1031.0, + 1403.0, + 1064.0, + 295.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1057.0, + 1405.0, + 1057.0, + 1405.0, + 1097.0, + 293.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1086.0, + 1410.0, + 1086.0, + 1410.0, + 1131.0, + 292.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1123.0, + 711.0, + 1123.0, + 711.0, + 1157.0, + 293.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1781.0, + 1405.0, + 1781.0, + 1405.0, + 1817.0, + 295.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1807.0, + 558.0, + 1807.0, + 558.0, + 1852.0, + 480.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1807.0, + 1059.0, + 1807.0, + 1059.0, + 1852.0, + 596.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1807.0, + 1409.0, + 1807.0, + 1409.0, + 1852.0, + 1091.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1842.0, + 433.0, + 1842.0, + 433.0, + 1879.0, + 296.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1842.0, + 1406.0, + 1842.0, + 1406.0, + 1879.0, + 465.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1874.0, + 1406.0, + 1874.0, + 1406.0, + 1909.0, + 294.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1904.0, + 460.0, + 1904.0, + 460.0, + 1940.0, + 293.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1904.0, + 599.0, + 1904.0, + 599.0, + 1940.0, + 497.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1904.0, + 664.0, + 1904.0, + 664.0, + 1940.0, + 619.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 1904.0, + 749.0, + 1904.0, + 749.0, + 1940.0, + 692.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1904.0, + 896.0, + 1904.0, + 896.0, + 1940.0, + 768.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1904.0, + 1050.0, + 1904.0, + 1050.0, + 1940.0, + 972.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1904.0, + 1153.0, + 1904.0, + 1153.0, + 1940.0, + 1104.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1904.0, + 1408.0, + 1904.0, + 1408.0, + 1940.0, + 1179.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1937.0, + 341.0, + 1937.0, + 341.0, + 1970.0, + 295.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1937.0, + 428.0, + 1937.0, + 428.0, + 1970.0, + 383.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1937.0, + 1403.0, + 1937.0, + 1403.0, + 1970.0, + 448.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1967.0, + 519.0, + 1967.0, + 519.0, + 2000.0, + 295.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1967.0, + 855.0, + 1967.0, + 855.0, + 2000.0, + 538.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1967.0, + 1403.0, + 1967.0, + 1403.0, + 2000.0, + 930.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1997.0, + 374.0, + 1997.0, + 374.0, + 2030.0, + 296.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1997.0, + 684.0, + 1997.0, + 684.0, + 2030.0, + 403.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1052.0, + 262.0, + 1052.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 355.0, + 368.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 349.0, + 1407.0, + 349.0, + 1407.0, + 385.0, + 394.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 382.0, + 1406.0, + 382.0, + 1406.0, + 417.0, + 395.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 413.0, + 1277.0, + 413.0, + 1277.0, + 445.0, + 394.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 491.0, + 389.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 488.0, + 1404.0, + 488.0, + 1404.0, + 520.0, + 393.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 517.0, + 1408.0, + 517.0, + 1408.0, + 553.0, + 394.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 549.0, + 1407.0, + 549.0, + 1407.0, + 582.0, + 395.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 592.0, + 1402.0, + 592.0, + 1402.0, + 625.0, + 392.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 625.0, + 1404.0, + 625.0, + 1404.0, + 657.0, + 395.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 654.0, + 992.0, + 654.0, + 992.0, + 686.0, + 395.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 698.0, + 1405.0, + 698.0, + 1405.0, + 730.0, + 382.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 727.0, + 1407.0, + 727.0, + 1407.0, + 763.0, + 394.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 757.0, + 1405.0, + 757.0, + 1405.0, + 794.0, + 393.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 791.0, + 1404.0, + 791.0, + 1404.0, + 820.0, + 395.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 819.0, + 1400.0, + 819.0, + 1400.0, + 852.0, + 393.0, + 852.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 544, + 229, + 1404, + 229, + 1404, + 626, + 544, + 626 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1308, + 1405, + 1308, + 1405, + 1493, + 298, + 1493 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 544, + 639, + 1405, + 639, + 1405, + 893, + 544, + 893 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1721, + 1404, + 1721, + 1404, + 1817, + 298, + 1817 + ], + "score": 0.967 + }, + { + "category_id": 2, + "poly": [ + 295, + 1837, + 1404, + 1837, + 1404, + 2034, + 295, + 2034 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 372, + 965, + 1322, + 965, + 1322, + 1048, + 372, + 1048 + ], + "score": 0.947 + }, + { + "category_id": 3, + "poly": [ + 300, + 230, + 519, + 230, + 519, + 533, + 300, + 533 + ], + "score": 0.946 + }, + { + "category_id": 4, + "poly": [ + 296, + 553, + 522, + 553, + 522, + 860, + 296, + 860 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 301, + 1128, + 1403, + 1128, + 1403, + 1192, + 301, + 1192 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 320, + 1205, + 1339, + 1205, + 1339, + 1292, + 320, + 1292 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 293, + 893, + 1401, + 893, + 1401, + 954, + 293, + 954 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 300, + 1675, + 1171, + 1675, + 1171, + 1709, + 300, + 1709 + ], + "score": 0.923 + }, + { + "category_id": 8, + "poly": [ + 360, + 1585, + 1332, + 1585, + 1332, + 1660, + 360, + 1660 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 294, + 1507, + 1403, + 1507, + 1403, + 1572, + 294, + 1572 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1368, + 1603, + 1401, + 1603, + 1401, + 1634, + 1368, + 1634 + ], + "score": 0.823 + }, + { + "category_id": 9, + "poly": [ + 1368, + 1230, + 1400, + 1230, + 1400, + 1260, + 1368, + 1260 + ], + "score": 0.813 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 827, + 76, + 827, + 104, + 300, + 104 + ], + "score": 0.783 + }, + { + "category_id": 9, + "poly": [ + 1368, + 982, + 1401, + 982, + 1401, + 1012, + 1368, + 1012 + ], + "score": 0.721 + }, + { + "category_id": 0, + "poly": [ + 297, + 1082, + 1097, + 1082, + 1097, + 1115, + 297, + 1115 + ], + "score": 0.695 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.656 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.426 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 827, + 76, + 827, + 104, + 300, + 104 + ], + "score": 0.345 + }, + { + "category_id": 1, + "poly": [ + 297, + 1082, + 1097, + 1082, + 1097, + 1115, + 297, + 1115 + ], + "score": 0.25 + }, + { + "category_id": 4, + "poly": [ + 297, + 553, + 522, + 553, + 522, + 861, + 297, + 861 + ], + "score": 0.152 + }, + { + "category_id": 14, + "poly": [ + 360, + 1584, + 1338, + 1584, + 1338, + 1664, + 360, + 1664 + ], + "score": 0.94, + "latex": "I ( A ^ { A } ; A ^ { B } | z ) = \\mathbb { E } _ { \\tau } \\Big [ D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | A ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] \\Big | z \\Big ] \\approx \\frac { 1 } { N } \\sum _ { n } D _ { \\mathrm { K L } } \\big [ p ( A ^ { B } | a _ { n } ^ { A } , z ) \\big \\| p ( A ^ { B } | z ) \\big ] ." + }, + { + "category_id": 13, + "poly": [ + 1000, + 533, + 1232, + 533, + 1232, + 568, + 1000, + 568 + ], + "score": 0.94, + "latex": "p ( a _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , s _ { t } ^ { B } , u _ { t } ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 806, + 640, + 948, + 640, + 948, + 676, + 806, + 676 + ], + "score": 0.93, + "latex": "\\boldsymbol { z } _ { t } = \\langle \\boldsymbol { u } _ { t } ^ { B } , \\boldsymbol { s } _ { t } ^ { B } \\rangle" + }, + { + "category_id": 14, + "poly": [ + 322, + 1203, + 1345, + 1203, + 1345, + 1295, + 322, + 1295 + ], + "score": 0.93, + "latex": "I ( A ^ { B } ; A ^ { A } | z ) = \\sum _ { a ^ { A } , a ^ { B } } p ( a ^ { B } , a ^ { A } | z ) \\mathrm { l o g } \\frac { p ( a ^ { B } , a ^ { A } | z ) } { p ( a ^ { B } | z ) p ( a ^ { A } | z ) } { = \\sum _ { a ^ { A } } p ( a ^ { A } | z ) D _ { \\mathrm { K L } } \\left[ p ( a ^ { B } | a ^ { A } , z ) \\right] \\left| p ( a ^ { B } | z ) \\right] } ," + }, + { + "category_id": 13, + "poly": [ + 434, + 1538, + 521, + 1538, + 521, + 1572, + 434, + 1572 + ], + "score": 0.92, + "latex": "\\overset { \\cdot } { p } ( a ^ { A } | z )" + }, + { + "category_id": 13, + "poly": [ + 1032, + 701, + 1373, + 701, + 1373, + 737, + 1032, + 737 + ], + "score": 0.92, + "latex": "p ( \\dot { a } _ { t } ^ { B } | d o ( \\tilde { a } _ { t } ^ { A } ) , z _ { t } ) \\equiv p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { \\tilde { A } } , z _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 374, + 965, + 1325, + 965, + 1325, + 1051, + 374, + 1051 + ], + "score": 0.92, + "latex": "{ I _ { t } ^ { A } } = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } \\mid z _ { t } , \\tilde { a } _ { t } ^ { A } ) p ( \\tilde { a } _ { t } ^ { A } \\mid z _ { t } ) \\right] = D _ { K L } \\left[ p ( a _ { t } ^ { B } \\mid a _ { t } ^ { A } , z _ { t } ) \\right| \\left| p ( a _ { t } ^ { B } \\mid z _ { t } ) \\right] ." + }, + { + "category_id": 13, + "poly": [ + 628, + 319, + 819, + 319, + 819, + 355, + 628, + 355 + ], + "score": 0.92, + "latex": "p \\big ( a _ { t } ^ { B } | a _ { t } ^ { A } , \\overline { { s _ { t } ^ { B } } } , u _ { t } ^ { B } \\big )" + }, + { + "category_id": 13, + "poly": [ + 1235, + 442, + 1271, + 442, + 1271, + 477, + 1235, + 477 + ], + "score": 0.91, + "latex": "u _ { t } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 531, + 1310, + 591, + 1310, + 591, + 1340, + 531, + 1340 + ], + "score": 0.91, + "latex": "D _ { K L }" + }, + { + "category_id": 13, + "poly": [ + 1225, + 351, + 1262, + 351, + 1262, + 385, + 1225, + 385 + ], + "score": 0.91, + "latex": "u _ { t } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 1084, + 2006, + 1114, + 2006, + 1114, + 2034, + 1084, + 2034 + ], + "score": 0.91, + "latex": "a _ { t } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1507, + 1281, + 1507, + 1281, + 1542, + 1247, + 1542 + ], + "score": 0.9, + "latex": "a _ { n } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 580, + 503, + 614, + 503, + 614, + 537, + 580, + 537 + ], + "score": 0.9, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1922, + 422, + 1922, + 422, + 1952, + 298, + 1952 + ], + "score": 0.9, + "latex": "( \\in [ 1 , N - 1 ] )" + }, + { + "category_id": 13, + "poly": [ + 373, + 2003, + 406, + 2003, + 406, + 2034, + 373, + 2034 + ], + "score": 0.9, + "latex": "a _ { t } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 1095, + 504, + 1174, + 504, + 1174, + 534, + 1095, + 534 + ], + "score": 0.89, + "latex": "d o \\dot { ( a _ { t } ^ { A } ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 2003, + 330, + 2003, + 330, + 2034, + 298, + 2034 + ], + "score": 0.89, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 486, + 795, + 522, + 795, + 522, + 830, + 486, + 830 + ], + "score": 0.89, + "latex": "a _ { t } ^ { \\dot { A } }" + }, + { + "category_id": 13, + "poly": [ + 683, + 793, + 1091, + 793, + 1091, + 835, + 683, + 835 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { p ( a _ { t } ^ { B } | \\tilde { z _ { t } } ) = \\sum _ { \\tilde { a } _ { t } ^ { A } } p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , z _ { t } ) p ( \\tilde { a } _ { t } ^ { A } | z _ { t } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 545, + 289, + 579, + 289, + 579, + 325, + 545, + 325 + ], + "score": 0.89, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 507, + 1462, + 552, + 1462, + 552, + 1489, + 507, + 1489 + ], + "score": 0.88, + "latex": "B ^ { \\prime } s" + }, + { + "category_id": 13, + "poly": [ + 1089, + 764, + 1123, + 764, + 1123, + 797, + 1089, + 797 + ], + "score": 0.87, + "latex": "\\tilde { a } _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 749, + 1514, + 778, + 1514, + 778, + 1539, + 749, + 1539 + ], + "score": 0.85, + "latex": "\\tau _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1208, + 648, + 1234, + 648, + 1234, + 674, + 1208, + 674 + ], + "score": 0.85, + "latex": "z _ { t }" + }, + { + "category_id": 13, + "poly": [ + 803, + 863, + 828, + 863, + 828, + 889, + 803, + 889 + ], + "score": 0.84, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 369, + 1161, + 393, + 1161, + 393, + 1186, + 369, + 1186 + ], + "score": 0.84, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1308, + 1843, + 1330, + 1843, + 1330, + 1863, + 1308, + 1863 + ], + "score": 0.84, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1261, + 894, + 1283, + 894, + 1283, + 918, + 1261, + 918 + ], + "score": 0.83, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 298, + 1162, + 321, + 1162, + 321, + 1186, + 298, + 1186 + ], + "score": 0.82, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 920, + 1842, + 942, + 1842, + 942, + 1863, + 920, + 1863 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 788, + 293, + 813, + 293, + 813, + 318, + 788, + 318 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1245, + 1842, + 1267, + 1842, + 1267, + 1863, + 1245, + 1863 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1326, + 323, + 1351, + 323, + 1351, + 349, + 1326, + 349 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1286, + 799, + 1311, + 799, + 1311, + 823, + 1286, + 823 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1343, + 384, + 1368, + 384, + 1368, + 410, + 1343, + 410 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1331, + 834, + 1355, + 834, + 1355, + 858, + 1331, + 858 + ], + "score": 0.81, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1064, + 569, + 1089, + 569, + 1089, + 593, + 1064, + 593 + ], + "score": 0.81, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 846, + 924, + 869, + 924, + 869, + 950, + 846, + 950 + ], + "score": 0.81, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 695, + 893, + 720, + 893, + 720, + 919, + 695, + 919 + ], + "score": 0.8, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1196, + 385, + 1219, + 385, + 1219, + 410, + 1196, + 410 + ], + "score": 0.8, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 893, + 863, + 916, + 863, + 916, + 889, + 893, + 889 + ], + "score": 0.8, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 895, + 1786, + 919, + 1786, + 919, + 1812, + 895, + 1812 + ], + "score": 0.8, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 358, + 893, + 383, + 893, + 383, + 919, + 358, + 919 + ], + "score": 0.8, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1433, + 1158, + 1433, + 1158, + 1458, + 1135, + 1458 + ], + "score": 0.8, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 969, + 1786, + 992, + 1786, + 992, + 1812, + 969, + 1812 + ], + "score": 0.79, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 448, + 1511, + 475, + 1511, + 475, + 1535, + 448, + 1535 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 781, + 263, + 804, + 263, + 804, + 289, + 781, + 289 + ], + "score": 0.79, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 1056, + 262, + 1082, + 262, + 1082, + 289, + 1056, + 289 + ], + "score": 0.78, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 857, + 262, + 881, + 262, + 881, + 289, + 857, + 289 + ], + "score": 0.77, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1085, + 1841, + 1106, + 1841, + 1106, + 1863, + 1085, + 1863 + ], + "score": 0.76, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 298, + 893, + 322, + 893, + 322, + 919, + 298, + 919 + ], + "score": 0.75, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 393, + 1868, + 414, + 1868, + 414, + 1892, + 393, + 1892 + ], + "score": 0.75, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 740, + 740, + 759, + 740, + 759, + 762, + 740, + 762 + ], + "score": 0.75, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 568, + 831, + 591, + 831, + 591, + 858, + 568, + 858 + ], + "score": 0.74, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 662, + 1842, + 683, + 1842, + 683, + 1864, + 662, + 1864 + ], + "score": 0.72, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 1384, + 265, + 1398, + 265, + 1398, + 289, + 1384, + 289 + ], + "score": 0.7, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1022, + 1985, + 1040, + 1985, + 1040, + 2002, + 1022, + 2002 + ], + "score": 0.68, + "latex": "\\mathrm { c }" + }, + { + "category_id": 13, + "poly": [ + 1181, + 263, + 1205, + 263, + 1205, + 289, + 1181, + 289 + ], + "score": 0.67, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 647, + 796, + 672, + 796, + 672, + 824, + 647, + 824 + ], + "score": 0.67, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 783, + 353, + 817, + 353, + 817, + 386, + 783, + 386 + ], + "score": 0.65, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 297, + 771, + 318, + 771, + 318, + 795, + 297, + 795 + ], + "score": 0.65, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 477, + 616, + 503, + 616, + 503, + 643, + 477, + 643 + ], + "score": 0.62, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 785, + 705, + 816, + 705, + 816, + 732, + 785, + 732 + ], + "score": 0.59, + "latex": "d o" + }, + { + "category_id": 13, + "poly": [ + 828, + 351, + 860, + 351, + 860, + 386, + 828, + 386 + ], + "score": 0.59, + "latex": "s _ { t } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 298, + 615, + 321, + 615, + 321, + 643, + 298, + 643 + ], + "score": 0.51, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 782, + 351, + 861, + 351, + 861, + 386, + 782, + 386 + ], + "score": 0.43, + "latex": "\\bar { a } _ { t } ^ { A } , s _ { t } ^ { B }" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1835.0, + 661.0, + 1835.0, + 661.0, + 1873.0, + 332.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1835.0, + 919.0, + 1835.0, + 919.0, + 1873.0, + 684.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1835.0, + 1084.0, + 1835.0, + 1084.0, + 1873.0, + 943.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1835.0, + 1244.0, + 1835.0, + 1244.0, + 1873.0, + 1107.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1835.0, + 1307.0, + 1835.0, + 1307.0, + 1873.0, + 1268.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1835.0, + 1407.0, + 1835.0, + 1407.0, + 1873.0, + 1331.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1867.0, + 392.0, + 1867.0, + 392.0, + 1900.0, + 295.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1867.0, + 1405.0, + 1867.0, + 1405.0, + 1900.0, + 415.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1404.0, + 1893.0, + 1404.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1920.0, + 297.0, + 1920.0, + 297.0, + 1957.0, + 294.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1920.0, + 1406.0, + 1920.0, + 1406.0, + 1957.0, + 423.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1947.0, + 1406.0, + 1947.0, + 1406.0, + 1982.0, + 293.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1977.0, + 1021.0, + 1977.0, + 1021.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1977.0, + 1405.0, + 1977.0, + 1405.0, + 2010.0, + 1041.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2002.0, + 372.0, + 2002.0, + 372.0, + 2039.0, + 331.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 2002.0, + 1083.0, + 2002.0, + 1083.0, + 2039.0, + 407.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 2002.0, + 1342.0, + 2002.0, + 1342.0, + 2039.0, + 1115.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 249.0, + 410.0, + 249.0, + 410.0, + 289.0, + 369.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 388.0, + 413.0, + 388.0, + 413.0, + 456.0, + 349.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.25, + 315.5, + 352.25, + 315.5, + 352.25, + 375.5, + 322.25, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 310.5, + 460.0, + 310.5, + 460.0, + 375.5, + 430.0, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.25, + 385.5, + 501.25, + 385.5, + 501.25, + 451.0, + 466.25, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 455.5, + 460.0, + 455.5, + 460.0, + 525.5, + 424.0, + 525.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.75, + 471.5, + 390.75, + 471.5, + 390.75, + 530.5, + 292.75, + 530.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 551.0, + 525.0, + 551.0, + 525.0, + 586.0, + 293.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 583.0, + 526.0, + 583.0, + 526.0, + 617.0, + 296.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 612.0, + 476.0, + 612.0, + 476.0, + 646.0, + 322.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 612.0, + 526.0, + 612.0, + 526.0, + 646.0, + 504.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 526.0, + 643.0, + 526.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 676.0, + 525.0, + 676.0, + 525.0, + 707.0, + 295.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 526.0, + 703.0, + 526.0, + 737.0, + 294.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 735.0, + 527.0, + 735.0, + 527.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 764.0, + 296.0, + 764.0, + 296.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 764.0, + 528.0, + 764.0, + 528.0, + 801.0, + 319.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 791.0, + 485.0, + 791.0, + 485.0, + 835.0, + 291.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 824.0, + 376.0, + 824.0, + 376.0, + 864.0, + 292.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 296.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1081.0, + 1101.0, + 1081.0, + 1101.0, + 1119.0, + 293.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 296.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 525.0, + 551.0, + 525.0, + 586.0, + 294.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 581.0, + 526.0, + 581.0, + 526.0, + 618.0, + 294.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 612.0, + 476.0, + 612.0, + 476.0, + 646.0, + 322.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 612.0, + 525.0, + 612.0, + 525.0, + 646.0, + 504.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 644.0, + 525.0, + 644.0, + 525.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 676.0, + 525.0, + 676.0, + 525.0, + 707.0, + 295.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 526.0, + 703.0, + 526.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 734.0, + 527.0, + 734.0, + 527.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 764.0, + 296.0, + 764.0, + 296.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 764.0, + 528.0, + 764.0, + 528.0, + 801.0, + 319.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 791.0, + 485.0, + 791.0, + 485.0, + 836.0, + 288.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 791.0, + 528.0, + 791.0, + 528.0, + 836.0, + 523.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 824.0, + 376.0, + 824.0, + 376.0, + 864.0, + 293.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 266.0, + 541.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 262.0, + 780.0, + 262.0, + 780.0, + 295.0, + 543.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 262.0, + 856.0, + 262.0, + 856.0, + 295.0, + 805.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 262.0, + 1055.0, + 262.0, + 1055.0, + 295.0, + 882.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 262.0, + 1180.0, + 262.0, + 1180.0, + 295.0, + 1083.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 262.0, + 1383.0, + 262.0, + 1383.0, + 295.0, + 1206.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 262.0, + 1409.0, + 262.0, + 1409.0, + 295.0, + 1399.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 281.0, + 544.0, + 281.0, + 544.0, + 329.0, + 535.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 281.0, + 787.0, + 281.0, + 787.0, + 329.0, + 580.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 281.0, + 1413.0, + 281.0, + 1413.0, + 329.0, + 814.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 302.0, + 627.0, + 302.0, + 627.0, + 399.0, + 531.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 302.0, + 1224.0, + 302.0, + 1224.0, + 399.0, + 862.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 302.0, + 1325.0, + 302.0, + 1325.0, + 399.0, + 1263.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 302.0, + 1414.0, + 302.0, + 1414.0, + 399.0, + 1352.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 381.0, + 1195.0, + 381.0, + 1195.0, + 417.0, + 541.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 381.0, + 1342.0, + 381.0, + 1342.0, + 417.0, + 1220.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 417.0, + 1369.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 447.0, + 543.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 438.0, + 1234.0, + 438.0, + 1234.0, + 481.0, + 540.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 438.0, + 1408.0, + 438.0, + 1408.0, + 481.0, + 1272.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 474.0, + 1405.0, + 474.0, + 1405.0, + 507.0, + 541.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 502.0, + 579.0, + 502.0, + 579.0, + 537.0, + 541.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 502.0, + 1094.0, + 502.0, + 1094.0, + 537.0, + 615.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 537.0, + 1175.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 527.0, + 999.0, + 527.0, + 999.0, + 575.0, + 538.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 527.0, + 1409.0, + 527.0, + 1409.0, + 575.0, + 1233.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 564.0, + 1063.0, + 564.0, + 1063.0, + 599.0, + 541.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 564.0, + 1405.0, + 564.0, + 1405.0, + 599.0, + 1090.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 594.0, + 855.0, + 594.0, + 855.0, + 628.0, + 544.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 319.5, + 820.0, + 319.5, + 820.0, + 343.5, + 787.0, + 343.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1310.0, + 530.0, + 1310.0, + 530.0, + 1342.0, + 296.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1310.0, + 1403.0, + 1310.0, + 1403.0, + 1342.0, + 592.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1341.0, + 1405.0, + 1341.0, + 1405.0, + 1373.0, + 295.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1374.0, + 1403.0, + 1374.0, + 1403.0, + 1402.0, + 296.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1398.0, + 1406.0, + 1398.0, + 1406.0, + 1435.0, + 293.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1425.0, + 1134.0, + 1425.0, + 1134.0, + 1469.0, + 292.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1425.0, + 1407.0, + 1425.0, + 1407.0, + 1469.0, + 1159.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1462.0, + 506.0, + 1462.0, + 506.0, + 1496.0, + 294.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1462.0, + 641.0, + 1462.0, + 641.0, + 1496.0, + 553.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 633.0, + 805.0, + 633.0, + 805.0, + 684.0, + 538.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 633.0, + 1207.0, + 633.0, + 1207.0, + 684.0, + 949.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 633.0, + 1410.0, + 633.0, + 1410.0, + 684.0, + 1235.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 671.0, + 1408.0, + 671.0, + 1408.0, + 708.0, + 542.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 700.0, + 784.0, + 700.0, + 784.0, + 741.0, + 540.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 700.0, + 1031.0, + 700.0, + 1031.0, + 741.0, + 817.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 700.0, + 1409.0, + 700.0, + 1409.0, + 741.0, + 1374.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 732.0, + 739.0, + 732.0, + 739.0, + 768.0, + 541.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 732.0, + 1408.0, + 732.0, + 1408.0, + 768.0, + 760.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 761.0, + 1088.0, + 761.0, + 1088.0, + 799.0, + 541.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 761.0, + 1407.0, + 761.0, + 1407.0, + 799.0, + 1124.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 792.0, + 646.0, + 792.0, + 646.0, + 833.0, + 541.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 792.0, + 682.0, + 792.0, + 682.0, + 833.0, + 673.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 792.0, + 1285.0, + 792.0, + 1285.0, + 833.0, + 1092.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 792.0, + 1406.0, + 792.0, + 1406.0, + 833.0, + 1312.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 827.0, + 567.0, + 827.0, + 567.0, + 864.0, + 541.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 827.0, + 1330.0, + 827.0, + 1330.0, + 864.0, + 592.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 827.0, + 1405.0, + 827.0, + 1405.0, + 864.0, + 1356.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 861.0, + 802.0, + 861.0, + 802.0, + 895.0, + 543.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 861.0, + 892.0, + 861.0, + 892.0, + 895.0, + 829.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 861.0, + 1406.0, + 861.0, + 1406.0, + 895.0, + 917.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1720.0, + 1405.0, + 1720.0, + 1405.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1754.0, + 1405.0, + 1754.0, + 1405.0, + 1787.0, + 294.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1784.0, + 894.0, + 1784.0, + 894.0, + 1818.0, + 293.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1784.0, + 968.0, + 1784.0, + 968.0, + 1818.0, + 920.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1784.0, + 1407.0, + 1784.0, + 1407.0, + 1818.0, + 993.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1125.0, + 1403.0, + 1125.0, + 1403.0, + 1164.0, + 296.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1155.0, + 368.0, + 1155.0, + 368.0, + 1196.0, + 322.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1155.0, + 594.0, + 1155.0, + 594.0, + 1196.0, + 394.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 890.0, + 357.0, + 890.0, + 357.0, + 925.0, + 323.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 890.0, + 694.0, + 890.0, + 694.0, + 925.0, + 384.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 890.0, + 1260.0, + 890.0, + 1260.0, + 925.0, + 721.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 890.0, + 1407.0, + 890.0, + 1407.0, + 925.0, + 1284.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 921.0, + 845.0, + 921.0, + 845.0, + 956.0, + 295.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 921.0, + 899.0, + 921.0, + 899.0, + 956.0, + 870.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1673.0, + 1172.0, + 1673.0, + 1172.0, + 1713.0, + 296.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1503.0, + 447.0, + 1503.0, + 447.0, + 1547.0, + 293.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1503.0, + 748.0, + 1503.0, + 748.0, + 1547.0, + 476.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1503.0, + 1246.0, + 1503.0, + 1246.0, + 1547.0, + 779.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1503.0, + 1406.0, + 1503.0, + 1406.0, + 1547.0, + 1282.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1534.0, + 433.0, + 1534.0, + 433.0, + 1574.0, + 294.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1534.0, + 1406.0, + 1534.0, + 1406.0, + 1574.0, + 522.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1081.0, + 1101.0, + 1081.0, + 1101.0, + 1119.0, + 293.0, + 1119.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 427, + 1404, + 427, + 1404, + 824, + 297, + 824 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 838, + 1405, + 838, + 1405, + 1116, + 297, + 1116 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1680, + 1404, + 1680, + 1404, + 1958, + 298, + 1958 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1402, + 228, + 1402, + 352, + 299, + 352 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 838, + 1460, + 1398, + 1460, + 1398, + 1613, + 838, + 1613 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 295, + 1460, + 800, + 1460, + 800, + 1613, + 295, + 1613 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 848, + 1146, + 1391, + 1146, + 1391, + 1430, + 848, + 1430 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 299, + 1150, + 799, + 1150, + 799, + 1434, + 299, + 1434 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 294, + 1972, + 1401, + 1972, + 1401, + 2034, + 294, + 2034 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 827, + 76, + 827, + 104, + 300, + 104 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 303, + 1639, + 854, + 1639, + 854, + 1668, + 303, + 1668 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 301, + 384, + 819, + 384, + 819, + 414, + 301, + 414 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2111, + 841, + 2111 + ], + "score": 0.78 + }, + { + "category_id": 13, + "poly": [ + 361, + 1924, + 518, + 1924, + 518, + 1959, + 361, + 1959 + ], + "score": 0.93, + "latex": "p ( { \\pmb a } _ { t + 1 } | { \\pmb a } _ { t } , { \\pmb s } _ { t } ^ { A } )" + }, + { + "category_id": 13, + "poly": [ + 712, + 1710, + 856, + 1710, + 856, + 1746, + 712, + 1746 + ], + "score": 0.93, + "latex": "p ( a _ { t } ^ { B } | \\tilde { a } _ { t } ^ { A } , s _ { t } ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 368, + 669, + 593, + 669, + 593, + 704, + 368, + 704 + ], + "score": 0.93, + "latex": "\\pmb { m } _ { t } = [ m _ { t } ^ { 0 } , m _ { t } ^ { 1 } . . . m _ { t } ^ { N } ] ," + }, + { + "category_id": 13, + "poly": [ + 477, + 1051, + 522, + 1051, + 522, + 1085, + 477, + 1085 + ], + "score": 0.92, + "latex": "m _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 802, + 899, + 855, + 899, + 855, + 936, + 802, + 936 + ], + "score": 0.92, + "latex": "a _ { t + 1 } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 1309, + 870, + 1352, + 870, + 1352, + 904, + 1309, + 904 + ], + "score": 0.91, + "latex": "m _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 747, + 639, + 786, + 639, + 786, + 674, + 747, + 674 + ], + "score": 0.9, + "latex": "\\Dot { m } _ { t } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 1335, + 1527, + 1391, + 1527, + 1391, + 1554, + 1335, + 1554 + ], + "score": 0.88, + "latex": "\\mathbf { \\pmb { a } } _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 655, + 795, + 684, + 795, + 684, + 824, + 655, + 824 + ], + "score": 0.88, + "latex": "V _ { e }" + }, + { + "category_id": 13, + "poly": [ + 970, + 676, + 1010, + 676, + 1010, + 702, + 970, + 702 + ], + "score": 0.86, + "latex": "\\mathbf { \\nabla } m _ { t }" + }, + { + "category_id": 13, + "poly": [ + 577, + 797, + 607, + 797, + 607, + 824, + 577, + 824 + ], + "score": 0.86, + "latex": "\\pi _ { e }" + }, + { + "category_id": 13, + "poly": [ + 648, + 763, + 678, + 763, + 678, + 793, + 648, + 793 + ], + "score": 0.86, + "latex": "V _ { c }" + }, + { + "category_id": 13, + "poly": [ + 755, + 1558, + 795, + 1558, + 795, + 1583, + 755, + 1583 + ], + "score": 0.85, + "latex": "\\mathbf { \\nabla } m _ { t }" + }, + { + "category_id": 13, + "poly": [ + 716, + 1497, + 745, + 1497, + 745, + 1522, + 716, + 1522 + ], + "score": 0.85, + "latex": "\\pi _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1023, + 737, + 1054, + 737, + 1054, + 763, + 1023, + 763 + ], + "score": 0.84, + "latex": "\\pi _ { c }" + }, + { + "category_id": 13, + "poly": [ + 718, + 228, + 752, + 228, + 752, + 264, + 718, + 264 + ], + "score": 0.84, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 1178, + 1496, + 1208, + 1496, + 1208, + 1522, + 1178, + 1522 + ], + "score": 0.84, + "latex": "\\pi _ { e }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1557, + 326, + 1557, + 326, + 1583, + 297, + 1583 + ], + "score": 0.84, + "latex": "\\pi _ { c }" + }, + { + "category_id": 13, + "poly": [ + 639, + 672, + 667, + 672, + 667, + 699, + 639, + 699 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1244, + 933, + 1270, + 933, + 1270, + 959, + 1244, + 959 + ], + "score": 0.83, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 789, + 1085, + 815, + 1085, + 815, + 1111, + 789, + 1111 + ], + "score": 0.82, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1370, + 846, + 1398, + 846, + 1398, + 871, + 1370, + 871 + ], + "score": 0.81, + "latex": "\\pi _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1195, + 963, + 1221, + 963, + 1221, + 989, + 1195, + 989 + ], + "score": 0.79, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1055, + 1060, + 1055, + 1060, + 1081, + 1035, + 1081 + ], + "score": 0.79, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1383, + 612, + 1401, + 612, + 1401, + 638, + 1383, + 638 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1225, + 794, + 1249, + 794, + 1249, + 820, + 1225, + 820 + ], + "score": 0.77, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 298, + 1714, + 323, + 1714, + 323, + 1740, + 298, + 1740 + ], + "score": 0.71, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1154, + 1024, + 1178, + 1024, + 1178, + 1050, + 1154, + 1050 + ], + "score": 0.71, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 777, + 263, + 801, + 263, + 801, + 289, + 777, + 289 + ], + "score": 0.69, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 399, + 963, + 422, + 963, + 422, + 989, + 399, + 989 + ], + "score": 0.66, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 685, + 231, + 709, + 231, + 709, + 259, + 685, + 259 + ], + "score": 0.61, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 897, + 1338, + 923, + 1338, + 923, + 1366, + 897, + 1366 + ], + "score": 0.6, + "latex": "s _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 352, + 1363, + 381, + 1363, + 381, + 1394, + 352, + 1394 + ], + "score": 0.58, + "latex": "s _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 1256, + 1325, + 1388, + 1325, + 1388, + 1355, + 1256, + 1355 + ], + "score": 0.33, + "latex": "p ( \\pmb { a } _ { t + 1 } | \\pmb { a } _ { t } , \\pmb { s } _ { t } ^ { A } )" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1459.0, + 1399.0, + 1459.0, + 1399.0, + 1493.0, + 839.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1492.0, + 1177.0, + 1492.0, + 1177.0, + 1526.0, + 838.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1492.0, + 1398.0, + 1492.0, + 1398.0, + 1526.0, + 1209.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1517.0, + 1334.0, + 1517.0, + 1334.0, + 1559.0, + 836.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1392.0, + 1517.0, + 1403.0, + 1517.0, + 1403.0, + 1559.0, + 1392.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1553.0, + 1399.0, + 1553.0, + 1399.0, + 1584.0, + 839.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1584.0, + 1186.0, + 1584.0, + 1186.0, + 1614.0, + 839.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1459.0, + 799.0, + 1459.0, + 799.0, + 1493.0, + 296.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1489.0, + 715.0, + 1489.0, + 715.0, + 1526.0, + 294.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1489.0, + 800.0, + 1489.0, + 800.0, + 1526.0, + 746.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1522.0, + 802.0, + 1522.0, + 802.0, + 1555.0, + 294.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1549.0, + 754.0, + 1549.0, + 754.0, + 1590.0, + 327.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1549.0, + 800.0, + 1549.0, + 800.0, + 1590.0, + 796.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1584.0, + 546.0, + 1584.0, + 546.0, + 1615.0, + 295.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1187.0, + 1187.0, + 1187.0, + 1187.0, + 1241.0, + 1159.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1198.0, + 1081.0, + 1198.0, + 1081.0, + 1227.0, + 1059.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1202.0, + 1114.0, + 1202.0, + 1114.0, + 1215.0, + 1100.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1211.0, + 1115.0, + 1211.0, + 1115.0, + 1223.0, + 1100.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1231.0, + 1034.0, + 1231.0, + 1034.0, + 1309.0, + 1007.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1327.0, + 931.0, + 1327.0, + 931.0, + 1373.0, + 924.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1329.0, + 1079.0, + 1329.0, + 1079.0, + 1355.0, + 1060.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1332.0, + 1116.0, + 1332.0, + 1116.0, + 1347.0, + 1099.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1315.0, + 1188.0, + 1315.0, + 1188.0, + 1372.0, + 1157.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1318.0, + 1255.0, + 1318.0, + 1255.0, + 1362.0, + 1249.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1389.0, + 1318.0, + 1395.0, + 1318.0, + 1395.0, + 1362.0, + 1389.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1209.0, + 1284.0, + 1209.0, + 1284.0, + 1235.0, + 1254.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1404.0, + 1147.0, + 1404.0, + 1147.0, + 1431.0, + 1120.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1216.0, + 801.0, + 1216.0, + 801.0, + 1248.0, + 764.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1262.0, + 681.0, + 1262.0, + 681.0, + 1326.0, + 648.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1264.0, + 503.0, + 1264.0, + 503.0, + 1320.0, + 475.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 1266.0, + 598.0, + 1266.0, + 598.0, + 1317.0, + 544.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1274.0, + 548.0, + 1274.0, + 548.0, + 1311.0, + 518.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.25, + 1318.0, + 788.25, + 1318.0, + 788.25, + 1334.0, + 783.25, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1357.5, + 376.0, + 1357.5, + 376.0, + 1394.5, + 355.0, + 1394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.75, + 1375.5, + 796.75, + 1375.5, + 796.75, + 1401.5, + 767.75, + 1401.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.75, + 1396.5, + 620.75, + 1396.5, + 620.75, + 1425.5, + 579.75, + 1425.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 296.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1636.0, + 859.0, + 1636.0, + 859.0, + 1671.0, + 296.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 382.0, + 825.0, + 382.0, + 825.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 429.0, + 1404.0, + 429.0, + 1404.0, + 461.0, + 296.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 488.0, + 1404.0, + 488.0, + 1404.0, + 525.0, + 295.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 517.0, + 1405.0, + 517.0, + 1405.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 547.0, + 1405.0, + 547.0, + 1405.0, + 587.0, + 292.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 1406.0, + 580.0, + 1406.0, + 615.0, + 294.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 613.0, + 1382.0, + 613.0, + 1382.0, + 645.0, + 296.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 613.0, + 1405.0, + 613.0, + 1405.0, + 645.0, + 1402.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 636.0, + 746.0, + 636.0, + 746.0, + 680.0, + 292.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 636.0, + 1405.0, + 636.0, + 1405.0, + 680.0, + 787.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 670.0, + 367.0, + 670.0, + 367.0, + 710.0, + 292.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 670.0, + 638.0, + 670.0, + 638.0, + 710.0, + 594.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 670.0, + 969.0, + 670.0, + 969.0, + 710.0, + 668.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 670.0, + 1407.0, + 670.0, + 1407.0, + 710.0, + 1011.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 1404.0, + 701.0, + 1404.0, + 735.0, + 294.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 730.0, + 1022.0, + 730.0, + 1022.0, + 770.0, + 294.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 730.0, + 1406.0, + 730.0, + 1406.0, + 770.0, + 1055.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 763.0, + 647.0, + 763.0, + 647.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 763.0, + 1406.0, + 763.0, + 1406.0, + 799.0, + 679.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 791.0, + 576.0, + 791.0, + 576.0, + 827.0, + 295.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 791.0, + 654.0, + 791.0, + 654.0, + 827.0, + 608.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 791.0, + 1224.0, + 791.0, + 1224.0, + 827.0, + 685.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 791.0, + 1265.0, + 791.0, + 1265.0, + 827.0, + 1250.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 831.0, + 1369.0, + 831.0, + 1369.0, + 880.0, + 292.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 831.0, + 1411.0, + 831.0, + 1411.0, + 880.0, + 1399.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 864.0, + 801.0, + 864.0, + 801.0, + 952.0, + 285.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 864.0, + 1308.0, + 864.0, + 1308.0, + 952.0, + 856.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 864.0, + 1416.0, + 864.0, + 1416.0, + 952.0, + 1353.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 928.0, + 1243.0, + 928.0, + 1243.0, + 969.0, + 294.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 928.0, + 1407.0, + 928.0, + 1407.0, + 969.0, + 1271.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 398.0, + 961.0, + 398.0, + 998.0, + 294.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 961.0, + 1194.0, + 961.0, + 1194.0, + 998.0, + 423.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 961.0, + 1406.0, + 961.0, + 1406.0, + 998.0, + 1222.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1407.0, + 991.0, + 1407.0, + 1028.0, + 294.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1024.0, + 1153.0, + 1024.0, + 1153.0, + 1057.0, + 295.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1024.0, + 1406.0, + 1024.0, + 1406.0, + 1057.0, + 1179.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1054.0, + 476.0, + 1054.0, + 476.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1054.0, + 1034.0, + 1054.0, + 1034.0, + 1091.0, + 523.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1091.0, + 1061.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1085.0, + 788.0, + 1085.0, + 788.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1085.0, + 1129.0, + 1085.0, + 1129.0, + 1117.0, + 816.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1679.0, + 1406.0, + 1679.0, + 1406.0, + 1717.0, + 295.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1701.0, + 297.0, + 1701.0, + 297.0, + 1759.0, + 289.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1701.0, + 711.0, + 1701.0, + 711.0, + 1759.0, + 324.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1701.0, + 1413.0, + 1701.0, + 1413.0, + 1759.0, + 857.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1406.0, + 1773.0, + 1406.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1799.0, + 1405.0, + 1799.0, + 1405.0, + 1843.0, + 291.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1832.0, + 1405.0, + 1832.0, + 1405.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 1406.0, + 1864.0, + 1406.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1894.0, + 1405.0, + 1894.0, + 1405.0, + 1931.0, + 294.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1925.0, + 360.0, + 1925.0, + 360.0, + 1962.0, + 294.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1925.0, + 1359.0, + 1925.0, + 1359.0, + 1962.0, + 519.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 684.0, + 227.0, + 684.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 227.0, + 717.0, + 227.0, + 717.0, + 267.0, + 710.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 227.0, + 1408.0, + 227.0, + 1408.0, + 267.0, + 753.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 776.0, + 261.0, + 776.0, + 294.0, + 293.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 294.0, + 802.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1397.0, + 321.0, + 1397.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1369, + 1405, + 1369, + 1405, + 1646, + 296, + 1646 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 598, + 881, + 598, + 881, + 939, + 298, + 939 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1404, + 229, + 1404, + 383, + 297, + 383 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 399, + 1405, + 399, + 1405, + 554, + 298, + 554 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 954, + 880, + 954, + 880, + 1204, + 298, + 1204 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 910, + 595, + 1344, + 595, + 1344, + 899, + 910, + 899 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 313, + 1682, + 1393, + 1682, + 1393, + 1873, + 313, + 1873 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 902, + 923, + 1405, + 923, + 1405, + 1168, + 902, + 1168 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 299, + 1205, + 1405, + 1205, + 1405, + 1264, + 299, + 1264 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 299, + 1897, + 1399, + 1897, + 1399, + 1960, + 299, + 1960 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 301, + 76, + 827, + 76, + 827, + 104, + 301, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 301, + 1315, + 733, + 1315, + 733, + 1344, + 301, + 1344 + ], + "score": 0.839 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.77 + }, + { + "category_id": 2, + "poly": [ + 316, + 2004, + 1399, + 2004, + 1399, + 2034, + 316, + 2034 + ], + "score": 0.58 + }, + { + "category_id": 1, + "poly": [ + 316, + 2004, + 1399, + 2004, + 1399, + 2034, + 316, + 2034 + ], + "score": 0.32 + }, + { + "category_id": 13, + "poly": [ + 519, + 785, + 679, + 785, + 679, + 821, + 519, + 821 + ], + "score": 0.93, + "latex": "p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { \\bar { A } } , s _ { t } ^ { A } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1107, + 459, + 1107, + 459, + 1142, + 298, + 1142 + ], + "score": 0.92, + "latex": "p ( a _ { t + 1 } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { A } )" + }, + { + "category_id": 13, + "poly": [ + 401, + 661, + 561, + 661, + 561, + 702, + 401, + 702 + ], + "score": 0.92, + "latex": "p ( a _ { t + 1 } ^ { B } | \\underline { { a } } _ { t } ^ { A } , s _ { t } ^ { A } )" + }, + { + "category_id": 13, + "poly": [ + 633, + 875, + 686, + 875, + 686, + 913, + 633, + 913 + ], + "score": 0.92, + "latex": "u _ { t + 1 } ^ { B ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 650, + 627, + 703, + 627, + 703, + 664, + 650, + 664 + ], + "score": 0.92, + "latex": "a _ { t + 1 } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 398, + 724, + 451, + 724, + 451, + 761, + 398, + 761 + ], + "score": 0.92, + "latex": "a _ { t + 1 } ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 367, + 459, + 401, + 459, + 401, + 493, + 367, + 493 + ], + "score": 0.9, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 327, + 724, + 361, + 724, + 361, + 759, + 327, + 759 + ], + "score": 0.9, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 576, + 627, + 611, + 627, + 611, + 662, + 576, + 662 + ], + "score": 0.9, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 1021, + 984, + 1056, + 984, + 1056, + 1019, + 1021, + 1019 + ], + "score": 0.9, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 1320, + 1075, + 1355, + 1075, + 1355, + 1110, + 1320, + 1110 + ], + "score": 0.9, + "latex": "a _ { t } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 376, + 849, + 433, + 849, + 433, + 880, + 376, + 880 + ], + "score": 0.89, + "latex": "s _ { t + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 421, + 1144, + 453, + 1144, + 453, + 1170, + 421, + 1170 + ], + "score": 0.86, + "latex": "A ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 902, + 1111, + 928, + 1111, + 928, + 1138, + 902, + 1138 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 857, + 462, + 882, + 462, + 882, + 488, + 857, + 488 + ], + "score": 0.79, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 820, + 818, + 844, + 818, + 844, + 845, + 820, + 845 + ], + "score": 0.77, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 852, + 1108, + 879, + 1108, + 879, + 1136, + 852, + 1136 + ], + "score": 0.72, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1268, + 431, + 1292, + 431, + 1292, + 457, + 1268, + 457 + ], + "score": 0.66, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 1195, + 773, + 1229, + 773, + 1229, + 801, + 1195, + 801 + ], + "score": 0.26, + "latex": "s _ { t + } ^ { \\scriptscriptstyle D }" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 600.0, + 1057.0, + 600.0, + 1057.0, + 654.0, + 1017.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 612.0, + 1143.0, + 612.0, + 1143.0, + 650.0, + 1105.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 600.5, + 1225.0, + 600.5, + 1225.0, + 655.0, + 1199.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 600.0, + 1322.0, + 600.0, + 1322.0, + 654.0, + 1293.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 674.0, + 1222.0, + 674.0, + 1222.0, + 728.5, + 1191.0, + 728.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 684.0, + 1079.0, + 684.0, + 1079.0, + 730.5, + 1001.0, + 730.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 767.0, + 1150.0, + 767.0, + 1150.0, + 801.0, + 1099.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.75, + 758.0, + 1255.75, + 758.0, + 1255.75, + 811.0, + 1169.75, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 760.5, + 1347.0, + 760.5, + 1347.0, + 807.5, + 1263.0, + 807.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 842.0, + 1258.0, + 842.0, + 1258.0, + 897.0, + 1168.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1688.0, + 742.0, + 1688.0, + 742.0, + 1706.0, + 701.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1685.0, + 1284.0, + 1685.0, + 1284.0, + 1703.0, + 1241.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1712.0, + 423.0, + 1712.0, + 423.0, + 1754.0, + 311.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1730.0, + 819.0, + 1730.0, + 819.0, + 1756.0, + 735.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1700.0, + 1017.0, + 1700.0, + 1017.0, + 1837.0, + 863.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1726.0, + 1319.0, + 1726.0, + 1319.0, + 1748.0, + 1265.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1771.0, + 420.0, + 1771.0, + 420.0, + 1789.0, + 375.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1807.0, + 819.0, + 1807.0, + 819.0, + 1826.0, + 774.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1798.0, + 1390.0, + 1798.0, + 1390.0, + 1835.0, + 1265.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1853.0, + 1067.0, + 1853.0, + 1067.0, + 1875.0, + 997.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 921.0, + 1410.0, + 921.0, + 1410.0, + 959.0, + 899.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 954.0, + 1407.0, + 954.0, + 1407.0, + 988.0, + 901.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 985.0, + 1020.0, + 985.0, + 1020.0, + 1019.0, + 901.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 985.0, + 1404.0, + 985.0, + 1404.0, + 1019.0, + 1057.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1048.0, + 901.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1045.0, + 1402.0, + 1045.0, + 1402.0, + 1076.0, + 901.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1075.0, + 1319.0, + 1075.0, + 1319.0, + 1109.0, + 900.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1075.0, + 1404.0, + 1075.0, + 1404.0, + 1109.0, + 1356.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1107.0, + 1405.0, + 1107.0, + 1405.0, + 1141.0, + 929.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1135.0, + 1258.0, + 1135.0, + 1258.0, + 1172.0, + 899.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1895.0, + 1401.0, + 1895.0, + 1401.0, + 1931.0, + 295.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 1229.0, + 1923.0, + 1229.0, + 1965.0, + 293.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 297.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 738.0, + 1312.0, + 738.0, + 1351.0, + 294.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1997.0, + 1403.0, + 1997.0, + 1403.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1371.0, + 1402.0, + 1371.0, + 1402.0, + 1404.0, + 296.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 1408.0, + 1400.0, + 1408.0, + 1437.0, + 294.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1406.0, + 1432.0, + 1406.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1458.0, + 1406.0, + 1458.0, + 1406.0, + 1500.0, + 290.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1492.0, + 1406.0, + 1492.0, + 1406.0, + 1528.0, + 294.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1521.0, + 1407.0, + 1521.0, + 1407.0, + 1559.0, + 291.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1553.0, + 1405.0, + 1553.0, + 1405.0, + 1590.0, + 294.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1582.0, + 1408.0, + 1582.0, + 1408.0, + 1620.0, + 293.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1615.0, + 1002.0, + 1615.0, + 1002.0, + 1650.0, + 291.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 593.0, + 885.0, + 593.0, + 885.0, + 635.0, + 294.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 612.0, + 575.0, + 612.0, + 575.0, + 677.0, + 285.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 612.0, + 649.0, + 612.0, + 649.0, + 677.0, + 612.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 612.0, + 893.0, + 612.0, + 893.0, + 677.0, + 704.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 650.0, + 400.0, + 650.0, + 400.0, + 709.0, + 288.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 650.0, + 892.0, + 650.0, + 892.0, + 709.0, + 562.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 695.0, + 326.0, + 695.0, + 326.0, + 773.0, + 286.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 695.0, + 397.0, + 695.0, + 397.0, + 773.0, + 362.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 695.0, + 884.0, + 695.0, + 884.0, + 773.0, + 452.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 754.0, + 885.0, + 754.0, + 885.0, + 789.0, + 293.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 776.0, + 518.0, + 776.0, + 518.0, + 829.0, + 288.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 776.0, + 889.0, + 776.0, + 889.0, + 829.0, + 680.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 817.0, + 819.0, + 817.0, + 819.0, + 847.0, + 295.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 817.0, + 882.0, + 817.0, + 882.0, + 847.0, + 845.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 846.0, + 375.0, + 846.0, + 375.0, + 879.0, + 294.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 846.0, + 882.0, + 846.0, + 882.0, + 879.0, + 434.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 863.0, + 632.0, + 863.0, + 632.0, + 922.0, + 286.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 863.0, + 889.0, + 863.0, + 889.0, + 922.0, + 687.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 906.0, + 401.0, + 906.0, + 401.0, + 943.0, + 294.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 733.5, + 443.0, + 733.5, + 443.0, + 744.0, + 423.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1407.0, + 289.0, + 1407.0, + 327.0, + 292.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 317.0, + 1404.0, + 317.0, + 1404.0, + 358.0, + 291.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 354.0, + 746.0, + 354.0, + 746.0, + 387.0, + 296.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 398.0, + 1405.0, + 398.0, + 1405.0, + 435.0, + 295.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 430.0, + 1267.0, + 430.0, + 1267.0, + 463.0, + 296.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 430.0, + 1403.0, + 430.0, + 1403.0, + 463.0, + 1293.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 459.0, + 366.0, + 459.0, + 366.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 459.0, + 856.0, + 459.0, + 856.0, + 495.0, + 402.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 459.0, + 1403.0, + 459.0, + 1403.0, + 495.0, + 883.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 491.0, + 1403.0, + 491.0, + 1403.0, + 524.0, + 296.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 522.0, + 1408.0, + 522.0, + 1408.0, + 558.0, + 295.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 952.0, + 884.0, + 952.0, + 884.0, + 989.0, + 295.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 987.0, + 881.0, + 987.0, + 881.0, + 1017.0, + 296.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1015.0, + 881.0, + 1015.0, + 881.0, + 1046.0, + 295.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 883.0, + 1048.0, + 883.0, + 1077.0, + 294.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1077.0, + 883.0, + 1077.0, + 883.0, + 1107.0, + 295.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1102.0, + 297.0, + 1102.0, + 297.0, + 1146.0, + 293.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 1102.0, + 851.0, + 1102.0, + 851.0, + 1146.0, + 460.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1102.0, + 885.0, + 1102.0, + 885.0, + 1146.0, + 880.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1140.0, + 420.0, + 1140.0, + 420.0, + 1174.0, + 295.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1140.0, + 884.0, + 1140.0, + 884.0, + 1174.0, + 454.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1170.0, + 885.0, + 1170.0, + 885.0, + 1207.0, + 295.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1200.0, + 1404.0, + 1200.0, + 1404.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1235.0, + 1406.0, + 1235.0, + 1406.0, + 1266.0, + 297.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1997.0, + 1403.0, + 1997.0, + 1403.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 304, + 1404, + 304, + 1404, + 671, + 299, + 671 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1576, + 1404, + 1576, + 1404, + 1911, + 298, + 1911 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 687, + 1404, + 687, + 1404, + 963, + 298, + 963 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1285, + 1404, + 1285, + 1404, + 1560, + 298, + 1560 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 978, + 1403, + 978, + 1403, + 1131, + 299, + 1131 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1147, + 1404, + 1147, + 1404, + 1269, + 299, + 1269 + ], + "score": 0.973 + }, + { + "category_id": 2, + "poly": [ + 297, + 1977, + 1403, + 1977, + 1403, + 2033, + 297, + 2033 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 571, + 226, + 571, + 261, + 300, + 261 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.794 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 828, + 75, + 828, + 104, + 300, + 104 + ], + "score": 0.742 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 827, + 75, + 827, + 104, + 300, + 104 + ], + "score": 0.399 + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1968.0, + 1405.0, + 1968.0, + 1405.0, + 2015.0, + 328.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 603.0, + 2003.0, + 603.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 577.0, + 222.0, + 577.0, + 268.0, + 291.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 831.0, + 72.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 831.0, + 72.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 301.0, + 1409.0, + 301.0, + 1409.0, + 343.0, + 293.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 334.0, + 1406.0, + 334.0, + 1406.0, + 372.0, + 294.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 1405.0, + 367.0, + 1405.0, + 401.0, + 295.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 400.0, + 1401.0, + 400.0, + 1401.0, + 430.0, + 295.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 424.0, + 1405.0, + 424.0, + 1405.0, + 464.0, + 293.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 460.0, + 1404.0, + 460.0, + 1404.0, + 494.0, + 295.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 490.0, + 1404.0, + 490.0, + 1404.0, + 524.0, + 295.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 519.0, + 1405.0, + 519.0, + 1405.0, + 553.0, + 295.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 550.0, + 1405.0, + 550.0, + 1405.0, + 586.0, + 293.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 577.0, + 1408.0, + 577.0, + 1408.0, + 616.0, + 292.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 612.0, + 1404.0, + 612.0, + 1404.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 744.0, + 642.0, + 744.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1404.0, + 1574.0, + 1404.0, + 1611.0, + 293.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1606.0, + 1406.0, + 1606.0, + 1406.0, + 1641.0, + 294.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 1405.0, + 1637.0, + 1405.0, + 1672.0, + 294.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 1405.0, + 1668.0, + 1405.0, + 1703.0, + 294.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1695.0, + 1409.0, + 1695.0, + 1409.0, + 1735.0, + 292.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1730.0, + 1406.0, + 1730.0, + 1406.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1789.0, + 1406.0, + 1789.0, + 1406.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1408.0, + 1818.0, + 1408.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1408.0, + 1881.0, + 1408.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 685.0, + 1406.0, + 685.0, + 1406.0, + 723.0, + 293.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 719.0, + 1405.0, + 719.0, + 1405.0, + 752.0, + 294.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 745.0, + 1408.0, + 745.0, + 1408.0, + 785.0, + 292.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 775.0, + 1406.0, + 775.0, + 1406.0, + 816.0, + 293.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 811.0, + 1405.0, + 811.0, + 1405.0, + 843.0, + 293.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 839.0, + 1408.0, + 839.0, + 1408.0, + 876.0, + 294.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 868.0, + 1406.0, + 868.0, + 1406.0, + 907.0, + 293.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 902.0, + 1401.0, + 902.0, + 1401.0, + 935.0, + 296.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 933.0, + 1381.0, + 933.0, + 1381.0, + 966.0, + 294.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1287.0, + 1402.0, + 1287.0, + 1402.0, + 1320.0, + 296.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1316.0, + 1401.0, + 1316.0, + 1401.0, + 1348.0, + 296.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1346.0, + 1405.0, + 1346.0, + 1405.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1377.0, + 1406.0, + 1377.0, + 1406.0, + 1412.0, + 293.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1406.0, + 1405.0, + 1406.0, + 1405.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1436.0, + 1404.0, + 1436.0, + 1404.0, + 1473.0, + 293.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1501.0, + 296.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1408.0, + 1498.0, + 1408.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1531.0, + 1037.0, + 1531.0, + 1037.0, + 1563.0, + 296.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 979.0, + 1404.0, + 979.0, + 1404.0, + 1012.0, + 297.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1008.0, + 1405.0, + 1008.0, + 1405.0, + 1044.0, + 293.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1040.0, + 1407.0, + 1040.0, + 1407.0, + 1077.0, + 294.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 1408.0, + 1069.0, + 1408.0, + 1105.0, + 293.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1102.0, + 600.0, + 1102.0, + 600.0, + 1135.0, + 295.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1147.0, + 1404.0, + 1147.0, + 1404.0, + 1179.0, + 295.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1178.0, + 1406.0, + 1178.0, + 1406.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1209.0, + 1404.0, + 1209.0, + 1404.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1239.0, + 1041.0, + 1239.0, + 1041.0, + 1274.0, + 293.0, + 1274.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 290, + 1404, + 290, + 1404, + 535, + 297, + 535 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1722, + 1405, + 1722, + 1405, + 2029, + 298, + 2029 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 780, + 1404, + 780, + 1404, + 965, + 297, + 965 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 550, + 1404, + 550, + 1404, + 673, + 298, + 673 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 304, + 1007, + 1385, + 1007, + 1385, + 1583, + 304, + 1583 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 298, + 1609, + 1398, + 1609, + 1398, + 1701, + 298, + 1701 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 547, + 226, + 547, + 261, + 299, + 261 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 827, + 76, + 827, + 104, + 300, + 104 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 300, + 725, + 686, + 725, + 686, + 755, + 300, + 755 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2111, + 842, + 2111 + ], + "score": 0.751 + }, + { + "category_id": 13, + "poly": [ + 749, + 472, + 976, + 472, + 976, + 506, + 749, + 506 + ], + "score": 0.9, + "latex": "( C \\in [ 0 . 2 \\mathrm { - } 3 . 5 ] \\times \\bar { 1 } 0 ^ { 8 } )" + }, + { + "category_id": 13, + "poly": [ + 1125, + 1610, + 1197, + 1610, + 1197, + 1638, + 1125, + 1638 + ], + "score": 0.88, + "latex": "9 9 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 661, + 474, + 686, + 474, + 686, + 500, + 661, + 500 + ], + "score": 0.79, + "latex": "C" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1010.0, + 365.0, + 1010.0, + 365.0, + 1033.0, + 330.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1013.0, + 730.0, + 1013.0, + 730.0, + 1032.0, + 699.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1012.0, + 1102.0, + 1012.0, + 1102.0, + 1033.0, + 1067.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1041.0, + 362.0, + 1041.0, + 362.0, + 1165.0, + 314.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1029.0, + 575.0, + 1029.0, + 575.0, + 1068.0, + 413.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1033.0, + 730.0, + 1033.0, + 730.0, + 1165.0, + 680.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1029.0, + 893.0, + 1029.0, + 893.0, + 1069.0, + 781.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1043.0, + 1100.0, + 1043.0, + 1100.0, + 1166.0, + 1047.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1029.0, + 1249.0, + 1029.0, + 1249.0, + 1069.0, + 1150.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1073.0, + 362.0, + 1073.0, + 362.0, + 1095.0, + 332.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1068.0, + 483.0, + 1068.0, + 483.0, + 1087.0, + 415.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1080.0, + 730.0, + 1080.0, + 730.0, + 1099.0, + 699.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1068.0, + 895.0, + 1068.0, + 895.0, + 1087.0, + 782.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1075.0, + 1099.0, + 1075.0, + 1099.0, + 1097.0, + 1067.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 1068.0, + 1252.0, + 1068.0, + 1252.0, + 1087.0, + 1151.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1104.0, + 362.0, + 1104.0, + 362.0, + 1126.0, + 332.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1101.0, + 731.0, + 1101.0, + 731.0, + 1124.0, + 699.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1109.0, + 1099.0, + 1109.0, + 1099.0, + 1131.0, + 1067.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1125.0, + 731.0, + 1125.0, + 731.0, + 1147.0, + 699.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1136.0, + 361.0, + 1136.0, + 361.0, + 1156.0, + 338.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1148.0, + 729.0, + 1148.0, + 729.0, + 1168.0, + 707.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1140.0, + 1099.0, + 1140.0, + 1099.0, + 1162.0, + 1067.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1168.0, + 363.0, + 1168.0, + 363.0, + 1184.0, + 347.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1172.0, + 730.0, + 1172.0, + 730.0, + 1189.0, + 715.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1176.0, + 1099.0, + 1176.0, + 1099.0, + 1193.0, + 1084.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1202.0, + 384.0, + 1202.0, + 384.0, + 1220.0, + 357.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1199.0, + 427.0, + 1199.0, + 427.0, + 1222.0, + 398.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1199.0, + 472.0, + 1199.0, + 472.0, + 1222.0, + 442.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1199.0, + 516.0, + 1199.0, + 516.0, + 1222.0, + 487.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1199.0, + 559.0, + 1199.0, + 559.0, + 1222.0, + 530.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1199.0, + 605.0, + 1199.0, + 605.0, + 1222.0, + 574.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1199.0, + 648.0, + 1199.0, + 648.0, + 1222.0, + 617.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1202.0, + 751.0, + 1202.0, + 751.0, + 1220.0, + 725.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1201.0, + 794.0, + 1201.0, + 794.0, + 1220.0, + 768.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1201.0, + 838.0, + 1201.0, + 838.0, + 1219.0, + 813.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1201.0, + 883.0, + 1201.0, + 883.0, + 1219.0, + 857.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1199.0, + 928.0, + 1199.0, + 928.0, + 1222.0, + 899.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1199.0, + 972.0, + 1199.0, + 972.0, + 1222.0, + 942.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 1199.0, + 1017.0, + 1199.0, + 1017.0, + 1222.0, + 985.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1202.0, + 1114.0, + 1202.0, + 1114.0, + 1219.0, + 1100.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1202.0, + 1168.0, + 1202.0, + 1168.0, + 1219.0, + 1153.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1202.0, + 1221.0, + 1202.0, + 1221.0, + 1219.0, + 1207.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1202.0, + 1275.0, + 1202.0, + 1275.0, + 1219.0, + 1260.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1202.0, + 1330.0, + 1202.0, + 1330.0, + 1219.0, + 1314.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1202.0, + 1383.0, + 1202.0, + 1383.0, + 1218.0, + 1367.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1217.0, + 541.0, + 1217.0, + 541.0, + 1237.0, + 463.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1214.0, + 638.0, + 1214.0, + 638.0, + 1232.0, + 613.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1217.0, + 908.0, + 1217.0, + 908.0, + 1237.0, + 831.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1215.0, + 1004.0, + 1215.0, + 1004.0, + 1230.0, + 984.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1217.0, + 1280.0, + 1217.0, + 1280.0, + 1237.0, + 1202.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1214.0, + 1381.0, + 1214.0, + 1381.0, + 1232.0, + 1356.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1253.0, + 651.0, + 1253.0, + 651.0, + 1288.0, + 303.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1253.0, + 993.0, + 1253.0, + 993.0, + 1288.0, + 697.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1255.0, + 1386.0, + 1255.0, + 1386.0, + 1286.0, + 1040.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1310.0, + 719.0, + 1310.0, + 719.0, + 1331.0, + 680.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1310.0, + 1073.0, + 1310.0, + 1073.0, + 1331.0, + 1039.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1321.0, + 361.0, + 1321.0, + 361.0, + 1342.0, + 322.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1327.0, + 718.0, + 1327.0, + 718.0, + 1350.0, + 680.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1325.0, + 864.0, + 1325.0, + 864.0, + 1348.0, + 767.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1326.0, + 1342.0, + 1326.0, + 1342.0, + 1350.0, + 1242.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1346.0, + 360.0, + 1346.0, + 360.0, + 1460.0, + 303.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1373.0, + 575.0, + 1373.0, + 575.0, + 1416.0, + 408.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1347.0, + 724.0, + 1347.0, + 724.0, + 1463.0, + 662.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1345.0, + 880.0, + 1345.0, + 880.0, + 1367.0, + 769.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1343.0, + 1071.0, + 1343.0, + 1071.0, + 1464.0, + 1022.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1346.0, + 1349.0, + 1346.0, + 1349.0, + 1387.0, + 1243.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1406.0, + 482.0, + 1406.0, + 482.0, + 1435.0, + 408.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1404.0, + 719.0, + 1404.0, + 719.0, + 1423.0, + 688.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1413.0, + 1070.0, + 1413.0, + 1070.0, + 1432.0, + 1039.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1429.0, + 360.0, + 1429.0, + 360.0, + 1447.0, + 330.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1423.0, + 721.0, + 1423.0, + 721.0, + 1444.0, + 687.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1443.0, + 717.0, + 1443.0, + 717.0, + 1462.0, + 687.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1447.0, + 1070.0, + 1447.0, + 1070.0, + 1465.0, + 1039.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1464.0, + 360.0, + 1464.0, + 360.0, + 1481.0, + 330.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1460.0, + 718.0, + 1460.0, + 718.0, + 1483.0, + 686.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1480.0, + 717.0, + 1480.0, + 717.0, + 1499.0, + 688.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1481.0, + 1070.0, + 1481.0, + 1070.0, + 1499.0, + 1039.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1500.0, + 381.0, + 1500.0, + 381.0, + 1517.0, + 354.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1498.0, + 426.0, + 1498.0, + 426.0, + 1520.0, + 397.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1499.0, + 469.0, + 1499.0, + 469.0, + 1517.0, + 442.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1498.0, + 515.0, + 1498.0, + 515.0, + 1520.0, + 485.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1499.0, + 558.0, + 1499.0, + 558.0, + 1517.0, + 530.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1498.0, + 605.0, + 1498.0, + 605.0, + 1520.0, + 573.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1498.0, + 648.0, + 1498.0, + 648.0, + 1520.0, + 617.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1499.0, + 740.0, + 1499.0, + 740.0, + 1517.0, + 713.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1499.0, + 783.0, + 1499.0, + 783.0, + 1517.0, + 756.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1499.0, + 827.0, + 1499.0, + 827.0, + 1517.0, + 800.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1499.0, + 872.0, + 1499.0, + 872.0, + 1517.0, + 844.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1499.0, + 917.0, + 1499.0, + 917.0, + 1517.0, + 889.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1498.0, + 963.0, + 1498.0, + 963.0, + 1520.0, + 932.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1498.0, + 1007.0, + 1498.0, + 1007.0, + 1520.0, + 974.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1500.0, + 1086.0, + 1500.0, + 1086.0, + 1517.0, + 1071.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1501.0, + 1141.0, + 1501.0, + 1141.0, + 1517.0, + 1126.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1500.0, + 1196.0, + 1500.0, + 1196.0, + 1517.0, + 1180.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1500.0, + 1252.0, + 1500.0, + 1252.0, + 1517.0, + 1237.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1500.0, + 1308.0, + 1500.0, + 1308.0, + 1516.0, + 1294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1500.0, + 1362.0, + 1500.0, + 1362.0, + 1516.0, + 1348.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1515.0, + 538.0, + 1515.0, + 538.0, + 1535.0, + 461.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1512.0, + 638.0, + 1512.0, + 638.0, + 1530.0, + 613.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1515.0, + 897.0, + 1515.0, + 897.0, + 1535.0, + 820.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1512.0, + 997.0, + 1512.0, + 997.0, + 1530.0, + 971.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1515.0, + 1254.0, + 1515.0, + 1254.0, + 1535.0, + 1178.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 1514.0, + 1359.0, + 1514.0, + 1359.0, + 1529.0, + 1338.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1552.0, + 648.0, + 1552.0, + 648.0, + 1582.0, + 306.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1552.0, + 979.0, + 1552.0, + 979.0, + 1583.0, + 688.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1552.0, + 1364.0, + 1552.0, + 1364.0, + 1583.0, + 1023.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1046.0, + 1402.0, + 1046.0, + 1402.0, + 1115.5, + 1187.0, + 1115.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.25, + 1075.5, + 1023.25, + 1075.5, + 1023.25, + 1152.5, + 744.25, + 1152.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1108.5, + 1253.0, + 1108.5, + 1253.0, + 1178.5, + 1124.0, + 1178.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1293.5, + 1051.0, + 1293.5, + 1051.0, + 1461.5, + 669.0, + 1461.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1608.0, + 1124.0, + 1608.0, + 1124.0, + 1642.0, + 294.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1608.0, + 1404.0, + 1608.0, + 1404.0, + 1642.0, + 1198.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1639.0, + 1403.0, + 1639.0, + 1403.0, + 1673.0, + 295.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1669.0, + 1143.0, + 1669.0, + 1143.0, + 1703.0, + 295.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 222.0, + 551.0, + 222.0, + 551.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 296.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 725.0, + 689.0, + 725.0, + 689.0, + 757.0, + 296.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 289.0, + 1406.0, + 289.0, + 1406.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 320.0, + 1407.0, + 320.0, + 1407.0, + 357.0, + 294.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 350.0, + 1405.0, + 350.0, + 1405.0, + 386.0, + 292.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 380.0, + 1407.0, + 380.0, + 1407.0, + 415.0, + 294.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 410.0, + 1404.0, + 410.0, + 1404.0, + 446.0, + 294.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 1404.0, + 444.0, + 1404.0, + 477.0, + 295.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 471.0, + 660.0, + 471.0, + 660.0, + 508.0, + 294.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 471.0, + 748.0, + 471.0, + 748.0, + 508.0, + 687.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 471.0, + 1405.0, + 471.0, + 1405.0, + 508.0, + 977.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 502.0, + 741.0, + 502.0, + 741.0, + 538.0, + 295.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1723.0, + 1405.0, + 1723.0, + 1405.0, + 1755.0, + 296.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1752.0, + 1408.0, + 1752.0, + 1408.0, + 1789.0, + 293.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1784.0, + 1405.0, + 1784.0, + 1405.0, + 1818.0, + 293.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1814.0, + 1408.0, + 1814.0, + 1408.0, + 1850.0, + 295.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1872.0, + 1406.0, + 1872.0, + 1406.0, + 1914.0, + 290.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1906.0, + 1405.0, + 1906.0, + 1405.0, + 1942.0, + 295.0, + 1942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1937.0, + 1405.0, + 1937.0, + 1405.0, + 1972.0, + 293.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1967.0, + 1406.0, + 1967.0, + 1406.0, + 2003.0, + 295.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1996.0, + 1040.0, + 1996.0, + 1040.0, + 2033.0, + 293.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 782.0, + 1404.0, + 782.0, + 1404.0, + 814.0, + 296.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 811.0, + 1405.0, + 811.0, + 1405.0, + 847.0, + 294.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 842.0, + 1407.0, + 842.0, + 1407.0, + 876.0, + 292.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 907.0, + 296.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 902.0, + 1406.0, + 902.0, + 1406.0, + 939.0, + 294.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 932.0, + 643.0, + 932.0, + 643.0, + 969.0, + 294.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 549.0, + 1404.0, + 549.0, + 1404.0, + 584.0, + 293.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 580.0, + 1404.0, + 580.0, + 1404.0, + 614.0, + 293.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 1404.0, + 613.0, + 1404.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 704.0, + 640.0, + 704.0, + 675.0, + 293.0, + 675.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 545, + 230, + 1404, + 230, + 1404, + 597, + 545, + 597 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 545, + 612, + 1404, + 612, + 1404, + 857, + 545, + 857 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1190, + 1406, + 1190, + 1406, + 1621, + 297, + 1621 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 929, + 1405, + 929, + 1405, + 1174, + 297, + 1174 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 575, + 1638, + 1395, + 1638, + 1395, + 1878, + 575, + 1878 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 300, + 230, + 514, + 230, + 514, + 533, + 300, + 533 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 296, + 559, + 524, + 559, + 524, + 834, + 296, + 834 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 296, + 1637, + 552, + 1637, + 552, + 2033, + 296, + 2033 + ], + "score": 0.928 + }, + { + "category_id": 4, + "poly": [ + 570, + 1903, + 1408, + 1903, + 1408, + 1997, + 570, + 1997 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 826, + 76, + 826, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 302, + 888, + 820, + 888, + 820, + 917, + 302, + 917 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.795 + }, + { + "category_id": 13, + "poly": [ + 1101, + 1219, + 1208, + 1219, + 1208, + 1254, + 1101, + 1254 + ], + "score": 0.94, + "latex": "p ( a ^ { k } | m ^ { k } )" + }, + { + "category_id": 13, + "poly": [ + 1262, + 1220, + 1370, + 1220, + 1370, + 1254, + 1262, + 1254 + ], + "score": 0.94, + "latex": "p ( m ^ { k } | a ^ { k } )" + }, + { + "category_id": 13, + "poly": [ + 1255, + 476, + 1397, + 476, + 1397, + 508, + 1255, + 508 + ], + "score": 0.93, + "latex": "p ( a _ { t } ^ { B } | a _ { t } ^ { A } , s _ { t } ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 493, + 1370, + 686, + 1370, + 686, + 1408, + 493, + 1408 + ], + "score": 0.92, + "latex": "\\mathbf { I C } = I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 883, + 1404, + 1073, + 1404, + 1073, + 1442, + 883, + 1442 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\mathbf { I C } = I ( a _ { t } ^ { A } ; a _ { t + 1 } ^ { B } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 414, + 1084, + 478, + 1084, + 478, + 1111, + 414, + 1111 + ], + "score": 0.89, + "latex": "\\alpha = 0" + }, + { + "category_id": 13, + "poly": [ + 634, + 323, + 700, + 323, + 700, + 350, + 634, + 350 + ], + "score": 0.89, + "latex": "t = 4 9" + }, + { + "category_id": 13, + "poly": [ + 679, + 1220, + 759, + 1220, + 759, + 1254, + 679, + 1254 + ], + "score": 0.88, + "latex": "\\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 874, + 995, + 904, + 995, + 904, + 1021, + 874, + 1021 + ], + "score": 0.85, + "latex": "\\pi _ { c }" + }, + { + "category_id": 13, + "poly": [ + 576, + 1725, + 613, + 1725, + 613, + 1760, + 576, + 1760 + ], + "score": 0.62, + "latex": "\\sum _ { \\stackrel { \\scriptstyle \\in } { \\mathscr { n } } } ^ { \\stackrel { \\scriptstyle \\flat } { \\ v } } 1 . 0" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1637.0, + 614.0, + 1637.0, + 614.0, + 1657.0, + 589.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1646.0, + 835.0, + 1646.0, + 835.0, + 1670.0, + 652.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1667.0, + 940.0, + 1667.0, + 940.0, + 1688.0, + 654.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1686.0, + 614.0, + 1686.0, + 614.0, + 1704.0, + 590.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1686.0, + 863.0, + 1686.0, + 863.0, + 1707.0, + 652.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1702.0, + 812.0, + 1702.0, + 812.0, + 1728.0, + 654.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1721.0, + 943.0, + 1721.0, + 943.0, + 1765.0, + 652.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1779.0, + 617.0, + 1779.0, + 617.0, + 1801.0, + 587.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1827.0, + 616.0, + 1827.0, + 616.0, + 1848.0, + 588.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1838.0, + 853.0, + 1838.0, + 853.0, + 1862.0, + 671.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1837.0, + 1119.0, + 1837.0, + 1119.0, + 1861.0, + 894.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1837.0, + 1361.0, + 1837.0, + 1361.0, + 1861.0, + 1135.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1854.0, + 1070.0, + 1854.0, + 1070.0, + 1879.0, + 941.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1854.0, + 1310.0, + 1854.0, + 1310.0, + 1879.0, + 1189.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 310.0, + 440.0, + 310.0, + 440.0, + 328.0, + 372.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 501.0, + 487.0, + 501.0, + 487.0, + 519.0, + 413.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 557.0, + 531.0, + 557.0, + 531.0, + 594.0, + 294.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 588.0, + 529.0, + 588.0, + 529.0, + 622.0, + 293.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 619.0, + 529.0, + 619.0, + 529.0, + 654.0, + 294.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 647.0, + 530.0, + 647.0, + 530.0, + 687.0, + 292.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 680.0, + 528.0, + 680.0, + 528.0, + 714.0, + 293.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 711.0, + 526.0, + 711.0, + 526.0, + 743.0, + 293.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 740.0, + 526.0, + 740.0, + 526.0, + 773.0, + 292.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 769.0, + 527.0, + 769.0, + 527.0, + 808.0, + 292.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 800.0, + 452.0, + 800.0, + 452.0, + 837.0, + 294.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1636.0, + 551.0, + 1636.0, + 551.0, + 1670.0, + 296.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1665.0, + 554.0, + 1665.0, + 554.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1699.0, + 554.0, + 1699.0, + 554.0, + 1728.0, + 294.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1728.0, + 551.0, + 1728.0, + 551.0, + 1759.0, + 296.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1757.0, + 553.0, + 1757.0, + 553.0, + 1791.0, + 294.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1791.0, + 550.0, + 1791.0, + 550.0, + 1819.0, + 296.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 552.0, + 1820.0, + 552.0, + 1851.0, + 295.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1851.0, + 555.0, + 1851.0, + 555.0, + 1881.0, + 296.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1882.0, + 553.0, + 1882.0, + 553.0, + 1913.0, + 296.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1913.0, + 552.0, + 1913.0, + 552.0, + 1943.0, + 296.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 555.0, + 1942.0, + 555.0, + 1973.0, + 293.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 551.0, + 1973.0, + 551.0, + 2004.0, + 295.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 552.0, + 2002.0, + 552.0, + 2034.0, + 292.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1901.0, + 1407.0, + 1901.0, + 1407.0, + 1938.0, + 570.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1933.0, + 1407.0, + 1933.0, + 1407.0, + 1966.0, + 570.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1964.0, + 1367.0, + 1964.0, + 1367.0, + 1999.0, + 571.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 830.0, + 73.0, + 830.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 884.0, + 825.0, + 884.0, + 825.0, + 923.0, + 295.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 263.0, + 543.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 295.0, + 543.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 325.0, + 541.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 322.0, + 633.0, + 322.0, + 633.0, + 353.0, + 542.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 322.0, + 1403.0, + 322.0, + 1403.0, + 353.0, + 701.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 385.0, + 542.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 382.0, + 1404.0, + 382.0, + 1404.0, + 415.0, + 542.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 446.0, + 542.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 443.0, + 1407.0, + 443.0, + 1407.0, + 476.0, + 542.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 462.0, + 1254.0, + 462.0, + 1254.0, + 517.0, + 535.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 462.0, + 1412.0, + 462.0, + 1412.0, + 517.0, + 1398.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 502.0, + 1404.0, + 502.0, + 1404.0, + 539.0, + 542.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 533.0, + 1404.0, + 533.0, + 1404.0, + 568.0, + 543.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 566.0, + 1262.0, + 566.0, + 1262.0, + 600.0, + 542.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 612.0, + 1406.0, + 612.0, + 1406.0, + 645.0, + 542.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 642.0, + 1407.0, + 642.0, + 1407.0, + 675.0, + 543.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 705.0, + 542.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 702.0, + 1404.0, + 702.0, + 1404.0, + 736.0, + 543.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 734.0, + 1404.0, + 734.0, + 1404.0, + 768.0, + 543.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 765.0, + 1404.0, + 765.0, + 1404.0, + 798.0, + 543.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 796.0, + 1403.0, + 796.0, + 1403.0, + 830.0, + 543.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 826.0, + 1049.0, + 826.0, + 1049.0, + 859.0, + 542.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1191.0, + 1404.0, + 1191.0, + 1404.0, + 1223.0, + 296.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1215.0, + 678.0, + 1215.0, + 678.0, + 1259.0, + 292.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1215.0, + 1100.0, + 1215.0, + 1100.0, + 1259.0, + 760.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1215.0, + 1261.0, + 1215.0, + 1261.0, + 1259.0, + 1209.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1215.0, + 1408.0, + 1215.0, + 1408.0, + 1259.0, + 1371.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 1408.0, + 1251.0, + 1408.0, + 1287.0, + 295.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1282.0, + 1407.0, + 1282.0, + 1407.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1312.0, + 1406.0, + 1312.0, + 1406.0, + 1348.0, + 295.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1344.0, + 492.0, + 1344.0, + 492.0, + 1423.0, + 284.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1344.0, + 1414.0, + 1344.0, + 1414.0, + 1423.0, + 687.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1402.0, + 882.0, + 1402.0, + 882.0, + 1450.0, + 291.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1402.0, + 1408.0, + 1402.0, + 1408.0, + 1450.0, + 1074.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1436.0, + 1406.0, + 1436.0, + 1406.0, + 1473.0, + 295.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1469.0, + 1407.0, + 1469.0, + 1407.0, + 1502.0, + 294.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1404.0, + 1500.0, + 1404.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1529.0, + 1403.0, + 1529.0, + 1403.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 1406.0, + 1559.0, + 1406.0, + 1596.0, + 294.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1592.0, + 796.0, + 1592.0, + 796.0, + 1624.0, + 296.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 929.0, + 1406.0, + 929.0, + 1406.0, + 963.0, + 296.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 960.0, + 1406.0, + 960.0, + 1406.0, + 994.0, + 295.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 990.0, + 873.0, + 990.0, + 873.0, + 1024.0, + 295.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 990.0, + 1405.0, + 990.0, + 1405.0, + 1024.0, + 905.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 1405.0, + 1022.0, + 1405.0, + 1055.0, + 295.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1050.0, + 1405.0, + 1050.0, + 1405.0, + 1086.0, + 292.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1080.0, + 413.0, + 1080.0, + 413.0, + 1119.0, + 292.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1080.0, + 1406.0, + 1080.0, + 1406.0, + 1119.0, + 479.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1113.0, + 1406.0, + 1113.0, + 1406.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1140.0, + 1410.0, + 1140.0, + 1410.0, + 1181.0, + 292.0, + 1181.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 688, + 1405, + 688, + 1405, + 1025, + 298, + 1025 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1110, + 1404, + 1110, + 1404, + 1356, + 298, + 1356 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 367, + 1404, + 367, + 1404, + 674, + 298, + 674 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1370, + 1404, + 1370, + 1404, + 1616, + 298, + 1616 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1631, + 1405, + 1631, + 1405, + 1785, + 299, + 1785 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 352, + 299, + 352 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 303, + 1826, + 789, + 1826, + 789, + 1860, + 303, + 1860 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 301, + 76, + 827, + 76, + 827, + 104, + 301, + 104 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 0, + "poly": [ + 297, + 1064, + 855, + 1064, + 855, + 1093, + 297, + 1093 + ], + "score": 0.552 + }, + { + "category_id": 0, + "poly": [ + 297, + 1064, + 855, + 1064, + 855, + 1093, + 297, + 1093 + ], + "score": 0.395 + }, + { + "category_id": 13, + "poly": [ + 298, + 580, + 433, + 580, + 433, + 617, + 298, + 617 + ], + "score": 0.92, + "latex": "I ( m _ { t } ^ { A } ; a _ { t + 1 } ^ { B } )" + }, + { + "category_id": 13, + "poly": [ + 893, + 1171, + 958, + 1171, + 958, + 1202, + 893, + 1202 + ], + "score": 0.88, + "latex": "5 \\cdot 1 0 ^ { 8 }" + }, + { + "category_id": 13, + "poly": [ + 736, + 1171, + 802, + 1171, + 802, + 1201, + 736, + 1201 + ], + "score": 0.88, + "latex": "3 \\cdot 1 0 ^ { 8 }" + }, + { + "category_id": 13, + "poly": [ + 814, + 749, + 851, + 749, + 851, + 778, + 814, + 778 + ], + "score": 0.87, + "latex": "\\dot { E } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 297, + 874, + 381, + 874, + 381, + 905, + 297, + 905 + ], + "score": 0.87, + "latex": "\\rho = . 6 7" + }, + { + "category_id": 13, + "poly": [ + 647, + 874, + 731, + 874, + 731, + 905, + 647, + 905 + ], + "score": 0.87, + "latex": "\\rho = . 3 4" + }, + { + "category_id": 13, + "poly": [ + 1113, + 461, + 1165, + 461, + 1165, + 490, + 1113, + 490 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 392, + 873, + 489, + 873, + 489, + 905, + 392, + 905 + ], + "score": 0.84, + "latex": "p { < } 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 742, + 873, + 839, + 873, + 839, + 905, + 742, + 905 + ], + "score": 0.81, + "latex": "p { < } 0 . 0 0 1" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1823.0, + 796.0, + 1823.0, + 796.0, + 1867.0, + 292.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 297.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 858.0, + 1061.0, + 858.0, + 1096.0, + 295.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 858.0, + 1061.0, + 858.0, + 1096.0, + 295.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 689.0, + 1405.0, + 689.0, + 1405.0, + 724.0, + 295.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 720.0, + 1403.0, + 720.0, + 1403.0, + 755.0, + 295.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 746.0, + 813.0, + 746.0, + 813.0, + 786.0, + 292.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 746.0, + 1406.0, + 746.0, + 1406.0, + 786.0, + 852.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 1406.0, + 783.0, + 1406.0, + 814.0, + 296.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 1403.0, + 812.0, + 1403.0, + 843.0, + 295.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 840.0, + 1409.0, + 840.0, + 1409.0, + 878.0, + 293.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 871.0, + 296.0, + 871.0, + 296.0, + 907.0, + 292.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 871.0, + 391.0, + 871.0, + 391.0, + 907.0, + 382.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 871.0, + 646.0, + 871.0, + 646.0, + 907.0, + 490.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 871.0, + 741.0, + 871.0, + 741.0, + 907.0, + 732.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 871.0, + 1410.0, + 871.0, + 1410.0, + 907.0, + 840.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 902.0, + 1405.0, + 902.0, + 1405.0, + 940.0, + 293.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 935.0, + 1406.0, + 935.0, + 1406.0, + 966.0, + 296.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 962.0, + 1406.0, + 962.0, + 1406.0, + 998.0, + 293.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 996.0, + 559.0, + 996.0, + 559.0, + 1027.0, + 296.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1113.0, + 1402.0, + 1113.0, + 1402.0, + 1143.0, + 296.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1142.0, + 1405.0, + 1142.0, + 1405.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1171.0, + 735.0, + 1171.0, + 735.0, + 1208.0, + 293.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 1171.0, + 892.0, + 1171.0, + 892.0, + 1208.0, + 803.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1171.0, + 1408.0, + 1171.0, + 1408.0, + 1208.0, + 959.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1203.0, + 1405.0, + 1203.0, + 1405.0, + 1237.0, + 294.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 1405.0, + 1233.0, + 1405.0, + 1267.0, + 294.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1298.0, + 294.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1293.0, + 1404.0, + 1293.0, + 1404.0, + 1328.0, + 293.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1323.0, + 900.0, + 1323.0, + 900.0, + 1359.0, + 293.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 369.0, + 1402.0, + 369.0, + 1402.0, + 401.0, + 296.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1404.0, + 400.0, + 1404.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 428.0, + 1405.0, + 428.0, + 1405.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 458.0, + 1112.0, + 458.0, + 1112.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 458.0, + 1406.0, + 458.0, + 1406.0, + 496.0, + 1166.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 487.0, + 1405.0, + 487.0, + 1405.0, + 528.0, + 292.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 1405.0, + 519.0, + 1405.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1404.0, + 551.0, + 1404.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 573.0, + 297.0, + 573.0, + 297.0, + 630.0, + 288.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 573.0, + 1411.0, + 573.0, + 1411.0, + 630.0, + 434.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 608.0, + 1406.0, + 608.0, + 1406.0, + 652.0, + 292.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 1090.0, + 640.0, + 1090.0, + 678.0, + 293.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1372.0, + 1405.0, + 1372.0, + 1405.0, + 1405.0, + 296.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1404.0, + 1404.0, + 1404.0, + 1404.0, + 1434.0, + 296.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1433.0, + 1404.0, + 1433.0, + 1404.0, + 1467.0, + 294.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1463.0, + 1402.0, + 1463.0, + 1402.0, + 1497.0, + 294.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1492.0, + 1405.0, + 1492.0, + 1405.0, + 1528.0, + 293.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1523.0, + 1408.0, + 1523.0, + 1408.0, + 1560.0, + 293.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1553.0, + 1406.0, + 1553.0, + 1406.0, + 1590.0, + 293.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1584.0, + 919.0, + 1584.0, + 919.0, + 1622.0, + 292.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1629.0, + 1409.0, + 1629.0, + 1409.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1660.0, + 1403.0, + 1660.0, + 1403.0, + 1693.0, + 297.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1406.0, + 1693.0, + 1406.0, + 1727.0, + 295.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1720.0, + 1407.0, + 1720.0, + 1407.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1754.0, + 1342.0, + 1754.0, + 1342.0, + 1791.0, + 294.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 297.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1403.0, + 1911.0, + 1403.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2005.0, + 1313.0, + 2005.0, + 1313.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 294.0, + 295.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1406.0, + 290.0, + 1406.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 1104.0, + 320.0, + 1104.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 521, + 1404, + 521, + 1404, + 887, + 298, + 887 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1140, + 1404, + 1140, + 1404, + 1355, + 298, + 1355 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 903, + 1403, + 903, + 1403, + 1055, + 299, + 1055 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 505, + 298, + 505 + ], + "score": 0.962 + }, + { + "category_id": 0, + "poly": [ + 299, + 1397, + 480, + 1397, + 480, + 1429, + 299, + 1429 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 299, + 1091, + 540, + 1091, + 540, + 1121, + 299, + 1121 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 826, + 76, + 826, + 104, + 300, + 104 + ], + "score": 0.843 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 298, + 1448, + 1401, + 1448, + 1401, + 1513, + 298, + 1513 + ], + "score": 0.676 + }, + { + "category_id": 1, + "poly": [ + 295, + 1535, + 1399, + 1535, + 1399, + 1598, + 295, + 1598 + ], + "score": 0.52 + }, + { + "category_id": 1, + "poly": [ + 297, + 1620, + 1401, + 1620, + 1401, + 1685, + 297, + 1685 + ], + "score": 0.435 + }, + { + "category_id": 1, + "poly": [ + 297, + 1707, + 1399, + 1707, + 1399, + 1800, + 297, + 1800 + ], + "score": 0.429 + }, + { + "category_id": 1, + "poly": [ + 289, + 1824, + 1403, + 1824, + 1403, + 1888, + 289, + 1888 + ], + "score": 0.402 + }, + { + "category_id": 1, + "poly": [ + 298, + 1910, + 1406, + 1910, + 1406, + 2032, + 298, + 2032 + ], + "score": 0.317 + }, + { + "category_id": 2, + "poly": [ + 301, + 76, + 827, + 76, + 827, + 104, + 301, + 104 + ], + "score": 0.193 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1393.0, + 483.0, + 1393.0, + 483.0, + 1435.0, + 296.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1089.0, + 545.0, + 1089.0, + 545.0, + 1127.0, + 294.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 830.0, + 73.0, + 830.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 297.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 552.0, + 1404.0, + 552.0, + 1404.0, + 583.0, + 297.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 583.0, + 1404.0, + 583.0, + 1404.0, + 617.0, + 296.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 1404.0, + 613.0, + 1404.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 641.0, + 1406.0, + 641.0, + 1406.0, + 677.0, + 293.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 672.0, + 1405.0, + 672.0, + 1405.0, + 709.0, + 293.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 699.0, + 1404.0, + 699.0, + 1404.0, + 743.0, + 292.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 732.0, + 1405.0, + 732.0, + 1405.0, + 771.0, + 292.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 765.0, + 1404.0, + 765.0, + 1404.0, + 797.0, + 295.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 796.0, + 1408.0, + 796.0, + 1408.0, + 830.0, + 293.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 824.0, + 1405.0, + 824.0, + 1405.0, + 863.0, + 292.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 856.0, + 499.0, + 856.0, + 499.0, + 893.0, + 294.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1138.0, + 1408.0, + 1138.0, + 1408.0, + 1176.0, + 296.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1170.0, + 1405.0, + 1170.0, + 1405.0, + 1207.0, + 293.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1201.0, + 1404.0, + 1201.0, + 1404.0, + 1235.0, + 294.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1230.0, + 1405.0, + 1230.0, + 1405.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1262.0, + 1406.0, + 1262.0, + 1406.0, + 1297.0, + 296.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1293.0, + 1406.0, + 1293.0, + 1406.0, + 1326.0, + 293.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1325.0, + 953.0, + 1325.0, + 953.0, + 1356.0, + 296.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 902.0, + 1405.0, + 902.0, + 1405.0, + 938.0, + 295.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 933.0, + 1405.0, + 933.0, + 1405.0, + 967.0, + 294.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 963.0, + 1403.0, + 963.0, + 1403.0, + 996.0, + 295.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 993.0, + 1405.0, + 993.0, + 1405.0, + 1030.0, + 295.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1025.0, + 777.0, + 1025.0, + 777.0, + 1062.0, + 295.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 290.0, + 1406.0, + 290.0, + 1406.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 319.0, + 1407.0, + 319.0, + 1407.0, + 359.0, + 293.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 354.0, + 1404.0, + 354.0, + 1404.0, + 387.0, + 296.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 383.0, + 1406.0, + 383.0, + 1406.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 411.0, + 1406.0, + 411.0, + 1406.0, + 452.0, + 292.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 443.0, + 1406.0, + 443.0, + 1406.0, + 479.0, + 294.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 473.0, + 1347.0, + 473.0, + 1347.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1447.0, + 1405.0, + 1447.0, + 1405.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1481.0, + 1363.0, + 1481.0, + 1363.0, + 1514.0, + 323.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1535.0, + 1403.0, + 1535.0, + 1403.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1567.0, + 970.0, + 1567.0, + 970.0, + 1600.0, + 323.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1622.0, + 1405.0, + 1622.0, + 1405.0, + 1658.0, + 296.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1653.0, + 1145.0, + 1653.0, + 1145.0, + 1686.0, + 322.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 1403.0, + 1708.0, + 1403.0, + 1741.0, + 295.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1739.0, + 1401.0, + 1739.0, + 1401.0, + 1773.0, + 323.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1770.0, + 631.0, + 1770.0, + 631.0, + 1802.0, + 322.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1864.0, + 291.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1855.0, + 996.0, + 1855.0, + 996.0, + 1889.0, + 322.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1407.0, + 1912.0, + 1407.0, + 1944.0, + 293.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1943.0, + 1408.0, + 1943.0, + 1408.0, + 1976.0, + 321.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2008.0, + 322.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 496.0, + 2003.0, + 496.0, + 2032.0, + 321.0, + 2032.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1069, + 1400, + 1069, + 1400, + 1134, + 296, + 1134 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 298, + 1910, + 1404, + 1910, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 297, + 1459, + 1406, + 1459, + 1406, + 1584, + 297, + 1584 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 298, + 454, + 1405, + 454, + 1405, + 549, + 298, + 549 + ], + "score": 0.867 + }, + { + "category_id": 1, + "poly": [ + 297, + 567, + 1406, + 567, + 1406, + 662, + 297, + 662 + ], + "score": 0.86 + }, + { + "category_id": 1, + "poly": [ + 296, + 986, + 1401, + 986, + 1401, + 1050, + 296, + 1050 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 298, + 310, + 1405, + 310, + 1405, + 436, + 298, + 436 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 296, + 1797, + 1404, + 1797, + 1404, + 1891, + 296, + 1891 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 296, + 1601, + 1402, + 1601, + 1402, + 1696, + 296, + 1696 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 297, + 1316, + 1408, + 1316, + 1408, + 1442, + 297, + 1442 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 297, + 904, + 1401, + 904, + 1401, + 970, + 297, + 970 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 291, + 1715, + 1403, + 1715, + 1403, + 1780, + 291, + 1780 + ], + "score": 0.846 + }, + { + "category_id": 1, + "poly": [ + 296, + 761, + 1406, + 761, + 1406, + 887, + 296, + 887 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 295, + 680, + 1401, + 680, + 1401, + 744, + 295, + 744 + ], + "score": 0.825 + }, + { + "category_id": 1, + "poly": [ + 297, + 1151, + 1174, + 1151, + 1174, + 1187, + 297, + 1187 + ], + "score": 0.821 + }, + { + "category_id": 1, + "poly": [ + 297, + 1203, + 1408, + 1203, + 1408, + 1297, + 297, + 1297 + ], + "score": 0.818 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 827, + 75, + 827, + 105, + 300, + 105 + ], + "score": 0.817 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 862, + 2088, + 862, + 2113, + 836, + 2113 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 291, + 228, + 1400, + 228, + 1400, + 293, + 291, + 293 + ], + "score": 0.645 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 828, + 75, + 828, + 105, + 300, + 105 + ], + "score": 0.313 + }, + { + "category_id": 13, + "poly": [ + 821, + 1724, + 838, + 1724, + 838, + 1742, + 821, + 1742 + ], + "score": 0.33, + "latex": "\\therefore" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1068.0, + 1404.0, + 1068.0, + 1404.0, + 1108.0, + 295.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1100.0, + 1382.0, + 1100.0, + 1382.0, + 1136.0, + 322.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1908.0, + 1404.0, + 1908.0, + 1404.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 323.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 321.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2002.0, + 738.0, + 2002.0, + 738.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1459.0, + 1404.0, + 1459.0, + 1404.0, + 1492.0, + 295.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1490.0, + 1404.0, + 1490.0, + 1404.0, + 1525.0, + 323.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1520.0, + 1406.0, + 1520.0, + 1406.0, + 1558.0, + 320.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1553.0, + 675.0, + 1553.0, + 675.0, + 1586.0, + 323.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 453.0, + 1407.0, + 453.0, + 1407.0, + 492.0, + 292.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 487.0, + 1407.0, + 487.0, + 1407.0, + 521.0, + 322.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 515.0, + 1293.0, + 515.0, + 1293.0, + 551.0, + 321.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 565.0, + 1406.0, + 565.0, + 1406.0, + 605.0, + 291.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 600.0, + 1404.0, + 600.0, + 1404.0, + 634.0, + 324.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 629.0, + 778.0, + 629.0, + 778.0, + 663.0, + 324.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 983.0, + 1406.0, + 983.0, + 1406.0, + 1024.0, + 291.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1014.0, + 394.0, + 1014.0, + 394.0, + 1054.0, + 319.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 313.0, + 1405.0, + 313.0, + 1405.0, + 346.0, + 296.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 344.0, + 1405.0, + 344.0, + 1405.0, + 378.0, + 322.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 374.0, + 1406.0, + 374.0, + 1406.0, + 408.0, + 320.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 405.0, + 614.0, + 405.0, + 614.0, + 439.0, + 322.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1795.0, + 1405.0, + 1795.0, + 1405.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1829.0, + 1407.0, + 1829.0, + 1407.0, + 1863.0, + 322.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1857.0, + 393.0, + 1857.0, + 393.0, + 1891.0, + 322.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1603.0, + 1403.0, + 1603.0, + 1403.0, + 1638.0, + 296.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1629.0, + 1409.0, + 1629.0, + 1409.0, + 1675.0, + 319.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1663.0, + 568.0, + 1663.0, + 568.0, + 1697.0, + 324.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1315.0, + 1409.0, + 1315.0, + 1409.0, + 1351.0, + 292.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1350.0, + 1409.0, + 1350.0, + 1409.0, + 1383.0, + 323.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1375.0, + 1410.0, + 1375.0, + 1410.0, + 1416.0, + 320.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1411.0, + 1069.0, + 1411.0, + 1069.0, + 1444.0, + 324.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 903.0, + 1403.0, + 903.0, + 1403.0, + 942.0, + 293.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 937.0, + 1189.0, + 937.0, + 1189.0, + 973.0, + 324.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1715.0, + 820.0, + 1715.0, + 820.0, + 1751.0, + 295.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1715.0, + 1409.0, + 1715.0, + 1409.0, + 1751.0, + 839.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1748.0, + 1168.0, + 1748.0, + 1168.0, + 1778.0, + 323.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 763.0, + 1407.0, + 763.0, + 1407.0, + 796.0, + 294.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 795.0, + 1404.0, + 795.0, + 1404.0, + 828.0, + 322.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 824.0, + 1404.0, + 824.0, + 1404.0, + 859.0, + 320.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 856.0, + 736.0, + 856.0, + 736.0, + 889.0, + 322.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 677.0, + 1407.0, + 677.0, + 1407.0, + 718.0, + 292.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 711.0, + 1142.0, + 711.0, + 1142.0, + 745.0, + 322.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1150.0, + 1172.0, + 1150.0, + 1172.0, + 1190.0, + 297.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1201.0, + 1406.0, + 1201.0, + 1406.0, + 1243.0, + 291.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1237.0, + 1410.0, + 1237.0, + 1410.0, + 1271.0, + 323.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1266.0, + 958.0, + 1266.0, + 958.0, + 1300.0, + 324.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 228.0, + 1403.0, + 228.0, + 1403.0, + 264.0, + 298.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 259.0, + 790.0, + 259.0, + 790.0, + 294.0, + 323.0, + 294.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 827, + 75, + 827, + 105, + 299, + 105 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2113, + 837, + 2113 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 297, + 796, + 1404, + 796, + 1404, + 860, + 297, + 860 + ], + "score": 0.743 + }, + { + "category_id": 1, + "poly": [ + 296, + 1395, + 1402, + 1395, + 1402, + 1460, + 296, + 1460 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 296, + 881, + 1402, + 881, + 1402, + 946, + 296, + 946 + ], + "score": 0.72 + }, + { + "category_id": 1, + "poly": [ + 296, + 1595, + 1405, + 1595, + 1405, + 1750, + 296, + 1750 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 296, + 1480, + 1406, + 1480, + 1406, + 1574, + 296, + 1574 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 295, + 1311, + 1400, + 1311, + 1400, + 1376, + 295, + 1376 + ], + "score": 0.688 + }, + { + "category_id": 1, + "poly": [ + 296, + 597, + 1400, + 597, + 1400, + 662, + 296, + 662 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 296, + 1080, + 1405, + 1080, + 1405, + 1206, + 296, + 1206 + ], + "score": 0.683 + }, + { + "category_id": 1, + "poly": [ + 296, + 1226, + 1400, + 1226, + 1400, + 1291, + 296, + 1291 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 297, + 965, + 1402, + 965, + 1402, + 1062, + 297, + 1062 + ], + "score": 0.669 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1400, + 1971, + 1400, + 2035, + 298, + 2035 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1403, + 229, + 1403, + 294, + 293, + 294 + ], + "score": 0.633 + }, + { + "category_id": 1, + "poly": [ + 295, + 681, + 1407, + 681, + 1407, + 776, + 295, + 776 + ], + "score": 0.628 + }, + { + "category_id": 1, + "poly": [ + 299, + 1886, + 1399, + 1886, + 1399, + 1951, + 299, + 1951 + ], + "score": 0.608 + }, + { + "category_id": 1, + "poly": [ + 296, + 512, + 1401, + 512, + 1401, + 577, + 296, + 577 + ], + "score": 0.595 + }, + { + "category_id": 1, + "poly": [ + 301, + 1771, + 1400, + 1771, + 1400, + 1865, + 301, + 1865 + ], + "score": 0.581 + }, + { + "category_id": 1, + "poly": [ + 301, + 397, + 1401, + 397, + 1401, + 493, + 301, + 493 + ], + "score": 0.569 + }, + { + "category_id": 1, + "poly": [ + 293, + 313, + 1404, + 313, + 1404, + 378, + 293, + 378 + ], + "score": 0.552 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 828, + 75, + 828, + 105, + 300, + 105 + ], + "score": 0.166 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 795.0, + 1406.0, + 795.0, + 1406.0, + 833.0, + 292.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 824.0, + 395.0, + 824.0, + 395.0, + 864.0, + 320.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1395.0, + 1406.0, + 1395.0, + 1406.0, + 1431.0, + 295.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1427.0, + 1088.0, + 1427.0, + 1088.0, + 1460.0, + 325.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 881.0, + 1404.0, + 881.0, + 1404.0, + 917.0, + 294.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 911.0, + 551.0, + 911.0, + 551.0, + 948.0, + 323.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 1407.0, + 1593.0, + 1407.0, + 1633.0, + 295.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1623.0, + 574.0, + 1623.0, + 574.0, + 1662.0, + 323.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1626.0, + 1406.0, + 1626.0, + 1406.0, + 1662.0, + 606.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1691.0, + 322.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1686.0, + 1407.0, + 1686.0, + 1407.0, + 1724.0, + 323.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1714.0, + 1157.0, + 1714.0, + 1157.0, + 1756.0, + 322.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1481.0, + 1408.0, + 1481.0, + 1408.0, + 1517.0, + 293.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1512.0, + 1406.0, + 1512.0, + 1406.0, + 1546.0, + 323.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1543.0, + 831.0, + 1543.0, + 831.0, + 1577.0, + 327.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1310.0, + 1402.0, + 1310.0, + 1402.0, + 1349.0, + 295.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1342.0, + 1081.0, + 1342.0, + 1081.0, + 1377.0, + 322.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 599.0, + 1402.0, + 599.0, + 1402.0, + 635.0, + 295.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 631.0, + 946.0, + 631.0, + 946.0, + 663.0, + 321.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1079.0, + 1406.0, + 1079.0, + 1406.0, + 1117.0, + 293.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1112.0, + 1406.0, + 1112.0, + 1406.0, + 1148.0, + 322.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1141.0, + 1406.0, + 1141.0, + 1406.0, + 1180.0, + 320.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1175.0, + 492.0, + 1175.0, + 492.0, + 1209.0, + 323.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 1404.0, + 1228.0, + 1404.0, + 1264.0, + 294.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1257.0, + 1025.0, + 1257.0, + 1025.0, + 1292.0, + 321.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 963.0, + 1406.0, + 963.0, + 1406.0, + 1002.0, + 292.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 999.0, + 1406.0, + 999.0, + 1406.0, + 1033.0, + 324.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1028.0, + 990.0, + 1028.0, + 990.0, + 1063.0, + 324.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 1283.0, + 2003.0, + 1283.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1403.0, + 228.0, + 1403.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 260.0, + 881.0, + 260.0, + 881.0, + 296.0, + 324.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 677.0, + 1405.0, + 677.0, + 1405.0, + 722.0, + 292.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 711.0, + 1410.0, + 711.0, + 1410.0, + 750.0, + 321.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 739.0, + 400.0, + 739.0, + 400.0, + 779.0, + 319.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1886.0, + 1404.0, + 1886.0, + 1404.0, + 1922.0, + 295.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1919.0, + 1313.0, + 1919.0, + 1313.0, + 1952.0, + 323.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 508.0, + 1407.0, + 508.0, + 1407.0, + 551.0, + 293.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 544.0, + 496.0, + 544.0, + 496.0, + 578.0, + 324.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1808.0, + 295.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1839.0, + 323.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1834.0, + 684.0, + 1834.0, + 684.0, + 1868.0, + 323.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 395.0, + 1406.0, + 395.0, + 1406.0, + 438.0, + 295.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 429.0, + 1405.0, + 429.0, + 1405.0, + 466.0, + 322.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 461.0, + 1055.0, + 461.0, + 1055.0, + 495.0, + 325.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 311.0, + 1407.0, + 311.0, + 1407.0, + 353.0, + 292.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 344.0, + 1189.0, + 344.0, + 1189.0, + 379.0, + 322.0, + 379.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 228, + 914, + 228, + 914, + 263, + 297, + 263 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 827, + 74, + 827, + 105, + 299, + 105 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 297, + 364, + 1072, + 364, + 1072, + 398, + 297, + 398 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 296, + 282, + 1402, + 282, + 1402, + 345, + 296, + 345 + ], + "score": 0.81 + }, + { + "category_id": 1, + "poly": [ + 295, + 417, + 1402, + 417, + 1402, + 481, + 295, + 481 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 302, + 501, + 1404, + 501, + 1404, + 593, + 302, + 593 + ], + "score": 0.785 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 831.0, + 72.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2123.0, + 833.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 914.0, + 226.0, + 914.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 364.0, + 1072.0, + 364.0, + 1072.0, + 400.0, + 294.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 280.0, + 1406.0, + 280.0, + 1406.0, + 319.0, + 295.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 312.0, + 1397.0, + 312.0, + 1397.0, + 348.0, + 324.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 417.0, + 1404.0, + 417.0, + 1404.0, + 453.0, + 294.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 448.0, + 871.0, + 448.0, + 871.0, + 484.0, + 323.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 499.0, + 1408.0, + 499.0, + 1408.0, + 537.0, + 296.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 529.0, + 1408.0, + 529.0, + 1408.0, + 570.0, + 320.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 557.0, + 395.0, + 557.0, + 395.0, + 598.0, + 320.0, + 598.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1590, + 824, + 1590, + 824, + 1866, + 299, + 1866 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 322, + 1405, + 322, + 1405, + 538, + 297, + 538 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1331, + 826, + 1331, + 826, + 1573, + 299, + 1573 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 314, + 728, + 1394, + 728, + 1394, + 1154, + 314, + 1154 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1881, + 1405, + 1881, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 551, + 1404, + 551, + 1404, + 704, + 296, + 704 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 846, + 1752, + 1402, + 1752, + 1402, + 1846, + 846, + 1846 + ], + "score": 0.955 + }, + { + "category_id": 4, + "poly": [ + 298, + 1179, + 1406, + 1179, + 1406, + 1243, + 298, + 1243 + ], + "score": 0.945 + }, + { + "category_id": 3, + "poly": [ + 856, + 1358, + 1385, + 1358, + 1385, + 1728, + 856, + 1728 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 827, + 75, + 827, + 104, + 300, + 104 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 298, + 227, + 494, + 227, + 494, + 261, + 298, + 261 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 300, + 279, + 741, + 279, + 741, + 310, + 300, + 310 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 837, + 2089, + 863, + 2089, + 863, + 2112, + 837, + 2112 + ], + "score": 0.855 + }, + { + "category_id": 0, + "poly": [ + 296, + 1287, + 856, + 1287, + 856, + 1318, + 296, + 1318 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 296, + 1287, + 856, + 1287, + 856, + 1318, + 296, + 1318 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 737, + 1912, + 788, + 1912, + 788, + 1941, + 737, + 1941 + ], + "score": 0.86, + "latex": "88 \\%" + }, + { + "category_id": 13, + "poly": [ + 1262, + 324, + 1300, + 324, + 1300, + 352, + 1262, + 352 + ], + "score": 0.84, + "latex": "+ 1" + }, + { + "category_id": 13, + "poly": [ + 783, + 1652, + 824, + 1652, + 824, + 1681, + 783, + 1681 + ], + "score": 0.83, + "latex": "0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1244, + 507, + 1295, + 507, + 1295, + 535, + 1244, + 535 + ], + "score": 0.7, + "latex": "- 5 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 1775, + 379, + 1775, + 379, + 1803, + 297, + 1803 + ], + "score": 0.33, + "latex": "\\Upsilon \\mathrm { K z x I }" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 732.0, + 375.0, + 732.0, + 375.0, + 760.0, + 334.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 734.0, + 914.0, + 734.0, + 914.0, + 760.0, + 874.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 753.0, + 1038.0, + 753.0, + 1038.0, + 778.0, + 968.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 768.0, + 913.0, + 768.0, + 913.0, + 791.0, + 875.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 774.0, + 1037.0, + 774.0, + 1037.0, + 799.0, + 968.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 793.0, + 375.0, + 793.0, + 375.0, + 820.0, + 334.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 799.0, + 913.0, + 799.0, + 913.0, + 826.0, + 874.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 793.0, + 1054.0, + 793.0, + 1054.0, + 817.0, + 968.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 827.0, + 343.0, + 827.0, + 343.0, + 966.0, + 311.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 824.0, + 912.0, + 824.0, + 912.0, + 967.0, + 846.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 855.0, + 372.0, + 855.0, + 372.0, + 878.0, + 337.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 866.0, + 913.0, + 866.0, + 913.0, + 893.0, + 872.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 902.0, + 912.0, + 902.0, + 912.0, + 925.0, + 875.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 914.0, + 372.0, + 914.0, + 372.0, + 937.0, + 337.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 935.0, + 912.0, + 935.0, + 912.0, + 957.0, + 875.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 971.0, + 377.0, + 971.0, + 377.0, + 1000.0, + 341.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 978.0, + 812.0, + 978.0, + 812.0, + 999.0, + 746.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 966.0, + 913.0, + 966.0, + 913.0, + 993.0, + 874.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 994.0, + 813.0, + 994.0, + 813.0, + 1022.0, + 744.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1000.0, + 913.0, + 1000.0, + 913.0, + 1026.0, + 882.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1017.0, + 831.0, + 1017.0, + 831.0, + 1041.0, + 745.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1036.0, + 372.0, + 1036.0, + 372.0, + 1056.0, + 353.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1035.0, + 914.0, + 1035.0, + 914.0, + 1059.0, + 891.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1050.0, + 729.0, + 1050.0, + 729.0, + 1060.0, + 719.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1050.0, + 1156.0, + 1050.0, + 1156.0, + 1060.0, + 1147.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1052.0, + 1267.0, + 1052.0, + 1267.0, + 1059.0, + 1261.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1064.0, + 394.0, + 1064.0, + 394.0, + 1081.0, + 378.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1064.0, + 507.0, + 1064.0, + 507.0, + 1081.0, + 491.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1061.0, + 622.0, + 1061.0, + 622.0, + 1084.0, + 602.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1061.0, + 735.0, + 1061.0, + 735.0, + 1084.0, + 714.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1062.0, + 847.0, + 1062.0, + 847.0, + 1083.0, + 827.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1062.0, + 935.0, + 1062.0, + 935.0, + 1083.0, + 915.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1062.0, + 1048.0, + 1062.0, + 1048.0, + 1084.0, + 1030.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1061.0, + 1160.0, + 1061.0, + 1160.0, + 1083.0, + 1140.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1061.0, + 1273.0, + 1061.0, + 1273.0, + 1084.0, + 1253.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1065.0, + 1384.0, + 1065.0, + 1384.0, + 1081.0, + 1368.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1081.0, + 725.0, + 1081.0, + 725.0, + 1105.0, + 500.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1081.0, + 1263.0, + 1081.0, + 1263.0, + 1105.0, + 1038.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1119.0, + 643.0, + 1119.0, + 643.0, + 1161.0, + 517.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1121.0, + 1182.0, + 1121.0, + 1182.0, + 1157.0, + 1058.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1752.0, + 1404.0, + 1752.0, + 1404.0, + 1786.0, + 845.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1783.0, + 1403.0, + 1783.0, + 1403.0, + 1816.0, + 845.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1815.0, + 1276.0, + 1815.0, + 1276.0, + 1848.0, + 846.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1178.0, + 1404.0, + 1178.0, + 1404.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1210.0, + 1304.0, + 1210.0, + 1304.0, + 1246.0, + 295.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1458.0, + 1246.0, + 1458.0, + 1246.0, + 1510.0, + 1198.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 831.0, + 72.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 500.0, + 222.0, + 500.0, + 267.0, + 291.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 278.0, + 744.0, + 278.0, + 744.0, + 314.0, + 296.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 829.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1286.0, + 858.0, + 1286.0, + 858.0, + 1322.0, + 296.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1590.0, + 827.0, + 1590.0, + 827.0, + 1624.0, + 296.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1619.0, + 828.0, + 1619.0, + 828.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 782.0, + 1652.0, + 782.0, + 1684.0, + 295.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1682.0, + 827.0, + 1682.0, + 827.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1713.0, + 827.0, + 1713.0, + 827.0, + 1744.0, + 296.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 824.0, + 1742.0, + 824.0, + 1774.0, + 295.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1771.0, + 827.0, + 1771.0, + 827.0, + 1807.0, + 380.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1803.0, + 825.0, + 1803.0, + 825.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1831.0, + 411.0, + 1831.0, + 411.0, + 1873.0, + 295.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 319.0, + 1261.0, + 319.0, + 1261.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 360.0, + 1301.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 388.0, + 296.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 383.0, + 1408.0, + 383.0, + 1408.0, + 420.0, + 291.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 411.0, + 1408.0, + 411.0, + 1408.0, + 451.0, + 292.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 1406.0, + 444.0, + 1406.0, + 481.0, + 292.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 476.0, + 1405.0, + 476.0, + 1405.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1243.0, + 505.0, + 1243.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 505.0, + 1382.0, + 505.0, + 1382.0, + 540.0, + 1296.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1331.0, + 828.0, + 1331.0, + 828.0, + 1364.0, + 296.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1360.0, + 829.0, + 1360.0, + 829.0, + 1393.0, + 295.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1391.0, + 830.0, + 1391.0, + 830.0, + 1424.0, + 296.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 828.0, + 1419.0, + 828.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1451.0, + 827.0, + 1451.0, + 827.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1483.0, + 826.0, + 1483.0, + 826.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 826.0, + 1514.0, + 826.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1543.0, + 694.0, + 1543.0, + 694.0, + 1574.0, + 296.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 828.0, + 1882.0, + 828.0, + 1913.0, + 294.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 736.0, + 1910.0, + 736.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 789.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 1407.0, + 1943.0, + 1407.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1969.0, + 1407.0, + 1969.0, + 1407.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1107.0, + 2002.0, + 1107.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 1404.0, + 552.0, + 1404.0, + 585.0, + 295.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 581.0, + 1405.0, + 581.0, + 1405.0, + 618.0, + 295.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 610.0, + 1405.0, + 610.0, + 1405.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 638.0, + 1405.0, + 638.0, + 1405.0, + 683.0, + 291.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 561.0, + 672.0, + 561.0, + 709.0, + 294.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1286.0, + 858.0, + 1286.0, + 858.0, + 1322.0, + 296.0, + 1322.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1432, + 1404, + 1432, + 1404, + 1619, + 297, + 1619 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1702, + 1405, + 1702, + 1405, + 1888, + 298, + 1888 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1405, + 228, + 1405, + 383, + 298, + 383 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 328, + 413, + 823, + 413, + 823, + 768, + 328, + 768 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 301, + 1325, + 1399, + 1325, + 1399, + 1419, + 301, + 1419 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 298, + 1028, + 1404, + 1028, + 1404, + 1122, + 298, + 1122 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 300, + 1136, + 1402, + 1136, + 1402, + 1231, + 300, + 1231 + ], + "score": 0.961 + }, + { + "category_id": 3, + "poly": [ + 878, + 426, + 1376, + 426, + 1376, + 780, + 878, + 780 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 509, + 1240, + 1186, + 1240, + 1186, + 1317, + 509, + 1317 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 300, + 1971, + 1402, + 1971, + 1402, + 2035, + 300, + 2035 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 297, + 1651, + 943, + 1651, + 943, + 1682, + 297, + 1682 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 300, + 1919, + 823, + 1919, + 823, + 1950, + 300, + 1950 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 827, + 75, + 827, + 104, + 299, + 104 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 299, + 985, + 672, + 985, + 672, + 1014, + 299, + 1014 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1368, + 1265, + 1400, + 1265, + 1400, + 1294, + 1368, + 1294 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.857 + }, + { + "category_id": 4, + "poly": [ + 876, + 796, + 1376, + 796, + 1376, + 827, + 876, + 827 + ], + "score": 0.846 + }, + { + "category_id": 1, + "poly": [ + 295, + 787, + 853, + 787, + 853, + 843, + 295, + 843 + ], + "score": 0.672 + }, + { + "category_id": 4, + "poly": [ + 292, + 869, + 1400, + 869, + 1400, + 932, + 292, + 932 + ], + "score": 0.623 + }, + { + "category_id": 1, + "poly": [ + 292, + 869, + 1400, + 869, + 1400, + 932, + 292, + 932 + ], + "score": 0.435 + }, + { + "category_id": 14, + "poly": [ + 512, + 1240, + 1186, + 1240, + 1186, + 1317, + 512, + 1317 + ], + "score": 0.93, + "latex": "p m i ( a ^ { A } ; a ^ { B } | Z { = } z ) { = } \\log \\frac { p ( a ^ { B } | a ^ { A } , z ) } { p ( a ^ { B } | z ) } { = } \\mathrm { l o g } \\frac { p ( a ^ { A } , a ^ { B } | z ) } { p ( a ^ { A } | z ) p ( a ^ { B } | z ) } ." + }, + { + "category_id": 13, + "poly": [ + 996, + 1355, + 1125, + 1355, + 1125, + 1390, + 996, + 1390 + ], + "score": 0.91, + "latex": "p \\dot { ( } a ^ { B } , a ^ { A } | z )" + }, + { + "category_id": 13, + "poly": [ + 559, + 1092, + 638, + 1092, + 638, + 1122, + 559, + 1122 + ], + "score": 0.9, + "latex": "\\gamma = . 9 9" + }, + { + "category_id": 13, + "poly": [ + 887, + 1092, + 914, + 1092, + 914, + 1117, + 887, + 1117 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1072, + 1391, + 1095, + 1391, + 1095, + 1414, + 1072, + 1414 + ], + "score": 0.83, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1736, + 1240, + 1736, + 1240, + 1761, + 1216, + 1761 + ], + "score": 0.82, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1209, + 1706, + 1229, + 1706, + 1229, + 1735, + 1209, + 1735 + ], + "score": 0.81, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 738, + 1388, + 762, + 1388, + 762, + 1414, + 738, + 1414 + ], + "score": 0.81, + "latex": "A" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 420.0, + 353.0, + 420.0, + 353.0, + 722.0, + 329.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 452.0, + 374.0, + 452.0, + 374.0, + 469.0, + 354.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 503.0, + 377.0, + 503.0, + 377.0, + 528.0, + 352.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 556.0, + 377.0, + 556.0, + 377.0, + 581.0, + 352.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 614.0, + 375.0, + 614.0, + 375.0, + 635.0, + 352.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 672.0, + 374.0, + 672.0, + 374.0, + 686.0, + 361.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 672.0, + 472.0, + 672.0, + 472.0, + 679.0, + 464.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 669.0, + 486.0, + 669.0, + 486.0, + 678.0, + 478.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 673.0, + 538.0, + 673.0, + 538.0, + 680.0, + 508.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 671.0, + 809.0, + 671.0, + 809.0, + 682.0, + 790.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 726.0, + 397.0, + 726.0, + 397.0, + 742.0, + 380.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 729.0, + 441.0, + 729.0, + 441.0, + 751.0, + 421.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 729.0, + 484.0, + 729.0, + 484.0, + 750.0, + 462.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 725.0, + 527.0, + 725.0, + 527.0, + 750.0, + 503.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 727.0, + 568.0, + 727.0, + 568.0, + 749.0, + 546.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 727.0, + 612.0, + 727.0, + 612.0, + 751.0, + 587.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 726.0, + 652.0, + 726.0, + 652.0, + 751.0, + 627.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 726.0, + 695.0, + 726.0, + 695.0, + 751.0, + 669.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 726.0, + 736.0, + 726.0, + 736.0, + 751.0, + 713.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 726.0, + 778.0, + 726.0, + 778.0, + 752.0, + 755.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 745.0, + 657.0, + 745.0, + 657.0, + 770.0, + 543.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 665.5, + 464.0, + 665.5, + 464.0, + 681.0, + 439.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 439.0, + 950.0, + 439.0, + 950.0, + 466.0, + 896.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 428.0, + 1294.0, + 428.0, + 1294.0, + 456.0, + 1014.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 458.0, + 1353.0, + 458.0, + 1353.0, + 482.0, + 1291.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 476.0, + 987.0, + 476.0, + 987.0, + 499.0, + 957.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 492.0, + 989.0, + 492.0, + 989.0, + 499.0, + 979.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 494.0, + 950.0, + 494.0, + 950.0, + 521.0, + 896.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 517.0, + 1207.0, + 517.0, + 1207.0, + 549.0, + 1025.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 549.0, + 940.0, + 549.0, + 940.0, + 574.0, + 896.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 555.0, + 901.0, + 555.0, + 901.0, + 630.0, + 877.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 605.0, + 937.0, + 605.0, + 937.0, + 629.0, + 896.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 660.0, + 940.0, + 660.0, + 940.0, + 685.0, + 897.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 711.0, + 939.0, + 711.0, + 939.0, + 740.0, + 897.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 731.0, + 1034.0, + 731.0, + 1034.0, + 739.0, + 1027.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 731.0, + 1281.0, + 731.0, + 1281.0, + 740.0, + 1273.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 741.0, + 959.0, + 741.0, + 959.0, + 762.0, + 939.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 739.0, + 1044.0, + 739.0, + 1044.0, + 763.0, + 1017.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 739.0, + 1126.0, + 739.0, + 1126.0, + 763.0, + 1099.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 739.0, + 1210.0, + 739.0, + 1210.0, + 764.0, + 1182.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 739.0, + 1290.0, + 739.0, + 1290.0, + 763.0, + 1264.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 738.0, + 1377.0, + 738.0, + 1377.0, + 764.0, + 1341.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 755.0, + 1212.0, + 755.0, + 1212.0, + 785.0, + 1096.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.75, + 688.0, + 976.75, + 688.0, + 976.75, + 708.0, + 968.75, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1649.0, + 946.0, + 1649.0, + 946.0, + 1684.0, + 296.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 829.0, + 1918.0, + 829.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 831.0, + 72.0, + 831.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 981.0, + 676.0, + 981.0, + 676.0, + 1019.0, + 294.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 792.0, + 1380.0, + 792.0, + 1380.0, + 832.0, + 869.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 899.0, + 884.0, + 899.0, + 884.0, + 935.0, + 295.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1431.0, + 1405.0, + 1431.0, + 1405.0, + 1469.0, + 294.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1463.0, + 1405.0, + 1463.0, + 1405.0, + 1500.0, + 294.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1493.0, + 1406.0, + 1493.0, + 1406.0, + 1528.0, + 294.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1526.0, + 1405.0, + 1526.0, + 1405.0, + 1558.0, + 294.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1555.0, + 1406.0, + 1555.0, + 1406.0, + 1590.0, + 292.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1587.0, + 617.0, + 1587.0, + 617.0, + 1619.0, + 296.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1701.0, + 1208.0, + 1701.0, + 1208.0, + 1736.0, + 293.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1701.0, + 1406.0, + 1701.0, + 1406.0, + 1736.0, + 1230.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1729.0, + 1215.0, + 1729.0, + 1215.0, + 1771.0, + 292.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1729.0, + 1409.0, + 1729.0, + 1409.0, + 1771.0, + 1241.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1766.0, + 1405.0, + 1766.0, + 1405.0, + 1798.0, + 295.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1794.0, + 1405.0, + 1794.0, + 1405.0, + 1828.0, + 293.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1824.0, + 1406.0, + 1824.0, + 1406.0, + 1860.0, + 295.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1856.0, + 833.0, + 1856.0, + 833.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1403.0, + 231.0, + 1403.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 298.0, + 295.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 1407.0, + 291.0, + 1407.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 319.0, + 1407.0, + 319.0, + 1407.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 349.0, + 400.0, + 349.0, + 400.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1323.0, + 1405.0, + 1323.0, + 1405.0, + 1361.0, + 296.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1353.0, + 995.0, + 1353.0, + 995.0, + 1393.0, + 293.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1353.0, + 1407.0, + 1353.0, + 1407.0, + 1393.0, + 1126.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1387.0, + 737.0, + 1387.0, + 737.0, + 1421.0, + 296.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1387.0, + 1071.0, + 1387.0, + 1071.0, + 1421.0, + 763.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1387.0, + 1107.0, + 1387.0, + 1107.0, + 1421.0, + 1096.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1063.0, + 294.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1059.0, + 1407.0, + 1059.0, + 1407.0, + 1093.0, + 294.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1088.0, + 558.0, + 1088.0, + 558.0, + 1124.0, + 293.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1088.0, + 886.0, + 1088.0, + 886.0, + 1124.0, + 639.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1088.0, + 1056.0, + 1088.0, + 1056.0, + 1124.0, + 915.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1134.0, + 1406.0, + 1134.0, + 1406.0, + 1174.0, + 294.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 1406.0, + 1168.0, + 1406.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1196.0, + 946.0, + 1196.0, + 946.0, + 1236.0, + 294.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1970.0, + 1404.0, + 1970.0, + 1404.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 853.0, + 784.0, + 853.0, + 819.0, + 294.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 816.0, + 639.0, + 816.0, + 639.0, + 845.0, + 296.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 899.0, + 884.0, + 899.0, + 884.0, + 935.0, + 295.0, + 935.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 383, + 646, + 1316, + 646, + 1316, + 1071, + 383, + 1071 + ], + "score": 0.98, + "html": "
HyperparameterCleanupHarvest
A3C baselineComm. baselineInfluence comm.A3C Comm. baseline baselineInfluence comm.
Entropy reg..00176.000249.00305.000687.000174.00220
lr_init.00126.00223.00249.00136.00137.000413
lr_end.000012.000022.0000127.000028.0000127.000049
Influence weight β=02.752=04.825
Extrinsic reward01.0
weight α
Curriculum C1 KL8
Policy comparison.000789KL
Comm. entropy reg..0758.00208
Comm. loss weight9.0709
Symbol vocab size327
Comm. embedding16
" + }, + { + "category_id": 5, + "poly": [ + 328, + 222, + 1382, + 222, + 1382, + 537, + 328, + 537 + ], + "score": 0.979, + "html": "
HyperparameterCleanupHarvest
A3CVisibleactionsInfluenceA3C baselineVisibleactionsInfluence
Entropy reg.baseline .00176baseline .00176.000248.000687baseline .00184.00025
lr_init.00126.00126.00107.00136.00215.00107
lr_end.000012.000012.000042.000028.000013.000042
Number of influencers3133
Influence weight β0.1460.224
Curriculum C11401140
Policy comparisonJSDPMI
Influencee reward10
" + }, + { + "category_id": 5, + "poly": [ + 399, + 1181, + 1303, + 1181, + 1303, + 1522, + 399, + 1522 + ], + "score": 0.979, + "html": "
HyperparameterCleanupHarvest
A3C baselineMOA baselineInfluence MOAA3C baselineMOA baselineInfluence MOA
Entropy reg..00176.00176.00176.000687.00495.00223
lr_init.00126.00123.00123.00136.00206.00120
lr_end.000012.000012.000012.000028.000022.000044
Influence weight β0.620=02.521
MOA loss weight1.31215.007=1.71110.911
Curriculum C-40=1226
Policy comparison=KL==KL
Train MOA onlyFalseTrueFalseTrue
when visible
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1660, + 1403, + 1660, + 1403, + 1815, + 298, + 1815 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1402, + 1942, + 1402, + 2033, + 299, + 2033 + ], + "score": 0.972 + }, + { + "category_id": 6, + "poly": [ + 367, + 1091, + 1326, + 1091, + 1326, + 1124, + 367, + 1124 + ], + "score": 0.85 + }, + { + "category_id": 0, + "poly": [ + 298, + 1877, + 1017, + 1877, + 1017, + 1909, + 298, + 1909 + ], + "score": 0.847 + }, + { + "category_id": 2, + "poly": [ + 837, + 2089, + 864, + 2089, + 864, + 2112, + 837, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 828, + 76, + 828, + 104, + 300, + 104 + ], + "score": 0.831 + }, + { + "category_id": 6, + "poly": [ + 337, + 557, + 1356, + 557, + 1356, + 590, + 337, + 590 + ], + "score": 0.779 + }, + { + "category_id": 6, + "poly": [ + 296, + 1543, + 1401, + 1543, + 1401, + 1575, + 296, + 1575 + ], + "score": 0.353 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 827, + 76, + 827, + 104, + 300, + 104 + ], + "score": 0.245 + }, + { + "category_id": 13, + "poly": [ + 567, + 817, + 586, + 817, + 586, + 844, + 567, + 844 + ], + "score": 0.73, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 579, + 1351, + 598, + 1351, + 598, + 1379, + 579, + 1379 + ], + "score": 0.71, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 659, + 1759, + 680, + 1759, + 680, + 1782, + 659, + 1782 + ], + "score": 0.67, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 500, + 421, + 520, + 421, + 520, + 448, + 500, + 448 + ], + "score": 0.65, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 529, + 1406, + 553, + 1406, + 553, + 1430, + 529, + 1430 + ], + "score": 0.57, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 517, + 899, + 541, + 899, + 541, + 924, + 517, + 924 + ], + "score": 0.55, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 472, + 875, + 493, + 875, + 493, + 896, + 472, + 896 + ], + "score": 0.53, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 450, + 448, + 474, + 448, + 474, + 472, + 450, + 472 + ], + "score": 0.49, + "latex": "C" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1087.0, + 1332.0, + 1087.0, + 1332.0, + 1130.0, + 363.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1877.0, + 1021.0, + 1877.0, + 1021.0, + 1912.0, + 296.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 553.0, + 1362.0, + 553.0, + 1362.0, + 596.0, + 335.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1582.0, + 292.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 832.0, + 72.0, + 832.0, + 110.0, + 296.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1662.0, + 1404.0, + 1662.0, + 1404.0, + 1695.0, + 294.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1691.0, + 1404.0, + 1691.0, + 1404.0, + 1727.0, + 294.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1722.0, + 1404.0, + 1722.0, + 1404.0, + 1757.0, + 294.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1752.0, + 658.0, + 1752.0, + 658.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1752.0, + 1405.0, + 1752.0, + 1405.0, + 1789.0, + 681.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1785.0, + 611.0, + 1785.0, + 611.0, + 1815.0, + 297.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 759.0, + 2002.0, + 759.0, + 2035.0, + 294.0, + 2035.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 600, + 1275, + 1404, + 1275, + 1404, + 1490, + 600, + 1490 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1599, + 1404, + 1599, + 1404, + 1750, + 298, + 1750 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 846, + 1403, + 846, + 1403, + 940, + 300, + 940 + ], + "score": 0.973 + }, + { + "category_id": 5, + "poly": [ + 512, + 223, + 1190, + 223, + 1190, + 412, + 512, + 412 + ], + "score": 0.969, + "html": "
CleanupHarvest
A3Cbaseline89485
Inequity aversion (Hughes et al., 2018)275750
Influence - Basic1901073
Influence - Communication166951
Influence -Model of other agents392588
" + }, + { + "category_id": 1, + "poly": [ + 600, + 1014, + 1404, + 1014, + 1404, + 1262, + 600, + 1262 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1764, + 1402, + 1764, + 1402, + 1858, + 298, + 1858 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 599, + 1505, + 1403, + 1505, + 1403, + 1598, + 599, + 1598 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 708, + 1870, + 990, + 1870, + 990, + 1962, + 708, + 1962 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 295, + 1437, + 579, + 1437, + 579, + 1559, + 295, + 1559 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 512, + 746, + 1181, + 746, + 1181, + 835, + 512, + 835 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1404, + 1972, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 296, + 434, + 1407, + 434, + 1407, + 590, + 296, + 590 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 301, + 699, + 800, + 699, + 800, + 731, + 301, + 731 + ], + "score": 0.918 + }, + { + "category_id": 3, + "poly": [ + 300, + 1024, + 563, + 1024, + 563, + 1411, + 300, + 1411 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 299, + 646, + 708, + 646, + 708, + 677, + 299, + 677 + ], + "score": 0.91 + }, + { + "category_id": 9, + "poly": [ + 1368, + 776, + 1400, + 776, + 1400, + 805, + 1368, + 805 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1368, + 1903, + 1400, + 1903, + 1400, + 1932, + 1368, + 1932 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 837, + 2087, + 864, + 2087, + 864, + 2112, + 837, + 2112 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 828, + 75, + 828, + 105, + 299, + 105 + ], + "score": 0.783 + }, + { + "category_id": 0, + "poly": [ + 298, + 972, + 617, + 972, + 617, + 1000, + 298, + 1000 + ], + "score": 0.651 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 828, + 75, + 828, + 105, + 299, + 105 + ], + "score": 0.418 + }, + { + "category_id": 0, + "poly": [ + 299, + 972, + 618, + 972, + 618, + 1000, + 299, + 1000 + ], + "score": 0.416 + }, + { + "category_id": 3, + "poly": [ + 300, + 1024, + 563, + 1024, + 563, + 1412, + 300, + 1412 + ], + "score": 0.374 + }, + { + "category_id": 14, + "poly": [ + 708, + 1868, + 989, + 1868, + 989, + 1962, + 708, + 1962 + ], + "score": 0.95, + "latex": "G = \\frac { \\sum _ { i = 1 } ^ { N } \\sum _ { j = 1 } ^ { N } \\left| { r ^ { i } - r ^ { j } } \\right| } { 2 N \\sum _ { i = 1 } ^ { N } r ^ { i } } ," + }, + { + "category_id": 14, + "poly": [ + 512, + 741, + 1183, + 741, + 1183, + 838, + 512, + 838 + ], + "score": 0.94, + "latex": "\\sum _ { k = 1 } ^ { N } 0 . 5 [ \\sum _ { c } 1 - \\frac { H ( p ( a ^ { k } | m ^ { k } = c ) ) } { H _ { m a x } } + \\sum _ { a } 1 - \\frac { H ( p ( m ^ { k } | a ^ { k } = a ) ) } { H _ { m a x } } ] ," + }, + { + "category_id": 13, + "poly": [ + 702, + 848, + 771, + 848, + 771, + 878, + 702, + 878 + ], + "score": 0.92, + "latex": "H _ { m a x }" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1973, + 1208, + 1973, + 1208, + 2001, + 1143, + 2001 + ], + "score": 0.89, + "latex": "G = 0" + }, + { + "category_id": 13, + "poly": [ + 983, + 1972, + 1078, + 1972, + 1078, + 2005, + 983, + 2005 + ], + "score": 0.88, + "latex": "G \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 474, + 2004, + 543, + 2004, + 543, + 2032, + 474, + 2032 + ], + "score": 0.87, + "latex": "1 - G" + }, + { + "category_id": 13, + "poly": [ + 695, + 1828, + 723, + 1828, + 723, + 1854, + 695, + 1854 + ], + "score": 0.85, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 370, + 849, + 398, + 849, + 398, + 875, + 370, + 875 + ], + "score": 0.82, + "latex": "H" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1436.0, + 580.0, + 1436.0, + 580.0, + 1467.0, + 296.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1464.0, + 579.0, + 1464.0, + 579.0, + 1498.0, + 295.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1497.0, + 583.0, + 1497.0, + 583.0, + 1531.0, + 293.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1529.0, + 483.0, + 1529.0, + 483.0, + 1559.0, + 293.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1097.0, + 423.0, + 1097.0, + 423.0, + 1116.0, + 333.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1240.0, + 487.0, + 1240.0, + 487.0, + 1265.0, + 393.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 643.0, + 712.0, + 643.0, + 712.0, + 681.0, + 295.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 971.0, + 622.0, + 971.0, + 622.0, + 1003.0, + 296.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 971.0, + 621.0, + 971.0, + 621.0, + 1003.0, + 296.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1095.0, + 424.0, + 1095.0, + 424.0, + 1117.0, + 332.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1241.0, + 486.0, + 1241.0, + 486.0, + 1264.0, + 391.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1275.0, + 1405.0, + 1275.0, + 1405.0, + 1309.0, + 596.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1305.0, + 1408.0, + 1305.0, + 1408.0, + 1340.0, + 596.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1336.0, + 1405.0, + 1336.0, + 1405.0, + 1369.0, + 598.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1369.0, + 1404.0, + 1369.0, + 1404.0, + 1399.0, + 599.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1397.0, + 1405.0, + 1397.0, + 1405.0, + 1432.0, + 595.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1463.0, + 596.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1459.0, + 1040.0, + 1459.0, + 1040.0, + 1491.0, + 597.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 1406.0, + 1599.0, + 1406.0, + 1631.0, + 293.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1626.0, + 1406.0, + 1626.0, + 1406.0, + 1661.0, + 295.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1694.0, + 293.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1688.0, + 1405.0, + 1688.0, + 1405.0, + 1723.0, + 295.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1719.0, + 686.0, + 1719.0, + 686.0, + 1752.0, + 292.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 846.0, + 369.0, + 846.0, + 369.0, + 883.0, + 295.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 846.0, + 701.0, + 846.0, + 701.0, + 883.0, + 399.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 846.0, + 1405.0, + 846.0, + 1405.0, + 883.0, + 772.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 1407.0, + 877.0, + 1407.0, + 914.0, + 294.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 906.0, + 948.0, + 906.0, + 948.0, + 946.0, + 293.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1011.0, + 1405.0, + 1011.0, + 1405.0, + 1051.0, + 596.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1047.0, + 1405.0, + 1047.0, + 1405.0, + 1079.0, + 597.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1076.0, + 1405.0, + 1076.0, + 1405.0, + 1109.0, + 598.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1106.0, + 1404.0, + 1106.0, + 1404.0, + 1140.0, + 597.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1137.0, + 1405.0, + 1137.0, + 1405.0, + 1170.0, + 597.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1169.0, + 1406.0, + 1169.0, + 1406.0, + 1201.0, + 597.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1200.0, + 1403.0, + 1200.0, + 1403.0, + 1231.0, + 598.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1229.0, + 1216.0, + 1229.0, + 1216.0, + 1265.0, + 596.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1764.0, + 1407.0, + 1764.0, + 1407.0, + 1800.0, + 294.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1796.0, + 1404.0, + 1796.0, + 1404.0, + 1830.0, + 296.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1826.0, + 694.0, + 1826.0, + 694.0, + 1863.0, + 294.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1826.0, + 1007.0, + 1826.0, + 1007.0, + 1863.0, + 724.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1504.0, + 1405.0, + 1504.0, + 1405.0, + 1539.0, + 595.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1534.0, + 1409.0, + 1534.0, + 1409.0, + 1571.0, + 596.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1565.0, + 1407.0, + 1565.0, + 1407.0, + 1602.0, + 596.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 982.0, + 1972.0, + 982.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1972.0, + 1142.0, + 1972.0, + 1142.0, + 2007.0, + 1079.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2007.0, + 1209.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 473.0, + 2001.0, + 473.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2037.0, + 544.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 433.0, + 1407.0, + 433.0, + 1407.0, + 469.0, + 295.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 465.0, + 1403.0, + 465.0, + 1403.0, + 499.0, + 296.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 493.0, + 1405.0, + 493.0, + 1405.0, + 535.0, + 293.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 528.0, + 1405.0, + 528.0, + 1405.0, + 561.0, + 294.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 557.0, + 725.0, + 557.0, + 725.0, + 591.0, + 295.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 696.0, + 804.0, + 696.0, + 804.0, + 736.0, + 297.0, + 736.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 320, + 399, + 1393, + 399, + 1393, + 1641, + 320, + 1641 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1910, + 1406, + 1910, + 1406, + 2034, + 298, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 297, + 1664, + 1405, + 1664, + 1405, + 1786, + 297, + 1786 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 827, + 75, + 827, + 105, + 300, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 293, + 298, + 293 + ], + "score": 0.797 + }, + { + "category_id": 4, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 293, + 298, + 293 + ], + "score": 0.264 + }, + { + "category_id": 13, + "poly": [ + 811, + 1665, + 928, + 1665, + 928, + 1699, + 811, + 1699 + ], + "score": 0.93, + "latex": "R * ( 1 - G )" + }, + { + "category_id": 13, + "poly": [ + 315, + 1696, + 389, + 1696, + 389, + 1725, + 315, + 1725 + ], + "score": 0.91, + "latex": "9 9 . 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 423.0, + 394.0, + 423.0, + 394.0, + 448.0, + 348.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 423.0, + 952.0, + 423.0, + 952.0, + 448.0, + 907.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 444.0, + 611.0, + 444.0, + 611.0, + 478.0, + 468.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 442.0, + 1170.0, + 442.0, + 1170.0, + 478.0, + 1026.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 469.0, + 393.0, + 469.0, + 393.0, + 650.0, + 321.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 473.0, + 713.0, + 473.0, + 713.0, + 534.0, + 468.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 462.0, + 952.0, + 462.0, + 952.0, + 650.0, + 879.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 467.0, + 1276.0, + 467.0, + 1276.0, + 537.0, + 1021.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 550.0, + 951.0, + 550.0, + 951.0, + 575.0, + 907.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 561.0, + 391.0, + 561.0, + 391.0, + 587.0, + 350.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 591.0, + 951.0, + 591.0, + 951.0, + 620.0, + 906.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 608.0, + 391.0, + 608.0, + 391.0, + 634.0, + 359.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 622.0, + 489.0, + 622.0, + 489.0, + 658.0, + 435.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 636.0, + 950.0, + 636.0, + 950.0, + 661.0, + 906.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 655.0, + 394.0, + 655.0, + 394.0, + 679.0, + 372.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 675.0, + 952.0, + 675.0, + 952.0, + 705.0, + 904.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 704.0, + 425.0, + 704.0, + 425.0, + 731.0, + 386.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 703.0, + 491.0, + 703.0, + 491.0, + 733.0, + 448.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 704.0, + 556.0, + 704.0, + 556.0, + 731.0, + 517.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 701.0, + 622.0, + 701.0, + 622.0, + 732.0, + 580.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 704.0, + 687.0, + 704.0, + 687.0, + 731.0, + 646.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 703.0, + 754.0, + 703.0, + 754.0, + 733.0, + 710.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 704.0, + 817.0, + 704.0, + 817.0, + 731.0, + 778.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 701.0, + 983.0, + 701.0, + 983.0, + 733.0, + 943.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 703.0, + 1051.0, + 703.0, + 1051.0, + 733.0, + 1010.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 701.0, + 1120.0, + 701.0, + 1120.0, + 732.0, + 1076.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 703.0, + 1188.0, + 703.0, + 1188.0, + 733.0, + 1146.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 704.0, + 1255.0, + 704.0, + 1255.0, + 731.0, + 1215.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 701.0, + 1325.0, + 701.0, + 1325.0, + 733.0, + 1281.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 704.0, + 1391.0, + 704.0, + 1391.0, + 731.0, + 1350.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 728.0, + 658.0, + 728.0, + 658.0, + 756.0, + 544.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 724.0, + 806.0, + 724.0, + 806.0, + 749.0, + 769.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 728.0, + 1224.0, + 728.0, + 1224.0, + 756.0, + 1110.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 722.0, + 1380.0, + 722.0, + 1380.0, + 749.0, + 1343.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 777.0, + 735.0, + 777.0, + 735.0, + 807.0, + 391.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 778.0, + 1306.0, + 778.0, + 1306.0, + 809.0, + 966.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 838.0, + 394.0, + 838.0, + 394.0, + 863.0, + 348.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 837.0, + 962.0, + 837.0, + 962.0, + 865.0, + 906.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 882.0, + 394.0, + 882.0, + 394.0, + 1066.0, + 321.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 859.0, + 635.0, + 859.0, + 635.0, + 921.0, + 468.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 897.0, + 910.0, + 897.0, + 910.0, + 1066.0, + 879.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 870.0, + 964.0, + 870.0, + 964.0, + 900.0, + 917.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 863.0, + 1178.0, + 863.0, + 1178.0, + 912.0, + 1040.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 928.0, + 394.0, + 928.0, + 394.0, + 957.0, + 349.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 918.0, + 640.0, + 918.0, + 640.0, + 949.0, + 471.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 908.0, + 964.0, + 908.0, + 964.0, + 938.0, + 916.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 925.0, + 1203.0, + 925.0, + 1203.0, + 943.0, + 1138.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 946.0, + 962.0, + 946.0, + 962.0, + 975.0, + 917.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 977.0, + 391.0, + 977.0, + 391.0, + 1002.0, + 350.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 982.0, + 964.0, + 982.0, + 964.0, + 1012.0, + 916.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1023.0, + 390.0, + 1023.0, + 390.0, + 1049.0, + 359.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1019.0, + 962.0, + 1019.0, + 962.0, + 1048.0, + 917.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1056.0, + 961.0, + 1056.0, + 961.0, + 1082.0, + 917.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1072.0, + 391.0, + 1072.0, + 391.0, + 1091.0, + 373.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1091.0, + 962.0, + 1091.0, + 962.0, + 1121.0, + 916.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1119.0, + 426.0, + 1119.0, + 426.0, + 1146.0, + 387.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1116.0, + 491.0, + 1116.0, + 491.0, + 1147.0, + 448.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1116.0, + 556.0, + 1116.0, + 556.0, + 1147.0, + 515.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1116.0, + 622.0, + 1116.0, + 622.0, + 1147.0, + 580.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1118.0, + 690.0, + 1118.0, + 690.0, + 1147.0, + 645.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1116.0, + 754.0, + 1116.0, + 754.0, + 1147.0, + 710.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1118.0, + 818.0, + 1118.0, + 818.0, + 1149.0, + 776.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1116.0, + 995.0, + 1116.0, + 995.0, + 1147.0, + 952.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1116.0, + 1061.0, + 1116.0, + 1061.0, + 1147.0, + 1017.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1116.0, + 1127.0, + 1116.0, + 1127.0, + 1147.0, + 1085.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1116.0, + 1192.0, + 1116.0, + 1192.0, + 1147.0, + 1150.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1116.0, + 1260.0, + 1116.0, + 1260.0, + 1147.0, + 1216.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1116.0, + 1326.0, + 1116.0, + 1326.0, + 1149.0, + 1284.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1119.0, + 1391.0, + 1119.0, + 1391.0, + 1146.0, + 1351.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1142.0, + 659.0, + 1142.0, + 659.0, + 1170.0, + 545.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1139.0, + 807.0, + 1139.0, + 807.0, + 1164.0, + 769.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1143.0, + 1229.0, + 1143.0, + 1229.0, + 1171.0, + 1115.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1139.0, + 1378.0, + 1139.0, + 1378.0, + 1164.0, + 1343.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1192.0, + 713.0, + 1192.0, + 713.0, + 1227.0, + 417.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1193.0, + 1281.0, + 1193.0, + 1281.0, + 1224.0, + 992.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1252.0, + 383.0, + 1252.0, + 383.0, + 1281.0, + 336.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1252.0, + 952.0, + 1252.0, + 952.0, + 1281.0, + 907.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1280.0, + 383.0, + 1280.0, + 383.0, + 1312.0, + 336.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1277.0, + 600.0, + 1277.0, + 600.0, + 1308.0, + 458.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1287.0, + 952.0, + 1287.0, + 952.0, + 1316.0, + 904.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1304.0, + 607.0, + 1304.0, + 607.0, + 1336.0, + 458.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1312.0, + 345.0, + 1312.0, + 345.0, + 1415.0, + 311.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1312.0, + 381.0, + 1312.0, + 381.0, + 1342.0, + 333.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1312.0, + 910.0, + 1312.0, + 910.0, + 1481.0, + 879.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1326.0, + 948.0, + 1326.0, + 948.0, + 1351.0, + 909.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1335.0, + 569.0, + 1335.0, + 569.0, + 1361.0, + 462.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1342.0, + 383.0, + 1342.0, + 383.0, + 1371.0, + 336.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1344.0, + 576.0, + 1344.0, + 576.0, + 1354.0, + 560.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1372.0, + 383.0, + 1372.0, + 383.0, + 1402.0, + 336.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1362.0, + 950.0, + 1362.0, + 950.0, + 1388.0, + 906.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1400.0, + 381.0, + 1400.0, + 381.0, + 1481.0, + 312.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1400.0, + 950.0, + 1400.0, + 950.0, + 1425.0, + 906.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1434.0, + 380.0, + 1434.0, + 380.0, + 1460.0, + 339.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1435.0, + 950.0, + 1435.0, + 950.0, + 1465.0, + 904.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1424.0, + 1356.0, + 1424.0, + 1356.0, + 1455.0, + 1213.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1463.0, + 381.0, + 1463.0, + 381.0, + 1519.0, + 346.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1470.0, + 951.0, + 1470.0, + 951.0, + 1499.0, + 906.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1448.0, + 1370.0, + 1448.0, + 1370.0, + 1511.0, + 1212.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1509.0, + 951.0, + 1509.0, + 951.0, + 1534.0, + 906.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1537.0, + 405.0, + 1537.0, + 405.0, + 1561.0, + 384.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1537.0, + 487.0, + 1537.0, + 487.0, + 1561.0, + 466.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1536.0, + 569.0, + 1536.0, + 569.0, + 1561.0, + 546.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1536.0, + 652.0, + 1536.0, + 652.0, + 1561.0, + 630.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1536.0, + 735.0, + 1536.0, + 735.0, + 1561.0, + 714.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1536.0, + 817.0, + 1536.0, + 817.0, + 1561.0, + 794.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1537.0, + 974.0, + 1537.0, + 974.0, + 1561.0, + 952.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1540.0, + 1055.0, + 1540.0, + 1055.0, + 1558.0, + 1038.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1536.0, + 1140.0, + 1536.0, + 1140.0, + 1561.0, + 1117.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1536.0, + 1224.0, + 1536.0, + 1224.0, + 1561.0, + 1202.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1536.0, + 1306.0, + 1536.0, + 1306.0, + 1561.0, + 1285.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1536.0, + 1391.0, + 1536.0, + 1391.0, + 1561.0, + 1368.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1558.0, + 656.0, + 1558.0, + 656.0, + 1586.0, + 542.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1554.0, + 813.0, + 1554.0, + 813.0, + 1579.0, + 778.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1558.0, + 1227.0, + 1558.0, + 1227.0, + 1586.0, + 1115.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1554.0, + 1387.0, + 1554.0, + 1387.0, + 1581.0, + 1351.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1607.0, + 738.0, + 1607.0, + 738.0, + 1642.0, + 388.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1609.0, + 1306.0, + 1609.0, + 1306.0, + 1639.0, + 966.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.75, + 908.5, + 1152.75, + 908.5, + 1152.75, + 954.5, + 1025.75, + 954.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.25, + 972.5, + 1017.25, + 972.5, + 1017.25, + 1009.5, + 1011.25, + 1009.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 977.5, + 704.0, + 977.5, + 704.0, + 1069.5, + 393.0, + 1069.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1317.0, + 852.0, + 1317.0, + 852.0, + 1436.0, + 412.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1666.0, + 810.0, + 1666.0, + 810.0, + 1698.0, + 295.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1666.0, + 1405.0, + 1666.0, + 1405.0, + 1698.0, + 929.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1693.0, + 314.0, + 1693.0, + 314.0, + 1730.0, + 292.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1693.0, + 1405.0, + 1693.0, + 1405.0, + 1730.0, + 390.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1725.0, + 1406.0, + 1725.0, + 1406.0, + 1759.0, + 292.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1757.0, + 467.0, + 1757.0, + 467.0, + 1787.0, + 296.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2122.0, + 833.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 778.0, + 262.0, + 778.0, + 294.0, + 296.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1945.0, + 293.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1407.0, + 1974.0, + 1407.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1169.0, + 2002.0, + 1169.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 778.0, + 262.0, + 778.0, + 294.0, + 296.0, + 294.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1719, + 1406, + 1719, + 1406, + 1959, + 297, + 1959 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 306, + 217, + 1393, + 217, + 1393, + 1460, + 306, + 1460 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 296, + 1479, + 1407, + 1479, + 1407, + 1603, + 296, + 1603 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 296, + 1972, + 1404, + 1972, + 1404, + 2035, + 296, + 2035 + ], + "score": 0.936 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 826, + 76, + 826, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 1666, + 832, + 1666, + 832, + 1697, + 300, + 1697 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.854 + }, + { + "category_id": 13, + "poly": [ + 1066, + 1842, + 1204, + 1842, + 1204, + 1887, + 1066, + 1887 + ], + "score": 0.94, + "latex": "\\textstyle \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 536, + 1883, + 806, + 1883, + 806, + 1929, + 536, + 1929 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { { R ^ { k } } = E ^ { k } + \\eta \\sum _ { i = 1 , i \\neq k } ^ { N } E ^ { i } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1305, + 2004, + 1398, + 2004, + 1398, + 2035, + 1305, + 2035 + ], + "score": 0.9, + "latex": "\\eta = 0 . 8 5" + }, + { + "category_id": 13, + "poly": [ + 667, + 1510, + 740, + 1510, + 740, + 1539, + 667, + 1539 + ], + "score": 0.9, + "latex": "9 9 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 545, + 1845, + 582, + 1845, + 582, + 1875, + 545, + 1875 + ], + "score": 0.89, + "latex": "E ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 490, + 1891, + 508, + 1891, + 508, + 1917, + 490, + 1917 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1924, + 1335, + 1924, + 1335, + 1959, + 1246, + 1959 + ], + "score": 0.69, + "latex": "\\eta \\in [ 0 , 2 ]" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 236.0, + 396.0, + 236.0, + 396.0, + 265.0, + 347.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 236.0, + 962.0, + 236.0, + 962.0, + 265.0, + 905.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 276.0, + 394.0, + 276.0, + 394.0, + 307.0, + 347.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 276.0, + 965.0, + 276.0, + 965.0, + 307.0, + 917.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 296.0, + 911.0, + 296.0, + 911.0, + 461.0, + 879.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 299.0, + 336.0, + 299.0, + 336.0, + 461.0, + 309.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 319.0, + 394.0, + 319.0, + 394.0, + 349.0, + 347.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 319.0, + 964.0, + 319.0, + 964.0, + 349.0, + 915.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 332.0, + 1259.0, + 332.0, + 1259.0, + 363.0, + 1113.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 363.0, + 393.0, + 363.0, + 393.0, + 391.0, + 359.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 363.0, + 964.0, + 363.0, + 964.0, + 392.0, + 917.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 360.0, + 1362.0, + 360.0, + 1362.0, + 391.0, + 1115.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 386.0, + 1222.0, + 386.0, + 1222.0, + 422.0, + 1114.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 408.0, + 393.0, + 408.0, + 393.0, + 433.0, + 370.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 406.0, + 687.0, + 406.0, + 687.0, + 437.0, + 543.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 405.0, + 964.0, + 405.0, + 964.0, + 434.0, + 915.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 448.0, + 392.0, + 448.0, + 392.0, + 475.0, + 350.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 434.0, + 787.0, + 434.0, + 787.0, + 465.0, + 544.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 447.0, + 962.0, + 447.0, + 962.0, + 476.0, + 915.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 465.0, + 647.0, + 465.0, + 647.0, + 493.0, + 546.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 490.0, + 393.0, + 490.0, + 393.0, + 520.0, + 336.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 490.0, + 964.0, + 490.0, + 964.0, + 520.0, + 915.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 518.0, + 426.0, + 518.0, + 426.0, + 545.0, + 386.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 518.0, + 490.0, + 518.0, + 490.0, + 545.0, + 450.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 518.0, + 556.0, + 518.0, + 556.0, + 545.0, + 517.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 515.0, + 623.0, + 515.0, + 623.0, + 546.0, + 581.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 515.0, + 690.0, + 515.0, + 690.0, + 546.0, + 644.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 517.0, + 754.0, + 517.0, + 754.0, + 548.0, + 711.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 518.0, + 818.0, + 518.0, + 818.0, + 545.0, + 778.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 518.0, + 994.0, + 518.0, + 994.0, + 545.0, + 954.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 515.0, + 1061.0, + 515.0, + 1061.0, + 546.0, + 1018.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 515.0, + 1126.0, + 515.0, + 1126.0, + 546.0, + 1085.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 515.0, + 1193.0, + 515.0, + 1193.0, + 546.0, + 1151.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 515.0, + 1260.0, + 515.0, + 1260.0, + 546.0, + 1216.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 515.0, + 1327.0, + 515.0, + 1327.0, + 548.0, + 1283.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 518.0, + 1390.0, + 518.0, + 1390.0, + 545.0, + 1349.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 542.0, + 660.0, + 542.0, + 660.0, + 570.0, + 544.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 538.0, + 805.0, + 538.0, + 805.0, + 563.0, + 768.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 542.0, + 1229.0, + 542.0, + 1229.0, + 570.0, + 1113.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 538.0, + 1380.0, + 538.0, + 1380.0, + 563.0, + 1342.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 590.0, + 736.0, + 590.0, + 736.0, + 622.0, + 390.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 590.0, + 1307.0, + 590.0, + 1307.0, + 623.0, + 962.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 651.0, + 397.0, + 651.0, + 397.0, + 680.0, + 347.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 650.0, + 954.0, + 650.0, + 954.0, + 680.0, + 905.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 714.0, + 336.0, + 714.0, + 336.0, + 876.0, + 309.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 694.0, + 393.0, + 694.0, + 393.0, + 720.0, + 349.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 682.0, + 952.0, + 682.0, + 952.0, + 788.0, + 881.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 676.0, + 1195.0, + 676.0, + 1195.0, + 735.0, + 1028.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 735.0, + 393.0, + 735.0, + 393.0, + 764.0, + 347.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 732.0, + 1194.0, + 732.0, + 1194.0, + 763.0, + 1030.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 749.0, + 951.0, + 749.0, + 951.0, + 774.0, + 907.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 780.0, + 392.0, + 780.0, + 392.0, + 806.0, + 359.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 777.0, + 951.0, + 777.0, + 951.0, + 876.0, + 881.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 809.0, + 952.0, + 809.0, + 952.0, + 840.0, + 904.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 823.0, + 393.0, + 823.0, + 393.0, + 848.0, + 370.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 822.0, + 760.0, + 822.0, + 760.0, + 852.0, + 617.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 850.0, + 784.0, + 850.0, + 784.0, + 880.0, + 618.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 844.0, + 949.0, + 844.0, + 949.0, + 869.0, + 905.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 862.0, + 394.0, + 862.0, + 394.0, + 892.0, + 346.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 876.0, + 788.0, + 876.0, + 788.0, + 910.0, + 616.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 873.0, + 952.0, + 873.0, + 952.0, + 903.0, + 907.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 906.0, + 393.0, + 906.0, + 393.0, + 935.0, + 336.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 906.0, + 951.0, + 906.0, + 951.0, + 935.0, + 905.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 932.0, + 426.0, + 932.0, + 426.0, + 963.0, + 384.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 932.0, + 491.0, + 932.0, + 491.0, + 963.0, + 450.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 932.0, + 557.0, + 932.0, + 557.0, + 963.0, + 514.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 932.0, + 623.0, + 932.0, + 623.0, + 963.0, + 580.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 932.0, + 690.0, + 932.0, + 690.0, + 963.0, + 644.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 932.0, + 754.0, + 932.0, + 754.0, + 963.0, + 710.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 930.0, + 821.0, + 930.0, + 821.0, + 964.0, + 778.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 932.0, + 984.0, + 932.0, + 984.0, + 963.0, + 942.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 932.0, + 1052.0, + 932.0, + 1052.0, + 963.0, + 1009.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 932.0, + 1119.0, + 932.0, + 1119.0, + 963.0, + 1076.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 932.0, + 1188.0, + 932.0, + 1188.0, + 963.0, + 1145.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 932.0, + 1258.0, + 932.0, + 1258.0, + 963.0, + 1213.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 932.0, + 1325.0, + 932.0, + 1325.0, + 963.0, + 1280.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 932.0, + 1393.0, + 932.0, + 1393.0, + 963.0, + 1349.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 957.0, + 660.0, + 957.0, + 660.0, + 985.0, + 544.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 953.0, + 805.0, + 953.0, + 805.0, + 978.0, + 770.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 957.0, + 1225.0, + 957.0, + 1225.0, + 985.0, + 1111.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 953.0, + 1379.0, + 953.0, + 1379.0, + 980.0, + 1342.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1006.0, + 713.0, + 1006.0, + 713.0, + 1041.0, + 416.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1008.0, + 1282.0, + 1008.0, + 1282.0, + 1039.0, + 992.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1067.0, + 394.0, + 1067.0, + 394.0, + 1096.0, + 347.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1067.0, + 954.0, + 1067.0, + 954.0, + 1096.0, + 905.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1097.0, + 394.0, + 1097.0, + 394.0, + 1128.0, + 346.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1102.0, + 952.0, + 1102.0, + 952.0, + 1131.0, + 905.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1089.0, + 1355.0, + 1089.0, + 1355.0, + 1123.0, + 1214.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1128.0, + 336.0, + 1128.0, + 336.0, + 1290.0, + 309.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1129.0, + 394.0, + 1129.0, + 394.0, + 1159.0, + 347.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1127.0, + 952.0, + 1127.0, + 952.0, + 1293.0, + 879.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1115.0, + 1370.0, + 1115.0, + 1370.0, + 1177.0, + 1212.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1162.0, + 393.0, + 1162.0, + 393.0, + 1191.0, + 347.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1177.0, + 951.0, + 1177.0, + 951.0, + 1202.0, + 907.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1195.0, + 393.0, + 1195.0, + 393.0, + 1220.0, + 349.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1211.0, + 952.0, + 1211.0, + 952.0, + 1241.0, + 905.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1226.0, + 393.0, + 1226.0, + 393.0, + 1285.0, + 357.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1237.0, + 788.0, + 1237.0, + 788.0, + 1296.0, + 638.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1248.0, + 951.0, + 1248.0, + 951.0, + 1278.0, + 905.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1288.0, + 393.0, + 1288.0, + 393.0, + 1318.0, + 346.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1293.0, + 795.0, + 1293.0, + 795.0, + 1324.0, + 637.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1285.0, + 951.0, + 1285.0, + 951.0, + 1314.0, + 905.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1321.0, + 393.0, + 1321.0, + 393.0, + 1349.0, + 336.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1320.0, + 952.0, + 1320.0, + 952.0, + 1351.0, + 905.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1351.0, + 416.0, + 1351.0, + 416.0, + 1374.0, + 393.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1355.0, + 494.0, + 1355.0, + 494.0, + 1372.0, + 477.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1349.0, + 576.0, + 1349.0, + 576.0, + 1376.0, + 553.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1349.0, + 658.0, + 1349.0, + 658.0, + 1376.0, + 634.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1351.0, + 738.0, + 1351.0, + 738.0, + 1374.0, + 715.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1351.0, + 817.0, + 1351.0, + 817.0, + 1376.0, + 794.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1351.0, + 975.0, + 1351.0, + 975.0, + 1374.0, + 952.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1355.0, + 1054.0, + 1355.0, + 1054.0, + 1373.0, + 1038.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1351.0, + 1141.0, + 1351.0, + 1141.0, + 1376.0, + 1118.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1349.0, + 1225.0, + 1349.0, + 1225.0, + 1376.0, + 1202.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1352.0, + 1307.0, + 1352.0, + 1307.0, + 1374.0, + 1285.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 1351.0, + 1390.0, + 1351.0, + 1390.0, + 1376.0, + 1369.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1373.0, + 663.0, + 1373.0, + 663.0, + 1401.0, + 548.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1369.0, + 814.0, + 1369.0, + 814.0, + 1394.0, + 778.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1373.0, + 1228.0, + 1373.0, + 1228.0, + 1401.0, + 1113.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1369.0, + 1386.0, + 1369.0, + 1386.0, + 1394.0, + 1350.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1425.0, + 738.0, + 1425.0, + 738.0, + 1455.0, + 390.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1422.0, + 1309.0, + 1422.0, + 1309.0, + 1457.0, + 965.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 739.5, + 1409.0, + 739.5, + 1409.0, + 839.5, + 959.0, + 839.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1159.0, + 1381.0, + 1159.0, + 1381.0, + 1255.5, + 1113.0, + 1255.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1513.0, + 295.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1509.0, + 666.0, + 1509.0, + 666.0, + 1542.0, + 295.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1509.0, + 1407.0, + 1509.0, + 1407.0, + 1542.0, + 741.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1540.0, + 1409.0, + 1540.0, + 1409.0, + 1576.0, + 295.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1571.0, + 986.0, + 1571.0, + 986.0, + 1607.0, + 294.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 830.0, + 73.0, + 830.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1666.0, + 837.0, + 1666.0, + 837.0, + 1701.0, + 296.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1721.0, + 1404.0, + 1721.0, + 1404.0, + 1754.0, + 296.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1750.0, + 1404.0, + 1750.0, + 1404.0, + 1787.0, + 295.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1779.0, + 1411.0, + 1779.0, + 1411.0, + 1818.0, + 292.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1805.0, + 544.0, + 1805.0, + 544.0, + 1908.0, + 278.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1805.0, + 1065.0, + 1805.0, + 1065.0, + 1908.0, + 583.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1805.0, + 1407.0, + 1805.0, + 1407.0, + 1908.0, + 1205.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1841.0, + 1408.0, + 1841.0, + 1408.0, + 1883.0, + 1205.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1879.0, + 489.0, + 1879.0, + 489.0, + 1939.0, + 287.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1879.0, + 535.0, + 1879.0, + 535.0, + 1939.0, + 509.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1879.0, + 1411.0, + 1879.0, + 1411.0, + 1939.0, + 807.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1923.0, + 1245.0, + 1923.0, + 1245.0, + 1962.0, + 291.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1923.0, + 1345.0, + 1923.0, + 1345.0, + 1962.0, + 1336.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1882.0, + 723.0, + 1882.0, + 723.0, + 1906.0, + 679.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1304.0, + 2002.0, + 1304.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 2002.0, + 1407.0, + 2002.0, + 1407.0, + 2038.0, + 1399.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 774, + 1403, + 774, + 1403, + 928, + 298, + 928 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 297, + 622, + 1403, + 622, + 1403, + 715, + 297, + 715 + ], + "score": 0.962 + }, + { + "category_id": 3, + "poly": [ + 327, + 216, + 1374, + 216, + 1374, + 595, + 327, + 595 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 944, + 1401, + 944, + 1401, + 1038, + 299, + 1038 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 826, + 76, + 826, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 13, + "poly": [ + 1076, + 654, + 1148, + 654, + 1148, + 683, + 1076, + 683 + ], + "score": 0.87, + "latex": "9 9 . 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 620.0, + 1407.0, + 620.0, + 1407.0, + 658.0, + 295.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 654.0, + 1075.0, + 654.0, + 1075.0, + 688.0, + 295.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 654.0, + 1405.0, + 654.0, + 1405.0, + 688.0, + 1149.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 682.0, + 871.0, + 682.0, + 871.0, + 722.0, + 295.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 225.0, + 420.0, + 225.0, + 420.0, + 258.0, + 353.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 226.0, + 969.0, + 226.0, + 969.0, + 256.0, + 903.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 255.0, + 958.0, + 255.0, + 958.0, + 285.0, + 903.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 277.0, + 413.0, + 277.0, + 413.0, + 308.0, + 357.0, + 308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 283.0, + 959.0, + 283.0, + 959.0, + 313.0, + 902.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 288.0, + 905.0, + 288.0, + 905.0, + 449.0, + 876.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 291.0, + 355.0, + 291.0, + 355.0, + 448.0, + 328.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 311.0, + 959.0, + 311.0, + 959.0, + 341.0, + 903.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 328.0, + 412.0, + 328.0, + 412.0, + 358.0, + 356.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 339.0, + 959.0, + 339.0, + 959.0, + 369.0, + 903.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 377.0, + 412.0, + 377.0, + 412.0, + 408.0, + 365.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 367.0, + 958.0, + 367.0, + 958.0, + 398.0, + 904.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 394.0, + 958.0, + 394.0, + 958.0, + 424.0, + 903.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 429.0, + 418.0, + 429.0, + 418.0, + 458.0, + 387.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 424.0, + 957.0, + 424.0, + 957.0, + 452.0, + 913.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 443.0, + 976.0, + 443.0, + 976.0, + 455.0, + 968.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 450.0, + 958.0, + 450.0, + 958.0, + 482.0, + 911.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 459.0, + 976.0, + 459.0, + 976.0, + 472.0, + 968.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 479.0, + 419.0, + 479.0, + 419.0, + 509.0, + 355.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 479.0, + 962.0, + 479.0, + 962.0, + 510.0, + 911.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 499.0, + 436.0, + 499.0, + 436.0, + 531.0, + 395.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 498.0, + 500.0, + 498.0, + 500.0, + 530.0, + 460.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 498.0, + 565.0, + 498.0, + 565.0, + 530.0, + 523.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 498.0, + 629.0, + 498.0, + 629.0, + 530.0, + 588.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 496.0, + 696.0, + 496.0, + 696.0, + 532.0, + 653.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 498.0, + 759.0, + 498.0, + 759.0, + 530.0, + 717.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 500.0, + 821.0, + 500.0, + 821.0, + 529.0, + 781.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 500.0, + 981.0, + 500.0, + 981.0, + 529.0, + 942.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 500.0, + 1045.0, + 500.0, + 1045.0, + 529.0, + 1007.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 498.0, + 1112.0, + 498.0, + 1112.0, + 530.0, + 1072.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 498.0, + 1178.0, + 498.0, + 1178.0, + 530.0, + 1135.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 496.0, + 1243.0, + 496.0, + 1243.0, + 532.0, + 1201.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 498.0, + 1307.0, + 498.0, + 1307.0, + 530.0, + 1265.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 500.0, + 1370.0, + 500.0, + 1370.0, + 529.0, + 1332.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 522.0, + 665.0, + 522.0, + 665.0, + 554.0, + 552.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 519.0, + 810.0, + 519.0, + 810.0, + 546.0, + 772.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 522.0, + 1214.0, + 522.0, + 1214.0, + 554.0, + 1100.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 519.0, + 1359.0, + 519.0, + 1359.0, + 546.0, + 1322.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 560.0, + 638.0, + 560.0, + 638.0, + 606.0, + 511.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 564.0, + 1187.0, + 564.0, + 1187.0, + 601.0, + 1063.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.75, + 282.5, + 559.75, + 282.5, + 559.75, + 303.0, + 509.75, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 264.0, + 1353.0, + 264.0, + 1353.0, + 347.5, + 1008.0, + 347.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 831.0, + 73.0, + 831.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2121.0, + 831.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 774.0, + 1405.0, + 774.0, + 1405.0, + 811.0, + 294.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 807.0, + 1404.0, + 807.0, + 1404.0, + 840.0, + 296.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 836.0, + 1404.0, + 836.0, + 1404.0, + 870.0, + 296.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 866.0, + 1405.0, + 866.0, + 1405.0, + 903.0, + 296.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 900.0, + 704.0, + 900.0, + 704.0, + 930.0, + 294.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 945.0, + 1403.0, + 945.0, + 1403.0, + 979.0, + 295.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 974.0, + 1402.0, + 974.0, + 1402.0, + 1008.0, + 295.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1007.0, + 1008.0, + 1007.0, + 1008.0, + 1041.0, + 295.0, + 1041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/BJg15lrKvS/BJg15lrKvS.md b/parse/train/BJg15lrKvS/BJg15lrKvS.md new file mode 100644 index 0000000000000000000000000000000000000000..6c274b8785b36eb762db2d087052db97f2fdd82c --- /dev/null +++ b/parse/train/BJg15lrKvS/BJg15lrKvS.md @@ -0,0 +1,737 @@ +# TOWARDS UNDERSTANDING THE SPECTRAL BIAS OF DEEP LEARNING + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +An intriguing phenomenon observed during training neural networks is the spectral bias, where neural networks are biased towards learning less complex functions. The priority of learning functions with low complexity might be at the core of explaining generalization ability of neural network, and certain efforts have been made to provide theoretical explanation for spectral bias. However, there is still no satisfying theoretical result justifying the underlying mechanism of spectral bias. In this paper, we give a comprehensive and rigorous explanation for spectral bias and relate it with the neural tangent kernel function proposed in recent work. We prove that the training process of neural networks can be decomposed along different directions defined by the eigenfunctions of the neural tangent kernel, where each direction has its own convergence rate and the rate is determined by the corresponding eigenvalue. We then provide a case study when the input data is uniformly distributed over the unit sphere, and show that lower degree spherical harmonics are easier to be learned by over-parameterized neural networks. + +# 1 INTRODUCTION + +Over-parameterized neural networks have achieved great success in many applications such as computer vision (He et al., 2016), natural language processing (Collobert and Weston, 2008) and speech recognition (Hinton et al., 2012). It has been shown that over-parameterized neural networks can fit complicated target function or even randomly labeled data (Zhang et al., 2016) and still exhibit good generalization performance when trained with real labels. Intuitively, this is at odds with the traditional notion of generalization ability such as model complexity. In order to understand neural network training, a line of work (Soudry et al., 2017; Gunasekar et al., 2018b;a) has made efforts in the perspective of “implicit bias”, which states that training algorithms for deep learning implicitly pose an inductive bias onto the training process and lead to a solution with low complexity measured by certain norms in the parameter space of the neural network. + +Among many attempts to establish implicit bias, Rahaman et al. (2018) pointed out an intriguing phenomenon called spectral bias, which says that during training, neural networks tend to learn the components of lower complexity faster. The concept of spectral bias is appealing because this may intuitively explain why over-parameterized neural networks can achieve a good generalization performance without overfitting. During training, the networks fit the low complexity components first and thus lie in the concept class of low complexity. Arguments like this may lead to rigorous guarantee for generalization. + +Great efforts have been made in seek of explanations about the spectral bias. Rahaman et al. (2018) evaluated the Fourier spectrum of ReLU networks and empirically showed that the lower frequencies are learned first; also lower frequencies are more robust to random perturbation. Andoni et al. (2014) showed that for a sufficiently wide two-layer network, gradient descent with respect to the second layer can learn any low degree bounded polynomial. Xu (2018) provided Fourier analysis to twolayer networks and showed similar empirical results on one-dimensional functions and real data. Nakkiran et al. (2019) used information theoretical approach to show that networks obtained by stochastic gradient descent can be explained by a linear classifier during early training. All these studies provide certain explanations about why neural networks exhibit spectral bias in real tasks. But explanations in the theoretical aspect, if any, are to some extent restricted. For example, the popular Fourier analysis is usually done in the one-dimensional setting, and thus lacks generality. + +Meanwhile, a recent line of work (Jacot et al., 2018; Du et al., 2018b; Li and Liang, 2018; Chizat and Bach, 2018) has shed light on new approaches to analyze neural networks. In particular, they show that under certain over-parameterized condition, the neural network trained by gradient descent behaves similar to the kernel regression predictor using the neural tangent kernel (NTK) (Jacot et al., 2018). Du et al. (2018b) showed that the convergence is provably guaranteed under certain over-parameterization conditions determined by the smallest eigenvalue of NTK. Arora et al. (2019b) further gave a finer characterization of error convergence based on the eigenvalues of NTK’s Gram matrix. Su and Yang (2019) improved the convergence guarantee in terms of the $k$ -th largest eigenvalue for certain target functions. + +Inspired by these works mentioned above, we can present a theoretical explanation for spectral bias. Under NTK regime, we establish a precise characterization for the training process of neural networks. More specifically, we theoretically prove that over-parameterized neural networks’ training process can be controlled by the eigenvalues of the integrating operator defined by the NTK. Under the specific case of uniform distribution on unit sphere, we give an exact calculation for these eigenvalues and show that the lower frequencies have larger eigenvalues, which thus leads to faster convergence. We also conduct experiments to corroborate the theory we establish. + +Our contributions are highlighted as follows: + +1. We prove a generic theorem for arbitrary data distributions, which states that under certain sample complexity and over-parameterization conditions, the error term’s convergence along different directions actually relies on the corresponding eigenvalues. This theorem gives a more precise control on the regression residual than $\mathtt { S u }$ and Yang (2019), where the authors focused on the case when the labeling function is close to the subspace spanned by the first few eigenfunctions. +2. We present a more general result about the spectra of the neural tangent kernel. In particular, we show that the order of eigenvalues appears as $\mu _ { k } = \Omega ( \operatorname* { m a x } \{ k ^ { - d - 1 } , \check { d } ^ { - k + 1 } \} )$ . Our result is better than the bound $\Omega ( k ^ { - d - \overline { { 1 } } } )$ derived in Bietti and Mairal (2019) when $d \gg k$ , which is clearly a more practical setting. +3. We establish a rigorous explanation for the spectral bias based on the aforementioned theoretical results without any specific assumptions on the target function. We show that the error terms from different frequencies are provably controlled by the eigenvalues of the NTK, and the lowerfrequency components can be learned with less training examples and narrower networks with a faster convergence rate. As far as we know, this is the first attempt to give a comprehensive theory justifying the existence of spectral bias. + +# 1.1 ADDITIONAL RELATED WORK + +Recently, there is a rich literature about the property of neural tangent kernel. Jacot et al. (2018) first showed that during training, the network function follows a descent along the kernel gradient with respect to the Neural Tangent Kernel (NTK) under infinity width setting. Li and Liang (2018) and Du et al. (2018b) implicitly built connection between Neural Tangent Kernel and gradient descent by showing that GD can provably optimize sufficiently wide two-layer neural networks. In Du et al. (2018b), it is shown that gradient descent can achieve zero training loss at a linear convergence rate for training two-layer ReLU network with square loss. Allen-Zhu et al. (2018); Du et al. (2018a); Zou et al. (2018); Cao and Gu (2019b); Arora et al. (2019a); Zou and Gu (2019); Cao and Gu (2019a) further studied the optimization and generalization of deep neural networks. These papers are all in the so-called neural tangent kernel regime, and their requirements on the network width depend either implicitly or explicitly on the smallest eigenvalue of the kernel Gram matrix. Later, Su and Yang (2019) showed that this smallest eigenvalue actually scales in the number of samples $n$ and will eventually converge to 0. In order to obtain constant convergence rate, Su and Yang (2019) assumed that the target function ş $f ^ { * }$ can be approximated by the first few eigenfunctions of the integrating operator $L _ { \kappa } \bar { f } ( s ) : = \int _ { \mathbb { S } ^ { d } } \kappa \bar { ( } x , s ) f ( s ) d \bar { \tau } \bar { ( } s )$ where $\kappa ( \cdot , \cdot )$ is the NTK function and $\tau ( s )$ is the input distribution, and proved linear convergence rate up to the this approximation error. + +A few theoretical results have been established towards understanding the spectra of neural tangent kernels. Bach (2017) studied two-layer ReLU networks by relating it to kernel methods, and proposed a harmonic decomposition for the functions in the reproducing kernel Hilbert space which we utilize in our proof. Based on the technique in Bach (2017), Bietti and Mairal (2019) studied the eigenvalue decay of integrating operator $L _ { \kappa } f ( x )$ defined by NTK on unit sphere by using spherical harmonics. Vempala and Wilmes (2018) calculated the eigenvalues of NTK corresponding to two-layer neural networks with sigmoid activation function. Basri et al. (2019) established similar results as Bietti and Mairal (2019), but considered the case of training the first layer parameters of a two-layer networks with bias terms. Yang and Salman (2019) studied the the eigenvalues of integral operator with respect to the NTK on Boolean cube by Fourier analysis. + +The rest of the paper is organized as follows. We state the notation, problem setup and other preliminaries in Section 2 and present our main results in Section 3. In Section 4, we present experimental results to support our theory. Proofs of our main results can be found in the appendix. + +# 2 PRELIMINARIES + +In this section we introduce the basic problem setup including the neural network structure and the training algorithm, as well as some background on the neural tangent kernel proposed recently in Jacot et al. (2018) and the corresponding integral operator. + +# 2.1 NOTATION + +We use lower case, lower case bold face, and upper case bold face letters to denote scalars, vectors and matrices respectively. For a vector ř $\mathbf { v } = ( \bar { v _ { 1 } } , \ldots , v _ { d } ) ^ { T } \in \mathbb { R } ^ { d }$ and a number $1 \leqslant p < \infty$ , we denote its p´norm by $\begin{array} { r } { \| \mathbf { v } \| _ { p } = ( \sum _ { i = 1 } ^ { d } | v _ { i } | ^ { p } ) ^ { 1 / p } } \end{array}$ . We also define infinity norm by $\| \mathbf { v } \| _ { \infty } = \operatorname* { m a x } _ { i } | v _ { i } |$ . For a matrix ř $\mathbf { A } = ( A _ { i , j } ) _ { m \times n }$ , we use $\lVert \mathbf { A } \rVert _ { 0 }$ to denote the number of non-zero entries of $\mathbf { A }$ , and use $\textstyle \| \mathbf { A } \| _ { F } = ( \sum _ { i , j = 1 } ^ { d } A _ { i , j } ^ { 2 } ) ^ { 1 / 2 }$ to denote its Frobenius norm. Let $\begin{array} { r } { \| \mathbf { A } \| _ { p } = \operatorname* { m a x } _ { \| \mathbf { v } \| _ { p } \leqslant 1 } \| \mathbf { A } \mathbf { v } \| _ { p } } \end{array}$ for $p \geqslant 1$ , and $\| \mathbf { A } \| _ { \operatorname* { m a x } } = \operatorname* { m a x } _ { i , j } | A _ { i , j } |$ . For two matrices $\mathbf { A } , \mathbf { B } \in \mathbb { R } ^ { m \times n }$ , we define $\langle \mathbf { A } , \mathbf { B } \rangle = \operatorname { T r } ( \mathbf { A } ^ { \top } \mathbf { B } )$ . We use $\mathbf { A } \succeq \mathbf { B }$ if $\mathbf { A } - \mathbf { B }$ is positive semi-definite. In addition, we define the asymptotic notations $\mathcal { O } ( \cdot )$ , $\widetilde { \mathcal { O } } ( \cdot ) , \Omega ( \cdot )$ and $\widetilde { \Omega } ( \cdot )$ as follows. Suppose that $a _ { n }$ and $b _ { n }$ be two sequences. We write $a _ { n } = \mathcal { O } ( b _ { n } )$ if lim $\begin{array} { r } { \operatorname* { s u p } _ { n \infty } | a _ { n } / b _ { n } | < \infty } \end{array}$ , and $a _ { n } = \Omega ( b _ { n } )$ if $\begin{array} { r } { \operatorname* { l i m } \operatorname* { i n f } _ { n \to \infty } \left| a _ { n } / b _ { n } \right| > 0 } \end{array}$ . We use $\tilde { \mathcal { O } } ( \cdot )$ and $\widetilde { \Omega } ( \cdot )$ to hide the logarithmic factors in $\mathcal { O } ( \cdot )$ and $\Omega ( \cdot )$ . + +# 2.2 PROBLEM SETUP + +Here we introduce the basic problem setup. We consider two-layer fully connected neural networks of the form + +$$ +f _ { \mathbf { W } } ( \mathbf { x } ) = \sqrt { m } \cdot \mathbf { W } _ { 2 } \sigma ( \mathbf { W } _ { 1 } \mathbf { x } ) , +$$ + +where $\mathbf { W } _ { 1 } \in \mathbb { R } ^ { m \times ( d + 1 ) }$ , $\mathbf { W } _ { 2 } \in \mathbb { R } ^ { 1 \times m }$ are the first and second layer weight matrices respectively, and $\sigma ( \cdot ) = \operatorname* { m a x } \{ 0 , \cdot \}$ is the entry-wise ReLU activation function. The network is trained according to the square loss on $n$ training examples $S = \left\{ \left( \mathbf { x } _ { i } , y _ { i } \right) : i \in \left[ n \right] \right\}$ : + +$$ +L _ { S } ( \mathbf { W } ) = \frac { 1 } { n } \sum _ { ( \mathbf { x } _ { i } , y _ { i } ) \in S } \left( y _ { i } - \theta f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) \right) ^ { 2 } , +$$ + +where $\theta$ is a small coefficient to control the effect of initialization, and the data inputs $\left\{ { \bf x } _ { i } \right\} _ { i = 1 } ^ { n }$ is assumed to follow some unknown distribution $\tau$ on the unit sphere $\mathbb { S } ^ { d } \in \mathbb { R } ^ { d + 1 }$ . Without loss of generality, we also assume that $| y _ { i } | \leqslant 1$ . + +We first randomly initialize the parameters of the network, and then apply gradient descent to optimize both layers. We present our detailed neural network training algorithm in Algorithm 1. + +Algorithm 1 GD for DNNs starting at Gaussian initialization + +
Input: Number of iterations T,step size n. Generate each entry of W(0) and w)
from N(0,2/m) and N(0,1/m) respectively. 2 fort=0,1,...,T-1do
Update W(t+1) = W(t) - n · VwLs(W(t).
end for Output: W(T).
+ +The initialization scheme for $\mathbf { W } ^ { ( 0 ) }$ given in Algorithm 1 is known as He initialization (He et al., 2015). This scheme generates each entry of the weight matrices from a Gaussian distribution with mean zero. The variances of the Gaussian distributions in initialization are chosen following the principle that the initialization does not change the magnitudes of inputs in each layer. The second layer parameter is not associated with the ReLU activation function, thus it is initialized with variance $1 / m$ instead of $2 / m$ . + +# 2.3 NEURAL TANGENT KERNEL + +Many attempts have been made to study the convergence of gradient descent assuming the width of the network is extremely large (Du et al., 2018b; Li and Liang, 2018). When the width of the network goes to infinity, with certain initialization on parameters, the inner product of gradients of the output function would converge to a limiting kernel, i.e., Neural Tangent Kernel (Jacot et al., 2018). In this paper, we denote it by $\begin{array} { r } { \kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = \operatorname* { l i m } _ { m \infty } m ^ { - 1 } \langle \nabla \mathbf { w } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } ) , \nabla \mathbf { w } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } ^ { \prime } ) \rangle } \end{array}$ and we have + +$$ +\kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = \langle \mathbf { x } , \mathbf { x } ^ { \prime } \rangle \cdot \kappa _ { 1 } ( \mathbf { x } , \mathbf { x } ^ { \prime } ) + 2 \cdot \kappa _ { 2 } ( \mathbf { x } , \mathbf { x } ^ { \prime } ) , +$$ + +where + +$$ +\begin{array} { r l } & { \kappa _ { 1 } ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = \mathbb { E } _ { \mathbf { w } \sim N ( \mathbf { 0 } , \mathbf { I } ) } [ \sigma ^ { \prime } ( \langle \mathbf { w } , \mathbf { x } \rangle ) \sigma ^ { \prime } ( \langle \mathbf { w } , \mathbf { x } ^ { \prime } \rangle ) ] , } \\ & { \kappa _ { 2 } ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = \mathbb { E } _ { \mathbf { w } \sim N ( \mathbf { 0 } , \mathbf { I } ) } [ \sigma ( \langle \mathbf { w } , \mathbf { x } \rangle ) \sigma ( \langle \mathbf { w } , \mathbf { x } ^ { \prime } \rangle ) ] . } \end{array} +$$ + +Since we apply gradient descent to both layers, the Neural Tangent Kernel is the sum of two different kernel functions and clearly it can be reduced to one layer training setting. These two kernels are arc-cosine kernels of degree 0 and 1 (Cho and Saul, 2009). Their explicit expressions are given as + +$$ +\kappa _ { 1 } ( t ) = \frac { 1 } { 2 \pi } \left( \pi - \operatorname { a r c c o s } \left( t \right) \right) , \kappa _ { 2 } ( t ) = \frac { 1 } { 2 \pi } \left( t \cdot \left( \pi - \operatorname { a r c c o s } \left( t \right) \right) + \sqrt { 1 - t ^ { 2 } } \right) , +$$ + +where $t = \langle \mathbf { x } , \mathbf { x } ^ { \prime } \rangle / \left( \left\| \mathbf { x } \right\| \left\| \mathbf { x } ^ { \prime } \right\| \right)$ . + +# 2.4 INTEGRAL OPERATOR + +The theory of integral operator with respect to kernel function has been well studied in machine learning (Smale and Zhou, 2007; Rosasco et al., 2010) thus we only give a brief introduction here. Let $L _ { \tau } ^ { 2 } ( X )$ be the Hilbert space of square-integrable functions with respect to a Borel measure $\tau$ from $X \to \mathbb { R }$ . For any continuous kernel function $\kappa : X \times X \to \mathbb { R }$ and $\tau$ we can define an integral operator $L _ { \kappa }$ on $L _ { \tau } ^ { 2 } ( X )$ by + +$$ +L _ { \kappa } ( f ) ( { \bf x } ) = \int _ { X } \kappa ( { \bf x } , { \bf y } ) f ( { \bf y } ) d \tau ( { \bf y } ) , ~ { \bf x } \in X . +$$ + +It has been pointed out in Cho and Saul (2009) that arc-cosine kernels are positive semi-definite. Thus the kernel function $\kappa$ defined by (2.1) is positive semi-definite being a product and a sum of positive semi-definite kernels. Clearly this kernel is also continuous and symmetric. Thus we know that the neural tangent kernel $\kappa$ is a Mercer kernel. + +# 3 MAIN RESULTS + +In this section we present our main results. In Section 3.1, we give a general result on the convergence rate of gradient descent along different eigendirections of neural tangent kernel. Motivated by this result, in Section 3.2, we give a case study on the spectrum of $L _ { \kappa }$ when the input data are uniformly distributed over the unit sphere $\mathbb { S } ^ { d }$ . In Section 3.3, we combine the spectrum analysis with the general convergence result to give explicit convergence rate for uniformly distributed data on the unit sphere. + +# 3.1 CONVERGENCE ANALYSIS OF GRADIENT DESCENT + +In this section we study the convergence of Algorithm 1. Instead of studying the standard convergence of loss function value, we aim to provide a refined analysis on the speed of convergence along different directions defined by the eigenfunctions of $L _ { \kappa }$ . We first introduce the following definitions and notations. + +Let $\{ \lambda _ { i } \} _ { i \geqslant 1 }$ with $\lambda _ { 1 } \geqslant \lambda _ { 2 } \geqslant \cdots$ be the strictly positive eigenvalues of $L _ { \kappa }$ , and $\phi _ { 1 } ( \cdot ) , \phi _ { 2 } ( \cdot ) , \ldots$ be the corresponding orthonormal eigenfunctions. Set $\mathbf v _ { i } = n ^ { - 1 / 2 } ( \phi _ { i } ( { \mathbf { x } } _ { 1 } ) , \dots , \phi _ { i } ( { \mathbf { x } } _ { n } ) ) ^ { \top }$ , $i = 1 , 2 , \dots$ Note that $L _ { \kappa }$ may have eigenvalues with multiplicities larger than 1 and $\lambda _ { i }$ , $i \geqslant 1$ are not distinct. Therefore for any integer $k$ , we define $r _ { k }$ as the sum of the multiplicities of the first $k$ distinct eigenvalues of $L _ { \kappa }$ . Define $\mathbf { V } _ { r _ { k } } = \left( \mathbf { v } _ { 1 } , \ldots , \mathbf { v } _ { r _ { k } } \right)$ . By definition, $\mathbf { v } _ { i }$ , $i \in [ r _ { k } ]$ are rescaled restrictions of orthonormal functions in $L _ { \tau } ^ { 2 } ( \mathbb { S } ^ { d } )$ on the training examples. Therefore we can expect them to form a set of almost orthonomal bases in the vector space $\mathbb { R } ^ { n }$ . The following lemma follows by standard concentration inequality. + +Lemma 3.1. Suppose that $| \phi _ { i } ( \mathbf { x } ) | \leqslant M$ for all $\mathbf { x } \in \mathbb { S } ^ { d }$ and $i \in [ r _ { k } ]$ . For any $\delta > 0$ , with probability at least $1 - \delta$ , + +$$ +\| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { V } _ { r _ { k } } - \mathbf { I } \| _ { \operatorname* { m a x } } \leqslant C M ^ { 2 } \sqrt { \log ( r _ { k } / \delta ) / n } , +$$ + +where $C$ is an absolute constant. + +Denote $\mathbf { y } ~ = ~ ( y _ { 1 } , \ldots , y _ { n } ) ^ { \top }$ and $\widehat { \mathbf { y } } ^ { ( t ) } ~ = ~ \theta ~ \cdot ~ \left( f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { 1 } ) , \ldots , f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { n } ) \right)$ , $t ~ = ~ 0 , \ldots , T$ . Then Lemma 3.1 shows that the convergence rate of $\| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ) } ) \| _ { 2 }$ roughly represents the speed gradient descent learns the components of the target function corresponding to the first $r _ { k }$ eigenvalues. The following theorem gives the convergence guarantee of $\| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ) } ) \| _ { 2 }$ . + +Theorem 3.2. Suppose $| \phi _ { j } ( \mathbf { x } ) | \leqslant M$ for $j \in [ r _ { k } ]$ and $\mathbf { x } \in \mathbb { S } ^ { d }$ . For any $\epsilon , \delta > 0$ and integer $k$ , if $n \geqslant \widetilde { \Omega } ( \epsilon ^ { - 2 } \cdot \operatorname* { m a x } \{ ( \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } ) ^ { - 2 } , M ^ { 4 } r _ { k } ^ { 2 } \} )$ , $m \geqslant \widetilde { \Omega } ( \mathrm { p o l y } ( T , \lambda _ { r _ { k } } ^ { - 1 } , \epsilon ^ { - 1 } ) )$ , then with probability at least $1 - \delta$ , Algorithm 1 with $\eta = \tilde { \mathcal { O } } ( ( m \theta ^ { 2 } ) ^ { - 1 } )$ , $\theta = \widetilde { \mathcal { O } } ( \epsilon )$ satisfies + +$$ +n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } \leqslant 2 ( 1 - \lambda _ { r _ { k } } ) ^ { T } \cdot n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { y } \| _ { 2 } + \epsilon . +$$ + +Remark 3.3. Theorem 3.2 theoretically reveals the spectral bias of deep learning. Specifically, as long as the network is wide enough and the sample size is large enough, gradient descent first learns the target function along the the eigendirections of neural tangent kernel with larger eigenvalues, and then learns the rest components corresponding to smaller eigenvalues. Moreover, by showing that learning the components corresponding to larger eigenvalues can be done with smaller sample size and narrower networks, our theory pushes the study of neural networks in the NTK regime towards a more practical setting. For these reasons, we believe that Theorem 3.2 to certain extent explains the empirical observations given in Rahaman et al. (2018), and demonstrates that the difficulty of a function to be learned by neural network can be characterized in the eigenspace of neural tangent kernel: if the target function has a component corresponding to a small eigenvalue of neural tangent kernel, then learning this function up to good accuracy takes longer time, and requires more examples and a wider network. + +# 3.2 SPECTRAL ANALYSIS OF NEURAL TANGENT KERNEL FOR UNIFORM DISTRIBUTION + +After presenting a general theorem (without assumptions on data distribution) in the previous subsection, we now study the case when the data inputs are uniformly distributed over the unit sphere. We present our results (an extension of Proposition 5 in Bietti and Mairal (2019)) of spectral analysis of neural tangent kernel. We show Mercer decomposition of neural tangent kernel for two-layer setting. We give explicit expression of eigenvalues and show orders of eigenvalues in both cases when $d \gg k$ and $k \gg d$ . + +Theorem 3.4. For any $\mathbf { x } , \mathbf { x } ^ { \prime } \in \mathbb { S } ^ { d } \subset \mathbb { R } ^ { d + 1 }$ , we have the Mercer decomposition of the neural tangent kernel $\kappa : \mathbb { S } ^ { d } \times \mathbb { S } ^ { d } \xrightarrow { } \mathbb { R }$ , + +$$ +\kappa \left( \mathbf { x } , \mathbf { x } ^ { \prime } \right) = \sum _ { k = 0 } ^ { \infty } \mu _ { k } \sum _ { j = 1 } ^ { N \left( d , k \right) } Y _ { k , j } \left( \mathbf { x } \right) Y _ { k , j } \left( \mathbf { x } ^ { \prime } \right) , +$$ + +where $Y _ { k , j }$ for $j = 1 , \cdots , N ( d , k )$ are linearly independent spherical harmonics of degree ˘ $k$ in $d + 1$ variables with N pd, kq “ 2k\`d´1 k\`d´2 and orders of $\mu _ { k }$ are given by + +µ0 “ µ1 “ Ωp1q, µk “ 0, k “ 2j \` 1, +µ $\begin{array} { r l } & { \mathbf { \Phi } ^ { \circ _ { 0 } } \quad \mathsf { \Lambda } ^ { \circ _ { 1 } } \quad \cdots \mathbf { \Phi } ^ { \circ _ { k } } \quad \cdots \quad \mathbf { \Phi } ^ { \circ _ { j } , \mathsf { w } } \quad \cdots \quad \mathbf { \Phi } ^ { \circ _ { j } \quad \bullet _ { k } , \mathsf { s } } , } \\ & { \boldsymbol { \iota } _ { k } = \Omega \left( \operatorname* { m a x } \left\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \right\} \right) , k = 2 j , } \end{array}$ where $j \in \mathbb { N } ^ { + }$ . More specifically, we have $\mu _ { k } = \Omega \left( k ^ { - d - 1 } \right)$ when $k \gg d$ and $\mu _ { k } = \Omega \left( d ^ { - k + 1 } \right)$ when $d \gg k$ , $k = 2 , 4 , 6 , \ldots$ . + +Remark 3.5. In the above theorem, the coefficients $\mu _ { k }$ are actually different eigenvalues of the integral operator $L _ { \kappa }$ on $L _ { \tau _ { d } } ^ { 2 } ( \mathbb { S } ^ { d } )$ defined by + +$$ +L _ { \kappa } ( f ) ( \mathbf { y } ) = \int _ { \mathbb { S } ^ { d } } \kappa ( \mathbf { x } , \mathbf { y } ) f ( \mathbf { x } ) d \tau _ { d } ( \mathbf { x } ) , \quad f \in L _ { \tau _ { d } } ^ { 2 } ( \mathbb { S } ^ { d } ) , +$$ + +where $\tau _ { d }$ is the uniform probability measure on unit sphere $\mathbb { S } ^ { d }$ . Therefore the $\lambda _ { r _ { k } }$ in Theorem 3.2 is just $\mu _ { k - 1 }$ given in Theorem 3.4 when $\tau _ { d }$ is uniform distribution. + +Remark 3.6. Vempala and Wilmes (2018) studied two-layer neural networks with sigmoid activation function, and established guarantees to achieve $\epsilon _ { 0 } + \epsilon$ error with iteration complexity $T = ( d { + } 1 ) ^ { \mathcal { O } ( k ) \log { ( \| f ^ { * } \| _ { 2 } / \epsilon ) } }$ under the over-parameterization condition $m = ( d { + } 1 ) ^ { \mathcal { O } ( k ) \mathrm { p o l y } ( \| f ^ { * } \| _ { 2 } / \epsilon ) }$ , where $f ^ { * }$ is the target function, and $\epsilon _ { \mathrm { 0 } }$ is certain function approximation error. Another highly related work is Bietti and Mairal (2019), which gives $\mu _ { k } = \bar { \Omega ( k ^ { - } { } ^ { d - 1 } ) }$ . The order of eigenvalues we present appears as $\mu _ { k } = \Omega ( \operatorname* { m a x } ( k ^ { - d - 1 } , d ^ { - k + \bar { 1 } } ) )$ . This is better when $d \gg k$ , which is closer to the practical setting. + +# 3.3 EXPLICIT CONVERGENCE RATE FOR UNIFORMLY DISTRIBUTED DATA + +In this subsection, we combine our results in the previous two subsections and give explicit convergence rate for uniformly distributed data on the unit sphere. Note that the first $k$ distinct eigenvalues of NTK have spherical harmonics up to degree $k - 1$ as eigenfunctions. + +Corollary 3.7. Suppose that $k \gg d$ , and the sample $\left\{ { \bf x } _ { i } \right\} _ { i = 1 } ^ { n }$ follows the uniform distribution $\tau _ { d }$ on the unit sphere $\mathbb { S } ^ { d }$ . For any $\epsilon , \delta > 0$ and integer $k$ , if $n \geqslant \widetilde \Omega ( \epsilon ^ { - 2 } \cdot \operatorname* { m a x } \{ k ^ { 2 d + 2 } , k ^ { 2 d - 2 } r _ { k } ^ { 2 } \} )$ , $m \geqslant$ $\tilde { \Omega } ( \mathrm { p o l y } ( T , k ^ { d + 1 } , \epsilon ^ { - 1 } ) )$ , then with probability at least $1 - \delta$ , Algorithm 1 with $\eta = \tilde { \mathcal { O } } ( ( m \theta ^ { 2 } ) ^ { - 1 } )$ , $\theta = \widetilde { \mathcal { O } } ( \epsilon )$ satisfies + +$$ +n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } \leqslant 2 \left( 1 - \Omega \left( k ^ { - d - 1 } \right) \right) ^ { T } \cdot n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { y } \| _ { 2 } + \epsilon , +$$ + +where thono $\begin{array} { r } { r _ { k } \ = \ \sum _ { k ^ { \prime } = 0 } ^ { k - 1 } N ( d , k ^ { \prime } ) } \end{array}$ and s of d $\mathbf { V } _ { r _ { k } } \ = \ ( n ^ { - 1 / 2 } \phi _ { j } ( \mathbf { x } _ { i } ) ) _ { n \times r _ { k } }$ with $\phi _ { 1 } , \ldots , \phi _ { r _ { k } }$ being a set of or$k - 1$ + +Corollary 3.8. Suppose that $d \ \gg \ k$ , and the sample $\left\{ { \bf x } _ { i } \right\} _ { i = 1 } ^ { n }$ follows the uniform distribution $\tau _ { d }$ on the unit sphere $\mathbb { S } ^ { d }$ . For any $\epsilon , \delta \ > \ 0$ and integer $k$ , if $n \geqslant \widetilde \Omega ( \epsilon ^ { - 2 } d ^ { 2 k - 2 } r _ { k } ^ { 2 } )$ , $m \geqslant$ $\tilde { \Omega } ( \mathrm { p o l y } ( T , d ^ { k - 2 } , \epsilon ^ { - 1 } ) )$ , then with probability at least $1 - \delta$ , Algorithm 1 with $\eta = \tilde { \mathcal { O } } ( ( m \theta ^ { 2 } ) ^ { - 1 } )$ , $\theta = \widetilde { \mathcal { O } } ( \epsilon )$ satisfies + +$$ +n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } \leqslant 2 \left( 1 - \Omega \left( d ^ { - k + 2 } \right) \right) ^ { T } \cdot n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { y } \| _ { 2 } + \epsilon , +$$ + +where $\begin{array} { r } { r _ { k } \ = \ \sum _ { k ^ { \prime } = 0 } ^ { k - 1 } N ( d , k ^ { \prime } ) } \end{array}$ and $\mathbf { V } _ { r _ { k } } \ = \ ( n ^ { - 1 / 2 } \phi _ { j } ( \mathbf { x } _ { i } ) ) _ { n \times r _ { k } }$ with $\phi _ { 1 } , \ldots , \phi _ { r _ { k } }$ being a set of orthonomal spherical harmonics of degrees up to $k - 1$ . + +Corollaries 3.7 and 3.8 further illustrate the spectral bias of neural networks by providing exact calculations of $\lambda _ { r _ { k } }$ , $\mathbf { V } _ { r _ { k } }$ and $M$ in Theorem 3.2. They show that if the input distribution is uniform over unit sphere, then spherical harmonics with lower degrees are learned first by over-parameterized neural networks. + +Remark 3.9. In Corollaries 3.7 and 3.8, it shows that the conditions on $n$ and $m$ depend exponentially on either $k$ or $d$ . We would like to emphasize that such exponential dependency is reasonable and unavoidable. In our case, we can take the $d \gg k$ setting as an example. The exponential dependency in $k$ is a natural consequence of the fact that in high dimensional space, there are a large number of linearly independent polynomials even for very low degrees. It is apparently only reasonable to expect to learn less than $n$ independent components of the true function, which means that it is unavoidable to assume + +$$ +n \geqslant r _ { k } = \sum _ { k ^ { \prime } = 0 } ^ { k - 1 } N ( d , k ^ { \prime } ) = \sum _ { k ^ { \prime } = 0 } ^ { k - 1 } { \frac { 2 k ^ { \prime } + d - 1 } { k ^ { \prime } } } { \binom { k ^ { \prime } + d - 2 } { d - 1 } } = \sum _ { k ^ { \prime } = 0 } ^ { k - 1 } { \frac { 2 k ^ { \prime } + d - 1 } { k ^ { \prime } } } { \binom { k ^ { \prime } + d - 2 } { k ^ { \prime } - 1 } } = \Omega ( d ^ { k - 1 } ) . +$$ + +Similar arguments can apply to the requirement of $m$ and the $k \gg d$ setting. + +# 4 EXPERIMENTS + +In this section we illustrate our results by training neural networks on synthetic data. Across all tasks, we train a two-layer hidden neural networks with 4096 neurons and initialize it exactly as defined in the setup. The optimization method is vanilla full gradient descent. We sample 1000 training data which is uniformly sampled from the unit sphere in $\mathbb { R } ^ { 1 0 }$ . + +# 4.1 LEARNING COMBINATION OF SPHERICAL HARMONICS + +First, we show a result when the target function is exactly linear combination of spherical harmonics. The target function is explicitly defined as + +$$ +f ^ { * } ( { \bf x } ) = a _ { 1 } \cdot P _ { 1 } ( \langle \zeta _ { 1 } , { \bf x } \rangle ) + a _ { 2 } \cdot P _ { 2 } ( \langle \zeta _ { 2 } , { \bf x } \rangle ) + a _ { 4 } \cdot P _ { 4 } ( \langle \zeta _ { 4 } , { \bf x } \rangle ) , +$$ + +where the $P _ { k } ( t )$ is the Gegenbauer polynomial, and $\zeta _ { k } , k = 1 , 2 , 4$ are fixed vectors that are inde +pendently generated from uniformaccording to the addition formula $\begin{array} { r } { \sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \bf x } ) Y _ { k , j } ( { \bf y } ) = N ( d , k ) P _ { k } ( \langle { \bf x } , { \bf y } \rangle ) } \end{array}$ $\mathbb { R } ^ { 1 0 }$ riments. Note that, every normalized $f ^ { * } ( \mathbf { x } )$ +monics of order 1,2 and 4. The higher odd-order Gegenbauer polynomials are omitted because the +spectral analysis showed that $\mu _ { k } = 0$ for $k = 3 , 5 , 7 \dots$ . + +Following the setting in section 3.1, we denote $\mathbf v _ { k } = n ^ { - 1 / 2 } ( P _ { k } ( \mathbf x _ { 1 } ) , \dots , P _ { k } ( \mathbf x _ { n } ) )$ . By Lemma 3.2 $\mathbf { v } _ { k }$ ’s are almost orthonormal with high probability. So we can define the (approximate) projection length of residual $\mathbf { r } ^ { ( t ) }$ onto $\mathbf { v } _ { k }$ at step $t$ as + +$$ +\widehat { a } _ { k } = | \mathbf { v } _ { k } ^ { \top } \mathbf { r } ^ { ( t ) } | , +$$ + +where $\mathbf { r } ^ { ( t ) } = \left( f ^ { * } \big ( \mathbf { x } _ { 1 } \big ) - \theta f _ { \mathbf { W } ^ { ( t ) } } \big ( \mathbf { x } _ { 1 } \big ) , \ldots , f ^ { * } \big ( \mathbf { x } _ { n } \big ) - \theta f _ { \mathbf { W } ^ { ( t ) } } \big ( \mathbf { x } _ { n } \big ) \right)$ and $f _ { \mathbf { W } ( t ) } \left( \mathbf { x } \right)$ is the neural network function. + +Remark 4.1. Here $\widehat { a } _ { k }$ is the projection length onto an approximate vector. In the function space, we can also project the residual function $r ( \mathbf { x } ) \bar { = } f ^ { * } ( \mathbf { x } ) - \bar { \theta } \bar { f } _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } )$ onto the orthonormal Gegenbauer functions $P _ { k } ( { \bf x } )$ . Replacing the training data with randomly sampled data points $\mathbf { x } _ { i }$ can lead to a random estimate of the projection length in function space. We provide the corresponding result for freshly sampled points in Appendix E.1. + +![](images/2771d3c8fdd44ee955473f67bdb5bc958722cdc8a8e43ad240100a6f7d39531f.jpg) +Figure 1: Convergence curve for projection length onto different components. (a) shows the curve when the target function have different component with the same scale. (b) shows the curve when the higner-order components have larger scale. Both illustrate that the lower-order components are learned faster. Log-scale figures are shown in Appendix E.2. + +The results are showned in Figure 1. It can be seen that at the beginning of training, the residual at the lowest frequency ( $k = 1 .$ ) converges to zero first and then the second lowest $k = 2$ ). The highest frequency component is the last one to converge. Following the setting of Rahaman et al. (2018) we assign high frequencies a larger scale, expecting that larger scale will introduce a better descending speed. Still, the low frequencies are regressed first. + +Apart from the synthesized low frequency function, we also showed the dynamics of normal functions’ projection to $P _ { k } ( x )$ . These functions, though in a simple form, have non-zero components in almost all frequencies. In this subsection we further show our results still apply when all frequenciesř ř exist in the target function, which is given by $\begin{array} { r } { f ^ { * } ( \mathbf { x } ) = \sum _ { i } \cos ( a _ { i } \langle \zeta , \mathbf { x } \rangle ) } \end{array}$ or $\begin{array} { r } { f ^ { * } ( \mathbf { x } ) = \sum _ { i } \bar { \langle \zeta , \mathbf { x } \rangle } ^ { p _ { i } } } \end{array}$ , where $\zeta$ is a fixed unit vector. The coefficients of given components are calculated in the same way as in Section 4.1. + +![](images/edeb7fc8e44d199514a417b12ca7990246224162dede7c2aaebaca80a02339cc.jpg) +Figure 2: Convergence curve for different components. (a) shows the curve of a trigonometric function. (b) shows the curve of a polynomial with even degrees. Both exhibits similar tendency as combination of spherical harmonics. + +Figure 2 shows that even for arbitrarily chosen functions of simple form, the networks can still first learn the low frequency components of the target function. Notice that early in training not all the curves may descend, we believe this is due to the unseen components’ influence on the gradient. Again, as the training proceeds, the convergence is controlled at the predicted rate. + +Remark 4.2. The reason why we only use cosine function and even polynomial is that the only odd basis function with non-zero eigenvalue is $P _ { 1 } ( \mathbf { x } )$ . To show a general tendency it is better to restrict the target function in the even function space. + +# 5 CONCLUSION AND DISCUSSION + +In this paper, we give theoretical justification for spectral bias through a detailed analysis of the convergence behavior of two-layer neural networks with ReLU activation function. We show that the convergence of gradient descent in different directions depends on the corresponding eigenvalues and essentially exhibits different convergence rates. We show Mercer decomposition of neural tangent kernel and give explicit order of eigenvalues of integral operator with respect to the neural tangent kernel when the data is uniformly distributed on the unit sphere $\mathbb { S } ^ { d }$ . Combined with the convergence analysis, we give the exact order of convergence rate on different directions. We also conduct experiments on synthetic data to support our theoretical result. + +So far, we have considered the upper bound for convergence with respect to low frequency components and present comprehensive theorem to explain the spectral bias. One desired improvement is to give the lower bound of convergence with respect to high frequency components, which is essential to establish tighter characterization of spectral-biased optimization. It is also interesting to extend our result to other training algorithms like Adam, where the analysis in Wu et al. (2019); Zhou et al. (2018) might be implemented with a more careful quantification on the projection of residual along different directions. Another potential improvement is to generalize the result to multi-layer neural networks, which might require different techniques since our analysis heavily rely on exactly computing the eigenvalues of the neural tangent kernel. It is also an important direction to weaken the requirement on over-parameterization, or study the spectral bias in a non-NTK regime to furthur close the gap between theory and practice. + +REFERENCES +ALLEN-ZHU, Z., LI, Y. and SONG, Z. (2018). A convergence theory for deep learning via overparameterization. arXiv preprint arXiv:1811.03962 . +ANDONI, A., PANIGRAHY, R., VALIANT, G. and ZHANG, L. (2014). Learning polynomials with neural networks. In International Conference on Machine Learning. +ARORA, S., DU, S. S., HU, W., LI, Z., SALAKHUTDINOV, R. and WANG, R. (2019a). On exact computation with an infinitely wide neural net. arXiv preprint arXiv:1904.11955 . +ARORA, S., DU, S. S., HU, W., LI, Z. and WANG, R. (2019b). Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. arXiv preprint arXiv:1901.08584 . +ATKINSON, K. and HAN, W. (2012). Spherical harmonics and approximations on the unit sphere: an introduction, vol. 2044. Springer Science & Business Media. +BACH, F. (2017). Breaking the curse of dimensionality with convex neural networks. Journal of Machine Learning Research 18 1–53. +BASRI, R., JACOBS, D., KASTEN, Y. and KRITCHMAN, S. (2019). The convergence rate of neural networks for learned functions of different frequencies. arXiv preprint arXiv:1906.00425 . +BIETTI, A. and MAIRAL, J. (2019). On the inductive bias of neural tangent kernels. arXiv preprint arXiv:1905.12173 . +CAO, Y. and GU, Q. (2019a). Generalization bounds of stochastic gradient descent for wide and deep neural networks. arXiv preprint arXiv:1905.13210 . +CAO, Y. and GU, Q. (2019b). A generalization theory of gradient descent for learning overparameterized deep relu networks. arXiv preprint arXiv:1902.01384 . +CHIZAT, L. and BACH, F. (2018). A note on lazy training in supervised differentiable programming. arXiv preprint arXiv:1812.07956 . +CHO, Y. and SAUL, L. K. (2009). Kernel methods for deep learning. In Advances in neural information processing systems. +COLLOBERT, R. and WESTON, J. (2008). A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning. ACM. +DU, S. S., LEE, J. D., LI, H., WANG, L. and ZHAI, X. (2018a). Gradient descent finds global minima of deep neural networks. arXiv preprint arXiv:1811.03804 . +DU, S. S., ZHAI, X., POCZOS, B. and SINGH, A. (2018b). Gradient descent provably optimizes over-parameterized neural networks. arXiv preprint arXiv:1810.02054 . +FRYE, C. and EFTHIMIOU, C. J. (2012). Spherical harmonics in p dimensions. arXiv preprint arXiv:1205.3548 . +GUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018a). Characterizing implicit bias in terms of optimization geometry. arXiv preprint arXiv:1802.08246 . +GUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018b). Implicit bias of gradient descent on linear convolutional networks. arXiv preprint arXiv:1806.00468 . +HE, K., ZHANG, X., REN, S. and SUN, J. (2015). Delving deep into rectifiers: Surpassing humanlevel performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision. +HE, K., ZHANG, X., REN, S. and SUN, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition. +HINTON, G., DENG, L., YU, D., DAHL, G., MOHAMED, A.-R., JAITLY, N., SENIOR, A., VANHOUCKE, V., NGUYEN, P., KINGSBURY, B. ET AL. (2012). Deep neural networks for acoustic modeling in speech recognition. IEEE Signal processing magazine 29. +JACOT, A., GABRIEL, F. and HONGLER, C. (2018). Neural tangent kernel: Convergence and generalization in neural networks. arXiv preprint arXiv:1806.07572 . +LI, Y. and LIANG, Y. (2018). Learning overparameterized neural networks via stochastic gradient descent on structured data. arXiv preprint arXiv:1808.01204 . +NAKKIRAN, P., KAPLUN, G., KALIMERIS, D., YANG, T., EDELMAN, B. L., ZHANG, F. and BARAK, B. (2019). Sgd on neural networks learns functions of increasing complexity. arXiv preprint arXiv:1905.11604 . +RAHAMAN, N., BARATIN, A., ARPIT, D., DRAXLER, F., LIN, M., HAMPRECHT, F. A., BENGIO, Y. and COURVILLE, A. (2018). On the spectral bias of neural networks. arXiv preprint arXiv:1806.08734 . +ROSASCO, L., BELKIN, M. and VITO, E. D. (2010). On learning with integral operators. Journal of Machine Learning Research 11 905–934. +SMALE, S. and ZHOU, D.-X. (2007). Learning theory estimates via integral operators and their approximations. Constructive approximation 26 153–172. +SMALE, S. and ZHOU, D.-X. (2009). Geometry on probability spaces. Constructive Approximation 30 311. +SOUDRY, D., HOFFER, E. and SREBRO, N. (2017). The implicit bias of gradient descent on separable data. arXiv preprint arXiv:1710.10345 . +SU, L. and YANG, P. (2019). On learning over-parameterized neural networks: A functional approximation prospective. arXiv preprint arXiv:1905.10826 . +VEMPALA, S. and WILMES, J. (2018). Gradient descent for one-hidden-layer neural networks: Polynomial convergence and sq lower bounds. arXiv preprint arXiv:1805.02677 . +WU, X., DU, S. S. and WARD, R. (2019). Global convergence of adaptive gradient methods for an over-parameterized neural network. arXiv preprint arXiv:1902.07111 . +XU, Z. J. (2018). Understanding training and generalization in deep learning by fourier analysis. arXiv preprint arXiv:1808.04295 . +YANG, G. and SALMAN, H. (2019). A fine-grained spectral perspective on neural networks. arXiv preprint arXiv:1907.10599 . +ZHANG, C., BENGIO, S., HARDT, M., RECHT, B. and VINYALS, O. (2016). Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530 . +ZHOU, D., TANG, Y., YANG, Z., CAO, Y. and GU, Q. (2018). On the convergence of adaptive gradient methods for nonconvex optimization. arXiv preprint arXiv:1808.05671 . +ZOU, D., CAO, Y., ZHOU, D. and GU, Q. (2018). Stochastic gradient descent optimizes overparameterized deep relu networks. arXiv preprint arXiv:1811.08888 . +ZOU, D. and GU, Q. (2019). An improved analysis of training over-parameterized deep neural networks. arXiv preprint arXiv:1906.04688 . + +# A REVIEW ON SPHERICAL HARMONICS + +In this section, we give a brief review on relevant concepts in spherical harmonics. For more detials, see Bach (2017), Bietti and Mairal (2019), Frye and Efthimiou (2012) and Atkinson and Han (2012) for references. + +We consider the unit sphere $\mathbb { S } ^ { d } = \left\{ \mathbf { x } \in \mathbb { R } ^ { d + 1 } : \| \mathbf { x } \| _ { 2 } = 1 \right\}$ , whose surface area is given by $\omega _ { d } =$ $2 \pi ^ { ( d + 1 ) / 2 } / \Gamma ( ( d + 1 ) / 2 )$ and denote $\tau _ { d }$ the uniform measure on the sphere. + +For any ! $k \qquad \geqslant \qquad 1 ,$ we consider) a set of spherical harmonics $\begin{array} { r } { \{ Y _ { k , j } : \mathbb { S } ^ { d } \mathbb { R } | 1 \leqslant j \leqslant N ( d , k ) = \frac { 2 k + d - 1 } { k } \binom { k + d - 2 } { d - 1 } \} } \end{array}$ They form an orthonormal basis and satisfy the following equation $\langle Y _ { k i } , Y _ { s j } \rangle _ { \mathbb { S } ^ { d } } = \int _ { \mathbb { S } ^ { d } } Y _ { k i } ( x ) Y _ { s j } ( x ) d \tau _ { d } ( x ) = \delta _ { i j } \delta _ { s k }$ . Moreover, since they are homogeneous functions of degree $k$ , it is clear that $Y _ { k } ( x )$ has the same parity as $k$ . + +We have the addition formula + +$$ +\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \bf x } ) Y _ { k , j } ( { \bf y } ) = N ( d , k ) P _ { k } ( \langle { \bf x } , { \bf y } \rangle ) , +$$ + +where $P _ { k } ( t )$ is the Legendre polynomial of degree $k$ in $d + 1$ dimensions, explicitly given by (Rodrigues’ formula) + +$$ +P _ { k } ( t ) = \left( - { \frac { 1 } { 2 } } \right) ^ { k } { \frac { \Gamma \left( { \frac { d } { 2 } } \right) } { \Gamma \left( k + { \frac { d } { 2 } } \right) } } \left( 1 - t ^ { 2 } \right) ^ { { \frac { 2 - d } { 2 } } } \left( { \frac { d } { d t } } \right) ^ { k } \left( 1 - t ^ { 2 } \right) ^ { k + { \frac { d - 2 } { 2 } } } . +$$ + +We can also see that $P _ { k } ( t )$ , the Legendre polynomial of degree $k$ shares the same parity with $k$ . By the orthogonality and the addition formula (A.1) we have, + +$$ +\int _ { \mathbb { S } ^ { d } } P _ { j } ( \langle \mathbf { w } , \mathbf { x } \rangle ) P _ { k } ( \langle \mathbf { w } , \mathbf { y } \rangle ) d \tau _ { d } ( \mathbf { w } ) = \frac { \delta _ { j k } } { N ( d , k ) } P _ { k } ( \langle \mathbf { x } , \mathbf { y } \rangle ) . +$$ + +Further we have the recurrence relation for the Legendre polynomials, + +$$ +t P _ { k } ( t ) = \frac { k } { 2 k + d - 1 } P _ { k - 1 } ( t ) + \frac { k + d - 1 } { 2 k + d - 1 } P _ { k + 1 } ( t ) , +$$ + +for $k \geqslant 1$ and $t P _ { 0 } ( t ) = P _ { 1 } ( t )$ for $k = 0$ . + +The Hecke-Funk formula is given for a spherical harmonic $Y _ { k }$ of degree $k$ + +$$ +\int _ { \mathbb { S } ^ { d } } f ( \langle \mathbf { x } , \mathbf { y } \rangle ) Y _ { k } ( \mathbf { y } ) d t ^ { \prime } d ( \mathbf { y } ) = { \frac { \omega _ { d - 1 } } { \omega _ { d } } } Y _ { k } ( \mathbf { x } ) \int _ { - 1 } ^ { 1 } f ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t . +$$ + +# B PROOF OF MAIN THEOREMS + +# B.1 PROOF OF THEOREM 3.2 + +In this section we give the proof of Theorem 3.2. The core idea of our proof is to establish connections between neural network gradients throughout training and the neural tangent kernel. To do so, we first introduce the following definitions and notations. + +Define $\begin{array} { r l r } { \mathbf { K } ^ { ( 0 ) } } & { { } = } & { m ^ { - 1 } ( \left. \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { j } ) \right. ) _ { n \times n } } \end{array}$ , $\begin{array} { r l r } { { \bf K } ^ { ( \infty ) } } & { { } = } & { ( \kappa ( { \bf x } _ { i } , { \bf x } _ { j } ) ) _ { n \times n } \quad = } \end{array}$ $\mathrm { l i m } _ { m \infty } \mathbf { K } ^ { ( 0 ) }$ . Let $\{ \widehat \lambda _ { i } \} _ { i = 1 } ^ { n } , \widehat \lambda _ { 1 } \geqslant \ \cdot \cdot \ \geqslant \ \widehat \lambda _ { n }$ be the eigenvalues of $n ^ { - 1 } \mathbf { K } ^ { \infty }$ , and $\widehat { \mathbf { v } } _ { 1 } , \ldots , \widehat { \mathbf { v } } _ { n }$ be the corresponding eigenvectors. Set $\widehat { \mathbf { V } } _ { r _ { k } } = ( \widehat { \mathbf { v } } _ { 1 } , \ldots , \widehat { \mathbf { v } } _ { r _ { k } } ) , \widehat { \mathbf { V } } _ { r _ { k } } ^ { \perp } = ( \widehat { \mathbf { v } } _ { r _ { k } + 1 } , \ldots , \widehat { \mathbf { v } } _ { n } )$ . + +For notation simplicity, we denote ˇ $\nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \ = \ \left[ \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) \right] \big | _ { \mathbf { W } = \mathbf { W } ^ { ( 0 ) } } , \ \nabla _ { \mathbf { W } _ { l } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \ =$ $\left[ \nabla _ { \mathbf { W } _ { l } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) \right] \big | _ { \mathbf { W } = \mathbf { W } ^ { ( 0 ) } } , l = 1 , 2 .$ + +The following lemma is a direct application of Proposition 1 in Smale and Zhou (2009) or Proposition 10 in Rosasco et al. (2010). It bounds the difference between the eigenvalues of NTK and their finite-width counterparts. + +Lemma B.1. For any $\delta > 0$ , with probability at least $1 - \delta$ , + +$$ +| \lambda _ { i } - \widehat { \lambda } _ { i } | \leqslant \mathcal { O } ( \sqrt { \log ( 1 / \delta ) / n } ) . +$$ + +The following lemma is partly summarized from the proof of equation (44) in $\mathrm { S u }$ and Yang (2019). Its purpose is to further connect the eigenfunctions of NTK with their finite-width, finite-sample counterparts. + +Lemma B.2. Suppose that $| \phi _ { i } ( \mathbf { x } ) | \leqslant M$ for all $\mathbf { x } \in S ^ { d }$ . There exist absolute constants $C , C ^ { \prime } , c ^ { \prime \prime } >$ 0, such that for any $\delta > 0$ and integer $k$ with $r _ { k } \leqslant n$ , if $n \geqslant C ( \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } ) ^ { - 2 } \log ( 1 / \delta )$ , then with probability at least $1 - \delta$ , + +$$ +\begin{array} { r l } & { \| \mathbf { V } _ { r _ { k } } ^ { \top } \widehat { \mathbf { V } } _ { r _ { k } } ^ { \bot } \| _ { F } \leqslant C ^ { \prime } \frac { 1 } { \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } } \cdot \sqrt { \frac { \log ( 1 / \delta ) } { n } } , } \\ & { \| \mathbf { V } _ { r _ { k } } \mathbf { V } _ { r _ { k } } ^ { \top } - \widehat { \mathbf { V } } _ { r _ { k } } \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } \| _ { 2 } \leqslant C ^ { \prime \prime } \bigg [ \frac { 1 } { \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } } \cdot \sqrt { \frac { \log ( 1 / \delta ) } { n } } + M ^ { 2 } r _ { k } \sqrt { \frac { \log ( r _ { k } / \delta ) } { n } } \bigg ] . } \end{array} +$$ + +The following two lemmas are key to characterize the dynamics of the residual throughout training. These two results, especially Lemma B.4, are the ones that distinguish our analysis from previous works on neural network training in the NTK regime (Du et al., 2018b; Su and Yang, 2019), since our analysis provides more careful characterization on the residual along different directions. + +Lemma B.3. Suppose that the iterates of gradient descent $\mathbf { W } ^ { ( 0 ) } , \ldots , \mathbf { W } ^ { ( t ) }$ are inside the ball $B ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . If $\omega \leqslant \mathcal { O } ( [ \log ( m ) ] ^ { - 3 / 2 } )$ , then with probability at least $1 - \mathcal { O } ( n ^ { 2 } ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ , + +$$ +\mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ^ { \prime } + 1 ) } = [ \mathbf { I } - ( \eta m \theta ^ { 2 } / n ) \mathbf { K } ^ { \infty } ] ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ^ { \prime } ) } ) + \mathbf { e } ^ { ( t ) } , \ \| \mathbf { e } ^ { ( t ^ { \prime } ) } \| _ { 2 } \leqslant \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ^ { \prime } ) } \| _ { 2 } +$$ + +for all $t ^ { \prime } = 0 , \ldots , t - 1$ , where $\mathbf { y } = \left( y _ { 1 } , \ldots , y _ { n } \right) ^ { \top } , \widehat { \mathbf { y } } ^ { ( t ^ { \prime } ) } = \theta \cdot \left( f _ { \mathbf { W } ^ { ( t ^ { \prime } ) } } ( \mathbf { x } _ { 1 } ) , \ldots , f _ { \mathbf { W } ^ { ( t ^ { \prime } ) } } ( \mathbf { x } _ { n } ) \right) ^ { \top } .$ + +Lemma B.4. Suppose that the iterates of gradient descent $\mathbf { W } ^ { ( 0 ) } , \ldots , \mathbf { W } ^ { ( t ) }$ are inside the ball $B ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . If $\omega \leqslant \widetilde { \mathcal { O } } ( \operatorname* { m i n } \{ [ \log ( m ) ] ^ { - 3 / 2 } , \lambda _ { r _ { k } } ^ { 3 } , ( \eta m ) ^ { - 3 } \} )$ and $n \geqslant \widetilde { \mathcal { O } } ( \lambda _ { r _ { k } } ^ { - 2 } )$ , then with probability at least $1 - \mathcal { O } ( t ^ { 2 } n ^ { 2 } ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ , + +$$ +\begin{array} { r l r } { { ( \widehat { \mathbf { V } } _ { r k } ^ { \perp } ) ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ^ { \prime } ) } ) \| _ { 2 } \leqslant \| ( \widehat { \mathbf { V } } _ { r k } ^ { \perp } ) ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } + t ^ { \prime } \cdot \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } \cdot \sqrt { n } \cdot \widetilde { O } ( 1 + \omega \sqrt { m } ) \qquad ( \mathbb { B } . 1 ) } } \\ & { } & { \| \widehat { \mathbf { V } } _ { r k } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ^ { \prime } ) } ) \| _ { 2 } \leqslant \big ( 1 - \eta m \theta ^ { 2 } \lambda _ { r k } / 2 \big ) t ^ { \prime } \| \widehat { \mathbf { V } } _ { r k } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } + t ^ { \prime } \lambda _ { r k } ^ { - 1 } \cdot \omega ^ { 2 / 3 } \eta m \theta ^ { 2 } \cdot \sqrt { n } \cdot \widetilde { O } ( 1 + \omega \sqrt { n } ) } \\ & { } & { + \lambda _ { r k } ^ { - 1 } \cdot \widetilde { O } ( \omega ^ { 1 / 3 } ) \cdot \| ( \widehat { \mathbf { V } } _ { r k } ^ { \perp } ) ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } } \\ & { } & { \| \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ^ { \prime } ) } \| _ { 2 } \leqslant \widetilde { \mathcal { O } } ( \sqrt { n } ) \cdot \big ( 1 - \eta m \theta ^ { 2 } \lambda _ { r k } / 2 \big ) ^ { t ^ { \prime } } + \widetilde { \mathcal { O } } ( \sqrt { n } \cdot ( \eta m \theta ^ { 2 } \lambda _ { r k } ) ^ { - 1 } ) } \\ & { } & { + \lambda _ { r k } ^ { - 1 } t ^ { \prime } \omega ^ { 1 / 3 } \cdot \sqrt { n } \cdot \widetilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) \qquad ( \mathbb { B } . 3 ) } \end{array} +$$ + +for all $t ^ { \prime } = 0 , \ldots , t - 1$ + +Now we are ready to prove Theorem 3.2. + +Proof of Theorem 3.2. Define $\omega \ = \ \overline { { C } } T / ( \lambda _ { r _ { k } } \sqrt { m } )$ for some small enough absolute constant $\overline { { C } }$ . Then by union bound, as long as $m \geqslant \widetilde \Omega ( \mathrm { p o l y } ( T , \lambda _ { r _ { k } } ^ { - 1 } , \epsilon ^ { - 1 } ) )$ , the conditions on $\omega$ given in Lemmas D.2, D.4, D.5, B.3 and B.4 are all satisfied. + +We first show that all the iterates $\mathbf { W } ^ { ( 0 ) } , \ldots , \mathbf { W } ^ { ( T ) }$ are inside the ball $B ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . We prove this result by inductively show that $\mathbf { W } ^ { ( t ) } \in \mathcal { B } ( \mathbf { W } ^ { ( 0 ) } , \omega ) , t = 0 , \ldots , T$ . First of all, it is clear that ${ \bf W } ^ { ( 0 ) } \in$ $B ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . Suppose that $\mathbf { W } ^ { ( 0 ) } , \ldots , \mathbf { W } ^ { ( t ) } \in \mathcal { B } ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . Then the results of Lemmas D.2, D.4, + +D.5, B.3 and B.4 hold for $\mathbf { W } ^ { ( 0 ) } , \ldots , \mathbf { W } ^ { ( t ) }$ . Denote $\mathbf { u } ^ { ( t ) } = \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ) }$ , $t \in T$ . Then we have + +$$ +\begin{array} { r l } { \| \mathbf { W } _ { i } ^ { ( t + 1 ) } - \mathbf { W } _ { i } ^ { ( 0 ) } \| _ { F } \leqslant } & { \displaystyle \sum _ { \ell = 1 } ^ { t } | \mathbf { W } _ { i } ^ { ( t + 1 ) } - \mathbf { W } _ { i } ^ { ( t ^ { \prime } ) } | \| _ { F } } \\ & { \qquad = \displaystyle \prod _ { \ell = 0 } ^ { t } \left| \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \left( g _ { \ell } - \vartheta \cdot f _ { \mathbf { W } ^ { ( i ) } } ( \mathbf { x } _ { i } ) \right) \cdot \vartheta \cdot \nabla _ { \mathbf { W } _ { i } } f _ { \mathbf { W } ^ { ( i ) } } ( \mathbf { x } _ { i } ) \right| \Bigg | _ { F } } \\ & { \leqslant \eta \displaystyle \sum _ { \ell = 0 } ^ { t } \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \| g _ { \ell } - \vartheta \cdot f _ { \mathbf { W } ^ { ( i ) } } ( \mathbf { x } _ { i } ) \| \cdot \| \nabla _ { \mathbf { W } _ { i } } f _ { \mathbf { W } ^ { ( i ) } } ( \mathbf { x } _ { i } ) \| _ { F } } \\ & { \leqslant C _ { 1 } \eta \displaystyle \theta \sqrt { \eta } \displaystyle \sum _ { \ell = 0 } ^ { t } \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \| g _ { \ell } - \vartheta \cdot f _ { \mathbf { W } ^ { ( i ) } } ( \mathbf { x } _ { i } ) \| } \\ & { \leqslant C _ { 1 } \eta \displaystyle \theta \sqrt { \eta } \displaystyle \sum _ { \ell = 0 } ^ { t } \frac { 1 } { n - 1 } \| \mathbf { y } - \vartheta \cdot \nabla _ { \mathbf { W } ^ { ( i ) } } ( \mathbf { x } _ { i } ) \| } \\ & { \leqslant C _ { 1 } \eta \displaystyle \theta \sqrt { \eta } \displaystyle \sum _ { \ell = 0 } ^ { t } \| \mathbf { y } - \widehat { \mathbf { g } } ^ { ( \ell ^ { \prime } ) } \| _ { 2 } , } \end{array} +$$ + +where the second inequality follows by Lemma D.4. By Lemma B.4, we have + +$$ +\sum _ { t ^ { \prime } = 0 } ^ { t } \| \mathbf { y } - \hat { \mathbf { y } } ^ { ( t ^ { \prime } ) } \| _ { 2 } \leqslant \widetilde { \mathcal { O } } ( \sqrt { n } / ( \eta m \theta ^ { 2 } \lambda _ { r _ { k } } ) ) + \widetilde { \mathcal { O } } ( T \sqrt { n } / ( \eta m \theta ^ { 2 } \lambda _ { r _ { k } } ) ) + \lambda _ { r _ { k } } ^ { - 1 } T ^ { 2 } \omega ^ { 1 / 3 } \sqrt { n } \cdot \widetilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) +$$ + +It then follows by the choice $\omega \ = \ \overline { { C } } T / ( \theta \lambda _ { r _ { k } } \sqrt { m } )$ , $\eta ~ = ~ { \tilde { \mathcal { O } } } ( ( m \theta ^ { 2 } \lambda _ { r _ { k } } ) ^ { - 1 } )$ , $\theta \ : = \ : \widetilde { \mathcal { O } } ( \epsilon )$ and the assumption $m \geqslant \tilde { \mathcal { O } } ( \mathrm { p o l y } ( T , \lambda _ { r _ { k } } ^ { - 1 } , \epsilon ^ { - 1 } ) )$ that $\| \mathbf { W } _ { l } ^ { ( t + 1 ) } - \mathbf { W } _ { l } ^ { ( 0 ) } \| _ { F } \leqslant \omega$ , $l = 1 , 2$ . Therefore by induction, we see that with probability at least $1 { - } \mathcal { O } ( T ^ { 3 } n ^ { 2 } ) { \cdot } \exp [ { - \Omega ( m \omega ^ { 2 / 3 } ) } ]$ s, $\mathbf { W } ( 0 ) , \ldots , \mathbf { W } ( T ) \in$ $B ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . + +Applying Lemma B.4 then gives + +$$ +\begin{array} { r l } & { n ^ { - 1 / 2 } \cdot \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } \leqslant ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { T } \cdot n ^ { - 1 / 2 } \cdot \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } } \\ & { \qquad + T \lambda _ { r _ { k } } ^ { - 1 } \cdot \omega ^ { 2 / 3 } \eta m \theta ^ { 2 } \cdot \widetilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) } \\ & { \qquad + \lambda _ { r _ { k } } ^ { - 1 } \cdot \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } ) \cdot n ^ { - 1 / 2 } \cdot \| ( \widehat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } . } \end{array} +$$ + +Now by $\omega = \overline { { C } } T / ( \lambda _ { r _ { k } } \sqrt { m } )$ , $\eta = \widetilde { \mathcal { O } } ( \theta ^ { 2 } m ) ^ { - 1 }$ and the assumption that $m \geqslant m ^ { * } = \widetilde { \mathcal { O } } ( \lambda _ { r _ { k } } ^ { - 1 4 } \cdot \epsilon ^ { - 6 } )$ , we obtain + +$$ +\begin{array} { r } { n ^ { - 1 / 2 } \cdot \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } \leqslant ( 1 - \lambda _ { r _ { k } } ) ^ { T } \cdot n ^ { - 1 / 2 } \cdot \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } + \epsilon / 1 6 . } \end{array} +$$ + +By Lemma 3.1, $\theta = \widetilde { \mathcal { O } } ( \epsilon )$ and the assumptions $n \geqslant \widetilde \Omega ( \operatorname* { m a x } \{ \epsilon ^ { - 1 } ( \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } ) ^ { - 1 } , \epsilon ^ { - 2 } M ^ { 2 } r _ { k } ^ { 2 } \} )$ , the eigenvalues of $\mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { V } _ { r _ { k } }$ are all between $1 / \sqrt { 2 }$ and $\sqrt { 2 }$ . Therefore by Lemma B.2 we have + +$$ +\begin{array} { r l } & { \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \hat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } = \| \hat { \mathbf { V } } _ { r _ { k } } \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \hat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } } \\ & { \qquad \geqslant \| \mathbf { V } _ { r _ { k } } \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \hat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } - \| ( \mathbf { V } _ { r _ { k } } \mathbf { V } _ { r _ { k } } ^ { \top } - \hat { \mathbf { V } } _ { r _ { k } } \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } ) ( \mathbf { y } - \hat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } } \\ & { \qquad \geqslant \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \hat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } / \sqrt { 2 } - \sqrt { n } \cdot \mathcal { O } \biggl ( \frac { 1 } { \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } } \cdot \sqrt { \frac { \log ( 1 / \delta ) } { n } } + M r _ { k } \sqrt { \frac { \log ( r _ { k } } { n } } } \\ & { \qquad \geqslant \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \hat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } / \sqrt { 2 } - \epsilon \sqrt { n } / 1 6 . } \end{array} +$$ + +Similarly, + +$$ +\begin{array} { r l } & { \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } \leqslant \sqrt { 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( 0 ) } ) \| _ { 2 } + \epsilon \sqrt { n } / 1 6 } \\ & { \qquad \leqslant \sqrt { 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { y } \| _ { 2 } + \sqrt { 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } \widehat { \mathbf { y } } ^ { ( 0 ) } \| _ { 2 } + \epsilon \sqrt { n } / 1 6 . } \end{array} +$$ + +By Lemma 3.1, with probability at least $1 - \delta$ , $\| \mathbf { V } _ { r _ { k } } ^ { \top } \| _ { 2 } \leqslant 1 + C r _ { k } M ^ { 2 } \sqrt { \log ( r _ { k } / \delta ) / n }$ . Combining this result with Lemma D.3 gives $\| \mathbf { V } _ { r _ { k } } ^ { \top } \widehat { \mathbf { y } } ^ { ( 0 ) } \| _ { 2 } \leqslant \theta \mathcal { O } ( \sqrt { n \log \left( n / \delta \right) } ) \leqslant \epsilon \sqrt { n } / 8$ . Plugging the above estimates into (B.4) gives + +$$ +n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } ( \mathbf { y } - \widehat { \mathbf { y } } ^ { ( T ) } ) \| _ { 2 } \leqslant 2 ( 1 - \lambda _ { r _ { k } } ) ^ { T } \cdot n ^ { - 1 / 2 } \cdot \| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { y } \| _ { 2 } + \epsilon . +$$ + +Applying union bounds completes the proof. + +# B.2 PROOF OF THE THEOREM 3.4 + +Proof of the Theorem 3.4. The idea of the proof is close to that of Proposition 5 in (Bietti and Mairal, 2019) where they consider $k \gg d$ and we present a more general case including ş $k \gg d$ and $d \gg k$ . For any function $g : \mathbb { S } ^ { d } \mathbb { R }$ , by denoting $g _ { 0 } ( \mathbf x ) = \int _ { \mathbb { S } ^ { d } } g ( \mathbf { \bar { y } } ) d \tau _ { d } ( \mathbf y )$ , it can be decomposed as + +$$ +\begin{array} { l } { { \displaystyle g ( { \bf x } ) = \sum _ { k = 0 } ^ { \infty } g _ { k } ( { \bf x } ) = \sum _ { k = 0 } ^ { \infty } \sum _ { j = 1 } ^ { N ( d , k ) } \int _ { { \mathbb { S } } ^ { d } } Y _ { k j } ( { \bf y } ) Y _ { k j } ( { \bf x } ) g ( { \bf y } ) d \tau _ { d } ( { \bf y } ) } } \\ { { \displaystyle ~ = \sum _ { k = 0 } ^ { \infty } N ( d , k ) \int _ { { \mathbb { S } } ^ { d } } g ( { \bf y } ) P _ { k } ( \langle { \bf x } , { \bf y } \rangle ) d \tau _ { d } ( { \bf y } ) } , } \end{array} +$$ + +where we project function $g$ to spherical harmonics in the second equality and apply the addition equation in the last equality. + +For a positive-definite dot-product kernel $\kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } ) : \mathbb { S } ^ { d } \times \mathbb { S } ^ { d } \mathbb { R }$ which has the form $\kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } ) =$ $\widehat { \kappa } ( \langle \mathbf { x } , \mathbf { \bar { x } } ^ { \prime } \rangle )$ for $\widehat { \kappa } : [ - 1 , 1 ] \ \mathrm { ~ ~ } \mathbb { R }$ , we can present a decomposition by (B.5) if we consider $g ( \mathbf { x } ) =$ $\phi ( \langle \mathbf { x } , \mathbf { z } \rangle )$ for $\mathbf { z } \in \mathbb { S } ^ { \bar { d } }$ and $\bar { \phi } : [ - 1 , 1 ] \to \mathbb { R }$ , + +$$ +\begin{array} { l } { { \displaystyle \kappa ( { \bf x } , { \bf x } ^ { \prime } ) = \sum _ { k = 0 } ^ { \infty } N ( d , k ) \int _ { \mathbb { S } ^ { d } } \widehat { \kappa } ( \langle { \bf y } , { \bf x } ^ { \prime } \rangle ) P _ { k } ( \langle { \bf y } , { \bf x } \rangle ) d \tau _ { d } ( { \bf y } ) } \ ~ } \\ { { \displaystyle = \sum _ { k = 0 } ^ { \infty } N ( d , k ) \frac { \omega _ { d - 1 } } { \omega _ { d } } P _ { k } ( \langle { \bf x } , { \bf x } ^ { \prime } \rangle ) \int _ { - 1 } ^ { 1 } \widehat { \kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t } , } \end{array} +$$ + +where we apply the Hecke-Funk formula and addition formula. By denoting ş $\begin{array} { r l } { \lambda _ { k } } & { { } = } \end{array}$ $( \omega _ { d - 1 } / \omega _ { d } ) \int _ { - 1 } ^ { 1 } \widehat { \kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t$ and the addition formula, we have + +$$ +\kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = \sum _ { k = 0 } ^ { \infty } \mu _ { k } N ( d , k ) P _ { k } ( \langle \mathbf { x } , \mathbf { x } ^ { \prime } \rangle ) = \sum _ { k = 0 } ^ { \infty } \mu _ { k } \sum _ { j = 1 } ^ { N ( p , k ) } Y _ { k , j } ( \mathbf { x } ) Y _ { k , j } ( \mathbf { x } ^ { \prime } ) . +$$ + +This formula (B.6) is the Mercer decomposition for the kernel function $\kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } )$ and $\mu _ { k }$ is exactly the eigenvalue of the integral operator $L _ { K }$ on $L _ { 2 } ( \mathbb { S } ^ { d } )$ defined by + +$$ +L _ { \kappa } ( f ) ( \mathbf { y } ) = \int _ { \mathbb { S } ^ { d } } \kappa ( \mathbf { x } , \mathbf { y } ) f ( \mathbf { x } ) d \tau _ { d } ( \mathbf { x } ) , \quad f \in L _ { 2 } ( \mathbb { S } ^ { d } ) . +$$ + +By using same technique as $\kappa ( \mathbf { x } , \mathbf { x } ^ { \prime } )$ , we can derive a similar expression for $\begin{array} { r l } { \sigma ( \langle { \bf w } , { \bf x } \rangle ) } & { { } = } \end{array}$ max $\{ \langle \mathbf { w } , \mathbf { x } \rangle , 0 \}$ and $\sigma ^ { \prime } ( \langle { \bf w } , { \bf x } \rangle ) = \mathbb { 1 } \{ \langle { \bf w } , { \bf x } \rangle > 0 \}$ , since they are essentially dot-product function on $L _ { 2 } ( \mathbb { S } ^ { d } )$ . We deliver the expression below without presenting proofs. + +$$ +\begin{array} { l } { { \displaystyle \sigma ^ { \prime } ( \langle { \bf w } , { \bf x } \rangle ) = \sum _ { k = 0 } ^ { \infty } \beta _ { 1 , k } N ( d , k ) P _ { k } ( \langle { \bf w } , { \bf x } \rangle ) } , } \\ { { \displaystyle \sigma ( \langle { \bf w } , { \bf x } \rangle ) = \sum _ { k = 0 } ^ { \infty } \beta _ { 2 , k } N ( d , k ) P _ { k } ( \langle { \bf w } , { \bf x } \rangle ) } , } \end{array} +$$ + +where $\beta _ { 1 , k } = \left( \omega _ { d - 1 } / \omega _ { d } \right) \int _ { - 1 } ^ { 1 } \sigma ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t$ and $\beta _ { 2 , k } = ( \omega _ { d - 1 } / \omega _ { d } ) \int _ { - 1 } ^ { 1 } \sigma ^ { \prime } ( t ) P _ { k } ( t ) ( 1 -$ $t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t$ . We add more comments on the values of $\beta _ { 1 , k }$ and $\beta _ { 2 , k }$ . It has been pointed out in Bach (2017) that when $k > \alpha$ and when $k$ and $\alpha$ have same parity, we have $\beta _ { \alpha + 1 , k } = 0$ . This is because the Legendre polynomial $P _ { k } ( t )$ is orthogonal to any other polynomials of degree less than $k$ with respect to the density function $p ( t ) = ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 }$ . Then we clearly know that $\beta _ { 1 , k } = 0$ for $k = 2 j$ and $\beta _ { 2 , k } = 0$ for $k = 2 j + 1$ with $j \in \mathbb { N } ^ { + }$ . + +For two kernel function defined in (2.2), we have + +$$ +\begin{array} { r l } & { \kappa _ { 1 } ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = \mathbb { E } _ { \mathbf { w } \sim N ( \mathbf { 0 } , \mathbf { I } ) } \left[ \sigma ^ { \prime } ( \langle \mathbf { w } , \mathbf { x } \rangle ) \sigma ^ { \prime } ( \langle \mathbf { w } , \mathbf { x } ^ { \prime } \rangle ) \right] } \\ & { \quad \quad \quad \quad = \mathbb { E } _ { \mathbf { w } \sim N ( \mathbf { 0 } , \mathbf { I } ) } \left[ \sigma ^ { \prime } ( \langle \mathbf { w } / \left. \mathbf { w } \right. , \mathbf { x } \rangle ) \sigma ^ { \prime } ( \langle \mathbf { w } / \left. \mathbf { w } \right. , \mathbf { x } ^ { \prime } \rangle ) \right] } \\ & { \quad \quad \quad = \displaystyle \int _ { \mathbb { S } ^ { d } } \sigma ^ { \prime } ( \langle \mathbf { v } , \mathbf { x } \rangle ) \sigma ^ { \prime } ( \langle \mathbf { v } , \mathbf { x } ^ { \prime } \rangle ) d \tau _ { d } ( \mathbf { v } ) . } \end{array} +$$ + +The first equality holds because $\sigma ^ { \prime }$ is 0-homogeneous function and the second equality is true since the normalized direction of a multivariate Gaussian random variable satisfies uniform distribution on the unit sphere. Similarly we can derive + +$$ +\kappa _ { 2 } ( \mathbf { x } , \mathbf { x } ^ { \prime } ) = ( d + 1 ) \int _ { \mathbb { S } ^ { d } } \sigma ( \langle \mathbf { v } , \mathbf { x } \rangle ) \sigma ( \langle \mathbf { v } , \mathbf { x } ^ { \prime } \rangle ) d \tau _ { d } ( \mathbf { v } ) . +$$ + +By combing (A.2), (B.7), (B.8), (B.9) and (B.10), we can get + +$$ +\kappa _ { 1 } ( { \bf x } , { \bf x } ^ { \prime } ) = \sum _ { k = 0 } ^ { \infty } \beta _ { 1 , k } ^ { 2 } N ( d , k ) P _ { k } ( \langle { \bf x } , { \bf x } ^ { \prime } \rangle ) , +$$ + +and + +$$ +\kappa _ { 2 } ( { \bf x } , { \bf x } ^ { \prime } ) = ( d + 1 ) \sum _ { k = 0 } ^ { \infty } \beta _ { 2 , k } ^ { 2 } N ( d , k ) P _ { k } ( \langle { \bf x } , { \bf x } ^ { \prime } \rangle ) . +$$ + +Comparing (B.6), (B.11) and (B.12), we can easily show that + +$$ +\mu _ { 1 , k } = \beta _ { 1 , k } ^ { 2 } \mathrm { a n d } \mu _ { 2 , k } = ( d + 1 ) \beta _ { 2 , k } ^ { 2 } . +$$ + +In Bach (2017), explicit expressions for $\beta _ { 1 , k }$ and $\beta _ { 2 , k }$ for $k \geqslant \alpha + 1$ are presented by + +$$ +\beta _ { \alpha + 1 , k } = \frac { d - 1 } { 2 \pi } \frac { \alpha ! ( - 1 ) ^ { ( k - 1 - \alpha ) / 2 } } { 2 ^ { k } } \frac { \Gamma ( d / 2 ) \Gamma ( k - \alpha ) } { \Gamma ( \frac { k - \alpha + 1 } { 2 } ) \Gamma ( \frac { k + d + \alpha + 1 } { 2 } ) } . +$$ + +By Stirling formula $\Gamma ( x ) \approx x ^ { x - 1 / 2 } e ^ { - x } \sqrt { 2 \pi }$ , we have following expression of $\beta _ { \alpha + 1 , k }$ for $k \geqslant \alpha + 1$ $\beta _ { \alpha + 1 , k } = C ( \alpha ) \frac { ( d - 1 ) d ^ { \frac { d - 1 } { 2 } } ( k - \alpha ) ^ { k - \alpha - \frac 1 2 } } { ( k - \alpha + 1 ) ^ { \frac { k - \alpha } { 2 } } ( k + d + \alpha + 1 ) ^ { \frac { k + d + \alpha } { 2 } } } = \Omega \left( d ^ { \frac { d + 1 } { 2 } } k ^ { \frac { k - \alpha - 1 } { 2 } } ( k + d ) ^ { \frac { - k - d - \alpha } { 2 } } \right)$ where $\begin{array} { r l r l r l r l r l } { C ( \alpha ) } & { { } = } & { { \frac { \sqrt { 2 } \alpha ! } { 2 \pi } } \exp \{ \alpha \ + \ 1 \} . } & { } & { \mathrm { A l s o } \ \beta _ { \alpha + 1 , 0 } } & { { } = } & { { \frac { d - 1 } { 4 \pi } } { \frac { \Gamma \left( { \frac { \alpha + 1 } { 2 } } \right) \Gamma \left( { \frac { d } { 2 } } \right) } { \Gamma \left( { \frac { d + \alpha + 2 } { 2 } } \right) } } , } & { \beta _ { 1 , 1 } \beta _ { 1 , 2 } \gamma _ { 1 , 0 } } & { { } = } & { { \frac { d - 1 } { 4 \pi } } { \frac { \Gamma \left( { \frac { \alpha + 1 } { 2 } } \right) \Gamma \left( { \frac { d } { 2 } } \right) } { \Gamma \left( { \frac { d + \alpha + 1 } { 2 } } \right) } } } \end{array}$ $\textstyle { \frac { d - 1 } { 2 d \pi } }$ and $\begin{array} { r l r } { \beta _ { 2 , 1 } } & { { } = } & { \frac { d - 1 } { 4 \pi d } \frac { \Gamma \left( \frac { 1 } { 2 } \right) \Gamma \left( \frac { d + 2 } { 2 } \right) } { \Gamma \left( \frac { d + 3 } { 2 } \right) } } \end{array}$ . Thus combine (B.13) we know that µα\`1,k “ $\Omega \left( d ^ { d + 1 + \alpha } k ^ { k - \alpha - 1 } ( k + d ) ^ { - k - \bar { d } - \bar { \alpha } } \right)$ . + +By considering (A.3) and (B.6), we have + +$$ +\begin{array} { c } { { \mu _ { 0 } = \mu _ { 1 , 1 } + 2 \mu _ { 2 , 0 } , } } \\ { { { } } } \\ { { \mu _ { k ^ { \prime } } = 0 , k ^ { \prime } = 2 j + 1 , j \in \mathbb { N } ^ { + } , } } \end{array} +$$ + +and + +$$ +\mu _ { k } = \frac { k } { 2 k + d - 1 } \mu _ { 1 , k - 1 } + \frac { k + d - 1 } { 2 k + d - 1 } \mu _ { 1 , k + 1 } + 2 \mu _ { 2 , k } , +$$ + +for $k \geqslant 1$ and $k \neq k ^ { \prime }$ . From the discussion above, we thus know exactly that for $k \geqslant 1$ + +$$ +\mu _ { k } = \Omega \left( \operatorname * { m a x } \left\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \right\} \right) . +$$ + +This finishes the proof. + +# B.3 PROOF OF THE COROLLARIES 3.7 AND 3.8 + +Proof of the Corollaries 3.7 and 3.8. We only need to bound $| \phi _ { j } ( \mathbf { x } ) |$ for $j \in [ r _ { k } ]$ to finish the proof. Since now we assume input data follows uniform distribution on the unit sphere $\mathbb { S } ^ { d }$ , $\phi _ { j } ( \mathbf { x } )$ would be spherical harmonics of order at most $k$ for $j \in [ r _ { k } ]$ . For any spherical harmonics $Y _ { k }$ of order $k$ and any point on $\mathbb { S } ^ { d }$ , we have an upper bound (Proposition 4.16 in Frye and Efthimiou (2012)) + +$$ +| Y _ { k } ( { \bf x } ) | \leqslant \left( N ( d , k ) \int _ { { \mathbb S } ^ { d } } Y _ { k } ^ { 2 } ( { \bf y } ) d \tau _ { d } ( { \bf y } ) \right) ^ { \frac { 1 } { 2 } } . +$$ + +kno. For $| \phi _ { j } ( \mathbf { x } ) | \leqslant \sqrt { N ( d , k ) }$ $k \gg d$ $\begin{array} { r } { N ( d , k ) \ = \ \frac { 2 k + d - 1 } { k } \binom { k + d - 2 } { d - 1 } \ = \ } \end{array}$ $\mathcal { O } ( k ^ { d - 1 } )$ $d \gg k$ $\begin{array} { r } { N ( d , k ) = \frac { 2 k + d - 1 } { k } \binom { k + d - 2 } { d - 1 } = \mathcal { O } ( d ^ { k } ) } \end{array}$ + +# C PROOF OF TECHNICAL LEMMAS + +# C.1 PROOF OF LEMMA B.2 + +Proof of Lemma B.2. The first inequality directly follows by equation (44) in $\mathtt { S u }$ and Yang (2019). To prove the second bound, we write a $\begin{array} { r } { \dot { \mathbf { V } _ { r _ { k } } } = \hat { \mathbf { V } } _ { r _ { k } } \mathbf { \bar { A } } + \hat { \mathbf { V } } _ { r _ { k } } ^ { \perp } \mathbf { \bar { B } } } \end{array}$ , where a $\mathbf { A } \in \mathbb { R } ^ { r _ { k } \times r _ { k } }$ , $\mathbf { B } \in \mathbb { R } ^ { ( n - r _ { k } ) \times r _ { k } }$ . Let $\xi _ { 1 } = C ^ { \prime } ( \lambda _ { r _ { k } } - \lambda _ { r _ { k } + 1 } ) ^ { - 1 } \cdot \sqrt { \log ( 1 / \delta ) / n }$ , $\xi _ { 2 } = C ^ { \prime \prime \prime } M ^ { 2 } \sqrt { \log ( r _ { k } / \delta ) / n }$ be the bounds given in the first inequality and Lemma 3.1. By the first inequality, we have + +$$ +\| \mathbf { B } \| _ { F } = \| \mathbf { B } ^ { \top } \| _ { F } = \| \mathbf { V } _ { r _ { k } } ^ { \top } \widehat { \mathbf { V } } _ { r _ { k } } ^ { \perp } \| _ { F } \leqslant \xi _ { 1 } . +$$ + +Moreover, since $\mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { V } _ { r _ { k } } = \mathbf { A } ^ { \top } \mathbf { A } + \mathbf { B } ^ { \top } \mathbf { B }$ , by Lemma 3.1 we have + +$$ +\| \mathbf { A } \mathbf { A } ^ { \top } - \mathbf { I } \| _ { 2 } = \| \mathbf { A } ^ { \top } \mathbf { A } - \mathbf { I } \| _ { 2 } \leqslant \| \mathbf { V } _ { r _ { k } } ^ { \top } \mathbf { V } _ { r _ { k } } - \mathbf { I } \| _ { 2 } + \| \mathbf { B } ^ { \top } \mathbf { B } \| _ { 2 } \leqslant r _ { k } \xi _ { 2 } + \xi _ { 1 } ^ { 2 } . +$$ + +Therefore + +$$ +\begin{array} { r l } & { \| \mathbf { V } _ { r _ { k } } \mathbf { V } _ { r _ { k } } ^ { \top } - \hat { \mathbf { V } } _ { r _ { k } } \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \| _ { 2 } } \\ & { \qquad = \| \hat { \mathbf { V } } _ { r _ { k } } \mathbf { A } \mathbf { A } ^ { \top } \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } + \hat { \mathbf { V } } _ { r _ { k } } \mathbf { A } \mathbf { B } ^ { \top } ( \hat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } + \hat { \mathbf { V } } _ { r _ { k } } ^ { \bot } \mathbf { B } \mathbf { A } ^ { \top } \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } + \hat { \mathbf { V } } _ { r _ { k } } ^ { \bot } \mathbf { B } \mathbf { B } ^ { \top } ( \hat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } - \hat { \mathbf { V } } _ { r _ { k } } \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \| _ { 2 } } \\ & { \qquad \leqslant \| \hat { \mathbf { V } } _ { r _ { k } } \big ( \mathbf { A } \mathbf { A } ^ { \top } - \mathbf { I } \big ) \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \| _ { 2 } + \mathcal { O } ( \| \mathbf { B } \| _ { 2 } ) } \\ & { \qquad = \| \mathbf { A } \mathbf { A } ^ { \top } - \mathbf { I } \| _ { 2 } + \mathcal { O } ( \| \mathbf { B } \| _ { 2 } ) } \\ & { \qquad \leqslant \mathcal { O } ( r _ { k } \xi _ { 2 } + \xi _ { 1 } ) } \end{array} +$$ + +Plugging in the definition of $\xi _ { 1 }$ and $\xi _ { 2 }$ completes the proof. + +# C.2 PROOF OF LEMMA B.3 + +Proof of Lemma B.3. The gradient descent update formula gives + +$$ +\mathbf { W } ^ { ( t + 1 ) } = \mathbf { W } ^ { ( t ) } + \frac { 2 \eta } { n } \sum _ { i = 1 } ^ { n } \bigl ( y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) \bigr ) \cdot \theta \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) . +$$ + +For any $j \in [ n ]$ , subtracting $\mathbf { W } ^ { ( t ) }$ and applying inner product with $\theta \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } \left( \mathbf { x } _ { j } \right)$ on both sides gives + +$$ +\theta \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { j } ) , \mathbf { W } ^ { ( t + 1 ) } - \mathbf { W } ^ { ( t ) } \rangle = \frac { 2 \eta \theta ^ { 2 } } { n } \sum _ { i = 1 } ^ { n } ( y _ { i } - \hat { y } _ { i } ^ { ( t ) } ) \cdot \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { j } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) \rangle . +$$ + +Further rearranging terms then gives + +$$ +y _ { j } - ( \widehat { \mathbf { y } } ^ { ( t + 1 ) } ) _ { j } = y _ { j } - ( \widehat { \mathbf { y } } ^ { ( t ) } ) _ { j } - \frac { 2 \eta m \theta ^ { 2 } } { n } \sum _ { i = 1 } ^ { n } ( y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) ) \cdot \mathbf { K } _ { i , j } ^ { \infty } + I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t } , +$$ + +where + +$$ +\begin{array} { l } { { \displaystyle I _ { 1 , j , t } = - \frac { 2 \eta \theta ^ { 2 } } { n } \sum _ { i = 1 } ^ { n } ( y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) ) \cdot [ \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { j } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) \rangle - m \mathbf { K } _ { i , j } ^ { ( 0 ) } ] } , } \\ { { \displaystyle I _ { 2 , j , t } = - \frac { 2 \eta \eta \theta ^ { 2 } } { n } \sum _ { i = 1 } ^ { n } ( y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) ) \cdot ( \mathbf { K } _ { i , j } ^ { ( 0 ) } - \mathbf { K } _ { i , j } ^ { \infty } ) , } } \\ { { \displaystyle I _ { 3 , j , t } = - \theta \cdot [ f _ { \mathbf { W } ^ { ( t + 1 ) } } ( \mathbf { x } _ { j } ) - f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { j } ) - \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { j } ) , \mathbf { W } ^ { ( t + 1 ) } - \mathbf { W } ^ { ( t ) } \rangle ] } . } \end{array} +$$ + +For $I _ { 1 , j , t }$ , by Lemma D.6, we have + +$$ +| I _ { 1 , j , t } | \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \frac { 1 } { n } \sum _ { i = 1 } ^ { n } | y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) | \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ) } \| _ { 2 } / \sqrt { n } +$$ + +with probability at least $1 - \mathcal { O } ( n ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ . For $I _ { 2 , j , t }$ , by Bernstein inequality and union bound, with probability at least $1 - \mathcal { O } ( n ^ { 2 } ) \cdot \exp ( - \Omega ( m \omega ^ { 2 / 3 } ) )$ , we have + +$$ +\big | \mathbf { K } _ { i , j } ^ { \infty } - \mathbf { K } _ { i , j } ^ { ( 0 ) } \big | \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } ) +$$ + +for all $i , j \in [ n ]$ . Therefore + +$$ +| I _ { 2 , j , t } | \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \frac { 1 } { n } \sum _ { i = 1 } ^ { n } | y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) | \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ) } \| _ { 2 } / \sqrt { n } . +$$ + +For $I _ { 3 , j , t }$ , we have + +$$ +\begin{array} { r l r } { { I _ { 3 , j , t } \leqslant \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \sqrt { m } \theta ) \cdot \| \mathbf { W } _ { 1 } ^ { ( t + 1 ) } - \mathbf { W } _ { 1 } ^ { ( t ) } \| _ { 2 } } } \\ & { } & { \leqslant \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \sqrt { m } \theta ) \cdot \frac { 2 \eta } { n } \sum _ { i = 1 } ^ { n } | y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) | \cdot \theta \cdot \| \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) \| _ { 2 } } \\ & { } & { \leqslant \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \cfrac { 1 } { n } \sum _ { i = 1 } ^ { n } | y _ { i } - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } _ { i } ) | } \\ & { } & { \leqslant \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \mathbf { y } - \widehat { \mathbf { y } } ^ { ( t ) } \| _ { 2 } / \sqrt { n } , } \end{array} +$$ + +where the first inequality follows by Lemmas D.2, the second inequality is obtained from (C.1), and the third inequality follows by Lemma D.4. Setting the $j$ -th entry of $\mathbf { e } ^ { ( t ) }$ as $I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t }$ and writing (C.2) into matrix form completes the proof. □ + +# C.3 PROOF OF LEMMA B.4 + +Proof of Lemma B.4. Denote $\mathbf { u } ^ { ( t ) } = \mathbf { y } - \hat { \mathbf { y } } ^ { ( t ) }$ , $t \in T$ . Then we have + +$$ +\begin{array} { r l } & { \| ( \widehat { { \mathbf { V } } } _ { r _ { k } } ^ { \perp } ) ^ { \top } { \mathbf { u } } ^ { ( t ^ { \prime } + 1 ) } \| _ { 2 } \leqslant \| ( \widehat { { \mathbf { V } } } _ { r _ { k } } ^ { \perp } ) ^ { \top } [ { \mathbf { I } } - ( \eta m \theta ^ { 2 } / n ) { \mathbf { K } } ^ { \infty } ] { \mathbf { u } } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| { \mathbf { u } } ^ { ( t ^ { \prime } ) } \| _ { 2 } } \\ & { \qquad \leqslant \| ( \widehat { { \mathbf { V } } } _ { r _ { k } } ^ { \perp } ) ^ { \top } { \mathbf { u } } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \sqrt { n } \cdot \widetilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) , } \end{array} +$$ + +where the first inequality follows by Lemma B.3, and the second inequality follows by Lemma D.5. Therefore we have + +$$ +\| ( \widehat { \mathbf { V } } _ { r _ { k } } ^ { \perp } ) ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } \leqslant \| ( \widehat { \mathbf { V } } _ { r _ { k } } ^ { \perp } ) ^ { \top } \mathbf { u } ^ { ( 0 ) } \| _ { 2 } + t ^ { \prime } \cdot \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } \cdot \sqrt { n } \cdot \widetilde { O } ( 1 + \omega \sqrt { m } ) , +$$ + +for $t ^ { \prime } = 0 , \ldots , t$ . This completes the proof of (B.1). Similarly, we have + +$$ +\begin{array} { r l } { \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } + 1 ) } \| _ { 2 } \leqslant \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } [ \mathbf { I } - ( \eta m \theta ^ { 2 } / n ) \mathbf { K } ^ { \infty } ] \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } } & { } \\ { \leqslant ( 1 - \eta m \theta ^ { 2 } \widehat { \lambda } _ { r _ { k } } ) \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot ( \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \| ( \widehat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } ) } & { } \\ { \leqslant ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| ( \widehat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } } & { } \\ { \leqslant ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) \| \widehat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + t ^ { \prime } \cdot ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) ^ { 2 } \cdot \sqrt { n } \cdot \widetilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) } & { } \\ + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| ( \widehat { \mathbf { V } } _ { r _ { k } } ^ { \bot } \end{array} +$$ + +for $t ^ { \prime } = 0 , \ldots , t - 1$ , where the third inequality is by Lemma B.1 and the assumption that $\omega \ \leqslant$ $\widetilde { \mathcal { O } } ( \lambda _ { r _ { k } } ^ { 3 } )$ , $n \geqslant \widetilde { \mathcal { O } } ( \lambda _ { r _ { k } } ^ { - 2 } )$ , and the fourth inequality is by (B.1). Therefore we have + +$$ +\begin{array} { r l } & { \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } \leqslant ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { t } \| \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( 0 ) } \| _ { 2 } + t ^ { \prime } \cdot ( \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { - 1 } \cdot ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) ^ { 2 } \cdot \sqrt { n } \cdot \tilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) } \\ & { \phantom { \hat { \mathbf { V } } } + ( \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { - 1 } \cdot \tilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| ( \hat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( 0 ) } \| _ { 2 } } \\ & { \phantom { \hat { \mathbf { V } } } = ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { t } \| \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( 0 ) } \| _ { 2 } + t ^ { \prime } \lambda _ { r _ { k } } ^ { - 1 } \cdot \omega ^ { 2 / 3 } \eta m \theta ^ { 2 } \cdot \sqrt { n } \cdot \tilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) } \\ & { \phantom { \hat { \mathbf { V } } } + \lambda _ { r _ { k } } ^ { - 1 } \cdot \tilde { \mathcal { O } } ( \omega ^ { 1 / 3 } ) \cdot \| ( \hat { \mathbf { V } } _ { r _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( 0 ) } \| _ { 2 } } \\ & { \leqslant ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { t } \| \hat { \mathbf { V } } _ { r _ { k } } ^ { \top } \mathbf { u } ^ { ( 0 ) } \| _ { 2 } + t ^ { \prime } \lambda _ { r _ { k } } ^ { - 1 } \cdot \omega ^ { 2 / 3 } \eta m \theta ^ { 2 } \cdot \sqrt { n } \cdot \tilde { \mathcal { O } } ( 1 + \omega \sqrt { m } ) } \\ & \phantom { \hat { \mathbf { V } } } + \end{array} +$$ + +This completes the proof of (B.2). Finally, for (B.3), by assumption we have $\omega ^ { 1 / 3 } \eta m \theta ^ { 2 } \leqslant \widetilde { \mathcal { O } } ( 1 )$ . Therefore + +$$ +\begin{array} { r l } { \mathbf { u } ^ { ( t ^ { \prime } + 1 ) } \| _ { 2 } \leqslant \| [ \mathbf { I } - ( \eta m \theta ^ { 2 } / n ) \mathbf { K } ^ { \infty } ] \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \| [ \mathbf { I } - ( \eta m \theta ^ { 2 } / n ) \mathbf { K } ^ { \infty } ] \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \bot } ( \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } } & { } \\ { \quad \quad \quad \quad \quad \quad \quad \quad + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| ( \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } } & { } \\ { \leqslant ( 1 - \eta m \theta ^ { 2 } \hat { \lambda } _ { \mathbf { r } _ { k } } ) \| \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( \omega ^ { 1 / 3 } \eta m \theta ^ { 2 } ) \cdot \| \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } + \widetilde { \mathcal { O } } ( 1 ) \cdot \| ( \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \bot } ) ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } } & { } \\ \leqslant ( 1 - \eta m \theta ^ { 2 } \lambda _ { \mathbf { r } _ { k } } / 2 ) \| \hat { \mathbf { V } } _ { \mathbf { r } _ { k } } ^ { \top } \mathbf { u } ^ { ( t ^ { \prime } ) } \| _ { 2 } \end{array} +$$ + +for $t ^ { \prime } = 0 , \ldots , t - 1$ , where the third inequality is by Lemma B.1 and the assumption that $\omega \ \leqslant$ $\widetilde { \mathcal { O } } ( \lambda _ { r _ { k } } ^ { 3 } )$ , and the fourth inequality follows by (B.1). Therefore we have + +} u p t 1 q } 2 ď $\mathscr { O } ( \sqrt { n } ) \cdot ( 1 - \eta m \theta ^ { 2 } \lambda _ { r _ { k } } / 2 ) ^ { t ^ { \prime } } + \widetilde { \mathscr { O } } ( ( \eta m \theta ^ { 2 } \lambda _ { r _ { k } } ) ^ { - 1 } ) \cdot \Vert ( \widehat { { \mathbf { V } } } _ { r _ { k } } ^ { \perp } ) ^ { \top } { \mathbf { u } } ^ { ( 0 ) } \Vert _ { 2 } + \lambda _ { r _ { k } } ^ { - 1 } t ^ { \prime } \omega ^ { 1 / 3 } \sqrt { n } \cdot \widetilde { \mathscr { O } } ( 1 + \omega \sqrt { m } ) .$ This finishes the proof. + +# D AUXILIARY LEMMAS + +In this section we list several auxiliary lemmas on the properties of over-parameterized neural networks we need in our proof of Theorem 3.2. These results are mostly summarized from Allen-Zhu et al. (2018) and Cao and Gu (2019a). + +# D.1 AUXILIARY LEMMAS + +Denote + +$$ +\begin{array} { r l } & { \mathbf { D } _ { i } = \operatorname { d i a g } \bigl ( \mathbb { 1 } \{ ( \mathbf { W } _ { 1 } \mathbf { x } _ { i } ) _ { 1 } > 0 \} , \ldots , \mathbb { 1 } \{ ( \mathbf { W } _ { 1 } \mathbf { x } _ { i } ) _ { m } > 0 \} \bigr ) , } \\ & { \mathbf { D } _ { i } ^ { ( 0 ) } = \operatorname { d i a g } \bigl ( \mathbb { 1 } \{ ( \mathbf { W } _ { 1 } ^ { ( 0 ) } \mathbf { x } _ { i } ) _ { 1 } > 0 \} , \ldots , \mathbb { 1 } \{ ( \mathbf { W } _ { 1 } ^ { ( 0 ) } \mathbf { x } _ { i } ) _ { m } > 0 \} \bigr ) . } \end{array} +$$ + +Lemma D.1 (Allen-Zhu et al. (2018)). If $\omega \leqslant \mathcal { O } ( [ \log ( m ) ] ^ { - 3 / 2 } )$ , then with probability at least $1 - \mathcal { O } ( n ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ , + +$$ +\| \mathbf { D } _ { i } - \mathbf { D } _ { i } ^ { ( 0 ) } \| _ { 0 } \leqslant \mathcal { O } ( \omega ^ { 2 / 3 } m ) +$$ + +for all $\mathbf { W } \in { \cal B } ( \mathbf { W } ^ { ( 0 ) } , \omega )$ , $i \in [ n ]$ + +Lemma D.2 (Cao and Gu (2019a)). There exists an absolute constant $\kappa$ such that, with probability at least $1 - \mathcal { O } ( n ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ over the randomness of $\mathbf { W } ^ { ( 1 ) }$ , for all $i \in [ n ]$ and $\mathbf { W } , \mathbf { W } ^ { \prime } \in$ $B ( \mathbf { W } ^ { ( \mathbf { \theta } ) } , \omega )$ with $\omega \leqslant \kappa [ \log ( m ) ] ^ { - 3 / 2 }$ , it holds uniformly that + +$$ +| f _ { \mathbf { W } ^ { \prime } } ( \mathbf { x } _ { i } ) - f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) , \mathbf { W } ^ { \prime } - \mathbf { W } \rangle | \leqslant { \mathcal { O } } \Bigl ( \omega ^ { 1 / 3 } \sqrt { m \log ( m ) } \Bigr ) \cdot \| \mathbf { W } _ { 1 } ^ { \prime } - \mathbf { W } _ { 1 } \| _ { 2 } . +$$ + +Lemma D.3 (Cao and Gu (2019a)). For any $\delta > 0$ , if $m \geqslant C \log ( n / \delta )$ for a large enough absolute constant $C$ , then with probability at least $1 - \delta$ , $| f _ { \mathbf { W } ^ { ( 0 ) } } ( { \pmb x } _ { i } ) | \leqslant \mathcal { O } ( \sqrt { \log ( n / \delta ) } )$ for all $i \in [ n ]$ . + +Lemma D.4 (Cao and Gu (2019a)). There exists an absolute constant $C$ such that, with probability at least $1 - \mathcal { O } ( n ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ , for all $i \in [ n ] , l \in [ L ]$ and $\mathbf { W } \in { \boldsymbol { B } } ( \mathbf { W } ^ { ( 0 ) } , \omega )$ with $\omega \ \leqslant$ $C [ \log ( m ) ] ^ { - 3 }$ , it holds uniformly that + +$$ +\| \nabla _ { \mathbf { W } _ { l } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) \| _ { F } \leqslant \mathcal { O } ( \sqrt { m } ) . +$$ + +The following lemma provides a uniform bound of the neural network function value over $B ( \mathbf { W } ^ { ( 0 ) } , \omega )$ . + +Lemma D.5. Suppose that $m \geqslant \Omega ( \omega ^ { - 2 / 3 } \log ( n / \delta ) )$ and $\omega \leqslant \mathcal { O } ( [ \log ( m ) ] ^ { - 3 } )$ . Then with probability at least $1 - \delta$ , $| f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) | \leqslant \mathcal { O } ( \sqrt { \log ( n / \delta ) } + \omega \sqrt { m } )$ for all $\mathbf { W } \in { \boldsymbol { B } } ( \mathbf { W } ^ { ( 0 ) } , \omega ) \ i \in [ n ]$ . + +Lemma D.6. If $\omega \leqslant \mathcal { O } ( [ \log ( m ) ] ^ { - 3 / 2 } )$ , then with probability at least $1 - \mathcal { O } ( n ) \cdot \exp [ - \Omega ( m \omega ^ { 2 / 3 } ) ]$ + +$$ +\begin{array} { r l } & { \| \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \| _ { F } \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \sqrt { m } ) , } \\ & { | \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { j } ) \rangle - \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { j } ) \rangle | \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } m ) } \end{array} +$$ + +for all $\mathbf { W } \in { \cal B } ( \mathbf { W } ^ { ( 0 ) } , \omega )$ and $i \in [ n ]$ + +D.2 PROOFS OF LEMMAS D.5 AND D.6 + +Proof of Lemma D.5. By Lemmas D.2 and D.4, we have + +$$ +\begin{array} { r l } & { f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) | \leqslant \| \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \| _ { F } \| \mathbf { W } _ { 1 } - \mathbf { W } _ { 1 } ^ { ( 0 ) } \| _ { F } + \| \nabla _ { \mathbf { W } _ { 2 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \| _ { F } \| \mathbf { W } _ { 2 } - \mathbf { W } _ { 2 } ^ { ( 0 ) } \| _ { F } } \\ & { \qquad + \mathcal { O } ( \omega ^ { 1 / 3 } \sqrt { m \log ( m ) } ) \cdot \| \mathbf { W } _ { 1 } - \mathbf { W } _ { 1 } ^ { ( 0 ) } \| _ { 2 } } \\ & { \qquad \leqslant \mathcal { O } ( \omega \sqrt { m } ) , } \end{array} +$$ + +where the last inequality is by the assumption $\omega \leqslant [ \log ( m ) ] ^ { - 3 }$ . Applying triangle inequality and Lemma D.3 then gives + +$$ +\begin{array} { r l } & { | f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) | \leqslant | f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) | + | f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) | \leqslant \mathcal { O } ( \sqrt { \log ( n / \delta ) } ) + \mathcal { O } ( \omega \sqrt { m } ) } \\ & { \qquad = \mathcal { O } ( \sqrt { \log ( n / \delta ) } + \omega \sqrt { m } ) , } \end{array} +$$ + +This completes the proof. + +Proof of Lemma D.6. By direct calculation, we have + +$$ +\nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) = \sqrt { m } \cdot \mathbf { D } _ { i } ^ { ( 0 ) } \mathbf { W } _ { 2 } ^ { ( 0 ) \top } \mathbf { x } _ { i } ^ { \top } , \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) = \sqrt { m } \cdot \mathbf { D } _ { i } \mathbf { W } _ { 2 } ^ { \top } \mathbf { x } _ { i } ^ { \top } . +$$ + +Therefore we have + +$$ +\begin{array} { r l } { | \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) | | _ { F } = \sqrt { m } \cdot \| \mathbf { D } _ { i } \mathbf { W } _ { 2 } ^ { \top } \mathbf { x } _ { i } ^ { \top } - \mathbf { D } _ { i } ^ { ( 0 ) } \mathbf { W } _ { 2 } ^ { ( 0 ) \top } \mathbf { x } _ { i } ^ { \top } \| _ { F } } & { } \\ { = \sqrt { m } \cdot \| \mathbf { x } _ { i } \mathbf { W } _ { 2 } \mathbf { D } _ { i } - \mathbf { x } _ { i } \mathbf { W } _ { 2 } ^ { ( 0 ) } \mathbf { D } _ { i } ^ { ( 0 ) } \| _ { F } } & { } \\ { = \sqrt { m } \cdot \| \mathbf { W } _ { 2 } \mathbf { D } _ { i } - \mathbf { W } _ { 2 } ^ { ( 0 ) } \mathbf { D } _ { i } ^ { ( 0 ) } \| _ { F } } & { } \\ { \leqslant \sqrt { m } \cdot \| \mathbf { W } _ { 2 } ^ { ( 0 ) } ( \mathbf { D } _ { i } ^ { ( 0 ) } - \mathbf { D } _ { i } ) \| _ { F } + \sqrt { m } \cdot \| ( \mathbf { W } _ { 2 } ^ { ( 0 ) } - \mathbf { W } _ { 2 } ) \mathbf { D } _ { i } \| _ { F } } & { } \end{array} +$$ + +By Lemma 7.4 in Allen-Zhu et al. (2018) and Lemma D.1, with probability at least $1 - n \cdot$ $\exp [ - \Omega ( m ) ]$ , $\sqrt { m } \cdot \| \mathbf { W } _ { 2 } ^ { ( 0 ) } ( \mathbf { D } _ { i } ^ { ( 0 ) } - \mathbf { D } _ { i } ) \| _ { F } \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \sqrt { m } )$ for all $i \in [ n ]$ . Moreover, clearly $\| ( \mathbf { W } _ { 2 } ^ { ( 0 ) } - \mathbf { W } _ { 2 } ) \mathbf { D } _ { i } \| _ { F } \leqslant \| \mathbf { W } _ { 2 } ^ { ( 0 ) } - \mathbf { W } _ { 2 } \| _ { F } \leqslant \omega$ . Therefore + +$$ +\| \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - \nabla _ { \mathbf { W } _ { 1 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \| _ { F } \leqslant \mathcal { O } ( \omega ^ { 1 / 3 } \sqrt { m } ) +$$ + +for all $i \in [ n ]$ . This proves the bound for the first layer gradients. For the second layer gradients, we have + +$$ +\nabla _ { \mathbf { W } _ { 2 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) = \sqrt { m } \cdot [ \sigma ( \mathbf { W } _ { 1 } ^ { ( 0 ) } \mathbf { x } _ { i } ) ] ^ { \top } , \nabla _ { \mathbf { W } _ { 2 } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) = \sqrt { m } \cdot [ \sigma ( \mathbf { W } _ { 1 } \mathbf { x } _ { i } ) ] ^ { \top } +$$ + +It therefore follows by the 1-Lipschitz continuity of $\sigma ( \cdot )$ that + +$$ +\| \nabla _ { \mathbf { W } _ { 2 } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - \nabla _ { \mathbf { W } _ { 2 } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) \| _ { F } \leqslant { \sqrt { m } } \cdot \| \mathbf { W } _ { 1 } \mathbf { x } _ { i } - \mathbf { W } _ { 1 } ^ { ( 0 ) } \mathbf { x } _ { i } \| _ { F } \leqslant \omega { \sqrt { m } } \leqslant \omega ^ { 1 / 3 } { \sqrt { m } } . +$$ + +This completes the proof of the first inequality. + +The second inequality directly follows by triangle inequality and Lemma D.4: + +$$ +\begin{array} { r l } & { | \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { j } ) \rangle - m \mathbf { K } ^ { ( 0 ) } | \leqslant | \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { i } ) - \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { j } ) \rangle | } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad + | \langle \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { i } ) , \nabla _ { \mathbf { W } } f _ { \mathbf { W } } ( \mathbf { x } _ { j } ) - \nabla _ { \mathbf { W } } f _ { \mathbf { W } ^ { ( 0 ) } } ( \mathbf { x } _ { j } ) \rangle | } \\ & { \quad \quad \quad \quad \quad \quad \quad \leqslant { \mathcal { O } } ( \omega ^ { 1 / 3 } m ) . } \end{array} +$$ + +This finishes the proof. + +# E ADDITIONAL EXPERIMENTAL RESULTS + +# E.1 ESTIMATING THE PROJECTION LENGTH IN FUNCTION SPACE + +As mentioned in Section 4.1, when using freshly sampled points, we are actually estimating the projection length of residual function $r ( \mathbf { x } ) = f ^ { * } ( \mathbf { x } ) - \theta f _ { \mathbf { W } ^ { ( t ) } } ( \mathbf { x } )$ onto the given Gegenbauer polynomial $P _ { k } ( { \bf x } )$ . Here we present in Figure 3 a comparison between the projection length using training data and that using test data. An interesting phenomenon is that the network generalizes well on the lower-order Gegenbauer polynomial and along the highest-order Gegenbauer polynomial the network suffers overfitting. + +![](images/ede46709d91cb04a929d72bd1ecb66385cb7ce3faf11dd60ee6d67e97d71b25f.jpg) +Figure 3: Convergence curve for projection length onto vectors (determined by training data) and functions (estimated by test data). We can see that for low-order Gegenbauer polynomials, the network learns the function while for the high-order Gegenbauer polynomial, the network overfits the training data. + +# E.2 NEAR-LINEAR CONVERGENCE BEHAVIOR + +In this subsection, we present the same curve shown in Section 4.1 in logarithmic scale instead of linear scale. As shown in Figure 4 we can see that the convergence of different projection length is close to linear convergence, which is well-aligned with our theorem. Note that we performed a moving average of range 20 on these curves to avoid the heavy oscillation especially in late stage. + +![](images/d4b04e61fd87b82bba68d9aeb4f350e54f3f281cc4e3ecdbada8bf74711541a3.jpg) +Figure 4: Log-scale convergence curve for projection length onto different component. (a) shows the curve when the target function have different component with the same scale. (b) shows the curve when the higner-order components have larger scale. Both exhibit nearly linear convergence especially at late stage. \ No newline at end of file diff --git a/parse/train/BJg15lrKvS/BJg15lrKvS_content_list.json b/parse/train/BJg15lrKvS/BJg15lrKvS_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..969a65328522317e4bc6b232355d8268d3d63018 --- /dev/null +++ b/parse/train/BJg15lrKvS/BJg15lrKvS_content_list.json @@ -0,0 +1,3206 @@ +[ + { + "type": "text", + "text": "TOWARDS UNDERSTANDING THE SPECTRAL BIAS OF DEEP LEARNING ", + "text_level": 1, + "bbox": [ + 176, + 98, + 823, + 145 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 236, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An intriguing phenomenon observed during training neural networks is the spectral bias, where neural networks are biased towards learning less complex functions. The priority of learning functions with low complexity might be at the core of explaining generalization ability of neural network, and certain efforts have been made to provide theoretical explanation for spectral bias. However, there is still no satisfying theoretical result justifying the underlying mechanism of spectral bias. In this paper, we give a comprehensive and rigorous explanation for spectral bias and relate it with the neural tangent kernel function proposed in recent work. We prove that the training process of neural networks can be decomposed along different directions defined by the eigenfunctions of the neural tangent kernel, where each direction has its own convergence rate and the rate is determined by the corresponding eigenvalue. We then provide a case study when the input data is uniformly distributed over the unit sphere, and show that lower degree spherical harmonics are easier to be learned by over-parameterized neural networks. ", + "bbox": [ + 233, + 263, + 764, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 491, + 336, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over-parameterized neural networks have achieved great success in many applications such as computer vision (He et al., 2016), natural language processing (Collobert and Weston, 2008) and speech recognition (Hinton et al., 2012). It has been shown that over-parameterized neural networks can fit complicated target function or even randomly labeled data (Zhang et al., 2016) and still exhibit good generalization performance when trained with real labels. Intuitively, this is at odds with the traditional notion of generalization ability such as model complexity. In order to understand neural network training, a line of work (Soudry et al., 2017; Gunasekar et al., 2018b;a) has made efforts in the perspective of “implicit bias”, which states that training algorithms for deep learning implicitly pose an inductive bias onto the training process and lead to a solution with low complexity measured by certain norms in the parameter space of the neural network. ", + "bbox": [ + 174, + 520, + 825, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among many attempts to establish implicit bias, Rahaman et al. (2018) pointed out an intriguing phenomenon called spectral bias, which says that during training, neural networks tend to learn the components of lower complexity faster. The concept of spectral bias is appealing because this may intuitively explain why over-parameterized neural networks can achieve a good generalization performance without overfitting. During training, the networks fit the low complexity components first and thus lie in the concept class of low complexity. Arguments like this may lead to rigorous guarantee for generalization. ", + "bbox": [ + 174, + 666, + 825, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Great efforts have been made in seek of explanations about the spectral bias. Rahaman et al. (2018) evaluated the Fourier spectrum of ReLU networks and empirically showed that the lower frequencies are learned first; also lower frequencies are more robust to random perturbation. Andoni et al. (2014) showed that for a sufficiently wide two-layer network, gradient descent with respect to the second layer can learn any low degree bounded polynomial. Xu (2018) provided Fourier analysis to twolayer networks and showed similar empirical results on one-dimensional functions and real data. Nakkiran et al. (2019) used information theoretical approach to show that networks obtained by stochastic gradient descent can be explained by a linear classifier during early training. All these studies provide certain explanations about why neural networks exhibit spectral bias in real tasks. But explanations in the theoretical aspect, if any, are to some extent restricted. For example, the popular Fourier analysis is usually done in the one-dimensional setting, and thus lacks generality. ", + "bbox": [ + 174, + 771, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, a recent line of work (Jacot et al., 2018; Du et al., 2018b; Li and Liang, 2018; Chizat and Bach, 2018) has shed light on new approaches to analyze neural networks. In particular, they show that under certain over-parameterized condition, the neural network trained by gradient descent behaves similar to the kernel regression predictor using the neural tangent kernel (NTK) (Jacot et al., 2018). Du et al. (2018b) showed that the convergence is provably guaranteed under certain over-parameterization conditions determined by the smallest eigenvalue of NTK. Arora et al. (2019b) further gave a finer characterization of error convergence based on the eigenvalues of NTK’s Gram matrix. Su and Yang (2019) improved the convergence guarantee in terms of the $k$ -th largest eigenvalue for certain target functions. ", + "bbox": [ + 173, + 103, + 825, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by these works mentioned above, we can present a theoretical explanation for spectral bias. Under NTK regime, we establish a precise characterization for the training process of neural networks. More specifically, we theoretically prove that over-parameterized neural networks’ training process can be controlled by the eigenvalues of the integrating operator defined by the NTK. Under the specific case of uniform distribution on unit sphere, we give an exact calculation for these eigenvalues and show that the lower frequencies have larger eigenvalues, which thus leads to faster convergence. We also conduct experiments to corroborate the theory we establish. ", + "bbox": [ + 174, + 236, + 825, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are highlighted as follows: ", + "bbox": [ + 176, + 340, + 467, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We prove a generic theorem for arbitrary data distributions, which states that under certain sample complexity and over-parameterization conditions, the error term’s convergence along different directions actually relies on the corresponding eigenvalues. This theorem gives a more precise control on the regression residual than $\\mathtt { S u }$ and Yang (2019), where the authors focused on the case when the labeling function is close to the subspace spanned by the first few eigenfunctions. \n2. We present a more general result about the spectra of the neural tangent kernel. In particular, we show that the order of eigenvalues appears as $\\mu _ { k } = \\Omega ( \\operatorname* { m a x } \\{ k ^ { - d - 1 } , \\check { d } ^ { - k + 1 } \\} )$ . Our result is better than the bound $\\Omega ( k ^ { - d - \\overline { { 1 } } } )$ derived in Bietti and Mairal (2019) when $d \\gg k$ , which is clearly a more practical setting. \n3. We establish a rigorous explanation for the spectral bias based on the aforementioned theoretical results without any specific assumptions on the target function. We show that the error terms from different frequencies are provably controlled by the eigenvalues of the NTK, and the lowerfrequency components can be learned with less training examples and narrower networks with a faster convergence rate. As far as we know, this is the first attempt to give a comprehensive theory justifying the existence of spectral bias. ", + "bbox": [ + 173, + 367, + 826, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1 ADDITIONAL RELATED WORK ", + "text_level": 1, + "bbox": [ + 178, + 598, + 424, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, there is a rich literature about the property of neural tangent kernel. Jacot et al. (2018) first showed that during training, the network function follows a descent along the kernel gradient with respect to the Neural Tangent Kernel (NTK) under infinity width setting. Li and Liang (2018) and Du et al. (2018b) implicitly built connection between Neural Tangent Kernel and gradient descent by showing that GD can provably optimize sufficiently wide two-layer neural networks. In Du et al. (2018b), it is shown that gradient descent can achieve zero training loss at a linear convergence rate for training two-layer ReLU network with square loss. Allen-Zhu et al. (2018); Du et al. (2018a); Zou et al. (2018); Cao and Gu (2019b); Arora et al. (2019a); Zou and Gu (2019); Cao and Gu (2019a) further studied the optimization and generalization of deep neural networks. These papers are all in the so-called neural tangent kernel regime, and their requirements on the network width depend either implicitly or explicitly on the smallest eigenvalue of the kernel Gram matrix. Later, Su and Yang (2019) showed that this smallest eigenvalue actually scales in the number of samples $n$ and will eventually converge to 0. In order to obtain constant convergence rate, Su and Yang (2019) assumed that the target function ş $f ^ { * }$ can be approximated by the first few eigenfunctions of the integrating operator $L _ { \\kappa } \\bar { f } ( s ) : = \\int _ { \\mathbb { S } ^ { d } } \\kappa \\bar { ( } x , s ) f ( s ) d \\bar { \\tau } \\bar { ( } s )$ where $\\kappa ( \\cdot , \\cdot )$ is the NTK function and $\\tau ( s )$ is the input distribution, and proved linear convergence rate up to the this approximation error. ", + "bbox": [ + 173, + 625, + 825, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A few theoretical results have been established towards understanding the spectra of neural tangent kernels. Bach (2017) studied two-layer ReLU networks by relating it to kernel methods, and proposed a harmonic decomposition for the functions in the reproducing kernel Hilbert space which we utilize in our proof. Based on the technique in Bach (2017), Bietti and Mairal (2019) studied the eigenvalue decay of integrating operator $L _ { \\kappa } f ( x )$ defined by NTK on unit sphere by using spherical harmonics. Vempala and Wilmes (2018) calculated the eigenvalues of NTK corresponding to two-layer neural networks with sigmoid activation function. Basri et al. (2019) established similar results as Bietti and Mairal (2019), but considered the case of training the first layer parameters of a two-layer networks with bias terms. Yang and Salman (2019) studied the the eigenvalues of integral operator with respect to the NTK on Boolean cube by Fourier analysis. ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The rest of the paper is organized as follows. We state the notation, problem setup and other preliminaries in Section 2 and present our main results in Section 3. In Section 4, we present experimental results to support our theory. Proofs of our main results can be found in the appendix. ", + "bbox": [ + 174, + 180, + 825, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 174, + 237, + 339, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we introduce the basic problem setup including the neural network structure and the training algorithm, as well as some background on the neural tangent kernel proposed recently in Jacot et al. (2018) and the corresponding integral operator. ", + "bbox": [ + 174, + 266, + 825, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 NOTATION ", + "text_level": 1, + "bbox": [ + 174, + 320, + 287, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use lower case, lower case bold face, and upper case bold face letters to denote scalars, vectors and matrices respectively. For a vector ř $\\mathbf { v } = ( \\bar { v _ { 1 } } , \\ldots , v _ { d } ) ^ { T } \\in \\mathbb { R } ^ { d }$ and a number $1 \\leqslant p < \\infty$ , we denote its p´norm by $\\begin{array} { r } { \\| \\mathbf { v } \\| _ { p } = ( \\sum _ { i = 1 } ^ { d } | v _ { i } | ^ { p } ) ^ { 1 / p } } \\end{array}$ . We also define infinity norm by $\\| \\mathbf { v } \\| _ { \\infty } = \\operatorname* { m a x } _ { i } | v _ { i } |$ . For a matrix ř $\\mathbf { A } = ( A _ { i , j } ) _ { m \\times n }$ , we use $\\lVert \\mathbf { A } \\rVert _ { 0 }$ to denote the number of non-zero entries of $\\mathbf { A }$ , and use $\\textstyle \\| \\mathbf { A } \\| _ { F } = ( \\sum _ { i , j = 1 } ^ { d } A _ { i , j } ^ { 2 } ) ^ { 1 / 2 }$ to denote its Frobenius norm. Let $\\begin{array} { r } { \\| \\mathbf { A } \\| _ { p } = \\operatorname* { m a x } _ { \\| \\mathbf { v } \\| _ { p } \\leqslant 1 } \\| \\mathbf { A } \\mathbf { v } \\| _ { p } } \\end{array}$ for $p \\geqslant 1$ , and $\\| \\mathbf { A } \\| _ { \\operatorname* { m a x } } = \\operatorname* { m a x } _ { i , j } | A _ { i , j } |$ . For two matrices $\\mathbf { A } , \\mathbf { B } \\in \\mathbb { R } ^ { m \\times n }$ , we define $\\langle \\mathbf { A } , \\mathbf { B } \\rangle = \\operatorname { T r } ( \\mathbf { A } ^ { \\top } \\mathbf { B } )$ . We use $\\mathbf { A } \\succeq \\mathbf { B }$ if $\\mathbf { A } - \\mathbf { B }$ is positive semi-definite. In addition, we define the asymptotic notations $\\mathcal { O } ( \\cdot )$ , $\\widetilde { \\mathcal { O } } ( \\cdot ) , \\Omega ( \\cdot )$ and $\\widetilde { \\Omega } ( \\cdot )$ as follows. Suppose that $a _ { n }$ and $b _ { n }$ be two sequences. We write $a _ { n } = \\mathcal { O } ( b _ { n } )$ if lim $\\begin{array} { r } { \\operatorname* { s u p } _ { n \\infty } | a _ { n } / b _ { n } | < \\infty } \\end{array}$ , and $a _ { n } = \\Omega ( b _ { n } )$ if $\\begin{array} { r } { \\operatorname* { l i m } \\operatorname* { i n f } _ { n \\to \\infty } \\left| a _ { n } / b _ { n } \\right| > 0 } \\end{array}$ . We use $\\tilde { \\mathcal { O } } ( \\cdot )$ and $\\widetilde { \\Omega } ( \\cdot )$ to hide the logarithmic factors in $\\mathcal { O } ( \\cdot )$ and $\\Omega ( \\cdot )$ . ", + "bbox": [ + 173, + 347, + 826, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 PROBLEM SETUP ", + "text_level": 1, + "bbox": [ + 174, + 513, + 334, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we introduce the basic problem setup. We consider two-layer fully connected neural networks of the form ", + "bbox": [ + 173, + 541, + 823, + 570 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b7469340c88d73afd76013b416e3d358b8ac7151f7010c91046a7f361bb7b78d.jpg", + "text": "$$\nf _ { \\mathbf { W } } ( \\mathbf { x } ) = \\sqrt { m } \\cdot \\mathbf { W } _ { 2 } \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 397, + 574, + 598, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { m \\times ( d + 1 ) }$ , $\\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { 1 \\times m }$ are the first and second layer weight matrices respectively, and $\\sigma ( \\cdot ) = \\operatorname* { m a x } \\{ 0 , \\cdot \\}$ is the entry-wise ReLU activation function. The network is trained according to the square loss on $n$ training examples $S = \\left\\{ \\left( \\mathbf { x } _ { i } , y _ { i } \\right) : i \\in \\left[ n \\right] \\right\\}$ : ", + "bbox": [ + 173, + 598, + 825, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/c624949c8ad186b693f5a14747d129ee97378522590c276fe9947110a2a18031.jpg", + "text": "$$\nL _ { S } ( \\mathbf { W } ) = \\frac { 1 } { n } \\sum _ { ( \\mathbf { x } _ { i } , y _ { i } ) \\in S } \\left( y _ { i } - \\theta f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right) ^ { 2 } ,\n$$", + "text_format": "latex", + "bbox": [ + 362, + 646, + 633, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ is a small coefficient to control the effect of initialization, and the data inputs $\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }$ is assumed to follow some unknown distribution $\\tau$ on the unit sphere $\\mathbb { S } ^ { d } \\in \\mathbb { R } ^ { d + 1 }$ . Without loss of generality, we also assume that $| y _ { i } | \\leqslant 1$ . ", + "bbox": [ + 174, + 690, + 825, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first randomly initialize the parameters of the network, and then apply gradient descent to optimize both layers. We present our detailed neural network training algorithm in Algorithm 1. ", + "bbox": [ + 174, + 742, + 820, + 771 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/b9511ab019a200c3aca9055c161b6cd32086034bf8efbfaa8da1f39b2f3d56b3.jpg", + "table_caption": [ + "Algorithm 1 GD for DNNs starting at Gaussian initialization " + ], + "table_footnote": [], + "table_body": "
Input: Number of iterations T,step size n. Generate each entry of W(0) and w)
from N(0,2/m) and N(0,1/m) respectively. 2 fort=0,1,...,T-1do
Update W(t+1) = W(t) - n · VwLs(W(t).
end for Output: W(T).
", + "bbox": [ + 184, + 795, + 754, + 887 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initialization scheme for $\\mathbf { W } ^ { ( 0 ) }$ given in Algorithm 1 is known as He initialization (He et al., 2015). This scheme generates each entry of the weight matrices from a Gaussian distribution with mean zero. The variances of the Gaussian distributions in initialization are chosen following the principle that the initialization does not change the magnitudes of inputs in each layer. The second layer parameter is not associated with the ReLU activation function, thus it is initialized with variance $1 / m$ instead of $2 / m$ . ", + "bbox": [ + 174, + 102, + 825, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 NEURAL TANGENT KERNEL ", + "text_level": 1, + "bbox": [ + 176, + 199, + 408, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Many attempts have been made to study the convergence of gradient descent assuming the width of the network is extremely large (Du et al., 2018b; Li and Liang, 2018). When the width of the network goes to infinity, with certain initialization on parameters, the inner product of gradients of the output function would converge to a limiting kernel, i.e., Neural Tangent Kernel (Jacot et al., 2018). In this paper, we denote it by $\\begin{array} { r } { \\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\operatorname* { l i m } _ { m \\infty } m ^ { - 1 } \\langle \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ) , \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ^ { \\prime } ) \\rangle } \\end{array}$ and we have ", + "bbox": [ + 173, + 226, + 825, + 310 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/302018cca28021cbfcc51c593d3df79466e93feae8679c3a70370982ca9d6fcb.jpg", + "text": "$$\n\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle \\cdot \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) + 2 \\cdot \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 346, + 315, + 648, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 174, + 340, + 217, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/64a6c508918c60ae6b0f1d7ea501c2d922bcc832e1a0e99dfcf96f5e8fb3f56a.jpg", + "text": "$$\n\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] , } \\\\ & { \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 338, + 348, + 658, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we apply gradient descent to both layers, the Neural Tangent Kernel is the sum of two different kernel functions and clearly it can be reduced to one layer training setting. These two kernels are arc-cosine kernels of degree 0 and 1 (Cho and Saul, 2009). Their explicit expressions are given as ", + "bbox": [ + 174, + 391, + 826, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/1a185070145cdd03708c097caaf18bd78644f54477168358a065491a6cefce37.jpg", + "text": "$$\n\\kappa _ { 1 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) , \\kappa _ { 2 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( t \\cdot \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) + \\sqrt { 1 - t ^ { 2 } } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 235, + 439, + 761, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle / \\left( \\left\\| \\mathbf { x } \\right\\| \\left\\| \\mathbf { x } ^ { \\prime } \\right\\| \\right)$ . ", + "bbox": [ + 174, + 474, + 372, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 INTEGRAL OPERATOR", + "text_level": 1, + "bbox": [ + 176, + 502, + 367, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The theory of integral operator with respect to kernel function has been well studied in machine learning (Smale and Zhou, 2007; Rosasco et al., 2010) thus we only give a brief introduction here. Let $L _ { \\tau } ^ { 2 } ( X )$ be the Hilbert space of square-integrable functions with respect to a Borel measure $\\tau$ from $X \\to \\mathbb { R }$ . For any continuous kernel function $\\kappa : X \\times X \\to \\mathbb { R }$ and $\\tau$ we can define an integral operator $L _ { \\kappa }$ on $L _ { \\tau } ^ { 2 } ( X )$ by ", + "bbox": [ + 173, + 529, + 825, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/c171d3e3de6037e1cb648cff9aaf4c52d0260ff0cdab83a9f81089f35558335c.jpg", + "text": "$$\nL _ { \\kappa } ( f ) ( { \\bf x } ) = \\int _ { X } \\kappa ( { \\bf x } , { \\bf y } ) f ( { \\bf y } ) d \\tau ( { \\bf y } ) , ~ { \\bf x } \\in X .\n$$", + "text_format": "latex", + "bbox": [ + 348, + 606, + 648, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It has been pointed out in Cho and Saul (2009) that arc-cosine kernels are positive semi-definite. Thus the kernel function $\\kappa$ defined by (2.1) is positive semi-definite being a product and a sum of positive semi-definite kernels. Clearly this kernel is also continuous and symmetric. Thus we know that the neural tangent kernel $\\kappa$ is a Mercer kernel. ", + "bbox": [ + 173, + 645, + 825, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 MAIN RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 714, + 336, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section we present our main results. In Section 3.1, we give a general result on the convergence rate of gradient descent along different eigendirections of neural tangent kernel. Motivated by this result, in Section 3.2, we give a case study on the spectrum of $L _ { \\kappa }$ when the input data are uniformly distributed over the unit sphere $\\mathbb { S } ^ { d }$ . In Section 3.3, we combine the spectrum analysis with the general convergence result to give explicit convergence rate for uniformly distributed data on the unit sphere. ", + "bbox": [ + 173, + 744, + 825, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 CONVERGENCE ANALYSIS OF GRADIENT DESCENT ", + "text_level": 1, + "bbox": [ + 173, + 840, + 571, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section we study the convergence of Algorithm 1. Instead of studying the standard convergence of loss function value, we aim to provide a refined analysis on the speed of convergence along different directions defined by the eigenfunctions of $L _ { \\kappa }$ . We first introduce the following definitions and notations. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\{ \\lambda _ { i } \\} _ { i \\geqslant 1 }$ with $\\lambda _ { 1 } \\geqslant \\lambda _ { 2 } \\geqslant \\cdots$ be the strictly positive eigenvalues of $L _ { \\kappa }$ , and $\\phi _ { 1 } ( \\cdot ) , \\phi _ { 2 } ( \\cdot ) , \\ldots$ be the corresponding orthonormal eigenfunctions. Set $\\mathbf v _ { i } = n ^ { - 1 / 2 } ( \\phi _ { i } ( { \\mathbf { x } } _ { 1 } ) , \\dots , \\phi _ { i } ( { \\mathbf { x } } _ { n } ) ) ^ { \\top }$ , $i = 1 , 2 , \\dots$ Note that $L _ { \\kappa }$ may have eigenvalues with multiplicities larger than 1 and $\\lambda _ { i }$ , $i \\geqslant 1$ are not distinct. Therefore for any integer $k$ , we define $r _ { k }$ as the sum of the multiplicities of the first $k$ distinct eigenvalues of $L _ { \\kappa }$ . Define $\\mathbf { V } _ { r _ { k } } = \\left( \\mathbf { v } _ { 1 } , \\ldots , \\mathbf { v } _ { r _ { k } } \\right)$ . By definition, $\\mathbf { v } _ { i }$ , $i \\in [ r _ { k } ]$ are rescaled restrictions of orthonormal functions in $L _ { \\tau } ^ { 2 } ( \\mathbb { S } ^ { d } )$ on the training examples. Therefore we can expect them to form a set of almost orthonomal bases in the vector space $\\mathbb { R } ^ { n }$ . The following lemma follows by standard concentration inequality. ", + "bbox": [ + 173, + 102, + 825, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 3.1. Suppose that $| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M$ for all $\\mathbf { x } \\in \\mathbb { S } ^ { d }$ and $i \\in [ r _ { k } ]$ . For any $\\delta > 0$ , with probability at least $1 - \\delta$ , ", + "bbox": [ + 171, + 220, + 823, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0fe6bb516f706d4fdfcaff27a06116d620b9ad1365b8a859543d6c5e38cb71f2.jpg", + "text": "$$\n\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { \\operatorname* { m a x } } \\leqslant C M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n } ,\n$$", + "text_format": "latex", + "bbox": [ + 354, + 252, + 640, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is an absolute constant. ", + "bbox": [ + 174, + 275, + 387, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denote $\\mathbf { y } ~ = ~ ( y _ { 1 } , \\ldots , y _ { n } ) ^ { \\top }$ and $\\widehat { \\mathbf { y } } ^ { ( t ) } ~ = ~ \\theta ~ \\cdot ~ \\left( f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { n } ) \\right)$ , $t ~ = ~ 0 , \\ldots , T$ . Then Lemma 3.1 shows that the convergence rate of $\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }$ roughly represents the speed gradient descent learns the components of the target function corresponding to the first $r _ { k }$ eigenvalues. The following theorem gives the convergence guarantee of $\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }$ . ", + "bbox": [ + 173, + 297, + 825, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3.2. Suppose $| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant M$ for $j \\in [ r _ { k } ]$ and $\\mathbf { x } \\in \\mathbb { S } ^ { d }$ . For any $\\epsilon , \\delta > 0$ and integer $k$ , if $n \\geqslant \\widetilde { \\Omega } ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } , M ^ { 4 } r _ { k } ^ { 2 } \\} )$ , $m \\geqslant \\widetilde { \\Omega } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )$ , then with probability at least $1 - \\delta$ , Algorithm 1 with $\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )$ , $\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )$ satisfies ", + "bbox": [ + 173, + 364, + 825, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/aa0502d3b26d2cd35bc983232428a7e918076766a0a9baf9f51ce7b11fe0222e.jpg", + "text": "$$\nn ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon .\n$$", + "text_format": "latex", + "bbox": [ + 282, + 419, + 715, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 3.3. Theorem 3.2 theoretically reveals the spectral bias of deep learning. Specifically, as long as the network is wide enough and the sample size is large enough, gradient descent first learns the target function along the the eigendirections of neural tangent kernel with larger eigenvalues, and then learns the rest components corresponding to smaller eigenvalues. Moreover, by showing that learning the components corresponding to larger eigenvalues can be done with smaller sample size and narrower networks, our theory pushes the study of neural networks in the NTK regime towards a more practical setting. For these reasons, we believe that Theorem 3.2 to certain extent explains the empirical observations given in Rahaman et al. (2018), and demonstrates that the difficulty of a function to be learned by neural network can be characterized in the eigenspace of neural tangent kernel: if the target function has a component corresponding to a small eigenvalue of neural tangent kernel, then learning this function up to good accuracy takes longer time, and requires more examples and a wider network. ", + "bbox": [ + 173, + 441, + 826, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 SPECTRAL ANALYSIS OF NEURAL TANGENT KERNEL FOR UNIFORM DISTRIBUTION ", + "text_level": 1, + "bbox": [ + 174, + 621, + 794, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After presenting a general theorem (without assumptions on data distribution) in the previous subsection, we now study the case when the data inputs are uniformly distributed over the unit sphere. We present our results (an extension of Proposition 5 in Bietti and Mairal (2019)) of spectral analysis of neural tangent kernel. We show Mercer decomposition of neural tangent kernel for two-layer setting. We give explicit expression of eigenvalues and show orders of eigenvalues in both cases when $d \\gg k$ and $k \\gg d$ . ", + "bbox": [ + 173, + 647, + 825, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3.4. For any $\\mathbf { x } , \\mathbf { x } ^ { \\prime } \\in \\mathbb { S } ^ { d } \\subset \\mathbb { R } ^ { d + 1 }$ , we have the Mercer decomposition of the neural tangent kernel $\\kappa : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\xrightarrow { } \\mathbb { R }$ , ", + "bbox": [ + 173, + 733, + 823, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/135846f45981f6060733fed8b71d51a76f1ee29fe8fc71cc1aabc8fba3e89be2.jpg", + "text": "$$\n\\kappa \\left( \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\right) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N \\left( d , k \\right) } Y _ { k , j } \\left( \\mathbf { x } \\right) Y _ { k , j } \\left( \\mathbf { x } ^ { \\prime } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 351, + 766, + 645, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y _ { k , j }$ for $j = 1 , \\cdots , N ( d , k )$ are linearly independent spherical harmonics of degree ˘ $k$ in $d + 1$ variables with N pd, kq “ 2k\\`d´1 k\\`d´2 and orders of $\\mu _ { k }$ are given by ", + "bbox": [ + 169, + 814, + 823, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "µ0 “ µ1 “ Ωp1q, µk “ 0, k “ 2j \\` 1, \nµ $\\begin{array} { r l } & { \\mathbf { \\Phi } ^ { \\circ _ { 0 } } \\quad \\mathsf { \\Lambda } ^ { \\circ _ { 1 } } \\quad \\cdots \\mathbf { \\Phi } ^ { \\circ _ { k } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } , \\mathsf { w } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } \\quad \\bullet _ { k } , \\mathsf { s } } , } \\\\ & { \\boldsymbol { \\iota } _ { k } = \\Omega \\left( \\operatorname* { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) , k = 2 j , } \\end{array}$ where $j \\in \\mathbb { N } ^ { + }$ . More specifically, we have $\\mu _ { k } = \\Omega \\left( k ^ { - d - 1 } \\right)$ when $k \\gg d$ and $\\mu _ { k } = \\Omega \\left( d ^ { - k + 1 } \\right)$ when $d \\gg k$ , $k = 2 , 4 , 6 , \\ldots$ . ", + "bbox": [ + 173, + 852, + 839, + 925 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 3.5. In the above theorem, the coefficients $\\mu _ { k }$ are actually different eigenvalues of the integral operator $L _ { \\kappa }$ on $L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } )$ defined by ", + "bbox": [ + 169, + 103, + 823, + 133 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/7930b891e545c6fba1ef7aa60a6151f14fa8437ffc5f7d2bc9ead1041f3f8d04.jpg", + "text": "$$\nL _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 325, + 141, + 671, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau _ { d }$ is the uniform probability measure on unit sphere $\\mathbb { S } ^ { d }$ . Therefore the $\\lambda _ { r _ { k } }$ in Theorem 3.2 is just $\\mu _ { k - 1 }$ given in Theorem 3.4 when $\\tau _ { d }$ is uniform distribution. ", + "bbox": [ + 171, + 184, + 821, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 3.6. Vempala and Wilmes (2018) studied two-layer neural networks with sigmoid activation function, and established guarantees to achieve $\\epsilon _ { 0 } + \\epsilon$ error with iteration complexity $T = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\log { ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) } }$ under the over-parameterization condition $m = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\mathrm { p o l y } ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) }$ , where $f ^ { * }$ is the target function, and $\\epsilon _ { \\mathrm { 0 } }$ is certain function approximation error. Another highly related work is Bietti and Mairal (2019), which gives $\\mu _ { k } = \\bar { \\Omega ( k ^ { - } { } ^ { d - 1 } ) }$ . The order of eigenvalues we present appears as $\\mu _ { k } = \\Omega ( \\operatorname* { m a x } ( k ^ { - d - 1 } , d ^ { - k + \\bar { 1 } } ) )$ . This is better when $d \\gg k$ , which is closer to the practical setting. ", + "bbox": [ + 173, + 217, + 825, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 EXPLICIT CONVERGENCE RATE FOR UNIFORMLY DISTRIBUTED DATA ", + "text_level": 1, + "bbox": [ + 176, + 329, + 700, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we combine our results in the previous two subsections and give explicit convergence rate for uniformly distributed data on the unit sphere. Note that the first $k$ distinct eigenvalues of NTK have spherical harmonics up to degree $k - 1$ as eigenfunctions. ", + "bbox": [ + 173, + 357, + 825, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corollary 3.7. Suppose that $k \\gg d$ , and the sample $\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }$ follows the uniform distribution $\\tau _ { d }$ on the unit sphere $\\mathbb { S } ^ { d }$ . For any $\\epsilon , \\delta > 0$ and integer $k$ , if $n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ k ^ { 2 d + 2 } , k ^ { 2 d - 2 } r _ { k } ^ { 2 } \\} )$ , $m \\geqslant$ $\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , k ^ { d + 1 } , \\epsilon ^ { - 1 } ) )$ , then with probability at least $1 - \\delta$ , Algorithm 1 with $\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )$ , $\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )$ satisfies ", + "bbox": [ + 173, + 404, + 825, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/27f6edf86d644addcd7dcf90f8781ebeed175b895cdf5d0ffc184a0a7a5dc6c5.jpg", + "text": "$$\nn ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( k ^ { - d - 1 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ,\n$$", + "text_format": "latex", + "bbox": [ + 254, + 478, + 741, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where thono $\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}$ and s of d $\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }$ with $\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }$ being a set of or$k - 1$ ", + "bbox": [ + 173, + 511, + 825, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corollary 3.8. Suppose that $d \\ \\gg \\ k$ , and the sample $\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }$ follows the uniform distribution $\\tau _ { d }$ on the unit sphere $\\mathbb { S } ^ { d }$ . For any $\\epsilon , \\delta \\ > \\ 0$ and integer $k$ , if $n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } d ^ { 2 k - 2 } r _ { k } ^ { 2 } )$ , $m \\geqslant$ $\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , d ^ { k - 2 } , \\epsilon ^ { - 1 } ) )$ , then with probability at least $1 - \\delta$ , Algorithm 1 with $\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )$ , $\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )$ satisfies ", + "bbox": [ + 173, + 545, + 825, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/888ef890c7e65b65d51416cf26c74d8141f2f8a4fb56aa9afcf8781c812ac20a.jpg", + "text": "$$\nn ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( d ^ { - k + 2 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ,\n$$", + "text_format": "latex", + "bbox": [ + 254, + 619, + 743, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}$ and $\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }$ with $\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }$ being a set of orthonomal spherical harmonics of degrees up to $k - 1$ . ", + "bbox": [ + 171, + 652, + 823, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corollaries 3.7 and 3.8 further illustrate the spectral bias of neural networks by providing exact calculations of $\\lambda _ { r _ { k } }$ , $\\mathbf { V } _ { r _ { k } }$ and $M$ in Theorem 3.2. They show that if the input distribution is uniform over unit sphere, then spherical harmonics with lower degrees are learned first by over-parameterized neural networks. ", + "bbox": [ + 173, + 695, + 825, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 3.9. In Corollaries 3.7 and 3.8, it shows that the conditions on $n$ and $m$ depend exponentially on either $k$ or $d$ . We would like to emphasize that such exponential dependency is reasonable and unavoidable. In our case, we can take the $d \\gg k$ setting as an example. The exponential dependency in $k$ is a natural consequence of the fact that in high dimensional space, there are a large number of linearly independent polynomials even for very low degrees. It is apparently only reasonable to expect to learn less than $n$ independent components of the true function, which means that it is unavoidable to assume ", + "bbox": [ + 173, + 756, + 825, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/6c96a4d6ed0e203c822fe2ead8f55b3f95a11507788014b118543f9201920818.jpg", + "text": "$$\nn \\geqslant r _ { k } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { d - 1 } } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { k ^ { \\prime } - 1 } } = \\Omega ( d ^ { k - 1 } ) .\n$$", + "text_format": "latex", + "bbox": [ + 183, + 859, + 812, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar arguments can apply to the requirement of $m$ and the $k \\gg d$ setting. ", + "bbox": [ + 173, + 909, + 669, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 326, + 118 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we illustrate our results by training neural networks on synthetic data. Across all tasks, we train a two-layer hidden neural networks with 4096 neurons and initialize it exactly as defined in the setup. The optimization method is vanilla full gradient descent. We sample 1000 training data which is uniformly sampled from the unit sphere in $\\mathbb { R } ^ { 1 0 }$ . ", + "bbox": [ + 173, + 131, + 825, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 LEARNING COMBINATION OF SPHERICAL HARMONICS ", + "text_level": 1, + "bbox": [ + 173, + 200, + 588, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we show a result when the target function is exactly linear combination of spherical harmonics. The target function is explicitly defined as ", + "bbox": [ + 173, + 227, + 823, + 255 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/ac2cf3be7b3317edb1a1ebaf573042ae9e38d677bfaa4a01ed7fd8d8b6497d9a.jpg", + "text": "$$\nf ^ { * } ( { \\bf x } ) = a _ { 1 } \\cdot P _ { 1 } ( \\langle \\zeta _ { 1 } , { \\bf x } \\rangle ) + a _ { 2 } \\cdot P _ { 2 } ( \\langle \\zeta _ { 2 } , { \\bf x } \\rangle ) + a _ { 4 } \\cdot P _ { 4 } ( \\langle \\zeta _ { 4 } , { \\bf x } \\rangle ) ,\n$$", + "text_format": "latex", + "bbox": [ + 285, + 261, + 709, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the $P _ { k } ( t )$ is the Gegenbauer polynomial, and $\\zeta _ { k } , k = 1 , 2 , 4$ are fixed vectors that are inde \npendently generated from uniformaccording to the addition formula $\\begin{array} { r } { \\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) } \\end{array}$ $\\mathbb { R } ^ { 1 0 }$ riments. Note that, every normalized $f ^ { * } ( \\mathbf { x } )$ \nmonics of order 1,2 and 4. The higher odd-order Gegenbauer polynomials are omitted because the \nspectral analysis showed that $\\mu _ { k } = 0$ for $k = 3 , 5 , 7 \\dots$ . ", + "bbox": [ + 173, + 284, + 825, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the setting in section 3.1, we denote $\\mathbf v _ { k } = n ^ { - 1 / 2 } ( P _ { k } ( \\mathbf x _ { 1 } ) , \\dots , P _ { k } ( \\mathbf x _ { n } ) )$ . By Lemma 3.2 $\\mathbf { v } _ { k }$ ’s are almost orthonormal with high probability. So we can define the (approximate) projection length of residual $\\mathbf { r } ^ { ( t ) }$ onto $\\mathbf { v } _ { k }$ at step $t$ as ", + "bbox": [ + 174, + 381, + 825, + 426 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/80aa5faf8d40795eb671a99a14e992f04309702eafc8773d859aeab898cd2c85.jpg", + "text": "$$\n\\widehat { a } _ { k } = | \\mathbf { v } _ { k } ^ { \\top } \\mathbf { r } ^ { ( t ) } | ,\n$$", + "text_format": "latex", + "bbox": [ + 449, + 431, + 547, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf { r } ^ { ( t ) } = \\left( f ^ { * } \\big ( \\mathbf { x } _ { 1 } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { 1 } \\big ) , \\ldots , f ^ { * } \\big ( \\mathbf { x } _ { n } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { n } \\big ) \\right)$ and $f _ { \\mathbf { W } ( t ) } \\left( \\mathbf { x } \\right)$ is the neural network function. ", + "bbox": [ + 173, + 458, + 823, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remark 4.1. Here $\\widehat { a } _ { k }$ is the projection length onto an approximate vector. In the function space, we can also project the residual function $r ( \\mathbf { x } ) \\bar { = } f ^ { * } ( \\mathbf { x } ) - \\bar { \\theta } \\bar { f } _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )$ onto the orthonormal Gegenbauer functions $P _ { k } ( { \\bf x } )$ . Replacing the training data with randomly sampled data points $\\mathbf { x } _ { i }$ can lead to a random estimate of the projection length in function space. We provide the corresponding result for freshly sampled points in Appendix E.1. ", + "bbox": [ + 173, + 491, + 825, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2771d3c8fdd44ee955473f67bdb5bc958722cdc8a8e43ad240100a6f7d39531f.jpg", + "image_caption": [ + "Figure 1: Convergence curve for projection length onto different components. (a) shows the curve when the target function have different component with the same scale. (b) shows the curve when the higner-order components have larger scale. Both illustrate that the lower-order components are learned faster. Log-scale figures are shown in Appendix E.2. " + ], + "image_footnote": [], + "bbox": [ + 181, + 575, + 813, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are showned in Figure 1. It can be seen that at the beginning of training, the residual at the lowest frequency ( $k = 1 .$ ) converges to zero first and then the second lowest $k = 2$ ). The highest frequency component is the last one to converge. Following the setting of Rahaman et al. (2018) we assign high frequencies a larger scale, expecting that larger scale will introduce a better descending speed. Still, the low frequencies are regressed first. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Apart from the synthesized low frequency function, we also showed the dynamics of normal functions’ projection to $P _ { k } ( x )$ . These functions, though in a simple form, have non-zero components in almost all frequencies. In this subsection we further show our results still apply when all frequenciesř ř exist in the target function, which is given by $\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\cos ( a _ { i } \\langle \\zeta , \\mathbf { x } \\rangle ) } \\end{array}$ or $\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\bar { \\langle \\zeta , \\mathbf { x } \\rangle } ^ { p _ { i } } } \\end{array}$ , where $\\zeta$ is a fixed unit vector. The coefficients of given components are calculated in the same way as in Section 4.1. ", + "bbox": [ + 173, + 132, + 825, + 217 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edeb7fc8e44d199514a417b12ca7990246224162dede7c2aaebaca80a02339cc.jpg", + "image_caption": [ + "Figure 2: Convergence curve for different components. (a) shows the curve of a trigonometric function. (b) shows the curve of a polynomial with even degrees. Both exhibits similar tendency as combination of spherical harmonics. " + ], + "image_footnote": [], + "bbox": [ + 181, + 226, + 815, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 2 shows that even for arbitrarily chosen functions of simple form, the networks can still first learn the low frequency components of the target function. Notice that early in training not all the curves may descend, we believe this is due to the unseen components’ influence on the gradient. Again, as the training proceeds, the convergence is controlled at the predicted rate. ", + "bbox": [ + 173, + 489, + 825, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark 4.2. The reason why we only use cosine function and even polynomial is that the only odd basis function with non-zero eigenvalue is $P _ { 1 } ( \\mathbf { x } )$ . To show a general tendency it is better to restrict the target function in the even function space. ", + "bbox": [ + 174, + 555, + 825, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 CONCLUSION AND DISCUSSION ", + "text_level": 1, + "bbox": [ + 176, + 621, + 468, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we give theoretical justification for spectral bias through a detailed analysis of the convergence behavior of two-layer neural networks with ReLU activation function. We show that the convergence of gradient descent in different directions depends on the corresponding eigenvalues and essentially exhibits different convergence rates. We show Mercer decomposition of neural tangent kernel and give explicit order of eigenvalues of integral operator with respect to the neural tangent kernel when the data is uniformly distributed on the unit sphere $\\mathbb { S } ^ { d }$ . Combined with the convergence analysis, we give the exact order of convergence rate on different directions. We also conduct experiments on synthetic data to support our theoretical result. ", + "bbox": [ + 173, + 651, + 825, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "So far, we have considered the upper bound for convergence with respect to low frequency components and present comprehensive theorem to explain the spectral bias. One desired improvement is to give the lower bound of convergence with respect to high frequency components, which is essential to establish tighter characterization of spectral-biased optimization. It is also interesting to extend our result to other training algorithms like Adam, where the analysis in Wu et al. (2019); Zhou et al. (2018) might be implemented with a more careful quantification on the projection of residual along different directions. Another potential improvement is to generalize the result to multi-layer neural networks, which might require different techniques since our analysis heavily rely on exactly computing the eigenvalues of the neural tangent kernel. It is also an important direction to weaken the requirement on over-parameterization, or study the spectral bias in a non-NTK regime to furthur close the gap between theory and practice. ", + "bbox": [ + 173, + 771, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES \nALLEN-ZHU, Z., LI, Y. and SONG, Z. (2018). A convergence theory for deep learning via overparameterization. arXiv preprint arXiv:1811.03962 . \nANDONI, A., PANIGRAHY, R., VALIANT, G. and ZHANG, L. (2014). Learning polynomials with neural networks. In International Conference on Machine Learning. \nARORA, S., DU, S. S., HU, W., LI, Z., SALAKHUTDINOV, R. and WANG, R. (2019a). On exact computation with an infinitely wide neural net. arXiv preprint arXiv:1904.11955 . \nARORA, S., DU, S. S., HU, W., LI, Z. and WANG, R. (2019b). Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. arXiv preprint arXiv:1901.08584 . \nATKINSON, K. and HAN, W. (2012). Spherical harmonics and approximations on the unit sphere: an introduction, vol. 2044. Springer Science & Business Media. \nBACH, F. (2017). Breaking the curse of dimensionality with convex neural networks. Journal of Machine Learning Research 18 1–53. \nBASRI, R., JACOBS, D., KASTEN, Y. and KRITCHMAN, S. (2019). The convergence rate of neural networks for learned functions of different frequencies. arXiv preprint arXiv:1906.00425 . \nBIETTI, A. and MAIRAL, J. (2019). On the inductive bias of neural tangent kernels. arXiv preprint arXiv:1905.12173 . \nCAO, Y. and GU, Q. (2019a). Generalization bounds of stochastic gradient descent for wide and deep neural networks. arXiv preprint arXiv:1905.13210 . \nCAO, Y. and GU, Q. (2019b). A generalization theory of gradient descent for learning overparameterized deep relu networks. arXiv preprint arXiv:1902.01384 . \nCHIZAT, L. and BACH, F. (2018). A note on lazy training in supervised differentiable programming. arXiv preprint arXiv:1812.07956 . \nCHO, Y. and SAUL, L. K. (2009). Kernel methods for deep learning. In Advances in neural information processing systems. \nCOLLOBERT, R. and WESTON, J. (2008). A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning. ACM. \nDU, S. S., LEE, J. D., LI, H., WANG, L. and ZHAI, X. (2018a). Gradient descent finds global minima of deep neural networks. arXiv preprint arXiv:1811.03804 . \nDU, S. S., ZHAI, X., POCZOS, B. and SINGH, A. (2018b). Gradient descent provably optimizes over-parameterized neural networks. arXiv preprint arXiv:1810.02054 . \nFRYE, C. and EFTHIMIOU, C. J. (2012). Spherical harmonics in p dimensions. arXiv preprint arXiv:1205.3548 . \nGUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018a). Characterizing implicit bias in terms of optimization geometry. arXiv preprint arXiv:1802.08246 . \nGUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018b). Implicit bias of gradient descent on linear convolutional networks. arXiv preprint arXiv:1806.00468 . \nHE, K., ZHANG, X., REN, S. and SUN, J. (2015). Delving deep into rectifiers: Surpassing humanlevel performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision. \nHE, K., ZHANG, X., REN, S. and SUN, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition. \nHINTON, G., DENG, L., YU, D., DAHL, G., MOHAMED, A.-R., JAITLY, N., SENIOR, A., VANHOUCKE, V., NGUYEN, P., KINGSBURY, B. ET AL. (2012). Deep neural networks for acoustic modeling in speech recognition. IEEE Signal processing magazine 29. \nJACOT, A., GABRIEL, F. and HONGLER, C. (2018). Neural tangent kernel: Convergence and generalization in neural networks. arXiv preprint arXiv:1806.07572 . \nLI, Y. and LIANG, Y. (2018). Learning overparameterized neural networks via stochastic gradient descent on structured data. arXiv preprint arXiv:1808.01204 . \nNAKKIRAN, P., KAPLUN, G., KALIMERIS, D., YANG, T., EDELMAN, B. L., ZHANG, F. and BARAK, B. (2019). Sgd on neural networks learns functions of increasing complexity. arXiv preprint arXiv:1905.11604 . \nRAHAMAN, N., BARATIN, A., ARPIT, D., DRAXLER, F., LIN, M., HAMPRECHT, F. A., BENGIO, Y. and COURVILLE, A. (2018). On the spectral bias of neural networks. arXiv preprint arXiv:1806.08734 . \nROSASCO, L., BELKIN, M. and VITO, E. D. (2010). On learning with integral operators. Journal of Machine Learning Research 11 905–934. \nSMALE, S. and ZHOU, D.-X. (2007). Learning theory estimates via integral operators and their approximations. Constructive approximation 26 153–172. \nSMALE, S. and ZHOU, D.-X. (2009). Geometry on probability spaces. Constructive Approximation 30 311. \nSOUDRY, D., HOFFER, E. and SREBRO, N. (2017). The implicit bias of gradient descent on separable data. arXiv preprint arXiv:1710.10345 . \nSU, L. and YANG, P. (2019). On learning over-parameterized neural networks: A functional approximation prospective. arXiv preprint arXiv:1905.10826 . \nVEMPALA, S. and WILMES, J. (2018). Gradient descent for one-hidden-layer neural networks: Polynomial convergence and sq lower bounds. arXiv preprint arXiv:1805.02677 . \nWU, X., DU, S. S. and WARD, R. (2019). Global convergence of adaptive gradient methods for an over-parameterized neural network. arXiv preprint arXiv:1902.07111 . \nXU, Z. J. (2018). Understanding training and generalization in deep learning by fourier analysis. arXiv preprint arXiv:1808.04295 . \nYANG, G. and SALMAN, H. (2019). A fine-grained spectral perspective on neural networks. arXiv preprint arXiv:1907.10599 . \nZHANG, C., BENGIO, S., HARDT, M., RECHT, B. and VINYALS, O. (2016). Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530 . \nZHOU, D., TANG, Y., YANG, Z., CAO, Y. and GU, Q. (2018). On the convergence of adaptive gradient methods for nonconvex optimization. arXiv preprint arXiv:1808.05671 . \nZOU, D., CAO, Y., ZHOU, D. and GU, Q. (2018). Stochastic gradient descent optimizes overparameterized deep relu networks. arXiv preprint arXiv:1811.08888 . \nZOU, D. and GU, Q. (2019). An improved analysis of training over-parameterized deep neural networks. arXiv preprint arXiv:1906.04688 . ", + "bbox": [ + 171, + 68, + 826, + 926 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 60, + 826, + 820 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A REVIEW ON SPHERICAL HARMONICS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 516, + 118 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we give a brief review on relevant concepts in spherical harmonics. For more detials, see Bach (2017), Bietti and Mairal (2019), Frye and Efthimiou (2012) and Atkinson and Han (2012) for references. ", + "bbox": [ + 173, + 131, + 825, + 174 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We consider the unit sphere $\\mathbb { S } ^ { d } = \\left\\{ \\mathbf { x } \\in \\mathbb { R } ^ { d + 1 } : \\| \\mathbf { x } \\| _ { 2 } = 1 \\right\\}$ , whose surface area is given by $\\omega _ { d } =$ $2 \\pi ^ { ( d + 1 ) / 2 } / \\Gamma ( ( d + 1 ) / 2 )$ and denote $\\tau _ { d }$ the uniform measure on the sphere. ", + "bbox": [ + 174, + 179, + 821, + 213 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For any ! $k \\qquad \\geqslant \\qquad 1 ,$ we consider) a set of spherical harmonics $\\begin{array} { r } { \\{ Y _ { k , j } : \\mathbb { S } ^ { d } \\mathbb { R } | 1 \\leqslant j \\leqslant N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\} } \\end{array}$ They form an orthonormal basis and satisfy the following equation $\\langle Y _ { k i } , Y _ { s j } \\rangle _ { \\mathbb { S } ^ { d } } = \\int _ { \\mathbb { S } ^ { d } } Y _ { k i } ( x ) Y _ { s j } ( x ) d \\tau _ { d } ( x ) = \\delta _ { i j } \\delta _ { s k }$ . Moreover, since they are homogeneous functions of degree $k$ , it is clear that $Y _ { k } ( x )$ has the same parity as $k$ . ", + "bbox": [ + 173, + 218, + 826, + 287 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We have the addition formula ", + "bbox": [ + 174, + 294, + 369, + 308 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/e4e6a6a3007539b5cbae000e6184154a231ac93d9516c781bfe369055e2c35bb.jpg", + "text": "$$\n\\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) ,\n$$", + "text_format": "latex", + "bbox": [ + 349, + 311, + 647, + 358 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $P _ { k } ( t )$ is the Legendre polynomial of degree $k$ in $d + 1$ dimensions, explicitly given by (Rodrigues’ formula) ", + "bbox": [ + 173, + 363, + 823, + 392 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/884d10edd575ec5a70028342121de9ec7280e05c1976d7891734223d069066ae.jpg", + "text": "$$\nP _ { k } ( t ) = \\left( - { \\frac { 1 } { 2 } } \\right) ^ { k } { \\frac { \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( k + { \\frac { d } { 2 } } \\right) } } \\left( 1 - t ^ { 2 } \\right) ^ { { \\frac { 2 - d } { 2 } } } \\left( { \\frac { d } { d t } } \\right) ^ { k } \\left( 1 - t ^ { 2 } \\right) ^ { k + { \\frac { d - 2 } { 2 } } } .\n$$", + "text_format": "latex", + "bbox": [ + 281, + 397, + 717, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We can also see that $P _ { k } ( t )$ , the Legendre polynomial of degree $k$ shares the same parity with $k$ . By the orthogonality and the addition formula (A.1) we have, ", + "bbox": [ + 174, + 444, + 821, + 473 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/3b995ffb6893c07c99ad26afad82d3f241cd2dbd0583fab1d660f4a3aa09f22a.jpg", + "text": "$$\n\\int _ { \\mathbb { S } ^ { d } } P _ { j } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) P _ { k } ( \\langle \\mathbf { w } , \\mathbf { y } \\rangle ) d \\tau _ { d } ( \\mathbf { w } ) = \\frac { \\delta _ { j k } } { N ( d , k ) } P _ { k } ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) .\n$$", + "text_format": "latex", + "bbox": [ + 305, + 478, + 691, + 512 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Further we have the recurrence relation for the Legendre polynomials, ", + "bbox": [ + 174, + 525, + 633, + 540 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/5cd19607dcf4de36ab35db86e5993c63f7e15e6ed25976ca5eb0e00d04d64014.jpg", + "text": "$$\nt P _ { k } ( t ) = \\frac { k } { 2 k + d - 1 } P _ { k - 1 } ( t ) + \\frac { k + d - 1 } { 2 k + d - 1 } P _ { k + 1 } ( t ) ,\n$$", + "text_format": "latex", + "bbox": [ + 320, + 545, + 676, + 578 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "for $k \\geqslant 1$ and $t P _ { 0 } ( t ) = P _ { 1 } ( t )$ for $k = 0$ . ", + "bbox": [ + 173, + 583, + 441, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The Hecke-Funk formula is given for a spherical harmonic $Y _ { k }$ of degree $k$ ", + "bbox": [ + 174, + 604, + 660, + 619 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/27e6f855d432637c79a63af9f93cfed267456e457dca54264bdb452130f7b337.jpg", + "text": "$$\n\\int _ { \\mathbb { S } ^ { d } } f ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) Y _ { k } ( \\mathbf { y } ) d t ^ { \\prime } d ( \\mathbf { y } ) = { \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } } Y _ { k } ( \\mathbf { x } ) \\int _ { - 1 } ^ { 1 } f ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t .\n$$", + "text_format": "latex", + "bbox": [ + 250, + 625, + 746, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B PROOF OF MAIN THEOREMS ", + "text_level": 1, + "bbox": [ + 174, + 676, + 446, + 693 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B.1 PROOF OF THEOREM 3.2 ", + "text_level": 1, + "bbox": [ + 174, + 707, + 390, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section we give the proof of Theorem 3.2. The core idea of our proof is to establish connections between neural network gradients throughout training and the neural tangent kernel. To do so, we first introduce the following definitions and notations. ", + "bbox": [ + 173, + 734, + 825, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Define $\\begin{array} { r l r } { \\mathbf { K } ^ { ( 0 ) } } & { { } = } & { m ^ { - 1 } ( \\left. \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\right. ) _ { n \\times n } } \\end{array}$ , $\\begin{array} { r l r } { { \\bf K } ^ { ( \\infty ) } } & { { } = } & { ( \\kappa ( { \\bf x } _ { i } , { \\bf x } _ { j } ) ) _ { n \\times n } \\quad = } \\end{array}$ $\\mathrm { l i m } _ { m \\infty } \\mathbf { K } ^ { ( 0 ) }$ . Let $\\{ \\widehat \\lambda _ { i } \\} _ { i = 1 } ^ { n } , \\widehat \\lambda _ { 1 } \\geqslant \\ \\cdot \\cdot \\ \\geqslant \\ \\widehat \\lambda _ { n }$ be the eigenvalues of $n ^ { - 1 } \\mathbf { K } ^ { \\infty }$ , and $\\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n }$ be the corresponding eigenvectors. Set $\\widehat { \\mathbf { V } } _ { r _ { k } } = ( \\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { r _ { k } } ) , \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } = ( \\widehat { \\mathbf { v } } _ { r _ { k } + 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n } )$ . ", + "bbox": [ + 173, + 782, + 825, + 835 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For notation simplicity, we denote ˇ $\\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ = \\ \\left[ \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , \\ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ =$ $\\left[ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , l = 1 , 2 .$ ", + "bbox": [ + 174, + 842, + 823, + 878 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The following lemma is a direct application of Proposition 1 in Smale and Zhou (2009) or Proposition 10 in Rosasco et al. (2010). It bounds the difference between the eigenvalues of NTK and their finite-width counterparts. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lemma B.1. For any $\\delta > 0$ , with probability at least $1 - \\delta$ , ", + "bbox": [ + 173, + 102, + 568, + 119 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/12ac2b7883baa3ecb8bd06e1e53798cd5a9113a218a23e671436df30c82f5fb6.jpg", + "text": "$$\n| \\lambda _ { i } - \\widehat { \\lambda } _ { i } | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( 1 / \\delta ) / n } ) .\n$$", + "text_format": "latex", + "bbox": [ + 397, + 132, + 599, + 154 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The following lemma is partly summarized from the proof of equation (44) in $\\mathrm { S u }$ and Yang (2019). Its purpose is to further connect the eigenfunctions of NTK with their finite-width, finite-sample counterparts. ", + "bbox": [ + 173, + 174, + 825, + 217 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lemma B.2. Suppose that $| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M$ for all $\\mathbf { x } \\in S ^ { d }$ . There exist absolute constants $C , C ^ { \\prime } , c ^ { \\prime \\prime } >$ 0, such that for any $\\delta > 0$ and integer $k$ with $r _ { k } \\leqslant n$ , if $n \\geqslant C ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } \\log ( 1 / \\delta )$ , then with probability at least $1 - \\delta$ , ", + "bbox": [ + 173, + 223, + 825, + 268 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/a68d8f2a70e385c01fe2cca5a61b71ac85ee1277f50aebc5731f18c8e0346861.jpg", + "text": "$$\n\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\| _ { F } \\leqslant C ^ { \\prime } \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } , } \\\\ & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\widehat { \\mathbf { V } } _ { r _ { k } } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant C ^ { \\prime \\prime } \\bigg [ \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M ^ { 2 } r _ { k } \\sqrt { \\frac { \\log ( r _ { k } / \\delta ) } { n } } \\bigg ] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 227, + 281, + 771, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The following two lemmas are key to characterize the dynamics of the residual throughout training. These two results, especially Lemma B.4, are the ones that distinguish our analysis from previous works on neural network training in the NTK regime (Du et al., 2018b; Su and Yang, 2019), since our analysis provides more careful characterization on the residual along different directions. ", + "bbox": [ + 174, + 378, + 823, + 436 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lemma B.3. Suppose that the iterates of gradient descent $\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }$ are inside the ball $B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . If $\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )$ , then with probability at least $1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ , ", + "bbox": [ + 171, + 444, + 823, + 477 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/33ee9e550d82b16b8bcfad2d23c97bd171cd4e1bd0feb467f5f6c86f3c38ad34.jpg", + "text": "$$\n\\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } + 1 ) } = [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) + \\mathbf { e } ^ { ( t ) } , \\ \\| \\mathbf { e } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 186, + 489, + 810, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "for all $t ^ { \\prime } = 0 , \\ldots , t - 1$ , where $\\mathbf { y } = \\left( y _ { 1 } , \\ldots , y _ { n } \\right) ^ { \\top } , \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } = \\theta \\cdot \\left( f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { n } ) \\right) ^ { \\top } .$ ", + "bbox": [ + 171, + 525, + 795, + 545 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lemma B.4. Suppose that the iterates of gradient descent $\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }$ are inside the ball $B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . If $\\omega \\leqslant \\widetilde { \\mathcal { O } } ( \\operatorname* { m i n } \\{ [ \\log ( m ) ] ^ { - 3 / 2 } , \\lambda _ { r _ { k } } ^ { 3 } , ( \\eta m ) ^ { - 3 } \\} )$ and $n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )$ , then with probability at least $1 - \\mathcal { O } ( t ^ { 2 } n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ , ", + "bbox": [ + 171, + 551, + 825, + 603 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/82373c60e69c8dbaa0ed81f5468aea31025051dbee60a3ac6b8b195bb9e56718.jpg", + "text": "$$\n\\begin{array} { r l r } { { ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 1 ) } } \\\\ & { } & { \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) t ^ { \\prime } \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r k } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { n } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } \\cdot \\widetilde { O } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { } & { \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } ) \\cdot \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) ^ { t ^ { \\prime } } + \\widetilde { \\mathcal { O } } ( \\sqrt { n } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r k } ) ^ { - 1 } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 3 ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 613, + 839, + 727 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "for all $t ^ { \\prime } = 0 , \\ldots , t - 1$ ", + "bbox": [ + 174, + 739, + 333, + 755 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Now we are ready to prove Theorem 3.2. ", + "bbox": [ + 174, + 770, + 441, + 785 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proof of Theorem 3.2. Define $\\omega \\ = \\ \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )$ for some small enough absolute constant $\\overline { { C } }$ . Then by union bound, as long as $m \\geqslant \\widetilde \\Omega ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )$ , the conditions on $\\omega$ given in Lemmas D.2, D.4, D.5, B.3 and B.4 are all satisfied. ", + "bbox": [ + 173, + 824, + 825, + 872 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We first show that all the iterates $\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( T ) }$ are inside the ball $B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . We prove this result by inductively show that $\\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) , t = 0 , \\ldots , T$ . First of all, it is clear that ${ \\bf W } ^ { ( 0 ) } \\in$ $B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . Suppose that $\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . Then the results of Lemmas D.2, D.4, ", + "bbox": [ + 174, + 876, + 825, + 925 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D.5, B.3 and B.4 hold for $\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }$ . Denote $\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) }$ , $t \\in T$ . Then we have ", + "bbox": [ + 169, + 101, + 764, + 119 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/a4a45bbc2922014d5419d9875c28672e315f8cfe28ac3e78a5c0b5c817a6f601.jpg", + "text": "$$\n\\begin{array} { r l } { \\| \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( 0 ) } \\| _ { F } \\leqslant } & { \\displaystyle \\sum _ { \\ell = 1 } ^ { t } | \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( t ^ { \\prime } ) } | \\| _ { F } } \\\\ & { \\qquad = \\displaystyle \\prod _ { \\ell = 0 } ^ { t } \\left| \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\left( g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right) \\cdot \\vartheta \\cdot \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right| \\Bigg | _ { F } } \\\\ & { \\leqslant \\eta \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| \\cdot \\| \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n - 1 } \\| \\mathbf { y } - \\vartheta \\cdot \\nabla _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\| \\mathbf { y } - \\widehat { \\mathbf { g } } ^ { ( \\ell ^ { \\prime } ) } \\| _ { 2 } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 235, + 125, + 764, + 335 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where the second inequality follows by Lemma D.4. By Lemma B.4, we have ", + "bbox": [ + 173, + 337, + 686, + 353 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/9b48a0e78cb33a76c618dd9a713c233217aad8b7a4c0a81a2badba5110f5424c.jpg", + "text": "$$\n\\sum _ { t ^ { \\prime } = 0 } ^ { t } \\| \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\widetilde { \\mathcal { O } } ( T \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\lambda _ { r _ { k } } ^ { - 1 } T ^ { 2 } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } )\n$$", + "text_format": "latex", + "bbox": [ + 169, + 357, + 816, + 397 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It then follows by the choice $\\omega \\ = \\ \\overline { { C } } T / ( \\theta \\lambda _ { r _ { k } } \\sqrt { m } )$ , $\\eta ~ = ~ { \\tilde { \\mathcal { O } } } ( ( m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } )$ , $\\theta \\ : = \\ : \\widetilde { \\mathcal { O } } ( \\epsilon )$ and the assumption $m \\geqslant \\tilde { \\mathcal { O } } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )$ that $\\| \\mathbf { W } _ { l } ^ { ( t + 1 ) } - \\mathbf { W } _ { l } ^ { ( 0 ) } \\| _ { F } \\leqslant \\omega$ , $l = 1 , 2$ . Therefore by induction, we see that with probability at least $1 { - } \\mathcal { O } ( T ^ { 3 } n ^ { 2 } ) { \\cdot } \\exp [ { - \\Omega ( m \\omega ^ { 2 / 3 } ) } ]$ s, $\\mathbf { W } ( 0 ) , \\ldots , \\mathbf { W } ( T ) \\in$ $B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . ", + "bbox": [ + 173, + 404, + 825, + 472 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Applying Lemma B.4 then gives ", + "bbox": [ + 174, + 478, + 390, + 492 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/878d859ddd6754ed37a3a15388ddc44c999cba13c8b28854e79eac451201f04b.jpg", + "text": "$$\n\\begin{array} { r l } & { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { \\qquad + T \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\qquad + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot n ^ { - 1 / 2 } \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 240, + 494, + 756, + 563 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Now by $\\omega = \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )$ , $\\eta = \\widetilde { \\mathcal { O } } ( \\theta ^ { 2 } m ) ^ { - 1 }$ and the assumption that $m \\geqslant m ^ { * } = \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 1 4 } \\cdot \\epsilon ^ { - 6 } )$ , we obtain ", + "bbox": [ + 169, + 565, + 825, + 595 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/f67a04ba5eeaab3dc35a4c3983140fcfabb1c9fb2c9089446e0ef81420c23f3f.jpg", + "text": "$$\n\\begin{array} { r } { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon / 1 6 . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 245, + 597, + 758, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "By Lemma 3.1, $\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )$ and the assumptions $n \\geqslant \\widetilde \\Omega ( \\operatorname* { m a x } \\{ \\epsilon ^ { - 1 } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } , \\epsilon ^ { - 2 } M ^ { 2 } r _ { k } ^ { 2 } \\} )$ , the eigenvalues of $\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } }$ are all between $1 / \\sqrt { 2 }$ and $\\sqrt { 2 }$ . Therefore by Lemma B.2 we have ", + "bbox": [ + 173, + 625, + 830, + 659 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/d4a7d0dc2c257dba314e6d313f46ae5d2eb36d0af3c270e18e1598bf0a21dc51.jpg", + "text": "$$\n\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } - \\| ( \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ) ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\sqrt { n } \\cdot \\mathcal { O } \\biggl ( \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M r _ { k } \\sqrt { \\frac { \\log ( r _ { k } } { n } } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\epsilon \\sqrt { n } / 1 6 . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 661, + 839, + 767 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Similarly, ", + "bbox": [ + 173, + 768, + 240, + 782 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/bc0943e4d670ca144524e5231af3a4a90370f07e688b3492ef3c239e2ee492d4.jpg", + "text": "$$\n\\begin{array} { r l } & { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 } \\\\ & { \\qquad \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 276, + 786, + 722, + 830 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "By Lemma 3.1, with probability at least $1 - \\delta$ , $\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant 1 + C r _ { k } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }$ . Combining this result with Lemma D.3 gives $\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } \\leqslant \\theta \\mathcal { O } ( \\sqrt { n \\log \\left( n / \\delta \\right) } ) \\leqslant \\epsilon \\sqrt { n } / 8$ . Plugging the above estimates into (B.4) gives ", + "bbox": [ + 173, + 834, + 826, + 882 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/0a6b9784a608f97249de05f6d09bac2302a17889aafa041d0472362b42bf2ec2.jpg", + "text": "$$\nn ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon .\n$$", + "text_format": "latex", + "bbox": [ + 282, + 886, + 715, + 906 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Applying union bounds completes the proof. ", + "bbox": [ + 174, + 909, + 467, + 924 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.2 PROOF OF THE THEOREM 3.4 ", + "text_level": 1, + "bbox": [ + 176, + 103, + 421, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof of the Theorem 3.4. The idea of the proof is close to that of Proposition 5 in (Bietti and Mairal, 2019) where they consider $k \\gg d$ and we present a more general case including ş $k \\gg d$ and $d \\gg k$ . For any function $g : \\mathbb { S } ^ { d } \\mathbb { R }$ , by denoting $g _ { 0 } ( \\mathbf x ) = \\int _ { \\mathbb { S } ^ { d } } g ( \\mathbf { \\bar { y } } ) d \\tau _ { d } ( \\mathbf y )$ , it can be decomposed as ", + "bbox": [ + 174, + 130, + 823, + 174 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/3b9040066dffa2eae94da91164b6a92c93f503dd7cec4c4f37b87bbaf14d8f73.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle g ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } g _ { k } ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } \\sum _ { j = 1 } ^ { N ( d , k ) } \\int _ { { \\mathbb { S } } ^ { d } } Y _ { k j } ( { \\bf y } ) Y _ { k j } ( { \\bf x } ) g ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) } } \\\\ { { \\displaystyle ~ = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { { \\mathbb { S } } ^ { d } } g ( { \\bf y } ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 295, + 179, + 702, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where we project function $g$ to spherical harmonics in the second equality and apply the addition equation in the last equality. ", + "bbox": [ + 173, + 268, + 823, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For a positive-definite dot-product kernel $\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\mathbb { R }$ which has the form $\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) =$ $\\widehat { \\kappa } ( \\langle \\mathbf { x } , \\mathbf { \\bar { x } } ^ { \\prime } \\rangle )$ for $\\widehat { \\kappa } : [ - 1 , 1 ] \\ \\mathrm { ~ ~ } \\mathbb { R }$ , we can present a decomposition by (B.5) if we consider $g ( \\mathbf { x } ) =$ $\\phi ( \\langle \\mathbf { x } , \\mathbf { z } \\rangle )$ for $\\mathbf { z } \\in \\mathbb { S } ^ { \\bar { d } }$ and $\\bar { \\phi } : [ - 1 , 1 ] \\to \\mathbb { R }$ , ", + "bbox": [ + 176, + 297, + 823, + 340 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/51ac6887d63cfb5a9876061232d703931f01e89901461cb8b45eb821155905d5.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle \\kappa ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { \\mathbb { S } ^ { d } } \\widehat { \\kappa } ( \\langle { \\bf y } , { \\bf x } ^ { \\prime } \\rangle ) P _ { k } ( \\langle { \\bf y } , { \\bf x } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } \\ ~ } \\\\ { { \\displaystyle = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 254, + 344, + 741, + 429 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where we apply the Hecke-Funk formula and addition formula. By denoting ş $\\begin{array} { r l } { \\lambda _ { k } } & { { } = } \\end{array}$ $( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t$ and the addition formula, we have ", + "bbox": [ + 173, + 430, + 826, + 463 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/263c48b7608eb35c97c86c30186fa6fdc13a24a4e9d2c1a68805e929093c39f8.jpg", + "text": "$$\n\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } N ( d , k ) P _ { k } ( \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N ( p , k ) } Y _ { k , j } ( \\mathbf { x } ) Y _ { k , j } ( \\mathbf { x } ^ { \\prime } ) .\n$$", + "text_format": "latex", + "bbox": [ + 263, + 469, + 736, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This formula (B.6) is the Mercer decomposition for the kernel function $\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )$ and $\\mu _ { k }$ is exactly the eigenvalue of the integral operator $L _ { K }$ on $L _ { 2 } ( \\mathbb { S } ^ { d } )$ defined by ", + "bbox": [ + 173, + 518, + 823, + 547 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/41d6e45b4c87b313148a091f568503bd0d15e892dc6ebb4339bc7fd9bd8e4459.jpg", + "text": "$$\nL _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { 2 } ( \\mathbb { S } ^ { d } ) .\n$$", + "text_format": "latex", + "bbox": [ + 326, + 551, + 669, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "By using same technique as $\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )$ , we can derive a similar expression for $\\begin{array} { r l } { \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) } & { { } = } \\end{array}$ max $\\{ \\langle \\mathbf { w } , \\mathbf { x } \\rangle , 0 \\}$ and $\\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\mathbb { 1 } \\{ \\langle { \\bf w } , { \\bf x } \\rangle > 0 \\}$ , since they are essentially dot-product function on $L _ { 2 } ( \\mathbb { S } ^ { d } )$ . We deliver the expression below without presenting proofs. ", + "bbox": [ + 173, + 595, + 825, + 640 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/cc1a0b4f9a5e173b51af87672e7a5a208b1277bbeea1854cd0901795af785898.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle \\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\\\ { { \\displaystyle \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 356, + 643, + 640, + 727 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where $\\beta _ { 1 , k } = \\left( \\omega _ { d - 1 } / \\omega _ { d } \\right) \\int _ { - 1 } ^ { 1 } \\sigma ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t$ and $\\beta _ { 2 , k } = ( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\sigma ^ { \\prime } ( t ) P _ { k } ( t ) ( 1 -$ $t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t$ . We add more comments on the values of $\\beta _ { 1 , k }$ and $\\beta _ { 2 , k }$ . It has been pointed out in Bach (2017) that when $k > \\alpha$ and when $k$ and $\\alpha$ have same parity, we have $\\beta _ { \\alpha + 1 , k } = 0$ . This is because the Legendre polynomial $P _ { k } ( t )$ is orthogonal to any other polynomials of degree less than $k$ with respect to the density function $p ( t ) = ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 }$ . Then we clearly know that $\\beta _ { 1 , k } = 0$ for $k = 2 j$ and $\\beta _ { 2 , k } = 0$ for $k = 2 j + 1$ with $j \\in \\mathbb { N } ^ { + }$ . ", + "bbox": [ + 171, + 731, + 826, + 825 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For two kernel function defined in (2.2), we have ", + "bbox": [ + 174, + 830, + 495, + 844 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/95d2fe547af7b1da33eeceedf59749a8a73300ef5b14f44234e9e71094039a3d.jpg", + "text": "$$\n\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad = \\displaystyle \\int _ { \\mathbb { S } ^ { d } } \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 297, + 847, + 700, + 922 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The first equality holds because $\\sigma ^ { \\prime }$ is 0-homogeneous function and the second equality is true since the normalized direction of a multivariate Gaussian random variable satisfies uniform distribution on the unit sphere. Similarly we can derive ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/bfa930864378680abdd0a8715fda31c8c82605c1ade23e00d05c55a4f7a9a1a6.jpg", + "text": "$$\n\\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = ( d + 1 ) \\int _ { \\mathbb { S } ^ { d } } \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) .\n$$", + "text_format": "latex", + "bbox": [ + 325, + 148, + 671, + 183 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "By combing (A.2), (B.7), (B.8), (B.9) and (B.10), we can get ", + "bbox": [ + 173, + 185, + 575, + 202 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/f96ce4a551ce7c988dbaa8b4eb08c8f195e2044e850349e9ac27e025dcb5722c.jpg", + "text": "$$\n\\kappa _ { 1 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) ,\n$$", + "text_format": "latex", + "bbox": [ + 362, + 223, + 635, + 265 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "and ", + "bbox": [ + 173, + 268, + 202, + 282 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/c83b470578b34839f6ba7c569b2474967481e20c3a2b30522f8268939034e564.jpg", + "text": "$$\n\\kappa _ { 2 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = ( d + 1 ) \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) .\n$$", + "text_format": "latex", + "bbox": [ + 336, + 284, + 661, + 324 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Comparing (B.6), (B.11) and (B.12), we can easily show that ", + "bbox": [ + 173, + 327, + 575, + 343 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/18b01ecfed90ecccae99f94fd9765d2330aaa4e0f4a23546a3dc4ef531c7b042.jpg", + "text": "$$\n\\mu _ { 1 , k } = \\beta _ { 1 , k } ^ { 2 } \\mathrm { a n d } \\mu _ { 2 , k } = ( d + 1 ) \\beta _ { 2 , k } ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 372, + 345, + 625, + 366 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In Bach (2017), explicit expressions for $\\beta _ { 1 , k }$ and $\\beta _ { 2 , k }$ for $k \\geqslant \\alpha + 1$ are presented by ", + "bbox": [ + 171, + 367, + 735, + 383 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/30a45a792f4acf2923772703da93a5fe6febd0faa1f9d175c8c87f58d1b8fd28.jpg", + "text": "$$\n\\beta _ { \\alpha + 1 , k } = \\frac { d - 1 } { 2 \\pi } \\frac { \\alpha ! ( - 1 ) ^ { ( k - 1 - \\alpha ) / 2 } } { 2 ^ { k } } \\frac { \\Gamma ( d / 2 ) \\Gamma ( k - \\alpha ) } { \\Gamma ( \\frac { k - \\alpha + 1 } { 2 } ) \\Gamma ( \\frac { k + d + \\alpha + 1 } { 2 } ) } .\n$$", + "text_format": "latex", + "bbox": [ + 302, + 387, + 696, + 425 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "By Stirling formula $\\Gamma ( x ) \\approx x ^ { x - 1 / 2 } e ^ { - x } \\sqrt { 2 \\pi }$ , we have following expression of $\\beta _ { \\alpha + 1 , k }$ for $k \\geqslant \\alpha + 1$ $\\beta _ { \\alpha + 1 , k } = C ( \\alpha ) \\frac { ( d - 1 ) d ^ { \\frac { d - 1 } { 2 } } ( k - \\alpha ) ^ { k - \\alpha - \\frac 1 2 } } { ( k - \\alpha + 1 ) ^ { \\frac { k - \\alpha } { 2 } } ( k + d + \\alpha + 1 ) ^ { \\frac { k + d + \\alpha } { 2 } } } = \\Omega \\left( d ^ { \\frac { d + 1 } { 2 } } k ^ { \\frac { k - \\alpha - 1 } { 2 } } ( k + d ) ^ { \\frac { - k - d - \\alpha } { 2 } } \\right)$ where $\\begin{array} { r l r l r l r l r l } { C ( \\alpha ) } & { { } = } & { { \\frac { \\sqrt { 2 } \\alpha ! } { 2 \\pi } } \\exp \\{ \\alpha \\ + \\ 1 \\} . } & { } & { \\mathrm { A l s o } \\ \\beta _ { \\alpha + 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 2 } { 2 } } \\right) } } , } & { \\beta _ { 1 , 1 } \\beta _ { 1 , 2 } \\gamma _ { 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 1 } { 2 } } \\right) } } } \\end{array}$ $\\textstyle { \\frac { d - 1 } { 2 d \\pi } }$ and $\\begin{array} { r l r } { \\beta _ { 2 , 1 } } & { { } = } & { \\frac { d - 1 } { 4 \\pi d } \\frac { \\Gamma \\left( \\frac { 1 } { 2 } \\right) \\Gamma \\left( \\frac { d + 2 } { 2 } \\right) } { \\Gamma \\left( \\frac { d + 3 } { 2 } \\right) } } \\end{array}$ . Thus combine (B.13) we know that µα\\`1,k “ $\\Omega \\left( d ^ { d + 1 + \\alpha } k ^ { k - \\alpha - 1 } ( k + d ) ^ { - k - \\bar { d } - \\bar { \\alpha } } \\right)$ . ", + "bbox": [ + 176, + 430, + 823, + 446 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 450, + 826, + 568 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "By considering (A.3) and (B.6), we have ", + "bbox": [ + 176, + 566, + 441, + 580 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/1f84e94fdc943bd8f863779f3436508288c49a6811b49c5dfe282c367d942950.jpg", + "text": "$$\n\\begin{array} { c } { { \\mu _ { 0 } = \\mu _ { 1 , 1 } + 2 \\mu _ { 2 , 0 } , } } \\\\ { { { } } } \\\\ { { \\mu _ { k ^ { \\prime } } = 0 , k ^ { \\prime } = 2 j + 1 , j \\in \\mathbb { N } ^ { + } , } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 390, + 584, + 606, + 631 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "and ", + "bbox": [ + 173, + 635, + 202, + 648 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/1ef685e4e6ee1b9ee7f631d72757de9efb8356aaa395fec6722b1f2ea2d4274a.jpg", + "text": "$$\n\\mu _ { k } = \\frac { k } { 2 k + d - 1 } \\mu _ { 1 , k - 1 } + \\frac { k + d - 1 } { 2 k + d - 1 } \\mu _ { 1 , k + 1 } + 2 \\mu _ { 2 , k } ,\n$$", + "text_format": "latex", + "bbox": [ + 313, + 650, + 681, + 681 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "for $k \\geqslant 1$ and $k \\neq k ^ { \\prime }$ . From the discussion above, we thus know exactly that for $k \\geqslant 1$ ", + "bbox": [ + 173, + 685, + 743, + 700 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/a57494a710d77e73676c55d20bfe3d5c5c4d0c95af0797120095f2b993daaa6b.jpg", + "text": "$$\n\\mu _ { k } = \\Omega \\left( \\operatorname * { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 191, + 704, + 805, + 724 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This finishes the proof. ", + "bbox": [ + 174, + 727, + 326, + 742 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B.3 PROOF OF THE COROLLARIES 3.7 AND 3.8 ", + "text_level": 1, + "bbox": [ + 174, + 756, + 514, + 772 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Proof of the Corollaries 3.7 and 3.8. We only need to bound $| \\phi _ { j } ( \\mathbf { x } ) |$ for $j \\in [ r _ { k } ]$ to finish the proof. Since now we assume input data follows uniform distribution on the unit sphere $\\mathbb { S } ^ { d }$ , $\\phi _ { j } ( \\mathbf { x } )$ would be spherical harmonics of order at most $k$ for $j \\in [ r _ { k } ]$ . For any spherical harmonics $Y _ { k }$ of order $k$ and any point on $\\mathbb { S } ^ { d }$ , we have an upper bound (Proposition 4.16 in Frye and Efthimiou (2012)) ", + "bbox": [ + 173, + 784, + 825, + 842 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/f161691bd018cc5066f6944bde6803ca908681fcf7dd9913f8fd6946acb0ef11.jpg", + "text": "$$\n| Y _ { k } ( { \\bf x } ) | \\leqslant \\left( N ( d , k ) \\int _ { { \\mathbb S } ^ { d } } Y _ { k } ^ { 2 } ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) \\right) ^ { \\frac { 1 } { 2 } } .\n$$", + "text_format": "latex", + "bbox": [ + 354, + 845, + 640, + 885 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "kno. For $| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant \\sqrt { N ( d , k ) }$ $k \\gg d$ $\\begin{array} { r } { N ( d , k ) \\ = \\ \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\ = \\ } \\end{array}$ $\\mathcal { O } ( k ^ { d - 1 } )$ $d \\gg k$ $\\begin{array} { r } { N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } = \\mathcal { O } ( d ^ { k } ) } \\end{array}$ ", + "bbox": [ + 174, + 888, + 831, + 928 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C PROOF OF TECHNICAL LEMMAS ", + "text_level": 1, + "bbox": [ + 173, + 102, + 477, + 118 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C.1 PROOF OF LEMMA B.2 ", + "text_level": 1, + "bbox": [ + 174, + 132, + 377, + 147 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proof of Lemma B.2. The first inequality directly follows by equation (44) in $\\mathtt { S u }$ and Yang (2019). To prove the second bound, we write a $\\begin{array} { r } { \\dot { \\mathbf { V } _ { r _ { k } } } = \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { \\bar { A } } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\mathbf { \\bar { B } } } \\end{array}$ , where a $\\mathbf { A } \\in \\mathbb { R } ^ { r _ { k } \\times r _ { k } }$ , $\\mathbf { B } \\in \\mathbb { R } ^ { ( n - r _ { k } ) \\times r _ { k } }$ . Let $\\xi _ { 1 } = C ^ { \\prime } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } \\cdot \\sqrt { \\log ( 1 / \\delta ) / n }$ , $\\xi _ { 2 } = C ^ { \\prime \\prime \\prime } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }$ be the bounds given in the first inequality and Lemma 3.1. By the first inequality, we have ", + "bbox": [ + 173, + 159, + 825, + 223 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/238383ca1bab2f25bc29d9aa6e8654061b7546a13c93a6f85c13dc9ed9e90fc5.jpg", + "text": "$$\n\\| \\mathbf { B } \\| _ { F } = \\| \\mathbf { B } ^ { \\top } \\| _ { F } = \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\| _ { F } \\leqslant \\xi _ { 1 } .\n$$", + "text_format": "latex", + "bbox": [ + 370, + 229, + 625, + 251 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Moreover, since $\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } = \\mathbf { A } ^ { \\top } \\mathbf { A } + \\mathbf { B } ^ { \\top } \\mathbf { B }$ , by Lemma 3.1 we have ", + "bbox": [ + 173, + 256, + 619, + 275 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/4990612d02a94e3bcd38b1128dbcce7b646f16a699c39f5201cdcfc2ffabc926.jpg", + "text": "$$\n\\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } = \\| \\mathbf { A } ^ { \\top } \\mathbf { A } - \\mathbf { I } \\| _ { 2 } \\leqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { 2 } + \\| \\mathbf { B } ^ { \\top } \\mathbf { B } \\| _ { 2 } \\leqslant r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 251, + 281, + 745, + 303 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Therefore ", + "bbox": [ + 173, + 308, + 241, + 321 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/55a4cfe7456cd67732c11b9ed4790dbf305020c4c1ca74c6d38c790ded067c03.jpg", + "text": "$$\n\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\big ( \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\big ) \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad = \\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 327, + 818, + 430 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Plugging in the definition of $\\xi _ { 1 }$ and $\\xi _ { 2 }$ completes the proof. ", + "bbox": [ + 173, + 436, + 562, + 452 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C.2 PROOF OF LEMMA B.3 ", + "text_level": 1, + "bbox": [ + 174, + 468, + 375, + 483 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proof of Lemma B.3. The gradient descent update formula gives ", + "bbox": [ + 174, + 496, + 598, + 511 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/231555ff0f8d836ac406bb6f499944f0274e7bdf42eed188f8b8d7399f8e1504.jpg", + "text": "$$\n\\mathbf { W } ^ { ( t + 1 ) } = \\mathbf { W } ^ { ( t ) } + \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } \\bigl ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\bigr ) \\cdot \\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) .\n$$", + "text_format": "latex", + "bbox": [ + 289, + 517, + 709, + 558 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "For any $j \\in [ n ]$ , subtracting $\\mathbf { W } ^ { ( t ) }$ and applying inner product with $\\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } \\left( \\mathbf { x } _ { j } \\right)$ on both sides gives ", + "bbox": [ + 171, + 565, + 823, + 597 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/ad4be42d9c5e3936277a36f4dd0eca38247a415f59209be45b60b8c0eb6b8374.jpg", + "text": "$$\n\\theta \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle = \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\hat { y } _ { i } ^ { ( t ) } ) \\cdot \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle .\n$$", + "text_format": "latex", + "bbox": [ + 187, + 602, + 807, + 642 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Further rearranging terms then gives ", + "bbox": [ + 173, + 648, + 415, + 664 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/18d477a76c43e2cf16f7a081c1059891ce6315d76b3de2b8ef8fc8a0994d776d.jpg", + "text": "$$\ny _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t + 1 ) } ) _ { j } = y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t ) } ) _ { j } - \\frac { 2 \\eta m \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot \\mathbf { K } _ { i , j } ^ { \\infty } + I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t } ,\n$$", + "text_format": "latex", + "bbox": [ + 187, + 671, + 807, + 710 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 174, + 733, + 217, + 746 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/ffa1a81375895e365daa3b830eea52322209bf771b4c6f7e2475f3b4e664be88.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle I _ { 1 , j , t } = - \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot [ \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle - m \\mathbf { K } _ { i , j } ^ { ( 0 ) } ] } , } \\\\ { { \\displaystyle I _ { 2 , j , t } = - \\frac { 2 \\eta \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot ( \\mathbf { K } _ { i , j } ^ { ( 0 ) } - \\mathbf { K } _ { i , j } ^ { \\infty } ) , } } \\\\ { { \\displaystyle I _ { 3 , j , t } = - \\theta \\cdot [ f _ { \\mathbf { W } ^ { ( t + 1 ) } } ( \\mathbf { x } _ { j } ) - f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle ] } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 217, + 751, + 779, + 856 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "For $I _ { 1 , j , t }$ , by Lemma D.6, we have ", + "bbox": [ + 173, + 859, + 405, + 875 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/977341470ef42f16c69e2ff113183c02981a6ffeb527f185cc19c64faf950623.jpg", + "text": "$$\n| I _ { 1 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n }\n$$", + "text_format": "latex", + "bbox": [ + 215, + 882, + 781, + 921 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "with probability at least $1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ . For $I _ { 2 , j , t }$ , by Bernstein inequality and union bound, with probability at least $1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp ( - \\Omega ( m \\omega ^ { 2 / 3 } ) )$ , we have ", + "bbox": [ + 171, + 102, + 825, + 135 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/cfd0fda51ed2619710ce18718f21595620bcc96ab6b0b4a38a35057b8b0fd9ff.jpg", + "text": "$$\n\\big | \\mathbf { K } _ { i , j } ^ { \\infty } - \\mathbf { K } _ { i , j } ^ { ( 0 ) } \\big | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } )\n$$", + "text_format": "latex", + "bbox": [ + 411, + 141, + 584, + 165 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "for all $i , j \\in [ n ]$ . Therefore ", + "bbox": [ + 173, + 171, + 352, + 188 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/395d49404d7877bb95aeff7283b15ea83f17598ca772e7d629aab519fd328916.jpg", + "text": "$$\n| I _ { 2 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } .\n$$", + "text_format": "latex", + "bbox": [ + 212, + 194, + 784, + 234 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "For $I _ { 3 , j , t }$ , we have ", + "bbox": [ + 173, + 241, + 297, + 256 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f2d3f819faf36e8f88e3144eb94e09776542b236d5d876b9fe406ab7264350f1.jpg", + "text": "$$\n\\begin{array} { r l r } { { I _ { 3 , j , t } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { ( t + 1 ) } - \\mathbf { W } _ { 1 } ^ { ( t ) } \\| _ { 2 } } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\cdot \\theta \\cdot \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\| _ { 2 } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\cfrac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 258, + 262, + 738, + 390 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where the first inequality follows by Lemmas D.2, the second inequality is obtained from (C.1), and the third inequality follows by Lemma D.4. Setting the $j$ -th entry of $\\mathbf { e } ^ { ( t ) }$ as $I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t }$ and writing (C.2) into matrix form completes the proof. □ ", + "bbox": [ + 176, + 393, + 825, + 439 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "C.3 PROOF OF LEMMA B.4 ", + "text_level": 1, + "bbox": [ + 174, + 454, + 375, + 469 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Proof of Lemma B.4. Denote $\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ) }$ , $t \\in T$ . Then we have ", + "bbox": [ + 173, + 483, + 620, + 501 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/a867be7b2013f3586a54ea83ac85ede4de733ea548f4eb08bdf76beec60f6047.jpg", + "text": "$$\n\\begin{array} { r l } & { \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } [ { \\mathbf { I } } - ( \\eta m \\theta ^ { 2 } / n ) { \\mathbf { K } } ^ { \\infty } ] { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 220, + 505, + 777, + 553 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where the first inequality follows by Lemma B.3, and the second inequality follows by Lemma D.5. Therefore we have ", + "bbox": [ + 173, + 555, + 823, + 584 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/ba5485a85db3b64cd31cc1ac92622c425118a1ab47ffda441e72f7c50635a617.jpg", + "text": "$$\n\\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 258, + 588, + 738, + 611 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "for $t ^ { \\prime } = 0 , \\ldots , t$ . This completes the proof of (B.1). Similarly, we have ", + "bbox": [ + 173, + 617, + 640, + 632 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/731d3412adbd36dfe4e2b6a9ed64333eb07ee5ee4307a5c354064b826a5d2e7a.jpg", + "text": "$$\n\\begin{array} { r l } { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\widehat { \\lambda } _ { r _ { k } } ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot ( \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } ) } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } & { } \\\\ + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 637, + 825, + 751 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "for $t ^ { \\prime } = 0 , \\ldots , t - 1$ , where the third inequality is by Lemma B.1 and the assumption that $\\omega \\ \\leqslant$ $\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )$ , $n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )$ , and the fourth inequality is by (B.1). Therefore we have ", + "bbox": [ + 173, + 755, + 826, + 787 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/735e7e055b5ef8b4ae398e44d36ba8c817c1c2532ffe58bf80ef3a9706813ee7.jpg", + "text": "$$\n\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } = ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & \\phantom { \\hat { \\mathbf { V } } } + \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 794, + 839, + 930 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "This completes the proof of (B.2). Finally, for (B.3), by assumption we have $\\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( 1 )$ . Therefore ", + "bbox": [ + 171, + 102, + 823, + 132 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/68dc0a55babfb44f99376c17aa93747eca9f9e8d91ad988b871323165d10fe03.jpg", + "text": "$$\n\\begin{array} { r l } { \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\hat { \\lambda } _ { \\mathbf { r } _ { k } } ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( 1 ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { \\mathbf { r } _ { k } } / 2 ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 130, + 838, + 244 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "for $t ^ { \\prime } = 0 , \\ldots , t - 1$ , where the third inequality is by Lemma B.1 and the assumption that $\\omega \\ \\leqslant$ $\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )$ , and the fourth inequality follows by (B.1). Therefore we have ", + "bbox": [ + 173, + 244, + 823, + 276 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "} u p t 1 q } 2 ď $\\mathscr { O } ( \\sqrt { n } ) \\cdot ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t ^ { \\prime } } + \\widetilde { \\mathscr { O } } ( ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } ) \\cdot \\Vert ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( 0 ) } \\Vert _ { 2 } + \\lambda _ { r _ { k } } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathscr { O } } ( 1 + \\omega \\sqrt { m } ) .$ This finishes the proof. ", + "bbox": [ + 173, + 280, + 908, + 319 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "D AUXILIARY LEMMAS ", + "text_level": 1, + "bbox": [ + 174, + 335, + 387, + 352 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In this section we list several auxiliary lemmas on the properties of over-parameterized neural networks we need in our proof of Theorem 3.2. These results are mostly summarized from Allen-Zhu et al. (2018) and Cao and Gu (2019a). ", + "bbox": [ + 173, + 364, + 825, + 407 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "D.1 AUXILIARY LEMMAS ", + "text_level": 1, + "bbox": [ + 174, + 420, + 367, + 434 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Denote ", + "bbox": [ + 173, + 446, + 223, + 460 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/ff0e5791ea25ef3963ffa126e032454c12dce142473a05c97bd0ff79716c28f9.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbf { D } _ { i } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) , } \\\\ & { \\mathbf { D } _ { i } ^ { ( 0 ) } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 295, + 462, + 700, + 506 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Lemma D.1 (Allen-Zhu et al. (2018)). If $\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )$ , then with probability at least $1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ , ", + "bbox": [ + 171, + 508, + 821, + 541 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/a0fcca48470c6919a30c09b5d5f14a5b98e2173732ee39612c8e06ed71eaccae.jpg", + "text": "$$\n\\| \\mathbf { D } _ { i } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { 0 } \\leqslant \\mathcal { O } ( \\omega ^ { 2 / 3 } m )\n$$", + "text_format": "latex", + "bbox": [ + 403, + 542, + 593, + 565 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "for all $\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ , $i \\in [ n ]$ ", + "bbox": [ + 173, + 569, + 392, + 585 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Lemma D.2 (Cao and Gu (2019a)). There exists an absolute constant $\\kappa$ such that, with probability at least $1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ over the randomness of $\\mathbf { W } ^ { ( 1 ) }$ , for all $i \\in [ n ]$ and $\\mathbf { W } , \\mathbf { W } ^ { \\prime } \\in$ $B ( \\mathbf { W } ^ { ( \\mathbf { \\theta } ) } , \\omega )$ with $\\omega \\leqslant \\kappa [ \\log ( m ) ] ^ { - 3 / 2 }$ , it holds uniformly that ", + "bbox": [ + 173, + 588, + 825, + 633 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/3f1ac05583175ca408c01d620917fcde1ce626c7b4f02ef194ebe5ea1d4c11cd.jpg", + "text": "$$\n| f _ { \\mathbf { W } ^ { \\prime } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\mathbf { W } ^ { \\prime } - \\mathbf { W } \\rangle | \\leqslant { \\mathcal { O } } \\Bigl ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } \\Bigr ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { \\prime } - \\mathbf { W } _ { 1 } \\| _ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 196, + 636, + 800, + 664 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Lemma D.3 (Cao and Gu (2019a)). For any $\\delta > 0$ , if $m \\geqslant C \\log ( n / \\delta )$ for a large enough absolute constant $C$ , then with probability at least $1 - \\delta$ , $| f _ { \\mathbf { W } ^ { ( 0 ) } } ( { \\pmb x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } )$ for all $i \\in [ n ]$ . ", + "bbox": [ + 174, + 666, + 820, + 699 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Lemma D.4 (Cao and Gu (2019a)). There exists an absolute constant $C$ such that, with probability at least $1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ , for all $i \\in [ n ] , l \\in [ L ]$ and $\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ with $\\omega \\ \\leqslant$ $C [ \\log ( m ) ] ^ { - 3 }$ , it holds uniformly that ", + "bbox": [ + 174, + 700, + 820, + 744 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/3ea31864bd3ef411c12372d9e785cce422d2bd9da181b1ef7ea1d4762698acf8.jpg", + "text": "$$\n\\| \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\sqrt { m } ) .\n$$", + "text_format": "latex", + "bbox": [ + 400, + 746, + 596, + 765 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "The following lemma provides a uniform bound of the neural network function value over $B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ . ", + "bbox": [ + 174, + 773, + 823, + 804 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Lemma D.5. Suppose that $m \\geqslant \\Omega ( \\omega ^ { - 2 / 3 } \\log ( n / \\delta ) )$ and $\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 } )$ . Then with probability at least $1 - \\delta$ , $| f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } )$ for all $\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) \\ i \\in [ n ]$ . ", + "bbox": [ + 173, + 808, + 823, + 840 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Lemma D.6. If $\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )$ , then with probability at least $1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]$ ", + "bbox": [ + 169, + 843, + 823, + 861 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/b3a03af7ab37fba24e0f0a2b59e5305a6930a79ac7298eb27f92f385cdcffb08.jpg", + "text": "$$\n\\begin{array} { r l } & { \\| \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } ) , } \\\\ & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } m ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 233, + 863, + 761, + 905 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "for all $\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )$ and $i \\in [ n ]$ ", + "bbox": [ + 173, + 907, + 415, + 924 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "D.2 PROOFS OF LEMMAS D.5 AND D.6 ", + "bbox": [ + 176, + 103, + 459, + 118 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Proof of Lemma D.5. By Lemmas D.2 and D.4, we have ", + "bbox": [ + 173, + 128, + 547, + 146 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/9b0b16369284ee8c4572991ce051730bee7663841d031fa484660bf7dc5c2903.jpg", + "text": "$$\n\\begin{array} { r l } & { f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { F } + \\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 2 } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\| _ { F } } \\\\ & { \\qquad + \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } ) \\cdot \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( \\omega \\sqrt { m } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 150, + 838, + 215 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "where the last inequality is by the assumption $\\omega \\leqslant [ \\log ( m ) ] ^ { - 3 }$ . Applying triangle inequality and Lemma D.3 then gives ", + "bbox": [ + 173, + 219, + 825, + 250 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/ba830733a52927b5172128e16ae55612568311868beaa93938e576110a0bdc8c.jpg", + "text": "$$\n\\begin{array} { r l } & { | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant | f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | + | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } ) + \\mathcal { O } ( \\omega \\sqrt { m } ) } \\\\ & { \\qquad = \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 228, + 253, + 767, + 297 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "This completes the proof. ", + "bbox": [ + 174, + 301, + 343, + 316 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Proof of Lemma D.6. By direct calculation, we have ", + "bbox": [ + 173, + 330, + 519, + 347 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/32652103ff49fa8f8ca461812435f446cf7cd55d9e3e8559eb4e66cc5c9939dc.jpg", + "text": "$$\n\\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } .\n$$", + "text_format": "latex", + "bbox": [ + 250, + 352, + 746, + 373 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Therefore we have ", + "bbox": [ + 174, + 377, + 299, + 392 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/1167691bc5adceb2bf869d56c3d675bc732d2cec630f37b53c09c75ff88b73a7.jpg", + "text": "$$\n\\begin{array} { r l } { | \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | | _ { F } = \\sqrt { m } \\cdot \\| \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { \\leqslant \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } + \\sqrt { m } \\cdot \\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } } & { } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 393, + 826, + 486 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "By Lemma 7.4 in Allen-Zhu et al. (2018) and Lemma D.1, with probability at least $1 - n \\cdot$ $\\exp [ - \\Omega ( m ) ]$ , $\\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )$ for all $i \\in [ n ]$ . Moreover, clearly $\\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } \\leqslant \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } \\| _ { F } \\leqslant \\omega$ . Therefore ", + "bbox": [ + 173, + 487, + 825, + 539 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/82517f5e7b596fb3aed17f5a1d9746417259ebcfeaab15f43c40b34fb6100eb8.jpg", + "text": "$$\n\\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )\n$$", + "text_format": "latex", + "bbox": [ + 325, + 544, + 671, + 564 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "for all $i \\in [ n ]$ . This proves the bound for the first layer gradients. For the second layer gradients, we have ", + "bbox": [ + 173, + 569, + 823, + 598 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/56cf3fc46e0b3d361f160b309cf351e307ff1bca25947efa4ed56b26b269f77e.jpg", + "text": "$$\n\\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) ] ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) ] ^ { \\top }\n$$", + "text_format": "latex", + "bbox": [ + 245, + 599, + 753, + 621 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "It therefore follows by the 1-Lipschitz continuity of $\\sigma ( \\cdot )$ that ", + "bbox": [ + 173, + 626, + 575, + 641 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/3ddd54a319073f6f2b0f103a2fd96e83b20c05e896855a3d94a0990dedab790a.jpg", + "text": "$$\n\\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant { \\sqrt { m } } \\cdot \\| \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } \\| _ { F } \\leqslant \\omega { \\sqrt { m } } \\leqslant \\omega ^ { 1 / 3 } { \\sqrt { m } } .\n$$", + "text_format": "latex", + "bbox": [ + 207, + 646, + 790, + 667 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "This completes the proof of the first inequality. ", + "bbox": [ + 174, + 672, + 480, + 688 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "The second inequality directly follows by triangle inequality and Lemma D.4: ", + "bbox": [ + 174, + 694, + 684, + 709 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/17041c13493a2ab7b5f12096dece9a8b00cd58c4a276014fbb609c238691cec8.jpg", + "text": "$$\n\\begin{array} { r l } & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - m \\mathbf { K } ^ { ( 0 ) } | \\leqslant | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\leqslant { \\mathcal { O } } ( \\omega ^ { 1 / 3 } m ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 713, + 815, + 775 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "This finishes the proof. ", + "bbox": [ + 174, + 777, + 326, + 792 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E ADDITIONAL EXPERIMENTAL RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 809, + 537, + 827 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E.1 ESTIMATING THE PROJECTION LENGTH IN FUNCTION SPACE ", + "text_level": 1, + "bbox": [ + 174, + 840, + 632, + 856 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "As mentioned in Section 4.1, when using freshly sampled points, we are actually estimating the projection length of residual function $r ( \\mathbf { x } ) = f ^ { * } ( \\mathbf { x } ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )$ onto the given Gegenbauer polynomial $P _ { k } ( { \\bf x } )$ . Here we present in Figure 3 a comparison between the projection length using training data and that using test data. An interesting phenomenon is that the network generalizes well on the lower-order Gegenbauer polynomial and along the highest-order Gegenbauer polynomial the network suffers overfitting. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/ede46709d91cb04a929d72bd1ecb66385cb7ce3faf11dd60ee6d67e97d71b25f.jpg", + "image_caption": [ + "Figure 3: Convergence curve for projection length onto vectors (determined by training data) and functions (estimated by test data). We can see that for low-order Gegenbauer polynomials, the network learns the function while for the high-order Gegenbauer polynomial, the network overfits the training data. " + ], + "image_footnote": [], + "bbox": [ + 181, + 145, + 813, + 559 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "E.2 NEAR-LINEAR CONVERGENCE BEHAVIOR ", + "text_level": 1, + "bbox": [ + 173, + 647, + 503, + 661 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "In this subsection, we present the same curve shown in Section 4.1 in logarithmic scale instead of linear scale. As shown in Figure 4 we can see that the convergence of different projection length is close to linear convergence, which is well-aligned with our theorem. Note that we performed a moving average of range 20 on these curves to avoid the heavy oscillation especially in late stage. ", + "bbox": [ + 173, + 674, + 825, + 731 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/d4b04e61fd87b82bba68d9aeb4f350e54f3f281cc4e3ecdbada8bf74711541a3.jpg", + "image_caption": [ + "Figure 4: Log-scale convergence curve for projection length onto different component. (a) shows the curve when the target function have different component with the same scale. (b) shows the curve when the higner-order components have larger scale. Both exhibit nearly linear convergence especially at late stage. " + ], + "image_footnote": [], + "bbox": [ + 181, + 378, + 813, + 579 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/parse/train/BJg15lrKvS/BJg15lrKvS_middle.json b/parse/train/BJg15lrKvS/BJg15lrKvS_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..f1b3d913ba5172d9f92705caa91319a6f599d8f3 --- /dev/null +++ b/parse/train/BJg15lrKvS/BJg15lrKvS_middle.json @@ -0,0 +1,62430 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 506, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 506, + 98 + ], + "score": 1.0, + "content": "TOWARDS UNDERSTANDING THE SPECTRAL BIAS OF", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 233, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 233, + 118 + ], + "score": 1.0, + "content": "DEEP LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 209, + 468, + 374 + ], + "lines": [ + { + "bbox": [ + 141, + 209, + 469, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 209, + 469, + 222 + ], + "score": 1.0, + "content": "An intriguing phenomenon observed during training neural networks is the spec-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 221, + 469, + 233 + ], + "spans": [ + { + "bbox": [ + 141, + 221, + 469, + 233 + ], + "score": 1.0, + "content": "tral bias, where neural networks are biased towards learning less complex func-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "score": 1.0, + "content": "tions. The priority of learning functions with low complexity might be at the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "score": 1.0, + "content": "core of explaining generalization ability of neural network, and certain efforts", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "score": 1.0, + "content": "have been made to provide theoretical explanation for spectral bias. However,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "score": 1.0, + "content": "there is still no satisfying theoretical result justifying the underlying mechanism", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 275, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 470, + 288 + ], + "score": 1.0, + "content": "of spectral bias. In this paper, we give a comprehensive and rigorous explanation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 287, + 469, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 287, + 469, + 298 + ], + "score": 1.0, + "content": "for spectral bias and relate it with the neural tangent kernel function proposed in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 298, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 470, + 309 + ], + "score": 1.0, + "content": "recent work. We prove that the training process of neural networks can be de-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "score": 1.0, + "content": "composed along different directions defined by the eigenfunctions of the neural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 320, + 470, + 331 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 470, + 331 + ], + "score": 1.0, + "content": "tangent kernel, where each direction has its own convergence rate and the rate is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 469, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 469, + 342 + ], + "score": 1.0, + "content": "determined by the corresponding eigenvalue. We then provide a case study when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 342, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 342, + 470, + 353 + ], + "score": 1.0, + "content": "the input data is uniformly distributed over the unit sphere, and show that lower", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 352, + 470, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 470, + 365 + ], + "score": 1.0, + "content": "degree spherical harmonics are easier to be learned by over-parameterized neural", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 363, + 183, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 363, + 183, + 374 + ], + "score": 1.0, + "content": "networks.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 389, + 206, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 208, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 208, + 405 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Over-parameterized neural networks have achieved great success in many applications such as com-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "puter vision (He et al., 2016), natural language processing (Collobert and Weston, 2008) and speech", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "recognition (Hinton et al., 2012). It has been shown that over-parameterized neural networks can", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "fit complicated target function or even randomly labeled data (Zhang et al., 2016) and still exhibit", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "good generalization performance when trained with real labels. Intuitively, this is at odds with the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "traditional notion of generalization ability such as model complexity. In order to understand neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "network training, a line of work (Soudry et al., 2017; Gunasekar et al., 2018b;a) has made efforts in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "the perspective of “implicit bias”, which states that training algorithms for deep learning implicitly", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "pose an inductive bias onto the training process and lead to a solution with low complexity measured", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 511, + 358, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 358, + 523 + ], + "score": 1.0, + "content": "by certain norms in the parameter space of the neural network.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "Among many attempts to establish implicit bias, Rahaman et al. (2018) pointed out an intriguing", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "phenomenon called spectral bias, which says that during training, neural networks tend to learn", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "the components of lower complexity faster. The concept of spectral bias is appealing because this", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "may intuitively explain why over-parameterized neural networks can achieve a good generalization", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "performance without overfitting. During training, the networks fit the low complexity components", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "first and thus lie in the concept class of low complexity. Arguments like this may lead to rigorous", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 595, + 222, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 222, + 606 + ], + "score": 1.0, + "content": "guarantee for generalization.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Great efforts have been made in seek of explanations about the spectral bias. Rahaman et al. (2018)", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 622, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 633 + ], + "score": 1.0, + "content": "evaluated the Fourier spectrum of ReLU networks and empirically showed that the lower frequencies", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "are learned first; also lower frequencies are more robust to random perturbation. Andoni et al. (2014)", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "showed that for a sufficiently wide two-layer network, gradient descent with respect to the second", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "layer can learn any low degree bounded polynomial. Xu (2018) provided Fourier analysis to two-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "score": 1.0, + "content": "layer networks and showed similar empirical results on one-dimensional functions and real data.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "Nakkiran et al. (2019) used information theoretical approach to show that networks obtained by", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "stochastic gradient descent can be explained by a linear classifier during early training. All these", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "studies provide certain explanations about why neural networks exhibit spectral bias in real tasks.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "But explanations in the theoretical aspect, if any, are to some extent restricted. For example, the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 496, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 496, + 734 + ], + "score": 1.0, + "content": "popular Fourier analysis is usually done in the one-dimensional setting, and thus lacks generality.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 506, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 506, + 98 + ], + "score": 1.0, + "content": "TOWARDS UNDERSTANDING THE SPECTRAL BIAS OF", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 233, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 233, + 118 + ], + "score": 1.0, + "content": "DEEP LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 209, + 468, + 374 + ], + "lines": [ + { + "bbox": [ + 141, + 209, + 469, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 209, + 469, + 222 + ], + "score": 1.0, + "content": "An intriguing phenomenon observed during training neural networks is the spec-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 221, + 469, + 233 + ], + "spans": [ + { + "bbox": [ + 141, + 221, + 469, + 233 + ], + "score": 1.0, + "content": "tral bias, where neural networks are biased towards learning less complex func-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "score": 1.0, + "content": "tions. The priority of learning functions with low complexity might be at the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "score": 1.0, + "content": "core of explaining generalization ability of neural network, and certain efforts", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "score": 1.0, + "content": "have been made to provide theoretical explanation for spectral bias. However,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "score": 1.0, + "content": "there is still no satisfying theoretical result justifying the underlying mechanism", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 275, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 470, + 288 + ], + "score": 1.0, + "content": "of spectral bias. In this paper, we give a comprehensive and rigorous explanation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 287, + 469, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 287, + 469, + 298 + ], + "score": 1.0, + "content": "for spectral bias and relate it with the neural tangent kernel function proposed in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 298, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 470, + 309 + ], + "score": 1.0, + "content": "recent work. We prove that the training process of neural networks can be de-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "score": 1.0, + "content": "composed along different directions defined by the eigenfunctions of the neural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 320, + 470, + 331 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 470, + 331 + ], + "score": 1.0, + "content": "tangent kernel, where each direction has its own convergence rate and the rate is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 469, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 469, + 342 + ], + "score": 1.0, + "content": "determined by the corresponding eigenvalue. We then provide a case study when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 342, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 342, + 470, + 353 + ], + "score": 1.0, + "content": "the input data is uniformly distributed over the unit sphere, and show that lower", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 352, + 470, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 470, + 365 + ], + "score": 1.0, + "content": "degree spherical harmonics are easier to be learned by over-parameterized neural", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 363, + 183, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 363, + 183, + 374 + ], + "score": 1.0, + "content": "networks.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12, + "bbox_fs": [ + 141, + 209, + 470, + 374 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 389, + 206, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 208, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 208, + 405 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Over-parameterized neural networks have achieved great success in many applications such as com-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "puter vision (He et al., 2016), natural language processing (Collobert and Weston, 2008) and speech", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "recognition (Hinton et al., 2012). It has been shown that over-parameterized neural networks can", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "fit complicated target function or even randomly labeled data (Zhang et al., 2016) and still exhibit", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "good generalization performance when trained with real labels. Intuitively, this is at odds with the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "traditional notion of generalization ability such as model complexity. In order to understand neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "network training, a line of work (Soudry et al., 2017; Gunasekar et al., 2018b;a) has made efforts in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "the perspective of “implicit bias”, which states that training algorithms for deep learning implicitly", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "pose an inductive bias onto the training process and lead to a solution with low complexity measured", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 511, + 358, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 358, + 523 + ], + "score": 1.0, + "content": "by certain norms in the parameter space of the neural network.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 412, + 505, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "Among many attempts to establish implicit bias, Rahaman et al. (2018) pointed out an intriguing", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "phenomenon called spectral bias, which says that during training, neural networks tend to learn", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "the components of lower complexity faster. The concept of spectral bias is appealing because this", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "may intuitively explain why over-parameterized neural networks can achieve a good generalization", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "performance without overfitting. During training, the networks fit the low complexity components", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "first and thus lie in the concept class of low complexity. Arguments like this may lead to rigorous", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 595, + 222, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 222, + 606 + ], + "score": 1.0, + "content": "guarantee for generalization.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 527, + 505, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Great efforts have been made in seek of explanations about the spectral bias. Rahaman et al. (2018)", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 622, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 633 + ], + "score": 1.0, + "content": "evaluated the Fourier spectrum of ReLU networks and empirically showed that the lower frequencies", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "are learned first; also lower frequencies are more robust to random perturbation. Andoni et al. (2014)", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "showed that for a sufficiently wide two-layer network, gradient descent with respect to the second", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "layer can learn any low degree bounded polynomial. Xu (2018) provided Fourier analysis to two-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 504, + 677 + ], + "score": 1.0, + "content": "layer networks and showed similar empirical results on one-dimensional functions and real data.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "Nakkiran et al. (2019) used information theoretical approach to show that networks obtained by", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "stochastic gradient descent can be explained by a linear classifier during early training. All these", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "studies provide certain explanations about why neural networks exhibit spectral bias in real tasks.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "But explanations in the theoretical aspect, if any, are to some extent restricted. For example, the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 496, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 496, + 734 + ], + "score": 1.0, + "content": "popular Fourier analysis is usually done in the one-dimensional setting, and thus lacks generality.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 611, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Meanwhile, a recent line of work (Jacot et al., 2018; Du et al., 2018b; Li and Liang, 2018; Chizat", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 106 + ], + "score": 1.0, + "content": "and Bach, 2018) has shed light on new approaches to analyze neural networks. In particular, they", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 504, + 115 + ], + "score": 1.0, + "content": "show that under certain over-parameterized condition, the neural network trained by gradient de-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "scent behaves similar to the kernel regression predictor using the neural tangent kernel (NTK) (Jacot", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "score": 1.0, + "content": "et al., 2018). Du et al. (2018b) showed that the convergence is provably guaranteed under cer-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "score": 1.0, + "content": "tain over-parameterization conditions determined by the smallest eigenvalue of NTK. Arora et al.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "(2019b) further gave a finer characterization of error convergence based on the eigenvalues of NTK’s", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 457, + 172 + ], + "score": 1.0, + "content": "Gram matrix. Su and Yang (2019) improved the convergence guarantee in terms of the", + "type": "text" + }, + { + "bbox": [ + 457, + 160, + 464, + 169 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "-th largest", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 262, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 262, + 182 + ], + "score": 1.0, + "content": "eigenvalue for certain target functions.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Inspired by these works mentioned above, we can present a theoretical explanation for spectral bias.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "Under NTK regime, we establish a precise characterization for the training process of neural net-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "works. More specifically, we theoretically prove that over-parameterized neural networks’ training", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "process can be controlled by the eigenvalues of the integrating operator defined by the NTK. Un-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "der the specific case of uniform distribution on unit sphere, we give an exact calculation for these", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 243, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 254 + ], + "score": 1.0, + "content": "eigenvalues and show that the lower frequencies have larger eigenvalues, which thus leads to faster", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 436, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 436, + 265 + ], + "score": 1.0, + "content": "convergence. We also conduct experiments to corroborate the theory we establish.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 108, + 270, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 107, + 270, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 107, + 270, + 288, + 282 + ], + "score": 1.0, + "content": "Our contributions are highlighted as follows:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 291, + 506, + 465 + ], + "lines": [ + { + "bbox": [ + 107, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 107, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "1. We prove a generic theorem for arbitrary data distributions, which states that under certain sample", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 118, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 118, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "complexity and over-parameterization conditions, the error term’s convergence along different", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 119, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 119, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "directions actually relies on the corresponding eigenvalues. This theorem gives a more precise", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 118, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 118, + 324, + 279, + 336 + ], + "score": 1.0, + "content": "control on the regression residual than", + "type": "text" + }, + { + "bbox": [ + 279, + 324, + 291, + 334 + ], + "score": 0.48, + "content": "\\mathtt { S u }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "and Yang (2019), where the authors focused on the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 118, + 334, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 118, + 334, + 505, + 349 + ], + "score": 1.0, + "content": "case when the labeling function is close to the subspace spanned by the first few eigenfunctions.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 107, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "2. We present a more general result about the spectra of the neural tangent kernel. In particular, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 358, + 507, + 375 + ], + "spans": [ + { + "bbox": [ + 117, + 358, + 299, + 375 + ], + "score": 1.0, + "content": "show that the order of eigenvalues appears as", + "type": "text" + }, + { + "bbox": [ + 300, + 360, + 423, + 372 + ], + "score": 0.92, + "content": "\\mu _ { k } = \\Omega ( \\operatorname* { m a x } \\{ k ^ { - d - 1 } , \\check { d } ^ { - k + 1 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 358, + 507, + 375 + ], + "score": 1.0, + "content": ". Our result is better", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 118, + 370, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 118, + 370, + 183, + 386 + ], + "score": 1.0, + "content": "than the bound", + "type": "text" + }, + { + "bbox": [ + 183, + 371, + 225, + 384 + ], + "score": 0.93, + "content": "\\Omega ( k ^ { - d - \\overline { { 1 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 370, + 398, + 386 + ], + "score": 1.0, + "content": "derived in Bietti and Mairal (2019) when", + "type": "text" + }, + { + "bbox": [ + 398, + 372, + 426, + 382 + ], + "score": 0.88, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 370, + 506, + 386 + ], + "score": 1.0, + "content": ", which is clearly a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 118, + 382, + 209, + 397 + ], + "spans": [ + { + "bbox": [ + 118, + 382, + 209, + 397 + ], + "score": 1.0, + "content": "more practical setting.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "3. We establish a rigorous explanation for the spectral bias based on the aforementioned theoretical", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 118, + 409, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 118, + 409, + 506, + 422 + ], + "score": 1.0, + "content": "results without any specific assumptions on the target function. We show that the error terms", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 118, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 118, + 419, + 506, + 433 + ], + "score": 1.0, + "content": "from different frequencies are provably controlled by the eigenvalues of the NTK, and the lower-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 118, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 118, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "frequency components can be learned with less training examples and narrower networks with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 118, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 118, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "a faster convergence rate. As far as we know, this is the first attempt to give a comprehensive", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 118, + 453, + 307, + 465 + ], + "spans": [ + { + "bbox": [ + 118, + 453, + 307, + 465 + ], + "score": 1.0, + "content": "theory justifying the existence of spectral bias.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 109, + 474, + 260, + 485 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 262, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 262, + 486 + ], + "score": 1.0, + "content": "1.1 ADDITIONAL RELATED WORK", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 506, + 508 + ], + "score": 1.0, + "content": "Recently, there is a rich literature about the property of neural tangent kernel. Jacot et al. (2018) first", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "showed that during training, the network function follows a descent along the kernel gradient with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "respect to the Neural Tangent Kernel (NTK) under infinity width setting. Li and Liang (2018) and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 529, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 540 + ], + "score": 1.0, + "content": "Du et al. (2018b) implicitly built connection between Neural Tangent Kernel and gradient descent", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "by showing that GD can provably optimize sufficiently wide two-layer neural networks. In Du et al.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "(2018b), it is shown that gradient descent can achieve zero training loss at a linear convergence rate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "score": 1.0, + "content": "for training two-layer ReLU network with square loss. Allen-Zhu et al. (2018); Du et al. (2018a);", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 572, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 583 + ], + "score": 1.0, + "content": "Zou et al. (2018); Cao and Gu (2019b); Arora et al. (2019a); Zou and Gu (2019); Cao and Gu", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "(2019a) further studied the optimization and generalization of deep neural networks. These papers", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "are all in the so-called neural tangent kernel regime, and their requirements on the network width", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "depend either implicitly or explicitly on the smallest eigenvalue of the kernel Gram matrix. Later,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Su and Yang (2019) showed that this smallest eigenvalue actually scales in the number of samples", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 114, + 637 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "and will eventually converge to 0. In order to obtain constant convergence rate, Su and Yang", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 268, + 650 + ], + "score": 1.0, + "content": "(2019) assumed that the target function ş", + "type": "text" + }, + { + "bbox": [ + 268, + 638, + 280, + 649 + ], + "score": 0.87, + "content": "f ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "can be approximated by the first few eigenfunctions of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 648, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 201, + 662 + ], + "score": 1.0, + "content": "the integrating operator", + "type": "text" + }, + { + "bbox": [ + 202, + 649, + 333, + 662 + ], + "score": 0.92, + "content": "L _ { \\kappa } \\bar { f } ( s ) : = \\int _ { \\mathbb { S } ^ { d } } \\kappa \\bar { ( } x , s ) f ( s ) d \\bar { \\tau } \\bar { ( } s )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 648, + 360, + 662 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 361, + 649, + 385, + 661 + ], + "score": 0.93, + "content": "\\kappa ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 648, + 484, + 662 + ], + "score": 1.0, + "content": "is the NTK function and", + "type": "text" + }, + { + "bbox": [ + 485, + 649, + 504, + 661 + ], + "score": 0.91, + "content": "\\tau ( s )", + "type": "inline_equation" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 660, + 484, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 484, + 672 + ], + "score": 1.0, + "content": "is the input distribution, and proved linear convergence rate up to the this approximation error.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "A few theoretical results have been established towards understanding the spectra of neural tangent", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 504, + 700 + ], + "score": 1.0, + "content": "kernels. Bach (2017) studied two-layer ReLU networks by relating it to kernel methods, and pro-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "posed a harmonic decomposition for the functions in the reproducing kernel Hilbert space which we", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "utilize in our proof. Based on the technique in Bach (2017), Bietti and Mairal (2019) studied the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 272, + 734 + ], + "score": 1.0, + "content": "eigenvalue decay of integrating operator", + "type": "text" + }, + { + "bbox": [ + 273, + 721, + 305, + 732 + ], + "score": 0.93, + "content": "L _ { \\kappa } f ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "defined by NTK on unit sphere by using spheri-", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Meanwhile, a recent line of work (Jacot et al., 2018; Du et al., 2018b; Li and Liang, 2018; Chizat", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 106 + ], + "score": 1.0, + "content": "and Bach, 2018) has shed light on new approaches to analyze neural networks. In particular, they", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 504, + 115 + ], + "score": 1.0, + "content": "show that under certain over-parameterized condition, the neural network trained by gradient de-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "scent behaves similar to the kernel regression predictor using the neural tangent kernel (NTK) (Jacot", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "score": 1.0, + "content": "et al., 2018). Du et al. (2018b) showed that the convergence is provably guaranteed under cer-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "score": 1.0, + "content": "tain over-parameterization conditions determined by the smallest eigenvalue of NTK. Arora et al.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "(2019b) further gave a finer characterization of error convergence based on the eigenvalues of NTK’s", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 457, + 172 + ], + "score": 1.0, + "content": "Gram matrix. Su and Yang (2019) improved the convergence guarantee in terms of the", + "type": "text" + }, + { + "bbox": [ + 457, + 160, + 464, + 169 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "-th largest", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 262, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 262, + 182 + ], + "score": 1.0, + "content": "eigenvalue for certain target functions.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 82, + 505, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Inspired by these works mentioned above, we can present a theoretical explanation for spectral bias.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "Under NTK regime, we establish a precise characterization for the training process of neural net-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "works. More specifically, we theoretically prove that over-parameterized neural networks’ training", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "process can be controlled by the eigenvalues of the integrating operator defined by the NTK. Un-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "der the specific case of uniform distribution on unit sphere, we give an exact calculation for these", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 243, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 254 + ], + "score": 1.0, + "content": "eigenvalues and show that the lower frequencies have larger eigenvalues, which thus leads to faster", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 436, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 436, + 265 + ], + "score": 1.0, + "content": "convergence. We also conduct experiments to corroborate the theory we establish.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 187, + 505, + 265 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 270, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 107, + 270, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 107, + 270, + 288, + 282 + ], + "score": 1.0, + "content": "Our contributions are highlighted as follows:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 107, + 270, + 288, + 282 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 291, + 506, + 465 + ], + "lines": [ + { + "bbox": [ + 107, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 107, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "1. We prove a generic theorem for arbitrary data distributions, which states that under certain sample", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 118, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "complexity and over-parameterization conditions, the error term’s convergence along different", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 119, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 119, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "directions actually relies on the corresponding eigenvalues. This theorem gives a more precise", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 118, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 118, + 324, + 279, + 336 + ], + "score": 1.0, + "content": "control on the regression residual than", + "type": "text" + }, + { + "bbox": [ + 279, + 324, + 291, + 334 + ], + "score": 0.48, + "content": "\\mathtt { S u }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "and Yang (2019), where the authors focused on the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 118, + 334, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 118, + 334, + 505, + 349 + ], + "score": 1.0, + "content": "case when the labeling function is close to the subspace spanned by the first few eigenfunctions.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 107, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "2. We present a more general result about the spectra of the neural tangent kernel. In particular, we", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 358, + 507, + 375 + ], + "spans": [ + { + "bbox": [ + 117, + 358, + 299, + 375 + ], + "score": 1.0, + "content": "show that the order of eigenvalues appears as", + "type": "text" + }, + { + "bbox": [ + 300, + 360, + 423, + 372 + ], + "score": 0.92, + "content": "\\mu _ { k } = \\Omega ( \\operatorname* { m a x } \\{ k ^ { - d - 1 } , \\check { d } ^ { - k + 1 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 358, + 507, + 375 + ], + "score": 1.0, + "content": ". Our result is better", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 118, + 370, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 118, + 370, + 183, + 386 + ], + "score": 1.0, + "content": "than the bound", + "type": "text" + }, + { + "bbox": [ + 183, + 371, + 225, + 384 + ], + "score": 0.93, + "content": "\\Omega ( k ^ { - d - \\overline { { 1 } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 370, + 398, + 386 + ], + "score": 1.0, + "content": "derived in Bietti and Mairal (2019) when", + "type": "text" + }, + { + "bbox": [ + 398, + 372, + 426, + 382 + ], + "score": 0.88, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 370, + 506, + 386 + ], + "score": 1.0, + "content": ", which is clearly a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 118, + 382, + 209, + 397 + ], + "spans": [ + { + "bbox": [ + 118, + 382, + 209, + 397 + ], + "score": 1.0, + "content": "more practical setting.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "3. We establish a rigorous explanation for the spectral bias based on the aforementioned theoretical", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 409, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 118, + 409, + 506, + 422 + ], + "score": 1.0, + "content": "results without any specific assumptions on the target function. We show that the error terms", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 118, + 419, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 118, + 419, + 506, + 433 + ], + "score": 1.0, + "content": "from different frequencies are provably controlled by the eigenvalues of the NTK, and the lower-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 118, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 118, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "frequency components can be learned with less training examples and narrower networks with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 118, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 118, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "a faster convergence rate. As far as we know, this is the first attempt to give a comprehensive", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 118, + 453, + 307, + 465 + ], + "spans": [ + { + "bbox": [ + 118, + 453, + 307, + 465 + ], + "score": 1.0, + "content": "theory justifying the existence of spectral bias.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + } + ], + "index": 24, + "bbox_fs": [ + 106, + 291, + 507, + 465 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 474, + 260, + 485 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 262, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 262, + 486 + ], + "score": 1.0, + "content": "1.1 ADDITIONAL RELATED WORK", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 506, + 508 + ], + "score": 1.0, + "content": "Recently, there is a rich literature about the property of neural tangent kernel. Jacot et al. (2018) first", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "showed that during training, the network function follows a descent along the kernel gradient with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "respect to the Neural Tangent Kernel (NTK) under infinity width setting. Li and Liang (2018) and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 529, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 540 + ], + "score": 1.0, + "content": "Du et al. (2018b) implicitly built connection between Neural Tangent Kernel and gradient descent", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "by showing that GD can provably optimize sufficiently wide two-layer neural networks. In Du et al.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "(2018b), it is shown that gradient descent can achieve zero training loss at a linear convergence rate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "score": 1.0, + "content": "for training two-layer ReLU network with square loss. Allen-Zhu et al. (2018); Du et al. (2018a);", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 572, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 583 + ], + "score": 1.0, + "content": "Zou et al. (2018); Cao and Gu (2019b); Arora et al. (2019a); Zou and Gu (2019); Cao and Gu", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "(2019a) further studied the optimization and generalization of deep neural networks. These papers", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "are all in the so-called neural tangent kernel regime, and their requirements on the network width", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "depend either implicitly or explicitly on the smallest eigenvalue of the kernel Gram matrix. Later,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Su and Yang (2019) showed that this smallest eigenvalue actually scales in the number of samples", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 114, + 637 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "and will eventually converge to 0. In order to obtain constant convergence rate, Su and Yang", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 268, + 650 + ], + "score": 1.0, + "content": "(2019) assumed that the target function ş", + "type": "text" + }, + { + "bbox": [ + 268, + 638, + 280, + 649 + ], + "score": 0.87, + "content": "f ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "can be approximated by the first few eigenfunctions of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 648, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 201, + 662 + ], + "score": 1.0, + "content": "the integrating operator", + "type": "text" + }, + { + "bbox": [ + 202, + 649, + 333, + 662 + ], + "score": 0.92, + "content": "L _ { \\kappa } \\bar { f } ( s ) : = \\int _ { \\mathbb { S } ^ { d } } \\kappa \\bar { ( } x , s ) f ( s ) d \\bar { \\tau } \\bar { ( } s )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 648, + 360, + 662 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 361, + 649, + 385, + 661 + ], + "score": 0.93, + "content": "\\kappa ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 648, + 484, + 662 + ], + "score": 1.0, + "content": "is the NTK function and", + "type": "text" + }, + { + "bbox": [ + 485, + 649, + 504, + 661 + ], + "score": 0.91, + "content": "\\tau ( s )", + "type": "inline_equation" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 660, + 484, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 484, + 672 + ], + "score": 1.0, + "content": "is the input distribution, and proved linear convergence rate up to the this approximation error.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 495, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "A few theoretical results have been established towards understanding the spectra of neural tangent", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 504, + 700 + ], + "score": 1.0, + "content": "kernels. Bach (2017) studied two-layer ReLU networks by relating it to kernel methods, and pro-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "posed a harmonic decomposition for the functions in the reproducing kernel Hilbert space which we", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "utilize in our proof. Based on the technique in Bach (2017), Bietti and Mairal (2019) studied the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 272, + 734 + ], + "score": 1.0, + "content": "eigenvalue decay of integrating operator", + "type": "text" + }, + { + "bbox": [ + 273, + 721, + 305, + 732 + ], + "score": 0.93, + "content": "L _ { \\kappa } f ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "defined by NTK on unit sphere by using spheri-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "cal harmonics. Vempala and Wilmes (2018) calculated the eigenvalues of NTK corresponding to", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "two-layer neural networks with sigmoid activation function. Basri et al. (2019) established similar", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "score": 1.0, + "content": "results as Bietti and Mairal (2019), but considered the case of training the first layer parameters of a", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "two-layer networks with bias terms. Yang and Salman (2019) studied the the eigenvalues of integral", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 391, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 391, + 139 + ], + "score": 1.0, + "content": "operator with respect to the NTK on Boolean cube by Fourier analysis.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 51, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "cal harmonics. Vempala and Wilmes (2018) calculated the eigenvalues of NTK corresponding to", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "two-layer neural networks with sigmoid activation function. Basri et al. (2019) established similar", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "score": 1.0, + "content": "results as Bietti and Mairal (2019), but considered the case of training the first layer parameters of a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "two-layer networks with bias terms. Yang and Salman (2019) studied the the eigenvalues of integral", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 391, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 391, + 139 + ], + "score": 1.0, + "content": "operator with respect to the NTK on Boolean cube by Fourier analysis.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 504, + 155 + ], + "score": 1.0, + "content": "The rest of the paper is organized as follows. We state the notation, problem setup and other prelim-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "inaries in Section 2 and present our main results in Section 3. In Section 4, we present experimental", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 450, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 450, + 178 + ], + "score": 1.0, + "content": "results to support our theory. Proofs of our main results can be found in the appendix.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 107, + 188, + 208, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 209, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 209, + 203 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 211, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "In this section we introduce the basic problem setup including the neural network structure and the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "training algorithm, as well as some background on the neural tangent kernel proposed recently in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 232, + 342, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 342, + 246 + ], + "score": 1.0, + "content": "Jacot et al. (2018) and the corresponding integral operator.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 107, + 254, + 176, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 178, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 178, + 267 + ], + "score": 1.0, + "content": "2.1 NOTATION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 506, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "We use lower case, lower case bold face, and upper case bold face letters to denote scalars, vectors", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 268, + 300 + ], + "score": 1.0, + "content": "and matrices respectively. For a vector ř", + "type": "text" + }, + { + "bbox": [ + 269, + 286, + 373, + 299 + ], + "score": 0.92, + "content": "\\mathbf { v } = ( \\bar { v _ { 1 } } , \\ldots , v _ { d } ) ^ { T } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 285, + 434, + 300 + ], + "score": 1.0, + "content": "and a number", + "type": "text" + }, + { + "bbox": [ + 434, + 288, + 486, + 299 + ], + "score": 0.9, + "content": "1 \\leqslant p < \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 285, + 506, + 300 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 102, + 293, + 509, + 318 + ], + "spans": [ + { + "bbox": [ + 102, + 293, + 197, + 318 + ], + "score": 1.0, + "content": "denote its p´norm by", + "type": "text" + }, + { + "bbox": [ + 198, + 298, + 293, + 312 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\| \\mathbf { v } \\| _ { p } = ( \\sum _ { i = 1 } ^ { d } | v _ { i } | ^ { p } ) ^ { 1 / p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 293, + 427, + 318 + ], + "score": 1.0, + "content": ". We also define infinity norm by", + "type": "text" + }, + { + "bbox": [ + 428, + 300, + 501, + 312 + ], + "score": 0.9, + "content": "\\| \\mathbf { v } \\| _ { \\infty } = \\operatorname* { m a x } _ { i } | v _ { i } |", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 293, + 509, + 318 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 309, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 158, + 325 + ], + "score": 1.0, + "content": "For a matrix ř", + "type": "text" + }, + { + "bbox": [ + 159, + 312, + 225, + 323 + ], + "score": 0.91, + "content": "\\mathbf { A } = ( A _ { i , j } ) _ { m \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 309, + 259, + 325 + ], + "score": 1.0, + "content": ", we use", + "type": "text" + }, + { + "bbox": [ + 259, + 312, + 281, + 323 + ], + "score": 0.9, + "content": "\\lVert \\mathbf { A } \\rVert _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 309, + 459, + 325 + ], + "score": 1.0, + "content": "to denote the number of non-zero entries of", + "type": "text" + }, + { + "bbox": [ + 459, + 313, + 468, + 321 + ], + "score": 0.42, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 309, + 506, + 325 + ], + "score": 1.0, + "content": ", and use", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 317, + 510, + 344 + ], + "spans": [ + { + "bbox": [ + 107, + 323, + 211, + 338 + ], + "score": 0.92, + "content": "\\textstyle \\| \\mathbf { A } \\| _ { F } = ( \\sum _ { i , j = 1 } ^ { d } A _ { i , j } ^ { 2 } ) ^ { 1 / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 317, + 349, + 344 + ], + "score": 1.0, + "content": "to denote its Frobenius norm. Let", + "type": "text" + }, + { + "bbox": [ + 350, + 324, + 460, + 338 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\| \\mathbf { A } \\| _ { p } = \\operatorname* { m a x } _ { \\| \\mathbf { v } \\| _ { p } \\leqslant 1 } \\| \\mathbf { A } \\mathbf { v } \\| _ { p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 317, + 476, + 344 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 476, + 325, + 501, + 336 + ], + "score": 0.88, + "content": "p \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 317, + 510, + 344 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 124, + 351 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 339, + 222, + 351 + ], + "score": 0.9, + "content": "\\| \\mathbf { A } \\| _ { \\operatorname* { m a x } } = \\operatorname* { m a x } _ { i , j } | A _ { i , j } |", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 336, + 296, + 351 + ], + "score": 1.0, + "content": ". For two matrices", + "type": "text" + }, + { + "bbox": [ + 297, + 338, + 355, + 350 + ], + "score": 0.91, + "content": "\\mathbf { A } , \\mathbf { B } \\in \\mathbb { R } ^ { m \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 336, + 400, + 351 + ], + "score": 1.0, + "content": ", we define", + "type": "text" + }, + { + "bbox": [ + 400, + 338, + 485, + 350 + ], + "score": 0.92, + "content": "\\langle \\mathbf { A } , \\mathbf { B } \\rangle = \\operatorname { T r } ( \\mathbf { A } ^ { \\top } \\mathbf { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 336, + 506, + 351 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 348, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 122, + 363 + ], + "score": 1.0, + "content": "use", + "type": "text" + }, + { + "bbox": [ + 123, + 351, + 153, + 360 + ], + "score": 0.84, + "content": "\\mathbf { A } \\succeq \\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 348, + 163, + 363 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 163, + 351, + 192, + 360 + ], + "score": 0.83, + "content": "\\mathbf { A } - \\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 348, + 482, + 363 + ], + "score": 1.0, + "content": "is positive semi-definite. In addition, we define the asymptotic notations", + "type": "text" + }, + { + "bbox": [ + 482, + 350, + 501, + 362 + ], + "score": 0.9, + "content": "\\mathcal { O } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 348, + 506, + 363 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 360, + 507, + 377 + ], + "spans": [ + { + "bbox": [ + 107, + 361, + 149, + 375 + ], + "score": 0.69, + "content": "\\widetilde { \\mathcal { O } } ( \\cdot ) , \\Omega ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 360, + 167, + 377 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 168, + 361, + 186, + 375 + ], + "score": 0.91, + "content": "\\widetilde { \\Omega } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 360, + 288, + 377 + ], + "score": 1.0, + "content": "as follows. Suppose that", + "type": "text" + }, + { + "bbox": [ + 288, + 365, + 299, + 374 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 360, + 317, + 377 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 318, + 363, + 329, + 374 + ], + "score": 0.88, + "content": "b _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 360, + 444, + 377 + ], + "score": 1.0, + "content": "be two sequences. We write", + "type": "text" + }, + { + "bbox": [ + 444, + 363, + 495, + 374 + ], + "score": 0.91, + "content": "a _ { n } = \\mathcal { O } ( b _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 360, + 507, + 377 + ], + "score": 1.0, + "content": "if", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 374, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 123, + 388 + ], + "score": 1.0, + "content": "lim", + "type": "text" + }, + { + "bbox": [ + 123, + 376, + 214, + 388 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\operatorname* { s u p } _ { n \\infty } | a _ { n } / b _ { n } | < \\infty } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 375, + 235, + 388 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 236, + 375, + 286, + 388 + ], + "score": 0.92, + "content": "a _ { n } = \\Omega ( b _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 375, + 297, + 388 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 297, + 375, + 398, + 388 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\operatorname* { l i m } \\operatorname* { i n f } _ { n \\to \\infty } \\left| a _ { n } / b _ { n } \\right| > 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 375, + 435, + 388 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 435, + 374, + 455, + 388 + ], + "score": 0.91, + "content": "\\tilde { \\mathcal { O } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 375, + 474, + 388 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 474, + 374, + 493, + 388 + ], + "score": 0.91, + "content": "\\widetilde { \\Omega } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 386, + 290, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 229, + 400 + ], + "score": 1.0, + "content": "hide the logarithmic factors in", + "type": "text" + }, + { + "bbox": [ + 229, + 388, + 249, + 399 + ], + "score": 0.9, + "content": "\\mathcal { O } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 386, + 267, + 400 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 388, + 286, + 399 + ], + "score": 0.92, + "content": "\\Omega ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 386, + 290, + 400 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 407, + 205, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 205, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 205, + 421 + ], + "score": 1.0, + "content": "2.2 PROBLEM SETUP", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "Here we introduce the basic problem setup. We consider two-layer fully connected neural networks", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 439, + 155, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 155, + 453 + ], + "score": 1.0, + "content": "of the form", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 455, + 366, + 469 + ], + "lines": [ + { + "bbox": [ + 243, + 455, + 366, + 469 + ], + "spans": [ + { + "bbox": [ + 243, + 455, + 366, + 469 + ], + "score": 0.9, + "content": "f _ { \\mathbf { W } } ( \\mathbf { x } ) = \\sqrt { m } \\cdot \\mathbf { W } _ { 2 } \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } ) ,", + "type": "interline_equation", + "image_path": "b7469340c88d73afd76013b416e3d358b8ac7151f7010c91046a7f361bb7b78d.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 243, + 455, + 366, + 469 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 104, + 471, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 133, + 489 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 474, + 204, + 486 + ], + "score": 0.89, + "content": "\\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { m \\times ( d + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 471, + 208, + 489 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 208, + 474, + 266, + 486 + ], + "score": 0.9, + "content": "\\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { 1 \\times m }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 471, + 506, + 489 + ], + "score": 1.0, + "content": "are the first and second layer weight matrices respectively,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 485, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 123, + 500 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 486, + 194, + 498 + ], + "score": 0.91, + "content": "\\sigma ( \\cdot ) = \\operatorname* { m a x } \\{ 0 , \\cdot \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 485, + 506, + 500 + ], + "score": 1.0, + "content": "is the entry-wise ReLU activation function. The network is trained according", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 497, + 371, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 190, + 510 + ], + "score": 1.0, + "content": "to the square loss on", + "type": "text" + }, + { + "bbox": [ + 190, + 500, + 198, + 507 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 497, + 272, + 510 + ], + "score": 1.0, + "content": "training examples", + "type": "text" + }, + { + "bbox": [ + 272, + 497, + 367, + 510 + ], + "score": 0.9, + "content": "S = \\left\\{ \\left( \\mathbf { x } _ { i } , y _ { i } \\right) : i \\in \\left[ n \\right] \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 497, + 371, + 510 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 512, + 388, + 543 + ], + "lines": [ + { + "bbox": [ + 222, + 512, + 388, + 543 + ], + "spans": [ + { + "bbox": [ + 222, + 512, + 388, + 543 + ], + "score": 0.94, + "content": "L _ { S } ( \\mathbf { W } ) = \\frac { 1 } { n } \\sum _ { ( \\mathbf { x } _ { i } , y _ { i } ) \\in S } \\left( y _ { i } - \\theta f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right) ^ { 2 } ,", + "type": "interline_equation", + "image_path": "c624949c8ad186b693f5a14747d129ee97378522590c276fe9947110a2a18031.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 512, + 388, + 527.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 222, + 527.5, + 388, + 543.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 104, + 545, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 133, + 564 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 549, + 140, + 558 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 545, + 461, + 564 + ], + "score": 1.0, + "content": "is a small coefficient to control the effect of initialization, and the data inputs", + "type": "text" + }, + { + "bbox": [ + 461, + 547, + 494, + 561 + ], + "score": 0.92, + "content": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 545, + 507, + 564 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 559, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 298, + 573 + ], + "score": 1.0, + "content": "assumed to follow some unknown distribution", + "type": "text" + }, + { + "bbox": [ + 298, + 563, + 306, + 570 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 559, + 384, + 573 + ], + "score": 1.0, + "content": "on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 385, + 559, + 432, + 571 + ], + "score": 0.92, + "content": "\\mathbb { S } ^ { d } \\in \\mathbb { R } ^ { d + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 559, + 507, + 573 + ], + "score": 1.0, + "content": ". Without loss of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 270, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 232, + 584 + ], + "score": 1.0, + "content": "generality, we also assume that", + "type": "text" + }, + { + "bbox": [ + 233, + 571, + 266, + 583 + ], + "score": 0.92, + "content": "| y _ { i } | \\leqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 571, + 270, + 584 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 502, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 504, + 601 + ], + "score": 1.0, + "content": "We first randomly initialize the parameters of the network, and then apply gradient descent to opti-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 599, + 475, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 475, + 612 + ], + "score": 1.0, + "content": "mize both layers. We present our detailed neural network training algorithm in Algorithm 1.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "table", + "bbox": [ + 113, + 630, + 462, + 703 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 615, + 353, + 627 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 614, + 354, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 354, + 628 + ], + "score": 1.0, + "content": "Algorithm 1 GD for DNNs starting at Gaussian initialization", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "table_body", + "bbox": [ + 113, + 630, + 462, + 703 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 630, + 462, + 703 + ], + "spans": [ + { + "bbox": [ + 113, + 630, + 462, + 703 + ], + "score": 0.829, + "html": "
Input: Number of iterations T,step size n. Generate each entry of W(0) and w)
from N(0,2/m) and N(0,1/m) respectively. 2 fort=0,1,...,T-1do
Update W(t+1) = W(t) - n · VwLs(W(t).
end for Output: W(T).
", + "type": "table", + "image_path": "b9511ab019a200c3aca9055c161b6cd32086034bf8efbfaa8da1f39b2f3d56b3.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 113, + 630, + 462, + 654.3333333333334 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 113, + 654.3333333333334, + 462, + 678.6666666666667 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 113, + 678.6666666666667, + 462, + 703.0000000000001 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "index": 38.0 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 711, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 710, + 234, + 723 + ], + "score": 1.0, + "content": "1Here the dimension of input is", + "type": "text" + }, + { + "bbox": [ + 234, + 712, + 255, + 721 + ], + "score": 0.87, + "content": "d + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "since throughout this paper we assume that all training data lie in the", + "type": "text" + } + ] + }, + { + "bbox": [ + 107, + 720, + 245, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 722, + 113, + 730 + ], + "score": 0.75, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 720, + 202, + 732 + ], + "score": 1.0, + "content": "-dimensional unit sphere", + "type": "text" + }, + { + "bbox": [ + 202, + 721, + 242, + 731 + ], + "score": 0.91, + "content": "\\mathbb { S } ^ { d ^ { \\bullet } } \\in \\mathbb { R } ^ { d + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 720, + 245, + 732 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 81, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 504, + 155 + ], + "score": 1.0, + "content": "The rest of the paper is organized as follows. We state the notation, problem setup and other prelim-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "inaries in Section 2 and present our main results in Section 3. In Section 4, we present experimental", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 450, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 450, + 178 + ], + "score": 1.0, + "content": "results to support our theory. Proofs of our main results can be found in the appendix.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 143, + 505, + 178 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 188, + 208, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 209, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 209, + 203 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 211, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "In this section we introduce the basic problem setup including the neural network structure and the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "training algorithm, as well as some background on the neural tangent kernel proposed recently in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 232, + 342, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 342, + 246 + ], + "score": 1.0, + "content": "Jacot et al. (2018) and the corresponding integral operator.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 212, + 505, + 246 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 254, + 176, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 178, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 178, + 267 + ], + "score": 1.0, + "content": "2.1 NOTATION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 506, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "We use lower case, lower case bold face, and upper case bold face letters to denote scalars, vectors", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 268, + 300 + ], + "score": 1.0, + "content": "and matrices respectively. For a vector ř", + "type": "text" + }, + { + "bbox": [ + 269, + 286, + 373, + 299 + ], + "score": 0.92, + "content": "\\mathbf { v } = ( \\bar { v _ { 1 } } , \\ldots , v _ { d } ) ^ { T } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 285, + 434, + 300 + ], + "score": 1.0, + "content": "and a number", + "type": "text" + }, + { + "bbox": [ + 434, + 288, + 486, + 299 + ], + "score": 0.9, + "content": "1 \\leqslant p < \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 285, + 506, + 300 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 102, + 293, + 509, + 318 + ], + "spans": [ + { + "bbox": [ + 102, + 293, + 197, + 318 + ], + "score": 1.0, + "content": "denote its p´norm by", + "type": "text" + }, + { + "bbox": [ + 198, + 298, + 293, + 312 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\| \\mathbf { v } \\| _ { p } = ( \\sum _ { i = 1 } ^ { d } | v _ { i } | ^ { p } ) ^ { 1 / p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 293, + 427, + 318 + ], + "score": 1.0, + "content": ". We also define infinity norm by", + "type": "text" + }, + { + "bbox": [ + 428, + 300, + 501, + 312 + ], + "score": 0.9, + "content": "\\| \\mathbf { v } \\| _ { \\infty } = \\operatorname* { m a x } _ { i } | v _ { i } |", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 293, + 509, + 318 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 309, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 158, + 325 + ], + "score": 1.0, + "content": "For a matrix ř", + "type": "text" + }, + { + "bbox": [ + 159, + 312, + 225, + 323 + ], + "score": 0.91, + "content": "\\mathbf { A } = ( A _ { i , j } ) _ { m \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 309, + 259, + 325 + ], + "score": 1.0, + "content": ", we use", + "type": "text" + }, + { + "bbox": [ + 259, + 312, + 281, + 323 + ], + "score": 0.9, + "content": "\\lVert \\mathbf { A } \\rVert _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 309, + 459, + 325 + ], + "score": 1.0, + "content": "to denote the number of non-zero entries of", + "type": "text" + }, + { + "bbox": [ + 459, + 313, + 468, + 321 + ], + "score": 0.42, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 309, + 506, + 325 + ], + "score": 1.0, + "content": ", and use", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 317, + 510, + 344 + ], + "spans": [ + { + "bbox": [ + 107, + 323, + 211, + 338 + ], + "score": 0.92, + "content": "\\textstyle \\| \\mathbf { A } \\| _ { F } = ( \\sum _ { i , j = 1 } ^ { d } A _ { i , j } ^ { 2 } ) ^ { 1 / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 317, + 349, + 344 + ], + "score": 1.0, + "content": "to denote its Frobenius norm. Let", + "type": "text" + }, + { + "bbox": [ + 350, + 324, + 460, + 338 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\| \\mathbf { A } \\| _ { p } = \\operatorname* { m a x } _ { \\| \\mathbf { v } \\| _ { p } \\leqslant 1 } \\| \\mathbf { A } \\mathbf { v } \\| _ { p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 317, + 476, + 344 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 476, + 325, + 501, + 336 + ], + "score": 0.88, + "content": "p \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 317, + 510, + 344 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 124, + 351 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 339, + 222, + 351 + ], + "score": 0.9, + "content": "\\| \\mathbf { A } \\| _ { \\operatorname* { m a x } } = \\operatorname* { m a x } _ { i , j } | A _ { i , j } |", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 336, + 296, + 351 + ], + "score": 1.0, + "content": ". For two matrices", + "type": "text" + }, + { + "bbox": [ + 297, + 338, + 355, + 350 + ], + "score": 0.91, + "content": "\\mathbf { A } , \\mathbf { B } \\in \\mathbb { R } ^ { m \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 336, + 400, + 351 + ], + "score": 1.0, + "content": ", we define", + "type": "text" + }, + { + "bbox": [ + 400, + 338, + 485, + 350 + ], + "score": 0.92, + "content": "\\langle \\mathbf { A } , \\mathbf { B } \\rangle = \\operatorname { T r } ( \\mathbf { A } ^ { \\top } \\mathbf { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 336, + 506, + 351 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 348, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 122, + 363 + ], + "score": 1.0, + "content": "use", + "type": "text" + }, + { + "bbox": [ + 123, + 351, + 153, + 360 + ], + "score": 0.84, + "content": "\\mathbf { A } \\succeq \\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 348, + 163, + 363 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 163, + 351, + 192, + 360 + ], + "score": 0.83, + "content": "\\mathbf { A } - \\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 348, + 482, + 363 + ], + "score": 1.0, + "content": "is positive semi-definite. In addition, we define the asymptotic notations", + "type": "text" + }, + { + "bbox": [ + 482, + 350, + 501, + 362 + ], + "score": 0.9, + "content": "\\mathcal { O } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 348, + 506, + 363 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 360, + 507, + 377 + ], + "spans": [ + { + "bbox": [ + 107, + 361, + 149, + 375 + ], + "score": 0.69, + "content": "\\widetilde { \\mathcal { O } } ( \\cdot ) , \\Omega ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 360, + 167, + 377 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 168, + 361, + 186, + 375 + ], + "score": 0.91, + "content": "\\widetilde { \\Omega } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 360, + 288, + 377 + ], + "score": 1.0, + "content": "as follows. Suppose that", + "type": "text" + }, + { + "bbox": [ + 288, + 365, + 299, + 374 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 360, + 317, + 377 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 318, + 363, + 329, + 374 + ], + "score": 0.88, + "content": "b _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 360, + 444, + 377 + ], + "score": 1.0, + "content": "be two sequences. We write", + "type": "text" + }, + { + "bbox": [ + 444, + 363, + 495, + 374 + ], + "score": 0.91, + "content": "a _ { n } = \\mathcal { O } ( b _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 360, + 507, + 377 + ], + "score": 1.0, + "content": "if", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 374, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 123, + 388 + ], + "score": 1.0, + "content": "lim", + "type": "text" + }, + { + "bbox": [ + 123, + 376, + 214, + 388 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\operatorname* { s u p } _ { n \\infty } | a _ { n } / b _ { n } | < \\infty } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 375, + 235, + 388 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 236, + 375, + 286, + 388 + ], + "score": 0.92, + "content": "a _ { n } = \\Omega ( b _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 375, + 297, + 388 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 297, + 375, + 398, + 388 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\operatorname* { l i m } \\operatorname* { i n f } _ { n \\to \\infty } \\left| a _ { n } / b _ { n } \\right| > 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 375, + 435, + 388 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 435, + 374, + 455, + 388 + ], + "score": 0.91, + "content": "\\tilde { \\mathcal { O } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 375, + 474, + 388 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 474, + 374, + 493, + 388 + ], + "score": 0.91, + "content": "\\widetilde { \\Omega } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 386, + 290, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 229, + 400 + ], + "score": 1.0, + "content": "hide the logarithmic factors in", + "type": "text" + }, + { + "bbox": [ + 229, + 388, + 249, + 399 + ], + "score": 0.9, + "content": "\\mathcal { O } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 386, + 267, + 400 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 388, + 286, + 399 + ], + "score": 0.92, + "content": "\\Omega ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 386, + 290, + 400 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5, + "bbox_fs": [ + 102, + 275, + 510, + 400 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 407, + 205, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 205, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 205, + 421 + ], + "score": 1.0, + "content": "2.2 PROBLEM SETUP", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "Here we introduce the basic problem setup. We consider two-layer fully connected neural networks", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 439, + 155, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 155, + 453 + ], + "score": 1.0, + "content": "of the form", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 429, + 505, + 453 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 455, + 366, + 469 + ], + "lines": [ + { + "bbox": [ + 243, + 455, + 366, + 469 + ], + "spans": [ + { + "bbox": [ + 243, + 455, + 366, + 469 + ], + "score": 0.9, + "content": "f _ { \\mathbf { W } } ( \\mathbf { x } ) = \\sqrt { m } \\cdot \\mathbf { W } _ { 2 } \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } ) ,", + "type": "interline_equation", + "image_path": "b7469340c88d73afd76013b416e3d358b8ac7151f7010c91046a7f361bb7b78d.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 243, + 455, + 366, + 469 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 104, + 471, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 133, + 489 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 474, + 204, + 486 + ], + "score": 0.89, + "content": "\\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { m \\times ( d + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 471, + 208, + 489 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 208, + 474, + 266, + 486 + ], + "score": 0.9, + "content": "\\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { 1 \\times m }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 471, + 506, + 489 + ], + "score": 1.0, + "content": "are the first and second layer weight matrices respectively,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 485, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 123, + 500 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 486, + 194, + 498 + ], + "score": 0.91, + "content": "\\sigma ( \\cdot ) = \\operatorname* { m a x } \\{ 0 , \\cdot \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 485, + 506, + 500 + ], + "score": 1.0, + "content": "is the entry-wise ReLU activation function. The network is trained according", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 497, + 371, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 190, + 510 + ], + "score": 1.0, + "content": "to the square loss on", + "type": "text" + }, + { + "bbox": [ + 190, + 500, + 198, + 507 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 497, + 272, + 510 + ], + "score": 1.0, + "content": "training examples", + "type": "text" + }, + { + "bbox": [ + 272, + 497, + 367, + 510 + ], + "score": 0.9, + "content": "S = \\left\\{ \\left( \\mathbf { x } _ { i } , y _ { i } \\right) : i \\in \\left[ n \\right] \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 497, + 371, + 510 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 471, + 506, + 510 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 512, + 388, + 543 + ], + "lines": [ + { + "bbox": [ + 222, + 512, + 388, + 543 + ], + "spans": [ + { + "bbox": [ + 222, + 512, + 388, + 543 + ], + "score": 0.94, + "content": "L _ { S } ( \\mathbf { W } ) = \\frac { 1 } { n } \\sum _ { ( \\mathbf { x } _ { i } , y _ { i } ) \\in S } \\left( y _ { i } - \\theta f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right) ^ { 2 } ,", + "type": "interline_equation", + "image_path": "c624949c8ad186b693f5a14747d129ee97378522590c276fe9947110a2a18031.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 512, + 388, + 527.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 222, + 527.5, + 388, + 543.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 104, + 545, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 133, + 564 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 549, + 140, + 558 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 545, + 461, + 564 + ], + "score": 1.0, + "content": "is a small coefficient to control the effect of initialization, and the data inputs", + "type": "text" + }, + { + "bbox": [ + 461, + 547, + 494, + 561 + ], + "score": 0.92, + "content": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 545, + 507, + 564 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 559, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 298, + 573 + ], + "score": 1.0, + "content": "assumed to follow some unknown distribution", + "type": "text" + }, + { + "bbox": [ + 298, + 563, + 306, + 570 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 559, + 384, + 573 + ], + "score": 1.0, + "content": "on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 385, + 559, + 432, + 571 + ], + "score": 0.92, + "content": "\\mathbb { S } ^ { d } \\in \\mathbb { R } ^ { d + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 559, + 507, + 573 + ], + "score": 1.0, + "content": ". Without loss of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 270, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 232, + 584 + ], + "score": 1.0, + "content": "generality, we also assume that", + "type": "text" + }, + { + "bbox": [ + 233, + 571, + 266, + 583 + ], + "score": 0.92, + "content": "| y _ { i } | \\leqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 571, + 270, + 584 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 545, + 507, + 584 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 502, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 504, + 601 + ], + "score": 1.0, + "content": "We first randomly initialize the parameters of the network, and then apply gradient descent to opti-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 599, + 475, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 475, + 612 + ], + "score": 1.0, + "content": "mize both layers. We present our detailed neural network training algorithm in Algorithm 1.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 587, + 504, + 612 + ] + }, + { + "type": "table", + "bbox": [ + 113, + 630, + 462, + 703 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 615, + 353, + 627 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 614, + 354, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 354, + 628 + ], + "score": 1.0, + "content": "Algorithm 1 GD for DNNs starting at Gaussian initialization", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "table_body", + "bbox": [ + 113, + 630, + 462, + 703 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 630, + 462, + 703 + ], + "spans": [ + { + "bbox": [ + 113, + 630, + 462, + 703 + ], + "score": 0.829, + "html": "
Input: Number of iterations T,step size n. Generate each entry of W(0) and w)
from N(0,2/m) and N(0,1/m) respectively. 2 fort=0,1,...,T-1do
Update W(t+1) = W(t) - n · VwLs(W(t).
end for Output: W(T).
", + "type": "table", + "image_path": "b9511ab019a200c3aca9055c161b6cd32086034bf8efbfaa8da1f39b2f3d56b3.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 113, + 630, + 462, + 654.3333333333334 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 113, + 654.3333333333334, + 462, + 678.6666666666667 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 113, + 678.6666666666667, + 462, + 703.0000000000001 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "index": 38.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 226, + 96 + ], + "score": 1.0, + "content": "The initialization scheme for", + "type": "text" + }, + { + "bbox": [ + 227, + 81, + 250, + 93 + ], + "score": 0.83, + "content": "\\mathbf { W } ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "given in Algorithm 1 is known as He initialization (He et al.,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "2015). This scheme generates each entry of the weight matrices from a Gaussian distribution with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "mean zero. The variances of the Gaussian distributions in initialization are chosen following the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "principle that the initialization does not change the magnitudes of inputs in each layer. The sec-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "ond layer parameter is not associated with the ReLU activation function, thus it is initialized with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 227, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 142, + 150 + ], + "score": 1.0, + "content": "variance", + "type": "text" + }, + { + "bbox": [ + 142, + 137, + 162, + 149 + ], + "score": 0.9, + "content": "1 / m", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 137, + 204, + 150 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 205, + 137, + 223, + 149 + ], + "score": 0.9, + "content": "2 / m", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 137, + 227, + 150 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 158, + 250, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 158, + 252, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 252, + 171 + ], + "score": 1.0, + "content": "2.3 NEURAL TANGENT KERNEL", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 179, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "Many attempts have been made to study the convergence of gradient descent assuming the width", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "score": 1.0, + "content": "of the network is extremely large (Du et al., 2018b; Li and Liang, 2018). When the width of the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "network goes to infinity, with certain initialization on parameters, the inner product of gradients of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 212, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 504, + 225 + ], + "score": 1.0, + "content": "the output function would converge to a limiting kernel, i.e., Neural Tangent Kernel (Jacot et al.,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 222, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 254, + 237 + ], + "score": 1.0, + "content": "2018). In this paper, we denote it by", + "type": "text" + }, + { + "bbox": [ + 254, + 223, + 487, + 236 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\operatorname* { l i m } _ { m \\infty } m ^ { - 1 } \\langle \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ) , \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ^ { \\prime } ) \\rangle } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 222, + 506, + 237 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 233, + 143, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 143, + 247 + ], + "score": 1.0, + "content": "we have", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 250, + 397, + 265 + ], + "lines": [ + { + "bbox": [ + 212, + 250, + 397, + 265 + ], + "spans": [ + { + "bbox": [ + 212, + 250, + 397, + 265 + ], + "score": 0.92, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle \\cdot \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) + 2 \\cdot \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) ,", + "type": "interline_equation", + "image_path": "302018cca28021cbfcc51c593d3df79466e93feae8679c3a70370982ca9d6fcb.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 212, + 250, + 397, + 265 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 133, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 135, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 135, + 282 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 276, + 403, + 309 + ], + "lines": [ + { + "bbox": [ + 207, + 276, + 402, + 309 + ], + "spans": [ + { + "bbox": [ + 207, + 276, + 402, + 309 + ], + "score": 0.61, + "content": "\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] , } \\\\ & { \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] . } \\end{array}", + "type": "interline_equation", + "image_path": "64a6c508918c60ae6b0f1d7ea501c2d922bcc832e1a0e99dfcf96f5e8fb3f56a.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 276, + 403, + 292.5 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 207, + 292.5, + 403, + 309.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 506, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "Since we apply gradient descent to both layers, the Neural Tangent Kernel is the sum of two different", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "kernel functions and clearly it can be reduced to one layer training setting. These two kernels are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 499, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 499, + 345 + ], + "score": 1.0, + "content": "arc-cosine kernels of degree 0 and 1 (Cho and Saul, 2009). Their explicit expressions are given as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 348, + 466, + 372 + ], + "lines": [ + { + "bbox": [ + 144, + 348, + 466, + 372 + ], + "spans": [ + { + "bbox": [ + 144, + 348, + 466, + 372 + ], + "score": 0.91, + "content": "\\kappa _ { 1 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) , \\kappa _ { 2 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( t \\cdot \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) + \\sqrt { 1 - t ^ { 2 } } \\right) ,", + "type": "interline_equation", + "image_path": "1a185070145cdd03708c097caaf18bd78644f54477168358a065491a6cefce37.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 144, + 348, + 466, + 372 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 228, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 228, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 133, + 391 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 376, + 225, + 390 + ], + "score": 0.9, + "content": "t = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle / \\left( \\left\\| \\mathbf { x } \\right\\| \\left\\| \\mathbf { x } ^ { \\prime } \\right\\| \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 376, + 228, + 391 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 398, + 225, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 226, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 226, + 411 + ], + "score": 1.0, + "content": "2.4 INTEGRAL OPERATOR", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "The theory of integral operator with respect to kernel function has been well studied in machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "learning (Smale and Zhou, 2007; Rosasco et al., 2010) thus we only give a brief introduction here.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 441, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 123, + 455 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 442, + 153, + 454 + ], + "score": 0.91, + "content": "L _ { \\tau } ^ { 2 } ( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 441, + 497, + 455 + ], + "score": 1.0, + "content": "be the Hilbert space of square-integrable functions with respect to a Borel measure", + "type": "text" + }, + { + "bbox": [ + 497, + 444, + 504, + 452 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 128, + 465 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 454, + 161, + 462 + ], + "score": 0.89, + "content": "X \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 451, + 308, + 465 + ], + "score": 1.0, + "content": ". For any continuous kernel function", + "type": "text" + }, + { + "bbox": [ + 308, + 453, + 377, + 463 + ], + "score": 0.91, + "content": "\\kappa : X \\times X \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 451, + 394, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 455, + 402, + 462 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "we can define an integral", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 463, + 212, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 142, + 477 + ], + "score": 1.0, + "content": "operator", + "type": "text" + }, + { + "bbox": [ + 142, + 465, + 155, + 475 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 463, + 169, + 477 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 169, + 464, + 198, + 476 + ], + "score": 0.92, + "content": "L _ { \\tau } ^ { 2 } ( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 463, + 212, + 477 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 480, + 397, + 506 + ], + "lines": [ + { + "bbox": [ + 213, + 480, + 397, + 506 + ], + "spans": [ + { + "bbox": [ + 213, + 480, + 397, + 506 + ], + "score": 0.95, + "content": "L _ { \\kappa } ( f ) ( { \\bf x } ) = \\int _ { X } \\kappa ( { \\bf x } , { \\bf y } ) f ( { \\bf y } ) d \\tau ( { \\bf y } ) , ~ { \\bf x } \\in X .", + "type": "interline_equation", + "image_path": "c171d3e3de6037e1cb648cff9aaf4c52d0260ff0cdab83a9f81089f35558335c.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 213, + 480, + 397, + 506 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 511, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "It has been pointed out in Cho and Saul (2009) that arc-cosine kernels are positive semi-definite.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 208, + 534 + ], + "score": 1.0, + "content": "Thus the kernel function", + "type": "text" + }, + { + "bbox": [ + 209, + 524, + 216, + 532 + ], + "score": 0.77, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "defined by (2.1) is positive semi-definite being a product and a sum of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "positive semi-definite kernels. Clearly this kernel is also continuous and symmetric. Thus we know", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 311, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 225, + 555 + ], + "score": 1.0, + "content": "that the neural tangent kernel", + "type": "text" + }, + { + "bbox": [ + 225, + 546, + 232, + 554 + ], + "score": 0.78, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 544, + 311, + 555 + ], + "score": 1.0, + "content": "is a Mercer kernel.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 566, + 206, + 579 + ], + "lines": [ + { + "bbox": [ + 104, + 565, + 207, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 565, + 207, + 582 + ], + "score": 1.0, + "content": "3 MAIN RESULTS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 602 + ], + "score": 1.0, + "content": "In this section we present our main results. In Section 3.1, we give a general result on the conver-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "gence rate of gradient descent along different eigendirections of neural tangent kernel. Motivated", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 392, + 625 + ], + "score": 1.0, + "content": "by this result, in Section 3.2, we give a case study on the spectrum of", + "type": "text" + }, + { + "bbox": [ + 392, + 613, + 406, + 623 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "when the input data are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 280, + 636 + ], + "score": 1.0, + "content": "uniformly distributed over the unit sphere", + "type": "text" + }, + { + "bbox": [ + 280, + 623, + 291, + 633 + ], + "score": 0.88, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 622, + 505, + 636 + ], + "score": 1.0, + "content": ". In Section 3.3, we combine the spectrum analysis", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "with the general convergence result to give explicit convergence rate for uniformly distributed data", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 645, + 182, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 182, + 657 + ], + "score": 1.0, + "content": "on the unit sphere.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 106, + 666, + 350, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 351, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 351, + 678 + ], + "score": 1.0, + "content": "3.1 CONVERGENCE ANALYSIS OF GRADIENT DESCENT", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "In this section we study the convergence of Algorithm 1. Instead of studying the standard conver-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "gence of loss function value, we aim to provide a refined analysis on the speed of convergence along", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 315, + 722 + ], + "score": 1.0, + "content": "different directions defined by the eigenfunctions of", + "type": "text" + }, + { + "bbox": [ + 315, + 710, + 327, + 721 + ], + "score": 0.88, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 709, + 506, + 722 + ], + "score": 1.0, + "content": ". We first introduce the following definitions", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 165, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 165, + 732 + ], + "score": 1.0, + "content": "and notations.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 226, + 96 + ], + "score": 1.0, + "content": "The initialization scheme for", + "type": "text" + }, + { + "bbox": [ + 227, + 81, + 250, + 93 + ], + "score": 0.83, + "content": "\\mathbf { W } ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "given in Algorithm 1 is known as He initialization (He et al.,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "2015). This scheme generates each entry of the weight matrices from a Gaussian distribution with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "mean zero. The variances of the Gaussian distributions in initialization are chosen following the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "principle that the initialization does not change the magnitudes of inputs in each layer. The sec-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "ond layer parameter is not associated with the ReLU activation function, thus it is initialized with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 227, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 142, + 150 + ], + "score": 1.0, + "content": "variance", + "type": "text" + }, + { + "bbox": [ + 142, + 137, + 162, + 149 + ], + "score": 0.9, + "content": "1 / m", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 137, + 204, + 150 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 205, + 137, + 223, + 149 + ], + "score": 0.9, + "content": "2 / m", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 137, + 227, + 150 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 81, + 505, + 150 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 158, + 250, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 158, + 252, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 252, + 171 + ], + "score": 1.0, + "content": "2.3 NEURAL TANGENT KERNEL", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 179, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "Many attempts have been made to study the convergence of gradient descent assuming the width", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "score": 1.0, + "content": "of the network is extremely large (Du et al., 2018b; Li and Liang, 2018). When the width of the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "network goes to infinity, with certain initialization on parameters, the inner product of gradients of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 212, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 504, + 225 + ], + "score": 1.0, + "content": "the output function would converge to a limiting kernel, i.e., Neural Tangent Kernel (Jacot et al.,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 222, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 254, + 237 + ], + "score": 1.0, + "content": "2018). In this paper, we denote it by", + "type": "text" + }, + { + "bbox": [ + 254, + 223, + 487, + 236 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\operatorname* { l i m } _ { m \\infty } m ^ { - 1 } \\langle \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ) , \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ^ { \\prime } ) \\rangle } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 222, + 506, + 237 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 233, + 143, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 143, + 247 + ], + "score": 1.0, + "content": "we have", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 180, + 506, + 247 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 250, + 397, + 265 + ], + "lines": [ + { + "bbox": [ + 212, + 250, + 397, + 265 + ], + "spans": [ + { + "bbox": [ + 212, + 250, + 397, + 265 + ], + "score": 0.92, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle \\cdot \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) + 2 \\cdot \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) ,", + "type": "interline_equation", + "image_path": "302018cca28021cbfcc51c593d3df79466e93feae8679c3a70370982ca9d6fcb.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 212, + 250, + 397, + 265 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 133, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 135, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 135, + 282 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 268, + 135, + 282 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 276, + 403, + 309 + ], + "lines": [ + { + "bbox": [ + 207, + 276, + 402, + 309 + ], + "spans": [ + { + "bbox": [ + 207, + 276, + 402, + 309 + ], + "score": 0.61, + "content": "\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] , } \\\\ & { \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] . } \\end{array}", + "type": "interline_equation", + "image_path": "64a6c508918c60ae6b0f1d7ea501c2d922bcc832e1a0e99dfcf96f5e8fb3f56a.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 276, + 403, + 292.5 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 207, + 292.5, + 403, + 309.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 506, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "Since we apply gradient descent to both layers, the Neural Tangent Kernel is the sum of two different", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "kernel functions and clearly it can be reduced to one layer training setting. These two kernels are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 499, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 499, + 345 + ], + "score": 1.0, + "content": "arc-cosine kernels of degree 0 and 1 (Cho and Saul, 2009). Their explicit expressions are given as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 309, + 506, + 345 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 348, + 466, + 372 + ], + "lines": [ + { + "bbox": [ + 144, + 348, + 466, + 372 + ], + "spans": [ + { + "bbox": [ + 144, + 348, + 466, + 372 + ], + "score": 0.91, + "content": "\\kappa _ { 1 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) , \\kappa _ { 2 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( t \\cdot \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) + \\sqrt { 1 - t ^ { 2 } } \\right) ,", + "type": "interline_equation", + "image_path": "1a185070145cdd03708c097caaf18bd78644f54477168358a065491a6cefce37.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 144, + 348, + 466, + 372 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 228, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 228, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 133, + 391 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 376, + 225, + 390 + ], + "score": 0.9, + "content": "t = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle / \\left( \\left\\| \\mathbf { x } \\right\\| \\left\\| \\mathbf { x } ^ { \\prime } \\right\\| \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 376, + 228, + 391 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 376, + 228, + 391 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 398, + 225, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 226, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 226, + 411 + ], + "score": 1.0, + "content": "2.4 INTEGRAL OPERATOR", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "The theory of integral operator with respect to kernel function has been well studied in machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "learning (Smale and Zhou, 2007; Rosasco et al., 2010) thus we only give a brief introduction here.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 441, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 123, + 455 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 442, + 153, + 454 + ], + "score": 0.91, + "content": "L _ { \\tau } ^ { 2 } ( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 441, + 497, + 455 + ], + "score": 1.0, + "content": "be the Hilbert space of square-integrable functions with respect to a Borel measure", + "type": "text" + }, + { + "bbox": [ + 497, + 444, + 504, + 452 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 128, + 465 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 454, + 161, + 462 + ], + "score": 0.89, + "content": "X \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 451, + 308, + 465 + ], + "score": 1.0, + "content": ". For any continuous kernel function", + "type": "text" + }, + { + "bbox": [ + 308, + 453, + 377, + 463 + ], + "score": 0.91, + "content": "\\kappa : X \\times X \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 451, + 394, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 455, + 402, + 462 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "we can define an integral", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 463, + 212, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 142, + 477 + ], + "score": 1.0, + "content": "operator", + "type": "text" + }, + { + "bbox": [ + 142, + 465, + 155, + 475 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 463, + 169, + 477 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 169, + 464, + 198, + 476 + ], + "score": 0.92, + "content": "L _ { \\tau } ^ { 2 } ( X )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 463, + 212, + 477 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 420, + 505, + 477 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 480, + 397, + 506 + ], + "lines": [ + { + "bbox": [ + 213, + 480, + 397, + 506 + ], + "spans": [ + { + "bbox": [ + 213, + 480, + 397, + 506 + ], + "score": 0.95, + "content": "L _ { \\kappa } ( f ) ( { \\bf x } ) = \\int _ { X } \\kappa ( { \\bf x } , { \\bf y } ) f ( { \\bf y } ) d \\tau ( { \\bf y } ) , ~ { \\bf x } \\in X .", + "type": "interline_equation", + "image_path": "c171d3e3de6037e1cb648cff9aaf4c52d0260ff0cdab83a9f81089f35558335c.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 213, + 480, + 397, + 506 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 511, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "It has been pointed out in Cho and Saul (2009) that arc-cosine kernels are positive semi-definite.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 208, + 534 + ], + "score": 1.0, + "content": "Thus the kernel function", + "type": "text" + }, + { + "bbox": [ + 209, + 524, + 216, + 532 + ], + "score": 0.77, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "defined by (2.1) is positive semi-definite being a product and a sum of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "positive semi-definite kernels. Clearly this kernel is also continuous and symmetric. Thus we know", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 311, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 225, + 555 + ], + "score": 1.0, + "content": "that the neural tangent kernel", + "type": "text" + }, + { + "bbox": [ + 225, + 546, + 232, + 554 + ], + "score": 0.78, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 544, + 311, + 555 + ], + "score": 1.0, + "content": "is a Mercer kernel.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 510, + 506, + 555 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 566, + 206, + 579 + ], + "lines": [ + { + "bbox": [ + 104, + 565, + 207, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 565, + 207, + 582 + ], + "score": 1.0, + "content": "3 MAIN RESULTS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 602 + ], + "score": 1.0, + "content": "In this section we present our main results. In Section 3.1, we give a general result on the conver-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "gence rate of gradient descent along different eigendirections of neural tangent kernel. Motivated", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 392, + 625 + ], + "score": 1.0, + "content": "by this result, in Section 3.2, we give a case study on the spectrum of", + "type": "text" + }, + { + "bbox": [ + 392, + 613, + 406, + 623 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "when the input data are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 280, + 636 + ], + "score": 1.0, + "content": "uniformly distributed over the unit sphere", + "type": "text" + }, + { + "bbox": [ + 280, + 623, + 291, + 633 + ], + "score": 0.88, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 622, + 505, + 636 + ], + "score": 1.0, + "content": ". In Section 3.3, we combine the spectrum analysis", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "with the general convergence result to give explicit convergence rate for uniformly distributed data", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 645, + 182, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 182, + 657 + ], + "score": 1.0, + "content": "on the unit sphere.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 591, + 506, + 657 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 666, + 350, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 351, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 351, + 678 + ], + "score": 1.0, + "content": "3.1 CONVERGENCE ANALYSIS OF GRADIENT DESCENT", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "In this section we study the convergence of Algorithm 1. Instead of studying the standard conver-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "gence of loss function value, we aim to provide a refined analysis on the speed of convergence along", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 315, + 722 + ], + "score": 1.0, + "content": "different directions defined by the eigenfunctions of", + "type": "text" + }, + { + "bbox": [ + 315, + 710, + 327, + 721 + ], + "score": 0.88, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 709, + 506, + 722 + ], + "score": 1.0, + "content": ". We first introduce the following definitions", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 165, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 165, + 732 + ], + "score": 1.0, + "content": "and notations.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 123, + 96 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 82, + 155, + 95 + ], + "score": 0.93, + "content": "\\{ \\lambda _ { i } \\} _ { i \\geqslant 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 81, + 177, + 96 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 177, + 83, + 239, + 94 + ], + "score": 0.91, + "content": "\\lambda _ { 1 } \\geqslant \\lambda _ { 2 } \\geqslant \\cdots", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 81, + 393, + 96 + ], + "score": 1.0, + "content": "be the strictly positive eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 394, + 83, + 406, + 93 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 81, + 428, + 96 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 428, + 82, + 491, + 95 + ], + "score": 0.9, + "content": "\\phi _ { 1 } ( \\cdot ) , \\phi _ { 2 } ( \\cdot ) , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 503, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 309, + 108 + ], + "score": 1.0, + "content": "the corresponding orthonormal eigenfunctions. Set", + "type": "text" + }, + { + "bbox": [ + 310, + 93, + 449, + 107 + ], + "score": 0.89, + "content": "\\mathbf v _ { i } = n ^ { - 1 / 2 } ( \\phi _ { i } ( { \\mathbf { x } } _ { 1 } ) , \\dots , \\phi _ { i } ( { \\mathbf { x } } _ { n } ) ) ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 92, + 452, + 108 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 453, + 95, + 503, + 106 + ], + "score": 0.86, + "content": "i = 1 , 2 , \\dots", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 146, + 119 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 147, + 106, + 160, + 117 + ], + "score": 0.88, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 105, + 399, + 119 + ], + "score": 1.0, + "content": "may have eigenvalues with multiplicities larger than 1 and", + "type": "text" + }, + { + "bbox": [ + 399, + 107, + 410, + 117 + ], + "score": 0.82, + "content": "\\lambda _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 105, + 413, + 119 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 414, + 107, + 438, + 117 + ], + "score": 0.79, + "content": "i \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "are not distinct.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 214, + 129 + ], + "score": 1.0, + "content": "Therefore for any integer", + "type": "text" + }, + { + "bbox": [ + 214, + 117, + 221, + 127 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 115, + 270, + 129 + ], + "score": 1.0, + "content": ", we define", + "type": "text" + }, + { + "bbox": [ + 270, + 119, + 281, + 128 + ], + "score": 0.84, + "content": "r _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 115, + 463, + 129 + ], + "score": 1.0, + "content": "as the sum of the multiplicities of the first", + "type": "text" + }, + { + "bbox": [ + 464, + 117, + 471, + 127 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "distinct", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 166, + 141 + ], + "score": 1.0, + "content": "eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 167, + 128, + 179, + 139 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 127, + 213, + 141 + ], + "score": 1.0, + "content": ". Define", + "type": "text" + }, + { + "bbox": [ + 213, + 128, + 299, + 140 + ], + "score": 0.89, + "content": "\\mathbf { V } _ { r _ { k } } = \\left( \\mathbf { v } _ { 1 } , \\ldots , \\mathbf { v } _ { r _ { k } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 127, + 361, + 141 + ], + "score": 1.0, + "content": ". By definition,", + "type": "text" + }, + { + "bbox": [ + 361, + 129, + 371, + 139 + ], + "score": 0.56, + "content": "\\mathbf { v } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 127, + 375, + 141 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 375, + 128, + 407, + 140 + ], + "score": 0.78, + "content": "i \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "are rescaled restrictions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 217, + 152 + ], + "score": 1.0, + "content": "of orthonormal functions in", + "type": "text" + }, + { + "bbox": [ + 218, + 140, + 248, + 152 + ], + "score": 0.91, + "content": "L _ { \\tau } ^ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 138, + 505, + 152 + ], + "score": 1.0, + "content": "on the training examples. Therefore we can expect them to form", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 150, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 104, + 150, + 316, + 164 + ], + "score": 1.0, + "content": "a set of almost orthonomal bases in the vector space", + "type": "text" + }, + { + "bbox": [ + 317, + 151, + 330, + 161 + ], + "score": 0.86, + "content": "\\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 150, + 506, + 164 + ], + "score": 1.0, + "content": ". The following lemma follows by standard", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 162, + 207, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 207, + 174 + ], + "score": 1.0, + "content": "concentration inequality.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 105, + 175, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 104, + 173, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 104, + 173, + 215, + 190 + ], + "score": 1.0, + "content": "Lemma 3.1. Suppose that", + "type": "text" + }, + { + "bbox": [ + 215, + 176, + 269, + 188 + ], + "score": 0.93, + "content": "| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 173, + 296, + 190 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 296, + 175, + 324, + 186 + ], + "score": 0.9, + "content": "\\mathbf { x } \\in \\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 173, + 342, + 190 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 342, + 176, + 374, + 188 + ], + "score": 0.92, + "content": "i \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 173, + 411, + 190 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 411, + 176, + 435, + 186 + ], + "score": 0.89, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 173, + 505, + 190 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 164, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 137, + 198 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 137, + 187, + 160, + 197 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 186, + 164, + 198 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 200, + 392, + 217 + ], + "lines": [ + { + "bbox": [ + 217, + 200, + 392, + 217 + ], + "spans": [ + { + "bbox": [ + 217, + 200, + 392, + 217 + ], + "score": 0.89, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { \\operatorname* { m a x } } \\leqslant C M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n } ,", + "type": "interline_equation", + "image_path": "0fe6bb516f706d4fdfcaff27a06116d620b9ad1365b8a859543d6c5e38cb71f2.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 217, + 200, + 392, + 217 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 218, + 237, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 239, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 133, + 231 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 219, + 142, + 228 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 217, + 239, + 231 + ], + "score": 1.0, + "content": "is an absolute constant.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 104, + 235, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 139, + 253 + ], + "score": 1.0, + "content": "Denote", + "type": "text" + }, + { + "bbox": [ + 139, + 237, + 221, + 250 + ], + "score": 0.93, + "content": "\\mathbf { y } ~ = ~ ( y _ { 1 } , \\ldots , y _ { n } ) ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 235, + 243, + 253 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 237, + 408, + 250 + ], + "score": 0.87, + "content": "\\widehat { \\mathbf { y } } ^ { ( t ) } ~ = ~ \\theta ~ \\cdot ~ \\left( f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { n } ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 235, + 414, + 253 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 414, + 239, + 473, + 250 + ], + "score": 0.86, + "content": "t ~ = ~ 0 , \\ldots , T", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 235, + 506, + 253 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 249, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 249, + 295, + 266 + ], + "score": 1.0, + "content": "Lemma 3.1 shows that the convergence rate of", + "type": "text" + }, + { + "bbox": [ + 296, + 250, + 368, + 263 + ], + "score": 0.92, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 249, + 506, + 266 + ], + "score": 1.0, + "content": "roughly represents the speed gra-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 442, + 275 + ], + "score": 1.0, + "content": "dient descent learns the components of the target function corresponding to the first", + "type": "text" + }, + { + "bbox": [ + 442, + 263, + 452, + 273 + ], + "score": 0.85, + "content": "r _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "eigenvalues.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 420, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 343, + 288 + ], + "score": 1.0, + "content": "The following theorem gives the convergence guarantee of", + "type": "text" + }, + { + "bbox": [ + 343, + 273, + 416, + 287 + ], + "score": 0.93, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 273, + 420, + 288 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 204, + 303 + ], + "score": 1.0, + "content": "Theorem 3.2. Suppose", + "type": "text" + }, + { + "bbox": [ + 205, + 289, + 261, + 302 + ], + "score": 0.92, + "content": "| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant M", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 287, + 277, + 303 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 278, + 289, + 312, + 302 + ], + "score": 0.92, + "content": "j \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 287, + 331, + 303 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 331, + 289, + 361, + 300 + ], + "score": 0.9, + "content": "\\mathbf { x } \\in \\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 287, + 401, + 303 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 401, + 290, + 436, + 301 + ], + "score": 0.92, + "content": "\\epsilon , \\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 287, + 485, + 303 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 485, + 290, + 492, + 300 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 287, + 506, + 303 + ], + "score": 1.0, + "content": ", if", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 289, + 316 + ], + "score": 0.9, + "content": "n \\geqslant \\widetilde { \\Omega } ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } , M ^ { 4 } r _ { k } ^ { 2 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 300, + 293, + 318 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 294, + 302, + 404, + 316 + ], + "score": 0.91, + "content": "m \\geqslant \\widetilde { \\Omega } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 300, + 506, + 318 + ], + "score": 1.0, + "content": ", then with probability at", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 378, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 127, + 331 + ], + "score": 1.0, + "content": "least", + "type": "text" + }, + { + "bbox": [ + 127, + 318, + 150, + 328 + ], + "score": 0.82, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 315, + 226, + 331 + ], + "score": 1.0, + "content": ", Algorithm 1 with", + "type": "text" + }, + { + "bbox": [ + 226, + 316, + 297, + 330 + ], + "score": 0.91, + "content": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 315, + 302, + 331 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 302, + 316, + 341, + 330 + ], + "score": 0.91, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 315, + 378, + 331 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 332, + 438, + 348 + ], + "lines": [ + { + "bbox": [ + 173, + 332, + 438, + 348 + ], + "spans": [ + { + "bbox": [ + 173, + 332, + 438, + 348 + ], + "score": 0.86, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon .", + "type": "interline_equation", + "image_path": "aa0502d3b26d2cd35bc983232428a7e918076766a0a9baf9f51ce7b11fe0222e.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 173, + 332, + 438, + 348 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 350, + 506, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "Remark 3.3. Theorem 3.2 theoretically reveals the spectral bias of deep learning. Specifically, as", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "long as the network is wide enough and the sample size is large enough, gradient descent first learns", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "the target function along the the eigendirections of neural tangent kernel with larger eigenvalues, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "then learns the rest components corresponding to smaller eigenvalues. Moreover, by showing that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "learning the components corresponding to larger eigenvalues can be done with smaller sample size", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "and narrower networks, our theory pushes the study of neural networks in the NTK regime towards", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "a more practical setting. For these reasons, we believe that Theorem 3.2 to certain extent explains", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 428, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 438 + ], + "score": 1.0, + "content": "the empirical observations given in Rahaman et al. (2018), and demonstrates that the difficulty of a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "function to be learned by neural network can be characterized in the eigenspace of neural tangent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "kernel: if the target function has a component corresponding to a small eigenvalue of neural tan-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "gent kernel, then learning this function up to good accuracy takes longer time, and requires more", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 471, + 233, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 233, + 482 + ], + "score": 1.0, + "content": "examples and a wider network.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 107, + 492, + 486, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 486, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 486, + 504 + ], + "score": 1.0, + "content": "3.2 SPECTRAL ANALYSIS OF NEURAL TANGENT KERNEL FOR UNIFORM DISTRIBUTION", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "After presenting a general theorem (without assumptions on data distribution) in the previous sub-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 524, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 504, + 536 + ], + "score": 1.0, + "content": "section, we now study the case when the data inputs are uniformly distributed over the unit sphere.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 535, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 548 + ], + "score": 1.0, + "content": "We present our results (an extension of Proposition 5 in Bietti and Mairal (2019)) of spectral analy-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "sis of neural tangent kernel. We show Mercer decomposition of neural tangent kernel for two-layer", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "setting. We give explicit expression of eigenvalues and show orders of eigenvalues in both cases", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 568, + 203, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 130, + 580 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 568, + 156, + 579 + ], + "score": 0.9, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 568, + 174, + 580 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 568, + 199, + 578 + ], + "score": 0.89, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 568, + 203, + 580 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 200, + 596 + ], + "score": 1.0, + "content": "Theorem 3.4. For any", + "type": "text" + }, + { + "bbox": [ + 200, + 581, + 276, + 593 + ], + "score": 0.89, + "content": "\\mathbf { x } , \\mathbf { x } ^ { \\prime } \\in \\mathbb { S } ^ { d } \\subset \\mathbb { R } ^ { d + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 579, + 506, + 596 + ], + "score": 1.0, + "content": ", we have the Mercer decomposition of the neural tangent", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 590, + 210, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 134, + 606 + ], + "score": 1.0, + "content": "kernel", + "type": "text" + }, + { + "bbox": [ + 134, + 592, + 204, + 604 + ], + "score": 0.92, + "content": "\\kappa : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\xrightarrow { } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 590, + 210, + 606 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "interline_equation", + "bbox": [ + 215, + 607, + 395, + 642 + ], + "lines": [ + { + "bbox": [ + 215, + 607, + 395, + 642 + ], + "spans": [ + { + "bbox": [ + 215, + 607, + 395, + 642 + ], + "score": 0.94, + "content": "\\kappa \\left( \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\right) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N \\left( d , k \\right) } Y _ { k , j } \\left( \\mathbf { x } \\right) Y _ { k , j } \\left( \\mathbf { x } ^ { \\prime } \\right) ,", + "type": "interline_equation", + "image_path": "135846f45981f6060733fed8b71d51a76f1ee29fe8fc71cc1aabc8fba3e89be2.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 607, + 395, + 624.5 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 215, + 624.5, + 395, + 642.0 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 645, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 504, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 133, + 658 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 646, + 151, + 658 + ], + "score": 0.91, + "content": "Y _ { k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 645, + 165, + 658 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 166, + 645, + 245, + 657 + ], + "score": 0.91, + "content": "j = 1 , \\cdots , N ( d , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 645, + 465, + 658 + ], + "score": 1.0, + "content": "are linearly independent spherical harmonics of degree ˘", + "type": "text" + }, + { + "bbox": [ + 465, + 646, + 472, + 655 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 645, + 483, + 658 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 483, + 646, + 504, + 656 + ], + "score": 0.89, + "content": "d + 1", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 397, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 278, + 672 + ], + "score": 1.0, + "content": "variables with N pd, kq “ 2k`d´1 k`d´2", + "type": "text" + }, + { + "bbox": [ + 276, + 657, + 333, + 672 + ], + "score": 1.0, + "content": "and orders of", + "type": "text" + }, + { + "bbox": [ + 333, + 660, + 345, + 670 + ], + "score": 0.87, + "content": "\\mu _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 657, + 397, + 672 + ], + "score": 1.0, + "content": "are given by", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 675, + 514, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 675, + 266, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 266, + 688 + ], + "score": 1.0, + "content": "µ0 “ µ1 “ Ωp1q, µk “ 0, k “ 2j ` 1,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 103, + 683, + 513, + 708 + ], + "spans": [ + { + "bbox": [ + 103, + 687, + 111, + 708 + ], + "score": 1.0, + "content": "µ", + "type": "text" + }, + { + "bbox": [ + 111, + 683, + 513, + 705 + ], + "score": 0.78, + "content": "\\begin{array} { r l } & { \\mathbf { \\Phi } ^ { \\circ _ { 0 } } \\quad \\mathsf { \\Lambda } ^ { \\circ _ { 1 } } \\quad \\cdots \\mathbf { \\Phi } ^ { \\circ _ { k } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } , \\mathsf { w } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } \\quad \\bullet _ { k } , \\mathsf { s } } , } \\\\ & { \\boldsymbol { \\iota } _ { k } = \\Omega \\left( \\operatorname* { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) , k = 2 j , } \\end{array}", + "type": "inline_equation", + "image_path": "764077198e1da67ebfa2dce2d00cb57fae510fcfc669868444804ba085727982.jpg" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 707, + 504, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 134, + 724 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 709, + 167, + 721 + ], + "score": 0.91, + "content": "j \\in \\mathbb { N } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 707, + 285, + 724 + ], + "score": 1.0, + "content": ". More specifically, we have", + "type": "text" + }, + { + "bbox": [ + 286, + 708, + 357, + 723 + ], + "score": 0.93, + "content": "\\mu _ { k } = \\Omega \\left( k ^ { - d - 1 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 707, + 384, + 724 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 384, + 709, + 412, + 720 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 707, + 432, + 724 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 708, + 504, + 723 + ], + "score": 0.92, + "content": "\\mu _ { k } = \\Omega \\left( d ^ { - k + 1 } \\right)", + "type": "inline_equation" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 718, + 224, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 130, + 734 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 721, + 156, + 731 + ], + "score": 0.77, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 718, + 160, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 160, + 721, + 219, + 732 + ], + "score": 0.81, + "content": "k = 2 , 4 , 6 , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 718, + 224, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 123, + 96 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 82, + 155, + 95 + ], + "score": 0.93, + "content": "\\{ \\lambda _ { i } \\} _ { i \\geqslant 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 81, + 177, + 96 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 177, + 83, + 239, + 94 + ], + "score": 0.91, + "content": "\\lambda _ { 1 } \\geqslant \\lambda _ { 2 } \\geqslant \\cdots", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 81, + 393, + 96 + ], + "score": 1.0, + "content": "be the strictly positive eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 394, + 83, + 406, + 93 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 81, + 428, + 96 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 428, + 82, + 491, + 95 + ], + "score": 0.9, + "content": "\\phi _ { 1 } ( \\cdot ) , \\phi _ { 2 } ( \\cdot ) , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 503, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 309, + 108 + ], + "score": 1.0, + "content": "the corresponding orthonormal eigenfunctions. Set", + "type": "text" + }, + { + "bbox": [ + 310, + 93, + 449, + 107 + ], + "score": 0.89, + "content": "\\mathbf v _ { i } = n ^ { - 1 / 2 } ( \\phi _ { i } ( { \\mathbf { x } } _ { 1 } ) , \\dots , \\phi _ { i } ( { \\mathbf { x } } _ { n } ) ) ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 92, + 452, + 108 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 453, + 95, + 503, + 106 + ], + "score": 0.86, + "content": "i = 1 , 2 , \\dots", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 146, + 119 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 147, + 106, + 160, + 117 + ], + "score": 0.88, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 105, + 399, + 119 + ], + "score": 1.0, + "content": "may have eigenvalues with multiplicities larger than 1 and", + "type": "text" + }, + { + "bbox": [ + 399, + 107, + 410, + 117 + ], + "score": 0.82, + "content": "\\lambda _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 105, + 413, + 119 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 414, + 107, + 438, + 117 + ], + "score": 0.79, + "content": "i \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "are not distinct.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 214, + 129 + ], + "score": 1.0, + "content": "Therefore for any integer", + "type": "text" + }, + { + "bbox": [ + 214, + 117, + 221, + 127 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 115, + 270, + 129 + ], + "score": 1.0, + "content": ", we define", + "type": "text" + }, + { + "bbox": [ + 270, + 119, + 281, + 128 + ], + "score": 0.84, + "content": "r _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 115, + 463, + 129 + ], + "score": 1.0, + "content": "as the sum of the multiplicities of the first", + "type": "text" + }, + { + "bbox": [ + 464, + 117, + 471, + 127 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "distinct", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 166, + 141 + ], + "score": 1.0, + "content": "eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 167, + 128, + 179, + 139 + ], + "score": 0.89, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 127, + 213, + 141 + ], + "score": 1.0, + "content": ". Define", + "type": "text" + }, + { + "bbox": [ + 213, + 128, + 299, + 140 + ], + "score": 0.89, + "content": "\\mathbf { V } _ { r _ { k } } = \\left( \\mathbf { v } _ { 1 } , \\ldots , \\mathbf { v } _ { r _ { k } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 127, + 361, + 141 + ], + "score": 1.0, + "content": ". By definition,", + "type": "text" + }, + { + "bbox": [ + 361, + 129, + 371, + 139 + ], + "score": 0.56, + "content": "\\mathbf { v } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 127, + 375, + 141 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 375, + 128, + 407, + 140 + ], + "score": 0.78, + "content": "i \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "are rescaled restrictions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 217, + 152 + ], + "score": 1.0, + "content": "of orthonormal functions in", + "type": "text" + }, + { + "bbox": [ + 218, + 140, + 248, + 152 + ], + "score": 0.91, + "content": "L _ { \\tau } ^ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 138, + 505, + 152 + ], + "score": 1.0, + "content": "on the training examples. Therefore we can expect them to form", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 150, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 104, + 150, + 316, + 164 + ], + "score": 1.0, + "content": "a set of almost orthonomal bases in the vector space", + "type": "text" + }, + { + "bbox": [ + 317, + 151, + 330, + 161 + ], + "score": 0.86, + "content": "\\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 150, + 506, + 164 + ], + "score": 1.0, + "content": ". The following lemma follows by standard", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 162, + 207, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 207, + 174 + ], + "score": 1.0, + "content": "concentration inequality.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 81, + 506, + 174 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 175, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 104, + 173, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 104, + 173, + 215, + 190 + ], + "score": 1.0, + "content": "Lemma 3.1. Suppose that", + "type": "text" + }, + { + "bbox": [ + 215, + 176, + 269, + 188 + ], + "score": 0.93, + "content": "| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 173, + 296, + 190 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 296, + 175, + 324, + 186 + ], + "score": 0.9, + "content": "\\mathbf { x } \\in \\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 173, + 342, + 190 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 342, + 176, + 374, + 188 + ], + "score": 0.92, + "content": "i \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 173, + 411, + 190 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 411, + 176, + 435, + 186 + ], + "score": 0.89, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 173, + 505, + 190 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 164, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 137, + 198 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 137, + 187, + 160, + 197 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 186, + 164, + 198 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 173, + 505, + 198 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 200, + 392, + 217 + ], + "lines": [ + { + "bbox": [ + 217, + 200, + 392, + 217 + ], + "spans": [ + { + "bbox": [ + 217, + 200, + 392, + 217 + ], + "score": 0.89, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { \\operatorname* { m a x } } \\leqslant C M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n } ,", + "type": "interline_equation", + "image_path": "0fe6bb516f706d4fdfcaff27a06116d620b9ad1365b8a859543d6c5e38cb71f2.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 217, + 200, + 392, + 217 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 218, + 237, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 239, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 133, + 231 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 219, + 142, + 228 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 217, + 239, + 231 + ], + "score": 1.0, + "content": "is an absolute constant.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 217, + 239, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 104, + 235, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 139, + 253 + ], + "score": 1.0, + "content": "Denote", + "type": "text" + }, + { + "bbox": [ + 139, + 237, + 221, + 250 + ], + "score": 0.93, + "content": "\\mathbf { y } ~ = ~ ( y _ { 1 } , \\ldots , y _ { n } ) ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 235, + 243, + 253 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 237, + 408, + 250 + ], + "score": 0.87, + "content": "\\widehat { \\mathbf { y } } ^ { ( t ) } ~ = ~ \\theta ~ \\cdot ~ \\left( f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { n } ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 235, + 414, + 253 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 414, + 239, + 473, + 250 + ], + "score": 0.86, + "content": "t ~ = ~ 0 , \\ldots , T", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 235, + 506, + 253 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 249, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 249, + 295, + 266 + ], + "score": 1.0, + "content": "Lemma 3.1 shows that the convergence rate of", + "type": "text" + }, + { + "bbox": [ + 296, + 250, + 368, + 263 + ], + "score": 0.92, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 249, + 506, + 266 + ], + "score": 1.0, + "content": "roughly represents the speed gra-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 442, + 275 + ], + "score": 1.0, + "content": "dient descent learns the components of the target function corresponding to the first", + "type": "text" + }, + { + "bbox": [ + 442, + 263, + 452, + 273 + ], + "score": 0.85, + "content": "r _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "eigenvalues.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 420, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 343, + 288 + ], + "score": 1.0, + "content": "The following theorem gives the convergence guarantee of", + "type": "text" + }, + { + "bbox": [ + 343, + 273, + 416, + 287 + ], + "score": 0.93, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 273, + 420, + 288 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 235, + 506, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 204, + 303 + ], + "score": 1.0, + "content": "Theorem 3.2. Suppose", + "type": "text" + }, + { + "bbox": [ + 205, + 289, + 261, + 302 + ], + "score": 0.92, + "content": "| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant M", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 287, + 277, + 303 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 278, + 289, + 312, + 302 + ], + "score": 0.92, + "content": "j \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 287, + 331, + 303 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 331, + 289, + 361, + 300 + ], + "score": 0.9, + "content": "\\mathbf { x } \\in \\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 287, + 401, + 303 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 401, + 290, + 436, + 301 + ], + "score": 0.92, + "content": "\\epsilon , \\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 287, + 485, + 303 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 485, + 290, + 492, + 300 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 287, + 506, + 303 + ], + "score": 1.0, + "content": ", if", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 289, + 316 + ], + "score": 0.9, + "content": "n \\geqslant \\widetilde { \\Omega } ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } , M ^ { 4 } r _ { k } ^ { 2 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 300, + 293, + 318 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 294, + 302, + 404, + 316 + ], + "score": 0.91, + "content": "m \\geqslant \\widetilde { \\Omega } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 300, + 506, + 318 + ], + "score": 1.0, + "content": ", then with probability at", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 378, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 127, + 331 + ], + "score": 1.0, + "content": "least", + "type": "text" + }, + { + "bbox": [ + 127, + 318, + 150, + 328 + ], + "score": 0.82, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 315, + 226, + 331 + ], + "score": 1.0, + "content": ", Algorithm 1 with", + "type": "text" + }, + { + "bbox": [ + 226, + 316, + 297, + 330 + ], + "score": 0.91, + "content": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 315, + 302, + 331 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 302, + 316, + 341, + 330 + ], + "score": 0.91, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 315, + 378, + 331 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 287, + 506, + 331 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 332, + 438, + 348 + ], + "lines": [ + { + "bbox": [ + 173, + 332, + 438, + 348 + ], + "spans": [ + { + "bbox": [ + 173, + 332, + 438, + 348 + ], + "score": 0.86, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon .", + "type": "interline_equation", + "image_path": "aa0502d3b26d2cd35bc983232428a7e918076766a0a9baf9f51ce7b11fe0222e.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 173, + 332, + 438, + 348 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 350, + 506, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "Remark 3.3. Theorem 3.2 theoretically reveals the spectral bias of deep learning. Specifically, as", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "long as the network is wide enough and the sample size is large enough, gradient descent first learns", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "the target function along the the eigendirections of neural tangent kernel with larger eigenvalues, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "then learns the rest components corresponding to smaller eigenvalues. Moreover, by showing that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "learning the components corresponding to larger eigenvalues can be done with smaller sample size", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "and narrower networks, our theory pushes the study of neural networks in the NTK regime towards", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "a more practical setting. For these reasons, we believe that Theorem 3.2 to certain extent explains", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 428, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 438 + ], + "score": 1.0, + "content": "the empirical observations given in Rahaman et al. (2018), and demonstrates that the difficulty of a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "function to be learned by neural network can be characterized in the eigenspace of neural tangent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "kernel: if the target function has a component corresponding to a small eigenvalue of neural tan-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "gent kernel, then learning this function up to good accuracy takes longer time, and requires more", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 471, + 233, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 233, + 482 + ], + "score": 1.0, + "content": "examples and a wider network.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 349, + 506, + 482 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 492, + 486, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 486, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 486, + 504 + ], + "score": 1.0, + "content": "3.2 SPECTRAL ANALYSIS OF NEURAL TANGENT KERNEL FOR UNIFORM DISTRIBUTION", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "After presenting a general theorem (without assumptions on data distribution) in the previous sub-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 524, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 504, + 536 + ], + "score": 1.0, + "content": "section, we now study the case when the data inputs are uniformly distributed over the unit sphere.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 535, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 548 + ], + "score": 1.0, + "content": "We present our results (an extension of Proposition 5 in Bietti and Mairal (2019)) of spectral analy-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "sis of neural tangent kernel. We show Mercer decomposition of neural tangent kernel for two-layer", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "setting. We give explicit expression of eigenvalues and show orders of eigenvalues in both cases", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 568, + 203, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 130, + 580 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 568, + 156, + 579 + ], + "score": 0.9, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 568, + 174, + 580 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 568, + 199, + 578 + ], + "score": 0.89, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 568, + 203, + 580 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 513, + 505, + 580 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 200, + 596 + ], + "score": 1.0, + "content": "Theorem 3.4. For any", + "type": "text" + }, + { + "bbox": [ + 200, + 581, + 276, + 593 + ], + "score": 0.89, + "content": "\\mathbf { x } , \\mathbf { x } ^ { \\prime } \\in \\mathbb { S } ^ { d } \\subset \\mathbb { R } ^ { d + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 579, + 506, + 596 + ], + "score": 1.0, + "content": ", we have the Mercer decomposition of the neural tangent", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 590, + 210, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 134, + 606 + ], + "score": 1.0, + "content": "kernel", + "type": "text" + }, + { + "bbox": [ + 134, + 592, + 204, + 604 + ], + "score": 0.92, + "content": "\\kappa : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\xrightarrow { } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 590, + 210, + 606 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 579, + 506, + 606 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 215, + 607, + 395, + 642 + ], + "lines": [ + { + "bbox": [ + 215, + 607, + 395, + 642 + ], + "spans": [ + { + "bbox": [ + 215, + 607, + 395, + 642 + ], + "score": 0.94, + "content": "\\kappa \\left( \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\right) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N \\left( d , k \\right) } Y _ { k , j } \\left( \\mathbf { x } \\right) Y _ { k , j } \\left( \\mathbf { x } ^ { \\prime } \\right) ,", + "type": "interline_equation", + "image_path": "135846f45981f6060733fed8b71d51a76f1ee29fe8fc71cc1aabc8fba3e89be2.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 607, + 395, + 624.5 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 215, + 624.5, + 395, + 642.0 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 645, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 504, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 133, + 658 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 646, + 151, + 658 + ], + "score": 0.91, + "content": "Y _ { k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 645, + 165, + 658 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 166, + 645, + 245, + 657 + ], + "score": 0.91, + "content": "j = 1 , \\cdots , N ( d , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 645, + 465, + 658 + ], + "score": 1.0, + "content": "are linearly independent spherical harmonics of degree ˘", + "type": "text" + }, + { + "bbox": [ + 465, + 646, + 472, + 655 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 645, + 483, + 658 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 483, + 646, + 504, + 656 + ], + "score": 0.89, + "content": "d + 1", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 397, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 278, + 672 + ], + "score": 1.0, + "content": "variables with N pd, kq “ 2k`d´1 k`d´2", + "type": "text" + }, + { + "bbox": [ + 276, + 657, + 333, + 672 + ], + "score": 1.0, + "content": "and orders of", + "type": "text" + }, + { + "bbox": [ + 333, + 660, + 345, + 670 + ], + "score": 0.87, + "content": "\\mu _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 657, + 397, + 672 + ], + "score": 1.0, + "content": "are given by", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 645, + 504, + 672 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 675, + 514, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 675, + 266, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 266, + 688 + ], + "score": 1.0, + "content": "µ0 “ µ1 “ Ωp1q, µk “ 0, k “ 2j ` 1,", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 683, + 513, + 708 + ], + "spans": [ + { + "bbox": [ + 103, + 687, + 111, + 708 + ], + "score": 1.0, + "content": "µ", + "type": "text" + }, + { + "bbox": [ + 111, + 683, + 513, + 705 + ], + "score": 0.78, + "content": "\\begin{array} { r l } & { \\mathbf { \\Phi } ^ { \\circ _ { 0 } } \\quad \\mathsf { \\Lambda } ^ { \\circ _ { 1 } } \\quad \\cdots \\mathbf { \\Phi } ^ { \\circ _ { k } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } , \\mathsf { w } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } \\quad \\bullet _ { k } , \\mathsf { s } } , } \\\\ & { \\boldsymbol { \\iota } _ { k } = \\Omega \\left( \\operatorname* { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) , k = 2 j , } \\end{array}", + "type": "inline_equation", + "image_path": "764077198e1da67ebfa2dce2d00cb57fae510fcfc669868444804ba085727982.jpg" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 707, + 504, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 134, + 724 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 709, + 167, + 721 + ], + "score": 0.91, + "content": "j \\in \\mathbb { N } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 707, + 285, + 724 + ], + "score": 1.0, + "content": ". More specifically, we have", + "type": "text" + }, + { + "bbox": [ + 286, + 708, + 357, + 723 + ], + "score": 0.93, + "content": "\\mu _ { k } = \\Omega \\left( k ^ { - d - 1 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 707, + 384, + 724 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 384, + 709, + 412, + 720 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 707, + 432, + 724 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 708, + 504, + 723 + ], + "score": 0.92, + "content": "\\mu _ { k } = \\Omega \\left( d ^ { - k + 1 } \\right)", + "type": "inline_equation" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 718, + 224, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 130, + 734 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 721, + 156, + 731 + ], + "score": 0.77, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 718, + 160, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 160, + 721, + 219, + 732 + ], + "score": 0.81, + "content": "k = 2 , 4 , 6 , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 718, + 224, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 48, + "is_list_end_line": true + } + ], + "index": 46.5, + "bbox_fs": [ + 103, + 675, + 513, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 324, + 95 + ], + "score": 1.0, + "content": "Remark 3.5. In the above theorem, the coefficients", + "type": "text" + }, + { + "bbox": [ + 325, + 84, + 337, + 94 + ], + "score": 0.86, + "content": "\\mu _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "are actually different eigenvalues of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 281, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 175, + 107 + ], + "score": 1.0, + "content": "integral operator", + "type": "text" + }, + { + "bbox": [ + 175, + 94, + 188, + 105 + ], + "score": 0.88, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 92, + 201, + 107 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 202, + 93, + 236, + 106 + ], + "score": 0.93, + "content": "L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 92, + 281, + 107 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 112, + 411, + 138 + ], + "lines": [ + { + "bbox": [ + 199, + 112, + 411, + 138 + ], + "spans": [ + { + "bbox": [ + 199, + 112, + 411, + 138 + ], + "score": 0.94, + "content": "L _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } ) ,", + "type": "interline_equation", + "image_path": "7930b891e545c6fba1ef7aa60a6151f14fa8437ffc5f7d2bc9ead1041f3f8d04.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 199, + 112, + 411, + 138 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 146, + 503, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 144, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 133, + 159 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 148, + 144, + 157 + ], + "score": 0.84, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 144, + 343, + 159 + ], + "score": 1.0, + "content": "is the uniform probability measure on unit sphere", + "type": "text" + }, + { + "bbox": [ + 343, + 145, + 355, + 156 + ], + "score": 0.88, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 144, + 415, + 159 + ], + "score": 1.0, + "content": ". Therefore the", + "type": "text" + }, + { + "bbox": [ + 416, + 147, + 430, + 158 + ], + "score": 0.9, + "content": "\\lambda _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 144, + 506, + 159 + ], + "score": 1.0, + "content": "in Theorem 3.2 is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 157, + 365, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 123, + 169 + ], + "score": 1.0, + "content": "just", + "type": "text" + }, + { + "bbox": [ + 123, + 159, + 145, + 169 + ], + "score": 0.87, + "content": "\\mu _ { k - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 157, + 258, + 169 + ], + "score": 1.0, + "content": "given in Theorem 3.4 when", + "type": "text" + }, + { + "bbox": [ + 259, + 158, + 269, + 168 + ], + "score": 0.86, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 157, + 365, + 169 + ], + "score": 1.0, + "content": "is uniform distribution.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "Remark 3.6. Vempala and Wilmes (2018) studied two-layer neural networks with sigmoid ac-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 342, + 196 + ], + "score": 1.0, + "content": "tivation function, and established guarantees to achieve", + "type": "text" + }, + { + "bbox": [ + 342, + 185, + 372, + 195 + ], + "score": 0.89, + "content": "\\epsilon _ { 0 } + \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "error with iteration complexity", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 193, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 217, + 208 + ], + "score": 0.91, + "content": "T = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\log { ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 193, + 386, + 209 + ], + "score": 1.0, + "content": "under the over-parameterization condition", + "type": "text" + }, + { + "bbox": [ + 387, + 195, + 501, + 209 + ], + "score": 0.91, + "content": "m = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\mathrm { p o l y } ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 193, + 506, + 209 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 133, + 220 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 208, + 146, + 219 + ], + "score": 0.88, + "content": "f ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 207, + 253, + 220 + ], + "score": 1.0, + "content": "is the target function, and", + "type": "text" + }, + { + "bbox": [ + 254, + 209, + 263, + 218 + ], + "score": 0.85, + "content": "\\epsilon _ { \\mathrm { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "is certain function approximation error. Another highly re-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 216, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 104, + 216, + 315, + 232 + ], + "score": 1.0, + "content": "lated work is Bietti and Mairal (2019), which gives", + "type": "text" + }, + { + "bbox": [ + 316, + 218, + 383, + 230 + ], + "score": 0.92, + "content": "\\mu _ { k } = \\bar { \\Omega ( k ^ { - } { } ^ { d - 1 } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 216, + 506, + 232 + ], + "score": 1.0, + "content": ". The order of eigenvalues we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 227, + 507, + 243 + ], + "spans": [ + { + "bbox": [ + 104, + 227, + 181, + 243 + ], + "score": 1.0, + "content": "present appears as", + "type": "text" + }, + { + "bbox": [ + 182, + 229, + 304, + 241 + ], + "score": 0.91, + "content": "\\mu _ { k } = \\Omega ( \\operatorname* { m a x } ( k ^ { - d - 1 } , d ^ { - k + \\bar { 1 } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 227, + 388, + 243 + ], + "score": 1.0, + "content": ". This is better when", + "type": "text" + }, + { + "bbox": [ + 389, + 230, + 414, + 240 + ], + "score": 0.89, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 227, + 507, + 243 + ], + "score": 1.0, + "content": ", which is closer to the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 176, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 176, + 254 + ], + "score": 1.0, + "content": "practical setting.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 261, + 429, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 430, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 430, + 275 + ], + "score": 1.0, + "content": "3.3 EXPLICIT CONVERGENCE RATE FOR UNIFORMLY DISTRIBUTED DATA", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 505, + 317 + ], + "lines": [ + { + "bbox": [ + 104, + 282, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 104, + 282, + 505, + 297 + ], + "score": 1.0, + "content": "In this subsection, we combine our results in the previous two subsections and give explicit conver-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 416, + 308 + ], + "score": 1.0, + "content": "gence rate for uniformly distributed data on the unit sphere. Note that the first", + "type": "text" + }, + { + "bbox": [ + 417, + 295, + 424, + 304 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "distinct eigenvalues", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 306, + 395, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 295, + 318 + ], + "score": 1.0, + "content": "of NTK have spherical harmonics up to degree", + "type": "text" + }, + { + "bbox": [ + 295, + 306, + 319, + 316 + ], + "score": 0.88, + "content": "k - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 306, + 395, + 318 + ], + "score": 1.0, + "content": "as eigenfunctions.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 505, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 224, + 336 + ], + "score": 1.0, + "content": "Corollary 3.7. Suppose that", + "type": "text" + }, + { + "bbox": [ + 225, + 321, + 250, + 332 + ], + "score": 0.89, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 318, + 317, + 336 + ], + "score": 1.0, + "content": ", and the sample", + "type": "text" + }, + { + "bbox": [ + 317, + 320, + 349, + 334 + ], + "score": 0.92, + "content": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 318, + 481, + 336 + ], + "score": 1.0, + "content": "follows the uniform distribution", + "type": "text" + }, + { + "bbox": [ + 481, + 324, + 491, + 332 + ], + "score": 0.85, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 318, + 506, + 336 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 331, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 104, + 331, + 169, + 349 + ], + "score": 1.0, + "content": "the unit sphere", + "type": "text" + }, + { + "bbox": [ + 169, + 334, + 181, + 345 + ], + "score": 0.86, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 331, + 221, + 349 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 222, + 335, + 257, + 346 + ], + "score": 0.91, + "content": "\\epsilon , \\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 331, + 306, + 349 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 307, + 335, + 313, + 345 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 331, + 327, + 349 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 327, + 333, + 477, + 347 + ], + "score": 0.9, + "content": "n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ k ^ { 2 d + 2 } , k ^ { 2 d - 2 } r _ { k } ^ { 2 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 331, + 482, + 349 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 482, + 335, + 505, + 346 + ], + "score": 0.8, + "content": "m \\geqslant", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 345, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 107, + 346, + 199, + 360 + ], + "score": 0.91, + "content": "\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , k ^ { d + 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 345, + 324, + 362 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 324, + 348, + 348, + 359 + ], + "score": 0.83, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 345, + 426, + 362 + ], + "score": 1.0, + "content": ", Algorithm 1 with", + "type": "text" + }, + { + "bbox": [ + 427, + 347, + 501, + 360 + ], + "score": 0.91, + "content": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 345, + 505, + 362 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 359, + 182, + 375 + ], + "spans": [ + { + "bbox": [ + 107, + 360, + 146, + 374 + ], + "score": 0.92, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 359, + 182, + 375 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 379, + 454, + 398 + ], + "lines": [ + { + "bbox": [ + 156, + 379, + 454, + 398 + ], + "spans": [ + { + "bbox": [ + 156, + 379, + 454, + 398 + ], + "score": 0.9, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( k ^ { - d - 1 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ,", + "type": "interline_equation", + "image_path": "27f6edf86d644addcd7dcf90f8781ebeed175b895cdf5d0ffc184a0a7a5dc6c5.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 156, + 379, + 454, + 398 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 405, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 101, + 397, + 510, + 430 + ], + "spans": [ + { + "bbox": [ + 101, + 397, + 134, + 430 + ], + "score": 1.0, + "content": "where thono", + "type": "text" + }, + { + "bbox": [ + 134, + 405, + 224, + 419 + ], + "score": 0.93, + "content": "\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 397, + 244, + 430 + ], + "score": 1.0, + "content": "and s of d", + "type": "text" + }, + { + "bbox": [ + 245, + 405, + 357, + 419 + ], + "score": 0.94, + "content": "\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 397, + 381, + 430 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 381, + 407, + 429, + 419 + ], + "score": 0.92, + "content": "\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 397, + 510, + 430 + ], + "score": 1.0, + "content": "being a set of or-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 419, + 319, + 428 + ], + "spans": [ + { + "bbox": [ + 295, + 419, + 319, + 428 + ], + "score": 0.86, + "content": "k - 1", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 505, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 229, + 448 + ], + "score": 1.0, + "content": "Corollary 3.8. Suppose that", + "type": "text" + }, + { + "bbox": [ + 229, + 433, + 261, + 444 + ], + "score": 0.9, + "content": "d \\ \\gg \\ k", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 431, + 334, + 448 + ], + "score": 1.0, + "content": ", and the sample", + "type": "text" + }, + { + "bbox": [ + 335, + 433, + 367, + 446 + ], + "score": 0.93, + "content": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 431, + 507, + 448 + ], + "score": 1.0, + "content": "follows the uniform distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 444, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 117, + 458 + ], + "score": 0.78, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 445, + 202, + 460 + ], + "score": 1.0, + "content": "on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 202, + 446, + 214, + 457 + ], + "score": 0.86, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 445, + 262, + 460 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 262, + 447, + 303, + 459 + ], + "score": 0.91, + "content": "\\epsilon , \\delta \\ > \\ 0", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 445, + 357, + 460 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 357, + 447, + 364, + 457 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 445, + 381, + 460 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 381, + 444, + 473, + 459 + ], + "score": 0.92, + "content": "n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } d ^ { 2 k - 2 } r _ { k } ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 445, + 479, + 460 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 479, + 446, + 505, + 458 + ], + "score": 0.81, + "content": "m \\geqslant", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 457, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 199, + 472 + ], + "score": 0.9, + "content": "\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , d ^ { k - 2 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 457, + 324, + 475 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 324, + 460, + 348, + 471 + ], + "score": 0.82, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 457, + 426, + 475 + ], + "score": 1.0, + "content": ", Algorithm 1 with", + "type": "text" + }, + { + "bbox": [ + 427, + 459, + 500, + 473 + ], + "score": 0.91, + "content": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 457, + 505, + 475 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 471, + 182, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 146, + 486 + ], + "score": 0.93, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 471, + 182, + 487 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 491, + 455, + 510 + ], + "lines": [ + { + "bbox": [ + 156, + 491, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 156, + 491, + 455, + 510 + ], + "score": 0.9, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( d ^ { - k + 2 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ,", + "type": "interline_equation", + "image_path": "888ef890c7e65b65d51416cf26c74d8141f2f8a4fb56aa9afcf8781c812ac20a.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 156, + 491, + 455, + 510 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 517, + 504, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 510, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 134, + 540 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 516, + 224, + 532 + ], + "score": 0.93, + "content": "\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 509, + 244, + 540 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 245, + 517, + 357, + 531 + ], + "score": 0.92, + "content": "\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 509, + 381, + 540 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 381, + 519, + 429, + 531 + ], + "score": 0.92, + "content": "\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 509, + 510, + 540 + ], + "score": 1.0, + "content": "being a set of or-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 529, + 324, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 295, + 542 + ], + "score": 1.0, + "content": "thonomal spherical harmonics of degrees up to", + "type": "text" + }, + { + "bbox": [ + 295, + 531, + 319, + 540 + ], + "score": 0.86, + "content": "k - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 529, + 324, + 542 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "Corollaries 3.7 and 3.8 further illustrate the spectral bias of neural networks by providing exact", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 167, + 574 + ], + "score": 1.0, + "content": "calculations of", + "type": "text" + }, + { + "bbox": [ + 168, + 563, + 182, + 574 + ], + "score": 0.86, + "content": "\\lambda _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 562, + 186, + 574 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 187, + 562, + 204, + 574 + ], + "score": 0.82, + "content": "\\mathbf { V } _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 562, + 223, + 574 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 223, + 563, + 235, + 572 + ], + "score": 0.78, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "in Theorem 3.2. They show that if the input distribution is uniform", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "over unit sphere, then spherical harmonics with lower degrees are learned first by over-parameterized", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 585, + 174, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 174, + 595 + ], + "score": 1.0, + "content": "neural networks.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 599, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 397, + 612 + ], + "score": 1.0, + "content": "Remark 3.9. In Corollaries 3.7 and 3.8, it shows that the conditions on", + "type": "text" + }, + { + "bbox": [ + 397, + 602, + 404, + 609 + ], + "score": 0.7, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 599, + 422, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 423, + 601, + 433, + 610 + ], + "score": 0.65, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "depend exponen-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 167, + 622 + ], + "score": 1.0, + "content": "tially on either", + "type": "text" + }, + { + "bbox": [ + 167, + 611, + 174, + 620 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 609, + 186, + 622 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 187, + 611, + 193, + 620 + ], + "score": 0.74, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 609, + 505, + 622 + ], + "score": 1.0, + "content": ". We would like to emphasize that such exponential dependency is reasonable", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 297, + 634 + ], + "score": 1.0, + "content": "and unavoidable. In our case, we can take the", + "type": "text" + }, + { + "bbox": [ + 298, + 622, + 325, + 632 + ], + "score": 0.9, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "setting as an example. The exponential de-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 157, + 646 + ], + "score": 1.0, + "content": "pendency in", + "type": "text" + }, + { + "bbox": [ + 158, + 633, + 165, + 642 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "is a natural consequence of the fact that in high dimensional space, there are a large", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "number of linearly independent polynomials even for very low degrees. It is apparently only reason-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 233, + 667 + ], + "score": 1.0, + "content": "able to expect to learn less than", + "type": "text" + }, + { + "bbox": [ + 234, + 656, + 241, + 664 + ], + "score": 0.75, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "independent components of the true function, which means that it", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 208, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 208, + 677 + ], + "score": 1.0, + "content": "is unavoidable to assume", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 681, + 497, + 714 + ], + "lines": [ + { + "bbox": [ + 112, + 681, + 497, + 714 + ], + "spans": [ + { + "bbox": [ + 112, + 681, + 497, + 714 + ], + "score": 0.94, + "content": "n \\geqslant r _ { k } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { d - 1 } } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { k ^ { \\prime } - 1 } } = \\Omega ( d ^ { k - 1 } ) .", + "type": "interline_equation", + "image_path": "6c96a4d6ed0e203c822fe2ead8f55b3f95a11507788014b118543f9201920818.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 112, + 681, + 497, + 692.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 112, + 692.0, + 497, + 703.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 112, + 703.0, + 497, + 714.0 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 410, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 410, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 309, + 734 + ], + "score": 1.0, + "content": "Similar arguments can apply to the requirement of", + "type": "text" + }, + { + "bbox": [ + 310, + 724, + 319, + 730 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 720, + 352, + 734 + ], + "score": 1.0, + "content": "and the", + "type": "text" + }, + { + "bbox": [ + 352, + 721, + 378, + 731 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 720, + 410, + 734 + ], + "score": 1.0, + "content": "setting.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 324, + 95 + ], + "score": 1.0, + "content": "Remark 3.5. In the above theorem, the coefficients", + "type": "text" + }, + { + "bbox": [ + 325, + 84, + 337, + 94 + ], + "score": 0.86, + "content": "\\mu _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "are actually different eigenvalues of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 281, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 175, + 107 + ], + "score": 1.0, + "content": "integral operator", + "type": "text" + }, + { + "bbox": [ + 175, + 94, + 188, + 105 + ], + "score": 0.88, + "content": "L _ { \\kappa }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 92, + 201, + 107 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 202, + 93, + 236, + 106 + ], + "score": 0.93, + "content": "L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 92, + 281, + 107 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 107 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 112, + 411, + 138 + ], + "lines": [ + { + "bbox": [ + 199, + 112, + 411, + 138 + ], + "spans": [ + { + "bbox": [ + 199, + 112, + 411, + 138 + ], + "score": 0.94, + "content": "L _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } ) ,", + "type": "interline_equation", + "image_path": "7930b891e545c6fba1ef7aa60a6151f14fa8437ffc5f7d2bc9ead1041f3f8d04.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 199, + 112, + 411, + 138 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 146, + 503, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 144, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 133, + 159 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 148, + 144, + 157 + ], + "score": 0.84, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 144, + 343, + 159 + ], + "score": 1.0, + "content": "is the uniform probability measure on unit sphere", + "type": "text" + }, + { + "bbox": [ + 343, + 145, + 355, + 156 + ], + "score": 0.88, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 144, + 415, + 159 + ], + "score": 1.0, + "content": ". Therefore the", + "type": "text" + }, + { + "bbox": [ + 416, + 147, + 430, + 158 + ], + "score": 0.9, + "content": "\\lambda _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 144, + 506, + 159 + ], + "score": 1.0, + "content": "in Theorem 3.2 is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 157, + 365, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 123, + 169 + ], + "score": 1.0, + "content": "just", + "type": "text" + }, + { + "bbox": [ + 123, + 159, + 145, + 169 + ], + "score": 0.87, + "content": "\\mu _ { k - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 157, + 258, + 169 + ], + "score": 1.0, + "content": "given in Theorem 3.4 when", + "type": "text" + }, + { + "bbox": [ + 259, + 158, + 269, + 168 + ], + "score": 0.86, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 157, + 365, + 169 + ], + "score": 1.0, + "content": "is uniform distribution.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 144, + 506, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "Remark 3.6. Vempala and Wilmes (2018) studied two-layer neural networks with sigmoid ac-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 342, + 196 + ], + "score": 1.0, + "content": "tivation function, and established guarantees to achieve", + "type": "text" + }, + { + "bbox": [ + 342, + 185, + 372, + 195 + ], + "score": 0.89, + "content": "\\epsilon _ { 0 } + \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "error with iteration complexity", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 193, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 217, + 208 + ], + "score": 0.91, + "content": "T = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\log { ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 193, + 386, + 209 + ], + "score": 1.0, + "content": "under the over-parameterization condition", + "type": "text" + }, + { + "bbox": [ + 387, + 195, + 501, + 209 + ], + "score": 0.91, + "content": "m = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\mathrm { p o l y } ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 193, + 506, + 209 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 133, + 220 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 208, + 146, + 219 + ], + "score": 0.88, + "content": "f ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 207, + 253, + 220 + ], + "score": 1.0, + "content": "is the target function, and", + "type": "text" + }, + { + "bbox": [ + 254, + 209, + 263, + 218 + ], + "score": 0.85, + "content": "\\epsilon _ { \\mathrm { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "is certain function approximation error. Another highly re-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 216, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 104, + 216, + 315, + 232 + ], + "score": 1.0, + "content": "lated work is Bietti and Mairal (2019), which gives", + "type": "text" + }, + { + "bbox": [ + 316, + 218, + 383, + 230 + ], + "score": 0.92, + "content": "\\mu _ { k } = \\bar { \\Omega ( k ^ { - } { } ^ { d - 1 } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 216, + 506, + 232 + ], + "score": 1.0, + "content": ". The order of eigenvalues we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 227, + 507, + 243 + ], + "spans": [ + { + "bbox": [ + 104, + 227, + 181, + 243 + ], + "score": 1.0, + "content": "present appears as", + "type": "text" + }, + { + "bbox": [ + 182, + 229, + 304, + 241 + ], + "score": 0.91, + "content": "\\mu _ { k } = \\Omega ( \\operatorname* { m a x } ( k ^ { - d - 1 } , d ^ { - k + \\bar { 1 } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 227, + 388, + 243 + ], + "score": 1.0, + "content": ". This is better when", + "type": "text" + }, + { + "bbox": [ + 389, + 230, + 414, + 240 + ], + "score": 0.89, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 227, + 507, + 243 + ], + "score": 1.0, + "content": ", which is closer to the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 176, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 176, + 254 + ], + "score": 1.0, + "content": "practical setting.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 171, + 507, + 254 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 261, + 429, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 430, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 430, + 275 + ], + "score": 1.0, + "content": "3.3 EXPLICIT CONVERGENCE RATE FOR UNIFORMLY DISTRIBUTED DATA", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 505, + 317 + ], + "lines": [ + { + "bbox": [ + 104, + 282, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 104, + 282, + 505, + 297 + ], + "score": 1.0, + "content": "In this subsection, we combine our results in the previous two subsections and give explicit conver-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 416, + 308 + ], + "score": 1.0, + "content": "gence rate for uniformly distributed data on the unit sphere. Note that the first", + "type": "text" + }, + { + "bbox": [ + 417, + 295, + 424, + 304 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "distinct eigenvalues", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 306, + 395, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 295, + 318 + ], + "score": 1.0, + "content": "of NTK have spherical harmonics up to degree", + "type": "text" + }, + { + "bbox": [ + 295, + 306, + 319, + 316 + ], + "score": 0.88, + "content": "k - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 306, + 395, + 318 + ], + "score": 1.0, + "content": "as eigenfunctions.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 282, + 506, + 318 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 505, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 224, + 336 + ], + "score": 1.0, + "content": "Corollary 3.7. Suppose that", + "type": "text" + }, + { + "bbox": [ + 225, + 321, + 250, + 332 + ], + "score": 0.89, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 318, + 317, + 336 + ], + "score": 1.0, + "content": ", and the sample", + "type": "text" + }, + { + "bbox": [ + 317, + 320, + 349, + 334 + ], + "score": 0.92, + "content": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 318, + 481, + 336 + ], + "score": 1.0, + "content": "follows the uniform distribution", + "type": "text" + }, + { + "bbox": [ + 481, + 324, + 491, + 332 + ], + "score": 0.85, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 318, + 506, + 336 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 331, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 104, + 331, + 169, + 349 + ], + "score": 1.0, + "content": "the unit sphere", + "type": "text" + }, + { + "bbox": [ + 169, + 334, + 181, + 345 + ], + "score": 0.86, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 331, + 221, + 349 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 222, + 335, + 257, + 346 + ], + "score": 0.91, + "content": "\\epsilon , \\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 331, + 306, + 349 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 307, + 335, + 313, + 345 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 331, + 327, + 349 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 327, + 333, + 477, + 347 + ], + "score": 0.9, + "content": "n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ k ^ { 2 d + 2 } , k ^ { 2 d - 2 } r _ { k } ^ { 2 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 331, + 482, + 349 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 482, + 335, + 505, + 346 + ], + "score": 0.8, + "content": "m \\geqslant", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 345, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 107, + 346, + 199, + 360 + ], + "score": 0.91, + "content": "\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , k ^ { d + 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 345, + 324, + 362 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 324, + 348, + 348, + 359 + ], + "score": 0.83, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 345, + 426, + 362 + ], + "score": 1.0, + "content": ", Algorithm 1 with", + "type": "text" + }, + { + "bbox": [ + 427, + 347, + 501, + 360 + ], + "score": 0.91, + "content": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 345, + 505, + 362 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 359, + 182, + 375 + ], + "spans": [ + { + "bbox": [ + 107, + 360, + 146, + 374 + ], + "score": 0.92, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 359, + 182, + 375 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 318, + 506, + 375 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 379, + 454, + 398 + ], + "lines": [ + { + "bbox": [ + 156, + 379, + 454, + 398 + ], + "spans": [ + { + "bbox": [ + 156, + 379, + 454, + 398 + ], + "score": 0.9, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( k ^ { - d - 1 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ,", + "type": "interline_equation", + "image_path": "27f6edf86d644addcd7dcf90f8781ebeed175b895cdf5d0ffc184a0a7a5dc6c5.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 156, + 379, + 454, + 398 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 405, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 101, + 397, + 510, + 430 + ], + "spans": [ + { + "bbox": [ + 101, + 397, + 134, + 430 + ], + "score": 1.0, + "content": "where thono", + "type": "text" + }, + { + "bbox": [ + 134, + 405, + 224, + 419 + ], + "score": 0.93, + "content": "\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 397, + 244, + 430 + ], + "score": 1.0, + "content": "and s of d", + "type": "text" + }, + { + "bbox": [ + 245, + 405, + 357, + 419 + ], + "score": 0.94, + "content": "\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 397, + 381, + 430 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 381, + 407, + 429, + 419 + ], + "score": 0.92, + "content": "\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 397, + 510, + 430 + ], + "score": 1.0, + "content": "being a set of or-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 419, + 319, + 428 + ], + "spans": [ + { + "bbox": [ + 295, + 419, + 319, + 428 + ], + "score": 0.86, + "content": "k - 1", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 101, + 397, + 510, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 505, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 229, + 448 + ], + "score": 1.0, + "content": "Corollary 3.8. Suppose that", + "type": "text" + }, + { + "bbox": [ + 229, + 433, + 261, + 444 + ], + "score": 0.9, + "content": "d \\ \\gg \\ k", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 431, + 334, + 448 + ], + "score": 1.0, + "content": ", and the sample", + "type": "text" + }, + { + "bbox": [ + 335, + 433, + 367, + 446 + ], + "score": 0.93, + "content": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 431, + 507, + 448 + ], + "score": 1.0, + "content": "follows the uniform distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 444, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 117, + 458 + ], + "score": 0.78, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 445, + 202, + 460 + ], + "score": 1.0, + "content": "on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 202, + 446, + 214, + 457 + ], + "score": 0.86, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 445, + 262, + 460 + ], + "score": 1.0, + "content": ". For any", + "type": "text" + }, + { + "bbox": [ + 262, + 447, + 303, + 459 + ], + "score": 0.91, + "content": "\\epsilon , \\delta \\ > \\ 0", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 445, + 357, + 460 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 357, + 447, + 364, + 457 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 445, + 381, + 460 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 381, + 444, + 473, + 459 + ], + "score": 0.92, + "content": "n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } d ^ { 2 k - 2 } r _ { k } ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 445, + 479, + 460 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 479, + 446, + 505, + 458 + ], + "score": 0.81, + "content": "m \\geqslant", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 457, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 199, + 472 + ], + "score": 0.9, + "content": "\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , d ^ { k - 2 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 457, + 324, + 475 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 324, + 460, + 348, + 471 + ], + "score": 0.82, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 457, + 426, + 475 + ], + "score": 1.0, + "content": ", Algorithm 1 with", + "type": "text" + }, + { + "bbox": [ + 427, + 459, + 500, + 473 + ], + "score": 0.91, + "content": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 457, + 505, + 475 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 471, + 182, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 146, + 486 + ], + "score": 0.93, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 471, + 182, + 487 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 431, + 507, + 487 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 491, + 455, + 510 + ], + "lines": [ + { + "bbox": [ + 156, + 491, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 156, + 491, + 455, + 510 + ], + "score": 0.9, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( d ^ { - k + 2 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ,", + "type": "interline_equation", + "image_path": "888ef890c7e65b65d51416cf26c74d8141f2f8a4fb56aa9afcf8781c812ac20a.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 156, + 491, + 455, + 510 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 517, + 504, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 510, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 134, + 540 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 516, + 224, + 532 + ], + "score": 0.93, + "content": "\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 509, + 244, + 540 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 245, + 517, + 357, + 531 + ], + "score": 0.92, + "content": "\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 509, + 381, + 540 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 381, + 519, + 429, + 531 + ], + "score": 0.92, + "content": "\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 509, + 510, + 540 + ], + "score": 1.0, + "content": "being a set of or-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 529, + 324, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 295, + 542 + ], + "score": 1.0, + "content": "thonomal spherical harmonics of degrees up to", + "type": "text" + }, + { + "bbox": [ + 295, + 531, + 319, + 540 + ], + "score": 0.86, + "content": "k - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 529, + 324, + 542 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 509, + 510, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "Corollaries 3.7 and 3.8 further illustrate the spectral bias of neural networks by providing exact", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 167, + 574 + ], + "score": 1.0, + "content": "calculations of", + "type": "text" + }, + { + "bbox": [ + 168, + 563, + 182, + 574 + ], + "score": 0.86, + "content": "\\lambda _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 562, + 186, + 574 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 187, + 562, + 204, + 574 + ], + "score": 0.82, + "content": "\\mathbf { V } _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 562, + 223, + 574 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 223, + 563, + 235, + 572 + ], + "score": 0.78, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "in Theorem 3.2. They show that if the input distribution is uniform", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "over unit sphere, then spherical harmonics with lower degrees are learned first by over-parameterized", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 585, + 174, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 174, + 595 + ], + "score": 1.0, + "content": "neural networks.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 550, + 505, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 599, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 397, + 612 + ], + "score": 1.0, + "content": "Remark 3.9. In Corollaries 3.7 and 3.8, it shows that the conditions on", + "type": "text" + }, + { + "bbox": [ + 397, + 602, + 404, + 609 + ], + "score": 0.7, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 599, + 422, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 423, + 601, + 433, + 610 + ], + "score": 0.65, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "depend exponen-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 167, + 622 + ], + "score": 1.0, + "content": "tially on either", + "type": "text" + }, + { + "bbox": [ + 167, + 611, + 174, + 620 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 609, + 186, + 622 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 187, + 611, + 193, + 620 + ], + "score": 0.74, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 609, + 505, + 622 + ], + "score": 1.0, + "content": ". We would like to emphasize that such exponential dependency is reasonable", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 297, + 634 + ], + "score": 1.0, + "content": "and unavoidable. In our case, we can take the", + "type": "text" + }, + { + "bbox": [ + 298, + 622, + 325, + 632 + ], + "score": 0.9, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "setting as an example. The exponential de-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 157, + 646 + ], + "score": 1.0, + "content": "pendency in", + "type": "text" + }, + { + "bbox": [ + 158, + 633, + 165, + 642 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "is a natural consequence of the fact that in high dimensional space, there are a large", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "number of linearly independent polynomials even for very low degrees. It is apparently only reason-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 233, + 667 + ], + "score": 1.0, + "content": "able to expect to learn less than", + "type": "text" + }, + { + "bbox": [ + 234, + 656, + 241, + 664 + ], + "score": 0.75, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "independent components of the true function, which means that it", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 208, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 208, + 677 + ], + "score": 1.0, + "content": "is unavoidable to assume", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 599, + 506, + 677 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 681, + 497, + 714 + ], + "lines": [ + { + "bbox": [ + 112, + 681, + 497, + 714 + ], + "spans": [ + { + "bbox": [ + 112, + 681, + 497, + 714 + ], + "score": 0.94, + "content": "n \\geqslant r _ { k } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { d - 1 } } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { k ^ { \\prime } - 1 } } = \\Omega ( d ^ { k - 1 } ) .", + "type": "interline_equation", + "image_path": "6c96a4d6ed0e203c822fe2ead8f55b3f95a11507788014b118543f9201920818.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 112, + 681, + 497, + 692.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 112, + 692.0, + 497, + 703.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 112, + 703.0, + 497, + 714.0 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 410, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 410, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 309, + 734 + ], + "score": 1.0, + "content": "Similar arguments can apply to the requirement of", + "type": "text" + }, + { + "bbox": [ + 310, + 724, + 319, + 730 + ], + "score": 0.78, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 720, + 352, + 734 + ], + "score": 1.0, + "content": "and the", + "type": "text" + }, + { + "bbox": [ + 352, + 721, + 378, + 731 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 720, + 410, + 734 + ], + "score": 1.0, + "content": "setting.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 720, + 410, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 200, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 201, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 201, + 96 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "In this section we illustrate our results by training neural networks on synthetic data. Across all", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "tasks, we train a two-layer hidden neural networks with 4096 neurons and initialize it exactly as", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "defined in the setup. The optimization method is vanilla full gradient descent. We sample 1000", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 137, + 388, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 366, + 150 + ], + "score": 1.0, + "content": "training data which is uniformly sampled from the unit sphere in", + "type": "text" + }, + { + "bbox": [ + 366, + 137, + 383, + 147 + ], + "score": 0.89, + "content": "\\mathbb { R } ^ { 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 137, + 388, + 150 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 106, + 159, + 360, + 170 + ], + "lines": [ + { + "bbox": [ + 106, + 159, + 361, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 361, + 171 + ], + "score": 1.0, + "content": "4.1 LEARNING COMBINATION OF SPHERICAL HARMONICS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 504, + 202 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "First, we show a result when the target function is exactly linear combination of spherical harmonics.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 190, + 277, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 277, + 203 + ], + "score": 1.0, + "content": "The target function is explicitly defined as", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 175, + 207, + 434, + 221 + ], + "lines": [ + { + "bbox": [ + 175, + 207, + 434, + 221 + ], + "spans": [ + { + "bbox": [ + 175, + 207, + 434, + 221 + ], + "score": 0.84, + "content": "f ^ { * } ( { \\bf x } ) = a _ { 1 } \\cdot P _ { 1 } ( \\langle \\zeta _ { 1 } , { \\bf x } \\rangle ) + a _ { 2 } \\cdot P _ { 2 } ( \\langle \\zeta _ { 2 } , { \\bf x } \\rangle ) + a _ { 4 } \\cdot P _ { 4 } ( \\langle \\zeta _ { 4 } , { \\bf x } \\rangle ) ,", + "type": "interline_equation", + "image_path": "ac2cf3be7b3317edb1a1ebaf573042ae9e38d677bfaa4a01ed7fd8d8b6497d9a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 175, + 207, + 434, + 221 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 149, + 239 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 149, + 226, + 172, + 238 + ], + "score": 0.94, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 225, + 317, + 239 + ], + "score": 1.0, + "content": "is the Gegenbauer polynomial, and", + "type": "text" + }, + { + "bbox": [ + 320, + 226, + 379, + 237 + ], + "score": 0.89, + "content": "\\zeta _ { k } , k = 1 , 2 , 4", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "are fixed vectors that are inde-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 236, + 511, + 276 + ], + "spans": [ + { + "bbox": [ + 100, + 236, + 242, + 276 + ], + "score": 1.0, + "content": "pendently generated from uniformaccording to the addition formula", + "type": "text" + }, + { + "bbox": [ + 243, + 248, + 429, + 264 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 236, + 382, + 247 + ], + "score": 0.88, + "content": "\\mathbb { R } ^ { 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 236, + 511, + 276 + ], + "score": 1.0, + "content": "riments. Note that, every normalized", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 263, + 341, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 263, + 341, + 275 + ], + "score": 0.92, + "content": "f ^ { * } ( \\mathbf { x } )", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "monics of order 1,2 and 4. The higher odd-order Gegenbauer polynomials are omitted because the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 285, + 333, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 225, + 298 + ], + "score": 1.0, + "content": "spectral analysis showed that", + "type": "text" + }, + { + "bbox": [ + 225, + 286, + 255, + 297 + ], + "score": 0.91, + "content": "\\mu _ { k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 285, + 271, + 298 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 271, + 285, + 329, + 297 + ], + "score": 0.87, + "content": "k = 3 , 5 , 7 \\dots", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 285, + 333, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 104, + 300, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 104, + 300, + 297, + 316 + ], + "score": 1.0, + "content": "Following the setting in section 3.1, we denote", + "type": "text" + }, + { + "bbox": [ + 297, + 302, + 437, + 315 + ], + "score": 0.92, + "content": "\\mathbf v _ { k } = n ^ { - 1 / 2 } ( P _ { k } ( \\mathbf x _ { 1 } ) , \\dots , P _ { k } ( \\mathbf x _ { n } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 300, + 506, + 316 + ], + "score": 1.0, + "content": ". By Lemma 3.2", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 119, + 325 + ], + "score": 0.83, + "content": "\\mathbf { v } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "’s are almost orthonormal with high probability. So we can define the (approximate) projection", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 323, + 275, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 178, + 340 + ], + "score": 1.0, + "content": "length of residual", + "type": "text" + }, + { + "bbox": [ + 179, + 325, + 194, + 336 + ], + "score": 0.88, + "content": "\\mathbf { r } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 323, + 216, + 340 + ], + "score": 1.0, + "content": "onto", + "type": "text" + }, + { + "bbox": [ + 216, + 327, + 228, + 337 + ], + "score": 0.86, + "content": "\\mathbf { v } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 323, + 257, + 340 + ], + "score": 1.0, + "content": "at step", + "type": "text" + }, + { + "bbox": [ + 258, + 327, + 263, + 336 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 323, + 275, + 340 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 342, + 335, + 358 + ], + "lines": [ + { + "bbox": [ + 275, + 342, + 335, + 358 + ], + "spans": [ + { + "bbox": [ + 275, + 342, + 335, + 358 + ], + "score": 0.9, + "content": "\\widehat { a } _ { k } = | \\mathbf { v } _ { k } ^ { \\top } \\mathbf { r } ^ { ( t ) } | ,", + "type": "interline_equation", + "image_path": "80aa5faf8d40795eb671a99a14e992f04309702eafc8773d859aeab898cd2c85.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 275, + 342, + 335, + 358 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 363, + 504, + 386 + ], + "lines": [ + { + "bbox": [ + 104, + 360, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 133, + 379 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 363, + 363, + 376 + ], + "score": 0.8, + "content": "\\mathbf { r } ^ { ( t ) } = \\left( f ^ { * } \\big ( \\mathbf { x } _ { 1 } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { 1 } \\big ) , \\ldots , f ^ { * } \\big ( \\mathbf { x } _ { n } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { n } \\big ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 360, + 381, + 379 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 381, + 364, + 419, + 376 + ], + "score": 0.92, + "content": "f _ { \\mathbf { W } ( t ) } \\left( \\mathbf { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 360, + 506, + 379 + ], + "score": 1.0, + "content": "is the neural network", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 373, + 146, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 146, + 388 + ], + "score": 1.0, + "content": "function.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 185, + 402 + ], + "score": 1.0, + "content": "Remark 4.1. Here", + "type": "text" + }, + { + "bbox": [ + 185, + 390, + 196, + 401 + ], + "score": 0.88, + "content": "\\widehat { a } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "is the projection length onto an approximate vector. In the function space, we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 399, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 255, + 415 + ], + "score": 1.0, + "content": "can also project the residual function", + "type": "text" + }, + { + "bbox": [ + 255, + 401, + 367, + 413 + ], + "score": 0.93, + "content": "r ( \\mathbf { x } ) \\bar { = } f ^ { * } ( \\mathbf { x } ) - \\bar { \\theta } \\bar { f } _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 399, + 506, + 415 + ], + "score": 1.0, + "content": "onto the orthonormal Gegenbauer", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 146, + 424 + ], + "score": 1.0, + "content": "functions", + "type": "text" + }, + { + "bbox": [ + 147, + 412, + 173, + 424 + ], + "score": 0.92, + "content": "P _ { k } ( { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 411, + 437, + 424 + ], + "score": 1.0, + "content": ". Replacing the training data with randomly sampled data points", + "type": "text" + }, + { + "bbox": [ + 438, + 413, + 449, + 423 + ], + "score": 0.86, + "content": "\\mathbf { x } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "can lead to a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "random estimate of the projection length in function space. We provide the corresponding result for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 433, + 270, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 270, + 447 + ], + "score": 1.0, + "content": "freshly sampled points in Appendix E.1.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 111, + 456, + 498, + 615 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 456, + 498, + 615 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 456, + 498, + 615 + ], + "spans": [ + { + "bbox": [ + 111, + 456, + 498, + 615 + ], + "score": 0.967, + "type": "image", + "image_path": "2771d3c8fdd44ee955473f67bdb5bc958722cdc8a8e43ad240100a6f7d39531f.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 111, + 456, + 498, + 509.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 111, + 509.0, + 498, + 562.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 111, + 562.0, + 498, + 615.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 628, + 505, + 673 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "Figure 1: Convergence curve for projection length onto different components. (a) shows the curve", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "when the target function have different component with the same scale. (b) shows the curve when", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "the higner-order components have larger scale. Both illustrate that the lower-order components are", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 662, + 351, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 351, + 675 + ], + "score": 1.0, + "content": "learned faster. Log-scale figures are shown in Appendix E.2.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + } + ], + "index": 27.75 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "The results are showned in Figure 1. It can be seen that at the beginning of training, the residual at", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 194, + 700 + ], + "score": 1.0, + "content": "the lowest frequency (", + "type": "text" + }, + { + "bbox": [ + 194, + 688, + 219, + 698 + ], + "score": 0.86, + "content": "k = 1 .", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 687, + 423, + 700 + ], + "score": 1.0, + "content": ") converges to zero first and then the second lowest", + "type": "text" + }, + { + "bbox": [ + 423, + 688, + 449, + 698 + ], + "score": 0.86, + "content": "k = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "). The highest", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "frequency component is the last one to converge. Following the setting of Rahaman et al. (2018) we", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "assign high frequencies a larger scale, expecting that larger scale will introduce a better descending", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 311, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 311, + 733 + ], + "score": 1.0, + "content": "speed. Still, the low frequencies are regressed first.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 200, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 201, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 201, + 96 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "In this section we illustrate our results by training neural networks on synthetic data. Across all", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "tasks, we train a two-layer hidden neural networks with 4096 neurons and initialize it exactly as", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "defined in the setup. The optimization method is vanilla full gradient descent. We sample 1000", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 137, + 388, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 366, + 150 + ], + "score": 1.0, + "content": "training data which is uniformly sampled from the unit sphere in", + "type": "text" + }, + { + "bbox": [ + 366, + 137, + 383, + 147 + ], + "score": 0.89, + "content": "\\mathbb { R } ^ { 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 137, + 388, + 150 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 105, + 505, + 150 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 159, + 360, + 170 + ], + "lines": [ + { + "bbox": [ + 106, + 159, + 361, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 361, + 171 + ], + "score": 1.0, + "content": "4.1 LEARNING COMBINATION OF SPHERICAL HARMONICS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 504, + 202 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "First, we show a result when the target function is exactly linear combination of spherical harmonics.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 190, + 277, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 277, + 203 + ], + "score": 1.0, + "content": "The target function is explicitly defined as", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 179, + 505, + 203 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 175, + 207, + 434, + 221 + ], + "lines": [ + { + "bbox": [ + 175, + 207, + 434, + 221 + ], + "spans": [ + { + "bbox": [ + 175, + 207, + 434, + 221 + ], + "score": 0.84, + "content": "f ^ { * } ( { \\bf x } ) = a _ { 1 } \\cdot P _ { 1 } ( \\langle \\zeta _ { 1 } , { \\bf x } \\rangle ) + a _ { 2 } \\cdot P _ { 2 } ( \\langle \\zeta _ { 2 } , { \\bf x } \\rangle ) + a _ { 4 } \\cdot P _ { 4 } ( \\langle \\zeta _ { 4 } , { \\bf x } \\rangle ) ,", + "type": "interline_equation", + "image_path": "ac2cf3be7b3317edb1a1ebaf573042ae9e38d677bfaa4a01ed7fd8d8b6497d9a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 175, + 207, + 434, + 221 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "list", + "bbox": [ + 106, + 225, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 149, + 239 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 149, + 226, + 172, + 238 + ], + "score": 0.94, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 225, + 317, + 239 + ], + "score": 1.0, + "content": "is the Gegenbauer polynomial, and", + "type": "text" + }, + { + "bbox": [ + 320, + 226, + 379, + 237 + ], + "score": 0.89, + "content": "\\zeta _ { k } , k = 1 , 2 , 4", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "are fixed vectors that are inde-", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 100, + 236, + 511, + 276 + ], + "spans": [ + { + "bbox": [ + 100, + 236, + 242, + 276 + ], + "score": 1.0, + "content": "pendently generated from uniformaccording to the addition formula", + "type": "text" + }, + { + "bbox": [ + 243, + 248, + 429, + 264 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 236, + 382, + 247 + ], + "score": 0.88, + "content": "\\mathbb { R } ^ { 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 236, + 511, + 276 + ], + "score": 1.0, + "content": "riments. Note that, every normalized", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 315, + 263, + 341, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 263, + 341, + 275 + ], + "score": 0.92, + "content": "f ^ { * } ( \\mathbf { x } )", + "type": "inline_equation" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "monics of order 1,2 and 4. The higher odd-order Gegenbauer polynomials are omitted because the", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 285, + 333, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 225, + 298 + ], + "score": 1.0, + "content": "spectral analysis showed that", + "type": "text" + }, + { + "bbox": [ + 225, + 286, + 255, + 297 + ], + "score": 0.91, + "content": "\\mu _ { k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 285, + 271, + 298 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 271, + 285, + 329, + 297 + ], + "score": 0.87, + "content": "k = 3 , 5 , 7 \\dots", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 285, + 333, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 11, + "bbox_fs": [ + 100, + 225, + 511, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 104, + 300, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 104, + 300, + 297, + 316 + ], + "score": 1.0, + "content": "Following the setting in section 3.1, we denote", + "type": "text" + }, + { + "bbox": [ + 297, + 302, + 437, + 315 + ], + "score": 0.92, + "content": "\\mathbf v _ { k } = n ^ { - 1 / 2 } ( P _ { k } ( \\mathbf x _ { 1 } ) , \\dots , P _ { k } ( \\mathbf x _ { n } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 300, + 506, + 316 + ], + "score": 1.0, + "content": ". By Lemma 3.2", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 119, + 325 + ], + "score": 0.83, + "content": "\\mathbf { v } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "’s are almost orthonormal with high probability. So we can define the (approximate) projection", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 323, + 275, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 178, + 340 + ], + "score": 1.0, + "content": "length of residual", + "type": "text" + }, + { + "bbox": [ + 179, + 325, + 194, + 336 + ], + "score": 0.88, + "content": "\\mathbf { r } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 323, + 216, + 340 + ], + "score": 1.0, + "content": "onto", + "type": "text" + }, + { + "bbox": [ + 216, + 327, + 228, + 337 + ], + "score": 0.86, + "content": "\\mathbf { v } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 323, + 257, + 340 + ], + "score": 1.0, + "content": "at step", + "type": "text" + }, + { + "bbox": [ + 258, + 327, + 263, + 336 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 323, + 275, + 340 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 300, + 506, + 340 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 342, + 335, + 358 + ], + "lines": [ + { + "bbox": [ + 275, + 342, + 335, + 358 + ], + "spans": [ + { + "bbox": [ + 275, + 342, + 335, + 358 + ], + "score": 0.9, + "content": "\\widehat { a } _ { k } = | \\mathbf { v } _ { k } ^ { \\top } \\mathbf { r } ^ { ( t ) } | ,", + "type": "interline_equation", + "image_path": "80aa5faf8d40795eb671a99a14e992f04309702eafc8773d859aeab898cd2c85.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 275, + 342, + 335, + 358 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 363, + 504, + 386 + ], + "lines": [ + { + "bbox": [ + 104, + 360, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 133, + 379 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 363, + 363, + 376 + ], + "score": 0.8, + "content": "\\mathbf { r } ^ { ( t ) } = \\left( f ^ { * } \\big ( \\mathbf { x } _ { 1 } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { 1 } \\big ) , \\ldots , f ^ { * } \\big ( \\mathbf { x } _ { n } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { n } \\big ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 360, + 381, + 379 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 381, + 364, + 419, + 376 + ], + "score": 0.92, + "content": "f _ { \\mathbf { W } ( t ) } \\left( \\mathbf { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 360, + 506, + 379 + ], + "score": 1.0, + "content": "is the neural network", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 373, + 146, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 146, + 388 + ], + "score": 1.0, + "content": "function.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 360, + 506, + 388 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 185, + 402 + ], + "score": 1.0, + "content": "Remark 4.1. Here", + "type": "text" + }, + { + "bbox": [ + 185, + 390, + 196, + 401 + ], + "score": 0.88, + "content": "\\widehat { a } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "is the projection length onto an approximate vector. In the function space, we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 399, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 255, + 415 + ], + "score": 1.0, + "content": "can also project the residual function", + "type": "text" + }, + { + "bbox": [ + 255, + 401, + 367, + 413 + ], + "score": 0.93, + "content": "r ( \\mathbf { x } ) \\bar { = } f ^ { * } ( \\mathbf { x } ) - \\bar { \\theta } \\bar { f } _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 399, + 506, + 415 + ], + "score": 1.0, + "content": "onto the orthonormal Gegenbauer", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 146, + 424 + ], + "score": 1.0, + "content": "functions", + "type": "text" + }, + { + "bbox": [ + 147, + 412, + 173, + 424 + ], + "score": 0.92, + "content": "P _ { k } ( { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 411, + 437, + 424 + ], + "score": 1.0, + "content": ". Replacing the training data with randomly sampled data points", + "type": "text" + }, + { + "bbox": [ + 438, + 413, + 449, + 423 + ], + "score": 0.86, + "content": "\\mathbf { x } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "can lead to a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "random estimate of the projection length in function space. We provide the corresponding result for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 433, + 270, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 270, + 447 + ], + "score": 1.0, + "content": "freshly sampled points in Appendix E.1.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 390, + 506, + 447 + ] + }, + { + "type": "image", + "bbox": [ + 111, + 456, + 498, + 615 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 456, + 498, + 615 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 456, + 498, + 615 + ], + "spans": [ + { + "bbox": [ + 111, + 456, + 498, + 615 + ], + "score": 0.967, + "type": "image", + "image_path": "2771d3c8fdd44ee955473f67bdb5bc958722cdc8a8e43ad240100a6f7d39531f.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 111, + 456, + 498, + 509.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 111, + 509.0, + 498, + 562.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 111, + 562.0, + 498, + 615.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 628, + 505, + 673 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "Figure 1: Convergence curve for projection length onto different components. (a) shows the curve", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "when the target function have different component with the same scale. (b) shows the curve when", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "the higner-order components have larger scale. Both illustrate that the lower-order components are", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 662, + 351, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 351, + 675 + ], + "score": 1.0, + "content": "learned faster. Log-scale figures are shown in Appendix E.2.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + } + ], + "index": 27.75 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "The results are showned in Figure 1. It can be seen that at the beginning of training, the residual at", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 194, + 700 + ], + "score": 1.0, + "content": "the lowest frequency (", + "type": "text" + }, + { + "bbox": [ + 194, + 688, + 219, + 698 + ], + "score": 0.86, + "content": "k = 1 .", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 687, + 423, + 700 + ], + "score": 1.0, + "content": ") converges to zero first and then the second lowest", + "type": "text" + }, + { + "bbox": [ + 423, + 688, + 449, + 698 + ], + "score": 0.86, + "content": "k = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "). The highest", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "frequency component is the last one to converge. Following the setting of Rahaman et al. (2018) we", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "assign high frequencies a larger scale, expecting that larger scale will introduce a better descending", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 311, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 311, + 733 + ], + "score": 1.0, + "content": "speed. Still, the low frequencies are regressed first.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 677, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 504, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 504, + 118 + ], + "score": 1.0, + "content": "Apart from the synthesized low frequency function, we also showed the dynamics of normal func-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 185, + 129 + ], + "score": 1.0, + "content": "tions’ projection to", + "type": "text" + }, + { + "bbox": [ + 185, + 117, + 211, + 129 + ], + "score": 0.92, + "content": "P _ { k } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 116, + 505, + 129 + ], + "score": 1.0, + "content": ". These functions, though in a simple form, have non-zero components in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "almost all frequencies. In this subsection we further show our results still apply when all frequenciesř ř", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 293, + 151 + ], + "score": 1.0, + "content": "exist in the target function, which is given by", + "type": "text" + }, + { + "bbox": [ + 293, + 138, + 402, + 151 + ], + "score": 0.92, + "content": "\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\cos ( a _ { i } \\langle \\zeta , \\mathbf { x } \\rangle ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 138, + 415, + 151 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 416, + 138, + 501, + 151 + ], + "score": 0.92, + "content": "\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\bar { \\langle \\zeta , \\mathbf { x } \\rangle } ^ { p _ { i } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 138, + 505, + 151 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 133, + 162 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 150, + 141, + 161 + ], + "score": 0.83, + "content": "\\zeta", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "is a fixed unit vector. The coefficients of given components are calculated in the same way", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 177, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 177, + 172 + ], + "score": 1.0, + "content": "as in Section 4.1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 111, + 179, + 499, + 336 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 179, + 499, + 336 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 179, + 499, + 336 + ], + "spans": [ + { + "bbox": [ + 111, + 179, + 499, + 336 + ], + "score": 0.965, + "type": "image", + "image_path": "edeb7fc8e44d199514a417b12ca7990246224162dede7c2aaebaca80a02339cc.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 111, + 179, + 499, + 231.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 231.33333333333334, + 499, + 283.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 283.6666666666667, + 499, + 336.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 349, + 505, + 383 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "Figure 2: Convergence curve for different components. (a) shows the curve of a trigonometric", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "function. (b) shows the curve of a polynomial with even degrees. Both exhibits similar tendency as", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 371, + 255, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 255, + 384 + ], + "score": 1.0, + "content": "combination of spherical harmonics.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "Figure 2 shows that even for arbitrarily chosen functions of simple form, the networks can still first", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "score": 1.0, + "content": "learn the low frequency components of the target function. Notice that early in training not all the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "curves may descend, we believe this is due to the unseen components’ influence on the gradient.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 422, + 438, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 438, + 434 + ], + "score": 1.0, + "content": "Again, as the training proceeds, the convergence is controlled at the predicted rate.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Remark 4.2. The reason why we only use cosine function and even polynomial is that the only odd", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 277, + 465 + ], + "score": 1.0, + "content": "basis function with non-zero eigenvalue is", + "type": "text" + }, + { + "bbox": [ + 278, + 452, + 303, + 464 + ], + "score": 0.92, + "content": "P _ { 1 } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 452, + 505, + 465 + ], + "score": 1.0, + "content": ". To show a general tendency it is better to restrict", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 462, + 290, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 290, + 476 + ], + "score": 1.0, + "content": "the target function in the even function space.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 492, + 287, + 504 + ], + "lines": [ + { + "bbox": [ + 104, + 490, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 104, + 490, + 289, + 506 + ], + "score": 1.0, + "content": "5 CONCLUSION AND DISCUSSION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 530 + ], + "score": 1.0, + "content": "In this paper, we give theoretical justification for spectral bias through a detailed analysis of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "convergence behavior of two-layer neural networks with ReLU activation function. We show that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "the convergence of gradient descent in different directions depends on the corresponding eigenval-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "ues and essentially exhibits different convergence rates. We show Mercer decomposition of neural", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "score": 1.0, + "content": "tangent kernel and give explicit order of eigenvalues of integral operator with respect to the neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 405, + 585 + ], + "score": 1.0, + "content": "tangent kernel when the data is uniformly distributed on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 405, + 572, + 416, + 582 + ], + "score": 0.87, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 571, + 506, + 585 + ], + "score": 1.0, + "content": ". Combined with the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "convergence analysis, we give the exact order of convergence rate on different directions. We also", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "score": 1.0, + "content": "conduct experiments on synthetic data to support our theoretical result.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 611, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "So far, we have considered the upper bound for convergence with respect to low frequency com-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "ponents and present comprehensive theorem to explain the spectral bias. One desired improvement", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "is to give the lower bound of convergence with respect to high frequency components, which is es-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "sential to establish tighter characterization of spectral-biased optimization. It is also interesting to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "extend our result to other training algorithms like Adam, where the analysis in Wu et al. (2019); Zhou", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "et al. (2018) might be implemented with a more careful quantification on the projection of residual", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "along different directions. Another potential improvement is to generalize the result to multi-layer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "score": 1.0, + "content": "neural networks, which might require different techniques since our analysis heavily rely on exactly", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "computing the eigenvalues of the neural tangent kernel. It is also an important direction to weaken", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "the requirement on over-parameterization, or study the spectral bias in a non-NTK regime to furthur", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 721, + 278, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 278, + 733 + ], + "score": 1.0, + "content": "close the gap between theory and practice.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 82, + 303, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 306, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 306, + 95 + ], + "score": 1.0, + "content": "4.2 LEARNING FUNCTIONS OF SIMPLE FORM", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 504, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 504, + 118 + ], + "score": 1.0, + "content": "Apart from the synthesized low frequency function, we also showed the dynamics of normal func-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 185, + 129 + ], + "score": 1.0, + "content": "tions’ projection to", + "type": "text" + }, + { + "bbox": [ + 185, + 117, + 211, + 129 + ], + "score": 0.92, + "content": "P _ { k } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 116, + 505, + 129 + ], + "score": 1.0, + "content": ". These functions, though in a simple form, have non-zero components in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "almost all frequencies. In this subsection we further show our results still apply when all frequenciesř ř", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 293, + 151 + ], + "score": 1.0, + "content": "exist in the target function, which is given by", + "type": "text" + }, + { + "bbox": [ + 293, + 138, + 402, + 151 + ], + "score": 0.92, + "content": "\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\cos ( a _ { i } \\langle \\zeta , \\mathbf { x } \\rangle ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 138, + 415, + 151 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 416, + 138, + 501, + 151 + ], + "score": 0.92, + "content": "\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\bar { \\langle \\zeta , \\mathbf { x } \\rangle } ^ { p _ { i } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 138, + 505, + 151 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 133, + 162 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 150, + 141, + 161 + ], + "score": 0.83, + "content": "\\zeta", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "is a fixed unit vector. The coefficients of given components are calculated in the same way", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 177, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 177, + 172 + ], + "score": 1.0, + "content": "as in Section 4.1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 106, + 505, + 172 + ] + }, + { + "type": "image", + "bbox": [ + 111, + 179, + 499, + 336 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 179, + 499, + 336 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 179, + 499, + 336 + ], + "spans": [ + { + "bbox": [ + 111, + 179, + 499, + 336 + ], + "score": 0.965, + "type": "image", + "image_path": "edeb7fc8e44d199514a417b12ca7990246224162dede7c2aaebaca80a02339cc.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 111, + 179, + 499, + 231.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 231.33333333333334, + 499, + 283.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 283.6666666666667, + 499, + 336.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 349, + 505, + 383 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "Figure 2: Convergence curve for different components. (a) shows the curve of a trigonometric", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "function. (b) shows the curve of a polynomial with even degrees. Both exhibits similar tendency as", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 371, + 255, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 255, + 384 + ], + "score": 1.0, + "content": "combination of spherical harmonics.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "Figure 2 shows that even for arbitrarily chosen functions of simple form, the networks can still first", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 412 + ], + "score": 1.0, + "content": "learn the low frequency components of the target function. Notice that early in training not all the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "curves may descend, we believe this is due to the unseen components’ influence on the gradient.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 422, + 438, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 438, + 434 + ], + "score": 1.0, + "content": "Again, as the training proceeds, the convergence is controlled at the predicted rate.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 388, + 506, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Remark 4.2. The reason why we only use cosine function and even polynomial is that the only odd", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 277, + 465 + ], + "score": 1.0, + "content": "basis function with non-zero eigenvalue is", + "type": "text" + }, + { + "bbox": [ + 278, + 452, + 303, + 464 + ], + "score": 0.92, + "content": "P _ { 1 } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 452, + 505, + 465 + ], + "score": 1.0, + "content": ". To show a general tendency it is better to restrict", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 462, + 290, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 290, + 476 + ], + "score": 1.0, + "content": "the target function in the even function space.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 440, + 505, + 476 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 492, + 287, + 504 + ], + "lines": [ + { + "bbox": [ + 104, + 490, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 104, + 490, + 289, + 506 + ], + "score": 1.0, + "content": "5 CONCLUSION AND DISCUSSION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 530 + ], + "score": 1.0, + "content": "In this paper, we give theoretical justification for spectral bias through a detailed analysis of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "convergence behavior of two-layer neural networks with ReLU activation function. We show that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "the convergence of gradient descent in different directions depends on the corresponding eigenval-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "ues and essentially exhibits different convergence rates. We show Mercer decomposition of neural", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "score": 1.0, + "content": "tangent kernel and give explicit order of eigenvalues of integral operator with respect to the neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 571, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 405, + 585 + ], + "score": 1.0, + "content": "tangent kernel when the data is uniformly distributed on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 405, + 572, + 416, + 582 + ], + "score": 0.87, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 571, + 506, + 585 + ], + "score": 1.0, + "content": ". Combined with the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "convergence analysis, we give the exact order of convergence rate on different directions. We also", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "score": 1.0, + "content": "conduct experiments on synthetic data to support our theoretical result.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 516, + 506, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 611, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "So far, we have considered the upper bound for convergence with respect to low frequency com-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "ponents and present comprehensive theorem to explain the spectral bias. One desired improvement", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "is to give the lower bound of convergence with respect to high frequency components, which is es-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "sential to establish tighter characterization of spectral-biased optimization. It is also interesting to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "extend our result to other training algorithms like Adam, where the analysis in Wu et al. (2019); Zhou", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "et al. (2018) might be implemented with a more careful quantification on the projection of residual", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "along different directions. Another potential improvement is to generalize the result to multi-layer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "score": 1.0, + "content": "neural networks, which might require different techniques since our analysis heavily rely on exactly", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "computing the eigenvalues of the neural tangent kernel. It is also an important direction to weaken", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "the requirement on over-parameterization, or study the spectral bias in a non-NTK regime to furthur", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 721, + 278, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 278, + 733 + ], + "score": 1.0, + "content": "close the gap between theory and practice.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 611, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 54, + 506, + 734 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 177, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 177, + 96 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "ALLEN-ZHU, Z., LI, Y. and SONG, Z. (2018). A convergence theory for deep learning via over-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 116, + 329, + 129 + ], + "spans": [ + { + "bbox": [ + 114, + 116, + 329, + 129 + ], + "score": 1.0, + "content": "parameterization. arXiv preprint arXiv:1811.03962 .", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "score": 1.0, + "content": "ANDONI, A., PANIGRAHY, R., VALIANT, G. and ZHANG, L. (2014). Learning polynomials with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 144, + 392, + 162 + ], + "spans": [ + { + "bbox": [ + 113, + 144, + 392, + 162 + ], + "score": 1.0, + "content": "neural networks. In International Conference on Machine Learning.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "score": 1.0, + "content": "ARORA, S., DU, S. S., HU, W., LI, Z., SALAKHUTDINOV, R. and WANG, R. (2019a). On exact", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 176, + 446, + 189 + ], + "spans": [ + { + "bbox": [ + 114, + 176, + 446, + 189 + ], + "score": 1.0, + "content": "computation with an infinitely wide neural net. arXiv preprint arXiv:1904.11955 .", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 195, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 209 + ], + "score": 1.0, + "content": "ARORA, S., DU, S. S., HU, W., LI, Z. and WANG, R. (2019b). Fine-grained analysis of op-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 205, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 114, + 205, + 505, + 221 + ], + "score": 1.0, + "content": "timization and generalization for overparameterized two-layer neural networks. arXiv preprint", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 217, + 199, + 229 + ], + "spans": [ + { + "bbox": [ + 114, + 217, + 199, + 229 + ], + "score": 1.0, + "content": "arXiv:1901.08584 .", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "ATKINSON, K. and HAN, W. (2012). Spherical harmonics and approximations on the unit sphere:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 247, + 375, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 375, + 259 + ], + "score": 1.0, + "content": "an introduction, vol. 2044. Springer Science & Business Media.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "score": 1.0, + "content": "BACH, F. (2017). Breaking the curse of dimensionality with convex neural networks. Journal of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 278, + 270, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 278, + 270, + 290 + ], + "score": 1.0, + "content": "Machine Learning Research 18 1–53.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "BASRI, R., JACOBS, D., KASTEN, Y. and KRITCHMAN, S. (2019). The convergence rate of neural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 307, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 114, + 307, + 480, + 321 + ], + "score": 1.0, + "content": "networks for learned functions of different frequencies. arXiv preprint arXiv:1906.00425 .", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "BIETTI, A. and MAIRAL, J. (2019). On the inductive bias of neural tangent kernels. arXiv preprint", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 338, + 199, + 350 + ], + "spans": [ + { + "bbox": [ + 114, + 338, + 199, + 350 + ], + "score": 1.0, + "content": "arXiv:1905.12173 .", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "CAO, Y. and GU, Q. (2019a). Generalization bounds of stochastic gradient descent for wide and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 367, + 347, + 381 + ], + "spans": [ + { + "bbox": [ + 114, + 367, + 347, + 381 + ], + "score": 1.0, + "content": "deep neural networks. arXiv preprint arXiv:1905.13210 .", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 385, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 402 + ], + "score": 1.0, + "content": "CAO, Y. and GU, Q. (2019b). A generalization theory of gradient descent for learning over-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 398, + 398, + 411 + ], + "spans": [ + { + "bbox": [ + 114, + 398, + 398, + 411 + ], + "score": 1.0, + "content": "parameterized deep relu networks. arXiv preprint arXiv:1902.01384 .", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "CHIZAT, L. and BACH, F. (2018). A note on lazy training in supervised differentiable programming.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 427, + 258, + 441 + ], + "spans": [ + { + "bbox": [ + 114, + 427, + 258, + 441 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1812.07956 .", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "CHO, Y. and SAUL, L. K. (2009). Kernel methods for deep learning. In Advances in neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 459, + 248, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 248, + 471 + ], + "score": 1.0, + "content": "information processing systems.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 475, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 491 + ], + "score": 1.0, + "content": "COLLOBERT, R. and WESTON, J. (2008). A unified architecture for natural language processing:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 114, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Deep neural networks with multitask learning. In Proceedings of the 25th international confer-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 499, + 253, + 512 + ], + "spans": [ + { + "bbox": [ + 114, + 499, + 253, + 512 + ], + "score": 1.0, + "content": "ence on Machine learning. ACM.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 517, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 532 + ], + "score": 1.0, + "content": "DU, S. S., LEE, J. D., LI, H., WANG, L. and ZHAI, X. (2018a). Gradient descent finds global", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 528, + 392, + 543 + ], + "spans": [ + { + "bbox": [ + 114, + 528, + 392, + 543 + ], + "score": 1.0, + "content": "minima of deep neural networks. arXiv preprint arXiv:1811.03804 .", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 562 + ], + "score": 1.0, + "content": "DU, S. S., ZHAI, X., POCZOS, B. and SINGH, A. (2018b). Gradient descent provably optimizes", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 559, + 406, + 573 + ], + "spans": [ + { + "bbox": [ + 114, + 559, + 406, + 573 + ], + "score": 1.0, + "content": "over-parameterized neural networks. arXiv preprint arXiv:1810.02054 .", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "FRYE, C. and EFTHIMIOU, C. J. (2012). Spherical harmonics in p dimensions. arXiv preprint", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 590, + 193, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 590, + 193, + 601 + ], + "score": 1.0, + "content": "arXiv:1205.3548 .", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "score": 1.0, + "content": "GUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018a). Characterizing implicit bias in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 619, + 389, + 632 + ], + "spans": [ + { + "bbox": [ + 114, + 619, + 389, + 632 + ], + "score": 1.0, + "content": "terms of optimization geometry. arXiv preprint arXiv:1802.08246 .", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "GUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018b). Implicit bias of gradient descent", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 649, + 395, + 663 + ], + "spans": [ + { + "bbox": [ + 114, + 649, + 395, + 663 + ], + "score": 1.0, + "content": "on linear convolutional networks. arXiv preprint arXiv:1806.00468 .", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "HE, K., ZHANG, X., REN, S. and SUN, J. (2015). Delving deep into rectifiers: Surpassing human-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 114, + 679, + 505, + 693 + ], + "score": 1.0, + "content": "level performance on imagenet classification. In Proceedings of the IEEE international confer-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 692, + 218, + 704 + ], + "spans": [ + { + "bbox": [ + 114, + 692, + 218, + 704 + ], + "score": 1.0, + "content": "ence on computer vision.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 708, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 506, + 725 + ], + "score": 1.0, + "content": "HE, K., ZHANG, X., REN, S. and SUN, J. (2016). Deep residual learning for image recognition.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 720, + 452, + 734 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 452, + 734 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 21.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 54, + 506, + 734 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 177, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 177, + 96 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "ALLEN-ZHU, Z., LI, Y. and SONG, Z. (2018). A convergence theory for deep learning via over-", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 116, + 329, + 129 + ], + "spans": [ + { + "bbox": [ + 114, + 116, + 329, + 129 + ], + "score": 1.0, + "content": "parameterization. arXiv preprint arXiv:1811.03962 .", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "score": 1.0, + "content": "ANDONI, A., PANIGRAHY, R., VALIANT, G. and ZHANG, L. (2014). Learning polynomials with", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 144, + 392, + 162 + ], + "spans": [ + { + "bbox": [ + 113, + 144, + 392, + 162 + ], + "score": 1.0, + "content": "neural networks. In International Conference on Machine Learning.", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 179 + ], + "score": 1.0, + "content": "ARORA, S., DU, S. S., HU, W., LI, Z., SALAKHUTDINOV, R. and WANG, R. (2019a). On exact", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 176, + 446, + 189 + ], + "spans": [ + { + "bbox": [ + 114, + 176, + 446, + 189 + ], + "score": 1.0, + "content": "computation with an infinitely wide neural net. arXiv preprint arXiv:1904.11955 .", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 195, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 209 + ], + "score": 1.0, + "content": "ARORA, S., DU, S. S., HU, W., LI, Z. and WANG, R. (2019b). Fine-grained analysis of op-", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 205, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 114, + 205, + 505, + 221 + ], + "score": 1.0, + "content": "timization and generalization for overparameterized two-layer neural networks. arXiv preprint", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 217, + 199, + 229 + ], + "spans": [ + { + "bbox": [ + 114, + 217, + 199, + 229 + ], + "score": 1.0, + "content": "arXiv:1901.08584 .", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "ATKINSON, K. and HAN, W. (2012). Spherical harmonics and approximations on the unit sphere:", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 247, + 375, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 375, + 259 + ], + "score": 1.0, + "content": "an introduction, vol. 2044. Springer Science & Business Media.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 280 + ], + "score": 1.0, + "content": "BACH, F. (2017). Breaking the curse of dimensionality with convex neural networks. Journal of", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 278, + 270, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 278, + 270, + 290 + ], + "score": 1.0, + "content": "Machine Learning Research 18 1–53.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "BASRI, R., JACOBS, D., KASTEN, Y. and KRITCHMAN, S. (2019). The convergence rate of neural", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 307, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 114, + 307, + 480, + 321 + ], + "score": 1.0, + "content": "networks for learned functions of different frequencies. arXiv preprint arXiv:1906.00425 .", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "BIETTI, A. and MAIRAL, J. (2019). On the inductive bias of neural tangent kernels. arXiv preprint", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 338, + 199, + 350 + ], + "spans": [ + { + "bbox": [ + 114, + 338, + 199, + 350 + ], + "score": 1.0, + "content": "arXiv:1905.12173 .", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "CAO, Y. and GU, Q. (2019a). Generalization bounds of stochastic gradient descent for wide and", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 367, + 347, + 381 + ], + "spans": [ + { + "bbox": [ + 114, + 367, + 347, + 381 + ], + "score": 1.0, + "content": "deep neural networks. arXiv preprint arXiv:1905.13210 .", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 385, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 402 + ], + "score": 1.0, + "content": "CAO, Y. and GU, Q. (2019b). A generalization theory of gradient descent for learning over-", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 398, + 398, + 411 + ], + "spans": [ + { + "bbox": [ + 114, + 398, + 398, + 411 + ], + "score": 1.0, + "content": "parameterized deep relu networks. arXiv preprint arXiv:1902.01384 .", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "CHIZAT, L. and BACH, F. (2018). A note on lazy training in supervised differentiable programming.", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 427, + 258, + 441 + ], + "spans": [ + { + "bbox": [ + 114, + 427, + 258, + 441 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1812.07956 .", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "CHO, Y. and SAUL, L. K. (2009). Kernel methods for deep learning. In Advances in neural", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 459, + 248, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 248, + 471 + ], + "score": 1.0, + "content": "information processing systems.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 475, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 491 + ], + "score": 1.0, + "content": "COLLOBERT, R. and WESTON, J. (2008). A unified architecture for natural language processing:", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 114, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Deep neural networks with multitask learning. In Proceedings of the 25th international confer-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 499, + 253, + 512 + ], + "spans": [ + { + "bbox": [ + 114, + 499, + 253, + 512 + ], + "score": 1.0, + "content": "ence on Machine learning. ACM.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 517, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 532 + ], + "score": 1.0, + "content": "DU, S. S., LEE, J. D., LI, H., WANG, L. and ZHAI, X. (2018a). Gradient descent finds global", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 528, + 392, + 543 + ], + "spans": [ + { + "bbox": [ + 114, + 528, + 392, + 543 + ], + "score": 1.0, + "content": "minima of deep neural networks. arXiv preprint arXiv:1811.03804 .", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 546, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 562 + ], + "score": 1.0, + "content": "DU, S. S., ZHAI, X., POCZOS, B. and SINGH, A. (2018b). Gradient descent provably optimizes", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 559, + 406, + 573 + ], + "spans": [ + { + "bbox": [ + 114, + 559, + 406, + 573 + ], + "score": 1.0, + "content": "over-parameterized neural networks. arXiv preprint arXiv:1810.02054 .", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "FRYE, C. and EFTHIMIOU, C. J. (2012). Spherical harmonics in p dimensions. arXiv preprint", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 590, + 193, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 590, + 193, + 601 + ], + "score": 1.0, + "content": "arXiv:1205.3548 .", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "score": 1.0, + "content": "GUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018a). Characterizing implicit bias in", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 619, + 389, + 632 + ], + "spans": [ + { + "bbox": [ + 114, + 619, + 389, + 632 + ], + "score": 1.0, + "content": "terms of optimization geometry. arXiv preprint arXiv:1802.08246 .", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "GUNASEKAR, S., LEE, J., SOUDRY, D. and SREBRO, N. (2018b). Implicit bias of gradient descent", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 649, + 395, + 663 + ], + "spans": [ + { + "bbox": [ + 114, + 649, + 395, + 663 + ], + "score": 1.0, + "content": "on linear convolutional networks. arXiv preprint arXiv:1806.00468 .", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "HE, K., ZHANG, X., REN, S. and SUN, J. (2015). Delving deep into rectifiers: Surpassing human-", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 114, + 679, + 505, + 693 + ], + "score": 1.0, + "content": "level performance on imagenet classification. In Proceedings of the IEEE international confer-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 692, + 218, + 704 + ], + "spans": [ + { + "bbox": [ + 114, + 692, + 218, + 704 + ], + "score": 1.0, + "content": "ence on computer vision.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 708, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 506, + 725 + ], + "score": 1.0, + "content": "HE, K., ZHANG, X., REN, S. and SUN, J. (2016). Deep residual learning for image recognition.", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 720, + 452, + 734 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 452, + 734 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "HINTON, G., DENG, L., YU, D., DAHL, G., MOHAMED, A.-R., JAITLY, N., SENIOR, A., VAN-", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "HOUCKE, V., NGUYEN, P., KINGSBURY, B. ET AL. (2012). Deep neural networks for acoustic", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 400, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 400, + 118 + ], + "score": 1.0, + "content": "modeling in speech recognition. IEEE Signal processing magazine 29.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 122, + 507, + 137 + ], + "spans": [ + { + "bbox": [ + 104, + 122, + 507, + 137 + ], + "score": 1.0, + "content": "JACOT, A., GABRIEL, F. and HONGLER, C. (2018). Neural tangent kernel: Convergence and", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 134, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 114, + 134, + 395, + 148 + ], + "score": 1.0, + "content": "generalization in neural networks. arXiv preprint arXiv:1806.07572 .", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 153, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 153, + 507, + 168 + ], + "score": 1.0, + "content": "LI, Y. and LIANG, Y. (2018). Learning overparameterized neural networks via stochastic gradient", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 164, + 366, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 366, + 178 + ], + "score": 1.0, + "content": "descent on structured data. arXiv preprint arXiv:1808.01204 .", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 183, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 506, + 196 + ], + "score": 1.0, + "content": "NAKKIRAN, P., KAPLUN, G., KALIMERIS, D., YANG, T., EDELMAN, B. L., ZHANG, F. and", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "BARAK, B. (2019). Sgd on neural networks learns functions of increasing complexity. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 205, + 234, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 205, + 234, + 219 + ], + "score": 1.0, + "content": "preprint arXiv:1905.11604 .", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 223, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 238 + ], + "score": 1.0, + "content": "RAHAMAN, N., BARATIN, A., ARPIT, D., DRAXLER, F., LIN, M., HAMPRECHT, F. A., BEN-", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "GIO, Y. and COURVILLE, A. (2018). On the spectral bias of neural networks. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 199, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 199, + 259 + ], + "score": 1.0, + "content": "arXiv:1806.08734 .", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "ROSASCO, L., BELKIN, M. and VITO, E. D. (2010). On learning with integral operators. Journal", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 275, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 295, + 290 + ], + "score": 1.0, + "content": "of Machine Learning Research 11 905–934.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 294, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 309 + ], + "score": 1.0, + "content": "SMALE, S. and ZHOU, D.-X. (2007). Learning theory estimates via integral operators and their", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 305, + 352, + 320 + ], + "spans": [ + { + "bbox": [ + 113, + 305, + 352, + 320 + ], + "score": 1.0, + "content": "approximations. Constructive approximation 26 153–172.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 324, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 324, + 505, + 339 + ], + "score": 1.0, + "content": "SMALE, S. and ZHOU, D.-X. (2009). Geometry on probability spaces. Constructive Approximation", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 336, + 150, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 150, + 349 + ], + "score": 1.0, + "content": "30 311.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 353, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 370 + ], + "score": 1.0, + "content": "SOUDRY, D., HOFFER, E. and SREBRO, N. (2017). The implicit bias of gradient descent on sepa-", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 365, + 303, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 303, + 380 + ], + "score": 1.0, + "content": "rable data. arXiv preprint arXiv:1710.10345 .", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "SU, L. and YANG, P. (2019). On learning over-parameterized neural networks: A functional ap-", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 395, + 361, + 409 + ], + "spans": [ + { + "bbox": [ + 114, + 395, + 361, + 409 + ], + "score": 1.0, + "content": "proximation prospective. arXiv preprint arXiv:1905.10826 .", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "VEMPALA, S. and WILMES, J. (2018). Gradient descent for one-hidden-layer neural networks:", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 425, + 447, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 447, + 439 + ], + "score": 1.0, + "content": "Polynomial convergence and sq lower bounds. arXiv preprint arXiv:1805.02677 .", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "score": 1.0, + "content": "WU, X., DU, S. S. and WARD, R. (2019). Global convergence of adaptive gradient methods for an", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 455, + 402, + 469 + ], + "spans": [ + { + "bbox": [ + 113, + 455, + 402, + 469 + ], + "score": 1.0, + "content": "over-parameterized neural network. arXiv preprint arXiv:1902.07111 .", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "score": 1.0, + "content": "XU, Z. J. (2018). Understanding training and generalization in deep learning by fourier analysis.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 485, + 257, + 499 + ], + "spans": [ + { + "bbox": [ + 114, + 485, + 257, + 499 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1808.04295 .", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 502, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 507, + 519 + ], + "score": 1.0, + "content": "YANG, G. and SALMAN, H. (2019). A fine-grained spectral perspective on neural networks. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 515, + 233, + 528 + ], + "spans": [ + { + "bbox": [ + 113, + 515, + 233, + 528 + ], + "score": 1.0, + "content": "preprint arXiv:1907.10599 .", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 531, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 506, + 549 + ], + "score": 1.0, + "content": "ZHANG, C., BENGIO, S., HARDT, M., RECHT, B. and VINYALS, O. (2016). Understanding deep", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 545, + 433, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 433, + 559 + ], + "score": 1.0, + "content": "learning requires rethinking generalization. arXiv preprint arXiv:1611.03530 .", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 563, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 578 + ], + "score": 1.0, + "content": "ZHOU, D., TANG, Y., YANG, Z., CAO, Y. and GU, Q. (2018). On the convergence of adaptive", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 574, + 444, + 589 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 444, + 589 + ], + "score": 1.0, + "content": "gradient methods for nonconvex optimization. arXiv preprint arXiv:1808.05671 .", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "score": 1.0, + "content": "ZOU, D., CAO, Y., ZHOU, D. and GU, Q. (2018). Stochastic gradient descent optimizes over-", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 604, + 399, + 619 + ], + "spans": [ + { + "bbox": [ + 113, + 604, + 399, + 619 + ], + "score": 1.0, + "content": "parameterized deep relu networks. arXiv preprint arXiv:1811.08888 .", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 623, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 623, + 506, + 638 + ], + "score": 1.0, + "content": "ZOU, D. and GU, Q. (2019). An improved analysis of training over-parameterized deep neural", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 634, + 300, + 648 + ], + "spans": [ + { + "bbox": [ + 114, + 634, + 300, + 648 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1906.04688 .", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 82, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 48, + 506, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "HINTON, G., DENG, L., YU, D., DAHL, G., MOHAMED, A.-R., JAITLY, N., SENIOR, A., VAN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "HOUCKE, V., NGUYEN, P., KINGSBURY, B. ET AL. (2012). Deep neural networks for acoustic", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 400, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 400, + 118 + ], + "score": 1.0, + "content": "modeling in speech recognition. IEEE Signal processing magazine 29.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 122, + 507, + 137 + ], + "spans": [ + { + "bbox": [ + 104, + 122, + 507, + 137 + ], + "score": 1.0, + "content": "JACOT, A., GABRIEL, F. and HONGLER, C. (2018). Neural tangent kernel: Convergence and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 134, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 114, + 134, + 395, + 148 + ], + "score": 1.0, + "content": "generalization in neural networks. arXiv preprint arXiv:1806.07572 .", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 153, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 153, + 507, + 168 + ], + "score": 1.0, + "content": "LI, Y. and LIANG, Y. (2018). Learning overparameterized neural networks via stochastic gradient", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 366, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 366, + 178 + ], + "score": 1.0, + "content": "descent on structured data. arXiv preprint arXiv:1808.01204 .", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 183, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 506, + 196 + ], + "score": 1.0, + "content": "NAKKIRAN, P., KAPLUN, G., KALIMERIS, D., YANG, T., EDELMAN, B. L., ZHANG, F. and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "BARAK, B. (2019). Sgd on neural networks learns functions of increasing complexity. arXiv", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 205, + 234, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 205, + 234, + 219 + ], + "score": 1.0, + "content": "preprint arXiv:1905.11604 .", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 238 + ], + "score": 1.0, + "content": "RAHAMAN, N., BARATIN, A., ARPIT, D., DRAXLER, F., LIN, M., HAMPRECHT, F. A., BEN-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "GIO, Y. and COURVILLE, A. (2018). On the spectral bias of neural networks. arXiv preprint", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 199, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 199, + 259 + ], + "score": 1.0, + "content": "arXiv:1806.08734 .", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "ROSASCO, L., BELKIN, M. and VITO, E. D. (2010). On learning with integral operators. Journal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 275, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 295, + 290 + ], + "score": 1.0, + "content": "of Machine Learning Research 11 905–934.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 294, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 309 + ], + "score": 1.0, + "content": "SMALE, S. and ZHOU, D.-X. (2007). Learning theory estimates via integral operators and their", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 305, + 352, + 320 + ], + "spans": [ + { + "bbox": [ + 113, + 305, + 352, + 320 + ], + "score": 1.0, + "content": "approximations. Constructive approximation 26 153–172.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 324, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 324, + 505, + 339 + ], + "score": 1.0, + "content": "SMALE, S. and ZHOU, D.-X. (2009). Geometry on probability spaces. Constructive Approximation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 336, + 150, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 150, + 349 + ], + "score": 1.0, + "content": "30 311.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 353, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 370 + ], + "score": 1.0, + "content": "SOUDRY, D., HOFFER, E. and SREBRO, N. (2017). The implicit bias of gradient descent on sepa-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 303, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 303, + 380 + ], + "score": 1.0, + "content": "rable data. arXiv preprint arXiv:1710.10345 .", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "SU, L. and YANG, P. (2019). On learning over-parameterized neural networks: A functional ap-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 114, + 395, + 361, + 409 + ], + "spans": [ + { + "bbox": [ + 114, + 395, + 361, + 409 + ], + "score": 1.0, + "content": "proximation prospective. arXiv preprint arXiv:1905.10826 .", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "VEMPALA, S. and WILMES, J. (2018). Gradient descent for one-hidden-layer neural networks:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 425, + 447, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 447, + 439 + ], + "score": 1.0, + "content": "Polynomial convergence and sq lower bounds. arXiv preprint arXiv:1805.02677 .", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 459 + ], + "score": 1.0, + "content": "WU, X., DU, S. S. and WARD, R. (2019). Global convergence of adaptive gradient methods for an", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 113, + 455, + 402, + 469 + ], + "spans": [ + { + "bbox": [ + 113, + 455, + 402, + 469 + ], + "score": 1.0, + "content": "over-parameterized neural network. arXiv preprint arXiv:1902.07111 .", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "score": 1.0, + "content": "XU, Z. J. (2018). Understanding training and generalization in deep learning by fourier analysis.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 485, + 257, + 499 + ], + "spans": [ + { + "bbox": [ + 114, + 485, + 257, + 499 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1808.04295 .", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 502, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 507, + 519 + ], + "score": 1.0, + "content": "YANG, G. and SALMAN, H. (2019). A fine-grained spectral perspective on neural networks. arXiv", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 113, + 515, + 233, + 528 + ], + "spans": [ + { + "bbox": [ + 113, + 515, + 233, + 528 + ], + "score": 1.0, + "content": "preprint arXiv:1907.10599 .", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 531, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 506, + 549 + ], + "score": 1.0, + "content": "ZHANG, C., BENGIO, S., HARDT, M., RECHT, B. and VINYALS, O. (2016). Understanding deep", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 545, + 433, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 433, + 559 + ], + "score": 1.0, + "content": "learning requires rethinking generalization. arXiv preprint arXiv:1611.03530 .", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 563, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 578 + ], + "score": 1.0, + "content": "ZHOU, D., TANG, Y., YANG, Z., CAO, Y. and GU, Q. (2018). On the convergence of adaptive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 574, + 444, + 589 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 444, + 589 + ], + "score": 1.0, + "content": "gradient methods for nonconvex optimization. arXiv preprint arXiv:1808.05671 .", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "score": 1.0, + "content": "ZOU, D., CAO, Y., ZHOU, D. and GU, Q. (2018). Stochastic gradient descent optimizes over-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 113, + 604, + 399, + 619 + ], + "spans": [ + { + "bbox": [ + 113, + 604, + 399, + 619 + ], + "score": 1.0, + "content": "parameterized deep relu networks. arXiv preprint arXiv:1811.08888 .", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 623, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 623, + 506, + 638 + ], + "score": 1.0, + "content": "ZOU, D. and GU, Q. (2019). An improved analysis of training over-parameterized deep neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 634, + 300, + 648 + ], + "spans": [ + { + "bbox": [ + 114, + 634, + 300, + 648 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1906.04688 .", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 19 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 48, + 506, + 650 + ], + "lines": [], + "index": 19, + "bbox_fs": [ + 104, + 83, + 507, + 648 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 316, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 316, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 316, + 96 + ], + "score": 1.0, + "content": "A REVIEW ON SPHERICAL HARMONICS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "In this section, we give a brief review on relevant concepts in spherical harmonics. For more detials,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "see Bach (2017), Bietti and Mairal (2019), Frye and Efthimiou (2012) and Atkinson and Han (2012)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 167, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 167, + 139 + ], + "score": 1.0, + "content": "for references.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 142, + 503, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 223, + 159 + ], + "score": 1.0, + "content": "We consider the unit sphere", + "type": "text" + }, + { + "bbox": [ + 223, + 142, + 345, + 156 + ], + "score": 0.94, + "content": "\\mathbb { S } ^ { d } = \\left\\{ \\mathbf { x } \\in \\mathbb { R } ^ { d + 1 } : \\| \\mathbf { x } \\| _ { 2 } = 1 \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 140, + 479, + 159 + ], + "score": 1.0, + "content": ", whose surface area is given by", + "type": "text" + }, + { + "bbox": [ + 480, + 145, + 505, + 155 + ], + "score": 0.83, + "content": "\\omega _ { d } =", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 154, + 406, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 204, + 169 + ], + "score": 0.91, + "content": "2 \\pi ^ { ( d + 1 ) / 2 } / \\Gamma ( ( d + 1 ) / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 154, + 251, + 171 + ], + "score": 1.0, + "content": "and denote", + "type": "text" + }, + { + "bbox": [ + 252, + 159, + 262, + 168 + ], + "score": 0.86, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 154, + 406, + 171 + ], + "score": 1.0, + "content": "the uniform measure on the sphere.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 506, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 165, + 187 + ], + "score": 1.0, + "content": "For any !", + "type": "text" + }, + { + "bbox": [ + 165, + 174, + 242, + 185 + ], + "score": 0.57, + "content": "k \\qquad \\geqslant \\qquad 1 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 174, + 277, + 186 + ], + "score": 1.0, + "content": "we", + "type": "text" + }, + { + "bbox": [ + 286, + 173, + 326, + 186 + ], + "score": 1.0, + "content": "consider)", + "type": "text" + }, + { + "bbox": [ + 335, + 174, + 347, + 185 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 354, + 173, + 506, + 186 + ], + "score": 1.0, + "content": "set of spherical harmonics", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 185, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 327, + 205 + ], + "score": 0.82, + "content": "\\begin{array} { r } { \\{ Y _ { k , j } : \\mathbb { S } ^ { d } \\mathbb { R } | 1 \\leqslant j \\leqslant N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "They form an orthonormal basis and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 202, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 104, + 202, + 230, + 218 + ], + "score": 1.0, + "content": "satisfy the following equation", + "type": "text" + }, + { + "bbox": [ + 230, + 204, + 430, + 217 + ], + "score": 0.87, + "content": "\\langle Y _ { k i } , Y _ { s j } \\rangle _ { \\mathbb { S } ^ { d } } = \\int _ { \\mathbb { S } ^ { d } } Y _ { k i } ( x ) Y _ { s j } ( x ) d \\tau _ { d } ( x ) = \\delta _ { i j } \\delta _ { s k }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 202, + 506, + 218 + ], + "score": 1.0, + "content": ". Moreover, since", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 216, + 472, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 277, + 229 + ], + "score": 1.0, + "content": "they are homogeneous functions of degree", + "type": "text" + }, + { + "bbox": [ + 278, + 217, + 284, + 226 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 216, + 345, + 229 + ], + "score": 1.0, + "content": ", it is clear that", + "type": "text" + }, + { + "bbox": [ + 345, + 217, + 370, + 228 + ], + "score": 0.92, + "content": "Y _ { k } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 216, + 461, + 229 + ], + "score": 1.0, + "content": "has the same parity as", + "type": "text" + }, + { + "bbox": [ + 461, + 217, + 468, + 226 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 216, + 472, + 229 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 233, + 226, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 227, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 227, + 245 + ], + "score": 1.0, + "content": "We have the addition formula", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 247, + 396, + 284 + ], + "lines": [ + { + "bbox": [ + 214, + 247, + 396, + 284 + ], + "spans": [ + { + "bbox": [ + 214, + 247, + 396, + 284 + ], + "score": 0.94, + "content": "\\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) ,", + "type": "interline_equation", + "image_path": "e4e6a6a3007539b5cbae000e6184154a231ac93d9516c781bfe369055e2c35bb.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 247, + 396, + 265.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 214, + 265.5, + 396, + 284.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 504, + 311 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 133, + 302 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 288, + 157, + 300 + ], + "score": 0.91, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 286, + 311, + 302 + ], + "score": 1.0, + "content": "is the Legendre polynomial of degree", + "type": "text" + }, + { + "bbox": [ + 312, + 289, + 319, + 298 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 286, + 330, + 302 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 331, + 289, + 354, + 299 + ], + "score": 0.9, + "content": "d + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 286, + 505, + 302 + ], + "score": 1.0, + "content": "dimensions, explicitly given by (Ro-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 299, + 180, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 180, + 312 + ], + "score": 1.0, + "content": "drigues’ formula)", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 315, + 439, + 347 + ], + "lines": [ + { + "bbox": [ + 172, + 315, + 439, + 347 + ], + "spans": [ + { + "bbox": [ + 172, + 315, + 439, + 347 + ], + "score": 0.94, + "content": "P _ { k } ( t ) = \\left( - { \\frac { 1 } { 2 } } \\right) ^ { k } { \\frac { \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( k + { \\frac { d } { 2 } } \\right) } } \\left( 1 - t ^ { 2 } \\right) ^ { { \\frac { 2 - d } { 2 } } } \\left( { \\frac { d } { d t } } \\right) ^ { k } \\left( 1 - t ^ { 2 } \\right) ^ { k + { \\frac { d - 2 } { 2 } } } .", + "type": "interline_equation", + "image_path": "884d10edd575ec5a70028342121de9ec7280e05c1976d7891734223d069066ae.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 172, + 315, + 439, + 325.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 172, + 325.6666666666667, + 439, + 336.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 172, + 336.33333333333337, + 439, + 347.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 503, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 190, + 366 + ], + "score": 1.0, + "content": "We can also see that", + "type": "text" + }, + { + "bbox": [ + 190, + 352, + 213, + 364 + ], + "score": 0.92, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 350, + 360, + 366 + ], + "score": 1.0, + "content": ", the Legendre polynomial of degree", + "type": "text" + }, + { + "bbox": [ + 360, + 353, + 367, + 362 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 350, + 479, + 366 + ], + "score": 1.0, + "content": "shares the same parity with", + "type": "text" + }, + { + "bbox": [ + 480, + 353, + 486, + 362 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 350, + 505, + 366 + ], + "score": 1.0, + "content": ". By", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 340, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 340, + 377 + ], + "score": 1.0, + "content": "the orthogonality and the addition formula (A.1) we have,", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 379, + 423, + 406 + ], + "lines": [ + { + "bbox": [ + 187, + 379, + 423, + 406 + ], + "spans": [ + { + "bbox": [ + 187, + 379, + 423, + 406 + ], + "score": 0.94, + "content": "\\int _ { \\mathbb { S } ^ { d } } P _ { j } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) P _ { k } ( \\langle \\mathbf { w } , \\mathbf { y } \\rangle ) d \\tau _ { d } ( \\mathbf { w } ) = \\frac { \\delta _ { j k } } { N ( d , k ) } P _ { k } ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) .", + "type": "interline_equation", + "image_path": "3b995ffb6893c07c99ad26afad82d3f241cd2dbd0583fab1d660f4a3aa09f22a.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 187, + 379, + 423, + 406 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 388, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 389, + 430 + ], + "score": 1.0, + "content": "Further we have the recurrence relation for the Legendre polynomials,", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 432, + 414, + 458 + ], + "lines": [ + { + "bbox": [ + 196, + 432, + 414, + 458 + ], + "spans": [ + { + "bbox": [ + 196, + 432, + 414, + 458 + ], + "score": 0.92, + "content": "t P _ { k } ( t ) = \\frac { k } { 2 k + d - 1 } P _ { k - 1 } ( t ) + \\frac { k + d - 1 } { 2 k + d - 1 } P _ { k + 1 } ( t ) ,", + "type": "interline_equation", + "image_path": "5cd19607dcf4de36ab35db86e5993c63f7e15e6ed25976ca5eb0e00d04d64014.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 196, + 432, + 414, + 458 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 270, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 270, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 120, + 475 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 463, + 146, + 473 + ], + "score": 0.89, + "content": "k \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 462, + 164, + 475 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 462, + 226, + 475 + ], + "score": 0.93, + "content": "t P _ { 0 } ( t ) = P _ { 1 } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 462, + 241, + 475 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 242, + 463, + 267, + 473 + ], + "score": 0.89, + "content": "k = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 462, + 270, + 475 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 404, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 403, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 343, + 493 + ], + "score": 1.0, + "content": "The Hecke-Funk formula is given for a spherical harmonic", + "type": "text" + }, + { + "bbox": [ + 343, + 480, + 355, + 491 + ], + "score": 0.88, + "content": "Y _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 478, + 396, + 493 + ], + "score": 1.0, + "content": "of degree", + "type": "text" + }, + { + "bbox": [ + 396, + 480, + 403, + 489 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 495, + 457, + 525 + ], + "lines": [ + { + "bbox": [ + 153, + 495, + 457, + 525 + ], + "spans": [ + { + "bbox": [ + 153, + 495, + 457, + 525 + ], + "score": 0.92, + "content": "\\int _ { \\mathbb { S } ^ { d } } f ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) Y _ { k } ( \\mathbf { y } ) d t ^ { \\prime } d ( \\mathbf { y } ) = { \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } } Y _ { k } ( \\mathbf { x } ) \\int _ { - 1 } ^ { 1 } f ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t .", + "type": "interline_equation", + "image_path": "27e6f855d432637c79a63af9f93cfed267456e457dca54264bdb452130f7b337.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 153, + 495, + 457, + 525 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 536, + 273, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 274, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 274, + 551 + ], + "score": 1.0, + "content": "B PROOF OF MAIN THEOREMS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 560, + 239, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 240, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 240, + 573 + ], + "score": 1.0, + "content": "B.1 PROOF OF THEOREM 3.2", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "In this section we give the proof of Theorem 3.2. The core idea of our proof is to establish connec-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "score": 1.0, + "content": "tions between neural network gradients throughout training and the neural tangent kernel. To do so,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 604, + 338, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 338, + 617 + ], + "score": 1.0, + "content": "we first introduce the following definitions and notations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 620, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 104, + 616, + 504, + 637 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 139, + 637 + ], + "score": 1.0, + "content": "Define", + "type": "text" + }, + { + "bbox": [ + 140, + 621, + 364, + 634 + ], + "score": 0.85, + "content": "\\begin{array} { r l r } { \\mathbf { K } ^ { ( 0 ) } } & { { } = } & { m ^ { - 1 } ( \\left. \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\right. ) _ { n \\times n } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 616, + 371, + 637 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 372, + 621, + 504, + 635 + ], + "score": 0.87, + "content": "\\begin{array} { r l r } { { \\bf K } ^ { ( \\infty ) } } & { { } = } & { ( \\kappa ( { \\bf x } _ { i } , { \\bf x } _ { j } ) ) _ { n \\times n } \\quad = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 632, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 165, + 648 + ], + "score": 0.91, + "content": "\\mathrm { l i m } _ { m \\infty } \\mathbf { K } ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 632, + 190, + 650 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 190, + 634, + 296, + 648 + ], + "score": 0.82, + "content": "\\{ \\widehat \\lambda _ { i } \\} _ { i = 1 } ^ { n } , \\widehat \\lambda _ { 1 } \\geqslant \\ \\cdot \\cdot \\ \\geqslant \\ \\widehat \\lambda _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 632, + 388, + 650 + ], + "score": 1.0, + "content": "be the eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 388, + 635, + 422, + 646 + ], + "score": 0.9, + "content": "n ^ { - 1 } \\mathbf { K } ^ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 632, + 445, + 650 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 446, + 636, + 491, + 648 + ], + "score": 0.91, + "content": "\\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 632, + 506, + 650 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 646, + 444, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 250, + 664 + ], + "score": 1.0, + "content": "the corresponding eigenvectors. Set", + "type": "text" + }, + { + "bbox": [ + 251, + 647, + 439, + 662 + ], + "score": 0.72, + "content": "\\widehat { \\mathbf { V } } _ { r _ { k } } = ( \\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { r _ { k } } ) , \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } = ( \\widehat { \\mathbf { v } } _ { r _ { k } + 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 646, + 444, + 664 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 504, + 696 + ], + "lines": [ + { + "bbox": [ + 104, + 665, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 250, + 684 + ], + "score": 1.0, + "content": "For notation simplicity, we denote ˇ", + "type": "text" + }, + { + "bbox": [ + 251, + 667, + 504, + 682 + ], + "score": 0.33, + "content": "\\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ = \\ \\left[ \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , \\ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ =", + "type": "inline_equation", + "image_path": "80b2d5955eedce6e22f8e1920369c5a6d6a8847f94deb30f60e7105843684ca6.jpg" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 680, + 242, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 242, + 696 + ], + "score": 0.85, + "content": "\\left[ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , l = 1 , 2 .", + "type": "inline_equation" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "The following lemma is a direct application of Proposition 1 in Smale and Zhou (2009) or Proposi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "tion 10 in Rosasco et al. (2010). It bounds the difference between the eigenvalues of NTK and their", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 210, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 210, + 732 + ], + "score": 1.0, + "content": "finite-width counterparts.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 316, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 316, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 316, + 96 + ], + "score": 1.0, + "content": "A REVIEW ON SPHERICAL HARMONICS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "In this section, we give a brief review on relevant concepts in spherical harmonics. For more detials,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "see Bach (2017), Bietti and Mairal (2019), Frye and Efthimiou (2012) and Atkinson and Han (2012)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 167, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 167, + 139 + ], + "score": 1.0, + "content": "for references.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 104, + 505, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 142, + 503, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 223, + 159 + ], + "score": 1.0, + "content": "We consider the unit sphere", + "type": "text" + }, + { + "bbox": [ + 223, + 142, + 345, + 156 + ], + "score": 0.94, + "content": "\\mathbb { S } ^ { d } = \\left\\{ \\mathbf { x } \\in \\mathbb { R } ^ { d + 1 } : \\| \\mathbf { x } \\| _ { 2 } = 1 \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 140, + 479, + 159 + ], + "score": 1.0, + "content": ", whose surface area is given by", + "type": "text" + }, + { + "bbox": [ + 480, + 145, + 505, + 155 + ], + "score": 0.83, + "content": "\\omega _ { d } =", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 154, + 406, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 204, + 169 + ], + "score": 0.91, + "content": "2 \\pi ^ { ( d + 1 ) / 2 } / \\Gamma ( ( d + 1 ) / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 154, + 251, + 171 + ], + "score": 1.0, + "content": "and denote", + "type": "text" + }, + { + "bbox": [ + 252, + 159, + 262, + 168 + ], + "score": 0.86, + "content": "\\tau _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 154, + 406, + 171 + ], + "score": 1.0, + "content": "the uniform measure on the sphere.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 140, + 505, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 506, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 165, + 187 + ], + "score": 1.0, + "content": "For any !", + "type": "text" + }, + { + "bbox": [ + 165, + 174, + 242, + 185 + ], + "score": 0.57, + "content": "k \\qquad \\geqslant \\qquad 1 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 174, + 277, + 186 + ], + "score": 1.0, + "content": "we", + "type": "text" + }, + { + "bbox": [ + 286, + 173, + 326, + 186 + ], + "score": 1.0, + "content": "consider)", + "type": "text" + }, + { + "bbox": [ + 335, + 174, + 347, + 185 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 354, + 173, + 506, + 186 + ], + "score": 1.0, + "content": "set of spherical harmonics", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 185, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 327, + 205 + ], + "score": 0.82, + "content": "\\begin{array} { r } { \\{ Y _ { k , j } : \\mathbb { S } ^ { d } \\mathbb { R } | 1 \\leqslant j \\leqslant N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "They form an orthonormal basis and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 202, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 104, + 202, + 230, + 218 + ], + "score": 1.0, + "content": "satisfy the following equation", + "type": "text" + }, + { + "bbox": [ + 230, + 204, + 430, + 217 + ], + "score": 0.87, + "content": "\\langle Y _ { k i } , Y _ { s j } \\rangle _ { \\mathbb { S } ^ { d } } = \\int _ { \\mathbb { S } ^ { d } } Y _ { k i } ( x ) Y _ { s j } ( x ) d \\tau _ { d } ( x ) = \\delta _ { i j } \\delta _ { s k }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 202, + 506, + 218 + ], + "score": 1.0, + "content": ". Moreover, since", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 216, + 472, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 277, + 229 + ], + "score": 1.0, + "content": "they are homogeneous functions of degree", + "type": "text" + }, + { + "bbox": [ + 278, + 217, + 284, + 226 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 216, + 345, + 229 + ], + "score": 1.0, + "content": ", it is clear that", + "type": "text" + }, + { + "bbox": [ + 345, + 217, + 370, + 228 + ], + "score": 0.92, + "content": "Y _ { k } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 216, + 461, + 229 + ], + "score": 1.0, + "content": "has the same parity as", + "type": "text" + }, + { + "bbox": [ + 461, + 217, + 468, + 226 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 216, + 472, + 229 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 172, + 506, + 229 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 233, + 226, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 227, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 227, + 245 + ], + "score": 1.0, + "content": "We have the addition formula", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 232, + 227, + 245 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 247, + 396, + 284 + ], + "lines": [ + { + "bbox": [ + 214, + 247, + 396, + 284 + ], + "spans": [ + { + "bbox": [ + 214, + 247, + 396, + 284 + ], + "score": 0.94, + "content": "\\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) ,", + "type": "interline_equation", + "image_path": "e4e6a6a3007539b5cbae000e6184154a231ac93d9516c781bfe369055e2c35bb.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 247, + 396, + 265.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 214, + 265.5, + 396, + 284.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 504, + 311 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 133, + 302 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 288, + 157, + 300 + ], + "score": 0.91, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 286, + 311, + 302 + ], + "score": 1.0, + "content": "is the Legendre polynomial of degree", + "type": "text" + }, + { + "bbox": [ + 312, + 289, + 319, + 298 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 286, + 330, + 302 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 331, + 289, + 354, + 299 + ], + "score": 0.9, + "content": "d + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 286, + 505, + 302 + ], + "score": 1.0, + "content": "dimensions, explicitly given by (Ro-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 299, + 180, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 180, + 312 + ], + "score": 1.0, + "content": "drigues’ formula)", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 286, + 505, + 312 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 315, + 439, + 347 + ], + "lines": [ + { + "bbox": [ + 172, + 315, + 439, + 347 + ], + "spans": [ + { + "bbox": [ + 172, + 315, + 439, + 347 + ], + "score": 0.94, + "content": "P _ { k } ( t ) = \\left( - { \\frac { 1 } { 2 } } \\right) ^ { k } { \\frac { \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( k + { \\frac { d } { 2 } } \\right) } } \\left( 1 - t ^ { 2 } \\right) ^ { { \\frac { 2 - d } { 2 } } } \\left( { \\frac { d } { d t } } \\right) ^ { k } \\left( 1 - t ^ { 2 } \\right) ^ { k + { \\frac { d - 2 } { 2 } } } .", + "type": "interline_equation", + "image_path": "884d10edd575ec5a70028342121de9ec7280e05c1976d7891734223d069066ae.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 172, + 315, + 439, + 325.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 172, + 325.6666666666667, + 439, + 336.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 172, + 336.33333333333337, + 439, + 347.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 503, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 190, + 366 + ], + "score": 1.0, + "content": "We can also see that", + "type": "text" + }, + { + "bbox": [ + 190, + 352, + 213, + 364 + ], + "score": 0.92, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 350, + 360, + 366 + ], + "score": 1.0, + "content": ", the Legendre polynomial of degree", + "type": "text" + }, + { + "bbox": [ + 360, + 353, + 367, + 362 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 350, + 479, + 366 + ], + "score": 1.0, + "content": "shares the same parity with", + "type": "text" + }, + { + "bbox": [ + 480, + 353, + 486, + 362 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 350, + 505, + 366 + ], + "score": 1.0, + "content": ". By", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 340, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 340, + 377 + ], + "score": 1.0, + "content": "the orthogonality and the addition formula (A.1) we have,", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 350, + 505, + 377 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 379, + 423, + 406 + ], + "lines": [ + { + "bbox": [ + 187, + 379, + 423, + 406 + ], + "spans": [ + { + "bbox": [ + 187, + 379, + 423, + 406 + ], + "score": 0.94, + "content": "\\int _ { \\mathbb { S } ^ { d } } P _ { j } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) P _ { k } ( \\langle \\mathbf { w } , \\mathbf { y } \\rangle ) d \\tau _ { d } ( \\mathbf { w } ) = \\frac { \\delta _ { j k } } { N ( d , k ) } P _ { k } ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) .", + "type": "interline_equation", + "image_path": "3b995ffb6893c07c99ad26afad82d3f241cd2dbd0583fab1d660f4a3aa09f22a.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 187, + 379, + 423, + 406 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 388, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 389, + 430 + ], + "score": 1.0, + "content": "Further we have the recurrence relation for the Legendre polynomials,", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 414, + 389, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 432, + 414, + 458 + ], + "lines": [ + { + "bbox": [ + 196, + 432, + 414, + 458 + ], + "spans": [ + { + "bbox": [ + 196, + 432, + 414, + 458 + ], + "score": 0.92, + "content": "t P _ { k } ( t ) = \\frac { k } { 2 k + d - 1 } P _ { k - 1 } ( t ) + \\frac { k + d - 1 } { 2 k + d - 1 } P _ { k + 1 } ( t ) ,", + "type": "interline_equation", + "image_path": "5cd19607dcf4de36ab35db86e5993c63f7e15e6ed25976ca5eb0e00d04d64014.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 196, + 432, + 414, + 458 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 270, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 270, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 120, + 475 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 463, + 146, + 473 + ], + "score": 0.89, + "content": "k \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 462, + 164, + 475 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 462, + 226, + 475 + ], + "score": 0.93, + "content": "t P _ { 0 } ( t ) = P _ { 1 } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 462, + 241, + 475 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 242, + 463, + 267, + 473 + ], + "score": 0.89, + "content": "k = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 462, + 270, + 475 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 462, + 270, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 404, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 403, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 343, + 493 + ], + "score": 1.0, + "content": "The Hecke-Funk formula is given for a spherical harmonic", + "type": "text" + }, + { + "bbox": [ + 343, + 480, + 355, + 491 + ], + "score": 0.88, + "content": "Y _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 478, + 396, + 493 + ], + "score": 1.0, + "content": "of degree", + "type": "text" + }, + { + "bbox": [ + 396, + 480, + 403, + 489 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 478, + 403, + 493 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 495, + 457, + 525 + ], + "lines": [ + { + "bbox": [ + 153, + 495, + 457, + 525 + ], + "spans": [ + { + "bbox": [ + 153, + 495, + 457, + 525 + ], + "score": 0.92, + "content": "\\int _ { \\mathbb { S } ^ { d } } f ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) Y _ { k } ( \\mathbf { y } ) d t ^ { \\prime } d ( \\mathbf { y } ) = { \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } } Y _ { k } ( \\mathbf { x } ) \\int _ { - 1 } ^ { 1 } f ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t .", + "type": "interline_equation", + "image_path": "27e6f855d432637c79a63af9f93cfed267456e457dca54264bdb452130f7b337.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 153, + 495, + 457, + 525 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 536, + 273, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 274, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 274, + 551 + ], + "score": 1.0, + "content": "B PROOF OF MAIN THEOREMS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 560, + 239, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 240, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 240, + 573 + ], + "score": 1.0, + "content": "B.1 PROOF OF THEOREM 3.2", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "In this section we give the proof of Theorem 3.2. The core idea of our proof is to establish connec-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 607 + ], + "score": 1.0, + "content": "tions between neural network gradients throughout training and the neural tangent kernel. To do so,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 604, + 338, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 338, + 617 + ], + "score": 1.0, + "content": "we first introduce the following definitions and notations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 581, + 505, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 620, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 104, + 616, + 504, + 637 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 139, + 637 + ], + "score": 1.0, + "content": "Define", + "type": "text" + }, + { + "bbox": [ + 140, + 621, + 364, + 634 + ], + "score": 0.85, + "content": "\\begin{array} { r l r } { \\mathbf { K } ^ { ( 0 ) } } & { { } = } & { m ^ { - 1 } ( \\left. \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\right. ) _ { n \\times n } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 616, + 371, + 637 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 372, + 621, + 504, + 635 + ], + "score": 0.87, + "content": "\\begin{array} { r l r } { { \\bf K } ^ { ( \\infty ) } } & { { } = } & { ( \\kappa ( { \\bf x } _ { i } , { \\bf x } _ { j } ) ) _ { n \\times n } \\quad = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 632, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 165, + 648 + ], + "score": 0.91, + "content": "\\mathrm { l i m } _ { m \\infty } \\mathbf { K } ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 632, + 190, + 650 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 190, + 634, + 296, + 648 + ], + "score": 0.82, + "content": "\\{ \\widehat \\lambda _ { i } \\} _ { i = 1 } ^ { n } , \\widehat \\lambda _ { 1 } \\geqslant \\ \\cdot \\cdot \\ \\geqslant \\ \\widehat \\lambda _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 632, + 388, + 650 + ], + "score": 1.0, + "content": "be the eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 388, + 635, + 422, + 646 + ], + "score": 0.9, + "content": "n ^ { - 1 } \\mathbf { K } ^ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 632, + 445, + 650 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 446, + 636, + 491, + 648 + ], + "score": 0.91, + "content": "\\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 632, + 506, + 650 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 646, + 444, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 250, + 664 + ], + "score": 1.0, + "content": "the corresponding eigenvectors. Set", + "type": "text" + }, + { + "bbox": [ + 251, + 647, + 439, + 662 + ], + "score": 0.72, + "content": "\\widehat { \\mathbf { V } } _ { r _ { k } } = ( \\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { r _ { k } } ) , \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } = ( \\widehat { \\mathbf { v } } _ { r _ { k } + 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 646, + 444, + 664 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 616, + 506, + 664 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 504, + 696 + ], + "lines": [ + { + "bbox": [ + 104, + 665, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 250, + 684 + ], + "score": 1.0, + "content": "For notation simplicity, we denote ˇ", + "type": "text" + }, + { + "bbox": [ + 251, + 667, + 504, + 682 + ], + "score": 0.33, + "content": "\\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ = \\ \\left[ \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , \\ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ =", + "type": "inline_equation", + "image_path": "80b2d5955eedce6e22f8e1920369c5a6d6a8847f94deb30f60e7105843684ca6.jpg" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 680, + 242, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 242, + 696 + ], + "score": 0.85, + "content": "\\left[ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , l = 1 , 2 .", + "type": "inline_equation" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 665, + 504, + 696 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "The following lemma is a direct application of Proposition 1 in Smale and Zhou (2009) or Proposi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "tion 10 in Rosasco et al. (2010). It bounds the difference between the eigenvalues of NTK and their", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 210, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 210, + 732 + ], + "score": 1.0, + "content": "finite-width counterparts.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 348, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 348, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 196, + 96 + ], + "score": 1.0, + "content": "Lemma B.1. For any", + "type": "text" + }, + { + "bbox": [ + 196, + 83, + 220, + 93 + ], + "score": 0.9, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 81, + 321, + 96 + ], + "score": 1.0, + "content": ", with probability at least", + "type": "text" + }, + { + "bbox": [ + 321, + 83, + 344, + 93 + ], + "score": 0.82, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 81, + 348, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 105, + 367, + 122 + ], + "lines": [ + { + "bbox": [ + 243, + 105, + 367, + 122 + ], + "spans": [ + { + "bbox": [ + 243, + 105, + 367, + 122 + ], + "score": 0.91, + "content": "| \\lambda _ { i } - \\widehat { \\lambda } _ { i } | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( 1 / \\delta ) / n } ) .", + "type": "interline_equation", + "image_path": "12ac2b7883baa3ecb8bd06e1e53798cd5a9113a218a23e671436df30c82f5fb6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 243, + 105, + 367, + 122 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 138, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 420, + 151 + ], + "score": 1.0, + "content": "The following lemma is partly summarized from the proof of equation (44) in", + "type": "text" + }, + { + "bbox": [ + 420, + 140, + 432, + 149 + ], + "score": 0.32, + "content": "\\mathrm { S u }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 138, + 504, + 151 + ], + "score": 1.0, + "content": "and Yang (2019).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "Its purpose is to further connect the eigenfunctions of NTK with their finite-width, finite-sample", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 162, + 162, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 162, + 174 + ], + "score": 1.0, + "content": "counterparts.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 177, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 217, + 191 + ], + "score": 1.0, + "content": "Lemma B.2. Suppose that", + "type": "text" + }, + { + "bbox": [ + 218, + 179, + 271, + 191 + ], + "score": 0.92, + "content": "| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 177, + 299, + 191 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 299, + 178, + 329, + 189 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in S ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 177, + 456, + 191 + ], + "score": 1.0, + "content": ". There exist absolute constants", + "type": "text" + }, + { + "bbox": [ + 456, + 179, + 505, + 190 + ], + "score": 0.9, + "content": "C , C ^ { \\prime } , c ^ { \\prime \\prime } >", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 189, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 189, + 185, + 203 + ], + "score": 1.0, + "content": "0, such that for any", + "type": "text" + }, + { + "bbox": [ + 186, + 190, + 210, + 200 + ], + "score": 0.89, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 189, + 257, + 203 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 258, + 191, + 264, + 200 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 189, + 285, + 203 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 286, + 191, + 316, + 201 + ], + "score": 0.9, + "content": "r _ { k } \\leqslant n", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 189, + 327, + 203 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 328, + 190, + 461, + 203 + ], + "score": 0.92, + "content": "n \\geqslant C ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } \\log ( 1 / \\delta )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 189, + 506, + 203 + ], + "score": 1.0, + "content": ", then with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 211, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 183, + 213 + ], + "score": 1.0, + "content": "probability at least", + "type": "text" + }, + { + "bbox": [ + 183, + 201, + 206, + 212 + ], + "score": 0.86, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 201, + 211, + 213 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 139, + 223, + 472, + 286 + ], + "lines": [ + { + "bbox": [ + 139, + 223, + 472, + 286 + ], + "spans": [ + { + "bbox": [ + 139, + 223, + 472, + 286 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\| _ { F } \\leqslant C ^ { \\prime } \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } , } \\\\ & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\widehat { \\mathbf { V } } _ { r _ { k } } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant C ^ { \\prime \\prime } \\bigg [ \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M ^ { 2 } r _ { k } \\sqrt { \\frac { \\log ( r _ { k } / \\delta ) } { n } } \\bigg ] . } \\end{array}", + "type": "interline_equation", + "image_path": "a68d8f2a70e385c01fe2cca5a61b71ac85ee1277f50aebc5731f18c8e0346861.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 139, + 223, + 472, + 244.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 139, + 244.0, + 472, + 265.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 139, + 265.0, + 472, + 286.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 300, + 504, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 314 + ], + "score": 1.0, + "content": "The following two lemmas are key to characterize the dynamics of the residual throughout training.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "These two results, especially Lemma B.4, are the ones that distinguish our analysis from previous", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "works on neural network training in the NTK regime (Du et al., 2018b; Su and Yang, 2019), since", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 334, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 479, + 347 + ], + "score": 1.0, + "content": "our analysis provides more careful characterization on the residual along different directions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 352, + 504, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 355, + 366 + ], + "score": 1.0, + "content": "Lemma B.3. Suppose that the iterates of gradient descent", + "type": "text" + }, + { + "bbox": [ + 356, + 351, + 423, + 364 + ], + "score": 0.91, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 350, + 506, + 366 + ], + "score": 1.0, + "content": "are inside the ball", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 363, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 107, + 364, + 156, + 377 + ], + "score": 0.91, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 363, + 169, + 379 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 170, + 364, + 259, + 377 + ], + "score": 0.93, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 363, + 380, + 379 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 380, + 364, + 501, + 378 + ], + "score": 0.9, + "content": "1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 363, + 505, + 379 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 114, + 388, + 496, + 405 + ], + "lines": [ + { + "bbox": [ + 114, + 388, + 496, + 405 + ], + "spans": [ + { + "bbox": [ + 114, + 388, + 496, + 405 + ], + "score": 0.83, + "content": "\\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } + 1 ) } = [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) + \\mathbf { e } ^ { ( t ) } , \\ \\| \\mathbf { e } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 }", + "type": "interline_equation", + "image_path": "33ee9e550d82b16b8bcfad2d23c97bd171cd4e1bd0feb467f5f6c86f3c38ad34.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 388, + 496, + 405 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 416, + 487, + 432 + ], + "lines": [ + { + "bbox": [ + 104, + 416, + 486, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 133, + 434 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 419, + 201, + 430 + ], + "score": 0.93, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 416, + 232, + 434 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 232, + 416, + 486, + 432 + ], + "score": 0.68, + "content": "\\mathbf { y } = \\left( y _ { 1 } , \\ldots , y _ { n } \\right) ^ { \\top } , \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } = \\theta \\cdot \\left( f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { n } ) \\right) ^ { \\top } .", + "type": "inline_equation" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 437, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 104, + 435, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 355, + 451 + ], + "score": 1.0, + "content": "Lemma B.4. Suppose that the iterates of gradient descent", + "type": "text" + }, + { + "bbox": [ + 356, + 437, + 423, + 450 + ], + "score": 0.91, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 435, + 506, + 451 + ], + "score": 1.0, + "content": "are inside the ball", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 448, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 107, + 450, + 156, + 464 + ], + "score": 0.93, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 448, + 170, + 466 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 171, + 450, + 342, + 464 + ], + "score": 0.91, + "content": "\\omega \\leqslant \\widetilde { \\mathcal { O } } ( \\operatorname* { m i n } \\{ [ \\log ( m ) ] ^ { - 3 / 2 } , \\lambda _ { r _ { k } } ^ { 3 } , ( \\eta m ) ^ { - 3 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 448, + 361, + 466 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 361, + 450, + 414, + 465 + ], + "score": 0.93, + "content": "n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 448, + 506, + 466 + ], + "score": 1.0, + "content": ", then with probability", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 461, + 273, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 137, + 479 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 137, + 464, + 268, + 477 + ], + "score": 0.9, + "content": "1 - \\mathcal { O } ( t ^ { 2 } n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 461, + 273, + 479 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 486, + 514, + 576 + ], + "lines": [ + { + "bbox": [ + 111, + 486, + 514, + 576 + ], + "spans": [ + { + "bbox": [ + 111, + 486, + 514, + 576 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 1 ) } } \\\\ & { } & { \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) t ^ { \\prime } \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r k } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { n } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } \\cdot \\widetilde { O } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { } & { \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } ) \\cdot \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) ^ { t ^ { \\prime } } + \\widetilde { \\mathcal { O } } ( \\sqrt { n } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r k } ) ^ { - 1 } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 3 ) } \\end{array}", + "type": "interline_equation", + "image_path": "82373c60e69c8dbaa0ed81f5468aea31025051dbee60a3ac6b8b195bb9e56718.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 111, + 486, + 514, + 516.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 111, + 516.0, + 514, + 546.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 546.0, + 514, + 576.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 204, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 584, + 201, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 133, + 600 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 586, + 201, + 598 + ], + "score": 0.91, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 270, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 272, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 272, + 623 + ], + "score": 1.0, + "content": "Now we are ready to prove Theorem 3.2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 653, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 229, + 667 + ], + "score": 1.0, + "content": "Proof of Theorem 3.2. Define", + "type": "text" + }, + { + "bbox": [ + 230, + 654, + 315, + 666 + ], + "score": 0.91, + "content": "\\omega \\ = \\ \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 652, + 492, + 667 + ], + "score": 1.0, + "content": "for some small enough absolute constant", + "type": "text" + }, + { + "bbox": [ + 492, + 654, + 501, + 664 + ], + "score": 0.82, + "content": "\\overline { { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 652, + 505, + 667 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 664, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 244, + 684 + ], + "score": 1.0, + "content": "Then by union bound, as long as", + "type": "text" + }, + { + "bbox": [ + 244, + 667, + 356, + 681 + ], + "score": 0.91, + "content": "m \\geqslant \\widetilde \\Omega ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 664, + 435, + 684 + ], + "score": 1.0, + "content": ", the conditions on", + "type": "text" + }, + { + "bbox": [ + 435, + 670, + 443, + 678 + ], + "score": 0.73, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 664, + 507, + 684 + ], + "score": 1.0, + "content": "given in Lem-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 678, + 299, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 299, + 691 + ], + "score": 1.0, + "content": "mas D.2, D.4, D.5, B.3 and B.4 are all satisfied.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 694, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 104, + 691, + 506, + 710 + ], + "spans": [ + { + "bbox": [ + 104, + 691, + 243, + 710 + ], + "score": 1.0, + "content": "We first show that all the iterates", + "type": "text" + }, + { + "bbox": [ + 244, + 695, + 313, + 707 + ], + "score": 0.9, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 691, + 391, + 710 + ], + "score": 1.0, + "content": "are inside the ball", + "type": "text" + }, + { + "bbox": [ + 392, + 695, + 440, + 708 + ], + "score": 0.93, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 691, + 506, + 710 + ], + "score": 1.0, + "content": ". We prove this", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 706, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 230, + 720 + ], + "score": 1.0, + "content": "result by inductively show that", + "type": "text" + }, + { + "bbox": [ + 230, + 707, + 368, + 720 + ], + "score": 0.84, + "content": "\\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) , t = 0 , \\ldots , T", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 706, + 471, + 720 + ], + "score": 1.0, + "content": ". First of all, it is clear that", + "type": "text" + }, + { + "bbox": [ + 472, + 707, + 505, + 719 + ], + "score": 0.89, + "content": "{ \\bf W } ^ { ( 0 ) } \\in", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 717, + 507, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 719, + 156, + 732 + ], + "score": 0.93, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 717, + 216, + 734 + ], + "score": 1.0, + "content": ". Suppose that", + "type": "text" + }, + { + "bbox": [ + 216, + 720, + 344, + 732 + ], + "score": 0.9, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 717, + 507, + 734 + ], + "score": 1.0, + "content": ". Then the results of Lemmas D.2, D.4,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 348, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 348, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 196, + 96 + ], + "score": 1.0, + "content": "Lemma B.1. For any", + "type": "text" + }, + { + "bbox": [ + 196, + 83, + 220, + 93 + ], + "score": 0.9, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 81, + 321, + 96 + ], + "score": 1.0, + "content": ", with probability at least", + "type": "text" + }, + { + "bbox": [ + 321, + 83, + 344, + 93 + ], + "score": 0.82, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 81, + 348, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 348, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 105, + 367, + 122 + ], + "lines": [ + { + "bbox": [ + 243, + 105, + 367, + 122 + ], + "spans": [ + { + "bbox": [ + 243, + 105, + 367, + 122 + ], + "score": 0.91, + "content": "| \\lambda _ { i } - \\widehat { \\lambda } _ { i } | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( 1 / \\delta ) / n } ) .", + "type": "interline_equation", + "image_path": "12ac2b7883baa3ecb8bd06e1e53798cd5a9113a218a23e671436df30c82f5fb6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 243, + 105, + 367, + 122 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 138, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 420, + 151 + ], + "score": 1.0, + "content": "The following lemma is partly summarized from the proof of equation (44) in", + "type": "text" + }, + { + "bbox": [ + 420, + 140, + 432, + 149 + ], + "score": 0.32, + "content": "\\mathrm { S u }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 138, + 504, + 151 + ], + "score": 1.0, + "content": "and Yang (2019).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "Its purpose is to further connect the eigenfunctions of NTK with their finite-width, finite-sample", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 162, + 162, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 162, + 174 + ], + "score": 1.0, + "content": "counterparts.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 138, + 505, + 174 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 177, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 217, + 191 + ], + "score": 1.0, + "content": "Lemma B.2. Suppose that", + "type": "text" + }, + { + "bbox": [ + 218, + 179, + 271, + 191 + ], + "score": 0.92, + "content": "| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 177, + 299, + 191 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 299, + 178, + 329, + 189 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in S ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 177, + 456, + 191 + ], + "score": 1.0, + "content": ". There exist absolute constants", + "type": "text" + }, + { + "bbox": [ + 456, + 179, + 505, + 190 + ], + "score": 0.9, + "content": "C , C ^ { \\prime } , c ^ { \\prime \\prime } >", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 189, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 189, + 185, + 203 + ], + "score": 1.0, + "content": "0, such that for any", + "type": "text" + }, + { + "bbox": [ + 186, + 190, + 210, + 200 + ], + "score": 0.89, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 189, + 257, + 203 + ], + "score": 1.0, + "content": "and integer", + "type": "text" + }, + { + "bbox": [ + 258, + 191, + 264, + 200 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 189, + 285, + 203 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 286, + 191, + 316, + 201 + ], + "score": 0.9, + "content": "r _ { k } \\leqslant n", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 189, + 327, + 203 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 328, + 190, + 461, + 203 + ], + "score": 0.92, + "content": "n \\geqslant C ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } \\log ( 1 / \\delta )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 189, + 506, + 203 + ], + "score": 1.0, + "content": ", then with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 211, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 183, + 213 + ], + "score": 1.0, + "content": "probability at least", + "type": "text" + }, + { + "bbox": [ + 183, + 201, + 206, + 212 + ], + "score": 0.86, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 201, + 211, + 213 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 104, + 177, + 506, + 213 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 139, + 223, + 472, + 286 + ], + "lines": [ + { + "bbox": [ + 139, + 223, + 472, + 286 + ], + "spans": [ + { + "bbox": [ + 139, + 223, + 472, + 286 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\| _ { F } \\leqslant C ^ { \\prime } \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } , } \\\\ & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\widehat { \\mathbf { V } } _ { r _ { k } } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant C ^ { \\prime \\prime } \\bigg [ \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M ^ { 2 } r _ { k } \\sqrt { \\frac { \\log ( r _ { k } / \\delta ) } { n } } \\bigg ] . } \\end{array}", + "type": "interline_equation", + "image_path": "a68d8f2a70e385c01fe2cca5a61b71ac85ee1277f50aebc5731f18c8e0346861.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 139, + 223, + 472, + 244.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 139, + 244.0, + 472, + 265.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 139, + 265.0, + 472, + 286.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 300, + 504, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 314 + ], + "score": 1.0, + "content": "The following two lemmas are key to characterize the dynamics of the residual throughout training.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "These two results, especially Lemma B.4, are the ones that distinguish our analysis from previous", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "works on neural network training in the NTK regime (Du et al., 2018b; Su and Yang, 2019), since", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 334, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 479, + 347 + ], + "score": 1.0, + "content": "our analysis provides more careful characterization on the residual along different directions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 299, + 506, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 352, + 504, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 355, + 366 + ], + "score": 1.0, + "content": "Lemma B.3. Suppose that the iterates of gradient descent", + "type": "text" + }, + { + "bbox": [ + 356, + 351, + 423, + 364 + ], + "score": 0.91, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 350, + 506, + 366 + ], + "score": 1.0, + "content": "are inside the ball", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 363, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 107, + 364, + 156, + 377 + ], + "score": 0.91, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 363, + 169, + 379 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 170, + 364, + 259, + 377 + ], + "score": 0.93, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 363, + 380, + 379 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 380, + 364, + 501, + 378 + ], + "score": 0.9, + "content": "1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 363, + 505, + 379 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 350, + 506, + 379 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 114, + 388, + 496, + 405 + ], + "lines": [ + { + "bbox": [ + 114, + 388, + 496, + 405 + ], + "spans": [ + { + "bbox": [ + 114, + 388, + 496, + 405 + ], + "score": 0.83, + "content": "\\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } + 1 ) } = [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) + \\mathbf { e } ^ { ( t ) } , \\ \\| \\mathbf { e } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 }", + "type": "interline_equation", + "image_path": "33ee9e550d82b16b8bcfad2d23c97bd171cd4e1bd0feb467f5f6c86f3c38ad34.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 388, + 496, + 405 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 416, + 487, + 432 + ], + "lines": [ + { + "bbox": [ + 104, + 416, + 486, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 133, + 434 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 419, + 201, + 430 + ], + "score": 0.93, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 416, + 232, + 434 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 232, + 416, + 486, + 432 + ], + "score": 0.68, + "content": "\\mathbf { y } = \\left( y _ { 1 } , \\ldots , y _ { n } \\right) ^ { \\top } , \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } = \\theta \\cdot \\left( f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { n } ) \\right) ^ { \\top } .", + "type": "inline_equation" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 416, + 486, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 437, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 104, + 435, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 355, + 451 + ], + "score": 1.0, + "content": "Lemma B.4. Suppose that the iterates of gradient descent", + "type": "text" + }, + { + "bbox": [ + 356, + 437, + 423, + 450 + ], + "score": 0.91, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 435, + 506, + 451 + ], + "score": 1.0, + "content": "are inside the ball", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 448, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 107, + 450, + 156, + 464 + ], + "score": 0.93, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 448, + 170, + 466 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 171, + 450, + 342, + 464 + ], + "score": 0.91, + "content": "\\omega \\leqslant \\widetilde { \\mathcal { O } } ( \\operatorname* { m i n } \\{ [ \\log ( m ) ] ^ { - 3 / 2 } , \\lambda _ { r _ { k } } ^ { 3 } , ( \\eta m ) ^ { - 3 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 448, + 361, + 466 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 361, + 450, + 414, + 465 + ], + "score": 0.93, + "content": "n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 448, + 506, + 466 + ], + "score": 1.0, + "content": ", then with probability", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 461, + 273, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 137, + 479 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 137, + 464, + 268, + 477 + ], + "score": 0.9, + "content": "1 - \\mathcal { O } ( t ^ { 2 } n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 461, + 273, + 479 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 435, + 506, + 479 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 486, + 514, + 576 + ], + "lines": [ + { + "bbox": [ + 111, + 486, + 514, + 576 + ], + "spans": [ + { + "bbox": [ + 111, + 486, + 514, + 576 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 1 ) } } \\\\ & { } & { \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) t ^ { \\prime } \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r k } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { n } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } \\cdot \\widetilde { O } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { } & { \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } ) \\cdot \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) ^ { t ^ { \\prime } } + \\widetilde { \\mathcal { O } } ( \\sqrt { n } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r k } ) ^ { - 1 } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 3 ) } \\end{array}", + "type": "interline_equation", + "image_path": "82373c60e69c8dbaa0ed81f5468aea31025051dbee60a3ac6b8b195bb9e56718.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 111, + 486, + 514, + 516.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 111, + 516.0, + 514, + 546.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 546.0, + 514, + 576.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 204, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 584, + 201, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 133, + 600 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 586, + 201, + 598 + ], + "score": 0.91, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 584, + 201, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 270, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 272, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 272, + 623 + ], + "score": 1.0, + "content": "Now we are ready to prove Theorem 3.2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 610, + 272, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 653, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 229, + 667 + ], + "score": 1.0, + "content": "Proof of Theorem 3.2. Define", + "type": "text" + }, + { + "bbox": [ + 230, + 654, + 315, + 666 + ], + "score": 0.91, + "content": "\\omega \\ = \\ \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 652, + 492, + 667 + ], + "score": 1.0, + "content": "for some small enough absolute constant", + "type": "text" + }, + { + "bbox": [ + 492, + 654, + 501, + 664 + ], + "score": 0.82, + "content": "\\overline { { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 652, + 505, + 667 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 664, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 244, + 684 + ], + "score": 1.0, + "content": "Then by union bound, as long as", + "type": "text" + }, + { + "bbox": [ + 244, + 667, + 356, + 681 + ], + "score": 0.91, + "content": "m \\geqslant \\widetilde \\Omega ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 664, + 435, + 684 + ], + "score": 1.0, + "content": ", the conditions on", + "type": "text" + }, + { + "bbox": [ + 435, + 670, + 443, + 678 + ], + "score": 0.73, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 664, + 507, + 684 + ], + "score": 1.0, + "content": "given in Lem-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 678, + 299, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 299, + 691 + ], + "score": 1.0, + "content": "mas D.2, D.4, D.5, B.3 and B.4 are all satisfied.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 652, + 507, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 694, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 104, + 691, + 506, + 710 + ], + "spans": [ + { + "bbox": [ + 104, + 691, + 243, + 710 + ], + "score": 1.0, + "content": "We first show that all the iterates", + "type": "text" + }, + { + "bbox": [ + 244, + 695, + 313, + 707 + ], + "score": 0.9, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 691, + 391, + 710 + ], + "score": 1.0, + "content": "are inside the ball", + "type": "text" + }, + { + "bbox": [ + 392, + 695, + 440, + 708 + ], + "score": 0.93, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 691, + 506, + 710 + ], + "score": 1.0, + "content": ". We prove this", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 706, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 230, + 720 + ], + "score": 1.0, + "content": "result by inductively show that", + "type": "text" + }, + { + "bbox": [ + 230, + 707, + 368, + 720 + ], + "score": 0.84, + "content": "\\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) , t = 0 , \\ldots , T", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 706, + 471, + 720 + ], + "score": 1.0, + "content": ". First of all, it is clear that", + "type": "text" + }, + { + "bbox": [ + 472, + 707, + 505, + 719 + ], + "score": 0.89, + "content": "{ \\bf W } ^ { ( 0 ) } \\in", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 717, + 507, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 719, + 156, + 732 + ], + "score": 0.93, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 717, + 216, + 734 + ], + "score": 1.0, + "content": ". Suppose that", + "type": "text" + }, + { + "bbox": [ + 216, + 720, + 344, + 732 + ], + "score": 0.9, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 717, + 507, + 734 + ], + "score": 1.0, + "content": ". Then the results of Lemmas D.2, D.4,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 691, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 80, + 468, + 95 + ], + "lines": [ + { + "bbox": [ + 103, + 77, + 470, + 97 + ], + "spans": [ + { + "bbox": [ + 103, + 77, + 210, + 97 + ], + "score": 1.0, + "content": "D.5, B.3 and B.4 hold for", + "type": "text" + }, + { + "bbox": [ + 211, + 81, + 279, + 94 + ], + "score": 0.91, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 77, + 314, + 97 + ], + "score": 1.0, + "content": ". Denote", + "type": "text" + }, + { + "bbox": [ + 314, + 81, + 379, + 94 + ], + "score": 0.9, + "content": "\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 77, + 382, + 97 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 383, + 82, + 406, + 93 + ], + "score": 0.85, + "content": "t \\in T", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 77, + 470, + 97 + ], + "score": 1.0, + "content": ". Then we have", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 99, + 468, + 266 + ], + "lines": [ + { + "bbox": [ + 144, + 99, + 468, + 266 + ], + "spans": [ + { + "bbox": [ + 144, + 99, + 468, + 266 + ], + "score": 0.96, + "content": "\\begin{array} { r l } { \\| \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( 0 ) } \\| _ { F } \\leqslant } & { \\displaystyle \\sum _ { \\ell = 1 } ^ { t } | \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( t ^ { \\prime } ) } | \\| _ { F } } \\\\ & { \\qquad = \\displaystyle \\prod _ { \\ell = 0 } ^ { t } \\left| \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\left( g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right) \\cdot \\vartheta \\cdot \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right| \\Bigg | _ { F } } \\\\ & { \\leqslant \\eta \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| \\cdot \\| \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n - 1 } \\| \\mathbf { y } - \\vartheta \\cdot \\nabla _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\| \\mathbf { y } - \\widehat { \\mathbf { g } } ^ { ( \\ell ^ { \\prime } ) } \\| _ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "a4a45bbc2922014d5419d9875c28672e315f8cfe28ac3e78a5c0b5c817a6f601.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 144, + 99, + 468, + 154.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 144, + 154.66666666666666, + 468, + 210.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 144, + 210.33333333333331, + 468, + 266.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 420, + 280 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 420, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 420, + 281 + ], + "score": 1.0, + "content": "where the second inequality follows by Lemma D.4. By Lemma B.4, we have", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 104, + 283, + 500, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 283, + 500, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 283, + 500, + 315 + ], + "score": 0.83, + "content": "\\sum _ { t ^ { \\prime } = 0 } ^ { t } \\| \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\widetilde { \\mathcal { O } } ( T \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\lambda _ { r _ { k } } ^ { - 1 } T ^ { 2 } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } )", + "type": "interline_equation", + "image_path": "9b48a0e78cb33a76c618dd9a713c233217aad8b7a4c0a81a2badba5110f5424c.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 104, + 283, + 500, + 315 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 104, + 318, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 231, + 336 + ], + "score": 1.0, + "content": "It then follows by the choice", + "type": "text" + }, + { + "bbox": [ + 231, + 321, + 321, + 334 + ], + "score": 0.88, + "content": "\\omega \\ = \\ \\overline { { C } } T / ( \\theta \\lambda _ { r _ { k } } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 318, + 327, + 336 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 327, + 320, + 418, + 334 + ], + "score": 0.88, + "content": "\\eta ~ = ~ { \\tilde { \\mathcal { O } } } ( ( m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 318, + 424, + 336 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 425, + 320, + 469, + 334 + ], + "score": 0.9, + "content": "\\theta \\ : = \\ : \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 318, + 506, + 336 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 330, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 155, + 351 + ], + "score": 1.0, + "content": "assumption", + "type": "text" + }, + { + "bbox": [ + 156, + 334, + 271, + 349 + ], + "score": 0.93, + "content": "m \\geqslant \\tilde { \\mathcal { O } } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 330, + 293, + 351 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 293, + 334, + 400, + 348 + ], + "score": 0.88, + "content": "\\| \\mathbf { W } _ { l } ^ { ( t + 1 ) } - \\mathbf { W } _ { l } ^ { ( 0 ) } \\| _ { F } \\leqslant \\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 330, + 406, + 351 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 406, + 336, + 442, + 347 + ], + "score": 0.46, + "content": "l = 1 , 2", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 330, + 507, + 351 + ], + "score": 1.0, + "content": ". Therefore by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 347, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 290, + 363 + ], + "score": 1.0, + "content": "induction, we see that with probability at least", + "type": "text" + }, + { + "bbox": [ + 290, + 349, + 415, + 362 + ], + "score": 0.79, + "content": "1 { - } \\mathcal { O } ( T ^ { 3 } n ^ { 2 } ) { \\cdot } \\exp [ { - \\Omega ( m \\omega ^ { 2 / 3 } ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 347, + 420, + 363 + ], + "score": 1.0, + "content": "s,", + "type": "text" + }, + { + "bbox": [ + 421, + 348, + 505, + 361 + ], + "score": 0.74, + "content": "\\mathbf { W } ( 0 ) , \\ldots , \\mathbf { W } ( T ) \\in", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 358, + 164, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 156, + 374 + ], + "score": 0.9, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 358, + 164, + 376 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 239, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 239, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 239, + 392 + ], + "score": 1.0, + "content": "Applying Lemma B.4 then gives", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 392, + 463, + 446 + ], + "lines": [ + { + "bbox": [ + 147, + 392, + 463, + 446 + ], + "spans": [ + { + "bbox": [ + 147, + 392, + 463, + 446 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { \\qquad + T \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\qquad + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot n ^ { - 1 / 2 } \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } . } \\end{array}", + "type": "interline_equation", + "image_path": "878d859ddd6754ed37a3a15388ddc44c999cba13c8b28854e79eac451201f04b.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 147, + 392, + 463, + 410.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 147, + 410.0, + 463, + 428.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 147, + 428.0, + 463, + 446.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 448, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 104, + 447, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 104, + 447, + 141, + 464 + ], + "score": 1.0, + "content": "Now by", + "type": "text" + }, + { + "bbox": [ + 141, + 449, + 222, + 462 + ], + "score": 0.71, + "content": "\\omega = \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 447, + 225, + 464 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 226, + 448, + 290, + 462 + ], + "score": 0.84, + "content": "\\eta = \\widetilde { \\mathcal { O } } ( \\theta ^ { 2 } m ) ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 447, + 390, + 464 + ], + "score": 1.0, + "content": "and the assumption that", + "type": "text" + }, + { + "bbox": [ + 391, + 448, + 501, + 463 + ], + "score": 0.91, + "content": "m \\geqslant m ^ { * } = \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 1 4 } \\cdot \\epsilon ^ { - 6 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 447, + 506, + 464 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 460, + 149, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 149, + 474 + ], + "score": 1.0, + "content": "we obtain", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 473, + 464, + 491 + ], + "lines": [ + { + "bbox": [ + 150, + 473, + 464, + 491 + ], + "spans": [ + { + "bbox": [ + 150, + 473, + 464, + 491 + ], + "score": 0.89, + "content": "\\begin{array} { r } { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon / 1 6 . } \\end{array}", + "type": "interline_equation", + "image_path": "f67a04ba5eeaab3dc35a4c3983140fcfabb1c9fb2c9089446e0ef81420c23f3f.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 150, + 473, + 464, + 491 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 508, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 171, + 510 + ], + "score": 1.0, + "content": "By Lemma 3.1,", + "type": "text" + }, + { + "bbox": [ + 171, + 494, + 211, + 509 + ], + "score": 0.93, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 493, + 295, + 510 + ], + "score": 1.0, + "content": "and the assumptions", + "type": "text" + }, + { + "bbox": [ + 296, + 495, + 485, + 509 + ], + "score": 0.91, + "content": "n \\geqslant \\widetilde \\Omega ( \\operatorname* { m a x } \\{ \\epsilon ^ { - 1 } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } , \\epsilon ^ { - 2 } M ^ { 2 } r _ { k } ^ { 2 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 493, + 506, + 510 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 507, + 466, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 166, + 523 + ], + "score": 1.0, + "content": "eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 167, + 509, + 201, + 522 + ], + "score": 0.92, + "content": "\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 507, + 266, + 523 + ], + "score": 1.0, + "content": "are all between", + "type": "text" + }, + { + "bbox": [ + 266, + 508, + 290, + 521 + ], + "score": 0.92, + "content": "1 / \\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 507, + 308, + 523 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 508, + 322, + 520 + ], + "score": 0.89, + "content": "\\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 507, + 466, + 523 + ], + "score": 1.0, + "content": ". Therefore by Lemma B.2 we have", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 524, + 514, + 608 + ], + "lines": [ + { + "bbox": [ + 111, + 524, + 514, + 608 + ], + "spans": [ + { + "bbox": [ + 111, + 524, + 514, + 608 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } - \\| ( \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ) ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\sqrt { n } \\cdot \\mathcal { O } \\biggl ( \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M r _ { k } \\sqrt { \\frac { \\log ( r _ { k } } { n } } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\epsilon \\sqrt { n } / 1 6 . } \\end{array}", + "type": "interline_equation", + "image_path": "d4a7d0dc2c257dba314e6d313f46ae5d2eb36d0af3c270e18e1598bf0a21dc51.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 111, + 524, + 514, + 552.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 111, + 552.0, + 514, + 580.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 111, + 580.0, + 514, + 608.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 147, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 149, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 149, + 623 + ], + "score": 1.0, + "content": "Similarly,", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 623, + 442, + 658 + ], + "lines": [ + { + "bbox": [ + 169, + 623, + 442, + 658 + ], + "spans": [ + { + "bbox": [ + 169, + 623, + 442, + 658 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 } \\\\ & { \\qquad \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 . } \\end{array}", + "type": "interline_equation", + "image_path": "bc0943e4d670ca144524e5231af3a4a90370f07e688b3492ef3c239e2ee492d4.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 169, + 623, + 442, + 634.6666666666666 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 169, + 634.6666666666666, + 442, + 646.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 169, + 646.3333333333333, + 442, + 657.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 661, + 506, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 270, + 676 + ], + "score": 1.0, + "content": "By Lemma 3.1, with probability at least", + "type": "text" + }, + { + "bbox": [ + 270, + 662, + 293, + 673 + ], + "score": 0.64, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 659, + 298, + 676 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 298, + 661, + 452, + 675 + ], + "score": 0.91, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant 1 + C r _ { k } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 659, + 506, + 676 + ], + "score": 1.0, + "content": ". Combining", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 674, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 242, + 690 + ], + "score": 1.0, + "content": "this result with Lemma D.3 gives", + "type": "text" + }, + { + "bbox": [ + 243, + 675, + 420, + 689 + ], + "score": 0.91, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } \\leqslant \\theta \\mathcal { O } ( \\sqrt { n \\log \\left( n / \\delta \\right) } ) \\leqslant \\epsilon \\sqrt { n } / 8", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 674, + 506, + 690 + ], + "score": 1.0, + "content": ". Plugging the above", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 687, + 211, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 211, + 700 + ], + "score": 1.0, + "content": "estimates into (B.4) gives", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 702, + 438, + 718 + ], + "lines": [ + { + "bbox": [ + 173, + 702, + 438, + 718 + ], + "spans": [ + { + "bbox": [ + 173, + 702, + 438, + 718 + ], + "score": 0.89, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon .", + "type": "interline_equation", + "image_path": "0a6b9784a608f97249de05f6d09bac2302a17889aafa041d0472362b42bf2ec2.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 173, + 702, + 438, + 718 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 286, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 288, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 288, + 734 + ], + "score": 1.0, + "content": "Applying union bounds completes the proof.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 721, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 496, + 722, + 505, + 731 + ], + "spans": [ + { + "bbox": [ + 496, + 722, + 505, + 731 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 80, + 468, + 95 + ], + "lines": [ + { + "bbox": [ + 103, + 77, + 470, + 97 + ], + "spans": [ + { + "bbox": [ + 103, + 77, + 210, + 97 + ], + "score": 1.0, + "content": "D.5, B.3 and B.4 hold for", + "type": "text" + }, + { + "bbox": [ + 211, + 81, + 279, + 94 + ], + "score": 0.91, + "content": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 77, + 314, + 97 + ], + "score": 1.0, + "content": ". Denote", + "type": "text" + }, + { + "bbox": [ + 314, + 81, + 379, + 94 + ], + "score": 0.9, + "content": "\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 77, + 382, + 97 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 383, + 82, + 406, + 93 + ], + "score": 0.85, + "content": "t \\in T", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 77, + 470, + 97 + ], + "score": 1.0, + "content": ". Then we have", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 103, + 77, + 470, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 99, + 468, + 266 + ], + "lines": [ + { + "bbox": [ + 144, + 99, + 468, + 266 + ], + "spans": [ + { + "bbox": [ + 144, + 99, + 468, + 266 + ], + "score": 0.96, + "content": "\\begin{array} { r l } { \\| \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( 0 ) } \\| _ { F } \\leqslant } & { \\displaystyle \\sum _ { \\ell = 1 } ^ { t } | \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( t ^ { \\prime } ) } | \\| _ { F } } \\\\ & { \\qquad = \\displaystyle \\prod _ { \\ell = 0 } ^ { t } \\left| \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\left( g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right) \\cdot \\vartheta \\cdot \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right| \\Bigg | _ { F } } \\\\ & { \\leqslant \\eta \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| \\cdot \\| \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n - 1 } \\| \\mathbf { y } - \\vartheta \\cdot \\nabla _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\| \\mathbf { y } - \\widehat { \\mathbf { g } } ^ { ( \\ell ^ { \\prime } ) } \\| _ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "a4a45bbc2922014d5419d9875c28672e315f8cfe28ac3e78a5c0b5c817a6f601.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 144, + 99, + 468, + 154.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 144, + 154.66666666666666, + 468, + 210.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 144, + 210.33333333333331, + 468, + 266.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 420, + 280 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 420, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 420, + 281 + ], + "score": 1.0, + "content": "where the second inequality follows by Lemma D.4. By Lemma B.4, we have", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 267, + 420, + 281 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 104, + 283, + 500, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 283, + 500, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 283, + 500, + 315 + ], + "score": 0.83, + "content": "\\sum _ { t ^ { \\prime } = 0 } ^ { t } \\| \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\widetilde { \\mathcal { O } } ( T \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\lambda _ { r _ { k } } ^ { - 1 } T ^ { 2 } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } )", + "type": "interline_equation", + "image_path": "9b48a0e78cb33a76c618dd9a713c233217aad8b7a4c0a81a2badba5110f5424c.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 104, + 283, + 500, + 315 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 104, + 318, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 231, + 336 + ], + "score": 1.0, + "content": "It then follows by the choice", + "type": "text" + }, + { + "bbox": [ + 231, + 321, + 321, + 334 + ], + "score": 0.88, + "content": "\\omega \\ = \\ \\overline { { C } } T / ( \\theta \\lambda _ { r _ { k } } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 318, + 327, + 336 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 327, + 320, + 418, + 334 + ], + "score": 0.88, + "content": "\\eta ~ = ~ { \\tilde { \\mathcal { O } } } ( ( m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 318, + 424, + 336 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 425, + 320, + 469, + 334 + ], + "score": 0.9, + "content": "\\theta \\ : = \\ : \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 318, + 506, + 336 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 330, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 155, + 351 + ], + "score": 1.0, + "content": "assumption", + "type": "text" + }, + { + "bbox": [ + 156, + 334, + 271, + 349 + ], + "score": 0.93, + "content": "m \\geqslant \\tilde { \\mathcal { O } } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 330, + 293, + 351 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 293, + 334, + 400, + 348 + ], + "score": 0.88, + "content": "\\| \\mathbf { W } _ { l } ^ { ( t + 1 ) } - \\mathbf { W } _ { l } ^ { ( 0 ) } \\| _ { F } \\leqslant \\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 330, + 406, + 351 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 406, + 336, + 442, + 347 + ], + "score": 0.46, + "content": "l = 1 , 2", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 330, + 507, + 351 + ], + "score": 1.0, + "content": ". Therefore by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 347, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 290, + 363 + ], + "score": 1.0, + "content": "induction, we see that with probability at least", + "type": "text" + }, + { + "bbox": [ + 290, + 349, + 415, + 362 + ], + "score": 0.79, + "content": "1 { - } \\mathcal { O } ( T ^ { 3 } n ^ { 2 } ) { \\cdot } \\exp [ { - \\Omega ( m \\omega ^ { 2 / 3 } ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 347, + 420, + 363 + ], + "score": 1.0, + "content": "s,", + "type": "text" + }, + { + "bbox": [ + 421, + 348, + 505, + 361 + ], + "score": 0.74, + "content": "\\mathbf { W } ( 0 ) , \\ldots , \\mathbf { W } ( T ) \\in", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 358, + 164, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 156, + 374 + ], + "score": 0.9, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 358, + 164, + 376 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 318, + 507, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 239, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 239, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 239, + 392 + ], + "score": 1.0, + "content": "Applying Lemma B.4 then gives", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 378, + 239, + 392 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 392, + 463, + 446 + ], + "lines": [ + { + "bbox": [ + 147, + 392, + 463, + 446 + ], + "spans": [ + { + "bbox": [ + 147, + 392, + 463, + 446 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { \\qquad + T \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\qquad + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot n ^ { - 1 / 2 } \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } . } \\end{array}", + "type": "interline_equation", + "image_path": "878d859ddd6754ed37a3a15388ddc44c999cba13c8b28854e79eac451201f04b.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 147, + 392, + 463, + 410.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 147, + 410.0, + 463, + 428.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 147, + 428.0, + 463, + 446.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 448, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 104, + 447, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 104, + 447, + 141, + 464 + ], + "score": 1.0, + "content": "Now by", + "type": "text" + }, + { + "bbox": [ + 141, + 449, + 222, + 462 + ], + "score": 0.71, + "content": "\\omega = \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 447, + 225, + 464 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 226, + 448, + 290, + 462 + ], + "score": 0.84, + "content": "\\eta = \\widetilde { \\mathcal { O } } ( \\theta ^ { 2 } m ) ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 447, + 390, + 464 + ], + "score": 1.0, + "content": "and the assumption that", + "type": "text" + }, + { + "bbox": [ + 391, + 448, + 501, + 463 + ], + "score": 0.91, + "content": "m \\geqslant m ^ { * } = \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 1 4 } \\cdot \\epsilon ^ { - 6 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 447, + 506, + 464 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 460, + 149, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 149, + 474 + ], + "score": 1.0, + "content": "we obtain", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 447, + 506, + 474 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 473, + 464, + 491 + ], + "lines": [ + { + "bbox": [ + 150, + 473, + 464, + 491 + ], + "spans": [ + { + "bbox": [ + 150, + 473, + 464, + 491 + ], + "score": 0.89, + "content": "\\begin{array} { r } { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon / 1 6 . } \\end{array}", + "type": "interline_equation", + "image_path": "f67a04ba5eeaab3dc35a4c3983140fcfabb1c9fb2c9089446e0ef81420c23f3f.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 150, + 473, + 464, + 491 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 508, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 171, + 510 + ], + "score": 1.0, + "content": "By Lemma 3.1,", + "type": "text" + }, + { + "bbox": [ + 171, + 494, + 211, + 509 + ], + "score": 0.93, + "content": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 493, + 295, + 510 + ], + "score": 1.0, + "content": "and the assumptions", + "type": "text" + }, + { + "bbox": [ + 296, + 495, + 485, + 509 + ], + "score": 0.91, + "content": "n \\geqslant \\widetilde \\Omega ( \\operatorname* { m a x } \\{ \\epsilon ^ { - 1 } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } , \\epsilon ^ { - 2 } M ^ { 2 } r _ { k } ^ { 2 } \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 493, + 506, + 510 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 507, + 466, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 166, + 523 + ], + "score": 1.0, + "content": "eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 167, + 509, + 201, + 522 + ], + "score": 0.92, + "content": "\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 507, + 266, + 523 + ], + "score": 1.0, + "content": "are all between", + "type": "text" + }, + { + "bbox": [ + 266, + 508, + 290, + 521 + ], + "score": 0.92, + "content": "1 / \\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 507, + 308, + 523 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 508, + 322, + 520 + ], + "score": 0.89, + "content": "\\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 507, + 466, + 523 + ], + "score": 1.0, + "content": ". Therefore by Lemma B.2 we have", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 493, + 506, + 523 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 524, + 514, + 608 + ], + "lines": [ + { + "bbox": [ + 111, + 524, + 514, + 608 + ], + "spans": [ + { + "bbox": [ + 111, + 524, + 514, + 608 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } - \\| ( \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ) ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\sqrt { n } \\cdot \\mathcal { O } \\biggl ( \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M r _ { k } \\sqrt { \\frac { \\log ( r _ { k } } { n } } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\epsilon \\sqrt { n } / 1 6 . } \\end{array}", + "type": "interline_equation", + "image_path": "d4a7d0dc2c257dba314e6d313f46ae5d2eb36d0af3c270e18e1598bf0a21dc51.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 111, + 524, + 514, + 552.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 111, + 552.0, + 514, + 580.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 111, + 580.0, + 514, + 608.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 147, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 149, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 149, + 623 + ], + "score": 1.0, + "content": "Similarly,", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 606, + 149, + 623 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 623, + 442, + 658 + ], + "lines": [ + { + "bbox": [ + 169, + 623, + 442, + 658 + ], + "spans": [ + { + "bbox": [ + 169, + 623, + 442, + 658 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 } \\\\ & { \\qquad \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 . } \\end{array}", + "type": "interline_equation", + "image_path": "bc0943e4d670ca144524e5231af3a4a90370f07e688b3492ef3c239e2ee492d4.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 169, + 623, + 442, + 634.6666666666666 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 169, + 634.6666666666666, + 442, + 646.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 169, + 646.3333333333333, + 442, + 657.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 661, + 506, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 270, + 676 + ], + "score": 1.0, + "content": "By Lemma 3.1, with probability at least", + "type": "text" + }, + { + "bbox": [ + 270, + 662, + 293, + 673 + ], + "score": 0.64, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 659, + 298, + 676 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 298, + 661, + 452, + 675 + ], + "score": 0.91, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant 1 + C r _ { k } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 659, + 506, + 676 + ], + "score": 1.0, + "content": ". Combining", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 674, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 242, + 690 + ], + "score": 1.0, + "content": "this result with Lemma D.3 gives", + "type": "text" + }, + { + "bbox": [ + 243, + 675, + 420, + 689 + ], + "score": 0.91, + "content": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } \\leqslant \\theta \\mathcal { O } ( \\sqrt { n \\log \\left( n / \\delta \\right) } ) \\leqslant \\epsilon \\sqrt { n } / 8", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 674, + 506, + 690 + ], + "score": 1.0, + "content": ". Plugging the above", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 687, + 211, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 211, + 700 + ], + "score": 1.0, + "content": "estimates into (B.4) gives", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 659, + 506, + 700 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 702, + 438, + 718 + ], + "lines": [ + { + "bbox": [ + 173, + 702, + 438, + 718 + ], + "spans": [ + { + "bbox": [ + 173, + 702, + 438, + 718 + ], + "score": 0.89, + "content": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon .", + "type": "interline_equation", + "image_path": "0a6b9784a608f97249de05f6d09bac2302a17889aafa041d0472362b42bf2ec2.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 173, + 702, + 438, + 718 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 286, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 288, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 288, + 734 + ], + "score": 1.0, + "content": "Applying union bounds completes the proof.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 719, + 288, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 258, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 259, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 259, + 94 + ], + "score": 1.0, + "content": "B.2 PROOF OF THE THEOREM 3.4", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 504, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 1.0, + "content": "Proof of the Theorem 3.4. The idea of the proof is close to that of Proposition 5 in (Bietti and Mairal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 497, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 214, + 127 + ], + "score": 1.0, + "content": "2019) where they consider", + "type": "text" + }, + { + "bbox": [ + 215, + 115, + 240, + 125 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 114, + 425, + 127 + ], + "score": 1.0, + "content": "and we present a more general case including ş", + "type": "text" + }, + { + "bbox": [ + 425, + 115, + 451, + 125 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 114, + 468, + 127 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 469, + 115, + 493, + 125 + ], + "score": 0.89, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 114, + 497, + 127 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 175, + 139 + ], + "score": 1.0, + "content": "For any function", + "type": "text" + }, + { + "bbox": [ + 175, + 125, + 222, + 137 + ], + "score": 0.92, + "content": "g : \\mathbb { S } ^ { d } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 124, + 276, + 139 + ], + "score": 1.0, + "content": ", by denoting", + "type": "text" + }, + { + "bbox": [ + 276, + 125, + 375, + 139 + ], + "score": 0.93, + "content": "g _ { 0 } ( \\mathbf x ) = \\int _ { \\mathbb { S } ^ { d } } g ( \\mathbf { \\bar { y } } ) d \\tau _ { d } ( \\mathbf y )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 124, + 479, + 139 + ], + "score": 1.0, + "content": ", it can be decomposed as", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 181, + 142, + 430, + 212 + ], + "lines": [ + { + "bbox": [ + 181, + 142, + 430, + 212 + ], + "spans": [ + { + "bbox": [ + 181, + 142, + 430, + 212 + ], + "score": 0.95, + "content": "\\begin{array} { l } { { \\displaystyle g ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } g _ { k } ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } \\sum _ { j = 1 } ^ { N ( d , k ) } \\int _ { { \\mathbb { S } } ^ { d } } Y _ { k j } ( { \\bf y } ) Y _ { k j } ( { \\bf x } ) g ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) } } \\\\ { { \\displaystyle ~ = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { { \\mathbb { S } } ^ { d } } g ( { \\bf y } ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "3b9040066dffa2eae94da91164b6a92c93f503dd7cec4c4f37b87bbaf14d8f73.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 181, + 142, + 430, + 165.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 181, + 165.33333333333334, + 430, + 188.66666666666669 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 181, + 188.66666666666669, + 430, + 212.00000000000003 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 216, + 226 + ], + "score": 1.0, + "content": "where we project function", + "type": "text" + }, + { + "bbox": [ + 216, + 216, + 223, + 226 + ], + "score": 0.83, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "to spherical harmonics in the second equality and apply the addition", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 225, + 221, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 221, + 238 + ], + "score": 1.0, + "content": "equation in the last equality.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 108, + 236, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 504, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 275, + 249 + ], + "score": 1.0, + "content": "For a positive-definite dot-product kernel", + "type": "text" + }, + { + "bbox": [ + 275, + 235, + 376, + 248 + ], + "score": 0.92, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 233, + 459, + 249 + ], + "score": 1.0, + "content": "which has the form", + "type": "text" + }, + { + "bbox": [ + 459, + 236, + 504, + 247 + ], + "score": 0.88, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) =", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 245, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 148, + 258 + ], + "score": 0.9, + "content": "\\widehat { \\kappa } ( \\langle \\mathbf { x } , \\mathbf { \\bar { x } } ^ { \\prime } \\rangle )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 245, + 165, + 259 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 166, + 246, + 237, + 258 + ], + "score": 0.9, + "content": "\\widehat { \\kappa } : [ - 1 , 1 ] \\ \\mathrm { ~ ~ } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 245, + 472, + 259 + ], + "score": 1.0, + "content": ", we can present a decomposition by (B.5) if we consider", + "type": "text" + }, + { + "bbox": [ + 472, + 248, + 504, + 259 + ], + "score": 0.87, + "content": "g ( \\mathbf { x } ) =", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 257, + 279, + 270 + ], + "spans": [ + { + "bbox": [ + 107, + 258, + 145, + 270 + ], + "score": 0.91, + "content": "\\phi ( \\langle \\mathbf { x } , \\mathbf { z } \\rangle )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 257, + 160, + 270 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 258, + 188, + 268 + ], + "score": 0.86, + "content": "\\mathbf { z } \\in \\mathbb { S } ^ { \\bar { d } }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 257, + 206, + 270 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 207, + 258, + 273, + 270 + ], + "score": 0.91, + "content": "\\bar { \\phi } : [ - 1 , 1 ] \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 257, + 279, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 273, + 454, + 340 + ], + "lines": [ + { + "bbox": [ + 156, + 273, + 454, + 340 + ], + "spans": [ + { + "bbox": [ + 156, + 273, + 454, + 340 + ], + "score": 0.95, + "content": "\\begin{array} { l } { { \\displaystyle \\kappa ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { \\mathbb { S } ^ { d } } \\widehat { \\kappa } ( \\langle { \\bf y } , { \\bf x } ^ { \\prime } \\rangle ) P _ { k } ( \\langle { \\bf y } , { \\bf x } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } \\ ~ } \\\\ { { \\displaystyle = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t } , } \\end{array}", + "type": "interline_equation", + "image_path": "51ac6887d63cfb5a9876061232d703931f01e89901461cb8b45eb821155905d5.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 156, + 273, + 454, + 295.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 156, + 295.3333333333333, + 454, + 317.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 156, + 317.66666666666663, + 454, + 339.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 341, + 506, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 504, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 472, + 355 + ], + "score": 1.0, + "content": "where we apply the Hecke-Funk formula and addition formula. By denoting ş", + "type": "text" + }, + { + "bbox": [ + 473, + 341, + 504, + 353 + ], + "score": 0.83, + "content": "\\begin{array} { r l } { \\lambda _ { k } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 352, + 420, + 368 + ], + "spans": [ + { + "bbox": [ + 107, + 352, + 279, + 368 + ], + "score": 0.92, + "content": "( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 352, + 420, + 368 + ], + "score": 1.0, + "content": "and the addition formula, we have", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 372, + 451, + 407 + ], + "lines": [ + { + "bbox": [ + 161, + 372, + 451, + 407 + ], + "spans": [ + { + "bbox": [ + 161, + 372, + 451, + 407 + ], + "score": 0.93, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } N ( d , k ) P _ { k } ( \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N ( p , k ) } Y _ { k , j } ( \\mathbf { x } ) Y _ { k , j } ( \\mathbf { x } ^ { \\prime } ) .", + "type": "interline_equation", + "image_path": "263c48b7608eb35c97c86c30186fa6fdc13a24a4e9d2c1a68805e929093c39f8.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 161, + 372, + 451, + 383.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 161, + 383.6666666666667, + 451, + 395.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 161, + 395.33333333333337, + 451, + 407.00000000000006 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 504, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 504, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 397, + 425 + ], + "score": 1.0, + "content": "This formula (B.6) is the Mercer decomposition for the kernel function", + "type": "text" + }, + { + "bbox": [ + 398, + 411, + 431, + 423 + ], + "score": 0.93, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 409, + 450, + 425 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 450, + 413, + 462, + 423 + ], + "score": 0.86, + "content": "\\mu _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 409, + 504, + 425 + ], + "score": 1.0, + "content": "is exactly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 421, + 366, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 260, + 435 + ], + "score": 1.0, + "content": "the eigenvalue of the integral operator", + "type": "text" + }, + { + "bbox": [ + 260, + 423, + 276, + 433 + ], + "score": 0.89, + "content": "L _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 421, + 289, + 435 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 290, + 422, + 320, + 434 + ], + "score": 0.95, + "content": "L _ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 421, + 366, + 435 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 437, + 410, + 464 + ], + "lines": [ + { + "bbox": [ + 200, + 437, + 410, + 464 + ], + "spans": [ + { + "bbox": [ + 200, + 437, + 410, + 464 + ], + "score": 0.92, + "content": "L _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { 2 } ( \\mathbb { S } ^ { d } ) .", + "type": "interline_equation", + "image_path": "41d6e45b4c87b313148a091f568503bd0d15e892dc6ebb4339bc7fd9bd8e4459.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 200, + 437, + 410, + 464 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 472, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 232, + 486 + ], + "score": 1.0, + "content": "By using same technique as", + "type": "text" + }, + { + "bbox": [ + 233, + 473, + 266, + 484 + ], + "score": 0.91, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 472, + 446, + 486 + ], + "score": 1.0, + "content": ", we can derive a similar expression for", + "type": "text" + }, + { + "bbox": [ + 446, + 473, + 504, + 485 + ], + "score": 0.9, + "content": "\\begin{array} { r l } { \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 484, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 127, + 497 + ], + "score": 1.0, + "content": "max", + "type": "text" + }, + { + "bbox": [ + 127, + 484, + 173, + 496 + ], + "score": 0.87, + "content": "\\{ \\langle \\mathbf { w } , \\mathbf { x } \\rangle , 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 484, + 194, + 497 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 194, + 484, + 320, + 496 + ], + "score": 0.89, + "content": "\\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\mathbb { 1 } \\{ \\langle { \\bf w } , { \\bf x } \\rangle > 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 484, + 504, + 497 + ], + "score": 1.0, + "content": ", since they are essentially dot-product func-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 493, + 410, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 137, + 509 + ], + "score": 1.0, + "content": "tion on", + "type": "text" + }, + { + "bbox": [ + 137, + 496, + 167, + 507 + ], + "score": 0.9, + "content": "L _ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 493, + 410, + 509 + ], + "score": 1.0, + "content": ". We deliver the expression below without presenting proofs.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 510, + 392, + 576 + ], + "lines": [ + { + "bbox": [ + 218, + 510, + 392, + 576 + ], + "spans": [ + { + "bbox": [ + 218, + 510, + 392, + 576 + ], + "score": 0.92, + "content": "\\begin{array} { l } { { \\displaystyle \\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\\\ { { \\displaystyle \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "cc1a0b4f9a5e173b51af87672e7a5a208b1277bbeea1854cd0901795af785898.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 510, + 392, + 526.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 218, + 526.5, + 392, + 543.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 218, + 543.0, + 392, + 559.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 218, + 559.5, + 392, + 576.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 579, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 104, + 578, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 578, + 133, + 596 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 579, + 333, + 596 + ], + "score": 0.86, + "content": "\\beta _ { 1 , k } = \\left( \\omega _ { d - 1 } / \\omega _ { d } \\right) \\int _ { - 1 } ^ { 1 } \\sigma ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 578, + 351, + 596 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 351, + 579, + 506, + 596 + ], + "score": 0.92, + "content": "\\beta _ { 2 , k } = ( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\sigma ^ { \\prime } ( t ) P _ { k } ( t ) ( 1 -", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 592, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 157, + 608 + ], + "score": 0.9, + "content": "t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 592, + 322, + 610 + ], + "score": 1.0, + "content": ". We add more comments on the values of", + "type": "text" + }, + { + "bbox": [ + 323, + 596, + 340, + 608 + ], + "score": 0.92, + "content": "\\beta _ { 1 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 592, + 358, + 610 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 358, + 596, + 376, + 608 + ], + "score": 0.9, + "content": "\\beta _ { 2 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 592, + 506, + 610 + ], + "score": 1.0, + "content": ". It has been pointed out in Bach", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 177, + 619 + ], + "score": 1.0, + "content": "(2017) that when", + "type": "text" + }, + { + "bbox": [ + 177, + 607, + 205, + 617 + ], + "score": 0.92, + "content": "k > \\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 606, + 247, + 619 + ], + "score": 1.0, + "content": "and when", + "type": "text" + }, + { + "bbox": [ + 247, + 608, + 254, + 617 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 606, + 272, + 619 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 609, + 280, + 617 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 606, + 388, + 619 + ], + "score": 1.0, + "content": "have same parity, we have", + "type": "text" + }, + { + "bbox": [ + 389, + 607, + 437, + 619 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 606, + 505, + 619 + ], + "score": 1.0, + "content": ". This is because", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 210, + 629 + ], + "score": 1.0, + "content": "the Legendre polynomial", + "type": "text" + }, + { + "bbox": [ + 211, + 618, + 235, + 630 + ], + "score": 0.93, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 617, + 476, + 629 + ], + "score": 1.0, + "content": "is orthogonal to any other polynomials of degree less than", + "type": "text" + }, + { + "bbox": [ + 476, + 618, + 483, + 628 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 628, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 233, + 644 + ], + "score": 1.0, + "content": "respect to the density function", + "type": "text" + }, + { + "bbox": [ + 234, + 629, + 329, + 642 + ], + "score": 0.9, + "content": "p ( t ) = ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 628, + 449, + 644 + ], + "score": 1.0, + "content": ". Then we clearly know that", + "type": "text" + }, + { + "bbox": [ + 450, + 630, + 489, + 643 + ], + "score": 0.92, + "content": "\\beta _ { 1 , k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 628, + 506, + 644 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 640, + 308, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 136, + 653 + ], + "score": 0.91, + "content": "k = 2 j", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 640, + 154, + 654 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 641, + 191, + 654 + ], + "score": 0.92, + "content": "\\beta _ { 2 , k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 640, + 206, + 654 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 206, + 642, + 253, + 653 + ], + "score": 0.92, + "content": "k = 2 j + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 640, + 274, + 654 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 275, + 642, + 304, + 653 + ], + "score": 0.88, + "content": "j \\in \\mathbb { N } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 640, + 308, + 654 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 303, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 304, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 304, + 671 + ], + "score": 1.0, + "content": "For two kernel function defined in (2.2), we have", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 671, + 429, + 731 + ], + "lines": [ + { + "bbox": [ + 182, + 671, + 429, + 731 + ], + "spans": [ + { + "bbox": [ + 182, + 671, + 429, + 731 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad = \\displaystyle \\int _ { \\mathbb { S } ^ { d } } \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "95d2fe547af7b1da33eeceedf59749a8a73300ef5b14f44234e9e71094039a3d.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 182, + 671, + 429, + 691.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 182, + 691.0, + 429, + 711.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 182, + 711.0, + 429, + 731.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 258, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 259, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 259, + 94 + ], + "score": 1.0, + "content": "B.2 PROOF OF THE THEOREM 3.4", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 504, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 1.0, + "content": "Proof of the Theorem 3.4. The idea of the proof is close to that of Proposition 5 in (Bietti and Mairal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 497, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 214, + 127 + ], + "score": 1.0, + "content": "2019) where they consider", + "type": "text" + }, + { + "bbox": [ + 215, + 115, + 240, + 125 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 114, + 425, + 127 + ], + "score": 1.0, + "content": "and we present a more general case including ş", + "type": "text" + }, + { + "bbox": [ + 425, + 115, + 451, + 125 + ], + "score": 0.9, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 114, + 468, + 127 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 469, + 115, + 493, + 125 + ], + "score": 0.89, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 114, + 497, + 127 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 175, + 139 + ], + "score": 1.0, + "content": "For any function", + "type": "text" + }, + { + "bbox": [ + 175, + 125, + 222, + 137 + ], + "score": 0.92, + "content": "g : \\mathbb { S } ^ { d } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 124, + 276, + 139 + ], + "score": 1.0, + "content": ", by denoting", + "type": "text" + }, + { + "bbox": [ + 276, + 125, + 375, + 139 + ], + "score": 0.93, + "content": "g _ { 0 } ( \\mathbf x ) = \\int _ { \\mathbb { S } ^ { d } } g ( \\mathbf { \\bar { y } } ) d \\tau _ { d } ( \\mathbf y )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 124, + 479, + 139 + ], + "score": 1.0, + "content": ", it can be decomposed as", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 103, + 505, + 139 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 181, + 142, + 430, + 212 + ], + "lines": [ + { + "bbox": [ + 181, + 142, + 430, + 212 + ], + "spans": [ + { + "bbox": [ + 181, + 142, + 430, + 212 + ], + "score": 0.95, + "content": "\\begin{array} { l } { { \\displaystyle g ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } g _ { k } ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } \\sum _ { j = 1 } ^ { N ( d , k ) } \\int _ { { \\mathbb { S } } ^ { d } } Y _ { k j } ( { \\bf y } ) Y _ { k j } ( { \\bf x } ) g ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) } } \\\\ { { \\displaystyle ~ = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { { \\mathbb { S } } ^ { d } } g ( { \\bf y } ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "3b9040066dffa2eae94da91164b6a92c93f503dd7cec4c4f37b87bbaf14d8f73.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 181, + 142, + 430, + 165.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 181, + 165.33333333333334, + 430, + 188.66666666666669 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 181, + 188.66666666666669, + 430, + 212.00000000000003 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 216, + 226 + ], + "score": 1.0, + "content": "where we project function", + "type": "text" + }, + { + "bbox": [ + 216, + 216, + 223, + 226 + ], + "score": 0.83, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "to spherical harmonics in the second equality and apply the addition", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 225, + 221, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 221, + 238 + ], + "score": 1.0, + "content": "equation in the last equality.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 214, + 505, + 238 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 236, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 504, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 275, + 249 + ], + "score": 1.0, + "content": "For a positive-definite dot-product kernel", + "type": "text" + }, + { + "bbox": [ + 275, + 235, + 376, + 248 + ], + "score": 0.92, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 233, + 459, + 249 + ], + "score": 1.0, + "content": "which has the form", + "type": "text" + }, + { + "bbox": [ + 459, + 236, + 504, + 247 + ], + "score": 0.88, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) =", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 245, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 148, + 258 + ], + "score": 0.9, + "content": "\\widehat { \\kappa } ( \\langle \\mathbf { x } , \\mathbf { \\bar { x } } ^ { \\prime } \\rangle )", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 245, + 165, + 259 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 166, + 246, + 237, + 258 + ], + "score": 0.9, + "content": "\\widehat { \\kappa } : [ - 1 , 1 ] \\ \\mathrm { ~ ~ } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 245, + 472, + 259 + ], + "score": 1.0, + "content": ", we can present a decomposition by (B.5) if we consider", + "type": "text" + }, + { + "bbox": [ + 472, + 248, + 504, + 259 + ], + "score": 0.87, + "content": "g ( \\mathbf { x } ) =", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 257, + 279, + 270 + ], + "spans": [ + { + "bbox": [ + 107, + 258, + 145, + 270 + ], + "score": 0.91, + "content": "\\phi ( \\langle \\mathbf { x } , \\mathbf { z } \\rangle )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 257, + 160, + 270 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 258, + 188, + 268 + ], + "score": 0.86, + "content": "\\mathbf { z } \\in \\mathbb { S } ^ { \\bar { d } }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 257, + 206, + 270 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 207, + 258, + 273, + 270 + ], + "score": 0.91, + "content": "\\bar { \\phi } : [ - 1 , 1 ] \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 257, + 279, + 270 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 233, + 504, + 270 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 273, + 454, + 340 + ], + "lines": [ + { + "bbox": [ + 156, + 273, + 454, + 340 + ], + "spans": [ + { + "bbox": [ + 156, + 273, + 454, + 340 + ], + "score": 0.95, + "content": "\\begin{array} { l } { { \\displaystyle \\kappa ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { \\mathbb { S } ^ { d } } \\widehat { \\kappa } ( \\langle { \\bf y } , { \\bf x } ^ { \\prime } \\rangle ) P _ { k } ( \\langle { \\bf y } , { \\bf x } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } \\ ~ } \\\\ { { \\displaystyle = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t } , } \\end{array}", + "type": "interline_equation", + "image_path": "51ac6887d63cfb5a9876061232d703931f01e89901461cb8b45eb821155905d5.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 156, + 273, + 454, + 295.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 156, + 295.3333333333333, + 454, + 317.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 156, + 317.66666666666663, + 454, + 339.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 341, + 506, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 504, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 472, + 355 + ], + "score": 1.0, + "content": "where we apply the Hecke-Funk formula and addition formula. By denoting ş", + "type": "text" + }, + { + "bbox": [ + 473, + 341, + 504, + 353 + ], + "score": 0.83, + "content": "\\begin{array} { r l } { \\lambda _ { k } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 352, + 420, + 368 + ], + "spans": [ + { + "bbox": [ + 107, + 352, + 279, + 368 + ], + "score": 0.92, + "content": "( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 352, + 420, + 368 + ], + "score": 1.0, + "content": "and the addition formula, we have", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 340, + 504, + 368 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 372, + 451, + 407 + ], + "lines": [ + { + "bbox": [ + 161, + 372, + 451, + 407 + ], + "spans": [ + { + "bbox": [ + 161, + 372, + 451, + 407 + ], + "score": 0.93, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } N ( d , k ) P _ { k } ( \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N ( p , k ) } Y _ { k , j } ( \\mathbf { x } ) Y _ { k , j } ( \\mathbf { x } ^ { \\prime } ) .", + "type": "interline_equation", + "image_path": "263c48b7608eb35c97c86c30186fa6fdc13a24a4e9d2c1a68805e929093c39f8.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 161, + 372, + 451, + 383.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 161, + 383.6666666666667, + 451, + 395.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 161, + 395.33333333333337, + 451, + 407.00000000000006 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 504, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 504, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 397, + 425 + ], + "score": 1.0, + "content": "This formula (B.6) is the Mercer decomposition for the kernel function", + "type": "text" + }, + { + "bbox": [ + 398, + 411, + 431, + 423 + ], + "score": 0.93, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 409, + 450, + 425 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 450, + 413, + 462, + 423 + ], + "score": 0.86, + "content": "\\mu _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 409, + 504, + 425 + ], + "score": 1.0, + "content": "is exactly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 421, + 366, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 260, + 435 + ], + "score": 1.0, + "content": "the eigenvalue of the integral operator", + "type": "text" + }, + { + "bbox": [ + 260, + 423, + 276, + 433 + ], + "score": 0.89, + "content": "L _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 421, + 289, + 435 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 290, + 422, + 320, + 434 + ], + "score": 0.95, + "content": "L _ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 421, + 366, + 435 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 409, + 504, + 435 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 437, + 410, + 464 + ], + "lines": [ + { + "bbox": [ + 200, + 437, + 410, + 464 + ], + "spans": [ + { + "bbox": [ + 200, + 437, + 410, + 464 + ], + "score": 0.92, + "content": "L _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { 2 } ( \\mathbb { S } ^ { d } ) .", + "type": "interline_equation", + "image_path": "41d6e45b4c87b313148a091f568503bd0d15e892dc6ebb4339bc7fd9bd8e4459.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 200, + 437, + 410, + 464 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 472, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 232, + 486 + ], + "score": 1.0, + "content": "By using same technique as", + "type": "text" + }, + { + "bbox": [ + 233, + 473, + 266, + 484 + ], + "score": 0.91, + "content": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 472, + 446, + 486 + ], + "score": 1.0, + "content": ", we can derive a similar expression for", + "type": "text" + }, + { + "bbox": [ + 446, + 473, + 504, + 485 + ], + "score": 0.9, + "content": "\\begin{array} { r l } { \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 484, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 127, + 497 + ], + "score": 1.0, + "content": "max", + "type": "text" + }, + { + "bbox": [ + 127, + 484, + 173, + 496 + ], + "score": 0.87, + "content": "\\{ \\langle \\mathbf { w } , \\mathbf { x } \\rangle , 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 484, + 194, + 497 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 194, + 484, + 320, + 496 + ], + "score": 0.89, + "content": "\\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\mathbb { 1 } \\{ \\langle { \\bf w } , { \\bf x } \\rangle > 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 484, + 504, + 497 + ], + "score": 1.0, + "content": ", since they are essentially dot-product func-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 493, + 410, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 137, + 509 + ], + "score": 1.0, + "content": "tion on", + "type": "text" + }, + { + "bbox": [ + 137, + 496, + 167, + 507 + ], + "score": 0.9, + "content": "L _ { 2 } ( \\mathbb { S } ^ { d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 493, + 410, + 509 + ], + "score": 1.0, + "content": ". We deliver the expression below without presenting proofs.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 472, + 504, + 509 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 510, + 392, + 576 + ], + "lines": [ + { + "bbox": [ + 218, + 510, + 392, + 576 + ], + "spans": [ + { + "bbox": [ + 218, + 510, + 392, + 576 + ], + "score": 0.92, + "content": "\\begin{array} { l } { { \\displaystyle \\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\\\ { { \\displaystyle \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "cc1a0b4f9a5e173b51af87672e7a5a208b1277bbeea1854cd0901795af785898.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 510, + 392, + 526.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 218, + 526.5, + 392, + 543.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 218, + 543.0, + 392, + 559.5 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 218, + 559.5, + 392, + 576.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 579, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 104, + 578, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 578, + 133, + 596 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 579, + 333, + 596 + ], + "score": 0.86, + "content": "\\beta _ { 1 , k } = \\left( \\omega _ { d - 1 } / \\omega _ { d } \\right) \\int _ { - 1 } ^ { 1 } \\sigma ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 578, + 351, + 596 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 351, + 579, + 506, + 596 + ], + "score": 0.92, + "content": "\\beta _ { 2 , k } = ( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\sigma ^ { \\prime } ( t ) P _ { k } ( t ) ( 1 -", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 592, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 157, + 608 + ], + "score": 0.9, + "content": "t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 592, + 322, + 610 + ], + "score": 1.0, + "content": ". We add more comments on the values of", + "type": "text" + }, + { + "bbox": [ + 323, + 596, + 340, + 608 + ], + "score": 0.92, + "content": "\\beta _ { 1 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 592, + 358, + 610 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 358, + 596, + 376, + 608 + ], + "score": 0.9, + "content": "\\beta _ { 2 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 592, + 506, + 610 + ], + "score": 1.0, + "content": ". It has been pointed out in Bach", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 177, + 619 + ], + "score": 1.0, + "content": "(2017) that when", + "type": "text" + }, + { + "bbox": [ + 177, + 607, + 205, + 617 + ], + "score": 0.92, + "content": "k > \\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 606, + 247, + 619 + ], + "score": 1.0, + "content": "and when", + "type": "text" + }, + { + "bbox": [ + 247, + 608, + 254, + 617 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 606, + 272, + 619 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 609, + 280, + 617 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 606, + 388, + 619 + ], + "score": 1.0, + "content": "have same parity, we have", + "type": "text" + }, + { + "bbox": [ + 389, + 607, + 437, + 619 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 606, + 505, + 619 + ], + "score": 1.0, + "content": ". This is because", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 210, + 629 + ], + "score": 1.0, + "content": "the Legendre polynomial", + "type": "text" + }, + { + "bbox": [ + 211, + 618, + 235, + 630 + ], + "score": 0.93, + "content": "P _ { k } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 617, + 476, + 629 + ], + "score": 1.0, + "content": "is orthogonal to any other polynomials of degree less than", + "type": "text" + }, + { + "bbox": [ + 476, + 618, + 483, + 628 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 628, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 233, + 644 + ], + "score": 1.0, + "content": "respect to the density function", + "type": "text" + }, + { + "bbox": [ + 234, + 629, + 329, + 642 + ], + "score": 0.9, + "content": "p ( t ) = ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 628, + 449, + 644 + ], + "score": 1.0, + "content": ". Then we clearly know that", + "type": "text" + }, + { + "bbox": [ + 450, + 630, + 489, + 643 + ], + "score": 0.92, + "content": "\\beta _ { 1 , k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 628, + 506, + 644 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 640, + 308, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 136, + 653 + ], + "score": 0.91, + "content": "k = 2 j", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 640, + 154, + 654 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 641, + 191, + 654 + ], + "score": 0.92, + "content": "\\beta _ { 2 , k } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 640, + 206, + 654 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 206, + 642, + 253, + 653 + ], + "score": 0.92, + "content": "k = 2 j + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 640, + 274, + 654 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 275, + 642, + 304, + 653 + ], + "score": 0.88, + "content": "j \\in \\mathbb { N } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 640, + 308, + 654 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 104, + 578, + 506, + 654 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 303, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 304, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 304, + 671 + ], + "score": 1.0, + "content": "For two kernel function defined in (2.2), we have", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 657, + 304, + 671 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 671, + 429, + 731 + ], + "lines": [ + { + "bbox": [ + 182, + 671, + 429, + 731 + ], + "spans": [ + { + "bbox": [ + 182, + 671, + 429, + 731 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad = \\displaystyle \\int _ { \\mathbb { S } ^ { d } } \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "95d2fe547af7b1da33eeceedf59749a8a73300ef5b14f44234e9e71094039a3d.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 182, + 671, + 429, + 691.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 182, + 691.0, + 429, + 711.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 182, + 711.0, + 429, + 731.0 + ], + "spans": [], + "index": 39 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 235, + 95 + ], + "score": 1.0, + "content": "The first equality holds because", + "type": "text" + }, + { + "bbox": [ + 236, + 83, + 245, + 92 + ], + "score": 0.86, + "content": "\\sigma ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "is 0-homogeneous function and the second equality is true since", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "the normalized direction of a multivariate Gaussian random variable satisfies uniform distribution", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 280, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 280, + 117 + ], + "score": 1.0, + "content": "on the unit sphere. Similarly we can derive", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 118, + 411, + 145 + ], + "lines": [ + { + "bbox": [ + 199, + 118, + 411, + 145 + ], + "spans": [ + { + "bbox": [ + 199, + 118, + 411, + 145 + ], + "score": 0.92, + "content": "\\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = ( d + 1 ) \\int _ { \\mathbb { S } ^ { d } } \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) .", + "type": "interline_equation", + "image_path": "bfa930864378680abdd0a8715fda31c8c82605c1ade23e00d05c55a4f7a9a1a6.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 199, + 118, + 411, + 145 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 147, + 352, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 352, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 352, + 161 + ], + "score": 1.0, + "content": "By combing (A.2), (B.7), (B.8), (B.9) and (B.10), we can get", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 177, + 389, + 210 + ], + "lines": [ + { + "bbox": [ + 222, + 177, + 389, + 210 + ], + "spans": [ + { + "bbox": [ + 222, + 177, + 389, + 210 + ], + "score": 0.95, + "content": "\\kappa _ { 1 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) ,", + "type": "interline_equation", + "image_path": "f96ce4a551ce7c988dbaa8b4eb08c8f195e2044e850349e9ac27e025dcb5722c.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 177, + 389, + 193.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 222, + 193.5, + 389, + 210.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 124, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 213, + 123, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 123, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 225, + 405, + 257 + ], + "lines": [ + { + "bbox": [ + 206, + 225, + 405, + 257 + ], + "spans": [ + { + "bbox": [ + 206, + 225, + 405, + 257 + ], + "score": 0.93, + "content": "\\kappa _ { 2 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = ( d + 1 ) \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) .", + "type": "interline_equation", + "image_path": "c83b470578b34839f6ba7c569b2474967481e20c3a2b30522f8268939034e564.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 225, + 405, + 241.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 206, + 241.0, + 405, + 257.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 352, + 272 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 353, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 353, + 273 + ], + "score": 1.0, + "content": "Comparing (B.6), (B.11) and (B.12), we can easily show that", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 274, + 383, + 290 + ], + "lines": [ + { + "bbox": [ + 228, + 274, + 383, + 290 + ], + "spans": [ + { + "bbox": [ + 228, + 274, + 383, + 290 + ], + "score": 0.86, + "content": "\\mu _ { 1 , k } = \\beta _ { 1 , k } ^ { 2 } \\mathrm { a n d } \\mu _ { 2 , k } = ( d + 1 ) \\beta _ { 2 , k } ^ { 2 } .", + "type": "interline_equation", + "image_path": "18b01ecfed90ecccae99f94fd9765d2330aaa4e0f4a23546a3dc4ef531c7b042.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 228, + 274, + 383, + 290 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 450, + 304 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 449, + 306 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 267, + 306 + ], + "score": 1.0, + "content": "In Bach (2017), explicit expressions for", + "type": "text" + }, + { + "bbox": [ + 267, + 292, + 285, + 304 + ], + "score": 0.91, + "content": "\\beta _ { 1 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 289, + 303, + 306 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 303, + 292, + 321, + 304 + ], + "score": 0.92, + "content": "\\beta _ { 2 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 289, + 336, + 306 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 337, + 292, + 380, + 302 + ], + "score": 0.9, + "content": "k \\geqslant \\alpha + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 289, + 449, + 306 + ], + "score": 1.0, + "content": "are presented by", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 307, + 426, + 337 + ], + "lines": [ + { + "bbox": [ + 185, + 307, + 426, + 337 + ], + "spans": [ + { + "bbox": [ + 185, + 307, + 426, + 337 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k } = \\frac { d - 1 } { 2 \\pi } \\frac { \\alpha ! ( - 1 ) ^ { ( k - 1 - \\alpha ) / 2 } } { 2 ^ { k } } \\frac { \\Gamma ( d / 2 ) \\Gamma ( k - \\alpha ) } { \\Gamma ( \\frac { k - \\alpha + 1 } { 2 } ) \\Gamma ( \\frac { k + d + \\alpha + 1 } { 2 } ) } .", + "type": "interline_equation", + "image_path": "30a45a792f4acf2923772703da93a5fe6febd0faa1f9d175c8c87f58d1b8fd28.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 307, + 426, + 322.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 185, + 322.0, + 426, + 337.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 341, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 186, + 357 + ], + "score": 1.0, + "content": "By Stirling formula", + "type": "text" + }, + { + "bbox": [ + 187, + 340, + 283, + 354 + ], + "score": 0.93, + "content": "\\Gamma ( x ) \\approx x ^ { x - 1 / 2 } e ^ { - x } \\sqrt { 2 \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 338, + 417, + 357 + ], + "score": 1.0, + "content": ", we have following expression of", + "type": "text" + }, + { + "bbox": [ + 418, + 342, + 447, + 354 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 338, + 462, + 357 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 462, + 342, + 504, + 353 + ], + "score": 0.89, + "content": "k \\geqslant \\alpha + 1", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 506, + 450 + ], + "lines": [ + { + "bbox": [ + 118, + 357, + 491, + 389 + ], + "spans": [ + { + "bbox": [ + 118, + 357, + 491, + 389 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k } = C ( \\alpha ) \\frac { ( d - 1 ) d ^ { \\frac { d - 1 } { 2 } } ( k - \\alpha ) ^ { k - \\alpha - \\frac 1 2 } } { ( k - \\alpha + 1 ) ^ { \\frac { k - \\alpha } { 2 } } ( k + d + \\alpha + 1 ) ^ { \\frac { k + d + \\alpha } { 2 } } } = \\Omega \\left( d ^ { \\frac { d + 1 } { 2 } } k ^ { \\frac { k - \\alpha - 1 } { 2 } } ( k + d ) ^ { \\frac { - k - d - \\alpha } { 2 } } \\right)", + "type": "inline_equation", + "image_path": "08502b71a77ceacf52fe8864fcbef0c6911219ce1215d57c73641a41c389cf55.jpg" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 393, + 483, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 138, + 411 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 138, + 393, + 483, + 416 + ], + "score": 0.73, + "content": "\\begin{array} { r l r l r l r l r l } { C ( \\alpha ) } & { { } = } & { { \\frac { \\sqrt { 2 } \\alpha ! } { 2 \\pi } } \\exp \\{ \\alpha \\ + \\ 1 \\} . } & { } & { \\mathrm { A l s o } \\ \\beta _ { \\alpha + 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 2 } { 2 } } \\right) } } , } & { \\beta _ { 1 , 1 } \\beta _ { 1 , 2 } \\gamma _ { 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 1 } { 2 } } \\right) } } } \\end{array}", + "type": "inline_equation", + "image_path": "918d79b5b61bca32ed6a2b9f074659184cfa44c78f51ae777f222c2428e488cf.jpg" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 411, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 125, + 434 + ], + "score": 0.87, + "content": "\\textstyle { \\frac { d - 1 } { 2 d \\pi } }", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 411, + 151, + 439 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 151, + 414, + 265, + 437 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } { \\beta _ { 2 , 1 } } & { { } = } & { \\frac { d - 1 } { 4 \\pi d } \\frac { \\Gamma \\left( \\frac { 1 } { 2 } \\right) \\Gamma \\left( \\frac { d + 2 } { 2 } \\right) } { \\Gamma \\left( \\frac { d + 3 } { 2 } \\right) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 411, + 268, + 439 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 280, + 417, + 506, + 435 + ], + "score": 1.0, + "content": "Thus combine (B.13) we know that µα`1,k “", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 434, + 257, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 251, + 450 + ], + "score": 0.86, + "content": "\\Omega \\left( d ^ { d + 1 + \\alpha } k ^ { k - \\alpha - 1 } ( k + d ) ^ { - k - \\bar { d } - \\bar { \\alpha } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 434, + 257, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 108, + 449, + 270, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 271, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 271, + 461 + ], + "score": 1.0, + "content": "By considering (A.3) and (B.6), we have", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 463, + 371, + 500 + ], + "lines": [ + { + "bbox": [ + 239, + 463, + 371, + 500 + ], + "spans": [ + { + "bbox": [ + 239, + 463, + 371, + 500 + ], + "score": 0.26, + "content": "\\begin{array} { c } { { \\mu _ { 0 } = \\mu _ { 1 , 1 } + 2 \\mu _ { 2 , 0 } , } } \\\\ { { { } } } \\\\ { { \\mu _ { k ^ { \\prime } } = 0 , k ^ { \\prime } = 2 j + 1 , j \\in \\mathbb { N } ^ { + } , } } \\end{array}", + "type": "interline_equation", + "image_path": "1f84e94fdc943bd8f863779f3436508288c49a6811b49c5dfe282c367d942950.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 463, + 371, + 481.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 239, + 481.5, + 371, + 500.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 124, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 124, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 124, + 514 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 515, + 417, + 540 + ], + "lines": [ + { + "bbox": [ + 192, + 515, + 417, + 540 + ], + "spans": [ + { + "bbox": [ + 192, + 515, + 417, + 540 + ], + "score": 0.91, + "content": "\\mu _ { k } = \\frac { k } { 2 k + d - 1 } \\mu _ { 1 , k - 1 } + \\frac { k + d - 1 } { 2 k + d - 1 } \\mu _ { 1 , k + 1 } + 2 \\mu _ { 2 , k } ,", + "type": "interline_equation", + "image_path": "1ef685e4e6ee1b9ee7f631d72757de9efb8356aaa395fec6722b1f2ea2d4274a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 192, + 515, + 417, + 540 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 543, + 455, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 453, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 120, + 556 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 544, + 146, + 554 + ], + "score": 0.91, + "content": "k \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 542, + 163, + 556 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 543, + 192, + 554 + ], + "score": 0.91, + "content": "k \\neq k ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 542, + 428, + 556 + ], + "score": 1.0, + "content": ". From the discussion above, we thus know exactly that for", + "type": "text" + }, + { + "bbox": [ + 428, + 544, + 453, + 554 + ], + "score": 0.89, + "content": "k \\geqslant 1", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 558, + 493, + 574 + ], + "lines": [ + { + "bbox": [ + 117, + 558, + 493, + 574 + ], + "spans": [ + { + "bbox": [ + 117, + 558, + 493, + 574 + ], + "score": 0.91, + "content": "\\mu _ { k } = \\Omega \\left( \\operatorname * { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) .", + "type": "interline_equation", + "image_path": "a57494a710d77e73676c55d20bfe3d5c5c4d0c95af0797120095f2b993daaa6b.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 117, + 558, + 493, + 574 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 200, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 201, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 201, + 590 + ], + "score": 1.0, + "content": "This finishes the proof.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 599, + 315, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 315, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 315, + 612 + ], + "score": 1.0, + "content": "B.3 PROOF OF THE COROLLARIES 3.7 AND 3.8", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 350, + 634 + ], + "score": 1.0, + "content": "Proof of the Corollaries 3.7 and 3.8. We only need to bound", + "type": "text" + }, + { + "bbox": [ + 350, + 621, + 381, + 634 + ], + "score": 0.93, + "content": "| \\phi _ { j } ( \\mathbf { x } ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 621, + 395, + 634 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 396, + 621, + 429, + 633 + ], + "score": 0.92, + "content": "j \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "to finish the proof.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 424, + 646 + ], + "score": 1.0, + "content": "Since now we assume input data follows uniform distribution on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 425, + 633, + 436, + 644 + ], + "score": 0.81, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 633, + 439, + 646 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 440, + 633, + 465, + 646 + ], + "score": 0.82, + "content": "\\phi _ { j } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "would be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 255, + 657 + ], + "score": 1.0, + "content": "spherical harmonics of order at most", + "type": "text" + }, + { + "bbox": [ + 256, + 645, + 263, + 654 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 644, + 278, + 657 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 279, + 644, + 311, + 657 + ], + "score": 0.92, + "content": "j \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 644, + 432, + 657 + ], + "score": 1.0, + "content": ". For any spherical harmonics", + "type": "text" + }, + { + "bbox": [ + 432, + 645, + 444, + 656 + ], + "score": 0.88, + "content": "Y _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 644, + 480, + 657 + ], + "score": 1.0, + "content": "of order", + "type": "text" + }, + { + "bbox": [ + 480, + 645, + 487, + 654 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 655, + 468, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 158, + 668 + ], + "score": 1.0, + "content": "any point on", + "type": "text" + }, + { + "bbox": [ + 159, + 655, + 170, + 666 + ], + "score": 0.88, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 655, + 468, + 668 + ], + "score": 1.0, + "content": ", we have an upper bound (Proposition 4.16 in Frye and Efthimiou (2012))", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 670, + 392, + 701 + ], + "lines": [ + { + "bbox": [ + 217, + 670, + 392, + 701 + ], + "spans": [ + { + "bbox": [ + 217, + 670, + 392, + 701 + ], + "score": 0.93, + "content": "| Y _ { k } ( { \\bf x } ) | \\leqslant \\left( N ( d , k ) \\int _ { { \\mathbb S } ^ { d } } Y _ { k } ^ { 2 } ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) \\right) ^ { \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "f161691bd018cc5066f6944bde6803ca908681fcf7dd9913f8fd6946acb0ef11.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 670, + 392, + 685.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 217, + 685.5, + 392, + 701.0 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 704, + 509, + 735 + ], + "lines": [ + { + "bbox": [ + 144, + 704, + 504, + 736 + ], + "spans": [ + { + "bbox": [ + 144, + 704, + 164, + 736 + ], + "score": 1.0, + "content": "kno. For", + "type": "text" + }, + { + "bbox": [ + 189, + 704, + 279, + 719 + ], + "score": 0.93, + "content": "| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant \\sqrt { N ( d , k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 706, + 333, + 716 + ], + "score": 0.89, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 704, + 504, + 720 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N ( d , k ) \\ = \\ \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\ = \\ } \\end{array}", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 718, + 379, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 719, + 144, + 732 + ], + "score": 0.91, + "content": "\\mathcal { O } ( k ^ { d - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 720, + 189, + 731 + ], + "score": 0.89, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 718, + 379, + 734 + ], + "score": 0.93, + "content": "\\begin{array} { r } { N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } = \\mathcal { O } ( d ^ { k } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 577, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 495, + 578, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 495, + 578, + 505, + 588 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 235, + 95 + ], + "score": 1.0, + "content": "The first equality holds because", + "type": "text" + }, + { + "bbox": [ + 236, + 83, + 245, + 92 + ], + "score": 0.86, + "content": "\\sigma ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "is 0-homogeneous function and the second equality is true since", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "the normalized direction of a multivariate Gaussian random variable satisfies uniform distribution", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 280, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 280, + 117 + ], + "score": 1.0, + "content": "on the unit sphere. Similarly we can derive", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 118, + 411, + 145 + ], + "lines": [ + { + "bbox": [ + 199, + 118, + 411, + 145 + ], + "spans": [ + { + "bbox": [ + 199, + 118, + 411, + 145 + ], + "score": 0.92, + "content": "\\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = ( d + 1 ) \\int _ { \\mathbb { S } ^ { d } } \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) .", + "type": "interline_equation", + "image_path": "bfa930864378680abdd0a8715fda31c8c82605c1ade23e00d05c55a4f7a9a1a6.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 199, + 118, + 411, + 145 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 147, + 352, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 352, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 352, + 161 + ], + "score": 1.0, + "content": "By combing (A.2), (B.7), (B.8), (B.9) and (B.10), we can get", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 147, + 352, + 161 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 177, + 389, + 210 + ], + "lines": [ + { + "bbox": [ + 222, + 177, + 389, + 210 + ], + "spans": [ + { + "bbox": [ + 222, + 177, + 389, + 210 + ], + "score": 0.95, + "content": "\\kappa _ { 1 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) ,", + "type": "interline_equation", + "image_path": "f96ce4a551ce7c988dbaa8b4eb08c8f195e2044e850349e9ac27e025dcb5722c.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 177, + 389, + 193.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 222, + 193.5, + 389, + 210.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 124, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 213, + 123, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 123, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 213, + 123, + 223 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 225, + 405, + 257 + ], + "lines": [ + { + "bbox": [ + 206, + 225, + 405, + 257 + ], + "spans": [ + { + "bbox": [ + 206, + 225, + 405, + 257 + ], + "score": 0.93, + "content": "\\kappa _ { 2 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = ( d + 1 ) \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) .", + "type": "interline_equation", + "image_path": "c83b470578b34839f6ba7c569b2474967481e20c3a2b30522f8268939034e564.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 225, + 405, + 241.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 206, + 241.0, + 405, + 257.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 352, + 272 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 353, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 353, + 273 + ], + "score": 1.0, + "content": "Comparing (B.6), (B.11) and (B.12), we can easily show that", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 259, + 353, + 273 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 274, + 383, + 290 + ], + "lines": [ + { + "bbox": [ + 228, + 274, + 383, + 290 + ], + "spans": [ + { + "bbox": [ + 228, + 274, + 383, + 290 + ], + "score": 0.86, + "content": "\\mu _ { 1 , k } = \\beta _ { 1 , k } ^ { 2 } \\mathrm { a n d } \\mu _ { 2 , k } = ( d + 1 ) \\beta _ { 2 , k } ^ { 2 } .", + "type": "interline_equation", + "image_path": "18b01ecfed90ecccae99f94fd9765d2330aaa4e0f4a23546a3dc4ef531c7b042.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 228, + 274, + 383, + 290 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 450, + 304 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 449, + 306 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 267, + 306 + ], + "score": 1.0, + "content": "In Bach (2017), explicit expressions for", + "type": "text" + }, + { + "bbox": [ + 267, + 292, + 285, + 304 + ], + "score": 0.91, + "content": "\\beta _ { 1 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 289, + 303, + 306 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 303, + 292, + 321, + 304 + ], + "score": 0.92, + "content": "\\beta _ { 2 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 289, + 336, + 306 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 337, + 292, + 380, + 302 + ], + "score": 0.9, + "content": "k \\geqslant \\alpha + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 289, + 449, + 306 + ], + "score": 1.0, + "content": "are presented by", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 104, + 289, + 449, + 306 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 307, + 426, + 337 + ], + "lines": [ + { + "bbox": [ + 185, + 307, + 426, + 337 + ], + "spans": [ + { + "bbox": [ + 185, + 307, + 426, + 337 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k } = \\frac { d - 1 } { 2 \\pi } \\frac { \\alpha ! ( - 1 ) ^ { ( k - 1 - \\alpha ) / 2 } } { 2 ^ { k } } \\frac { \\Gamma ( d / 2 ) \\Gamma ( k - \\alpha ) } { \\Gamma ( \\frac { k - \\alpha + 1 } { 2 } ) \\Gamma ( \\frac { k + d + \\alpha + 1 } { 2 } ) } .", + "type": "interline_equation", + "image_path": "30a45a792f4acf2923772703da93a5fe6febd0faa1f9d175c8c87f58d1b8fd28.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 307, + 426, + 322.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 185, + 322.0, + 426, + 337.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 341, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 186, + 357 + ], + "score": 1.0, + "content": "By Stirling formula", + "type": "text" + }, + { + "bbox": [ + 187, + 340, + 283, + 354 + ], + "score": 0.93, + "content": "\\Gamma ( x ) \\approx x ^ { x - 1 / 2 } e ^ { - x } \\sqrt { 2 \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 338, + 417, + 357 + ], + "score": 1.0, + "content": ", we have following expression of", + "type": "text" + }, + { + "bbox": [ + 418, + 342, + 447, + 354 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 338, + 462, + 357 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 462, + 342, + 504, + 353 + ], + "score": 0.89, + "content": "k \\geqslant \\alpha + 1", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 118, + 357, + 491, + 389 + ], + "spans": [ + { + "bbox": [ + 118, + 357, + 491, + 389 + ], + "score": 0.92, + "content": "\\beta _ { \\alpha + 1 , k } = C ( \\alpha ) \\frac { ( d - 1 ) d ^ { \\frac { d - 1 } { 2 } } ( k - \\alpha ) ^ { k - \\alpha - \\frac 1 2 } } { ( k - \\alpha + 1 ) ^ { \\frac { k - \\alpha } { 2 } } ( k + d + \\alpha + 1 ) ^ { \\frac { k + d + \\alpha } { 2 } } } = \\Omega \\left( d ^ { \\frac { d + 1 } { 2 } } k ^ { \\frac { k - \\alpha - 1 } { 2 } } ( k + d ) ^ { \\frac { - k - d - \\alpha } { 2 } } \\right)", + "type": "inline_equation", + "image_path": "08502b71a77ceacf52fe8864fcbef0c6911219ce1215d57c73641a41c389cf55.jpg" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 393, + 483, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 138, + 411 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 138, + 393, + 483, + 416 + ], + "score": 0.73, + "content": "\\begin{array} { r l r l r l r l r l } { C ( \\alpha ) } & { { } = } & { { \\frac { \\sqrt { 2 } \\alpha ! } { 2 \\pi } } \\exp \\{ \\alpha \\ + \\ 1 \\} . } & { } & { \\mathrm { A l s o } \\ \\beta _ { \\alpha + 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 2 } { 2 } } \\right) } } , } & { \\beta _ { 1 , 1 } \\beta _ { 1 , 2 } \\gamma _ { 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 1 } { 2 } } \\right) } } } \\end{array}", + "type": "inline_equation", + "image_path": "918d79b5b61bca32ed6a2b9f074659184cfa44c78f51ae777f222c2428e488cf.jpg" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 411, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 125, + 434 + ], + "score": 0.87, + "content": "\\textstyle { \\frac { d - 1 } { 2 d \\pi } }", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 411, + 151, + 439 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 151, + 414, + 265, + 437 + ], + "score": 0.88, + "content": "\\begin{array} { r l r } { \\beta _ { 2 , 1 } } & { { } = } & { \\frac { d - 1 } { 4 \\pi d } \\frac { \\Gamma \\left( \\frac { 1 } { 2 } \\right) \\Gamma \\left( \\frac { d + 2 } { 2 } \\right) } { \\Gamma \\left( \\frac { d + 3 } { 2 } \\right) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 411, + 268, + 439 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 280, + 417, + 506, + 435 + ], + "score": 1.0, + "content": "Thus combine (B.13) we know that µα`1,k “", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 434, + 257, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 251, + 450 + ], + "score": 0.86, + "content": "\\Omega \\left( d ^ { d + 1 + \\alpha } k ^ { k - \\alpha - 1 } ( k + d ) ^ { - k - \\bar { d } - \\bar { \\alpha } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 434, + 257, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 338, + 504, + 357 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 506, + 450 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 105, + 357, + 506, + 451 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 449, + 270, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 271, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 271, + 461 + ], + "score": 1.0, + "content": "By considering (A.3) and (B.6), we have", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 448, + 271, + 461 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 463, + 371, + 500 + ], + "lines": [ + { + "bbox": [ + 239, + 463, + 371, + 500 + ], + "spans": [ + { + "bbox": [ + 239, + 463, + 371, + 500 + ], + "score": 0.26, + "content": "\\begin{array} { c } { { \\mu _ { 0 } = \\mu _ { 1 , 1 } + 2 \\mu _ { 2 , 0 } , } } \\\\ { { { } } } \\\\ { { \\mu _ { k ^ { \\prime } } = 0 , k ^ { \\prime } = 2 j + 1 , j \\in \\mathbb { N } ^ { + } , } } \\end{array}", + "type": "interline_equation", + "image_path": "1f84e94fdc943bd8f863779f3436508288c49a6811b49c5dfe282c367d942950.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 463, + 371, + 481.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 239, + 481.5, + 371, + 500.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 124, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 124, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 124, + 514 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 502, + 124, + 514 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 515, + 417, + 540 + ], + "lines": [ + { + "bbox": [ + 192, + 515, + 417, + 540 + ], + "spans": [ + { + "bbox": [ + 192, + 515, + 417, + 540 + ], + "score": 0.91, + "content": "\\mu _ { k } = \\frac { k } { 2 k + d - 1 } \\mu _ { 1 , k - 1 } + \\frac { k + d - 1 } { 2 k + d - 1 } \\mu _ { 1 , k + 1 } + 2 \\mu _ { 2 , k } ,", + "type": "interline_equation", + "image_path": "1ef685e4e6ee1b9ee7f631d72757de9efb8356aaa395fec6722b1f2ea2d4274a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 192, + 515, + 417, + 540 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 543, + 455, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 453, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 120, + 556 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 544, + 146, + 554 + ], + "score": 0.91, + "content": "k \\geqslant 1", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 542, + 163, + 556 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 543, + 192, + 554 + ], + "score": 0.91, + "content": "k \\neq k ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 542, + 428, + 556 + ], + "score": 1.0, + "content": ". From the discussion above, we thus know exactly that for", + "type": "text" + }, + { + "bbox": [ + 428, + 544, + 453, + 554 + ], + "score": 0.89, + "content": "k \\geqslant 1", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 542, + 453, + 556 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 558, + 493, + 574 + ], + "lines": [ + { + "bbox": [ + 117, + 558, + 493, + 574 + ], + "spans": [ + { + "bbox": [ + 117, + 558, + 493, + 574 + ], + "score": 0.91, + "content": "\\mu _ { k } = \\Omega \\left( \\operatorname * { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) .", + "type": "interline_equation", + "image_path": "a57494a710d77e73676c55d20bfe3d5c5c4d0c95af0797120095f2b993daaa6b.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 117, + 558, + 493, + 574 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 200, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 201, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 201, + 590 + ], + "score": 1.0, + "content": "This finishes the proof.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 574, + 201, + 590 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 599, + 315, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 315, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 315, + 612 + ], + "score": 1.0, + "content": "B.3 PROOF OF THE COROLLARIES 3.7 AND 3.8", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 350, + 634 + ], + "score": 1.0, + "content": "Proof of the Corollaries 3.7 and 3.8. We only need to bound", + "type": "text" + }, + { + "bbox": [ + 350, + 621, + 381, + 634 + ], + "score": 0.93, + "content": "| \\phi _ { j } ( \\mathbf { x } ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 621, + 395, + 634 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 396, + 621, + 429, + 633 + ], + "score": 0.92, + "content": "j \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "to finish the proof.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 424, + 646 + ], + "score": 1.0, + "content": "Since now we assume input data follows uniform distribution on the unit sphere", + "type": "text" + }, + { + "bbox": [ + 425, + 633, + 436, + 644 + ], + "score": 0.81, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 633, + 439, + 646 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 440, + 633, + 465, + 646 + ], + "score": 0.82, + "content": "\\phi _ { j } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "would be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 255, + 657 + ], + "score": 1.0, + "content": "spherical harmonics of order at most", + "type": "text" + }, + { + "bbox": [ + 256, + 645, + 263, + 654 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 644, + 278, + 657 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 279, + 644, + 311, + 657 + ], + "score": 0.92, + "content": "j \\in [ r _ { k } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 644, + 432, + 657 + ], + "score": 1.0, + "content": ". For any spherical harmonics", + "type": "text" + }, + { + "bbox": [ + 432, + 645, + 444, + 656 + ], + "score": 0.88, + "content": "Y _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 644, + 480, + 657 + ], + "score": 1.0, + "content": "of order", + "type": "text" + }, + { + "bbox": [ + 480, + 645, + 487, + 654 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 655, + 468, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 158, + 668 + ], + "score": 1.0, + "content": "any point on", + "type": "text" + }, + { + "bbox": [ + 159, + 655, + 170, + 666 + ], + "score": 0.88, + "content": "\\mathbb { S } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 655, + 468, + 668 + ], + "score": 1.0, + "content": ", we have an upper bound (Proposition 4.16 in Frye and Efthimiou (2012))", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 621, + 506, + 668 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 670, + 392, + 701 + ], + "lines": [ + { + "bbox": [ + 217, + 670, + 392, + 701 + ], + "spans": [ + { + "bbox": [ + 217, + 670, + 392, + 701 + ], + "score": 0.93, + "content": "| Y _ { k } ( { \\bf x } ) | \\leqslant \\left( N ( d , k ) \\int _ { { \\mathbb S } ^ { d } } Y _ { k } ^ { 2 } ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) \\right) ^ { \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "f161691bd018cc5066f6944bde6803ca908681fcf7dd9913f8fd6946acb0ef11.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 670, + 392, + 685.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 217, + 685.5, + 392, + 701.0 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 704, + 509, + 735 + ], + "lines": [ + { + "bbox": [ + 144, + 704, + 504, + 736 + ], + "spans": [ + { + "bbox": [ + 144, + 704, + 164, + 736 + ], + "score": 1.0, + "content": "kno. For", + "type": "text" + }, + { + "bbox": [ + 189, + 704, + 279, + 719 + ], + "score": 0.93, + "content": "| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant \\sqrt { N ( d , k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 706, + 333, + 716 + ], + "score": 0.89, + "content": "k \\gg d", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 704, + 504, + 720 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N ( d , k ) \\ = \\ \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\ = \\ } \\end{array}", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 718, + 379, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 719, + 144, + 732 + ], + "score": 0.91, + "content": "\\mathcal { O } ( k ^ { d - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 720, + 189, + 731 + ], + "score": 0.89, + "content": "d \\gg k", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 718, + 379, + 734 + ], + "score": 0.93, + "content": "\\begin{array} { r } { N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } = \\mathcal { O } ( d ^ { k } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 107, + 704, + 504, + 736 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 292, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 294, + 96 + ], + "score": 1.0, + "content": "C PROOF OF TECHNICAL LEMMAS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 231, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 232, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 232, + 118 + ], + "score": 1.0, + "content": "C.1 PROOF OF LEMMA B.2", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 419, + 140 + ], + "score": 1.0, + "content": "Proof of Lemma B.2. The first inequality directly follows by equation (44) in", + "type": "text" + }, + { + "bbox": [ + 419, + 128, + 432, + 137 + ], + "score": 0.34, + "content": "\\mathtt { S u }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "and Yang (2019).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 138, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 255, + 152 + ], + "score": 1.0, + "content": "To prove the second bound, we write a", + "type": "text" + }, + { + "bbox": [ + 255, + 138, + 349, + 153 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\dot { \\mathbf { V } _ { r _ { k } } } = \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { \\bar { A } } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\mathbf { \\bar { B } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 138, + 379, + 152 + ], + "score": 1.0, + "content": ", where a", + "type": "text" + }, + { + "bbox": [ + 380, + 140, + 429, + 151 + ], + "score": 0.82, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { r _ { k } \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 138, + 434, + 152 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 434, + 138, + 501, + 150 + ], + "score": 0.89, + "content": "\\mathbf { B } \\in \\mathbb { R } ^ { ( n - r _ { k } ) \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 138, + 506, + 152 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 152, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 123, + 167 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 152, + 291, + 166 + ], + "score": 0.78, + "content": "\\xi _ { 1 } = C ^ { \\prime } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } \\cdot \\sqrt { \\log ( 1 / \\delta ) / n }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 152, + 295, + 167 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 296, + 153, + 409, + 166 + ], + "score": 0.74, + "content": "\\xi _ { 2 } = C ^ { \\prime \\prime \\prime } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 152, + 506, + 167 + ], + "score": 1.0, + "content": "be the bounds given in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 164, + 376, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 376, + 177 + ], + "score": 1.0, + "content": "the first inequality and Lemma 3.1. By the first inequality, we have", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 182, + 383, + 199 + ], + "lines": [ + { + "bbox": [ + 227, + 182, + 383, + 199 + ], + "spans": [ + { + "bbox": [ + 227, + 182, + 383, + 199 + ], + "score": 0.92, + "content": "\\| \\mathbf { B } \\| _ { F } = \\| \\mathbf { B } ^ { \\top } \\| _ { F } = \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\| _ { F } \\leqslant \\xi _ { 1 } .", + "type": "interline_equation", + "image_path": "238383ca1bab2f25bc29d9aa6e8654061b7546a13c93a6f85c13dc9ed9e90fc5.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 227, + 182, + 383, + 199 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 379, + 218 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 380, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 173, + 219 + ], + "score": 1.0, + "content": "Moreover, since", + "type": "text" + }, + { + "bbox": [ + 173, + 204, + 280, + 218 + ], + "score": 0.94, + "content": "\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } = \\mathbf { A } ^ { \\top } \\mathbf { A } + \\mathbf { B } ^ { \\top } \\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 202, + 380, + 219 + ], + "score": 1.0, + "content": ", by Lemma 3.1 we have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 223, + 456, + 240 + ], + "lines": [ + { + "bbox": [ + 154, + 223, + 456, + 240 + ], + "spans": [ + { + "bbox": [ + 154, + 223, + 456, + 240 + ], + "score": 0.88, + "content": "\\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } = \\| \\mathbf { A } ^ { \\top } \\mathbf { A } - \\mathbf { I } \\| _ { 2 } \\leqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { 2 } + \\| \\mathbf { B } ^ { \\top } \\mathbf { B } \\| _ { 2 } \\leqslant r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ^ { 2 } .", + "type": "interline_equation", + "image_path": "4990612d02a94e3bcd38b1128dbcce7b646f16a699c39f5201cdcfc2ffabc926.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 154, + 223, + 456, + 240 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 148, + 255 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 148, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 148, + 257 + ], + "score": 1.0, + "content": "Therefore", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 259, + 501, + 341 + ], + "lines": [ + { + "bbox": [ + 111, + 259, + 501, + 341 + ], + "spans": [ + { + "bbox": [ + 111, + 259, + 501, + 341 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\big ( \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\big ) \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad = \\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "55a4cfe7456cd67732c11b9ed4790dbf305020c4c1ca74c6d38c790ded067c03.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 111, + 259, + 501, + 286.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 111, + 286.3333333333333, + 501, + 313.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 111, + 313.66666666666663, + 501, + 340.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 344, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 344, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 221, + 360 + ], + "score": 1.0, + "content": "Plugging in the definition of", + "type": "text" + }, + { + "bbox": [ + 222, + 347, + 231, + 358 + ], + "score": 0.89, + "content": "\\xi _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 345, + 249, + 360 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 250, + 347, + 259, + 358 + ], + "score": 0.89, + "content": "\\xi _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 345, + 344, + 360 + ], + "score": 1.0, + "content": "completes the proof.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 371, + 230, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 232, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 232, + 384 + ], + "score": 1.0, + "content": "C.2 PROOF OF LEMMA B.3", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 366, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 366, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 366, + 407 + ], + "score": 1.0, + "content": "Proof of Lemma B.3. The gradient descent update formula gives", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 410, + 434, + 442 + ], + "lines": [ + { + "bbox": [ + 177, + 410, + 434, + 442 + ], + "spans": [ + { + "bbox": [ + 177, + 410, + 434, + 442 + ], + "score": 0.94, + "content": "\\mathbf { W } ^ { ( t + 1 ) } = \\mathbf { W } ^ { ( t ) } + \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } \\bigl ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\bigr ) \\cdot \\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) .", + "type": "interline_equation", + "image_path": "231555ff0f8d836ac406bb6f499944f0274e7bdf42eed188f8b8d7399f8e1504.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 177, + 410, + 434, + 420.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 177, + 420.6666666666667, + 434, + 431.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 177, + 431.33333333333337, + 434, + 442.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 448, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 104, + 447, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 104, + 447, + 140, + 463 + ], + "score": 1.0, + "content": "For any", + "type": "text" + }, + { + "bbox": [ + 141, + 449, + 172, + 461 + ], + "score": 0.92, + "content": "j \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 447, + 223, + 463 + ], + "score": 1.0, + "content": ", subtracting", + "type": "text" + }, + { + "bbox": [ + 224, + 448, + 246, + 460 + ], + "score": 0.88, + "content": "\\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 447, + 381, + 463 + ], + "score": 1.0, + "content": "and applying inner product with", + "type": "text" + }, + { + "bbox": [ + 381, + 449, + 447, + 462 + ], + "score": 0.92, + "content": "\\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } \\left( \\mathbf { x } _ { j } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 447, + 506, + 463 + ], + "score": 1.0, + "content": "on both sides", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 460, + 131, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 460, + 131, + 475 + ], + "score": 1.0, + "content": "gives", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 477, + 494, + 509 + ], + "lines": [ + { + "bbox": [ + 115, + 477, + 494, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 477, + 494, + 509 + ], + "score": 0.94, + "content": "\\theta \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle = \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\hat { y } _ { i } ^ { ( t ) } ) \\cdot \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle .", + "type": "interline_equation", + "image_path": "ad4be42d9c5e3936277a36f4dd0eca38247a415f59209be45b60b8c0eb6b8374.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 115, + 477, + 494, + 487.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 115, + 487.6666666666667, + 494, + 498.33333333333337 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 115, + 498.33333333333337, + 494, + 509.00000000000006 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 254, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 254, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 254, + 528 + ], + "score": 1.0, + "content": "Further rearranging terms then gives", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 532, + 494, + 563 + ], + "lines": [ + { + "bbox": [ + 115, + 532, + 494, + 563 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 494, + 563 + ], + "score": 0.94, + "content": "y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t + 1 ) } ) _ { j } = y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t ) } ) _ { j } - \\frac { 2 \\eta m \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot \\mathbf { K } _ { i , j } ^ { \\infty } + I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t } ,", + "type": "interline_equation", + "image_path": "18d477a76c43e2cf16f7a081c1059891ce6315d76b3de2b8ef8fc8a0994d776d.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 115, + 532, + 494, + 542.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 115, + 542.3333333333334, + 494, + 552.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 115, + 552.6666666666667, + 494, + 563.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 133, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 135, + 593 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 595, + 477, + 678 + ], + "lines": [ + { + "bbox": [ + 133, + 595, + 477, + 678 + ], + "spans": [ + { + "bbox": [ + 133, + 595, + 477, + 678 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle I _ { 1 , j , t } = - \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot [ \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle - m \\mathbf { K } _ { i , j } ^ { ( 0 ) } ] } , } \\\\ { { \\displaystyle I _ { 2 , j , t } = - \\frac { 2 \\eta \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot ( \\mathbf { K } _ { i , j } ^ { ( 0 ) } - \\mathbf { K } _ { i , j } ^ { \\infty } ) , } } \\\\ { { \\displaystyle I _ { 3 , j , t } = - \\theta \\cdot [ f _ { \\mathbf { W } ^ { ( t + 1 ) } } ( \\mathbf { x } _ { j } ) - f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle ] } . } \\end{array}", + "type": "interline_equation", + "image_path": "ffa1a81375895e365daa3b830eea52322209bf771b4c6f7e2475f3b4e664be88.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 133, + 595, + 477, + 622.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 133, + 622.6666666666666, + 477, + 650.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 133, + 650.3333333333333, + 477, + 677.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 248, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 248, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 122, + 695 + ], + "score": 1.0, + "content": "For", + "type": "text" + }, + { + "bbox": [ + 123, + 682, + 144, + 694 + ], + "score": 0.93, + "content": "I _ { 1 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 680, + 248, + 695 + ], + "score": 1.0, + "content": ", by Lemma D.6, we have", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "interline_equation", + "bbox": [ + 132, + 699, + 478, + 730 + ], + "lines": [ + { + "bbox": [ + 132, + 699, + 478, + 730 + ], + "spans": [ + { + "bbox": [ + 132, + 699, + 478, + 730 + ], + "score": 0.92, + "content": "| I _ { 1 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n }", + "type": "interline_equation", + "image_path": "977341470ef42f16c69e2ff113183c02981a6ffeb527f185cc19c64faf950623.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 132, + 699, + 478, + 709.3333333333334 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 132, + 709.3333333333334, + 478, + 719.6666666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 132, + 719.6666666666667, + 478, + 730.0000000000001 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 347, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 496, + 348, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 496, + 348, + 505, + 358 + ], + "score": 1.0, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 292, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 294, + 96 + ], + "score": 1.0, + "content": "C PROOF OF TECHNICAL LEMMAS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 231, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 232, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 232, + 118 + ], + "score": 1.0, + "content": "C.1 PROOF OF LEMMA B.2", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 419, + 140 + ], + "score": 1.0, + "content": "Proof of Lemma B.2. The first inequality directly follows by equation (44) in", + "type": "text" + }, + { + "bbox": [ + 419, + 128, + 432, + 137 + ], + "score": 0.34, + "content": "\\mathtt { S u }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "and Yang (2019).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 138, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 255, + 152 + ], + "score": 1.0, + "content": "To prove the second bound, we write a", + "type": "text" + }, + { + "bbox": [ + 255, + 138, + 349, + 153 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\dot { \\mathbf { V } _ { r _ { k } } } = \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { \\bar { A } } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\mathbf { \\bar { B } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 138, + 379, + 152 + ], + "score": 1.0, + "content": ", where a", + "type": "text" + }, + { + "bbox": [ + 380, + 140, + 429, + 151 + ], + "score": 0.82, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { r _ { k } \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 138, + 434, + 152 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 434, + 138, + 501, + 150 + ], + "score": 0.89, + "content": "\\mathbf { B } \\in \\mathbb { R } ^ { ( n - r _ { k } ) \\times r _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 138, + 506, + 152 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 152, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 123, + 167 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 152, + 291, + 166 + ], + "score": 0.78, + "content": "\\xi _ { 1 } = C ^ { \\prime } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } \\cdot \\sqrt { \\log ( 1 / \\delta ) / n }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 152, + 295, + 167 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 296, + 153, + 409, + 166 + ], + "score": 0.74, + "content": "\\xi _ { 2 } = C ^ { \\prime \\prime \\prime } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 152, + 506, + 167 + ], + "score": 1.0, + "content": "be the bounds given in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 164, + 376, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 376, + 177 + ], + "score": 1.0, + "content": "the first inequality and Lemma 3.1. By the first inequality, we have", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 126, + 506, + 177 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 182, + 383, + 199 + ], + "lines": [ + { + "bbox": [ + 227, + 182, + 383, + 199 + ], + "spans": [ + { + "bbox": [ + 227, + 182, + 383, + 199 + ], + "score": 0.92, + "content": "\\| \\mathbf { B } \\| _ { F } = \\| \\mathbf { B } ^ { \\top } \\| _ { F } = \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\| _ { F } \\leqslant \\xi _ { 1 } .", + "type": "interline_equation", + "image_path": "238383ca1bab2f25bc29d9aa6e8654061b7546a13c93a6f85c13dc9ed9e90fc5.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 227, + 182, + 383, + 199 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 379, + 218 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 380, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 173, + 219 + ], + "score": 1.0, + "content": "Moreover, since", + "type": "text" + }, + { + "bbox": [ + 173, + 204, + 280, + 218 + ], + "score": 0.94, + "content": "\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } = \\mathbf { A } ^ { \\top } \\mathbf { A } + \\mathbf { B } ^ { \\top } \\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 202, + 380, + 219 + ], + "score": 1.0, + "content": ", by Lemma 3.1 we have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 202, + 380, + 219 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 223, + 456, + 240 + ], + "lines": [ + { + "bbox": [ + 154, + 223, + 456, + 240 + ], + "spans": [ + { + "bbox": [ + 154, + 223, + 456, + 240 + ], + "score": 0.88, + "content": "\\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } = \\| \\mathbf { A } ^ { \\top } \\mathbf { A } - \\mathbf { I } \\| _ { 2 } \\leqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { 2 } + \\| \\mathbf { B } ^ { \\top } \\mathbf { B } \\| _ { 2 } \\leqslant r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ^ { 2 } .", + "type": "interline_equation", + "image_path": "4990612d02a94e3bcd38b1128dbcce7b646f16a699c39f5201cdcfc2ffabc926.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 154, + 223, + 456, + 240 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 148, + 255 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 148, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 148, + 257 + ], + "score": 1.0, + "content": "Therefore", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 244, + 148, + 257 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 259, + 501, + 341 + ], + "lines": [ + { + "bbox": [ + 111, + 259, + 501, + 341 + ], + "spans": [ + { + "bbox": [ + 111, + 259, + 501, + 341 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\big ( \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\big ) \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad = \\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "55a4cfe7456cd67732c11b9ed4790dbf305020c4c1ca74c6d38c790ded067c03.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 111, + 259, + 501, + 286.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 111, + 286.3333333333333, + 501, + 313.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 111, + 313.66666666666663, + 501, + 340.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 344, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 344, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 221, + 360 + ], + "score": 1.0, + "content": "Plugging in the definition of", + "type": "text" + }, + { + "bbox": [ + 222, + 347, + 231, + 358 + ], + "score": 0.89, + "content": "\\xi _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 345, + 249, + 360 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 250, + 347, + 259, + 358 + ], + "score": 0.89, + "content": "\\xi _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 345, + 344, + 360 + ], + "score": 1.0, + "content": "completes the proof.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 345, + 344, + 360 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 371, + 230, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 232, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 232, + 384 + ], + "score": 1.0, + "content": "C.2 PROOF OF LEMMA B.3", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 366, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 366, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 366, + 407 + ], + "score": 1.0, + "content": "Proof of Lemma B.3. The gradient descent update formula gives", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 392, + 366, + 407 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 410, + 434, + 442 + ], + "lines": [ + { + "bbox": [ + 177, + 410, + 434, + 442 + ], + "spans": [ + { + "bbox": [ + 177, + 410, + 434, + 442 + ], + "score": 0.94, + "content": "\\mathbf { W } ^ { ( t + 1 ) } = \\mathbf { W } ^ { ( t ) } + \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } \\bigl ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\bigr ) \\cdot \\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) .", + "type": "interline_equation", + "image_path": "231555ff0f8d836ac406bb6f499944f0274e7bdf42eed188f8b8d7399f8e1504.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 177, + 410, + 434, + 420.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 177, + 420.6666666666667, + 434, + 431.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 177, + 431.33333333333337, + 434, + 442.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 448, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 104, + 447, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 104, + 447, + 140, + 463 + ], + "score": 1.0, + "content": "For any", + "type": "text" + }, + { + "bbox": [ + 141, + 449, + 172, + 461 + ], + "score": 0.92, + "content": "j \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 447, + 223, + 463 + ], + "score": 1.0, + "content": ", subtracting", + "type": "text" + }, + { + "bbox": [ + 224, + 448, + 246, + 460 + ], + "score": 0.88, + "content": "\\mathbf { W } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 447, + 381, + 463 + ], + "score": 1.0, + "content": "and applying inner product with", + "type": "text" + }, + { + "bbox": [ + 381, + 449, + 447, + 462 + ], + "score": 0.92, + "content": "\\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } \\left( \\mathbf { x } _ { j } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 447, + 506, + 463 + ], + "score": 1.0, + "content": "on both sides", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 460, + 131, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 460, + 131, + 475 + ], + "score": 1.0, + "content": "gives", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 447, + 506, + 475 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 477, + 494, + 509 + ], + "lines": [ + { + "bbox": [ + 115, + 477, + 494, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 477, + 494, + 509 + ], + "score": 0.94, + "content": "\\theta \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle = \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\hat { y } _ { i } ^ { ( t ) } ) \\cdot \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle .", + "type": "interline_equation", + "image_path": "ad4be42d9c5e3936277a36f4dd0eca38247a415f59209be45b60b8c0eb6b8374.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 115, + 477, + 494, + 487.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 115, + 487.6666666666667, + 494, + 498.33333333333337 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 115, + 498.33333333333337, + 494, + 509.00000000000006 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 254, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 254, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 254, + 528 + ], + "score": 1.0, + "content": "Further rearranging terms then gives", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 513, + 254, + 528 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 532, + 494, + 563 + ], + "lines": [ + { + "bbox": [ + 115, + 532, + 494, + 563 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 494, + 563 + ], + "score": 0.94, + "content": "y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t + 1 ) } ) _ { j } = y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t ) } ) _ { j } - \\frac { 2 \\eta m \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot \\mathbf { K } _ { i , j } ^ { \\infty } + I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t } ,", + "type": "interline_equation", + "image_path": "18d477a76c43e2cf16f7a081c1059891ce6315d76b3de2b8ef8fc8a0994d776d.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 115, + 532, + 494, + 542.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 115, + 542.3333333333334, + 494, + 552.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 115, + 552.6666666666667, + 494, + 563.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 133, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 135, + 593 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 579, + 135, + 593 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 595, + 477, + 678 + ], + "lines": [ + { + "bbox": [ + 133, + 595, + 477, + 678 + ], + "spans": [ + { + "bbox": [ + 133, + 595, + 477, + 678 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle I _ { 1 , j , t } = - \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot [ \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle - m \\mathbf { K } _ { i , j } ^ { ( 0 ) } ] } , } \\\\ { { \\displaystyle I _ { 2 , j , t } = - \\frac { 2 \\eta \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot ( \\mathbf { K } _ { i , j } ^ { ( 0 ) } - \\mathbf { K } _ { i , j } ^ { \\infty } ) , } } \\\\ { { \\displaystyle I _ { 3 , j , t } = - \\theta \\cdot [ f _ { \\mathbf { W } ^ { ( t + 1 ) } } ( \\mathbf { x } _ { j } ) - f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle ] } . } \\end{array}", + "type": "interline_equation", + "image_path": "ffa1a81375895e365daa3b830eea52322209bf771b4c6f7e2475f3b4e664be88.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 133, + 595, + 477, + 622.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 133, + 622.6666666666666, + 477, + 650.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 133, + 650.3333333333333, + 477, + 677.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 248, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 248, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 122, + 695 + ], + "score": 1.0, + "content": "For", + "type": "text" + }, + { + "bbox": [ + 123, + 682, + 144, + 694 + ], + "score": 0.93, + "content": "I _ { 1 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 680, + 248, + 695 + ], + "score": 1.0, + "content": ", by Lemma D.6, we have", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 680, + 248, + 695 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 132, + 699, + 478, + 730 + ], + "lines": [ + { + "bbox": [ + 132, + 699, + 478, + 730 + ], + "spans": [ + { + "bbox": [ + 132, + 699, + 478, + 730 + ], + "score": 0.92, + "content": "| I _ { 1 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n }", + "type": "interline_equation", + "image_path": "977341470ef42f16c69e2ff113183c02981a6ffeb527f185cc19c64faf950623.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 132, + 699, + 478, + 709.3333333333334 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 132, + 709.3333333333334, + 478, + 719.6666666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 132, + 719.6666666666667, + 478, + 730.0000000000001 + ], + "spans": [], + "index": 35 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 203, + 96 + ], + "score": 1.0, + "content": "with probability at least", + "type": "text" + }, + { + "bbox": [ + 204, + 81, + 322, + 95 + ], + "score": 0.9, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 81, + 343, + 96 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 343, + 83, + 363, + 95 + ], + "score": 0.9, + "content": "I _ { 2 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 81, + 505, + 96 + ], + "score": 1.0, + "content": ", by Bernstein inequality and union", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 397, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 233, + 109 + ], + "score": 1.0, + "content": "bound, with probability at least", + "type": "text" + }, + { + "bbox": [ + 234, + 95, + 356, + 107 + ], + "score": 0.87, + "content": "1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp ( - \\Omega ( m \\omega ^ { 2 / 3 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 93, + 397, + 109 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 112, + 358, + 131 + ], + "lines": [ + { + "bbox": [ + 252, + 112, + 358, + 131 + ], + "spans": [ + { + "bbox": [ + 252, + 112, + 358, + 131 + ], + "score": 0.92, + "content": "\\big | \\mathbf { K } _ { i , j } ^ { \\infty } - \\mathbf { K } _ { i , j } ^ { ( 0 ) } \\big | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } )", + "type": "interline_equation", + "image_path": "cfd0fda51ed2619710ce18718f21595620bcc96ab6b0b4a38a35057b8b0fd9ff.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 252, + 112, + 358, + 131 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 136, + 216, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 135, + 217, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 133, + 150 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 136, + 170, + 149 + ], + "score": 0.92, + "content": "i , j \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 135, + 217, + 150 + ], + "score": 1.0, + "content": ". Therefore", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 154, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 130, + 154, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 480, + 186 + ], + "score": 0.94, + "content": "| I _ { 2 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } .", + "type": "interline_equation", + "image_path": "395d49404d7877bb95aeff7283b15ea83f17598ca772e7d629aab519fd328916.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 130, + 154, + 480, + 164.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 130, + 164.66666666666666, + 480, + 175.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 130, + 175.33333333333331, + 480, + 185.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 182, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 189, + 183, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 122, + 205 + ], + "score": 1.0, + "content": "For", + "type": "text" + }, + { + "bbox": [ + 123, + 191, + 144, + 204 + ], + "score": 0.92, + "content": "I _ { 3 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 189, + 183, + 205 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 208, + 452, + 309 + ], + "lines": [ + { + "bbox": [ + 158, + 208, + 452, + 309 + ], + "spans": [ + { + "bbox": [ + 158, + 208, + 452, + 309 + ], + "score": 0.95, + "content": "\\begin{array} { r l r } { { I _ { 3 , j , t } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { ( t + 1 ) } - \\mathbf { W } _ { 1 } ^ { ( t ) } \\| _ { 2 } } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\cdot \\theta \\cdot \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\| _ { 2 } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\cfrac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } , } \\end{array}", + "type": "interline_equation", + "image_path": "f2d3f819faf36e8f88e3144eb94e09776542b236d5d876b9fe406ab7264350f1.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 158, + 208, + 452, + 241.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 158, + 241.66666666666666, + 452, + 275.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 158, + 275.3333333333333, + 452, + 309.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 312, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "where the first inequality follows by Lemmas D.2, the second inequality is obtained from (C.1), and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 326, + 339 + ], + "score": 1.0, + "content": "the third inequality follows by Lemma D.4. Setting the", + "type": "text" + }, + { + "bbox": [ + 326, + 325, + 332, + 336 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 321, + 378, + 339 + ], + "score": 1.0, + "content": "-th entry of", + "type": "text" + }, + { + "bbox": [ + 378, + 323, + 393, + 335 + ], + "score": 0.89, + "content": "\\mathbf { e } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 321, + 405, + 339 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 405, + 324, + 487, + 337 + ], + "score": 0.91, + "content": "I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 321, + 506, + 339 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 314, + 349 + ], + "score": 1.0, + "content": "writing (C.2) into matrix form completes the proof.", + "type": "text" + }, + { + "bbox": [ + 495, + 336, + 505, + 347 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 107, + 360, + 230, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 232, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 232, + 372 + ], + "score": 1.0, + "content": "C.3 PROOF OF LEMMA B.4", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 380, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 380, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 226, + 398 + ], + "score": 1.0, + "content": "Proof of Lemma B.4. Denote", + "type": "text" + }, + { + "bbox": [ + 226, + 383, + 290, + 396 + ], + "score": 0.87, + "content": "\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 381, + 294, + 398 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 294, + 384, + 317, + 395 + ], + "score": 0.81, + "content": "t \\in T", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 381, + 380, + 398 + ], + "score": 1.0, + "content": ". Then we have", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 135, + 400, + 476, + 438 + ], + "lines": [ + { + "bbox": [ + 135, + 400, + 476, + 438 + ], + "spans": [ + { + "bbox": [ + 135, + 400, + 476, + 438 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } [ { \\mathbf { I } } - ( \\eta m \\theta ^ { 2 } / n ) { \\mathbf { K } } ^ { \\infty } ] { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "a867be7b2013f3586a54ea83ac85ede4de733ea548f4eb08bdf76beec60f6047.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 135, + 400, + 476, + 412.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 135, + 412.6666666666667, + 476, + 425.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 135, + 425.33333333333337, + 476, + 438.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 504, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "where the first inequality follows by Lemma B.3, and the second inequality follows by Lemma D.5.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 451, + 183, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 183, + 464 + ], + "score": 1.0, + "content": "Therefore we have", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 466, + 452, + 484 + ], + "lines": [ + { + "bbox": [ + 158, + 466, + 452, + 484 + ], + "spans": [ + { + "bbox": [ + 158, + 466, + 452, + 484 + ], + "score": 0.89, + "content": "\\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) ,", + "type": "interline_equation", + "image_path": "ba5485a85db3b64cd31cc1ac92622c425118a1ab47ffda441e72f7c50635a617.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 158, + 466, + 452, + 484 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 392, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 392, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 120, + 503 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 489, + 172, + 501 + ], + "score": 0.91, + "content": "t ^ { \\prime } = 0 , \\ldots , t", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 488, + 392, + 503 + ], + "score": 1.0, + "content": ". This completes the proof of (B.1). Similarly, we have", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 505, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 111, + 505, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 111, + 505, + 505, + 595 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\widehat { \\lambda } _ { r _ { k } } ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot ( \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } ) } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } & { } \\\\ + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\end{array}", + "type": "interline_equation", + "image_path": "731d3412adbd36dfe4e2b6a9ed64333eb07ee5ee4307a5c354064b826a5d2e7a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 111, + 505, + 505, + 535.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 535.0, + 505, + 565.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 111, + 565.0, + 505, + 595.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 506, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 121, + 613 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 599, + 195, + 610 + ], + "score": 0.9, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 597, + 483, + 613 + ], + "score": 1.0, + "content": ", where the third inequality is by Lemma B.1 and the assumption that", + "type": "text" + }, + { + "bbox": [ + 484, + 599, + 505, + 610 + ], + "score": 0.82, + "content": "\\omega \\ \\leqslant", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 609, + 425, + 626 + ], + "spans": [ + { + "bbox": [ + 107, + 610, + 137, + 625 + ], + "score": 0.34, + "content": "\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 609, + 141, + 626 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 142, + 610, + 194, + 625 + ], + "score": 0.74, + "content": "n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 609, + 425, + 626 + ], + "score": 1.0, + "content": ", and the fourth inequality is by (B.1). Therefore we have", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 629, + 514, + 737 + ], + "lines": [ + { + "bbox": [ + 111, + 629, + 514, + 737 + ], + "spans": [ + { + "bbox": [ + 111, + 629, + 514, + 737 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } = ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & \\phantom { \\hat { \\mathbf { V } } } + \\end{array}", + "type": "interline_equation", + "image_path": "735e7e055b5ef8b4ae398e44d36ba8c817c1c2532ffe58bf80ef3a9706813ee7.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 111, + 629, + 514, + 665.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 111, + 665.0, + 514, + 701.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 111, + 701.0, + 514, + 737.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 203, + 96 + ], + "score": 1.0, + "content": "with probability at least", + "type": "text" + }, + { + "bbox": [ + 204, + 81, + 322, + 95 + ], + "score": 0.9, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 81, + 343, + 96 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 343, + 83, + 363, + 95 + ], + "score": 0.9, + "content": "I _ { 2 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 81, + 505, + 96 + ], + "score": 1.0, + "content": ", by Bernstein inequality and union", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 397, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 233, + 109 + ], + "score": 1.0, + "content": "bound, with probability at least", + "type": "text" + }, + { + "bbox": [ + 234, + 95, + 356, + 107 + ], + "score": 0.87, + "content": "1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp ( - \\Omega ( m \\omega ^ { 2 / 3 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 93, + 397, + 109 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 109 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 112, + 358, + 131 + ], + "lines": [ + { + "bbox": [ + 252, + 112, + 358, + 131 + ], + "spans": [ + { + "bbox": [ + 252, + 112, + 358, + 131 + ], + "score": 0.92, + "content": "\\big | \\mathbf { K } _ { i , j } ^ { \\infty } - \\mathbf { K } _ { i , j } ^ { ( 0 ) } \\big | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } )", + "type": "interline_equation", + "image_path": "cfd0fda51ed2619710ce18718f21595620bcc96ab6b0b4a38a35057b8b0fd9ff.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 252, + 112, + 358, + 131 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 136, + 216, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 135, + 217, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 133, + 150 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 136, + 170, + 149 + ], + "score": 0.92, + "content": "i , j \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 135, + 217, + 150 + ], + "score": 1.0, + "content": ". Therefore", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 135, + 217, + 150 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 154, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 130, + 154, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 480, + 186 + ], + "score": 0.94, + "content": "| I _ { 2 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } .", + "type": "interline_equation", + "image_path": "395d49404d7877bb95aeff7283b15ea83f17598ca772e7d629aab519fd328916.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 130, + 154, + 480, + 164.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 130, + 164.66666666666666, + 480, + 175.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 130, + 175.33333333333331, + 480, + 185.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 182, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 189, + 183, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 122, + 205 + ], + "score": 1.0, + "content": "For", + "type": "text" + }, + { + "bbox": [ + 123, + 191, + 144, + 204 + ], + "score": 0.92, + "content": "I _ { 3 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 189, + 183, + 205 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 189, + 183, + 205 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 208, + 452, + 309 + ], + "lines": [ + { + "bbox": [ + 158, + 208, + 452, + 309 + ], + "spans": [ + { + "bbox": [ + 158, + 208, + 452, + 309 + ], + "score": 0.95, + "content": "\\begin{array} { r l r } { { I _ { 3 , j , t } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { ( t + 1 ) } - \\mathbf { W } _ { 1 } ^ { ( t ) } \\| _ { 2 } } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\cdot \\theta \\cdot \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\| _ { 2 } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\cfrac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } , } \\end{array}", + "type": "interline_equation", + "image_path": "f2d3f819faf36e8f88e3144eb94e09776542b236d5d876b9fe406ab7264350f1.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 158, + 208, + 452, + 241.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 158, + 241.66666666666666, + 452, + 275.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 158, + 275.3333333333333, + 452, + 309.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 312, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "where the first inequality follows by Lemmas D.2, the second inequality is obtained from (C.1), and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 326, + 339 + ], + "score": 1.0, + "content": "the third inequality follows by Lemma D.4. Setting the", + "type": "text" + }, + { + "bbox": [ + 326, + 325, + 332, + 336 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 321, + 378, + 339 + ], + "score": 1.0, + "content": "-th entry of", + "type": "text" + }, + { + "bbox": [ + 378, + 323, + 393, + 335 + ], + "score": 0.89, + "content": "\\mathbf { e } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 321, + 405, + 339 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 405, + 324, + 487, + 337 + ], + "score": 0.91, + "content": "I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 321, + 506, + 339 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 314, + 349 + ], + "score": 1.0, + "content": "writing (C.2) into matrix form completes the proof.", + "type": "text" + }, + { + "bbox": [ + 495, + 336, + 505, + 347 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 312, + 506, + 349 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 360, + 230, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 232, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 232, + 372 + ], + "score": 1.0, + "content": "C.3 PROOF OF LEMMA B.4", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 380, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 380, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 226, + 398 + ], + "score": 1.0, + "content": "Proof of Lemma B.4. Denote", + "type": "text" + }, + { + "bbox": [ + 226, + 383, + 290, + 396 + ], + "score": 0.87, + "content": "\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 381, + 294, + 398 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 294, + 384, + 317, + 395 + ], + "score": 0.81, + "content": "t \\in T", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 381, + 380, + 398 + ], + "score": 1.0, + "content": ". Then we have", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 381, + 380, + 398 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 135, + 400, + 476, + 438 + ], + "lines": [ + { + "bbox": [ + 135, + 400, + 476, + 438 + ], + "spans": [ + { + "bbox": [ + 135, + 400, + 476, + 438 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } [ { \\mathbf { I } } - ( \\eta m \\theta ^ { 2 } / n ) { \\mathbf { K } } ^ { \\infty } ] { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "a867be7b2013f3586a54ea83ac85ede4de733ea548f4eb08bdf76beec60f6047.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 135, + 400, + 476, + 412.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 135, + 412.6666666666667, + 476, + 425.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 135, + 425.33333333333337, + 476, + 438.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 504, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "where the first inequality follows by Lemma B.3, and the second inequality follows by Lemma D.5.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 451, + 183, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 183, + 464 + ], + "score": 1.0, + "content": "Therefore we have", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 440, + 505, + 464 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 466, + 452, + 484 + ], + "lines": [ + { + "bbox": [ + 158, + 466, + 452, + 484 + ], + "spans": [ + { + "bbox": [ + 158, + 466, + 452, + 484 + ], + "score": 0.89, + "content": "\\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) ,", + "type": "interline_equation", + "image_path": "ba5485a85db3b64cd31cc1ac92622c425118a1ab47ffda441e72f7c50635a617.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 158, + 466, + 452, + 484 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 392, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 392, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 120, + 503 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 489, + 172, + 501 + ], + "score": 0.91, + "content": "t ^ { \\prime } = 0 , \\ldots , t", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 488, + 392, + 503 + ], + "score": 1.0, + "content": ". This completes the proof of (B.1). Similarly, we have", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 488, + 392, + 503 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 505, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 111, + 505, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 111, + 505, + 505, + 595 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\widehat { \\lambda } _ { r _ { k } } ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot ( \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } ) } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } & { } \\\\ + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\end{array}", + "type": "interline_equation", + "image_path": "731d3412adbd36dfe4e2b6a9ed64333eb07ee5ee4307a5c354064b826a5d2e7a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 111, + 505, + 505, + 535.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 535.0, + 505, + 565.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 111, + 565.0, + 505, + 595.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 506, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 121, + 613 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 599, + 195, + 610 + ], + "score": 0.9, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 597, + 483, + 613 + ], + "score": 1.0, + "content": ", where the third inequality is by Lemma B.1 and the assumption that", + "type": "text" + }, + { + "bbox": [ + 484, + 599, + 505, + 610 + ], + "score": 0.82, + "content": "\\omega \\ \\leqslant", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 609, + 425, + 626 + ], + "spans": [ + { + "bbox": [ + 107, + 610, + 137, + 625 + ], + "score": 0.34, + "content": "\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 609, + 141, + 626 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 142, + 610, + 194, + 625 + ], + "score": 0.74, + "content": "n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 609, + 425, + 626 + ], + "score": 1.0, + "content": ", and the fourth inequality is by (B.1). Therefore we have", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 597, + 505, + 626 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 629, + 514, + 737 + ], + "lines": [ + { + "bbox": [ + 111, + 629, + 514, + 737 + ], + "spans": [ + { + "bbox": [ + 111, + 629, + 514, + 737 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } = ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & \\phantom { \\hat { \\mathbf { V } } } + \\end{array}", + "type": "interline_equation", + "image_path": "735e7e055b5ef8b4ae398e44d36ba8c817c1c2532ffe58bf80ef3a9706813ee7.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 111, + 629, + 514, + 665.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 111, + 665.0, + 514, + 701.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 111, + 701.0, + 514, + 737.0 + ], + "spans": [], + "index": 30 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 422, + 97 + ], + "score": 1.0, + "content": "This completes the proof of (B.2). Finally, for (B.3), by assumption we have", + "type": "text" + }, + { + "bbox": [ + 423, + 80, + 501, + 95 + ], + "score": 0.92, + "content": "\\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 78, + 505, + 97 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 91, + 149, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 149, + 106 + ], + "score": 1.0, + "content": "Therefore", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 103, + 513, + 194 + ], + "lines": [ + { + "bbox": [ + 111, + 103, + 513, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 513, + 194 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\hat { \\lambda } _ { \\mathbf { r } _ { k } } ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( 1 ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { \\mathbf { r } _ { k } } / 2 ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\end{array}", + "type": "interline_equation", + "image_path": "68dc0a55babfb44f99376c17aa93747eca9f9e8d91ad988b871323165d10fe03.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 111, + 103, + 513, + 133.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 133.33333333333334, + 513, + 163.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 163.66666666666669, + 513, + 194.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 121, + 208 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 194, + 195, + 206 + ], + "score": 0.92, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 192, + 483, + 208 + ], + "score": 1.0, + "content": ", where the third inequality is by Lemma B.1 and the assumption that", + "type": "text" + }, + { + "bbox": [ + 484, + 195, + 505, + 205 + ], + "score": 0.8, + "content": "\\omega \\ \\leqslant", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 205, + 392, + 221 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 137, + 220 + ], + "score": 0.92, + "content": "\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 205, + 392, + 221 + ], + "score": 1.0, + "content": ", and the fourth inequality follows by (B.1). Therefore we have", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 556, + 253 + ], + "lines": [ + { + "bbox": [ + 104, + 222, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 151, + 241 + ], + "score": 1.0, + "content": "} u p t 1 q } 2 ď", + "type": "text" + }, + { + "bbox": [ + 151, + 222, + 554, + 239 + ], + "score": 0.74, + "content": "\\mathscr { O } ( \\sqrt { n } ) \\cdot ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t ^ { \\prime } } + \\widetilde { \\mathscr { O } } ( ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } ) \\cdot \\Vert ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( 0 ) } \\Vert _ { 2 } + \\lambda _ { r _ { k } } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathscr { O } } ( 1 + \\omega \\sqrt { m } ) .", + "type": "inline_equation", + "image_path": "05aa08f491626174face9a4f84a97a152c1ba9d0b5df65ad28ac77e38b675cb6.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 239, + 201, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 201, + 254 + ], + "score": 1.0, + "content": "This finishes the proof.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 107, + 266, + 237, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 238, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 238, + 280 + ], + "score": 1.0, + "content": "D AUXILIARY LEMMAS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 323 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "In this section we list several auxiliary lemmas on the properties of over-parameterized neural net-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 301, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 312 + ], + "score": 1.0, + "content": "works we need in our proof of Theorem 3.2. These results are mostly summarized from Allen-Zhu", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 311, + 260, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 260, + 323 + ], + "score": 1.0, + "content": "et al. (2018) and Cao and Gu (2019a).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 333, + 225, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 225, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 225, + 345 + ], + "score": 1.0, + "content": "D.1 AUXILIARY LEMMAS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 137, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 138, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 138, + 366 + ], + "score": 1.0, + "content": "Denote", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 181, + 366, + 429, + 401 + ], + "lines": [ + { + "bbox": [ + 181, + 366, + 429, + 401 + ], + "spans": [ + { + "bbox": [ + 181, + 366, + 429, + 401 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\mathbf { D } _ { i } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) , } \\\\ & { \\mathbf { D } _ { i } ^ { ( 0 ) } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ff0e5791ea25ef3963ffa126e032454c12dce142473a05c97bd0ff79716c28f9.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 181, + 366, + 429, + 377.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 181, + 377.6666666666667, + 429, + 389.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 181, + 389.33333333333337, + 429, + 401.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 403, + 503, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 282, + 418 + ], + "score": 1.0, + "content": "Lemma D.1 (Allen-Zhu et al. (2018)). If", + "type": "text" + }, + { + "bbox": [ + 282, + 403, + 377, + 416 + ], + "score": 0.92, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 402, + 506, + 418 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 415, + 230, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 226, + 429 + ], + "score": 0.89, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 415, + 230, + 429 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 430, + 363, + 448 + ], + "lines": [ + { + "bbox": [ + 247, + 430, + 363, + 448 + ], + "spans": [ + { + "bbox": [ + 247, + 430, + 363, + 448 + ], + "score": 0.9, + "content": "\\| \\mathbf { D } _ { i } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { 0 } \\leqslant \\mathcal { O } ( \\omega ^ { 2 / 3 } m )", + "type": "interline_equation", + "image_path": "a0fcca48470c6919a30c09b5d5f14a5b98e2173732ee39612c8e06ed71eaccae.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 247, + 430, + 363, + 448 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 240, + 464 + ], + "lines": [ + { + "bbox": [ + 104, + 448, + 237, + 468 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 133, + 468 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 451, + 205, + 464 + ], + "score": 0.82, + "content": "\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 448, + 209, + 468 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 209, + 452, + 237, + 464 + ], + "score": 0.58, + "content": "i \\in [ n ]", + "type": "inline_equation" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 389, + 478 + ], + "score": 1.0, + "content": "Lemma D.2 (Cao and Gu (2019a)). There exists an absolute constant", + "type": "text" + }, + { + "bbox": [ + 389, + 468, + 396, + 476 + ], + "score": 0.79, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "such that, with probability", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 475, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 104, + 475, + 138, + 491 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 138, + 477, + 258, + 490 + ], + "score": 0.88, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 475, + 357, + 491 + ], + "score": 1.0, + "content": "over the randomness of", + "type": "text" + }, + { + "bbox": [ + 358, + 477, + 381, + 488 + ], + "score": 0.89, + "content": "\\mathbf { W } ^ { ( 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 475, + 413, + 491 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 414, + 477, + 443, + 490 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 475, + 462, + 491 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 463, + 478, + 505, + 489 + ], + "score": 0.82, + "content": "\\mathbf { W } , \\mathbf { W } ^ { \\prime } \\in", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 488, + 355, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 156, + 502 + ], + "score": 0.92, + "content": "B ( \\mathbf { W } ^ { ( \\mathbf { \\theta } ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 488, + 177, + 504 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 178, + 489, + 258, + 502 + ], + "score": 0.9, + "content": "\\omega \\leqslant \\kappa [ \\log ( m ) ] ^ { - 3 / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 488, + 355, + 504 + ], + "score": 1.0, + "content": ", it holds uniformly that", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 120, + 504, + 490, + 526 + ], + "lines": [ + { + "bbox": [ + 120, + 504, + 490, + 526 + ], + "spans": [ + { + "bbox": [ + 120, + 504, + 490, + 526 + ], + "score": 0.92, + "content": "| f _ { \\mathbf { W } ^ { \\prime } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\mathbf { W } ^ { \\prime } - \\mathbf { W } \\rangle | \\leqslant { \\mathcal { O } } \\Bigl ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } \\Bigr ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { \\prime } - \\mathbf { W } _ { 1 } \\| _ { 2 } .", + "type": "interline_equation", + "image_path": "3f1ac05583175ca408c01d620917fcde1ce626c7b4f02ef194ebe5ea1d4c11cd.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 120, + 504, + 490, + 526 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 502, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 287, + 541 + ], + "score": 1.0, + "content": "Lemma D.3 (Cao and Gu (2019a)). For any", + "type": "text" + }, + { + "bbox": [ + 287, + 529, + 312, + 539 + ], + "score": 0.9, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 528, + 324, + 541 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 325, + 528, + 392, + 540 + ], + "score": 0.89, + "content": "m \\geqslant C \\log ( n / \\delta )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "for a large enough absolute", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 540, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 142, + 554 + ], + "score": 1.0, + "content": "constant", + "type": "text" + }, + { + "bbox": [ + 142, + 542, + 151, + 551 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 541, + 271, + 554 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 272, + 541, + 295, + 552 + ], + "score": 0.54, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 541, + 298, + 554 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 298, + 540, + 421, + 554 + ], + "score": 0.89, + "content": "| f _ { \\mathbf { W } ^ { ( 0 ) } } ( { \\pmb x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 541, + 449, + 554 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 450, + 541, + 478, + 553 + ], + "score": 0.91, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 541, + 481, + 554 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 502, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 388, + 568 + ], + "score": 1.0, + "content": "Lemma D.4 (Cao and Gu (2019a)). There exists an absolute constant", + "type": "text" + }, + { + "bbox": [ + 388, + 555, + 397, + 565 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 552, + 504, + 568 + ], + "score": 1.0, + "content": "such that, with probability", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 565, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 139, + 580 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 139, + 566, + 261, + 579 + ], + "score": 0.91, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 565, + 294, + 580 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 295, + 566, + 364, + 579 + ], + "score": 0.86, + "content": "i \\in [ n ] , l \\in [ L ]", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 565, + 384, + 580 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 566, + 460, + 579 + ], + "score": 0.93, + "content": "\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 565, + 483, + 580 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 484, + 567, + 505, + 578 + ], + "score": 0.82, + "content": "\\omega \\ \\leqslant", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 577, + 259, + 591 + ], + "spans": [ + { + "bbox": [ + 107, + 578, + 162, + 590 + ], + "score": 0.92, + "content": "C [ \\log ( m ) ] ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 577, + 259, + 591 + ], + "score": 1.0, + "content": ", it holds uniformly that", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 591, + 365, + 606 + ], + "lines": [ + { + "bbox": [ + 245, + 591, + 365, + 606 + ], + "spans": [ + { + "bbox": [ + 245, + 591, + 365, + 606 + ], + "score": 0.89, + "content": "\\| \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\sqrt { m } ) .", + "type": "interline_equation", + "image_path": "3ea31864bd3ef411c12372d9e785cce422d2bd9da181b1ef7ea1d4762698acf8.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 245, + 591, + 365, + 606 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 504, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 627 + ], + "score": 1.0, + "content": "The following lemma provides a uniform bound of the neural network function value over", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 622, + 162, + 640 + ], + "spans": [ + { + "bbox": [ + 107, + 625, + 156, + 639 + ], + "score": 0.92, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 622, + 162, + 640 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 104, + 638, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 638, + 217, + 655 + ], + "score": 1.0, + "content": "Lemma D.5. Suppose that", + "type": "text" + }, + { + "bbox": [ + 218, + 640, + 316, + 653 + ], + "score": 0.9, + "content": "m \\geqslant \\Omega ( \\omega ^ { - 2 / 3 } \\log ( n / \\delta ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 638, + 335, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 335, + 640, + 419, + 653 + ], + "score": 0.91, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 638, + 505, + 655 + ], + "score": 1.0, + "content": ". Then with probabil-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 651, + 460, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 651, + 150, + 668 + ], + "score": 1.0, + "content": "ity at least", + "type": "text" + }, + { + "bbox": [ + 150, + 654, + 173, + 665 + ], + "score": 0.44, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 651, + 176, + 668 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 176, + 653, + 326, + 667 + ], + "score": 0.91, + "content": "| f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 651, + 354, + 668 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 354, + 653, + 456, + 666 + ], + "score": 0.92, + "content": "\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) \\ i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 651, + 460, + 668 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 104, + 668, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 501, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 172, + 685 + ], + "score": 1.0, + "content": "Lemma D.6. If", + "type": "text" + }, + { + "bbox": [ + 173, + 668, + 263, + 682 + ], + "score": 0.93, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 666, + 383, + 685 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 384, + 668, + 501, + 682 + ], + "score": 0.91, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 684, + 466, + 717 + ], + "lines": [ + { + "bbox": [ + 143, + 684, + 466, + 717 + ], + "spans": [ + { + "bbox": [ + 143, + 684, + 466, + 717 + ], + "score": 0.84, + "content": "\\begin{array} { r l } & { \\| \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } ) , } \\\\ & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } m ) } \\end{array}", + "type": "interline_equation", + "image_path": "b3a03af7ab37fba24e0f0a2b59e5305a6930a79ac7298eb27f92f385cdcffb08.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 143, + 684, + 466, + 695.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 143, + 695.0, + 466, + 706.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 143, + 706.0, + 466, + 717.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 719, + 254, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 716, + 252, + 736 + ], + "spans": [ + { + "bbox": [ + 104, + 716, + 133, + 736 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 719, + 205, + 732 + ], + "score": 0.94, + "content": "\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 716, + 223, + 736 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 223, + 720, + 252, + 732 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 495, + 241, + 504, + 251 + ], + "lines": [ + { + "bbox": [ + 495, + 243, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 495, + 243, + 504, + 251 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 422, + 97 + ], + "score": 1.0, + "content": "This completes the proof of (B.2). Finally, for (B.3), by assumption we have", + "type": "text" + }, + { + "bbox": [ + 423, + 80, + 501, + 95 + ], + "score": 0.92, + "content": "\\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 78, + 505, + 97 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 91, + 149, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 149, + 106 + ], + "score": 1.0, + "content": "Therefore", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 104, + 78, + 505, + 106 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 103, + 513, + 194 + ], + "lines": [ + { + "bbox": [ + 111, + 103, + 513, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 513, + 194 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\hat { \\lambda } _ { \\mathbf { r } _ { k } } ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( 1 ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { \\mathbf { r } _ { k } } / 2 ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\end{array}", + "type": "interline_equation", + "image_path": "68dc0a55babfb44f99376c17aa93747eca9f9e8d91ad988b871323165d10fe03.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 111, + 103, + 513, + 133.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 133.33333333333334, + 513, + 163.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 163.66666666666669, + 513, + 194.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 121, + 208 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 194, + 195, + 206 + ], + "score": 0.92, + "content": "t ^ { \\prime } = 0 , \\ldots , t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 192, + 483, + 208 + ], + "score": 1.0, + "content": ", where the third inequality is by Lemma B.1 and the assumption that", + "type": "text" + }, + { + "bbox": [ + 484, + 195, + 505, + 205 + ], + "score": 0.8, + "content": "\\omega \\ \\leqslant", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 205, + 392, + 221 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 137, + 220 + ], + "score": 0.92, + "content": "\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 205, + 392, + 221 + ], + "score": 1.0, + "content": ", and the fourth inequality follows by (B.1). Therefore we have", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 192, + 505, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 556, + 253 + ], + "lines": [ + { + "bbox": [ + 104, + 222, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 151, + 241 + ], + "score": 1.0, + "content": "} u p t 1 q } 2 ď", + "type": "text" + }, + { + "bbox": [ + 151, + 222, + 554, + 239 + ], + "score": 0.74, + "content": "\\mathscr { O } ( \\sqrt { n } ) \\cdot ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t ^ { \\prime } } + \\widetilde { \\mathscr { O } } ( ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } ) \\cdot \\Vert ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( 0 ) } \\Vert _ { 2 } + \\lambda _ { r _ { k } } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathscr { O } } ( 1 + \\omega \\sqrt { m } ) .", + "type": "inline_equation", + "image_path": "05aa08f491626174face9a4f84a97a152c1ba9d0b5df65ad28ac77e38b675cb6.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 239, + 201, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 201, + 254 + ], + "score": 1.0, + "content": "This finishes the proof.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 222, + 554, + 254 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 266, + 237, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 238, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 238, + 280 + ], + "score": 1.0, + "content": "D AUXILIARY LEMMAS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 323 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "In this section we list several auxiliary lemmas on the properties of over-parameterized neural net-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 301, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 312 + ], + "score": 1.0, + "content": "works we need in our proof of Theorem 3.2. These results are mostly summarized from Allen-Zhu", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 311, + 260, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 260, + 323 + ], + "score": 1.0, + "content": "et al. (2018) and Cao and Gu (2019a).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 288, + 505, + 323 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 333, + 225, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 225, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 225, + 345 + ], + "score": 1.0, + "content": "D.1 AUXILIARY LEMMAS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 137, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 138, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 138, + 366 + ], + "score": 1.0, + "content": "Denote", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 354, + 138, + 366 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 181, + 366, + 429, + 401 + ], + "lines": [ + { + "bbox": [ + 181, + 366, + 429, + 401 + ], + "spans": [ + { + "bbox": [ + 181, + 366, + 429, + 401 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\mathbf { D } _ { i } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) , } \\\\ & { \\mathbf { D } _ { i } ^ { ( 0 ) } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ff0e5791ea25ef3963ffa126e032454c12dce142473a05c97bd0ff79716c28f9.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 181, + 366, + 429, + 377.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 181, + 377.6666666666667, + 429, + 389.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 181, + 389.33333333333337, + 429, + 401.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 403, + 503, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 282, + 418 + ], + "score": 1.0, + "content": "Lemma D.1 (Allen-Zhu et al. (2018)). If", + "type": "text" + }, + { + "bbox": [ + 282, + 403, + 377, + 416 + ], + "score": 0.92, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 402, + 506, + 418 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 415, + 230, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 226, + 429 + ], + "score": 0.89, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 415, + 230, + 429 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 402, + 506, + 429 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 430, + 363, + 448 + ], + "lines": [ + { + "bbox": [ + 247, + 430, + 363, + 448 + ], + "spans": [ + { + "bbox": [ + 247, + 430, + 363, + 448 + ], + "score": 0.9, + "content": "\\| \\mathbf { D } _ { i } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { 0 } \\leqslant \\mathcal { O } ( \\omega ^ { 2 / 3 } m )", + "type": "interline_equation", + "image_path": "a0fcca48470c6919a30c09b5d5f14a5b98e2173732ee39612c8e06ed71eaccae.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 247, + 430, + 363, + 448 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 240, + 464 + ], + "lines": [ + { + "bbox": [ + 104, + 448, + 237, + 468 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 133, + 468 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 451, + 205, + 464 + ], + "score": 0.82, + "content": "\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 448, + 209, + 468 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 209, + 452, + 237, + 464 + ], + "score": 0.58, + "content": "i \\in [ n ]", + "type": "inline_equation" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 448, + 237, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 389, + 478 + ], + "score": 1.0, + "content": "Lemma D.2 (Cao and Gu (2019a)). There exists an absolute constant", + "type": "text" + }, + { + "bbox": [ + 389, + 468, + 396, + 476 + ], + "score": 0.79, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "such that, with probability", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 475, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 104, + 475, + 138, + 491 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 138, + 477, + 258, + 490 + ], + "score": 0.88, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 475, + 357, + 491 + ], + "score": 1.0, + "content": "over the randomness of", + "type": "text" + }, + { + "bbox": [ + 358, + 477, + 381, + 488 + ], + "score": 0.89, + "content": "\\mathbf { W } ^ { ( 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 475, + 413, + 491 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 414, + 477, + 443, + 490 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 475, + 462, + 491 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 463, + 478, + 505, + 489 + ], + "score": 0.82, + "content": "\\mathbf { W } , \\mathbf { W } ^ { \\prime } \\in", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 488, + 355, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 156, + 502 + ], + "score": 0.92, + "content": "B ( \\mathbf { W } ^ { ( \\mathbf { \\theta } ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 488, + 177, + 504 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 178, + 489, + 258, + 502 + ], + "score": 0.9, + "content": "\\omega \\leqslant \\kappa [ \\log ( m ) ] ^ { - 3 / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 488, + 355, + 504 + ], + "score": 1.0, + "content": ", it holds uniformly that", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 465, + 505, + 504 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 120, + 504, + 490, + 526 + ], + "lines": [ + { + "bbox": [ + 120, + 504, + 490, + 526 + ], + "spans": [ + { + "bbox": [ + 120, + 504, + 490, + 526 + ], + "score": 0.92, + "content": "| f _ { \\mathbf { W } ^ { \\prime } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\mathbf { W } ^ { \\prime } - \\mathbf { W } \\rangle | \\leqslant { \\mathcal { O } } \\Bigl ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } \\Bigr ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { \\prime } - \\mathbf { W } _ { 1 } \\| _ { 2 } .", + "type": "interline_equation", + "image_path": "3f1ac05583175ca408c01d620917fcde1ce626c7b4f02ef194ebe5ea1d4c11cd.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 120, + 504, + 490, + 526 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 502, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 287, + 541 + ], + "score": 1.0, + "content": "Lemma D.3 (Cao and Gu (2019a)). For any", + "type": "text" + }, + { + "bbox": [ + 287, + 529, + 312, + 539 + ], + "score": 0.9, + "content": "\\delta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 528, + 324, + 541 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 325, + 528, + 392, + 540 + ], + "score": 0.89, + "content": "m \\geqslant C \\log ( n / \\delta )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "for a large enough absolute", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 540, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 142, + 554 + ], + "score": 1.0, + "content": "constant", + "type": "text" + }, + { + "bbox": [ + 142, + 542, + 151, + 551 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 541, + 271, + 554 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 272, + 541, + 295, + 552 + ], + "score": 0.54, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 541, + 298, + 554 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 298, + 540, + 421, + 554 + ], + "score": 0.89, + "content": "| f _ { \\mathbf { W } ^ { ( 0 ) } } ( { \\pmb x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 541, + 449, + 554 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 450, + 541, + 478, + 553 + ], + "score": 0.91, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 541, + 481, + 554 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 528, + 505, + 554 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 502, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 388, + 568 + ], + "score": 1.0, + "content": "Lemma D.4 (Cao and Gu (2019a)). There exists an absolute constant", + "type": "text" + }, + { + "bbox": [ + 388, + 555, + 397, + 565 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 552, + 504, + 568 + ], + "score": 1.0, + "content": "such that, with probability", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 565, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 139, + 580 + ], + "score": 1.0, + "content": "at least", + "type": "text" + }, + { + "bbox": [ + 139, + 566, + 261, + 579 + ], + "score": 0.91, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 565, + 294, + 580 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 295, + 566, + 364, + 579 + ], + "score": 0.86, + "content": "i \\in [ n ] , l \\in [ L ]", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 565, + 384, + 580 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 384, + 566, + 460, + 579 + ], + "score": 0.93, + "content": "\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 565, + 483, + 580 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 484, + 567, + 505, + 578 + ], + "score": 0.82, + "content": "\\omega \\ \\leqslant", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 577, + 259, + 591 + ], + "spans": [ + { + "bbox": [ + 107, + 578, + 162, + 590 + ], + "score": 0.92, + "content": "C [ \\log ( m ) ] ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 577, + 259, + 591 + ], + "score": 1.0, + "content": ", it holds uniformly that", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 552, + 505, + 591 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 591, + 365, + 606 + ], + "lines": [ + { + "bbox": [ + 245, + 591, + 365, + 606 + ], + "spans": [ + { + "bbox": [ + 245, + 591, + 365, + 606 + ], + "score": 0.89, + "content": "\\| \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\sqrt { m } ) .", + "type": "interline_equation", + "image_path": "3ea31864bd3ef411c12372d9e785cce422d2bd9da181b1ef7ea1d4762698acf8.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 245, + 591, + 365, + 606 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 504, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 627 + ], + "score": 1.0, + "content": "The following lemma provides a uniform bound of the neural network function value over", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 622, + 162, + 640 + ], + "spans": [ + { + "bbox": [ + 107, + 625, + 156, + 639 + ], + "score": 0.92, + "content": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 622, + 162, + 640 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 612, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 104, + 638, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 638, + 217, + 655 + ], + "score": 1.0, + "content": "Lemma D.5. Suppose that", + "type": "text" + }, + { + "bbox": [ + 218, + 640, + 316, + 653 + ], + "score": 0.9, + "content": "m \\geqslant \\Omega ( \\omega ^ { - 2 / 3 } \\log ( n / \\delta ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 638, + 335, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 335, + 640, + 419, + 653 + ], + "score": 0.91, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 638, + 505, + 655 + ], + "score": 1.0, + "content": ". Then with probabil-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 651, + 460, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 651, + 150, + 668 + ], + "score": 1.0, + "content": "ity at least", + "type": "text" + }, + { + "bbox": [ + 150, + 654, + 173, + 665 + ], + "score": 0.44, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 651, + 176, + 668 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 176, + 653, + 326, + 667 + ], + "score": 0.91, + "content": "| f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 651, + 354, + 668 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 354, + 653, + 456, + 666 + ], + "score": 0.92, + "content": "\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) \\ i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 651, + 460, + 668 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 638, + 505, + 668 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 668, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 501, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 172, + 685 + ], + "score": 1.0, + "content": "Lemma D.6. If", + "type": "text" + }, + { + "bbox": [ + 173, + 668, + 263, + 682 + ], + "score": 0.93, + "content": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 666, + 383, + 685 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 384, + 668, + 501, + 682 + ], + "score": 0.91, + "content": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]", + "type": "inline_equation" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 666, + 501, + 685 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 684, + 466, + 717 + ], + "lines": [ + { + "bbox": [ + 143, + 684, + 466, + 717 + ], + "spans": [ + { + "bbox": [ + 143, + 684, + 466, + 717 + ], + "score": 0.84, + "content": "\\begin{array} { r l } & { \\| \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } ) , } \\\\ & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } m ) } \\end{array}", + "type": "interline_equation", + "image_path": "b3a03af7ab37fba24e0f0a2b59e5305a6930a79ac7298eb27f92f385cdcffb08.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 143, + 684, + 466, + 695.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 143, + 695.0, + 466, + 706.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 143, + 706.0, + 466, + 717.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 719, + 254, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 716, + 252, + 736 + ], + "spans": [ + { + "bbox": [ + 104, + 716, + 133, + 736 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 719, + 205, + 732 + ], + "score": 0.94, + "content": "\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 716, + 223, + 736 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 223, + 720, + 252, + 732 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 716, + 252, + 736 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 281, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 282, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 282, + 95 + ], + "score": 1.0, + "content": "D.2 PROOFS OF LEMMAS D.5 AND D.6", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 335, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 336, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 336, + 117 + ], + "score": 1.0, + "content": "Proof of Lemma D.5. By Lemmas D.2 and D.4, we have", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 119, + 513, + 171 + ], + "lines": [ + { + "bbox": [ + 111, + 119, + 513, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 119, + 513, + 171 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { F } + \\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 2 } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\| _ { F } } \\\\ & { \\qquad + \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } ) \\cdot \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( \\omega \\sqrt { m } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "9b0b16369284ee8c4572991ce051730bee7663841d031fa484660bf7dc5c2903.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 111, + 119, + 513, + 136.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 136.33333333333334, + 513, + 153.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 153.66666666666669, + 513, + 171.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 295, + 189 + ], + "score": 1.0, + "content": "where the last inequality is by the assumption", + "type": "text" + }, + { + "bbox": [ + 295, + 174, + 365, + 187 + ], + "score": 0.93, + "content": "\\omega \\leqslant [ \\log ( m ) ] ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 173, + 506, + 189 + ], + "score": 1.0, + "content": ". Applying triangle inequality and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 184, + 200, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 200, + 199 + ], + "score": 1.0, + "content": "Lemma D.3 then gives", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 201, + 470, + 236 + ], + "lines": [ + { + "bbox": [ + 140, + 201, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 470, + 236 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant | f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | + | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } ) + \\mathcal { O } ( \\omega \\sqrt { m } ) } \\\\ & { \\qquad = \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "ba830733a52927b5172128e16ae55612568311868beaa93938e576110a0bdc8c.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 140, + 201, + 470, + 212.66666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 140, + 212.66666666666666, + 470, + 224.33333333333331 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 140, + 224.33333333333331, + 470, + 235.99999999999997 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 210, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 210, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 210, + 253 + ], + "score": 1.0, + "content": "This completes the proof.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 318, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 262, + 319, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 319, + 276 + ], + "score": 1.0, + "content": "Proof of Lemma D.6. By direct calculation, we have", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 279, + 457, + 296 + ], + "lines": [ + { + "bbox": [ + 153, + 279, + 457, + 296 + ], + "spans": [ + { + "bbox": [ + 153, + 279, + 457, + 296 + ], + "score": 0.89, + "content": "\\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } .", + "type": "interline_equation", + "image_path": "32652103ff49fa8f8ca461812435f446cf7cd55d9e3e8559eb4e66cc5c9939dc.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 153, + 279, + 457, + 296 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 299, + 183, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 183, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 183, + 312 + ], + "score": 1.0, + "content": "Therefore we have", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 312, + 506, + 385 + ], + "lines": [ + { + "bbox": [ + 111, + 312, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 111, + 312, + 506, + 385 + ], + "score": 0.92, + "content": "\\begin{array} { r l } { | \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | | _ { F } = \\sqrt { m } \\cdot \\| \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { \\leqslant \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } + \\sqrt { m } \\cdot \\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } } & { } \\end{array}", + "type": "interline_equation", + "image_path": "1167691bc5adceb2bf869d56c3d675bc732d2cec630f37b53c09c75ff88b73a7.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 111, + 312, + 506, + 336.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 111, + 336.3333333333333, + 506, + 360.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 111, + 360.66666666666663, + 506, + 384.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 503, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 470, + 400 + ], + "score": 1.0, + "content": "By Lemma 7.4 in Allen-Zhu et al. (2018) and Lemma D.1, with probability at least", + "type": "text" + }, + { + "bbox": [ + 470, + 387, + 503, + 398 + ], + "score": 0.79, + "content": "1 - n \\cdot", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 394, + 507, + 416 + ], + "spans": [ + { + "bbox": [ + 107, + 399, + 161, + 413 + ], + "score": 0.69, + "content": "\\exp [ - \\Omega ( m ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 394, + 166, + 416 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 167, + 398, + 356, + 413 + ], + "score": 0.88, + "content": "\\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 394, + 387, + 416 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 388, + 400, + 421, + 413 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 394, + 507, + 416 + ], + "score": 1.0, + "content": ". Moreover, clearly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 408, + 343, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 293, + 427 + ], + "score": 0.89, + "content": "\\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } \\leqslant \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } \\| _ { F } \\leqslant \\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 408, + 343, + 430 + ], + "score": 1.0, + "content": ". Therefore", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 431, + 411, + 447 + ], + "lines": [ + { + "bbox": [ + 199, + 431, + 411, + 447 + ], + "spans": [ + { + "bbox": [ + 199, + 431, + 411, + 447 + ], + "score": 0.87, + "content": "\\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )", + "type": "interline_equation", + "image_path": "82517f5e7b596fb3aed17f5a1d9746417259ebcfeaab15f43c40b34fb6100eb8.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 199, + 431, + 411, + 447 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 133, + 465 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 451, + 161, + 463 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 449, + 505, + 465 + ], + "score": 1.0, + "content": ". This proves the bound for the first layer gradients. For the second layer gradients, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 460, + 129, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 460, + 129, + 475 + ], + "score": 1.0, + "content": "have", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 475, + 461, + 492 + ], + "lines": [ + { + "bbox": [ + 150, + 475, + 461, + 492 + ], + "spans": [ + { + "bbox": [ + 150, + 475, + 461, + 492 + ], + "score": 0.91, + "content": "\\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) ] ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) ] ^ { \\top }", + "type": "interline_equation", + "image_path": "56cf3fc46e0b3d361f160b309cf351e307ff1bca25947efa4ed56b26b269f77e.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 150, + 475, + 461, + 492 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 352, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 351, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 314, + 510 + ], + "score": 1.0, + "content": "It therefore follows by the 1-Lipschitz continuity of", + "type": "text" + }, + { + "bbox": [ + 315, + 496, + 332, + 508 + ], + "score": 0.91, + "content": "\\sigma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 494, + 351, + 510 + ], + "score": 1.0, + "content": "that", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 512, + 484, + 529 + ], + "lines": [ + { + "bbox": [ + 127, + 512, + 484, + 529 + ], + "spans": [ + { + "bbox": [ + 127, + 512, + 484, + 529 + ], + "score": 0.91, + "content": "\\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant { \\sqrt { m } } \\cdot \\| \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } \\| _ { F } \\leqslant \\omega { \\sqrt { m } } \\leqslant \\omega ^ { 1 / 3 } { \\sqrt { m } } .", + "type": "interline_equation", + "image_path": "3ddd54a319073f6f2b0f103a2fd96e83b20c05e896855a3d94a0990dedab790a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 127, + 512, + 484, + 529 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 294, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 296, + 547 + ], + "score": 1.0, + "content": "This completes the proof of the first inequality.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 419, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 420, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 420, + 564 + ], + "score": 1.0, + "content": "The second inequality directly follows by triangle inequality and Lemma D.4:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 565, + 499, + 614 + ], + "lines": [ + { + "bbox": [ + 112, + 565, + 499, + 614 + ], + "spans": [ + { + "bbox": [ + 112, + 565, + 499, + 614 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - m \\mathbf { K } ^ { ( 0 ) } | \\leqslant | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\leqslant { \\mathcal { O } } ( \\omega ^ { 1 / 3 } m ) . } \\end{array}", + "type": "interline_equation", + "image_path": "17041c13493a2ab7b5f12096dece9a8b00cd58c4a276014fbb609c238691cec8.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 112, + 565, + 499, + 581.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 112, + 581.3333333333334, + 499, + 597.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 112, + 597.6666666666667, + 499, + 614.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 200, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 201, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 201, + 629 + ], + "score": 1.0, + "content": "This finishes the proof.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 641, + 329, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 641, + 329, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 329, + 656 + ], + "score": 1.0, + "content": "E ADDITIONAL EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 107, + 666, + 387, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 387, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 387, + 679 + ], + "score": 1.0, + "content": "E.1 ESTIMATING THE PROJECTION LENGTH IN FUNCTION SPACE", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "As mentioned in Section 4.1, when using freshly sampled points, we are actually estimating the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 258, + 712 + ], + "score": 1.0, + "content": "projection length of residual function", + "type": "text" + }, + { + "bbox": [ + 258, + 699, + 370, + 711 + ], + "score": 0.91, + "content": "r ( \\mathbf { x } ) = f ^ { * } ( \\mathbf { x } ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "onto the given Gegenbauer poly-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 137, + 723 + ], + "score": 1.0, + "content": "nomial", + "type": "text" + }, + { + "bbox": [ + 137, + 709, + 163, + 722 + ], + "score": 0.92, + "content": "P _ { k } ( { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 710, + 505, + 723 + ], + "score": 1.0, + "content": ". Here we present in Figure 3 a comparison between the projection length using train-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "ing data and that using test data. An interesting phenomenon is that the network generalizes well", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 239, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 496, + 241, + 504, + 249 + ], + "spans": [ + { + "bbox": [ + 496, + 241, + 504, + 249 + ], + "score": 0.995, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 617, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 495, + 618, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 495, + 618, + 505, + 628 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 281, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 282, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 282, + 95 + ], + "score": 1.0, + "content": "D.2 PROOFS OF LEMMAS D.5 AND D.6", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 82, + 282, + 95 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 335, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 336, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 336, + 117 + ], + "score": 1.0, + "content": "Proof of Lemma D.5. By Lemmas D.2 and D.4, we have", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 102, + 336, + 117 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 119, + 513, + 171 + ], + "lines": [ + { + "bbox": [ + 111, + 119, + 513, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 119, + 513, + 171 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { F } + \\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 2 } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\| _ { F } } \\\\ & { \\qquad + \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } ) \\cdot \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( \\omega \\sqrt { m } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "9b0b16369284ee8c4572991ce051730bee7663841d031fa484660bf7dc5c2903.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 111, + 119, + 513, + 136.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 136.33333333333334, + 513, + 153.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 153.66666666666669, + 513, + 171.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 295, + 189 + ], + "score": 1.0, + "content": "where the last inequality is by the assumption", + "type": "text" + }, + { + "bbox": [ + 295, + 174, + 365, + 187 + ], + "score": 0.93, + "content": "\\omega \\leqslant [ \\log ( m ) ] ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 173, + 506, + 189 + ], + "score": 1.0, + "content": ". Applying triangle inequality and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 184, + 200, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 200, + 199 + ], + "score": 1.0, + "content": "Lemma D.3 then gives", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 173, + 506, + 199 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 201, + 470, + 236 + ], + "lines": [ + { + "bbox": [ + 140, + 201, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 470, + 236 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant | f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | + | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } ) + \\mathcal { O } ( \\omega \\sqrt { m } ) } \\\\ & { \\qquad = \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "ba830733a52927b5172128e16ae55612568311868beaa93938e576110a0bdc8c.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 140, + 201, + 470, + 212.66666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 140, + 212.66666666666666, + 470, + 224.33333333333331 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 140, + 224.33333333333331, + 470, + 235.99999999999997 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 210, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 210, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 210, + 253 + ], + "score": 1.0, + "content": "This completes the proof.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 237, + 210, + 253 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 318, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 262, + 319, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 319, + 276 + ], + "score": 1.0, + "content": "Proof of Lemma D.6. By direct calculation, we have", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 262, + 319, + 276 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 279, + 457, + 296 + ], + "lines": [ + { + "bbox": [ + 153, + 279, + 457, + 296 + ], + "spans": [ + { + "bbox": [ + 153, + 279, + 457, + 296 + ], + "score": 0.89, + "content": "\\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } .", + "type": "interline_equation", + "image_path": "32652103ff49fa8f8ca461812435f446cf7cd55d9e3e8559eb4e66cc5c9939dc.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 153, + 279, + 457, + 296 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 299, + 183, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 183, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 183, + 312 + ], + "score": 1.0, + "content": "Therefore we have", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 298, + 183, + 312 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 312, + 506, + 385 + ], + "lines": [ + { + "bbox": [ + 111, + 312, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 111, + 312, + 506, + 385 + ], + "score": 0.92, + "content": "\\begin{array} { r l } { | \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | | _ { F } = \\sqrt { m } \\cdot \\| \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { \\leqslant \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } + \\sqrt { m } \\cdot \\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } } & { } \\end{array}", + "type": "interline_equation", + "image_path": "1167691bc5adceb2bf869d56c3d675bc732d2cec630f37b53c09c75ff88b73a7.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 111, + 312, + 506, + 336.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 111, + 336.3333333333333, + 506, + 360.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 111, + 360.66666666666663, + 506, + 384.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 503, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 470, + 400 + ], + "score": 1.0, + "content": "By Lemma 7.4 in Allen-Zhu et al. (2018) and Lemma D.1, with probability at least", + "type": "text" + }, + { + "bbox": [ + 470, + 387, + 503, + 398 + ], + "score": 0.79, + "content": "1 - n \\cdot", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 394, + 507, + 416 + ], + "spans": [ + { + "bbox": [ + 107, + 399, + 161, + 413 + ], + "score": 0.69, + "content": "\\exp [ - \\Omega ( m ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 394, + 166, + 416 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 167, + 398, + 356, + 413 + ], + "score": 0.88, + "content": "\\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 394, + 387, + 416 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 388, + 400, + 421, + 413 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 394, + 507, + 416 + ], + "score": 1.0, + "content": ". Moreover, clearly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 408, + 343, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 293, + 427 + ], + "score": 0.89, + "content": "\\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } \\leqslant \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } \\| _ { F } \\leqslant \\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 408, + 343, + 430 + ], + "score": 1.0, + "content": ". Therefore", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 385, + 507, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 431, + 411, + 447 + ], + "lines": [ + { + "bbox": [ + 199, + 431, + 411, + 447 + ], + "spans": [ + { + "bbox": [ + 199, + 431, + 411, + 447 + ], + "score": 0.87, + "content": "\\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )", + "type": "interline_equation", + "image_path": "82517f5e7b596fb3aed17f5a1d9746417259ebcfeaab15f43c40b34fb6100eb8.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 199, + 431, + 411, + 447 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 133, + 465 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 451, + 161, + 463 + ], + "score": 0.92, + "content": "i \\in [ n ]", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 449, + 505, + 465 + ], + "score": 1.0, + "content": ". This proves the bound for the first layer gradients. For the second layer gradients, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 460, + 129, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 460, + 129, + 475 + ], + "score": 1.0, + "content": "have", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 449, + 505, + 475 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 475, + 461, + 492 + ], + "lines": [ + { + "bbox": [ + 150, + 475, + 461, + 492 + ], + "spans": [ + { + "bbox": [ + 150, + 475, + 461, + 492 + ], + "score": 0.91, + "content": "\\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) ] ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) ] ^ { \\top }", + "type": "interline_equation", + "image_path": "56cf3fc46e0b3d361f160b309cf351e307ff1bca25947efa4ed56b26b269f77e.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 150, + 475, + 461, + 492 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 352, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 351, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 314, + 510 + ], + "score": 1.0, + "content": "It therefore follows by the 1-Lipschitz continuity of", + "type": "text" + }, + { + "bbox": [ + 315, + 496, + 332, + 508 + ], + "score": 0.91, + "content": "\\sigma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 494, + 351, + 510 + ], + "score": 1.0, + "content": "that", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 494, + 351, + 510 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 512, + 484, + 529 + ], + "lines": [ + { + "bbox": [ + 127, + 512, + 484, + 529 + ], + "spans": [ + { + "bbox": [ + 127, + 512, + 484, + 529 + ], + "score": 0.91, + "content": "\\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant { \\sqrt { m } } \\cdot \\| \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } \\| _ { F } \\leqslant \\omega { \\sqrt { m } } \\leqslant \\omega ^ { 1 / 3 } { \\sqrt { m } } .", + "type": "interline_equation", + "image_path": "3ddd54a319073f6f2b0f103a2fd96e83b20c05e896855a3d94a0990dedab790a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 127, + 512, + 484, + 529 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 294, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 296, + 547 + ], + "score": 1.0, + "content": "This completes the proof of the first inequality.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 531, + 296, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 419, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 420, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 420, + 564 + ], + "score": 1.0, + "content": "The second inequality directly follows by triangle inequality and Lemma D.4:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 549, + 420, + 564 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 565, + 499, + 614 + ], + "lines": [ + { + "bbox": [ + 112, + 565, + 499, + 614 + ], + "spans": [ + { + "bbox": [ + 112, + 565, + 499, + 614 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - m \\mathbf { K } ^ { ( 0 ) } | \\leqslant | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\leqslant { \\mathcal { O } } ( \\omega ^ { 1 / 3 } m ) . } \\end{array}", + "type": "interline_equation", + "image_path": "17041c13493a2ab7b5f12096dece9a8b00cd58c4a276014fbb609c238691cec8.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 112, + 565, + 499, + 581.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 112, + 581.3333333333334, + 499, + 597.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 112, + 597.6666666666667, + 499, + 614.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 200, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 201, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 201, + 629 + ], + "score": 1.0, + "content": "This finishes the proof.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 614, + 201, + 629 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 641, + 329, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 641, + 329, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 329, + 656 + ], + "score": 1.0, + "content": "E ADDITIONAL EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 107, + 666, + 387, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 387, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 387, + 679 + ], + "score": 1.0, + "content": "E.1 ESTIMATING THE PROJECTION LENGTH IN FUNCTION SPACE", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "As mentioned in Section 4.1, when using freshly sampled points, we are actually estimating the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 258, + 712 + ], + "score": 1.0, + "content": "projection length of residual function", + "type": "text" + }, + { + "bbox": [ + 258, + 699, + 370, + 711 + ], + "score": 0.91, + "content": "r ( \\mathbf { x } ) = f ^ { * } ( \\mathbf { x } ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "onto the given Gegenbauer poly-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 137, + 723 + ], + "score": 1.0, + "content": "nomial", + "type": "text" + }, + { + "bbox": [ + 137, + 709, + 163, + 722 + ], + "score": 0.92, + "content": "P _ { k } ( { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 710, + 505, + 723 + ], + "score": 1.0, + "content": ". Here we present in Figure 3 a comparison between the projection length using train-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "ing data and that using test data. An interesting phenomenon is that the network generalizes well", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "on the lower-order Gegenbauer polynomial and along the highest-order Gegenbauer polynomial the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 217, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 217, + 107 + ], + "score": 1.0, + "content": "network suffers overfitting.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 687, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "on the lower-order Gegenbauer polynomial and along the highest-order Gegenbauer polynomial the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 217, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 217, + 107 + ], + "score": 1.0, + "content": "network suffers overfitting.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "image", + "bbox": [ + 111, + 115, + 498, + 443 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 115, + 498, + 443 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 115, + 498, + 443 + ], + "spans": [ + { + "bbox": [ + 111, + 115, + 498, + 443 + ], + "score": 0.975, + "type": "image", + "image_path": "ede46709d91cb04a929d72bd1ecb66385cb7ce3faf11dd60ee6d67e97d71b25f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 111, + 115, + 498, + 224.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 224.33333333333331, + 498, + 333.66666666666663 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 333.66666666666663, + 498, + 442.99999999999994 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 455, + 505, + 500 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "Figure 3: Convergence curve for projection length onto vectors (determined by training data) and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "functions (estimated by test data). We can see that for low-order Gegenbauer polynomials, the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "network learns the function while for the high-order Gegenbauer polynomial, the network overfits", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 489, + 176, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 176, + 501 + ], + "score": 1.0, + "content": "the training data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + } + ], + "index": 4.75 + }, + { + "type": "title", + "bbox": [ + 106, + 513, + 308, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 309, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 309, + 524 + ], + "score": 1.0, + "content": "E.2 NEAR-LINEAR CONVERGENCE BEHAVIOR", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "score": 1.0, + "content": "In this subsection, we present the same curve shown in Section 4.1 in logarithmic scale instead of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "linear scale. As shown in Figure 4 we can see that the convergence of different projection length", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "is close to linear convergence, which is well-aligned with our theorem. Note that we performed a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 567, + 499, + 581 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 499, + 581 + ], + "score": 1.0, + "content": "moving average of range 20 on these curves to avoid the heavy oscillation especially in late stage.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 107 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 111, + 115, + 498, + 443 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 115, + 498, + 443 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 115, + 498, + 443 + ], + "spans": [ + { + "bbox": [ + 111, + 115, + 498, + 443 + ], + "score": 0.975, + "type": "image", + "image_path": "ede46709d91cb04a929d72bd1ecb66385cb7ce3faf11dd60ee6d67e97d71b25f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 111, + 115, + 498, + 224.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 224.33333333333331, + 498, + 333.66666666666663 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 333.66666666666663, + 498, + 442.99999999999994 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 455, + 505, + 500 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "Figure 3: Convergence curve for projection length onto vectors (determined by training data) and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "functions (estimated by test data). We can see that for low-order Gegenbauer polynomials, the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "network learns the function while for the high-order Gegenbauer polynomial, the network overfits", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 489, + 176, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 176, + 501 + ], + "score": 1.0, + "content": "the training data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + } + ], + "index": 4.75 + }, + { + "type": "title", + "bbox": [ + 106, + 513, + 308, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 309, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 309, + 524 + ], + "score": 1.0, + "content": "E.2 NEAR-LINEAR CONVERGENCE BEHAVIOR", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "score": 1.0, + "content": "In this subsection, we present the same curve shown in Section 4.1 in logarithmic scale instead of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "linear scale. As shown in Figure 4 we can see that the convergence of different projection length", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "is close to linear convergence, which is well-aligned with our theorem. Note that we performed a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 567, + 499, + 581 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 499, + 581 + ], + "score": 1.0, + "content": "moving average of range 20 on these curves to avoid the heavy oscillation especially in late stage.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 534, + 506, + 581 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 300, + 498, + 459 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 300, + 498, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 300, + 498, + 459 + ], + "spans": [ + { + "bbox": [ + 111, + 300, + 498, + 459 + ], + "score": 0.958, + "type": "image", + "image_path": "d4b04e61fd87b82bba68d9aeb4f350e54f3f281cc4e3ecdbada8bf74711541a3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 300, + 498, + 353.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 353.0, + 498, + 406.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 406.0, + 498, + 459.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 471, + 505, + 516 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "Figure 4: Log-scale convergence curve for projection length onto different component. (a) shows", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "the curve when the target function have different component with the same scale. (b) shows the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "curve when the higner-order components have larger scale. Both exhibit nearly linear convergence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 504, + 201, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 201, + 519 + ], + "score": 1.0, + "content": "especially at late stage.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 300, + 498, + 459 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 300, + 498, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 300, + 498, + 459 + ], + "spans": [ + { + "bbox": [ + 111, + 300, + 498, + 459 + ], + "score": 0.958, + "type": "image", + "image_path": "d4b04e61fd87b82bba68d9aeb4f350e54f3f281cc4e3ecdbada8bf74711541a3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 300, + 498, + 353.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 353.0, + 498, + 406.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 406.0, + 498, + 459.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 471, + 505, + 516 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "Figure 4: Log-scale convergence curve for projection length onto different component. (a) shows", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "the curve when the target function have different component with the same scale. (b) shows the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "curve when the higner-order components have larger scale. Both exhibit nearly linear convergence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 504, + 201, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 201, + 519 + ], + "score": 1.0, + "content": "especially at late stage.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/BJg15lrKvS/BJg15lrKvS_model.json b/parse/train/BJg15lrKvS/BJg15lrKvS_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9d5c0eb7141079a9a91b21f6e9173054fb86eae0 --- /dev/null +++ b/parse/train/BJg15lrKvS/BJg15lrKvS_model.json @@ -0,0 +1,33058 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 583, + 1302, + 583, + 1302, + 1039, + 398, + 1039 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1147, + 1403, + 1147, + 1403, + 1452, + 298, + 1452 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1698, + 1404, + 1698, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1468, + 1403, + 1468, + 1403, + 1682, + 299, + 1682 + ], + "score": 0.979 + }, + { + "category_id": 0, + "poly": [ + 300, + 219, + 1400, + 219, + 1400, + 322, + 300, + 322 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 302, + 1081, + 573, + 1081, + 573, + 1117, + 302, + 1117 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 773, + 520, + 926, + 520, + 926, + 553, + 773, + 553 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.635 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 216.0, + 1406.0, + 216.0, + 1406.0, + 273.0, + 293.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 274.0, + 649.0, + 274.0, + 649.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1078.0, + 579.0, + 1078.0, + 579.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 519.0, + 932.0, + 519.0, + 932.0, + 556.0, + 769.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 582.0, + 1304.0, + 582.0, + 1304.0, + 619.0, + 394.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 615.0, + 1303.0, + 615.0, + 1303.0, + 649.0, + 394.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 644.0, + 1306.0, + 644.0, + 1306.0, + 679.0, + 394.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 676.0, + 1306.0, + 676.0, + 1306.0, + 709.0, + 392.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 704.0, + 1304.0, + 704.0, + 1304.0, + 741.0, + 393.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 734.0, + 1306.0, + 734.0, + 1306.0, + 771.0, + 393.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 765.0, + 1306.0, + 765.0, + 1306.0, + 802.0, + 393.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 798.0, + 1305.0, + 798.0, + 1305.0, + 828.0, + 393.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 828.0, + 1306.0, + 828.0, + 1306.0, + 861.0, + 392.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 857.0, + 1307.0, + 857.0, + 1307.0, + 892.0, + 393.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 890.0, + 1306.0, + 890.0, + 1306.0, + 921.0, + 394.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 918.0, + 1305.0, + 918.0, + 1305.0, + 952.0, + 394.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 950.0, + 1306.0, + 950.0, + 1306.0, + 982.0, + 395.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 978.0, + 1307.0, + 978.0, + 1307.0, + 1015.0, + 392.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1011.0, + 511.0, + 1011.0, + 511.0, + 1041.0, + 394.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1145.0, + 1404.0, + 1145.0, + 1404.0, + 1182.0, + 294.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 1405.0, + 1178.0, + 1405.0, + 1213.0, + 294.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1243.0, + 293.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 1405.0, + 1239.0, + 1405.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1301.0, + 1404.0, + 1301.0, + 1404.0, + 1332.0, + 296.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 1405.0, + 1330.0, + 1405.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1360.0, + 1403.0, + 1360.0, + 1403.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1392.0, + 1405.0, + 1392.0, + 1405.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1421.0, + 995.0, + 1421.0, + 995.0, + 1455.0, + 296.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1698.0, + 1404.0, + 1698.0, + 1404.0, + 1732.0, + 296.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1730.0, + 1404.0, + 1730.0, + 1404.0, + 1761.0, + 296.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1758.0, + 1404.0, + 1758.0, + 1404.0, + 1793.0, + 293.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1788.0, + 1403.0, + 1788.0, + 1403.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1816.0, + 1405.0, + 1816.0, + 1405.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1852.0, + 1402.0, + 1852.0, + 1402.0, + 1883.0, + 296.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 1404.0, + 1878.0, + 1404.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1404.0, + 1911.0, + 1404.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 1380.0, + 2002.0, + 1380.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1465.0, + 1405.0, + 1465.0, + 1405.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1404.0, + 1528.0, + 1404.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1562.0, + 1404.0, + 1562.0, + 1404.0, + 1596.0, + 295.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1591.0, + 1403.0, + 1591.0, + 1403.0, + 1625.0, + 295.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1618.0, + 1404.0, + 1618.0, + 1404.0, + 1658.0, + 293.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1654.0, + 619.0, + 1654.0, + 619.0, + 1685.0, + 295.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1377, + 1403, + 1377, + 1403, + 1867, + 297, + 1867 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 521, + 1403, + 521, + 1403, + 735, + 298, + 735 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1400, + 1881, + 1400, + 2035, + 299, + 2035 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 297, + 230, + 1403, + 230, + 1403, + 506, + 297, + 506 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 302, + 750, + 795, + 750, + 795, + 782, + 302, + 782 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 303, + 1319, + 723, + 1319, + 723, + 1349, + 303, + 1349 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 297, + 809, + 1406, + 809, + 1406, + 1292, + 297, + 1292 + ], + "score": 0.71 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.695 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.159 + }, + { + "category_id": 13, + "poly": [ + 759, + 2005, + 848, + 2005, + 848, + 2036, + 759, + 2036 + ], + "score": 0.93, + "latex": "L _ { \\kappa } f ( x )" + }, + { + "category_id": 13, + "poly": [ + 510, + 1032, + 627, + 1032, + 627, + 1067, + 510, + 1067 + ], + "score": 0.93, + "latex": "\\Omega ( k ^ { - d - \\overline { { 1 } } } )" + }, + { + "category_id": 13, + "poly": [ + 1003, + 1804, + 1071, + 1804, + 1071, + 1837, + 1003, + 1837 + ], + "score": 0.93, + "latex": "\\kappa ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 562, + 1803, + 926, + 1803, + 926, + 1839, + 562, + 1839 + ], + "score": 0.92, + "latex": "L _ { \\kappa } \\bar { f } ( s ) : = \\int _ { \\mathbb { S } ^ { d } } \\kappa \\bar { ( } x , s ) f ( s ) d \\bar { \\tau } \\bar { ( } s )" + }, + { + "category_id": 13, + "poly": [ + 834, + 1001, + 1176, + 1001, + 1176, + 1036, + 834, + 1036 + ], + "score": 0.92, + "latex": "\\mu _ { k } = \\Omega ( \\operatorname* { m a x } \\{ k ^ { - d - 1 } , \\check { d } ^ { - k + 1 } \\} )" + }, + { + "category_id": 13, + "poly": [ + 1348, + 1804, + 1400, + 1804, + 1400, + 1837, + 1348, + 1837 + ], + "score": 0.91, + "latex": "\\tau ( s )" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1036, + 1184, + 1036, + 1184, + 1063, + 1107, + 1063 + ], + "score": 0.88, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 746, + 1774, + 779, + 1774, + 779, + 1804, + 746, + 1804 + ], + "score": 0.87, + "latex": "f ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1271, + 445, + 1289, + 445, + 1289, + 471, + 1271, + 471 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 297, + 1748, + 318, + 1748, + 318, + 1770, + 297, + 1770 + ], + "score": 0.72, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 776, + 902, + 810, + 902, + 810, + 929, + 776, + 929 + ], + "score": 0.48, + "latex": "\\mathtt { S u }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1317.0, + 728.0, + 1317.0, + 728.0, + 1352.0, + 297.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1375.0, + 1407.0, + 1375.0, + 1407.0, + 1413.0, + 295.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1407.0, + 1404.0, + 1407.0, + 1404.0, + 1441.0, + 295.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1472.0, + 293.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 1404.0, + 1470.0, + 1404.0, + 1500.0, + 296.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1500.0, + 1401.0, + 1500.0, + 1401.0, + 1531.0, + 296.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1527.0, + 1404.0, + 1527.0, + 1404.0, + 1563.0, + 296.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1562.0, + 1403.0, + 1562.0, + 1403.0, + 1592.0, + 296.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1622.0, + 296.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1618.0, + 1405.0, + 1618.0, + 1405.0, + 1655.0, + 295.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 1404.0, + 1650.0, + 1404.0, + 1685.0, + 293.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1715.0, + 295.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1739.0, + 296.0, + 1739.0, + 296.0, + 1779.0, + 291.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1739.0, + 1407.0, + 1739.0, + 1407.0, + 1779.0, + 319.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 745.0, + 1771.0, + 745.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1808.0, + 780.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1800.0, + 561.0, + 1800.0, + 561.0, + 1841.0, + 292.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1800.0, + 1002.0, + 1800.0, + 1002.0, + 1841.0, + 927.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1800.0, + 1347.0, + 1800.0, + 1347.0, + 1841.0, + 1072.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1800.0, + 1404.0, + 1800.0, + 1404.0, + 1841.0, + 1401.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1835.0, + 1346.0, + 1835.0, + 1346.0, + 1869.0, + 293.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1404.0, + 520.0, + 1404.0, + 558.0, + 294.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 588.0, + 294.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 581.0, + 1404.0, + 581.0, + 1404.0, + 618.0, + 293.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 648.0, + 293.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 1405.0, + 643.0, + 1405.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 676.0, + 1405.0, + 676.0, + 1405.0, + 707.0, + 296.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 704.0, + 1212.0, + 704.0, + 1212.0, + 738.0, + 293.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1402.0, + 1909.0, + 1402.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 297.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 758.0, + 2002.0, + 758.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2039.0, + 849.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 262.0, + 296.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 257.0, + 1404.0, + 257.0, + 1404.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 290.0, + 1400.0, + 290.0, + 1400.0, + 322.0, + 296.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 355.0, + 292.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 348.0, + 1405.0, + 348.0, + 1405.0, + 388.0, + 292.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 382.0, + 1401.0, + 382.0, + 1401.0, + 415.0, + 295.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 446.0, + 296.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 442.0, + 1270.0, + 442.0, + 1270.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 442.0, + 1405.0, + 442.0, + 1405.0, + 479.0, + 1290.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 475.0, + 729.0, + 475.0, + 729.0, + 508.0, + 296.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 750.0, + 800.0, + 750.0, + 800.0, + 786.0, + 298.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 810.0, + 1403.0, + 810.0, + 1403.0, + 842.0, + 299.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 838.0, + 1406.0, + 838.0, + 1406.0, + 875.0, + 329.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 905.0, + 331.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 900.0, + 775.0, + 900.0, + 775.0, + 936.0, + 328.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 900.0, + 1404.0, + 900.0, + 1404.0, + 936.0, + 811.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 928.0, + 1404.0, + 928.0, + 1404.0, + 970.0, + 328.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 974.0, + 1406.0, + 974.0, + 1406.0, + 1010.0, + 299.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 996.0, + 833.0, + 996.0, + 833.0, + 1044.0, + 326.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 996.0, + 1409.0, + 996.0, + 1409.0, + 1044.0, + 1177.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1028.0, + 509.0, + 1028.0, + 509.0, + 1073.0, + 328.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1028.0, + 1106.0, + 1028.0, + 1106.0, + 1073.0, + 628.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1028.0, + 1408.0, + 1028.0, + 1408.0, + 1073.0, + 1185.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1063.0, + 583.0, + 1063.0, + 583.0, + 1103.0, + 329.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1105.0, + 1406.0, + 1105.0, + 1406.0, + 1142.0, + 297.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1138.0, + 1406.0, + 1138.0, + 1406.0, + 1174.0, + 329.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1165.0, + 1406.0, + 1165.0, + 1406.0, + 1205.0, + 328.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1198.0, + 1404.0, + 1198.0, + 1404.0, + 1234.0, + 328.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1230.0, + 1404.0, + 1230.0, + 1404.0, + 1263.0, + 329.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1260.0, + 854.0, + 1260.0, + 854.0, + 1293.0, + 330.0, + 1293.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 765, + 1406, + 765, + 1406, + 1110, + 297, + 1110 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1522, + 1404, + 1522, + 1404, + 1622, + 298, + 1622 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1403, + 228, + 1403, + 385, + 299, + 385 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1318, + 1405, + 1318, + 1405, + 1416, + 297, + 1416 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 398, + 1403, + 398, + 1403, + 492, + 299, + 492 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 588, + 1404, + 588, + 1404, + 682, + 299, + 682 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 619, + 1426, + 1081, + 1426, + 1081, + 1509, + 619, + 1509 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 299, + 1635, + 1397, + 1635, + 1397, + 1699, + 299, + 1699 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 1192, + 1402, + 1192, + 1402, + 1256, + 297, + 1256 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 679, + 1265, + 1021, + 1265, + 1021, + 1305, + 679, + 1305 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 299, + 1133, + 570, + 1133, + 570, + 1165, + 299, + 1165 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 299, + 523, + 579, + 523, + 579, + 558, + 299, + 558 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 298, + 707, + 491, + 707, + 491, + 739, + 298, + 739 + ], + "score": 0.902 + }, + { + "category_id": 5, + "poly": [ + 314, + 1750, + 1285, + 1750, + 1285, + 1955, + 314, + 1955 + ], + "score": 0.829, + "html": "
Input: Number of iterations T,step size n. Generate each entry of W(0) and w)
from N(0,2/m) and N(0,1/m) respectively. 2 fort=0,1,...,T-1do
Update W(t+1) = W(t) - n · VwLs(W(t).
end for Output: W(T).
" + }, + { + "category_id": 2, + "poly": [ + 301, + 1977, + 1402, + 1977, + 1402, + 2034, + 301, + 2034 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2113, + 841, + 2113 + ], + "score": 0.683 + }, + { + "category_id": 6, + "poly": [ + 296, + 1709, + 983, + 1709, + 983, + 1743, + 296, + 1743 + ], + "score": 0.626 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 860, + 2088, + 860, + 2112, + 841, + 2112 + ], + "score": 0.463 + }, + { + "category_id": 7, + "poly": [ + 301, + 1977, + 1402, + 1977, + 1402, + 2034, + 301, + 2034 + ], + "score": 0.151 + }, + { + "category_id": 1, + "poly": [ + 314, + 1750, + 1285, + 1750, + 1285, + 1955, + 314, + 1955 + ], + "score": 0.099 + }, + { + "category_id": 14, + "poly": [ + 617, + 1423, + 1079, + 1423, + 1079, + 1511, + 617, + 1511 + ], + "score": 0.94, + "latex": "L _ { S } ( \\mathbf { W } ) = \\frac { 1 } { n } \\sum _ { ( \\mathbf { x } _ { i } , y _ { i } ) \\in S } \\left( y _ { i } - \\theta f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right) ^ { 2 } ," + }, + { + "category_id": 13, + "poly": [ + 1008, + 1790, + 1132, + 1790, + 1132, + 1824, + 1008, + 1824 + ], + "score": 0.94, + "latex": "N ( 0 , 1 / m )" + }, + { + "category_id": 13, + "poly": [ + 831, + 1789, + 956, + 1789, + 956, + 1824, + 831, + 1824 + ], + "score": 0.93, + "latex": "N ( 0 , 2 / m )" + }, + { + "category_id": 13, + "poly": [ + 550, + 829, + 814, + 829, + 814, + 869, + 550, + 869 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\| \\mathbf { v } \\| _ { p } = ( \\sum _ { i = 1 } ^ { d } | v _ { i } | ^ { p } ) ^ { 1 / p } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 748, + 797, + 1037, + 797, + 1037, + 832, + 748, + 832 + ], + "score": 0.92, + "latex": "\\mathbf { v } = ( \\bar { v _ { 1 } } , \\ldots , v _ { d } ) ^ { T } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 1522, + 1373, + 1522, + 1373, + 1559, + 1283, + 1559 + ], + "score": 0.92, + "latex": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 648, + 1588, + 739, + 1588, + 739, + 1622, + 648, + 1622 + ], + "score": 0.92, + "latex": "| y _ { i } | \\leqslant 1" + }, + { + "category_id": 13, + "poly": [ + 298, + 898, + 587, + 898, + 587, + 941, + 298, + 941 + ], + "score": 0.92, + "latex": "\\textstyle \\| \\mathbf { A } \\| _ { F } = ( \\sum _ { i , j = 1 } ^ { d } A _ { i , j } ^ { 2 } ) ^ { 1 / 2 }" + }, + { + "category_id": 13, + "poly": [ + 744, + 1078, + 795, + 1078, + 795, + 1110, + 744, + 1110 + ], + "score": 0.92, + "latex": "\\Omega ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1113, + 941, + 1349, + 941, + 1349, + 974, + 1113, + 974 + ], + "score": 0.92, + "latex": "\\langle \\mathbf { A } , \\mathbf { B } \\rangle = \\operatorname { T r } ( \\mathbf { A } ^ { \\top } \\mathbf { B } )" + }, + { + "category_id": 13, + "poly": [ + 656, + 1044, + 796, + 1044, + 796, + 1078, + 656, + 1078 + ], + "score": 0.92, + "latex": "a _ { n } = \\Omega ( b _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 973, + 902, + 1280, + 902, + 1280, + 940, + 973, + 940 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\| \\mathbf { A } \\| _ { p } = \\operatorname* { m a x } _ { \\| \\mathbf { v } \\| _ { p } \\leqslant 1 } \\| \\mathbf { A } \\mathbf { v } \\| _ { p } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1070, + 1555, + 1202, + 1555, + 1202, + 1587, + 1070, + 1587 + ], + "score": 0.92, + "latex": "\\mathbb { S } ^ { d } \\in \\mathbb { R } ^ { d + 1 }" + }, + { + "category_id": 13, + "poly": [ + 584, + 1783, + 649, + 1783, + 649, + 1824, + 584, + 1824 + ], + "score": 0.91, + "latex": "\\mathbf { W } _ { 1 } ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 563, + 2003, + 674, + 2003, + 674, + 2031, + 563, + 2031 + ], + "score": 0.91, + "latex": "\\mathbb { S } ^ { d ^ { \\bullet } } \\in \\mathbb { R } ^ { d + 1 }" + }, + { + "category_id": 13, + "poly": [ + 442, + 867, + 626, + 867, + 626, + 899, + 442, + 899 + ], + "score": 0.91, + "latex": "\\mathbf { A } = ( A _ { i , j } ) _ { m \\times n }" + }, + { + "category_id": 13, + "poly": [ + 1211, + 1040, + 1265, + 1040, + 1265, + 1079, + 1211, + 1079 + ], + "score": 0.91, + "latex": "\\tilde { \\mathcal { O } } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 701, + 1783, + 766, + 1783, + 766, + 1825, + 701, + 1825 + ], + "score": 0.91, + "latex": "\\mathbf { W } _ { 2 } ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 467, + 1004, + 518, + 1004, + 518, + 1043, + 467, + 1043 + ], + "score": 0.91, + "latex": "\\widetilde { \\Omega } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1319, + 1041, + 1371, + 1041, + 1371, + 1079, + 1319, + 1079 + ], + "score": 0.91, + "latex": "\\widetilde { \\Omega } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1009, + 1375, + 1009, + 1375, + 1041, + 1235, + 1041 + ], + "score": 0.91, + "latex": "a _ { n } = \\mathcal { O } ( b _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 345, + 1352, + 539, + 1352, + 539, + 1385, + 345, + 1385 + ], + "score": 0.91, + "latex": "\\sigma ( \\cdot ) = \\operatorname* { m a x } \\{ 0 , \\cdot \\}" + }, + { + "category_id": 13, + "poly": [ + 825, + 940, + 987, + 940, + 987, + 973, + 825, + 973 + ], + "score": 0.91, + "latex": "\\mathbf { A } , \\mathbf { B } \\in \\mathbb { R } ^ { m \\times n }" + }, + { + "category_id": 14, + "poly": [ + 677, + 1264, + 1019, + 1264, + 1019, + 1305, + 677, + 1305 + ], + "score": 0.9, + "latex": "f _ { \\mathbf { W } } ( \\mathbf { x } ) = \\sqrt { m } \\cdot \\mathbf { W } _ { 2 } \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } ) ," + }, + { + "category_id": 13, + "poly": [ + 638, + 1078, + 692, + 1078, + 692, + 1110, + 638, + 1110 + ], + "score": 0.9, + "latex": "\\mathcal { O } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1207, + 800, + 1352, + 800, + 1352, + 831, + 1207, + 831 + ], + "score": 0.9, + "latex": "1 \\leqslant p < \\infty" + }, + { + "category_id": 13, + "poly": [ + 721, + 868, + 782, + 868, + 782, + 899, + 721, + 899 + ], + "score": 0.9, + "latex": "\\lVert \\mathbf { A } \\rVert _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 758, + 1382, + 1022, + 1382, + 1022, + 1417, + 758, + 1417 + ], + "score": 0.9, + "latex": "S = \\left\\{ \\left( \\mathbf { x } _ { i } , y _ { i } \\right) : i \\in \\left[ n \\right] \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 580, + 1319, + 741, + 1319, + 741, + 1352, + 580, + 1352 + ], + "score": 0.9, + "latex": "\\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { 1 \\times m }" + }, + { + "category_id": 13, + "poly": [ + 346, + 942, + 617, + 942, + 617, + 976, + 346, + 976 + ], + "score": 0.9, + "latex": "\\| \\mathbf { A } \\| _ { \\operatorname* { m a x } } = \\operatorname* { m a x } _ { i , j } | A _ { i , j } |" + }, + { + "category_id": 13, + "poly": [ + 1189, + 834, + 1393, + 834, + 1393, + 869, + 1189, + 869 + ], + "score": 0.9, + "latex": "\\| \\mathbf { v } \\| _ { \\infty } = \\operatorname* { m a x } _ { i } | v _ { i } |" + }, + { + "category_id": 13, + "poly": [ + 1340, + 973, + 1392, + 973, + 1392, + 1006, + 1340, + 1006 + ], + "score": 0.9, + "latex": "\\mathcal { O } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1318, + 567, + 1318, + 567, + 1351, + 373, + 1351 + ], + "score": 0.89, + "latex": "\\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { m \\times ( d + 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 827, + 1044, + 1106, + 1044, + 1106, + 1079, + 827, + 1079 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\operatorname* { l i m } \\operatorname* { i n f } _ { n \\to \\infty } \\left| a _ { n } / b _ { n } \\right| > 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1324, + 903, + 1393, + 903, + 1393, + 936, + 1324, + 936 + ], + "score": 0.88, + "latex": "p \\geqslant 1" + }, + { + "category_id": 13, + "poly": [ + 884, + 1010, + 914, + 1010, + 914, + 1040, + 884, + 1040 + ], + "score": 0.88, + "latex": "b _ { n }" + }, + { + "category_id": 13, + "poly": [ + 441, + 1849, + 866, + 1849, + 866, + 1885, + 441, + 1885 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\dot { \\mathbf { W } } ^ { ( t + 1 ) } = \\mathbf { W } ^ { ( t ) } - \\eta \\cdot \\nabla _ { \\mathbf { W } } L _ { S } ( \\mathbf { W } ^ { ( t ) } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 343, + 1045, + 595, + 1045, + 595, + 1079, + 343, + 1079 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\operatorname* { s u p } _ { n \\infty } | a _ { n } / b _ { n } | < \\infty } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 651, + 1979, + 710, + 1979, + 710, + 2003, + 651, + 2003 + ], + "score": 0.87, + "latex": "d + 1" + }, + { + "category_id": 13, + "poly": [ + 801, + 1014, + 833, + 1014, + 833, + 1040, + 801, + 1040 + ], + "score": 0.86, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1822, + 586, + 1822, + 586, + 1852, + 366, + 1852 + ], + "score": 0.85, + "latex": "t = 0 , 1 , \\dots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 342, + 975, + 427, + 975, + 427, + 1002, + 342, + 1002 + ], + "score": 0.84, + "latex": "\\mathbf { A } \\succeq \\mathbf { B }" + }, + { + "category_id": 13, + "poly": [ + 455, + 976, + 534, + 976, + 534, + 1001, + 455, + 1001 + ], + "score": 0.83, + "latex": "\\mathbf { A } - \\mathbf { B }" + }, + { + "category_id": 13, + "poly": [ + 429, + 1909, + 499, + 1909, + 499, + 1942, + 429, + 1942 + ], + "score": 0.82, + "latex": "\\mathbf { W } ^ { ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1525, + 391, + 1525, + 391, + 1551, + 373, + 1551 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 830, + 1565, + 850, + 1565, + 850, + 1586, + 830, + 1586 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 530, + 1389, + 550, + 1389, + 550, + 1410, + 530, + 1410 + ], + "score": 0.77, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 646, + 1754, + 669, + 1754, + 669, + 1780, + 646, + 1780 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 782, + 1760, + 800, + 1760, + 800, + 1784, + 782, + 1784 + ], + "score": 0.76, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 298, + 2006, + 314, + 2006, + 314, + 2030, + 298, + 2030 + ], + "score": 0.75, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 298, + 1003, + 416, + 1003, + 416, + 1044, + 298, + 1044 + ], + "score": 0.69, + "latex": "\\widetilde { \\mathcal { O } } ( \\cdot ) , \\Omega ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1276, + 870, + 1302, + 870, + 1302, + 893, + 1276, + 893 + ], + "score": 0.42, + "latex": "\\mathbf { A }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1003, + 352, + 1003, + 352, + 1043, + 298, + 1043 + ], + "score": 0.37, + "latex": "\\tilde { \\mathcal { O } } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 362, + 1005, + 415, + 1005, + 415, + 1043, + 362, + 1043 + ], + "score": 0.26, + "latex": "\\Omega ( \\cdot )" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 572.0, + 1130.0, + 572.0, + 1170.0, + 294.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 520.0, + 582.0, + 520.0, + 582.0, + 564.0, + 292.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 704.0, + 495.0, + 704.0, + 495.0, + 744.0, + 293.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1974.0, + 650.0, + 1974.0, + 650.0, + 2009.0, + 332.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2009.0, + 711.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 297.0, + 2002.0, + 297.0, + 2034.0, + 292.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 2002.0, + 562.0, + 2002.0, + 562.0, + 2034.0, + 315.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 2002.0, + 683.0, + 2002.0, + 683.0, + 2034.0, + 675.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1707.0, + 984.0, + 1707.0, + 984.0, + 1747.0, + 297.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 837.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1974.0, + 650.0, + 1974.0, + 650.0, + 2009.0, + 332.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2009.0, + 711.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 297.0, + 2002.0, + 297.0, + 2034.0, + 292.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 2002.0, + 562.0, + 2002.0, + 562.0, + 2034.0, + 315.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 2002.0, + 683.0, + 2002.0, + 683.0, + 2034.0, + 675.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 764.0, + 1406.0, + 764.0, + 1406.0, + 803.0, + 295.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 792.0, + 747.0, + 792.0, + 747.0, + 835.0, + 292.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 792.0, + 1206.0, + 792.0, + 1206.0, + 835.0, + 1038.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 792.0, + 1407.0, + 792.0, + 1407.0, + 835.0, + 1353.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 814.0, + 549.0, + 814.0, + 549.0, + 886.0, + 284.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 814.0, + 1188.0, + 814.0, + 1188.0, + 886.0, + 815.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 814.0, + 1414.0, + 814.0, + 1414.0, + 886.0, + 1394.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 860.0, + 441.0, + 860.0, + 441.0, + 905.0, + 291.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 860.0, + 720.0, + 860.0, + 720.0, + 905.0, + 627.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 860.0, + 1275.0, + 860.0, + 1275.0, + 905.0, + 783.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 860.0, + 1407.0, + 860.0, + 1407.0, + 905.0, + 1303.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 881.0, + 297.0, + 881.0, + 297.0, + 956.0, + 284.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 881.0, + 972.0, + 881.0, + 972.0, + 956.0, + 588.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 881.0, + 1323.0, + 881.0, + 1323.0, + 956.0, + 1281.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 881.0, + 1417.0, + 881.0, + 1417.0, + 956.0, + 1394.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 936.0, + 345.0, + 936.0, + 345.0, + 977.0, + 292.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 936.0, + 824.0, + 936.0, + 824.0, + 977.0, + 618.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 936.0, + 1112.0, + 936.0, + 1112.0, + 977.0, + 988.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 936.0, + 1408.0, + 936.0, + 1408.0, + 977.0, + 1350.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 968.0, + 341.0, + 968.0, + 341.0, + 1010.0, + 291.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 968.0, + 454.0, + 968.0, + 454.0, + 1010.0, + 428.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 968.0, + 1339.0, + 968.0, + 1339.0, + 1010.0, + 535.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 968.0, + 1406.0, + 968.0, + 1406.0, + 1010.0, + 1393.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1001.0, + 297.0, + 1001.0, + 297.0, + 1049.0, + 292.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1001.0, + 466.0, + 1001.0, + 466.0, + 1049.0, + 417.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1001.0, + 800.0, + 1001.0, + 800.0, + 1049.0, + 519.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1001.0, + 883.0, + 1001.0, + 883.0, + 1049.0, + 834.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1001.0, + 1234.0, + 1001.0, + 1234.0, + 1049.0, + 915.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1001.0, + 1409.0, + 1001.0, + 1409.0, + 1049.0, + 1376.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1044.0, + 342.0, + 1044.0, + 342.0, + 1080.0, + 292.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1044.0, + 655.0, + 1044.0, + 655.0, + 1080.0, + 596.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1044.0, + 826.0, + 1044.0, + 826.0, + 1080.0, + 797.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1044.0, + 1210.0, + 1044.0, + 1210.0, + 1080.0, + 1107.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1044.0, + 1318.0, + 1044.0, + 1318.0, + 1080.0, + 1266.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1044.0, + 1406.0, + 1044.0, + 1406.0, + 1080.0, + 1372.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1074.0, + 637.0, + 1074.0, + 637.0, + 1113.0, + 292.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1074.0, + 743.0, + 1074.0, + 743.0, + 1113.0, + 693.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1074.0, + 806.0, + 1074.0, + 806.0, + 1113.0, + 796.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1516.0, + 372.0, + 1516.0, + 372.0, + 1567.0, + 291.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1516.0, + 1282.0, + 1516.0, + 1282.0, + 1567.0, + 392.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 1516.0, + 1411.0, + 1516.0, + 1411.0, + 1567.0, + 1374.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1553.0, + 829.0, + 1553.0, + 829.0, + 1592.0, + 292.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1553.0, + 1069.0, + 1553.0, + 1069.0, + 1592.0, + 851.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1553.0, + 1409.0, + 1553.0, + 1409.0, + 1592.0, + 1203.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1588.0, + 647.0, + 1588.0, + 647.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1588.0, + 750.0, + 1588.0, + 750.0, + 1623.0, + 740.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 293.0, + 295.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 288.0, + 1407.0, + 288.0, + 1407.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 317.0, + 1405.0, + 317.0, + 1405.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1088.0, + 351.0, + 1088.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1311.0, + 372.0, + 1311.0, + 372.0, + 1361.0, + 291.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1311.0, + 579.0, + 1311.0, + 579.0, + 1361.0, + 568.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1311.0, + 1406.0, + 1311.0, + 1406.0, + 1361.0, + 742.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1348.0, + 344.0, + 1348.0, + 344.0, + 1389.0, + 292.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1348.0, + 1406.0, + 1348.0, + 1406.0, + 1389.0, + 540.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1381.0, + 529.0, + 1381.0, + 529.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1381.0, + 757.0, + 1381.0, + 757.0, + 1419.0, + 551.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1381.0, + 1033.0, + 1381.0, + 1033.0, + 1419.0, + 1023.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 399.0, + 1402.0, + 399.0, + 1402.0, + 433.0, + 297.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 428.0, + 1404.0, + 428.0, + 1404.0, + 464.0, + 294.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 461.0, + 1251.0, + 461.0, + 1251.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 589.0, + 1404.0, + 589.0, + 1404.0, + 623.0, + 294.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 622.0, + 1404.0, + 622.0, + 1404.0, + 652.0, + 297.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 647.0, + 951.0, + 647.0, + 951.0, + 685.0, + 293.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 1401.0, + 1632.0, + 1401.0, + 1670.0, + 295.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1666.0, + 1322.0, + 1666.0, + 1322.0, + 1700.0, + 294.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1192.0, + 1404.0, + 1192.0, + 1404.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1220.0, + 432.0, + 1220.0, + 432.0, + 1259.0, + 293.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1745.0, + 645.0, + 1745.0, + 645.0, + 1789.0, + 321.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1745.0, + 781.0, + 1745.0, + 781.0, + 1789.0, + 670.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1745.0, + 813.0, + 1745.0, + 813.0, + 1789.0, + 801.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1771.0, + 365.0, + 1771.0, + 365.0, + 1854.0, + 317.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1771.0, + 700.0, + 1771.0, + 700.0, + 1854.0, + 650.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1771.0, + 830.0, + 1771.0, + 830.0, + 1854.0, + 767.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1771.0, + 1007.0, + 1771.0, + 1007.0, + 1854.0, + 957.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1771.0, + 1282.0, + 1771.0, + 1282.0, + 1854.0, + 1133.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1841.0, + 440.0, + 1841.0, + 440.0, + 1887.0, + 344.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1841.0, + 875.0, + 1841.0, + 875.0, + 1887.0, + 867.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1879.0, + 422.0, + 1879.0, + 422.0, + 1914.0, + 320.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1908.0, + 428.0, + 1908.0, + 428.0, + 1944.0, + 318.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1908.0, + 510.0, + 1908.0, + 510.0, + 1944.0, + 500.0, + 1944.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1640, + 1404, + 1640, + 1404, + 1826, + 297, + 1826 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 499, + 1404, + 499, + 1404, + 685, + 297, + 685 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 227, + 1403, + 227, + 1403, + 415, + 298, + 415 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1165, + 1403, + 1165, + 1403, + 1322, + 297, + 1322 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1420, + 1403, + 1420, + 1403, + 1544, + 297, + 1544 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1910, + 1401, + 1910, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 862, + 1406, + 862, + 1406, + 956, + 299, + 956 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 594, + 1337, + 1103, + 1337, + 1103, + 1409, + 594, + 1409 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 593, + 699, + 1103, + 699, + 1103, + 739, + 593, + 739 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 397, + 970, + 1298, + 970, + 1298, + 1036, + 397, + 1036 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 300, + 1574, + 573, + 1574, + 573, + 1610, + 300, + 1610 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 298, + 752, + 370, + 752, + 370, + 782, + 298, + 782 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 300, + 1107, + 625, + 1107, + 625, + 1139, + 300, + 1139 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 297, + 1850, + 973, + 1850, + 973, + 1884, + 297, + 1884 + ], + "score": 0.918 + }, + { + "category_id": 9, + "poly": [ + 1345, + 1356, + 1400, + 1356, + 1400, + 1388, + 1345, + 1388 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 301, + 440, + 696, + 440, + 696, + 472, + 301, + 472 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1345, + 799, + 1400, + 799, + 1400, + 830, + 1345, + 830 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 299, + 1047, + 635, + 1047, + 635, + 1084, + 299, + 1084 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1345, + 702, + 1400, + 702, + 1400, + 733, + 1345, + 733 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1344, + 986, + 1400, + 986, + 1400, + 1018, + 1344, + 1018 + ], + "score": 0.874 + }, + { + "category_id": 8, + "poly": [ + 576, + 773, + 1119, + 773, + 1119, + 857, + 576, + 857 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.752 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.471 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.46 + }, + { + "category_id": 8, + "poly": [ + 579, + 820, + 1108, + 820, + 1108, + 857, + 579, + 857 + ], + "score": 0.229 + }, + { + "category_id": 8, + "poly": [ + 582, + 773, + 1118, + 773, + 1118, + 814, + 582, + 814 + ], + "score": 0.115 + }, + { + "category_id": 14, + "poly": [ + 593, + 1334, + 1105, + 1334, + 1105, + 1408, + 593, + 1408 + ], + "score": 0.95, + "latex": "L _ { \\kappa } ( f ) ( { \\bf x } ) = \\int _ { X } \\kappa ( { \\bf x } , { \\bf y } ) f ( { \\bf y } ) d \\tau ( { \\bf y } ) , ~ { \\bf x } \\in X ." + }, + { + "category_id": 13, + "poly": [ + 472, + 1291, + 552, + 1291, + 552, + 1323, + 472, + 1323 + ], + "score": 0.92, + "latex": "L _ { \\tau } ^ { 2 } ( X )" + }, + { + "category_id": 14, + "poly": [ + 591, + 697, + 1105, + 697, + 1105, + 738, + 591, + 738 + ], + "score": 0.92, + "latex": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle \\cdot \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) + 2 \\cdot \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) ," + }, + { + "category_id": 13, + "poly": [ + 858, + 1259, + 1048, + 1259, + 1048, + 1287, + 858, + 1287 + ], + "score": 0.91, + "latex": "\\kappa : X \\times X \\to \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 344, + 1229, + 425, + 1229, + 425, + 1262, + 344, + 1262 + ], + "score": 0.91, + "latex": "L _ { \\tau } ^ { 2 } ( X )" + }, + { + "category_id": 14, + "poly": [ + 402, + 968, + 1295, + 968, + 1295, + 1035, + 402, + 1035 + ], + "score": 0.91, + "latex": "\\kappa _ { 1 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) , \\kappa _ { 2 } ( t ) = \\frac { 1 } { 2 \\pi } \\left( t \\cdot \\left( \\pi - \\operatorname { a r c c o s } \\left( t \\right) \\right) + \\sqrt { 1 - t ^ { 2 } } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 372, + 1047, + 627, + 1047, + 627, + 1084, + 372, + 1084 + ], + "score": 0.9, + "latex": "t = \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle / \\left( \\left\\| \\mathbf { x } \\right\\| \\left\\| \\mathbf { x } ^ { \\prime } \\right\\| \\right)" + }, + { + "category_id": 13, + "poly": [ + 397, + 383, + 451, + 383, + 451, + 415, + 397, + 415 + ], + "score": 0.9, + "latex": "1 / m" + }, + { + "category_id": 13, + "poly": [ + 570, + 382, + 622, + 382, + 622, + 416, + 570, + 416 + ], + "score": 0.9, + "latex": "2 / m" + }, + { + "category_id": 13, + "poly": [ + 359, + 1262, + 449, + 1262, + 449, + 1286, + 359, + 1286 + ], + "score": 0.89, + "latex": "X \\to \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 708, + 621, + 1355, + 621, + 1355, + 657, + 708, + 657 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\operatorname* { l i m } _ { m \\infty } m ^ { - 1 } \\langle \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ) , \\nabla \\mathbf { w } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } ^ { \\prime } ) \\rangle } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1091, + 1704, + 1128, + 1704, + 1128, + 1733, + 1091, + 1733 + ], + "score": 0.89, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 397, + 1292, + 432, + 1292, + 432, + 1320, + 397, + 1320 + ], + "score": 0.89, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 876, + 1974, + 911, + 1974, + 911, + 2003, + 876, + 2003 + ], + "score": 0.88, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 779, + 1731, + 811, + 1731, + 811, + 1761, + 779, + 1761 + ], + "score": 0.88, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 631, + 227, + 697, + 227, + 697, + 260, + 631, + 260 + ], + "score": 0.83, + "latex": "\\mathbf { W } ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 627, + 1517, + 646, + 1517, + 646, + 1540, + 627, + 1540 + ], + "score": 0.78, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 1098, + 1264, + 1117, + 1264, + 1117, + 1286, + 1098, + 1286 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 581, + 1457, + 601, + 1457, + 601, + 1479, + 581, + 1479 + ], + "score": 0.77, + "latex": "\\kappa" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1235, + 1401, + 1235, + 1401, + 1256, + 1382, + 1256 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 14, + "poly": [ + 575, + 768, + 1117, + 768, + 1117, + 860, + 575, + 860 + ], + "score": 0.61, + "latex": "\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] , } \\\\ & { \\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 580, + 772, + 1121, + 772, + 1121, + 814, + 580, + 814 + ], + "score": 0.39, + "latex": "\\begin{array} { r } { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 577, + 818, + 1111, + 818, + 1111, + 858, + 577, + 858 + ], + "score": 0.32, + "latex": "\\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } [ \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) ] ." + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1571.0, + 576.0, + 1571.0, + 576.0, + 1618.0, + 291.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1107.0, + 630.0, + 1107.0, + 630.0, + 1143.0, + 295.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 977.0, + 1850.0, + 977.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 440.0, + 702.0, + 440.0, + 702.0, + 476.0, + 295.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 1404.0, + 1642.0, + 1404.0, + 1674.0, + 295.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1672.0, + 1405.0, + 1672.0, + 1405.0, + 1705.0, + 291.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 1090.0, + 1702.0, + 1090.0, + 1738.0, + 294.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1702.0, + 1406.0, + 1702.0, + 1406.0, + 1738.0, + 1129.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1729.0, + 778.0, + 1729.0, + 778.0, + 1767.0, + 292.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1729.0, + 1404.0, + 1729.0, + 1404.0, + 1767.0, + 812.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1801.0, + 292.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1794.0, + 507.0, + 1794.0, + 507.0, + 1827.0, + 295.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 500.0, + 1404.0, + 500.0, + 1404.0, + 532.0, + 295.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 529.0, + 1406.0, + 529.0, + 1406.0, + 564.0, + 294.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 560.0, + 1406.0, + 560.0, + 1406.0, + 596.0, + 294.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 591.0, + 1402.0, + 591.0, + 1402.0, + 626.0, + 294.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 618.0, + 707.0, + 618.0, + 707.0, + 661.0, + 291.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 618.0, + 1407.0, + 618.0, + 1407.0, + 661.0, + 1356.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 649.0, + 398.0, + 649.0, + 398.0, + 687.0, + 293.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 630.0, + 225.0, + 630.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 225.0, + 1404.0, + 225.0, + 1404.0, + 267.0, + 698.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 294.0, + 296.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1404.0, + 352.0, + 1404.0, + 384.0, + 294.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 381.0, + 396.0, + 381.0, + 396.0, + 417.0, + 296.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 381.0, + 569.0, + 381.0, + 569.0, + 417.0, + 452.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 381.0, + 633.0, + 381.0, + 633.0, + 417.0, + 623.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1167.0, + 1404.0, + 1167.0, + 1404.0, + 1201.0, + 296.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1233.0, + 292.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1226.0, + 343.0, + 1226.0, + 343.0, + 1264.0, + 293.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1226.0, + 1381.0, + 1226.0, + 1381.0, + 1264.0, + 426.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1226.0, + 1407.0, + 1226.0, + 1407.0, + 1264.0, + 1402.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1254.0, + 358.0, + 1254.0, + 358.0, + 1294.0, + 292.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1254.0, + 857.0, + 1254.0, + 857.0, + 1294.0, + 450.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1254.0, + 1097.0, + 1254.0, + 1097.0, + 1294.0, + 1049.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1254.0, + 1405.0, + 1254.0, + 1405.0, + 1294.0, + 1118.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1288.0, + 396.0, + 1288.0, + 396.0, + 1326.0, + 295.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1288.0, + 471.0, + 1288.0, + 471.0, + 1326.0, + 433.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1288.0, + 591.0, + 1288.0, + 591.0, + 1326.0, + 553.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1419.0, + 1403.0, + 1419.0, + 1403.0, + 1454.0, + 292.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1449.0, + 580.0, + 1449.0, + 580.0, + 1486.0, + 294.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1449.0, + 1407.0, + 1449.0, + 1407.0, + 1486.0, + 602.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1482.0, + 1406.0, + 1482.0, + 1406.0, + 1517.0, + 292.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1512.0, + 626.0, + 1512.0, + 626.0, + 1544.0, + 293.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1512.0, + 864.0, + 1512.0, + 864.0, + 1544.0, + 647.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 875.0, + 1970.0, + 875.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2008.0, + 912.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 460.0, + 2000.0, + 460.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 861.0, + 1407.0, + 861.0, + 1407.0, + 897.0, + 293.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 894.0, + 1404.0, + 894.0, + 1404.0, + 928.0, + 296.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 922.0, + 1387.0, + 922.0, + 1387.0, + 960.0, + 294.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 747.0, + 375.0, + 747.0, + 375.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1046.0, + 371.0, + 1046.0, + 371.0, + 1087.0, + 297.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1046.0, + 636.0, + 1046.0, + 636.0, + 1087.0, + 628.0, + 1087.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 973, + 1406, + 973, + 1406, + 1341, + 296, + 1341 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 1405, + 227, + 1405, + 483, + 296, + 483 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1425, + 1404, + 1425, + 1404, + 1610, + 297, + 1610 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 296, + 658, + 1404, + 658, + 1404, + 797, + 296, + 797 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 296, + 804, + 1405, + 804, + 1405, + 916, + 296, + 916 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 598, + 1690, + 1100, + 1690, + 1100, + 1783, + 598, + 1783 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 289, + 1793, + 1401, + 1793, + 1401, + 1866, + 289, + 1866 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 610, + 559, + 1093, + 559, + 1093, + 603, + 610, + 603 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 296, + 1877, + 1430, + 1877, + 1430, + 2037, + 296, + 2037 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 294, + 487, + 1401, + 487, + 1401, + 552, + 294, + 552 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 297, + 1615, + 1402, + 1615, + 1402, + 1680, + 297, + 1680 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 608, + 661, + 608, + 661, + 640, + 298, + 640 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 477, + 924, + 1217, + 924, + 1217, + 969, + 477, + 969 + ], + "score": 0.928 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.915 + }, + { + "category_id": 9, + "poly": [ + 1344, + 1720, + 1400, + 1720, + 1400, + 1751, + 1344, + 1751 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 298, + 1367, + 1351, + 1367, + 1351, + 1400, + 298, + 1400 + ], + "score": 0.773 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 859, + 2087, + 859, + 2112, + 840, + 2112 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 298, + 1367, + 1351, + 1367, + 1351, + 1400, + 298, + 1400 + ], + "score": 0.164 + }, + { + "category_id": 14, + "poly": [ + 598, + 1687, + 1099, + 1687, + 1099, + 1785, + 598, + 1785 + ], + "score": 0.94, + "latex": "\\kappa \\left( \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\right) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N \\left( d , k \\right) } Y _ { k , j } \\left( \\mathbf { x } \\right) Y _ { k , j } \\left( \\mathbf { x } ^ { \\prime } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 795, + 1969, + 993, + 1969, + 993, + 2009, + 795, + 2009 + ], + "score": 0.93, + "latex": "\\mu _ { k } = \\Omega \\left( k ^ { - d - 1 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 955, + 759, + 1156, + 759, + 1156, + 798, + 955, + 798 + ], + "score": 0.93, + "latex": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 343, + 230, + 432, + 230, + 432, + 264, + 343, + 264 + ], + "score": 0.93, + "latex": "\\{ \\lambda _ { i } \\} _ { i \\geqslant 1 }" + }, + { + "category_id": 13, + "poly": [ + 388, + 661, + 616, + 661, + 616, + 697, + 388, + 697 + ], + "score": 0.93, + "latex": "\\mathbf { y } ~ = ~ ( y _ { 1 } , \\ldots , y _ { n } ) ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 599, + 489, + 748, + 489, + 748, + 523, + 599, + 523 + ], + "score": 0.93, + "latex": "| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M" + }, + { + "category_id": 13, + "poly": [ + 374, + 1647, + 569, + 1647, + 569, + 1678, + 374, + 1678 + ], + "score": 0.92, + "latex": "\\kappa : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\xrightarrow { } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1201, + 1969, + 1400, + 1969, + 1400, + 2009, + 1201, + 2009 + ], + "score": 0.92, + "latex": "\\mu _ { k } = \\Omega \\left( d ^ { - k + 1 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 823, + 697, + 1023, + 697, + 1023, + 733, + 823, + 733 + ], + "score": 0.92, + "latex": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } ) \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 952, + 489, + 1039, + 489, + 1039, + 523, + 952, + 523 + ], + "score": 0.92, + "latex": "i \\in [ r _ { k } ]" + }, + { + "category_id": 13, + "poly": [ + 1116, + 807, + 1212, + 807, + 1212, + 838, + 1116, + 838 + ], + "score": 0.92, + "latex": "\\epsilon , \\delta > 0" + }, + { + "category_id": 13, + "poly": [ + 773, + 805, + 869, + 805, + 869, + 840, + 773, + 840 + ], + "score": 0.92, + "latex": "j \\in [ r _ { k } ]" + }, + { + "category_id": 13, + "poly": [ + 570, + 805, + 727, + 805, + 727, + 840, + 570, + 840 + ], + "score": 0.92, + "latex": "| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant M" + }, + { + "category_id": 13, + "poly": [ + 374, + 1972, + 464, + 1972, + 464, + 2003, + 374, + 2003 + ], + "score": 0.91, + "latex": "j \\in \\mathbb { N } ^ { + }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1796, + 420, + 1796, + 420, + 1829, + 371, + 1829 + ], + "score": 0.91, + "latex": "Y _ { k , j }" + }, + { + "category_id": 13, + "poly": [ + 841, + 878, + 949, + 878, + 949, + 917, + 841, + 917 + ], + "score": 0.91, + "latex": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 606, + 390, + 691, + 390, + 691, + 423, + 606, + 423 + ], + "score": 0.91, + "latex": "L _ { \\tau } ^ { 2 } ( \\mathbb { S } ^ { d } )" + }, + { + "category_id": 13, + "poly": [ + 462, + 1794, + 683, + 1794, + 683, + 1827, + 462, + 1827 + ], + "score": 0.91, + "latex": "j = 1 , \\cdots , N ( d , k )" + }, + { + "category_id": 13, + "poly": [ + 629, + 878, + 826, + 878, + 826, + 917, + 629, + 917 + ], + "score": 0.91, + "latex": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 461, + 1827, + 768, + 1827, + 768, + 1869, + 461, + 1869 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 817, + 839, + 1123, + 839, + 1123, + 879, + 817, + 879 + ], + "score": 0.91, + "latex": "m \\geqslant \\widetilde { \\Omega } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )" + }, + { + "category_id": 13, + "poly": [ + 493, + 231, + 666, + 231, + 666, + 262, + 493, + 262 + ], + "score": 0.91, + "latex": "\\lambda _ { 1 } \\geqslant \\lambda _ { 2 } \\geqslant \\cdots" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1972, + 1146, + 1972, + 1146, + 2002, + 1069, + 2002 + ], + "score": 0.9, + "latex": "k \\gg d" + }, + { + "category_id": 13, + "poly": [ + 297, + 839, + 803, + 839, + 803, + 879, + 297, + 879 + ], + "score": 0.9, + "latex": "n \\geqslant \\widetilde { \\Omega } ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } , M ^ { 4 } r _ { k } ^ { 2 } \\} )" + }, + { + "category_id": 13, + "poly": [ + 922, + 804, + 1005, + 804, + 1005, + 835, + 922, + 835 + ], + "score": 0.9, + "latex": "\\mathbf { x } \\in \\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 824, + 487, + 902, + 487, + 902, + 518, + 824, + 518 + ], + "score": 0.9, + "latex": "\\mathbf { x } \\in \\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 364, + 1580, + 434, + 1580, + 434, + 1609, + 364, + 1609 + ], + "score": 0.9, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 1191, + 229, + 1366, + 229, + 1366, + 264, + 1191, + 264 + ], + "score": 0.9, + "latex": "\\phi _ { 1 } ( \\cdot ) , \\phi _ { 2 } ( \\cdot ) , \\ldots" + }, + { + "category_id": 13, + "poly": [ + 862, + 261, + 1249, + 261, + 1249, + 298, + 862, + 298 + ], + "score": 0.89, + "latex": "\\mathbf v _ { i } = n ^ { - 1 / 2 } ( \\phi _ { i } ( { \\mathbf { x } } _ { 1 } ) , \\dots , \\phi _ { i } ( { \\mathbf { x } } _ { n } ) ) ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 557, + 1615, + 768, + 1615, + 768, + 1649, + 557, + 1649 + ], + "score": 0.89, + "latex": "\\mathbf { x } , \\mathbf { x } ^ { \\prime } \\in \\mathbb { S } ^ { d } \\subset \\mathbb { R } ^ { d + 1 }" + }, + { + "category_id": 13, + "poly": [ + 485, + 1580, + 554, + 1580, + 554, + 1608, + 485, + 1608 + ], + "score": 0.89, + "latex": "k \\gg d" + }, + { + "category_id": 13, + "poly": [ + 593, + 357, + 831, + 357, + 831, + 390, + 593, + 390 + ], + "score": 0.89, + "latex": "\\mathbf { V } _ { r _ { k } } = \\left( \\mathbf { v } _ { 1 } , \\ldots , \\mathbf { v } _ { r _ { k } } \\right)" + }, + { + "category_id": 14, + "poly": [ + 605, + 557, + 1091, + 557, + 1091, + 603, + 605, + 603 + ], + "score": 0.89, + "latex": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { \\operatorname* { m a x } } \\leqslant C M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n } ," + }, + { + "category_id": 13, + "poly": [ + 1143, + 491, + 1210, + 491, + 1210, + 518, + 1143, + 518 + ], + "score": 0.89, + "latex": "\\delta > 0" + }, + { + "category_id": 13, + "poly": [ + 1095, + 232, + 1130, + 232, + 1130, + 261, + 1095, + 261 + ], + "score": 0.89, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 464, + 358, + 498, + 358, + 498, + 387, + 464, + 387 + ], + "score": 0.89, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 1343, + 1795, + 1401, + 1795, + 1401, + 1823, + 1343, + 1823 + ], + "score": 0.89, + "latex": "d + 1" + }, + { + "category_id": 13, + "poly": [ + 409, + 297, + 445, + 297, + 445, + 326, + 409, + 326 + ], + "score": 0.88, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 926, + 1834, + 959, + 1834, + 959, + 1863, + 926, + 1863 + ], + "score": 0.87, + "latex": "\\mu _ { k }" + }, + { + "category_id": 13, + "poly": [ + 678, + 660, + 1136, + 660, + 1136, + 697, + 678, + 697 + ], + "score": 0.87, + "latex": "\\widehat { \\mathbf { y } } ^ { ( t ) } ~ = ~ \\theta ~ \\cdot ~ \\left( f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { n } ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 382, + 520, + 445, + 520, + 445, + 549, + 382, + 549 + ], + "score": 0.87, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 881, + 422, + 919, + 422, + 919, + 449, + 881, + 449 + ], + "score": 0.86, + "latex": "\\mathbb { R } ^ { n }" + }, + { + "category_id": 14, + "poly": [ + 481, + 924, + 1217, + 924, + 1217, + 969, + 481, + 969 + ], + "score": 0.86, + "latex": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 1152, + 664, + 1314, + 664, + 1314, + 696, + 1152, + 696 + ], + "score": 0.86, + "latex": "t ~ = ~ 0 , \\ldots , T" + }, + { + "category_id": 13, + "poly": [ + 1259, + 265, + 1398, + 265, + 1398, + 296, + 1259, + 296 + ], + "score": 0.86, + "latex": "i = 1 , 2 , \\dots" + }, + { + "category_id": 13, + "poly": [ + 1229, + 733, + 1257, + 733, + 1257, + 759, + 1229, + 759 + ], + "score": 0.85, + "latex": "r _ { k }" + }, + { + "category_id": 13, + "poly": [ + 372, + 610, + 397, + 610, + 397, + 636, + 372, + 636 + ], + "score": 0.84, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 751, + 331, + 781, + 331, + 781, + 357, + 751, + 357 + ], + "score": 0.84, + "latex": "r _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1294, + 1796, + 1312, + 1796, + 1312, + 1822, + 1294, + 1822 + ], + "score": 0.84, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 355, + 884, + 419, + 884, + 419, + 913, + 355, + 913 + ], + "score": 0.82, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 1111, + 298, + 1139, + 298, + 1139, + 326, + 1111, + 326 + ], + "score": 0.82, + "latex": "\\lambda _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1349, + 808, + 1368, + 808, + 1368, + 834, + 1349, + 834 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 446, + 2003, + 610, + 2003, + 610, + 2034, + 446, + 2034 + ], + "score": 0.81, + "latex": "k = 2 , 4 , 6 , \\ldots" + }, + { + "category_id": 13, + "poly": [ + 1289, + 326, + 1310, + 326, + 1310, + 354, + 1289, + 354 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1150, + 298, + 1219, + 298, + 1219, + 325, + 1150, + 325 + ], + "score": 0.79, + "latex": "i \\geqslant 1" + }, + { + "category_id": 13, + "poly": [ + 1044, + 356, + 1132, + 356, + 1132, + 390, + 1044, + 390 + ], + "score": 0.78, + "latex": "i \\in [ r _ { k } ]" + }, + { + "category_id": 14, + "poly": [ + 310, + 1899, + 1425, + 1899, + 1425, + 1960, + 310, + 1960 + ], + "score": 0.78, + "latex": "\\begin{array} { r l } & { \\mathbf { \\Phi } ^ { \\circ _ { 0 } } \\quad \\mathsf { \\Lambda } ^ { \\circ _ { 1 } } \\quad \\cdots \\mathbf { \\Phi } ^ { \\circ _ { k } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } , \\mathsf { w } } \\quad \\cdots \\quad \\mathbf { \\Phi } ^ { \\circ _ { j } \\quad \\bullet _ { k } , \\mathsf { s } } , } \\\\ & { \\boldsymbol { \\iota } _ { k } = \\Omega \\left( \\operatorname* { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) , k = 2 j , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 597, + 327, + 616, + 327, + 616, + 354, + 597, + 354 + ], + "score": 0.78, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 364, + 2004, + 434, + 2004, + 434, + 2032, + 364, + 2032 + ], + "score": 0.77, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 1004, + 360, + 1033, + 360, + 1033, + 388, + 1004, + 388 + ], + "score": 0.56, + "latex": "\\mathbf { v } _ { i }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1365.0, + 1352.0, + 1365.0, + 1352.0, + 1400.0, + 296.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 971.0, + 1406.0, + 971.0, + 1406.0, + 1011.0, + 293.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1005.0, + 1406.0, + 1005.0, + 1406.0, + 1040.0, + 293.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1034.0, + 1404.0, + 1034.0, + 1404.0, + 1070.0, + 293.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1066.0, + 1406.0, + 1066.0, + 1406.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1096.0, + 1404.0, + 1096.0, + 1404.0, + 1130.0, + 294.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1129.0, + 1403.0, + 1129.0, + 1403.0, + 1159.0, + 295.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1189.0, + 1404.0, + 1189.0, + 1404.0, + 1219.0, + 295.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1217.0, + 1406.0, + 1217.0, + 1406.0, + 1253.0, + 293.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1283.0, + 293.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1279.0, + 1406.0, + 1279.0, + 1406.0, + 1315.0, + 293.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1311.0, + 648.0, + 1311.0, + 648.0, + 1341.0, + 295.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 342.0, + 226.0, + 342.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 226.0, + 492.0, + 226.0, + 492.0, + 268.0, + 433.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 226.0, + 1094.0, + 226.0, + 1094.0, + 268.0, + 667.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 226.0, + 1190.0, + 226.0, + 1190.0, + 268.0, + 1131.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 226.0, + 1406.0, + 226.0, + 1406.0, + 268.0, + 1367.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 257.0, + 861.0, + 257.0, + 861.0, + 302.0, + 291.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 257.0, + 1258.0, + 257.0, + 1258.0, + 302.0, + 1250.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 257.0, + 1403.0, + 257.0, + 1403.0, + 302.0, + 1399.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 293.0, + 408.0, + 293.0, + 408.0, + 332.0, + 293.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 293.0, + 1110.0, + 293.0, + 1110.0, + 332.0, + 446.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 293.0, + 1149.0, + 293.0, + 1149.0, + 332.0, + 1140.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 332.0, + 1220.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 322.0, + 596.0, + 322.0, + 596.0, + 361.0, + 293.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 322.0, + 750.0, + 322.0, + 750.0, + 361.0, + 617.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 322.0, + 1288.0, + 322.0, + 1288.0, + 361.0, + 782.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 361.0, + 1311.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 355.0, + 463.0, + 355.0, + 463.0, + 393.0, + 294.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 355.0, + 592.0, + 355.0, + 592.0, + 393.0, + 499.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 355.0, + 1003.0, + 355.0, + 1003.0, + 393.0, + 832.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 355.0, + 1043.0, + 355.0, + 1043.0, + 393.0, + 1034.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 393.0, + 1133.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 386.0, + 605.0, + 386.0, + 605.0, + 424.0, + 293.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 386.0, + 1405.0, + 386.0, + 1405.0, + 424.0, + 692.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 417.0, + 880.0, + 417.0, + 880.0, + 456.0, + 290.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 417.0, + 1406.0, + 417.0, + 1406.0, + 456.0, + 920.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 451.0, + 576.0, + 451.0, + 576.0, + 486.0, + 295.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 1404.0, + 1426.0, + 1404.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1458.0, + 1401.0, + 1458.0, + 1401.0, + 1490.0, + 295.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1487.0, + 1404.0, + 1487.0, + 1404.0, + 1523.0, + 295.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1516.0, + 1405.0, + 1516.0, + 1405.0, + 1554.0, + 294.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 1405.0, + 1548.0, + 1405.0, + 1584.0, + 294.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1580.0, + 363.0, + 1580.0, + 363.0, + 1612.0, + 296.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1580.0, + 484.0, + 1580.0, + 484.0, + 1612.0, + 435.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1580.0, + 565.0, + 1580.0, + 565.0, + 1612.0, + 555.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 655.0, + 387.0, + 655.0, + 387.0, + 703.0, + 291.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 655.0, + 677.0, + 655.0, + 677.0, + 703.0, + 617.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 655.0, + 1151.0, + 655.0, + 1151.0, + 703.0, + 1137.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 655.0, + 1406.0, + 655.0, + 1406.0, + 703.0, + 1315.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 693.0, + 822.0, + 693.0, + 822.0, + 739.0, + 290.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 693.0, + 1406.0, + 693.0, + 1406.0, + 739.0, + 1024.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 726.0, + 1228.0, + 726.0, + 1228.0, + 764.0, + 294.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 726.0, + 1404.0, + 726.0, + 1404.0, + 764.0, + 1258.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 759.0, + 954.0, + 759.0, + 954.0, + 801.0, + 294.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 759.0, + 1168.0, + 759.0, + 1168.0, + 801.0, + 1157.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 799.0, + 569.0, + 799.0, + 569.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 799.0, + 772.0, + 799.0, + 772.0, + 844.0, + 728.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 799.0, + 921.0, + 799.0, + 921.0, + 844.0, + 870.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 799.0, + 1115.0, + 799.0, + 1115.0, + 844.0, + 1006.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 799.0, + 1348.0, + 799.0, + 1348.0, + 844.0, + 1213.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 799.0, + 1408.0, + 799.0, + 1408.0, + 844.0, + 1369.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 836.0, + 296.0, + 836.0, + 296.0, + 885.0, + 290.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 836.0, + 816.0, + 836.0, + 816.0, + 885.0, + 804.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 836.0, + 1408.0, + 836.0, + 1408.0, + 885.0, + 1124.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 876.0, + 354.0, + 876.0, + 354.0, + 921.0, + 293.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 876.0, + 628.0, + 876.0, + 628.0, + 921.0, + 420.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 876.0, + 840.0, + 876.0, + 840.0, + 921.0, + 827.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 876.0, + 1052.0, + 876.0, + 1052.0, + 921.0, + 950.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1792.0, + 370.0, + 1792.0, + 370.0, + 1829.0, + 296.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1792.0, + 461.0, + 1792.0, + 461.0, + 1829.0, + 421.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1792.0, + 1293.0, + 1792.0, + 1293.0, + 1829.0, + 684.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1792.0, + 1342.0, + 1792.0, + 1342.0, + 1829.0, + 1313.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1825.0, + 925.0, + 1825.0, + 925.0, + 1867.0, + 769.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1825.0, + 1104.0, + 1825.0, + 1104.0, + 1867.0, + 960.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1820.0, + 773.0, + 1820.0, + 773.0, + 1867.5, + 292.0, + 1867.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1876.0, + 741.0, + 1876.0, + 741.0, + 1913.0, + 293.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1910.0, + 309.0, + 1910.0, + 309.0, + 1968.0, + 288.0, + 1968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1426.0, + 1910.0, + 1436.0, + 1910.0, + 1436.0, + 1968.0, + 1426.0, + 1968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1964.0, + 373.0, + 1964.0, + 373.0, + 2012.0, + 292.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1964.0, + 794.0, + 1964.0, + 794.0, + 2012.0, + 465.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1964.0, + 1068.0, + 1964.0, + 1068.0, + 2012.0, + 994.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1964.0, + 1200.0, + 1964.0, + 1200.0, + 2012.0, + 1147.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1964.0, + 1405.0, + 1964.0, + 1405.0, + 2012.0, + 1401.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1997.0, + 363.0, + 1997.0, + 363.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1997.0, + 445.0, + 1997.0, + 445.0, + 2040.0, + 435.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1997.0, + 624.0, + 1997.0, + 624.0, + 2040.0, + 611.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 481.0, + 598.0, + 481.0, + 598.0, + 529.0, + 291.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 481.0, + 823.0, + 481.0, + 823.0, + 529.0, + 749.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 481.0, + 951.0, + 481.0, + 951.0, + 529.0, + 903.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 481.0, + 1142.0, + 481.0, + 1142.0, + 529.0, + 1040.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 481.0, + 1405.0, + 481.0, + 1405.0, + 529.0, + 1211.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 517.0, + 381.0, + 517.0, + 381.0, + 550.0, + 293.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 517.0, + 458.0, + 517.0, + 458.0, + 550.0, + 446.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1609.0, + 556.0, + 1609.0, + 556.0, + 1657.0, + 292.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1609.0, + 1407.0, + 1609.0, + 1407.0, + 1657.0, + 769.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1640.0, + 373.0, + 1640.0, + 373.0, + 1684.0, + 292.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1640.0, + 584.0, + 1640.0, + 584.0, + 1684.0, + 570.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 604.0, + 371.0, + 604.0, + 371.0, + 643.0, + 296.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 604.0, + 664.0, + 604.0, + 664.0, + 643.0, + 398.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1365.0, + 1352.0, + 1365.0, + 1352.0, + 1400.0, + 296.0, + 1400.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 479, + 1405, + 479, + 1405, + 703, + 296, + 703 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 295, + 1664, + 1405, + 1664, + 1405, + 1881, + 295, + 1881 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 1202, + 1404, + 1202, + 1404, + 1348, + 296, + 1348 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 890, + 1404, + 890, + 1404, + 1037, + 296, + 1037 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 787, + 1404, + 787, + 1404, + 883, + 297, + 883 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 1531, + 1404, + 1531, + 1404, + 1655, + 296, + 1655 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 291, + 228, + 1401, + 228, + 1401, + 295, + 291, + 295 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 294, + 406, + 1398, + 406, + 1398, + 470, + 294, + 470 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 554, + 315, + 1145, + 315, + 1145, + 386, + 554, + 386 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 294, + 1438, + 1401, + 1438, + 1401, + 1506, + 294, + 1506 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 296, + 1125, + 1403, + 1125, + 1403, + 1194, + 296, + 1194 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 431, + 1057, + 1264, + 1057, + 1264, + 1108, + 431, + 1108 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 2002, + 1140, + 2002, + 1140, + 2036, + 297, + 2036 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.924 + }, + { + "category_id": 8, + "poly": [ + 433, + 1367, + 1265, + 1367, + 1265, + 1419, + 433, + 1419 + ], + "score": 0.918 + }, + { + "category_id": 8, + "poly": [ + 311, + 1894, + 1385, + 1894, + 1385, + 1986, + 311, + 1986 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 301, + 727, + 1192, + 727, + 1192, + 762, + 301, + 762 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2113, + 840, + 2113 + ], + "score": 0.816 + }, + { + "category_id": 14, + "poly": [ + 313, + 1892, + 1383, + 1892, + 1383, + 1986, + 313, + 1986 + ], + "score": 0.94, + "latex": "n \\geqslant r _ { k } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { d - 1 } } = \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } { \\frac { 2 k ^ { \\prime } + d - 1 } { k ^ { \\prime } } } { \\binom { k ^ { \\prime } + d - 2 } { k ^ { \\prime } - 1 } } = \\Omega ( d ^ { k - 1 } ) ." + }, + { + "category_id": 14, + "poly": [ + 553, + 312, + 1143, + 312, + 1143, + 386, + 553, + 386 + ], + "score": 0.94, + "latex": "L _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } ) ," + }, + { + "category_id": 13, + "poly": [ + 681, + 1126, + 994, + 1126, + 994, + 1165, + 681, + 1165 + ], + "score": 0.94, + "latex": "\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 562, + 259, + 656, + 259, + 656, + 297, + 562, + 297 + ], + "score": 0.93, + "latex": "L _ { \\tau _ { d } } ^ { 2 } ( \\mathbb { S } ^ { d } )" + }, + { + "category_id": 13, + "poly": [ + 931, + 1203, + 1021, + 1203, + 1021, + 1240, + 931, + 1240 + ], + "score": 0.93, + "latex": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1125, + 624, + 1125, + 624, + 1166, + 374, + 1166 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1312, + 406, + 1312, + 406, + 1350, + 297, + 1350 + ], + "score": 0.93, + "latex": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 374, + 1436, + 624, + 1436, + 624, + 1478, + 374, + 1478 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { r _ { k } \\ = \\ \\sum _ { k ^ { \\prime } = 0 } ^ { k - 1 } N ( d , k ^ { \\prime } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 882, + 891, + 971, + 891, + 971, + 928, + 882, + 928 + ], + "score": 0.92, + "latex": "\\left\\{ { \\bf x } _ { i } \\right\\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 878, + 607, + 1065, + 607, + 1065, + 641, + 878, + 641 + ], + "score": 0.92, + "latex": "\\mu _ { k } = \\bar { \\Omega ( k ^ { - } { } ^ { d - 1 } ) }" + }, + { + "category_id": 13, + "poly": [ + 681, + 1438, + 994, + 1438, + 994, + 1476, + 681, + 1476 + ], + "score": 0.92, + "latex": "\\mathbf { V } _ { r _ { k } } \\ = \\ ( n ^ { - 1 / 2 } \\phi _ { j } ( \\mathbf { x } _ { i } ) ) _ { n \\times r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1001, + 406, + 1001, + 406, + 1039, + 298, + 1039 + ], + "score": 0.92, + "latex": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1132, + 1192, + 1132, + 1192, + 1165, + 1060, + 1165 + ], + "score": 0.92, + "latex": "\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1443, + 1192, + 1443, + 1192, + 1476, + 1060, + 1476 + ], + "score": 0.92, + "latex": "\\phi _ { 1 } , \\ldots , \\phi _ { r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 1061, + 1235, + 1314, + 1235, + 1314, + 1276, + 1061, + 1276 + ], + "score": 0.92, + "latex": "n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } d ^ { 2 k - 2 } r _ { k } ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 542, + 605, + 542, + 605, + 580, + 297, + 580 + ], + "score": 0.91, + "latex": "T = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\log { ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) } }" + }, + { + "category_id": 13, + "poly": [ + 617, + 931, + 715, + 931, + 715, + 963, + 617, + 963 + ], + "score": 0.91, + "latex": "\\epsilon , \\delta > 0" + }, + { + "category_id": 13, + "poly": [ + 1187, + 1276, + 1390, + 1276, + 1390, + 1314, + 1187, + 1314 + ], + "score": 0.91, + "latex": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 1187, + 965, + 1392, + 965, + 1392, + 1002, + 1187, + 1002 + ], + "score": 0.91, + "latex": "\\eta = \\tilde { \\mathcal { O } } ( ( m \\theta ^ { 2 } ) ^ { - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 729, + 1242, + 843, + 1242, + 843, + 1275, + 729, + 1275 + ], + "score": 0.91, + "latex": "\\epsilon , \\delta \\ > \\ 0" + }, + { + "category_id": 13, + "poly": [ + 298, + 963, + 554, + 963, + 554, + 1001, + 298, + 1001 + ], + "score": 0.91, + "latex": "\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , k ^ { d + 1 } , \\epsilon ^ { - 1 } ) )" + }, + { + "category_id": 13, + "poly": [ + 1075, + 542, + 1393, + 542, + 1393, + 581, + 1075, + 581 + ], + "score": 0.91, + "latex": "m = ( d { + } 1 ) ^ { \\mathcal { O } ( k ) \\mathrm { p o l y } ( \\| f ^ { * } \\| _ { 2 } / \\epsilon ) }" + }, + { + "category_id": 13, + "poly": [ + 506, + 638, + 847, + 638, + 847, + 672, + 506, + 672 + ], + "score": 0.91, + "latex": "\\mu _ { k } = \\Omega ( \\operatorname* { m a x } ( k ^ { - d - 1 } , d ^ { - k + \\bar { 1 } } ) )" + }, + { + "category_id": 13, + "poly": [ + 638, + 1204, + 726, + 1204, + 726, + 1234, + 638, + 1234 + ], + "score": 0.9, + "latex": "d \\ \\gg \\ k" + }, + { + "category_id": 13, + "poly": [ + 1156, + 409, + 1197, + 409, + 1197, + 441, + 1156, + 441 + ], + "score": 0.9, + "latex": "\\lambda _ { r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 828, + 1728, + 905, + 1728, + 905, + 1756, + 828, + 1756 + ], + "score": 0.9, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 911, + 926, + 1327, + 926, + 1327, + 965, + 911, + 965 + ], + "score": 0.9, + "latex": "n \\geqslant \\widetilde \\Omega ( \\epsilon ^ { - 2 } \\cdot \\operatorname* { m a x } \\{ k ^ { 2 d + 2 } , k ^ { 2 d - 2 } r _ { k } ^ { 2 } \\} )" + }, + { + "category_id": 14, + "poly": [ + 434, + 1366, + 1264, + 1366, + 1264, + 1418, + 434, + 1418 + ], + "score": 0.9, + "latex": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( d ^ { - k + 2 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ," + }, + { + "category_id": 13, + "poly": [ + 980, + 2003, + 1050, + 2003, + 1050, + 2031, + 980, + 2031 + ], + "score": 0.9, + "latex": "k \\gg d" + }, + { + "category_id": 14, + "poly": [ + 434, + 1055, + 1263, + 1055, + 1263, + 1107, + 434, + 1107 + ], + "score": 0.9, + "latex": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 \\left( 1 - \\Omega \\left( k ^ { - d - 1 } \\right) \\right) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ," + }, + { + "category_id": 13, + "poly": [ + 299, + 1274, + 554, + 1274, + 554, + 1313, + 299, + 1313 + ], + "score": 0.9, + "latex": "\\tilde { \\Omega } ( \\mathrm { p o l y } ( T , d ^ { k - 2 } , \\epsilon ^ { - 1 } ) )" + }, + { + "category_id": 13, + "poly": [ + 952, + 515, + 1034, + 515, + 1034, + 543, + 952, + 543 + ], + "score": 0.89, + "latex": "\\epsilon _ { 0 } + \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 625, + 894, + 696, + 894, + 696, + 923, + 625, + 923 + ], + "score": 0.89, + "latex": "k \\gg d" + }, + { + "category_id": 13, + "poly": [ + 1081, + 639, + 1150, + 639, + 1150, + 667, + 1081, + 667 + ], + "score": 0.89, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 373, + 579, + 407, + 579, + 407, + 610, + 373, + 610 + ], + "score": 0.88, + "latex": "f ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 488, + 262, + 524, + 262, + 524, + 292, + 488, + 292 + ], + "score": 0.88, + "latex": "L _ { \\kappa }" + }, + { + "category_id": 13, + "poly": [ + 955, + 405, + 987, + 405, + 987, + 435, + 955, + 435 + ], + "score": 0.88, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 821, + 851, + 888, + 851, + 888, + 880, + 821, + 880 + ], + "score": 0.88, + "latex": "k - 1" + }, + { + "category_id": 13, + "poly": [ + 344, + 442, + 405, + 442, + 405, + 471, + 344, + 471 + ], + "score": 0.87, + "latex": "\\mu _ { k - 1 }" + }, + { + "category_id": 13, + "poly": [ + 563, + 1239, + 595, + 1239, + 595, + 1271, + 563, + 1271 + ], + "score": 0.86, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 822, + 1476, + 888, + 1476, + 888, + 1502, + 822, + 1502 + ], + "score": 0.86, + "latex": "k - 1" + }, + { + "category_id": 13, + "poly": [ + 472, + 929, + 504, + 929, + 504, + 960, + 472, + 960 + ], + "score": 0.86, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 822, + 1165, + 888, + 1165, + 888, + 1191, + 822, + 1191 + ], + "score": 0.86, + "latex": "k - 1" + }, + { + "category_id": 13, + "poly": [ + 467, + 1564, + 507, + 1564, + 507, + 1595, + 467, + 1595 + ], + "score": 0.86, + "latex": "\\lambda _ { r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 903, + 236, + 937, + 236, + 937, + 263, + 903, + 263 + ], + "score": 0.86, + "latex": "\\mu _ { k }" + }, + { + "category_id": 13, + "poly": [ + 720, + 441, + 749, + 441, + 749, + 468, + 720, + 468 + ], + "score": 0.86, + "latex": "\\tau _ { d }" + }, + { + "category_id": 13, + "poly": [ + 706, + 581, + 733, + 581, + 733, + 608, + 706, + 608 + ], + "score": 0.85, + "latex": "\\epsilon _ { \\mathrm { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1338, + 900, + 1366, + 900, + 1366, + 924, + 1338, + 924 + ], + "score": 0.85, + "latex": "\\tau _ { d }" + }, + { + "category_id": 13, + "poly": [ + 372, + 413, + 400, + 413, + 400, + 438, + 372, + 438 + ], + "score": 0.84, + "latex": "\\tau _ { d }" + }, + { + "category_id": 13, + "poly": [ + 902, + 969, + 968, + 969, + 968, + 998, + 902, + 998 + ], + "score": 0.83, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 902, + 1280, + 969, + 1280, + 969, + 1310, + 902, + 1310 + ], + "score": 0.82, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 439, + 1759, + 459, + 1759, + 459, + 1785, + 439, + 1785 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 520, + 1563, + 569, + 1563, + 569, + 1596, + 520, + 1596 + ], + "score": 0.82, + "latex": "\\mathbf { V } _ { r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 1332, + 1241, + 1403, + 1241, + 1403, + 1274, + 1332, + 1274 + ], + "score": 0.81, + "latex": "m \\geqslant" + }, + { + "category_id": 13, + "poly": [ + 466, + 1698, + 486, + 1698, + 486, + 1724, + 466, + 1724 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1340, + 931, + 1404, + 931, + 1404, + 963, + 1340, + 963 + ], + "score": 0.8, + "latex": "m \\geqslant" + }, + { + "category_id": 13, + "poly": [ + 993, + 1244, + 1012, + 1244, + 1012, + 1271, + 993, + 1271 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1159, + 821, + 1178, + 821, + 1178, + 847, + 1159, + 847 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 853, + 932, + 872, + 932, + 872, + 960, + 853, + 960 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 622, + 1564, + 655, + 1564, + 655, + 1591, + 622, + 1591 + ], + "score": 0.78, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 862, + 2012, + 888, + 2012, + 888, + 2030, + 862, + 2030 + ], + "score": 0.78, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 297, + 1247, + 326, + 1247, + 326, + 1273, + 297, + 1273 + ], + "score": 0.78, + "latex": "\\tau _ { d }" + }, + { + "category_id": 13, + "poly": [ + 650, + 1824, + 670, + 1824, + 670, + 1846, + 650, + 1846 + ], + "score": 0.75, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 520, + 1698, + 538, + 1698, + 538, + 1724, + 520, + 1724 + ], + "score": 0.74, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1104, + 1673, + 1124, + 1673, + 1124, + 1694, + 1104, + 1694 + ], + "score": 0.7, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1175, + 1672, + 1204, + 1672, + 1204, + 1695, + 1175, + 1695 + ], + "score": 0.65, + "latex": "m" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 727.0, + 1196.0, + 727.0, + 1196.0, + 765.0, + 293.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 477.0, + 1403.0, + 477.0, + 1403.0, + 514.0, + 293.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 511.0, + 951.0, + 511.0, + 951.0, + 547.0, + 295.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 511.0, + 1403.0, + 511.0, + 1403.0, + 547.0, + 1035.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 537.0, + 296.0, + 537.0, + 296.0, + 583.0, + 290.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 537.0, + 1074.0, + 537.0, + 1074.0, + 583.0, + 606.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 537.0, + 1407.0, + 537.0, + 1407.0, + 583.0, + 1394.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 575.0, + 372.0, + 575.0, + 372.0, + 613.0, + 294.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 575.0, + 705.0, + 575.0, + 705.0, + 613.0, + 408.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 575.0, + 1405.0, + 575.0, + 1405.0, + 613.0, + 734.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 602.0, + 877.0, + 602.0, + 877.0, + 645.0, + 291.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 602.0, + 1408.0, + 602.0, + 1408.0, + 645.0, + 1066.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 632.0, + 505.0, + 632.0, + 505.0, + 676.0, + 289.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 632.0, + 1080.0, + 632.0, + 1080.0, + 676.0, + 848.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 632.0, + 1410.0, + 632.0, + 1410.0, + 676.0, + 1151.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 664.0, + 489.0, + 664.0, + 489.0, + 708.0, + 292.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1664.0, + 1103.0, + 1664.0, + 1103.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1664.0, + 1174.0, + 1664.0, + 1174.0, + 1701.0, + 1125.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1664.0, + 1403.0, + 1664.0, + 1403.0, + 1701.0, + 1205.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1694.0, + 465.0, + 1694.0, + 465.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1694.0, + 519.0, + 1694.0, + 519.0, + 1730.0, + 487.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1694.0, + 1403.0, + 1694.0, + 1403.0, + 1730.0, + 539.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1728.0, + 827.0, + 1728.0, + 827.0, + 1763.0, + 295.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1728.0, + 1403.0, + 1728.0, + 1403.0, + 1763.0, + 906.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 438.0, + 1756.0, + 438.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1795.0, + 460.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1787.0, + 1403.0, + 1787.0, + 1403.0, + 1822.0, + 294.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 649.0, + 1818.0, + 649.0, + 1853.0, + 294.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1818.0, + 1407.0, + 1818.0, + 1407.0, + 1853.0, + 671.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 580.0, + 1848.0, + 580.0, + 1883.0, + 292.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 637.0, + 1198.0, + 637.0, + 1246.0, + 293.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1198.0, + 930.0, + 1198.0, + 930.0, + 1246.0, + 727.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1198.0, + 1409.0, + 1198.0, + 1409.0, + 1246.0, + 1022.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1237.0, + 296.0, + 1237.0, + 296.0, + 1279.0, + 290.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1237.0, + 562.0, + 1237.0, + 562.0, + 1279.0, + 327.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1237.0, + 728.0, + 1237.0, + 728.0, + 1279.0, + 596.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1237.0, + 992.0, + 1237.0, + 992.0, + 1279.0, + 844.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1237.0, + 1060.0, + 1237.0, + 1060.0, + 1279.0, + 1013.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1237.0, + 1331.0, + 1237.0, + 1331.0, + 1279.0, + 1315.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1272.0, + 298.0, + 1272.0, + 298.0, + 1320.0, + 291.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1272.0, + 901.0, + 1272.0, + 901.0, + 1320.0, + 555.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1272.0, + 1186.0, + 1272.0, + 1186.0, + 1320.0, + 970.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 1272.0, + 1405.0, + 1272.0, + 1405.0, + 1320.0, + 1391.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1309.0, + 296.0, + 1309.0, + 296.0, + 1353.0, + 292.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1309.0, + 508.0, + 1309.0, + 508.0, + 1353.0, + 407.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 886.0, + 624.0, + 886.0, + 624.0, + 935.0, + 294.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 886.0, + 881.0, + 886.0, + 881.0, + 935.0, + 697.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 886.0, + 1337.0, + 886.0, + 1337.0, + 935.0, + 972.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 886.0, + 1407.0, + 886.0, + 1407.0, + 935.0, + 1367.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 921.0, + 471.0, + 921.0, + 471.0, + 970.0, + 291.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 921.0, + 616.0, + 921.0, + 616.0, + 970.0, + 505.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 921.0, + 852.0, + 921.0, + 852.0, + 970.0, + 716.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 921.0, + 910.0, + 921.0, + 910.0, + 970.0, + 873.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 921.0, + 1339.0, + 921.0, + 1339.0, + 970.0, + 1328.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 960.0, + 297.0, + 960.0, + 297.0, + 1006.0, + 291.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 960.0, + 901.0, + 960.0, + 901.0, + 1006.0, + 555.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 960.0, + 1186.0, + 960.0, + 1186.0, + 1006.0, + 969.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 1006.0, + 1393.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 998.0, + 297.0, + 998.0, + 297.0, + 1042.0, + 292.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 998.0, + 508.0, + 998.0, + 508.0, + 1042.0, + 407.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 786.0, + 1404.0, + 786.0, + 1404.0, + 826.0, + 291.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 819.0, + 1158.0, + 819.0, + 1158.0, + 856.0, + 292.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 819.0, + 1406.0, + 819.0, + 1406.0, + 856.0, + 1179.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 851.0, + 820.0, + 851.0, + 820.0, + 885.0, + 295.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 851.0, + 1099.0, + 851.0, + 1099.0, + 885.0, + 889.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1529.0, + 1405.0, + 1529.0, + 1405.0, + 1567.0, + 295.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1563.0, + 466.0, + 1563.0, + 466.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1563.0, + 519.0, + 1563.0, + 519.0, + 1595.0, + 508.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1563.0, + 621.0, + 1563.0, + 621.0, + 1595.0, + 570.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1563.0, + 1404.0, + 1563.0, + 1404.0, + 1595.0, + 656.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1593.0, + 1404.0, + 1593.0, + 1404.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1625.0, + 486.0, + 1625.0, + 486.0, + 1655.0, + 295.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 902.0, + 226.0, + 902.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 265.0, + 938.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 256.0, + 487.0, + 256.0, + 487.0, + 299.0, + 293.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 256.0, + 561.0, + 256.0, + 561.0, + 299.0, + 525.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 256.0, + 783.0, + 256.0, + 783.0, + 299.0, + 657.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 402.0, + 371.0, + 402.0, + 371.0, + 443.0, + 293.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 402.0, + 954.0, + 402.0, + 954.0, + 443.0, + 401.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 402.0, + 1155.0, + 402.0, + 1155.0, + 443.0, + 988.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 402.0, + 1407.0, + 402.0, + 1407.0, + 443.0, + 1198.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 437.0, + 343.0, + 437.0, + 343.0, + 471.0, + 292.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 437.0, + 719.0, + 437.0, + 719.0, + 471.0, + 406.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 437.0, + 1014.0, + 437.0, + 1014.0, + 471.0, + 750.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1414.0, + 373.0, + 1414.0, + 373.0, + 1501.0, + 293.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1414.0, + 680.0, + 1414.0, + 680.0, + 1501.0, + 625.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1414.0, + 1059.0, + 1414.0, + 1059.0, + 1501.0, + 995.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1414.0, + 1418.0, + 1414.0, + 1418.0, + 1501.0, + 1193.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1471.0, + 821.0, + 1471.0, + 821.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1471.0, + 902.0, + 1471.0, + 902.0, + 1507.0, + 889.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 281.0, + 1104.0, + 373.0, + 1104.0, + 373.0, + 1196.0, + 281.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1104.0, + 680.0, + 1104.0, + 680.0, + 1196.0, + 625.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1104.0, + 1059.0, + 1104.0, + 1059.0, + 1196.0, + 995.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1104.0, + 1417.0, + 1104.0, + 1417.0, + 1196.0, + 1193.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2000.0, + 861.0, + 2000.0, + 861.0, + 2040.0, + 296.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 2000.0, + 979.0, + 2000.0, + 979.0, + 2040.0, + 889.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 2000.0, + 1141.0, + 2000.0, + 1141.0, + 2040.0, + 1051.0, + 2040.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 627, + 1405, + 627, + 1405, + 827, + 297, + 827 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 291, + 1404, + 291, + 1404, + 416, + 297, + 416 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1083, + 1404, + 1083, + 1404, + 1238, + 297, + 1238 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1881, + 1403, + 1881, + 1403, + 2035, + 298, + 2035 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 840, + 1404, + 840, + 1404, + 940, + 298, + 940 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 310, + 1269, + 1385, + 1269, + 1385, + 1711, + 310, + 1711 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 296, + 1747, + 1404, + 1747, + 1404, + 1872, + 296, + 1872 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 295, + 500, + 1400, + 500, + 1400, + 563, + 295, + 563 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 486, + 577, + 1210, + 577, + 1210, + 617, + 486, + 617 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 762, + 953, + 933, + 953, + 933, + 996, + 762, + 996 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 297, + 443, + 1002, + 443, + 1002, + 474, + 297, + 474 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 558, + 226, + 558, + 262, + 299, + 262 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 297, + 1010, + 1400, + 1010, + 1400, + 1074, + 297, + 1074 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.699 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.191 + }, + { + "category_id": 13, + "poly": [ + 415, + 628, + 480, + 628, + 480, + 662, + 415, + 662 + ], + "score": 0.94, + "latex": "P _ { k } ( t )" + }, + { + "category_id": 13, + "poly": [ + 711, + 1114, + 1020, + 1114, + 1020, + 1148, + 711, + 1148 + ], + "score": 0.93, + "latex": "r ( \\mathbf { x } ) \\bar { = } f ^ { * } ( \\mathbf { x } ) - \\bar { \\theta } \\bar { f } _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )" + }, + { + "category_id": 13, + "poly": [ + 409, + 1145, + 481, + 1145, + 481, + 1178, + 409, + 1178 + ], + "score": 0.92, + "latex": "P _ { k } ( { \\bf x } )" + }, + { + "category_id": 13, + "poly": [ + 827, + 839, + 1215, + 839, + 1215, + 877, + 827, + 877 + ], + "score": 0.92, + "latex": "\\mathbf v _ { k } = n ^ { - 1 / 2 } ( P _ { k } ( \\mathbf x _ { 1 } ) , \\dots , P _ { k } ( \\mathbf x _ { n } ) )" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1013, + 1166, + 1013, + 1166, + 1046, + 1060, + 1046 + ], + "score": 0.92, + "latex": "f _ { \\mathbf { W } ( t ) } \\left( \\mathbf { x } \\right)" + }, + { + "category_id": 13, + "poly": [ + 877, + 733, + 948, + 733, + 948, + 765, + 877, + 765 + ], + "score": 0.92, + "latex": "f ^ { * } ( \\mathbf { x } )" + }, + { + "category_id": 13, + "poly": [ + 675, + 690, + 1194, + 690, + 1194, + 735, + 675, + 735 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 626, + 795, + 710, + 795, + 710, + 825, + 626, + 825 + ], + "score": 0.91, + "latex": "\\mu _ { k } = 0" + }, + { + "category_id": 14, + "poly": [ + 764, + 951, + 933, + 951, + 933, + 996, + 764, + 996 + ], + "score": 0.9, + "latex": "\\widehat { a } _ { k } = | \\mathbf { v } _ { k } ^ { \\top } \\mathbf { r } ^ { ( t ) } | ," + }, + { + "category_id": 13, + "poly": [ + 1019, + 382, + 1065, + 382, + 1065, + 411, + 1019, + 411 + ], + "score": 0.89, + "latex": "\\mathbb { R } ^ { 1 0 }" + }, + { + "category_id": 13, + "poly": [ + 889, + 629, + 1053, + 629, + 1053, + 660, + 889, + 660 + ], + "score": 0.89, + "latex": "\\zeta _ { k } , k = 1 , 2 , 4" + }, + { + "category_id": 13, + "poly": [ + 515, + 1085, + 547, + 1085, + 547, + 1115, + 515, + 1115 + ], + "score": 0.88, + "latex": "\\widehat { a } _ { k }" + }, + { + "category_id": 13, + "poly": [ + 498, + 904, + 540, + 904, + 540, + 935, + 498, + 935 + ], + "score": 0.88, + "latex": "\\mathbf { r } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 1017, + 658, + 1063, + 658, + 1063, + 687, + 1017, + 687 + ], + "score": 0.88, + "latex": "\\mathbb { R } ^ { 1 0 }" + }, + { + "category_id": 13, + "poly": [ + 754, + 794, + 916, + 794, + 916, + 826, + 754, + 826 + ], + "score": 0.87, + "latex": "k = 3 , 5 , 7 \\dots" + }, + { + "category_id": 13, + "poly": [ + 540, + 1912, + 611, + 1912, + 611, + 1941, + 540, + 1941 + ], + "score": 0.86, + "latex": "k = 1 ." + }, + { + "category_id": 13, + "poly": [ + 601, + 911, + 634, + 911, + 634, + 938, + 601, + 938 + ], + "score": 0.86, + "latex": "\\mathbf { v } _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1177, + 1913, + 1249, + 1913, + 1249, + 1941, + 1177, + 1941 + ], + "score": 0.86, + "latex": "k = 2" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1149, + 1248, + 1149, + 1248, + 1176, + 1217, + 1176 + ], + "score": 0.86, + "latex": "\\mathbf { x } _ { i }" + }, + { + "category_id": 14, + "poly": [ + 488, + 577, + 1208, + 577, + 1208, + 615, + 488, + 615 + ], + "score": 0.84, + "latex": "f ^ { * } ( { \\bf x } ) = a _ { 1 } \\cdot P _ { 1 } ( \\langle \\zeta _ { 1 } , { \\bf x } \\rangle ) + a _ { 2 } \\cdot P _ { 2 } ( \\langle \\zeta _ { 2 } , { \\bf x } \\rangle ) + a _ { 4 } \\cdot P _ { 4 } ( \\langle \\zeta _ { 4 } , { \\bf x } \\rangle ) ," + }, + { + "category_id": 13, + "poly": [ + 296, + 877, + 331, + 877, + 331, + 905, + 296, + 905 + ], + "score": 0.83, + "latex": "\\mathbf { v } _ { k }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1009, + 1010, + 1009, + 1010, + 1046, + 371, + 1046 + ], + "score": 0.8, + "latex": "\\mathbf { r } ^ { ( t ) } = \\left( f ^ { * } \\big ( \\mathbf { x } _ { 1 } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { 1 } \\big ) , \\ldots , f ^ { * } \\big ( \\mathbf { x } _ { n } \\big ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } \\big ( \\mathbf { x } _ { n } \\big ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 717, + 911, + 731, + 911, + 731, + 935, + 717, + 935 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 884, + 629, + 914, + 629, + 914, + 661, + 884, + 661 + ], + "score": 0.48, + "latex": "\\zeta _ { k }" + }, + { + "category_id": 14, + "poly": [ + 959, + 1272, + 1298, + 1272, + 1298, + 1305, + 959, + 1305 + ], + "score": 0.32, + "latex": "f ^ { * } ( x ) = 1 P _ { 1 } ( x ) + 3 P _ { 2 } ( x ) + 5 P _ { 4 } ( x )" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1271.0, + 757.0, + 1271.0, + 757.0, + 1306.0, + 419.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1305.0, + 380.0, + 1305.0, + 380.0, + 1333.0, + 342.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1310.0, + 825.0, + 1310.0, + 825.0, + 1349.0, + 767.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1312.0, + 910.0, + 1312.0, + 910.0, + 1332.0, + 890.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1315.0, + 1371.0, + 1315.0, + 1371.0, + 1348.0, + 1314.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1324.0, + 362.0, + 1324.0, + 362.0, + 1594.0, + 299.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1328.0, + 904.0, + 1328.0, + 904.0, + 1589.0, + 849.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1337.0, + 827.0, + 1337.0, + 827.0, + 1379.0, + 766.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1339.0, + 1373.0, + 1339.0, + 1373.0, + 1379.0, + 1311.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1353.0, + 379.0, + 1353.0, + 379.0, + 1380.0, + 340.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1363.0, + 904.0, + 1363.0, + 904.0, + 1375.0, + 894.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1370.0, + 825.0, + 1370.0, + 825.0, + 1404.0, + 768.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1370.0, + 1371.0, + 1370.0, + 1371.0, + 1404.0, + 1314.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1399.0, + 379.0, + 1399.0, + 379.0, + 1426.0, + 341.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1410.0, + 905.0, + 1410.0, + 905.0, + 1420.0, + 894.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1446.0, + 379.0, + 1446.0, + 379.0, + 1474.0, + 341.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1456.0, + 905.0, + 1456.0, + 905.0, + 1468.0, + 894.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1496.0, + 378.0, + 1496.0, + 378.0, + 1519.0, + 342.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1503.0, + 905.0, + 1503.0, + 905.0, + 1514.0, + 894.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1538.0, + 378.0, + 1538.0, + 378.0, + 1569.0, + 344.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1588.0, + 379.0, + 1588.0, + 379.0, + 1615.0, + 342.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1593.0, + 907.0, + 1593.0, + 907.0, + 1610.0, + 892.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1618.0, + 407.0, + 1618.0, + 407.0, + 1639.0, + 389.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 1616.0, + 526.0, + 1616.0, + 526.0, + 1640.0, + 428.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1615.0, + 838.0, + 1615.0, + 838.0, + 1641.0, + 528.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1621.0, + 941.0, + 1621.0, + 941.0, + 1638.0, + 925.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1615.0, + 1385.0, + 1615.0, + 1385.0, + 1642.0, + 963.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1630.0, + 652.0, + 1630.0, + 652.0, + 1671.0, + 561.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1632.0, + 1191.0, + 1632.0, + 1191.0, + 1669.0, + 1103.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1680.0, + 760.0, + 1680.0, + 760.0, + 1711.0, + 393.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1680.0, + 1306.0, + 1680.0, + 1306.0, + 1710.0, + 940.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1749.0, + 1404.0, + 1749.0, + 1404.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1777.0, + 1405.0, + 1777.0, + 1405.0, + 1814.0, + 293.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1808.0, + 1404.0, + 1808.0, + 1404.0, + 1843.0, + 294.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1839.0, + 975.0, + 1839.0, + 975.0, + 1875.0, + 294.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 442.0, + 1004.0, + 442.0, + 1004.0, + 476.0, + 295.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 561.0, + 224.0, + 561.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 627.0, + 414.0, + 627.0, + 414.0, + 664.0, + 295.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 627.0, + 883.0, + 627.0, + 883.0, + 664.0, + 481.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 627.0, + 1403.0, + 627.0, + 1403.0, + 664.0, + 1054.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 657.0, + 674.0, + 657.0, + 674.0, + 769.0, + 278.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 657.0, + 1420.0, + 657.0, + 1420.0, + 769.0, + 1195.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 763.0, + 1405.0, + 763.0, + 1405.0, + 797.0, + 295.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 793.0, + 625.0, + 793.0, + 625.0, + 829.0, + 294.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 793.0, + 753.0, + 793.0, + 753.0, + 829.0, + 711.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 793.0, + 925.0, + 793.0, + 925.0, + 829.0, + 917.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 323.0, + 1405.0, + 323.0, + 1405.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 387.0, + 292.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 382.0, + 1018.0, + 382.0, + 1018.0, + 419.0, + 290.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 382.0, + 1078.0, + 382.0, + 1078.0, + 419.0, + 1066.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1085.0, + 514.0, + 1085.0, + 514.0, + 1118.0, + 296.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1085.0, + 1405.0, + 1085.0, + 1405.0, + 1118.0, + 548.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1110.0, + 710.0, + 1110.0, + 710.0, + 1153.0, + 292.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1110.0, + 1406.0, + 1110.0, + 1406.0, + 1153.0, + 1021.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1143.0, + 408.0, + 1143.0, + 408.0, + 1180.0, + 295.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1143.0, + 1216.0, + 1143.0, + 1216.0, + 1180.0, + 482.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1143.0, + 1406.0, + 1143.0, + 1406.0, + 1180.0, + 1249.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 1404.0, + 1176.0, + 1404.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 751.0, + 1204.0, + 751.0, + 1242.0, + 294.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1914.0, + 297.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 539.0, + 1910.0, + 539.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1910.0, + 1176.0, + 1910.0, + 1176.0, + 1947.0, + 612.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 1250.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 865.0, + 2002.0, + 865.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 836.0, + 826.0, + 836.0, + 826.0, + 880.0, + 291.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 836.0, + 1406.0, + 836.0, + 1406.0, + 880.0, + 1216.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 871.0, + 1404.0, + 871.0, + 1404.0, + 909.0, + 332.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 497.0, + 899.0, + 497.0, + 946.0, + 292.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 899.0, + 600.0, + 899.0, + 600.0, + 946.0, + 541.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 899.0, + 716.0, + 899.0, + 716.0, + 946.0, + 635.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 899.0, + 766.0, + 899.0, + 766.0, + 946.0, + 732.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 498.0, + 1404.0, + 498.0, + 1404.0, + 537.0, + 294.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 529.0, + 771.0, + 529.0, + 771.0, + 566.0, + 295.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1001.0, + 370.0, + 1001.0, + 370.0, + 1054.0, + 291.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1001.0, + 1059.0, + 1001.0, + 1059.0, + 1054.0, + 1011.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1001.0, + 1407.0, + 1001.0, + 1407.0, + 1054.0, + 1167.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1038.0, + 406.0, + 1038.0, + 406.0, + 1079.0, + 293.0, + 1079.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1698, + 1404, + 1698, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1436, + 1403, + 1436, + 1403, + 1683, + 297, + 1683 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 294, + 1404, + 294, + 1404, + 478, + 297, + 478 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1078, + 1403, + 1078, + 1403, + 1204, + 297, + 1204 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1224, + 1403, + 1224, + 1403, + 1319, + 299, + 1319 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 311, + 498, + 1387, + 498, + 1387, + 936, + 311, + 936 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 296, + 970, + 1405, + 970, + 1405, + 1064, + 296, + 1064 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 301, + 1367, + 799, + 1367, + 799, + 1401, + 301, + 1401 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.796 + }, + { + "category_id": 2, + "poly": [ + 299, + 230, + 844, + 230, + 844, + 261, + 299, + 261 + ], + "score": 0.586 + }, + { + "category_id": 0, + "poly": [ + 299, + 230, + 844, + 230, + 844, + 261, + 299, + 261 + ], + "score": 0.324 + }, + { + "category_id": 13, + "poly": [ + 516, + 325, + 587, + 325, + 587, + 359, + 516, + 359 + ], + "score": 0.92, + "latex": "P _ { k } ( x )" + }, + { + "category_id": 13, + "poly": [ + 773, + 1256, + 843, + 1256, + 843, + 1290, + 773, + 1290 + ], + "score": 0.92, + "latex": "P _ { 1 } ( \\mathbf { x } )" + }, + { + "category_id": 13, + "poly": [ + 1156, + 386, + 1393, + 386, + 1393, + 421, + 1156, + 421 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\bar { \\langle \\zeta , \\mathbf { x } \\rangle } ^ { p _ { i } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 816, + 386, + 1118, + 386, + 1118, + 421, + 816, + 421 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { f ^ { * } ( \\mathbf { x } ) = \\sum _ { i } \\cos ( a _ { i } \\langle \\zeta , \\mathbf { x } \\rangle ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1589, + 1158, + 1589, + 1158, + 1619, + 1127, + 1619 + ], + "score": 0.87, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 372, + 417, + 392, + 417, + 392, + 449, + 372, + 449 + ], + "score": 0.83, + "latex": "\\zeta" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 496.0, + 718.0, + 496.0, + 718.0, + 531.0, + 457.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 497.0, + 1264.0, + 497.0, + 1264.0, + 530.0, + 1008.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 518.0, + 931.0, + 518.0, + 931.0, + 544.0, + 888.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 537.0, + 824.0, + 537.0, + 824.0, + 570.0, + 768.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 537.0, + 1371.0, + 537.0, + 1371.0, + 572.0, + 1313.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 564.0, + 826.0, + 564.0, + 826.0, + 598.0, + 768.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 558.0, + 932.0, + 558.0, + 932.0, + 602.0, + 854.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 566.0, + 1372.0, + 566.0, + 1372.0, + 599.0, + 1314.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 609.0, + 379.0, + 609.0, + 379.0, + 635.0, + 342.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 592.0, + 826.0, + 592.0, + 826.0, + 627.0, + 768.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 589.0, + 930.0, + 589.0, + 930.0, + 811.0, + 856.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 595.0, + 1371.0, + 595.0, + 1371.0, + 629.0, + 1311.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 646.0, + 929.0, + 646.0, + 929.0, + 668.0, + 889.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 662.0, + 377.0, + 662.0, + 377.0, + 685.0, + 344.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 688.0, + 929.0, + 688.0, + 929.0, + 710.0, + 889.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 712.0, + 377.0, + 712.0, + 377.0, + 735.0, + 343.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 730.0, + 929.0, + 730.0, + 929.0, + 752.0, + 891.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 764.0, + 375.0, + 764.0, + 375.0, + 787.0, + 344.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 771.0, + 929.0, + 771.0, + 929.0, + 796.0, + 888.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 814.0, + 379.0, + 814.0, + 379.0, + 837.0, + 344.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 813.0, + 931.0, + 813.0, + 931.0, + 838.0, + 888.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 842.0, + 408.0, + 842.0, + 408.0, + 862.0, + 388.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 839.0, + 838.0, + 839.0, + 838.0, + 866.0, + 426.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 842.0, + 962.0, + 842.0, + 962.0, + 862.0, + 942.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 838.0, + 1382.0, + 838.0, + 1382.0, + 865.0, + 976.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 856.0, + 651.0, + 856.0, + 651.0, + 894.0, + 562.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 856.0, + 1201.0, + 856.0, + 1201.0, + 892.0, + 1112.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 901.0, + 777.0, + 901.0, + 777.0, + 937.0, + 377.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 901.0, + 1329.0, + 901.0, + 1329.0, + 937.0, + 916.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 971.0, + 1405.0, + 971.0, + 1405.0, + 1005.0, + 295.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1002.0, + 1405.0, + 1002.0, + 1405.0, + 1036.0, + 295.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1033.0, + 709.0, + 1033.0, + 709.0, + 1067.0, + 295.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1362.0, + 805.0, + 1362.0, + 805.0, + 1408.0, + 291.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 850.0, + 229.0, + 850.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 850.0, + 229.0, + 850.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1699.0, + 1404.0, + 1699.0, + 1404.0, + 1734.0, + 295.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1729.0, + 1406.0, + 1729.0, + 1406.0, + 1765.0, + 291.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1758.0, + 1405.0, + 1758.0, + 1405.0, + 1797.0, + 292.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1788.0, + 1406.0, + 1788.0, + 1406.0, + 1824.0, + 292.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1854.0, + 292.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1402.0, + 1910.0, + 1402.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 774.0, + 2003.0, + 774.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1435.0, + 1407.0, + 1435.0, + 1407.0, + 1474.0, + 292.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1469.0, + 1405.0, + 1469.0, + 1405.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1502.0, + 1401.0, + 1502.0, + 1401.0, + 1532.0, + 296.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1528.0, + 1407.0, + 1528.0, + 1407.0, + 1566.0, + 293.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1560.0, + 1406.0, + 1560.0, + 1406.0, + 1594.0, + 292.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1588.0, + 1126.0, + 1588.0, + 1126.0, + 1625.0, + 292.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1588.0, + 1406.0, + 1588.0, + 1406.0, + 1625.0, + 1159.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1622.0, + 1405.0, + 1622.0, + 1405.0, + 1653.0, + 293.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1652.0, + 1089.0, + 1652.0, + 1089.0, + 1685.0, + 293.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 296.0, + 1402.0, + 296.0, + 1402.0, + 328.0, + 296.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 324.0, + 515.0, + 324.0, + 515.0, + 361.0, + 292.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 324.0, + 1404.0, + 324.0, + 1404.0, + 361.0, + 588.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 356.0, + 1405.0, + 356.0, + 1405.0, + 391.0, + 295.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 385.0, + 815.0, + 385.0, + 815.0, + 421.0, + 292.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 385.0, + 1155.0, + 385.0, + 1155.0, + 421.0, + 1119.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 385.0, + 1404.0, + 385.0, + 1404.0, + 421.0, + 1394.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 414.0, + 371.0, + 414.0, + 371.0, + 452.0, + 294.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 414.0, + 1404.0, + 414.0, + 1404.0, + 452.0, + 393.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 447.0, + 493.0, + 447.0, + 493.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1080.0, + 1405.0, + 1080.0, + 1405.0, + 1113.0, + 295.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1108.0, + 1407.0, + 1108.0, + 1407.0, + 1146.0, + 292.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1142.0, + 1403.0, + 1142.0, + 1403.0, + 1175.0, + 293.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1173.0, + 1219.0, + 1173.0, + 1219.0, + 1206.0, + 296.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 1405.0, + 1224.0, + 1405.0, + 1262.0, + 294.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1256.0, + 772.0, + 1256.0, + 772.0, + 1292.0, + 294.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1292.0, + 844.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1286.0, + 808.0, + 1286.0, + 808.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 106, + 298, + 106 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 262, + 300, + 262 + ], + "score": 0.818 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 293, + 152, + 1408, + 152, + 1408, + 2041, + 293, + 2041 + ], + "score": 0.594 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 109.0, + 298.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 494.0, + 228.0, + 494.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 330.0, + 294.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 324.0, + 916.0, + 324.0, + 916.0, + 359.0, + 317.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 373.0, + 1407.0, + 373.0, + 1407.0, + 415.0, + 294.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 402.0, + 1091.0, + 402.0, + 1091.0, + 450.0, + 315.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 456.0, + 1407.0, + 456.0, + 1407.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 489.0, + 1241.0, + 489.0, + 1241.0, + 527.0, + 317.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 1405.0, + 543.0, + 1405.0, + 583.0, + 294.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 572.0, + 1405.0, + 572.0, + 1405.0, + 614.0, + 317.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 604.0, + 555.0, + 604.0, + 555.0, + 638.0, + 317.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 657.0, + 1405.0, + 657.0, + 1405.0, + 697.0, + 294.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 688.0, + 1042.0, + 688.0, + 1042.0, + 721.0, + 321.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 738.0, + 1407.0, + 738.0, + 1407.0, + 780.0, + 292.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 773.0, + 751.0, + 773.0, + 751.0, + 806.0, + 321.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 858.0, + 296.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 854.0, + 1334.0, + 854.0, + 1334.0, + 893.0, + 319.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 906.0, + 1407.0, + 906.0, + 1407.0, + 947.0, + 292.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 941.0, + 553.0, + 941.0, + 553.0, + 974.0, + 319.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1407.0, + 991.0, + 1407.0, + 1030.0, + 294.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1022.0, + 964.0, + 1022.0, + 964.0, + 1061.0, + 319.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1117.0, + 294.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1107.0, + 1108.0, + 1107.0, + 1108.0, + 1144.0, + 317.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1188.0, + 717.0, + 1188.0, + 717.0, + 1227.0, + 317.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 1405.0, + 1239.0, + 1405.0, + 1279.0, + 294.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1277.0, + 690.0, + 1277.0, + 690.0, + 1310.0, + 321.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1322.0, + 1405.0, + 1322.0, + 1405.0, + 1366.0, + 292.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1353.0, + 1403.0, + 1353.0, + 1403.0, + 1393.0, + 319.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1387.0, + 705.0, + 1387.0, + 705.0, + 1424.0, + 317.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1438.0, + 1407.0, + 1438.0, + 1407.0, + 1478.0, + 294.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1469.0, + 1089.0, + 1469.0, + 1089.0, + 1509.0, + 319.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1519.0, + 1407.0, + 1519.0, + 1407.0, + 1563.0, + 292.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1554.0, + 1129.0, + 1554.0, + 1129.0, + 1592.0, + 317.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1606.0, + 1407.0, + 1606.0, + 1407.0, + 1646.0, + 294.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1639.0, + 538.0, + 1639.0, + 538.0, + 1672.0, + 321.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1691.0, + 1407.0, + 1691.0, + 1407.0, + 1730.0, + 294.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1722.0, + 1082.0, + 1722.0, + 1082.0, + 1757.0, + 317.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1776.0, + 1405.0, + 1776.0, + 1405.0, + 1809.0, + 296.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1805.0, + 1099.0, + 1805.0, + 1099.0, + 1844.0, + 319.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1857.0, + 1403.0, + 1857.0, + 1403.0, + 1896.0, + 294.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1888.0, + 1403.0, + 1888.0, + 1403.0, + 1927.0, + 319.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1923.0, + 608.0, + 1923.0, + 608.0, + 1956.0, + 319.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1967.0, + 1407.0, + 1967.0, + 1407.0, + 2014.0, + 289.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 2002.0, + 1258.0, + 2002.0, + 1258.0, + 2041.0, + 319.0, + 2041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.845 + }, + { + "category_id": 1, + "poly": [ + 294, + 134, + 1408, + 134, + 1408, + 1808, + 294, + 1808 + ], + "score": 0.635 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 232.0, + 1403.0, + 232.0, + 1403.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 263.0, + 1405.0, + 263.0, + 1405.0, + 298.0, + 322.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 293.0, + 1113.0, + 293.0, + 1113.0, + 328.0, + 322.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 341.0, + 1409.0, + 341.0, + 1409.0, + 383.0, + 290.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 374.0, + 1099.0, + 374.0, + 1099.0, + 413.0, + 318.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 426.0, + 1409.0, + 426.0, + 1409.0, + 467.0, + 291.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 457.0, + 1018.0, + 457.0, + 1018.0, + 496.0, + 320.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 511.0, + 1407.0, + 511.0, + 1407.0, + 546.0, + 295.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 542.0, + 1405.0, + 542.0, + 1405.0, + 577.0, + 324.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 570.0, + 651.0, + 570.0, + 651.0, + 609.0, + 316.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 622.0, + 1405.0, + 622.0, + 1405.0, + 662.0, + 293.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 653.0, + 1407.0, + 653.0, + 1407.0, + 694.0, + 320.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 685.0, + 553.0, + 685.0, + 553.0, + 720.0, + 322.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 736.0, + 1407.0, + 736.0, + 1407.0, + 777.0, + 293.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 766.0, + 822.0, + 766.0, + 822.0, + 806.0, + 318.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 818.0, + 1407.0, + 818.0, + 1407.0, + 860.0, + 293.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 849.0, + 978.0, + 849.0, + 978.0, + 891.0, + 316.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 943.0, + 291.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 934.0, + 419.0, + 934.0, + 419.0, + 971.0, + 320.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 982.0, + 1403.0, + 982.0, + 1403.0, + 1028.0, + 292.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1015.0, + 842.0, + 1015.0, + 842.0, + 1056.0, + 320.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1067.0, + 1403.0, + 1067.0, + 1403.0, + 1110.0, + 291.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1099.0, + 1003.0, + 1099.0, + 1003.0, + 1137.0, + 318.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1150.0, + 1405.0, + 1150.0, + 1405.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1182.0, + 1244.0, + 1182.0, + 1244.0, + 1222.0, + 320.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1232.0, + 1407.0, + 1232.0, + 1407.0, + 1276.0, + 293.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1265.0, + 1118.0, + 1265.0, + 1118.0, + 1304.0, + 316.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1318.0, + 1405.0, + 1318.0, + 1405.0, + 1359.0, + 293.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1348.0, + 714.0, + 1348.0, + 714.0, + 1387.0, + 318.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1409.0, + 1396.0, + 1409.0, + 1444.0, + 293.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1433.0, + 648.0, + 1433.0, + 648.0, + 1468.0, + 316.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1477.0, + 1407.0, + 1477.0, + 1407.0, + 1527.0, + 290.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1514.0, + 1204.0, + 1514.0, + 1204.0, + 1555.0, + 322.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1564.0, + 1407.0, + 1564.0, + 1407.0, + 1607.0, + 291.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1597.0, + 1236.0, + 1597.0, + 1236.0, + 1638.0, + 320.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1647.0, + 1407.0, + 1647.0, + 1407.0, + 1690.0, + 291.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1679.0, + 1109.0, + 1679.0, + 1109.0, + 1721.0, + 316.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1732.0, + 1407.0, + 1732.0, + 1407.0, + 1773.0, + 291.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1762.0, + 834.0, + 1762.0, + 834.0, + 1801.0, + 318.0, + 1801.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 482, + 1406, + 482, + 1406, + 636, + 297, + 636 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1724, + 1404, + 1724, + 1404, + 1841, + 297, + 1841 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 297, + 291, + 1405, + 291, + 1405, + 385, + 297, + 385 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 1617, + 1404, + 1617, + 1404, + 1712, + 297, + 1712 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1941, + 1403, + 1941, + 1403, + 2036, + 298, + 2036 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 429, + 1380, + 1271, + 1380, + 1271, + 1457, + 429, + 1457 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 599, + 692, + 1101, + 692, + 1101, + 786, + 599, + 786 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 476, + 879, + 1221, + 879, + 1221, + 966, + 476, + 966 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 546, + 1204, + 1151, + 1204, + 1151, + 1274, + 546, + 1274 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 299, + 395, + 1398, + 395, + 1398, + 471, + 299, + 471 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 525, + 1056, + 1176, + 1056, + 1176, + 1130, + 525, + 1130 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 296, + 800, + 1402, + 800, + 1402, + 866, + 296, + 866 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 299, + 1854, + 1399, + 1854, + 1399, + 1934, + 299, + 1934 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 298, + 1156, + 1079, + 1156, + 1079, + 1190, + 298, + 1190 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 298, + 1332, + 1123, + 1332, + 1123, + 1366, + 298, + 1366 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 1490, + 760, + 1490, + 760, + 1527, + 299, + 1527 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 299, + 648, + 630, + 648, + 630, + 679, + 299, + 679 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 297, + 1284, + 750, + 1284, + 750, + 1320, + 297, + 1320 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 1558, + 665, + 1558, + 665, + 1591, + 299, + 1591 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1338, + 723, + 1401, + 723, + 1401, + 755, + 1338, + 755 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 299, + 227, + 880, + 227, + 880, + 262, + 299, + 262 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1337, + 1222, + 1401, + 1222, + 1401, + 1255, + 1337, + 1255 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1337, + 1075, + 1401, + 1075, + 1401, + 1107, + 1337, + 1107 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 298, + 978, + 1399, + 978, + 1399, + 1043, + 298, + 1043 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 863, + 2087, + 863, + 2113, + 836, + 2113 + ], + "score": 0.856 + }, + { + "category_id": 14, + "poly": [ + 478, + 877, + 1222, + 877, + 1222, + 966, + 478, + 966 + ], + "score": 0.94, + "latex": "P _ { k } ( t ) = \\left( - { \\frac { 1 } { 2 } } \\right) ^ { k } { \\frac { \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( k + { \\frac { d } { 2 } } \\right) } } \\left( 1 - t ^ { 2 } \\right) ^ { { \\frac { 2 - d } { 2 } } } \\left( { \\frac { d } { d t } } \\right) ^ { k } \\left( 1 - t ^ { 2 } \\right) ^ { k + { \\frac { d - 2 } { 2 } } } ." + }, + { + "category_id": 13, + "poly": [ + 622, + 397, + 961, + 397, + 961, + 436, + 622, + 436 + ], + "score": 0.94, + "latex": "\\mathbb { S } ^ { d } = \\left\\{ \\mathbf { x } \\in \\mathbb { R } ^ { d + 1 } : \\| \\mathbf { x } \\| _ { 2 } = 1 \\right\\}" + }, + { + "category_id": 14, + "poly": [ + 522, + 1054, + 1175, + 1054, + 1175, + 1129, + 522, + 1129 + ], + "score": 0.94, + "latex": "\\int _ { \\mathbb { S } ^ { d } } P _ { j } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) P _ { k } ( \\langle \\mathbf { w } , \\mathbf { y } \\rangle ) d \\tau _ { d } ( \\mathbf { w } ) = \\frac { \\delta _ { j k } } { N ( d , k ) } P _ { k } ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) ." + }, + { + "category_id": 14, + "poly": [ + 595, + 688, + 1100, + 688, + 1100, + 789, + 595, + 789 + ], + "score": 0.94, + "latex": "\\sum _ { j = 1 } ^ { N ( d , k ) } Y _ { k , j } ( { \\bf x } ) Y _ { k , j } ( { \\bf y } ) = N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) ," + }, + { + "category_id": 13, + "poly": [ + 457, + 1285, + 629, + 1285, + 629, + 1320, + 457, + 1320 + ], + "score": 0.93, + "latex": "t P _ { 0 } ( t ) = P _ { 1 } ( t )" + }, + { + "category_id": 13, + "poly": [ + 530, + 979, + 594, + 979, + 594, + 1013, + 530, + 1013 + ], + "score": 0.92, + "latex": "P _ { k } ( t )" + }, + { + "category_id": 13, + "poly": [ + 960, + 603, + 1028, + 603, + 1028, + 635, + 960, + 635 + ], + "score": 0.92, + "latex": "Y _ { k } ( x )" + }, + { + "category_id": 14, + "poly": [ + 425, + 1376, + 1270, + 1376, + 1270, + 1459, + 425, + 1459 + ], + "score": 0.92, + "latex": "\\int _ { \\mathbb { S } ^ { d } } f ( \\langle \\mathbf { x } , \\mathbf { y } \\rangle ) Y _ { k } ( \\mathbf { y } ) d t ^ { \\prime } d ( \\mathbf { y } ) = { \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } } Y _ { k } ( \\mathbf { x } ) \\int _ { - 1 } ^ { 1 } f ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t ." + }, + { + "category_id": 14, + "poly": [ + 547, + 1201, + 1150, + 1201, + 1150, + 1273, + 547, + 1273 + ], + "score": 0.92, + "latex": "t P _ { k } ( t ) = \\frac { k } { 2 k + d - 1 } P _ { k - 1 } ( t ) + \\frac { k + d - 1 } { 2 k + d - 1 } P _ { k + 1 } ( t ) ," + }, + { + "category_id": 13, + "poly": [ + 297, + 433, + 567, + 433, + 567, + 472, + 297, + 472 + ], + "score": 0.91, + "latex": "2 \\pi ^ { ( d + 1 ) / 2 } / \\Gamma ( ( d + 1 ) / 2 )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1763, + 459, + 1763, + 459, + 1800, + 297, + 1800 + ], + "score": 0.91, + "latex": "\\mathrm { l i m } _ { m \\infty } \\mathbf { K } ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 373, + 802, + 438, + 802, + 438, + 836, + 373, + 836 + ], + "score": 0.91, + "latex": "P _ { k } ( t )" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1767, + 1365, + 1767, + 1365, + 1800, + 1239, + 1800 + ], + "score": 0.91, + "latex": "\\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 920, + 803, + 986, + 803, + 986, + 832, + 920, + 832 + ], + "score": 0.9, + "latex": "d + 1" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1766, + 1174, + 1766, + 1174, + 1796, + 1080, + 1796 + ], + "score": 0.9, + "latex": "n ^ { - 1 } \\mathbf { K } ^ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 337, + 1287, + 406, + 1287, + 406, + 1316, + 337, + 1316 + ], + "score": 0.89, + "latex": "k \\geqslant 1" + }, + { + "category_id": 13, + "poly": [ + 673, + 1287, + 742, + 1287, + 742, + 1314, + 673, + 1314 + ], + "score": 0.89, + "latex": "k = 0" + }, + { + "category_id": 13, + "poly": [ + 954, + 1334, + 987, + 1334, + 987, + 1364, + 954, + 1364 + ], + "score": 0.88, + "latex": "Y _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1034, + 1725, + 1400, + 1725, + 1400, + 1765, + 1034, + 1765 + ], + "score": 0.87, + "latex": "\\begin{array} { r l r } { { \\bf K } ^ { ( \\infty ) } } & { { } = } & { ( \\kappa ( { \\bf x } _ { i } , { \\bf x } _ { j } ) ) _ { n \\times n } \\quad = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 640, + 567, + 1195, + 567, + 1195, + 604, + 640, + 604 + ], + "score": 0.87, + "latex": "\\langle Y _ { k i } , Y _ { s j } \\rangle _ { \\mathbb { S } ^ { d } } = \\int _ { \\mathbb { S } ^ { d } } Y _ { k i } ( x ) Y _ { s j } ( x ) d \\tau _ { d } ( x ) = \\delta _ { i j } \\delta _ { s k }" + }, + { + "category_id": 13, + "poly": [ + 700, + 442, + 729, + 442, + 729, + 468, + 700, + 468 + ], + "score": 0.86, + "latex": "\\tau _ { d }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1891, + 674, + 1891, + 674, + 1934, + 297, + 1934 + ], + "score": 0.85, + "latex": "\\left[ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , l = 1 , 2 ." + }, + { + "category_id": 13, + "poly": [ + 389, + 1725, + 1012, + 1725, + 1012, + 1762, + 389, + 1762 + ], + "score": 0.85, + "latex": "\\begin{array} { r l r } { \\mathbf { K } ^ { ( 0 ) } } & { { } = } & { m ^ { - 1 } ( \\left. \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\right. ) _ { n \\times n } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 867, + 804, + 887, + 804, + 887, + 830, + 867, + 830 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1334, + 404, + 1403, + 404, + 1403, + 433, + 1334, + 433 + ], + "score": 0.83, + "latex": "\\omega _ { d } =" + }, + { + "category_id": 13, + "poly": [ + 1002, + 981, + 1020, + 981, + 1020, + 1007, + 1002, + 1007 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 530, + 1763, + 823, + 1763, + 823, + 1802, + 530, + 1802 + ], + "score": 0.82, + "latex": "\\{ \\widehat \\lambda _ { i } \\} _ { i = 1 } ^ { n } , \\widehat \\lambda _ { 1 } \\geqslant \\ \\cdot \\cdot \\ \\geqslant \\ \\widehat \\lambda _ { n }" + }, + { + "category_id": 13, + "poly": [ + 297, + 516, + 910, + 516, + 910, + 570, + 297, + 570 + ], + "score": 0.82, + "latex": "\\begin{array} { r } { \\{ Y _ { k , j } : \\mathbb { S } ^ { d } \\mathbb { R } | 1 \\leqslant j \\leqslant N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1283, + 603, + 1300, + 603, + 1300, + 629, + 1283, + 629 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1334, + 982, + 1352, + 982, + 1352, + 1007, + 1334, + 1007 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 773, + 605, + 791, + 605, + 791, + 629, + 773, + 629 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1102, + 1335, + 1120, + 1335, + 1120, + 1361, + 1102, + 1361 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 700, + 1801, + 937, + 1801, + 937, + 1840, + 700, + 1840 + ], + "score": 0.76, + "latex": "\\hat { \\mathbf { V } } _ { r _ { k } } = ( \\hat { \\mathbf { v } } _ { 1 } , \\hdots , \\hat { \\mathbf { v } } _ { r _ { k } } )" + }, + { + "category_id": 13, + "poly": [ + 698, + 1854, + 1177, + 1854, + 1177, + 1896, + 698, + 1896 + ], + "score": 0.72, + "latex": "\\nabla _ { \\bf W } f _ { \\bf W ^ { ( 0 ) } } \\left( { \\bf x } _ { i } \\right) \\ = \\ \\left[ \\nabla _ { \\bf W } f _ { \\bf W } \\left( { \\bf x } _ { i } \\right) \\right] \\big | _ { \\bf W = \\bf W ^ { ( 0 ) } }" + }, + { + "category_id": 13, + "poly": [ + 698, + 1799, + 1222, + 1799, + 1222, + 1841, + 698, + 1841 + ], + "score": 0.72, + "latex": "\\widehat { \\mathbf { V } } _ { r _ { k } } = ( \\widehat { \\mathbf { v } } _ { 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { r _ { k } } ) , \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } = ( \\widehat { \\mathbf { v } } _ { r _ { k } + 1 } , \\ldots , \\widehat { \\mathbf { v } } _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 461, + 484, + 673, + 484, + 673, + 514, + 461, + 514 + ], + "score": 0.57, + "latex": "k \\qquad \\geqslant \\qquad 1 ," + }, + { + "category_id": 13, + "poly": [ + 529, + 1762, + 617, + 1762, + 617, + 1803, + 529, + 1803 + ], + "score": 0.45, + "latex": "\\{ \\widehat { \\lambda } _ { i } \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 461, + 485, + 483, + 485, + 483, + 513, + 461, + 513 + ], + "score": 0.38, + "latex": "k" + }, + { + "category_id": 14, + "poly": [ + 698, + 1853, + 1401, + 1853, + 1401, + 1896, + 698, + 1896 + ], + "score": 0.33, + "latex": "\\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ = \\ \\left[ \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\right] \\big | _ { \\mathbf { W } = \\mathbf { W } ^ { ( 0 ) } } , \\ \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\ =" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1486.0, + 763.0, + 1486.0, + 763.0, + 1533.0, + 292.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1555.0, + 667.0, + 1555.0, + 667.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 880.0, + 225.0, + 880.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 833.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 460.0, + 478.0, + 460.0, + 521.0, + 294.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 481.0, + 684.0, + 481.0, + 684.0, + 521.0, + 674.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 485.0, + 770.0, + 485.0, + 770.0, + 518.0, + 722.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 482.0, + 907.0, + 482.0, + 907.0, + 517.0, + 797.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 485.0, + 966.0, + 485.0, + 966.0, + 515.0, + 933.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 481.0, + 1406.0, + 481.0, + 1406.0, + 518.0, + 985.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 513.0, + 296.0, + 513.0, + 296.0, + 570.0, + 292.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 522.0, + 1404.0, + 522.0, + 1404.0, + 559.0, + 945.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 562.0, + 639.0, + 562.0, + 639.0, + 608.0, + 291.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 562.0, + 1406.0, + 562.0, + 1406.0, + 608.0, + 1196.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 600.0, + 772.0, + 600.0, + 772.0, + 637.0, + 295.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 600.0, + 959.0, + 600.0, + 959.0, + 637.0, + 792.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 600.0, + 1282.0, + 600.0, + 1282.0, + 637.0, + 1029.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 600.0, + 1312.0, + 600.0, + 1312.0, + 637.0, + 1301.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1712.0, + 388.0, + 1712.0, + 388.0, + 1772.0, + 289.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1712.0, + 1033.0, + 1712.0, + 1033.0, + 1772.0, + 1013.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1712.0, + 1409.0, + 1712.0, + 1409.0, + 1772.0, + 1401.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1757.0, + 296.0, + 1757.0, + 296.0, + 1807.0, + 292.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 1757.0, + 528.0, + 1757.0, + 528.0, + 1807.0, + 460.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1757.0, + 1079.0, + 1757.0, + 1079.0, + 1807.0, + 824.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1757.0, + 1238.0, + 1757.0, + 1238.0, + 1807.0, + 1175.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 1757.0, + 1407.0, + 1757.0, + 1407.0, + 1807.0, + 1366.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1797.0, + 697.0, + 1797.0, + 697.0, + 1845.0, + 294.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 1797.0, + 1236.0, + 1797.0, + 1236.0, + 1845.0, + 1223.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1403.0, + 291.0, + 1403.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1403.0, + 321.0, + 1403.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 350.0, + 464.0, + 350.0, + 464.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1616.0, + 1404.0, + 1616.0, + 1404.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1645.0, + 1405.0, + 1645.0, + 1405.0, + 1688.0, + 292.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 940.0, + 1678.0, + 940.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1403.0, + 1941.0, + 1403.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 585.0, + 2001.0, + 585.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 389.0, + 621.0, + 389.0, + 621.0, + 443.0, + 293.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 389.0, + 1333.0, + 389.0, + 1333.0, + 443.0, + 962.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 429.0, + 296.0, + 429.0, + 296.0, + 476.0, + 293.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 429.0, + 699.0, + 429.0, + 699.0, + 476.0, + 568.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 429.0, + 1130.0, + 429.0, + 1130.0, + 476.0, + 730.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 797.0, + 372.0, + 797.0, + 372.0, + 840.0, + 293.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 797.0, + 866.0, + 797.0, + 866.0, + 840.0, + 439.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 797.0, + 919.0, + 797.0, + 919.0, + 840.0, + 888.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 797.0, + 1403.0, + 797.0, + 1403.0, + 840.0, + 987.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 832.0, + 500.0, + 832.0, + 500.0, + 869.0, + 295.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1849.0, + 697.0, + 1849.0, + 697.0, + 1901.0, + 291.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1849.0, + 1407.0, + 1849.0, + 1407.0, + 1901.0, + 1402.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1886.0, + 296.0, + 1886.0, + 296.0, + 1940.0, + 291.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1886.0, + 686.0, + 1886.0, + 686.0, + 1940.0, + 675.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1152.0, + 1081.0, + 1152.0, + 1081.0, + 1195.0, + 293.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 953.0, + 1328.0, + 953.0, + 1370.0, + 294.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1328.0, + 1101.0, + 1328.0, + 1101.0, + 1370.0, + 988.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1328.0, + 1124.0, + 1328.0, + 1124.0, + 1370.0, + 1121.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 646.0, + 631.0, + 646.0, + 631.0, + 682.0, + 297.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1284.0, + 336.0, + 1284.0, + 336.0, + 1321.0, + 296.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1284.0, + 456.0, + 1284.0, + 456.0, + 1321.0, + 407.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1284.0, + 672.0, + 1284.0, + 672.0, + 1321.0, + 630.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1284.0, + 751.0, + 1284.0, + 751.0, + 1321.0, + 743.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 974.0, + 529.0, + 974.0, + 529.0, + 1017.0, + 294.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 974.0, + 1001.0, + 974.0, + 1001.0, + 1017.0, + 595.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 974.0, + 1333.0, + 974.0, + 1333.0, + 1017.0, + 1021.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 974.0, + 1404.0, + 974.0, + 1404.0, + 1017.0, + 1353.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 945.0, + 1007.0, + 945.0, + 1048.0, + 294.0, + 1048.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 836, + 1402, + 836, + 1402, + 962, + 299, + 962 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 385, + 1405, + 385, + 1405, + 480, + 297, + 480 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 296, + 494, + 1405, + 494, + 1405, + 593, + 296, + 593 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 294, + 1216, + 1405, + 1216, + 1405, + 1328, + 294, + 1328 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 1814, + 1405, + 1814, + 1405, + 1920, + 297, + 1920 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 386, + 623, + 1310, + 623, + 1310, + 794, + 386, + 794 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1929, + 1404, + 1929, + 1404, + 2037, + 298, + 2037 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 294, + 978, + 1400, + 978, + 1400, + 1050, + 294, + 1050 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 676, + 294, + 1022, + 294, + 1022, + 341, + 676, + 341 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 401, + 1505, + 1201, + 1505, + 1201, + 1596, + 401, + 1596 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 295, + 227, + 967, + 227, + 967, + 264, + 295, + 264 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 298, + 1628, + 569, + 1628, + 569, + 1662, + 298, + 1662 + ], + "score": 0.911 + }, + { + "category_id": 1, + "poly": [ + 298, + 1696, + 752, + 1696, + 752, + 1729, + 298, + 1729 + ], + "score": 0.908 + }, + { + "category_id": 8, + "poly": [ + 309, + 1078, + 1378, + 1078, + 1378, + 1127, + 309, + 1127 + ], + "score": 0.903 + }, + { + "category_id": 8, + "poly": [ + 300, + 1358, + 1255, + 1358, + 1255, + 1403, + 300, + 1403 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1367, + 1401, + 1367, + 1401, + 1399, + 1340, + 1399 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1464, + 1401, + 1464, + 1401, + 1495, + 1340, + 1495 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1559, + 1401, + 1559, + 1401, + 1590, + 1340, + 1590 + ], + "score": 0.887 + }, + { + "category_id": 8, + "poly": [ + 326, + 1408, + 1332, + 1408, + 1332, + 1452, + 326, + 1452 + ], + "score": 0.87 + }, + { + "category_id": 8, + "poly": [ + 574, + 1456, + 1033, + 1456, + 1033, + 1500, + 574, + 1500 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.845 + }, + { + "category_id": 1, + "poly": [ + 293, + 1157, + 1353, + 1157, + 1353, + 1202, + 293, + 1202 + ], + "score": 0.661 + }, + { + "category_id": 8, + "poly": [ + 293, + 1157, + 1353, + 1157, + 1353, + 1202, + 293, + 1202 + ], + "score": 0.23 + }, + { + "category_id": 14, + "poly": [ + 387, + 622, + 1313, + 622, + 1313, + 796, + 387, + 796 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\| _ { F } \\leqslant C ^ { \\prime } \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } , } \\\\ & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\widehat { \\mathbf { V } } _ { r _ { k } } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant C ^ { \\prime \\prime } \\bigg [ \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M ^ { 2 } r _ { k } \\sqrt { \\frac { \\log ( r _ { k } / \\delta ) } { n } } \\bigg ] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1089, + 1931, + 1224, + 1931, + 1224, + 1968, + 1089, + 1968 + ], + "score": 0.93, + "latex": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1251, + 434, + 1251, + 434, + 1289, + 298, + 1289 + ], + "score": 0.93, + "latex": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1164, + 561, + 1164, + 561, + 1197, + 372, + 1197 + ], + "score": 0.93, + "latex": "t ^ { \\prime } = 0 , \\ldots , t - 1" + }, + { + "category_id": 13, + "poly": [ + 1005, + 1251, + 1152, + 1251, + 1152, + 1292, + 1005, + 1292 + ], + "score": 0.93, + "latex": "n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )" + }, + { + "category_id": 13, + "poly": [ + 473, + 1012, + 721, + 1012, + 721, + 1049, + 473, + 1049 + ], + "score": 0.93, + "latex": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )" + }, + { + "category_id": 14, + "poly": [ + 310, + 1352, + 1428, + 1352, + 1428, + 1601, + 310, + 1601 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { { ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 1 ) } } \\\\ & { } & { \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) \\| _ { 2 } \\leqslant \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) t ^ { \\prime } \\| \\widehat { \\mathbf { V } } _ { r k } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r k } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { n } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } \\cdot \\widetilde { O } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r k } ^ { \\perp } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { } & { \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } ) \\cdot \\big ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r k } / 2 \\big ) ^ { t ^ { \\prime } } + \\widetilde { \\mathcal { O } } ( \\sqrt { n } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r k } ) ^ { - 1 } ) } \\\\ & { } & { + \\lambda _ { r k } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) \\qquad ( \\mathbb { B } . 3 ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1999, + 434, + 1999, + 434, + 2036, + 298, + 2036 + ], + "score": 0.93, + "latex": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 606, + 498, + 755, + 498, + 755, + 531, + 606, + 531 + ], + "score": 0.92, + "latex": "| \\phi _ { i } ( \\mathbf { x } ) | \\leqslant M" + }, + { + "category_id": 13, + "poly": [ + 912, + 528, + 1283, + 528, + 1283, + 564, + 912, + 564 + ], + "score": 0.92, + "latex": "n \\geqslant C ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 2 } \\log ( 1 / \\delta )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1012, + 434, + 1012, + 434, + 1049, + 298, + 1049 + ], + "score": 0.91, + "latex": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 371, + 1629, + 561, + 1629, + 561, + 1662, + 371, + 1662 + ], + "score": 0.91, + "latex": "t ^ { \\prime } = 0 , \\ldots , t - 1" + }, + { + "category_id": 13, + "poly": [ + 475, + 1250, + 952, + 1250, + 952, + 1291, + 475, + 1291 + ], + "score": 0.91, + "latex": "\\omega \\leqslant \\widetilde { \\mathcal { O } } ( \\operatorname* { m i n } \\{ [ \\log ( m ) ] ^ { - 3 / 2 } , \\lambda _ { r _ { k } } ^ { 3 } , ( \\eta m ) ^ { - 3 } \\} )" + }, + { + "category_id": 13, + "poly": [ + 680, + 1853, + 991, + 1853, + 991, + 1892, + 680, + 1892 + ], + "score": 0.91, + "latex": "m \\geqslant \\widetilde \\Omega ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )" + }, + { + "category_id": 13, + "poly": [ + 989, + 977, + 1177, + 977, + 1177, + 1013, + 989, + 1013 + ], + "score": 0.91, + "latex": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 639, + 1817, + 876, + 1817, + 876, + 1852, + 639, + 1852 + ], + "score": 0.91, + "latex": "\\omega \\ = \\ \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )" + }, + { + "category_id": 13, + "poly": [ + 989, + 1215, + 1176, + 1215, + 1176, + 1250, + 989, + 1250 + ], + "score": 0.91, + "latex": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 833, + 496, + 915, + 496, + 915, + 527, + 833, + 527 + ], + "score": 0.91, + "latex": "\\mathbf { x } \\in S ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 675, + 293, + 1022, + 293, + 1022, + 340, + 675, + 340 + ], + "score": 0.91, + "latex": "| \\lambda _ { i } - \\widehat { \\lambda } _ { i } | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( 1 / \\delta ) / n } ) ." + }, + { + "category_id": 13, + "poly": [ + 1057, + 1012, + 1393, + 1012, + 1393, + 1050, + 1057, + 1050 + ], + "score": 0.9, + "latex": "1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 678, + 1931, + 871, + 1931, + 871, + 1965, + 678, + 1965 + ], + "score": 0.9, + "latex": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 498, + 1403, + 498, + 1403, + 530, + 1269, + 530 + ], + "score": 0.9, + "latex": "C , C ^ { \\prime } , c ^ { \\prime \\prime } >" + }, + { + "category_id": 13, + "poly": [ + 601, + 2001, + 957, + 2001, + 957, + 2036, + 601, + 2036 + ], + "score": 0.9, + "latex": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 547, + 231, + 613, + 231, + 613, + 259, + 547, + 259 + ], + "score": 0.9, + "latex": "\\delta > 0" + }, + { + "category_id": 13, + "poly": [ + 795, + 532, + 878, + 532, + 878, + 560, + 795, + 560 + ], + "score": 0.9, + "latex": "r _ { k } \\leqslant n" + }, + { + "category_id": 13, + "poly": [ + 382, + 1291, + 746, + 1291, + 746, + 1326, + 382, + 1326 + ], + "score": 0.9, + "latex": "1 - \\mathcal { O } ( t ^ { 2 } n ^ { 2 } ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 517, + 530, + 584, + 530, + 584, + 558, + 517, + 558 + ], + "score": 0.89, + "latex": "\\delta > 0" + }, + { + "category_id": 13, + "poly": [ + 1312, + 1965, + 1405, + 1965, + 1405, + 1998, + 1312, + 1998 + ], + "score": 0.89, + "latex": "{ \\bf W } ^ { ( 0 ) } \\in" + }, + { + "category_id": 13, + "poly": [ + 511, + 561, + 574, + 561, + 574, + 589, + 511, + 589 + ], + "score": 0.86, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 640, + 1966, + 1023, + 1966, + 1023, + 2000, + 640, + 2000 + ], + "score": 0.84, + "latex": "\\mathbf { W } ^ { ( t ) } \\in \\mathcal { B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) , t = 0 , \\ldots , T" + }, + { + "category_id": 14, + "poly": [ + 317, + 1079, + 1380, + 1079, + 1380, + 1125, + 317, + 1125 + ], + "score": 0.83, + "latex": "\\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } + 1 ) } = [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } ) + \\mathbf { e } ^ { ( t ) } , \\ \\| \\mathbf { e } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1369, + 1817, + 1393, + 1817, + 1393, + 1847, + 1369, + 1847 + ], + "score": 0.82, + "latex": "\\overline { { C } }" + }, + { + "category_id": 13, + "poly": [ + 893, + 232, + 957, + 232, + 957, + 260, + 893, + 260 + ], + "score": 0.82, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 717, + 532, + 736, + 532, + 736, + 557, + 717, + 557 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1211, + 1862, + 1233, + 1862, + 1233, + 1884, + 1211, + 1884 + ], + "score": 0.73, + "latex": "\\omega" + }, + { + "category_id": 13, + "poly": [ + 647, + 1158, + 1350, + 1158, + 1350, + 1200, + 647, + 1200 + ], + "score": 0.68, + "latex": "\\mathbf { y } = \\left( y _ { 1 } , \\ldots , y _ { n } \\right) ^ { \\top } , \\widehat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } = \\theta \\cdot \\left( f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { 1 } ) , \\ldots , f _ { \\mathbf { W } ^ { ( t ^ { \\prime } ) } } ( \\mathbf { x } _ { n } ) \\right) ^ { \\top } ." + }, + { + "category_id": 13, + "poly": [ + 647, + 1162, + 854, + 1162, + 854, + 1198, + 647, + 1198 + ], + "score": 0.45, + "latex": "\\mathbf { y } = ( y _ { 1 } , \\ldots , y _ { n } ) ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 389, + 1202, + 389, + 1202, + 416, + 1168, + 416 + ], + "score": 0.32, + "latex": "\\mathrm { S u }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 833.0, + 1403.0, + 833.0, + 1403.0, + 874.0, + 293.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 868.0, + 1404.0, + 868.0, + 1404.0, + 901.0, + 297.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 897.0, + 1407.0, + 897.0, + 1407.0, + 934.0, + 293.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 928.0, + 1333.0, + 928.0, + 1333.0, + 964.0, + 294.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 386.0, + 1167.0, + 386.0, + 1167.0, + 420.0, + 295.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 386.0, + 1400.0, + 386.0, + 1400.0, + 420.0, + 1203.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 417.0, + 1403.0, + 417.0, + 1403.0, + 451.0, + 292.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 450.0, + 450.0, + 450.0, + 485.0, + 294.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 494.0, + 605.0, + 494.0, + 605.0, + 532.0, + 293.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 494.0, + 832.0, + 494.0, + 832.0, + 532.0, + 756.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 494.0, + 1268.0, + 494.0, + 1268.0, + 532.0, + 916.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 525.0, + 516.0, + 525.0, + 516.0, + 566.0, + 290.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 525.0, + 716.0, + 525.0, + 716.0, + 566.0, + 585.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 525.0, + 794.0, + 525.0, + 794.0, + 566.0, + 737.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 525.0, + 911.0, + 525.0, + 911.0, + 566.0, + 879.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 525.0, + 1406.0, + 525.0, + 1406.0, + 566.0, + 1284.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 560.0, + 510.0, + 560.0, + 510.0, + 593.0, + 294.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 560.0, + 588.0, + 560.0, + 588.0, + 593.0, + 575.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1211.0, + 988.0, + 1211.0, + 988.0, + 1253.0, + 291.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1211.0, + 1406.0, + 1211.0, + 1406.0, + 1253.0, + 1177.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1247.0, + 297.0, + 1247.0, + 297.0, + 1296.0, + 292.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1247.0, + 474.0, + 1247.0, + 474.0, + 1296.0, + 435.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1247.0, + 1004.0, + 1247.0, + 1004.0, + 1296.0, + 953.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1247.0, + 1406.0, + 1247.0, + 1406.0, + 1296.0, + 1153.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1281.0, + 381.0, + 1281.0, + 381.0, + 1332.0, + 292.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1281.0, + 759.0, + 1281.0, + 759.0, + 1332.0, + 747.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1813.0, + 638.0, + 1813.0, + 638.0, + 1853.0, + 294.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1813.0, + 1368.0, + 1813.0, + 1368.0, + 1853.0, + 877.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1813.0, + 1403.0, + 1813.0, + 1403.0, + 1853.0, + 1394.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1845.0, + 679.0, + 1845.0, + 679.0, + 1902.0, + 290.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1845.0, + 1210.0, + 1845.0, + 1210.0, + 1902.0, + 992.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1845.0, + 1410.0, + 1845.0, + 1410.0, + 1902.0, + 1234.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1885.0, + 833.0, + 1885.0, + 833.0, + 1920.0, + 292.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1922.0, + 677.0, + 1922.0, + 677.0, + 1973.0, + 291.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1922.0, + 1088.0, + 1922.0, + 1088.0, + 1973.0, + 872.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1922.0, + 1408.0, + 1922.0, + 1408.0, + 1973.0, + 1225.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1962.0, + 639.0, + 1962.0, + 639.0, + 2002.0, + 292.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1962.0, + 1311.0, + 1962.0, + 1311.0, + 2002.0, + 1024.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1993.0, + 297.0, + 1993.0, + 297.0, + 2041.0, + 293.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1993.0, + 600.0, + 1993.0, + 600.0, + 2041.0, + 435.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1993.0, + 1409.0, + 1993.0, + 1409.0, + 2041.0, + 958.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 974.0, + 988.0, + 974.0, + 988.0, + 1017.0, + 292.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 974.0, + 1407.0, + 974.0, + 1407.0, + 1017.0, + 1178.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1009.0, + 297.0, + 1009.0, + 297.0, + 1053.0, + 293.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1009.0, + 472.0, + 1009.0, + 472.0, + 1053.0, + 435.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1009.0, + 1056.0, + 1009.0, + 1056.0, + 1053.0, + 722.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1053.0, + 1394.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 546.0, + 226.0, + 546.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 226.0, + 892.0, + 226.0, + 892.0, + 268.0, + 614.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 226.0, + 969.0, + 226.0, + 969.0, + 268.0, + 958.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1624.0, + 370.0, + 1624.0, + 370.0, + 1667.0, + 294.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1624.0, + 573.0, + 1624.0, + 573.0, + 1667.0, + 562.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1695.0, + 756.0, + 1695.0, + 756.0, + 1731.0, + 296.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1156.0, + 371.0, + 1156.0, + 371.0, + 1206.0, + 290.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1156.0, + 646.0, + 1156.0, + 646.0, + 1206.0, + 562.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1156.0, + 1355.0, + 1156.0, + 1355.0, + 1206.0, + 1351.0, + 1206.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1837, + 1406, + 1837, + 1406, + 1944, + 297, + 1944 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 409, + 1092, + 1287, + 1092, + 1287, + 1237, + 409, + 1237 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 889, + 1405, + 889, + 1405, + 1040, + 297, + 1040 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 401, + 276, + 1298, + 276, + 1298, + 737, + 401, + 737 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 470, + 1733, + 1229, + 1733, + 1229, + 1827, + 470, + 1827 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 295, + 1376, + 1412, + 1376, + 1412, + 1450, + 295, + 1450 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 296, + 1460, + 1492, + 1460, + 1492, + 1688, + 296, + 1688 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 290, + 1246, + 1404, + 1246, + 1404, + 1313, + 290, + 1313 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 476, + 1951, + 1217, + 1951, + 1217, + 1998, + 476, + 1998 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 298, + 2002, + 797, + 2002, + 797, + 2036, + 298, + 2036 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 296, + 1692, + 410, + 1692, + 410, + 1724, + 296, + 1724 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 298, + 1053, + 665, + 1053, + 665, + 1086, + 298, + 1086 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 107, + 297, + 107 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 295, + 744, + 1168, + 744, + 1168, + 780, + 295, + 780 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 289, + 223, + 1302, + 223, + 1302, + 265, + 289, + 265 + ], + "score": 0.922 + }, + { + "category_id": 8, + "poly": [ + 414, + 1318, + 1282, + 1318, + 1282, + 1366, + 414, + 1366 + ], + "score": 0.919 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1326, + 1401, + 1326, + 1401, + 1358, + 1339, + 1358 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 1374, + 2003, + 1403, + 2003, + 1403, + 2032, + 1374, + 2032 + ], + "score": 0.832 + }, + { + "category_id": 8, + "poly": [ + 304, + 798, + 1415, + 798, + 1415, + 877, + 304, + 877 + ], + "score": 0.717 + }, + { + "category_id": 8, + "poly": [ + 297, + 1460, + 1491, + 1460, + 1491, + 1688, + 297, + 1688 + ], + "score": 0.112 + }, + { + "category_id": 14, + "poly": [ + 402, + 275, + 1300, + 275, + 1300, + 741, + 402, + 741 + ], + "score": 0.96, + "latex": "\\begin{array} { r l } { \\| \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( 0 ) } \\| _ { F } \\leqslant } & { \\displaystyle \\sum _ { \\ell = 1 } ^ { t } | \\mathbf { W } _ { i } ^ { ( t + 1 ) } - \\mathbf { W } _ { i } ^ { ( t ^ { \\prime } ) } | \\| _ { F } } \\\\ & { \\qquad = \\displaystyle \\prod _ { \\ell = 0 } ^ { t } \\left| \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\left( g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right) \\cdot \\vartheta \\cdot \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\right| \\Bigg | _ { F } } \\\\ & { \\leqslant \\eta \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| \\cdot \\| \\nabla _ { \\mathbf { W } _ { i } } f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\| g _ { \\ell } - \\vartheta \\cdot f _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\frac { 1 } { n - 1 } \\| \\mathbf { y } - \\vartheta \\cdot \\nabla _ { \\mathbf { W } ^ { ( i ) } } ( \\mathbf { x } _ { i } ) \\| } \\\\ & { \\leqslant C _ { 1 } \\eta \\displaystyle \\theta \\sqrt { \\eta } \\displaystyle \\sum _ { \\ell = 0 } ^ { t } \\| \\mathbf { y } - \\widehat { \\mathbf { g } } ^ { ( \\ell ^ { \\prime } ) } \\| _ { 2 } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 477, + 1374, + 587, + 1374, + 587, + 1414, + 477, + 1414 + ], + "score": 0.93, + "latex": "\\theta = \\widetilde { \\mathcal { O } } ( \\epsilon )" + }, + { + "category_id": 14, + "poly": [ + 310, + 1458, + 1428, + 1458, + 1428, + 1690, + 310, + 1690 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } - \\| ( \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ) ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\sqrt { n } \\cdot \\mathcal { O } \\biggl ( \\frac { 1 } { \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } } \\cdot \\sqrt { \\frac { \\log ( 1 / \\delta ) } { n } } + M r _ { k } \\sqrt { \\frac { \\log ( r _ { k } } { n } } } \\\\ & { \\qquad \\geqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } / \\sqrt { 2 } - \\epsilon \\sqrt { n } / 1 6 . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 434, + 928, + 754, + 928, + 754, + 971, + 434, + 971 + ], + "score": 0.93, + "latex": "m \\geqslant \\tilde { \\mathcal { O } } ( \\mathrm { p o l y } ( T , \\lambda _ { r _ { k } } ^ { - 1 } , \\epsilon ^ { - 1 } ) )" + }, + { + "category_id": 14, + "poly": [ + 409, + 1091, + 1287, + 1091, + 1287, + 1239, + 409, + 1239 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } } \\\\ & { \\qquad + T \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\qquad + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot n ^ { - 1 / 2 } \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 464, + 1414, + 560, + 1414, + 560, + 1452, + 464, + 1452 + ], + "score": 0.92, + "latex": "\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 740, + 1412, + 806, + 1412, + 806, + 1449, + 740, + 1449 + ], + "score": 0.92, + "latex": "1 / \\sqrt { 2 }" + }, + { + "category_id": 14, + "poly": [ + 471, + 1732, + 1229, + 1732, + 1229, + 1830, + 471, + 1830 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 } \\\\ & { \\qquad \\leqslant \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\sqrt { 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } + \\epsilon \\sqrt { n } / 1 6 . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 823, + 1375, + 1349, + 1375, + 1349, + 1415, + 823, + 1415 + ], + "score": 0.91, + "latex": "n \\geqslant \\widetilde \\Omega ( \\operatorname* { m a x } \\{ \\epsilon ^ { - 1 } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } , \\epsilon ^ { - 2 } M ^ { 2 } r _ { k } ^ { 2 } \\} )" + }, + { + "category_id": 13, + "poly": [ + 587, + 226, + 775, + 226, + 775, + 263, + 587, + 263 + ], + "score": 0.91, + "latex": "\\mathbf { W } ^ { ( 0 ) } , \\ldots , \\mathbf { W } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 1087, + 1246, + 1394, + 1246, + 1394, + 1288, + 1087, + 1288 + ], + "score": 0.91, + "latex": "m \\geqslant m ^ { * } = \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 1 4 } \\cdot \\epsilon ^ { - 6 } )" + }, + { + "category_id": 13, + "poly": [ + 675, + 1877, + 1167, + 1877, + 1167, + 1915, + 675, + 1915 + ], + "score": 0.91, + "latex": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { y } } ^ { ( 0 ) } \\| _ { 2 } \\leqslant \\theta \\mathcal { O } ( \\sqrt { n \\log \\left( n / \\delta \\right) } ) \\leqslant \\epsilon \\sqrt { n } / 8" + }, + { + "category_id": 13, + "poly": [ + 830, + 1838, + 1258, + 1838, + 1258, + 1877, + 830, + 1877 + ], + "score": 0.91, + "latex": "\\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\| _ { 2 } \\leqslant 1 + C r _ { k } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1002, + 436, + 1002, + 436, + 1041, + 297, + 1041 + ], + "score": 0.9, + "latex": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 1181, + 889, + 1305, + 889, + 1305, + 929, + 1181, + 929 + ], + "score": 0.9, + "latex": "\\theta \\ : = \\ : \\widetilde { \\mathcal { O } } ( \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 874, + 226, + 1053, + 226, + 1053, + 263, + 874, + 263 + ], + "score": 0.9, + "latex": "\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 857, + 1413, + 896, + 1413, + 896, + 1446, + 857, + 1446 + ], + "score": 0.89, + "latex": "\\sqrt { 2 }" + }, + { + "category_id": 14, + "poly": [ + 418, + 1316, + 1290, + 1316, + 1290, + 1365, + 418, + 1365 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( 0 ) } ) \\| _ { 2 } + \\epsilon / 1 6 . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 481, + 1951, + 1219, + 1951, + 1219, + 1996, + 481, + 1996 + ], + "score": 0.89, + "latex": "n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } ( \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( T ) } ) \\| _ { 2 } \\leqslant 2 ( 1 - \\lambda _ { r _ { k } } ) ^ { T } \\cdot n ^ { - 1 / 2 } \\cdot \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { y } \\| _ { 2 } + \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 910, + 889, + 1163, + 889, + 1163, + 929, + 910, + 929 + ], + "score": 0.88, + "latex": "\\eta ~ = ~ { \\tilde { \\mathcal { O } } } ( ( m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 644, + 893, + 893, + 893, + 893, + 929, + 644, + 929 + ], + "score": 0.88, + "latex": "\\omega \\ = \\ \\overline { { C } } T / ( \\theta \\lambda _ { r _ { k } } \\sqrt { m } )" + }, + { + "category_id": 13, + "poly": [ + 816, + 929, + 1113, + 929, + 1113, + 969, + 816, + 969 + ], + "score": 0.88, + "latex": "\\| \\mathbf { W } _ { l } ^ { ( t + 1 ) } - \\mathbf { W } _ { l } ^ { ( 0 ) } \\| _ { F } \\leqslant \\omega" + }, + { + "category_id": 13, + "poly": [ + 1064, + 230, + 1128, + 230, + 1128, + 260, + 1064, + 260 + ], + "score": 0.85, + "latex": "t \\in T" + }, + { + "category_id": 13, + "poly": [ + 628, + 1246, + 808, + 1246, + 808, + 1286, + 628, + 1286 + ], + "score": 0.84, + "latex": "\\eta = \\widetilde { \\mathcal { O } } ( \\theta ^ { 2 } m ) ^ { - 1 }" + }, + { + "category_id": 14, + "poly": [ + 291, + 788, + 1390, + 788, + 1390, + 876, + 291, + 876 + ], + "score": 0.83, + "latex": "\\sum _ { t ^ { \\prime } = 0 } ^ { t } \\| \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\widetilde { \\mathcal { O } } ( T \\sqrt { n } / ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ) + \\lambda _ { r _ { k } } ^ { - 1 } T ^ { 2 } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } )" + }, + { + "category_id": 13, + "poly": [ + 807, + 970, + 1154, + 970, + 1154, + 1006, + 807, + 1006 + ], + "score": 0.79, + "latex": "1 { - } \\mathcal { O } ( T ^ { 3 } n ^ { 2 } ) { \\cdot } \\exp [ { - \\Omega ( m \\omega ^ { 2 / 3 } ) } ]" + }, + { + "category_id": 13, + "poly": [ + 1170, + 969, + 1403, + 969, + 1403, + 1005, + 1170, + 1005 + ], + "score": 0.74, + "latex": "\\mathbf { W } ( 0 ) , \\ldots , \\mathbf { W } ( T ) \\in" + }, + { + "category_id": 13, + "poly": [ + 394, + 1249, + 617, + 1249, + 617, + 1286, + 394, + 1286 + ], + "score": 0.71, + "latex": "\\omega = \\overline { { C } } T / ( \\lambda _ { r _ { k } } \\sqrt { m } )" + }, + { + "category_id": 13, + "poly": [ + 752, + 1841, + 816, + 1841, + 816, + 1872, + 752, + 1872 + ], + "score": 0.64, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 1130, + 934, + 1229, + 934, + 1229, + 966, + 1130, + 966 + ], + "score": 0.46, + "latex": "l = 1 , 2" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 2007.0, + 1403.0, + 2007.0, + 1403.0, + 2033.0, + 1378.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1832.0, + 751.0, + 1832.0, + 751.0, + 1880.0, + 292.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1832.0, + 829.0, + 1832.0, + 829.0, + 1880.0, + 817.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1832.0, + 1406.0, + 1832.0, + 1406.0, + 1880.0, + 1259.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1874.0, + 674.0, + 1874.0, + 674.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1874.0, + 1407.0, + 1874.0, + 1407.0, + 1918.0, + 1168.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 588.0, + 1909.0, + 588.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 886.0, + 643.0, + 886.0, + 643.0, + 934.0, + 290.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 886.0, + 909.0, + 886.0, + 909.0, + 934.0, + 894.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 886.0, + 1180.0, + 886.0, + 1180.0, + 934.0, + 1164.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 886.0, + 1407.0, + 886.0, + 1407.0, + 934.0, + 1306.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 919.0, + 433.0, + 919.0, + 433.0, + 976.0, + 290.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 919.0, + 815.0, + 919.0, + 815.0, + 976.0, + 755.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 919.0, + 1129.0, + 919.0, + 1129.0, + 976.0, + 1114.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 919.0, + 1410.0, + 919.0, + 1410.0, + 976.0, + 1230.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 966.0, + 806.0, + 966.0, + 806.0, + 1010.0, + 292.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 966.0, + 1169.0, + 966.0, + 1169.0, + 1010.0, + 1155.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 996.0, + 296.0, + 996.0, + 296.0, + 1046.0, + 293.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 996.0, + 456.0, + 996.0, + 456.0, + 1046.0, + 437.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1372.0, + 476.0, + 1372.0, + 476.0, + 1418.0, + 293.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1372.0, + 822.0, + 1372.0, + 822.0, + 1418.0, + 588.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1372.0, + 1407.0, + 1372.0, + 1407.0, + 1418.0, + 1350.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1409.0, + 463.0, + 1409.0, + 463.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1409.0, + 739.0, + 1409.0, + 739.0, + 1455.0, + 561.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1409.0, + 856.0, + 1409.0, + 856.0, + 1455.0, + 807.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1409.0, + 1297.0, + 1409.0, + 1297.0, + 1455.0, + 897.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1242.0, + 393.0, + 1242.0, + 393.0, + 1291.0, + 291.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1242.0, + 627.0, + 1242.0, + 627.0, + 1291.0, + 618.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1242.0, + 1086.0, + 1242.0, + 1086.0, + 1291.0, + 809.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1242.0, + 1406.0, + 1242.0, + 1406.0, + 1291.0, + 1395.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 416.0, + 1280.0, + 416.0, + 1317.0, + 293.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 800.0, + 1999.0, + 800.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1686.0, + 414.0, + 1686.0, + 414.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1050.0, + 666.0, + 1050.0, + 666.0, + 1090.0, + 295.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 743.0, + 1167.0, + 743.0, + 1167.0, + 783.0, + 296.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 216.0, + 586.0, + 216.0, + 586.0, + 271.0, + 288.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 216.0, + 873.0, + 216.0, + 873.0, + 271.0, + 776.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 216.0, + 1063.0, + 216.0, + 1063.0, + 271.0, + 1054.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 216.0, + 1307.0, + 216.0, + 1307.0, + 271.0, + 1129.0, + 271.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 294, + 1610, + 1407, + 1610, + 1407, + 1817, + 294, + 1817 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1313, + 1404, + 1313, + 1404, + 1409, + 297, + 1409 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 287, + 1401, + 287, + 1401, + 386, + 299, + 386 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 502, + 399, + 1197, + 399, + 1197, + 584, + 502, + 584 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 507, + 1868, + 1191, + 1868, + 1191, + 2031, + 507, + 2031 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 434, + 764, + 1264, + 764, + 1264, + 938, + 434, + 938 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 296, + 1142, + 1400, + 1142, + 1400, + 1208, + 296, + 1208 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 446, + 1037, + 1252, + 1037, + 1252, + 1131, + 446, + 1131 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 557, + 1217, + 1139, + 1217, + 1139, + 1289, + 557, + 1289 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 607, + 1514, + 1083, + 1514, + 1083, + 1599, + 607, + 1599 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 296, + 948, + 1407, + 948, + 1407, + 1021, + 296, + 1021 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 607, + 1424, + 1090, + 1424, + 1090, + 1507, + 607, + 1507 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 302, + 657, + 1401, + 657, + 1401, + 751, + 302, + 751 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 298, + 1828, + 843, + 1828, + 843, + 1861, + 298, + 1861 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.92 + }, + { + "category_id": 9, + "poly": [ + 1340, + 524, + 1401, + 524, + 1401, + 556, + 1340, + 556 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 296, + 594, + 1401, + 594, + 1401, + 656, + 296, + 656 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1537, + 1401, + 1537, + 1401, + 1569, + 1339, + 1569 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1978, + 1401, + 1978, + 1401, + 2009, + 1339, + 2009 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 717, + 228, + 717, + 262, + 300, + 262 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1068, + 1401, + 1068, + 1401, + 1100, + 1340, + 1100 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1446, + 1401, + 1446, + 1401, + 1478, + 1340, + 1478 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.871 + }, + { + "category_id": 14, + "poly": [ + 434, + 760, + 1262, + 760, + 1262, + 945, + 434, + 945 + ], + "score": 0.95, + "latex": "\\begin{array} { l } { { \\displaystyle \\kappa ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { \\mathbb { S } ^ { d } } \\widehat { \\kappa } ( \\langle { \\bf y } , { \\bf x } ^ { \\prime } \\rangle ) P _ { k } ( \\langle { \\bf y } , { \\bf x } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } \\ ~ } \\\\ { { \\displaystyle = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\frac { \\omega _ { d - 1 } } { \\omega _ { d } } P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 503, + 397, + 1196, + 397, + 1196, + 590, + 503, + 590 + ], + "score": 0.95, + "latex": "\\begin{array} { l } { { \\displaystyle g ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } g _ { k } ( { \\bf x } ) = \\sum _ { k = 0 } ^ { \\infty } \\sum _ { j = 1 } ^ { N ( d , k ) } \\int _ { { \\mathbb { S } } ^ { d } } Y _ { k j } ( { \\bf y } ) Y _ { k j } ( { \\bf x } ) g ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) } } \\\\ { { \\displaystyle ~ = \\sum _ { k = 0 } ^ { \\infty } N ( d , k ) \\int _ { { \\mathbb { S } } ^ { d } } g ( { \\bf y } ) P _ { k } ( \\langle { \\bf x } , { \\bf y } \\rangle ) d \\tau _ { d } ( { \\bf y } ) } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 806, + 1173, + 889, + 1173, + 889, + 1208, + 806, + 1208 + ], + "score": 0.95, + "latex": "L _ { 2 } ( \\mathbb { S } ^ { d } )" + }, + { + "category_id": 14, + "poly": [ + 508, + 1866, + 1192, + 1866, + 1192, + 2033, + 508, + 2033 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\kappa _ { 1 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad = \\mathbb { E } _ { \\mathbf { w } \\sim N ( \\mathbf { 0 } , \\mathbf { I } ) } \\left[ \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { w } / \\left. \\mathbf { w } \\right. , \\mathbf { x } ^ { \\prime } \\rangle ) \\right] } \\\\ & { \\quad \\quad \\quad = \\displaystyle \\int _ { \\mathbb { S } ^ { d } } \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ^ { \\prime } ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 448, + 1034, + 1253, + 1034, + 1253, + 1133, + 448, + 1133 + ], + "score": 0.93, + "latex": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } N ( d , k ) P _ { k } ( \\langle \\mathbf { x } , \\mathbf { x } ^ { \\prime } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\mu _ { k } \\sum _ { j = 1 } ^ { N ( p , k ) } Y _ { k , j } ( \\mathbf { x } ) Y _ { k , j } ( \\mathbf { x } ^ { \\prime } ) ." + }, + { + "category_id": 13, + "poly": [ + 1106, + 1143, + 1198, + 1143, + 1198, + 1177, + 1106, + 1177 + ], + "score": 0.93, + "latex": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 769, + 349, + 1042, + 349, + 1042, + 387, + 769, + 387 + ], + "score": 0.93, + "latex": "g _ { 0 } ( \\mathbf x ) = \\int _ { \\mathbb { S } ^ { d } } g ( \\mathbf { \\bar { y } } ) d \\tau _ { d } ( \\mathbf y )" + }, + { + "category_id": 13, + "poly": [ + 587, + 1718, + 653, + 1718, + 653, + 1751, + 587, + 1751 + ], + "score": 0.93, + "latex": "P _ { k } ( t )" + }, + { + "category_id": 14, + "poly": [ + 607, + 1418, + 1090, + 1418, + 1090, + 1601, + 607, + 1601 + ], + "score": 0.92, + "latex": "\\begin{array} { l } { { \\displaystyle \\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\\\ { { \\displaystyle \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } N ( d , k ) P _ { k } ( \\langle { \\bf w } , { \\bf x } \\rangle ) } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1081, + 1688, + 1214, + 1688, + 1214, + 1721, + 1081, + 1721 + ], + "score": 0.92, + "latex": "\\beta _ { \\alpha + 1 , k } = 0" + }, + { + "category_id": 14, + "poly": [ + 558, + 1216, + 1140, + 1216, + 1140, + 1290, + 558, + 1290 + ], + "score": 0.92, + "latex": "L _ { \\kappa } ( f ) ( \\mathbf { y } ) = \\int _ { \\mathbb { S } ^ { d } } \\kappa ( \\mathbf { x } , \\mathbf { y } ) f ( \\mathbf { x } ) d \\tau _ { d } ( \\mathbf { x } ) , \\quad f \\in L _ { 2 } ( \\mathbb { S } ^ { d } ) ." + }, + { + "category_id": 13, + "poly": [ + 494, + 1688, + 570, + 1688, + 570, + 1716, + 494, + 1716 + ], + "score": 0.92, + "latex": "k > \\alpha" + }, + { + "category_id": 13, + "poly": [ + 298, + 979, + 776, + 979, + 776, + 1024, + 298, + 1024 + ], + "score": 0.92, + "latex": "( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\widehat { \\kappa } ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t" + }, + { + "category_id": 13, + "poly": [ + 430, + 1783, + 531, + 1783, + 531, + 1817, + 430, + 1817 + ], + "score": 0.92, + "latex": "\\beta _ { 2 , k } = 0" + }, + { + "category_id": 13, + "poly": [ + 977, + 1610, + 1407, + 1610, + 1407, + 1656, + 977, + 1656 + ], + "score": 0.92, + "latex": "\\beta _ { 2 , k } = ( \\omega _ { d - 1 } / \\omega _ { d } ) \\int _ { - 1 } ^ { 1 } \\sigma ^ { \\prime } ( t ) P _ { k } ( t ) ( 1 -" + }, + { + "category_id": 13, + "poly": [ + 488, + 349, + 619, + 349, + 619, + 383, + 488, + 383 + ], + "score": 0.92, + "latex": "g : \\mathbb { S } ^ { d } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 765, + 653, + 1046, + 653, + 1046, + 690, + 765, + 690 + ], + "score": 0.92, + "latex": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) : \\mathbb { S } ^ { d } \\times \\mathbb { S } ^ { d } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1752, + 1359, + 1752, + 1359, + 1787, + 1250, + 1787 + ], + "score": 0.92, + "latex": "\\beta _ { 1 , k } = 0" + }, + { + "category_id": 13, + "poly": [ + 574, + 1784, + 704, + 1784, + 704, + 1815, + 574, + 1815 + ], + "score": 0.92, + "latex": "k = 2 j + 1" + }, + { + "category_id": 13, + "poly": [ + 898, + 1657, + 947, + 1657, + 947, + 1691, + 898, + 1691 + ], + "score": 0.92, + "latex": "\\beta _ { 1 , k }" + }, + { + "category_id": 13, + "poly": [ + 575, + 717, + 760, + 717, + 760, + 751, + 575, + 751 + ], + "score": 0.91, + "latex": "\\bar { \\phi } : [ - 1 , 1 ] \\to \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1783, + 380, + 1783, + 380, + 1815, + 297, + 1815 + ], + "score": 0.91, + "latex": "k = 2 j" + }, + { + "category_id": 13, + "poly": [ + 648, + 1315, + 741, + 1315, + 741, + 1347, + 648, + 1347 + ], + "score": 0.91, + "latex": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 719, + 404, + 719, + 404, + 751, + 298, + 751 + ], + "score": 0.91, + "latex": "\\phi ( \\langle \\mathbf { x } , \\mathbf { z } \\rangle )" + }, + { + "category_id": 13, + "poly": [ + 462, + 686, + 659, + 686, + 659, + 719, + 462, + 719 + ], + "score": 0.9, + "latex": "\\widehat { \\kappa } : [ - 1 , 1 ] \\ \\mathrm { ~ ~ } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 996, + 1657, + 1046, + 1657, + 1046, + 1691, + 996, + 1691 + ], + "score": 0.9, + "latex": "\\beta _ { 2 , k }" + }, + { + "category_id": 13, + "poly": [ + 650, + 1749, + 916, + 1749, + 916, + 1784, + 650, + 1784 + ], + "score": 0.9, + "latex": "p ( t ) = ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 }" + }, + { + "category_id": 13, + "poly": [ + 1182, + 320, + 1253, + 320, + 1253, + 348, + 1182, + 348 + ], + "score": 0.9, + "latex": "k \\gg d" + }, + { + "category_id": 13, + "poly": [ + 295, + 1652, + 437, + 1652, + 437, + 1690, + 295, + 1690 + ], + "score": 0.9, + "latex": "t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t" + }, + { + "category_id": 13, + "poly": [ + 297, + 686, + 413, + 686, + 413, + 719, + 297, + 719 + ], + "score": 0.9, + "latex": "\\widehat { \\kappa } ( \\langle \\mathbf { x } , \\mathbf { \\bar { x } } ^ { \\prime } \\rangle )" + }, + { + "category_id": 13, + "poly": [ + 382, + 1378, + 465, + 1378, + 465, + 1410, + 382, + 1410 + ], + "score": 0.9, + "latex": "L _ { 2 } ( \\mathbb { S } ^ { d } )" + }, + { + "category_id": 13, + "poly": [ + 598, + 320, + 669, + 320, + 669, + 348, + 598, + 348 + ], + "score": 0.9, + "latex": "k \\gg d" + }, + { + "category_id": 13, + "poly": [ + 1241, + 1314, + 1401, + 1314, + 1401, + 1349, + 1241, + 1349 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } { \\sigma ( \\langle { \\bf w } , { \\bf x } \\rangle ) } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1303, + 320, + 1372, + 320, + 1372, + 348, + 1303, + 348 + ], + "score": 0.89, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 540, + 1347, + 890, + 1347, + 890, + 1380, + 540, + 1380 + ], + "score": 0.89, + "latex": "\\sigma ^ { \\prime } ( \\langle { \\bf w } , { \\bf x } \\rangle ) = \\mathbb { 1 } \\{ \\langle { \\bf w } , { \\bf x } \\rangle > 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 724, + 1176, + 767, + 1176, + 767, + 1205, + 724, + 1205 + ], + "score": 0.89, + "latex": "L _ { K }" + }, + { + "category_id": 13, + "poly": [ + 1276, + 656, + 1402, + 656, + 1402, + 688, + 1276, + 688 + ], + "score": 0.88, + "latex": "\\kappa ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) =" + }, + { + "category_id": 13, + "poly": [ + 764, + 1784, + 846, + 1784, + 846, + 1815, + 764, + 1815 + ], + "score": 0.88, + "latex": "j \\in \\mathbb { N } ^ { + }" + }, + { + "category_id": 13, + "poly": [ + 1313, + 689, + 1402, + 689, + 1402, + 720, + 1313, + 720 + ], + "score": 0.87, + "latex": "g ( \\mathbf { x } ) =" + }, + { + "category_id": 13, + "poly": [ + 355, + 1346, + 483, + 1346, + 483, + 1379, + 355, + 1379 + ], + "score": 0.87, + "latex": "\\{ \\langle \\mathbf { w } , \\mathbf { x } \\rangle , 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 448, + 717, + 523, + 717, + 523, + 746, + 448, + 746 + ], + "score": 0.86, + "latex": "\\mathbf { z } \\in \\mathbb { S } ^ { \\bar { d } }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1611, + 927, + 1611, + 927, + 1656, + 371, + 1656 + ], + "score": 0.86, + "latex": "\\beta _ { 1 , k } = \\left( \\omega _ { d - 1 } / \\omega _ { d } \\right) \\int _ { - 1 } ^ { 1 } \\sigma ( t ) P _ { k } ( t ) ( 1 - t ^ { 2 } ) ^ { ( d - 2 ) / 2 } d t" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1148, + 1286, + 1148, + 1286, + 1176, + 1252, + 1176 + ], + "score": 0.86, + "latex": "\\mu _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1314, + 949, + 1402, + 949, + 1402, + 982, + 1314, + 982 + ], + "score": 0.83, + "latex": "\\begin{array} { r l } { \\lambda _ { k } } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 602, + 602, + 620, + 602, + 620, + 628, + 602, + 628 + ], + "score": 0.83, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 688, + 1689, + 706, + 1689, + 706, + 1715, + 688, + 1715 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1324, + 1719, + 1343, + 1719, + 1343, + 1745, + 1324, + 1745 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 758, + 1693, + 779, + 1693, + 779, + 1715, + 758, + 1715 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 720.0, + 228.0, + 720.0, + 263.0, + 294.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1606.0, + 370.0, + 1606.0, + 370.0, + 1657.0, + 290.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1606.0, + 976.0, + 1606.0, + 976.0, + 1657.0, + 928.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1645.0, + 294.0, + 1645.0, + 294.0, + 1696.0, + 291.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1645.0, + 897.0, + 1645.0, + 897.0, + 1696.0, + 438.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1645.0, + 995.0, + 1645.0, + 995.0, + 1696.0, + 948.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1645.0, + 1408.0, + 1645.0, + 1408.0, + 1696.0, + 1047.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1685.0, + 493.0, + 1685.0, + 493.0, + 1722.0, + 295.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1685.0, + 687.0, + 1685.0, + 687.0, + 1722.0, + 571.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1685.0, + 757.0, + 1685.0, + 757.0, + 1722.0, + 707.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1685.0, + 1080.0, + 1685.0, + 1080.0, + 1722.0, + 780.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1722.0, + 1215.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1716.0, + 586.0, + 1716.0, + 586.0, + 1749.0, + 296.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1716.0, + 1323.0, + 1716.0, + 1323.0, + 1749.0, + 654.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 1716.0, + 1403.0, + 1716.0, + 1403.0, + 1749.0, + 1344.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1745.0, + 649.0, + 1745.0, + 649.0, + 1789.0, + 289.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1745.0, + 1249.0, + 1745.0, + 1249.0, + 1789.0, + 917.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 1745.0, + 1408.0, + 1745.0, + 1408.0, + 1789.0, + 1360.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1779.0, + 296.0, + 1779.0, + 296.0, + 1817.0, + 292.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1779.0, + 429.0, + 1779.0, + 429.0, + 1817.0, + 381.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1779.0, + 573.0, + 1779.0, + 573.0, + 1817.0, + 532.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1779.0, + 763.0, + 1779.0, + 763.0, + 1817.0, + 705.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1779.0, + 858.0, + 1779.0, + 858.0, + 1817.0, + 847.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1313.0, + 647.0, + 1313.0, + 647.0, + 1351.0, + 295.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1313.0, + 1240.0, + 1313.0, + 1240.0, + 1351.0, + 742.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1346.0, + 354.0, + 1346.0, + 354.0, + 1381.0, + 295.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1346.0, + 539.0, + 1346.0, + 539.0, + 1381.0, + 484.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1346.0, + 1402.0, + 1346.0, + 1402.0, + 1381.0, + 891.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1372.0, + 381.0, + 1372.0, + 381.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1372.0, + 1141.0, + 1372.0, + 1141.0, + 1414.0, + 466.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 1403.0, + 287.0, + 1403.0, + 322.0, + 295.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 317.0, + 597.0, + 317.0, + 597.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 317.0, + 1181.0, + 317.0, + 1181.0, + 354.0, + 670.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 317.0, + 1302.0, + 317.0, + 1302.0, + 354.0, + 1254.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 317.0, + 1382.0, + 317.0, + 1382.0, + 354.0, + 1373.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 346.0, + 487.0, + 346.0, + 487.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 346.0, + 768.0, + 346.0, + 768.0, + 388.0, + 620.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 346.0, + 1331.0, + 346.0, + 1331.0, + 388.0, + 1043.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1137.0, + 1105.0, + 1137.0, + 1105.0, + 1182.0, + 294.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1137.0, + 1251.0, + 1137.0, + 1251.0, + 1182.0, + 1199.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1137.0, + 1402.0, + 1137.0, + 1402.0, + 1182.0, + 1287.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1171.0, + 723.0, + 1171.0, + 723.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1171.0, + 805.0, + 1171.0, + 805.0, + 1211.0, + 768.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1171.0, + 1017.0, + 1171.0, + 1017.0, + 1211.0, + 890.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 945.0, + 1313.0, + 945.0, + 1313.0, + 987.0, + 295.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 978.0, + 297.0, + 978.0, + 297.0, + 1024.0, + 293.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 978.0, + 1168.0, + 978.0, + 1168.0, + 1024.0, + 777.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 648.0, + 764.0, + 648.0, + 764.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 648.0, + 1275.0, + 648.0, + 1275.0, + 693.0, + 1047.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 648.0, + 1406.0, + 648.0, + 1406.0, + 693.0, + 1403.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 682.0, + 296.0, + 682.0, + 296.0, + 722.0, + 293.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 682.0, + 461.0, + 682.0, + 461.0, + 722.0, + 414.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 682.0, + 1312.0, + 682.0, + 1312.0, + 722.0, + 660.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 715.0, + 297.0, + 715.0, + 297.0, + 752.0, + 294.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 715.0, + 447.0, + 715.0, + 447.0, + 752.0, + 405.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 715.0, + 574.0, + 715.0, + 574.0, + 752.0, + 524.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 715.0, + 775.0, + 715.0, + 775.0, + 752.0, + 761.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1825.0, + 847.0, + 1825.0, + 847.0, + 1865.0, + 293.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 595.0, + 601.0, + 595.0, + 601.0, + 630.0, + 296.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 595.0, + 1405.0, + 595.0, + 1405.0, + 630.0, + 621.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 625.0, + 615.0, + 625.0, + 615.0, + 662.0, + 295.0, + 662.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1725, + 1405, + 1725, + 1405, + 1855, + 297, + 1855 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1402, + 228, + 1402, + 323, + 298, + 323 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 616, + 496, + 1083, + 496, + 1083, + 583, + 616, + 583 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 515, + 854, + 1182, + 854, + 1182, + 937, + 515, + 937 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 556, + 334, + 1142, + 334, + 1142, + 403, + 556, + 403 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 572, + 630, + 1123, + 630, + 1123, + 714, + 572, + 714 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 534, + 1431, + 1162, + 1431, + 1162, + 1503, + 534, + 1503 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 606, + 1867, + 1091, + 1867, + 1091, + 1948, + 606, + 1948 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 296, + 410, + 978, + 410, + 978, + 445, + 296, + 445 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 635, + 764, + 1060, + 764, + 1060, + 805, + 635, + 805 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 298, + 1666, + 875, + 1666, + 875, + 1700, + 298, + 1700 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 296, + 722, + 979, + 722, + 979, + 756, + 296, + 756 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 296, + 1511, + 1265, + 1511, + 1265, + 1544, + 296, + 1544 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 294, + 810, + 1250, + 810, + 1250, + 845, + 294, + 845 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 296, + 592, + 345, + 592, + 345, + 623, + 296, + 623 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 300, + 948, + 1401, + 948, + 1401, + 986, + 300, + 986 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1326, + 520, + 1401, + 520, + 1401, + 554, + 1326, + 554 + ], + "score": 0.911 + }, + { + "category_id": 1, + "poly": [ + 298, + 1601, + 557, + 1601, + 557, + 1635, + 298, + 1635 + ], + "score": 0.91 + }, + { + "category_id": 9, + "poly": [ + 1326, + 766, + 1401, + 766, + 1401, + 800, + 1326, + 800 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 296, + 1399, + 345, + 1399, + 345, + 1428, + 296, + 1428 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1326, + 351, + 1401, + 351, + 1401, + 383, + 1326, + 383 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1326, + 652, + 1401, + 652, + 1401, + 686, + 1326, + 686 + ], + "score": 0.902 + }, + { + "category_id": 8, + "poly": [ + 661, + 1292, + 1037, + 1292, + 1037, + 1391, + 661, + 1391 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 298, + 1957, + 1414, + 1957, + 1414, + 2042, + 298, + 2042 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2114, + 835, + 2114 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1603, + 1403, + 1603, + 1403, + 1632, + 1374, + 1632 + ], + "score": 0.821 + }, + { + "category_id": 1, + "poly": [ + 296, + 995, + 1407, + 995, + 1407, + 1252, + 296, + 1252 + ], + "score": 0.569 + }, + { + "category_id": 8, + "poly": [ + 308, + 1551, + 1363, + 1551, + 1363, + 1593, + 308, + 1593 + ], + "score": 0.564 + }, + { + "category_id": 1, + "poly": [ + 308, + 1551, + 1363, + 1551, + 1363, + 1593, + 308, + 1593 + ], + "score": 0.456 + }, + { + "category_id": 8, + "poly": [ + 296, + 995, + 1407, + 995, + 1407, + 1252, + 296, + 1252 + ], + "score": 0.25 + }, + { + "category_id": 1, + "poly": [ + 300, + 1249, + 752, + 1249, + 752, + 1279, + 300, + 1279 + ], + "score": 0.233 + }, + { + "category_id": 8, + "poly": [ + 664, + 1349, + 1036, + 1349, + 1036, + 1390, + 664, + 1390 + ], + "score": 0.19 + }, + { + "category_id": 8, + "poly": [ + 736, + 1293, + 961, + 1293, + 961, + 1329, + 736, + 1329 + ], + "score": 0.158 + }, + { + "category_id": 14, + "poly": [ + 617, + 493, + 1081, + 493, + 1081, + 584, + 617, + 584 + ], + "score": 0.95, + "latex": "\\kappa _ { 1 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 1 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) ," + }, + { + "category_id": 14, + "poly": [ + 605, + 1862, + 1091, + 1862, + 1091, + 1949, + 605, + 1949 + ], + "score": 0.93, + "latex": "| Y _ { k } ( { \\bf x } ) | \\leqslant \\left( N ( d , k ) \\int _ { { \\mathbb S } ^ { d } } Y _ { k } ^ { 2 } ( { \\bf y } ) d \\tau _ { d } ( { \\bf y } ) \\right) ^ { \\frac { 1 } { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 520, + 947, + 788, + 947, + 788, + 985, + 520, + 985 + ], + "score": 0.93, + "latex": "\\Gamma ( x ) \\approx x ^ { x - 1 / 2 } e ^ { - x } \\sqrt { 2 \\pi }" + }, + { + "category_id": 13, + "poly": [ + 527, + 1958, + 776, + 1958, + 776, + 1998, + 527, + 1998 + ], + "score": 0.93, + "latex": "| \\phi _ { j } ( \\mathbf { x } ) | \\leqslant \\sqrt { N ( d , k ) }" + }, + { + "category_id": 13, + "poly": [ + 636, + 1997, + 1055, + 1997, + 1055, + 2041, + 636, + 2041 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { N ( d , k ) = \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } = \\mathcal { O } ( d ^ { k } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 573, + 626, + 1127, + 626, + 1127, + 716, + 573, + 716 + ], + "score": 0.93, + "latex": "\\kappa _ { 2 } ( { \\bf x } , { \\bf x } ^ { \\prime } ) = ( d + 1 ) \\sum _ { k = 0 } ^ { \\infty } \\beta _ { 2 , k } ^ { 2 } N ( d , k ) P _ { k } ( \\langle { \\bf x } , { \\bf x } ^ { \\prime } \\rangle ) ." + }, + { + "category_id": 13, + "poly": [ + 974, + 1727, + 1059, + 1727, + 1059, + 1762, + 974, + 1762 + ], + "score": 0.93, + "latex": "| \\phi _ { j } ( \\mathbf { x } ) |" + }, + { + "category_id": 14, + "poly": [ + 514, + 854, + 1184, + 854, + 1184, + 938, + 514, + 938 + ], + "score": 0.92, + "latex": "\\beta _ { \\alpha + 1 , k } = \\frac { d - 1 } { 2 \\pi } \\frac { \\alpha ! ( - 1 ) ^ { ( k - 1 - \\alpha ) / 2 } } { 2 ^ { k } } \\frac { \\Gamma ( d / 2 ) \\Gamma ( k - \\alpha ) } { \\Gamma ( \\frac { k - \\alpha + 1 } { 2 } ) \\Gamma ( \\frac { k + d + \\alpha + 1 } { 2 } ) } ." + }, + { + "category_id": 14, + "poly": [ + 555, + 330, + 1143, + 330, + 1143, + 405, + 555, + 405 + ], + "score": 0.92, + "latex": "\\kappa _ { 2 } ( \\mathbf { x } , \\mathbf { x } ^ { \\prime } ) = ( d + 1 ) \\int _ { \\mathbb { S } ^ { d } } \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } \\rangle ) \\sigma ( \\langle \\mathbf { v } , \\mathbf { x } ^ { \\prime } \\rangle ) d \\tau _ { d } ( \\mathbf { v } ) ." + }, + { + "category_id": 13, + "poly": [ + 775, + 1791, + 866, + 1791, + 866, + 1825, + 775, + 1825 + ], + "score": 0.92, + "latex": "j \\in [ r _ { k } ]" + }, + { + "category_id": 13, + "poly": [ + 1162, + 952, + 1243, + 952, + 1243, + 985, + 1162, + 985 + ], + "score": 0.92, + "latex": "\\beta _ { \\alpha + 1 , k }" + }, + { + "category_id": 13, + "poly": [ + 1045, + 1957, + 1400, + 1957, + 1400, + 2002, + 1045, + 2002 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { N ( d , k ) \\ = \\ \\frac { 2 k + d - 1 } { k } \\binom { k + d - 2 } { d - 1 } \\ = \\ } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 844, + 812, + 893, + 812, + 893, + 846, + 844, + 846 + ], + "score": 0.92, + "latex": "\\beta _ { 2 , k }" + }, + { + "category_id": 14, + "poly": [ + 330, + 994, + 1366, + 994, + 1366, + 1082, + 330, + 1082 + ], + "score": 0.92, + "latex": "\\beta _ { \\alpha + 1 , k } = C ( \\alpha ) \\frac { ( d - 1 ) d ^ { \\frac { d - 1 } { 2 } } ( k - \\alpha ) ^ { k - \\alpha - \\frac 1 2 } } { ( k - \\alpha + 1 ) ^ { \\frac { k - \\alpha } { 2 } } ( k + d + \\alpha + 1 ) ^ { \\frac { k + d + \\alpha } { 2 } } } = \\Omega \\left( d ^ { \\frac { d + 1 } { 2 } } k ^ { \\frac { k - \\alpha - 1 } { 2 } } ( k + d ) ^ { \\frac { - k - d - \\alpha } { 2 } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1727, + 1192, + 1727, + 1192, + 1761, + 1100, + 1761 + ], + "score": 0.92, + "latex": "j \\in [ r _ { k } ]" + }, + { + "category_id": 14, + "poly": [ + 536, + 1431, + 1161, + 1431, + 1161, + 1502, + 536, + 1502 + ], + "score": 0.91, + "latex": "\\mu _ { k } = \\frac { k } { 2 k + d - 1 } \\mu _ { 1 , k - 1 } + \\frac { k + d - 1 } { 2 k + d - 1 } \\mu _ { 1 , k + 1 } + 2 \\mu _ { 2 , k } ," + }, + { + "category_id": 13, + "poly": [ + 298, + 1999, + 400, + 1999, + 400, + 2036, + 298, + 2036 + ], + "score": 0.91, + "latex": "\\mathcal { O } ( k ^ { d - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 743, + 812, + 793, + 812, + 793, + 846, + 743, + 846 + ], + "score": 0.91, + "latex": "\\beta _ { 1 , k }" + }, + { + "category_id": 13, + "poly": [ + 456, + 1511, + 534, + 1511, + 534, + 1541, + 456, + 1541 + ], + "score": 0.91, + "latex": "k \\neq k ^ { \\prime }" + }, + { + "category_id": 14, + "poly": [ + 326, + 1550, + 1370, + 1550, + 1370, + 1595, + 326, + 1595 + ], + "score": 0.91, + "latex": "\\mu _ { k } = \\Omega \\left( \\operatorname * { m a x } \\left\\{ d ^ { d + 1 } k ^ { k - 1 } ( k + d ) ^ { - k - d } , d ^ { d + 1 } k ^ { k } ( k + d ) ^ { - k - d - 1 } , d ^ { d + 2 } k ^ { k - 2 } ( k + d ) ^ { - k - d - 1 } \\right\\} \\right) ." + }, + { + "category_id": 13, + "poly": [ + 337, + 1512, + 406, + 1512, + 406, + 1541, + 337, + 1541 + ], + "score": 0.91, + "latex": "k \\geqslant 1" + }, + { + "category_id": 13, + "poly": [ + 937, + 812, + 1058, + 812, + 1058, + 841, + 937, + 841 + ], + "score": 0.9, + "latex": "k \\geqslant \\alpha + 1" + }, + { + "category_id": 13, + "poly": [ + 1191, + 1512, + 1259, + 1512, + 1259, + 1541, + 1191, + 1541 + ], + "score": 0.89, + "latex": "k \\geqslant 1" + }, + { + "category_id": 13, + "poly": [ + 846, + 1962, + 927, + 1962, + 927, + 1991, + 846, + 1991 + ], + "score": 0.89, + "latex": "k \\gg d" + }, + { + "category_id": 13, + "poly": [ + 1285, + 952, + 1401, + 952, + 1401, + 981, + 1285, + 981 + ], + "score": 0.89, + "latex": "k \\geqslant \\alpha + 1" + }, + { + "category_id": 13, + "poly": [ + 458, + 2002, + 527, + 2002, + 527, + 2032, + 458, + 2032 + ], + "score": 0.89, + "latex": "d \\gg k" + }, + { + "category_id": 13, + "poly": [ + 422, + 1152, + 737, + 1152, + 737, + 1215, + 422, + 1215 + ], + "score": 0.88, + "latex": "\\begin{array} { r l r } { \\beta _ { 2 , 1 } } & { { } = } & { \\frac { d - 1 } { 4 \\pi d } \\frac { \\Gamma \\left( \\frac { 1 } { 2 } \\right) \\Gamma \\left( \\frac { d + 2 } { 2 } \\right) } { \\Gamma \\left( \\frac { d + 3 } { 2 } \\right) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 442, + 1821, + 473, + 1821, + 473, + 1850, + 442, + 1850 + ], + "score": 0.88, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1794, + 1234, + 1794, + 1234, + 1823, + 1202, + 1823 + ], + "score": 0.88, + "latex": "Y _ { k }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1163, + 348, + 1163, + 348, + 1206, + 299, + 1206 + ], + "score": 0.87, + "latex": "\\textstyle { \\frac { d - 1 } { 2 d \\pi } }" + }, + { + "category_id": 13, + "poly": [ + 656, + 231, + 683, + 231, + 683, + 258, + 656, + 258 + ], + "score": 0.86, + "latex": "\\sigma ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1212, + 699, + 1212, + 699, + 1252, + 297, + 1252 + ], + "score": 0.86, + "latex": "\\Omega \\left( d ^ { d + 1 + \\alpha } k ^ { k - \\alpha - 1 } ( k + d ) ^ { - k - \\bar { d } - \\bar { \\alpha } } \\right)" + }, + { + "category_id": 14, + "poly": [ + 634, + 762, + 1064, + 762, + 1064, + 806, + 634, + 806 + ], + "score": 0.86, + "latex": "\\mu _ { 1 , k } = \\beta _ { 1 , k } ^ { 2 } \\mathrm { a n d } \\mu _ { 2 , k } = ( d + 1 ) \\beta _ { 2 , k } ^ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 387, + 1098, + 766, + 1098, + 766, + 1145, + 387, + 1145 + ], + "score": 0.83, + "latex": "\\begin{array} { r l r } { C ( \\alpha ) } & { { } = } & { \\frac { \\sqrt { 2 } \\alpha ! } { 2 \\pi } \\exp \\{ \\alpha \\ + \\ 1 \\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1223, + 1760, + 1292, + 1760, + 1292, + 1795, + 1223, + 1795 + ], + "score": 0.82, + "latex": "\\phi _ { j } ( \\mathbf { x } )" + }, + { + "category_id": 13, + "poly": [ + 1335, + 1793, + 1354, + 1793, + 1354, + 1819, + 1335, + 1819 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 712, + 1793, + 731, + 1793, + 731, + 1819, + 712, + 1819 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1181, + 1760, + 1213, + 1760, + 1213, + 1791, + 1181, + 1791 + ], + "score": 0.81, + "latex": "\\mathbb { S } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 385, + 1092, + 1342, + 1092, + 1342, + 1156, + 385, + 1156 + ], + "score": 0.73, + "latex": "\\begin{array} { r l r l r l r l r l } { C ( \\alpha ) } & { { } = } & { { \\frac { \\sqrt { 2 } \\alpha ! } { 2 \\pi } } \\exp \\{ \\alpha \\ + \\ 1 \\} . } & { } & { \\mathrm { A l s o } \\ \\beta _ { \\alpha + 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 2 } { 2 } } \\right) } } , } & { \\beta _ { 1 , 1 } \\beta _ { 1 , 2 } \\gamma _ { 1 , 0 } } & { { } = } & { { \\frac { d - 1 } { 4 \\pi } } { \\frac { \\Gamma \\left( { \\frac { \\alpha + 1 } { 2 } } \\right) \\Gamma \\left( { \\frac { d } { 2 } } \\right) } { \\Gamma \\left( { \\frac { d + \\alpha + 1 } { 2 } } \\right) } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 664, + 1349, + 1033, + 1349, + 1033, + 1388, + 664, + 1388 + ], + "score": 0.66, + "latex": "\\mu _ { k ^ { \\prime } } = 0 , k ^ { \\prime } = 2 j + 1 , j \\in \\mathbb { N } ^ { + } ," + }, + { + "category_id": 14, + "poly": [ + 738, + 1291, + 956, + 1291, + 956, + 1328, + 738, + 1328 + ], + "score": 0.49, + "latex": "\\mu _ { 0 } = \\mu _ { 1 , 1 } + 2 \\mu _ { 2 , 0 } ," + }, + { + "category_id": 14, + "poly": [ + 665, + 1288, + 1033, + 1288, + 1033, + 1391, + 665, + 1391 + ], + "score": 0.26, + "latex": "\\begin{array} { c } { { \\mu _ { 0 } = \\mu _ { 1 , 1 } + 2 \\mu _ { 2 , 0 } , } } \\\\ { { { } } } \\\\ { { \\mu _ { k ^ { \\prime } } = 0 , k ^ { \\prime } = 2 j + 1 , j \\in \\mathbb { N } ^ { + } , } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 877.0, + 1665.0, + 877.0, + 1702.0, + 294.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1607.0, + 1403.0, + 1607.0, + 1403.0, + 1634.0, + 1377.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1725.0, + 973.0, + 1725.0, + 973.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1725.0, + 1099.0, + 1725.0, + 1099.0, + 1762.0, + 1060.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1725.0, + 1403.0, + 1725.0, + 1403.0, + 1762.0, + 1193.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1180.0, + 1760.0, + 1180.0, + 1797.0, + 294.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1760.0, + 1222.0, + 1760.0, + 1222.0, + 1797.0, + 1214.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1797.0, + 1293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1790.0, + 711.0, + 1790.0, + 711.0, + 1826.0, + 295.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1790.0, + 774.0, + 1790.0, + 774.0, + 1826.0, + 732.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1790.0, + 1201.0, + 1790.0, + 1201.0, + 1826.0, + 867.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1790.0, + 1334.0, + 1790.0, + 1334.0, + 1826.0, + 1235.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1790.0, + 1406.0, + 1790.0, + 1406.0, + 1826.0, + 1355.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 441.0, + 1821.0, + 441.0, + 1857.0, + 292.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1821.0, + 1301.0, + 1821.0, + 1301.0, + 1857.0, + 474.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 655.0, + 228.0, + 655.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 684.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 294.0, + 296.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 780.0, + 291.0, + 780.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 409.0, + 980.0, + 409.0, + 980.0, + 449.0, + 295.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 720.0, + 981.0, + 720.0, + 981.0, + 759.0, + 295.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1508.0, + 336.0, + 1508.0, + 336.0, + 1546.0, + 294.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1508.0, + 455.0, + 1508.0, + 455.0, + 1546.0, + 407.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1508.0, + 1190.0, + 1508.0, + 1190.0, + 1546.0, + 535.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1508.0, + 1263.0, + 1508.0, + 1263.0, + 1546.0, + 1260.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 804.0, + 742.0, + 804.0, + 742.0, + 852.0, + 290.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 804.0, + 843.0, + 804.0, + 843.0, + 852.0, + 794.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 804.0, + 936.0, + 804.0, + 936.0, + 852.0, + 894.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 804.0, + 1249.0, + 804.0, + 1249.0, + 852.0, + 1059.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 592.0, + 344.0, + 592.0, + 344.0, + 622.0, + 294.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 940.0, + 519.0, + 940.0, + 519.0, + 993.0, + 293.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 940.0, + 1161.0, + 940.0, + 1161.0, + 993.0, + 789.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 940.0, + 1284.0, + 940.0, + 1284.0, + 993.0, + 1244.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 940.0, + 1407.0, + 940.0, + 1407.0, + 993.0, + 1402.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1597.0, + 559.0, + 1597.0, + 559.0, + 1639.0, + 294.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1397.0, + 345.0, + 1397.0, + 345.0, + 1429.0, + 293.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1956.0, + 297.0, + 1956.0, + 297.0, + 2045.0, + 290.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1956.0, + 457.0, + 1956.0, + 457.0, + 2045.0, + 401.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1956.0, + 1415.0, + 1956.0, + 1415.0, + 2045.0, + 1401.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1100.0, + 384.0, + 1100.0, + 384.0, + 1143.0, + 293.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1094.0, + 1383.0, + 1094.0, + 1383.0, + 1152.0, + 1343.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1142.0, + 298.0, + 1142.0, + 298.0, + 1221.0, + 286.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1142.0, + 421.0, + 1142.0, + 421.0, + 1221.0, + 349.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1142.0, + 747.0, + 1142.0, + 747.0, + 1221.0, + 738.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1159.0, + 1407.0, + 1159.0, + 1407.0, + 1210.0, + 780.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1207.0, + 296.0, + 1207.0, + 296.0, + 1255.0, + 291.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1207.0, + 715.0, + 1207.0, + 715.0, + 1255.0, + 700.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 998.5, + 1375.0, + 998.5, + 1375.0, + 1097.5, + 491.0, + 1097.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.25, + 1092.0, + 1057.25, + 1092.0, + 1057.25, + 1151.5, + 537.25, + 1151.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.75, + 1092.0, + 1157.75, + 1092.0, + 1157.75, + 1129.0, + 1069.75, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1543.0, + 325.0, + 1543.0, + 325.0, + 1602.0, + 316.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1245.0, + 755.0, + 1245.0, + 755.0, + 1281.0, + 295.0, + 1281.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 351, + 1405, + 351, + 1405, + 492, + 296, + 492 + ], + "score": 0.977 + }, + { + "category_id": 8, + "poly": [ + 306, + 724, + 1394, + 724, + 1394, + 949, + 306, + 949 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 373, + 1658, + 1327, + 1658, + 1327, + 1882, + 373, + 1882 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 494, + 1143, + 1204, + 1143, + 1204, + 1227, + 494, + 1227 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 293, + 1246, + 1401, + 1246, + 1401, + 1315, + 293, + 1315 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 633, + 507, + 1063, + 507, + 1063, + 555, + 633, + 555 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 321, + 1329, + 1375, + 1329, + 1375, + 1413, + 321, + 1413 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 365, + 1945, + 1334, + 1945, + 1334, + 2029, + 365, + 2029 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 298, + 1092, + 1018, + 1092, + 1018, + 1127, + 298, + 1127 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 295, + 566, + 1055, + 566, + 1055, + 607, + 295, + 607 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 297, + 962, + 958, + 962, + 958, + 997, + 297, + 997 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 296, + 1892, + 689, + 1892, + 689, + 1927, + 296, + 1927 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 297, + 1429, + 706, + 1429, + 706, + 1462, + 297, + 1462 + ], + "score": 0.928 + }, + { + "category_id": 8, + "poly": [ + 432, + 622, + 1267, + 622, + 1267, + 669, + 432, + 669 + ], + "score": 0.926 + }, + { + "category_id": 8, + "poly": [ + 316, + 1479, + 1375, + 1479, + 1375, + 1563, + 316, + 1563 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 107, + 297, + 107 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 298, + 1614, + 371, + 1614, + 371, + 1643, + 298, + 1643 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 297, + 679, + 413, + 679, + 413, + 711, + 297, + 711 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 298, + 1032, + 641, + 1032, + 641, + 1065, + 298, + 1065 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 297, + 225, + 813, + 225, + 813, + 263, + 297, + 263 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1166, + 1401, + 1166, + 1401, + 1199, + 1340, + 1199 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1340, + 1562, + 1401, + 1562, + 1401, + 1593, + 1340, + 1593 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 866, + 2087, + 866, + 2114, + 836, + 2114 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 1374, + 964, + 1404, + 964, + 1404, + 993, + 1374, + 993 + ], + "score": 0.829 + }, + { + "category_id": 0, + "poly": [ + 299, + 293, + 642, + 293, + 642, + 327, + 299, + 327 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 299, + 293, + 642, + 293, + 642, + 327, + 299, + 327 + ], + "score": 0.212 + }, + { + "category_id": 14, + "poly": [ + 492, + 1140, + 1207, + 1140, + 1207, + 1229, + 492, + 1229 + ], + "score": 0.94, + "latex": "\\mathbf { W } ^ { ( t + 1 ) } = \\mathbf { W } ^ { ( t ) } + \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } \\bigl ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\bigr ) \\cdot \\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ." + }, + { + "category_id": 14, + "poly": [ + 322, + 1327, + 1374, + 1327, + 1374, + 1414, + 322, + 1414 + ], + "score": 0.94, + "latex": "\\theta \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle = \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\hat { y } _ { i } ^ { ( t ) } ) \\cdot \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle ." + }, + { + "category_id": 14, + "poly": [ + 322, + 1478, + 1374, + 1478, + 1374, + 1564, + 322, + 1564 + ], + "score": 0.94, + "latex": "y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t + 1 ) } ) _ { j } = y _ { j } - ( \\widehat { \\mathbf { y } } ^ { ( t ) } ) _ { j } - \\frac { 2 \\eta m \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot \\mathbf { K } _ { i , j } ^ { \\infty } + I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t } ," + }, + { + "category_id": 13, + "poly": [ + 482, + 568, + 779, + 568, + 779, + 608, + 482, + 608 + ], + "score": 0.94, + "latex": "\\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } = \\mathbf { A } ^ { \\top } \\mathbf { A } + \\mathbf { B } ^ { \\top } \\mathbf { B }" + }, + { + "category_id": 14, + "poly": [ + 370, + 1655, + 1325, + 1655, + 1325, + 1884, + 370, + 1884 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { { \\displaystyle I _ { 1 , j , t } = - \\frac { 2 \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot [ \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\rangle - m \\mathbf { K } _ { i , j } ^ { ( 0 ) } ] } , } \\\\ { { \\displaystyle I _ { 2 , j , t } = - \\frac { 2 \\eta \\eta \\theta ^ { 2 } } { n } \\sum _ { i = 1 } ^ { n } ( y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) ) \\cdot ( \\mathbf { K } _ { i , j } ^ { ( 0 ) } - \\mathbf { K } _ { i , j } ^ { \\infty } ) , } } \\\\ { { \\displaystyle I _ { 3 , j , t } = - \\theta \\cdot [ f _ { \\mathbf { W } ^ { ( t + 1 ) } } ( \\mathbf { x } _ { j } ) - f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { j } ) , \\mathbf { W } ^ { ( t + 1 ) } - \\mathbf { W } ^ { ( t ) } \\rangle ] } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 342, + 1895, + 400, + 1895, + 400, + 1928, + 342, + 1928 + ], + "score": 0.93, + "latex": "I _ { 1 , j , t }" + }, + { + "category_id": 14, + "poly": [ + 310, + 721, + 1393, + 721, + 1393, + 949, + 310, + 949 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\| \\mathbf { V } _ { r _ { k } } \\mathbf { V } _ { r _ { k } } ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad = \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { A } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { A } ^ { \\top } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\mathbf { B } \\mathbf { B } ^ { \\top } ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } - \\hat { \\mathbf { V } } _ { r _ { k } } \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| \\hat { \\mathbf { V } } _ { r _ { k } } \\big ( \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\big ) \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad = \\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } + \\mathcal { O } ( \\| \\mathbf { B } \\| _ { 2 } ) } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 710, + 384, + 971, + 384, + 971, + 425, + 710, + 425 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\dot { \\mathbf { V } _ { r _ { k } } } = \\hat { \\mathbf { V } } _ { r _ { k } } \\mathbf { \\bar { A } } + \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\mathbf { \\bar { B } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 632, + 506, + 1066, + 506, + 1066, + 553, + 632, + 553 + ], + "score": 0.92, + "latex": "\\| \\mathbf { B } \\| _ { F } = \\| \\mathbf { B } ^ { \\top } \\| _ { F } = \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } \\| _ { F } \\leqslant \\xi _ { 1 } ." + }, + { + "category_id": 14, + "poly": [ + 368, + 1943, + 1330, + 1943, + 1330, + 2030, + 368, + 2030 + ], + "score": 0.92, + "latex": "| I _ { 1 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n }" + }, + { + "category_id": 13, + "poly": [ + 1061, + 1249, + 1243, + 1249, + 1243, + 1284, + 1061, + 1284 + ], + "score": 0.92, + "latex": "\\theta \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( t ) } } \\left( \\mathbf { x } _ { j } \\right)" + }, + { + "category_id": 13, + "poly": [ + 392, + 1249, + 478, + 1249, + 478, + 1283, + 392, + 1283 + ], + "score": 0.92, + "latex": "j \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 617, + 965, + 644, + 965, + 644, + 995, + 617, + 995 + ], + "score": 0.89, + "latex": "\\xi _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 695, + 965, + 722, + 965, + 722, + 995, + 695, + 995 + ], + "score": 0.89, + "latex": "\\xi _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1207, + 386, + 1392, + 386, + 1392, + 419, + 1207, + 419 + ], + "score": 0.89, + "latex": "\\mathbf { B } \\in \\mathbb { R } ^ { ( n - r _ { k } ) \\times r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 623, + 1246, + 685, + 1246, + 685, + 1279, + 623, + 1279 + ], + "score": 0.88, + "latex": "\\mathbf { W } ^ { ( t ) }" + }, + { + "category_id": 14, + "poly": [ + 430, + 622, + 1268, + 622, + 1268, + 667, + 430, + 667 + ], + "score": 0.88, + "latex": "\\| \\mathbf { A } \\mathbf { A } ^ { \\top } - \\mathbf { I } \\| _ { 2 } = \\| \\mathbf { A } ^ { \\top } \\mathbf { A } - \\mathbf { I } \\| _ { 2 } \\leqslant \\| \\mathbf { V } _ { r _ { k } } ^ { \\top } \\mathbf { V } _ { r _ { k } } - \\mathbf { I } \\| _ { 2 } + \\| \\mathbf { B } ^ { \\top } \\mathbf { B } \\| _ { 2 } \\leqslant r _ { k } \\xi _ { 2 } + \\xi _ { 1 } ^ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 1056, + 389, + 1194, + 389, + 1194, + 420, + 1056, + 420 + ], + "score": 0.82, + "latex": "\\mathbf { A } \\in \\mathbb { R } ^ { r _ { k } \\times r _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 343, + 424, + 811, + 424, + 811, + 463, + 343, + 463 + ], + "score": 0.78, + "latex": "\\xi _ { 1 } = C ^ { \\prime } ( \\lambda _ { r _ { k } } - \\lambda _ { r _ { k } + 1 } ) ^ { - 1 } \\cdot \\sqrt { \\log ( 1 / \\delta ) / n }" + }, + { + "category_id": 13, + "poly": [ + 823, + 425, + 1138, + 425, + 1138, + 463, + 823, + 463 + ], + "score": 0.74, + "latex": "\\xi _ { 2 } = C ^ { \\prime \\prime \\prime } M ^ { 2 } \\sqrt { \\log ( r _ { k } / \\delta ) / n }" + }, + { + "category_id": 13, + "poly": [ + 1166, + 356, + 1200, + 356, + 1200, + 383, + 1166, + 383 + ], + "score": 0.34, + "latex": "\\mathtt { S u }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 645.0, + 1030.0, + 645.0, + 1069.0, + 295.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 222.0, + 817.0, + 222.0, + 817.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 968.0, + 1403.0, + 968.0, + 1403.0, + 995.0, + 1379.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 645.0, + 292.0, + 645.0, + 329.0, + 295.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 350.0, + 1165.0, + 350.0, + 1165.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 350.0, + 1403.0, + 350.0, + 1403.0, + 389.0, + 1201.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 385.0, + 709.0, + 385.0, + 709.0, + 424.0, + 291.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 385.0, + 1055.0, + 385.0, + 1055.0, + 424.0, + 972.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 385.0, + 1206.0, + 385.0, + 1206.0, + 424.0, + 1195.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 385.0, + 1406.0, + 385.0, + 1406.0, + 424.0, + 1393.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 423.0, + 342.0, + 423.0, + 342.0, + 465.0, + 291.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 423.0, + 822.0, + 423.0, + 822.0, + 465.0, + 812.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 423.0, + 1406.0, + 423.0, + 1406.0, + 465.0, + 1139.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 458.0, + 1046.0, + 458.0, + 1046.0, + 493.0, + 295.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1242.0, + 391.0, + 1242.0, + 391.0, + 1287.0, + 291.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1242.0, + 622.0, + 1242.0, + 622.0, + 1287.0, + 479.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1242.0, + 1060.0, + 1242.0, + 1060.0, + 1287.0, + 686.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1242.0, + 1406.0, + 1242.0, + 1406.0, + 1287.0, + 1244.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1278.0, + 366.0, + 1278.0, + 366.0, + 1320.0, + 291.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1089.0, + 1018.0, + 1089.0, + 1018.0, + 1131.0, + 293.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 562.0, + 481.0, + 562.0, + 481.0, + 610.0, + 292.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 562.0, + 1058.0, + 562.0, + 1058.0, + 610.0, + 780.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 961.0, + 616.0, + 961.0, + 616.0, + 1001.0, + 295.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 961.0, + 694.0, + 961.0, + 694.0, + 1001.0, + 645.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 961.0, + 958.0, + 961.0, + 958.0, + 1001.0, + 723.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1891.0, + 341.0, + 1891.0, + 341.0, + 1931.0, + 294.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1891.0, + 691.0, + 1891.0, + 691.0, + 1931.0, + 401.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 708.0, + 1426.0, + 708.0, + 1467.0, + 294.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1610.0, + 376.0, + 1610.0, + 376.0, + 1648.0, + 293.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 680.0, + 413.0, + 680.0, + 413.0, + 714.0, + 297.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 645.0, + 292.0, + 645.0, + 329.0, + 295.0, + 329.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 301, + 868, + 1404, + 868, + 1404, + 967, + 301, + 967 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 300, + 1405, + 1398, + 1405, + 1398, + 1651, + 300, + 1651 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 440, + 579, + 1261, + 579, + 1261, + 856, + 440, + 856 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 294, + 1749, + 1481, + 1749, + 1481, + 2043, + 294, + 2043 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 294, + 226, + 1403, + 226, + 1403, + 299, + 294, + 299 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 703, + 315, + 996, + 315, + 996, + 366, + 703, + 366 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 377, + 1116, + 1322, + 1116, + 1322, + 1213, + 377, + 1213 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 295, + 1663, + 1407, + 1663, + 1407, + 1735, + 295, + 1735 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 1224, + 1401, + 1224, + 1401, + 1288, + 296, + 1288 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 363, + 432, + 1333, + 432, + 1333, + 515, + 363, + 515 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 296, + 1064, + 1056, + 1064, + 1056, + 1103, + 296, + 1103 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 296, + 378, + 601, + 378, + 601, + 414, + 296, + 414 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 295, + 531, + 507, + 531, + 507, + 565, + 295, + 565 + ], + "score": 0.925 + }, + { + "category_id": 8, + "poly": [ + 438, + 1297, + 1258, + 1297, + 1258, + 1345, + 438, + 1345 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 298, + 1002, + 641, + 1002, + 641, + 1034, + 298, + 1034 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 295, + 1359, + 1090, + 1359, + 1090, + 1394, + 295, + 1394 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 14, + "poly": [ + 440, + 580, + 1258, + 580, + 1258, + 859, + 440, + 859 + ], + "score": 0.95, + "latex": "\\begin{array} { r l r } { { I _ { 3 , j , t } \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { ( t + 1 ) } - \\mathbf { W } _ { 1 } ^ { ( t ) } \\| _ { 2 } } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\sqrt { m } \\theta ) \\cdot \\frac { 2 \\eta } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\cdot \\theta \\cdot \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) \\| _ { 2 } } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\cfrac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | } \\\\ & { } & { \\leqslant \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 1405, + 1403, + 1405, + 1403, + 1653, + 310, + 1653 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } { \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\widehat { \\lambda } _ { r _ { k } } ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot ( \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } ) } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) \\| \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } & { } \\\\ + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 363, + 429, + 1334, + 429, + 1334, + 517, + 363, + 517 + ], + "score": 0.94, + "latex": "| I _ { 2 , j , t } | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } | y _ { i } - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\mathbf { y } - \\widehat { \\mathbf { y } } ^ { ( t ) } \\| _ { 2 } / \\sqrt { n } ." + }, + { + "category_id": 14, + "poly": [ + 375, + 1113, + 1324, + 1113, + 1324, + 1217, + 375, + 1217 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } [ { \\mathbf { I } } - ( \\eta m \\theta ^ { 2 } / n ) { \\mathbf { K } } ^ { \\infty } ] { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\| ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\sqrt { n } \\cdot \\widetilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 1748, + 1428, + 1748, + 1428, + 2049, + 310, + 2049 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } = ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & { \\phantom { \\hat { \\mathbf { V } } } + \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\tilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t } \\| \\hat { \\mathbf { V } } _ { r _ { k } } ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\lambda _ { r _ { k } } ^ { - 1 } \\cdot \\omega ^ { 2 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\tilde { \\mathcal { O } } ( 1 + \\omega \\sqrt { m } ) } \\\\ & \\phantom { \\hat { \\mathbf { V } } } + \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 372, + 380, + 474, + 380, + 474, + 414, + 372, + 414 + ], + "score": 0.92, + "latex": "i , j \\in [ n ]" + }, + { + "category_id": 14, + "poly": [ + 701, + 313, + 997, + 313, + 997, + 365, + 701, + 365 + ], + "score": 0.92, + "latex": "\\big | \\mathbf { K } _ { i , j } ^ { \\infty } - \\mathbf { K } _ { i , j } ^ { ( 0 ) } \\big | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } )" + }, + { + "category_id": 13, + "poly": [ + 342, + 533, + 400, + 533, + 400, + 567, + 342, + 567 + ], + "score": 0.92, + "latex": "I _ { 3 , j , t }" + }, + { + "category_id": 13, + "poly": [ + 1127, + 902, + 1354, + 902, + 1354, + 938, + 1127, + 938 + ], + "score": 0.91, + "latex": "I _ { 1 , j , t } + I _ { 2 , j , t } + I _ { 3 , j , t }" + }, + { + "category_id": 13, + "poly": [ + 337, + 1360, + 480, + 1360, + 480, + 1393, + 337, + 1393 + ], + "score": 0.91, + "latex": "t ^ { \\prime } = 0 , \\ldots , t" + }, + { + "category_id": 13, + "poly": [ + 339, + 1665, + 542, + 1665, + 542, + 1696, + 339, + 1696 + ], + "score": 0.9, + "latex": "t ^ { \\prime } = 0 , \\ldots , t - 1" + }, + { + "category_id": 13, + "poly": [ + 567, + 227, + 896, + 227, + 896, + 264, + 567, + 264 + ], + "score": 0.9, + "latex": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 954, + 231, + 1011, + 231, + 1011, + 265, + 954, + 265 + ], + "score": 0.9, + "latex": "I _ { 2 , j , t }" + }, + { + "category_id": 14, + "poly": [ + 439, + 1296, + 1256, + 1296, + 1256, + 1345, + 439, + 1345 + ], + "score": 0.89, + "latex": "\\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\leqslant \\| ( \\widehat { \\mathbf { V } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } \\mathbf { u } ^ { ( 0 ) } \\| _ { 2 } + t ^ { \\prime } \\cdot \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\cdot \\sqrt { n } \\cdot \\widetilde { O } ( 1 + \\omega \\sqrt { m } ) ," + }, + { + "category_id": 13, + "poly": [ + 1051, + 899, + 1094, + 899, + 1094, + 931, + 1051, + 931 + ], + "score": 0.89, + "latex": "\\mathbf { e } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 629, + 1065, + 807, + 1065, + 807, + 1102, + 629, + 1102 + ], + "score": 0.87, + "latex": "\\mathbf { u } ^ { ( t ) } = \\mathbf { y } - \\hat { \\mathbf { y } } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 650, + 265, + 991, + 265, + 991, + 299, + 650, + 299 + ], + "score": 0.87, + "latex": "1 - \\mathcal { O } ( n ^ { 2 } ) \\cdot \\exp ( - \\Omega ( m \\omega ^ { 2 / 3 } ) )" + }, + { + "category_id": 13, + "poly": [ + 907, + 904, + 924, + 904, + 924, + 935, + 907, + 935 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1666, + 1405, + 1666, + 1405, + 1696, + 1345, + 1696 + ], + "score": 0.82, + "latex": "\\omega \\ \\leqslant" + }, + { + "category_id": 13, + "poly": [ + 818, + 1068, + 882, + 1068, + 882, + 1099, + 818, + 1099 + ], + "score": 0.81, + "latex": "t \\in T" + }, + { + "category_id": 13, + "poly": [ + 395, + 1696, + 541, + 1696, + 541, + 1737, + 395, + 1737 + ], + "score": 0.74, + "latex": "n \\geqslant \\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { - 2 } )" + }, + { + "category_id": 13, + "poly": [ + 299, + 1696, + 383, + 1696, + 383, + 1737, + 299, + 1737 + ], + "score": 0.34, + "latex": "\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1000.0, + 645.0, + 1000.0, + 645.0, + 1036.0, + 294.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 905.0, + 295.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 893.0, + 906.0, + 893.0, + 906.0, + 944.0, + 292.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 893.0, + 1050.0, + 893.0, + 1050.0, + 944.0, + 925.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 893.0, + 1126.0, + 893.0, + 1126.0, + 944.0, + 1095.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 893.0, + 1408.0, + 893.0, + 1408.0, + 944.0, + 1355.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 931.0, + 873.0, + 931.0, + 873.0, + 970.0, + 296.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 936.0, + 1403.0, + 936.0, + 1403.0, + 964.0, + 1376.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 566.0, + 225.0, + 566.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 225.0, + 953.0, + 225.0, + 953.0, + 269.0, + 897.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 225.0, + 1405.0, + 225.0, + 1405.0, + 269.0, + 1012.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 649.0, + 261.0, + 649.0, + 303.0, + 293.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 261.0, + 1105.0, + 261.0, + 1105.0, + 303.0, + 992.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1659.0, + 338.0, + 1659.0, + 338.0, + 1703.0, + 293.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1659.0, + 1344.0, + 1659.0, + 1344.0, + 1703.0, + 543.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1694.0, + 298.0, + 1694.0, + 298.0, + 1739.0, + 294.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1694.0, + 394.0, + 1694.0, + 394.0, + 1739.0, + 384.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1694.0, + 1182.0, + 1694.0, + 1182.0, + 1739.0, + 542.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1223.0, + 1403.0, + 1223.0, + 1403.0, + 1259.0, + 296.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1253.0, + 511.0, + 1253.0, + 511.0, + 1289.0, + 296.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1061.0, + 628.0, + 1061.0, + 628.0, + 1108.0, + 293.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1061.0, + 817.0, + 1061.0, + 817.0, + 1108.0, + 808.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1061.0, + 1058.0, + 1061.0, + 1058.0, + 1108.0, + 883.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 375.0, + 371.0, + 375.0, + 371.0, + 417.0, + 295.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 375.0, + 603.0, + 375.0, + 603.0, + 417.0, + 475.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 527.0, + 341.0, + 527.0, + 341.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 527.0, + 510.0, + 527.0, + 510.0, + 570.0, + 401.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 336.0, + 1358.0, + 336.0, + 1398.0, + 295.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1358.0, + 1090.0, + 1358.0, + 1090.0, + 1398.0, + 481.0, + 1398.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1295, + 1405, + 1295, + 1405, + 1397, + 296, + 1397 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 296, + 805, + 1403, + 805, + 1403, + 899, + 296, + 899 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 299, + 1542, + 1396, + 1542, + 1396, + 1639, + 299, + 1639 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 686, + 1645, + 1013, + 1645, + 1013, + 1687, + 686, + 1687 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 292, + 1121, + 1399, + 1121, + 1399, + 1194, + 292, + 1194 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 690, + 1201, + 1008, + 1201, + 1008, + 1246, + 690, + 1246 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 296, + 294, + 1448, + 294, + 1448, + 535, + 296, + 535 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 299, + 1704, + 1402, + 1704, + 1402, + 1772, + 299, + 1772 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 502, + 1021, + 1189, + 1021, + 1189, + 1113, + 502, + 1113 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 296, + 540, + 1400, + 540, + 1400, + 610, + 296, + 610 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 293, + 226, + 1400, + 226, + 1400, + 294, + 293, + 294 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 296, + 1779, + 1400, + 1779, + 1400, + 1852, + 296, + 1852 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 1467, + 1396, + 1467, + 1396, + 1539, + 298, + 1539 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 859, + 74, + 859, + 106, + 297, + 106 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 299, + 739, + 661, + 739, + 661, + 776, + 299, + 776 + ], + "score": 0.918 + }, + { + "category_id": 8, + "poly": [ + 336, + 1406, + 1359, + 1406, + 1359, + 1459, + 336, + 1459 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 296, + 1254, + 669, + 1254, + 669, + 1289, + 296, + 1289 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 298, + 925, + 625, + 925, + 625, + 957, + 298, + 957 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 296, + 985, + 383, + 985, + 383, + 1015, + 296, + 1015 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 289, + 1858, + 1401, + 1858, + 1401, + 1896, + 289, + 1896 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 296, + 1999, + 708, + 1999, + 708, + 2036, + 296, + 2036 + ], + "score": 0.801 + }, + { + "category_id": 1, + "poly": [ + 295, + 619, + 1546, + 619, + 1546, + 704, + 295, + 704 + ], + "score": 0.675 + }, + { + "category_id": 2, + "poly": [ + 1375, + 672, + 1402, + 672, + 1402, + 699, + 1375, + 699 + ], + "score": 0.342 + }, + { + "category_id": 8, + "poly": [ + 396, + 1903, + 1297, + 1903, + 1297, + 1993, + 396, + 1993 + ], + "score": 0.334 + }, + { + "category_id": 13, + "poly": [ + 372, + 1999, + 572, + 1999, + 572, + 2036, + 372, + 2036 + ], + "score": 0.94, + "latex": "\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1573, + 1280, + 1573, + 1280, + 1609, + 1069, + 1609 + ], + "score": 0.93, + "latex": "\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 481, + 1858, + 731, + 1858, + 731, + 1895, + 481, + 1895 + ], + "score": 0.93, + "latex": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )" + }, + { + "category_id": 13, + "poly": [ + 785, + 1121, + 1049, + 1121, + 1049, + 1158, + 785, + 1158 + ], + "score": 0.92, + "latex": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 / 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1327, + 1233, + 1327, + 1233, + 1362, + 1150, + 1362 + ], + "score": 0.92, + "latex": "i \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 298, + 572, + 383, + 572, + 383, + 613, + 298, + 613 + ], + "score": 0.92, + "latex": "\\widetilde { \\mathcal { O } } ( \\lambda _ { r _ { k } } ^ { 3 } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1737, + 434, + 1737, + 434, + 1775, + 298, + 1775 + ], + "score": 0.92, + "latex": "B ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 14, + "poly": [ + 503, + 1019, + 1193, + 1019, + 1193, + 1116, + 503, + 1116 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { \\mathbf { D } _ { i } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) , } \\\\ & { \\mathbf { D } _ { i } ^ { ( 0 ) } = \\operatorname { d i a g } \\bigl ( \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { 1 } > 0 \\} , \\ldots , \\mathbb { 1 } \\{ ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) _ { m } > 0 \\} \\bigr ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1175, + 224, + 1393, + 224, + 1393, + 264, + 1175, + 264 + ], + "score": 0.92, + "latex": "\\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } \\leqslant \\widetilde { \\mathcal { O } } ( 1 )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1606, + 451, + 1606, + 451, + 1640, + 298, + 1640 + ], + "score": 0.92, + "latex": "C [ \\log ( m ) ] ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1360, + 434, + 1360, + 434, + 1397, + 297, + 1397 + ], + "score": 0.92, + "latex": "B ( \\mathbf { W } ^ { ( \\mathbf { \\theta } ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 622, + 2002, + 700, + 2002, + 700, + 2036, + 622, + 2036 + ], + "score": 0.92, + "latex": "i \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 985, + 1816, + 1267, + 1816, + 1267, + 1852, + 985, + 1852 + ], + "score": 0.92, + "latex": "\\mathbf { W } \\in { \\boldsymbol { B } } ( \\mathbf { W } ^ { ( 0 ) } , \\omega ) \\ i \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 338, + 541, + 542, + 541, + 542, + 574, + 338, + 574 + ], + "score": 0.92, + "latex": "t ^ { \\prime } = 0 , \\ldots , t - 1" + }, + { + "category_id": 14, + "poly": [ + 335, + 1402, + 1363, + 1402, + 1363, + 1462, + 335, + 1462 + ], + "score": 0.92, + "latex": "| f _ { \\mathbf { W } ^ { \\prime } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\mathbf { W } ^ { \\prime } - \\mathbf { W } \\rangle | \\leqslant { \\mathcal { O } } \\Bigl ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } \\Bigr ) \\cdot \\| \\mathbf { W } _ { 1 } ^ { \\prime } - \\mathbf { W } _ { 1 } \\| _ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 932, + 1780, + 1164, + 1780, + 1164, + 1816, + 932, + 1816 + ], + "score": 0.91, + "latex": "\\omega \\leqslant \\mathcal { O } ( [ \\log ( m ) ] ^ { - 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1503, + 1328, + 1503, + 1328, + 1538, + 1250, + 1538 + ], + "score": 0.91, + "latex": "i \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 388, + 1573, + 725, + 1573, + 725, + 1609, + 388, + 1609 + ], + "score": 0.91, + "latex": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 491, + 1815, + 907, + 1815, + 907, + 1853, + 491, + 1853 + ], + "score": 0.91, + "latex": "| f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } )" + }, + { + "category_id": 14, + "poly": [ + 310, + 288, + 1427, + 288, + 1427, + 540, + 310, + 540 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { \\mathbf { u } ^ { ( t ^ { \\prime } + 1 ) } \\| _ { 2 } \\leqslant \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\| [ \\mathbf { I } - ( \\eta m \\theta ^ { 2 } / n ) \\mathbf { K } ^ { \\infty } ] \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ { \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\hat { \\lambda } _ { \\mathbf { r } _ { k } } ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( \\omega ^ { 1 / 3 } \\eta m \\theta ^ { 2 } ) \\cdot \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } + \\widetilde { \\mathcal { O } } ( 1 ) \\cdot \\| ( \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\bot } ) ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } } & { } \\\\ \\leqslant ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { \\mathbf { r } _ { k } } / 2 ) \\| \\hat { \\mathbf { V } } _ { \\mathbf { r } _ { k } } ^ { \\top } \\mathbf { u } ^ { ( t ^ { \\prime } ) } \\| _ { 2 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1067, + 1858, + 1393, + 1858, + 1393, + 1895, + 1067, + 1895 + ], + "score": 0.91, + "latex": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 606, + 1779, + 878, + 1779, + 878, + 1814, + 606, + 1814 + ], + "score": 0.9, + "latex": "m \\geqslant \\Omega ( \\omega ^ { - 2 / 3 } \\log ( n / \\delta ) )" + }, + { + "category_id": 13, + "poly": [ + 799, + 1470, + 867, + 1470, + 867, + 1499, + 799, + 1499 + ], + "score": 0.9, + "latex": "\\delta > 0" + }, + { + "category_id": 14, + "poly": [ + 688, + 1197, + 1010, + 1197, + 1010, + 1245, + 688, + 1245 + ], + "score": 0.9, + "latex": "\\| \\mathbf { D } _ { i } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { 0 } \\leqslant \\mathcal { O } ( \\omega ^ { 2 / 3 } m )" + }, + { + "category_id": 13, + "poly": [ + 495, + 1361, + 718, + 1361, + 718, + 1397, + 495, + 1397 + ], + "score": 0.9, + "latex": "\\omega \\leqslant \\kappa [ \\log ( m ) ] ^ { - 3 / 2 }" + }, + { + "category_id": 13, + "poly": [ + 829, + 1502, + 1171, + 1502, + 1171, + 1539, + 829, + 1539 + ], + "score": 0.89, + "latex": "| f _ { \\mathbf { W } ^ { ( 0 ) } } ( { \\pmb x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } )" + }, + { + "category_id": 13, + "poly": [ + 903, + 1469, + 1091, + 1469, + 1091, + 1502, + 903, + 1502 + ], + "score": 0.89, + "latex": "m \\geqslant C \\log ( n / \\delta )" + }, + { + "category_id": 14, + "poly": [ + 683, + 1643, + 1015, + 1643, + 1015, + 1684, + 683, + 1684 + ], + "score": 0.89, + "latex": "\\| \\nabla _ { \\mathbf { W } _ { l } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\sqrt { m } ) ." + }, + { + "category_id": 13, + "poly": [ + 995, + 1325, + 1060, + 1325, + 1060, + 1357, + 995, + 1357 + ], + "score": 0.89, + "latex": "\\mathbf { W } ^ { ( 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1155, + 628, + 1155, + 628, + 1194, + 296, + 1194 + ], + "score": 0.89, + "latex": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 385, + 1327, + 719, + 1327, + 719, + 1362, + 385, + 1362 + ], + "score": 0.88, + "latex": "1 - \\mathcal { O } ( n ) \\cdot \\exp [ - \\Omega ( m \\omega ^ { 2 / 3 } ) ]" + }, + { + "category_id": 13, + "poly": [ + 821, + 1574, + 1012, + 1574, + 1012, + 1609, + 821, + 1609 + ], + "score": 0.86, + "latex": "i \\in [ n ] , l \\in [ L ]" + }, + { + "category_id": 14, + "poly": [ + 399, + 1900, + 1296, + 1900, + 1296, + 1993, + 399, + 1993 + ], + "score": 0.84, + "latex": "\\begin{array} { r l } & { \\| \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } ) , } \\\\ & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } m ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1544, + 1103, + 1544, + 1103, + 1570, + 1079, + 1570 + ], + "score": 0.84, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1577, + 1404, + 1577, + 1404, + 1607, + 1345, + 1607 + ], + "score": 0.82, + "latex": "\\omega \\ \\leqslant" + }, + { + "category_id": 13, + "poly": [ + 371, + 1254, + 571, + 1254, + 571, + 1291, + 371, + 1291 + ], + "score": 0.82, + "latex": "\\mathbf { W } \\in { \\cal B } ( \\mathbf { W } ^ { ( 0 ) } , \\omega )" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1328, + 1404, + 1328, + 1404, + 1360, + 1287, + 1360 + ], + "score": 0.82, + "latex": "\\mathbf { W } , \\mathbf { W } ^ { \\prime } \\in" + }, + { + "category_id": 13, + "poly": [ + 397, + 1506, + 421, + 1506, + 421, + 1532, + 397, + 1532 + ], + "score": 0.82, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1345, + 543, + 1405, + 543, + 1405, + 572, + 1345, + 572 + ], + "score": 0.8, + "latex": "\\omega \\ \\leqslant" + }, + { + "category_id": 13, + "poly": [ + 1083, + 1302, + 1102, + 1302, + 1102, + 1323, + 1083, + 1323 + ], + "score": 0.79, + "latex": "\\kappa" + }, + { + "category_id": 14, + "poly": [ + 422, + 618, + 1539, + 618, + 1539, + 666, + 422, + 666 + ], + "score": 0.74, + "latex": "\\mathscr { O } ( \\sqrt { n } ) \\cdot ( 1 - \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } / 2 ) ^ { t ^ { \\prime } } + \\widetilde { \\mathscr { O } } ( ( \\eta m \\theta ^ { 2 } \\lambda _ { r _ { k } } ) ^ { - 1 } ) \\cdot \\Vert ( \\widehat { { \\mathbf { V } } } _ { r _ { k } } ^ { \\perp } ) ^ { \\top } { \\mathbf { u } } ^ { ( 0 ) } \\Vert _ { 2 } + \\lambda _ { r _ { k } } ^ { - 1 } t ^ { \\prime } \\omega ^ { 1 / 3 } \\sqrt { n } \\cdot \\widetilde { \\mathscr { O } } ( 1 + \\omega \\sqrt { m } ) ." + }, + { + "category_id": 13, + "poly": [ + 583, + 1256, + 660, + 1256, + 660, + 1291, + 583, + 1291 + ], + "score": 0.58, + "latex": "i \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 756, + 1504, + 820, + 1504, + 820, + 1536, + 756, + 1536 + ], + "score": 0.54, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 419, + 1817, + 482, + 1817, + 482, + 1848, + 419, + 1848 + ], + "score": 0.44, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 819, + 1575, + 908, + 1575, + 908, + 1609, + 819, + 1609 + ], + "score": 0.29, + "latex": "i \\in [ n ]" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 737.0, + 663.0, + 737.0, + 663.0, + 780.0, + 293.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 922.0, + 627.0, + 922.0, + 627.0, + 961.0, + 294.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 676.0, + 1402.0, + 676.0, + 1402.0, + 701.0, + 1379.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1292.0, + 1082.0, + 1292.0, + 1082.0, + 1330.0, + 294.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1292.0, + 1403.0, + 1292.0, + 1403.0, + 1330.0, + 1103.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1322.0, + 384.0, + 1322.0, + 384.0, + 1365.0, + 291.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1322.0, + 994.0, + 1322.0, + 994.0, + 1365.0, + 720.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1322.0, + 1149.0, + 1322.0, + 1149.0, + 1365.0, + 1061.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1322.0, + 1286.0, + 1322.0, + 1286.0, + 1365.0, + 1234.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1357.0, + 494.0, + 1357.0, + 494.0, + 1400.0, + 435.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1357.0, + 988.0, + 1357.0, + 988.0, + 1400.0, + 719.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 802.0, + 1404.0, + 802.0, + 1404.0, + 841.0, + 293.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 837.0, + 1404.0, + 837.0, + 1404.0, + 868.0, + 296.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 865.0, + 723.0, + 865.0, + 723.0, + 899.0, + 294.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1536.0, + 1078.0, + 1536.0, + 1078.0, + 1578.0, + 293.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1536.0, + 1401.0, + 1536.0, + 1401.0, + 1578.0, + 1104.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1571.0, + 387.0, + 1571.0, + 387.0, + 1613.0, + 293.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1571.0, + 818.0, + 1571.0, + 818.0, + 1613.0, + 726.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1571.0, + 1068.0, + 1571.0, + 1068.0, + 1613.0, + 1013.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1571.0, + 1344.0, + 1571.0, + 1344.0, + 1613.0, + 1281.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1604.0, + 720.0, + 1604.0, + 720.0, + 1642.0, + 452.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1117.0, + 784.0, + 1117.0, + 784.0, + 1163.0, + 292.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1117.0, + 1408.0, + 1117.0, + 1408.0, + 1163.0, + 1050.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1155.0, + 295.0, + 1155.0, + 295.0, + 1194.0, + 292.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1155.0, + 639.0, + 1155.0, + 639.0, + 1194.0, + 629.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1700.0, + 1406.0, + 1700.0, + 1406.0, + 1743.0, + 293.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1730.0, + 297.0, + 1730.0, + 297.0, + 1780.0, + 292.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1730.0, + 451.0, + 1730.0, + 451.0, + 1780.0, + 435.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 337.0, + 535.0, + 337.0, + 579.0, + 292.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 535.0, + 1344.0, + 535.0, + 1344.0, + 579.0, + 543.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 571.0, + 1089.0, + 571.0, + 1089.0, + 616.0, + 384.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 219.0, + 1174.0, + 219.0, + 1174.0, + 271.0, + 291.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 219.0, + 1404.0, + 219.0, + 1404.0, + 271.0, + 1394.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 254.0, + 416.0, + 254.0, + 416.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1774.0, + 605.0, + 1774.0, + 605.0, + 1821.0, + 291.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1774.0, + 931.0, + 1774.0, + 931.0, + 1821.0, + 879.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1821.0, + 1165.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1809.0, + 418.0, + 1809.0, + 418.0, + 1857.0, + 291.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1809.0, + 490.0, + 1809.0, + 490.0, + 1857.0, + 483.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1809.0, + 984.0, + 1809.0, + 984.0, + 1857.0, + 908.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1809.0, + 1279.0, + 1809.0, + 1279.0, + 1857.0, + 1268.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1467.0, + 798.0, + 1467.0, + 798.0, + 1504.0, + 295.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1467.0, + 902.0, + 1467.0, + 902.0, + 1504.0, + 868.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1467.0, + 1404.0, + 1467.0, + 1404.0, + 1504.0, + 1092.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1503.0, + 396.0, + 1503.0, + 396.0, + 1541.0, + 294.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1503.0, + 755.0, + 1503.0, + 755.0, + 1541.0, + 422.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1503.0, + 828.0, + 1503.0, + 828.0, + 1541.0, + 821.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1503.0, + 1249.0, + 1503.0, + 1249.0, + 1541.0, + 1172.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1503.0, + 1337.0, + 1503.0, + 1337.0, + 1541.0, + 1329.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1245.0, + 370.0, + 1245.0, + 370.0, + 1301.0, + 291.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1245.0, + 582.0, + 1245.0, + 582.0, + 1301.0, + 572.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1245.0, + 674.0, + 1245.0, + 674.0, + 1301.0, + 661.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 384.0, + 984.0, + 384.0, + 1017.0, + 295.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1851.0, + 480.0, + 1851.0, + 480.0, + 1903.0, + 292.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1851.0, + 1066.0, + 1851.0, + 1066.0, + 1903.0, + 732.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1903.0, + 1394.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1991.0, + 371.0, + 1991.0, + 371.0, + 2045.0, + 291.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1991.0, + 621.0, + 1991.0, + 621.0, + 2045.0, + 573.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1991.0, + 713.0, + 1991.0, + 713.0, + 2045.0, + 701.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 617.0, + 421.0, + 617.0, + 421.0, + 670.0, + 291.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1540.0, + 617.0, + 1544.0, + 617.0, + 1544.0, + 670.0, + 1540.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 560.0, + 666.0, + 560.0, + 706.0, + 294.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 675.0, + 1402.0, + 675.0, + 1402.0, + 698.0, + 1376.0, + 698.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1074, + 1404, + 1074, + 1404, + 1188, + 297, + 1188 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1909, + 1404, + 1909, + 1404, + 2036, + 298, + 2036 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 296, + 335, + 1426, + 335, + 1426, + 472, + 296, + 472 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 314, + 1574, + 1386, + 1574, + 1386, + 1700, + 314, + 1700 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 299, + 872, + 1403, + 872, + 1403, + 1065, + 299, + 1065 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 390, + 565, + 1307, + 565, + 1307, + 654, + 390, + 654 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 296, + 1253, + 1400, + 1253, + 1400, + 1318, + 296, + 1318 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 296, + 485, + 1405, + 485, + 1405, + 552, + 296, + 552 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 556, + 1201, + 1144, + 1201, + 1144, + 1244, + 556, + 1244 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 422, + 776, + 1270, + 776, + 1270, + 824, + 422, + 824 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 1712, + 556, + 1712, + 556, + 1745, + 298, + 1745 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 297, + 730, + 886, + 730, + 886, + 764, + 297, + 764 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 299, + 665, + 584, + 665, + 584, + 698, + 299, + 698 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 298, + 1528, + 1166, + 1528, + 1166, + 1563, + 298, + 1563 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 298, + 1482, + 819, + 1482, + 819, + 1515, + 298, + 1515 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 296, + 286, + 933, + 286, + 933, + 323, + 296, + 323 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 298, + 833, + 510, + 833, + 510, + 864, + 298, + 864 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 298, + 1783, + 915, + 1783, + 915, + 1822, + 298, + 1822 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 297, + 1378, + 978, + 1378, + 978, + 1412, + 297, + 1412 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 299, + 1851, + 1075, + 1851, + 1075, + 1885, + 299, + 1885 + ], + "score": 0.892 + }, + { + "category_id": 8, + "poly": [ + 352, + 1426, + 1344, + 1426, + 1344, + 1473, + 352, + 1473 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.864 + }, + { + "category_id": 8, + "poly": [ + 404, + 1324, + 1281, + 1324, + 1281, + 1370, + 404, + 1370 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 1374, + 665, + 1404, + 665, + 1404, + 696, + 1374, + 696 + ], + "score": 0.821 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1714, + 1403, + 1714, + 1403, + 1744, + 1374, + 1744 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 783, + 229, + 783, + 262, + 300, + 262 + ], + "score": 0.561 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 783, + 229, + 783, + 262, + 300, + 262 + ], + "score": 0.384 + }, + { + "category_id": 13, + "poly": [ + 821, + 486, + 1015, + 486, + 1015, + 522, + 821, + 522 + ], + "score": 0.93, + "latex": "\\omega \\leqslant [ \\log ( m ) ] ^ { - 3 }" + }, + { + "category_id": 14, + "poly": [ + 310, + 331, + 1425, + 331, + 1425, + 476, + 310, + 476 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { F } + \\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\| \\mathbf { W } _ { 2 } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\| _ { F } } \\\\ & { \\qquad + \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m \\log ( m ) } ) \\cdot \\| \\mathbf { W } _ { 1 } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\| _ { 2 } } \\\\ & { \\qquad \\leqslant \\mathcal { O } ( \\omega \\sqrt { m } ) , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 313, + 1571, + 1388, + 1571, + 1388, + 1706, + 313, + 1706 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle - m \\mathbf { K } ^ { ( 0 ) } | \\leqslant | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + | \\langle \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) , \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { j } ) - \\nabla _ { \\mathbf { W } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { j } ) \\rangle | } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\leqslant { \\mathcal { O } } ( \\omega ^ { 1 / 3 } m ) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 868, + 1407, + 868, + 1407, + 1071, + 310, + 1071 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } { | \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | | _ { F } = \\sqrt { m } \\cdot \\| \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } - \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { x } _ { i } \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { = \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } \\mathbf { D } _ { i } - \\mathbf { W } _ { 2 } ^ { ( 0 ) } \\mathbf { D } _ { i } ^ { ( 0 ) } \\| _ { F } } & { } \\\\ { \\leqslant \\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } + \\sqrt { m } \\cdot \\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } } & { } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1113, + 1171, + 1113, + 1171, + 1148, + 1078, + 1148 + ], + "score": 0.92, + "latex": "i \\in [ n ]" + }, + { + "category_id": 13, + "poly": [ + 382, + 1972, + 453, + 1972, + 453, + 2006, + 382, + 2006 + ], + "score": 0.92, + "latex": "P _ { k } ( { \\bf x } )" + }, + { + "category_id": 13, + "poly": [ + 371, + 1254, + 448, + 1254, + 448, + 1288, + 371, + 1288 + ], + "score": 0.92, + "latex": "i \\in [ n ]" + }, + { + "category_id": 14, + "poly": [ + 353, + 1424, + 1345, + 1424, + 1345, + 1471, + 353, + 1471 + ], + "score": 0.91, + "latex": "\\| \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant { \\sqrt { m } } \\cdot \\| \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } - \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } \\| _ { F } \\leqslant \\omega { \\sqrt { m } } \\leqslant \\omega ^ { 1 / 3 } { \\sqrt { m } } ." + }, + { + "category_id": 13, + "poly": [ + 718, + 1942, + 1030, + 1942, + 1030, + 1976, + 718, + 1976 + ], + "score": 0.91, + "latex": "r ( \\mathbf { x } ) = f ^ { * } ( \\mathbf { x } ) - \\theta f _ { \\mathbf { W } ^ { ( t ) } } ( \\mathbf { x } )" + }, + { + "category_id": 13, + "poly": [ + 875, + 1379, + 923, + 1379, + 923, + 1413, + 875, + 1413 + ], + "score": 0.91, + "latex": "\\sigma ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 417, + 1321, + 1282, + 1321, + 1282, + 1368, + 417, + 1368 + ], + "score": 0.91, + "latex": "\\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } ^ { ( 0 ) } \\mathbf { x } _ { i } ) ] ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 2 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot [ \\sigma ( \\mathbf { W } _ { 1 } \\mathbf { x } _ { i } ) ] ^ { \\top }" + }, + { + "category_id": 14, + "poly": [ + 391, + 560, + 1308, + 560, + 1308, + 658, + 391, + 658 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) | \\leqslant | f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | + | f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) | \\leqslant \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } ) + \\mathcal { O } ( \\omega \\sqrt { m } ) } \\\\ & { \\qquad = \\mathcal { O } ( \\sqrt { \\log ( n / \\delta ) } + \\omega \\sqrt { m } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1149, + 815, + 1149, + 815, + 1188, + 297, + 1188 + ], + "score": 0.89, + "latex": "\\| ( \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } ) \\mathbf { D } _ { i } \\| _ { F } \\leqslant \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } - \\mathbf { W } _ { 2 } \\| _ { F } \\leqslant \\omega" + }, + { + "category_id": 14, + "poly": [ + 425, + 775, + 1270, + 775, + 1270, + 823, + 425, + 823 + ], + "score": 0.89, + "latex": "\\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } ^ { ( 0 ) } \\mathbf { W } _ { 2 } ^ { ( 0 ) \\top } \\mathbf { x } _ { i } ^ { \\top } , \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) = \\sqrt { m } \\cdot \\mathbf { D } _ { i } \\mathbf { W } _ { 2 } ^ { \\top } \\mathbf { x } _ { i } ^ { \\top } ." + }, + { + "category_id": 13, + "poly": [ + 464, + 1106, + 989, + 1106, + 989, + 1149, + 464, + 1149 + ], + "score": 0.88, + "latex": "\\sqrt { m } \\cdot \\| \\mathbf { W } _ { 2 } ^ { ( 0 ) } ( \\mathbf { D } _ { i } ^ { ( 0 ) } - \\mathbf { D } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )" + }, + { + "category_id": 14, + "poly": [ + 555, + 1199, + 1144, + 1199, + 1144, + 1242, + 555, + 1242 + ], + "score": 0.87, + "latex": "\\| \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } } ( \\mathbf { x } _ { i } ) - \\nabla _ { \\mathbf { W } _ { 1 } } f _ { \\mathbf { W } ^ { ( 0 ) } } ( \\mathbf { x } _ { i } ) \\| _ { F } \\leqslant \\mathcal { O } ( \\omega ^ { 1 / 3 } \\sqrt { m } )" + }, + { + "category_id": 13, + "poly": [ + 1307, + 1077, + 1398, + 1077, + 1398, + 1106, + 1307, + 1106 + ], + "score": 0.79, + "latex": "1 - n \\cdot" + }, + { + "category_id": 13, + "poly": [ + 298, + 1111, + 449, + 1111, + 449, + 1148, + 298, + 1148 + ], + "score": 0.69, + "latex": "\\exp [ - \\Omega ( m ) ]" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1783.0, + 914.0, + 1783.0, + 914.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1077.0, + 1851.0, + 1077.0, + 1887.0, + 295.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1380.0, + 671.0, + 1402.0, + 671.0, + 1402.0, + 694.0, + 1380.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1719.0, + 1403.0, + 1719.0, + 1403.0, + 1746.0, + 1377.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 785.0, + 228.0, + 785.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1071.0, + 1306.0, + 1071.0, + 1306.0, + 1112.0, + 294.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1071.0, + 1404.0, + 1071.0, + 1404.0, + 1112.0, + 1399.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1095.0, + 297.0, + 1095.0, + 297.0, + 1158.0, + 287.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1095.0, + 463.0, + 1095.0, + 463.0, + 1158.0, + 450.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1095.0, + 1077.0, + 1095.0, + 1077.0, + 1158.0, + 990.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1095.0, + 1410.0, + 1095.0, + 1410.0, + 1158.0, + 1172.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1136.0, + 296.0, + 1136.0, + 296.0, + 1195.0, + 288.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1136.0, + 953.0, + 1136.0, + 953.0, + 1195.0, + 816.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 717.0, + 1939.0, + 717.0, + 1980.0, + 292.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1980.0, + 1031.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 381.0, + 1973.0, + 381.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2009.0, + 454.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1249.0, + 370.0, + 1249.0, + 370.0, + 1292.0, + 292.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1249.0, + 1404.0, + 1249.0, + 1404.0, + 1292.0, + 449.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1278.0, + 361.0, + 1278.0, + 361.0, + 1322.0, + 290.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 482.0, + 820.0, + 482.0, + 820.0, + 526.0, + 294.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 482.0, + 1406.0, + 482.0, + 1406.0, + 526.0, + 1016.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 513.0, + 557.0, + 513.0, + 557.0, + 555.0, + 293.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 559.0, + 1708.0, + 559.0, + 1749.0, + 295.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 728.0, + 887.0, + 728.0, + 887.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 661.0, + 586.0, + 661.0, + 586.0, + 703.0, + 295.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1525.0, + 1169.0, + 1525.0, + 1169.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1477.0, + 824.0, + 1477.0, + 824.0, + 1521.0, + 294.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 285.0, + 934.0, + 285.0, + 934.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 829.0, + 511.0, + 829.0, + 511.0, + 868.0, + 295.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 874.0, + 1374.0, + 874.0, + 1417.0, + 294.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1374.0, + 977.0, + 1374.0, + 977.0, + 1417.0, + 924.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 785.0, + 228.0, + 785.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 311, + 321, + 1386, + 321, + 1386, + 1232, + 311, + 1232 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1484, + 1405, + 1484, + 1405, + 1609, + 297, + 1609 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 295, + 1265, + 1405, + 1265, + 1405, + 1389, + 295, + 1389 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 295, + 229, + 1402, + 229, + 1402, + 294, + 295, + 294 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 297, + 1425, + 857, + 1425, + 857, + 1457, + 297, + 1457 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 866, + 2088, + 866, + 2113, + 834, + 2113 + ], + "score": 0.849 + }, + { + "category_id": 15, + "poly": [ + 419.0, + 318.0, + 758.0, + 318.0, + 758.0, + 354.0, + 419.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 318.0, + 1303.0, + 318.0, + 1303.0, + 354.0, + 965.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 344.0, + 927.0, + 344.0, + 927.0, + 371.0, + 888.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 352.0, + 380.0, + 352.0, + 380.0, + 380.0, + 343.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 361.0, + 1371.0, + 361.0, + 1371.0, + 396.0, + 1314.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 362.0, + 822.0, + 362.0, + 822.0, + 393.0, + 766.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 372.0, + 906.0, + 372.0, + 906.0, + 641.0, + 846.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 376.0, + 356.0, + 376.0, + 356.0, + 638.0, + 301.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 386.0, + 826.0, + 386.0, + 826.0, + 425.0, + 766.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 386.0, + 1372.0, + 386.0, + 1372.0, + 427.0, + 1311.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 393.0, + 924.0, + 393.0, + 924.0, + 419.0, + 888.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 402.0, + 376.0, + 402.0, + 376.0, + 425.0, + 343.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 417.0, + 825.0, + 417.0, + 825.0, + 452.0, + 768.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 417.0, + 1371.0, + 417.0, + 1371.0, + 452.0, + 1312.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 445.0, + 377.0, + 445.0, + 377.0, + 474.0, + 345.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 441.0, + 923.0, + 441.0, + 923.0, + 468.0, + 889.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 496.0, + 377.0, + 496.0, + 377.0, + 519.0, + 343.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 490.0, + 923.0, + 490.0, + 923.0, + 516.0, + 889.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 543.0, + 377.0, + 543.0, + 377.0, + 566.0, + 343.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 541.0, + 922.0, + 541.0, + 922.0, + 564.0, + 890.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 591.0, + 375.0, + 591.0, + 375.0, + 615.0, + 344.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 587.0, + 923.0, + 587.0, + 923.0, + 614.0, + 889.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 636.0, + 379.0, + 636.0, + 379.0, + 663.0, + 343.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 636.0, + 924.0, + 636.0, + 924.0, + 662.0, + 888.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 665.0, + 408.0, + 665.0, + 408.0, + 687.0, + 388.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 663.0, + 475.0, + 663.0, + 475.0, + 689.0, + 427.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 663.0, + 527.0, + 663.0, + 527.0, + 689.0, + 476.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 662.0, + 839.0, + 662.0, + 839.0, + 689.0, + 529.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 666.0, + 954.0, + 666.0, + 954.0, + 687.0, + 934.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 663.0, + 1020.0, + 663.0, + 1020.0, + 689.0, + 972.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 663.0, + 1072.0, + 663.0, + 1072.0, + 689.0, + 1024.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 662.0, + 1385.0, + 662.0, + 1385.0, + 689.0, + 1074.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 679.0, + 651.0, + 679.0, + 651.0, + 717.0, + 562.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 679.0, + 1197.0, + 679.0, + 1197.0, + 717.0, + 1108.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 728.0, + 743.0, + 728.0, + 743.0, + 762.0, + 413.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 728.0, + 1270.0, + 728.0, + 1270.0, + 762.0, + 979.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 790.0, + 752.0, + 790.0, + 752.0, + 825.0, + 412.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 790.0, + 1298.0, + 790.0, + 1298.0, + 825.0, + 959.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 827.0, + 934.0, + 827.0, + 934.0, + 1124.0, + 841.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 832.0, + 363.0, + 832.0, + 363.0, + 848.0, + 345.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 832.0, + 824.0, + 832.0, + 824.0, + 868.0, + 768.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 832.0, + 1370.0, + 832.0, + 1370.0, + 868.0, + 1314.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 849.0, + 360.0, + 849.0, + 360.0, + 1107.0, + 306.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 858.0, + 825.0, + 858.0, + 825.0, + 897.0, + 766.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 858.0, + 1372.0, + 858.0, + 1372.0, + 898.0, + 1312.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 888.0, + 824.0, + 888.0, + 824.0, + 922.0, + 769.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 911.0, + 905.0, + 911.0, + 905.0, + 922.0, + 895.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 888.0, + 1370.0, + 888.0, + 1370.0, + 922.0, + 1314.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 930.0, + 359.0, + 930.0, + 359.0, + 939.0, + 349.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 961.0, + 904.0, + 961.0, + 904.0, + 973.0, + 895.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 973.0, + 361.0, + 973.0, + 361.0, + 988.0, + 347.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1013.0, + 904.0, + 1013.0, + 904.0, + 1026.0, + 895.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1022.0, + 359.0, + 1022.0, + 359.0, + 1032.0, + 348.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1112.0, + 361.0, + 1112.0, + 361.0, + 1129.0, + 347.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1112.0, + 907.0, + 1112.0, + 907.0, + 1129.0, + 891.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1136.0, + 397.0, + 1136.0, + 397.0, + 1158.0, + 379.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1133.0, + 467.0, + 1133.0, + 467.0, + 1162.0, + 416.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1135.0, + 517.0, + 1135.0, + 517.0, + 1159.0, + 470.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1134.0, + 837.0, + 1134.0, + 837.0, + 1160.0, + 522.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1136.0, + 943.0, + 1136.0, + 943.0, + 1158.0, + 923.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1133.0, + 1015.0, + 1133.0, + 1015.0, + 1162.0, + 962.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1135.0, + 1063.0, + 1135.0, + 1063.0, + 1159.0, + 1016.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1134.0, + 1383.0, + 1134.0, + 1383.0, + 1160.0, + 1069.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1151.0, + 645.0, + 1151.0, + 645.0, + 1187.0, + 556.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1151.0, + 1191.0, + 1151.0, + 1191.0, + 1188.0, + 1102.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1198.0, + 741.0, + 1198.0, + 741.0, + 1233.0, + 414.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1197.0, + 1269.0, + 1197.0, + 1269.0, + 1231.0, + 977.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1264.0, + 1406.0, + 1264.0, + 1406.0, + 1301.0, + 293.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1295.0, + 1405.0, + 1295.0, + 1405.0, + 1331.0, + 293.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1327.0, + 1405.0, + 1327.0, + 1405.0, + 1363.0, + 292.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1359.0, + 491.0, + 1359.0, + 491.0, + 1392.0, + 294.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1426.0, + 859.0, + 1426.0, + 859.0, + 1457.0, + 296.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2123.0, + 828.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1484.0, + 1406.0, + 1484.0, + 1406.0, + 1517.0, + 294.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1516.0, + 1403.0, + 1516.0, + 1403.0, + 1549.0, + 295.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1579.0, + 294.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1575.0, + 1387.0, + 1575.0, + 1387.0, + 1616.0, + 291.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 257.0, + 605.0, + 257.0, + 605.0, + 299.0, + 294.0, + 299.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 311, + 835, + 1386, + 835, + 1386, + 1276, + 311, + 1276 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 296, + 1311, + 1405, + 1311, + 1405, + 1436, + 296, + 1436 + ], + "score": 0.95 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 859, + 75, + 859, + 106, + 297, + 106 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.83 + }, + { + "category_id": 13, + "poly": [ + 970, + 842, + 1307, + 842, + 1307, + 869, + 970, + 869 + ], + "score": 0.27, + "latex": "f ^ { * } ( x ) = 1 P _ { 1 } ( x ) + 3 P _ { 2 } ( x ) + 5 P _ { 4 } ( x )" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 832.0, + 762.0, + 832.0, + 762.0, + 870.0, + 424.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 873.0, + 386.0, + 873.0, + 386.0, + 901.0, + 349.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 877.0, + 824.0, + 877.0, + 824.0, + 911.0, + 769.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 862.0, + 933.0, + 862.0, + 933.0, + 890.0, + 895.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 877.0, + 1370.0, + 877.0, + 1370.0, + 911.0, + 1314.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 896.0, + 352.0, + 896.0, + 352.0, + 1150.0, + 309.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 905.0, + 824.0, + 905.0, + 824.0, + 938.0, + 768.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 894.0, + 933.0, + 894.0, + 933.0, + 925.0, + 891.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 905.0, + 1371.0, + 905.0, + 1371.0, + 940.0, + 1314.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 929.0, + 381.0, + 929.0, + 381.0, + 952.0, + 345.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 935.0, + 823.0, + 935.0, + 823.0, + 965.0, + 770.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 925.0, + 929.0, + 925.0, + 929.0, + 1150.0, + 858.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 934.0, + 1370.0, + 934.0, + 1370.0, + 967.0, + 1311.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 966.0, + 927.0, + 966.0, + 927.0, + 987.0, + 888.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 981.0, + 382.0, + 981.0, + 382.0, + 1004.0, + 347.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1035.0, + 928.0, + 1035.0, + 928.0, + 1057.0, + 891.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1088.0, + 381.0, + 1088.0, + 381.0, + 1106.0, + 344.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1103.0, + 926.0, + 1103.0, + 926.0, + 1121.0, + 889.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1136.0, + 385.0, + 1136.0, + 385.0, + 1162.0, + 344.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1136.0, + 927.0, + 1136.0, + 927.0, + 1158.0, + 891.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1182.0, + 418.0, + 1182.0, + 418.0, + 1202.0, + 398.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1179.0, + 839.0, + 1179.0, + 839.0, + 1205.0, + 432.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1182.0, + 964.0, + 1182.0, + 964.0, + 1202.0, + 944.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1179.0, + 1385.0, + 1179.0, + 1385.0, + 1205.0, + 979.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1197.0, + 655.0, + 1197.0, + 655.0, + 1231.0, + 567.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1196.0, + 1202.0, + 1196.0, + 1202.0, + 1232.0, + 1113.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1244.0, + 760.0, + 1244.0, + 760.0, + 1275.0, + 393.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1243.0, + 1306.0, + 1243.0, + 1306.0, + 1274.0, + 940.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 906.0, + 912.0, + 906.0, + 912.0, + 935.0, + 846.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 999.5, + 931.0, + 999.5, + 931.0, + 1021.5, + 889.0, + 1021.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1031.5, + 386.0, + 1031.5, + 386.0, + 1055.5, + 339.0, + 1055.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1068.0, + 929.0, + 1068.0, + 929.0, + 1090.0, + 889.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1313.0, + 1403.0, + 1313.0, + 1403.0, + 1345.0, + 295.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1342.0, + 1405.0, + 1342.0, + 1405.0, + 1375.0, + 296.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1372.0, + 1405.0, + 1372.0, + 1405.0, + 1407.0, + 294.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1401.0, + 560.0, + 1401.0, + 560.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/BJxLH2AcYX/BJxLH2AcYX.md b/parse/train/BJxLH2AcYX/BJxLH2AcYX.md new file mode 100644 index 0000000000000000000000000000000000000000..34b6af46a06593beea7f0fce36749f41f51b5934 --- /dev/null +++ b/parse/train/BJxLH2AcYX/BJxLH2AcYX.md @@ -0,0 +1,369 @@ +# UNSUPERVISED MULTI-TARGET DOMAIN ADAPTATION:AN INFORMATION THEORETIC APPROACH + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Unsupervised domain adaptation (uDA) models focus on pairwise adaptation settings where there is a single, labeled, source and a single target domain. However, in many real-world settings one seeks to adapt to multiple, but somewhat similar, target domains. Applying pairwise adaptation approaches to this setting may be suboptimal, as they fail to leverage shared information among multiple domains. In this work, we propose an information theoretic approach for domain adaptation in the novel context of multiple target domains with unlabeled instances and one source domain with labeled instances. Our model aims to find a shared latent space common to all domains, while simultaneously accounting for the remaining private, domain-specific factors. Disentanglement of shared and private information is accomplished using a unified information-theoretic approach, which also serves to establish a stronger link between the latent representations and the observed data. The resulting model, accompanied by an efficient optimization algorithm, allows simultaneous adaptation from a single source to multiple target domains. We test our approach on three challenging publicly-available datasets, showing that it outperforms several popular domain adaptation methods. + +# 1 INTRODUCTION + +In real-world data, the training and test instances often do not come from the same underlying distribution (Sun et al. (2016)). For example, in the task of object recognition/classification from image data, this may be due to the image noise, changes in the object view, etc., which induce different biases in the observed data sampled during the training and test stage. Consequently, assumptions made by traditional learning algorithms are often violated, resulting in degradation of the algorithms’ performance during inference of test data. Domain Adaptation (DA) approaches (Fernando et al. (2013); Gong et al. (2012); Kodirov et al. (2015); Yoo et al. (2016)) aim to tackle this by transferring knowledge from a source domain (training data) to an unlabeled target domain (test data) to reduce the discrepancy between the source and target data distributions, typically by exploring domain-invariant data structures. + +Existing DA methods tackle the adaptation problems in one of the two settings: (semi)supervised DA and unsupervised DA (Csurka (2017)). The former assume that in addition to the labeled data of the source domain, some labeled data from the target domain are also available for training/adapting the classifiers. In contrast, the latter does not require any labels from the target domain but rather explores the similarity in the data distributions of the two domains. In this work, we focus on the unsupervised DA (uDA) scenario, which is more challenging due to the lack of correspondences in source and target labels. + +Most works on uDA today focus on a single-source-single-target-domain scenario. However, in many real-world applications, unlabeled data may come from different domains, thus, with different statistical properties but with common task-related content. For instance, we may have access to images of the same class of objects (e.g., cars) recorded by various types of cameras, and/or under different camera views and at different times, rendering multiple different domains (e.g., datasets). Likewise, facial expressions of emotions, such as joy and surprise, shown by different people and recorded under different views, result in multiple domains with varying data distributions. In most cases, these domains have similar underlying data distributions, which can be leveraged to build more effective and robust classifiers for tasks such as the object or emotion recognition across multiple datasets/domains. To this end, traditional uDA methods focus on the single-source-single-target DA scenario. However, in the presence of multiple domains, as typically encountered in real-world settings, this pair-wise adaptation approach may be suboptimal as it fails to leverage simultaneously the knowledge shared across multiple task-related domains. Recently, Zhao et al. (2017) showed that by having access to multiple source domains can facilitate better adaptation to a single target domain, when compared to the pair-wise DA approach. While this is intuitive due to the access to multiple labelled source domains, offering more adaptation flexibility for the target domain (i.e., by efficiently exploring the data labels across multiple source domains that are most related to the target domain), it comes at the expense of the data labelling in multiple source domains, which can be costly and time-consuming. In either case, a single source domain or readily available multiple source domains, to the best of our knowledge, a simultaneous adaptation to multiple and unlabelled target domains remains an unexplored DA scenario. However, this DA scenario is important as we usually have access to multiple unlabeled domains; yet, the adaptation process is also more challenging due to the lack of supervision in the target domains. Still, multi-target DA can have advantages over a single-target DA when: (i) there is direct knowledge sharing between the source and multiple target domains (fig. 1a), and (ii) the source and a target domain are related through another target domain (fig. 1b). While this seems intuitive, it is critical how the data from multiple unlabelled target domains are leveraged within the multi-target DA approach, in order to improve its performance over the single target DA approaches and naive fusion of multiple target domains. + +![](images/36886e0179e0304a2140f74c7b598abbe23c7d9c3021a983addc10b904e84af3.jpg) +Figure 1: Illustration of domains with common (a) and pairwise-shared spaces (b). We tackle the domain adaptation task when all domains share a common task/space, which is then leveraged to transfer knowledge across multiple target domains. + +To this end, we propose a Multi-Target DA-Information-Theoretic-Approach (MTDA-ITA) for single-sourcemulti-target DA. We exploit a single source domain and focus on multiple target domains to investigate the effects of multi-target DA; however, the proposed approach can easily be extended to multiple source domains. This approach leverages the data from multiple target domains to improve performance compared to individually learning from pair-wise source-target domains. Specifically, we simultaneously factorize the information from each available target domain and learn separate subspaces for modeling the shared (i.e., correlated across the domains) and private (i.e., independent between the domains) subspaces of the data (Salzmann et al. (2010)). To this end, we employ deep learning to derive an information theoretic approach where we jointly maximize the mutual information between the domain labels and private (domain-specific) features, while minimizing the mutual information between the the domain labels and the shared (domaininvariant) features. Consequently, the more robust feature representations are learned for each target domain by exploiting dependencies between multiple target domains. We show on benchmark datasets for DA that this approach leads to overall improved performance on each target domain, compared to independent DA for each pair of source-target domains, or the naive combination of multiple target domains, and state-of-the-art models applicable to the target task. + +# 2 THE PROPOSED METHOD + +Without loss of generality, we consider a multi-class ( $K$ -class) classification problem as the running example. Furthermore, let $( \mathbf { X } , \mathbf { Y } , \mathbf { D } ) = \{ ( \pmb { x } _ { i } , \pmb { y } _ { i } , \pmb { d } _ { i } ) \} _ { i = 0 } ^ { N }$ be a collection of $M$ domains (a labeled source domain, and $M - 1$ unlabeled target domains), where $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { i } }$ denotes the $i$ -th sample, and $\pmb { y } _ { i } = [ y _ { i } ^ { 1 } , y _ { i } ^ { 2 } , . . . , y _ { i } ^ { K } ]$ and $d _ { i } = [ d _ { i } ^ { 1 } , d _ { i } ^ { 2 } , . . . , d _ { i } ^ { M } ]$ are the $K$ -D and $M$ -D encoding of the class and domain labels for $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { i } }$ , respectively. Note that the class labels are only available for the source samples. + +The latent space representation of the data point $_ { \textbf { \em x } }$ is denoted as $\boldsymbol { z } = [ z _ { s } , z _ { p } ]$ , where $z _ { s }$ and $z _ { p }$ are the (latent) shared and private features of the data point $_ { \textbf { \em x } }$ , respectively. By factorizing the joint distribution $p ( { \pmb x } , { \pmb y } , d , z _ { s } , z _ { p } )$ as + +$$ +\begin{array} { r } { p ( { \pmb x } , { \pmb y } , { \pmb d } , z _ { s } , z _ { p } ) = p ( { \pmb x } ) p ( { \pmb d } ) p ( z _ { s } | { \pmb x } ) p ( z _ { p } | { \pmb x } ) p ( { \pmb y } | z _ { s } ) , } \end{array} +$$ + +we propose to maximize the following objective function: + +$$ +\begin{array} { r } { \boldsymbol { \mathcal { L } } ( \boldsymbol { \theta } _ { s } , \boldsymbol { \theta } _ { p } , \boldsymbol { \theta } _ { c } ; \boldsymbol { x } , \boldsymbol { y } , \boldsymbol { d } ) = \lambda _ { r } \boldsymbol { I } ( \boldsymbol { x } ; \boldsymbol { z } ) + \lambda _ { c } \boldsymbol { I } ( \boldsymbol { y } ; \boldsymbol { z } _ { s } ) + \lambda _ { d } \big ( \boldsymbol { I } ( \boldsymbol { d } ; \boldsymbol { z } _ { p } ) - \boldsymbol { I } ( \boldsymbol { d } ; \boldsymbol { z } _ { s } ) \big ) , } \end{array} +$$ + +where $p ( { \pmb x } )$ and $p ( d )$ denote the underlying (true) data distribution and domain label distribution, respectively, $I ( { \pmb x } ; { \pmb y } )$ denotes the Mutual Information between the random variables $_ { \textbf { \em x } }$ and $\textbf { { y } }$ . $\lambda _ { r } , \lambda _ { c }$ and $\lambda _ { d }$ denote the hyper-parameters controlling the weights of the objective terms. The proposed objective function (2) maximizes the three terms described below: + +• $I ( { \pmb x } ; z )$ : encourages the latent features (both shared and private) to preserve information about the data samples (that can be used to reconstruct $_ { \textbf { \em x } }$ from $_ { z }$ ). +• $I ( \pmb { y } ; \pmb { z } _ { s } )$ : enables to correctly predict the true class label of the samples out of their common shared features. +• $I ( d ; z _ { p } ) - I ( d ; z _ { s } )$ : encourages the latent private features to preserve the information about the domain label and penalizes the latent shared features to be domain informative. This not only reduces the redundancy in the shared and private features, but also, penalizes the redundancy of different private spaces, while preserving the shared information. + +An additional term could be used to minimize the mutual information between the shared $( z _ { s } )$ and private $( z _ { p } )$ features. However, computing the mutual information (even approximating it) is intractable due to the highly complex joint distribution $p ( z _ { s } , z _ { p } )$ . Since we want $z _ { s }$ and $z _ { p }$ features to encode different aspects of $_ { \textbf { \em x } }$ , we enforce such constraint by jointly maximizing the term: $I ( d ; \dot { z } _ { p } ) - I ( d ; z _ { s } )$ . + +# 2.1 OPTIMIZATION + +The following lower bound for mutual information is derived using the non-negativity of KL-divergence (Barber & Agakov (2003)); i.e., $\begin{array} { r } { \Sigma _ { { \pmb x } } p ( { \pmb x } | z ) \ln \frac { p ( { \pmb x } | z ) } { q ( { \pmb x } | z ) } \ge 0 } \end{array}$ gives: + +$$ +I ( \pmb { x } ; z ) \geq H ( \pmb { x } ) + \mathbb { E } _ { p ( \pmb { x } , z ) } [ \ln q ( \pmb { x } | z ; \phi ) ] +$$ + +where $H ( { \pmb x } )$ denotes the Shanon Entropy (Lin (1991)) of the random variable $_ { \textbf { \em x } }$ . $q ( { \pmb x } | { \pmb z } ; { \pmb \phi } )$ is any arbitrary distribution parameterized by $\phi$ . We need a variational distribution $q ( { \pmb x } | { \pmb z } ; \phi )$ because the posterior distribution $p ( \pmb { x } | \pmb { z } ) = p ( \pmb { z } | \pmb { x } ) p ( \pmb { x } ) / p ( \pmb { z } )$ is intractable since the true data distribution $p ( { \pmb x } )$ is assumed to be unknown. Similarly, we can derive lower bounds for $I ( d ; z _ { p } ) \geq H ( d ) + \mathbb { E } _ { p ( d , z _ { p } ) } [ \ln q ( d | z _ { p } ; \psi ) ]$ and $I ( d ; z _ { s } ) \geq$ + +![](images/8e3477bd318324bc4827d471de794c2c830ef3f1de87446abdf6c005292de940.jpg) +Figure 2: MDTA-ITA: The encoder $E _ { s } ( { \pmb x } )$ captures the feature representations $( z _ { s } )$ for a given input sample $_ { \pmb { x } }$ that are shared among domains. $E _ { p } ( { \pmb x } )$ captures domain–specific private features $( z _ { p } )$ using the shared private encoder. The shared decoder $F \big ( z _ { p } , z _ { s } \big )$ learns to reconstruct the input sample by using both the private and shared features. The domain classifier $D$ learns to correctly predict the domain labels of the actual samples from both their shared and private features while the classifier $C$ learns to correctly predict the class labels from the shared features. + +$H ( d ) + \mathbb { E } _ { p ( d , z _ { s } ) } [ \ln q ( d | z _ { s } ; \psi ) ]$ , where $q ( d | z _ { p } ; \psi )$ is any arbitrary distribution parameterized by $\psi$ . 1 We further drive lower bound for $I ( \pmb { y } ; \pmb { z } _ { s } )$ as $I ( \pmb { y } ; \pmb { z _ { s } } ) \geq H ( \pmb { y } ) + \mathbb { E } _ { p ( \pmb { y } , \pmb { z _ { s } } ) } [ \ln q ( \pmb { y } | \pmb { z _ { s } } ; \pmb { \theta _ { c } } ) ]$ , where $q ( \pmb { y } | \pmb { z } _ { s } ; \theta _ { c } )$ is a variational distribution parameterized by $\theta _ { c }$ approximating $p ( \pmb { y } | \boldsymbol { z } _ { s } )$ . + +Let next $E _ { s } ( \pmb { x } ; \theta _ { s } )$ be a function parameterized by $\theta _ { s }$ that maps a sample $_ { \textbf { \em x } }$ to its corresponding shared feature $z _ { s }$ , and $E _ { p } ( \pmb { x } ; \theta _ { p } )$ be an analogous function which maps $_ { \textbf { \em x } }$ to $z _ { p }$ , the feature that is private to each domain (fig. 2). We also define $F ( z _ { s } , z _ { p } ; \phi )$ as a decoding function mapping the concatenation of the latent features $z _ { s }$ and $z _ { p }$ to a sample reconstruction $\hat { \pmb x }$ , and $D ( z ; \psi )$ as a decoding function mapping $z _ { s }$ and $z _ { p }$ to a $M$ -dimensional vector: the predictions of the domain label $\hat { d }$ . Finally, $C ( z _ { s } ; \theta _ { c } )$ is a task-specific function mapping $z _ { s }$ to a $K$ -dimensional probability vector of the class label $\hat { y }$ . + +We represent $\begin{array} { r } { \frac { 1 } { N } \sum _ { i = 1 } ^ { \bar { N } } \delta ( { \bf d } - { \bf d } _ { i } ) ) } \end{array}$ $p ( { \pmb d } ) , p ( { \pmb x } ) , p ( { \pmb y } )$ as in the case of variational autoencoders (VAE) (Abbasnejad et al. (2017); as the empirical distribution of a finite training set (e.g. $\begin{array} { r l } { p ( d ) } & { { } = } \end{array}$ $\mathrm { P u }$ et al. (2017)), $p ( \pmb { z } _ { s } | \pmb { x } ) , p ( \pmb { z } _ { p } | \pmb { x } )$ as deterministic functions of $_ { \textbf { \em x } }$ as $p ( z _ { s } | \pmb { x } ) = \delta \big ( z _ { s } - E _ { s } ( \pmb { x } ; \theta _ { s } ) \big )$ and $p ( z _ { p } | \pmb { x } ) = \delta \big ( z _ { p } - E _ { p } ( \pmb { x } ; \theta _ { p } ) \big )$ , and the variational distributions $q ( \pmb { y } | \pmb { z } _ { s } )$ , $q ( { \pmb x } | z )$ and , $q ( d | z )$ as + +$$ +q ( \boldsymbol { y } | \boldsymbol { z } _ { s } ) = \mathrm { S o f t M a x } ( C ( \boldsymbol { z } _ { s } ; \boldsymbol { \theta } _ { c } ) ) , q ( d | \boldsymbol { z } ) = \mathrm { S o f t M a x } ( D ( \boldsymbol { z } ; \boldsymbol { \psi } ) ) , q ( \boldsymbol { x } | \boldsymbol { z } ; \boldsymbol { \phi } ) \propto \exp ( \| \boldsymbol { x } - \boldsymbol { F } ( \boldsymbol { z } ; \boldsymbol { \phi } ) \| _ { 1 } ) +$$ + +where Softmax $( \cdot )$ denotes the softmax or normalized exponential function (Bridle (1990)), and $\left. . \right. _ { 1 }$ denotes the $L _ { 1 }$ norm. Then, the optimization task can be posed as a minimax saddle point problem, where we use adversarial training to maximize (2) w.r.t. the parameters $( \theta _ { s } , \theta _ { p } , \theta _ { c } )$ , and to minimize (2) w.r.t. the parameters $( \phi , \psi )$ , using Stochastic Gradient Descent (SGD). + +# Optimizing the parameters $\phi$ of the decoder $F$ + +$$ +\hat { \phi } = \underset { \phi } { \arg \operatorname* { m i n } } \ : \mathcal { L } _ { F } = \frac { \lambda _ { r } } { N } \sum _ { i = 1 } ^ { N } \| \pmb { x } _ { i } - F \big ( E _ { s } ( \pmb { x } _ { i } ) , E _ { p } ( \pmb { x } _ { i } ) \big ) \| _ { 1 } . +$$ + +The decoder $F ( z _ { s } , z _ { p } ; \phi )$ is trained in such a way so as to minimize the difference between original input $_ { \textbf { \em x } }$ and its decoding from corresponding shared and private features via the decoder $F$ . + +# Optimizing the parameters $\psi$ of the domain classifier $D$ + +$$ +\hat { \psi } = \underset { \psi } { \arg \operatorname* { m i n } } \ : \mathcal { L } _ { D } = - \frac { \lambda _ { d } } { N } \sum _ { i = 1 } ^ { N } d _ { i } ^ { \top } \ln D \big ( E _ { s } ( \pmb { x } _ { i } ) \big ) - \frac { \lambda _ { d } } { N } \sum _ { i = 1 } ^ { N } d _ { i } ^ { \top } \ln D \big ( E _ { p } ( \pmb { x } _ { i } ) \big ) . +$$ + +$D ( z ; \psi )$ can be considered as a classifier whose task is to distinguish between the shared/private features of the different domains. More precisely, the two terms in Eq. 6 encourage $D$ to correctly predict the domain labels from the shared and private features, respectively. + +# Optimizing the parameters $\theta _ { c }$ of the label classifier $C$ + +$$ +\hat { \theta } _ { c } = \underset { \theta _ { c } } { \arg \operatorname* { m i n } } \big \{ - H ( \pmb { y } ) - \mathbb { E } _ { p ( \pmb { y } , z _ { s } ) } \big [ \ln q ( \pmb { y } | z _ { s } ) \big ] \big \} . +$$ + +Since we have access to the source labels, $H ( \pmb { y } )$ is a constant for source samples. we can approximate $H [ \pmb { y } ]$ for the target samples using the output of the classifier $C$ , leading to the following optimization problem: + +$$ +\begin{array} { r l } & { \hat { \theta } _ { c } = \underset { \theta _ { c } } { \arg \operatorname* { m i n } } \ \mathcal { L } _ { C } = - \ \frac { 1 } { N } \displaystyle \sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \ln C \big ( E _ { s } ( x _ { i } ) \big ) - \frac { \lambda _ { c } } { N - N _ { s } } \displaystyle \sum _ { i = N _ { s } + 1 } ^ { N } C \big ( E _ { s } ( x _ { i } ) \big ) ^ { \top } \ln C \big ( E _ { s } ( x _ { i } ) \big ) } \\ & { \qquad + \ \frac { \lambda _ { c } } { N - N _ { s } } \displaystyle \sum _ { i = N _ { s } + 1 } ^ { N } C \big ( E _ { s } ( x _ { i } ) \big ) ^ { \top } \ln \bigg ( \displaystyle \frac { 1 } { N - N _ { s } } \displaystyle \sum _ { i = N _ { s } + 1 } ^ { N } C \big ( E _ { s } ( x _ { i } ) \big ) \bigg ) , } \end{array} +$$ + +where $N _ { s }$ denotes the number of source samples. Intuitively, we enforce the classifier $C$ to correctly predict the class labels of the source samples by the first term in Eq. 8. We use the second term to minimize the entropy of $q ( \pmb { y } | \pmb { z } _ { s } )$ for the target samples; effectively, reducing the effects of "confusing" labels of target samples, as given by $p ( \pmb { y } | \pmb { z } _ { s } )$ that leads to decision boundaries occur far away from target data-dense regions in the feature space. The intuition behind the last term is that by minimizing only the entropy (second term), we may arrive at a degenerate solution where every target point $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ is assigned to the same class. Hence, the last term encourages the classifier $C$ to have balanced labeling for the target samples where it reaches its minimum, $\ln K$ , when each class is selected with uniform probability. + +# Optimizing the parameter $\theta _ { p }$ of the private encoder $E _ { p }$ + +$$ +\hat { \theta _ { p } } = \underset { \theta _ { p } } { \operatorname { a r g m i n } } \ \mathcal { L } _ { P } = \frac { \lambda _ { r } } { N } \sum _ { i = 1 } ^ { N } \| \boldsymbol { x } _ { i } - F \big ( E _ { s } ( \boldsymbol { x } _ { i } ) , E _ { p } ( \boldsymbol { x } _ { i } ) \big ) \| _ { 1 } - \frac { \lambda _ { d } } { N } \sum _ { i = 1 } ^ { N } d _ { i } ^ { \top } D \big ( E _ { p } ( \boldsymbol { x } _ { i } ) \big ) . +$$ + +The first term in Eq. 9 encourages the private encoder $E _ { p }$ to preserve the recovery ability of the private features. The second term enforces distinct private features be produced for each domain by penalizing the representation redundancy in different private spaces. This, in turn, encourages moving this common information from multiple domains to their shared space. + +# Optimizing the parameter $\theta _ { s }$ of the shared encoder $E _ { s }$ + +$$ +\begin{array} { c } { \displaystyle \hat { \theta _ { s } } = \arg \operatorname* { m i n } _ { \theta _ { s } } \mathcal { L } _ { S } = \frac { \lambda _ { r } } { N } \sum _ { i = 1 } ^ { N } \| \boldsymbol { x } - F \big ( E _ { s } ( \boldsymbol { x } _ { i } ) , E _ { p } ( \boldsymbol { x } _ { i } ) \big ) \| _ { 1 } - \frac { \lambda _ { c } } { N } \sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \ln C \big ( E _ { s } ( \boldsymbol { x } _ { i } ) \big ) } \\ { \displaystyle - \frac { \lambda _ { d } } { N } \sum _ { i = 1 } ^ { N } d _ { i } ^ { \top } \ln D \big ( E _ { s } ( \boldsymbol { x } _ { i } ) \big ) - \frac { \lambda _ { c } } { N - N _ { s } } \sum _ { i = N _ { s } + 1 } ^ { N } C \big ( E _ { s } ( \boldsymbol { x } _ { i } ) \big ) ^ { \top } \ln C \big ( E _ { s } ( \boldsymbol { x } _ { i } ) \big ) } \\ { \displaystyle + \frac { \lambda _ { c } } { N - N _ { s } } \sum _ { i = N _ { s } + 1 } ^ { N } C \big ( E _ { s } ( \boldsymbol { x } _ { i } ) \big ) ^ { \top } \ln \bigg ( \frac { 1 } { N - N _ { s } } \sum _ { i = N _ { s } + 1 } ^ { N } C \big ( E _ { s } ( \boldsymbol { x } _ { i } ) \big ) \bigg ) . } \end{array} +$$ + +The first term in Eq. 10 encourages the shared encoder $E _ { s }$ to preserve the recovery ability of the shared features. The second term is the source domain classification loss penalty that encourages $E _ { s }$ to produce discriminative features for the labeled source samples. The third term simulates the adversarial training by trying to fool the domain classifier $D$ when predicting the domain labels $^ d$ , given the shared features $z _ { s }$ . The effect of this is two-fold: (i) the rendered shared features are more distinct from the corresponding private features, (ii) the shared features of different domains are encouraged to be similar to each other. The last two terms encourage $E _ { s }$ to produce the shared features for target samples so that the classifier is confident on the unlabeled target data, driving the shared features away from the decision boundaries. To train our model, we alternate between updating the shared encoder $E _ { s }$ , the private encoder $E _ { p }$ , the decoder $F$ , the classifier $C$ and the domain classifier $D$ using the SGD algorithm (see Algorithm 1 in Appendix E for more details). + +# 3 RELATED WORK + +There has been extensive prior work on domain adaptation (Csurka (2017)). Recent papers have focused on transferring deep neural network representations from a labeled source dataset to an unlabeled target domain, where the main strategy is to find a feature space such that the confusion between source and target distributions in that space is maximized ( Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al. (2017); Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017); Yan et al. (2017); Bousmalis et al. (2017)). For this, it is critical to first define a measure of divergence between source and target distributions. For instance, several methods have used the Maximum Mean Discrepancy (MMD) loss for this purpose (Bousmalis et al. (2017); Zellinger et al. (2017); Long et al. (2014)). MMD computes the norm of the difference between two domain means in the reproducing Kernel Hilbert Space (RKHS) induced by a pre-specified kernel. The Deep Adaptation Network (DAN) (Long et al. (2015)) applied MMD to layers embedded in a RKHS, effectively matching higher order statistics of the two distributions. The deep Correlation Alignment (CORAL) method (Sun & Saenko (2016)) attempts to match the mean and covariance of the two distributions. Deep Transfer Network (DTN) (Zhang et al. (2015)) achieved source/target distribution alignment via two types of network layers based on MMD distance: the shared feature extraction layer, which learns a subspace that matches the marginal distributions of the source and the target samples, and the discrimination layer, which matches the conditional distributions by classifier transduction. + +Recently proposed unsupervised DA methods (Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al. (2017); Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017)) operate by training deep neural networks using adversarial training, which allows the learning of feature representations that are simultaneously discriminative of source labels, and indistinguishable between the source and target domain. For instance, Ganin & Lempitsky (2015) proposed a DA mechanism called Domain-Adversarial Training of Neural Networks (DANN), which enables the network to learn domain invariant representations in an adversarial way by adding a domain classifier and back-propagating inverse gradients. Adversarial Discriminative Domain Adaptation (ADDA) (Tzeng et al. (2017)) learns a discriminative feature subspace using the source labels, followed by a separate encoding of the target data to this subspace using an asymmetric mapping learned through a domain-adversarial loss. Liu et al. (2017) makes a shared-latent space assumption and proposes an unsupervised image-to-image translation (UNIT) framework based on Coupled GANs (Liu & Tuzel (2016)). Another example is the pixel-level domain adaptation models that perform the distribution alignment not in the feature space but directly in raw pixel space. PixelDA (Bousmalis et al. (2017)) uses adversarial approaches to adapt source-domain images as if drawn from the target domain while maintaining the original content. + +While these approaches have shown success in DA tasks with single source-target domains, they are not designed to leverage information from multiple domains simultaneously. More recently, Zhao et al. (2017) introduced an adversarial framework called MDAN for multiple source single target domain adaptation where a domain classifier, induced by minimizing the H-divergence between multiple source and a target domain, is used to align their feature distributions in a shared space. Instead, in our approach we focus on multi-target DA where we perform adaptation of multiple unlabelled target domains. Although both our model and MDAN use the similar notion of the domain classifier to minimize the domain mismatch in shared space, the domain classifier induced by our information-theoretic (IT) loss also acts to separate domains in the private space (see Eqs. $6 \& 9$ for more details), improving the essential reconstruction ability, similar to (Bousmalis et al. (2016)). We provided how our model is related to IT representation learning approaches, and multiple domain transfer networks in Appendices A and B respectively. In Appendix C, we also clearly contrasted our model with DSN model which also uses the notion of auto-encoders to explicitly separate the feature representations private to each source/target domain from those that are shared between the domains. + +# 4 EXPERIMENTAL RESULTS + +We compare the proposed method with state-of-the-art methods on standard benchmark datasets: a digit classification task that includes 4 datasets: MNIST (LeCun et al. (1998)), MNIST-M (Ganin et al. (2016)), SVHN (Netzer et al. (2011)), USPS (Tzeng et al. (2017)), Multi-PIE expression recognition dataset2, and PACS multi-domain image recognition benchmark (Li et al. (2017)), a new dataset designed for the crossdomain recognition problems (the details for this experiment is available in Appendix G). Fig. 3 illustrates image samples from different datasets and domains. We evaluate the performance of all methods with classification accuracy metric. We repeated each experiment 5 times and report the average and the standard deviation of the accuracy. + +We used ADAM (Kingma & Ba (2015)) for training; the learning rate was set to 0.0002 and momentum parameters to 0.5 and 0.999. We used batches of size 16 from each domain, and the input images were meancentered/rescaled to $[ - 1 , 1 ]$ . The hyper-parameters are empirically set as $\lambda _ { r } = 1 . 0 , \lambda _ { c } = 0 . 0 1$ , $\lambda _ { d } = 0 . 2 0$ . For the network architecture, our private/shared encoders consisted of three convolutional layers as the front-end and four basic residual blocks as the back-end. The decoder consisted of four basic residual blocks as the front-end and four transposed convolutional layers as the back-end. The discriminator and the classifier consisted of stacks of convolutional layers. We used ReLU for nonlinearity. Tanh function is used as the activation function of the last layer in the decoder $F$ for scaling the output pixels to $[ - 1 , 1 ]$ . The details of the networks are given in Appendix D. + +The quantitative evaluation involves a comparison of the performance of our model to previous work and to Source Only and 1-NN baselines that do not use any domain adaptation. For Source Only baseline, we train our model only on the unaltered source training data and evaluate on the target test data. We compare the proposed method MTDA-ITA with several related methods designed for pair-wise source-target adaptation: CORAL (Sun & Saenko (2016)), DANN (Ganin & Lempitsky (2015)), ADDA (Tzeng et al. (2017)), DTN (Zhang et al. (2015)), UNIT (Liu et al. (2017)), PixelDA (Bousmalis et al. (2017)), and DSN (Bousmalis et al. (2016)). We reported the results of two following baselines: (i) one is to combine all the target domains into a single one and train it using MTDA-ITA, which we denote as (c-MTDA-ITA). (ii) the other one is to train multiple MTDA-ITA separately, where each one corresponds to a source-target pair which we denote as (s-MTDA-ITA). For completeness, we reported the results of the competing methods by combining all the target domains into a single one (denoted by c-DTN, c-ADDA, and c-DSN) as well. We also extend DSN to multiple domains by (i) having one private encoder for all domains denoted by (1p-DSN), (ii) adding multiple private encoders to it denoted by (mp-DSN) and contrast them with our model. + +# 4.1 DIGITS DATASETS + +We combine four popular digits datasets (MNIST, MNIST-M, SVHN, and USPS) to build the multitarget domain dataset. All images were uniformly rescaled to $3 2 \times 3 2$ . We take each of MNIST-M, + +![](images/c17f5da419bdc0deb50b99f30b19766556f98c280148111ece40ad50497e48c6.jpg) +Figure 3: Exemplary images from different datasets. a) Digits datasets, b) PACS datatset (first row: Art-painting, second row: Cartoon, Third row: Photo, last row: Sketch), c) Multi-PIE dataset (each row corresponds to a different camera angle and each subject depicts an expression(normal, smile, surprise, squint, disgust, scream) at every camera position). + +SVHN, USPS, and MNIST as source domain in turn, and the rest as targets. We use all labeled source images and all unlabeled target images, following the standard evaluation protocol for unsupervised domain adaptation (Ganin et al. (2016); Long et al. (2016). We show the accuracy of different methods in Table 1. Additional results are available in Appendix F. The results show that first of all cMTDA-ITA has worse performance than sMTDA-ITA and MTDA-ITA. We have similar observations for ADDA, DTN, and DSN that demonstrates a naive combination of different target datasets can sometimes even decrease the performance of the competing methods. Furthermore, MTDA-ITA outperforms the state-of-the-art methods in most of domain transformations. The higher performance of MTDA-ITA compared to other methods is mainly attributed to the joint adaptation of related domains where each domain could benefit of other related domains. Furthermore, from the results obtained, we see that it is beneficial to use information coming from unlabeled target data (see Eq. 8 for updating the classifier $C$ ) during the learning process, compared to when no data from target domain is used (See the ablation study section for more information). Indeed, using our scheme, we find a representation space in which embeds the knowledge from the target domain into the learned classifier. By contrast, the competing methods do not provide a principled way of sharing information across all domains, leading to overall lower performance. The results also verify the superiority of MTDA-ITA over both mp-DSN, and 1p-DSN. This can be due to (i) having multiple private encoders increase the number of parameters that may lead to mp-DSN overfitting, (ii) superiority of the MTDA-ITA’s domain adversarial loss over the DSN’s MMD loss to separate the shared and private features, (iii) utilization of the unlabeled target data to regularize the classifier in MTDA-ITA. + +# 4.2 MULTI-PIE DATASET + +The Multi-PIE dataset includes face images of 337 individuals captured from different expressions, views, and illumination conditions (fig. 3(c)). For this experiment, we use 5 different camera views (positions) $C 0 5$ , $C 0 8$ , $C 0 9$ , $C 1 3$ , and $C 1 4$ as different domains (Fig. 3(c)) and the face expressions (normal, smile, surprise, squint, disgust, scream) as labels. Each domain contains 27120 images of size $6 4 \times 6 4 \times 3$ . We used each view as the source domain, in turn, and the rest as targets. We expect the face inclination angle to reflect the complexity of transfer learning. Tab. 2 shows the classification accuracy for $C 1 3$ and $C 1 4$ as source domain (the results for views $C 0 5$ , $C 0 8$ and $C 0 9$ as source domain are available in Appendix F). As can be seen, MTDA-ITA achieves the best performances as well as the best scores in most settings that verifies the effectiveness of MTDA-ITA for multi-target domain adaptation. Clearly, with the increasing camera angle, the image structure changes up to a certain extent (the views become heterogeneous). However, our method produces better results even under such very challenging conditions. + +Table 1: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown in red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDAITA where each one correspond to a source-target pair. 1p-DSN: extended DSN with single private encoder.mp-DSN: extended DSN with multiple private encoder. Last column shows the average rank of each method over all adaptation pairs. $\bf \Pi ^ { * } U N I T$ trains with the extended SVHN $> 5 0 0 \mathrm { K }$ images vs ours 72K). \*PixelDA uses $( \approx 1 , 0 0 0 )$ of labeled target domain data as a validation set for tuning the hyper-parameters. + +
methodS→MS→MMS→UM→SM→MMM→UAve.ranking
Source Only62.10±0.6040.43±0.7039.90± 0.6030.29± 0.5955.98 ± 0.4878.30±0.3814.00
1-NN35.8618.2129.3128.0112.5841.2215.00
CORAL (Sun & Saenko (2016))63.10 ± 0.6154.37 ± 0.5350.15 ± 0.6333.40 ±0.7457.70 ± 0.6981.05 ± 0.8011.33
DANN (Ganin et al. (2016))73.80 ±0.4961.05±0.8062.54 ± 0.9135.50 ± 0.6577.40±0.7381.60 ± 0.608.75
ADDA(Tzeng et al. (2017))77.68 ± 0.9264.23±0.7064.10±0.7930.04± 0.9891.47 ±1.090.51 ± 0.806.43
C-ADDA80.10 ± 0.6956.80 ± 0.7964.80±0.8827.50 ± 0.8683.30 ±0.9084.10 ±0.988.95
DTN (Zhang et al. (2015))81.40± 0.4263.70±0.3960.12 ± 0.5240.40±0.5085.70±0.3985.80 ± 0.466.04
c-DTN PixelDA (Bousmalis et al. (2017))82.10 ±0.6259.30 ± 0.5956.87 ± 0.6538.32 ± 0.5080.90±0.8079.31 ±0.787.96
UNIT (Liu et al. (2017))98.10*94.10*1
DSN (Bousmalis et al. (2016))90.6*92.901
c-DSN82.70±0.37 83.10 ±0.2064.80± 0.4065.30± 0.2849.30±0.3083.20±0.3091.65 ± 0.402.85
1p-DSN81.00 ± 0.4760.56 ± 0.36 58.22 ± 0.6860.35 ± 0.5946.80 ± 0.4580.49±0.4088.21 ±0.384.84
mp-DSN83.40±0.3061.00±0.5058.06 ±0.48 58.10 ±0.6445.11 ± 0.3377.33 ± 0.52 79.30±0.5985.16 ± 0.634.90
s-MTDA-ITA82.90±0.1363.10±0.2863.54±0.3047.35 ± 0.4082.42± 0.1986.45 ± 0.715.33
c-MTDA-ITA79.20±0.2859.90 ± 0.3063.70±0.2649.60± 0.25 45.30 ±0.3077.12 ± 0.2289.21±0.282.88
MTDA-ITA84.60±0.2465.30 ± 0.1570.03 ±0.2052.01 ± 0.2185.50 ±0.1887.47 ± 0.25 94.20±0.204.25
1.16
+ +
methodC13→C05C13→C08C13→C09C13→C14C14→C05C14→C08C14→C09C14→C13Ave.ranking
Source Only50.79± 0.3345.90± 0.5040.04±0.4059.68± 0.2960.03±0.5536.80± 0.6140.11 ±0.5060.57± 0.3616.08
1-NN CORAL33.21 54.89 ± 0.5237.0134.4548.79 68.90± 0.3547.4428.2430.8644.8617.00
DANN57.86± 0.4148.90±0.48 50.30 ± 0.4340.30±0.53 45.30± 0.5070.68 ± 0.3559.98 ± 0.45 57.20± 0.4540.63±0.55 40.22±0.5540.80 ±0.53 40.77 ± 0.4565.11 ± 0.45 70.50± 0.5511.95 9.92
ADDA64.83±0.6963.20±0.4555.48 ±0.6574.25±0.5573.62 ± 0.7543.56±0.9538.68 ±0.9572.84±0.759.33
C-ADDA59.20±0.2530.70±0.6353.20±0.4068.33±0.3565.88±0.3830.60±0.6145.34±0.4864.30±0.4011.50
DTN63.78±0.2960.45±0.3560.55±0.3572.60±0.2570.67±0.3041.55±0.6541.45 ± 0.4570.67±0.45
c-DTN57.53 ± 0.4255.24 ± 0.4557.14 ± 0.3965.16 ± 0.3563.80 ±0.4238.97 ± 0.7139.80 ± 0.6562.10 ± 0.458.75 10.92
PixelDA45.68± 0.5244.95± 0.4244.45±0.5590.50± 0.2546.28±0.6045.89±0.6144.45± 0.5169.15 ± 0.459.95
UNIT44.14±0.1044.47 ± 0.1144.21 ±0.1244.47 ± 0.1143.03 ±0.144.44 ±0.1544.47 ± 0.1544.47 ± 0.0511.07
DSN64.15± 0.3057.70±0.3849.15± 0.4580.75±0.2782.20±0.2838.75±0.5345.00± 0.2580.50± 0.35
c-DSN57.34±0.4531.63 ±0.6051.17 ±0.4074.52 ±0.3782.01 ± 0.3534.25±0.5842.63 ± 0.5579.42 ±0.355.15 8.20
1p-DSN55.84± 0.5030.03 ± 0.5049.06 ± 0.3872.11 ± 0.5081.22 ±0.4533.33 ± 0.5842.03 ±0.2478.78±0.578.63
mp-DSN55.20±0.4630.40±0.5047.80±0.3575.30 ± 0.2580.75±0.2030.20 ±0.5543.00±0.3579.02±0.408.88
s-MTDA-ITA70.10± 0.2758.90±0.2558.10±0.2780.12±0.1582.05±0.1845.90±0.3052.67±0.3081.60±0.24
c-MTDA-ITA60.34 ± 0.1755.67 ± 0.2157.10 ± 0.2373.50 ± 0.2076.80 ±0.1043.10 ±0.1248.10 ± 0.1480.90 ± 0.113.65
MTDA-ITA78.40 ± 0.266.70 ± 0.1770.30± 0.1485.49 ± 0.1187.20±0.1061.40±0.1460.05 ±0.1386.70 ± 0.105.01 1.20
+ +Table 2: Classification results on Multi-PIE dataset. Last column shows the average rank of each method over all adaptation pairs. The best is shown in red. + +# 4.3 ABLATION STUDIES + +We performed an ablation study on the proposed model measuring impact of various terms on the model’s performance. To this end, we conducted additional experiments for the digit datasets with different components ablation, i.e., training without the reconstruction loss (denoted as MTDA-woR) by setting $\lambda _ { r } = 0$ , training without the classifier entropy loss (denoted as MTDA-woE) by setting $\lambda _ { c } = 0$ , training without the multidomain separation loss (denoted as MTDA-woD) by setting $\lambda _ { d } = 0$ . + +As can be seen from fig. 4, disabling each of the above components leads to degraded performance. More precisely, the average drop by disabling the classifier entropy loss is $\approx 3 . 5 \%$ . Similarly, by disabling the reconstruction loss and the multi-domain separation loss, we have $\approx 4 . 5 \%$ and $\approx 2 2 \%$ average drop in performance, respectively. Clearly, by disabling the multi-domain separation loss, the accuracy drops significantly due to the severe data distribution mismatch between different domains. The figure also demonstrates that leveraging the unlabeled data from multiple target domains during training enhances the generalization ability of the model that leads to higher performance. In addition, the performance drop caused by removing the reconstruction loss , i.e., without the private encoder/decoder, indicates (i) the benefit of modeling the latent features as the combination of shared and private features, (ii) the ability of the model’s domain adversarial loss to effectively learn those features. + +In order to examine the effect of the private features on the model’s classification performance, we took the MTDA-ITA and trained it without the private encoder (denoted as MTDA-woP). As fig. 4 shows, without the private features, the model performed consistently worse $\approx 2 \%$ average drop in performance) in all scenarios. This demonstrates explicitly modeling what is unique to each domain can improve the model’s ability to extract domain–invariant features. In summary, this ablation study showed that the individual components bring complimentary information to achieve the best classification results. + +# 4.4 FEATURE VISUALIZATION + +We use t-SNE (Maaten & Hinton (2008)) on Digit dataset to visualize shared and private feature representations from different domains. Fig. 6 shows shared and private features from source (SVHN) and target domains before (a),(b) and after adaptation (c),(d). MTDA-ITA significantly reduces the domain mismatch for the shared features (circle markers in fig. 6d, strong mixing of domain labels in this cluster, fig. 6c) and increases it for the private features (triangle markers, pure and well-separated domain clusters in fig. 6c). This is partially due to the proposed multi-domain separation loss through the use of the domain classifier $D$ , which penalizes the domain mismatch for the shared features and rewards the mismatch for the private features. Moreover, as supported by the quantitative results in tab. 1, joint adaptation of related domains and the classifier, accomplished through the model, leads to superior class separability, compared to original features. This is depicted in fig. 6d, where the points in the shared space (large cluster) are grouped into class-specific subgroups (color indicates class label), while they are mixed in private spaces (smaller clusters). This is in contrast to fig. 6b, where original features show no class-specificity. + +We also show the learned shared and private features for the models MTDA-woE, MTDA-woP, MTDAwoR, and MTDA-woD, in figs. 6e to 6l. Note that since the private encoder $E _ { p }$ is disabled for MTDA-woR, and MTDA-woP, no private features are depicted in figs. $6 \mathrm { g }$ to 6j. The class label separation in the shared space for MTDA-woE, MTDA-woP, and MTDA-woR, figs. 6f, 6h and 6j, is still evident but not as strong as in the full model, fig. 6d, corroborating the small loss in classification accuracy observed in fig. 4a. On the other hand, MTDA-woD has significant mixing of class labels in the shared space, fig. 6l, more so than MTDA-woE, MTDA-woR, and MTDA-woP, implying worse classification prediction in fig. 4a due to the severe mismatch between different domains. Since our model uses one private encoder for all target domains, we also contrasted the visualization of DSN model with one private encoder 1p-DSN in Appendix H. + +# 5 CONCLUSION + +This paper presented an information theoretic end-to-end approach to uDA in the context of a single source and multiple target domains that share a common task or properties. The proposed method learns feature representations invariant under multiple domain shifts and simultaneously discriminative for the learning task. This is accomplished by explicitly separating representations private to each domain and shared between source and target domains using a novel discrimination strategy. Our use of a single private domain encoder results in a highly scalable model, easily optimized using established back-propagation approaches. Results on three benchmark datasets for image classification show superiority of the proposed method compared to the state-of-the-art methods for unsupervised domain adaptation of visual domain categories. + +# REFERENCES + +M Ehsan Abbasnejad, Anthony Dick, and Anton van den Hengel. Infinite variational autoencoder for semi-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 781–790. IEEE, 2017. + +Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy computation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018. + +Alexander Alemi, Ben Poole, Ian Fischer, Joshua Dillon, Rif A Saurous, and Kevin Murphy. Fixing a broken elbo. In International Conference on Machine Learning, pp. 159–168, 2018a. + +Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. International Conference on Learning Representation (ICLR), 2016. + +Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018b. + +Asha Anoosheh, Eirikur Agustsson, Radu Timofte, and Luc Van Gool. Combogan: Unrestrained scalability for image domain translation. ICLR Workshops, 2017. + +David Barber and Felix Agakov. The im algorithm: a variational approach to information maximization. In Advances in Neural Information Processing Systems (NIPS), pp. 201–208. MIT Press, 2003. + +Sagie Benaim and Lior Wolf. One-sided unsupervised domain mapping. In Advances in Neural Information Processing Systems (NIPS), pp. 752–762, 2017. + +Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, and Dumitru Erhan. Domain separation networks. In Advances in Neural Information Processing Systems (NIPS), pp. 343–351, 2016. + +Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pp. 7, 2017. + +Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio. Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349, 2015. + +John S Bridle. Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In Neurocomputing, pp. 227–236. Springer, 1990. + +Matthew Chalk, Olivier Marre, and Gasper Tkacik. Relevant sparse codes with variational information bottleneck. In Advances in Neural Information Processing Systems (NIPS), pp. 1957–1965, 2016. + +Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1711, 2017. + +Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. In Advances in Neural Information Processing Systems (NIPS), pp. 3733–3742, 2017. + +Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. pp. 1–35. Springer, 2017. + +Basura Fernando, Amaury Habrard, Marc Sebban, and Tinne Tuytelaars. Unsupervised visual domain adaptation using subspace alignment. In IEEE International Conference on Computer Vision (ICCV), pp. 2960–2967, 2013. + +Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. International Conference on Machine Learning (ICML), 2015. + +Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096–2030, 2016. + +Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2066–2073, 2012. + +Guang-Yuan Hao, Hong-Xing Yu, and Wei-Shi Zheng. Mixgan: Learning concepts from different domains for mixture generation. arXiv preprint arXiv:1807.01659, 2018. + +Hirokazu Kameoka, Takuhiro Kaneko, Kou Tanaka, and Nobukatsu Hojo. Stargan-vc: Non-parallel manyto-many voice conversion with star generative adversarial networks. arXiv preprint arXiv:1806.02169, 2018. + +Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representation (ICLR), 2015. + +Elyor Kodirov, Tao Xiang, Zhenyong Fu, and Shaogang Gong. Unsupervised domain adaptation for zero-shot learning. In IEEE International Conference on Computer Vision (ICCV), pp. 2452–2460, 2015. + +Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. + +Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In IEEE International Conference on Computer Vision (ICCV), pp. 5543–5551. IEEE, 2017. + +Jianhua Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information theory, 37(1):145–151, 1991. + +Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems (NIPS), pp. 469–477. Curran Associates, Inc., 2016. + +Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Advances in Neural Information Processing Systems (NIPS), pp. 700–708, 2017. + +Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1410–1417, 2014. + +Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I Jordan. Learning transferable features with deep adaptation networks. International Conference on Machine Learning (ICML), 2015. + +Mingsheng Long, Jianmin Wang, Yue Cao, Jiaguang Sun, and S Yu Philip. Deep learning of transferable representation for scalable domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 28(8):2027–2040, 2016. + +Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(Nov):2579–2605, 2008. + +Shakir Mohamed and Danilo Jimenez Rezende. Variational information maximisation for intrinsically motivated reinforcement learning. In Advances in neural information processing systems (NIPS), pp. 2125–2133, 2015. + +Saeid Motiian, Quinn Jones, Seyed Iranmanesh, and Gianfranco Doretto. Few-shot adversarial domain adaptation. In Advances in Neural Information Processing Systems (NIPS), pp. 6673–6683, 2017. + +Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, volume 2011, pp. 5, 2011. + +Yuchen Pu, Weiyao Wang, Ricardo Henao, Liqun Chen, Zhe Gan, Chunyuan Li, and Lawrence Carin. Adversarial symmetric variational autoencoder. In Advances in Neural Information Processing Systems (NIPS), pp. 4330–4339, 2017. + +Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In Advances in Neural Information Processing Systems (NIPS), pp. 506–516, 2017. + +Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain adaptation. International Conference on Machine Learning (ICML), 2017. + +Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces. In International Conference on Artificial Intelligence and Statistics (AISTATS), pp. 701–708, 2010. + +Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision (ECCV), pp. 443–450. Springer, 2016. + +Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. Thirtieth AAAI Conference on Artificial Intelligence, 2016. + +Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In IEEE Information Theory Workshop (ITW), pp. 1–5. IEEE, 2015. + +Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pp. 4, 2017. + +Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. + +Donggeun Yoo, Namil Kim, Sunggyun Park, Anthony S Paek, and In So Kweon. Pixel-level domain transfer. In European Conference on Computer Vision (ECCV), pp. 517–532, 2016. + +Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschläger, and Susanne Saminger-Platz. Central moment discrepancy (cmd) for domain-invariant representation learning. International Conference on Learning Representation (ICLR), 2017. + +Jing Zhang, Wanqing Li, and Philip Ogunbona. Joint geometrical and statistical alignment for visual domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. + +Xu Zhang, Felix Xinnan Yu, Shih-Fu Chang, and Shengjin Wang. Deep transfer network: Unsupervised domain adaptation. arXiv preprint arXiv:1503.00591, 2015. + +Han Zhao, Shanghang Zhang, Guanhang Wu, João P. Costeira, Jose Moura, and Geoffrey J. Gordon. Multiple source domain adaptation with adversarial training of neural networks. International Conference on Learning Representation (ICLR) workshops, 05 2017. + +# A CONNECTION TO INFORMATION THEORETIC REPRESENTATION LEARNING + +The idea of using information theoretic (IT) objectives for representation learning was originally introduced in (Tishby & Zaslavsky (2015)). Since their approach for optimizing the IT objective functions relied on the iterative Blahut Arimoto algorithm (Tishby & Zaslavsky (2015)), it is not feasible to apply to deep neural network (DNN) frameworks. Similar to our approach, there have been some recent works (Mohamed & Rezende (2015); Chalk et al. (2016); Alemi et al. (2018b; 2016; 2018a)) to approximate the MI by applying variational bounds on MI, though not in the context of domain adaptation. + +Mohamed & Rezende (2015) utilized the variational bounds on MI, and apply it to DNNs in the context of reinforcement learning. Chalk et al. (2016) and Alemi et al. (2016), developed the same variational lower bound In the context of Information Bottleneck (IB) principle (Tishby & Zaslavsky (2015)), where the former applied it to sparse coding problems, and used the kernel trick to achieve nonlinear mappings, whereas the latter applied it to DNNs to handle large datasets thanks to the SGD algorithm. Achille & Soatto (2018) proposed a variational bound on the MI in the context of IB, from the perspective of variational dropout and demonstrated its utility in learning disentangled representations for variational autoencoders. + +The main difference between our method and the above methods is that these methods throw away the information in the data not related to the task by minimizing the mutual information between the data points and the latent representations that may lead to ignoring the individual characteristics (private features) of the datasets in a multiple dataset regime, whereas our method explicitly models what is unique to each domain (dataset) that improves the model’s ability to extract domain–invariant features. + +In the unsupervised representation learning literature, our work is also related to the VAE-based models (Bowman et al. (2015)). However, we propose to tackle the task using our IT approach using deterministic mappings instead of the traditional evidence lower bound (ELBO) optimization with stochastic mappings. In contrast to the unsupervised representation learning approaches, our setting also allows us to further improve the latent representation using the labeled data in the source domain while leveraging the sharing of dependencies across different target domains. + +# B CONNECTION TO MULTIPLE DOMAIN TRANSFER NETWORKS + +Recent studies have shown remarkable success in multiple domain transfer (MDT) (Choi et al. (2017); Anoosheh et al. (2017); Kameoka et al. (2018); Hao et al. (2018)) though not in the context of the image classification, rather in the context of image generation. Choi et al. (2017) proposed StarGAN, a generative adversarial network capable of learning mappings among multiple domains in the contest of image to image translation framework. The goal of StarGAN is to train a single generator $G$ though this requires passing in a vector along with each input to the generator specifying the output domain desired, that learns mappings among multiple domains. To achieve this, $G$ is trained to translate an input image $_ { \textbf { \em x } }$ into an output image $\mathbf { x } ^ { \prime }$ conditioned on the target domain label $^ d$ , $G ( \pmb { x } , \pmb { d } ) \pmb { x } ^ { \prime }$ . Similar to our domain classifier module $D$ , they introduce an auxiliary classifier that allows a single discriminator to control multiple domains. + +Anoosheh et al. (2017) introduced ComboGAN, which decouples the domains and networks from each other. Similar to our encoder/decoder modules, ComboGAN’s generator networks contain encoder/decoders assigning each encoder and decoder to a domain. They combine the encoders and decoders of the trained model like building blocks, taking as input any domain and outputting any other. For example during inference, to transform an image $_ { \textbf { \em x } }$ from an arbitrary domain $\mathbf { X }$ to $\mathbf { x } ^ { \prime }$ from domain $\mathbf { X } ^ { \prime }$ , they simply perform $\pmb { x } ^ { \prime } = G _ { \mathbf { X } ^ { \prime } , \mathbf { X } } ( \pmb { x } ) = D e c o d e r _ { X ^ { \prime } } ( E n c o d e r _ { X } ( \pmb { x } ) )$ . The result of $E n c o d e r _ { X } ( { \pmb x } )$ can even be cached when translating to other domains as not to repeat computation. + +The main differences between the MDT methods and ours is that, unlike our method which does domain alignment in feature space, MDT methods adapt representations not in feature space but rather in raw pixel space; translating samples from one domain to the “style” of a other domains. This works well for limited domain shifts where the domains are similar in pixel-space, but can be too limiting for settings with larger domain shifts that results in poor performance in significant structural change of the samples in different domains. + +# C CONNECTION TO DOMAIN SEPARATION NETWORKS + +The method closest to our work is Domain Separation Networks (DSN) (Bousmalis et al. (2016)), which use the notion of auto-encoders to explicitly separate the feature representations private to each source/target domain from those that are shared between the domains. Although extending DSN to multiple domains might seem trivial, DSN requires an autoencoder per domain, making the model impractical in the case of more than a couple of domains. + +The overall loss of DSN consists of a reconstruction loss for each domain modeled by a shared decoder, a similarity loss such as MMD, which encourages domain invariance modeled by a shared encoder, and a dissimilarity loss modeled by two private encoders: one for the source domain and one for the target domain. While one could attempt to generalize DSN to multiple target domains by having individual per-target domain private encoders, doing so would prove problematic when the number of target domains is large — each private encoder would require a large "private" dataset to learn the private parameters. Precisely, for multiple $\bar { ( \cal M ) }$ target domains, we could train a DSN model with one shared encoder, $M + 1$ private encoder (one for each domain), and one shared decoder. This leads to $M + 3$ models to train that implies the number of models increases linearly with the number of domains, as does the required training time. Second, DSN uses an orthogonality constraint among the shared and the private representations which may not be strong enough to remove redundancy and enforce disentangling among different private spaces. Precisely, DSN defines the loss via a soft subspace orthogonality constraint between the private and shared representation of each domain. However, it does not enforce the private representation of different domains to be different that may result in redundancy of different private spaces. + +In addition, DSN enforces separation of spaces using the notion of Euclidean orthogonality, e.g., $\| z _ { s } - z _ { p } \| ^ { 2 }$ . In case of multiple target domains, this would result in learning of all pairs of private spaces independently. To address those deficiencies, we first explicitly couple different private encoders into a single private encoder model, $E _ { \theta _ { p } }$ of fig. 2 , which allows us to generalize to an arbitrary number of target domains. To assure that the information among the private and shared spaces is not shared (i.e., "orthogonal"), we define an informationtheoretic criteria enforced by a domain classifier, $\mathcal { D } _ { \psi }$ of fig. 2, which aims to segment the private space into clusters that correspond to individual target domains. By using $D _ { \psi }$ within the adversarial framework, MTDA-ITA learns simultaneously the shared and private features from different domains (see fig. 6). We showed in Sec. 4 that our model performs better than the trivial extension of DSNs to the multi-domain case. + +# D NETWORK ARCHITECTURE + +The network architecture used for the experiments is given in tab. 3. We use the following abbreviation for ease of presentation: ${ \bf N } =$ Neurons, $\mathbf { K } { = }$ Kernel size, ${ \mathrm { S } } { = } { }$ Stride size, $\begin{array} { r } { \boldsymbol { \mathrm { D } } \boldsymbol { = } } \end{array}$ Number of Domains, $\mathbf { C } =$ number of Classes. The transposed convolutional layer is denoted by DCONV. The residual basic block is denoted as RESBLK. + +Table 3: Network architecture for the experiments. + +
LayerEncoders (shared,private)
1CONV-(N16,K7,S1), ReLU
2CONV-(N32,K3,S2), ReLU
3CONV-(N64,K3,S2), ReLU
4RESBLK-(N64,K3,S1)
5RESBLK-(N64,K3,S1)
6RESBLK-(N64,K3,S1)
7RESBLK-(N64,K3,S1)
LayerDecoder
1RESBLK-(N64,K3,S1)
2RESBLK-(N64,K3,S1)
3RESBLK-(N64,K3,S1)
4RESBLK-(N64,K3,S1)
5DCONV-(N32,K3,S2), ReLU
6DCONV-(N16,K3,S2), ReLU
7DCONV-(N1,K1,S1), Tanh
LayerDiscriminator
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1),ReLU
3CONV-(N16,K3,S1),ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(ND), Softmax
LayerClassifier
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1), ReLU
3CONV-(N16,K3,S1), ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(NC), Softmax
+ +# E PROPOSED MODEL’S ALGORITHM + +The detailed optimization process of the proposed model is shown in Algorithm 1. + +# F ADDITIONAL EXPERIMENTS FOR DIGIT AND MULTI-PIE DATASETS + +The additional experiments for Digit dataset where we set MNIST-M and USPS as source domain is available in tab. 4. The additional experiments for Multi-PIE dataset where we set $C 0 5$ , $C 0 8$ and $C 0 9$ as source domain is available in tabs. $5 \& 6$ . + +# G PACS DATASET + +This dataset contains 9991 images $2 2 7 \times 2 2 7 \times 3$ dimension) across 7 categories (‘dog’, ‘elephant’, ‘giraffe’,‘guitar’, ‘house’, ‘horse’ and ‘person’) and 4 domains of different stylistic depictions (‘Photo’, ‘Art painting’, ‘Cartoon’ and ‘Sketch’). The very diverse depiction styles provide a significant gap between domains, coupled with the small number of data samples, making it extremely challenging for domain adaptation. Consequently, the dataset was originally used for multi-source to single target domain adaptation (Li et al. (2017)). Instead, we tackle a significantly more challenging problem of single-source to multiple target adaptation. Tab. 7 shows the classification accuracy of various methods. MTDA-ITA consistently + +![](images/18facd59a83865bc6e26c98e1f1fdea84883eb0fbbf03a8444befb1f847d7fe4.jpg) +Figure 4: Ablation of MTDA-ITA on Digit dataset. We show that each component of our method, Reconstruction loss, Classifier entropy loss with separating shared/private features, contributes to the overall performance. + +Table 4: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown in red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDA-ITA where each one correspond to a source-target pair. mp-DSN: extended DSN with multiple private encoder. $\bf \Pi ^ { * } U N I T$ trains with the extended SVHN $> 5 0 0 \mathrm { K }$ images vs ours 72K). $^ *$ PixelDA uses $( \approx 1 , 0 0 0 )$ of labeled target domain data as a validation set for tuning the hyper-parameters. + +
methodMM→SMM→MMM→UU→SU→MU→MM
Source Only40.00±0.6184.46 ± 0.2980.43± 0.5023.41± 0.5250.64 ± 0.3741.45±0.38
1-NN21.4582.1336.9015.3438.4518.54
CORAL (Sun & Saenko (2016))40.20 ±0.6084.90 ± 0.7087.54 ± 0.4438.90 ± 0.9685.01 ± 0.6160.45 ± 0.70
DANN (Ganin et al. (2016))51.80 ± 0.9161.05 ± 0.7185.34 ± 0.6435.50± 0.8477.40 ± 0.6461.60 ± 0.64
ADDA (Tzeng et al. (2017))40.64 ± 0.8692.82 ± 0.4880.70±0.4841.23 ± 0.7890.10± 0.5856.21 ± 0.79
c-ADDA35.43 ± 0.9488.47 ± 0.6174.19 ± 0.5839.36 ± 0.9984.67 ± 0.9452.54 ± 0.88
DTN (Zhang et al. (2015)) c-DTN48.80± 0.6688.80±0.3890.68 ± 0.3542.43 ± 0.6189.04 ± 0.3655.78 ± 0.40
UNIT (Liu et al. (2017))44.21 ± 0.6183.60 ± 0.5484.98 ± 0.4139.75 ± 0.6485.04 ± 0.4548.86 ± 0.54
DSN (Bousmalis et al. (2016))1 51.50 ± 0.64/90.60
c-DSN47.10 ± 0.5090.20± 0.3189.95± 0.2948.20 ± 0.5991.40 ± 0.3060.45 ± 0.35
1p-DSN45.00 ± 0.6084.60 ± 0.4084.80 ±0.3940.50 ± 0.6186.05 ± 0.4656.25 ± 0.50
mp-DSN47.15 ± 0.6481.96 ± 0.6083.03 ±0.4939.30 ± 0.5184.55 ± 0.5655.03 ± 0.60
s-MTDA-ITA85.51 ± 0.5483.24± 0.2438.30 ± 0.7487.40 ± 0.3555.47 ± 0.44
c-MTDA-ITA50.55± 0.18 47.32±0.1994.82±0.2189.05 ±0.2840.13±0.3087.10 ± 0.2561.01 ± 0.24
MTDA-ITA53.50 ± 0.2290.20 ±0.30 98.20 ± 0.1090.01 ± 0.24 94.10 ± 0.1141.10 ± 0.35 46.00 ± 0.4885.35 ± 0.28 91.50 ± 0.2360.31± 0.34 67.30 ± 0.15
+ +Table 5: Classification results on Multi-PIE dataset. The best is shown in red. + +
methodC05 →C08C05→C09C05→C13C05→C14
Source Only31.56 ± 0.4040.67 ± 0.3639.89 ± 0.2254.70± 0.25
1-NN CORAL (Sun & Saenko (2016))27.28 36.55 ± 0.6631.22 38.60 ± 0.6733.6647.04 55.29 ± 0.47
40.60 ± 0.58
DANN (citeganin2016domain)40.30 ± 0.6041.20 ± 0.6540.12 ± 0.6058.90 ± 0.38
ADDA (Tzeng et al. (2017)) C-ADDA33.21 ± 0.8130.86 ± 0.9052.44± 0.8070.18 ± 0.60
46.88 ± 0.6536.38 ± 0.8839.14 ± 0.8565.41 ± 0.69
DTN (Zhang et al. (2015))38.50 ± 0.5130.56 ± 0.4655.78 ± 0.3668.90 ± 0.31
c-DTN41.70 ± 0.4231.10 ± 0.4850.19 ± 0.4560.34 ± 0.35
PixelDA (Bousmalis et al. (2017))44.93 ±0.4244.75 ± 0.4545.18 ± 0.4546.88 ± 0.49
UNIT (Liu et al. (2017))44.47 ± 0.2144.47 ± 0.2144.47 ± 0.2044.51 ± 0.28
DSN (Bousmalis et al. (2016))45.12 ± 0.4644.35± 0.4948.12 ± 0.5375.00 ± 0.39
c-DSN42.52 ± 0.4838.54 ± 0.6434.15 ± 0.6469.45 ± 0.55
1p-DSN41.64 ± 0.5837.84 ± 0.6334.65 ± 0.4468.75 ± 0.85
mp-DSN41.30 ± 0.2835.14 ± 0.3534.40 ± 0.3565.70 ± 0.27
s-MTDA-ITA44.40 ± 0.2344.60 ± 0.2547.65 ± 0.2780.20 ± 0.13
c-MTDA-ITA40.49 ± 0.2540.70 ± 0.2542.80 ± 0.2571.60 ± 0.10
MTDA-ITA49.01 ± 0.20
48.20± 0.2753.13 ±0.2284.29 9±0.10
+ +# Algorithm 1 MDTA-ITA Algorithm + +Require: $\{ \mathbf { X } , \mathbf { Y } , \mathbf { D } \} { : \mathrm { M } }$ domain datasets. +$\lambda _ { r } , \lambda _ { c } , \lambda _ { d }$ : Model hyper-parameters. +Ensure: $\theta _ { s } , \theta _ { p } , \theta _ { c } , \phi , \psi$ : Model parameters. +1: Initialize $\bar { \theta } _ { s } , \theta _ { p } , \theta _ { c } , \phi , \psi$ ; +2: repeat 3: Sample a mini-batch from each of source/target domain datasets. +4: Update $\{ \theta _ { s } \}$ by minimizing $\mathcal { L } _ { s }$ in Eq.(10) through the gradient descent: $\begin{array} { r } { \theta _ { s } = \theta _ { s } - \eta \frac { \partial \mathcal { L } _ { s } } { \partial \theta _ { s } } } \end{array}$ . +5: Update {θp} by minimizing Lp in Eq.(9) through the gradient descent:θp = θp − η ∂Lp∂θp . +6: Update $\{ \theta _ { c } \}$ by minimizing $\mathcal { L } _ { c }$ in Eq.(8) through the gradient descent: $\begin{array} { r } { \theta _ { c } = \theta _ { c } - \eta \frac { \partial \mathcal { L } _ { c } } { \partial \theta _ { c } } } \end{array}$ 7: Update $\{ \phi \}$ by minimizing $\mathcal { L } _ { \phi }$ in Eq.(6) through the gradient descent: $\phi = \phi - \eta \frac { \partial \mathcal { L } _ { \phi } } { \partial \phi }$ . +8: Update $\{ \psi \}$ by minimizing $\mathcal { L } _ { \psi }$ in Eq.(5) through the gradient descent: $\begin{array} { r } { \psi = \psi - \eta \frac { \partial \mathcal { L } _ { s } } { \partial \psi } } \end{array}$ . +9: until Convergence; +10: return $\{ \theta _ { s } , \theta _ { p } , \theta _ { c } , \phi , \psi \}$ . + +achieves the best performance for all transfer tasks. Evaluations were obtained by training all models (ADDA, DSN, and ours) from scratch on the PACS dataset. Note that the overall performance figures are low due to the extreme difficulty of the transfer task, induced by large differences among domains. + +# H ANALYSIS OF SHARED/PRIVATE SPACE EMBEDDING + +In the experiments conducted, we showed that our approach is able to achieve better performance than the competing methods including the extended DSN with one private encoder (1p-DSN) which is the most similar method to ours. + +![](images/07d4618a501cd10a00bd625bd5df1054e718feb2cb44cd9581d0cca9092af8f0.jpg) +Figure 5: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features. (c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters in the right column represent points from the shared space, while the smaller ones are from the private spaces. (e),(f) learned features for 1p-DSN. + +Indeed, fig. 5 depicts the embedding of the MTDA-ITA learned private/shared features, those of 1p-DSN and the original features from different domains for Digit datasets (SVHN is the source). + +Notice that both MTDA-ITA and 1p-DSN reduces the domain mismatch for the shared features (circle markers in fig. 5) and separate the shared features from private features. On the other hand, MTDA-ITA increases the domain separation for the private features (triangle markers, pure and well-separated domain clusters in fig. 5c) while 1p-DSN is unable to enforce the private representation of different domains to be different (fig. 5e) that may result in redundancy of different private spaces. This is partially due to the proposed multi-domain separation loss through the use of the domain classifier $D$ , which penalizes the domain mismatch for the shared features and rewards the mismatch for the private features, something the 1p-DSN fails to account for. Moreover, as supported by the quantitative results in tab. 1, the class label separation in + +
methodC08→C05C08→C09C08→C13C08→C14C09→C05C09→C08C09→C13C09→C14
Source Only33.70±0.3350.10±0.2850.80±0.3240.13±0.2633.32±0.4448.24±0.3249.24±0.3036.19±0.27
1-NN28.7535.3939.7932.1326.8235.3034.2628.41
CORAL (Sun & Saenko (2016))35.89 ± 0.4455.79±0.5060.00 ±0.2940.67 ± 0.4835.89 ± 0.4051.56±0.4350.45 ± 0.4140.67± 0.35
DANN (Ganin et al. (2016))40.20±0.5056.89 ± 0.3955.83 ±0.4043.25±0.4150.63±0.3858.40 ± 0.5155.81 ± 0.5348.90 ±0.43
ADDA(Tzeng et al. (2017))37.40±0.6858.40±0.7360.40±0.8342.10±0.4829.40±0.7053.30±0.4945.30± 0.5338.30±0.63
c-ADDA41.60±0.6439.65 ±0.7050.00 ±0.5246.25±0.5245.01±0.6352.14 ± 0.5337.43 ± 0.6043.26±0.58
DTN (Zhang et al. (2015))44.13 ± 0.4157.42 ± 0.4255.89±0.4845.76± 0.3944.53± 0.4957.34±0.3552.43±0.3851.55 ± 0.40
c-DTN45.10 ±0.4449.78±0.5047.43 ±0.4645.79 ±0.4849.80±0.4055.69 ± 0.3550.10 ± 0.3852.31 ±0.29
PixelDA (Bousmalis et al. (2017))46.45 ± 0.4544.33±0.3844.87 ± 0.4146.83±0.2945.63±0.3416.37 ± 0.2745.43 ± 0.3547.00± 0.49
UNIT (Liu et al. (2017))43.88±0.1843.99±0.2344.47±0.1944.47 ± 0.2444.47 ± 0.1743.95± 0.2144.64±0.2244.47 ±0.19
DSN (Bousmalis et al. (2016)) c-DSN46.25± 0.5347.50±0.6062.15 ± 0.5839.72± 0.5545.85±0.4856.65±0.5056.5±0.3842.87±0.43
45.82 ± 0.5344.64± 0.4245.60 ± 0.4846.32±0.5245.18 ± 0.4745.52 ± 0.5544.79 ± 0.5347.37 ± 0.48
1p-DSN44.12 ± 0.7344.14 ± 0.2045.00 ±0.3845.62 ± 0.4244.78 ±0.4745.02 ±0.6544.21 ± 0.4846.97 ± 0.38
mp-DSN42.19±0.4644.70±0.5342.47 ± 0.4840.50±0.3945.00±0.5143.80±0.5045.79±0.4842.39 ±0.49
s-MTDA-ITA44.77±0.1945.61±0.1860.00±0.2746.70±0.2849.06±0.2455.33±0.2259.90±0.3050.64± 0.26
c-MTDA-ITA MTDA-ITA44.35± 0.2742.67 ± 0.2458.90 ±0.2644.32 ± 0.2646.74 ±0.2254.11 ± 0.2156.89 ± 0.2349.64 ± 0.19
46.30 ± 0.2560.60 ± 0.1860.50±0.1950.40±0.2055.59± 0.2557.80±0.2164.20 ±0.1856.34 ± 0.20
+ +Table 6: Classification results on Multi-PIE dataset. The best (red). + +
methodP→AP→CP→SA→PA→CA→S
1-NN15.2818.1625.6022.7019.7522.70
ADDA(Tzeng et al. (2017))24.35 ± 2.3720.12 ± 2.5022.45 ± 2.1132.57 ± 2.7017.68 ± 2.0418.90 ± 2.48
DSN (Bousmalis et al. (2016))28.42 ±2.1221.14 ± 2.082.04 ± 1.9029.54 ± 1.9525.89 ± 1.8824.69 ± 2.08
s-MTDA-ITA28.02 ± 1.5921.64±1.2426.24 ± 1.6031.06 ± 1.5025.09±1.4025.89±1.03
c-MTDA-ITA25.35 ± 1.8020.24 ± 1.3923.64 ± 1.6026.54 ± 1.3320.30 ± 1.2922.38 ± 1.45
MTDA-ITA31.40 ±1.5523.05 ±1.0428.24 ±1.7835.74 ± 1.5027.00 ±1.2528.90 ±1.60
+ +Table 7: Classification results on PACS dataset classification. A:Art-painting, C:Cartoon, S:Sketch, P:Photo. The best (red). + +the shared space for 1p-DSN, fig. 5f, is still evident but not as strong as in the MTDA-ITA, fig. 5d. This can be attributed to the lack of redundancy in the private space that helps MTDA-ITA to learn more disentangled shared features and usage of the target samples during training, something the 1p-DSN fails to account for. + +![](images/f7a643eee7751386eefa467035b5bfae216f67ec9df6317cb9b21d6dd8832516.jpg) +Figure 6: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features. (c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters21 in the right column represent points from the shared space, while the smaller ones are from the private spaces. The remaining figures depict the learned features without: (e),(f) the classifier entropy loss, MTDA-woE; (g),(h) the private encoder, MTDA-woP; (i),(j) the reconstruction loss/decoder, MTDA-woR; and (k),(l) the multi-domain separation loss, MTDA-woD. \ No newline at end of file diff --git a/parse/train/BJxLH2AcYX/BJxLH2AcYX_content_list.json b/parse/train/BJxLH2AcYX/BJxLH2AcYX_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..542cab1b6d7bb36968e1211221acbfb605a7e699 --- /dev/null +++ b/parse/train/BJxLH2AcYX/BJxLH2AcYX_content_list.json @@ -0,0 +1,1941 @@ +[ + { + "type": "text", + "text": "UNSUPERVISED MULTI-TARGET DOMAIN ADAPTATION:AN INFORMATION THEORETIC APPROACH", + "text_level": 1, + "bbox": [ + 148, + 114, + 852, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 151, + 188, + 367, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 251, + 544, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised domain adaptation (uDA) models focus on pairwise adaptation settings where there is a single, labeled, source and a single target domain. However, in many real-world settings one seeks to adapt to multiple, but somewhat similar, target domains. Applying pairwise adaptation approaches to this setting may be suboptimal, as they fail to leverage shared information among multiple domains. In this work, we propose an information theoretic approach for domain adaptation in the novel context of multiple target domains with unlabeled instances and one source domain with labeled instances. Our model aims to find a shared latent space common to all domains, while simultaneously accounting for the remaining private, domain-specific factors. Disentanglement of shared and private information is accomplished using a unified information-theoretic approach, which also serves to establish a stronger link between the latent representations and the observed data. The resulting model, accompanied by an efficient optimization algorithm, allows simultaneous adaptation from a single source to multiple target domains. We test our approach on three challenging publicly-available datasets, showing that it outperforms several popular domain adaptation methods. ", + "bbox": [ + 207, + 281, + 792, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 150, + 512, + 310, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world data, the training and test instances often do not come from the same underlying distribution (Sun et al. (2016)). For example, in the task of object recognition/classification from image data, this may be due to the image noise, changes in the object view, etc., which induce different biases in the observed data sampled during the training and test stage. Consequently, assumptions made by traditional learning algorithms are often violated, resulting in degradation of the algorithms’ performance during inference of test data. Domain Adaptation (DA) approaches (Fernando et al. (2013); Gong et al. (2012); Kodirov et al. (2015); Yoo et al. (2016)) aim to tackle this by transferring knowledge from a source domain (training data) to an unlabeled target domain (test data) to reduce the discrepancy between the source and target data distributions, typically by exploring domain-invariant data structures. ", + "bbox": [ + 148, + 542, + 851, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing DA methods tackle the adaptation problems in one of the two settings: (semi)supervised DA and unsupervised DA (Csurka (2017)). The former assume that in addition to the labeled data of the source domain, some labeled data from the target domain are also available for training/adapting the classifiers. In contrast, the latter does not require any labels from the target domain but rather explores the similarity in the data distributions of the two domains. In this work, we focus on the unsupervised DA (uDA) scenario, which is more challenging due to the lack of correspondences in source and target labels. ", + "bbox": [ + 148, + 675, + 849, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most works on uDA today focus on a single-source-single-target-domain scenario. However, in many real-world applications, unlabeled data may come from different domains, thus, with different statistical properties but with common task-related content. For instance, we may have access to images of the same class of objects (e.g., cars) recorded by various types of cameras, and/or under different camera views and at different times, rendering multiple different domains (e.g., datasets). Likewise, facial expressions of emotions, such as joy and surprise, shown by different people and recorded under different views, result in multiple domains with varying data distributions. In most cases, these domains have similar underlying data distributions, which can be leveraged to build more effective and robust classifiers for tasks such as the object or emotion recognition across multiple datasets/domains. To this end, traditional uDA methods focus on the single-source-single-target DA scenario. However, in the presence of multiple domains, as typically encountered in real-world settings, this pair-wise adaptation approach may be suboptimal as it fails to leverage simultaneously the knowledge shared across multiple task-related domains. Recently, Zhao et al. (2017) showed that by having access to multiple source domains can facilitate better adaptation to a single target domain, when compared to the pair-wise DA approach. While this is intuitive due to the access to multiple labelled source domains, offering more adaptation flexibility for the target domain (i.e., by efficiently exploring the data labels across multiple source domains that are most related to the target domain), it comes at the expense of the data labelling in multiple source domains, which can be costly and time-consuming. In either case, a single source domain or readily available multiple source domains, to the best of our knowledge, a simultaneous adaptation to multiple and unlabelled target domains remains an unexplored DA scenario. However, this DA scenario is important as we usually have access to multiple unlabeled domains; yet, the adaptation process is also more challenging due to the lack of supervision in the target domains. Still, multi-target DA can have advantages over a single-target DA when: (i) there is direct knowledge sharing between the source and multiple target domains (fig. 1a), and (ii) the source and a target domain are related through another target domain (fig. 1b). While this seems intuitive, it is critical how the data from multiple unlabelled target domains are leveraged within the multi-target DA approach, in order to improve its performance over the single target DA approaches and naive fusion of multiple target domains. ", + "bbox": [ + 148, + 766, + 849, + 821 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36886e0179e0304a2140f74c7b598abbe23c7d9c3021a983addc10b904e84af3.jpg", + "image_caption": [ + "Figure 1: Illustration of domains with common (a) and pairwise-shared spaces (b). We tackle the domain adaptation task when all domains share a common task/space, which is then leveraged to transfer knowledge across multiple target domains. " + ], + "image_footnote": [], + "bbox": [ + 210, + 106, + 784, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 329, + 851, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a Multi-Target DA-Information-Theoretic-Approach (MTDA-ITA) for single-sourcemulti-target DA. We exploit a single source domain and focus on multiple target domains to investigate the effects of multi-target DA; however, the proposed approach can easily be extended to multiple source domains. This approach leverages the data from multiple target domains to improve performance compared to individually learning from pair-wise source-target domains. Specifically, we simultaneously factorize the information from each available target domain and learn separate subspaces for modeling the shared (i.e., correlated across the domains) and private (i.e., independent between the domains) subspaces of the data (Salzmann et al. (2010)). To this end, we employ deep learning to derive an information theoretic approach where we jointly maximize the mutual information between the domain labels and private (domain-specific) features, while minimizing the mutual information between the the domain labels and the shared (domaininvariant) features. Consequently, the more robust feature representations are learned for each target domain by exploiting dependencies between multiple target domains. We show on benchmark datasets for DA that this approach leads to overall improved performance on each target domain, compared to independent DA for each pair of source-target domains, or the naive combination of multiple target domains, and state-of-the-art models applicable to the target task. ", + "bbox": [ + 148, + 641, + 852, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 145, + 119, + 852, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 THE PROPOSED METHOD ", + "text_level": 1, + "bbox": [ + 148, + 169, + 390, + 185 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without loss of generality, we consider a multi-class ( $K$ -class) classification problem as the running example. Furthermore, let $( \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } ) = \\{ ( \\pmb { x } _ { i } , \\pmb { y } _ { i } , \\pmb { d } _ { i } ) \\} _ { i = 0 } ^ { N }$ be a collection of $M$ domains (a labeled source domain, and $M - 1$ unlabeled target domains), where $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }$ denotes the $i$ -th sample, and $\\pmb { y } _ { i } = [ y _ { i } ^ { 1 } , y _ { i } ^ { 2 } , . . . , y _ { i } ^ { K } ]$ and $d _ { i } = [ d _ { i } ^ { 1 } , d _ { i } ^ { 2 } , . . . , d _ { i } ^ { M } ]$ are the $K$ -D and $M$ -D encoding of the class and domain labels for $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }$ , respectively. Note that the class labels are only available for the source samples. ", + "bbox": [ + 147, + 200, + 852, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The latent space representation of the data point $_ { \\textbf { \\em x } }$ is denoted as $\\boldsymbol { z } = [ z _ { s } , z _ { p } ]$ , where $z _ { s }$ and $z _ { p }$ are the (latent) shared and private features of the data point $_ { \\textbf { \\em x } }$ , respectively. By factorizing the joint distribution $p ( { \\pmb x } , { \\pmb y } , d , z _ { s } , z _ { p } )$ as ", + "bbox": [ + 148, + 281, + 849, + 323 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/6ba6e0a7950e1dd2dbca048af80b4411ac3bcf9c16cad5bded0e0a26e8982113.jpg", + "text": "$$\n\\begin{array} { r } { p ( { \\pmb x } , { \\pmb y } , { \\pmb d } , z _ { s } , z _ { p } ) = p ( { \\pmb x } ) p ( { \\pmb d } ) p ( z _ { s } | { \\pmb x } ) p ( z _ { p } | { \\pmb x } ) p ( { \\pmb y } | z _ { s } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 316, + 321, + 679, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we propose to maximize the following objective function: ", + "bbox": [ + 148, + 343, + 526, + 358 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/2ebeb1570173651be3e9d961e9dfdaa8a2afd212b856849a8d63e5244997c423.jpg", + "text": "$$\n\\begin{array} { r } { \\boldsymbol { \\mathcal { L } } ( \\boldsymbol { \\theta } _ { s } , \\boldsymbol { \\theta } _ { p } , \\boldsymbol { \\theta } _ { c } ; \\boldsymbol { x } , \\boldsymbol { y } , \\boldsymbol { d } ) = \\lambda _ { r } \\boldsymbol { I } ( \\boldsymbol { x } ; \\boldsymbol { z } ) + \\lambda _ { c } \\boldsymbol { I } ( \\boldsymbol { y } ; \\boldsymbol { z } _ { s } ) + \\lambda _ { d } \\big ( \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { p } ) - \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { s } ) \\big ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 246, + 366, + 750, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p ( { \\pmb x } )$ and $p ( d )$ denote the underlying (true) data distribution and domain label distribution, respectively, $I ( { \\pmb x } ; { \\pmb y } )$ denotes the Mutual Information between the random variables $_ { \\textbf { \\em x } }$ and $\\textbf { { y } }$ . $\\lambda _ { r } , \\lambda _ { c }$ and $\\lambda _ { d }$ denote the hyper-parameters controlling the weights of the objective terms. The proposed objective function (2) maximizes the three terms described below: ", + "bbox": [ + 148, + 393, + 851, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "• $I ( { \\pmb x } ; z )$ : encourages the latent features (both shared and private) to preserve information about the data samples (that can be used to reconstruct $_ { \\textbf { \\em x } }$ from $_ { z }$ ). \n• $I ( \\pmb { y } ; \\pmb { z } _ { s } )$ : enables to correctly predict the true class label of the samples out of their common shared features. \n• $I ( d ; z _ { p } ) - I ( d ; z _ { s } )$ : encourages the latent private features to preserve the information about the domain label and penalizes the latent shared features to be domain informative. This not only reduces the redundancy in the shared and private features, but also, penalizes the redundancy of different private spaces, while preserving the shared information. ", + "bbox": [ + 189, + 462, + 851, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An additional term could be used to minimize the mutual information between the shared $( z _ { s } )$ and private $( z _ { p } )$ features. However, computing the mutual information (even approximating it) is intractable due to the highly complex joint distribution $p ( z _ { s } , z _ { p } )$ . Since we want $z _ { s }$ and $z _ { p }$ features to encode different aspects of $_ { \\textbf { \\em x } }$ , we enforce such constraint by jointly maximizing the term: $I ( d ; \\dot { z } _ { p } ) - I ( d ; z _ { s } )$ . ", + "bbox": [ + 147, + 597, + 851, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 OPTIMIZATION ", + "text_level": 1, + "bbox": [ + 148, + 671, + 294, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following lower bound for mutual information is derived using the non-negativity of KL-divergence (Barber & Agakov (2003)); i.e., $\\begin{array} { r } { \\Sigma _ { { \\pmb x } } p ( { \\pmb x } | z ) \\ln \\frac { p ( { \\pmb x } | z ) } { q ( { \\pmb x } | z ) } \\ge 0 } \\end{array}$ gives: ", + "bbox": [ + 147, + 698, + 851, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/a13d30e2817b0209208e09198a760859aaa7aca32fa8311d73bdbdb853ec071e.jpg", + "text": "$$\nI ( \\pmb { x } ; z ) \\geq H ( \\pmb { x } ) + \\mathbb { E } _ { p ( \\pmb { x } , z ) } [ \\ln q ( \\pmb { x } | z ; \\phi ) ]\n$$", + "text_format": "latex", + "bbox": [ + 364, + 739, + 633, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H ( { \\pmb x } )$ denotes the Shanon Entropy (Lin (1991)) of the random variable $_ { \\textbf { \\em x } }$ . $q ( { \\pmb x } | { \\pmb z } ; { \\pmb \\phi } )$ is any arbitrary distribution parameterized by $\\phi$ . We need a variational distribution $q ( { \\pmb x } | { \\pmb z } ; \\phi )$ because the posterior distribution $p ( \\pmb { x } | \\pmb { z } ) = p ( \\pmb { z } | \\pmb { x } ) p ( \\pmb { x } ) / p ( \\pmb { z } )$ is intractable since the true data distribution $p ( { \\pmb x } )$ is assumed to be unknown. Similarly, we can derive lower bounds for $I ( d ; z _ { p } ) \\geq H ( d ) + \\mathbb { E } _ { p ( d , z _ { p } ) } [ \\ln q ( d | z _ { p } ; \\psi ) ]$ and $I ( d ; z _ { s } ) \\geq$ ", + "bbox": [ + 148, + 765, + 851, + 824 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8e3477bd318324bc4827d471de794c2c830ef3f1de87446abdf6c005292de940.jpg", + "image_caption": [ + "Figure 2: MDTA-ITA: The encoder $E _ { s } ( { \\pmb x } )$ captures the feature representations $( z _ { s } )$ for a given input sample $_ { \\pmb { x } }$ that are shared among domains. $E _ { p } ( { \\pmb x } )$ captures domain–specific private features $( z _ { p } )$ using the shared private encoder. The shared decoder $F \\big ( z _ { p } , z _ { s } \\big )$ learns to reconstruct the input sample by using both the private and shared features. The domain classifier $D$ learns to correctly predict the domain labels of the actual samples from both their shared and private features while the classifier $C$ learns to correctly predict the class labels from the shared features. " + ], + "image_footnote": [], + "bbox": [ + 336, + 92, + 671, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$H ( d ) + \\mathbb { E } _ { p ( d , z _ { s } ) } [ \\ln q ( d | z _ { s } ; \\psi ) ]$ , where $q ( d | z _ { p } ; \\psi )$ is any arbitrary distribution parameterized by $\\psi$ . 1 We further drive lower bound for $I ( \\pmb { y } ; \\pmb { z } _ { s } )$ as $I ( \\pmb { y } ; \\pmb { z _ { s } } ) \\geq H ( \\pmb { y } ) + \\mathbb { E } _ { p ( \\pmb { y } , \\pmb { z _ { s } } ) } [ \\ln q ( \\pmb { y } | \\pmb { z _ { s } } ; \\pmb { \\theta _ { c } } ) ]$ , where $q ( \\pmb { y } | \\pmb { z } _ { s } ; \\theta _ { c } )$ is a variational distribution parameterized by $\\theta _ { c }$ approximating $p ( \\pmb { y } | \\boldsymbol { z } _ { s } )$ . ", + "bbox": [ + 147, + 371, + 852, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let next $E _ { s } ( \\pmb { x } ; \\theta _ { s } )$ be a function parameterized by $\\theta _ { s }$ that maps a sample $_ { \\textbf { \\em x } }$ to its corresponding shared feature $z _ { s }$ , and $E _ { p } ( \\pmb { x } ; \\theta _ { p } )$ be an analogous function which maps $_ { \\textbf { \\em x } }$ to $z _ { p }$ , the feature that is private to each domain (fig. 2). We also define $F ( z _ { s } , z _ { p } ; \\phi )$ as a decoding function mapping the concatenation of the latent features $z _ { s }$ and $z _ { p }$ to a sample reconstruction $\\hat { \\pmb x }$ , and $D ( z ; \\psi )$ as a decoding function mapping $z _ { s }$ and $z _ { p }$ to a $M$ -dimensional vector: the predictions of the domain label $\\hat { d }$ . Finally, $C ( z _ { s } ; \\theta _ { c } )$ is a task-specific function mapping $z _ { s }$ to a $K$ -dimensional probability vector of the class label $\\hat { y }$ . ", + "bbox": [ + 147, + 422, + 851, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We represent $\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { \\bar { N } } \\delta ( { \\bf d } - { \\bf d } _ { i } ) ) } \\end{array}$ $p ( { \\pmb d } ) , p ( { \\pmb x } ) , p ( { \\pmb y } )$ as in the case of variational autoencoders (VAE) (Abbasnejad et al. (2017); as the empirical distribution of a finite training set (e.g. $\\begin{array} { r l } { p ( d ) } & { { } = } \\end{array}$ $\\mathrm { P u }$ et al. (2017)), $p ( \\pmb { z } _ { s } | \\pmb { x } ) , p ( \\pmb { z } _ { p } | \\pmb { x } )$ as deterministic functions of $_ { \\textbf { \\em x } }$ as $p ( z _ { s } | \\pmb { x } ) = \\delta \\big ( z _ { s } - E _ { s } ( \\pmb { x } ; \\theta _ { s } ) \\big )$ and $p ( z _ { p } | \\pmb { x } ) = \\delta \\big ( z _ { p } - E _ { p } ( \\pmb { x } ; \\theta _ { p } ) \\big )$ , and the variational distributions $q ( \\pmb { y } | \\pmb { z } _ { s } )$ , $q ( { \\pmb x } | z )$ and , $q ( d | z )$ as ", + "bbox": [ + 147, + 516, + 852, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/8e79b2c4472bab8e005ba23f78447ebc5b45bb38aeee0613256586456be89703.jpg", + "text": "$$\nq ( \\boldsymbol { y } | \\boldsymbol { z } _ { s } ) = \\mathrm { S o f t M a x } ( C ( \\boldsymbol { z } _ { s } ; \\boldsymbol { \\theta } _ { c } ) ) , q ( d | \\boldsymbol { z } ) = \\mathrm { S o f t M a x } ( D ( \\boldsymbol { z } ; \\boldsymbol { \\psi } ) ) , q ( \\boldsymbol { x } | \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\propto \\exp ( \\| \\boldsymbol { x } - \\boldsymbol { F } ( \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\| _ { 1 } )\n$$", + "text_format": "latex", + "bbox": [ + 160, + 589, + 816, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Softmax $( \\cdot )$ denotes the softmax or normalized exponential function (Bridle (1990)), and $\\left. . \\right. _ { 1 }$ denotes the $L _ { 1 }$ norm. Then, the optimization task can be posed as a minimax saddle point problem, where we use adversarial training to maximize (2) w.r.t. the parameters $( \\theta _ { s } , \\theta _ { p } , \\theta _ { c } )$ , and to minimize (2) w.r.t. the parameters $( \\phi , \\psi )$ , using Stochastic Gradient Descent (SGD). ", + "bbox": [ + 147, + 621, + 852, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimizing the parameters $\\phi$ of the decoder $F$ ", + "text_level": 1, + "bbox": [ + 148, + 684, + 470, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/1668b7a1b50aa39a5963118532a442c2a3e05b841494ce364eb9929b9bed6412.jpg", + "text": "$$\n\\hat { \\phi } = \\underset { \\phi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { F } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\pmb { x } _ { i } - F \\big ( E _ { s } ( \\pmb { x } _ { i } ) , E _ { p } ( \\pmb { x } _ { i } ) \\big ) \\| _ { 1 } .\n$$", + "text_format": "latex", + "bbox": [ + 307, + 705, + 691, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoder $F ( z _ { s } , z _ { p } ; \\phi )$ is trained in such a way so as to minimize the difference between original input $_ { \\textbf { \\em x } }$ and its decoding from corresponding shared and private features via the decoder $F$ . ", + "bbox": [ + 147, + 756, + 849, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimizing the parameters $\\psi$ of the domain classifier $D$ ", + "text_level": 1, + "bbox": [ + 148, + 119, + 537, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/21e90163bfe21babae9cdc9f8986f7fd1d54c2568bd6a5fd21c5c4b3297d1048.jpg", + "text": "$$\n\\hat { \\psi } = \\underset { \\psi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { D } = - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\pmb { x } _ { i } ) \\big ) - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { p } ( \\pmb { x } _ { i } ) \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 243, + 138, + 754, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$D ( z ; \\psi )$ can be considered as a classifier whose task is to distinguish between the shared/private features of the different domains. More precisely, the two terms in Eq. 6 encourage $D$ to correctly predict the domain labels from the shared and private features, respectively. ", + "bbox": [ + 147, + 188, + 851, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing the parameters $\\theta _ { c }$ of the label classifier $C$ ", + "text_level": 1, + "bbox": [ + 147, + 236, + 519, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/8efe30add848b2c029835c54b2d4e7f3d10df2c29d3d034dfe1abb77b7bca699.jpg", + "text": "$$\n\\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\big \\{ - H ( \\pmb { y } ) - \\mathbb { E } _ { p ( \\pmb { y } , z _ { s } ) } \\big [ \\ln q ( \\pmb { y } | z _ { s } ) \\big ] \\big \\} .\n$$", + "text_format": "latex", + "bbox": [ + 330, + 256, + 668, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we have access to the source labels, $H ( \\pmb { y } )$ is a constant for source samples. we can approximate $H [ \\pmb { y } ]$ for the target samples using the output of the classifier $C$ , leading to the following optimization problem: ", + "bbox": [ + 150, + 290, + 851, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/7dadae5cfa60703855a0e15efbcd5d3ff98636b04efe028d1e689cf41d3c6c58.jpg", + "text": "$$\n\\begin{array} { r l } & { \\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { C } = - \\ \\frac { 1 } { N } \\displaystyle \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) } \\\\ & { \\qquad + \\ \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\displaystyle \\frac { 1 } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) \\bigg ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 323, + 815, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N _ { s }$ denotes the number of source samples. Intuitively, we enforce the classifier $C$ to correctly predict the class labels of the source samples by the first term in Eq. 8. We use the second term to minimize the entropy of $q ( \\pmb { y } | \\pmb { z } _ { s } )$ for the target samples; effectively, reducing the effects of \"confusing\" labels of target samples, as given by $p ( \\pmb { y } | \\pmb { z } _ { s } )$ that leads to decision boundaries occur far away from target data-dense regions in the feature space. The intuition behind the last term is that by minimizing only the entropy (second term), we may arrive at a degenerate solution where every target point $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ is assigned to the same class. Hence, the last term encourages the classifier $C$ to have balanced labeling for the target samples where it reaches its minimum, $\\ln K$ , when each class is selected with uniform probability. ", + "bbox": [ + 147, + 416, + 852, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing the parameter $\\theta _ { p }$ of the private encoder $E _ { p }$ ", + "text_level": 1, + "bbox": [ + 148, + 536, + 531, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/7ae28fd7f8bbf8c8f1b3af03a71b8eaa401fada8a1326906381d325d31c0b30a.jpg", + "text": "$$\n\\hat { \\theta _ { p } } = \\underset { \\theta _ { p } } { \\operatorname { a r g m i n } } \\ \\mathcal { L } _ { P } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } _ { i } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } D \\big ( E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 223, + 556, + 774, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first term in Eq. 9 encourages the private encoder $E _ { p }$ to preserve the recovery ability of the private features. The second term enforces distinct private features be produced for each domain by penalizing the representation redundancy in different private spaces. This, in turn, encourages moving this common information from multiple domains to their shared space. ", + "bbox": [ + 147, + 603, + 849, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing the parameter $\\theta _ { s }$ of the shared encoder $E _ { s }$ ", + "text_level": 1, + "bbox": [ + 147, + 666, + 527, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a13c7de608ab3bf0d8508ee0dcc844ccff26109d6c0a9c40a3e43e86f9cbf02e.jpg", + "text": "$$\n\\begin{array} { c } { \\displaystyle \\hat { \\theta _ { s } } = \\arg \\operatorname* { m i n } _ { \\theta _ { s } } \\mathcal { L } _ { S } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { c } } { N } \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle + \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\frac { 1 } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) \\bigg ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 685, + 816, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first term in Eq. 10 encourages the shared encoder $E _ { s }$ to preserve the recovery ability of the shared features. The second term is the source domain classification loss penalty that encourages $E _ { s }$ to produce discriminative features for the labeled source samples. The third term simulates the adversarial training by trying to fool the domain classifier $D$ when predicting the domain labels $^ d$ , given the shared features $z _ { s }$ . The effect of this is two-fold: (i) the rendered shared features are more distinct from the corresponding private features, (ii) the shared features of different domains are encouraged to be similar to each other. The last two terms encourage $E _ { s }$ to produce the shared features for target samples so that the classifier is confident on the unlabeled target data, driving the shared features away from the decision boundaries. To train our model, we alternate between updating the shared encoder $E _ { s }$ , the private encoder $E _ { p }$ , the decoder $F$ , the classifier $C$ and the domain classifier $D$ using the SGD algorithm (see Algorithm 1 in Appendix E for more details). ", + "bbox": [ + 148, + 119, + 851, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 RELATED WORK ", + "text_level": 1, + "bbox": [ + 148, + 286, + 318, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There has been extensive prior work on domain adaptation (Csurka (2017)). Recent papers have focused on transferring deep neural network representations from a labeled source dataset to an unlabeled target domain, where the main strategy is to find a feature space such that the confusion between source and target distributions in that space is maximized ( Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al. (2017); Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017); Yan et al. (2017); Bousmalis et al. (2017)). For this, it is critical to first define a measure of divergence between source and target distributions. For instance, several methods have used the Maximum Mean Discrepancy (MMD) loss for this purpose (Bousmalis et al. (2017); Zellinger et al. (2017); Long et al. (2014)). MMD computes the norm of the difference between two domain means in the reproducing Kernel Hilbert Space (RKHS) induced by a pre-specified kernel. The Deep Adaptation Network (DAN) (Long et al. (2015)) applied MMD to layers embedded in a RKHS, effectively matching higher order statistics of the two distributions. The deep Correlation Alignment (CORAL) method (Sun & Saenko (2016)) attempts to match the mean and covariance of the two distributions. Deep Transfer Network (DTN) (Zhang et al. (2015)) achieved source/target distribution alignment via two types of network layers based on MMD distance: the shared feature extraction layer, which learns a subspace that matches the marginal distributions of the source and the target samples, and the discrimination layer, which matches the conditional distributions by classifier transduction. ", + "bbox": [ + 147, + 321, + 851, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently proposed unsupervised DA methods (Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al. (2017); Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017)) operate by training deep neural networks using adversarial training, which allows the learning of feature representations that are simultaneously discriminative of source labels, and indistinguishable between the source and target domain. For instance, Ganin & Lempitsky (2015) proposed a DA mechanism called Domain-Adversarial Training of Neural Networks (DANN), which enables the network to learn domain invariant representations in an adversarial way by adding a domain classifier and back-propagating inverse gradients. Adversarial Discriminative Domain Adaptation (ADDA) (Tzeng et al. (2017)) learns a discriminative feature subspace using the source labels, followed by a separate encoding of the target data to this subspace using an asymmetric mapping learned through a domain-adversarial loss. Liu et al. (2017) makes a shared-latent space assumption and proposes an unsupervised image-to-image translation (UNIT) framework based on Coupled GANs (Liu & Tuzel (2016)). Another example is the pixel-level domain adaptation models that perform the distribution alignment not in the feature space but directly in raw pixel space. PixelDA (Bousmalis et al. (2017)) uses adversarial approaches to adapt source-domain images as if drawn from the target domain while maintaining the original content. ", + "bbox": [ + 148, + 550, + 851, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While these approaches have shown success in DA tasks with single source-target domains, they are not designed to leverage information from multiple domains simultaneously. More recently, Zhao et al. (2017) introduced an adversarial framework called MDAN for multiple source single target domain adaptation where a domain classifier, induced by minimizing the H-divergence between multiple source and a target domain, is used to align their feature distributions in a shared space. Instead, in our approach we focus on multi-target DA where we perform adaptation of multiple unlabelled target domains. Although both our model and MDAN use the similar notion of the domain classifier to minimize the domain mismatch in shared space, the domain classifier induced by our information-theoretic (IT) loss also acts to separate domains in the private space (see Eqs. $6 \\& 9$ for more details), improving the essential reconstruction ability, similar to (Bousmalis et al. (2016)). We provided how our model is related to IT representation learning approaches, and multiple domain transfer networks in Appendices A and B respectively. In Appendix C, we also clearly contrasted our model with DSN model which also uses the notion of auto-encoders to explicitly separate the feature representations private to each source/target domain from those that are shared between the domains. ", + "bbox": [ + 148, + 766, + 851, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 119, + 851, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 EXPERIMENTAL RESULTS ", + "text_level": 1, + "bbox": [ + 150, + 265, + 392, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the proposed method with state-of-the-art methods on standard benchmark datasets: a digit classification task that includes 4 datasets: MNIST (LeCun et al. (1998)), MNIST-M (Ganin et al. (2016)), SVHN (Netzer et al. (2011)), USPS (Tzeng et al. (2017)), Multi-PIE expression recognition dataset2, and PACS multi-domain image recognition benchmark (Li et al. (2017)), a new dataset designed for the crossdomain recognition problems (the details for this experiment is available in Appendix G). Fig. 3 illustrates image samples from different datasets and domains. We evaluate the performance of all methods with classification accuracy metric. We repeated each experiment 5 times and report the average and the standard deviation of the accuracy. ", + "bbox": [ + 148, + 296, + 852, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We used ADAM (Kingma & Ba (2015)) for training; the learning rate was set to 0.0002 and momentum parameters to 0.5 and 0.999. We used batches of size 16 from each domain, and the input images were meancentered/rescaled to $[ - 1 , 1 ]$ . The hyper-parameters are empirically set as $\\lambda _ { r } = 1 . 0 , \\lambda _ { c } = 0 . 0 1$ , $\\lambda _ { d } = 0 . 2 0$ . For the network architecture, our private/shared encoders consisted of three convolutional layers as the front-end and four basic residual blocks as the back-end. The decoder consisted of four basic residual blocks as the front-end and four transposed convolutional layers as the back-end. The discriminator and the classifier consisted of stacks of convolutional layers. We used ReLU for nonlinearity. Tanh function is used as the activation function of the last layer in the decoder $F$ for scaling the output pixels to $[ - 1 , 1 ]$ . The details of the networks are given in Appendix D. ", + "bbox": [ + 148, + 415, + 852, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative evaluation involves a comparison of the performance of our model to previous work and to Source Only and 1-NN baselines that do not use any domain adaptation. For Source Only baseline, we train our model only on the unaltered source training data and evaluate on the target test data. We compare the proposed method MTDA-ITA with several related methods designed for pair-wise source-target adaptation: CORAL (Sun & Saenko (2016)), DANN (Ganin & Lempitsky (2015)), ADDA (Tzeng et al. (2017)), DTN (Zhang et al. (2015)), UNIT (Liu et al. (2017)), PixelDA (Bousmalis et al. (2017)), and DSN (Bousmalis et al. (2016)). We reported the results of two following baselines: (i) one is to combine all the target domains into a single one and train it using MTDA-ITA, which we denote as (c-MTDA-ITA). (ii) the other one is to train multiple MTDA-ITA separately, where each one corresponds to a source-target pair which we denote as (s-MTDA-ITA). For completeness, we reported the results of the competing methods by combining all the target domains into a single one (denoted by c-DTN, c-ADDA, and c-DSN) as well. We also extend DSN to multiple domains by (i) having one private encoder for all domains denoted by (1p-DSN), (ii) adding multiple private encoders to it denoted by (mp-DSN) and contrast them with our model. ", + "bbox": [ + 147, + 547, + 851, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 DIGITS DATASETS ", + "text_level": 1, + "bbox": [ + 148, + 744, + 313, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We combine four popular digits datasets (MNIST, MNIST-M, SVHN, and USPS) to build the multitarget domain dataset. All images were uniformly rescaled to $3 2 \\times 3 2$ . We take each of MNIST-M, ", + "bbox": [ + 150, + 770, + 851, + 799 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c17f5da419bdc0deb50b99f30b19766556f98c280148111ece40ad50497e48c6.jpg", + "image_caption": [ + "Figure 3: Exemplary images from different datasets. a) Digits datasets, b) PACS datatset (first row: Art-painting, second row: Cartoon, Third row: Photo, last row: Sketch), c) Multi-PIE dataset (each row corresponds to a different camera angle and each subject depicts an expression(normal, smile, surprise, squint, disgust, scream) at every camera position). " + ], + "image_footnote": [], + "bbox": [ + 215, + 90, + 777, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SVHN, USPS, and MNIST as source domain in turn, and the rest as targets. We use all labeled source images and all unlabeled target images, following the standard evaluation protocol for unsupervised domain adaptation (Ganin et al. (2016); Long et al. (2016). We show the accuracy of different methods in Table 1. Additional results are available in Appendix F. The results show that first of all cMTDA-ITA has worse performance than sMTDA-ITA and MTDA-ITA. We have similar observations for ADDA, DTN, and DSN that demonstrates a naive combination of different target datasets can sometimes even decrease the performance of the competing methods. Furthermore, MTDA-ITA outperforms the state-of-the-art methods in most of domain transformations. The higher performance of MTDA-ITA compared to other methods is mainly attributed to the joint adaptation of related domains where each domain could benefit of other related domains. Furthermore, from the results obtained, we see that it is beneficial to use information coming from unlabeled target data (see Eq. 8 for updating the classifier $C$ ) during the learning process, compared to when no data from target domain is used (See the ablation study section for more information). Indeed, using our scheme, we find a representation space in which embeds the knowledge from the target domain into the learned classifier. By contrast, the competing methods do not provide a principled way of sharing information across all domains, leading to overall lower performance. The results also verify the superiority of MTDA-ITA over both mp-DSN, and 1p-DSN. This can be due to (i) having multiple private encoders increase the number of parameters that may lead to mp-DSN overfitting, (ii) superiority of the MTDA-ITA’s domain adversarial loss over the DSN’s MMD loss to separate the shared and private features, (iii) utilization of the unlabeled target data to regularize the classifier in MTDA-ITA. ", + "bbox": [ + 147, + 313, + 851, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 MULTI-PIE DATASET ", + "text_level": 1, + "bbox": [ + 148, + 630, + 333, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Multi-PIE dataset includes face images of 337 individuals captured from different expressions, views, and illumination conditions (fig. 3(c)). For this experiment, we use 5 different camera views (positions) $C 0 5$ , $C 0 8$ , $C 0 9$ , $C 1 3$ , and $C 1 4$ as different domains (Fig. 3(c)) and the face expressions (normal, smile, surprise, squint, disgust, scream) as labels. Each domain contains 27120 images of size $6 4 \\times 6 4 \\times 3$ . We used each view as the source domain, in turn, and the rest as targets. We expect the face inclination angle to reflect the complexity of transfer learning. Tab. 2 shows the classification accuracy for $C 1 3$ and $C 1 4$ as source domain (the results for views $C 0 5$ , $C 0 8$ and $C 0 9$ as source domain are available in Appendix F). As can be seen, MTDA-ITA achieves the best performances as well as the best scores in most settings that verifies the effectiveness of MTDA-ITA for multi-target domain adaptation. Clearly, with the increasing camera angle, the image structure changes up to a certain extent (the views become heterogeneous). However, our method produces better results even under such very challenging conditions. ", + "bbox": [ + 147, + 669, + 852, + 821 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8fd442a94d88d4c3187f10668a4eaf7e8e6a87c4149cab928e7e3f5d35839d41.jpg", + "table_caption": [ + "Table 1: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown in red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDAITA where each one correspond to a source-target pair. 1p-DSN: extended DSN with single private encoder.mp-DSN: extended DSN with multiple private encoder. Last column shows the average rank of each method over all adaptation pairs. $\\bf \\Pi ^ { * } U N I T$ trains with the extended SVHN $> 5 0 0 \\mathrm { K }$ images vs ours 72K). \\*PixelDA uses $( \\approx 1 , 0 0 0 )$ of labeled target domain data as a validation set for tuning the hyper-parameters. " + ], + "table_footnote": [], + "table_body": "
methodS→MS→MMS→UM→SM→MMM→UAve.ranking
Source Only62.10±0.6040.43±0.7039.90± 0.6030.29± 0.5955.98 ± 0.4878.30±0.3814.00
1-NN35.8618.2129.3128.0112.5841.2215.00
CORAL (Sun & Saenko (2016))63.10 ± 0.6154.37 ± 0.5350.15 ± 0.6333.40 ±0.7457.70 ± 0.6981.05 ± 0.8011.33
DANN (Ganin et al. (2016))73.80 ±0.4961.05±0.8062.54 ± 0.9135.50 ± 0.6577.40±0.7381.60 ± 0.608.75
ADDA(Tzeng et al. (2017))77.68 ± 0.9264.23±0.7064.10±0.7930.04± 0.9891.47 ±1.090.51 ± 0.806.43
C-ADDA80.10 ± 0.6956.80 ± 0.7964.80±0.8827.50 ± 0.8683.30 ±0.9084.10 ±0.988.95
DTN (Zhang et al. (2015))81.40± 0.4263.70±0.3960.12 ± 0.5240.40±0.5085.70±0.3985.80 ± 0.466.04
c-DTN PixelDA (Bousmalis et al. (2017))82.10 ±0.6259.30 ± 0.5956.87 ± 0.6538.32 ± 0.5080.90±0.8079.31 ±0.787.96
UNIT (Liu et al. (2017))98.10*94.10*1
DSN (Bousmalis et al. (2016))90.6*92.901
c-DSN82.70±0.37 83.10 ±0.2064.80± 0.4065.30± 0.2849.30±0.3083.20±0.3091.65 ± 0.402.85
1p-DSN81.00 ± 0.4760.56 ± 0.36 58.22 ± 0.6860.35 ± 0.5946.80 ± 0.4580.49±0.4088.21 ±0.384.84
mp-DSN83.40±0.3061.00±0.5058.06 ±0.48 58.10 ±0.6445.11 ± 0.3377.33 ± 0.52 79.30±0.5985.16 ± 0.634.90
s-MTDA-ITA82.90±0.1363.10±0.2863.54±0.3047.35 ± 0.4082.42± 0.1986.45 ± 0.715.33
c-MTDA-ITA79.20±0.2859.90 ± 0.3063.70±0.2649.60± 0.25 45.30 ±0.3077.12 ± 0.2289.21±0.282.88
MTDA-ITA84.60±0.2465.30 ± 0.1570.03 ±0.2052.01 ± 0.2185.50 ±0.1887.47 ± 0.25 94.20±0.204.25
1.16
", + "bbox": [ + 147, + 117, + 849, + 314 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5c32f244dced015904d1dfdfb8c3e5deb43fab55ecc31c215787a0f0641e50d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodC13→C05C13→C08C13→C09C13→C14C14→C05C14→C08C14→C09C14→C13Ave.ranking
Source Only50.79± 0.3345.90± 0.5040.04±0.4059.68± 0.2960.03±0.5536.80± 0.6140.11 ±0.5060.57± 0.3616.08
1-NN CORAL33.21 54.89 ± 0.5237.0134.4548.79 68.90± 0.3547.4428.2430.8644.8617.00
DANN57.86± 0.4148.90±0.48 50.30 ± 0.4340.30±0.53 45.30± 0.5070.68 ± 0.3559.98 ± 0.45 57.20± 0.4540.63±0.55 40.22±0.5540.80 ±0.53 40.77 ± 0.4565.11 ± 0.45 70.50± 0.5511.95 9.92
ADDA64.83±0.6963.20±0.4555.48 ±0.6574.25±0.5573.62 ± 0.7543.56±0.9538.68 ±0.9572.84±0.759.33
C-ADDA59.20±0.2530.70±0.6353.20±0.4068.33±0.3565.88±0.3830.60±0.6145.34±0.4864.30±0.4011.50
DTN63.78±0.2960.45±0.3560.55±0.3572.60±0.2570.67±0.3041.55±0.6541.45 ± 0.4570.67±0.45
c-DTN57.53 ± 0.4255.24 ± 0.4557.14 ± 0.3965.16 ± 0.3563.80 ±0.4238.97 ± 0.7139.80 ± 0.6562.10 ± 0.458.75 10.92
PixelDA45.68± 0.5244.95± 0.4244.45±0.5590.50± 0.2546.28±0.6045.89±0.6144.45± 0.5169.15 ± 0.459.95
UNIT44.14±0.1044.47 ± 0.1144.21 ±0.1244.47 ± 0.1143.03 ±0.144.44 ±0.1544.47 ± 0.1544.47 ± 0.0511.07
DSN64.15± 0.3057.70±0.3849.15± 0.4580.75±0.2782.20±0.2838.75±0.5345.00± 0.2580.50± 0.35
c-DSN57.34±0.4531.63 ±0.6051.17 ±0.4074.52 ±0.3782.01 ± 0.3534.25±0.5842.63 ± 0.5579.42 ±0.355.15 8.20
1p-DSN55.84± 0.5030.03 ± 0.5049.06 ± 0.3872.11 ± 0.5081.22 ±0.4533.33 ± 0.5842.03 ±0.2478.78±0.578.63
mp-DSN55.20±0.4630.40±0.5047.80±0.3575.30 ± 0.2580.75±0.2030.20 ±0.5543.00±0.3579.02±0.408.88
s-MTDA-ITA70.10± 0.2758.90±0.2558.10±0.2780.12±0.1582.05±0.1845.90±0.3052.67±0.3081.60±0.24
c-MTDA-ITA60.34 ± 0.1755.67 ± 0.2157.10 ± 0.2373.50 ± 0.2076.80 ±0.1043.10 ±0.1248.10 ± 0.1480.90 ± 0.113.65
MTDA-ITA78.40 ± 0.266.70 ± 0.1770.30± 0.1485.49 ± 0.1187.20±0.1061.40±0.1460.05 ±0.1386.70 ± 0.105.01 1.20
", + "bbox": [ + 147, + 428, + 849, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Classification results on Multi-PIE dataset. Last column shows the average rank of each method over all adaptation pairs. The best is shown in red. ", + "bbox": [ + 145, + 631, + 849, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 ABLATION STUDIES ", + "text_level": 1, + "bbox": [ + 148, + 667, + 326, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We performed an ablation study on the proposed model measuring impact of various terms on the model’s performance. To this end, we conducted additional experiments for the digit datasets with different components ablation, i.e., training without the reconstruction loss (denoted as MTDA-woR) by setting $\\lambda _ { r } = 0$ , training without the classifier entropy loss (denoted as MTDA-woE) by setting $\\lambda _ { c } = 0$ , training without the multidomain separation loss (denoted as MTDA-woD) by setting $\\lambda _ { d } = 0$ . ", + "bbox": [ + 148, + 696, + 851, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As can be seen from fig. 4, disabling each of the above components leads to degraded performance. More precisely, the average drop by disabling the classifier entropy loss is $\\approx 3 . 5 \\%$ . Similarly, by disabling the reconstruction loss and the multi-domain separation loss, we have $\\approx 4 . 5 \\%$ and $\\approx 2 2 \\%$ average drop in performance, respectively. Clearly, by disabling the multi-domain separation loss, the accuracy drops significantly due to the severe data distribution mismatch between different domains. The figure also demonstrates that leveraging the unlabeled data from multiple target domains during training enhances the generalization ability of the model that leads to higher performance. In addition, the performance drop caused by removing the reconstruction loss , i.e., without the private encoder/decoder, indicates (i) the benefit of modeling the latent features as the combination of shared and private features, (ii) the ability of the model’s domain adversarial loss to effectively learn those features. ", + "bbox": [ + 148, + 767, + 849, + 823 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 148, + 121, + 851, + 203 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In order to examine the effect of the private features on the model’s classification performance, we took the MTDA-ITA and trained it without the private encoder (denoted as MTDA-woP). As fig. 4 shows, without the private features, the model performed consistently worse $\\approx 2 \\%$ average drop in performance) in all scenarios. This demonstrates explicitly modeling what is unique to each domain can improve the model’s ability to extract domain–invariant features. In summary, this ablation study showed that the individual components bring complimentary information to achieve the best classification results. ", + "bbox": [ + 148, + 204, + 851, + 286 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 FEATURE VISUALIZATION ", + "text_level": 1, + "bbox": [ + 150, + 314, + 369, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use t-SNE (Maaten & Hinton (2008)) on Digit dataset to visualize shared and private feature representations from different domains. Fig. 6 shows shared and private features from source (SVHN) and target domains before (a),(b) and after adaptation (c),(d). MTDA-ITA significantly reduces the domain mismatch for the shared features (circle markers in fig. 6d, strong mixing of domain labels in this cluster, fig. 6c) and increases it for the private features (triangle markers, pure and well-separated domain clusters in fig. 6c). This is partially due to the proposed multi-domain separation loss through the use of the domain classifier $D$ , which penalizes the domain mismatch for the shared features and rewards the mismatch for the private features. Moreover, as supported by the quantitative results in tab. 1, joint adaptation of related domains and the classifier, accomplished through the model, leads to superior class separability, compared to original features. This is depicted in fig. 6d, where the points in the shared space (large cluster) are grouped into class-specific subgroups (color indicates class label), while they are mixed in private spaces (smaller clusters). This is in contrast to fig. 6b, where original features show no class-specificity. ", + "bbox": [ + 148, + 343, + 851, + 511 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also show the learned shared and private features for the models MTDA-woE, MTDA-woP, MTDAwoR, and MTDA-woD, in figs. 6e to 6l. Note that since the private encoder $E _ { p }$ is disabled for MTDA-woR, and MTDA-woP, no private features are depicted in figs. $6 \\mathrm { g }$ to 6j. The class label separation in the shared space for MTDA-woE, MTDA-woP, and MTDA-woR, figs. 6f, 6h and 6j, is still evident but not as strong as in the full model, fig. 6d, corroborating the small loss in classification accuracy observed in fig. 4a. On the other hand, MTDA-woD has significant mixing of class labels in the shared space, fig. 6l, more so than MTDA-woE, MTDA-woR, and MTDA-woP, implying worse classification prediction in fig. 4a due to the severe mismatch between different domains. Since our model uses one private encoder for all target domains, we also contrasted the visualization of DSN model with one private encoder 1p-DSN in Appendix H. ", + "bbox": [ + 148, + 517, + 852, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 148, + 672, + 292, + 689 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This paper presented an information theoretic end-to-end approach to uDA in the context of a single source and multiple target domains that share a common task or properties. The proposed method learns feature representations invariant under multiple domain shifts and simultaneously discriminative for the learning task. This is accomplished by explicitly separating representations private to each domain and shared between source and target domains using a novel discrimination strategy. Our use of a single private domain encoder results in a highly scalable model, easily optimized using established back-propagation approaches. Results on three benchmark datasets for image classification show superiority of the proposed method compared to the state-of-the-art methods for unsupervised domain adaptation of visual domain categories. ", + "bbox": [ + 148, + 710, + 851, + 821 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 148, + 119, + 261, + 135 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "M Ehsan Abbasnejad, Anthony Dick, and Anton van den Hengel. Infinite variational autoencoder for semi-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 781–790. IEEE, 2017. ", + "bbox": [ + 145, + 142, + 854, + 185 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy computation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018. ", + "bbox": [ + 143, + 195, + 846, + 226 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexander Alemi, Ben Poole, Ian Fischer, Joshua Dillon, Rif A Saurous, and Kevin Murphy. Fixing a broken elbo. In International Conference on Machine Learning, pp. 159–168, 2018a. ", + "bbox": [ + 150, + 234, + 849, + 266 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. International Conference on Learning Representation (ICLR), 2016. ", + "bbox": [ + 148, + 275, + 849, + 305 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018b. ", + "bbox": [ + 145, + 315, + 851, + 344 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Asha Anoosheh, Eirikur Agustsson, Radu Timofte, and Luc Van Gool. Combogan: Unrestrained scalability for image domain translation. ICLR Workshops, 2017. ", + "bbox": [ + 145, + 354, + 851, + 385 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Barber and Felix Agakov. The im algorithm: a variational approach to information maximization. In Advances in Neural Information Processing Systems (NIPS), pp. 201–208. MIT Press, 2003. ", + "bbox": [ + 147, + 395, + 849, + 424 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sagie Benaim and Lior Wolf. One-sided unsupervised domain mapping. In Advances in Neural Information Processing Systems (NIPS), pp. 752–762, 2017. ", + "bbox": [ + 145, + 434, + 846, + 464 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, and Dumitru Erhan. Domain separation networks. In Advances in Neural Information Processing Systems (NIPS), pp. 343–351, 2016. ", + "bbox": [ + 148, + 473, + 849, + 503 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pp. 7, 2017. ", + "bbox": [ + 148, + 513, + 848, + 558 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio. Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349, 2015. ", + "bbox": [ + 143, + 566, + 849, + 597 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John S Bridle. Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In Neurocomputing, pp. 227–236. Springer, 1990. ", + "bbox": [ + 147, + 606, + 846, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Matthew Chalk, Olivier Marre, and Gasper Tkacik. Relevant sparse codes with variational information bottleneck. In Advances in Neural Information Processing Systems (NIPS), pp. 1957–1965, 2016. ", + "bbox": [ + 145, + 646, + 844, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1711, 2017. ", + "bbox": [ + 148, + 686, + 848, + 729 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation for domain adaptation. In Advances in Neural Information Processing Systems (NIPS), pp. 3733–3742, 2017. ", + "bbox": [ + 147, + 739, + 849, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. pp. 1–35. Springer, 2017. ", + "bbox": [ + 145, + 794, + 849, + 821 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Basura Fernando, Amaury Habrard, Marc Sebban, and Tinne Tuytelaars. Unsupervised visual domain adaptation using subspace alignment. In IEEE International Conference on Computer Vision (ICCV), pp. 2960–2967, 2013. ", + "bbox": [ + 147, + 119, + 852, + 162 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. International Conference on Machine Learning (ICML), 2015. ", + "bbox": [ + 147, + 171, + 848, + 200 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096–2030, 2016. ", + "bbox": [ + 148, + 208, + 852, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2066–2073, 2012. ", + "bbox": [ + 148, + 260, + 852, + 303 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Guang-Yuan Hao, Hong-Xing Yu, and Wei-Shi Zheng. Mixgan: Learning concepts from different domains for mixture generation. arXiv preprint arXiv:1807.01659, 2018. ", + "bbox": [ + 150, + 310, + 852, + 342 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hirokazu Kameoka, Takuhiro Kaneko, Kou Tanaka, and Nobukatsu Hojo. Stargan-vc: Non-parallel manyto-many voice conversion with star generative adversarial networks. arXiv preprint arXiv:1806.02169, 2018. ", + "bbox": [ + 147, + 348, + 854, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representation (ICLR), 2015. ", + "bbox": [ + 147, + 400, + 849, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Elyor Kodirov, Tao Xiang, Zhenyong Fu, and Shaogang Gong. Unsupervised domain adaptation for zero-shot learning. In IEEE International Conference on Computer Vision (ICCV), pp. 2452–2460, 2015. ", + "bbox": [ + 151, + 438, + 849, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ", + "bbox": [ + 148, + 474, + 851, + 505 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In IEEE International Conference on Computer Vision (ICCV), pp. 5543–5551. IEEE, 2017. ", + "bbox": [ + 147, + 512, + 852, + 555 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jianhua Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information theory, 37(1):145–151, 1991. ", + "bbox": [ + 145, + 564, + 851, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems (NIPS), pp. 469–477. Curran Associates, Inc., 2016. ", + "bbox": [ + 148, + 602, + 851, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Advances in Neural Information Processing Systems (NIPS), pp. 700–708, 2017. ", + "bbox": [ + 142, + 652, + 851, + 683 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1410–1417, 2014. ", + "bbox": [ + 145, + 690, + 849, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I Jordan. Learning transferable features with deep adaptation networks. International Conference on Machine Learning (ICML), 2015. ", + "bbox": [ + 143, + 742, + 851, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mingsheng Long, Jianmin Wang, Yue Cao, Jiaguang Sun, and S Yu Philip. Deep learning of transferable representation for scalable domain adaptation. IEEE Transactions on Knowledge and Data Engineering, 28(8):2027–2040, 2016. ", + "bbox": [ + 147, + 780, + 854, + 821 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(Nov):2579–2605, 2008. ", + "bbox": [ + 145, + 119, + 849, + 148 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Shakir Mohamed and Danilo Jimenez Rezende. Variational information maximisation for intrinsically motivated reinforcement learning. In Advances in neural information processing systems (NIPS), pp. 2125–2133, 2015. ", + "bbox": [ + 147, + 157, + 852, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Saeid Motiian, Quinn Jones, Seyed Iranmanesh, and Gianfranco Doretto. Few-shot adversarial domain adaptation. In Advances in Neural Information Processing Systems (NIPS), pp. 6673–6683, 2017. ", + "bbox": [ + 147, + 208, + 848, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, volume 2011, pp. 5, 2011. ", + "bbox": [ + 145, + 246, + 849, + 290 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Yuchen Pu, Weiyao Wang, Ricardo Henao, Liqun Chen, Zhe Gan, Chunyuan Li, and Lawrence Carin. Adversarial symmetric variational autoencoder. In Advances in Neural Information Processing Systems (NIPS), pp. 4330–4339, 2017. ", + "bbox": [ + 147, + 297, + 852, + 342 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In Advances in Neural Information Processing Systems (NIPS), pp. 506–516, 2017. ", + "bbox": [ + 148, + 349, + 851, + 380 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain adaptation. International Conference on Machine Learning (ICML), 2017. ", + "bbox": [ + 150, + 387, + 851, + 417 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces. In International Conference on Artificial Intelligence and Statistics (AISTATS), pp. 701–708, 2010. ", + "bbox": [ + 147, + 425, + 848, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European Conference on Computer Vision (ECCV), pp. 443–450. Springer, 2016. ", + "bbox": [ + 148, + 463, + 851, + 493 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. Thirtieth AAAI Conference on Artificial Intelligence, 2016. ", + "bbox": [ + 147, + 501, + 849, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In IEEE Information Theory Workshop (ITW), pp. 1–5. IEEE, 2015. ", + "bbox": [ + 143, + 539, + 851, + 569 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pp. 4, 2017. ", + "bbox": [ + 143, + 575, + 851, + 607 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. ", + "bbox": [ + 148, + 614, + 852, + 659 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Donggeun Yoo, Namil Kim, Sunggyun Park, Anthony S Paek, and In So Kweon. Pixel-level domain transfer. In European Conference on Computer Vision (ECCV), pp. 517–532, 2016. ", + "bbox": [ + 145, + 666, + 851, + 695 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschläger, and Susanne Saminger-Platz. Central moment discrepancy (cmd) for domain-invariant representation learning. International Conference on Learning Representation (ICLR), 2017. ", + "bbox": [ + 150, + 704, + 849, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jing Zhang, Wanqing Li, and Philip Ogunbona. Joint geometrical and statistical alignment for visual domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. ", + "bbox": [ + 143, + 755, + 849, + 785 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xu Zhang, Felix Xinnan Yu, Shih-Fu Chang, and Shengjin Wang. Deep transfer network: Unsupervised domain adaptation. arXiv preprint arXiv:1503.00591, 2015. ", + "bbox": [ + 147, + 794, + 848, + 823 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Han Zhao, Shanghang Zhang, Guanhang Wu, João P. Costeira, Jose Moura, and Geoffrey J. Gordon. Multiple source domain adaptation with adversarial training of neural networks. International Conference on Learning Representation (ICLR) workshops, 05 2017. ", + "bbox": [ + 148, + 119, + 852, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A CONNECTION TO INFORMATION THEORETIC REPRESENTATION LEARNING ", + "text_level": 1, + "bbox": [ + 148, + 191, + 802, + 208 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The idea of using information theoretic (IT) objectives for representation learning was originally introduced in (Tishby & Zaslavsky (2015)). Since their approach for optimizing the IT objective functions relied on the iterative Blahut Arimoto algorithm (Tishby & Zaslavsky (2015)), it is not feasible to apply to deep neural network (DNN) frameworks. Similar to our approach, there have been some recent works (Mohamed & Rezende (2015); Chalk et al. (2016); Alemi et al. (2018b; 2016; 2018a)) to approximate the MI by applying variational bounds on MI, though not in the context of domain adaptation. ", + "bbox": [ + 148, + 223, + 851, + 308 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Mohamed & Rezende (2015) utilized the variational bounds on MI, and apply it to DNNs in the context of reinforcement learning. Chalk et al. (2016) and Alemi et al. (2016), developed the same variational lower bound In the context of Information Bottleneck (IB) principle (Tishby & Zaslavsky (2015)), where the former applied it to sparse coding problems, and used the kernel trick to achieve nonlinear mappings, whereas the latter applied it to DNNs to handle large datasets thanks to the SGD algorithm. Achille & Soatto (2018) proposed a variational bound on the MI in the context of IB, from the perspective of variational dropout and demonstrated its utility in learning disentangled representations for variational autoencoders. ", + "bbox": [ + 148, + 314, + 851, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The main difference between our method and the above methods is that these methods throw away the information in the data not related to the task by minimizing the mutual information between the data points and the latent representations that may lead to ignoring the individual characteristics (private features) of the datasets in a multiple dataset regime, whereas our method explicitly models what is unique to each domain (dataset) that improves the model’s ability to extract domain–invariant features. ", + "bbox": [ + 148, + 419, + 849, + 488 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In the unsupervised representation learning literature, our work is also related to the VAE-based models (Bowman et al. (2015)). However, we propose to tackle the task using our IT approach using deterministic mappings instead of the traditional evidence lower bound (ELBO) optimization with stochastic mappings. In contrast to the unsupervised representation learning approaches, our setting also allows us to further improve the latent representation using the labeled data in the source domain while leveraging the sharing of dependencies across different target domains. ", + "bbox": [ + 148, + 496, + 851, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B CONNECTION TO MULTIPLE DOMAIN TRANSFER NETWORKS ", + "text_level": 1, + "bbox": [ + 148, + 601, + 691, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Recent studies have shown remarkable success in multiple domain transfer (MDT) (Choi et al. (2017); Anoosheh et al. (2017); Kameoka et al. (2018); Hao et al. (2018)) though not in the context of the image classification, rather in the context of image generation. Choi et al. (2017) proposed StarGAN, a generative adversarial network capable of learning mappings among multiple domains in the contest of image to image translation framework. The goal of StarGAN is to train a single generator $G$ though this requires passing in a vector along with each input to the generator specifying the output domain desired, that learns mappings among multiple domains. To achieve this, $G$ is trained to translate an input image $_ { \\textbf { \\em x } }$ into an output image $\\mathbf { x } ^ { \\prime }$ conditioned on the target domain label $^ d$ , $G ( \\pmb { x } , \\pmb { d } ) \\pmb { x } ^ { \\prime }$ . Similar to our domain classifier module $D$ , they introduce an auxiliary classifier that allows a single discriminator to control multiple domains. ", + "bbox": [ + 148, + 633, + 851, + 760 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Anoosheh et al. (2017) introduced ComboGAN, which decouples the domains and networks from each other. Similar to our encoder/decoder modules, ComboGAN’s generator networks contain encoder/decoders assigning each encoder and decoder to a domain. They combine the encoders and decoders of the trained model like building blocks, taking as input any domain and outputting any other. For example during inference, to transform an image $_ { \\textbf { \\em x } }$ from an arbitrary domain $\\mathbf { X }$ to $\\mathbf { x } ^ { \\prime }$ from domain $\\mathbf { X } ^ { \\prime }$ , they simply perform $\\pmb { x } ^ { \\prime } = G _ { \\mathbf { X } ^ { \\prime } , \\mathbf { X } } ( \\pmb { x } ) = D e c o d e r _ { X ^ { \\prime } } ( E n c o d e r _ { X } ( \\pmb { x } ) )$ . The result of $E n c o d e r _ { X } ( { \\pmb x } )$ can even be cached when translating to other domains as not to repeat computation. ", + "bbox": [ + 150, + 766, + 849, + 821 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 148, + 119, + 849, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The main differences between the MDT methods and ours is that, unlike our method which does domain alignment in feature space, MDT methods adapt representations not in feature space but rather in raw pixel space; translating samples from one domain to the “style” of a other domains. This works well for limited domain shifts where the domains are similar in pixel-space, but can be too limiting for settings with larger domain shifts that results in poor performance in significant structural change of the samples in different domains. ", + "bbox": [ + 148, + 170, + 851, + 252 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C CONNECTION TO DOMAIN SEPARATION NETWORKS ", + "text_level": 1, + "bbox": [ + 148, + 275, + 614, + 291 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The method closest to our work is Domain Separation Networks (DSN) (Bousmalis et al. (2016)), which use the notion of auto-encoders to explicitly separate the feature representations private to each source/target domain from those that are shared between the domains. Although extending DSN to multiple domains might seem trivial, DSN requires an autoencoder per domain, making the model impractical in the case of more than a couple of domains. ", + "bbox": [ + 148, + 308, + 851, + 377 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The overall loss of DSN consists of a reconstruction loss for each domain modeled by a shared decoder, a similarity loss such as MMD, which encourages domain invariance modeled by a shared encoder, and a dissimilarity loss modeled by two private encoders: one for the source domain and one for the target domain. While one could attempt to generalize DSN to multiple target domains by having individual per-target domain private encoders, doing so would prove problematic when the number of target domains is large — each private encoder would require a large \"private\" dataset to learn the private parameters. Precisely, for multiple $\\bar { ( \\cal M ) }$ target domains, we could train a DSN model with one shared encoder, $M + 1$ private encoder (one for each domain), and one shared decoder. This leads to $M + 3$ models to train that implies the number of models increases linearly with the number of domains, as does the required training time. Second, DSN uses an orthogonality constraint among the shared and the private representations which may not be strong enough to remove redundancy and enforce disentangling among different private spaces. Precisely, DSN defines the loss via a soft subspace orthogonality constraint between the private and shared representation of each domain. However, it does not enforce the private representation of different domains to be different that may result in redundancy of different private spaces. ", + "bbox": [ + 148, + 385, + 851, + 579 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In addition, DSN enforces separation of spaces using the notion of Euclidean orthogonality, e.g., $\\| z _ { s } - z _ { p } \\| ^ { 2 }$ . In case of multiple target domains, this would result in learning of all pairs of private spaces independently. To address those deficiencies, we first explicitly couple different private encoders into a single private encoder model, $E _ { \\theta _ { p } }$ of fig. 2 , which allows us to generalize to an arbitrary number of target domains. To assure that the information among the private and shared spaces is not shared (i.e., \"orthogonal\"), we define an informationtheoretic criteria enforced by a domain classifier, $\\mathcal { D } _ { \\psi }$ of fig. 2, which aims to segment the private space into clusters that correspond to individual target domains. By using $D _ { \\psi }$ within the adversarial framework, MTDA-ITA learns simultaneously the shared and private features from different domains (see fig. 6). We showed in Sec. 4 that our model performs better than the trivial extension of DSNs to the multi-domain case. ", + "bbox": [ + 148, + 585, + 852, + 712 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D NETWORK ARCHITECTURE ", + "text_level": 1, + "bbox": [ + 150, + 733, + 408, + 750 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The network architecture used for the experiments is given in tab. 3. We use the following abbreviation for ease of presentation: ${ \\bf N } =$ Neurons, $\\mathbf { K } { = }$ Kernel size, ${ \\mathrm { S } } { = } { }$ Stride size, $\\begin{array} { r } { \\boldsymbol { \\mathrm { D } } \\boldsymbol { = } } \\end{array}$ Number of Domains, $\\mathbf { C } =$ number of Classes. The transposed convolutional layer is denoted by DCONV. The residual basic block is denoted as RESBLK. ", + "bbox": [ + 148, + 766, + 851, + 821 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/50d3352cf70c1ded69dc8863026d49852f9c33a698de56302a1a3fb842a20ce7.jpg", + "table_caption": [ + "Table 3: Network architecture for the experiments. " + ], + "table_footnote": [], + "table_body": "
LayerEncoders (shared,private)
1CONV-(N16,K7,S1), ReLU
2CONV-(N32,K3,S2), ReLU
3CONV-(N64,K3,S2), ReLU
4RESBLK-(N64,K3,S1)
5RESBLK-(N64,K3,S1)
6RESBLK-(N64,K3,S1)
7RESBLK-(N64,K3,S1)
LayerDecoder
1RESBLK-(N64,K3,S1)
2RESBLK-(N64,K3,S1)
3RESBLK-(N64,K3,S1)
4RESBLK-(N64,K3,S1)
5DCONV-(N32,K3,S2), ReLU
6DCONV-(N16,K3,S2), ReLU
7DCONV-(N1,K1,S1), Tanh
LayerDiscriminator
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1),ReLU
3CONV-(N16,K3,S1),ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(ND), Softmax
LayerClassifier
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1), ReLU
3CONV-(N16,K3,S1), ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(NC), Softmax
", + "bbox": [ + 279, + 118, + 717, + 549 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "E PROPOSED MODEL’S ALGORITHM ", + "text_level": 1, + "bbox": [ + 148, + 580, + 464, + 597 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The detailed optimization process of the proposed model is shown in Algorithm 1. ", + "bbox": [ + 148, + 613, + 684, + 628 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "F ADDITIONAL EXPERIMENTS FOR DIGIT AND MULTI-PIE DATASETS ", + "text_level": 1, + "bbox": [ + 147, + 648, + 740, + 666 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The additional experiments for Digit dataset where we set MNIST-M and USPS as source domain is available in tab. 4. The additional experiments for Multi-PIE dataset where we set $C 0 5$ , $C 0 8$ and $C 0 9$ as source domain is available in tabs. $5 \\& 6$ . ", + "bbox": [ + 148, + 681, + 849, + 724 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "G PACS DATASET ", + "text_level": 1, + "bbox": [ + 148, + 747, + 313, + 763 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "This dataset contains 9991 images $2 2 7 \\times 2 2 7 \\times 3$ dimension) across 7 categories (‘dog’, ‘elephant’, ‘giraffe’,‘guitar’, ‘house’, ‘horse’ and ‘person’) and 4 domains of different stylistic depictions (‘Photo’, ‘Art painting’, ‘Cartoon’ and ‘Sketch’). The very diverse depiction styles provide a significant gap between domains, coupled with the small number of data samples, making it extremely challenging for domain adaptation. Consequently, the dataset was originally used for multi-source to single target domain adaptation (Li et al. (2017)). Instead, we tackle a significantly more challenging problem of single-source to multiple target adaptation. Tab. 7 shows the classification accuracy of various methods. MTDA-ITA consistently ", + "bbox": [ + 148, + 780, + 852, + 821 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/18facd59a83865bc6e26c98e1f1fdea84883eb0fbbf03a8444befb1f847d7fe4.jpg", + "image_caption": [ + "Figure 4: Ablation of MTDA-ITA on Digit dataset. We show that each component of our method, Reconstruction loss, Classifier entropy loss with separating shared/private features, contributes to the overall performance. " + ], + "image_footnote": [], + "bbox": [ + 233, + 102, + 759, + 402 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/ca7be10fae8c750a1749320d86fa67cb898aefc922aeca1ed95a61a028bdff54.jpg", + "table_caption": [ + "Table 4: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown in red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDA-ITA where each one correspond to a source-target pair. mp-DSN: extended DSN with multiple private encoder. $\\bf \\Pi ^ { * } U N I T$ trains with the extended SVHN $> 5 0 0 \\mathrm { K }$ images vs ours 72K). $^ *$ PixelDA uses $( \\approx 1 , 0 0 0 )$ of labeled target domain data as a validation set for tuning the hyper-parameters. " + ], + "table_footnote": [], + "table_body": "
methodMM→SMM→MMM→UU→SU→MU→MM
Source Only40.00±0.6184.46 ± 0.2980.43± 0.5023.41± 0.5250.64 ± 0.3741.45±0.38
1-NN21.4582.1336.9015.3438.4518.54
CORAL (Sun & Saenko (2016))40.20 ±0.6084.90 ± 0.7087.54 ± 0.4438.90 ± 0.9685.01 ± 0.6160.45 ± 0.70
DANN (Ganin et al. (2016))51.80 ± 0.9161.05 ± 0.7185.34 ± 0.6435.50± 0.8477.40 ± 0.6461.60 ± 0.64
ADDA (Tzeng et al. (2017))40.64 ± 0.8692.82 ± 0.4880.70±0.4841.23 ± 0.7890.10± 0.5856.21 ± 0.79
c-ADDA35.43 ± 0.9488.47 ± 0.6174.19 ± 0.5839.36 ± 0.9984.67 ± 0.9452.54 ± 0.88
DTN (Zhang et al. (2015)) c-DTN48.80± 0.6688.80±0.3890.68 ± 0.3542.43 ± 0.6189.04 ± 0.3655.78 ± 0.40
UNIT (Liu et al. (2017))44.21 ± 0.6183.60 ± 0.5484.98 ± 0.4139.75 ± 0.6485.04 ± 0.4548.86 ± 0.54
DSN (Bousmalis et al. (2016))1 51.50 ± 0.64/90.60
c-DSN47.10 ± 0.5090.20± 0.3189.95± 0.2948.20 ± 0.5991.40 ± 0.3060.45 ± 0.35
1p-DSN45.00 ± 0.6084.60 ± 0.4084.80 ±0.3940.50 ± 0.6186.05 ± 0.4656.25 ± 0.50
mp-DSN47.15 ± 0.6481.96 ± 0.6083.03 ±0.4939.30 ± 0.5184.55 ± 0.5655.03 ± 0.60
s-MTDA-ITA85.51 ± 0.5483.24± 0.2438.30 ± 0.7487.40 ± 0.3555.47 ± 0.44
c-MTDA-ITA50.55± 0.18 47.32±0.1994.82±0.2189.05 ±0.2840.13±0.3087.10 ± 0.2561.01 ± 0.24
MTDA-ITA53.50 ± 0.2290.20 ±0.30 98.20 ± 0.1090.01 ± 0.24 94.10 ± 0.1141.10 ± 0.35 46.00 ± 0.4885.35 ± 0.28 91.50 ± 0.2360.31± 0.34 67.30 ± 0.15
", + "bbox": [ + 147, + 462, + 848, + 672 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "", + "bbox": [ + 148, + 766, + 851, + 821 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/e5083e1d87737e6b3895ff8dbf3fa881fb2f3211b52e8393a6236ec40c8fe0f1.jpg", + "table_caption": [ + "Table 5: Classification results on Multi-PIE dataset. The best is shown in red. " + ], + "table_footnote": [], + "table_body": "
methodC05 →C08C05→C09C05→C13C05→C14
Source Only31.56 ± 0.4040.67 ± 0.3639.89 ± 0.2254.70± 0.25
1-NN CORAL (Sun & Saenko (2016))27.28 36.55 ± 0.6631.22 38.60 ± 0.6733.6647.04 55.29 ± 0.47
40.60 ± 0.58
DANN (citeganin2016domain)40.30 ± 0.6041.20 ± 0.6540.12 ± 0.6058.90 ± 0.38
ADDA (Tzeng et al. (2017)) C-ADDA33.21 ± 0.8130.86 ± 0.9052.44± 0.8070.18 ± 0.60
46.88 ± 0.6536.38 ± 0.8839.14 ± 0.8565.41 ± 0.69
DTN (Zhang et al. (2015))38.50 ± 0.5130.56 ± 0.4655.78 ± 0.3668.90 ± 0.31
c-DTN41.70 ± 0.4231.10 ± 0.4850.19 ± 0.4560.34 ± 0.35
PixelDA (Bousmalis et al. (2017))44.93 ±0.4244.75 ± 0.4545.18 ± 0.4546.88 ± 0.49
UNIT (Liu et al. (2017))44.47 ± 0.2144.47 ± 0.2144.47 ± 0.2044.51 ± 0.28
DSN (Bousmalis et al. (2016))45.12 ± 0.4644.35± 0.4948.12 ± 0.5375.00 ± 0.39
c-DSN42.52 ± 0.4838.54 ± 0.6434.15 ± 0.6469.45 ± 0.55
1p-DSN41.64 ± 0.5837.84 ± 0.6334.65 ± 0.4468.75 ± 0.85
mp-DSN41.30 ± 0.2835.14 ± 0.3534.40 ± 0.3565.70 ± 0.27
s-MTDA-ITA44.40 ± 0.2344.60 ± 0.2547.65 ± 0.2780.20 ± 0.13
c-MTDA-ITA40.49 ± 0.2540.70 ± 0.2542.80 ± 0.2571.60 ± 0.10
MTDA-ITA49.01 ± 0.20
48.20± 0.2753.13 ±0.2284.29 9±0.10
", + "bbox": [ + 147, + 116, + 848, + 404 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Algorithm 1 MDTA-ITA Algorithm ", + "text_level": 1, + "bbox": [ + 148, + 445, + 388, + 460 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Require: $\\{ \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } \\} { : \\mathrm { M } }$ domain datasets. \n$\\lambda _ { r } , \\lambda _ { c } , \\lambda _ { d }$ : Model hyper-parameters. \nEnsure: $\\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi$ : Model parameters. \n1: Initialize $\\bar { \\theta } _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi$ ; \n2: repeat 3: Sample a mini-batch from each of source/target domain datasets. \n4: Update $\\{ \\theta _ { s } \\}$ by minimizing $\\mathcal { L } _ { s }$ in Eq.(10) through the gradient descent: $\\begin{array} { r } { \\theta _ { s } = \\theta _ { s } - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\theta _ { s } } } \\end{array}$ . \n5: Update {θp} by minimizing Lp in Eq.(9) through the gradient descent:θp = θp − η ∂Lp∂θp . \n6: Update $\\{ \\theta _ { c } \\}$ by minimizing $\\mathcal { L } _ { c }$ in Eq.(8) through the gradient descent: $\\begin{array} { r } { \\theta _ { c } = \\theta _ { c } - \\eta \\frac { \\partial \\mathcal { L } _ { c } } { \\partial \\theta _ { c } } } \\end{array}$ 7: Update $\\{ \\phi \\}$ by minimizing $\\mathcal { L } _ { \\phi }$ in Eq.(6) through the gradient descent: $\\phi = \\phi - \\eta \\frac { \\partial \\mathcal { L } _ { \\phi } } { \\partial \\phi }$ . \n8: Update $\\{ \\psi \\}$ by minimizing $\\mathcal { L } _ { \\psi }$ in Eq.(5) through the gradient descent: $\\begin{array} { r } { \\psi = \\psi - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\psi } } \\end{array}$ . \n9: until Convergence; \n10: return $\\{ \\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi \\}$ . ", + "bbox": [ + 150, + 464, + 782, + 674 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "achieves the best performance for all transfer tasks. Evaluations were obtained by training all models (ADDA, DSN, and ours) from scratch on the PACS dataset. Note that the overall performance figures are low due to the extreme difficulty of the transfer task, induced by large differences among domains. ", + "bbox": [ + 148, + 684, + 852, + 727 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "H ANALYSIS OF SHARED/PRIVATE SPACE EMBEDDING ", + "text_level": 1, + "bbox": [ + 147, + 748, + 609, + 763 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In the experiments conducted, we showed that our approach is able to achieve better performance than the competing methods including the extended DSN with one private encoder (1p-DSN) which is the most similar method to ours. ", + "bbox": [ + 148, + 780, + 851, + 821 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/07d4618a501cd10a00bd625bd5df1054e718feb2cb44cd9581d0cca9092af8f0.jpg", + "image_caption": [ + "Figure 5: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features. (c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters in the right column represent points from the shared space, while the smaller ones are from the private spaces. (e),(f) learned features for 1p-DSN. " + ], + "image_footnote": [], + "bbox": [ + 168, + 114, + 769, + 582 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Indeed, fig. 5 depicts the embedding of the MTDA-ITA learned private/shared features, those of 1p-DSN and the original features from different domains for Digit datasets (SVHN is the source). ", + "bbox": [ + 147, + 675, + 849, + 704 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Notice that both MTDA-ITA and 1p-DSN reduces the domain mismatch for the shared features (circle markers in fig. 5) and separate the shared features from private features. On the other hand, MTDA-ITA increases the domain separation for the private features (triangle markers, pure and well-separated domain clusters in fig. 5c) while 1p-DSN is unable to enforce the private representation of different domains to be different (fig. 5e) that may result in redundancy of different private spaces. This is partially due to the proposed multi-domain separation loss through the use of the domain classifier $D$ , which penalizes the domain mismatch for the shared features and rewards the mismatch for the private features, something the 1p-DSN fails to account for. Moreover, as supported by the quantitative results in tab. 1, the class label separation in ", + "bbox": [ + 148, + 710, + 851, + 823 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/c85b4ef391c96ca0791a2ae020ce98ec467cb561c381b65c00be093640811879.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodC08→C05C08→C09C08→C13C08→C14C09→C05C09→C08C09→C13C09→C14
Source Only33.70±0.3350.10±0.2850.80±0.3240.13±0.2633.32±0.4448.24±0.3249.24±0.3036.19±0.27
1-NN28.7535.3939.7932.1326.8235.3034.2628.41
CORAL (Sun & Saenko (2016))35.89 ± 0.4455.79±0.5060.00 ±0.2940.67 ± 0.4835.89 ± 0.4051.56±0.4350.45 ± 0.4140.67± 0.35
DANN (Ganin et al. (2016))40.20±0.5056.89 ± 0.3955.83 ±0.4043.25±0.4150.63±0.3858.40 ± 0.5155.81 ± 0.5348.90 ±0.43
ADDA(Tzeng et al. (2017))37.40±0.6858.40±0.7360.40±0.8342.10±0.4829.40±0.7053.30±0.4945.30± 0.5338.30±0.63
c-ADDA41.60±0.6439.65 ±0.7050.00 ±0.5246.25±0.5245.01±0.6352.14 ± 0.5337.43 ± 0.6043.26±0.58
DTN (Zhang et al. (2015))44.13 ± 0.4157.42 ± 0.4255.89±0.4845.76± 0.3944.53± 0.4957.34±0.3552.43±0.3851.55 ± 0.40
c-DTN45.10 ±0.4449.78±0.5047.43 ±0.4645.79 ±0.4849.80±0.4055.69 ± 0.3550.10 ± 0.3852.31 ±0.29
PixelDA (Bousmalis et al. (2017))46.45 ± 0.4544.33±0.3844.87 ± 0.4146.83±0.2945.63±0.3416.37 ± 0.2745.43 ± 0.3547.00± 0.49
UNIT (Liu et al. (2017))43.88±0.1843.99±0.2344.47±0.1944.47 ± 0.2444.47 ± 0.1743.95± 0.2144.64±0.2244.47 ±0.19
DSN (Bousmalis et al. (2016)) c-DSN46.25± 0.5347.50±0.6062.15 ± 0.5839.72± 0.5545.85±0.4856.65±0.5056.5±0.3842.87±0.43
45.82 ± 0.5344.64± 0.4245.60 ± 0.4846.32±0.5245.18 ± 0.4745.52 ± 0.5544.79 ± 0.5347.37 ± 0.48
1p-DSN44.12 ± 0.7344.14 ± 0.2045.00 ±0.3845.62 ± 0.4244.78 ±0.4745.02 ±0.6544.21 ± 0.4846.97 ± 0.38
mp-DSN42.19±0.4644.70±0.5342.47 ± 0.4840.50±0.3945.00±0.5143.80±0.5045.79±0.4842.39 ±0.49
s-MTDA-ITA44.77±0.1945.61±0.1860.00±0.2746.70±0.2849.06±0.2455.33±0.2259.90±0.3050.64± 0.26
c-MTDA-ITA MTDA-ITA44.35± 0.2742.67 ± 0.2458.90 ±0.2644.32 ± 0.2646.74 ±0.2254.11 ± 0.2156.89 ± 0.2349.64 ± 0.19
46.30 ± 0.2560.60 ± 0.1860.50±0.1950.40±0.2055.59± 0.2557.80±0.2164.20 ±0.1856.34 ± 0.20
", + "bbox": [ + 147, + 92, + 848, + 268 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/5010897f659b60be96fdf8d14f7bef3429234f0e1fd20936d49e7805613ab995.jpg", + "table_caption": [ + "Table 6: Classification results on Multi-PIE dataset. The best (red). " + ], + "table_footnote": [], + "table_body": "
methodP→AP→CP→SA→PA→CA→S
1-NN15.2818.1625.6022.7019.7522.70
ADDA(Tzeng et al. (2017))24.35 ± 2.3720.12 ± 2.5022.45 ± 2.1132.57 ± 2.7017.68 ± 2.0418.90 ± 2.48
DSN (Bousmalis et al. (2016))28.42 ±2.1221.14 ± 2.082.04 ± 1.9029.54 ± 1.9525.89 ± 1.8824.69 ± 2.08
s-MTDA-ITA28.02 ± 1.5921.64±1.2426.24 ± 1.6031.06 ± 1.5025.09±1.4025.89±1.03
c-MTDA-ITA25.35 ± 1.8020.24 ± 1.3923.64 ± 1.6026.54 ± 1.3320.30 ± 1.2922.38 ± 1.45
MTDA-ITA31.40 ±1.5523.05 ±1.0428.24 ±1.7835.74 ± 1.5027.00 ±1.2528.90 ±1.60
", + "bbox": [ + 148, + 306, + 849, + 400 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Table 7: Classification results on PACS dataset classification. A:Art-painting, C:Cartoon, S:Sketch, P:Photo. The best (red). ", + "bbox": [ + 143, + 417, + 849, + 444 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "the shared space for 1p-DSN, fig. 5f, is still evident but not as strong as in the MTDA-ITA, fig. 5d. This can be attributed to the lack of redundancy in the private space that helps MTDA-ITA to learn more disentangled shared features and usage of the target samples during training, something the 1p-DSN fails to account for. ", + "bbox": [ + 143, + 452, + 851, + 494 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/f7a643eee7751386eefa467035b5bfae216f67ec9df6317cb9b21d6dd8832516.jpg", + "image_caption": [ + "Figure 6: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features. (c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters21 in the right column represent points from the shared space, while the smaller ones are from the private spaces. The remaining figures depict the learned features without: (e),(f) the classifier entropy loss, MTDA-woE; (g),(h) the private encoder, MTDA-woP; (i),(j) the reconstruction loss/decoder, MTDA-woR; and (k),(l) the multi-domain separation loss, MTDA-woD. " + ], + "image_footnote": [], + "bbox": [ + 263, + 102, + 730, + 814 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/parse/train/BJxLH2AcYX/BJxLH2AcYX_middle.json b/parse/train/BJxLH2AcYX/BJxLH2AcYX_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..a0d523864f5e281a13165a0b5f31dd0ea311127e --- /dev/null +++ b/parse/train/BJxLH2AcYX/BJxLH2AcYX_middle.json @@ -0,0 +1,47147 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 91, + 91, + 522, + 129 + ], + "lines": [ + { + "bbox": [ + 90, + 91, + 524, + 111 + ], + "spans": [ + { + "bbox": [ + 90, + 91, + 524, + 111 + ], + "score": 1.0, + "content": "UNSUPERVISED MULTI-TARGET DOMAIN ADAPTATION:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 111, + 403, + 131 + ], + "spans": [ + { + "bbox": [ + 89, + 111, + 403, + 131 + ], + "score": 1.0, + "content": "AN INFORMATION THEORETIC APPROACH", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 93, + 149, + 225, + 170 + ], + "lines": [ + { + "bbox": [ + 94, + 149, + 181, + 160 + ], + "spans": [ + { + "bbox": [ + 94, + 149, + 181, + 160 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 159, + 226, + 171 + ], + "spans": [ + { + "bbox": [ + 92, + 159, + 226, + 171 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 199, + 333, + 212 + ], + "lines": [ + { + "bbox": [ + 277, + 199, + 335, + 213 + ], + "spans": [ + { + "bbox": [ + 277, + 199, + 335, + 213 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 127, + 223, + 485, + 387 + ], + "lines": [ + { + "bbox": [ + 126, + 222, + 486, + 236 + ], + "spans": [ + { + "bbox": [ + 126, + 222, + 486, + 236 + ], + "score": 1.0, + "content": "Unsupervised domain adaptation (uDA) models focus on pairwise adaptation settings", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 234, + 486, + 247 + ], + "spans": [ + { + "bbox": [ + 124, + 234, + 486, + 247 + ], + "score": 1.0, + "content": "where there is a single, labeled, source and a single target domain. However, in many", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 245, + 486, + 257 + ], + "spans": [ + { + "bbox": [ + 126, + 245, + 486, + 257 + ], + "score": 1.0, + "content": "real-world settings one seeks to adapt to multiple, but somewhat similar, target domains.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 256, + 486, + 269 + ], + "spans": [ + { + "bbox": [ + 126, + 256, + 486, + 269 + ], + "score": 1.0, + "content": "Applying pairwise adaptation approaches to this setting may be suboptimal, as they fail", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 267, + 486, + 279 + ], + "spans": [ + { + "bbox": [ + 125, + 267, + 486, + 279 + ], + "score": 1.0, + "content": "to leverage shared information among multiple domains. In this work, we propose an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 277, + 486, + 290 + ], + "spans": [ + { + "bbox": [ + 125, + 277, + 486, + 290 + ], + "score": 1.0, + "content": "information theoretic approach for domain adaptation in the novel context of multiple target", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 289, + 486, + 301 + ], + "spans": [ + { + "bbox": [ + 126, + 289, + 486, + 301 + ], + "score": 1.0, + "content": "domains with unlabeled instances and one source domain with labeled instances. Our", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 300, + 485, + 312 + ], + "spans": [ + { + "bbox": [ + 126, + 300, + 485, + 312 + ], + "score": 1.0, + "content": "model aims to find a shared latent space common to all domains, while simultaneously", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 311, + 485, + 323 + ], + "spans": [ + { + "bbox": [ + 126, + 311, + 485, + 323 + ], + "score": 1.0, + "content": "accounting for the remaining private, domain-specific factors. Disentanglement of shared", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 322, + 486, + 334 + ], + "spans": [ + { + "bbox": [ + 126, + 322, + 486, + 334 + ], + "score": 1.0, + "content": "and private information is accomplished using a unified information-theoretic approach,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 333, + 486, + 345 + ], + "spans": [ + { + "bbox": [ + 126, + 333, + 486, + 345 + ], + "score": 1.0, + "content": "which also serves to establish a stronger link between the latent representations and the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 124, + 343, + 486, + 356 + ], + "spans": [ + { + "bbox": [ + 124, + 343, + 486, + 356 + ], + "score": 1.0, + "content": "observed data. The resulting model, accompanied by an efficient optimization algorithm,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 354, + 486, + 367 + ], + "spans": [ + { + "bbox": [ + 125, + 354, + 486, + 367 + ], + "score": 1.0, + "content": "allows simultaneous adaptation from a single source to multiple target domains. We test", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 366, + 486, + 379 + ], + "spans": [ + { + "bbox": [ + 125, + 366, + 486, + 379 + ], + "score": 1.0, + "content": "our approach on three challenging publicly-available datasets, showing that it outperforms", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 377, + 304, + 389 + ], + "spans": [ + { + "bbox": [ + 126, + 377, + 304, + 389 + ], + "score": 1.0, + "content": "several popular domain adaptation methods.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 92, + 406, + 190, + 419 + ], + "lines": [ + { + "bbox": [ + 89, + 405, + 192, + 422 + ], + "spans": [ + { + "bbox": [ + 89, + 405, + 192, + 422 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 91, + 430, + 521, + 529 + ], + "lines": [ + { + "bbox": [ + 90, + 431, + 521, + 443 + ], + "spans": [ + { + "bbox": [ + 90, + 431, + 521, + 443 + ], + "score": 1.0, + "content": "In real-world data, the training and test instances often do not come from the same underlying distribution", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 441, + 522, + 456 + ], + "spans": [ + { + "bbox": [ + 88, + 441, + 522, + 456 + ], + "score": 1.0, + "content": "(Sun et al. (2016)). For example, in the task of object recognition/classification from image data, this may", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 453, + 522, + 465 + ], + "spans": [ + { + "bbox": [ + 90, + 453, + 522, + 465 + ], + "score": 1.0, + "content": "be due to the image noise, changes in the object view, etc., which induce different biases in the observed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 88, + 462, + 522, + 478 + ], + "spans": [ + { + "bbox": [ + 88, + 462, + 522, + 478 + ], + "score": 1.0, + "content": "data sampled during the training and test stage. Consequently, assumptions made by traditional learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 475, + 522, + 488 + ], + "spans": [ + { + "bbox": [ + 89, + 475, + 522, + 488 + ], + "score": 1.0, + "content": "algorithms are often violated, resulting in degradation of the algorithms’ performance during inference of test", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 484, + 523, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 484, + 523, + 499 + ], + "score": 1.0, + "content": "data. Domain Adaptation (DA) approaches (Fernando et al. (2013); Gong et al. (2012); Kodirov et al. (2015);", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 495, + 522, + 510 + ], + "spans": [ + { + "bbox": [ + 89, + 495, + 522, + 510 + ], + "score": 1.0, + "content": "Yoo et al. (2016)) aim to tackle this by transferring knowledge from a source domain (training data) to an", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 506, + 524, + 522 + ], + "spans": [ + { + "bbox": [ + 88, + 506, + 524, + 522 + ], + "score": 1.0, + "content": "unlabeled target domain (test data) to reduce the discrepancy between the source and target data distributions,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 519, + 314, + 532 + ], + "spans": [ + { + "bbox": [ + 89, + 519, + 314, + 532 + ], + "score": 1.0, + "content": "typically by exploring domain-invariant data structures.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 91, + 535, + 520, + 602 + ], + "lines": [ + { + "bbox": [ + 90, + 535, + 521, + 546 + ], + "spans": [ + { + "bbox": [ + 90, + 535, + 521, + 546 + ], + "score": 1.0, + "content": "Existing DA methods tackle the adaptation problems in one of the two settings: (semi)supervised DA and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 88, + 546, + 522, + 559 + ], + "spans": [ + { + "bbox": [ + 88, + 546, + 522, + 559 + ], + "score": 1.0, + "content": "unsupervised DA (Csurka (2017)). The former assume that in addition to the labeled data of the source", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 90, + 558, + 522, + 569 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 522, + 569 + ], + "score": 1.0, + "content": "domain, some labeled data from the target domain are also available for training/adapting the classifiers. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 90, + 569, + 522, + 581 + ], + "spans": [ + { + "bbox": [ + 90, + 569, + 522, + 581 + ], + "score": 1.0, + "content": "contrast, the latter does not require any labels from the target domain but rather explores the similarity in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 579, + 522, + 592 + ], + "spans": [ + { + "bbox": [ + 89, + 579, + 522, + 592 + ], + "score": 1.0, + "content": "data distributions of the two domains. In this work, we focus on the unsupervised DA (uDA) scenario, which", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 88, + 590, + 421, + 603 + ], + "spans": [ + { + "bbox": [ + 88, + 590, + 421, + 603 + ], + "score": 1.0, + "content": "is more challenging due to the lack of correspondences in source and target labels.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 520, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 607, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 607, + 522, + 621 + ], + "score": 1.0, + "content": "Most works on uDA today focus on a single-source-single-target-domain scenario. However, in many", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 617, + 520, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 617, + 520, + 630 + ], + "score": 1.0, + "content": "real-world applications, unlabeled data may come from different domains, thus, with different statistical", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 88, + 629, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 88, + 629, + 522, + 642 + ], + "score": 1.0, + "content": "properties but with common task-related content. For instance, we may have access to images of the same", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 90, + 639, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 90, + 639, + 522, + 653 + ], + "score": 1.0, + "content": "class of objects (e.g., cars) recorded by various types of cameras, and/or under different camera views and", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 291, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 671, + 293, + 682 + ], + "spans": [ + { + "bbox": [ + 285, + 671, + 293, + 682 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 91, + 91, + 522, + 129 + ], + "lines": [ + { + "bbox": [ + 90, + 91, + 524, + 111 + ], + "spans": [ + { + "bbox": [ + 90, + 91, + 524, + 111 + ], + "score": 1.0, + "content": "UNSUPERVISED MULTI-TARGET DOMAIN ADAPTATION:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 111, + 403, + 131 + ], + "spans": [ + { + "bbox": [ + 89, + 111, + 403, + 131 + ], + "score": 1.0, + "content": "AN INFORMATION THEORETIC APPROACH", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 93, + 149, + 225, + 170 + ], + "lines": [ + { + "bbox": [ + 94, + 149, + 181, + 160 + ], + "spans": [ + { + "bbox": [ + 94, + 149, + 181, + 160 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 159, + 226, + 171 + ], + "spans": [ + { + "bbox": [ + 92, + 159, + 226, + 171 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 92, + 149, + 226, + 171 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 199, + 333, + 212 + ], + "lines": [ + { + "bbox": [ + 277, + 199, + 335, + 213 + ], + "spans": [ + { + "bbox": [ + 277, + 199, + 335, + 213 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 127, + 223, + 485, + 387 + ], + "lines": [ + { + "bbox": [ + 126, + 222, + 486, + 236 + ], + "spans": [ + { + "bbox": [ + 126, + 222, + 486, + 236 + ], + "score": 1.0, + "content": "Unsupervised domain adaptation (uDA) models focus on pairwise adaptation settings", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 234, + 486, + 247 + ], + "spans": [ + { + "bbox": [ + 124, + 234, + 486, + 247 + ], + "score": 1.0, + "content": "where there is a single, labeled, source and a single target domain. However, in many", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 245, + 486, + 257 + ], + "spans": [ + { + "bbox": [ + 126, + 245, + 486, + 257 + ], + "score": 1.0, + "content": "real-world settings one seeks to adapt to multiple, but somewhat similar, target domains.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 256, + 486, + 269 + ], + "spans": [ + { + "bbox": [ + 126, + 256, + 486, + 269 + ], + "score": 1.0, + "content": "Applying pairwise adaptation approaches to this setting may be suboptimal, as they fail", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 267, + 486, + 279 + ], + "spans": [ + { + "bbox": [ + 125, + 267, + 486, + 279 + ], + "score": 1.0, + "content": "to leverage shared information among multiple domains. In this work, we propose an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 277, + 486, + 290 + ], + "spans": [ + { + "bbox": [ + 125, + 277, + 486, + 290 + ], + "score": 1.0, + "content": "information theoretic approach for domain adaptation in the novel context of multiple target", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 289, + 486, + 301 + ], + "spans": [ + { + "bbox": [ + 126, + 289, + 486, + 301 + ], + "score": 1.0, + "content": "domains with unlabeled instances and one source domain with labeled instances. Our", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 300, + 485, + 312 + ], + "spans": [ + { + "bbox": [ + 126, + 300, + 485, + 312 + ], + "score": 1.0, + "content": "model aims to find a shared latent space common to all domains, while simultaneously", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 311, + 485, + 323 + ], + "spans": [ + { + "bbox": [ + 126, + 311, + 485, + 323 + ], + "score": 1.0, + "content": "accounting for the remaining private, domain-specific factors. Disentanglement of shared", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 322, + 486, + 334 + ], + "spans": [ + { + "bbox": [ + 126, + 322, + 486, + 334 + ], + "score": 1.0, + "content": "and private information is accomplished using a unified information-theoretic approach,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 333, + 486, + 345 + ], + "spans": [ + { + "bbox": [ + 126, + 333, + 486, + 345 + ], + "score": 1.0, + "content": "which also serves to establish a stronger link between the latent representations and the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 124, + 343, + 486, + 356 + ], + "spans": [ + { + "bbox": [ + 124, + 343, + 486, + 356 + ], + "score": 1.0, + "content": "observed data. The resulting model, accompanied by an efficient optimization algorithm,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 354, + 486, + 367 + ], + "spans": [ + { + "bbox": [ + 125, + 354, + 486, + 367 + ], + "score": 1.0, + "content": "allows simultaneous adaptation from a single source to multiple target domains. We test", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 366, + 486, + 379 + ], + "spans": [ + { + "bbox": [ + 125, + 366, + 486, + 379 + ], + "score": 1.0, + "content": "our approach on three challenging publicly-available datasets, showing that it outperforms", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 377, + 304, + 389 + ], + "spans": [ + { + "bbox": [ + 126, + 377, + 304, + 389 + ], + "score": 1.0, + "content": "several popular domain adaptation methods.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12, + "bbox_fs": [ + 124, + 222, + 486, + 389 + ] + }, + { + "type": "title", + "bbox": [ + 92, + 406, + 190, + 419 + ], + "lines": [ + { + "bbox": [ + 89, + 405, + 192, + 422 + ], + "spans": [ + { + "bbox": [ + 89, + 405, + 192, + 422 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 91, + 430, + 521, + 529 + ], + "lines": [ + { + "bbox": [ + 90, + 431, + 521, + 443 + ], + "spans": [ + { + "bbox": [ + 90, + 431, + 521, + 443 + ], + "score": 1.0, + "content": "In real-world data, the training and test instances often do not come from the same underlying distribution", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 441, + 522, + 456 + ], + "spans": [ + { + "bbox": [ + 88, + 441, + 522, + 456 + ], + "score": 1.0, + "content": "(Sun et al. (2016)). For example, in the task of object recognition/classification from image data, this may", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 453, + 522, + 465 + ], + "spans": [ + { + "bbox": [ + 90, + 453, + 522, + 465 + ], + "score": 1.0, + "content": "be due to the image noise, changes in the object view, etc., which induce different biases in the observed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 88, + 462, + 522, + 478 + ], + "spans": [ + { + "bbox": [ + 88, + 462, + 522, + 478 + ], + "score": 1.0, + "content": "data sampled during the training and test stage. Consequently, assumptions made by traditional learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 475, + 522, + 488 + ], + "spans": [ + { + "bbox": [ + 89, + 475, + 522, + 488 + ], + "score": 1.0, + "content": "algorithms are often violated, resulting in degradation of the algorithms’ performance during inference of test", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 484, + 523, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 484, + 523, + 499 + ], + "score": 1.0, + "content": "data. Domain Adaptation (DA) approaches (Fernando et al. (2013); Gong et al. (2012); Kodirov et al. (2015);", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 495, + 522, + 510 + ], + "spans": [ + { + "bbox": [ + 89, + 495, + 522, + 510 + ], + "score": 1.0, + "content": "Yoo et al. (2016)) aim to tackle this by transferring knowledge from a source domain (training data) to an", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 506, + 524, + 522 + ], + "spans": [ + { + "bbox": [ + 88, + 506, + 524, + 522 + ], + "score": 1.0, + "content": "unlabeled target domain (test data) to reduce the discrepancy between the source and target data distributions,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 519, + 314, + 532 + ], + "spans": [ + { + "bbox": [ + 89, + 519, + 314, + 532 + ], + "score": 1.0, + "content": "typically by exploring domain-invariant data structures.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25, + "bbox_fs": [ + 88, + 431, + 524, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 535, + 520, + 602 + ], + "lines": [ + { + "bbox": [ + 90, + 535, + 521, + 546 + ], + "spans": [ + { + "bbox": [ + 90, + 535, + 521, + 546 + ], + "score": 1.0, + "content": "Existing DA methods tackle the adaptation problems in one of the two settings: (semi)supervised DA and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 88, + 546, + 522, + 559 + ], + "spans": [ + { + "bbox": [ + 88, + 546, + 522, + 559 + ], + "score": 1.0, + "content": "unsupervised DA (Csurka (2017)). The former assume that in addition to the labeled data of the source", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 90, + 558, + 522, + 569 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 522, + 569 + ], + "score": 1.0, + "content": "domain, some labeled data from the target domain are also available for training/adapting the classifiers. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 90, + 569, + 522, + 581 + ], + "spans": [ + { + "bbox": [ + 90, + 569, + 522, + 581 + ], + "score": 1.0, + "content": "contrast, the latter does not require any labels from the target domain but rather explores the similarity in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 579, + 522, + 592 + ], + "spans": [ + { + "bbox": [ + 89, + 579, + 522, + 592 + ], + "score": 1.0, + "content": "data distributions of the two domains. In this work, we focus on the unsupervised DA (uDA) scenario, which", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 88, + 590, + 421, + 603 + ], + "spans": [ + { + "bbox": [ + 88, + 590, + 421, + 603 + ], + "score": 1.0, + "content": "is more challenging due to the lack of correspondences in source and target labels.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 88, + 535, + 522, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 520, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 607, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 607, + 522, + 621 + ], + "score": 1.0, + "content": "Most works on uDA today focus on a single-source-single-target-domain scenario. However, in many", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 617, + 520, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 617, + 520, + 630 + ], + "score": 1.0, + "content": "real-world applications, unlabeled data may come from different domains, thus, with different statistical", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 88, + 629, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 88, + 629, + 522, + 642 + ], + "score": 1.0, + "content": "properties but with common task-related content. For instance, we may have access to images of the same", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 90, + 639, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 90, + 639, + 522, + 653 + ], + "score": 1.0, + "content": "class of objects (e.g., cars) recorded by various types of cameras, and/or under different camera views and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 262, + 522, + 274 + ], + "spans": [ + { + "bbox": [ + 89, + 262, + 522, + 274 + ], + "score": 1.0, + "content": "at different times, rendering multiple different domains (e.g., datasets). Likewise, facial expressions of", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 273, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 273, + 522, + 285 + ], + "score": 1.0, + "content": "emotions, such as joy and surprise, shown by different people and recorded under different views, result", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 282, + 522, + 298 + ], + "spans": [ + { + "bbox": [ + 88, + 282, + 522, + 298 + ], + "score": 1.0, + "content": "in multiple domains with varying data distributions. In most cases, these domains have similar underlying", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 295, + 522, + 307 + ], + "spans": [ + { + "bbox": [ + 89, + 295, + 522, + 307 + ], + "score": 1.0, + "content": "data distributions, which can be leveraged to build more effective and robust classifiers for tasks such as", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 306, + 522, + 317 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 522, + 317 + ], + "score": 1.0, + "content": "the object or emotion recognition across multiple datasets/domains. To this end, traditional uDA methods", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 315, + 523, + 331 + ], + "spans": [ + { + "bbox": [ + 88, + 315, + 523, + 331 + ], + "score": 1.0, + "content": "focus on the single-source-single-target DA scenario. However, in the presence of multiple domains, as", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 327, + 523, + 341 + ], + "spans": [ + { + "bbox": [ + 89, + 327, + 523, + 341 + ], + "score": 1.0, + "content": "typically encountered in real-world settings, this pair-wise adaptation approach may be suboptimal as it", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 337, + 522, + 351 + ], + "spans": [ + { + "bbox": [ + 88, + 337, + 522, + 351 + ], + "score": 1.0, + "content": "fails to leverage simultaneously the knowledge shared across multiple task-related domains. Recently, Zhao", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "score": 1.0, + "content": "et al. (2017) showed that by having access to multiple source domains can facilitate better adaptation to", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 361, + 522, + 372 + ], + "spans": [ + { + "bbox": [ + 89, + 361, + 522, + 372 + ], + "score": 1.0, + "content": "a single target domain, when compared to the pair-wise DA approach. While this is intuitive due to the", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 371, + 522, + 383 + ], + "spans": [ + { + "bbox": [ + 89, + 371, + 522, + 383 + ], + "score": 1.0, + "content": "access to multiple labelled source domains, offering more adaptation flexibility for the target domain (i.e.,", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 381, + 522, + 395 + ], + "spans": [ + { + "bbox": [ + 89, + 381, + 522, + 395 + ], + "score": 1.0, + "content": "by efficiently exploring the data labels across multiple source domains that are most related to the target", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 393, + 522, + 406 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 522, + 406 + ], + "score": 1.0, + "content": "domain), it comes at the expense of the data labelling in multiple source domains, which can be costly and", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 403, + 522, + 417 + ], + "spans": [ + { + "bbox": [ + 88, + 403, + 522, + 417 + ], + "score": 1.0, + "content": "time-consuming. In either case, a single source domain or readily available multiple source domains, to", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "spans": [ + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "score": 1.0, + "content": "the best of our knowledge, a simultaneous adaptation to multiple and unlabelled target domains remains", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 426, + 522, + 439 + ], + "spans": [ + { + "bbox": [ + 88, + 426, + 522, + 439 + ], + "score": 1.0, + "content": "an unexplored DA scenario. However, this DA scenario is important as we usually have access to multiple", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 437, + 522, + 450 + ], + "spans": [ + { + "bbox": [ + 89, + 437, + 522, + 450 + ], + "score": 1.0, + "content": "unlabeled domains; yet, the adaptation process is also more challenging due to the lack of supervision in the", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 448, + 522, + 461 + ], + "spans": [ + { + "bbox": [ + 89, + 448, + 522, + 461 + ], + "score": 1.0, + "content": "target domains. Still, multi-target DA can have advantages over a single-target DA when: (i) there is direct", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "spans": [ + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "score": 1.0, + "content": "knowledge sharing between the source and multiple target domains (fig. 1a), and (ii) the source and a target", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "spans": [ + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "score": 1.0, + "content": "domain are related through another target domain (fig. 1b). While this seems intuitive, it is critical how the", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 479, + 522, + 494 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 522, + 494 + ], + "score": 1.0, + "content": "data from multiple unlabelled target domains are leveraged within the multi-target DA approach, in order to", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 492, + 519, + 505 + ], + "spans": [ + { + "bbox": [ + 89, + 492, + 519, + 505 + ], + "score": 1.0, + "content": "improve its performance over the single target DA approaches and naive fusion of multiple target domains.", + "type": "text", + "cross_page": true + } + ], + "index": 27 + } + ], + "index": 37.5, + "bbox_fs": [ + 88, + 607, + 522, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 84, + 480, + 209 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 84, + 480, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 84, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 129, + 84, + 480, + 209 + ], + "score": 0.971, + "type": "image", + "image_path": "36886e0179e0304a2140f74c7b598abbe23c7d9c3021a983addc10b904e84af3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 84, + 480, + 125.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 125.66666666666666, + 480, + 167.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 167.33333333333331, + 480, + 208.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 218, + 521, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 218, + 521, + 230 + ], + "spans": [ + { + "bbox": [ + 90, + 218, + 521, + 230 + ], + "score": 1.0, + "content": "Figure 1: Illustration of domains with common (a) and pairwise-shared spaces (b). We tackle the domain adaptation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 227, + 522, + 241 + ], + "spans": [ + { + "bbox": [ + 89, + 227, + 522, + 241 + ], + "score": 1.0, + "content": "task when all domains share a common task/space, which is then leveraged to transfer knowledge across multiple target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 239, + 125, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 239, + 125, + 249 + ], + "score": 1.0, + "content": "domains.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 90, + 261, + 521, + 503 + ], + "lines": [ + { + "bbox": [ + 89, + 262, + 522, + 274 + ], + "spans": [ + { + "bbox": [ + 89, + 262, + 522, + 274 + ], + "score": 1.0, + "content": "at different times, rendering multiple different domains (e.g., datasets). Likewise, facial expressions of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 273, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 273, + 522, + 285 + ], + "score": 1.0, + "content": "emotions, such as joy and surprise, shown by different people and recorded under different views, result", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 282, + 522, + 298 + ], + "spans": [ + { + "bbox": [ + 88, + 282, + 522, + 298 + ], + "score": 1.0, + "content": "in multiple domains with varying data distributions. In most cases, these domains have similar underlying", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 295, + 522, + 307 + ], + "spans": [ + { + "bbox": [ + 89, + 295, + 522, + 307 + ], + "score": 1.0, + "content": "data distributions, which can be leveraged to build more effective and robust classifiers for tasks such as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 306, + 522, + 317 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 522, + 317 + ], + "score": 1.0, + "content": "the object or emotion recognition across multiple datasets/domains. To this end, traditional uDA methods", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 315, + 523, + 331 + ], + "spans": [ + { + "bbox": [ + 88, + 315, + 523, + 331 + ], + "score": 1.0, + "content": "focus on the single-source-single-target DA scenario. However, in the presence of multiple domains, as", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 327, + 523, + 341 + ], + "spans": [ + { + "bbox": [ + 89, + 327, + 523, + 341 + ], + "score": 1.0, + "content": "typically encountered in real-world settings, this pair-wise adaptation approach may be suboptimal as it", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 337, + 522, + 351 + ], + "spans": [ + { + "bbox": [ + 88, + 337, + 522, + 351 + ], + "score": 1.0, + "content": "fails to leverage simultaneously the knowledge shared across multiple task-related domains. Recently, Zhao", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "score": 1.0, + "content": "et al. (2017) showed that by having access to multiple source domains can facilitate better adaptation to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 361, + 522, + 372 + ], + "spans": [ + { + "bbox": [ + 89, + 361, + 522, + 372 + ], + "score": 1.0, + "content": "a single target domain, when compared to the pair-wise DA approach. While this is intuitive due to the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 371, + 522, + 383 + ], + "spans": [ + { + "bbox": [ + 89, + 371, + 522, + 383 + ], + "score": 1.0, + "content": "access to multiple labelled source domains, offering more adaptation flexibility for the target domain (i.e.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 381, + 522, + 395 + ], + "spans": [ + { + "bbox": [ + 89, + 381, + 522, + 395 + ], + "score": 1.0, + "content": "by efficiently exploring the data labels across multiple source domains that are most related to the target", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 393, + 522, + 406 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 522, + 406 + ], + "score": 1.0, + "content": "domain), it comes at the expense of the data labelling in multiple source domains, which can be costly and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 403, + 522, + 417 + ], + "spans": [ + { + "bbox": [ + 88, + 403, + 522, + 417 + ], + "score": 1.0, + "content": "time-consuming. In either case, a single source domain or readily available multiple source domains, to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "spans": [ + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "score": 1.0, + "content": "the best of our knowledge, a simultaneous adaptation to multiple and unlabelled target domains remains", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 426, + 522, + 439 + ], + "spans": [ + { + "bbox": [ + 88, + 426, + 522, + 439 + ], + "score": 1.0, + "content": "an unexplored DA scenario. However, this DA scenario is important as we usually have access to multiple", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 437, + 522, + 450 + ], + "spans": [ + { + "bbox": [ + 89, + 437, + 522, + 450 + ], + "score": 1.0, + "content": "unlabeled domains; yet, the adaptation process is also more challenging due to the lack of supervision in the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 448, + 522, + 461 + ], + "spans": [ + { + "bbox": [ + 89, + 448, + 522, + 461 + ], + "score": 1.0, + "content": "target domains. Still, multi-target DA can have advantages over a single-target DA when: (i) there is direct", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "spans": [ + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "score": 1.0, + "content": "knowledge sharing between the source and multiple target domains (fig. 1a), and (ii) the source and a target", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "spans": [ + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "score": 1.0, + "content": "domain are related through another target domain (fig. 1b). While this seems intuitive, it is critical how the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 479, + 522, + 494 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 522, + 494 + ], + "score": 1.0, + "content": "data from multiple unlabelled target domains are leveraged within the multi-target DA approach, in order to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 492, + 519, + 505 + ], + "spans": [ + { + "bbox": [ + 89, + 492, + 519, + 505 + ], + "score": 1.0, + "content": "improve its performance over the single target DA approaches and naive fusion of multiple target domains.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 91, + 508, + 522, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 509, + 523, + 522 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 523, + 522 + ], + "score": 1.0, + "content": "To this end, we propose a Multi-Target DA-Information-Theoretic-Approach (MTDA-ITA) for single-source-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 519, + 522, + 533 + ], + "spans": [ + { + "bbox": [ + 89, + 519, + 522, + 533 + ], + "score": 1.0, + "content": "multi-target DA. We exploit a single source domain and focus on multiple target domains to investigate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 529, + 522, + 544 + ], + "spans": [ + { + "bbox": [ + 89, + 529, + 522, + 544 + ], + "score": 1.0, + "content": "the effects of multi-target DA; however, the proposed approach can easily be extended to multiple source", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "score": 1.0, + "content": "domains. This approach leverages the data from multiple target domains to improve performance compared", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 552, + 522, + 565 + ], + "spans": [ + { + "bbox": [ + 89, + 552, + 522, + 565 + ], + "score": 1.0, + "content": "to individually learning from pair-wise source-target domains. Specifically, we simultaneously factorize the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 563, + 523, + 577 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 523, + 577 + ], + "score": 1.0, + "content": "information from each available target domain and learn separate subspaces for modeling the shared (i.e.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 575, + 522, + 586 + ], + "spans": [ + { + "bbox": [ + 89, + 575, + 522, + 586 + ], + "score": 1.0, + "content": "correlated across the domains) and private (i.e., independent between the domains) subspaces of the data", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 586, + 522, + 599 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 522, + 599 + ], + "score": 1.0, + "content": "(Salzmann et al. (2010)). To this end, we employ deep learning to derive an information theoretic approach", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "score": 1.0, + "content": "where we jointly maximize the mutual information between the domain labels and private (domain-specific)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 608, + 523, + 619 + ], + "spans": [ + { + "bbox": [ + 89, + 608, + 523, + 619 + ], + "score": 1.0, + "content": "features, while minimizing the mutual information between the the domain labels and the shared (domain-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 90, + 619, + 522, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 619, + 522, + 630 + ], + "score": 1.0, + "content": "invariant) features. Consequently, the more robust feature representations are learned for each target domain", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 90, + 629, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 629, + 522, + 642 + ], + "score": 1.0, + "content": "by exploiting dependencies between multiple target domains. We show on benchmark datasets for DA that", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 640, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 640, + 522, + 653 + ], + "score": 1.0, + "content": "this approach leads to overall improved performance on each target domain, compared to independent DA for", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 292, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 670, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 285, + 670, + 294, + 683 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 84, + 480, + 209 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 84, + 480, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 84, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 129, + 84, + 480, + 209 + ], + "score": 0.971, + "type": "image", + "image_path": "36886e0179e0304a2140f74c7b598abbe23c7d9c3021a983addc10b904e84af3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 84, + 480, + 125.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 125.66666666666666, + 480, + 167.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 167.33333333333331, + 480, + 208.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 218, + 521, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 218, + 521, + 230 + ], + "spans": [ + { + "bbox": [ + 90, + 218, + 521, + 230 + ], + "score": 1.0, + "content": "Figure 1: Illustration of domains with common (a) and pairwise-shared spaces (b). We tackle the domain adaptation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 227, + 522, + 241 + ], + "spans": [ + { + "bbox": [ + 89, + 227, + 522, + 241 + ], + "score": 1.0, + "content": "task when all domains share a common task/space, which is then leveraged to transfer knowledge across multiple target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 239, + 125, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 239, + 125, + 249 + ], + "score": 1.0, + "content": "domains.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 90, + 261, + 521, + 503 + ], + "lines": [], + "index": 16.5, + "bbox_fs": [ + 88, + 262, + 523, + 505 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 91, + 508, + 522, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 509, + 523, + 522 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 523, + 522 + ], + "score": 1.0, + "content": "To this end, we propose a Multi-Target DA-Information-Theoretic-Approach (MTDA-ITA) for single-source-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 519, + 522, + 533 + ], + "spans": [ + { + "bbox": [ + 89, + 519, + 522, + 533 + ], + "score": 1.0, + "content": "multi-target DA. We exploit a single source domain and focus on multiple target domains to investigate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 529, + 522, + 544 + ], + "spans": [ + { + "bbox": [ + 89, + 529, + 522, + 544 + ], + "score": 1.0, + "content": "the effects of multi-target DA; however, the proposed approach can easily be extended to multiple source", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "score": 1.0, + "content": "domains. This approach leverages the data from multiple target domains to improve performance compared", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 552, + 522, + 565 + ], + "spans": [ + { + "bbox": [ + 89, + 552, + 522, + 565 + ], + "score": 1.0, + "content": "to individually learning from pair-wise source-target domains. Specifically, we simultaneously factorize the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 563, + 523, + 577 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 523, + 577 + ], + "score": 1.0, + "content": "information from each available target domain and learn separate subspaces for modeling the shared (i.e.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 575, + 522, + 586 + ], + "spans": [ + { + "bbox": [ + 89, + 575, + 522, + 586 + ], + "score": 1.0, + "content": "correlated across the domains) and private (i.e., independent between the domains) subspaces of the data", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 586, + 522, + 599 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 522, + 599 + ], + "score": 1.0, + "content": "(Salzmann et al. (2010)). To this end, we employ deep learning to derive an information theoretic approach", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "score": 1.0, + "content": "where we jointly maximize the mutual information between the domain labels and private (domain-specific)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 608, + 523, + 619 + ], + "spans": [ + { + "bbox": [ + 89, + 608, + 523, + 619 + ], + "score": 1.0, + "content": "features, while minimizing the mutual information between the the domain labels and the shared (domain-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 90, + 619, + 522, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 619, + 522, + 630 + ], + "score": 1.0, + "content": "invariant) features. Consequently, the more robust feature representations are learned for each target domain", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 90, + 629, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 629, + 522, + 642 + ], + "score": 1.0, + "content": "by exploiting dependencies between multiple target domains. We show on benchmark datasets for DA that", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 640, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 640, + 522, + 653 + ], + "score": 1.0, + "content": "this approach leads to overall improved performance on each target domain, compared to independent DA for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 95, + 522, + 108 + ], + "spans": [ + { + "bbox": [ + 89, + 95, + 522, + 108 + ], + "score": 1.0, + "content": "each pair of source-target domains, or the naive combination of multiple target domains, and state-of-the-art", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 106, + 236, + 119 + ], + "spans": [ + { + "bbox": [ + 89, + 106, + 236, + 119 + ], + "score": 1.0, + "content": "models applicable to the target task.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 34, + "bbox_fs": [ + 89, + 509, + 523, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 89, + 95, + 522, + 118 + ], + "lines": [ + { + "bbox": [ + 89, + 95, + 522, + 108 + ], + "spans": [ + { + "bbox": [ + 89, + 95, + 522, + 108 + ], + "score": 1.0, + "content": "each pair of source-target domains, or the naive combination of multiple target domains, and state-of-the-art", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 106, + 236, + 119 + ], + "spans": [ + { + "bbox": [ + 89, + 106, + 236, + 119 + ], + "score": 1.0, + "content": "models applicable to the target task.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 91, + 134, + 239, + 147 + ], + "lines": [ + { + "bbox": [ + 88, + 133, + 240, + 149 + ], + "spans": [ + { + "bbox": [ + 88, + 133, + 240, + 149 + ], + "score": 1.0, + "content": "2 THE PROPOSED METHOD", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 90, + 159, + 522, + 218 + ], + "lines": [ + { + "bbox": [ + 90, + 160, + 522, + 172 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 301, + 172 + ], + "score": 1.0, + "content": "Without loss of generality, we consider a multi-class (", + "type": "text" + }, + { + "bbox": [ + 302, + 161, + 311, + 170 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 160, + 522, + 172 + ], + "score": 1.0, + "content": "-class) classification problem as the running example.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 170, + 523, + 187 + ], + "spans": [ + { + "bbox": [ + 88, + 170, + 160, + 187 + ], + "score": 1.0, + "content": "Furthermore, let", + "type": "text" + }, + { + "bbox": [ + 160, + 171, + 287, + 184 + ], + "score": 0.93, + "content": "( \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } ) = \\{ ( \\pmb { x } _ { i } , \\pmb { y } _ { i } , \\pmb { d } _ { i } ) \\} _ { i = 0 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 170, + 363, + 187 + ], + "score": 1.0, + "content": "be a collection of", + "type": "text" + }, + { + "bbox": [ + 363, + 172, + 375, + 182 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 170, + 523, + 187 + ], + "score": 1.0, + "content": "domains (a labeled source domain,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 182, + 522, + 196 + ], + "spans": [ + { + "bbox": [ + 90, + 183, + 108, + 196 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 108, + 183, + 138, + 194 + ], + "score": 0.86, + "content": "M - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 183, + 281, + 196 + ], + "score": 1.0, + "content": "unlabeled target domains), where", + "type": "text" + }, + { + "bbox": [ + 282, + 185, + 293, + 194 + ], + "score": 0.86, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 183, + 345, + 196 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 345, + 184, + 350, + 193 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 183, + 416, + 196 + ], + "score": 1.0, + "content": "-th sample, and", + "type": "text" + }, + { + "bbox": [ + 416, + 182, + 502, + 195 + ], + "score": 0.93, + "content": "\\pmb { y } _ { i } = [ y _ { i } ^ { 1 } , y _ { i } ^ { 2 } , . . . , y _ { i } ^ { K } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 183, + 522, + 196 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 192, + 523, + 209 + ], + "spans": [ + { + "bbox": [ + 91, + 195, + 174, + 207 + ], + "score": 0.92, + "content": "d _ { i } = [ d _ { i } ^ { 1 } , d _ { i } ^ { 2 } , . . . , d _ { i } ^ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 192, + 207, + 209 + ], + "score": 1.0, + "content": "are the", + "type": "text" + }, + { + "bbox": [ + 207, + 195, + 217, + 205 + ], + "score": 0.72, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 192, + 248, + 209 + ], + "score": 1.0, + "content": "-D and", + "type": "text" + }, + { + "bbox": [ + 248, + 195, + 259, + 205 + ], + "score": 0.7, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 192, + 454, + 209 + ], + "score": 1.0, + "content": "-D encoding of the class and domain labels for", + "type": "text" + }, + { + "bbox": [ + 454, + 197, + 465, + 206 + ], + "score": 0.84, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 192, + 523, + 209 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 205, + 358, + 218 + ], + "spans": [ + { + "bbox": [ + 89, + 205, + 358, + 218 + ], + "score": 1.0, + "content": "Note that the class labels are only available for the source samples.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 91, + 223, + 520, + 256 + ], + "lines": [ + { + "bbox": [ + 89, + 223, + 522, + 236 + ], + "spans": [ + { + "bbox": [ + 89, + 223, + 293, + 236 + ], + "score": 1.0, + "content": "The latent space representation of the data point", + "type": "text" + }, + { + "bbox": [ + 293, + 226, + 301, + 233 + ], + "score": 0.77, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 223, + 360, + 236 + ], + "score": 1.0, + "content": "is denoted as", + "type": "text" + }, + { + "bbox": [ + 360, + 223, + 413, + 236 + ], + "score": 0.94, + "content": "\\boldsymbol { z } = [ z _ { s } , z _ { p } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 223, + 446, + 236 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 446, + 225, + 457, + 234 + ], + "score": 0.86, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 223, + 477, + 236 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 477, + 225, + 488, + 235 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 223, + 522, + 236 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 233, + 522, + 248 + ], + "spans": [ + { + "bbox": [ + 89, + 233, + 306, + 248 + ], + "score": 1.0, + "content": "(latent) shared and private features of the data point", + "type": "text" + }, + { + "bbox": [ + 306, + 236, + 313, + 244 + ], + "score": 0.78, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 233, + 522, + 248 + ], + "score": 1.0, + "content": ", respectively. By factorizing the joint distribution", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 245, + 173, + 258 + ], + "spans": [ + { + "bbox": [ + 91, + 245, + 160, + 258 + ], + "score": 0.91, + "content": "p ( { \\pmb x } , { \\pmb y } , d , z _ { s } , z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 245, + 173, + 258 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 255, + 416, + 270 + ], + "lines": [ + { + "bbox": [ + 194, + 255, + 416, + 270 + ], + "spans": [ + { + "bbox": [ + 194, + 255, + 416, + 270 + ], + "score": 0.88, + "content": "\\begin{array} { r } { p ( { \\pmb x } , { \\pmb y } , { \\pmb d } , z _ { s } , z _ { p } ) = p ( { \\pmb x } ) p ( { \\pmb d } ) p ( z _ { s } | { \\pmb x } ) p ( z _ { p } | { \\pmb x } ) p ( { \\pmb y } | z _ { s } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "6ba6e0a7950e1dd2dbca048af80b4411ac3bcf9c16cad5bded0e0a26e8982113.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 194, + 255, + 416, + 270 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 272, + 322, + 284 + ], + "lines": [ + { + "bbox": [ + 89, + 272, + 322, + 286 + ], + "spans": [ + { + "bbox": [ + 89, + 272, + 322, + 286 + ], + "score": 1.0, + "content": "we propose to maximize the following objective function:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 151, + 290, + 459, + 306 + ], + "lines": [ + { + "bbox": [ + 151, + 290, + 459, + 306 + ], + "spans": [ + { + "bbox": [ + 151, + 290, + 459, + 306 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\boldsymbol { \\mathcal { L } } ( \\boldsymbol { \\theta } _ { s } , \\boldsymbol { \\theta } _ { p } , \\boldsymbol { \\theta } _ { c } ; \\boldsymbol { x } , \\boldsymbol { y } , \\boldsymbol { d } ) = \\lambda _ { r } \\boldsymbol { I } ( \\boldsymbol { x } ; \\boldsymbol { z } ) + \\lambda _ { c } \\boldsymbol { I } ( \\boldsymbol { y } ; \\boldsymbol { z } _ { s } ) + \\lambda _ { d } \\big ( \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { p } ) - \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { s } ) \\big ) , } \\end{array}", + "type": "interline_equation", + "image_path": "2ebeb1570173651be3e9d961e9dfdaa8a2afd212b856849a8d63e5244997c423.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 151, + 290, + 459, + 306 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 312, + 521, + 357 + ], + "lines": [ + { + "bbox": [ + 89, + 312, + 522, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 312, + 116, + 326 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 116, + 313, + 136, + 325 + ], + "score": 0.92, + "content": "p ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 312, + 154, + 326 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 313, + 174, + 325 + ], + "score": 0.92, + "content": "p ( d )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 312, + 522, + 326 + ], + "score": 1.0, + "content": "denote the underlying (true) data distribution and domain label distribution, respectively,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 91, + 323, + 522, + 337 + ], + "spans": [ + { + "bbox": [ + 91, + 324, + 122, + 336 + ], + "score": 0.92, + "content": "I ( { \\pmb x } ; { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 323, + 387, + 337 + ], + "score": 1.0, + "content": "denotes the Mutual Information between the random variables", + "type": "text" + }, + { + "bbox": [ + 388, + 326, + 396, + 334 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 323, + 416, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 416, + 325, + 424, + 335 + ], + "score": 0.64, + "content": "\\textbf { { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 323, + 431, + 337 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 431, + 324, + 457, + 335 + ], + "score": 0.9, + "content": "\\lambda _ { r } , \\lambda _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 323, + 477, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 478, + 324, + 489, + 335 + ], + "score": 0.89, + "content": "\\lambda _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 323, + 522, + 337 + ], + "score": 1.0, + "content": "denote", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 333, + 522, + 348 + ], + "spans": [ + { + "bbox": [ + 89, + 333, + 522, + 348 + ], + "score": 1.0, + "content": "the hyper-parameters controlling the weights of the objective terms. The proposed objective function (2)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 345, + 268, + 358 + ], + "spans": [ + { + "bbox": [ + 89, + 345, + 268, + 358 + ], + "score": 1.0, + "content": "maximizes the three terms described below:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 116, + 366, + 521, + 465 + ], + "lines": [ + { + "bbox": [ + 116, + 366, + 521, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 126, + 379 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 126, + 367, + 158, + 379 + ], + "score": 0.89, + "content": "I ( { \\pmb x } ; z )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 366, + 521, + 379 + ], + "score": 1.0, + "content": ": encourages the latent features (both shared and private) to preserve information about the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 378, + 351, + 390 + ], + "spans": [ + { + "bbox": [ + 125, + 378, + 306, + 390 + ], + "score": 1.0, + "content": "data samples (that can be used to reconstruct", + "type": "text" + }, + { + "bbox": [ + 307, + 380, + 315, + 388 + ], + "score": 0.69, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 378, + 337, + 390 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 338, + 380, + 344, + 388 + ], + "score": 0.67, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 378, + 351, + 390 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 392, + 521, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 126, + 406 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 126, + 393, + 160, + 405 + ], + "score": 0.92, + "content": "I ( \\pmb { y } ; \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 392, + 521, + 406 + ], + "score": 1.0, + "content": ": enables to correctly predict the true class label of the samples out of their common shared", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 404, + 163, + 416 + ], + "spans": [ + { + "bbox": [ + 125, + 404, + 163, + 416 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 419, + 522, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 126, + 433 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 126, + 419, + 208, + 432 + ], + "score": 0.9, + "content": "I ( d ; z _ { p } ) - I ( d ; z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 419, + 522, + 433 + ], + "score": 1.0, + "content": ": encourages the latent private features to preserve the information about the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 124, + 430, + 521, + 443 + ], + "spans": [ + { + "bbox": [ + 124, + 430, + 521, + 443 + ], + "score": 1.0, + "content": "domain label and penalizes the latent shared features to be domain informative. This not only reduces", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 442, + 521, + 453 + ], + "spans": [ + { + "bbox": [ + 126, + 442, + 521, + 453 + ], + "score": 1.0, + "content": "the redundancy in the shared and private features, but also, penalizes the redundancy of different", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 124, + 453, + 351, + 465 + ], + "spans": [ + { + "bbox": [ + 124, + 453, + 351, + 465 + ], + "score": 1.0, + "content": "private spaces, while preserving the shared information.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 90, + 473, + 521, + 519 + ], + "lines": [ + { + "bbox": [ + 89, + 474, + 521, + 487 + ], + "spans": [ + { + "bbox": [ + 89, + 474, + 455, + 487 + ], + "score": 1.0, + "content": "An additional term could be used to minimize the mutual information between the shared", + "type": "text" + }, + { + "bbox": [ + 455, + 474, + 473, + 485 + ], + "score": 0.84, + "content": "( z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 474, + 521, + 487 + ], + "score": 1.0, + "content": "and private", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 484, + 522, + 498 + ], + "spans": [ + { + "bbox": [ + 90, + 486, + 108, + 497 + ], + "score": 0.86, + "content": "( z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 108, + 484, + 522, + 498 + ], + "score": 1.0, + "content": "features. However, computing the mutual information (even approximating it) is intractable due to the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 496, + 522, + 509 + ], + "spans": [ + { + "bbox": [ + 90, + 496, + 223, + 509 + ], + "score": 1.0, + "content": "highly complex joint distribution", + "type": "text" + }, + { + "bbox": [ + 224, + 496, + 261, + 508 + ], + "score": 0.92, + "content": "p ( z _ { s } , z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 496, + 326, + 509 + ], + "score": 1.0, + "content": ". Since we want", + "type": "text" + }, + { + "bbox": [ + 326, + 497, + 337, + 507 + ], + "score": 0.87, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 496, + 355, + 509 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 497, + 367, + 507 + ], + "score": 0.85, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 496, + 522, + 509 + ], + "score": 1.0, + "content": "features to encode different aspects of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 91, + 506, + 425, + 521 + ], + "spans": [ + { + "bbox": [ + 91, + 509, + 98, + 517 + ], + "score": 0.64, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 98, + 506, + 340, + 521 + ], + "score": 1.0, + "content": ", we enforce such constraint by jointly maximizing the term:", + "type": "text" + }, + { + "bbox": [ + 341, + 507, + 421, + 519 + ], + "score": 0.92, + "content": "I ( d ; \\dot { z } _ { p } ) - I ( d ; z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 506, + 425, + 521 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 91, + 532, + 180, + 544 + ], + "lines": [ + { + "bbox": [ + 89, + 531, + 181, + 545 + ], + "spans": [ + { + "bbox": [ + 89, + 531, + 181, + 545 + ], + "score": 1.0, + "content": "2.1 OPTIMIZATION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 90, + 553, + 521, + 581 + ], + "lines": [ + { + "bbox": [ + 88, + 552, + 521, + 567 + ], + "spans": [ + { + "bbox": [ + 88, + 552, + 521, + 567 + ], + "score": 1.0, + "content": "The following lower bound for mutual information is derived using the non-negativity of KL-divergence", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 87, + 564, + 343, + 583 + ], + "spans": [ + { + "bbox": [ + 87, + 564, + 219, + 583 + ], + "score": 1.0, + "content": "(Barber & Agakov (2003)); i.e.,", + "type": "text" + }, + { + "bbox": [ + 219, + 564, + 316, + 581 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\Sigma _ { { \\pmb x } } p ( { \\pmb x } | z ) \\ln \\frac { p ( { \\pmb x } | z ) } { q ( { \\pmb x } | z ) } \\ge 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 564, + 343, + 583 + ], + "score": 1.0, + "content": "gives:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 586, + 388, + 601 + ], + "lines": [ + { + "bbox": [ + 223, + 586, + 388, + 601 + ], + "spans": [ + { + "bbox": [ + 223, + 586, + 388, + 601 + ], + "score": 0.92, + "content": "I ( \\pmb { x } ; z ) \\geq H ( \\pmb { x } ) + \\mathbb { E } _ { p ( \\pmb { x } , z ) } [ \\ln q ( \\pmb { x } | z ; \\phi ) ]", + "type": "interline_equation", + "image_path": "a13d30e2817b0209208e09198a760859aaa7aca32fa8311d73bdbdb853ec071e.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 223, + 586, + 388, + 601 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 606, + 521, + 653 + ], + "lines": [ + { + "bbox": [ + 90, + 607, + 521, + 620 + ], + "spans": [ + { + "bbox": [ + 90, + 607, + 117, + 620 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 117, + 608, + 141, + 619 + ], + "score": 0.92, + "content": "H ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 607, + 405, + 620 + ], + "score": 1.0, + "content": "denotes the Shanon Entropy (Lin (1991)) of the random variable", + "type": "text" + }, + { + "bbox": [ + 405, + 609, + 413, + 617 + ], + "score": 0.65, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 607, + 417, + 620 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 418, + 607, + 457, + 619 + ], + "score": 0.91, + "content": "q ( { \\pmb x } | { \\pmb z } ; { \\pmb \\phi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 607, + 521, + 620 + ], + "score": 1.0, + "content": "is any arbitrary", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 90, + 618, + 521, + 631 + ], + "spans": [ + { + "bbox": [ + 90, + 618, + 207, + 631 + ], + "score": 1.0, + "content": "distribution parameterized by", + "type": "text" + }, + { + "bbox": [ + 207, + 619, + 213, + 630 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 618, + 350, + 631 + ], + "score": 1.0, + "content": ". We need a variational distribution", + "type": "text" + }, + { + "bbox": [ + 350, + 618, + 389, + 630 + ], + "score": 0.92, + "content": "q ( { \\pmb x } | { \\pmb z } ; \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 618, + 521, + 631 + ], + "score": 1.0, + "content": "because the posterior distribution", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 91, + 628, + 523, + 643 + ], + "spans": [ + { + "bbox": [ + 91, + 629, + 205, + 641 + ], + "score": 0.93, + "content": "p ( \\pmb { x } | \\pmb { z } ) = p ( \\pmb { z } | \\pmb { x } ) p ( \\pmb { x } ) / p ( \\pmb { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 628, + 388, + 643 + ], + "score": 1.0, + "content": "is intractable since the true data distribution", + "type": "text" + }, + { + "bbox": [ + 388, + 630, + 408, + 641 + ], + "score": 0.9, + "content": "p ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 628, + 523, + 643 + ], + "score": 1.0, + "content": "is assumed to be unknown.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 639, + 521, + 655 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 271, + 655 + ], + "score": 1.0, + "content": "Similarly, we can derive lower bounds for", + "type": "text" + }, + { + "bbox": [ + 272, + 640, + 452, + 653 + ], + "score": 0.9, + "content": "I ( d ; z _ { p } ) \\geq H ( d ) + \\mathbb { E } _ { p ( d , z _ { p } ) } [ \\ln q ( d | z _ { p } ; \\psi ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 639, + 473, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 473, + 640, + 521, + 652 + ], + "score": 0.91, + "content": "I ( d ; z _ { s } ) \\geq", + "type": "inline_equation" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 39, + 292, + 51 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 293, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 670, + 293, + 681 + ], + "spans": [ + { + "bbox": [ + 285, + 670, + 293, + 681 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 89, + 95, + 522, + 118 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 89, + 95, + 522, + 119 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 91, + 134, + 239, + 147 + ], + "lines": [ + { + "bbox": [ + 88, + 133, + 240, + 149 + ], + "spans": [ + { + "bbox": [ + 88, + 133, + 240, + 149 + ], + "score": 1.0, + "content": "2 THE PROPOSED METHOD", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 90, + 159, + 522, + 218 + ], + "lines": [ + { + "bbox": [ + 90, + 160, + 522, + 172 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 301, + 172 + ], + "score": 1.0, + "content": "Without loss of generality, we consider a multi-class (", + "type": "text" + }, + { + "bbox": [ + 302, + 161, + 311, + 170 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 160, + 522, + 172 + ], + "score": 1.0, + "content": "-class) classification problem as the running example.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 170, + 523, + 187 + ], + "spans": [ + { + "bbox": [ + 88, + 170, + 160, + 187 + ], + "score": 1.0, + "content": "Furthermore, let", + "type": "text" + }, + { + "bbox": [ + 160, + 171, + 287, + 184 + ], + "score": 0.93, + "content": "( \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } ) = \\{ ( \\pmb { x } _ { i } , \\pmb { y } _ { i } , \\pmb { d } _ { i } ) \\} _ { i = 0 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 170, + 363, + 187 + ], + "score": 1.0, + "content": "be a collection of", + "type": "text" + }, + { + "bbox": [ + 363, + 172, + 375, + 182 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 170, + 523, + 187 + ], + "score": 1.0, + "content": "domains (a labeled source domain,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 182, + 522, + 196 + ], + "spans": [ + { + "bbox": [ + 90, + 183, + 108, + 196 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 108, + 183, + 138, + 194 + ], + "score": 0.86, + "content": "M - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 183, + 281, + 196 + ], + "score": 1.0, + "content": "unlabeled target domains), where", + "type": "text" + }, + { + "bbox": [ + 282, + 185, + 293, + 194 + ], + "score": 0.86, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 183, + 345, + 196 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 345, + 184, + 350, + 193 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 183, + 416, + 196 + ], + "score": 1.0, + "content": "-th sample, and", + "type": "text" + }, + { + "bbox": [ + 416, + 182, + 502, + 195 + ], + "score": 0.93, + "content": "\\pmb { y } _ { i } = [ y _ { i } ^ { 1 } , y _ { i } ^ { 2 } , . . . , y _ { i } ^ { K } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 183, + 522, + 196 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 192, + 523, + 209 + ], + "spans": [ + { + "bbox": [ + 91, + 195, + 174, + 207 + ], + "score": 0.92, + "content": "d _ { i } = [ d _ { i } ^ { 1 } , d _ { i } ^ { 2 } , . . . , d _ { i } ^ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 192, + 207, + 209 + ], + "score": 1.0, + "content": "are the", + "type": "text" + }, + { + "bbox": [ + 207, + 195, + 217, + 205 + ], + "score": 0.72, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 192, + 248, + 209 + ], + "score": 1.0, + "content": "-D and", + "type": "text" + }, + { + "bbox": [ + 248, + 195, + 259, + 205 + ], + "score": 0.7, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 192, + 454, + 209 + ], + "score": 1.0, + "content": "-D encoding of the class and domain labels for", + "type": "text" + }, + { + "bbox": [ + 454, + 197, + 465, + 206 + ], + "score": 0.84, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 192, + 523, + 209 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 205, + 358, + 218 + ], + "spans": [ + { + "bbox": [ + 89, + 205, + 358, + 218 + ], + "score": 1.0, + "content": "Note that the class labels are only available for the source samples.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 88, + 160, + 523, + 218 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 223, + 520, + 256 + ], + "lines": [ + { + "bbox": [ + 89, + 223, + 522, + 236 + ], + "spans": [ + { + "bbox": [ + 89, + 223, + 293, + 236 + ], + "score": 1.0, + "content": "The latent space representation of the data point", + "type": "text" + }, + { + "bbox": [ + 293, + 226, + 301, + 233 + ], + "score": 0.77, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 223, + 360, + 236 + ], + "score": 1.0, + "content": "is denoted as", + "type": "text" + }, + { + "bbox": [ + 360, + 223, + 413, + 236 + ], + "score": 0.94, + "content": "\\boldsymbol { z } = [ z _ { s } , z _ { p } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 223, + 446, + 236 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 446, + 225, + 457, + 234 + ], + "score": 0.86, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 223, + 477, + 236 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 477, + 225, + 488, + 235 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 223, + 522, + 236 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 233, + 522, + 248 + ], + "spans": [ + { + "bbox": [ + 89, + 233, + 306, + 248 + ], + "score": 1.0, + "content": "(latent) shared and private features of the data point", + "type": "text" + }, + { + "bbox": [ + 306, + 236, + 313, + 244 + ], + "score": 0.78, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 233, + 522, + 248 + ], + "score": 1.0, + "content": ", respectively. By factorizing the joint distribution", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 245, + 173, + 258 + ], + "spans": [ + { + "bbox": [ + 91, + 245, + 160, + 258 + ], + "score": 0.91, + "content": "p ( { \\pmb x } , { \\pmb y } , d , z _ { s } , z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 245, + 173, + 258 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 89, + 223, + 522, + 258 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 255, + 416, + 270 + ], + "lines": [ + { + "bbox": [ + 194, + 255, + 416, + 270 + ], + "spans": [ + { + "bbox": [ + 194, + 255, + 416, + 270 + ], + "score": 0.88, + "content": "\\begin{array} { r } { p ( { \\pmb x } , { \\pmb y } , { \\pmb d } , z _ { s } , z _ { p } ) = p ( { \\pmb x } ) p ( { \\pmb d } ) p ( z _ { s } | { \\pmb x } ) p ( z _ { p } | { \\pmb x } ) p ( { \\pmb y } | z _ { s } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "6ba6e0a7950e1dd2dbca048af80b4411ac3bcf9c16cad5bded0e0a26e8982113.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 194, + 255, + 416, + 270 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 272, + 322, + 284 + ], + "lines": [ + { + "bbox": [ + 89, + 272, + 322, + 286 + ], + "spans": [ + { + "bbox": [ + 89, + 272, + 322, + 286 + ], + "score": 1.0, + "content": "we propose to maximize the following objective function:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 89, + 272, + 322, + 286 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 151, + 290, + 459, + 306 + ], + "lines": [ + { + "bbox": [ + 151, + 290, + 459, + 306 + ], + "spans": [ + { + "bbox": [ + 151, + 290, + 459, + 306 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\boldsymbol { \\mathcal { L } } ( \\boldsymbol { \\theta } _ { s } , \\boldsymbol { \\theta } _ { p } , \\boldsymbol { \\theta } _ { c } ; \\boldsymbol { x } , \\boldsymbol { y } , \\boldsymbol { d } ) = \\lambda _ { r } \\boldsymbol { I } ( \\boldsymbol { x } ; \\boldsymbol { z } ) + \\lambda _ { c } \\boldsymbol { I } ( \\boldsymbol { y } ; \\boldsymbol { z } _ { s } ) + \\lambda _ { d } \\big ( \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { p } ) - \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { s } ) \\big ) , } \\end{array}", + "type": "interline_equation", + "image_path": "2ebeb1570173651be3e9d961e9dfdaa8a2afd212b856849a8d63e5244997c423.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 151, + 290, + 459, + 306 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 312, + 521, + 357 + ], + "lines": [ + { + "bbox": [ + 89, + 312, + 522, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 312, + 116, + 326 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 116, + 313, + 136, + 325 + ], + "score": 0.92, + "content": "p ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 312, + 154, + 326 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 313, + 174, + 325 + ], + "score": 0.92, + "content": "p ( d )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 312, + 522, + 326 + ], + "score": 1.0, + "content": "denote the underlying (true) data distribution and domain label distribution, respectively,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 91, + 323, + 522, + 337 + ], + "spans": [ + { + "bbox": [ + 91, + 324, + 122, + 336 + ], + "score": 0.92, + "content": "I ( { \\pmb x } ; { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 323, + 387, + 337 + ], + "score": 1.0, + "content": "denotes the Mutual Information between the random variables", + "type": "text" + }, + { + "bbox": [ + 388, + 326, + 396, + 334 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 323, + 416, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 416, + 325, + 424, + 335 + ], + "score": 0.64, + "content": "\\textbf { { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 323, + 431, + 337 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 431, + 324, + 457, + 335 + ], + "score": 0.9, + "content": "\\lambda _ { r } , \\lambda _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 323, + 477, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 478, + 324, + 489, + 335 + ], + "score": 0.89, + "content": "\\lambda _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 323, + 522, + 337 + ], + "score": 1.0, + "content": "denote", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 333, + 522, + 348 + ], + "spans": [ + { + "bbox": [ + 89, + 333, + 522, + 348 + ], + "score": 1.0, + "content": "the hyper-parameters controlling the weights of the objective terms. The proposed objective function (2)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 345, + 268, + 358 + ], + "spans": [ + { + "bbox": [ + 89, + 345, + 268, + 358 + ], + "score": 1.0, + "content": "maximizes the three terms described below:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 89, + 312, + 522, + 358 + ] + }, + { + "type": "list", + "bbox": [ + 116, + 366, + 521, + 465 + ], + "lines": [ + { + "bbox": [ + 116, + 366, + 521, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 126, + 379 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 126, + 367, + 158, + 379 + ], + "score": 0.89, + "content": "I ( { \\pmb x } ; z )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 366, + 521, + 379 + ], + "score": 1.0, + "content": ": encourages the latent features (both shared and private) to preserve information about the", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 378, + 351, + 390 + ], + "spans": [ + { + "bbox": [ + 125, + 378, + 306, + 390 + ], + "score": 1.0, + "content": "data samples (that can be used to reconstruct", + "type": "text" + }, + { + "bbox": [ + 307, + 380, + 315, + 388 + ], + "score": 0.69, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 378, + 337, + 390 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 338, + 380, + 344, + 388 + ], + "score": 0.67, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 378, + 351, + 390 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 392, + 521, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 126, + 406 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 126, + 393, + 160, + 405 + ], + "score": 0.92, + "content": "I ( \\pmb { y } ; \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 392, + 521, + 406 + ], + "score": 1.0, + "content": ": enables to correctly predict the true class label of the samples out of their common shared", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 404, + 163, + 416 + ], + "spans": [ + { + "bbox": [ + 125, + 404, + 163, + 416 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 115, + 419, + 522, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 126, + 433 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 126, + 419, + 208, + 432 + ], + "score": 0.9, + "content": "I ( d ; z _ { p } ) - I ( d ; z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 419, + 522, + 433 + ], + "score": 1.0, + "content": ": encourages the latent private features to preserve the information about the", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 430, + 521, + 443 + ], + "spans": [ + { + "bbox": [ + 124, + 430, + 521, + 443 + ], + "score": 1.0, + "content": "domain label and penalizes the latent shared features to be domain informative. This not only reduces", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 442, + 521, + 453 + ], + "spans": [ + { + "bbox": [ + 126, + 442, + 521, + 453 + ], + "score": 1.0, + "content": "the redundancy in the shared and private features, but also, penalizes the redundancy of different", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 124, + 453, + 351, + 465 + ], + "spans": [ + { + "bbox": [ + 124, + 453, + 351, + 465 + ], + "score": 1.0, + "content": "private spaces, while preserving the shared information.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + } + ], + "index": 21.5, + "bbox_fs": [ + 115, + 366, + 522, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 473, + 521, + 519 + ], + "lines": [ + { + "bbox": [ + 89, + 474, + 521, + 487 + ], + "spans": [ + { + "bbox": [ + 89, + 474, + 455, + 487 + ], + "score": 1.0, + "content": "An additional term could be used to minimize the mutual information between the shared", + "type": "text" + }, + { + "bbox": [ + 455, + 474, + 473, + 485 + ], + "score": 0.84, + "content": "( z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 474, + 521, + 487 + ], + "score": 1.0, + "content": "and private", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 484, + 522, + 498 + ], + "spans": [ + { + "bbox": [ + 90, + 486, + 108, + 497 + ], + "score": 0.86, + "content": "( z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 108, + 484, + 522, + 498 + ], + "score": 1.0, + "content": "features. However, computing the mutual information (even approximating it) is intractable due to the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 496, + 522, + 509 + ], + "spans": [ + { + "bbox": [ + 90, + 496, + 223, + 509 + ], + "score": 1.0, + "content": "highly complex joint distribution", + "type": "text" + }, + { + "bbox": [ + 224, + 496, + 261, + 508 + ], + "score": 0.92, + "content": "p ( z _ { s } , z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 496, + 326, + 509 + ], + "score": 1.0, + "content": ". Since we want", + "type": "text" + }, + { + "bbox": [ + 326, + 497, + 337, + 507 + ], + "score": 0.87, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 496, + 355, + 509 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 497, + 367, + 507 + ], + "score": 0.85, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 496, + 522, + 509 + ], + "score": 1.0, + "content": "features to encode different aspects of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 91, + 506, + 425, + 521 + ], + "spans": [ + { + "bbox": [ + 91, + 509, + 98, + 517 + ], + "score": 0.64, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 98, + 506, + 340, + 521 + ], + "score": 1.0, + "content": ", we enforce such constraint by jointly maximizing the term:", + "type": "text" + }, + { + "bbox": [ + 341, + 507, + 421, + 519 + ], + "score": 0.92, + "content": "I ( d ; \\dot { z } _ { p } ) - I ( d ; z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 506, + 425, + 521 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 89, + 474, + 522, + 521 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 532, + 180, + 544 + ], + "lines": [ + { + "bbox": [ + 89, + 531, + 181, + 545 + ], + "spans": [ + { + "bbox": [ + 89, + 531, + 181, + 545 + ], + "score": 1.0, + "content": "2.1 OPTIMIZATION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 90, + 553, + 521, + 581 + ], + "lines": [ + { + "bbox": [ + 88, + 552, + 521, + 567 + ], + "spans": [ + { + "bbox": [ + 88, + 552, + 521, + 567 + ], + "score": 1.0, + "content": "The following lower bound for mutual information is derived using the non-negativity of KL-divergence", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 87, + 564, + 343, + 583 + ], + "spans": [ + { + "bbox": [ + 87, + 564, + 219, + 583 + ], + "score": 1.0, + "content": "(Barber & Agakov (2003)); i.e.,", + "type": "text" + }, + { + "bbox": [ + 219, + 564, + 316, + 581 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\Sigma _ { { \\pmb x } } p ( { \\pmb x } | z ) \\ln \\frac { p ( { \\pmb x } | z ) } { q ( { \\pmb x } | z ) } \\ge 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 564, + 343, + 583 + ], + "score": 1.0, + "content": "gives:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 87, + 552, + 521, + 583 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 586, + 388, + 601 + ], + "lines": [ + { + "bbox": [ + 223, + 586, + 388, + 601 + ], + "spans": [ + { + "bbox": [ + 223, + 586, + 388, + 601 + ], + "score": 0.92, + "content": "I ( \\pmb { x } ; z ) \\geq H ( \\pmb { x } ) + \\mathbb { E } _ { p ( \\pmb { x } , z ) } [ \\ln q ( \\pmb { x } | z ; \\phi ) ]", + "type": "interline_equation", + "image_path": "a13d30e2817b0209208e09198a760859aaa7aca32fa8311d73bdbdb853ec071e.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 223, + 586, + 388, + 601 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 606, + 521, + 653 + ], + "lines": [ + { + "bbox": [ + 90, + 607, + 521, + 620 + ], + "spans": [ + { + "bbox": [ + 90, + 607, + 117, + 620 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 117, + 608, + 141, + 619 + ], + "score": 0.92, + "content": "H ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 607, + 405, + 620 + ], + "score": 1.0, + "content": "denotes the Shanon Entropy (Lin (1991)) of the random variable", + "type": "text" + }, + { + "bbox": [ + 405, + 609, + 413, + 617 + ], + "score": 0.65, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 607, + 417, + 620 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 418, + 607, + 457, + 619 + ], + "score": 0.91, + "content": "q ( { \\pmb x } | { \\pmb z } ; { \\pmb \\phi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 607, + 521, + 620 + ], + "score": 1.0, + "content": "is any arbitrary", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 90, + 618, + 521, + 631 + ], + "spans": [ + { + "bbox": [ + 90, + 618, + 207, + 631 + ], + "score": 1.0, + "content": "distribution parameterized by", + "type": "text" + }, + { + "bbox": [ + 207, + 619, + 213, + 630 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 618, + 350, + 631 + ], + "score": 1.0, + "content": ". We need a variational distribution", + "type": "text" + }, + { + "bbox": [ + 350, + 618, + 389, + 630 + ], + "score": 0.92, + "content": "q ( { \\pmb x } | { \\pmb z } ; \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 618, + 521, + 631 + ], + "score": 1.0, + "content": "because the posterior distribution", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 91, + 628, + 523, + 643 + ], + "spans": [ + { + "bbox": [ + 91, + 629, + 205, + 641 + ], + "score": 0.93, + "content": "p ( \\pmb { x } | \\pmb { z } ) = p ( \\pmb { z } | \\pmb { x } ) p ( \\pmb { x } ) / p ( \\pmb { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 628, + 388, + 643 + ], + "score": 1.0, + "content": "is intractable since the true data distribution", + "type": "text" + }, + { + "bbox": [ + 388, + 630, + 408, + 641 + ], + "score": 0.9, + "content": "p ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 628, + 523, + 643 + ], + "score": 1.0, + "content": "is assumed to be unknown.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 639, + 521, + 655 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 271, + 655 + ], + "score": 1.0, + "content": "Similarly, we can derive lower bounds for", + "type": "text" + }, + { + "bbox": [ + 272, + 640, + 452, + 653 + ], + "score": 0.9, + "content": "I ( d ; z _ { p } ) \\geq H ( d ) + \\mathbb { E } _ { p ( d , z _ { p } ) } [ \\ln q ( d | z _ { p } ; \\psi ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 639, + 473, + 655 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 473, + 640, + 521, + 652 + ], + "score": 0.91, + "content": "I ( d ; z _ { s } ) \\geq", + "type": "inline_equation" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 89, + 607, + 523, + 655 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 206, + 73, + 411, + 229 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 206, + 73, + 411, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 73, + 411, + 229 + ], + "spans": [ + { + "bbox": [ + 206, + 73, + 411, + 229 + ], + "score": 0.972, + "type": "image", + "image_path": "8e3477bd318324bc4827d471de794c2c830ef3f1de87446abdf6c005292de940.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 73, + 411, + 86.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 206, + 86.0, + 411, + 99.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 206, + 99.0, + 411, + 112.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 206, + 112.0, + 411, + 125.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 206, + 125.0, + 411, + 138.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 206, + 138.0, + 411, + 151.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 206, + 151.0, + 411, + 164.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 206, + 164.0, + 411, + 177.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 206, + 177.0, + 411, + 190.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 206, + 190.0, + 411, + 203.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 206, + 203.0, + 411, + 216.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 206, + 216.0, + 411, + 229.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 237, + 522, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 237, + 522, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 237, + 224, + 249 + ], + "score": 1.0, + "content": "Figure 2: MDTA-ITA: The encoder", + "type": "text" + }, + { + "bbox": [ + 224, + 238, + 249, + 249 + ], + "score": 0.92, + "content": "E _ { s } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 237, + 379, + 249 + ], + "score": 1.0, + "content": "captures the feature representations", + "type": "text" + }, + { + "bbox": [ + 379, + 239, + 394, + 248 + ], + "score": 0.82, + "content": "( z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 237, + 483, + 249 + ], + "score": 1.0, + "content": "for a given input sample", + "type": "text" + }, + { + "bbox": [ + 484, + 240, + 491, + 247 + ], + "score": 0.76, + "content": "_ { \\pmb { x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 237, + 522, + 249 + ], + "score": 1.0, + "content": "that are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 248, + 522, + 259 + ], + "spans": [ + { + "bbox": [ + 90, + 248, + 181, + 259 + ], + "score": 1.0, + "content": "shared among domains.", + "type": "text" + }, + { + "bbox": [ + 181, + 248, + 207, + 259 + ], + "score": 0.92, + "content": "E _ { p } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 248, + 362, + 259 + ], + "score": 1.0, + "content": "captures domain–specific private features", + "type": "text" + }, + { + "bbox": [ + 363, + 248, + 378, + 258 + ], + "score": 0.82, + "content": "( z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 248, + 522, + 259 + ], + "score": 1.0, + "content": "using the shared private encoder. The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 258, + 522, + 269 + ], + "spans": [ + { + "bbox": [ + 90, + 258, + 145, + 269 + ], + "score": 1.0, + "content": "shared decoder", + "type": "text" + }, + { + "bbox": [ + 145, + 258, + 182, + 269 + ], + "score": 0.93, + "content": "F \\big ( z _ { p } , z _ { s } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 258, + 522, + 269 + ], + "score": 1.0, + "content": "learns to reconstruct the input sample by using both the private and shared features. The domain", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 267, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 90, + 267, + 124, + 279 + ], + "score": 1.0, + "content": "classifier", + "type": "text" + }, + { + "bbox": [ + 124, + 268, + 133, + 277 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 267, + 522, + 279 + ], + "score": 1.0, + "content": "learns to correctly predict the domain labels of the actual samples from both their shared and private features", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 277, + 410, + 289 + ], + "spans": [ + { + "bbox": [ + 89, + 277, + 159, + 289 + ], + "score": 1.0, + "content": "while the classifier", + "type": "text" + }, + { + "bbox": [ + 160, + 278, + 168, + 287 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 277, + 410, + 289 + ], + "score": 1.0, + "content": "learns to correctly predict the class labels from the shared features.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + } + ], + "index": 9.75 + }, + { + "type": "text", + "bbox": [ + 90, + 294, + 522, + 331 + ], + "lines": [ + { + "bbox": [ + 91, + 293, + 523, + 308 + ], + "spans": [ + { + "bbox": [ + 91, + 294, + 218, + 308 + ], + "score": 0.91, + "content": "H ( d ) + \\mathbb { E } _ { p ( d , z _ { s } ) } [ \\ln q ( d | z _ { s } ; \\psi ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 293, + 251, + 308 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 252, + 295, + 295, + 307 + ], + "score": 0.93, + "content": "q ( d | z _ { p } ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 293, + 487, + 308 + ], + "score": 1.0, + "content": "is any arbitrary distribution parameterized by", + "type": "text" + }, + { + "bbox": [ + 487, + 295, + 495, + 306 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 293, + 523, + 308 + ], + "score": 1.0, + "content": ". 1 We", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 306, + 522, + 321 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 206, + 321 + ], + "score": 1.0, + "content": "further drive lower bound for", + "type": "text" + }, + { + "bbox": [ + 207, + 307, + 241, + 319 + ], + "score": 0.93, + "content": "I ( \\pmb { y } ; \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 306, + 253, + 321 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 253, + 307, + 428, + 320 + ], + "score": 0.91, + "content": "I ( \\pmb { y } ; \\pmb { z _ { s } } ) \\geq H ( \\pmb { y } ) + \\mathbb { E } _ { p ( \\pmb { y } , \\pmb { z _ { s } } ) } [ \\ln q ( \\pmb { y } | \\pmb { z _ { s } } ; \\pmb { \\theta _ { c } } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 306, + 459, + 321 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 459, + 307, + 504, + 319 + ], + "score": 0.93, + "content": "q ( \\pmb { y } | \\pmb { z } _ { s } ; \\theta _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 504, + 306, + 522, + 321 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 318, + 363, + 333 + ], + "spans": [ + { + "bbox": [ + 89, + 318, + 254, + 333 + ], + "score": 1.0, + "content": "variational distribution parameterized by", + "type": "text" + }, + { + "bbox": [ + 254, + 320, + 264, + 330 + ], + "score": 0.87, + "content": "\\theta _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 318, + 326, + 333 + ], + "score": 1.0, + "content": "approximating", + "type": "text" + }, + { + "bbox": [ + 326, + 320, + 358, + 331 + ], + "score": 0.91, + "content": "p ( \\pmb { y } | \\boldsymbol { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 318, + 363, + 333 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 90, + 335, + 521, + 406 + ], + "lines": [ + { + "bbox": [ + 90, + 336, + 522, + 348 + ], + "spans": [ + { + "bbox": [ + 90, + 336, + 127, + 348 + ], + "score": 1.0, + "content": "Let next", + "type": "text" + }, + { + "bbox": [ + 128, + 336, + 168, + 348 + ], + "score": 0.92, + "content": "E _ { s } ( \\pmb { x } ; \\theta _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 336, + 301, + 348 + ], + "score": 1.0, + "content": "be a function parameterized by", + "type": "text" + }, + { + "bbox": [ + 302, + 336, + 312, + 347 + ], + "score": 0.87, + "content": "\\theta _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 336, + 396, + 348 + ], + "score": 1.0, + "content": "that maps a sample", + "type": "text" + }, + { + "bbox": [ + 397, + 338, + 404, + 346 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 336, + 522, + 348 + ], + "score": 1.0, + "content": "to its corresponding shared", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 347, + 522, + 360 + ], + "spans": [ + { + "bbox": [ + 90, + 347, + 120, + 359 + ], + "score": 1.0, + "content": "feature", + "type": "text" + }, + { + "bbox": [ + 121, + 349, + 132, + 358 + ], + "score": 0.86, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 347, + 153, + 359 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 153, + 347, + 194, + 359 + ], + "score": 0.93, + "content": "E _ { p } ( \\pmb { x } ; \\theta _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 347, + 351, + 359 + ], + "score": 1.0, + "content": "be an analogous function which maps", + "type": "text" + }, + { + "bbox": [ + 351, + 349, + 359, + 357 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 347, + 370, + 359 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 370, + 348, + 381, + 360 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 347, + 522, + 359 + ], + "score": 1.0, + "content": ", the feature that is private to each", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 357, + 523, + 372 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 216, + 372 + ], + "score": 1.0, + "content": "domain (fig. 2). We also define", + "type": "text" + }, + { + "bbox": [ + 217, + 358, + 268, + 370 + ], + "score": 0.93, + "content": "F ( z _ { s } , z _ { p } ; \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 357, + 523, + 372 + ], + "score": 1.0, + "content": "as a decoding function mapping the concatenation of the latent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 369, + 523, + 382 + ], + "spans": [ + { + "bbox": [ + 90, + 369, + 123, + 382 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 124, + 370, + 135, + 380 + ], + "score": 0.85, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 369, + 152, + 382 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 153, + 371, + 164, + 381 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 369, + 271, + 382 + ], + "score": 1.0, + "content": "to a sample reconstruction", + "type": "text" + }, + { + "bbox": [ + 271, + 370, + 279, + 379 + ], + "score": 0.82, + "content": "\\hat { \\pmb x }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 369, + 299, + 382 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 299, + 369, + 334, + 380 + ], + "score": 0.94, + "content": "D ( z ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 369, + 462, + 382 + ], + "score": 1.0, + "content": "as a decoding function mapping", + "type": "text" + }, + { + "bbox": [ + 462, + 370, + 474, + 380 + ], + "score": 0.85, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 369, + 491, + 382 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 491, + 370, + 503, + 381 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 369, + 523, + 382 + ], + "score": 1.0, + "content": "to a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 91, + 380, + 521, + 394 + ], + "spans": [ + { + "bbox": [ + 91, + 382, + 102, + 392 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 102, + 381, + 330, + 394 + ], + "score": 1.0, + "content": "-dimensional vector: the predictions of the domain label", + "type": "text" + }, + { + "bbox": [ + 330, + 380, + 338, + 392 + ], + "score": 0.81, + "content": "\\hat { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 381, + 375, + 394 + ], + "score": 1.0, + "content": ". Finally,", + "type": "text" + }, + { + "bbox": [ + 375, + 381, + 414, + 394 + ], + "score": 0.94, + "content": "C ( z _ { s } ; \\theta _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 381, + 521, + 394 + ], + "score": 1.0, + "content": "is a task-specific function", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 393, + 375, + 406 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 127, + 406 + ], + "score": 1.0, + "content": "mapping", + "type": "text" + }, + { + "bbox": [ + 128, + 395, + 138, + 404 + ], + "score": 0.85, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 393, + 157, + 406 + ], + "score": 1.0, + "content": "to a", + "type": "text" + }, + { + "bbox": [ + 157, + 394, + 167, + 403 + ], + "score": 0.85, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 393, + 363, + 406 + ], + "score": 1.0, + "content": "-dimensional probability vector of the class label", + "type": "text" + }, + { + "bbox": [ + 363, + 394, + 371, + 405 + ], + "score": 0.84, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 393, + 375, + 406 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 90, + 409, + 522, + 462 + ], + "lines": [ + { + "bbox": [ + 89, + 408, + 520, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 408, + 151, + 423 + ], + "score": 1.0, + "content": "We represent", + "type": "text" + }, + { + "bbox": [ + 91, + 421, + 174, + 436 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { \\bar { N } } \\delta ( { \\bf d } - { \\bf d } _ { i } ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 410, + 218, + 422 + ], + "score": 0.92, + "content": "p ( { \\pmb d } ) , p ( { \\pmb x } ) , p ( { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 414, + 507, + 440 + ], + "score": 1.0, + "content": "as in the case of variational autoencoders (VAE) (Abbasnejad et al. (2017);", + "type": "text" + }, + { + "bbox": [ + 219, + 408, + 483, + 423 + ], + "score": 1.0, + "content": "as the empirical distribution of a finite training set (e.g.", + "type": "text" + }, + { + "bbox": [ + 484, + 410, + 520, + 423 + ], + "score": 0.88, + "content": "\\begin{array} { r l } { p ( d ) } & { { } = } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 507, + 424, + 520, + 434 + ], + "score": 0.27, + "content": "\\mathrm { P u }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 435, + 522, + 449 + ], + "spans": [ + { + "bbox": [ + 88, + 435, + 153, + 449 + ], + "score": 1.0, + "content": "et al. (2017)),", + "type": "text" + }, + { + "bbox": [ + 153, + 436, + 222, + 448 + ], + "score": 0.92, + "content": "p ( \\pmb { z } _ { s } | \\pmb { x } ) , p ( \\pmb { z } _ { p } | \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 435, + 349, + 449 + ], + "score": 1.0, + "content": "as deterministic functions of", + "type": "text" + }, + { + "bbox": [ + 349, + 438, + 357, + 447 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 435, + 372, + 449 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 372, + 435, + 501, + 449 + ], + "score": 0.92, + "content": "p ( z _ { s } | \\pmb { x } ) = \\delta \\big ( z _ { s } - E _ { s } ( \\pmb { x } ; \\theta _ { s } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 435, + 522, + 449 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 448, + 475, + 462 + ], + "spans": [ + { + "bbox": [ + 90, + 449, + 212, + 462 + ], + "score": 0.9, + "content": "p ( z _ { p } | \\pmb { x } ) = \\delta \\big ( z _ { p } - E _ { p } ( \\pmb { x } ; \\theta _ { p } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 448, + 345, + 462 + ], + "score": 1.0, + "content": ", and the variational distributions", + "type": "text" + }, + { + "bbox": [ + 346, + 449, + 378, + 462 + ], + "score": 0.58, + "content": "q ( \\pmb { y } | \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 448, + 382, + 462 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 382, + 449, + 411, + 462 + ], + "score": 0.56, + "content": "q ( { \\pmb x } | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 448, + 434, + 462 + ], + "score": 1.0, + "content": "and ,", + "type": "text" + }, + { + "bbox": [ + 434, + 450, + 462, + 461 + ], + "score": 0.93, + "content": "q ( d | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 448, + 475, + 462 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 98, + 467, + 500, + 481 + ], + "lines": [ + { + "bbox": [ + 98, + 467, + 500, + 481 + ], + "spans": [ + { + "bbox": [ + 98, + 467, + 500, + 481 + ], + "score": 0.88, + "content": "q ( \\boldsymbol { y } | \\boldsymbol { z } _ { s } ) = \\mathrm { S o f t M a x } ( C ( \\boldsymbol { z } _ { s } ; \\boldsymbol { \\theta } _ { c } ) ) , q ( d | \\boldsymbol { z } ) = \\mathrm { S o f t M a x } ( D ( \\boldsymbol { z } ; \\boldsymbol { \\psi } ) ) , q ( \\boldsymbol { x } | \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\propto \\exp ( \\| \\boldsymbol { x } - \\boldsymbol { F } ( \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\| _ { 1 } )", + "type": "interline_equation", + "image_path": "8e79b2c4472bab8e005ba23f78447ebc5b45bb38aeee0613256586456be89703.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 98, + 467, + 500, + 481 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 492, + 522, + 537 + ], + "lines": [ + { + "bbox": [ + 89, + 492, + 522, + 506 + ], + "spans": [ + { + "bbox": [ + 89, + 492, + 151, + 506 + ], + "score": 1.0, + "content": "where Softmax", + "type": "text" + }, + { + "bbox": [ + 151, + 492, + 162, + 504 + ], + "score": 0.39, + "content": "( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 492, + 469, + 506 + ], + "score": 1.0, + "content": "denotes the softmax or normalized exponential function (Bridle (1990)), and", + "type": "text" + }, + { + "bbox": [ + 469, + 492, + 487, + 505 + ], + "score": 0.91, + "content": "\\left. . \\right. _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 492, + 522, + 506 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 503, + 521, + 516 + ], + "spans": [ + { + "bbox": [ + 89, + 503, + 105, + 516 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 105, + 504, + 118, + 515 + ], + "score": 0.88, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 503, + 521, + 516 + ], + "score": 1.0, + "content": "norm. Then, the optimization task can be posed as a minimax saddle point problem, where we use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 514, + 522, + 527 + ], + "spans": [ + { + "bbox": [ + 89, + 514, + 313, + 527 + ], + "score": 1.0, + "content": "adversarial training to maximize (2) w.r.t. the parameters", + "type": "text" + }, + { + "bbox": [ + 314, + 515, + 357, + 527 + ], + "score": 0.93, + "content": "( \\theta _ { s } , \\theta _ { p } , \\theta _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 514, + 522, + 527 + ], + "score": 1.0, + "content": ", and to minimize (2) w.r.t. the parameters", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 90, + 525, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 90, + 526, + 113, + 537 + ], + "score": 0.89, + "content": "( \\phi , \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 525, + 288, + 538 + ], + "score": 1.0, + "content": ", using Stochastic Gradient Descent (SGD).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 91, + 542, + 288, + 554 + ], + "lines": [ + { + "bbox": [ + 89, + 542, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 89, + 542, + 208, + 555 + ], + "score": 1.0, + "content": "Optimizing the parameters", + "type": "text" + }, + { + "bbox": [ + 208, + 543, + 215, + 554 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 542, + 279, + 555 + ], + "score": 1.0, + "content": "of the decoder", + "type": "text" + }, + { + "bbox": [ + 279, + 543, + 288, + 552 + ], + "score": 0.81, + "content": "F", + "type": "inline_equation" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 559, + 423, + 594 + ], + "lines": [ + { + "bbox": [ + 188, + 559, + 423, + 594 + ], + "spans": [ + { + "bbox": [ + 188, + 559, + 423, + 594 + ], + "score": 0.94, + "content": "\\hat { \\phi } = \\underset { \\phi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { F } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\pmb { x } _ { i } - F \\big ( E _ { s } ( \\pmb { x } _ { i } ) , E _ { p } ( \\pmb { x } _ { i } ) \\big ) \\| _ { 1 } .", + "type": "interline_equation", + "image_path": "1668b7a1b50aa39a5963118532a442c2a3e05b841494ce364eb9929b9bed6412.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 559, + 423, + 576.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 188, + 576.5, + 423, + 594.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 599, + 520, + 623 + ], + "lines": [ + { + "bbox": [ + 89, + 598, + 520, + 612 + ], + "spans": [ + { + "bbox": [ + 89, + 598, + 142, + 612 + ], + "score": 1.0, + "content": "The decoder", + "type": "text" + }, + { + "bbox": [ + 142, + 599, + 193, + 612 + ], + "score": 0.93, + "content": "F ( z _ { s } , z _ { p } ; \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 598, + 512, + 612 + ], + "score": 1.0, + "content": "is trained in such a way so as to minimize the difference between original input", + "type": "text" + }, + { + "bbox": [ + 513, + 602, + 520, + 609 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 89, + 610, + 425, + 622 + ], + "spans": [ + { + "bbox": [ + 89, + 610, + 412, + 622 + ], + "score": 1.0, + "content": "and its decoding from corresponding shared and private features via the decoder", + "type": "text" + }, + { + "bbox": [ + 412, + 611, + 421, + 621 + ], + "score": 0.84, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 610, + 425, + 622 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 89, + 631, + 520, + 652 + ], + "lines": [ + { + "bbox": [ + 102, + 628, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 102, + 628, + 286, + 643 + ], + "score": 1.0, + "content": "1Note that, for simplicity, we shared the parameters", + "type": "text" + }, + { + "bbox": [ + 286, + 633, + 293, + 641 + ], + "score": 0.89, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 628, + 465, + 643 + ], + "score": 1.0, + "content": "between the approximate posterior distributions", + "type": "text" + }, + { + "bbox": [ + 465, + 631, + 505, + 642 + ], + "score": 0.92, + "content": "q ( d | z _ { s } , \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 505, + 628, + 522, + 643 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ] + }, + { + "bbox": [ + 91, + 639, + 136, + 654 + ], + "spans": [ + { + "bbox": [ + 91, + 641, + 131, + 652 + ], + "score": 0.89, + "content": "q ( d | z _ { p } ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 639, + 136, + 654 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 292, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 670, + 294, + 682 + ], + "spans": [ + { + "bbox": [ + 285, + 670, + 294, + 682 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 206, + 73, + 411, + 229 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 206, + 73, + 411, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 73, + 411, + 229 + ], + "spans": [ + { + "bbox": [ + 206, + 73, + 411, + 229 + ], + "score": 0.972, + "type": "image", + "image_path": "8e3477bd318324bc4827d471de794c2c830ef3f1de87446abdf6c005292de940.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 73, + 411, + 86.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 206, + 86.0, + 411, + 99.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 206, + 99.0, + 411, + 112.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 206, + 112.0, + 411, + 125.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 206, + 125.0, + 411, + 138.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 206, + 138.0, + 411, + 151.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 206, + 151.0, + 411, + 164.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 206, + 164.0, + 411, + 177.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 206, + 177.0, + 411, + 190.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 206, + 190.0, + 411, + 203.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 206, + 203.0, + 411, + 216.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 206, + 216.0, + 411, + 229.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 237, + 522, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 237, + 522, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 237, + 224, + 249 + ], + "score": 1.0, + "content": "Figure 2: MDTA-ITA: The encoder", + "type": "text" + }, + { + "bbox": [ + 224, + 238, + 249, + 249 + ], + "score": 0.92, + "content": "E _ { s } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 237, + 379, + 249 + ], + "score": 1.0, + "content": "captures the feature representations", + "type": "text" + }, + { + "bbox": [ + 379, + 239, + 394, + 248 + ], + "score": 0.82, + "content": "( z _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 237, + 483, + 249 + ], + "score": 1.0, + "content": "for a given input sample", + "type": "text" + }, + { + "bbox": [ + 484, + 240, + 491, + 247 + ], + "score": 0.76, + "content": "_ { \\pmb { x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 237, + 522, + 249 + ], + "score": 1.0, + "content": "that are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 248, + 522, + 259 + ], + "spans": [ + { + "bbox": [ + 90, + 248, + 181, + 259 + ], + "score": 1.0, + "content": "shared among domains.", + "type": "text" + }, + { + "bbox": [ + 181, + 248, + 207, + 259 + ], + "score": 0.92, + "content": "E _ { p } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 248, + 362, + 259 + ], + "score": 1.0, + "content": "captures domain–specific private features", + "type": "text" + }, + { + "bbox": [ + 363, + 248, + 378, + 258 + ], + "score": 0.82, + "content": "( z _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 248, + 522, + 259 + ], + "score": 1.0, + "content": "using the shared private encoder. The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 258, + 522, + 269 + ], + "spans": [ + { + "bbox": [ + 90, + 258, + 145, + 269 + ], + "score": 1.0, + "content": "shared decoder", + "type": "text" + }, + { + "bbox": [ + 145, + 258, + 182, + 269 + ], + "score": 0.93, + "content": "F \\big ( z _ { p } , z _ { s } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 258, + 522, + 269 + ], + "score": 1.0, + "content": "learns to reconstruct the input sample by using both the private and shared features. The domain", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 267, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 90, + 267, + 124, + 279 + ], + "score": 1.0, + "content": "classifier", + "type": "text" + }, + { + "bbox": [ + 124, + 268, + 133, + 277 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 267, + 522, + 279 + ], + "score": 1.0, + "content": "learns to correctly predict the domain labels of the actual samples from both their shared and private features", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 277, + 410, + 289 + ], + "spans": [ + { + "bbox": [ + 89, + 277, + 159, + 289 + ], + "score": 1.0, + "content": "while the classifier", + "type": "text" + }, + { + "bbox": [ + 160, + 278, + 168, + 287 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 277, + 410, + 289 + ], + "score": 1.0, + "content": "learns to correctly predict the class labels from the shared features.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + } + ], + "index": 9.75 + }, + { + "type": "text", + "bbox": [ + 90, + 294, + 522, + 331 + ], + "lines": [ + { + "bbox": [ + 91, + 293, + 523, + 308 + ], + "spans": [ + { + "bbox": [ + 91, + 294, + 218, + 308 + ], + "score": 0.91, + "content": "H ( d ) + \\mathbb { E } _ { p ( d , z _ { s } ) } [ \\ln q ( d | z _ { s } ; \\psi ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 293, + 251, + 308 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 252, + 295, + 295, + 307 + ], + "score": 0.93, + "content": "q ( d | z _ { p } ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 293, + 487, + 308 + ], + "score": 1.0, + "content": "is any arbitrary distribution parameterized by", + "type": "text" + }, + { + "bbox": [ + 487, + 295, + 495, + 306 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 293, + 523, + 308 + ], + "score": 1.0, + "content": ". 1 We", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 306, + 522, + 321 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 206, + 321 + ], + "score": 1.0, + "content": "further drive lower bound for", + "type": "text" + }, + { + "bbox": [ + 207, + 307, + 241, + 319 + ], + "score": 0.93, + "content": "I ( \\pmb { y } ; \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 306, + 253, + 321 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 253, + 307, + 428, + 320 + ], + "score": 0.91, + "content": "I ( \\pmb { y } ; \\pmb { z _ { s } } ) \\geq H ( \\pmb { y } ) + \\mathbb { E } _ { p ( \\pmb { y } , \\pmb { z _ { s } } ) } [ \\ln q ( \\pmb { y } | \\pmb { z _ { s } } ; \\pmb { \\theta _ { c } } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 306, + 459, + 321 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 459, + 307, + 504, + 319 + ], + "score": 0.93, + "content": "q ( \\pmb { y } | \\pmb { z } _ { s } ; \\theta _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 504, + 306, + 522, + 321 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 318, + 363, + 333 + ], + "spans": [ + { + "bbox": [ + 89, + 318, + 254, + 333 + ], + "score": 1.0, + "content": "variational distribution parameterized by", + "type": "text" + }, + { + "bbox": [ + 254, + 320, + 264, + 330 + ], + "score": 0.87, + "content": "\\theta _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 318, + 326, + 333 + ], + "score": 1.0, + "content": "approximating", + "type": "text" + }, + { + "bbox": [ + 326, + 320, + 358, + 331 + ], + "score": 0.91, + "content": "p ( \\pmb { y } | \\boldsymbol { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 318, + 363, + 333 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 89, + 293, + 523, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 335, + 521, + 406 + ], + "lines": [ + { + "bbox": [ + 90, + 336, + 522, + 348 + ], + "spans": [ + { + "bbox": [ + 90, + 336, + 127, + 348 + ], + "score": 1.0, + "content": "Let next", + "type": "text" + }, + { + "bbox": [ + 128, + 336, + 168, + 348 + ], + "score": 0.92, + "content": "E _ { s } ( \\pmb { x } ; \\theta _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 336, + 301, + 348 + ], + "score": 1.0, + "content": "be a function parameterized by", + "type": "text" + }, + { + "bbox": [ + 302, + 336, + 312, + 347 + ], + "score": 0.87, + "content": "\\theta _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 336, + 396, + 348 + ], + "score": 1.0, + "content": "that maps a sample", + "type": "text" + }, + { + "bbox": [ + 397, + 338, + 404, + 346 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 336, + 522, + 348 + ], + "score": 1.0, + "content": "to its corresponding shared", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 347, + 522, + 360 + ], + "spans": [ + { + "bbox": [ + 90, + 347, + 120, + 359 + ], + "score": 1.0, + "content": "feature", + "type": "text" + }, + { + "bbox": [ + 121, + 349, + 132, + 358 + ], + "score": 0.86, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 347, + 153, + 359 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 153, + 347, + 194, + 359 + ], + "score": 0.93, + "content": "E _ { p } ( \\pmb { x } ; \\theta _ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 347, + 351, + 359 + ], + "score": 1.0, + "content": "be an analogous function which maps", + "type": "text" + }, + { + "bbox": [ + 351, + 349, + 359, + 357 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 347, + 370, + 359 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 370, + 348, + 381, + 360 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 347, + 522, + 359 + ], + "score": 1.0, + "content": ", the feature that is private to each", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 357, + 523, + 372 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 216, + 372 + ], + "score": 1.0, + "content": "domain (fig. 2). We also define", + "type": "text" + }, + { + "bbox": [ + 217, + 358, + 268, + 370 + ], + "score": 0.93, + "content": "F ( z _ { s } , z _ { p } ; \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 357, + 523, + 372 + ], + "score": 1.0, + "content": "as a decoding function mapping the concatenation of the latent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 369, + 523, + 382 + ], + "spans": [ + { + "bbox": [ + 90, + 369, + 123, + 382 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 124, + 370, + 135, + 380 + ], + "score": 0.85, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 369, + 152, + 382 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 153, + 371, + 164, + 381 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 369, + 271, + 382 + ], + "score": 1.0, + "content": "to a sample reconstruction", + "type": "text" + }, + { + "bbox": [ + 271, + 370, + 279, + 379 + ], + "score": 0.82, + "content": "\\hat { \\pmb x }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 369, + 299, + 382 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 299, + 369, + 334, + 380 + ], + "score": 0.94, + "content": "D ( z ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 369, + 462, + 382 + ], + "score": 1.0, + "content": "as a decoding function mapping", + "type": "text" + }, + { + "bbox": [ + 462, + 370, + 474, + 380 + ], + "score": 0.85, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 369, + 491, + 382 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 491, + 370, + 503, + 381 + ], + "score": 0.87, + "content": "z _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 369, + 523, + 382 + ], + "score": 1.0, + "content": "to a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 91, + 380, + 521, + 394 + ], + "spans": [ + { + "bbox": [ + 91, + 382, + 102, + 392 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 102, + 381, + 330, + 394 + ], + "score": 1.0, + "content": "-dimensional vector: the predictions of the domain label", + "type": "text" + }, + { + "bbox": [ + 330, + 380, + 338, + 392 + ], + "score": 0.81, + "content": "\\hat { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 381, + 375, + 394 + ], + "score": 1.0, + "content": ". Finally,", + "type": "text" + }, + { + "bbox": [ + 375, + 381, + 414, + 394 + ], + "score": 0.94, + "content": "C ( z _ { s } ; \\theta _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 381, + 521, + 394 + ], + "score": 1.0, + "content": "is a task-specific function", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 393, + 375, + 406 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 127, + 406 + ], + "score": 1.0, + "content": "mapping", + "type": "text" + }, + { + "bbox": [ + 128, + 395, + 138, + 404 + ], + "score": 0.85, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 393, + 157, + 406 + ], + "score": 1.0, + "content": "to a", + "type": "text" + }, + { + "bbox": [ + 157, + 394, + 167, + 403 + ], + "score": 0.85, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 393, + 363, + 406 + ], + "score": 1.0, + "content": "-dimensional probability vector of the class label", + "type": "text" + }, + { + "bbox": [ + 363, + 394, + 371, + 405 + ], + "score": 0.84, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 393, + 375, + 406 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 89, + 336, + 523, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 409, + 522, + 462 + ], + "lines": [ + { + "bbox": [ + 89, + 408, + 520, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 408, + 151, + 423 + ], + "score": 1.0, + "content": "We represent", + "type": "text" + }, + { + "bbox": [ + 91, + 421, + 174, + 436 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { \\bar { N } } \\delta ( { \\bf d } - { \\bf d } _ { i } ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 410, + 218, + 422 + ], + "score": 0.92, + "content": "p ( { \\pmb d } ) , p ( { \\pmb x } ) , p ( { \\pmb y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 414, + 507, + 440 + ], + "score": 1.0, + "content": "as in the case of variational autoencoders (VAE) (Abbasnejad et al. (2017);", + "type": "text" + }, + { + "bbox": [ + 219, + 408, + 483, + 423 + ], + "score": 1.0, + "content": "as the empirical distribution of a finite training set (e.g.", + "type": "text" + }, + { + "bbox": [ + 484, + 410, + 520, + 423 + ], + "score": 0.88, + "content": "\\begin{array} { r l } { p ( d ) } & { { } = } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 507, + 424, + 520, + 434 + ], + "score": 0.27, + "content": "\\mathrm { P u }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 435, + 522, + 449 + ], + "spans": [ + { + "bbox": [ + 88, + 435, + 153, + 449 + ], + "score": 1.0, + "content": "et al. (2017)),", + "type": "text" + }, + { + "bbox": [ + 153, + 436, + 222, + 448 + ], + "score": 0.92, + "content": "p ( \\pmb { z } _ { s } | \\pmb { x } ) , p ( \\pmb { z } _ { p } | \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 435, + 349, + 449 + ], + "score": 1.0, + "content": "as deterministic functions of", + "type": "text" + }, + { + "bbox": [ + 349, + 438, + 357, + 447 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 435, + 372, + 449 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 372, + 435, + 501, + 449 + ], + "score": 0.92, + "content": "p ( z _ { s } | \\pmb { x } ) = \\delta \\big ( z _ { s } - E _ { s } ( \\pmb { x } ; \\theta _ { s } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 435, + 522, + 449 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 448, + 475, + 462 + ], + "spans": [ + { + "bbox": [ + 90, + 449, + 212, + 462 + ], + "score": 0.9, + "content": "p ( z _ { p } | \\pmb { x } ) = \\delta \\big ( z _ { p } - E _ { p } ( \\pmb { x } ; \\theta _ { p } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 448, + 345, + 462 + ], + "score": 1.0, + "content": ", and the variational distributions", + "type": "text" + }, + { + "bbox": [ + 346, + 449, + 378, + 462 + ], + "score": 0.58, + "content": "q ( \\pmb { y } | \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 448, + 382, + 462 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 382, + 449, + 411, + 462 + ], + "score": 0.56, + "content": "q ( { \\pmb x } | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 448, + 434, + 462 + ], + "score": 1.0, + "content": "and ,", + "type": "text" + }, + { + "bbox": [ + 434, + 450, + 462, + 461 + ], + "score": 0.93, + "content": "q ( d | z )", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 448, + 475, + 462 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 88, + 408, + 522, + 462 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 98, + 467, + 500, + 481 + ], + "lines": [ + { + "bbox": [ + 98, + 467, + 500, + 481 + ], + "spans": [ + { + "bbox": [ + 98, + 467, + 500, + 481 + ], + "score": 0.88, + "content": "q ( \\boldsymbol { y } | \\boldsymbol { z } _ { s } ) = \\mathrm { S o f t M a x } ( C ( \\boldsymbol { z } _ { s } ; \\boldsymbol { \\theta } _ { c } ) ) , q ( d | \\boldsymbol { z } ) = \\mathrm { S o f t M a x } ( D ( \\boldsymbol { z } ; \\boldsymbol { \\psi } ) ) , q ( \\boldsymbol { x } | \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\propto \\exp ( \\| \\boldsymbol { x } - \\boldsymbol { F } ( \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\| _ { 1 } )", + "type": "interline_equation", + "image_path": "8e79b2c4472bab8e005ba23f78447ebc5b45bb38aeee0613256586456be89703.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 98, + 467, + 500, + 481 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 492, + 522, + 537 + ], + "lines": [ + { + "bbox": [ + 89, + 492, + 522, + 506 + ], + "spans": [ + { + "bbox": [ + 89, + 492, + 151, + 506 + ], + "score": 1.0, + "content": "where Softmax", + "type": "text" + }, + { + "bbox": [ + 151, + 492, + 162, + 504 + ], + "score": 0.39, + "content": "( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 492, + 469, + 506 + ], + "score": 1.0, + "content": "denotes the softmax or normalized exponential function (Bridle (1990)), and", + "type": "text" + }, + { + "bbox": [ + 469, + 492, + 487, + 505 + ], + "score": 0.91, + "content": "\\left. . \\right. _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 492, + 522, + 506 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 503, + 521, + 516 + ], + "spans": [ + { + "bbox": [ + 89, + 503, + 105, + 516 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 105, + 504, + 118, + 515 + ], + "score": 0.88, + "content": "L _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 503, + 521, + 516 + ], + "score": 1.0, + "content": "norm. Then, the optimization task can be posed as a minimax saddle point problem, where we use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 514, + 522, + 527 + ], + "spans": [ + { + "bbox": [ + 89, + 514, + 313, + 527 + ], + "score": 1.0, + "content": "adversarial training to maximize (2) w.r.t. the parameters", + "type": "text" + }, + { + "bbox": [ + 314, + 515, + 357, + 527 + ], + "score": 0.93, + "content": "( \\theta _ { s } , \\theta _ { p } , \\theta _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 514, + 522, + 527 + ], + "score": 1.0, + "content": ", and to minimize (2) w.r.t. the parameters", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 90, + 525, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 90, + 526, + 113, + 537 + ], + "score": 0.89, + "content": "( \\phi , \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 525, + 288, + 538 + ], + "score": 1.0, + "content": ", using Stochastic Gradient Descent (SGD).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 89, + 492, + 522, + 538 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 542, + 288, + 554 + ], + "lines": [ + { + "bbox": [ + 89, + 542, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 89, + 542, + 208, + 555 + ], + "score": 1.0, + "content": "Optimizing the parameters", + "type": "text" + }, + { + "bbox": [ + 208, + 543, + 215, + 554 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 542, + 279, + 555 + ], + "score": 1.0, + "content": "of the decoder", + "type": "text" + }, + { + "bbox": [ + 279, + 543, + 288, + 552 + ], + "score": 0.81, + "content": "F", + "type": "inline_equation" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 559, + 423, + 594 + ], + "lines": [ + { + "bbox": [ + 188, + 559, + 423, + 594 + ], + "spans": [ + { + "bbox": [ + 188, + 559, + 423, + 594 + ], + "score": 0.94, + "content": "\\hat { \\phi } = \\underset { \\phi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { F } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\pmb { x } _ { i } - F \\big ( E _ { s } ( \\pmb { x } _ { i } ) , E _ { p } ( \\pmb { x } _ { i } ) \\big ) \\| _ { 1 } .", + "type": "interline_equation", + "image_path": "1668b7a1b50aa39a5963118532a442c2a3e05b841494ce364eb9929b9bed6412.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 559, + 423, + 576.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 188, + 576.5, + 423, + 594.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 599, + 520, + 623 + ], + "lines": [ + { + "bbox": [ + 89, + 598, + 520, + 612 + ], + "spans": [ + { + "bbox": [ + 89, + 598, + 142, + 612 + ], + "score": 1.0, + "content": "The decoder", + "type": "text" + }, + { + "bbox": [ + 142, + 599, + 193, + 612 + ], + "score": 0.93, + "content": "F ( z _ { s } , z _ { p } ; \\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 598, + 512, + 612 + ], + "score": 1.0, + "content": "is trained in such a way so as to minimize the difference between original input", + "type": "text" + }, + { + "bbox": [ + 513, + 602, + 520, + 609 + ], + "score": 0.73, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 89, + 610, + 425, + 622 + ], + "spans": [ + { + "bbox": [ + 89, + 610, + 412, + 622 + ], + "score": 1.0, + "content": "and its decoding from corresponding shared and private features via the decoder", + "type": "text" + }, + { + "bbox": [ + 412, + 611, + 421, + 621 + ], + "score": 0.84, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 610, + 425, + 622 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 89, + 598, + 520, + 622 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 91, + 95, + 329, + 108 + ], + "lines": [ + { + "bbox": [ + 90, + 95, + 328, + 109 + ], + "spans": [ + { + "bbox": [ + 90, + 95, + 208, + 109 + ], + "score": 1.0, + "content": "Optimizing the parameters", + "type": "text" + }, + { + "bbox": [ + 208, + 97, + 216, + 108 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 95, + 318, + 109 + ], + "score": 1.0, + "content": "of the domain classifier", + "type": "text" + }, + { + "bbox": [ + 318, + 97, + 328, + 106 + ], + "score": 0.77, + "content": "D", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 110, + 462, + 145 + ], + "lines": [ + { + "bbox": [ + 149, + 110, + 462, + 145 + ], + "spans": [ + { + "bbox": [ + 149, + 110, + 462, + 145 + ], + "score": 0.94, + "content": "\\hat { \\psi } = \\underset { \\psi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { D } = - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\pmb { x } _ { i } ) \\big ) - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { p } ( \\pmb { x } _ { i } ) \\big ) .", + "type": "interline_equation", + "image_path": "21e90163bfe21babae9cdc9f8986f7fd1d54c2568bd6a5fd21c5c4b3297d1048.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 149, + 110, + 462, + 121.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 149, + 121.66666666666667, + 462, + 133.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 149, + 133.33333333333334, + 462, + 145.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 149, + 521, + 183 + ], + "lines": [ + { + "bbox": [ + 91, + 147, + 521, + 161 + ], + "spans": [ + { + "bbox": [ + 91, + 149, + 125, + 161 + ], + "score": 0.91, + "content": "D ( z ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 147, + 521, + 161 + ], + "score": 1.0, + "content": "can be considered as a classifier whose task is to distinguish between the shared/private features of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 160, + 520, + 172 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 383, + 172 + ], + "score": 1.0, + "content": "the different domains. More precisely, the two terms in Eq. 6 encourage", + "type": "text" + }, + { + "bbox": [ + 383, + 160, + 393, + 170 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 160, + 520, + 172 + ], + "score": 1.0, + "content": "to correctly predict the domain", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 169, + 316, + 184 + ], + "spans": [ + { + "bbox": [ + 88, + 169, + 316, + 184 + ], + "score": 1.0, + "content": "labels from the shared and private features, respectively.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 90, + 187, + 318, + 200 + ], + "lines": [ + { + "bbox": [ + 90, + 187, + 317, + 201 + ], + "spans": [ + { + "bbox": [ + 90, + 187, + 208, + 201 + ], + "score": 1.0, + "content": "Optimizing the parameters", + "type": "text" + }, + { + "bbox": [ + 208, + 188, + 218, + 199 + ], + "score": 0.88, + "content": "\\theta _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 187, + 308, + 201 + ], + "score": 1.0, + "content": "of the label classifier", + "type": "text" + }, + { + "bbox": [ + 308, + 189, + 317, + 198 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 202, + 203, + 409, + 226 + ], + "lines": [ + { + "bbox": [ + 202, + 203, + 409, + 226 + ], + "spans": [ + { + "bbox": [ + 202, + 203, + 409, + 226 + ], + "score": 0.9, + "content": "\\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\big \\{ - H ( \\pmb { y } ) - \\mathbb { E } _ { p ( \\pmb { y } , z _ { s } ) } \\big [ \\ln q ( \\pmb { y } | z _ { s } ) \\big ] \\big \\} .", + "type": "interline_equation", + "image_path": "8efe30add848b2c029835c54b2d4e7f3d10df2c29d3d034dfe1abb77b7bca699.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 202, + 203, + 409, + 226 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 92, + 230, + 521, + 253 + ], + "lines": [ + { + "bbox": [ + 89, + 228, + 520, + 244 + ], + "spans": [ + { + "bbox": [ + 89, + 228, + 258, + 244 + ], + "score": 1.0, + "content": "Since we have access to the source labels,", + "type": "text" + }, + { + "bbox": [ + 259, + 230, + 283, + 242 + ], + "score": 0.92, + "content": "H ( \\pmb { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 228, + 498, + 244 + ], + "score": 1.0, + "content": "is a constant for source samples. we can approximate", + "type": "text" + }, + { + "bbox": [ + 498, + 230, + 520, + 242 + ], + "score": 0.9, + "content": "H [ \\pmb { y } ]", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 240, + 510, + 254 + ], + "spans": [ + { + "bbox": [ + 89, + 240, + 309, + 254 + ], + "score": 1.0, + "content": "for the target samples using the output of the classifier", + "type": "text" + }, + { + "bbox": [ + 309, + 242, + 317, + 251 + ], + "score": 0.83, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 240, + 510, + 254 + ], + "score": 1.0, + "content": ", leading to the following optimization problem:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 256, + 499, + 329 + ], + "lines": [ + { + "bbox": [ + 112, + 256, + 499, + 329 + ], + "spans": [ + { + "bbox": [ + 112, + 256, + 499, + 329 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { C } = - \\ \\frac { 1 } { N } \\displaystyle \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) } \\\\ & { \\qquad + \\ \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\displaystyle \\frac { 1 } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) \\bigg ) , } \\end{array}", + "type": "interline_equation", + "image_path": "7dadae5cfa60703855a0e15efbcd5d3ff98636b04efe028d1e689cf41d3c6c58.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 112, + 256, + 499, + 280.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 112, + 280.3333333333333, + 499, + 304.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 112, + 304.66666666666663, + 499, + 328.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 330, + 522, + 420 + ], + "lines": [ + { + "bbox": [ + 90, + 331, + 522, + 344 + ], + "spans": [ + { + "bbox": [ + 90, + 331, + 117, + 344 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 117, + 332, + 130, + 343 + ], + "score": 0.88, + "content": "N _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 331, + 433, + 344 + ], + "score": 1.0, + "content": "denotes the number of source samples. Intuitively, we enforce the classifier", + "type": "text" + }, + { + "bbox": [ + 433, + 332, + 442, + 341 + ], + "score": 0.83, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 331, + 522, + 344 + ], + "score": 1.0, + "content": "to correctly predict", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 342, + 522, + 355 + ], + "spans": [ + { + "bbox": [ + 90, + 342, + 522, + 355 + ], + "score": 1.0, + "content": "the class labels of the source samples by the first term in Eq. 8. We use the second term to minimize the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 353, + 522, + 366 + ], + "spans": [ + { + "bbox": [ + 90, + 354, + 135, + 366 + ], + "score": 1.0, + "content": "entropy of", + "type": "text" + }, + { + "bbox": [ + 136, + 353, + 168, + 365 + ], + "score": 0.92, + "content": "q ( \\pmb { y } | \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 354, + 522, + 366 + ], + "score": 1.0, + "content": "for the target samples; effectively, reducing the effects of \"confusing\" labels of target", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 364, + 522, + 377 + ], + "spans": [ + { + "bbox": [ + 89, + 364, + 174, + 377 + ], + "score": 1.0, + "content": "samples, as given by", + "type": "text" + }, + { + "bbox": [ + 174, + 365, + 206, + 376 + ], + "score": 0.92, + "content": "p ( \\pmb { y } | \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 364, + 522, + 377 + ], + "score": 1.0, + "content": "that leads to decision boundaries occur far away from target data-dense regions", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 88, + 374, + 522, + 388 + ], + "spans": [ + { + "bbox": [ + 88, + 374, + 522, + 388 + ], + "score": 1.0, + "content": "in the feature space. The intuition behind the last term is that by minimizing only the entropy (second term),", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 387, + 522, + 399 + ], + "spans": [ + { + "bbox": [ + 89, + 387, + 344, + 399 + ], + "score": 1.0, + "content": "we may arrive at a degenerate solution where every target point", + "type": "text" + }, + { + "bbox": [ + 345, + 388, + 356, + 397 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 387, + 522, + 399 + ], + "score": 1.0, + "content": "is assigned to the same class. Hence, the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 396, + 522, + 410 + ], + "spans": [ + { + "bbox": [ + 88, + 396, + 231, + 410 + ], + "score": 1.0, + "content": "last term encourages the classifier", + "type": "text" + }, + { + "bbox": [ + 231, + 398, + 240, + 407 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 396, + 522, + 410 + ], + "score": 1.0, + "content": "to have balanced labeling for the target samples where it reaches its", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 407, + 371, + 422 + ], + "spans": [ + { + "bbox": [ + 88, + 407, + 133, + 422 + ], + "score": 1.0, + "content": "minimum,", + "type": "text" + }, + { + "bbox": [ + 134, + 408, + 154, + 419 + ], + "score": 0.82, + "content": "\\ln K", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 407, + 371, + 422 + ], + "score": 1.0, + "content": ", when each class is selected with uniform probability.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 91, + 425, + 325, + 437 + ], + "lines": [ + { + "bbox": [ + 90, + 423, + 324, + 441 + ], + "spans": [ + { + "bbox": [ + 90, + 423, + 204, + 441 + ], + "score": 1.0, + "content": "Optimizing the parameter", + "type": "text" + }, + { + "bbox": [ + 204, + 425, + 214, + 438 + ], + "score": 0.88, + "content": "\\theta _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 423, + 311, + 441 + ], + "score": 1.0, + "content": "of the private encoder", + "type": "text" + }, + { + "bbox": [ + 312, + 426, + 324, + 438 + ], + "score": 0.85, + "content": "E _ { p }", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 137, + 441, + 474, + 475 + ], + "lines": [ + { + "bbox": [ + 137, + 441, + 474, + 475 + ], + "spans": [ + { + "bbox": [ + 137, + 441, + 474, + 475 + ], + "score": 0.94, + "content": "\\hat { \\theta _ { p } } = \\underset { \\theta _ { p } } { \\operatorname { a r g m i n } } \\ \\mathcal { L } _ { P } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } _ { i } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } D \\big ( E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) .", + "type": "interline_equation", + "image_path": "7ae28fd7f8bbf8c8f1b3af03a71b8eaa401fada8a1326906381d325d31c0b30a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 137, + 441, + 474, + 452.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 137, + 452.3333333333333, + 474, + 463.66666666666663 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 137, + 463.66666666666663, + 474, + 474.99999999999994 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 478, + 520, + 523 + ], + "lines": [ + { + "bbox": [ + 89, + 478, + 522, + 492 + ], + "spans": [ + { + "bbox": [ + 89, + 478, + 317, + 492 + ], + "score": 1.0, + "content": "The first term in Eq. 9 encourages the private encoder", + "type": "text" + }, + { + "bbox": [ + 317, + 479, + 330, + 491 + ], + "score": 0.9, + "content": "E _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 478, + 522, + 492 + ], + "score": 1.0, + "content": "to preserve the recovery ability of the private", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 488, + 522, + 504 + ], + "spans": [ + { + "bbox": [ + 88, + 488, + 522, + 504 + ], + "score": 1.0, + "content": "features. The second term enforces distinct private features be produced for each domain by penalizing", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 500, + 521, + 514 + ], + "spans": [ + { + "bbox": [ + 89, + 500, + 521, + 514 + ], + "score": 1.0, + "content": "the representation redundancy in different private spaces. This, in turn, encourages moving this common", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 510, + 320, + 525 + ], + "spans": [ + { + "bbox": [ + 89, + 510, + 320, + 525 + ], + "score": 1.0, + "content": "information from multiple domains to their shared space.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 90, + 528, + 323, + 540 + ], + "lines": [ + { + "bbox": [ + 90, + 527, + 322, + 542 + ], + "spans": [ + { + "bbox": [ + 90, + 527, + 204, + 542 + ], + "score": 1.0, + "content": "Optimizing the parameter", + "type": "text" + }, + { + "bbox": [ + 204, + 529, + 214, + 540 + ], + "score": 0.88, + "content": "\\theta _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 527, + 309, + 542 + ], + "score": 1.0, + "content": "of the shared encoder", + "type": "text" + }, + { + "bbox": [ + 309, + 529, + 322, + 540 + ], + "score": 0.87, + "content": "E _ { s }", + "type": "inline_equation" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 543, + 500, + 652 + ], + "lines": [ + { + "bbox": [ + 112, + 543, + 500, + 652 + ], + "spans": [ + { + "bbox": [ + 112, + 543, + 500, + 652 + ], + "score": 0.94, + "content": "\\begin{array} { c } { \\displaystyle \\hat { \\theta _ { s } } = \\arg \\operatorname* { m i n } _ { \\theta _ { s } } \\mathcal { L } _ { S } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { c } } { N } \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle + \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\frac { 1 } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) \\bigg ) . } \\end{array}", + "type": "interline_equation", + "image_path": "a13c7de608ab3bf0d8508ee0dcc844ccff26109d6c0a9c40a3e43e86f9cbf02e.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 112, + 543, + 500, + 579.3333333333334 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 112, + 579.3333333333334, + 500, + 615.6666666666667 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 112, + 615.6666666666667, + 500, + 652.0000000000001 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 39, + 292, + 51 + ], + "lines": [ + { + "bbox": [ + 90, + 38, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 38, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 293, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 671, + 293, + 682 + ], + "spans": [ + { + "bbox": [ + 285, + 671, + 293, + 682 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 91, + 95, + 329, + 108 + ], + "lines": [ + { + "bbox": [ + 90, + 95, + 328, + 109 + ], + "spans": [ + { + "bbox": [ + 90, + 95, + 208, + 109 + ], + "score": 1.0, + "content": "Optimizing the parameters", + "type": "text" + }, + { + "bbox": [ + 208, + 97, + 216, + 108 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 95, + 318, + 109 + ], + "score": 1.0, + "content": "of the domain classifier", + "type": "text" + }, + { + "bbox": [ + 318, + 97, + 328, + 106 + ], + "score": 0.77, + "content": "D", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 110, + 462, + 145 + ], + "lines": [ + { + "bbox": [ + 149, + 110, + 462, + 145 + ], + "spans": [ + { + "bbox": [ + 149, + 110, + 462, + 145 + ], + "score": 0.94, + "content": "\\hat { \\psi } = \\underset { \\psi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { D } = - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\pmb { x } _ { i } ) \\big ) - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { p } ( \\pmb { x } _ { i } ) \\big ) .", + "type": "interline_equation", + "image_path": "21e90163bfe21babae9cdc9f8986f7fd1d54c2568bd6a5fd21c5c4b3297d1048.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 149, + 110, + 462, + 121.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 149, + 121.66666666666667, + 462, + 133.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 149, + 133.33333333333334, + 462, + 145.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 149, + 521, + 183 + ], + "lines": [ + { + "bbox": [ + 91, + 147, + 521, + 161 + ], + "spans": [ + { + "bbox": [ + 91, + 149, + 125, + 161 + ], + "score": 0.91, + "content": "D ( z ; \\psi )", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 147, + 521, + 161 + ], + "score": 1.0, + "content": "can be considered as a classifier whose task is to distinguish between the shared/private features of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 160, + 520, + 172 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 383, + 172 + ], + "score": 1.0, + "content": "the different domains. More precisely, the two terms in Eq. 6 encourage", + "type": "text" + }, + { + "bbox": [ + 383, + 160, + 393, + 170 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 160, + 520, + 172 + ], + "score": 1.0, + "content": "to correctly predict the domain", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 169, + 316, + 184 + ], + "spans": [ + { + "bbox": [ + 88, + 169, + 316, + 184 + ], + "score": 1.0, + "content": "labels from the shared and private features, respectively.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 88, + 147, + 521, + 184 + ] + }, + { + "type": "title", + "bbox": [ + 90, + 187, + 318, + 200 + ], + "lines": [ + { + "bbox": [ + 90, + 187, + 317, + 201 + ], + "spans": [ + { + "bbox": [ + 90, + 187, + 208, + 201 + ], + "score": 1.0, + "content": "Optimizing the parameters", + "type": "text" + }, + { + "bbox": [ + 208, + 188, + 218, + 199 + ], + "score": 0.88, + "content": "\\theta _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 187, + 308, + 201 + ], + "score": 1.0, + "content": "of the label classifier", + "type": "text" + }, + { + "bbox": [ + 308, + 189, + 317, + 198 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 202, + 203, + 409, + 226 + ], + "lines": [ + { + "bbox": [ + 202, + 203, + 409, + 226 + ], + "spans": [ + { + "bbox": [ + 202, + 203, + 409, + 226 + ], + "score": 0.9, + "content": "\\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\big \\{ - H ( \\pmb { y } ) - \\mathbb { E } _ { p ( \\pmb { y } , z _ { s } ) } \\big [ \\ln q ( \\pmb { y } | z _ { s } ) \\big ] \\big \\} .", + "type": "interline_equation", + "image_path": "8efe30add848b2c029835c54b2d4e7f3d10df2c29d3d034dfe1abb77b7bca699.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 202, + 203, + 409, + 226 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 92, + 230, + 521, + 253 + ], + "lines": [ + { + "bbox": [ + 89, + 228, + 520, + 244 + ], + "spans": [ + { + "bbox": [ + 89, + 228, + 258, + 244 + ], + "score": 1.0, + "content": "Since we have access to the source labels,", + "type": "text" + }, + { + "bbox": [ + 259, + 230, + 283, + 242 + ], + "score": 0.92, + "content": "H ( \\pmb { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 228, + 498, + 244 + ], + "score": 1.0, + "content": "is a constant for source samples. we can approximate", + "type": "text" + }, + { + "bbox": [ + 498, + 230, + 520, + 242 + ], + "score": 0.9, + "content": "H [ \\pmb { y } ]", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 240, + 510, + 254 + ], + "spans": [ + { + "bbox": [ + 89, + 240, + 309, + 254 + ], + "score": 1.0, + "content": "for the target samples using the output of the classifier", + "type": "text" + }, + { + "bbox": [ + 309, + 242, + 317, + 251 + ], + "score": 0.83, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 240, + 510, + 254 + ], + "score": 1.0, + "content": ", leading to the following optimization problem:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 89, + 228, + 520, + 254 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 256, + 499, + 329 + ], + "lines": [ + { + "bbox": [ + 112, + 256, + 499, + 329 + ], + "spans": [ + { + "bbox": [ + 112, + 256, + 499, + 329 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { C } = - \\ \\frac { 1 } { N } \\displaystyle \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) } \\\\ & { \\qquad + \\ \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\displaystyle \\frac { 1 } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) \\bigg ) , } \\end{array}", + "type": "interline_equation", + "image_path": "7dadae5cfa60703855a0e15efbcd5d3ff98636b04efe028d1e689cf41d3c6c58.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 112, + 256, + 499, + 280.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 112, + 280.3333333333333, + 499, + 304.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 112, + 304.66666666666663, + 499, + 328.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 330, + 522, + 420 + ], + "lines": [ + { + "bbox": [ + 90, + 331, + 522, + 344 + ], + "spans": [ + { + "bbox": [ + 90, + 331, + 117, + 344 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 117, + 332, + 130, + 343 + ], + "score": 0.88, + "content": "N _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 331, + 433, + 344 + ], + "score": 1.0, + "content": "denotes the number of source samples. Intuitively, we enforce the classifier", + "type": "text" + }, + { + "bbox": [ + 433, + 332, + 442, + 341 + ], + "score": 0.83, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 331, + 522, + 344 + ], + "score": 1.0, + "content": "to correctly predict", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 342, + 522, + 355 + ], + "spans": [ + { + "bbox": [ + 90, + 342, + 522, + 355 + ], + "score": 1.0, + "content": "the class labels of the source samples by the first term in Eq. 8. We use the second term to minimize the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 353, + 522, + 366 + ], + "spans": [ + { + "bbox": [ + 90, + 354, + 135, + 366 + ], + "score": 1.0, + "content": "entropy of", + "type": "text" + }, + { + "bbox": [ + 136, + 353, + 168, + 365 + ], + "score": 0.92, + "content": "q ( \\pmb { y } | \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 354, + 522, + 366 + ], + "score": 1.0, + "content": "for the target samples; effectively, reducing the effects of \"confusing\" labels of target", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 364, + 522, + 377 + ], + "spans": [ + { + "bbox": [ + 89, + 364, + 174, + 377 + ], + "score": 1.0, + "content": "samples, as given by", + "type": "text" + }, + { + "bbox": [ + 174, + 365, + 206, + 376 + ], + "score": 0.92, + "content": "p ( \\pmb { y } | \\pmb { z } _ { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 364, + 522, + 377 + ], + "score": 1.0, + "content": "that leads to decision boundaries occur far away from target data-dense regions", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 88, + 374, + 522, + 388 + ], + "spans": [ + { + "bbox": [ + 88, + 374, + 522, + 388 + ], + "score": 1.0, + "content": "in the feature space. The intuition behind the last term is that by minimizing only the entropy (second term),", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 387, + 522, + 399 + ], + "spans": [ + { + "bbox": [ + 89, + 387, + 344, + 399 + ], + "score": 1.0, + "content": "we may arrive at a degenerate solution where every target point", + "type": "text" + }, + { + "bbox": [ + 345, + 388, + 356, + 397 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 387, + 522, + 399 + ], + "score": 1.0, + "content": "is assigned to the same class. Hence, the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 396, + 522, + 410 + ], + "spans": [ + { + "bbox": [ + 88, + 396, + 231, + 410 + ], + "score": 1.0, + "content": "last term encourages the classifier", + "type": "text" + }, + { + "bbox": [ + 231, + 398, + 240, + 407 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 396, + 522, + 410 + ], + "score": 1.0, + "content": "to have balanced labeling for the target samples where it reaches its", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 407, + 371, + 422 + ], + "spans": [ + { + "bbox": [ + 88, + 407, + 133, + 422 + ], + "score": 1.0, + "content": "minimum,", + "type": "text" + }, + { + "bbox": [ + 134, + 408, + 154, + 419 + ], + "score": 0.82, + "content": "\\ln K", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 407, + 371, + 422 + ], + "score": 1.0, + "content": ", when each class is selected with uniform probability.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5, + "bbox_fs": [ + 88, + 331, + 522, + 422 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 425, + 325, + 437 + ], + "lines": [ + { + "bbox": [ + 90, + 423, + 324, + 441 + ], + "spans": [ + { + "bbox": [ + 90, + 423, + 204, + 441 + ], + "score": 1.0, + "content": "Optimizing the parameter", + "type": "text" + }, + { + "bbox": [ + 204, + 425, + 214, + 438 + ], + "score": 0.88, + "content": "\\theta _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 423, + 311, + 441 + ], + "score": 1.0, + "content": "of the private encoder", + "type": "text" + }, + { + "bbox": [ + 312, + 426, + 324, + 438 + ], + "score": 0.85, + "content": "E _ { p }", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 137, + 441, + 474, + 475 + ], + "lines": [ + { + "bbox": [ + 137, + 441, + 474, + 475 + ], + "spans": [ + { + "bbox": [ + 137, + 441, + 474, + 475 + ], + "score": 0.94, + "content": "\\hat { \\theta _ { p } } = \\underset { \\theta _ { p } } { \\operatorname { a r g m i n } } \\ \\mathcal { L } _ { P } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } _ { i } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } D \\big ( E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) .", + "type": "interline_equation", + "image_path": "7ae28fd7f8bbf8c8f1b3af03a71b8eaa401fada8a1326906381d325d31c0b30a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 137, + 441, + 474, + 452.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 137, + 452.3333333333333, + 474, + 463.66666666666663 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 137, + 463.66666666666663, + 474, + 474.99999999999994 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 90, + 478, + 520, + 523 + ], + "lines": [ + { + "bbox": [ + 89, + 478, + 522, + 492 + ], + "spans": [ + { + "bbox": [ + 89, + 478, + 317, + 492 + ], + "score": 1.0, + "content": "The first term in Eq. 9 encourages the private encoder", + "type": "text" + }, + { + "bbox": [ + 317, + 479, + 330, + 491 + ], + "score": 0.9, + "content": "E _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 478, + 522, + 492 + ], + "score": 1.0, + "content": "to preserve the recovery ability of the private", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 488, + 522, + 504 + ], + "spans": [ + { + "bbox": [ + 88, + 488, + 522, + 504 + ], + "score": 1.0, + "content": "features. The second term enforces distinct private features be produced for each domain by penalizing", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 500, + 521, + 514 + ], + "spans": [ + { + "bbox": [ + 89, + 500, + 521, + 514 + ], + "score": 1.0, + "content": "the representation redundancy in different private spaces. This, in turn, encourages moving this common", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 510, + 320, + 525 + ], + "spans": [ + { + "bbox": [ + 89, + 510, + 320, + 525 + ], + "score": 1.0, + "content": "information from multiple domains to their shared space.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 88, + 478, + 522, + 525 + ] + }, + { + "type": "title", + "bbox": [ + 90, + 528, + 323, + 540 + ], + "lines": [ + { + "bbox": [ + 90, + 527, + 322, + 542 + ], + "spans": [ + { + "bbox": [ + 90, + 527, + 204, + 542 + ], + "score": 1.0, + "content": "Optimizing the parameter", + "type": "text" + }, + { + "bbox": [ + 204, + 529, + 214, + 540 + ], + "score": 0.88, + "content": "\\theta _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 527, + 309, + 542 + ], + "score": 1.0, + "content": "of the shared encoder", + "type": "text" + }, + { + "bbox": [ + 309, + 529, + 322, + 540 + ], + "score": 0.87, + "content": "E _ { s }", + "type": "inline_equation" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 543, + 500, + 652 + ], + "lines": [ + { + "bbox": [ + 112, + 543, + 500, + 652 + ], + "spans": [ + { + "bbox": [ + 112, + 543, + 500, + 652 + ], + "score": 0.94, + "content": "\\begin{array} { c } { \\displaystyle \\hat { \\theta _ { s } } = \\arg \\operatorname* { m i n } _ { \\theta _ { s } } \\mathcal { L } _ { S } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { c } } { N } \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle + \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\frac { 1 } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) \\bigg ) . } \\end{array}", + "type": "interline_equation", + "image_path": "a13c7de608ab3bf0d8508ee0dcc844ccff26109d6c0a9c40a3e43e86f9cbf02e.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 112, + 543, + 500, + 579.3333333333334 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 112, + 579.3333333333334, + 500, + 615.6666666666667 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 112, + 615.6666666666667, + 500, + 652.0000000000001 + ], + "spans": [], + "index": 33 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 95, + 521, + 206 + ], + "lines": [ + { + "bbox": [ + 89, + 95, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 89, + 95, + 319, + 109 + ], + "score": 1.0, + "content": "The first term in Eq. 10 encourages the shared encoder", + "type": "text" + }, + { + "bbox": [ + 320, + 96, + 332, + 107 + ], + "score": 0.89, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 95, + 522, + 109 + ], + "score": 1.0, + "content": "to preserve the recovery ability of the shared", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 106, + 522, + 119 + ], + "spans": [ + { + "bbox": [ + 89, + 106, + 461, + 119 + ], + "score": 1.0, + "content": "features. The second term is the source domain classification loss penalty that encourages", + "type": "text" + }, + { + "bbox": [ + 461, + 107, + 474, + 118 + ], + "score": 0.89, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 106, + 522, + 119 + ], + "score": 1.0, + "content": "to produce", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 116, + 522, + 131 + ], + "spans": [ + { + "bbox": [ + 88, + 116, + 522, + 131 + ], + "score": 1.0, + "content": "discriminative features for the labeled source samples. The third term simulates the adversarial training by", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 128, + 522, + 142 + ], + "spans": [ + { + "bbox": [ + 89, + 128, + 228, + 142 + ], + "score": 1.0, + "content": "trying to fool the domain classifier", + "type": "text" + }, + { + "bbox": [ + 228, + 129, + 238, + 139 + ], + "score": 0.78, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 128, + 377, + 142 + ], + "score": 1.0, + "content": "when predicting the domain labels", + "type": "text" + }, + { + "bbox": [ + 378, + 129, + 385, + 138 + ], + "score": 0.69, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 128, + 488, + 142 + ], + "score": 1.0, + "content": ", given the shared features", + "type": "text" + }, + { + "bbox": [ + 488, + 130, + 499, + 140 + ], + "score": 0.86, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 128, + 522, + 142 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 138, + 522, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 138, + 522, + 152 + ], + "score": 1.0, + "content": "effect of this is two-fold: (i) the rendered shared features are more distinct from the corresponding private", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 150, + 522, + 163 + ], + "spans": [ + { + "bbox": [ + 89, + 150, + 522, + 163 + ], + "score": 1.0, + "content": "features, (ii) the shared features of different domains are encouraged to be similar to each other. The last two", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 162, + 522, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 158, + 174 + ], + "score": 1.0, + "content": "terms encourage", + "type": "text" + }, + { + "bbox": [ + 158, + 162, + 170, + 172 + ], + "score": 0.89, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 162, + 522, + 174 + ], + "score": 1.0, + "content": "to produce the shared features for target samples so that the classifier is confident on the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 172, + 522, + 185 + ], + "spans": [ + { + "bbox": [ + 90, + 172, + 522, + 185 + ], + "score": 1.0, + "content": "unlabeled target data, driving the shared features away from the decision boundaries. To train our model, we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 182, + 519, + 196 + ], + "spans": [ + { + "bbox": [ + 89, + 182, + 280, + 196 + ], + "score": 1.0, + "content": "alternate between updating the shared encoder", + "type": "text" + }, + { + "bbox": [ + 281, + 184, + 293, + 194 + ], + "score": 0.88, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 182, + 376, + 196 + ], + "score": 1.0, + "content": ", the private encoder", + "type": "text" + }, + { + "bbox": [ + 377, + 184, + 390, + 195 + ], + "score": 0.9, + "content": "E _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 182, + 443, + 196 + ], + "score": 1.0, + "content": ", the decoder", + "type": "text" + }, + { + "bbox": [ + 443, + 184, + 452, + 193 + ], + "score": 0.8, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 182, + 509, + 196 + ], + "score": 1.0, + "content": ", the classifier", + "type": "text" + }, + { + "bbox": [ + 510, + 184, + 519, + 193 + ], + "score": 0.79, + "content": "C", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 193, + 507, + 208 + ], + "spans": [ + { + "bbox": [ + 88, + 193, + 192, + 208 + ], + "score": 1.0, + "content": "and the domain classifier", + "type": "text" + }, + { + "bbox": [ + 192, + 195, + 202, + 204 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 193, + 507, + 208 + ], + "score": 1.0, + "content": "using the SGD algorithm (see Algorithm 1 in Appendix E for more details).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 91, + 227, + 195, + 240 + ], + "lines": [ + { + "bbox": [ + 88, + 226, + 197, + 242 + ], + "spans": [ + { + "bbox": [ + 88, + 226, + 197, + 242 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 90, + 255, + 521, + 430 + ], + "lines": [ + { + "bbox": [ + 90, + 256, + 522, + 268 + ], + "spans": [ + { + "bbox": [ + 90, + 256, + 522, + 268 + ], + "score": 1.0, + "content": "There has been extensive prior work on domain adaptation (Csurka (2017)). Recent papers have focused", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 266, + 521, + 279 + ], + "spans": [ + { + "bbox": [ + 89, + 266, + 521, + 279 + ], + "score": 1.0, + "content": "on transferring deep neural network representations from a labeled source dataset to an unlabeled target", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 277, + 522, + 291 + ], + "spans": [ + { + "bbox": [ + 89, + 277, + 522, + 291 + ], + "score": 1.0, + "content": "domain, where the main strategy is to find a feature space such that the confusion between source and target", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 288, + 523, + 301 + ], + "spans": [ + { + "bbox": [ + 89, + 288, + 523, + 301 + ], + "score": 1.0, + "content": "distributions in that space is maximized ( Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al. (2017);", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 299, + 523, + 312 + ], + "spans": [ + { + "bbox": [ + 89, + 299, + 523, + 312 + ], + "score": 1.0, + "content": "Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017); Yan et al. (2017); Bousmalis et al. (2017)). For", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 88, + 309, + 523, + 324 + ], + "spans": [ + { + "bbox": [ + 88, + 309, + 523, + 324 + ], + "score": 1.0, + "content": "this, it is critical to first define a measure of divergence between source and target distributions. For instance,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 322, + 523, + 334 + ], + "spans": [ + { + "bbox": [ + 89, + 322, + 523, + 334 + ], + "score": 1.0, + "content": "several methods have used the Maximum Mean Discrepancy (MMD) loss for this purpose (Bousmalis et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 332, + 523, + 345 + ], + "spans": [ + { + "bbox": [ + 89, + 332, + 523, + 345 + ], + "score": 1.0, + "content": "(2017); Zellinger et al. (2017); Long et al. (2014)). MMD computes the norm of the difference between two", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 342, + 522, + 357 + ], + "spans": [ + { + "bbox": [ + 88, + 342, + 522, + 357 + ], + "score": 1.0, + "content": "domain means in the reproducing Kernel Hilbert Space (RKHS) induced by a pre-specified kernel. The Deep", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 354, + 522, + 367 + ], + "spans": [ + { + "bbox": [ + 88, + 354, + 522, + 367 + ], + "score": 1.0, + "content": "Adaptation Network (DAN) (Long et al. (2015)) applied MMD to layers embedded in a RKHS, effectively", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 366, + 522, + 377 + ], + "spans": [ + { + "bbox": [ + 89, + 366, + 522, + 377 + ], + "score": 1.0, + "content": "matching higher order statistics of the two distributions. The deep Correlation Alignment (CORAL) method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 375, + 522, + 389 + ], + "spans": [ + { + "bbox": [ + 88, + 375, + 522, + 389 + ], + "score": 1.0, + "content": "(Sun & Saenko (2016)) attempts to match the mean and covariance of the two distributions. Deep Transfer", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 88, + 386, + 522, + 400 + ], + "spans": [ + { + "bbox": [ + 88, + 386, + 522, + 400 + ], + "score": 1.0, + "content": "Network (DTN) (Zhang et al. (2015)) achieved source/target distribution alignment via two types of network", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 398, + 521, + 410 + ], + "spans": [ + { + "bbox": [ + 89, + 398, + 521, + 410 + ], + "score": 1.0, + "content": "layers based on MMD distance: the shared feature extraction layer, which learns a subspace that matches the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 88, + 409, + 522, + 421 + ], + "spans": [ + { + "bbox": [ + 88, + 409, + 522, + 421 + ], + "score": 1.0, + "content": "marginal distributions of the source and the target samples, and the discrimination layer, which matches the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 420, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 89, + 420, + 294, + 432 + ], + "score": 1.0, + "content": "conditional distributions by classifier transduction.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 91, + 436, + 521, + 601 + ], + "lines": [ + { + "bbox": [ + 90, + 437, + 523, + 449 + ], + "spans": [ + { + "bbox": [ + 90, + 437, + 523, + 449 + ], + "score": 1.0, + "content": "Recently proposed unsupervised DA methods (Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 448, + 522, + 460 + ], + "spans": [ + { + "bbox": [ + 90, + 448, + 522, + 460 + ], + "score": 1.0, + "content": "(2017); Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017)) operate by training deep neural networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "spans": [ + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "score": 1.0, + "content": "using adversarial training, which allows the learning of feature representations that are simultaneously dis-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "spans": [ + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "score": 1.0, + "content": "criminative of source labels, and indistinguishable between the source and target domain. For instance, Ganin", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 481, + 521, + 493 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 521, + 493 + ], + "score": 1.0, + "content": "& Lempitsky (2015) proposed a DA mechanism called Domain-Adversarial Training of Neural Networks", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 491, + 522, + 505 + ], + "spans": [ + { + "bbox": [ + 88, + 491, + 522, + 505 + ], + "score": 1.0, + "content": "(DANN), which enables the network to learn domain invariant representations in an adversarial way by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 502, + 522, + 516 + ], + "spans": [ + { + "bbox": [ + 89, + 502, + 522, + 516 + ], + "score": 1.0, + "content": "adding a domain classifier and back-propagating inverse gradients. Adversarial Discriminative Domain", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 88, + 513, + 523, + 527 + ], + "spans": [ + { + "bbox": [ + 88, + 513, + 523, + 527 + ], + "score": 1.0, + "content": "Adaptation (ADDA) (Tzeng et al. (2017)) learns a discriminative feature subspace using the source labels,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 524, + 522, + 538 + ], + "spans": [ + { + "bbox": [ + 89, + 524, + 522, + 538 + ], + "score": 1.0, + "content": "followed by a separate encoding of the target data to this subspace using an asymmetric mapping learned", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 535, + 523, + 549 + ], + "spans": [ + { + "bbox": [ + 89, + 535, + 523, + 549 + ], + "score": 1.0, + "content": "through a domain-adversarial loss. Liu et al. (2017) makes a shared-latent space assumption and proposes an", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 546, + 523, + 559 + ], + "spans": [ + { + "bbox": [ + 89, + 546, + 523, + 559 + ], + "score": 1.0, + "content": "unsupervised image-to-image translation (UNIT) framework based on Coupled GANs (Liu & Tuzel (2016)).", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 90, + 558, + 522, + 570 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 522, + 570 + ], + "score": 1.0, + "content": "Another example is the pixel-level domain adaptation models that perform the distribution alignment not", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 89, + 568, + 522, + 582 + ], + "spans": [ + { + "bbox": [ + 89, + 568, + 522, + 582 + ], + "score": 1.0, + "content": "in the feature space but directly in raw pixel space. PixelDA (Bousmalis et al. (2017)) uses adversarial", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 88, + 579, + 522, + 593 + ], + "spans": [ + { + "bbox": [ + 88, + 579, + 522, + 593 + ], + "score": 1.0, + "content": "approaches to adapt source-domain images as if drawn from the target domain while maintaining the original", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 591, + 124, + 603 + ], + "spans": [ + { + "bbox": [ + 89, + 591, + 124, + 603 + ], + "score": 1.0, + "content": "content.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "score": 1.0, + "content": "While these approaches have shown success in DA tasks with single source-target domains, they are not", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 90, + 618, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 90, + 618, + 522, + 631 + ], + "score": 1.0, + "content": "designed to leverage information from multiple domains simultaneously. More recently, Zhao et al. (2017)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 90, + 630, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 630, + 521, + 642 + ], + "score": 1.0, + "content": "introduced an adversarial framework called MDAN for multiple source single target domain adaptation where", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "score": 1.0, + "content": "a domain classifier, induced by minimizing the H-divergence between multiple source and a target domain, is", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 672, + 292, + 680 + ], + "lines": [ + { + "bbox": [ + 286, + 671, + 293, + 681 + ], + "spans": [ + { + "bbox": [ + 286, + 671, + 293, + 681 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 95, + 521, + 206 + ], + "lines": [ + { + "bbox": [ + 89, + 95, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 89, + 95, + 319, + 109 + ], + "score": 1.0, + "content": "The first term in Eq. 10 encourages the shared encoder", + "type": "text" + }, + { + "bbox": [ + 320, + 96, + 332, + 107 + ], + "score": 0.89, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 95, + 522, + 109 + ], + "score": 1.0, + "content": "to preserve the recovery ability of the shared", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 106, + 522, + 119 + ], + "spans": [ + { + "bbox": [ + 89, + 106, + 461, + 119 + ], + "score": 1.0, + "content": "features. The second term is the source domain classification loss penalty that encourages", + "type": "text" + }, + { + "bbox": [ + 461, + 107, + 474, + 118 + ], + "score": 0.89, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 106, + 522, + 119 + ], + "score": 1.0, + "content": "to produce", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 116, + 522, + 131 + ], + "spans": [ + { + "bbox": [ + 88, + 116, + 522, + 131 + ], + "score": 1.0, + "content": "discriminative features for the labeled source samples. The third term simulates the adversarial training by", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 128, + 522, + 142 + ], + "spans": [ + { + "bbox": [ + 89, + 128, + 228, + 142 + ], + "score": 1.0, + "content": "trying to fool the domain classifier", + "type": "text" + }, + { + "bbox": [ + 228, + 129, + 238, + 139 + ], + "score": 0.78, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 128, + 377, + 142 + ], + "score": 1.0, + "content": "when predicting the domain labels", + "type": "text" + }, + { + "bbox": [ + 378, + 129, + 385, + 138 + ], + "score": 0.69, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 128, + 488, + 142 + ], + "score": 1.0, + "content": ", given the shared features", + "type": "text" + }, + { + "bbox": [ + 488, + 130, + 499, + 140 + ], + "score": 0.86, + "content": "z _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 128, + 522, + 142 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 138, + 522, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 138, + 522, + 152 + ], + "score": 1.0, + "content": "effect of this is two-fold: (i) the rendered shared features are more distinct from the corresponding private", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 150, + 522, + 163 + ], + "spans": [ + { + "bbox": [ + 89, + 150, + 522, + 163 + ], + "score": 1.0, + "content": "features, (ii) the shared features of different domains are encouraged to be similar to each other. The last two", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 162, + 522, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 158, + 174 + ], + "score": 1.0, + "content": "terms encourage", + "type": "text" + }, + { + "bbox": [ + 158, + 162, + 170, + 172 + ], + "score": 0.89, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 162, + 522, + 174 + ], + "score": 1.0, + "content": "to produce the shared features for target samples so that the classifier is confident on the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 172, + 522, + 185 + ], + "spans": [ + { + "bbox": [ + 90, + 172, + 522, + 185 + ], + "score": 1.0, + "content": "unlabeled target data, driving the shared features away from the decision boundaries. To train our model, we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 182, + 519, + 196 + ], + "spans": [ + { + "bbox": [ + 89, + 182, + 280, + 196 + ], + "score": 1.0, + "content": "alternate between updating the shared encoder", + "type": "text" + }, + { + "bbox": [ + 281, + 184, + 293, + 194 + ], + "score": 0.88, + "content": "E _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 182, + 376, + 196 + ], + "score": 1.0, + "content": ", the private encoder", + "type": "text" + }, + { + "bbox": [ + 377, + 184, + 390, + 195 + ], + "score": 0.9, + "content": "E _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 182, + 443, + 196 + ], + "score": 1.0, + "content": ", the decoder", + "type": "text" + }, + { + "bbox": [ + 443, + 184, + 452, + 193 + ], + "score": 0.8, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 182, + 509, + 196 + ], + "score": 1.0, + "content": ", the classifier", + "type": "text" + }, + { + "bbox": [ + 510, + 184, + 519, + 193 + ], + "score": 0.79, + "content": "C", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 193, + 507, + 208 + ], + "spans": [ + { + "bbox": [ + 88, + 193, + 192, + 208 + ], + "score": 1.0, + "content": "and the domain classifier", + "type": "text" + }, + { + "bbox": [ + 192, + 195, + 202, + 204 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 193, + 507, + 208 + ], + "score": 1.0, + "content": "using the SGD algorithm (see Algorithm 1 in Appendix E for more details).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 88, + 95, + 522, + 208 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 227, + 195, + 240 + ], + "lines": [ + { + "bbox": [ + 88, + 226, + 197, + 242 + ], + "spans": [ + { + "bbox": [ + 88, + 226, + 197, + 242 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 90, + 255, + 521, + 430 + ], + "lines": [ + { + "bbox": [ + 90, + 256, + 522, + 268 + ], + "spans": [ + { + "bbox": [ + 90, + 256, + 522, + 268 + ], + "score": 1.0, + "content": "There has been extensive prior work on domain adaptation (Csurka (2017)). Recent papers have focused", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 266, + 521, + 279 + ], + "spans": [ + { + "bbox": [ + 89, + 266, + 521, + 279 + ], + "score": 1.0, + "content": "on transferring deep neural network representations from a labeled source dataset to an unlabeled target", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 277, + 522, + 291 + ], + "spans": [ + { + "bbox": [ + 89, + 277, + 522, + 291 + ], + "score": 1.0, + "content": "domain, where the main strategy is to find a feature space such that the confusion between source and target", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 288, + 523, + 301 + ], + "spans": [ + { + "bbox": [ + 89, + 288, + 523, + 301 + ], + "score": 1.0, + "content": "distributions in that space is maximized ( Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al. (2017);", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 299, + 523, + 312 + ], + "spans": [ + { + "bbox": [ + 89, + 299, + 523, + 312 + ], + "score": 1.0, + "content": "Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017); Yan et al. (2017); Bousmalis et al. (2017)). For", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 88, + 309, + 523, + 324 + ], + "spans": [ + { + "bbox": [ + 88, + 309, + 523, + 324 + ], + "score": 1.0, + "content": "this, it is critical to first define a measure of divergence between source and target distributions. For instance,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 322, + 523, + 334 + ], + "spans": [ + { + "bbox": [ + 89, + 322, + 523, + 334 + ], + "score": 1.0, + "content": "several methods have used the Maximum Mean Discrepancy (MMD) loss for this purpose (Bousmalis et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 332, + 523, + 345 + ], + "spans": [ + { + "bbox": [ + 89, + 332, + 523, + 345 + ], + "score": 1.0, + "content": "(2017); Zellinger et al. (2017); Long et al. (2014)). MMD computes the norm of the difference between two", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 342, + 522, + 357 + ], + "spans": [ + { + "bbox": [ + 88, + 342, + 522, + 357 + ], + "score": 1.0, + "content": "domain means in the reproducing Kernel Hilbert Space (RKHS) induced by a pre-specified kernel. The Deep", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 354, + 522, + 367 + ], + "spans": [ + { + "bbox": [ + 88, + 354, + 522, + 367 + ], + "score": 1.0, + "content": "Adaptation Network (DAN) (Long et al. (2015)) applied MMD to layers embedded in a RKHS, effectively", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 366, + 522, + 377 + ], + "spans": [ + { + "bbox": [ + 89, + 366, + 522, + 377 + ], + "score": 1.0, + "content": "matching higher order statistics of the two distributions. The deep Correlation Alignment (CORAL) method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 375, + 522, + 389 + ], + "spans": [ + { + "bbox": [ + 88, + 375, + 522, + 389 + ], + "score": 1.0, + "content": "(Sun & Saenko (2016)) attempts to match the mean and covariance of the two distributions. Deep Transfer", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 88, + 386, + 522, + 400 + ], + "spans": [ + { + "bbox": [ + 88, + 386, + 522, + 400 + ], + "score": 1.0, + "content": "Network (DTN) (Zhang et al. (2015)) achieved source/target distribution alignment via two types of network", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 398, + 521, + 410 + ], + "spans": [ + { + "bbox": [ + 89, + 398, + 521, + 410 + ], + "score": 1.0, + "content": "layers based on MMD distance: the shared feature extraction layer, which learns a subspace that matches the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 88, + 409, + 522, + 421 + ], + "spans": [ + { + "bbox": [ + 88, + 409, + 522, + 421 + ], + "score": 1.0, + "content": "marginal distributions of the source and the target samples, and the discrimination layer, which matches the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 420, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 89, + 420, + 294, + 432 + ], + "score": 1.0, + "content": "conditional distributions by classifier transduction.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18.5, + "bbox_fs": [ + 88, + 256, + 523, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 436, + 521, + 601 + ], + "lines": [ + { + "bbox": [ + 90, + 437, + 523, + 449 + ], + "spans": [ + { + "bbox": [ + 90, + 437, + 523, + 449 + ], + "score": 1.0, + "content": "Recently proposed unsupervised DA methods (Rebuffi et al. (2017); Benaim & Wolf (2017); Courty et al.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 448, + 522, + 460 + ], + "spans": [ + { + "bbox": [ + 90, + 448, + 522, + 460 + ], + "score": 1.0, + "content": "(2017); Motiian et al. (2017); Saito et al. (2017); Zhang et al. (2017)) operate by training deep neural networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "spans": [ + { + "bbox": [ + 89, + 459, + 522, + 471 + ], + "score": 1.0, + "content": "using adversarial training, which allows the learning of feature representations that are simultaneously dis-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "spans": [ + { + "bbox": [ + 89, + 470, + 522, + 482 + ], + "score": 1.0, + "content": "criminative of source labels, and indistinguishable between the source and target domain. For instance, Ganin", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 481, + 521, + 493 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 521, + 493 + ], + "score": 1.0, + "content": "& Lempitsky (2015) proposed a DA mechanism called Domain-Adversarial Training of Neural Networks", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 491, + 522, + 505 + ], + "spans": [ + { + "bbox": [ + 88, + 491, + 522, + 505 + ], + "score": 1.0, + "content": "(DANN), which enables the network to learn domain invariant representations in an adversarial way by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 502, + 522, + 516 + ], + "spans": [ + { + "bbox": [ + 89, + 502, + 522, + 516 + ], + "score": 1.0, + "content": "adding a domain classifier and back-propagating inverse gradients. Adversarial Discriminative Domain", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 88, + 513, + 523, + 527 + ], + "spans": [ + { + "bbox": [ + 88, + 513, + 523, + 527 + ], + "score": 1.0, + "content": "Adaptation (ADDA) (Tzeng et al. (2017)) learns a discriminative feature subspace using the source labels,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 524, + 522, + 538 + ], + "spans": [ + { + "bbox": [ + 89, + 524, + 522, + 538 + ], + "score": 1.0, + "content": "followed by a separate encoding of the target data to this subspace using an asymmetric mapping learned", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 535, + 523, + 549 + ], + "spans": [ + { + "bbox": [ + 89, + 535, + 523, + 549 + ], + "score": 1.0, + "content": "through a domain-adversarial loss. Liu et al. (2017) makes a shared-latent space assumption and proposes an", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 546, + 523, + 559 + ], + "spans": [ + { + "bbox": [ + 89, + 546, + 523, + 559 + ], + "score": 1.0, + "content": "unsupervised image-to-image translation (UNIT) framework based on Coupled GANs (Liu & Tuzel (2016)).", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 90, + 558, + 522, + 570 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 522, + 570 + ], + "score": 1.0, + "content": "Another example is the pixel-level domain adaptation models that perform the distribution alignment not", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 89, + 568, + 522, + 582 + ], + "spans": [ + { + "bbox": [ + 89, + 568, + 522, + 582 + ], + "score": 1.0, + "content": "in the feature space but directly in raw pixel space. PixelDA (Bousmalis et al. (2017)) uses adversarial", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 88, + 579, + 522, + 593 + ], + "spans": [ + { + "bbox": [ + 88, + 579, + 522, + 593 + ], + "score": 1.0, + "content": "approaches to adapt source-domain images as if drawn from the target domain while maintaining the original", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 591, + 124, + 603 + ], + "spans": [ + { + "bbox": [ + 89, + 591, + 124, + 603 + ], + "score": 1.0, + "content": "content.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 34, + "bbox_fs": [ + 88, + 437, + 523, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "score": 1.0, + "content": "While these approaches have shown success in DA tasks with single source-target domains, they are not", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 90, + 618, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 90, + 618, + 522, + 631 + ], + "score": 1.0, + "content": "designed to leverage information from multiple domains simultaneously. More recently, Zhao et al. (2017)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 90, + 630, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 630, + 521, + 642 + ], + "score": 1.0, + "content": "introduced an adversarial framework called MDAN for multiple source single target domain adaptation where", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "score": 1.0, + "content": "a domain classifier, induced by minimizing the H-divergence between multiple source and a target domain, is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 89, + 96, + 521, + 108 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 521, + 108 + ], + "score": 1.0, + "content": "used to align their feature distributions in a shared space. Instead, in our approach we focus on multi-target", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 107, + 521, + 119 + ], + "spans": [ + { + "bbox": [ + 89, + 107, + 521, + 119 + ], + "score": 1.0, + "content": "DA where we perform adaptation of multiple unlabelled target domains. Although both our model and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 117, + 521, + 130 + ], + "spans": [ + { + "bbox": [ + 89, + 117, + 521, + 130 + ], + "score": 1.0, + "content": "MDAN use the similar notion of the domain classifier to minimize the domain mismatch in shared space, the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 129, + 522, + 141 + ], + "spans": [ + { + "bbox": [ + 90, + 129, + 522, + 141 + ], + "score": 1.0, + "content": "domain classifier induced by our information-theoretic (IT) loss also acts to separate domains in the private", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 139, + 522, + 153 + ], + "spans": [ + { + "bbox": [ + 88, + 139, + 153, + 153 + ], + "score": 1.0, + "content": "space (see Eqs.", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 153, + 139, + 177, + 150 + ], + "score": 0.51, + "content": "6 \\& 9", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 178, + 139, + 522, + 153 + ], + "score": 1.0, + "content": "for more details), improving the essential reconstruction ability, similar to (Bousmalis", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 150, + 521, + 164 + ], + "spans": [ + { + "bbox": [ + 88, + 150, + 521, + 164 + ], + "score": 1.0, + "content": "et al. (2016)). We provided how our model is related to IT representation learning approaches, and multiple", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 162, + 521, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 521, + 174 + ], + "score": 1.0, + "content": "domain transfer networks in Appendices A and B respectively. In Appendix C, we also clearly contrasted", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "spans": [ + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "score": 1.0, + "content": "our model with DSN model which also uses the notion of auto-encoders to explicitly separate the feature", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 183, + 495, + 196 + ], + "spans": [ + { + "bbox": [ + 88, + 183, + 495, + 196 + ], + "score": 1.0, + "content": "representations private to each source/target domain from those that are shared between the domains.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 43.5, + "bbox_fs": [ + 88, + 606, + 522, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 90, + 95, + 521, + 195 + ], + "lines": [ + { + "bbox": [ + 89, + 96, + 521, + 108 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 521, + 108 + ], + "score": 1.0, + "content": "used to align their feature distributions in a shared space. Instead, in our approach we focus on multi-target", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 107, + 521, + 119 + ], + "spans": [ + { + "bbox": [ + 89, + 107, + 521, + 119 + ], + "score": 1.0, + "content": "DA where we perform adaptation of multiple unlabelled target domains. Although both our model and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 117, + 521, + 130 + ], + "spans": [ + { + "bbox": [ + 89, + 117, + 521, + 130 + ], + "score": 1.0, + "content": "MDAN use the similar notion of the domain classifier to minimize the domain mismatch in shared space, the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 129, + 522, + 141 + ], + "spans": [ + { + "bbox": [ + 90, + 129, + 522, + 141 + ], + "score": 1.0, + "content": "domain classifier induced by our information-theoretic (IT) loss also acts to separate domains in the private", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 139, + 522, + 153 + ], + "spans": [ + { + "bbox": [ + 88, + 139, + 153, + 153 + ], + "score": 1.0, + "content": "space (see Eqs.", + "type": "text" + }, + { + "bbox": [ + 153, + 139, + 177, + 150 + ], + "score": 0.51, + "content": "6 \\& 9", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 139, + 522, + 153 + ], + "score": 1.0, + "content": "for more details), improving the essential reconstruction ability, similar to (Bousmalis", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 150, + 521, + 164 + ], + "spans": [ + { + "bbox": [ + 88, + 150, + 521, + 164 + ], + "score": 1.0, + "content": "et al. (2016)). We provided how our model is related to IT representation learning approaches, and multiple", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 162, + 521, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 521, + 174 + ], + "score": 1.0, + "content": "domain transfer networks in Appendices A and B respectively. In Appendix C, we also clearly contrasted", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "spans": [ + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "score": 1.0, + "content": "our model with DSN model which also uses the notion of auto-encoders to explicitly separate the feature", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 183, + 495, + 196 + ], + "spans": [ + { + "bbox": [ + 88, + 183, + 495, + 196 + ], + "score": 1.0, + "content": "representations private to each source/target domain from those that are shared between the domains.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 92, + 210, + 240, + 223 + ], + "lines": [ + { + "bbox": [ + 89, + 210, + 241, + 225 + ], + "spans": [ + { + "bbox": [ + 89, + 210, + 241, + 225 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 91, + 235, + 522, + 324 + ], + "lines": [ + { + "bbox": [ + 90, + 235, + 521, + 247 + ], + "spans": [ + { + "bbox": [ + 90, + 235, + 521, + 247 + ], + "score": 1.0, + "content": "We compare the proposed method with state-of-the-art methods on standard benchmark datasets: a digit", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 246, + 522, + 259 + ], + "spans": [ + { + "bbox": [ + 90, + 246, + 522, + 259 + ], + "score": 1.0, + "content": "classification task that includes 4 datasets: MNIST (LeCun et al. (1998)), MNIST-M (Ganin et al. (2016)),", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 257, + 522, + 271 + ], + "spans": [ + { + "bbox": [ + 89, + 257, + 522, + 271 + ], + "score": 1.0, + "content": "SVHN (Netzer et al. (2011)), USPS (Tzeng et al. (2017)), Multi-PIE expression recognition dataset2, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 267, + 523, + 281 + ], + "spans": [ + { + "bbox": [ + 88, + 267, + 523, + 281 + ], + "score": 1.0, + "content": "PACS multi-domain image recognition benchmark (Li et al. (2017)), a new dataset designed for the cross-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 280, + 522, + 292 + ], + "spans": [ + { + "bbox": [ + 90, + 280, + 522, + 292 + ], + "score": 1.0, + "content": "domain recognition problems (the details for this experiment is available in Appendix G). Fig. 3 illustrates", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 290, + 521, + 303 + ], + "spans": [ + { + "bbox": [ + 89, + 290, + 521, + 303 + ], + "score": 1.0, + "content": "image samples from different datasets and domains. We evaluate the performance of all methods with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 301, + 520, + 313 + ], + "spans": [ + { + "bbox": [ + 90, + 301, + 520, + 313 + ], + "score": 1.0, + "content": "classification accuracy metric. We repeated each experiment 5 times and report the average and the standard", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 311, + 195, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 311, + 195, + 326 + ], + "score": 1.0, + "content": "deviation of the accuracy.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 91, + 329, + 522, + 428 + ], + "lines": [ + { + "bbox": [ + 89, + 329, + 522, + 341 + ], + "spans": [ + { + "bbox": [ + 89, + 329, + 522, + 341 + ], + "score": 1.0, + "content": "We used ADAM (Kingma & Ba (2015)) for training; the learning rate was set to 0.0002 and momentum", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 339, + 524, + 353 + ], + "spans": [ + { + "bbox": [ + 88, + 339, + 524, + 353 + ], + "score": 1.0, + "content": "parameters to 0.5 and 0.999. We used batches of size 16 from each domain, and the input images were mean-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 350, + 523, + 365 + ], + "spans": [ + { + "bbox": [ + 88, + 350, + 174, + 365 + ], + "score": 1.0, + "content": "centered/rescaled to", + "type": "text" + }, + { + "bbox": [ + 174, + 351, + 203, + 363 + ], + "score": 0.81, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 350, + 389, + 365 + ], + "score": 1.0, + "content": ". The hyper-parameters are empirically set as", + "type": "text" + }, + { + "bbox": [ + 389, + 351, + 473, + 362 + ], + "score": 0.71, + "content": "\\lambda _ { r } = 1 . 0 , \\lambda _ { c } = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 350, + 476, + 365 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 476, + 351, + 519, + 362 + ], + "score": 0.72, + "content": "\\lambda _ { d } = 0 . 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 519, + 350, + 523, + 365 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 362, + 522, + 374 + ], + "spans": [ + { + "bbox": [ + 90, + 362, + 522, + 374 + ], + "score": 1.0, + "content": "For the network architecture, our private/shared encoders consisted of three convolutional layers as the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 90, + 374, + 521, + 384 + ], + "spans": [ + { + "bbox": [ + 90, + 374, + 521, + 384 + ], + "score": 1.0, + "content": "front-end and four basic residual blocks as the back-end. The decoder consisted of four basic residual blocks", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 384, + 522, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 384, + 522, + 396 + ], + "score": 1.0, + "content": "as the front-end and four transposed convolutional layers as the back-end. The discriminator and the classifier", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 90, + 396, + 520, + 406 + ], + "spans": [ + { + "bbox": [ + 90, + 396, + 520, + 406 + ], + "score": 1.0, + "content": "consisted of stacks of convolutional layers. We used ReLU for nonlinearity. Tanh function is used as the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 406, + 521, + 418 + ], + "spans": [ + { + "bbox": [ + 89, + 406, + 286, + 418 + ], + "score": 1.0, + "content": "activation function of the last layer in the decoder", + "type": "text" + }, + { + "bbox": [ + 286, + 406, + 295, + 416 + ], + "score": 0.8, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 406, + 417, + 418 + ], + "score": 1.0, + "content": "for scaling the output pixels to", + "type": "text" + }, + { + "bbox": [ + 418, + 406, + 446, + 418 + ], + "score": 0.88, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 406, + 521, + 418 + ], + "score": 1.0, + "content": ". The details of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 88, + 417, + 232, + 429 + ], + "spans": [ + { + "bbox": [ + 88, + 417, + 232, + 429 + ], + "score": 1.0, + "content": "networks are given in Appendix D.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 90, + 434, + 521, + 576 + ], + "lines": [ + { + "bbox": [ + 89, + 434, + 522, + 447 + ], + "spans": [ + { + "bbox": [ + 89, + 434, + 522, + 447 + ], + "score": 1.0, + "content": "The quantitative evaluation involves a comparison of the performance of our model to previous work and to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 444, + 522, + 457 + ], + "spans": [ + { + "bbox": [ + 90, + 444, + 522, + 457 + ], + "score": 1.0, + "content": "Source Only and 1-NN baselines that do not use any domain adaptation. For Source Only baseline, we train", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 456, + 521, + 469 + ], + "spans": [ + { + "bbox": [ + 89, + 456, + 521, + 469 + ], + "score": 1.0, + "content": "our model only on the unaltered source training data and evaluate on the target test data. We compare the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 466, + 523, + 480 + ], + "spans": [ + { + "bbox": [ + 89, + 466, + 523, + 480 + ], + "score": 1.0, + "content": "proposed method MTDA-ITA with several related methods designed for pair-wise source-target adaptation:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 478, + 522, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 478, + 522, + 490 + ], + "score": 1.0, + "content": "CORAL (Sun & Saenko (2016)), DANN (Ganin & Lempitsky (2015)), ADDA (Tzeng et al. (2017)), DTN", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 488, + 521, + 501 + ], + "spans": [ + { + "bbox": [ + 89, + 488, + 521, + 501 + ], + "score": 1.0, + "content": "(Zhang et al. (2015)), UNIT (Liu et al. (2017)), PixelDA (Bousmalis et al. (2017)), and DSN (Bousmalis", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 499, + 522, + 513 + ], + "spans": [ + { + "bbox": [ + 89, + 499, + 522, + 513 + ], + "score": 1.0, + "content": "et al. (2016)). We reported the results of two following baselines: (i) one is to combine all the target domains", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 510, + 522, + 523 + ], + "spans": [ + { + "bbox": [ + 89, + 510, + 522, + 523 + ], + "score": 1.0, + "content": "into a single one and train it using MTDA-ITA, which we denote as (c-MTDA-ITA). (ii) the other one is to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 521, + 523, + 535 + ], + "spans": [ + { + "bbox": [ + 89, + 521, + 523, + 535 + ], + "score": 1.0, + "content": "train multiple MTDA-ITA separately, where each one corresponds to a source-target pair which we denote as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 532, + 521, + 546 + ], + "spans": [ + { + "bbox": [ + 89, + 532, + 521, + 546 + ], + "score": 1.0, + "content": "(s-MTDA-ITA). For completeness, we reported the results of the competing methods by combining all the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 543, + 521, + 556 + ], + "spans": [ + { + "bbox": [ + 89, + 543, + 521, + 556 + ], + "score": 1.0, + "content": "target domains into a single one (denoted by c-DTN, c-ADDA, and c-DSN) as well. We also extend DSN to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 90, + 555, + 521, + 567 + ], + "spans": [ + { + "bbox": [ + 90, + 555, + 521, + 567 + ], + "score": 1.0, + "content": "multiple domains by (i) having one private encoder for all domains denoted by (1p-DSN), (ii) adding multiple", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 89, + 565, + 408, + 578 + ], + "spans": [ + { + "bbox": [ + 89, + 565, + 408, + 578 + ], + "score": 1.0, + "content": "private encoders to it denoted by (mp-DSN) and contrast them with our model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 91, + 590, + 192, + 601 + ], + "lines": [ + { + "bbox": [ + 90, + 589, + 194, + 603 + ], + "spans": [ + { + "bbox": [ + 90, + 589, + 194, + 603 + ], + "score": 1.0, + "content": "4.1 DIGITS DATASETS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 92, + 610, + 521, + 633 + ], + "lines": [ + { + "bbox": [ + 90, + 610, + 523, + 623 + ], + "spans": [ + { + "bbox": [ + 90, + 610, + 523, + 623 + ], + "score": 1.0, + "content": "We combine four popular digits datasets (MNIST, MNIST-M, SVHN, and USPS) to build the multi-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 89, + 622, + 523, + 633 + ], + "spans": [ + { + "bbox": [ + 89, + 622, + 356, + 633 + ], + "score": 1.0, + "content": "target domain dataset. All images were uniformly rescaled to", + "type": "text" + }, + { + "bbox": [ + 356, + 622, + 392, + 632 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 622, + 523, + 633 + ], + "score": 1.0, + "content": ". We take each of MNIST-M,", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 102, + 641, + 371, + 651 + ], + "lines": [ + { + "bbox": [ + 103, + 639, + 372, + 653 + ], + "spans": [ + { + "bbox": [ + 103, + 639, + 372, + 653 + ], + "score": 1.0, + "content": "2http://www.cs.cmu.edu/afs/cs/project/PIE/MultiPie/Multi-Pie/Home.html", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 292, + 679 + ], + "lines": [ + { + "bbox": [ + 285, + 670, + 293, + 682 + ], + "spans": [ + { + "bbox": [ + 285, + 670, + 293, + 682 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 90, + 95, + 521, + 195 + ], + "lines": [], + "index": 4, + "bbox_fs": [ + 88, + 96, + 522, + 196 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 92, + 210, + 240, + 223 + ], + "lines": [ + { + "bbox": [ + 89, + 210, + 241, + 225 + ], + "spans": [ + { + "bbox": [ + 89, + 210, + 241, + 225 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 91, + 235, + 522, + 324 + ], + "lines": [ + { + "bbox": [ + 90, + 235, + 521, + 247 + ], + "spans": [ + { + "bbox": [ + 90, + 235, + 521, + 247 + ], + "score": 1.0, + "content": "We compare the proposed method with state-of-the-art methods on standard benchmark datasets: a digit", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 246, + 522, + 259 + ], + "spans": [ + { + "bbox": [ + 90, + 246, + 522, + 259 + ], + "score": 1.0, + "content": "classification task that includes 4 datasets: MNIST (LeCun et al. (1998)), MNIST-M (Ganin et al. (2016)),", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 257, + 522, + 271 + ], + "spans": [ + { + "bbox": [ + 89, + 257, + 522, + 271 + ], + "score": 1.0, + "content": "SVHN (Netzer et al. (2011)), USPS (Tzeng et al. (2017)), Multi-PIE expression recognition dataset2, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 267, + 523, + 281 + ], + "spans": [ + { + "bbox": [ + 88, + 267, + 523, + 281 + ], + "score": 1.0, + "content": "PACS multi-domain image recognition benchmark (Li et al. (2017)), a new dataset designed for the cross-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 280, + 522, + 292 + ], + "spans": [ + { + "bbox": [ + 90, + 280, + 522, + 292 + ], + "score": 1.0, + "content": "domain recognition problems (the details for this experiment is available in Appendix G). Fig. 3 illustrates", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 290, + 521, + 303 + ], + "spans": [ + { + "bbox": [ + 89, + 290, + 521, + 303 + ], + "score": 1.0, + "content": "image samples from different datasets and domains. We evaluate the performance of all methods with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 301, + 520, + 313 + ], + "spans": [ + { + "bbox": [ + 90, + 301, + 520, + 313 + ], + "score": 1.0, + "content": "classification accuracy metric. We repeated each experiment 5 times and report the average and the standard", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 311, + 195, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 311, + 195, + 326 + ], + "score": 1.0, + "content": "deviation of the accuracy.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 88, + 235, + 523, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 329, + 522, + 428 + ], + "lines": [ + { + "bbox": [ + 89, + 329, + 522, + 341 + ], + "spans": [ + { + "bbox": [ + 89, + 329, + 522, + 341 + ], + "score": 1.0, + "content": "We used ADAM (Kingma & Ba (2015)) for training; the learning rate was set to 0.0002 and momentum", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 339, + 524, + 353 + ], + "spans": [ + { + "bbox": [ + 88, + 339, + 524, + 353 + ], + "score": 1.0, + "content": "parameters to 0.5 and 0.999. We used batches of size 16 from each domain, and the input images were mean-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 350, + 523, + 365 + ], + "spans": [ + { + "bbox": [ + 88, + 350, + 174, + 365 + ], + "score": 1.0, + "content": "centered/rescaled to", + "type": "text" + }, + { + "bbox": [ + 174, + 351, + 203, + 363 + ], + "score": 0.81, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 350, + 389, + 365 + ], + "score": 1.0, + "content": ". The hyper-parameters are empirically set as", + "type": "text" + }, + { + "bbox": [ + 389, + 351, + 473, + 362 + ], + "score": 0.71, + "content": "\\lambda _ { r } = 1 . 0 , \\lambda _ { c } = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 350, + 476, + 365 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 476, + 351, + 519, + 362 + ], + "score": 0.72, + "content": "\\lambda _ { d } = 0 . 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 519, + 350, + 523, + 365 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 362, + 522, + 374 + ], + "spans": [ + { + "bbox": [ + 90, + 362, + 522, + 374 + ], + "score": 1.0, + "content": "For the network architecture, our private/shared encoders consisted of three convolutional layers as the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 90, + 374, + 521, + 384 + ], + "spans": [ + { + "bbox": [ + 90, + 374, + 521, + 384 + ], + "score": 1.0, + "content": "front-end and four basic residual blocks as the back-end. The decoder consisted of four basic residual blocks", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 384, + 522, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 384, + 522, + 396 + ], + "score": 1.0, + "content": "as the front-end and four transposed convolutional layers as the back-end. The discriminator and the classifier", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 90, + 396, + 520, + 406 + ], + "spans": [ + { + "bbox": [ + 90, + 396, + 520, + 406 + ], + "score": 1.0, + "content": "consisted of stacks of convolutional layers. We used ReLU for nonlinearity. Tanh function is used as the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 406, + 521, + 418 + ], + "spans": [ + { + "bbox": [ + 89, + 406, + 286, + 418 + ], + "score": 1.0, + "content": "activation function of the last layer in the decoder", + "type": "text" + }, + { + "bbox": [ + 286, + 406, + 295, + 416 + ], + "score": 0.8, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 406, + 417, + 418 + ], + "score": 1.0, + "content": "for scaling the output pixels to", + "type": "text" + }, + { + "bbox": [ + 418, + 406, + 446, + 418 + ], + "score": 0.88, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 406, + 521, + 418 + ], + "score": 1.0, + "content": ". The details of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 88, + 417, + 232, + 429 + ], + "spans": [ + { + "bbox": [ + 88, + 417, + 232, + 429 + ], + "score": 1.0, + "content": "networks are given in Appendix D.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 88, + 329, + 524, + 429 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 434, + 521, + 576 + ], + "lines": [ + { + "bbox": [ + 89, + 434, + 522, + 447 + ], + "spans": [ + { + "bbox": [ + 89, + 434, + 522, + 447 + ], + "score": 1.0, + "content": "The quantitative evaluation involves a comparison of the performance of our model to previous work and to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 444, + 522, + 457 + ], + "spans": [ + { + "bbox": [ + 90, + 444, + 522, + 457 + ], + "score": 1.0, + "content": "Source Only and 1-NN baselines that do not use any domain adaptation. For Source Only baseline, we train", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 456, + 521, + 469 + ], + "spans": [ + { + "bbox": [ + 89, + 456, + 521, + 469 + ], + "score": 1.0, + "content": "our model only on the unaltered source training data and evaluate on the target test data. We compare the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 466, + 523, + 480 + ], + "spans": [ + { + "bbox": [ + 89, + 466, + 523, + 480 + ], + "score": 1.0, + "content": "proposed method MTDA-ITA with several related methods designed for pair-wise source-target adaptation:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 478, + 522, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 478, + 522, + 490 + ], + "score": 1.0, + "content": "CORAL (Sun & Saenko (2016)), DANN (Ganin & Lempitsky (2015)), ADDA (Tzeng et al. (2017)), DTN", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 488, + 521, + 501 + ], + "spans": [ + { + "bbox": [ + 89, + 488, + 521, + 501 + ], + "score": 1.0, + "content": "(Zhang et al. (2015)), UNIT (Liu et al. (2017)), PixelDA (Bousmalis et al. (2017)), and DSN (Bousmalis", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 499, + 522, + 513 + ], + "spans": [ + { + "bbox": [ + 89, + 499, + 522, + 513 + ], + "score": 1.0, + "content": "et al. (2016)). We reported the results of two following baselines: (i) one is to combine all the target domains", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 510, + 522, + 523 + ], + "spans": [ + { + "bbox": [ + 89, + 510, + 522, + 523 + ], + "score": 1.0, + "content": "into a single one and train it using MTDA-ITA, which we denote as (c-MTDA-ITA). (ii) the other one is to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 521, + 523, + 535 + ], + "spans": [ + { + "bbox": [ + 89, + 521, + 523, + 535 + ], + "score": 1.0, + "content": "train multiple MTDA-ITA separately, where each one corresponds to a source-target pair which we denote as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 532, + 521, + 546 + ], + "spans": [ + { + "bbox": [ + 89, + 532, + 521, + 546 + ], + "score": 1.0, + "content": "(s-MTDA-ITA). For completeness, we reported the results of the competing methods by combining all the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 89, + 543, + 521, + 556 + ], + "spans": [ + { + "bbox": [ + 89, + 543, + 521, + 556 + ], + "score": 1.0, + "content": "target domains into a single one (denoted by c-DTN, c-ADDA, and c-DSN) as well. We also extend DSN to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 90, + 555, + 521, + 567 + ], + "spans": [ + { + "bbox": [ + 90, + 555, + 521, + 567 + ], + "score": 1.0, + "content": "multiple domains by (i) having one private encoder for all domains denoted by (1p-DSN), (ii) adding multiple", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 89, + 565, + 408, + 578 + ], + "spans": [ + { + "bbox": [ + 89, + 565, + 408, + 578 + ], + "score": 1.0, + "content": "private encoders to it denoted by (mp-DSN) and contrast them with our model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33, + "bbox_fs": [ + 89, + 434, + 523, + 578 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 590, + 192, + 601 + ], + "lines": [ + { + "bbox": [ + 90, + 589, + 194, + 603 + ], + "spans": [ + { + "bbox": [ + 90, + 589, + 194, + 603 + ], + "score": 1.0, + "content": "4.1 DIGITS DATASETS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 92, + 610, + 521, + 633 + ], + "lines": [ + { + "bbox": [ + 90, + 610, + 523, + 623 + ], + "spans": [ + { + "bbox": [ + 90, + 610, + 523, + 623 + ], + "score": 1.0, + "content": "We combine four popular digits datasets (MNIST, MNIST-M, SVHN, and USPS) to build the multi-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 89, + 622, + 523, + 633 + ], + "spans": [ + { + "bbox": [ + 89, + 622, + 356, + 633 + ], + "score": 1.0, + "content": "target domain dataset. All images were uniformly rescaled to", + "type": "text" + }, + { + "bbox": [ + 356, + 622, + 392, + 632 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 622, + 523, + 633 + ], + "score": 1.0, + "content": ". We take each of MNIST-M,", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 89, + 610, + 523, + 633 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 72, + 476, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 132, + 72, + 476, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 72, + 476, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 476, + 187 + ], + "score": 0.976, + "type": "image", + "image_path": "c17f5da419bdc0deb50b99f30b19766556f98c280148111ece40ad50497e48c6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 132, + 72, + 476, + 110.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 132, + 110.33333333333334, + 476, + 148.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 132, + 148.66666666666669, + 476, + 187.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 91, + 190, + 520, + 230 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 188, + 523, + 203 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 523, + 203 + ], + "score": 1.0, + "content": "Figure 3: Exemplary images from different datasets. a) Digits datasets, b) PACS datatset (first row: Art-painting,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 199, + 521, + 211 + ], + "spans": [ + { + "bbox": [ + 89, + 199, + 521, + 211 + ], + "score": 1.0, + "content": "second row: Cartoon, Third row: Photo, last row: Sketch), c) Multi-PIE dataset (each row corresponds to a different", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 209, + 522, + 222 + ], + "spans": [ + { + "bbox": [ + 89, + 209, + 522, + 222 + ], + "score": 1.0, + "content": "camera angle and each subject depicts an expression(normal, smile, surprise, squint, disgust, scream) at every camera", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 219, + 127, + 231 + ], + "spans": [ + { + "bbox": [ + 88, + 219, + 127, + 231 + ], + "score": 1.0, + "content": "position).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 90, + 248, + 521, + 458 + ], + "lines": [ + { + "bbox": [ + 88, + 248, + 522, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 248, + 522, + 262 + ], + "score": 1.0, + "content": "SVHN, USPS, and MNIST as source domain in turn, and the rest as targets. We use all labeled source", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 259, + 522, + 273 + ], + "spans": [ + { + "bbox": [ + 88, + 259, + 522, + 273 + ], + "score": 1.0, + "content": "images and all unlabeled target images, following the standard evaluation protocol for unsupervised domain", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 271, + 523, + 284 + ], + "spans": [ + { + "bbox": [ + 88, + 271, + 523, + 284 + ], + "score": 1.0, + "content": "adaptation (Ganin et al. (2016); Long et al. (2016). We show the accuracy of different methods in Table 1.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 281, + 522, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 281, + 522, + 295 + ], + "score": 1.0, + "content": "Additional results are available in Appendix F. The results show that first of all cMTDA-ITA has worse", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 292, + 522, + 305 + ], + "spans": [ + { + "bbox": [ + 88, + 292, + 522, + 305 + ], + "score": 1.0, + "content": "performance than sMTDA-ITA and MTDA-ITA. We have similar observations for ADDA, DTN, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 303, + 522, + 316 + ], + "spans": [ + { + "bbox": [ + 89, + 303, + 522, + 316 + ], + "score": 1.0, + "content": "DSN that demonstrates a naive combination of different target datasets can sometimes even decrease the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 315, + 522, + 327 + ], + "spans": [ + { + "bbox": [ + 88, + 315, + 522, + 327 + ], + "score": 1.0, + "content": "performance of the competing methods. Furthermore, MTDA-ITA outperforms the state-of-the-art methods", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 326, + 522, + 338 + ], + "spans": [ + { + "bbox": [ + 89, + 326, + 522, + 338 + ], + "score": 1.0, + "content": "in most of domain transformations. The higher performance of MTDA-ITA compared to other methods is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 336, + 522, + 349 + ], + "spans": [ + { + "bbox": [ + 88, + 336, + 522, + 349 + ], + "score": 1.0, + "content": "mainly attributed to the joint adaptation of related domains where each domain could benefit of other related", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 346, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 89, + 346, + 522, + 362 + ], + "score": 1.0, + "content": "domains. Furthermore, from the results obtained, we see that it is beneficial to use information coming", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 359, + 521, + 371 + ], + "spans": [ + { + "bbox": [ + 89, + 359, + 349, + 371 + ], + "score": 1.0, + "content": "from unlabeled target data (see Eq. 8 for updating the classifier", + "type": "text" + }, + { + "bbox": [ + 349, + 359, + 358, + 369 + ], + "score": 0.73, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 359, + 521, + 371 + ], + "score": 1.0, + "content": ") during the learning process, compared", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 369, + 523, + 383 + ], + "spans": [ + { + "bbox": [ + 89, + 369, + 523, + 383 + ], + "score": 1.0, + "content": "to when no data from target domain is used (See the ablation study section for more information). Indeed,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 380, + 522, + 393 + ], + "spans": [ + { + "bbox": [ + 89, + 380, + 522, + 393 + ], + "score": 1.0, + "content": "using our scheme, we find a representation space in which embeds the knowledge from the target domain", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 89, + 391, + 522, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 391, + 522, + 405 + ], + "score": 1.0, + "content": "into the learned classifier. By contrast, the competing methods do not provide a principled way of sharing", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 402, + 521, + 416 + ], + "spans": [ + { + "bbox": [ + 88, + 402, + 521, + 416 + ], + "score": 1.0, + "content": "information across all domains, leading to overall lower performance. The results also verify the superiority", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 413, + 522, + 426 + ], + "spans": [ + { + "bbox": [ + 89, + 413, + 522, + 426 + ], + "score": 1.0, + "content": "of MTDA-ITA over both mp-DSN, and 1p-DSN. This can be due to (i) having multiple private encoders", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 425, + 521, + 437 + ], + "spans": [ + { + "bbox": [ + 89, + 425, + 521, + 437 + ], + "score": 1.0, + "content": "increase the number of parameters that may lead to mp-DSN overfitting, (ii) superiority of the MTDA-ITA’s", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 435, + 522, + 448 + ], + "spans": [ + { + "bbox": [ + 89, + 435, + 522, + 448 + ], + "score": 1.0, + "content": "domain adversarial loss over the DSN’s MMD loss to separate the shared and private features, (iii) utilization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 88, + 446, + 372, + 459 + ], + "spans": [ + { + "bbox": [ + 88, + 446, + 372, + 459 + ], + "score": 1.0, + "content": "of the unlabeled target data to regularize the classifier in MTDA-ITA.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 91, + 499, + 204, + 510 + ], + "lines": [ + { + "bbox": [ + 89, + 497, + 207, + 512 + ], + "spans": [ + { + "bbox": [ + 89, + 497, + 207, + 512 + ], + "score": 1.0, + "content": "4.2 MULTI-PIE DATASET", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 90, + 530, + 522, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 530, + 523, + 543 + ], + "spans": [ + { + "bbox": [ + 89, + 530, + 523, + 543 + ], + "score": 1.0, + "content": "The Multi-PIE dataset includes face images of 337 individuals captured from different expressions, views,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 500, + 554 + ], + "score": 1.0, + "content": "and illumination conditions (fig. 3(c)). For this experiment, we use 5 different camera views (positions)", + "type": "text" + }, + { + "bbox": [ + 500, + 542, + 519, + 552 + ], + "score": 0.81, + "content": "C 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 519, + 541, + 522, + 554 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 90, + 551, + 523, + 565 + ], + "spans": [ + { + "bbox": [ + 90, + 552, + 109, + 563 + ], + "score": 0.64, + "content": "C 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 110, + 551, + 113, + 565 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 113, + 552, + 132, + 563 + ], + "score": 0.67, + "content": "C 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 551, + 135, + 565 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 136, + 552, + 155, + 563 + ], + "score": 0.59, + "content": "C 1 3", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 551, + 174, + 565 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 175, + 552, + 194, + 563 + ], + "score": 0.83, + "content": "C 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 551, + 523, + 565 + ], + "score": 1.0, + "content": "as different domains (Fig. 3(c)) and the face expressions (normal, smile, surprise,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 564, + 522, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 564, + 409, + 576 + ], + "score": 1.0, + "content": "squint, disgust, scream) as labels. Each domain contains 27120 images of size", + "type": "text" + }, + { + "bbox": [ + 410, + 564, + 460, + 574 + ], + "score": 0.92, + "content": "6 4 \\times 6 4 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 564, + 522, + 576 + ], + "score": 1.0, + "content": ". We used each", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 574, + 522, + 587 + ], + "spans": [ + { + "bbox": [ + 89, + 574, + 522, + 587 + ], + "score": 1.0, + "content": "view as the source domain, in turn, and the rest as targets. We expect the face inclination angle to reflect", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 90, + 586, + 522, + 598 + ], + "spans": [ + { + "bbox": [ + 90, + 586, + 421, + 598 + ], + "score": 1.0, + "content": "the complexity of transfer learning. Tab. 2 shows the classification accuracy for", + "type": "text" + }, + { + "bbox": [ + 422, + 586, + 441, + 596 + ], + "score": 0.81, + "content": "C 1 3", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 586, + 460, + 598 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 460, + 586, + 479, + 596 + ], + "score": 0.84, + "content": "C 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 586, + 522, + 598 + ], + "score": 1.0, + "content": "as source", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 88, + 595, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 88, + 595, + 209, + 609 + ], + "score": 1.0, + "content": "domain (the results for views", + "type": "text" + }, + { + "bbox": [ + 209, + 596, + 228, + 607 + ], + "score": 0.8, + "content": "C 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 595, + 231, + 609 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 231, + 596, + 251, + 607 + ], + "score": 0.81, + "content": "C 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 595, + 268, + 609 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 268, + 596, + 288, + 606 + ], + "score": 0.85, + "content": "C 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 595, + 522, + 609 + ], + "score": 1.0, + "content": "as source domain are available in Appendix F). As can be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "score": 1.0, + "content": "seen, MTDA-ITA achieves the best performances as well as the best scores in most settings that verifies the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 618, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 522, + 631 + ], + "score": 1.0, + "content": "effectiveness of MTDA-ITA for multi-target domain adaptation. Clearly, with the increasing camera angle,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 90, + 630, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 630, + 522, + 642 + ], + "score": 1.0, + "content": "the image structure changes up to a certain extent (the views become heterogeneous). However, our method", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 88, + 640, + 365, + 653 + ], + "spans": [ + { + "bbox": [ + 88, + 640, + 365, + 653 + ], + "score": 1.0, + "content": "produces better results even under such very challenging conditions.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 292, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 670, + 293, + 681 + ], + "spans": [ + { + "bbox": [ + 285, + 670, + 293, + 681 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 72, + 476, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 132, + 72, + 476, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 72, + 476, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 476, + 187 + ], + "score": 0.976, + "type": "image", + "image_path": "c17f5da419bdc0deb50b99f30b19766556f98c280148111ece40ad50497e48c6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 132, + 72, + 476, + 110.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 132, + 110.33333333333334, + 476, + 148.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 132, + 148.66666666666669, + 476, + 187.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 91, + 190, + 520, + 230 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 188, + 523, + 203 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 523, + 203 + ], + "score": 1.0, + "content": "Figure 3: Exemplary images from different datasets. a) Digits datasets, b) PACS datatset (first row: Art-painting,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 199, + 521, + 211 + ], + "spans": [ + { + "bbox": [ + 89, + 199, + 521, + 211 + ], + "score": 1.0, + "content": "second row: Cartoon, Third row: Photo, last row: Sketch), c) Multi-PIE dataset (each row corresponds to a different", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 209, + 522, + 222 + ], + "spans": [ + { + "bbox": [ + 89, + 209, + 522, + 222 + ], + "score": 1.0, + "content": "camera angle and each subject depicts an expression(normal, smile, surprise, squint, disgust, scream) at every camera", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 219, + 127, + 231 + ], + "spans": [ + { + "bbox": [ + 88, + 219, + 127, + 231 + ], + "score": 1.0, + "content": "position).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 90, + 248, + 521, + 458 + ], + "lines": [ + { + "bbox": [ + 88, + 248, + 522, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 248, + 522, + 262 + ], + "score": 1.0, + "content": "SVHN, USPS, and MNIST as source domain in turn, and the rest as targets. We use all labeled source", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 259, + 522, + 273 + ], + "spans": [ + { + "bbox": [ + 88, + 259, + 522, + 273 + ], + "score": 1.0, + "content": "images and all unlabeled target images, following the standard evaluation protocol for unsupervised domain", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 271, + 523, + 284 + ], + "spans": [ + { + "bbox": [ + 88, + 271, + 523, + 284 + ], + "score": 1.0, + "content": "adaptation (Ganin et al. (2016); Long et al. (2016). We show the accuracy of different methods in Table 1.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 281, + 522, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 281, + 522, + 295 + ], + "score": 1.0, + "content": "Additional results are available in Appendix F. The results show that first of all cMTDA-ITA has worse", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 292, + 522, + 305 + ], + "spans": [ + { + "bbox": [ + 88, + 292, + 522, + 305 + ], + "score": 1.0, + "content": "performance than sMTDA-ITA and MTDA-ITA. We have similar observations for ADDA, DTN, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 303, + 522, + 316 + ], + "spans": [ + { + "bbox": [ + 89, + 303, + 522, + 316 + ], + "score": 1.0, + "content": "DSN that demonstrates a naive combination of different target datasets can sometimes even decrease the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 315, + 522, + 327 + ], + "spans": [ + { + "bbox": [ + 88, + 315, + 522, + 327 + ], + "score": 1.0, + "content": "performance of the competing methods. Furthermore, MTDA-ITA outperforms the state-of-the-art methods", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 326, + 522, + 338 + ], + "spans": [ + { + "bbox": [ + 89, + 326, + 522, + 338 + ], + "score": 1.0, + "content": "in most of domain transformations. The higher performance of MTDA-ITA compared to other methods is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 336, + 522, + 349 + ], + "spans": [ + { + "bbox": [ + 88, + 336, + 522, + 349 + ], + "score": 1.0, + "content": "mainly attributed to the joint adaptation of related domains where each domain could benefit of other related", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 346, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 89, + 346, + 522, + 362 + ], + "score": 1.0, + "content": "domains. Furthermore, from the results obtained, we see that it is beneficial to use information coming", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 359, + 521, + 371 + ], + "spans": [ + { + "bbox": [ + 89, + 359, + 349, + 371 + ], + "score": 1.0, + "content": "from unlabeled target data (see Eq. 8 for updating the classifier", + "type": "text" + }, + { + "bbox": [ + 349, + 359, + 358, + 369 + ], + "score": 0.73, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 359, + 521, + 371 + ], + "score": 1.0, + "content": ") during the learning process, compared", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 369, + 523, + 383 + ], + "spans": [ + { + "bbox": [ + 89, + 369, + 523, + 383 + ], + "score": 1.0, + "content": "to when no data from target domain is used (See the ablation study section for more information). Indeed,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 380, + 522, + 393 + ], + "spans": [ + { + "bbox": [ + 89, + 380, + 522, + 393 + ], + "score": 1.0, + "content": "using our scheme, we find a representation space in which embeds the knowledge from the target domain", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 89, + 391, + 522, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 391, + 522, + 405 + ], + "score": 1.0, + "content": "into the learned classifier. By contrast, the competing methods do not provide a principled way of sharing", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 88, + 402, + 521, + 416 + ], + "spans": [ + { + "bbox": [ + 88, + 402, + 521, + 416 + ], + "score": 1.0, + "content": "information across all domains, leading to overall lower performance. The results also verify the superiority", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 413, + 522, + 426 + ], + "spans": [ + { + "bbox": [ + 89, + 413, + 522, + 426 + ], + "score": 1.0, + "content": "of MTDA-ITA over both mp-DSN, and 1p-DSN. This can be due to (i) having multiple private encoders", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 425, + 521, + 437 + ], + "spans": [ + { + "bbox": [ + 89, + 425, + 521, + 437 + ], + "score": 1.0, + "content": "increase the number of parameters that may lead to mp-DSN overfitting, (ii) superiority of the MTDA-ITA’s", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 435, + 522, + 448 + ], + "spans": [ + { + "bbox": [ + 89, + 435, + 522, + 448 + ], + "score": 1.0, + "content": "domain adversarial loss over the DSN’s MMD loss to separate the shared and private features, (iii) utilization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 88, + 446, + 372, + 459 + ], + "spans": [ + { + "bbox": [ + 88, + 446, + 372, + 459 + ], + "score": 1.0, + "content": "of the unlabeled target data to regularize the classifier in MTDA-ITA.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 16, + "bbox_fs": [ + 88, + 248, + 523, + 459 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 499, + 204, + 510 + ], + "lines": [ + { + "bbox": [ + 89, + 497, + 207, + 512 + ], + "spans": [ + { + "bbox": [ + 89, + 497, + 207, + 512 + ], + "score": 1.0, + "content": "4.2 MULTI-PIE DATASET", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 90, + 530, + 522, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 530, + 523, + 543 + ], + "spans": [ + { + "bbox": [ + 89, + 530, + 523, + 543 + ], + "score": 1.0, + "content": "The Multi-PIE dataset includes face images of 337 individuals captured from different expressions, views,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 500, + 554 + ], + "score": 1.0, + "content": "and illumination conditions (fig. 3(c)). For this experiment, we use 5 different camera views (positions)", + "type": "text" + }, + { + "bbox": [ + 500, + 542, + 519, + 552 + ], + "score": 0.81, + "content": "C 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 519, + 541, + 522, + 554 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 90, + 551, + 523, + 565 + ], + "spans": [ + { + "bbox": [ + 90, + 552, + 109, + 563 + ], + "score": 0.64, + "content": "C 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 110, + 551, + 113, + 565 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 113, + 552, + 132, + 563 + ], + "score": 0.67, + "content": "C 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 551, + 135, + 565 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 136, + 552, + 155, + 563 + ], + "score": 0.59, + "content": "C 1 3", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 551, + 174, + 565 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 175, + 552, + 194, + 563 + ], + "score": 0.83, + "content": "C 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 551, + 523, + 565 + ], + "score": 1.0, + "content": "as different domains (Fig. 3(c)) and the face expressions (normal, smile, surprise,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 564, + 522, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 564, + 409, + 576 + ], + "score": 1.0, + "content": "squint, disgust, scream) as labels. Each domain contains 27120 images of size", + "type": "text" + }, + { + "bbox": [ + 410, + 564, + 460, + 574 + ], + "score": 0.92, + "content": "6 4 \\times 6 4 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 564, + 522, + 576 + ], + "score": 1.0, + "content": ". We used each", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 574, + 522, + 587 + ], + "spans": [ + { + "bbox": [ + 89, + 574, + 522, + 587 + ], + "score": 1.0, + "content": "view as the source domain, in turn, and the rest as targets. We expect the face inclination angle to reflect", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 90, + 586, + 522, + 598 + ], + "spans": [ + { + "bbox": [ + 90, + 586, + 421, + 598 + ], + "score": 1.0, + "content": "the complexity of transfer learning. Tab. 2 shows the classification accuracy for", + "type": "text" + }, + { + "bbox": [ + 422, + 586, + 441, + 596 + ], + "score": 0.81, + "content": "C 1 3", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 586, + 460, + 598 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 460, + 586, + 479, + 596 + ], + "score": 0.84, + "content": "C 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 586, + 522, + 598 + ], + "score": 1.0, + "content": "as source", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 88, + 595, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 88, + 595, + 209, + 609 + ], + "score": 1.0, + "content": "domain (the results for views", + "type": "text" + }, + { + "bbox": [ + 209, + 596, + 228, + 607 + ], + "score": 0.8, + "content": "C 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 595, + 231, + 609 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 231, + 596, + 251, + 607 + ], + "score": 0.81, + "content": "C 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 595, + 268, + 609 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 268, + 596, + 288, + 606 + ], + "score": 0.85, + "content": "C 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 595, + 522, + 609 + ], + "score": 1.0, + "content": "as source domain are available in Appendix F). As can be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "score": 1.0, + "content": "seen, MTDA-ITA achieves the best performances as well as the best scores in most settings that verifies the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 618, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 522, + 631 + ], + "score": 1.0, + "content": "effectiveness of MTDA-ITA for multi-target domain adaptation. Clearly, with the increasing camera angle,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 90, + 630, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 630, + 522, + 642 + ], + "score": 1.0, + "content": "the image structure changes up to a certain extent (the views become heterogeneous). However, our method", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 88, + 640, + 365, + 653 + ], + "spans": [ + { + "bbox": [ + 88, + 640, + 365, + 653 + ], + "score": 1.0, + "content": "produces better results even under such very challenging conditions.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32, + "bbox_fs": [ + 88, + 530, + 523, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 93, + 520, + 249 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 93, + 520, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 93, + 520, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 93, + 520, + 249 + ], + "score": 0.981, + "html": "
methodS→MS→MMS→UM→SM→MMM→UAve.ranking
Source Only62.10±0.6040.43±0.7039.90± 0.6030.29± 0.5955.98 ± 0.4878.30±0.3814.00
1-NN35.8618.2129.3128.0112.5841.2215.00
CORAL (Sun & Saenko (2016))63.10 ± 0.6154.37 ± 0.5350.15 ± 0.6333.40 ±0.7457.70 ± 0.6981.05 ± 0.8011.33
DANN (Ganin et al. (2016))73.80 ±0.4961.05±0.8062.54 ± 0.9135.50 ± 0.6577.40±0.7381.60 ± 0.608.75
ADDA(Tzeng et al. (2017))77.68 ± 0.9264.23±0.7064.10±0.7930.04± 0.9891.47 ±1.090.51 ± 0.806.43
C-ADDA80.10 ± 0.6956.80 ± 0.7964.80±0.8827.50 ± 0.8683.30 ±0.9084.10 ±0.988.95
DTN (Zhang et al. (2015))81.40± 0.4263.70±0.3960.12 ± 0.5240.40±0.5085.70±0.3985.80 ± 0.466.04
c-DTN PixelDA (Bousmalis et al. (2017))82.10 ±0.6259.30 ± 0.5956.87 ± 0.6538.32 ± 0.5080.90±0.8079.31 ±0.787.96
UNIT (Liu et al. (2017))98.10*94.10*1
DSN (Bousmalis et al. (2016))90.6*92.901
c-DSN82.70±0.37 83.10 ±0.2064.80± 0.4065.30± 0.2849.30±0.3083.20±0.3091.65 ± 0.402.85
1p-DSN81.00 ± 0.4760.56 ± 0.36 58.22 ± 0.6860.35 ± 0.5946.80 ± 0.4580.49±0.4088.21 ±0.384.84
mp-DSN83.40±0.3061.00±0.5058.06 ±0.48 58.10 ±0.6445.11 ± 0.3377.33 ± 0.52 79.30±0.5985.16 ± 0.634.90
s-MTDA-ITA82.90±0.1363.10±0.2863.54±0.3047.35 ± 0.4082.42± 0.1986.45 ± 0.715.33
c-MTDA-ITA79.20±0.2859.90 ± 0.3063.70±0.2649.60± 0.25 45.30 ±0.3077.12 ± 0.2289.21±0.282.88
MTDA-ITA84.60±0.2465.30 ± 0.1570.03 ±0.2052.01 ± 0.2185.50 ±0.1887.47 ± 0.25 94.20±0.204.25
1.16
", + "type": "table", + "image_path": "8fd442a94d88d4c3187f10668a4eaf7e8e6a87c4149cab928e7e3f5d35839d41.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 90, + 93, + 520, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 90, + 145.0, + 520, + 197.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 90, + 197.0, + 520, + 249.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 90, + 263, + 522, + 325 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 263, + 522, + 276 + ], + "spans": [ + { + "bbox": [ + 89, + 263, + 522, + 276 + ], + "score": 1.0, + "content": "Table 1: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 273, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 273, + 523, + 285 + ], + "score": 1.0, + "content": "in red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDA-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 283, + 523, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 283, + 523, + 295 + ], + "score": 1.0, + "content": "ITA where each one correspond to a source-target pair. 1p-DSN: extended DSN with single private encoder.mp-DSN:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 294, + 521, + 305 + ], + "spans": [ + { + "bbox": [ + 90, + 294, + 521, + 305 + ], + "score": 1.0, + "content": "extended DSN with multiple private encoder. Last column shows the average rank of each method over all adaptation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 304, + 522, + 316 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 110, + 316 + ], + "score": 1.0, + "content": "pairs.", + "type": "text" + }, + { + "bbox": [ + 110, + 304, + 138, + 314 + ], + "score": 0.37, + "content": "\\bf \\Pi ^ { * } U N I T", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 304, + 255, + 316 + ], + "score": 1.0, + "content": "trains with the extended SVHN", + "type": "text" + }, + { + "bbox": [ + 255, + 304, + 287, + 314 + ], + "score": 0.85, + "content": "> 5 0 0 \\mathrm { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 304, + 421, + 316 + ], + "score": 1.0, + "content": "images vs ours 72K). *PixelDA uses", + "type": "text" + }, + { + "bbox": [ + 421, + 304, + 460, + 315 + ], + "score": 0.85, + "content": "( \\approx 1 , 0 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 304, + 522, + 316 + ], + "score": 1.0, + "content": "of labeled target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 313, + 321, + 325 + ], + "spans": [ + { + "bbox": [ + 89, + 313, + 321, + 325 + ], + "score": 1.0, + "content": "domain data as a validation set for tuning the hyper-parameters.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 90, + 339, + 520, + 484 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 339, + 520, + 484 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 90, + 339, + 520, + 484 + ], + "spans": [ + { + "bbox": [ + 90, + 339, + 520, + 484 + ], + "score": 0.978, + "html": "
methodC13→C05C13→C08C13→C09C13→C14C14→C05C14→C08C14→C09C14→C13Ave.ranking
Source Only50.79± 0.3345.90± 0.5040.04±0.4059.68± 0.2960.03±0.5536.80± 0.6140.11 ±0.5060.57± 0.3616.08
1-NN CORAL33.21 54.89 ± 0.5237.0134.4548.79 68.90± 0.3547.4428.2430.8644.8617.00
DANN57.86± 0.4148.90±0.48 50.30 ± 0.4340.30±0.53 45.30± 0.5070.68 ± 0.3559.98 ± 0.45 57.20± 0.4540.63±0.55 40.22±0.5540.80 ±0.53 40.77 ± 0.4565.11 ± 0.45 70.50± 0.5511.95 9.92
ADDA64.83±0.6963.20±0.4555.48 ±0.6574.25±0.5573.62 ± 0.7543.56±0.9538.68 ±0.9572.84±0.759.33
C-ADDA59.20±0.2530.70±0.6353.20±0.4068.33±0.3565.88±0.3830.60±0.6145.34±0.4864.30±0.4011.50
DTN63.78±0.2960.45±0.3560.55±0.3572.60±0.2570.67±0.3041.55±0.6541.45 ± 0.4570.67±0.45
c-DTN57.53 ± 0.4255.24 ± 0.4557.14 ± 0.3965.16 ± 0.3563.80 ±0.4238.97 ± 0.7139.80 ± 0.6562.10 ± 0.458.75 10.92
PixelDA45.68± 0.5244.95± 0.4244.45±0.5590.50± 0.2546.28±0.6045.89±0.6144.45± 0.5169.15 ± 0.459.95
UNIT44.14±0.1044.47 ± 0.1144.21 ±0.1244.47 ± 0.1143.03 ±0.144.44 ±0.1544.47 ± 0.1544.47 ± 0.0511.07
DSN64.15± 0.3057.70±0.3849.15± 0.4580.75±0.2782.20±0.2838.75±0.5345.00± 0.2580.50± 0.35
c-DSN57.34±0.4531.63 ±0.6051.17 ±0.4074.52 ±0.3782.01 ± 0.3534.25±0.5842.63 ± 0.5579.42 ±0.355.15 8.20
1p-DSN55.84± 0.5030.03 ± 0.5049.06 ± 0.3872.11 ± 0.5081.22 ±0.4533.33 ± 0.5842.03 ±0.2478.78±0.578.63
mp-DSN55.20±0.4630.40±0.5047.80±0.3575.30 ± 0.2580.75±0.2030.20 ±0.5543.00±0.3579.02±0.408.88
s-MTDA-ITA70.10± 0.2758.90±0.2558.10±0.2780.12±0.1582.05±0.1845.90±0.3052.67±0.3081.60±0.24
c-MTDA-ITA60.34 ± 0.1755.67 ± 0.2157.10 ± 0.2373.50 ± 0.2076.80 ±0.1043.10 ±0.1248.10 ± 0.1480.90 ± 0.113.65
MTDA-ITA78.40 ± 0.266.70 ± 0.1770.30± 0.1485.49 ± 0.1187.20±0.1061.40±0.1460.05 ±0.1386.70 ± 0.105.01 1.20
", + "type": "table", + "image_path": "5c32f244dced015904d1dfdfb8c3e5deb43fab55ecc31c215787a0f0641e50d0.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 90, + 339, + 520, + 387.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 90, + 387.3333333333333, + 520, + 435.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 90, + 435.66666666666663, + 520, + 483.99999999999994 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 89, + 500, + 520, + 520 + ], + "lines": [ + { + "bbox": [ + 90, + 500, + 520, + 510 + ], + "spans": [ + { + "bbox": [ + 90, + 500, + 520, + 510 + ], + "score": 1.0, + "content": "Table 2: Classification results on Multi-PIE dataset. Last column shows the average rank of each method over all", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 509, + 244, + 520 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 244, + 520 + ], + "score": 1.0, + "content": "adaptation pairs. The best is shown in red.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 91, + 529, + 200, + 541 + ], + "lines": [ + { + "bbox": [ + 90, + 528, + 202, + 542 + ], + "spans": [ + { + "bbox": [ + 90, + 528, + 202, + 542 + ], + "score": 1.0, + "content": "4.3 ABLATION STUDIES", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 91, + 552, + 521, + 608 + ], + "lines": [ + { + "bbox": [ + 90, + 551, + 521, + 565 + ], + "spans": [ + { + "bbox": [ + 90, + 551, + 521, + 565 + ], + "score": 1.0, + "content": "We performed an ablation study on the proposed model measuring impact of various terms on the model’s", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 563, + 521, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 521, + 576 + ], + "score": 1.0, + "content": "performance. To this end, we conducted additional experiments for the digit datasets with different components", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 573, + 521, + 588 + ], + "spans": [ + { + "bbox": [ + 89, + 573, + 453, + 588 + ], + "score": 1.0, + "content": "ablation, i.e., training without the reconstruction loss (denoted as MTDA-woR) by setting", + "type": "text" + }, + { + "bbox": [ + 454, + 574, + 483, + 586 + ], + "score": 0.92, + "content": "\\lambda _ { r } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 573, + 521, + 588 + ], + "score": 1.0, + "content": ", training", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 585, + 522, + 599 + ], + "spans": [ + { + "bbox": [ + 89, + 585, + 379, + 599 + ], + "score": 1.0, + "content": "without the classifier entropy loss (denoted as MTDA-woE) by setting", + "type": "text" + }, + { + "bbox": [ + 379, + 586, + 408, + 596 + ], + "score": 0.91, + "content": "\\lambda _ { c } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 585, + 522, + 599 + ], + "score": 1.0, + "content": ", training without the multi-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 596, + 366, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 332, + 609 + ], + "score": 1.0, + "content": "domain separation loss (denoted as MTDA-woD) by setting", + "type": "text" + }, + { + "bbox": [ + 332, + 597, + 362, + 607 + ], + "score": 0.91, + "content": "\\lambda _ { d } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 596, + 366, + 609 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 91, + 608, + 520, + 652 + ], + "lines": [ + { + "bbox": [ + 90, + 608, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 90, + 608, + 522, + 620 + ], + "score": 1.0, + "content": "As can be seen from fig. 4, disabling each of the above components leads to degraded performance. More", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 617, + 521, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 617, + 380, + 631 + ], + "score": 1.0, + "content": "precisely, the average drop by disabling the classifier entropy loss is", + "type": "text" + }, + { + "bbox": [ + 380, + 618, + 416, + 629 + ], + "score": 0.87, + "content": "\\approx 3 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 617, + 521, + 631 + ], + "score": 1.0, + "content": ". Similarly, by disabling", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 627, + 523, + 643 + ], + "spans": [ + { + "bbox": [ + 88, + 627, + 379, + 643 + ], + "score": 1.0, + "content": "the reconstruction loss and the multi-domain separation loss, we have", + "type": "text" + }, + { + "bbox": [ + 380, + 630, + 414, + 640 + ], + "score": 0.82, + "content": "\\approx 4 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 627, + 432, + 643 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 433, + 629, + 464, + 640 + ], + "score": 0.89, + "content": "\\approx 2 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 627, + 523, + 643 + ], + "score": 1.0, + "content": "average drop", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 88, + 639, + 522, + 654 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 654 + ], + "score": 1.0, + "content": "in performance, respectively. Clearly, by disabling the multi-domain separation loss, the accuracy drops", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 286, + 671, + 293, + 680 + ], + "lines": [ + { + "bbox": [ + 285, + 670, + 293, + 681 + ], + "spans": [ + { + "bbox": [ + 285, + 670, + 293, + 681 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 93, + 520, + 249 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 93, + 520, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 93, + 520, + 249 + ], + "spans": [ + { + "bbox": [ + 90, + 93, + 520, + 249 + ], + "score": 0.981, + "html": "
methodS→MS→MMS→UM→SM→MMM→UAve.ranking
Source Only62.10±0.6040.43±0.7039.90± 0.6030.29± 0.5955.98 ± 0.4878.30±0.3814.00
1-NN35.8618.2129.3128.0112.5841.2215.00
CORAL (Sun & Saenko (2016))63.10 ± 0.6154.37 ± 0.5350.15 ± 0.6333.40 ±0.7457.70 ± 0.6981.05 ± 0.8011.33
DANN (Ganin et al. (2016))73.80 ±0.4961.05±0.8062.54 ± 0.9135.50 ± 0.6577.40±0.7381.60 ± 0.608.75
ADDA(Tzeng et al. (2017))77.68 ± 0.9264.23±0.7064.10±0.7930.04± 0.9891.47 ±1.090.51 ± 0.806.43
C-ADDA80.10 ± 0.6956.80 ± 0.7964.80±0.8827.50 ± 0.8683.30 ±0.9084.10 ±0.988.95
DTN (Zhang et al. (2015))81.40± 0.4263.70±0.3960.12 ± 0.5240.40±0.5085.70±0.3985.80 ± 0.466.04
c-DTN PixelDA (Bousmalis et al. (2017))82.10 ±0.6259.30 ± 0.5956.87 ± 0.6538.32 ± 0.5080.90±0.8079.31 ±0.787.96
UNIT (Liu et al. (2017))98.10*94.10*1
DSN (Bousmalis et al. (2016))90.6*92.901
c-DSN82.70±0.37 83.10 ±0.2064.80± 0.4065.30± 0.2849.30±0.3083.20±0.3091.65 ± 0.402.85
1p-DSN81.00 ± 0.4760.56 ± 0.36 58.22 ± 0.6860.35 ± 0.5946.80 ± 0.4580.49±0.4088.21 ±0.384.84
mp-DSN83.40±0.3061.00±0.5058.06 ±0.48 58.10 ±0.6445.11 ± 0.3377.33 ± 0.52 79.30±0.5985.16 ± 0.634.90
s-MTDA-ITA82.90±0.1363.10±0.2863.54±0.3047.35 ± 0.4082.42± 0.1986.45 ± 0.715.33
c-MTDA-ITA79.20±0.2859.90 ± 0.3063.70±0.2649.60± 0.25 45.30 ±0.3077.12 ± 0.2289.21±0.282.88
MTDA-ITA84.60±0.2465.30 ± 0.1570.03 ±0.2052.01 ± 0.2185.50 ±0.1887.47 ± 0.25 94.20±0.204.25
1.16
", + "type": "table", + "image_path": "8fd442a94d88d4c3187f10668a4eaf7e8e6a87c4149cab928e7e3f5d35839d41.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 90, + 93, + 520, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 90, + 145.0, + 520, + 197.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 90, + 197.0, + 520, + 249.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 90, + 263, + 522, + 325 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 263, + 522, + 276 + ], + "spans": [ + { + "bbox": [ + 89, + 263, + 522, + 276 + ], + "score": 1.0, + "content": "Table 1: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 273, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 273, + 523, + 285 + ], + "score": 1.0, + "content": "in red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDA-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 283, + 523, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 283, + 523, + 295 + ], + "score": 1.0, + "content": "ITA where each one correspond to a source-target pair. 1p-DSN: extended DSN with single private encoder.mp-DSN:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 294, + 521, + 305 + ], + "spans": [ + { + "bbox": [ + 90, + 294, + 521, + 305 + ], + "score": 1.0, + "content": "extended DSN with multiple private encoder. Last column shows the average rank of each method over all adaptation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 304, + 522, + 316 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 110, + 316 + ], + "score": 1.0, + "content": "pairs.", + "type": "text" + }, + { + "bbox": [ + 110, + 304, + 138, + 314 + ], + "score": 0.37, + "content": "\\bf \\Pi ^ { * } U N I T", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 304, + 255, + 316 + ], + "score": 1.0, + "content": "trains with the extended SVHN", + "type": "text" + }, + { + "bbox": [ + 255, + 304, + 287, + 314 + ], + "score": 0.85, + "content": "> 5 0 0 \\mathrm { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 304, + 421, + 316 + ], + "score": 1.0, + "content": "images vs ours 72K). *PixelDA uses", + "type": "text" + }, + { + "bbox": [ + 421, + 304, + 460, + 315 + ], + "score": 0.85, + "content": "( \\approx 1 , 0 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 304, + 522, + 316 + ], + "score": 1.0, + "content": "of labeled target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 313, + 321, + 325 + ], + "spans": [ + { + "bbox": [ + 89, + 313, + 321, + 325 + ], + "score": 1.0, + "content": "domain data as a validation set for tuning the hyper-parameters.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 90, + 339, + 520, + 484 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 339, + 520, + 484 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 90, + 339, + 520, + 484 + ], + "spans": [ + { + "bbox": [ + 90, + 339, + 520, + 484 + ], + "score": 0.978, + "html": "
methodC13→C05C13→C08C13→C09C13→C14C14→C05C14→C08C14→C09C14→C13Ave.ranking
Source Only50.79± 0.3345.90± 0.5040.04±0.4059.68± 0.2960.03±0.5536.80± 0.6140.11 ±0.5060.57± 0.3616.08
1-NN CORAL33.21 54.89 ± 0.5237.0134.4548.79 68.90± 0.3547.4428.2430.8644.8617.00
DANN57.86± 0.4148.90±0.48 50.30 ± 0.4340.30±0.53 45.30± 0.5070.68 ± 0.3559.98 ± 0.45 57.20± 0.4540.63±0.55 40.22±0.5540.80 ±0.53 40.77 ± 0.4565.11 ± 0.45 70.50± 0.5511.95 9.92
ADDA64.83±0.6963.20±0.4555.48 ±0.6574.25±0.5573.62 ± 0.7543.56±0.9538.68 ±0.9572.84±0.759.33
C-ADDA59.20±0.2530.70±0.6353.20±0.4068.33±0.3565.88±0.3830.60±0.6145.34±0.4864.30±0.4011.50
DTN63.78±0.2960.45±0.3560.55±0.3572.60±0.2570.67±0.3041.55±0.6541.45 ± 0.4570.67±0.45
c-DTN57.53 ± 0.4255.24 ± 0.4557.14 ± 0.3965.16 ± 0.3563.80 ±0.4238.97 ± 0.7139.80 ± 0.6562.10 ± 0.458.75 10.92
PixelDA45.68± 0.5244.95± 0.4244.45±0.5590.50± 0.2546.28±0.6045.89±0.6144.45± 0.5169.15 ± 0.459.95
UNIT44.14±0.1044.47 ± 0.1144.21 ±0.1244.47 ± 0.1143.03 ±0.144.44 ±0.1544.47 ± 0.1544.47 ± 0.0511.07
DSN64.15± 0.3057.70±0.3849.15± 0.4580.75±0.2782.20±0.2838.75±0.5345.00± 0.2580.50± 0.35
c-DSN57.34±0.4531.63 ±0.6051.17 ±0.4074.52 ±0.3782.01 ± 0.3534.25±0.5842.63 ± 0.5579.42 ±0.355.15 8.20
1p-DSN55.84± 0.5030.03 ± 0.5049.06 ± 0.3872.11 ± 0.5081.22 ±0.4533.33 ± 0.5842.03 ±0.2478.78±0.578.63
mp-DSN55.20±0.4630.40±0.5047.80±0.3575.30 ± 0.2580.75±0.2030.20 ±0.5543.00±0.3579.02±0.408.88
s-MTDA-ITA70.10± 0.2758.90±0.2558.10±0.2780.12±0.1582.05±0.1845.90±0.3052.67±0.3081.60±0.24
c-MTDA-ITA60.34 ± 0.1755.67 ± 0.2157.10 ± 0.2373.50 ± 0.2076.80 ±0.1043.10 ±0.1248.10 ± 0.1480.90 ± 0.113.65
MTDA-ITA78.40 ± 0.266.70 ± 0.1770.30± 0.1485.49 ± 0.1187.20±0.1061.40±0.1460.05 ±0.1386.70 ± 0.105.01 1.20
", + "type": "table", + "image_path": "5c32f244dced015904d1dfdfb8c3e5deb43fab55ecc31c215787a0f0641e50d0.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 90, + 339, + 520, + 387.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 90, + 387.3333333333333, + 520, + 435.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 90, + 435.66666666666663, + 520, + 483.99999999999994 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 89, + 500, + 520, + 520 + ], + "lines": [ + { + "bbox": [ + 90, + 500, + 520, + 510 + ], + "spans": [ + { + "bbox": [ + 90, + 500, + 520, + 510 + ], + "score": 1.0, + "content": "Table 2: Classification results on Multi-PIE dataset. Last column shows the average rank of each method over all", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 509, + 244, + 520 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 244, + 520 + ], + "score": 1.0, + "content": "adaptation pairs. The best is shown in red.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 89, + 500, + 520, + 520 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 529, + 200, + 541 + ], + "lines": [ + { + "bbox": [ + 90, + 528, + 202, + 542 + ], + "spans": [ + { + "bbox": [ + 90, + 528, + 202, + 542 + ], + "score": 1.0, + "content": "4.3 ABLATION STUDIES", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 91, + 552, + 521, + 608 + ], + "lines": [ + { + "bbox": [ + 90, + 551, + 521, + 565 + ], + "spans": [ + { + "bbox": [ + 90, + 551, + 521, + 565 + ], + "score": 1.0, + "content": "We performed an ablation study on the proposed model measuring impact of various terms on the model’s", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 563, + 521, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 521, + 576 + ], + "score": 1.0, + "content": "performance. To this end, we conducted additional experiments for the digit datasets with different components", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 573, + 521, + 588 + ], + "spans": [ + { + "bbox": [ + 89, + 573, + 453, + 588 + ], + "score": 1.0, + "content": "ablation, i.e., training without the reconstruction loss (denoted as MTDA-woR) by setting", + "type": "text" + }, + { + "bbox": [ + 454, + 574, + 483, + 586 + ], + "score": 0.92, + "content": "\\lambda _ { r } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 573, + 521, + 588 + ], + "score": 1.0, + "content": ", training", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 585, + 522, + 599 + ], + "spans": [ + { + "bbox": [ + 89, + 585, + 379, + 599 + ], + "score": 1.0, + "content": "without the classifier entropy loss (denoted as MTDA-woE) by setting", + "type": "text" + }, + { + "bbox": [ + 379, + 586, + 408, + 596 + ], + "score": 0.91, + "content": "\\lambda _ { c } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 585, + 522, + 599 + ], + "score": 1.0, + "content": ", training without the multi-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 596, + 366, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 332, + 609 + ], + "score": 1.0, + "content": "domain separation loss (denoted as MTDA-woD) by setting", + "type": "text" + }, + { + "bbox": [ + 332, + 597, + 362, + 607 + ], + "score": 0.91, + "content": "\\lambda _ { d } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 596, + 366, + 609 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 89, + 551, + 522, + 609 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 608, + 520, + 652 + ], + "lines": [ + { + "bbox": [ + 90, + 608, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 90, + 608, + 522, + 620 + ], + "score": 1.0, + "content": "As can be seen from fig. 4, disabling each of the above components leads to degraded performance. More", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 617, + 521, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 617, + 380, + 631 + ], + "score": 1.0, + "content": "precisely, the average drop by disabling the classifier entropy loss is", + "type": "text" + }, + { + "bbox": [ + 380, + 618, + 416, + 629 + ], + "score": 0.87, + "content": "\\approx 3 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 617, + 521, + 631 + ], + "score": 1.0, + "content": ". Similarly, by disabling", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 88, + 627, + 523, + 643 + ], + "spans": [ + { + "bbox": [ + 88, + 627, + 379, + 643 + ], + "score": 1.0, + "content": "the reconstruction loss and the multi-domain separation loss, we have", + "type": "text" + }, + { + "bbox": [ + 380, + 630, + 414, + 640 + ], + "score": 0.82, + "content": "\\approx 4 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 627, + 432, + 643 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 433, + 629, + 464, + 640 + ], + "score": 0.89, + "content": "\\approx 2 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 627, + 523, + 643 + ], + "score": 1.0, + "content": "average drop", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 88, + 639, + 522, + 654 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 654 + ], + "score": 1.0, + "content": "in performance, respectively. Clearly, by disabling the multi-domain separation loss, the accuracy drops", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 90, + 95, + 521, + 108 + ], + "spans": [ + { + "bbox": [ + 90, + 95, + 521, + 108 + ], + "score": 1.0, + "content": "significantly due to the severe data distribution mismatch between different domains. The figure also", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 107, + 521, + 118 + ], + "spans": [ + { + "bbox": [ + 90, + 107, + 521, + 118 + ], + "score": 1.0, + "content": "demonstrates that leveraging the unlabeled data from multiple target domains during training enhances the", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 117, + 523, + 131 + ], + "spans": [ + { + "bbox": [ + 88, + 117, + 523, + 131 + ], + "score": 1.0, + "content": "generalization ability of the model that leads to higher performance. In addition, the performance drop caused", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 129, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 90, + 129, + 522, + 140 + ], + "score": 1.0, + "content": "by removing the reconstruction loss , i.e., without the private encoder/decoder, indicates (i) the benefit of", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 139, + 522, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 139, + 522, + 152 + ], + "score": 1.0, + "content": "modeling the latent features as the combination of shared and private features, (ii) the ability of the model’s", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 150, + 324, + 163 + ], + "spans": [ + { + "bbox": [ + 90, + 150, + 324, + 163 + ], + "score": 1.0, + "content": "domain adversarial loss to effectively learn those features.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 21.5, + "bbox_fs": [ + 88, + 608, + 523, + 654 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 96, + 521, + 161 + ], + "lines": [ + { + "bbox": [ + 90, + 95, + 521, + 108 + ], + "spans": [ + { + "bbox": [ + 90, + 95, + 521, + 108 + ], + "score": 1.0, + "content": "significantly due to the severe data distribution mismatch between different domains. The figure also", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 107, + 521, + 118 + ], + "spans": [ + { + "bbox": [ + 90, + 107, + 521, + 118 + ], + "score": 1.0, + "content": "demonstrates that leveraging the unlabeled data from multiple target domains during training enhances the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 117, + 523, + 131 + ], + "spans": [ + { + "bbox": [ + 88, + 117, + 523, + 131 + ], + "score": 1.0, + "content": "generalization ability of the model that leads to higher performance. In addition, the performance drop caused", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 129, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 90, + 129, + 522, + 140 + ], + "score": 1.0, + "content": "by removing the reconstruction loss , i.e., without the private encoder/decoder, indicates (i) the benefit of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 139, + 522, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 139, + 522, + 152 + ], + "score": 1.0, + "content": "modeling the latent features as the combination of shared and private features, (ii) the ability of the model’s", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 150, + 324, + 163 + ], + "spans": [ + { + "bbox": [ + 90, + 150, + 324, + 163 + ], + "score": 1.0, + "content": "domain adversarial loss to effectively learn those features.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 91, + 162, + 521, + 227 + ], + "lines": [ + { + "bbox": [ + 90, + 162, + 521, + 173 + ], + "spans": [ + { + "bbox": [ + 90, + 162, + 521, + 173 + ], + "score": 1.0, + "content": "In order to examine the effect of the private features on the model’s classification performance, we took the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "spans": [ + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "score": 1.0, + "content": "MTDA-ITA and trained it without the private encoder (denoted as MTDA-woP). As fig. 4 shows, without the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 183, + 523, + 196 + ], + "spans": [ + { + "bbox": [ + 89, + 183, + 316, + 196 + ], + "score": 1.0, + "content": "private features, the model performed consistently worse", + "type": "text" + }, + { + "bbox": [ + 316, + 183, + 342, + 194 + ], + "score": 0.88, + "content": "\\approx 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 183, + 523, + 196 + ], + "score": 1.0, + "content": "average drop in performance) in all scenarios.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 194, + 522, + 207 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 522, + 207 + ], + "score": 1.0, + "content": "This demonstrates explicitly modeling what is unique to each domain can improve the model’s ability to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 206, + 521, + 217 + ], + "spans": [ + { + "bbox": [ + 90, + 206, + 521, + 217 + ], + "score": 1.0, + "content": "extract domain–invariant features. In summary, this ablation study showed that the individual components", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 216, + 388, + 229 + ], + "spans": [ + { + "bbox": [ + 90, + 216, + 388, + 229 + ], + "score": 1.0, + "content": "bring complimentary information to achieve the best classification results.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 92, + 249, + 226, + 260 + ], + "lines": [ + { + "bbox": [ + 89, + 247, + 227, + 262 + ], + "spans": [ + { + "bbox": [ + 89, + 247, + 227, + 262 + ], + "score": 1.0, + "content": "4.4 FEATURE VISUALIZATION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 91, + 272, + 521, + 405 + ], + "lines": [ + { + "bbox": [ + 90, + 273, + 521, + 285 + ], + "spans": [ + { + "bbox": [ + 90, + 273, + 521, + 285 + ], + "score": 1.0, + "content": "We use t-SNE (Maaten & Hinton (2008)) on Digit dataset to visualize shared and private feature representations", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 283, + 521, + 296 + ], + "spans": [ + { + "bbox": [ + 89, + 283, + 521, + 296 + ], + "score": 1.0, + "content": "from different domains. Fig. 6 shows shared and private features from source (SVHN) and target domains", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 294, + 522, + 308 + ], + "spans": [ + { + "bbox": [ + 89, + 294, + 522, + 308 + ], + "score": 1.0, + "content": "before (a),(b) and after adaptation (c),(d). MTDA-ITA significantly reduces the domain mismatch for the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 305, + 522, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 305, + 522, + 319 + ], + "score": 1.0, + "content": "shared features (circle markers in fig. 6d, strong mixing of domain labels in this cluster, fig. 6c) and increases", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 317, + 523, + 330 + ], + "spans": [ + { + "bbox": [ + 89, + 317, + 523, + 330 + ], + "score": 1.0, + "content": "it for the private features (triangle markers, pure and well-separated domain clusters in fig. 6c). This is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 88, + 327, + 522, + 340 + ], + "spans": [ + { + "bbox": [ + 88, + 327, + 481, + 340 + ], + "score": 1.0, + "content": "partially due to the proposed multi-domain separation loss through the use of the domain classifier", + "type": "text" + }, + { + "bbox": [ + 481, + 328, + 491, + 338 + ], + "score": 0.77, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 327, + 522, + 340 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 338, + 524, + 352 + ], + "spans": [ + { + "bbox": [ + 88, + 338, + 524, + 352 + ], + "score": 1.0, + "content": "penalizes the domain mismatch for the shared features and rewards the mismatch for the private features.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 90, + 350, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 90, + 350, + 522, + 362 + ], + "score": 1.0, + "content": "Moreover, as supported by the quantitative results in tab. 1, joint adaptation of related domains and the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 360, + 523, + 374 + ], + "spans": [ + { + "bbox": [ + 89, + 360, + 523, + 374 + ], + "score": 1.0, + "content": "classifier, accomplished through the model, leads to superior class separability, compared to original features.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 371, + 521, + 385 + ], + "spans": [ + { + "bbox": [ + 89, + 371, + 521, + 385 + ], + "score": 1.0, + "content": "This is depicted in fig. 6d, where the points in the shared space (large cluster) are grouped into class-specific", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 382, + 522, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 382, + 522, + 396 + ], + "score": 1.0, + "content": "subgroups (color indicates class label), while they are mixed in private spaces (smaller clusters). This is in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 394, + 362, + 407 + ], + "spans": [ + { + "bbox": [ + 89, + 394, + 362, + 407 + ], + "score": 1.0, + "content": "contrast to fig. 6b, where original features show no class-specificity.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 91, + 410, + 522, + 509 + ], + "lines": [ + { + "bbox": [ + 89, + 410, + 523, + 422 + ], + "spans": [ + { + "bbox": [ + 89, + 410, + 523, + 422 + ], + "score": 1.0, + "content": "We also show the learned shared and private features for the models MTDA-woE, MTDA-woP, MTDA-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 90, + 421, + 522, + 434 + ], + "spans": [ + { + "bbox": [ + 90, + 421, + 392, + 433 + ], + "score": 1.0, + "content": "woR, and MTDA-woD, in figs. 6e to 6l. Note that since the private encoder", + "type": "text" + }, + { + "bbox": [ + 392, + 422, + 405, + 434 + ], + "score": 0.9, + "content": "E _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 421, + 522, + 433 + ], + "score": 1.0, + "content": "is disabled for MTDA-woR,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 433, + 522, + 444 + ], + "spans": [ + { + "bbox": [ + 90, + 433, + 323, + 444 + ], + "score": 1.0, + "content": "and MTDA-woP, no private features are depicted in figs.", + "type": "text" + }, + { + "bbox": [ + 323, + 433, + 335, + 444 + ], + "score": 0.35, + "content": "6 \\mathrm { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 433, + 522, + 444 + ], + "score": 1.0, + "content": "to 6j. The class label separation in the shared", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 442, + 523, + 458 + ], + "spans": [ + { + "bbox": [ + 88, + 442, + 523, + 458 + ], + "score": 1.0, + "content": "space for MTDA-woE, MTDA-woP, and MTDA-woR, figs. 6f, 6h and 6j, is still evident but not as strong", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 455, + 522, + 468 + ], + "spans": [ + { + "bbox": [ + 89, + 455, + 522, + 468 + ], + "score": 1.0, + "content": "as in the full model, fig. 6d, corroborating the small loss in classification accuracy observed in fig. 4a. On", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 464, + 522, + 478 + ], + "spans": [ + { + "bbox": [ + 89, + 464, + 522, + 478 + ], + "score": 1.0, + "content": "the other hand, MTDA-woD has significant mixing of class labels in the shared space, fig. 6l, more so than", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 476, + 522, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 522, + 489 + ], + "score": 1.0, + "content": "MTDA-woE, MTDA-woR, and MTDA-woP, implying worse classification prediction in fig. 4a due to the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 486, + 523, + 501 + ], + "spans": [ + { + "bbox": [ + 88, + 486, + 523, + 501 + ], + "score": 1.0, + "content": "severe mismatch between different domains. Since our model uses one private encoder for all target domains,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 498, + 496, + 511 + ], + "spans": [ + { + "bbox": [ + 89, + 498, + 496, + 511 + ], + "score": 1.0, + "content": "we also contrasted the visualization of DSN model with one private encoder 1p-DSN in Appendix H.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 91, + 533, + 179, + 546 + ], + "lines": [ + { + "bbox": [ + 88, + 532, + 181, + 549 + ], + "spans": [ + { + "bbox": [ + 88, + 532, + 181, + 549 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 91, + 563, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 563, + 522, + 577 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 522, + 577 + ], + "score": 1.0, + "content": "This paper presented an information theoretic end-to-end approach to uDA in the context of a single source", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 574, + 521, + 587 + ], + "spans": [ + { + "bbox": [ + 89, + 574, + 521, + 587 + ], + "score": 1.0, + "content": "and multiple target domains that share a common task or properties. The proposed method learns feature", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 88, + 585, + 523, + 599 + ], + "spans": [ + { + "bbox": [ + 88, + 585, + 523, + 599 + ], + "score": 1.0, + "content": "representations invariant under multiple domain shifts and simultaneously discriminative for the learning task.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "score": 1.0, + "content": "This is accomplished by explicitly separating representations private to each domain and shared between", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 89, + 608, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 89, + 608, + 522, + 620 + ], + "score": 1.0, + "content": "source and target domains using a novel discrimination strategy. Our use of a single private domain encoder", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 618, + 522, + 632 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 522, + 632 + ], + "score": 1.0, + "content": "results in a highly scalable model, easily optimized using established back-propagation approaches. Results", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 629, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 522, + 643 + ], + "score": 1.0, + "content": "on three benchmark datasets for image classification show superiority of the proposed method compared to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 89, + 640, + 462, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 640, + 462, + 653 + ], + "score": 1.0, + "content": "the state-of-the-art methods for unsupervised domain adaptation of visual domain categories.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 96, + 521, + 161 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 88, + 95, + 523, + 163 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 91, + 162, + 521, + 227 + ], + "lines": [ + { + "bbox": [ + 90, + 162, + 521, + 173 + ], + "spans": [ + { + "bbox": [ + 90, + 162, + 521, + 173 + ], + "score": 1.0, + "content": "In order to examine the effect of the private features on the model’s classification performance, we took the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "spans": [ + { + "bbox": [ + 89, + 172, + 521, + 185 + ], + "score": 1.0, + "content": "MTDA-ITA and trained it without the private encoder (denoted as MTDA-woP). As fig. 4 shows, without the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 183, + 523, + 196 + ], + "spans": [ + { + "bbox": [ + 89, + 183, + 316, + 196 + ], + "score": 1.0, + "content": "private features, the model performed consistently worse", + "type": "text" + }, + { + "bbox": [ + 316, + 183, + 342, + 194 + ], + "score": 0.88, + "content": "\\approx 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 183, + 523, + 196 + ], + "score": 1.0, + "content": "average drop in performance) in all scenarios.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 194, + 522, + 207 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 522, + 207 + ], + "score": 1.0, + "content": "This demonstrates explicitly modeling what is unique to each domain can improve the model’s ability to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 206, + 521, + 217 + ], + "spans": [ + { + "bbox": [ + 90, + 206, + 521, + 217 + ], + "score": 1.0, + "content": "extract domain–invariant features. In summary, this ablation study showed that the individual components", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 216, + 388, + 229 + ], + "spans": [ + { + "bbox": [ + 90, + 216, + 388, + 229 + ], + "score": 1.0, + "content": "bring complimentary information to achieve the best classification results.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 89, + 162, + 523, + 229 + ] + }, + { + "type": "title", + "bbox": [ + 92, + 249, + 226, + 260 + ], + "lines": [ + { + "bbox": [ + 89, + 247, + 227, + 262 + ], + "spans": [ + { + "bbox": [ + 89, + 247, + 227, + 262 + ], + "score": 1.0, + "content": "4.4 FEATURE VISUALIZATION", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 91, + 272, + 521, + 405 + ], + "lines": [ + { + "bbox": [ + 90, + 273, + 521, + 285 + ], + "spans": [ + { + "bbox": [ + 90, + 273, + 521, + 285 + ], + "score": 1.0, + "content": "We use t-SNE (Maaten & Hinton (2008)) on Digit dataset to visualize shared and private feature representations", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 283, + 521, + 296 + ], + "spans": [ + { + "bbox": [ + 89, + 283, + 521, + 296 + ], + "score": 1.0, + "content": "from different domains. Fig. 6 shows shared and private features from source (SVHN) and target domains", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 294, + 522, + 308 + ], + "spans": [ + { + "bbox": [ + 89, + 294, + 522, + 308 + ], + "score": 1.0, + "content": "before (a),(b) and after adaptation (c),(d). MTDA-ITA significantly reduces the domain mismatch for the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 305, + 522, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 305, + 522, + 319 + ], + "score": 1.0, + "content": "shared features (circle markers in fig. 6d, strong mixing of domain labels in this cluster, fig. 6c) and increases", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 317, + 523, + 330 + ], + "spans": [ + { + "bbox": [ + 89, + 317, + 523, + 330 + ], + "score": 1.0, + "content": "it for the private features (triangle markers, pure and well-separated domain clusters in fig. 6c). This is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 88, + 327, + 522, + 340 + ], + "spans": [ + { + "bbox": [ + 88, + 327, + 481, + 340 + ], + "score": 1.0, + "content": "partially due to the proposed multi-domain separation loss through the use of the domain classifier", + "type": "text" + }, + { + "bbox": [ + 481, + 328, + 491, + 338 + ], + "score": 0.77, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 327, + 522, + 340 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 338, + 524, + 352 + ], + "spans": [ + { + "bbox": [ + 88, + 338, + 524, + 352 + ], + "score": 1.0, + "content": "penalizes the domain mismatch for the shared features and rewards the mismatch for the private features.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 90, + 350, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 90, + 350, + 522, + 362 + ], + "score": 1.0, + "content": "Moreover, as supported by the quantitative results in tab. 1, joint adaptation of related domains and the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 360, + 523, + 374 + ], + "spans": [ + { + "bbox": [ + 89, + 360, + 523, + 374 + ], + "score": 1.0, + "content": "classifier, accomplished through the model, leads to superior class separability, compared to original features.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 371, + 521, + 385 + ], + "spans": [ + { + "bbox": [ + 89, + 371, + 521, + 385 + ], + "score": 1.0, + "content": "This is depicted in fig. 6d, where the points in the shared space (large cluster) are grouped into class-specific", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 382, + 522, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 382, + 522, + 396 + ], + "score": 1.0, + "content": "subgroups (color indicates class label), while they are mixed in private spaces (smaller clusters). This is in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 394, + 362, + 407 + ], + "spans": [ + { + "bbox": [ + 89, + 394, + 362, + 407 + ], + "score": 1.0, + "content": "contrast to fig. 6b, where original features show no class-specificity.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5, + "bbox_fs": [ + 88, + 273, + 524, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 410, + 522, + 509 + ], + "lines": [ + { + "bbox": [ + 89, + 410, + 523, + 422 + ], + "spans": [ + { + "bbox": [ + 89, + 410, + 523, + 422 + ], + "score": 1.0, + "content": "We also show the learned shared and private features for the models MTDA-woE, MTDA-woP, MTDA-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 90, + 421, + 522, + 434 + ], + "spans": [ + { + "bbox": [ + 90, + 421, + 392, + 433 + ], + "score": 1.0, + "content": "woR, and MTDA-woD, in figs. 6e to 6l. Note that since the private encoder", + "type": "text" + }, + { + "bbox": [ + 392, + 422, + 405, + 434 + ], + "score": 0.9, + "content": "E _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 421, + 522, + 433 + ], + "score": 1.0, + "content": "is disabled for MTDA-woR,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 433, + 522, + 444 + ], + "spans": [ + { + "bbox": [ + 90, + 433, + 323, + 444 + ], + "score": 1.0, + "content": "and MTDA-woP, no private features are depicted in figs.", + "type": "text" + }, + { + "bbox": [ + 323, + 433, + 335, + 444 + ], + "score": 0.35, + "content": "6 \\mathrm { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 433, + 522, + 444 + ], + "score": 1.0, + "content": "to 6j. The class label separation in the shared", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 442, + 523, + 458 + ], + "spans": [ + { + "bbox": [ + 88, + 442, + 523, + 458 + ], + "score": 1.0, + "content": "space for MTDA-woE, MTDA-woP, and MTDA-woR, figs. 6f, 6h and 6j, is still evident but not as strong", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 455, + 522, + 468 + ], + "spans": [ + { + "bbox": [ + 89, + 455, + 522, + 468 + ], + "score": 1.0, + "content": "as in the full model, fig. 6d, corroborating the small loss in classification accuracy observed in fig. 4a. On", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 464, + 522, + 478 + ], + "spans": [ + { + "bbox": [ + 89, + 464, + 522, + 478 + ], + "score": 1.0, + "content": "the other hand, MTDA-woD has significant mixing of class labels in the shared space, fig. 6l, more so than", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 476, + 522, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 522, + 489 + ], + "score": 1.0, + "content": "MTDA-woE, MTDA-woR, and MTDA-woP, implying worse classification prediction in fig. 4a due to the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 486, + 523, + 501 + ], + "spans": [ + { + "bbox": [ + 88, + 486, + 523, + 501 + ], + "score": 1.0, + "content": "severe mismatch between different domains. Since our model uses one private encoder for all target domains,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 498, + 496, + 511 + ], + "spans": [ + { + "bbox": [ + 89, + 498, + 496, + 511 + ], + "score": 1.0, + "content": "we also contrasted the visualization of DSN model with one private encoder 1p-DSN in Appendix H.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29, + "bbox_fs": [ + 88, + 410, + 523, + 511 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 533, + 179, + 546 + ], + "lines": [ + { + "bbox": [ + 88, + 532, + 181, + 549 + ], + "spans": [ + { + "bbox": [ + 88, + 532, + 181, + 549 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 91, + 563, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 563, + 522, + 577 + ], + "spans": [ + { + "bbox": [ + 89, + 563, + 522, + 577 + ], + "score": 1.0, + "content": "This paper presented an information theoretic end-to-end approach to uDA in the context of a single source", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 574, + 521, + 587 + ], + "spans": [ + { + "bbox": [ + 89, + 574, + 521, + 587 + ], + "score": 1.0, + "content": "and multiple target domains that share a common task or properties. The proposed method learns feature", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 88, + 585, + 523, + 599 + ], + "spans": [ + { + "bbox": [ + 88, + 585, + 523, + 599 + ], + "score": 1.0, + "content": "representations invariant under multiple domain shifts and simultaneously discriminative for the learning task.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "score": 1.0, + "content": "This is accomplished by explicitly separating representations private to each domain and shared between", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 89, + 608, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 89, + 608, + 522, + 620 + ], + "score": 1.0, + "content": "source and target domains using a novel discrimination strategy. Our use of a single private domain encoder", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 618, + 522, + 632 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 522, + 632 + ], + "score": 1.0, + "content": "results in a highly scalable model, easily optimized using established back-propagation approaches. Results", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 629, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 522, + 643 + ], + "score": 1.0, + "content": "on three benchmark datasets for image classification show superiority of the proposed method compared to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 89, + 640, + 462, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 640, + 462, + 653 + ], + "score": 1.0, + "content": "the state-of-the-art methods for unsupervised domain adaptation of visual domain categories.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5, + "bbox_fs": [ + 88, + 563, + 523, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 91, + 95, + 160, + 107 + ], + "lines": [ + { + "bbox": [ + 90, + 96, + 160, + 108 + ], + "spans": [ + { + "bbox": [ + 90, + 96, + 160, + 108 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 89, + 113, + 523, + 147 + ], + "lines": [ + { + "bbox": [ + 89, + 113, + 522, + 127 + ], + "spans": [ + { + "bbox": [ + 89, + 113, + 522, + 127 + ], + "score": 1.0, + "content": "M Ehsan Abbasnejad, Anthony Dick, and Anton van den Hengel. Infinite variational autoencoder for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 98, + 124, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 98, + 124, + 523, + 140 + ], + "score": 1.0, + "content": "semi-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 136, + 191, + 148 + ], + "spans": [ + { + "bbox": [ + 100, + 136, + 191, + 148 + ], + "score": 1.0, + "content": "781–790. IEEE, 2017.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 88, + 155, + 518, + 179 + ], + "lines": [ + { + "bbox": [ + 90, + 155, + 520, + 169 + ], + "spans": [ + { + "bbox": [ + 90, + 155, + 520, + 169 + ], + "score": 1.0, + "content": "Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 167, + 446, + 179 + ], + "spans": [ + { + "bbox": [ + 99, + 167, + 446, + 179 + ], + "score": 1.0, + "content": "computation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 92, + 186, + 520, + 211 + ], + "lines": [ + { + "bbox": [ + 89, + 186, + 521, + 200 + ], + "spans": [ + { + "bbox": [ + 89, + 186, + 521, + 200 + ], + "score": 1.0, + "content": "Alexander Alemi, Ben Poole, Ian Fischer, Joshua Dillon, Rif A Saurous, and Kevin Murphy. Fixing a broken", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 198, + 413, + 211 + ], + "spans": [ + { + "bbox": [ + 99, + 198, + 413, + 211 + ], + "score": 1.0, + "content": "elbo. In International Conference on Machine Learning, pp. 159–168, 2018a.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 91, + 218, + 520, + 242 + ], + "lines": [ + { + "bbox": [ + 90, + 219, + 522, + 232 + ], + "spans": [ + { + "bbox": [ + 90, + 219, + 522, + 232 + ], + "score": 1.0, + "content": "Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 230, + 376, + 243 + ], + "spans": [ + { + "bbox": [ + 99, + 230, + 376, + 243 + ], + "score": 1.0, + "content": "International Conference on Learning Representation (ICLR), 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 89, + 250, + 521, + 273 + ], + "lines": [ + { + "bbox": [ + 90, + 250, + 523, + 263 + ], + "spans": [ + { + "bbox": [ + 90, + 250, + 523, + 263 + ], + "score": 1.0, + "content": "Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 262, + 267, + 273 + ], + "spans": [ + { + "bbox": [ + 100, + 262, + 267, + 273 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1807.00906, 2018b.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 89, + 281, + 521, + 305 + ], + "lines": [ + { + "bbox": [ + 89, + 280, + 521, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 280, + 521, + 295 + ], + "score": 1.0, + "content": "Asha Anoosheh, Eirikur Agustsson, Radu Timofte, and Luc Van Gool. Combogan: Unrestrained scalability", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 100, + 293, + 320, + 305 + ], + "spans": [ + { + "bbox": [ + 100, + 293, + 320, + 305 + ], + "score": 1.0, + "content": "for image domain translation. ICLR Workshops, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 90, + 313, + 520, + 336 + ], + "lines": [ + { + "bbox": [ + 90, + 314, + 522, + 326 + ], + "spans": [ + { + "bbox": [ + 90, + 314, + 522, + 326 + ], + "score": 1.0, + "content": "David Barber and Felix Agakov. The im algorithm: a variational approach to information maximization. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 97, + 325, + 470, + 336 + ], + "spans": [ + { + "bbox": [ + 97, + 325, + 470, + 336 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems (NIPS), pp. 201–208. MIT Press, 2003.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 89, + 344, + 518, + 368 + ], + "lines": [ + { + "bbox": [ + 90, + 344, + 520, + 357 + ], + "spans": [ + { + "bbox": [ + 90, + 344, + 520, + 357 + ], + "score": 1.0, + "content": "Sagie Benaim and Lior Wolf. One-sided unsupervised domain mapping. In Advances in Neural Information", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 99, + 355, + 293, + 367 + ], + "spans": [ + { + "bbox": [ + 99, + 355, + 293, + 367 + ], + "score": 1.0, + "content": "Processing Systems (NIPS), pp. 752–762, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 91, + 375, + 520, + 399 + ], + "lines": [ + { + "bbox": [ + 90, + 376, + 521, + 388 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 521, + 388 + ], + "score": 1.0, + "content": "Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, and Dumitru Erhan. Domain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 99, + 387, + 521, + 399 + ], + "spans": [ + { + "bbox": [ + 99, + 387, + 521, + 399 + ], + "score": 1.0, + "content": "separation networks. In Advances in Neural Information Processing Systems (NIPS), pp. 343–351, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 91, + 407, + 519, + 442 + ], + "lines": [ + { + "bbox": [ + 90, + 407, + 522, + 421 + ], + "spans": [ + { + "bbox": [ + 90, + 407, + 522, + 421 + ], + "score": 1.0, + "content": "Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 100, + 418, + 522, + 432 + ], + "spans": [ + { + "bbox": [ + 100, + 418, + 522, + 432 + ], + "score": 1.0, + "content": "pixel-level domain adaptation with generative adversarial networks. In IEEE Conference on Computer", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 100, + 429, + 357, + 442 + ], + "spans": [ + { + "bbox": [ + 100, + 429, + 357, + 442 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition (CVPR), volume 1, pp. 7, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 88, + 449, + 520, + 473 + ], + "lines": [ + { + "bbox": [ + 89, + 449, + 521, + 463 + ], + "spans": [ + { + "bbox": [ + 89, + 449, + 521, + 463 + ], + "score": 1.0, + "content": "Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 461, + 453, + 474 + ], + "spans": [ + { + "bbox": [ + 99, + 461, + 453, + 474 + ], + "score": 1.0, + "content": "Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 90, + 480, + 518, + 505 + ], + "lines": [ + { + "bbox": [ + 89, + 480, + 520, + 494 + ], + "spans": [ + { + "bbox": [ + 89, + 480, + 520, + 494 + ], + "score": 1.0, + "content": "John S Bridle. Probabilistic interpretation of feedforward classification network outputs, with relationships to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 100, + 492, + 423, + 505 + ], + "spans": [ + { + "bbox": [ + 100, + 492, + 423, + 505 + ], + "score": 1.0, + "content": "statistical pattern recognition. In Neurocomputing, pp. 227–236. Springer, 1990.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 89, + 512, + 517, + 536 + ], + "lines": [ + { + "bbox": [ + 90, + 513, + 519, + 525 + ], + "spans": [ + { + "bbox": [ + 90, + 513, + 519, + 525 + ], + "score": 1.0, + "content": "Matthew Chalk, Olivier Marre, and Gasper Tkacik. Relevant sparse codes with variational information", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 100, + 523, + 492, + 537 + ], + "spans": [ + { + "bbox": [ + 100, + 523, + 492, + 537 + ], + "score": 1.0, + "content": "bottleneck. In Advances in Neural Information Processing Systems (NIPS), pp. 1957–1965, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 91, + 544, + 519, + 578 + ], + "lines": [ + { + "bbox": [ + 90, + 544, + 521, + 558 + ], + "spans": [ + { + "bbox": [ + 90, + 544, + 521, + 558 + ], + "score": 1.0, + "content": "Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 99, + 554, + 522, + 569 + ], + "spans": [ + { + "bbox": [ + 99, + 554, + 522, + 569 + ], + "score": 1.0, + "content": "Unified generative adversarial networks for multi-domain image-to-image translation. IEEE Conference on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 100, + 566, + 356, + 579 + ], + "spans": [ + { + "bbox": [ + 100, + 566, + 356, + 579 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition (CVPR), 1711, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 90, + 586, + 520, + 620 + ], + "lines": [ + { + "bbox": [ + 89, + 586, + 522, + 600 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 522, + 600 + ], + "score": 1.0, + "content": "Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 98, + 596, + 522, + 612 + ], + "spans": [ + { + "bbox": [ + 98, + 596, + 522, + 612 + ], + "score": 1.0, + "content": "transportation for domain adaptation. In Advances in Neural Information Processing Systems (NIPS), pp.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 99, + 608, + 175, + 621 + ], + "spans": [ + { + "bbox": [ + 99, + 608, + 175, + 621 + ], + "score": 1.0, + "content": "3733–3742, 2017.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 89, + 629, + 520, + 651 + ], + "lines": [ + { + "bbox": [ + 88, + 627, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 88, + 627, + 522, + 643 + ], + "score": 1.0, + "content": "Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. pp. 1–35. Springer,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 99, + 639, + 127, + 652 + ], + "spans": [ + { + "bbox": [ + 99, + 639, + 127, + 652 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 51 + ], + "lines": [ + { + "bbox": [ + 90, + 38, + 293, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 38, + 293, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 295, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 91, + 95, + 160, + 107 + ], + "lines": [ + { + "bbox": [ + 90, + 96, + 160, + 108 + ], + "spans": [ + { + "bbox": [ + 90, + 96, + 160, + 108 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 89, + 113, + 523, + 147 + ], + "lines": [ + { + "bbox": [ + 89, + 113, + 522, + 127 + ], + "spans": [ + { + "bbox": [ + 89, + 113, + 522, + 127 + ], + "score": 1.0, + "content": "M Ehsan Abbasnejad, Anthony Dick, and Anton van den Hengel. Infinite variational autoencoder for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 98, + 124, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 98, + 124, + 523, + 140 + ], + "score": 1.0, + "content": "semi-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 136, + 191, + 148 + ], + "spans": [ + { + "bbox": [ + 100, + 136, + 191, + 148 + ], + "score": 1.0, + "content": "781–790. IEEE, 2017.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 89, + 113, + 523, + 148 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 155, + 518, + 179 + ], + "lines": [ + { + "bbox": [ + 90, + 155, + 520, + 169 + ], + "spans": [ + { + "bbox": [ + 90, + 155, + 520, + 169 + ], + "score": 1.0, + "content": "Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 167, + 446, + 179 + ], + "spans": [ + { + "bbox": [ + 99, + 167, + 446, + 179 + ], + "score": 1.0, + "content": "computation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 90, + 155, + 520, + 179 + ] + }, + { + "type": "text", + "bbox": [ + 92, + 186, + 520, + 211 + ], + "lines": [ + { + "bbox": [ + 89, + 186, + 521, + 200 + ], + "spans": [ + { + "bbox": [ + 89, + 186, + 521, + 200 + ], + "score": 1.0, + "content": "Alexander Alemi, Ben Poole, Ian Fischer, Joshua Dillon, Rif A Saurous, and Kevin Murphy. Fixing a broken", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 198, + 413, + 211 + ], + "spans": [ + { + "bbox": [ + 99, + 198, + 413, + 211 + ], + "score": 1.0, + "content": "elbo. In International Conference on Machine Learning, pp. 159–168, 2018a.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 89, + 186, + 521, + 211 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 218, + 520, + 242 + ], + "lines": [ + { + "bbox": [ + 90, + 219, + 522, + 232 + ], + "spans": [ + { + "bbox": [ + 90, + 219, + 522, + 232 + ], + "score": 1.0, + "content": "Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 230, + 376, + 243 + ], + "spans": [ + { + "bbox": [ + 99, + 230, + 376, + 243 + ], + "score": 1.0, + "content": "International Conference on Learning Representation (ICLR), 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 90, + 219, + 522, + 243 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 250, + 521, + 273 + ], + "lines": [ + { + "bbox": [ + 90, + 250, + 523, + 263 + ], + "spans": [ + { + "bbox": [ + 90, + 250, + 523, + 263 + ], + "score": 1.0, + "content": "Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 262, + 267, + 273 + ], + "spans": [ + { + "bbox": [ + 100, + 262, + 267, + 273 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1807.00906, 2018b.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 90, + 250, + 523, + 273 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 281, + 521, + 305 + ], + "lines": [ + { + "bbox": [ + 89, + 280, + 521, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 280, + 521, + 295 + ], + "score": 1.0, + "content": "Asha Anoosheh, Eirikur Agustsson, Radu Timofte, and Luc Van Gool. Combogan: Unrestrained scalability", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 100, + 293, + 320, + 305 + ], + "spans": [ + { + "bbox": [ + 100, + 293, + 320, + 305 + ], + "score": 1.0, + "content": "for image domain translation. ICLR Workshops, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 89, + 280, + 521, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 313, + 520, + 336 + ], + "lines": [ + { + "bbox": [ + 90, + 314, + 522, + 326 + ], + "spans": [ + { + "bbox": [ + 90, + 314, + 522, + 326 + ], + "score": 1.0, + "content": "David Barber and Felix Agakov. The im algorithm: a variational approach to information maximization. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 97, + 325, + 470, + 336 + ], + "spans": [ + { + "bbox": [ + 97, + 325, + 470, + 336 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems (NIPS), pp. 201–208. MIT Press, 2003.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 90, + 314, + 522, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 344, + 518, + 368 + ], + "lines": [ + { + "bbox": [ + 90, + 344, + 520, + 357 + ], + "spans": [ + { + "bbox": [ + 90, + 344, + 520, + 357 + ], + "score": 1.0, + "content": "Sagie Benaim and Lior Wolf. One-sided unsupervised domain mapping. In Advances in Neural Information", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 99, + 355, + 293, + 367 + ], + "spans": [ + { + "bbox": [ + 99, + 355, + 293, + 367 + ], + "score": 1.0, + "content": "Processing Systems (NIPS), pp. 752–762, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 90, + 344, + 520, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 375, + 520, + 399 + ], + "lines": [ + { + "bbox": [ + 90, + 376, + 521, + 388 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 521, + 388 + ], + "score": 1.0, + "content": "Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, and Dumitru Erhan. Domain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 99, + 387, + 521, + 399 + ], + "spans": [ + { + "bbox": [ + 99, + 387, + 521, + 399 + ], + "score": 1.0, + "content": "separation networks. In Advances in Neural Information Processing Systems (NIPS), pp. 343–351, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 90, + 376, + 521, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 407, + 519, + 442 + ], + "lines": [ + { + "bbox": [ + 90, + 407, + 522, + 421 + ], + "spans": [ + { + "bbox": [ + 90, + 407, + 522, + 421 + ], + "score": 1.0, + "content": "Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 100, + 418, + 522, + 432 + ], + "spans": [ + { + "bbox": [ + 100, + 418, + 522, + 432 + ], + "score": 1.0, + "content": "pixel-level domain adaptation with generative adversarial networks. In IEEE Conference on Computer", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 100, + 429, + 357, + 442 + ], + "spans": [ + { + "bbox": [ + 100, + 429, + 357, + 442 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition (CVPR), volume 1, pp. 7, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 90, + 407, + 522, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 449, + 520, + 473 + ], + "lines": [ + { + "bbox": [ + 89, + 449, + 521, + 463 + ], + "spans": [ + { + "bbox": [ + 89, + 449, + 521, + 463 + ], + "score": 1.0, + "content": "Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 461, + 453, + 474 + ], + "spans": [ + { + "bbox": [ + 99, + 461, + 453, + 474 + ], + "score": 1.0, + "content": "Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 89, + 449, + 521, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 480, + 518, + 505 + ], + "lines": [ + { + "bbox": [ + 89, + 480, + 520, + 494 + ], + "spans": [ + { + "bbox": [ + 89, + 480, + 520, + 494 + ], + "score": 1.0, + "content": "John S Bridle. Probabilistic interpretation of feedforward classification network outputs, with relationships to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 100, + 492, + 423, + 505 + ], + "spans": [ + { + "bbox": [ + 100, + 492, + 423, + 505 + ], + "score": 1.0, + "content": "statistical pattern recognition. In Neurocomputing, pp. 227–236. Springer, 1990.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 89, + 480, + 520, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 512, + 517, + 536 + ], + "lines": [ + { + "bbox": [ + 90, + 513, + 519, + 525 + ], + "spans": [ + { + "bbox": [ + 90, + 513, + 519, + 525 + ], + "score": 1.0, + "content": "Matthew Chalk, Olivier Marre, and Gasper Tkacik. Relevant sparse codes with variational information", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 100, + 523, + 492, + 537 + ], + "spans": [ + { + "bbox": [ + 100, + 523, + 492, + 537 + ], + "score": 1.0, + "content": "bottleneck. In Advances in Neural Information Processing Systems (NIPS), pp. 1957–1965, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 90, + 513, + 519, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 544, + 519, + 578 + ], + "lines": [ + { + "bbox": [ + 90, + 544, + 521, + 558 + ], + "spans": [ + { + "bbox": [ + 90, + 544, + 521, + 558 + ], + "score": 1.0, + "content": "Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 99, + 554, + 522, + 569 + ], + "spans": [ + { + "bbox": [ + 99, + 554, + 522, + 569 + ], + "score": 1.0, + "content": "Unified generative adversarial networks for multi-domain image-to-image translation. IEEE Conference on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 100, + 566, + 356, + 579 + ], + "spans": [ + { + "bbox": [ + 100, + 566, + 356, + 579 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition (CVPR), 1711, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 90, + 544, + 522, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 586, + 520, + 620 + ], + "lines": [ + { + "bbox": [ + 89, + 586, + 522, + 600 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 522, + 600 + ], + "score": 1.0, + "content": "Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 98, + 596, + 522, + 612 + ], + "spans": [ + { + "bbox": [ + 98, + 596, + 522, + 612 + ], + "score": 1.0, + "content": "transportation for domain adaptation. In Advances in Neural Information Processing Systems (NIPS), pp.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 99, + 608, + 175, + 621 + ], + "spans": [ + { + "bbox": [ + 99, + 608, + 175, + 621 + ], + "score": 1.0, + "content": "3733–3742, 2017.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 89, + 586, + 522, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 629, + 520, + 651 + ], + "lines": [ + { + "bbox": [ + 88, + 627, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 88, + 627, + 522, + 643 + ], + "score": 1.0, + "content": "Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. pp. 1–35. Springer,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 99, + 639, + 127, + 652 + ], + "spans": [ + { + "bbox": [ + 99, + 639, + 127, + 652 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 88, + 627, + 522, + 652 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 90, + 95, + 522, + 129 + ], + "lines": [ + { + "bbox": [ + 89, + 95, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 89, + 95, + 522, + 109 + ], + "score": 1.0, + "content": "Basura Fernando, Amaury Habrard, Marc Sebban, and Tinne Tuytelaars. Unsupervised visual domain", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 105, + 523, + 122 + ], + "spans": [ + { + "bbox": [ + 98, + 105, + 523, + 122 + ], + "score": 1.0, + "content": "adaptation using subspace alignment. In IEEE International Conference on Computer Vision (ICCV), pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 117, + 175, + 130 + ], + "spans": [ + { + "bbox": [ + 100, + 117, + 175, + 130 + ], + "score": 1.0, + "content": "2960–2967, 2013.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 90, + 136, + 519, + 159 + ], + "lines": [ + { + "bbox": [ + 90, + 135, + 521, + 149 + ], + "spans": [ + { + "bbox": [ + 90, + 135, + 521, + 149 + ], + "score": 1.0, + "content": "Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. International", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 147, + 298, + 159 + ], + "spans": [ + { + "bbox": [ + 99, + 147, + 298, + 159 + ], + "score": 1.0, + "content": "Conference on Machine Learning (ICML), 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 91, + 165, + 522, + 200 + ], + "lines": [ + { + "bbox": [ + 90, + 165, + 523, + 179 + ], + "spans": [ + { + "bbox": [ + 90, + 165, + 523, + 179 + ], + "score": 1.0, + "content": "Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 176, + 522, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 176, + 522, + 190 + ], + "score": 1.0, + "content": "Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 188, + 317, + 200 + ], + "spans": [ + { + "bbox": [ + 98, + 188, + 317, + 200 + ], + "score": 1.0, + "content": "Machine Learning Research, 17(1):2096–2030, 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 91, + 206, + 522, + 240 + ], + "lines": [ + { + "bbox": [ + 90, + 206, + 522, + 219 + ], + "spans": [ + { + "bbox": [ + 90, + 206, + 522, + 219 + ], + "score": 1.0, + "content": "Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 217, + 523, + 231 + ], + "spans": [ + { + "bbox": [ + 99, + 217, + 523, + 231 + ], + "score": 1.0, + "content": "adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2066–2073,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 228, + 125, + 241 + ], + "spans": [ + { + "bbox": [ + 99, + 228, + 125, + 241 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 92, + 246, + 522, + 271 + ], + "lines": [ + { + "bbox": [ + 89, + 246, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 89, + 246, + 522, + 261 + ], + "score": 1.0, + "content": "Guang-Yuan Hao, Hong-Xing Yu, and Wei-Shi Zheng. Mixgan: Learning concepts from different domains", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 258, + 359, + 271 + ], + "spans": [ + { + "bbox": [ + 99, + 258, + 359, + 271 + ], + "score": 1.0, + "content": "for mixture generation. arXiv preprint arXiv:1807.01659, 2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 90, + 276, + 523, + 310 + ], + "lines": [ + { + "bbox": [ + 89, + 276, + 523, + 290 + ], + "spans": [ + { + "bbox": [ + 89, + 276, + 523, + 290 + ], + "score": 1.0, + "content": "Hirokazu Kameoka, Takuhiro Kaneko, Kou Tanaka, and Nobukatsu Hojo. Stargan-vc: Non-parallel many-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 99, + 287, + 522, + 301 + ], + "spans": [ + { + "bbox": [ + 99, + 287, + 522, + 301 + ], + "score": 1.0, + "content": "to-many voice conversion with star generative adversarial networks. arXiv preprint arXiv:1806.02169,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 99, + 297, + 126, + 312 + ], + "spans": [ + { + "bbox": [ + 99, + 297, + 126, + 312 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 90, + 317, + 520, + 340 + ], + "lines": [ + { + "bbox": [ + 88, + 316, + 522, + 331 + ], + "spans": [ + { + "bbox": [ + 88, + 316, + 522, + 331 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 99, + 329, + 272, + 340 + ], + "spans": [ + { + "bbox": [ + 99, + 329, + 272, + 340 + ], + "score": 1.0, + "content": "on Learning Representation (ICLR), 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 93, + 347, + 520, + 371 + ], + "lines": [ + { + "bbox": [ + 91, + 347, + 522, + 360 + ], + "spans": [ + { + "bbox": [ + 91, + 347, + 522, + 360 + ], + "score": 1.0, + "content": "Elyor Kodirov, Tao Xiang, Zhenyong Fu, and Shaogang Gong. Unsupervised domain adaptation for zero-shot", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 99, + 358, + 483, + 371 + ], + "spans": [ + { + "bbox": [ + 99, + 358, + 483, + 371 + ], + "score": 1.0, + "content": "learning. In IEEE International Conference on Computer Vision (ICCV), pp. 2452–2460, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 91, + 376, + 521, + 400 + ], + "lines": [ + { + "bbox": [ + 90, + 376, + 521, + 390 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 521, + 390 + ], + "score": 1.0, + "content": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 99, + 388, + 358, + 399 + ], + "spans": [ + { + "bbox": [ + 99, + 388, + 358, + 399 + ], + "score": 1.0, + "content": "recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 90, + 406, + 522, + 440 + ], + "lines": [ + { + "bbox": [ + 90, + 406, + 521, + 419 + ], + "spans": [ + { + "bbox": [ + 90, + 406, + 521, + 419 + ], + "score": 1.0, + "content": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 98, + 417, + 523, + 431 + ], + "spans": [ + { + "bbox": [ + 98, + 417, + 523, + 431 + ], + "score": 1.0, + "content": "generalization. In IEEE International Conference on Computer Vision (ICCV), pp. 5543–5551. IEEE,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 428, + 126, + 441 + ], + "spans": [ + { + "bbox": [ + 99, + 428, + 126, + 441 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 89, + 447, + 521, + 470 + ], + "lines": [ + { + "bbox": [ + 88, + 446, + 523, + 461 + ], + "spans": [ + { + "bbox": [ + 88, + 446, + 523, + 461 + ], + "score": 1.0, + "content": "Jianhua Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information theory,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 99, + 459, + 190, + 471 + ], + "spans": [ + { + "bbox": [ + 99, + 459, + 190, + 471 + ], + "score": 1.0, + "content": "37(1):145–151, 1991.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 91, + 477, + 521, + 511 + ], + "lines": [ + { + "bbox": [ + 90, + 477, + 523, + 489 + ], + "spans": [ + { + "bbox": [ + 90, + 477, + 523, + 489 + ], + "score": 1.0, + "content": "Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In D. D. Lee, M. Sugiyama, U. V.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 98, + 487, + 523, + 502 + ], + "spans": [ + { + "bbox": [ + 98, + 487, + 523, + 502 + ], + "score": 1.0, + "content": "Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems (NIPS), pp.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 99, + 500, + 262, + 511 + ], + "spans": [ + { + "bbox": [ + 99, + 500, + 262, + 511 + ], + "score": 1.0, + "content": "469–477. Curran Associates, Inc., 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 87, + 517, + 521, + 541 + ], + "lines": [ + { + "bbox": [ + 90, + 518, + 522, + 531 + ], + "spans": [ + { + "bbox": [ + 90, + 518, + 522, + 531 + ], + "score": 1.0, + "content": "Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 97, + 528, + 424, + 541 + ], + "spans": [ + { + "bbox": [ + 97, + 528, + 424, + 541 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems (NIPS), pp. 700–708, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 89, + 547, + 520, + 581 + ], + "lines": [ + { + "bbox": [ + 89, + 546, + 522, + 562 + ], + "spans": [ + { + "bbox": [ + 89, + 546, + 522, + 562 + ], + "score": 1.0, + "content": "Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 100, + 558, + 522, + 572 + ], + "spans": [ + { + "bbox": [ + 100, + 558, + 522, + 572 + ], + "score": 1.0, + "content": "for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 99, + 569, + 226, + 583 + ], + "spans": [ + { + "bbox": [ + 99, + 569, + 226, + 583 + ], + "score": 1.0, + "content": "(CVPR), pp. 1410–1417, 2014.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 88, + 588, + 521, + 611 + ], + "lines": [ + { + "bbox": [ + 89, + 587, + 522, + 602 + ], + "spans": [ + { + "bbox": [ + 89, + 587, + 522, + 602 + ], + "score": 1.0, + "content": "Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I Jordan. Learning transferable features with deep", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 99, + 599, + 439, + 612 + ], + "spans": [ + { + "bbox": [ + 99, + 599, + 439, + 612 + ], + "score": 1.0, + "content": "adaptation networks. International Conference on Machine Learning (ICML), 2015.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 90, + 618, + 523, + 651 + ], + "lines": [ + { + "bbox": [ + 90, + 618, + 521, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 618, + 521, + 630 + ], + "score": 1.0, + "content": "Mingsheng Long, Jianmin Wang, Yue Cao, Jiaguang Sun, and S Yu Philip. Deep learning of transferable", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 98, + 628, + 523, + 644 + ], + "spans": [ + { + "bbox": [ + 98, + 628, + 523, + 644 + ], + "score": 1.0, + "content": "representation for scalable domain adaptation. IEEE Transactions on Knowledge and Data Engineering,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 100, + 640, + 199, + 652 + ], + "spans": [ + { + "bbox": [ + 100, + 640, + 199, + 652 + ], + "score": 1.0, + "content": "28(8):2027–2040, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 51 + ], + "lines": [ + { + "bbox": [ + 90, + 38, + 293, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 38, + 293, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 90, + 95, + 522, + 129 + ], + "lines": [ + { + "bbox": [ + 89, + 95, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 89, + 95, + 522, + 109 + ], + "score": 1.0, + "content": "Basura Fernando, Amaury Habrard, Marc Sebban, and Tinne Tuytelaars. Unsupervised visual domain", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 105, + 523, + 122 + ], + "spans": [ + { + "bbox": [ + 98, + 105, + 523, + 122 + ], + "score": 1.0, + "content": "adaptation using subspace alignment. In IEEE International Conference on Computer Vision (ICCV), pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 117, + 175, + 130 + ], + "spans": [ + { + "bbox": [ + 100, + 117, + 175, + 130 + ], + "score": 1.0, + "content": "2960–2967, 2013.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 89, + 95, + 523, + 130 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 136, + 519, + 159 + ], + "lines": [ + { + "bbox": [ + 90, + 135, + 521, + 149 + ], + "spans": [ + { + "bbox": [ + 90, + 135, + 521, + 149 + ], + "score": 1.0, + "content": "Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. International", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 147, + 298, + 159 + ], + "spans": [ + { + "bbox": [ + 99, + 147, + 298, + 159 + ], + "score": 1.0, + "content": "Conference on Machine Learning (ICML), 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 90, + 135, + 521, + 159 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 165, + 522, + 200 + ], + "lines": [ + { + "bbox": [ + 90, + 165, + 523, + 179 + ], + "spans": [ + { + "bbox": [ + 90, + 165, + 523, + 179 + ], + "score": 1.0, + "content": "Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 176, + 522, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 176, + 522, + 190 + ], + "score": 1.0, + "content": "Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 188, + 317, + 200 + ], + "spans": [ + { + "bbox": [ + 98, + 188, + 317, + 200 + ], + "score": 1.0, + "content": "Machine Learning Research, 17(1):2096–2030, 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 90, + 165, + 523, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 206, + 522, + 240 + ], + "lines": [ + { + "bbox": [ + 90, + 206, + 522, + 219 + ], + "spans": [ + { + "bbox": [ + 90, + 206, + 522, + 219 + ], + "score": 1.0, + "content": "Boqing Gong, Yuan Shi, Fei Sha, and Kristen Grauman. Geodesic flow kernel for unsupervised domain", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 217, + 523, + 231 + ], + "spans": [ + { + "bbox": [ + 99, + 217, + 523, + 231 + ], + "score": 1.0, + "content": "adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2066–2073,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 228, + 125, + 241 + ], + "spans": [ + { + "bbox": [ + 99, + 228, + 125, + 241 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 90, + 206, + 523, + 241 + ] + }, + { + "type": "text", + "bbox": [ + 92, + 246, + 522, + 271 + ], + "lines": [ + { + "bbox": [ + 89, + 246, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 89, + 246, + 522, + 261 + ], + "score": 1.0, + "content": "Guang-Yuan Hao, Hong-Xing Yu, and Wei-Shi Zheng. Mixgan: Learning concepts from different domains", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 258, + 359, + 271 + ], + "spans": [ + { + "bbox": [ + 99, + 258, + 359, + 271 + ], + "score": 1.0, + "content": "for mixture generation. arXiv preprint arXiv:1807.01659, 2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 89, + 246, + 522, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 276, + 523, + 310 + ], + "lines": [ + { + "bbox": [ + 89, + 276, + 523, + 290 + ], + "spans": [ + { + "bbox": [ + 89, + 276, + 523, + 290 + ], + "score": 1.0, + "content": "Hirokazu Kameoka, Takuhiro Kaneko, Kou Tanaka, and Nobukatsu Hojo. Stargan-vc: Non-parallel many-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 99, + 287, + 522, + 301 + ], + "spans": [ + { + "bbox": [ + 99, + 287, + 522, + 301 + ], + "score": 1.0, + "content": "to-many voice conversion with star generative adversarial networks. arXiv preprint arXiv:1806.02169,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 99, + 297, + 126, + 312 + ], + "spans": [ + { + "bbox": [ + 99, + 297, + 126, + 312 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 89, + 276, + 523, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 317, + 520, + 340 + ], + "lines": [ + { + "bbox": [ + 88, + 316, + 522, + 331 + ], + "spans": [ + { + "bbox": [ + 88, + 316, + 522, + 331 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 99, + 329, + 272, + 340 + ], + "spans": [ + { + "bbox": [ + 99, + 329, + 272, + 340 + ], + "score": 1.0, + "content": "on Learning Representation (ICLR), 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 88, + 316, + 522, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 93, + 347, + 520, + 371 + ], + "lines": [ + { + "bbox": [ + 91, + 347, + 522, + 360 + ], + "spans": [ + { + "bbox": [ + 91, + 347, + 522, + 360 + ], + "score": 1.0, + "content": "Elyor Kodirov, Tao Xiang, Zhenyong Fu, and Shaogang Gong. Unsupervised domain adaptation for zero-shot", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 99, + 358, + 483, + 371 + ], + "spans": [ + { + "bbox": [ + 99, + 358, + 483, + 371 + ], + "score": 1.0, + "content": "learning. In IEEE International Conference on Computer Vision (ICCV), pp. 2452–2460, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 91, + 347, + 522, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 376, + 521, + 400 + ], + "lines": [ + { + "bbox": [ + 90, + 376, + 521, + 390 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 521, + 390 + ], + "score": 1.0, + "content": "Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 99, + 388, + 358, + 399 + ], + "spans": [ + { + "bbox": [ + 99, + 388, + 358, + 399 + ], + "score": 1.0, + "content": "recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 90, + 376, + 521, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 406, + 522, + 440 + ], + "lines": [ + { + "bbox": [ + 90, + 406, + 521, + 419 + ], + "spans": [ + { + "bbox": [ + 90, + 406, + 521, + 419 + ], + "score": 1.0, + "content": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 98, + 417, + 523, + 431 + ], + "spans": [ + { + "bbox": [ + 98, + 417, + 523, + 431 + ], + "score": 1.0, + "content": "generalization. In IEEE International Conference on Computer Vision (ICCV), pp. 5543–5551. IEEE,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 428, + 126, + 441 + ], + "spans": [ + { + "bbox": [ + 99, + 428, + 126, + 441 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 90, + 406, + 523, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 447, + 521, + 470 + ], + "lines": [ + { + "bbox": [ + 88, + 446, + 523, + 461 + ], + "spans": [ + { + "bbox": [ + 88, + 446, + 523, + 461 + ], + "score": 1.0, + "content": "Jianhua Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information theory,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 99, + 459, + 190, + 471 + ], + "spans": [ + { + "bbox": [ + 99, + 459, + 190, + 471 + ], + "score": 1.0, + "content": "37(1):145–151, 1991.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 88, + 446, + 523, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 477, + 521, + 511 + ], + "lines": [ + { + "bbox": [ + 90, + 477, + 523, + 489 + ], + "spans": [ + { + "bbox": [ + 90, + 477, + 523, + 489 + ], + "score": 1.0, + "content": "Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In D. D. Lee, M. Sugiyama, U. V.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 98, + 487, + 523, + 502 + ], + "spans": [ + { + "bbox": [ + 98, + 487, + 523, + 502 + ], + "score": 1.0, + "content": "Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems (NIPS), pp.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 99, + 500, + 262, + 511 + ], + "spans": [ + { + "bbox": [ + 99, + 500, + 262, + 511 + ], + "score": 1.0, + "content": "469–477. Curran Associates, Inc., 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 90, + 477, + 523, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 87, + 517, + 521, + 541 + ], + "lines": [ + { + "bbox": [ + 90, + 518, + 522, + 531 + ], + "spans": [ + { + "bbox": [ + 90, + 518, + 522, + 531 + ], + "score": 1.0, + "content": "Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 97, + 528, + 424, + 541 + ], + "spans": [ + { + "bbox": [ + 97, + 528, + 424, + 541 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems (NIPS), pp. 700–708, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 90, + 518, + 522, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 547, + 520, + 581 + ], + "lines": [ + { + "bbox": [ + 89, + 546, + 522, + 562 + ], + "spans": [ + { + "bbox": [ + 89, + 546, + 522, + 562 + ], + "score": 1.0, + "content": "Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 100, + 558, + 522, + 572 + ], + "spans": [ + { + "bbox": [ + 100, + 558, + 522, + 572 + ], + "score": 1.0, + "content": "for unsupervised domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 99, + 569, + 226, + 583 + ], + "spans": [ + { + "bbox": [ + 99, + 569, + 226, + 583 + ], + "score": 1.0, + "content": "(CVPR), pp. 1410–1417, 2014.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 89, + 546, + 522, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 588, + 521, + 611 + ], + "lines": [ + { + "bbox": [ + 89, + 587, + 522, + 602 + ], + "spans": [ + { + "bbox": [ + 89, + 587, + 522, + 602 + ], + "score": 1.0, + "content": "Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I Jordan. Learning transferable features with deep", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 99, + 599, + 439, + 612 + ], + "spans": [ + { + "bbox": [ + 99, + 599, + 439, + 612 + ], + "score": 1.0, + "content": "adaptation networks. International Conference on Machine Learning (ICML), 2015.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 89, + 587, + 522, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 618, + 523, + 651 + ], + "lines": [ + { + "bbox": [ + 90, + 618, + 521, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 618, + 521, + 630 + ], + "score": 1.0, + "content": "Mingsheng Long, Jianmin Wang, Yue Cao, Jiaguang Sun, and S Yu Philip. Deep learning of transferable", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 98, + 628, + 523, + 644 + ], + "spans": [ + { + "bbox": [ + 98, + 628, + 523, + 644 + ], + "score": 1.0, + "content": "representation for scalable domain adaptation. IEEE Transactions on Knowledge and Data Engineering,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 100, + 640, + 199, + 652 + ], + "spans": [ + { + "bbox": [ + 100, + 640, + 199, + 652 + ], + "score": 1.0, + "content": "28(8):2027–2040, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 90, + 618, + 523, + 652 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 89, + 95, + 520, + 118 + ], + "lines": [ + { + "bbox": [ + 88, + 93, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 88, + 93, + 522, + 109 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 107, + 248, + 117 + ], + "spans": [ + { + "bbox": [ + 99, + 107, + 248, + 117 + ], + "score": 1.0, + "content": "Research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 90, + 125, + 522, + 159 + ], + "lines": [ + { + "bbox": [ + 88, + 123, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 123, + 522, + 140 + ], + "score": 1.0, + "content": "Shakir Mohamed and Danilo Jimenez Rezende. Variational information maximisation for intrinsically", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 135, + 523, + 151 + ], + "spans": [ + { + "bbox": [ + 98, + 135, + 523, + 151 + ], + "score": 1.0, + "content": "motivated reinforcement learning. In Advances in neural information processing systems (NIPS), pp.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 146, + 175, + 160 + ], + "spans": [ + { + "bbox": [ + 99, + 146, + 175, + 160 + ], + "score": 1.0, + "content": "2125–2133, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 90, + 165, + 519, + 189 + ], + "lines": [ + { + "bbox": [ + 91, + 167, + 520, + 178 + ], + "spans": [ + { + "bbox": [ + 91, + 167, + 520, + 178 + ], + "score": 1.0, + "content": "Saeid Motiian, Quinn Jones, Seyed Iranmanesh, and Gianfranco Doretto. Few-shot adversarial domain", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 177, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 177, + 493, + 190 + ], + "score": 1.0, + "content": "adaptation. In Advances in Neural Information Processing Systems (NIPS), pp. 6673–6683, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 89, + 195, + 520, + 230 + ], + "lines": [ + { + "bbox": [ + 90, + 195, + 522, + 209 + ], + "spans": [ + { + "bbox": [ + 90, + 195, + 522, + 209 + ], + "score": 1.0, + "content": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 207, + 522, + 220 + ], + "spans": [ + { + "bbox": [ + 99, + 207, + 522, + 220 + ], + "score": 1.0, + "content": "natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 218, + 279, + 231 + ], + "spans": [ + { + "bbox": [ + 99, + 218, + 279, + 231 + ], + "score": 1.0, + "content": "feature learning, volume 2011, pp. 5, 2011.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 90, + 236, + 522, + 271 + ], + "lines": [ + { + "bbox": [ + 90, + 236, + 523, + 250 + ], + "spans": [ + { + "bbox": [ + 90, + 236, + 523, + 250 + ], + "score": 1.0, + "content": "Yuchen Pu, Weiyao Wang, Ricardo Henao, Liqun Chen, Zhe Gan, Chunyuan Li, and Lawrence Carin.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 247, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 99, + 247, + 522, + 261 + ], + "score": 1.0, + "content": "Adversarial symmetric variational autoencoder. In Advances in Neural Information Processing Systems", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 259, + 222, + 272 + ], + "spans": [ + { + "bbox": [ + 99, + 259, + 222, + 272 + ], + "score": 1.0, + "content": "(NIPS), pp. 4330–4339, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 91, + 277, + 521, + 301 + ], + "lines": [ + { + "bbox": [ + 90, + 278, + 521, + 290 + ], + "spans": [ + { + "bbox": [ + 90, + 278, + 521, + 290 + ], + "score": 1.0, + "content": "Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 99, + 289, + 474, + 301 + ], + "spans": [ + { + "bbox": [ + 99, + 289, + 474, + 301 + ], + "score": 1.0, + "content": "adapters. In Advances in Neural Information Processing Systems (NIPS), pp. 506–516, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 92, + 307, + 521, + 331 + ], + "lines": [ + { + "bbox": [ + 89, + 306, + 521, + 322 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 521, + 322 + ], + "score": 1.0, + "content": "Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 100, + 319, + 401, + 332 + ], + "spans": [ + { + "bbox": [ + 100, + 319, + 401, + 332 + ], + "score": 1.0, + "content": "adaptation. International Conference on Machine Learning (ICML), 2017.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 90, + 337, + 519, + 361 + ], + "lines": [ + { + "bbox": [ + 89, + 337, + 522, + 351 + ], + "spans": [ + { + "bbox": [ + 89, + 337, + 522, + 351 + ], + "score": 1.0, + "content": "Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 99, + 349, + 497, + 361 + ], + "spans": [ + { + "bbox": [ + 99, + 349, + 497, + 361 + ], + "score": 1.0, + "content": "In International Conference on Artificial Intelligence and Statistics (AISTATS), pp. 701–708, 2010.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 91, + 367, + 521, + 391 + ], + "lines": [ + { + "bbox": [ + 89, + 367, + 521, + 381 + ], + "spans": [ + { + "bbox": [ + 89, + 367, + 521, + 381 + ], + "score": 1.0, + "content": "Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 100, + 378, + 386, + 391 + ], + "spans": [ + { + "bbox": [ + 100, + 378, + 386, + 391 + ], + "score": 1.0, + "content": "Conference on Computer Vision (ECCV), pp. 443–450. Springer, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 90, + 397, + 520, + 420 + ], + "lines": [ + { + "bbox": [ + 89, + 398, + 522, + 411 + ], + "spans": [ + { + "bbox": [ + 89, + 398, + 522, + 411 + ], + "score": 1.0, + "content": "Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. Thirtieth AAAI", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 100, + 409, + 275, + 420 + ], + "spans": [ + { + "bbox": [ + 100, + 409, + 275, + 420 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 88, + 427, + 521, + 451 + ], + "lines": [ + { + "bbox": [ + 89, + 427, + 522, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 427, + 522, + 440 + ], + "score": 1.0, + "content": "Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In IEEE", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 438, + 338, + 450 + ], + "spans": [ + { + "bbox": [ + 99, + 438, + 338, + 450 + ], + "score": 1.0, + "content": "Information Theory Workshop (ITW), pp. 1–5. IEEE, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 88, + 456, + 521, + 481 + ], + "lines": [ + { + "bbox": [ + 90, + 458, + 523, + 470 + ], + "spans": [ + { + "bbox": [ + 90, + 458, + 523, + 470 + ], + "score": 1.0, + "content": "Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 99, + 468, + 496, + 481 + ], + "spans": [ + { + "bbox": [ + 99, + 468, + 496, + 481 + ], + "score": 1.0, + "content": "In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pp. 4, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 91, + 487, + 522, + 522 + ], + "lines": [ + { + "bbox": [ + 90, + 487, + 522, + 501 + ], + "spans": [ + { + "bbox": [ + 90, + 487, + 522, + 501 + ], + "score": 1.0, + "content": "Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 99, + 497, + 522, + 512 + ], + "spans": [ + { + "bbox": [ + 99, + 497, + 522, + 512 + ], + "score": 1.0, + "content": "bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In IEEE Conference on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 99, + 508, + 350, + 523 + ], + "spans": [ + { + "bbox": [ + 99, + 508, + 350, + 523 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 89, + 528, + 521, + 551 + ], + "lines": [ + { + "bbox": [ + 89, + 528, + 523, + 541 + ], + "spans": [ + { + "bbox": [ + 89, + 528, + 523, + 541 + ], + "score": 1.0, + "content": "Donggeun Yoo, Namil Kim, Sunggyun Park, Anthony S Paek, and In So Kweon. Pixel-level domain transfer.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 99, + 539, + 399, + 552 + ], + "spans": [ + { + "bbox": [ + 99, + 539, + 399, + 552 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision (ECCV), pp. 517–532, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 92, + 558, + 520, + 592 + ], + "lines": [ + { + "bbox": [ + 90, + 558, + 522, + 571 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 522, + 571 + ], + "score": 1.0, + "content": "Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschläger, and Susanne Saminger-Platz.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 100, + 569, + 521, + 582 + ], + "spans": [ + { + "bbox": [ + 100, + 569, + 521, + 582 + ], + "score": 1.0, + "content": "Central moment discrepancy (cmd) for domain-invariant representation learning. International Conference", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 100, + 581, + 272, + 593 + ], + "spans": [ + { + "bbox": [ + 100, + 581, + 272, + 593 + ], + "score": 1.0, + "content": "on Learning Representation (ICLR), 2017.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 88, + 598, + 520, + 622 + ], + "lines": [ + { + "bbox": [ + 90, + 599, + 522, + 611 + ], + "spans": [ + { + "bbox": [ + 90, + 599, + 522, + 611 + ], + "score": 1.0, + "content": "Jing Zhang, Wanqing Li, and Philip Ogunbona. Joint geometrical and statistical alignment for visual domain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 99, + 610, + 493, + 622 + ], + "spans": [ + { + "bbox": [ + 99, + 610, + 493, + 622 + ], + "score": 1.0, + "content": "adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 90, + 629, + 519, + 652 + ], + "lines": [ + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "score": 1.0, + "content": "Xu Zhang, Felix Xinnan Yu, Shih-Fu Chang, and Shengjin Wang. Deep transfer network: Unsupervised", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 99, + 640, + 343, + 652 + ], + "spans": [ + { + "bbox": [ + 99, + 640, + 343, + 652 + ], + "score": 1.0, + "content": "domain adaptation. arXiv preprint arXiv:1503.00591, 2015.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 51 + ], + "lines": [ + { + "bbox": [ + 90, + 38, + 293, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 38, + 293, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 89, + 95, + 520, + 118 + ], + "lines": [ + { + "bbox": [ + 88, + 93, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 88, + 93, + 522, + 109 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 107, + 248, + 117 + ], + "spans": [ + { + "bbox": [ + 99, + 107, + 248, + 117 + ], + "score": 1.0, + "content": "Research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 88, + 93, + 522, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 125, + 522, + 159 + ], + "lines": [ + { + "bbox": [ + 88, + 123, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 123, + 522, + 140 + ], + "score": 1.0, + "content": "Shakir Mohamed and Danilo Jimenez Rezende. Variational information maximisation for intrinsically", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 135, + 523, + 151 + ], + "spans": [ + { + "bbox": [ + 98, + 135, + 523, + 151 + ], + "score": 1.0, + "content": "motivated reinforcement learning. In Advances in neural information processing systems (NIPS), pp.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 146, + 175, + 160 + ], + "spans": [ + { + "bbox": [ + 99, + 146, + 175, + 160 + ], + "score": 1.0, + "content": "2125–2133, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 88, + 123, + 523, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 165, + 519, + 189 + ], + "lines": [ + { + "bbox": [ + 91, + 167, + 520, + 178 + ], + "spans": [ + { + "bbox": [ + 91, + 167, + 520, + 178 + ], + "score": 1.0, + "content": "Saeid Motiian, Quinn Jones, Seyed Iranmanesh, and Gianfranco Doretto. Few-shot adversarial domain", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 177, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 177, + 493, + 190 + ], + "score": 1.0, + "content": "adaptation. In Advances in Neural Information Processing Systems (NIPS), pp. 6673–6683, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 91, + 167, + 520, + 190 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 195, + 520, + 230 + ], + "lines": [ + { + "bbox": [ + 90, + 195, + 522, + 209 + ], + "spans": [ + { + "bbox": [ + 90, + 195, + 522, + 209 + ], + "score": 1.0, + "content": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 207, + 522, + 220 + ], + "spans": [ + { + "bbox": [ + 99, + 207, + 522, + 220 + ], + "score": 1.0, + "content": "natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 218, + 279, + 231 + ], + "spans": [ + { + "bbox": [ + 99, + 218, + 279, + 231 + ], + "score": 1.0, + "content": "feature learning, volume 2011, pp. 5, 2011.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 90, + 195, + 522, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 236, + 522, + 271 + ], + "lines": [ + { + "bbox": [ + 90, + 236, + 523, + 250 + ], + "spans": [ + { + "bbox": [ + 90, + 236, + 523, + 250 + ], + "score": 1.0, + "content": "Yuchen Pu, Weiyao Wang, Ricardo Henao, Liqun Chen, Zhe Gan, Chunyuan Li, and Lawrence Carin.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 247, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 99, + 247, + 522, + 261 + ], + "score": 1.0, + "content": "Adversarial symmetric variational autoencoder. In Advances in Neural Information Processing Systems", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 259, + 222, + 272 + ], + "spans": [ + { + "bbox": [ + 99, + 259, + 222, + 272 + ], + "score": 1.0, + "content": "(NIPS), pp. 4330–4339, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 90, + 236, + 523, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 277, + 521, + 301 + ], + "lines": [ + { + "bbox": [ + 90, + 278, + 521, + 290 + ], + "spans": [ + { + "bbox": [ + 90, + 278, + 521, + 290 + ], + "score": 1.0, + "content": "Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 99, + 289, + 474, + 301 + ], + "spans": [ + { + "bbox": [ + 99, + 289, + 474, + 301 + ], + "score": 1.0, + "content": "adapters. In Advances in Neural Information Processing Systems (NIPS), pp. 506–516, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 90, + 278, + 521, + 301 + ] + }, + { + "type": "text", + "bbox": [ + 92, + 307, + 521, + 331 + ], + "lines": [ + { + "bbox": [ + 89, + 306, + 521, + 322 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 521, + 322 + ], + "score": 1.0, + "content": "Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 100, + 319, + 401, + 332 + ], + "spans": [ + { + "bbox": [ + 100, + 319, + 401, + 332 + ], + "score": 1.0, + "content": "adaptation. International Conference on Machine Learning (ICML), 2017.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 89, + 306, + 521, + 332 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 337, + 519, + 361 + ], + "lines": [ + { + "bbox": [ + 89, + 337, + 522, + 351 + ], + "spans": [ + { + "bbox": [ + 89, + 337, + 522, + 351 + ], + "score": 1.0, + "content": "Mathieu Salzmann, Carl Henrik Ek, Raquel Urtasun, and Trevor Darrell. Factorized orthogonal latent spaces.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 99, + 349, + 497, + 361 + ], + "spans": [ + { + "bbox": [ + 99, + 349, + 497, + 361 + ], + "score": 1.0, + "content": "In International Conference on Artificial Intelligence and Statistics (AISTATS), pp. 701–708, 2010.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 89, + 337, + 522, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 367, + 521, + 391 + ], + "lines": [ + { + "bbox": [ + 89, + 367, + 521, + 381 + ], + "spans": [ + { + "bbox": [ + 89, + 367, + 521, + 381 + ], + "score": 1.0, + "content": "Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 100, + 378, + 386, + 391 + ], + "spans": [ + { + "bbox": [ + 100, + 378, + 386, + 391 + ], + "score": 1.0, + "content": "Conference on Computer Vision (ECCV), pp. 443–450. Springer, 2016.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 89, + 367, + 521, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 397, + 520, + 420 + ], + "lines": [ + { + "bbox": [ + 89, + 398, + 522, + 411 + ], + "spans": [ + { + "bbox": [ + 89, + 398, + 522, + 411 + ], + "score": 1.0, + "content": "Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. Thirtieth AAAI", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 100, + 409, + 275, + 420 + ], + "spans": [ + { + "bbox": [ + 100, + 409, + 275, + 420 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 89, + 398, + 522, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 427, + 521, + 451 + ], + "lines": [ + { + "bbox": [ + 89, + 427, + 522, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 427, + 522, + 440 + ], + "score": 1.0, + "content": "Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In IEEE", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 438, + 338, + 450 + ], + "spans": [ + { + "bbox": [ + 99, + 438, + 338, + 450 + ], + "score": 1.0, + "content": "Information Theory Workshop (ITW), pp. 1–5. IEEE, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 89, + 427, + 522, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 456, + 521, + 481 + ], + "lines": [ + { + "bbox": [ + 90, + 458, + 523, + 470 + ], + "spans": [ + { + "bbox": [ + 90, + 458, + 523, + 470 + ], + "score": 1.0, + "content": "Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 99, + 468, + 496, + 481 + ], + "spans": [ + { + "bbox": [ + 99, + 468, + 496, + 481 + ], + "score": 1.0, + "content": "In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pp. 4, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 90, + 458, + 523, + 481 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 487, + 522, + 522 + ], + "lines": [ + { + "bbox": [ + 90, + 487, + 522, + 501 + ], + "spans": [ + { + "bbox": [ + 90, + 487, + 522, + 501 + ], + "score": 1.0, + "content": "Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 99, + 497, + 522, + 512 + ], + "spans": [ + { + "bbox": [ + 99, + 497, + 522, + 512 + ], + "score": 1.0, + "content": "bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In IEEE Conference on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 99, + 508, + 350, + 523 + ], + "spans": [ + { + "bbox": [ + 99, + 508, + 350, + 523 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 90, + 487, + 522, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 89, + 528, + 521, + 551 + ], + "lines": [ + { + "bbox": [ + 89, + 528, + 523, + 541 + ], + "spans": [ + { + "bbox": [ + 89, + 528, + 523, + 541 + ], + "score": 1.0, + "content": "Donggeun Yoo, Namil Kim, Sunggyun Park, Anthony S Paek, and In So Kweon. Pixel-level domain transfer.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 99, + 539, + 399, + 552 + ], + "spans": [ + { + "bbox": [ + 99, + 539, + 399, + 552 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision (ECCV), pp. 517–532, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 89, + 528, + 523, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 92, + 558, + 520, + 592 + ], + "lines": [ + { + "bbox": [ + 90, + 558, + 522, + 571 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 522, + 571 + ], + "score": 1.0, + "content": "Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschläger, and Susanne Saminger-Platz.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 100, + 569, + 521, + 582 + ], + "spans": [ + { + "bbox": [ + 100, + 569, + 521, + 582 + ], + "score": 1.0, + "content": "Central moment discrepancy (cmd) for domain-invariant representation learning. International Conference", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 100, + 581, + 272, + 593 + ], + "spans": [ + { + "bbox": [ + 100, + 581, + 272, + 593 + ], + "score": 1.0, + "content": "on Learning Representation (ICLR), 2017.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 90, + 558, + 522, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 598, + 520, + 622 + ], + "lines": [ + { + "bbox": [ + 90, + 599, + 522, + 611 + ], + "spans": [ + { + "bbox": [ + 90, + 599, + 522, + 611 + ], + "score": 1.0, + "content": "Jing Zhang, Wanqing Li, and Philip Ogunbona. Joint geometrical and statistical alignment for visual domain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 99, + 610, + 493, + 622 + ], + "spans": [ + { + "bbox": [ + 99, + 610, + 493, + 622 + ], + "score": 1.0, + "content": "adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 90, + 599, + 522, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 90, + 629, + 519, + 652 + ], + "lines": [ + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "score": 1.0, + "content": "Xu Zhang, Felix Xinnan Yu, Shih-Fu Chang, and Shengjin Wang. Deep transfer network: Unsupervised", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 99, + 640, + 343, + 652 + ], + "spans": [ + { + "bbox": [ + 99, + 640, + 343, + 652 + ], + "score": 1.0, + "content": "domain adaptation. arXiv preprint arXiv:1503.00591, 2015.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 90, + 629, + 521, + 652 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 95, + 522, + 129 + ], + "lines": [ + { + "bbox": [ + 90, + 95, + 523, + 109 + ], + "spans": [ + { + "bbox": [ + 90, + 95, + 523, + 109 + ], + "score": 1.0, + "content": "Han Zhao, Shanghang Zhang, Guanhang Wu, João P. Costeira, Jose Moura, and Geoffrey J. Gordon.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 105, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 98, + 105, + 522, + 121 + ], + "score": 1.0, + "content": "Multiple source domain adaptation with adversarial training of neural networks. International Conference", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 117, + 330, + 131 + ], + "spans": [ + { + "bbox": [ + 99, + 117, + 330, + 131 + ], + "score": 1.0, + "content": "on Learning Representation (ICLR) workshops, 05 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 91, + 152, + 491, + 165 + ], + "lines": [ + { + "bbox": [ + 90, + 151, + 491, + 165 + ], + "spans": [ + { + "bbox": [ + 90, + 151, + 491, + 165 + ], + "score": 1.0, + "content": "A CONNECTION TO INFORMATION THEORETIC REPRESENTATION LEARNING", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 91, + 177, + 521, + 244 + ], + "lines": [ + { + "bbox": [ + 88, + 176, + 522, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 176, + 522, + 191 + ], + "score": 1.0, + "content": "The idea of using information theoretic (IT) objectives for representation learning was originally introduced", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 188, + 522, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 522, + 201 + ], + "score": 1.0, + "content": "in (Tishby & Zaslavsky (2015)). Since their approach for optimizing the IT objective functions relied on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 200, + 521, + 212 + ], + "spans": [ + { + "bbox": [ + 90, + 200, + 521, + 212 + ], + "score": 1.0, + "content": "iterative Blahut Arimoto algorithm (Tishby & Zaslavsky (2015)), it is not feasible to apply to deep neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 211, + 522, + 224 + ], + "spans": [ + { + "bbox": [ + 89, + 211, + 522, + 224 + ], + "score": 1.0, + "content": "network (DNN) frameworks. Similar to our approach, there have been some recent works (Mohamed &", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 220, + 522, + 236 + ], + "spans": [ + { + "bbox": [ + 88, + 220, + 522, + 236 + ], + "score": 1.0, + "content": "Rezende (2015); Chalk et al. (2016); Alemi et al. (2018b; 2016; 2018a)) to approximate the MI by applying", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 232, + 388, + 246 + ], + "spans": [ + { + "bbox": [ + 89, + 232, + 388, + 246 + ], + "score": 1.0, + "content": "variational bounds on MI, though not in the context of domain adaptation.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 91, + 249, + 521, + 327 + ], + "lines": [ + { + "bbox": [ + 92, + 250, + 522, + 262 + ], + "spans": [ + { + "bbox": [ + 92, + 250, + 522, + 262 + ], + "score": 1.0, + "content": "Mohamed & Rezende (2015) utilized the variational bounds on MI, and apply it to DNNs in the context of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 261, + 522, + 273 + ], + "spans": [ + { + "bbox": [ + 90, + 261, + 522, + 273 + ], + "score": 1.0, + "content": "reinforcement learning. Chalk et al. (2016) and Alemi et al. (2016), developed the same variational lower", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 270, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 270, + 522, + 285 + ], + "score": 1.0, + "content": "bound In the context of Information Bottleneck (IB) principle (Tishby & Zaslavsky (2015)), where the former", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 283, + 522, + 296 + ], + "spans": [ + { + "bbox": [ + 90, + 283, + 522, + 296 + ], + "score": 1.0, + "content": "applied it to sparse coding problems, and used the kernel trick to achieve nonlinear mappings, whereas the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 293, + 522, + 306 + ], + "spans": [ + { + "bbox": [ + 88, + 293, + 522, + 306 + ], + "score": 1.0, + "content": "latter applied it to DNNs to handle large datasets thanks to the SGD algorithm. Achille & Soatto (2018)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 304, + 522, + 318 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 522, + 318 + ], + "score": 1.0, + "content": "proposed a variational bound on the MI in the context of IB, from the perspective of variational dropout and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 315, + 462, + 328 + ], + "spans": [ + { + "bbox": [ + 90, + 315, + 462, + 328 + ], + "score": 1.0, + "content": "demonstrated its utility in learning disentangled representations for variational autoencoders.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 91, + 332, + 520, + 387 + ], + "lines": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "spans": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "score": 1.0, + "content": "The main difference between our method and the above methods is that these methods throw away the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 90, + 344, + 522, + 356 + ], + "spans": [ + { + "bbox": [ + 90, + 344, + 522, + 356 + ], + "score": 1.0, + "content": "information in the data not related to the task by minimizing the mutual information between the data points", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 354, + 522, + 366 + ], + "spans": [ + { + "bbox": [ + 90, + 354, + 522, + 366 + ], + "score": 1.0, + "content": "and the latent representations that may lead to ignoring the individual characteristics (private features) of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 90, + 366, + 522, + 378 + ], + "spans": [ + { + "bbox": [ + 90, + 366, + 522, + 378 + ], + "score": 1.0, + "content": "datasets in a multiple dataset regime, whereas our method explicitly models what is unique to each domain", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 377, + 408, + 389 + ], + "spans": [ + { + "bbox": [ + 90, + 377, + 408, + 389 + ], + "score": 1.0, + "content": "(dataset) that improves the model’s ability to extract domain–invariant features.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 91, + 393, + 521, + 459 + ], + "lines": [ + { + "bbox": [ + 89, + 393, + 522, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 522, + 405 + ], + "score": 1.0, + "content": "In the unsupervised representation learning literature, our work is also related to the VAE-based models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 404, + 521, + 417 + ], + "spans": [ + { + "bbox": [ + 89, + 404, + 521, + 417 + ], + "score": 1.0, + "content": "(Bowman et al. (2015)). However, we propose to tackle the task using our IT approach using deterministic", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 88, + 415, + 523, + 428 + ], + "spans": [ + { + "bbox": [ + 88, + 415, + 523, + 428 + ], + "score": 1.0, + "content": "mappings instead of the traditional evidence lower bound (ELBO) optimization with stochastic mappings.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 426, + 523, + 439 + ], + "spans": [ + { + "bbox": [ + 89, + 426, + 523, + 439 + ], + "score": 1.0, + "content": "In contrast to the unsupervised representation learning approaches, our setting also allows us to further", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 90, + 438, + 522, + 449 + ], + "spans": [ + { + "bbox": [ + 90, + 438, + 522, + 449 + ], + "score": 1.0, + "content": "improve the latent representation using the labeled data in the source domain while leveraging the sharing of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 449, + 273, + 460 + ], + "spans": [ + { + "bbox": [ + 90, + 449, + 273, + 460 + ], + "score": 1.0, + "content": "dependencies across different target domains.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 91, + 476, + 423, + 489 + ], + "lines": [ + { + "bbox": [ + 88, + 475, + 424, + 492 + ], + "spans": [ + { + "bbox": [ + 88, + 475, + 424, + 492 + ], + "score": 1.0, + "content": "B CONNECTION TO MULTIPLE DOMAIN TRANSFER NETWORKS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 91, + 502, + 521, + 602 + ], + "lines": [ + { + "bbox": [ + 90, + 503, + 522, + 515 + ], + "spans": [ + { + "bbox": [ + 90, + 503, + 522, + 515 + ], + "score": 1.0, + "content": "Recent studies have shown remarkable success in multiple domain transfer (MDT) (Choi et al. (2017);", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 513, + 522, + 527 + ], + "spans": [ + { + "bbox": [ + 89, + 513, + 522, + 527 + ], + "score": 1.0, + "content": "Anoosheh et al. (2017); Kameoka et al. (2018); Hao et al. (2018)) though not in the context of the image", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 90, + 525, + 522, + 537 + ], + "spans": [ + { + "bbox": [ + 90, + 525, + 522, + 537 + ], + "score": 1.0, + "content": "classification, rather in the context of image generation. Choi et al. (2017) proposed StarGAN, a generative", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 534, + 522, + 550 + ], + "spans": [ + { + "bbox": [ + 88, + 534, + 522, + 550 + ], + "score": 1.0, + "content": "adversarial network capable of learning mappings among multiple domains in the contest of image to image", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 546, + 522, + 560 + ], + "spans": [ + { + "bbox": [ + 89, + 546, + 387, + 560 + ], + "score": 1.0, + "content": "translation framework. The goal of StarGAN is to train a single generator", + "type": "text" + }, + { + "bbox": [ + 388, + 547, + 397, + 557 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 546, + 522, + 560 + ], + "score": 1.0, + "content": "though this requires passing in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 88, + 556, + 522, + 572 + ], + "spans": [ + { + "bbox": [ + 88, + 556, + 522, + 572 + ], + "score": 1.0, + "content": "a vector along with each input to the generator specifying the output domain desired, that learns mappings", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 568, + 520, + 581 + ], + "spans": [ + { + "bbox": [ + 89, + 568, + 259, + 581 + ], + "score": 1.0, + "content": "among multiple domains. To achieve this,", + "type": "text" + }, + { + "bbox": [ + 259, + 569, + 268, + 578 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 568, + 416, + 581 + ], + "score": 1.0, + "content": "is trained to translate an input image", + "type": "text" + }, + { + "bbox": [ + 416, + 570, + 424, + 578 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 568, + 509, + 581 + ], + "score": 1.0, + "content": "into an output image", + "type": "text" + }, + { + "bbox": [ + 509, + 569, + 520, + 578 + ], + "score": 0.85, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 578, + 522, + 594 + ], + "spans": [ + { + "bbox": [ + 89, + 578, + 249, + 594 + ], + "score": 1.0, + "content": "conditioned on the target domain label", + "type": "text" + }, + { + "bbox": [ + 250, + 580, + 257, + 589 + ], + "score": 0.48, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 578, + 261, + 594 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 261, + 579, + 320, + 591 + ], + "score": 0.91, + "content": "G ( \\pmb { x } , \\pmb { d } ) \\pmb { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 578, + 488, + 594 + ], + "score": 1.0, + "content": ". Similar to our domain classifier module", + "type": "text" + }, + { + "bbox": [ + 488, + 580, + 497, + 589 + ], + "score": 0.79, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 498, + 578, + 522, + 594 + ], + "score": 1.0, + "content": ", they", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 591, + 467, + 603 + ], + "spans": [ + { + "bbox": [ + 90, + 591, + 467, + 603 + ], + "score": 1.0, + "content": "introduce an auxiliary classifier that allows a single discriminator to control multiple domains.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 92, + 607, + 520, + 651 + ], + "lines": [ + { + "bbox": [ + 92, + 607, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 92, + 607, + 522, + 620 + ], + "score": 1.0, + "content": "Anoosheh et al. (2017) introduced ComboGAN, which decouples the domains and networks from each", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 90, + 617, + 522, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 617, + 522, + 630 + ], + "score": 1.0, + "content": "other. Similar to our encoder/decoder modules, ComboGAN’s generator networks contain encoder/decoders", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 629, + 522, + 641 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 522, + 641 + ], + "score": 1.0, + "content": "assigning each encoder and decoder to a domain. They combine the encoders and decoders of the trained", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 639, + 522, + 654 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 522, + 654 + ], + "score": 1.0, + "content": "model like building blocks, taking as input any domain and outputting any other. For example during", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 283, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 283, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 95, + 522, + 129 + ], + "lines": [ + { + "bbox": [ + 90, + 95, + 523, + 109 + ], + "spans": [ + { + "bbox": [ + 90, + 95, + 523, + 109 + ], + "score": 1.0, + "content": "Han Zhao, Shanghang Zhang, Guanhang Wu, João P. Costeira, Jose Moura, and Geoffrey J. Gordon.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 105, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 98, + 105, + 522, + 121 + ], + "score": 1.0, + "content": "Multiple source domain adaptation with adversarial training of neural networks. International Conference", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 117, + 330, + 131 + ], + "spans": [ + { + "bbox": [ + 99, + 117, + 330, + 131 + ], + "score": 1.0, + "content": "on Learning Representation (ICLR) workshops, 05 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 90, + 95, + 523, + 131 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 152, + 491, + 165 + ], + "lines": [ + { + "bbox": [ + 90, + 151, + 491, + 165 + ], + "spans": [ + { + "bbox": [ + 90, + 151, + 491, + 165 + ], + "score": 1.0, + "content": "A CONNECTION TO INFORMATION THEORETIC REPRESENTATION LEARNING", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 91, + 177, + 521, + 244 + ], + "lines": [ + { + "bbox": [ + 88, + 176, + 522, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 176, + 522, + 191 + ], + "score": 1.0, + "content": "The idea of using information theoretic (IT) objectives for representation learning was originally introduced", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 188, + 522, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 522, + 201 + ], + "score": 1.0, + "content": "in (Tishby & Zaslavsky (2015)). Since their approach for optimizing the IT objective functions relied on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 200, + 521, + 212 + ], + "spans": [ + { + "bbox": [ + 90, + 200, + 521, + 212 + ], + "score": 1.0, + "content": "iterative Blahut Arimoto algorithm (Tishby & Zaslavsky (2015)), it is not feasible to apply to deep neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 211, + 522, + 224 + ], + "spans": [ + { + "bbox": [ + 89, + 211, + 522, + 224 + ], + "score": 1.0, + "content": "network (DNN) frameworks. Similar to our approach, there have been some recent works (Mohamed &", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 220, + 522, + 236 + ], + "spans": [ + { + "bbox": [ + 88, + 220, + 522, + 236 + ], + "score": 1.0, + "content": "Rezende (2015); Chalk et al. (2016); Alemi et al. (2018b; 2016; 2018a)) to approximate the MI by applying", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 232, + 388, + 246 + ], + "spans": [ + { + "bbox": [ + 89, + 232, + 388, + 246 + ], + "score": 1.0, + "content": "variational bounds on MI, though not in the context of domain adaptation.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 88, + 176, + 522, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 249, + 521, + 327 + ], + "lines": [ + { + "bbox": [ + 92, + 250, + 522, + 262 + ], + "spans": [ + { + "bbox": [ + 92, + 250, + 522, + 262 + ], + "score": 1.0, + "content": "Mohamed & Rezende (2015) utilized the variational bounds on MI, and apply it to DNNs in the context of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 261, + 522, + 273 + ], + "spans": [ + { + "bbox": [ + 90, + 261, + 522, + 273 + ], + "score": 1.0, + "content": "reinforcement learning. Chalk et al. (2016) and Alemi et al. (2016), developed the same variational lower", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 270, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 270, + 522, + 285 + ], + "score": 1.0, + "content": "bound In the context of Information Bottleneck (IB) principle (Tishby & Zaslavsky (2015)), where the former", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 283, + 522, + 296 + ], + "spans": [ + { + "bbox": [ + 90, + 283, + 522, + 296 + ], + "score": 1.0, + "content": "applied it to sparse coding problems, and used the kernel trick to achieve nonlinear mappings, whereas the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 293, + 522, + 306 + ], + "spans": [ + { + "bbox": [ + 88, + 293, + 522, + 306 + ], + "score": 1.0, + "content": "latter applied it to DNNs to handle large datasets thanks to the SGD algorithm. Achille & Soatto (2018)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 304, + 522, + 318 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 522, + 318 + ], + "score": 1.0, + "content": "proposed a variational bound on the MI in the context of IB, from the perspective of variational dropout and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 315, + 462, + 328 + ], + "spans": [ + { + "bbox": [ + 90, + 315, + 462, + 328 + ], + "score": 1.0, + "content": "demonstrated its utility in learning disentangled representations for variational autoencoders.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 88, + 250, + 522, + 328 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 332, + 520, + 387 + ], + "lines": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "spans": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "score": 1.0, + "content": "The main difference between our method and the above methods is that these methods throw away the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 90, + 344, + 522, + 356 + ], + "spans": [ + { + "bbox": [ + 90, + 344, + 522, + 356 + ], + "score": 1.0, + "content": "information in the data not related to the task by minimizing the mutual information between the data points", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 354, + 522, + 366 + ], + "spans": [ + { + "bbox": [ + 90, + 354, + 522, + 366 + ], + "score": 1.0, + "content": "and the latent representations that may lead to ignoring the individual characteristics (private features) of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 90, + 366, + 522, + 378 + ], + "spans": [ + { + "bbox": [ + 90, + 366, + 522, + 378 + ], + "score": 1.0, + "content": "datasets in a multiple dataset regime, whereas our method explicitly models what is unique to each domain", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 377, + 408, + 389 + ], + "spans": [ + { + "bbox": [ + 90, + 377, + 408, + 389 + ], + "score": 1.0, + "content": "(dataset) that improves the model’s ability to extract domain–invariant features.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 89, + 332, + 522, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 393, + 521, + 459 + ], + "lines": [ + { + "bbox": [ + 89, + 393, + 522, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 522, + 405 + ], + "score": 1.0, + "content": "In the unsupervised representation learning literature, our work is also related to the VAE-based models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 404, + 521, + 417 + ], + "spans": [ + { + "bbox": [ + 89, + 404, + 521, + 417 + ], + "score": 1.0, + "content": "(Bowman et al. (2015)). However, we propose to tackle the task using our IT approach using deterministic", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 88, + 415, + 523, + 428 + ], + "spans": [ + { + "bbox": [ + 88, + 415, + 523, + 428 + ], + "score": 1.0, + "content": "mappings instead of the traditional evidence lower bound (ELBO) optimization with stochastic mappings.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 426, + 523, + 439 + ], + "spans": [ + { + "bbox": [ + 89, + 426, + 523, + 439 + ], + "score": 1.0, + "content": "In contrast to the unsupervised representation learning approaches, our setting also allows us to further", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 90, + 438, + 522, + 449 + ], + "spans": [ + { + "bbox": [ + 90, + 438, + 522, + 449 + ], + "score": 1.0, + "content": "improve the latent representation using the labeled data in the source domain while leveraging the sharing of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 449, + 273, + 460 + ], + "spans": [ + { + "bbox": [ + 90, + 449, + 273, + 460 + ], + "score": 1.0, + "content": "dependencies across different target domains.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 88, + 393, + 523, + 460 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 476, + 423, + 489 + ], + "lines": [ + { + "bbox": [ + 88, + 475, + 424, + 492 + ], + "spans": [ + { + "bbox": [ + 88, + 475, + 424, + 492 + ], + "score": 1.0, + "content": "B CONNECTION TO MULTIPLE DOMAIN TRANSFER NETWORKS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 91, + 502, + 521, + 602 + ], + "lines": [ + { + "bbox": [ + 90, + 503, + 522, + 515 + ], + "spans": [ + { + "bbox": [ + 90, + 503, + 522, + 515 + ], + "score": 1.0, + "content": "Recent studies have shown remarkable success in multiple domain transfer (MDT) (Choi et al. (2017);", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 89, + 513, + 522, + 527 + ], + "spans": [ + { + "bbox": [ + 89, + 513, + 522, + 527 + ], + "score": 1.0, + "content": "Anoosheh et al. (2017); Kameoka et al. (2018); Hao et al. (2018)) though not in the context of the image", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 90, + 525, + 522, + 537 + ], + "spans": [ + { + "bbox": [ + 90, + 525, + 522, + 537 + ], + "score": 1.0, + "content": "classification, rather in the context of image generation. Choi et al. (2017) proposed StarGAN, a generative", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 534, + 522, + 550 + ], + "spans": [ + { + "bbox": [ + 88, + 534, + 522, + 550 + ], + "score": 1.0, + "content": "adversarial network capable of learning mappings among multiple domains in the contest of image to image", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 546, + 522, + 560 + ], + "spans": [ + { + "bbox": [ + 89, + 546, + 387, + 560 + ], + "score": 1.0, + "content": "translation framework. The goal of StarGAN is to train a single generator", + "type": "text" + }, + { + "bbox": [ + 388, + 547, + 397, + 557 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 546, + 522, + 560 + ], + "score": 1.0, + "content": "though this requires passing in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 88, + 556, + 522, + 572 + ], + "spans": [ + { + "bbox": [ + 88, + 556, + 522, + 572 + ], + "score": 1.0, + "content": "a vector along with each input to the generator specifying the output domain desired, that learns mappings", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 568, + 520, + 581 + ], + "spans": [ + { + "bbox": [ + 89, + 568, + 259, + 581 + ], + "score": 1.0, + "content": "among multiple domains. To achieve this,", + "type": "text" + }, + { + "bbox": [ + 259, + 569, + 268, + 578 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 568, + 416, + 581 + ], + "score": 1.0, + "content": "is trained to translate an input image", + "type": "text" + }, + { + "bbox": [ + 416, + 570, + 424, + 578 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 568, + 509, + 581 + ], + "score": 1.0, + "content": "into an output image", + "type": "text" + }, + { + "bbox": [ + 509, + 569, + 520, + 578 + ], + "score": 0.85, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 578, + 522, + 594 + ], + "spans": [ + { + "bbox": [ + 89, + 578, + 249, + 594 + ], + "score": 1.0, + "content": "conditioned on the target domain label", + "type": "text" + }, + { + "bbox": [ + 250, + 580, + 257, + 589 + ], + "score": 0.48, + "content": "^ d", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 578, + 261, + 594 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 261, + 579, + 320, + 591 + ], + "score": 0.91, + "content": "G ( \\pmb { x } , \\pmb { d } ) \\pmb { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 578, + 488, + 594 + ], + "score": 1.0, + "content": ". Similar to our domain classifier module", + "type": "text" + }, + { + "bbox": [ + 488, + 580, + 497, + 589 + ], + "score": 0.79, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 498, + 578, + 522, + 594 + ], + "score": 1.0, + "content": ", they", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 591, + 467, + 603 + ], + "spans": [ + { + "bbox": [ + 90, + 591, + 467, + 603 + ], + "score": 1.0, + "content": "introduce an auxiliary classifier that allows a single discriminator to control multiple domains.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33, + "bbox_fs": [ + 88, + 503, + 522, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 92, + 607, + 520, + 651 + ], + "lines": [ + { + "bbox": [ + 92, + 607, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 92, + 607, + 522, + 620 + ], + "score": 1.0, + "content": "Anoosheh et al. (2017) introduced ComboGAN, which decouples the domains and networks from each", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 90, + 617, + 522, + 630 + ], + "spans": [ + { + "bbox": [ + 90, + 617, + 522, + 630 + ], + "score": 1.0, + "content": "other. Similar to our encoder/decoder modules, ComboGAN’s generator networks contain encoder/decoders", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 629, + 522, + 641 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 522, + 641 + ], + "score": 1.0, + "content": "assigning each encoder and decoder to a domain. They combine the encoders and decoders of the trained", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 89, + 639, + 522, + 654 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 522, + 654 + ], + "score": 1.0, + "content": "model like building blocks, taking as input any domain and outputting any other. For example during", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 88, + 93, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 88, + 93, + 222, + 109 + ], + "score": 1.0, + "content": "inference, to transform an image", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 222, + 98, + 230, + 106 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 230, + 93, + 334, + 109 + ], + "score": 1.0, + "content": "from an arbitrary domain", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 334, + 96, + 344, + 106 + ], + "score": 0.52, + "content": "\\mathbf { X }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 344, + 93, + 355, + 109 + ], + "score": 1.0, + "content": "to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 355, + 96, + 366, + 106 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 366, + 93, + 421, + 109 + ], + "score": 1.0, + "content": "from domain", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 421, + 96, + 434, + 106 + ], + "score": 0.85, + "content": "\\mathbf { X } ^ { \\prime }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 434, + 93, + 522, + 109 + ], + "score": 1.0, + "content": ", they simply perform", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 106, + 522, + 120 + ], + "spans": [ + { + "bbox": [ + 90, + 106, + 288, + 119 + ], + "score": 0.88, + "content": "\\pmb { x } ^ { \\prime } = G _ { \\mathbf { X } ^ { \\prime } , \\mathbf { X } } ( \\pmb { x } ) = D e c o d e r _ { X ^ { \\prime } } ( E n c o d e r _ { X } ( \\pmb { x } ) )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 288, + 106, + 351, + 120 + ], + "score": 1.0, + "content": ". The result of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 351, + 107, + 411, + 119 + ], + "score": 0.65, + "content": "E n c o d e r _ { X } ( { \\pmb x } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 411, + 106, + 522, + 120 + ], + "score": 1.0, + "content": "can even be cached when", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 117, + 322, + 131 + ], + "spans": [ + { + "bbox": [ + 89, + 117, + 322, + 131 + ], + "score": 1.0, + "content": "translating to other domains as not to repeat computation.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 39.5, + "bbox_fs": [ + 89, + 607, + 522, + 654 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 95, + 520, + 129 + ], + "lines": [ + { + "bbox": [ + 88, + 93, + 522, + 109 + ], + "spans": [ + { + "bbox": [ + 88, + 93, + 222, + 109 + ], + "score": 1.0, + "content": "inference, to transform an image", + "type": "text" + }, + { + "bbox": [ + 222, + 98, + 230, + 106 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 93, + 334, + 109 + ], + "score": 1.0, + "content": "from an arbitrary domain", + "type": "text" + }, + { + "bbox": [ + 334, + 96, + 344, + 106 + ], + "score": 0.52, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 93, + 355, + 109 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 355, + 96, + 366, + 106 + ], + "score": 0.87, + "content": "\\mathbf { x } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 93, + 421, + 109 + ], + "score": 1.0, + "content": "from domain", + "type": "text" + }, + { + "bbox": [ + 421, + 96, + 434, + 106 + ], + "score": 0.85, + "content": "\\mathbf { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 93, + 522, + 109 + ], + "score": 1.0, + "content": ", they simply perform", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 106, + 522, + 120 + ], + "spans": [ + { + "bbox": [ + 90, + 106, + 288, + 119 + ], + "score": 0.88, + "content": "\\pmb { x } ^ { \\prime } = G _ { \\mathbf { X } ^ { \\prime } , \\mathbf { X } } ( \\pmb { x } ) = D e c o d e r _ { X ^ { \\prime } } ( E n c o d e r _ { X } ( \\pmb { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 106, + 351, + 120 + ], + "score": 1.0, + "content": ". The result of", + "type": "text" + }, + { + "bbox": [ + 351, + 107, + 411, + 119 + ], + "score": 0.65, + "content": "E n c o d e r _ { X } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 106, + 522, + 120 + ], + "score": 1.0, + "content": "can even be cached when", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 117, + 322, + 131 + ], + "spans": [ + { + "bbox": [ + 89, + 117, + 322, + 131 + ], + "score": 1.0, + "content": "translating to other domains as not to repeat computation.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 91, + 135, + 521, + 200 + ], + "lines": [ + { + "bbox": [ + 90, + 135, + 521, + 146 + ], + "spans": [ + { + "bbox": [ + 90, + 135, + 521, + 146 + ], + "score": 1.0, + "content": "The main differences between the MDT methods and ours is that, unlike our method which does domain", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 146, + 521, + 157 + ], + "spans": [ + { + "bbox": [ + 90, + 146, + 521, + 157 + ], + "score": 1.0, + "content": "alignment in feature space, MDT methods adapt representations not in feature space but rather in raw pixel", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 156, + 522, + 169 + ], + "spans": [ + { + "bbox": [ + 88, + 156, + 522, + 169 + ], + "score": 1.0, + "content": "space; translating samples from one domain to the “style” of a other domains. This works well for limited", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 166, + 522, + 180 + ], + "spans": [ + { + "bbox": [ + 89, + 166, + 522, + 180 + ], + "score": 1.0, + "content": "domain shifts where the domains are similar in pixel-space, but can be too limiting for settings with larger", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 177, + 523, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 177, + 523, + 191 + ], + "score": 1.0, + "content": "domain shifts that results in poor performance in significant structural change of the samples in different", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 190, + 129, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 190, + 129, + 201 + ], + "score": 1.0, + "content": "domains.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 91, + 218, + 376, + 231 + ], + "lines": [ + { + "bbox": [ + 89, + 217, + 378, + 233 + ], + "spans": [ + { + "bbox": [ + 89, + 217, + 378, + 233 + ], + "score": 1.0, + "content": "C CONNECTION TO DOMAIN SEPARATION NETWORKS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 91, + 244, + 521, + 299 + ], + "lines": [ + { + "bbox": [ + 90, + 245, + 521, + 257 + ], + "spans": [ + { + "bbox": [ + 90, + 245, + 521, + 257 + ], + "score": 1.0, + "content": "The method closest to our work is Domain Separation Networks (DSN) (Bousmalis et al. (2016)), which use", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 255, + 522, + 267 + ], + "spans": [ + { + "bbox": [ + 90, + 255, + 522, + 267 + ], + "score": 1.0, + "content": "the notion of auto-encoders to explicitly separate the feature representations private to each source/target", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 265, + 522, + 280 + ], + "spans": [ + { + "bbox": [ + 88, + 265, + 522, + 280 + ], + "score": 1.0, + "content": "domain from those that are shared between the domains. Although extending DSN to multiple domains might", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 277, + 522, + 290 + ], + "spans": [ + { + "bbox": [ + 88, + 277, + 522, + 290 + ], + "score": 1.0, + "content": "seem trivial, DSN requires an autoencoder per domain, making the model impractical in the case of more", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 289, + 196, + 300 + ], + "spans": [ + { + "bbox": [ + 90, + 289, + 196, + 300 + ], + "score": 1.0, + "content": "than a couple of domains.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 91, + 305, + 521, + 459 + ], + "lines": [ + { + "bbox": [ + 89, + 304, + 523, + 318 + ], + "spans": [ + { + "bbox": [ + 89, + 304, + 523, + 318 + ], + "score": 1.0, + "content": "The overall loss of DSN consists of a reconstruction loss for each domain modeled by a shared decoder, a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 316, + 522, + 328 + ], + "spans": [ + { + "bbox": [ + 89, + 316, + 522, + 328 + ], + "score": 1.0, + "content": "similarity loss such as MMD, which encourages domain invariance modeled by a shared encoder, and a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 327, + 523, + 339 + ], + "spans": [ + { + "bbox": [ + 90, + 327, + 523, + 339 + ], + "score": 1.0, + "content": "dissimilarity loss modeled by two private encoders: one for the source domain and one for the target domain.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 338, + 522, + 352 + ], + "spans": [ + { + "bbox": [ + 89, + 338, + 522, + 352 + ], + "score": 1.0, + "content": "While one could attempt to generalize DSN to multiple target domains by having individual per-target domain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "score": 1.0, + "content": "private encoders, doing so would prove problematic when the number of target domains is large — each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 360, + 522, + 373 + ], + "spans": [ + { + "bbox": [ + 88, + 360, + 522, + 373 + ], + "score": 1.0, + "content": "private encoder would require a large \"private\" dataset to learn the private parameters. Precisely, for multiple", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 370, + 522, + 384 + ], + "spans": [ + { + "bbox": [ + 90, + 371, + 109, + 383 + ], + "score": 0.71, + "content": "\\bar { ( \\cal M ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 109, + 370, + 392, + 384 + ], + "score": 1.0, + "content": "target domains, we could train a DSN model with one shared encoder,", + "type": "text" + }, + { + "bbox": [ + 393, + 371, + 422, + 381 + ], + "score": 0.9, + "content": "M + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 370, + 522, + 384 + ], + "score": 1.0, + "content": "private encoder (one for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 382, + 522, + 394 + ], + "spans": [ + { + "bbox": [ + 89, + 382, + 296, + 394 + ], + "score": 1.0, + "content": "each domain), and one shared decoder. This leads to", + "type": "text" + }, + { + "bbox": [ + 296, + 382, + 324, + 393 + ], + "score": 0.89, + "content": "M + 3", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 382, + 522, + 394 + ], + "score": 1.0, + "content": "models to train that implies the number of models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 393, + 521, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 521, + 405 + ], + "score": 1.0, + "content": "increases linearly with the number of domains, as does the required training time. Second, DSN uses an", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 403, + 522, + 417 + ], + "spans": [ + { + "bbox": [ + 89, + 403, + 522, + 417 + ], + "score": 1.0, + "content": "orthogonality constraint among the shared and the private representations which may not be strong enough to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "spans": [ + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "score": 1.0, + "content": "remove redundancy and enforce disentangling among different private spaces. Precisely, DSN defines the loss", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 425, + 522, + 438 + ], + "spans": [ + { + "bbox": [ + 89, + 425, + 522, + 438 + ], + "score": 1.0, + "content": "via a soft subspace orthogonality constraint between the private and shared representation of each domain.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 437, + 521, + 450 + ], + "spans": [ + { + "bbox": [ + 90, + 437, + 521, + 450 + ], + "score": 1.0, + "content": "However, it does not enforce the private representation of different domains to be different that may result in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 447, + 247, + 461 + ], + "spans": [ + { + "bbox": [ + 88, + 447, + 247, + 461 + ], + "score": 1.0, + "content": "redundancy of different private spaces.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 91, + 464, + 522, + 564 + ], + "lines": [ + { + "bbox": [ + 88, + 463, + 522, + 478 + ], + "spans": [ + { + "bbox": [ + 88, + 463, + 473, + 478 + ], + "score": 1.0, + "content": "In addition, DSN enforces separation of spaces using the notion of Euclidean orthogonality, e.g.,", + "type": "text" + }, + { + "bbox": [ + 474, + 464, + 519, + 477 + ], + "score": 0.92, + "content": "\\| z _ { s } - z _ { p } \\| ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 519, + 463, + 522, + 478 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 88, + 475, + 523, + 489 + ], + "spans": [ + { + "bbox": [ + 88, + 475, + 523, + 489 + ], + "score": 1.0, + "content": "In case of multiple target domains, this would result in learning of all pairs of private spaces independently.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 486, + 522, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 486, + 522, + 499 + ], + "score": 1.0, + "content": "To address those deficiencies, we first explicitly couple different private encoders into a single private encoder", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 90, + 498, + 522, + 510 + ], + "spans": [ + { + "bbox": [ + 90, + 498, + 119, + 510 + ], + "score": 1.0, + "content": "model,", + "type": "text" + }, + { + "bbox": [ + 119, + 498, + 136, + 510 + ], + "score": 0.9, + "content": "E _ { \\theta _ { p } }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 498, + 522, + 510 + ], + "score": 1.0, + "content": "of fig. 2 , which allows us to generalize to an arbitrary number of target domains. To assure that the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 508, + 523, + 521 + ], + "spans": [ + { + "bbox": [ + 89, + 508, + 523, + 521 + ], + "score": 1.0, + "content": "information among the private and shared spaces is not shared (i.e., \"orthogonal\"), we define an information-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 519, + 522, + 532 + ], + "spans": [ + { + "bbox": [ + 89, + 519, + 293, + 532 + ], + "score": 1.0, + "content": "theoretic criteria enforced by a domain classifier,", + "type": "text" + }, + { + "bbox": [ + 294, + 519, + 309, + 532 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 519, + 522, + 532 + ], + "score": 1.0, + "content": "of fig. 2, which aims to segment the private space", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 90, + 531, + 522, + 542 + ], + "spans": [ + { + "bbox": [ + 90, + 531, + 366, + 542 + ], + "score": 1.0, + "content": "into clusters that correspond to individual target domains. By using", + "type": "text" + }, + { + "bbox": [ + 366, + 531, + 381, + 542 + ], + "score": 0.9, + "content": "D _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 531, + 522, + 542 + ], + "score": 1.0, + "content": "within the adversarial framework,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "score": 1.0, + "content": "MTDA-ITA learns simultaneously the shared and private features from different domains (see fig. 6). We", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 552, + 523, + 564 + ], + "spans": [ + { + "bbox": [ + 90, + 552, + 523, + 564 + ], + "score": 1.0, + "content": "showed in Sec. 4 that our model performs better than the trivial extension of DSNs to the multi-domain case.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 92, + 581, + 250, + 594 + ], + "lines": [ + { + "bbox": [ + 89, + 580, + 252, + 596 + ], + "spans": [ + { + "bbox": [ + 89, + 580, + 252, + 596 + ], + "score": 1.0, + "content": "D NETWORK ARCHITECTURE", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 606, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 620 + ], + "score": 1.0, + "content": "The network architecture used for the experiments is given in tab. 3. We use the following abbreviation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 618, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 190, + 631 + ], + "score": 1.0, + "content": "for ease of presentation:", + "type": "text" + }, + { + "bbox": [ + 190, + 618, + 205, + 629 + ], + "score": 0.48, + "content": "{ \\bf N } =", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 618, + 242, + 631 + ], + "score": 1.0, + "content": "Neurons,", + "type": "text" + }, + { + "bbox": [ + 243, + 618, + 257, + 628 + ], + "score": 0.61, + "content": "\\mathbf { K } { = }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 618, + 306, + 631 + ], + "score": 1.0, + "content": "Kernel size,", + "type": "text" + }, + { + "bbox": [ + 306, + 619, + 319, + 628 + ], + "score": 0.38, + "content": "{ \\mathrm { S } } { = } { }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 618, + 363, + 631 + ], + "score": 1.0, + "content": "Stride size,", + "type": "text" + }, + { + "bbox": [ + 364, + 619, + 379, + 628 + ], + "score": 0.5, + "content": "\\begin{array} { r } { \\boldsymbol { \\mathrm { D } } \\boldsymbol { = } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 618, + 464, + 631 + ], + "score": 1.0, + "content": "Number of Domains,", + "type": "text" + }, + { + "bbox": [ + 465, + 618, + 478, + 628 + ], + "score": 0.73, + "content": "\\mathbf { C } =", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 618, + 522, + 631 + ], + "score": 1.0, + "content": "number of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "score": 1.0, + "content": "Classes. The transposed convolutional layer is denoted by DCONV. The residual basic block is denoted as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 89, + 640, + 133, + 652 + ], + "spans": [ + { + "bbox": [ + 89, + 640, + 133, + 652 + ], + "score": 1.0, + "content": "RESBLK.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 91, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 91, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 283, + 670, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 283, + 670, + 295, + 684 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 91, + 95, + 520, + 129 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 88, + 93, + 522, + 131 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 91, + 135, + 521, + 200 + ], + "lines": [ + { + "bbox": [ + 90, + 135, + 521, + 146 + ], + "spans": [ + { + "bbox": [ + 90, + 135, + 521, + 146 + ], + "score": 1.0, + "content": "The main differences between the MDT methods and ours is that, unlike our method which does domain", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 146, + 521, + 157 + ], + "spans": [ + { + "bbox": [ + 90, + 146, + 521, + 157 + ], + "score": 1.0, + "content": "alignment in feature space, MDT methods adapt representations not in feature space but rather in raw pixel", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 156, + 522, + 169 + ], + "spans": [ + { + "bbox": [ + 88, + 156, + 522, + 169 + ], + "score": 1.0, + "content": "space; translating samples from one domain to the “style” of a other domains. This works well for limited", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 166, + 522, + 180 + ], + "spans": [ + { + "bbox": [ + 89, + 166, + 522, + 180 + ], + "score": 1.0, + "content": "domain shifts where the domains are similar in pixel-space, but can be too limiting for settings with larger", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 177, + 523, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 177, + 523, + 191 + ], + "score": 1.0, + "content": "domain shifts that results in poor performance in significant structural change of the samples in different", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 190, + 129, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 190, + 129, + 201 + ], + "score": 1.0, + "content": "domains.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 88, + 135, + 523, + 201 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 218, + 376, + 231 + ], + "lines": [ + { + "bbox": [ + 89, + 217, + 378, + 233 + ], + "spans": [ + { + "bbox": [ + 89, + 217, + 378, + 233 + ], + "score": 1.0, + "content": "C CONNECTION TO DOMAIN SEPARATION NETWORKS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 91, + 244, + 521, + 299 + ], + "lines": [ + { + "bbox": [ + 90, + 245, + 521, + 257 + ], + "spans": [ + { + "bbox": [ + 90, + 245, + 521, + 257 + ], + "score": 1.0, + "content": "The method closest to our work is Domain Separation Networks (DSN) (Bousmalis et al. (2016)), which use", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 255, + 522, + 267 + ], + "spans": [ + { + "bbox": [ + 90, + 255, + 522, + 267 + ], + "score": 1.0, + "content": "the notion of auto-encoders to explicitly separate the feature representations private to each source/target", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 265, + 522, + 280 + ], + "spans": [ + { + "bbox": [ + 88, + 265, + 522, + 280 + ], + "score": 1.0, + "content": "domain from those that are shared between the domains. Although extending DSN to multiple domains might", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 277, + 522, + 290 + ], + "spans": [ + { + "bbox": [ + 88, + 277, + 522, + 290 + ], + "score": 1.0, + "content": "seem trivial, DSN requires an autoencoder per domain, making the model impractical in the case of more", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 90, + 289, + 196, + 300 + ], + "spans": [ + { + "bbox": [ + 90, + 289, + 196, + 300 + ], + "score": 1.0, + "content": "than a couple of domains.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 88, + 245, + 522, + 300 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 305, + 521, + 459 + ], + "lines": [ + { + "bbox": [ + 89, + 304, + 523, + 318 + ], + "spans": [ + { + "bbox": [ + 89, + 304, + 523, + 318 + ], + "score": 1.0, + "content": "The overall loss of DSN consists of a reconstruction loss for each domain modeled by a shared decoder, a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 316, + 522, + 328 + ], + "spans": [ + { + "bbox": [ + 89, + 316, + 522, + 328 + ], + "score": 1.0, + "content": "similarity loss such as MMD, which encourages domain invariance modeled by a shared encoder, and a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 327, + 523, + 339 + ], + "spans": [ + { + "bbox": [ + 90, + 327, + 523, + 339 + ], + "score": 1.0, + "content": "dissimilarity loss modeled by two private encoders: one for the source domain and one for the target domain.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 338, + 522, + 352 + ], + "spans": [ + { + "bbox": [ + 89, + 338, + 522, + 352 + ], + "score": 1.0, + "content": "While one could attempt to generalize DSN to multiple target domains by having individual per-target domain", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "spans": [ + { + "bbox": [ + 88, + 349, + 522, + 362 + ], + "score": 1.0, + "content": "private encoders, doing so would prove problematic when the number of target domains is large — each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 360, + 522, + 373 + ], + "spans": [ + { + "bbox": [ + 88, + 360, + 522, + 373 + ], + "score": 1.0, + "content": "private encoder would require a large \"private\" dataset to learn the private parameters. Precisely, for multiple", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 370, + 522, + 384 + ], + "spans": [ + { + "bbox": [ + 90, + 371, + 109, + 383 + ], + "score": 0.71, + "content": "\\bar { ( \\cal M ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 109, + 370, + 392, + 384 + ], + "score": 1.0, + "content": "target domains, we could train a DSN model with one shared encoder,", + "type": "text" + }, + { + "bbox": [ + 393, + 371, + 422, + 381 + ], + "score": 0.9, + "content": "M + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 370, + 522, + 384 + ], + "score": 1.0, + "content": "private encoder (one for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 382, + 522, + 394 + ], + "spans": [ + { + "bbox": [ + 89, + 382, + 296, + 394 + ], + "score": 1.0, + "content": "each domain), and one shared decoder. This leads to", + "type": "text" + }, + { + "bbox": [ + 296, + 382, + 324, + 393 + ], + "score": 0.89, + "content": "M + 3", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 382, + 522, + 394 + ], + "score": 1.0, + "content": "models to train that implies the number of models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 393, + 521, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 393, + 521, + 405 + ], + "score": 1.0, + "content": "increases linearly with the number of domains, as does the required training time. Second, DSN uses an", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 403, + 522, + 417 + ], + "spans": [ + { + "bbox": [ + 89, + 403, + 522, + 417 + ], + "score": 1.0, + "content": "orthogonality constraint among the shared and the private representations which may not be strong enough to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "spans": [ + { + "bbox": [ + 88, + 414, + 522, + 428 + ], + "score": 1.0, + "content": "remove redundancy and enforce disentangling among different private spaces. Precisely, DSN defines the loss", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 425, + 522, + 438 + ], + "spans": [ + { + "bbox": [ + 89, + 425, + 522, + 438 + ], + "score": 1.0, + "content": "via a soft subspace orthogonality constraint between the private and shared representation of each domain.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 90, + 437, + 521, + 450 + ], + "spans": [ + { + "bbox": [ + 90, + 437, + 521, + 450 + ], + "score": 1.0, + "content": "However, it does not enforce the private representation of different domains to be different that may result in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 447, + 247, + 461 + ], + "spans": [ + { + "bbox": [ + 88, + 447, + 247, + 461 + ], + "score": 1.0, + "content": "redundancy of different private spaces.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5, + "bbox_fs": [ + 88, + 304, + 523, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 464, + 522, + 564 + ], + "lines": [ + { + "bbox": [ + 88, + 463, + 522, + 478 + ], + "spans": [ + { + "bbox": [ + 88, + 463, + 473, + 478 + ], + "score": 1.0, + "content": "In addition, DSN enforces separation of spaces using the notion of Euclidean orthogonality, e.g.,", + "type": "text" + }, + { + "bbox": [ + 474, + 464, + 519, + 477 + ], + "score": 0.92, + "content": "\\| z _ { s } - z _ { p } \\| ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 519, + 463, + 522, + 478 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 88, + 475, + 523, + 489 + ], + "spans": [ + { + "bbox": [ + 88, + 475, + 523, + 489 + ], + "score": 1.0, + "content": "In case of multiple target domains, this would result in learning of all pairs of private spaces independently.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 486, + 522, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 486, + 522, + 499 + ], + "score": 1.0, + "content": "To address those deficiencies, we first explicitly couple different private encoders into a single private encoder", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 90, + 498, + 522, + 510 + ], + "spans": [ + { + "bbox": [ + 90, + 498, + 119, + 510 + ], + "score": 1.0, + "content": "model,", + "type": "text" + }, + { + "bbox": [ + 119, + 498, + 136, + 510 + ], + "score": 0.9, + "content": "E _ { \\theta _ { p } }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 498, + 522, + 510 + ], + "score": 1.0, + "content": "of fig. 2 , which allows us to generalize to an arbitrary number of target domains. To assure that the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 508, + 523, + 521 + ], + "spans": [ + { + "bbox": [ + 89, + 508, + 523, + 521 + ], + "score": 1.0, + "content": "information among the private and shared spaces is not shared (i.e., \"orthogonal\"), we define an information-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 519, + 522, + 532 + ], + "spans": [ + { + "bbox": [ + 89, + 519, + 293, + 532 + ], + "score": 1.0, + "content": "theoretic criteria enforced by a domain classifier,", + "type": "text" + }, + { + "bbox": [ + 294, + 519, + 309, + 532 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 519, + 522, + 532 + ], + "score": 1.0, + "content": "of fig. 2, which aims to segment the private space", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 90, + 531, + 522, + 542 + ], + "spans": [ + { + "bbox": [ + 90, + 531, + 366, + 542 + ], + "score": 1.0, + "content": "into clusters that correspond to individual target domains. By using", + "type": "text" + }, + { + "bbox": [ + 366, + 531, + 381, + 542 + ], + "score": 0.9, + "content": "D _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 531, + 522, + 542 + ], + "score": 1.0, + "content": "within the adversarial framework,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 522, + 554 + ], + "score": 1.0, + "content": "MTDA-ITA learns simultaneously the shared and private features from different domains (see fig. 6). We", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 552, + 523, + 564 + ], + "spans": [ + { + "bbox": [ + 90, + 552, + 523, + 564 + ], + "score": 1.0, + "content": "showed in Sec. 4 that our model performs better than the trivial extension of DSNs to the multi-domain case.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33, + "bbox_fs": [ + 88, + 463, + 523, + 564 + ] + }, + { + "type": "title", + "bbox": [ + 92, + 581, + 250, + 594 + ], + "lines": [ + { + "bbox": [ + 89, + 580, + 252, + 596 + ], + "spans": [ + { + "bbox": [ + 89, + 580, + 252, + 596 + ], + "score": 1.0, + "content": "D NETWORK ARCHITECTURE", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 606, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 620 + ], + "score": 1.0, + "content": "The network architecture used for the experiments is given in tab. 3. We use the following abbreviation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 618, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 190, + 631 + ], + "score": 1.0, + "content": "for ease of presentation:", + "type": "text" + }, + { + "bbox": [ + 190, + 618, + 205, + 629 + ], + "score": 0.48, + "content": "{ \\bf N } =", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 618, + 242, + 631 + ], + "score": 1.0, + "content": "Neurons,", + "type": "text" + }, + { + "bbox": [ + 243, + 618, + 257, + 628 + ], + "score": 0.61, + "content": "\\mathbf { K } { = }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 618, + 306, + 631 + ], + "score": 1.0, + "content": "Kernel size,", + "type": "text" + }, + { + "bbox": [ + 306, + 619, + 319, + 628 + ], + "score": 0.38, + "content": "{ \\mathrm { S } } { = } { }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 618, + 363, + 631 + ], + "score": 1.0, + "content": "Stride size,", + "type": "text" + }, + { + "bbox": [ + 364, + 619, + 379, + 628 + ], + "score": 0.5, + "content": "\\begin{array} { r } { \\boldsymbol { \\mathrm { D } } \\boldsymbol { = } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 618, + 464, + 631 + ], + "score": 1.0, + "content": "Number of Domains,", + "type": "text" + }, + { + "bbox": [ + 465, + 618, + 478, + 628 + ], + "score": 0.73, + "content": "\\mathbf { C } =", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 618, + 522, + 631 + ], + "score": 1.0, + "content": "number of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 90, + 629, + 521, + 642 + ], + "score": 1.0, + "content": "Classes. The transposed convolutional layer is denoted by DCONV. The residual basic block is denoted as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 89, + 640, + 133, + 652 + ], + "spans": [ + { + "bbox": [ + 89, + 640, + 133, + 652 + ], + "score": 1.0, + "content": "RESBLK.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 89, + 606, + 522, + 652 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 171, + 94, + 439, + 435 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 171, + 94, + 439, + 435 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 171, + 94, + 439, + 435 + ], + "spans": [ + { + "bbox": [ + 171, + 94, + 439, + 435 + ], + "score": 0.972, + "html": "
LayerEncoders (shared,private)
1CONV-(N16,K7,S1), ReLU
2CONV-(N32,K3,S2), ReLU
3CONV-(N64,K3,S2), ReLU
4RESBLK-(N64,K3,S1)
5RESBLK-(N64,K3,S1)
6RESBLK-(N64,K3,S1)
7RESBLK-(N64,K3,S1)
LayerDecoder
1RESBLK-(N64,K3,S1)
2RESBLK-(N64,K3,S1)
3RESBLK-(N64,K3,S1)
4RESBLK-(N64,K3,S1)
5DCONV-(N32,K3,S2), ReLU
6DCONV-(N16,K3,S2), ReLU
7DCONV-(N1,K1,S1), Tanh
LayerDiscriminator
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1),ReLU
3CONV-(N16,K3,S1),ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(ND), Softmax
LayerClassifier
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1), ReLU
3CONV-(N16,K3,S1), ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(NC), Softmax
", + "type": "table", + "image_path": "50d3352cf70c1ded69dc8863026d49852f9c33a698de56302a1a3fb842a20ce7.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 171, + 94, + 439, + 109.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 171, + 109.5, + 439, + 125.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 171, + 125.0, + 439, + 140.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 171, + 140.5, + 439, + 156.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 171, + 156.0, + 439, + 171.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 171, + 171.5, + 439, + 187.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 171, + 187.0, + 439, + 202.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 171, + 202.5, + 439, + 218.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 171, + 218.0, + 439, + 233.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 171, + 233.5, + 439, + 249.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 171, + 249.0, + 439, + 264.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 171, + 264.5, + 439, + 280.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 171, + 280.0, + 439, + 295.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 171, + 295.5, + 439, + 311.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 171, + 311.0, + 439, + 326.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 171, + 326.5, + 439, + 342.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 171, + 342.0, + 439, + 357.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 171, + 357.5, + 439, + 373.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 171, + 373.0, + 439, + 388.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 171, + 388.5, + 439, + 404.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 171, + 404.0, + 439, + 419.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 171, + 419.5, + 439, + 435.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 213, + 443, + 397, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 443, + 398, + 455 + ], + "spans": [ + { + "bbox": [ + 212, + 443, + 398, + 455 + ], + "score": 1.0, + "content": "Table 3: Network architecture for the experiments.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 16.25 + }, + { + "type": "title", + "bbox": [ + 91, + 460, + 284, + 473 + ], + "lines": [ + { + "bbox": [ + 88, + 458, + 285, + 475 + ], + "spans": [ + { + "bbox": [ + 88, + 458, + 285, + 475 + ], + "score": 1.0, + "content": "E PROPOSED MODEL’S ALGORITHM", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 91, + 486, + 419, + 498 + ], + "lines": [ + { + "bbox": [ + 89, + 485, + 421, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 485, + 421, + 499 + ], + "score": 1.0, + "content": "The detailed optimization process of the proposed model is shown in Algorithm 1.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 90, + 514, + 453, + 528 + ], + "lines": [ + { + "bbox": [ + 88, + 514, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 88, + 514, + 455, + 529 + ], + "score": 1.0, + "content": "F ADDITIONAL EXPERIMENTS FOR DIGIT AND MULTI-PIE DATASETS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 91, + 540, + 520, + 574 + ], + "lines": [ + { + "bbox": [ + 89, + 540, + 522, + 553 + ], + "spans": [ + { + "bbox": [ + 89, + 540, + 522, + 553 + ], + "score": 1.0, + "content": "The additional experiments for Digit dataset where we set MNIST-M and USPS as source domain is available", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 551, + 522, + 564 + ], + "spans": [ + { + "bbox": [ + 88, + 551, + 372, + 564 + ], + "score": 1.0, + "content": "in tab. 4. The additional experiments for Multi-PIE dataset where we set", + "type": "text" + }, + { + "bbox": [ + 372, + 552, + 392, + 563 + ], + "score": 0.78, + "content": "C 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 551, + 395, + 564 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 395, + 552, + 414, + 563 + ], + "score": 0.73, + "content": "C 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 551, + 431, + 564 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 552, + 451, + 562 + ], + "score": 0.88, + "content": "C 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 551, + 522, + 564 + ], + "score": 1.0, + "content": "as source domain", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 563, + 196, + 574 + ], + "spans": [ + { + "bbox": [ + 90, + 564, + 168, + 574 + ], + "score": 1.0, + "content": "is available in tabs.", + "type": "text" + }, + { + "bbox": [ + 169, + 563, + 193, + 574 + ], + "score": 0.63, + "content": "5 \\& 6", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 564, + 196, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 91, + 592, + 192, + 605 + ], + "lines": [ + { + "bbox": [ + 89, + 590, + 194, + 607 + ], + "spans": [ + { + "bbox": [ + 89, + 590, + 194, + 607 + ], + "score": 1.0, + "content": "G PACS DATASET", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 91, + 618, + 522, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 616, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 616, + 241, + 631 + ], + "score": 1.0, + "content": "This dataset contains 9991 images", + "type": "text" + }, + { + "bbox": [ + 242, + 618, + 308, + 629 + ], + "score": 0.88, + "content": "2 2 7 \\times 2 2 7 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 616, + 522, + 631 + ], + "score": 1.0, + "content": "dimension) across 7 categories (‘dog’, ‘elephant’,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 628, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 89, + 628, + 522, + 643 + ], + "score": 1.0, + "content": "‘giraffe’,‘guitar’, ‘house’, ‘horse’ and ‘person’) and 4 domains of different stylistic depictions (‘Photo’, ‘Art", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 639, + 522, + 655 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 655 + ], + "score": 1.0, + "content": "painting’, ‘Cartoon’ and ‘Sketch’). The very diverse depiction styles provide a significant gap between", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 171, + 94, + 439, + 435 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 171, + 94, + 439, + 435 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 171, + 94, + 439, + 435 + ], + "spans": [ + { + "bbox": [ + 171, + 94, + 439, + 435 + ], + "score": 0.972, + "html": "
LayerEncoders (shared,private)
1CONV-(N16,K7,S1), ReLU
2CONV-(N32,K3,S2), ReLU
3CONV-(N64,K3,S2), ReLU
4RESBLK-(N64,K3,S1)
5RESBLK-(N64,K3,S1)
6RESBLK-(N64,K3,S1)
7RESBLK-(N64,K3,S1)
LayerDecoder
1RESBLK-(N64,K3,S1)
2RESBLK-(N64,K3,S1)
3RESBLK-(N64,K3,S1)
4RESBLK-(N64,K3,S1)
5DCONV-(N32,K3,S2), ReLU
6DCONV-(N16,K3,S2), ReLU
7DCONV-(N1,K1,S1), Tanh
LayerDiscriminator
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1),ReLU
3CONV-(N16,K3,S1),ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(ND), Softmax
LayerClassifier
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1), ReLU
3CONV-(N16,K3,S1), ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(NC), Softmax
", + "type": "table", + "image_path": "50d3352cf70c1ded69dc8863026d49852f9c33a698de56302a1a3fb842a20ce7.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 171, + 94, + 439, + 109.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 171, + 109.5, + 439, + 125.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 171, + 125.0, + 439, + 140.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 171, + 140.5, + 439, + 156.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 171, + 156.0, + 439, + 171.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 171, + 171.5, + 439, + 187.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 171, + 187.0, + 439, + 202.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 171, + 202.5, + 439, + 218.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 171, + 218.0, + 439, + 233.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 171, + 233.5, + 439, + 249.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 171, + 249.0, + 439, + 264.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 171, + 264.5, + 439, + 280.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 171, + 280.0, + 439, + 295.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 171, + 295.5, + 439, + 311.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 171, + 311.0, + 439, + 326.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 171, + 326.5, + 439, + 342.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 171, + 342.0, + 439, + 357.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 171, + 357.5, + 439, + 373.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 171, + 373.0, + 439, + 388.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 171, + 388.5, + 439, + 404.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 171, + 404.0, + 439, + 419.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 171, + 419.5, + 439, + 435.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 213, + 443, + 397, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 443, + 398, + 455 + ], + "spans": [ + { + "bbox": [ + 212, + 443, + 398, + 455 + ], + "score": 1.0, + "content": "Table 3: Network architecture for the experiments.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 16.25 + }, + { + "type": "title", + "bbox": [ + 91, + 460, + 284, + 473 + ], + "lines": [ + { + "bbox": [ + 88, + 458, + 285, + 475 + ], + "spans": [ + { + "bbox": [ + 88, + 458, + 285, + 475 + ], + "score": 1.0, + "content": "E PROPOSED MODEL’S ALGORITHM", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 91, + 486, + 419, + 498 + ], + "lines": [ + { + "bbox": [ + 89, + 485, + 421, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 485, + 421, + 499 + ], + "score": 1.0, + "content": "The detailed optimization process of the proposed model is shown in Algorithm 1.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 89, + 485, + 421, + 499 + ] + }, + { + "type": "title", + "bbox": [ + 90, + 514, + 453, + 528 + ], + "lines": [ + { + "bbox": [ + 88, + 514, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 88, + 514, + 455, + 529 + ], + "score": 1.0, + "content": "F ADDITIONAL EXPERIMENTS FOR DIGIT AND MULTI-PIE DATASETS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 91, + 540, + 520, + 574 + ], + "lines": [ + { + "bbox": [ + 89, + 540, + 522, + 553 + ], + "spans": [ + { + "bbox": [ + 89, + 540, + 522, + 553 + ], + "score": 1.0, + "content": "The additional experiments for Digit dataset where we set MNIST-M and USPS as source domain is available", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 551, + 522, + 564 + ], + "spans": [ + { + "bbox": [ + 88, + 551, + 372, + 564 + ], + "score": 1.0, + "content": "in tab. 4. The additional experiments for Multi-PIE dataset where we set", + "type": "text" + }, + { + "bbox": [ + 372, + 552, + 392, + 563 + ], + "score": 0.78, + "content": "C 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 551, + 395, + 564 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 395, + 552, + 414, + 563 + ], + "score": 0.73, + "content": "C 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 551, + 431, + 564 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 552, + 451, + 562 + ], + "score": 0.88, + "content": "C 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 551, + 522, + 564 + ], + "score": 1.0, + "content": "as source domain", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 90, + 563, + 196, + 574 + ], + "spans": [ + { + "bbox": [ + 90, + 564, + 168, + 574 + ], + "score": 1.0, + "content": "is available in tabs.", + "type": "text" + }, + { + "bbox": [ + 169, + 563, + 193, + 574 + ], + "score": 0.63, + "content": "5 \\& 6", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 564, + 196, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 88, + 540, + 522, + 574 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 592, + 192, + 605 + ], + "lines": [ + { + "bbox": [ + 89, + 590, + 194, + 607 + ], + "spans": [ + { + "bbox": [ + 89, + 590, + 194, + 607 + ], + "score": 1.0, + "content": "G PACS DATASET", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 91, + 618, + 522, + 651 + ], + "lines": [ + { + "bbox": [ + 89, + 616, + 522, + 631 + ], + "spans": [ + { + "bbox": [ + 89, + 616, + 241, + 631 + ], + "score": 1.0, + "content": "This dataset contains 9991 images", + "type": "text" + }, + { + "bbox": [ + 242, + 618, + 308, + 629 + ], + "score": 0.88, + "content": "2 2 7 \\times 2 2 7 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 616, + 522, + 631 + ], + "score": 1.0, + "content": "dimension) across 7 categories (‘dog’, ‘elephant’,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 628, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 89, + 628, + 522, + 643 + ], + "score": 1.0, + "content": "‘giraffe’,‘guitar’, ‘house’, ‘horse’ and ‘person’) and 4 domains of different stylistic depictions (‘Photo’, ‘Art", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 639, + 522, + 655 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 655 + ], + "score": 1.0, + "content": "painting’, ‘Cartoon’ and ‘Sketch’). The very diverse depiction styles provide a significant gap between", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 90, + 607, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 90, + 607, + 522, + 620 + ], + "score": 1.0, + "content": "domains, coupled with the small number of data samples, making it extremely challenging for domain", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 618, + 521, + 630 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 521, + 630 + ], + "score": 1.0, + "content": "adaptation. Consequently, the dataset was originally used for multi-source to single target domain adaptation", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 628, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 89, + 628, + 522, + 642 + ], + "score": 1.0, + "content": "(Li et al. (2017)). Instead, we tackle a significantly more challenging problem of single-source to multiple", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 639, + 521, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 521, + 653 + ], + "score": 1.0, + "content": "target adaptation. Tab. 7 shows the classification accuracy of various methods. MTDA-ITA consistently", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 31, + "bbox_fs": [ + 88, + 616, + 522, + 655 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 81, + 465, + 319 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 143, + 81, + 465, + 319 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 81, + 465, + 319 + ], + "spans": [ + { + "bbox": [ + 143, + 81, + 465, + 319 + ], + "score": 0.967, + "type": "image", + "image_path": "18facd59a83865bc6e26c98e1f1fdea84883eb0fbbf03a8444befb1f847d7fe4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 143, + 81, + 465, + 160.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 143, + 160.33333333333331, + 465, + 239.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 143, + 239.66666666666663, + 465, + 318.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 88, + 332, + 520, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "spans": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "score": 1.0, + "content": "Figure 4: Ablation of MTDA-ITA on Digit dataset. We show that each component of our method, Reconstruction loss,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 343, + 456, + 354 + ], + "spans": [ + { + "bbox": [ + 90, + 343, + 456, + 354 + ], + "score": 1.0, + "content": "Classifier entropy loss with separating shared/private features, contributes to the overall performance.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 90, + 366, + 519, + 533 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 366, + 519, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 366, + 519, + 533 + ], + "spans": [ + { + "bbox": [ + 90, + 366, + 519, + 533 + ], + "score": 0.98, + "html": "
methodMM→SMM→MMM→UU→SU→MU→MM
Source Only40.00±0.6184.46 ± 0.2980.43± 0.5023.41± 0.5250.64 ± 0.3741.45±0.38
1-NN21.4582.1336.9015.3438.4518.54
CORAL (Sun & Saenko (2016))40.20 ±0.6084.90 ± 0.7087.54 ± 0.4438.90 ± 0.9685.01 ± 0.6160.45 ± 0.70
DANN (Ganin et al. (2016))51.80 ± 0.9161.05 ± 0.7185.34 ± 0.6435.50± 0.8477.40 ± 0.6461.60 ± 0.64
ADDA (Tzeng et al. (2017))40.64 ± 0.8692.82 ± 0.4880.70±0.4841.23 ± 0.7890.10± 0.5856.21 ± 0.79
c-ADDA35.43 ± 0.9488.47 ± 0.6174.19 ± 0.5839.36 ± 0.9984.67 ± 0.9452.54 ± 0.88
DTN (Zhang et al. (2015)) c-DTN48.80± 0.6688.80±0.3890.68 ± 0.3542.43 ± 0.6189.04 ± 0.3655.78 ± 0.40
UNIT (Liu et al. (2017))44.21 ± 0.6183.60 ± 0.5484.98 ± 0.4139.75 ± 0.6485.04 ± 0.4548.86 ± 0.54
DSN (Bousmalis et al. (2016))1 51.50 ± 0.64/90.60
c-DSN47.10 ± 0.5090.20± 0.3189.95± 0.2948.20 ± 0.5991.40 ± 0.3060.45 ± 0.35
1p-DSN45.00 ± 0.6084.60 ± 0.4084.80 ±0.3940.50 ± 0.6186.05 ± 0.4656.25 ± 0.50
mp-DSN47.15 ± 0.6481.96 ± 0.6083.03 ±0.4939.30 ± 0.5184.55 ± 0.5655.03 ± 0.60
s-MTDA-ITA85.51 ± 0.5483.24± 0.2438.30 ± 0.7487.40 ± 0.3555.47 ± 0.44
c-MTDA-ITA50.55± 0.18 47.32±0.1994.82±0.2189.05 ±0.2840.13±0.3087.10 ± 0.2561.01 ± 0.24
MTDA-ITA53.50 ± 0.2290.20 ±0.30 98.20 ± 0.1090.01 ± 0.24 94.10 ± 0.1141.10 ± 0.35 46.00 ± 0.4885.35 ± 0.28 91.50 ± 0.2360.31± 0.34 67.30 ± 0.15
", + "type": "table", + "image_path": "ca7be10fae8c750a1749320d86fa67cb898aefc922aeca1ed95a61a028bdff54.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 90, + 366, + 519, + 421.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 90, + 421.6666666666667, + 519, + 477.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 90, + 477.33333333333337, + 519, + 533.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 90, + 547, + 521, + 598 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 547, + 522, + 559 + ], + "spans": [ + { + "bbox": [ + 89, + 547, + 522, + 559 + ], + "score": 1.0, + "content": "Table 4: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 558, + 521, + 569 + ], + "spans": [ + { + "bbox": [ + 89, + 558, + 521, + 569 + ], + "score": 1.0, + "content": "red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDA-ITA", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 567, + 522, + 580 + ], + "spans": [ + { + "bbox": [ + 89, + 567, + 470, + 580 + ], + "score": 1.0, + "content": "where each one correspond to a source-target pair. mp-DSN: extended DSN with multiple private encoder.", + "type": "text" + }, + { + "bbox": [ + 470, + 568, + 499, + 578 + ], + "score": 0.62, + "content": "\\bf \\Pi ^ { * } U N I T", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 567, + 522, + 580 + ], + "score": 1.0, + "content": "trains", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 577, + 523, + 590 + ], + "spans": [ + { + "bbox": [ + 88, + 577, + 186, + 590 + ], + "score": 1.0, + "content": "with the extended SVHN", + "type": "text" + }, + { + "bbox": [ + 187, + 578, + 219, + 588 + ], + "score": 0.82, + "content": "> 5 0 0 \\mathrm { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 577, + 299, + 590 + ], + "score": 1.0, + "content": "images vs ours 72K).", + "type": "text" + }, + { + "bbox": [ + 299, + 578, + 305, + 587 + ], + "score": 0.39, + "content": "^ *", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 577, + 356, + 590 + ], + "score": 1.0, + "content": "PixelDA uses", + "type": "text" + }, + { + "bbox": [ + 356, + 578, + 395, + 588 + ], + "score": 0.83, + "content": "( \\approx 1 , 0 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 577, + 523, + 590 + ], + "score": 1.0, + "content": "of labeled target domain data as a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 587, + 259, + 600 + ], + "spans": [ + { + "bbox": [ + 89, + 587, + 259, + 600 + ], + "score": 1.0, + "content": "validation set for tuning the hyper-parameters.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + } + ], + "index": 8.0 + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 90, + 607, + 522, + 620 + ], + "spans": [ + { + "bbox": [ + 90, + 607, + 522, + 620 + ], + "score": 1.0, + "content": "domains, coupled with the small number of data samples, making it extremely challenging for domain", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 618, + 521, + 630 + ], + "spans": [ + { + "bbox": [ + 89, + 618, + 521, + 630 + ], + "score": 1.0, + "content": "adaptation. Consequently, the dataset was originally used for multi-source to single target domain adaptation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 628, + 522, + 642 + ], + "spans": [ + { + "bbox": [ + 89, + 628, + 522, + 642 + ], + "score": 1.0, + "content": "(Li et al. (2017)). Instead, we tackle a significantly more challenging problem of single-source to multiple", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 639, + 521, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 521, + 653 + ], + "score": 1.0, + "content": "target adaptation. Tab. 7 shows the classification accuracy of various methods. MTDA-ITA consistently", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 81, + 465, + 319 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 143, + 81, + 465, + 319 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 81, + 465, + 319 + ], + "spans": [ + { + "bbox": [ + 143, + 81, + 465, + 319 + ], + "score": 0.967, + "type": "image", + "image_path": "18facd59a83865bc6e26c98e1f1fdea84883eb0fbbf03a8444befb1f847d7fe4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 143, + 81, + 465, + 160.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 143, + 160.33333333333331, + 465, + 239.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 143, + 239.66666666666663, + 465, + 318.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 88, + 332, + 520, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "spans": [ + { + "bbox": [ + 89, + 332, + 522, + 344 + ], + "score": 1.0, + "content": "Figure 4: Ablation of MTDA-ITA on Digit dataset. We show that each component of our method, Reconstruction loss,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 343, + 456, + 354 + ], + "spans": [ + { + "bbox": [ + 90, + 343, + 456, + 354 + ], + "score": 1.0, + "content": "Classifier entropy loss with separating shared/private features, contributes to the overall performance.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 90, + 366, + 519, + 533 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 366, + 519, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 366, + 519, + 533 + ], + "spans": [ + { + "bbox": [ + 90, + 366, + 519, + 533 + ], + "score": 0.98, + "html": "
methodMM→SMM→MMM→UU→SU→MU→MM
Source Only40.00±0.6184.46 ± 0.2980.43± 0.5023.41± 0.5250.64 ± 0.3741.45±0.38
1-NN21.4582.1336.9015.3438.4518.54
CORAL (Sun & Saenko (2016))40.20 ±0.6084.90 ± 0.7087.54 ± 0.4438.90 ± 0.9685.01 ± 0.6160.45 ± 0.70
DANN (Ganin et al. (2016))51.80 ± 0.9161.05 ± 0.7185.34 ± 0.6435.50± 0.8477.40 ± 0.6461.60 ± 0.64
ADDA (Tzeng et al. (2017))40.64 ± 0.8692.82 ± 0.4880.70±0.4841.23 ± 0.7890.10± 0.5856.21 ± 0.79
c-ADDA35.43 ± 0.9488.47 ± 0.6174.19 ± 0.5839.36 ± 0.9984.67 ± 0.9452.54 ± 0.88
DTN (Zhang et al. (2015)) c-DTN48.80± 0.6688.80±0.3890.68 ± 0.3542.43 ± 0.6189.04 ± 0.3655.78 ± 0.40
UNIT (Liu et al. (2017))44.21 ± 0.6183.60 ± 0.5484.98 ± 0.4139.75 ± 0.6485.04 ± 0.4548.86 ± 0.54
DSN (Bousmalis et al. (2016))1 51.50 ± 0.64/90.60
c-DSN47.10 ± 0.5090.20± 0.3189.95± 0.2948.20 ± 0.5991.40 ± 0.3060.45 ± 0.35
1p-DSN45.00 ± 0.6084.60 ± 0.4084.80 ±0.3940.50 ± 0.6186.05 ± 0.4656.25 ± 0.50
mp-DSN47.15 ± 0.6481.96 ± 0.6083.03 ±0.4939.30 ± 0.5184.55 ± 0.5655.03 ± 0.60
s-MTDA-ITA85.51 ± 0.5483.24± 0.2438.30 ± 0.7487.40 ± 0.3555.47 ± 0.44
c-MTDA-ITA50.55± 0.18 47.32±0.1994.82±0.2189.05 ±0.2840.13±0.3087.10 ± 0.2561.01 ± 0.24
MTDA-ITA53.50 ± 0.2290.20 ±0.30 98.20 ± 0.1090.01 ± 0.24 94.10 ± 0.1141.10 ± 0.35 46.00 ± 0.4885.35 ± 0.28 91.50 ± 0.2360.31± 0.34 67.30 ± 0.15
", + "type": "table", + "image_path": "ca7be10fae8c750a1749320d86fa67cb898aefc922aeca1ed95a61a028bdff54.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 90, + 366, + 519, + 421.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 90, + 421.6666666666667, + 519, + 477.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 90, + 477.33333333333337, + 519, + 533.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 90, + 547, + 521, + 598 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 89, + 547, + 522, + 559 + ], + "spans": [ + { + "bbox": [ + 89, + 547, + 522, + 559 + ], + "score": 1.0, + "content": "Table 4: Classification results on digit datasets. M: MNIST; MM: MNIST-M, S: SVHN, U: USPS. The best is shown in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 558, + 521, + 569 + ], + "spans": [ + { + "bbox": [ + 89, + 558, + 521, + 569 + ], + "score": 1.0, + "content": "red. c-X: combining all target domains into a single one and train it using X. s-MTDA-ITA: training multiple MTDA-ITA", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 567, + 522, + 580 + ], + "spans": [ + { + "bbox": [ + 89, + 567, + 470, + 580 + ], + "score": 1.0, + "content": "where each one correspond to a source-target pair. mp-DSN: extended DSN with multiple private encoder.", + "type": "text" + }, + { + "bbox": [ + 470, + 568, + 499, + 578 + ], + "score": 0.62, + "content": "\\bf \\Pi ^ { * } U N I T", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 567, + 522, + 580 + ], + "score": 1.0, + "content": "trains", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 577, + 523, + 590 + ], + "spans": [ + { + "bbox": [ + 88, + 577, + 186, + 590 + ], + "score": 1.0, + "content": "with the extended SVHN", + "type": "text" + }, + { + "bbox": [ + 187, + 578, + 219, + 588 + ], + "score": 0.82, + "content": "> 5 0 0 \\mathrm { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 577, + 299, + 590 + ], + "score": 1.0, + "content": "images vs ours 72K).", + "type": "text" + }, + { + "bbox": [ + 299, + 578, + 305, + 587 + ], + "score": 0.39, + "content": "^ *", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 577, + 356, + 590 + ], + "score": 1.0, + "content": "PixelDA uses", + "type": "text" + }, + { + "bbox": [ + 356, + 578, + 395, + 588 + ], + "score": 0.83, + "content": "( \\approx 1 , 0 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 577, + 523, + 590 + ], + "score": 1.0, + "content": "of labeled target domain data as a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 587, + 259, + 600 + ], + "spans": [ + { + "bbox": [ + 89, + 587, + 259, + 600 + ], + "score": 1.0, + "content": "validation set for tuning the hyper-parameters.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + } + ], + "index": 8.0 + }, + { + "type": "text", + "bbox": [ + 91, + 607, + 521, + 651 + ], + "lines": [], + "index": 14.5, + "bbox_fs": [ + 89, + 607, + 522, + 653 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 92, + 519, + 320 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 92, + 519, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 92, + 519, + 320 + ], + "spans": [ + { + "bbox": [ + 90, + 92, + 519, + 320 + ], + "score": 0.982, + "html": "
methodC05 →C08C05→C09C05→C13C05→C14
Source Only31.56 ± 0.4040.67 ± 0.3639.89 ± 0.2254.70± 0.25
1-NN CORAL (Sun & Saenko (2016))27.28 36.55 ± 0.6631.22 38.60 ± 0.6733.6647.04 55.29 ± 0.47
40.60 ± 0.58
DANN (citeganin2016domain)40.30 ± 0.6041.20 ± 0.6540.12 ± 0.6058.90 ± 0.38
ADDA (Tzeng et al. (2017)) C-ADDA33.21 ± 0.8130.86 ± 0.9052.44± 0.8070.18 ± 0.60
46.88 ± 0.6536.38 ± 0.8839.14 ± 0.8565.41 ± 0.69
DTN (Zhang et al. (2015))38.50 ± 0.5130.56 ± 0.4655.78 ± 0.3668.90 ± 0.31
c-DTN41.70 ± 0.4231.10 ± 0.4850.19 ± 0.4560.34 ± 0.35
PixelDA (Bousmalis et al. (2017))44.93 ±0.4244.75 ± 0.4545.18 ± 0.4546.88 ± 0.49
UNIT (Liu et al. (2017))44.47 ± 0.2144.47 ± 0.2144.47 ± 0.2044.51 ± 0.28
DSN (Bousmalis et al. (2016))45.12 ± 0.4644.35± 0.4948.12 ± 0.5375.00 ± 0.39
c-DSN42.52 ± 0.4838.54 ± 0.6434.15 ± 0.6469.45 ± 0.55
1p-DSN41.64 ± 0.5837.84 ± 0.6334.65 ± 0.4468.75 ± 0.85
mp-DSN41.30 ± 0.2835.14 ± 0.3534.40 ± 0.3565.70 ± 0.27
s-MTDA-ITA44.40 ± 0.2344.60 ± 0.2547.65 ± 0.2780.20 ± 0.13
c-MTDA-ITA40.49 ± 0.2540.70 ± 0.2542.80 ± 0.2571.60 ± 0.10
MTDA-ITA49.01 ± 0.20
48.20± 0.2753.13 ±0.2284.29 9±0.10
", + "type": "table", + "image_path": "e5083e1d87737e6b3895ff8dbf3fa881fb2f3211b52e8393a6236ec40c8fe0f1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 90, + 92, + 519, + 168.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 90, + 168.0, + 519, + 244.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 90, + 244.0, + 519, + 320.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 165, + 334, + 445, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 164, + 333, + 446, + 345 + ], + "spans": [ + { + "bbox": [ + 164, + 333, + 446, + 345 + ], + "score": 1.0, + "content": "Table 5: Classification results on Multi-PIE dataset. The best is shown in red.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 91, + 353, + 238, + 365 + ], + "lines": [ + { + "bbox": [ + 90, + 352, + 237, + 367 + ], + "spans": [ + { + "bbox": [ + 90, + 352, + 237, + 367 + ], + "score": 1.0, + "content": "Algorithm 1 MDTA-ITA Algorithm", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 92, + 368, + 479, + 534 + ], + "lines": [ + { + "bbox": [ + 90, + 367, + 261, + 380 + ], + "spans": [ + { + "bbox": [ + 90, + 367, + 133, + 380 + ], + "score": 1.0, + "content": "Require:", + "type": "text" + }, + { + "bbox": [ + 133, + 367, + 191, + 379 + ], + "score": 0.9, + "content": "\\{ \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } \\} { : \\mathrm { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 367, + 261, + 380 + ], + "score": 1.0, + "content": "domain datasets.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 378, + 273, + 391 + ], + "spans": [ + { + "bbox": [ + 125, + 379, + 165, + 390 + ], + "score": 0.88, + "content": "\\lambda _ { r } , \\lambda _ { c } , \\lambda _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 378, + 273, + 391 + ], + "score": 1.0, + "content": ": Model hyper-parameters.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 389, + 269, + 402 + ], + "spans": [ + { + "bbox": [ + 90, + 389, + 129, + 402 + ], + "score": 1.0, + "content": "Ensure:", + "type": "text" + }, + { + "bbox": [ + 129, + 390, + 187, + 401 + ], + "score": 0.89, + "content": "\\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 389, + 269, + 402 + ], + "score": 1.0, + "content": ": Model parameters.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 399, + 208, + 413 + ], + "spans": [ + { + "bbox": [ + 93, + 399, + 145, + 413 + ], + "score": 1.0, + "content": "1: Initialize", + "type": "text" + }, + { + "bbox": [ + 145, + 401, + 204, + 412 + ], + "score": 0.92, + "content": "\\bar { \\theta } _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 399, + 208, + 413 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 410, + 137, + 424 + ], + "spans": [ + { + "bbox": [ + 93, + 410, + 137, + 424 + ], + "score": 1.0, + "content": "2: repeat", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 422, + 378, + 435 + ], + "spans": [ + { + "bbox": [ + 94, + 422, + 105, + 434 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 116, + 422, + 378, + 435 + ], + "score": 1.0, + "content": "Sample a mini-batch from each of source/target domain datasets.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 94, + 432, + 474, + 451 + ], + "spans": [ + { + "bbox": [ + 94, + 435, + 105, + 446 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 116, + 432, + 149, + 451 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 434, + 168, + 447 + ], + "score": 0.91, + "content": "\\{ \\theta _ { s } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 432, + 230, + 451 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 231, + 434, + 243, + 446 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 432, + 406, + 451 + ], + "score": 1.0, + "content": "in Eq.(10) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 407, + 433, + 472, + 448 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\theta _ { s } = \\theta _ { s } - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\theta _ { s } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 432, + 474, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 94, + 444, + 471, + 468 + ], + "spans": [ + { + "bbox": [ + 94, + 450, + 105, + 462 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 114, + 444, + 471, + 468 + ], + "score": 1.0, + "content": "Update {θp} by minimizing Lp in Eq.(9) through the gradient descent:θp = θp − η ∂Lp∂θp .", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 94, + 464, + 463, + 479 + ], + "spans": [ + { + "bbox": [ + 94, + 466, + 105, + 477 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 116, + 464, + 149, + 479 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 465, + 168, + 477 + ], + "score": 0.91, + "content": "\\{ \\theta _ { c } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 464, + 230, + 479 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 231, + 465, + 243, + 477 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 464, + 398, + 479 + ], + "score": 1.0, + "content": "in Eq.(8) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 398, + 464, + 463, + 479 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\theta _ { c } = \\theta _ { c } - \\eta \\frac { \\partial \\mathcal { L } _ { c } } { \\partial \\theta _ { c } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 95, + 478, + 459, + 496 + ], + "spans": [ + { + "bbox": [ + 95, + 480, + 105, + 492 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 116, + 478, + 149, + 496 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 480, + 165, + 493 + ], + "score": 0.9, + "content": "\\{ \\phi \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 478, + 227, + 496 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 228, + 480, + 241, + 493 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 478, + 397, + 496 + ], + "score": 1.0, + "content": "in Eq.(6) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 397, + 479, + 457, + 494 + ], + "score": 0.87, + "content": "\\phi = \\phi - \\eta \\frac { \\partial \\mathcal { L } _ { \\phi } } { \\partial \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 478, + 459, + 496 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 95, + 491, + 462, + 511 + ], + "spans": [ + { + "bbox": [ + 95, + 496, + 105, + 507 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 116, + 491, + 149, + 511 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 495, + 167, + 507 + ], + "score": 0.91, + "content": "\\{ \\psi \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 491, + 228, + 511 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 228, + 495, + 243, + 507 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 491, + 398, + 511 + ], + "score": 1.0, + "content": "in Eq.(5) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 398, + 494, + 460, + 510 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\psi = \\psi - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\psi } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 491, + 462, + 511 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 95, + 507, + 186, + 520 + ], + "spans": [ + { + "bbox": [ + 95, + 509, + 105, + 519 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 106, + 507, + 186, + 520 + ], + "score": 1.0, + "content": "until Convergence;", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 91, + 519, + 206, + 532 + ], + "spans": [ + { + "bbox": [ + 91, + 519, + 134, + 532 + ], + "score": 1.0, + "content": "10: return", + "type": "text" + }, + { + "bbox": [ + 134, + 519, + 202, + 531 + ], + "score": 0.92, + "content": "\\{ \\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 519, + 206, + 532 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 91, + 542, + 522, + 576 + ], + "lines": [ + { + "bbox": [ + 90, + 541, + 522, + 555 + ], + "spans": [ + { + "bbox": [ + 90, + 541, + 522, + 555 + ], + "score": 1.0, + "content": "achieves the best performance for all transfer tasks. Evaluations were obtained by training all models (ADDA,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 552, + 522, + 567 + ], + "spans": [ + { + "bbox": [ + 89, + 552, + 522, + 567 + ], + "score": 1.0, + "content": "DSN, and ours) from scratch on the PACS dataset. Note that the overall performance figures are low due to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 90, + 564, + 441, + 577 + ], + "spans": [ + { + "bbox": [ + 90, + 564, + 441, + 577 + ], + "score": 1.0, + "content": "the extreme difficulty of the transfer task, induced by large differences among domains.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 90, + 593, + 373, + 605 + ], + "lines": [ + { + "bbox": [ + 89, + 592, + 374, + 607 + ], + "spans": [ + { + "bbox": [ + 89, + 592, + 374, + 607 + ], + "score": 1.0, + "content": "H ANALYSIS OF SHARED/PRIVATE SPACE EMBEDDING", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 91, + 618, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 88, + 616, + 522, + 632 + ], + "spans": [ + { + "bbox": [ + 88, + 616, + 522, + 632 + ], + "score": 1.0, + "content": "In the experiments conducted, we showed that our approach is able to achieve better performance than the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 88, + 628, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 88, + 628, + 522, + 643 + ], + "score": 1.0, + "content": "competing methods including the extended DSN with one private encoder (1p-DSN) which is the most similar", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 639, + 155, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 155, + 653 + ], + "score": 1.0, + "content": "method to ours.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 683 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 683 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 92, + 519, + 320 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 92, + 519, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 92, + 519, + 320 + ], + "spans": [ + { + "bbox": [ + 90, + 92, + 519, + 320 + ], + "score": 0.982, + "html": "
methodC05 →C08C05→C09C05→C13C05→C14
Source Only31.56 ± 0.4040.67 ± 0.3639.89 ± 0.2254.70± 0.25
1-NN CORAL (Sun & Saenko (2016))27.28 36.55 ± 0.6631.22 38.60 ± 0.6733.6647.04 55.29 ± 0.47
40.60 ± 0.58
DANN (citeganin2016domain)40.30 ± 0.6041.20 ± 0.6540.12 ± 0.6058.90 ± 0.38
ADDA (Tzeng et al. (2017)) C-ADDA33.21 ± 0.8130.86 ± 0.9052.44± 0.8070.18 ± 0.60
46.88 ± 0.6536.38 ± 0.8839.14 ± 0.8565.41 ± 0.69
DTN (Zhang et al. (2015))38.50 ± 0.5130.56 ± 0.4655.78 ± 0.3668.90 ± 0.31
c-DTN41.70 ± 0.4231.10 ± 0.4850.19 ± 0.4560.34 ± 0.35
PixelDA (Bousmalis et al. (2017))44.93 ±0.4244.75 ± 0.4545.18 ± 0.4546.88 ± 0.49
UNIT (Liu et al. (2017))44.47 ± 0.2144.47 ± 0.2144.47 ± 0.2044.51 ± 0.28
DSN (Bousmalis et al. (2016))45.12 ± 0.4644.35± 0.4948.12 ± 0.5375.00 ± 0.39
c-DSN42.52 ± 0.4838.54 ± 0.6434.15 ± 0.6469.45 ± 0.55
1p-DSN41.64 ± 0.5837.84 ± 0.6334.65 ± 0.4468.75 ± 0.85
mp-DSN41.30 ± 0.2835.14 ± 0.3534.40 ± 0.3565.70 ± 0.27
s-MTDA-ITA44.40 ± 0.2344.60 ± 0.2547.65 ± 0.2780.20 ± 0.13
c-MTDA-ITA40.49 ± 0.2540.70 ± 0.2542.80 ± 0.2571.60 ± 0.10
MTDA-ITA49.01 ± 0.20
48.20± 0.2753.13 ±0.2284.29 9±0.10
", + "type": "table", + "image_path": "e5083e1d87737e6b3895ff8dbf3fa881fb2f3211b52e8393a6236ec40c8fe0f1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 90, + 92, + 519, + 168.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 90, + 168.0, + 519, + 244.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 90, + 244.0, + 519, + 320.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 165, + 334, + 445, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 164, + 333, + 446, + 345 + ], + "spans": [ + { + "bbox": [ + 164, + 333, + 446, + 345 + ], + "score": 1.0, + "content": "Table 5: Classification results on Multi-PIE dataset. The best is shown in red.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 91, + 353, + 238, + 365 + ], + "lines": [ + { + "bbox": [ + 90, + 352, + 237, + 367 + ], + "spans": [ + { + "bbox": [ + 90, + 352, + 237, + 367 + ], + "score": 1.0, + "content": "Algorithm 1 MDTA-ITA Algorithm", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "list", + "bbox": [ + 92, + 368, + 479, + 534 + ], + "lines": [ + { + "bbox": [ + 90, + 367, + 261, + 380 + ], + "spans": [ + { + "bbox": [ + 90, + 367, + 133, + 380 + ], + "score": 1.0, + "content": "Require:", + "type": "text" + }, + { + "bbox": [ + 133, + 367, + 191, + 379 + ], + "score": 0.9, + "content": "\\{ \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } \\} { : \\mathrm { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 367, + 261, + 380 + ], + "score": 1.0, + "content": "domain datasets.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 125, + 378, + 273, + 391 + ], + "spans": [ + { + "bbox": [ + 125, + 379, + 165, + 390 + ], + "score": 0.88, + "content": "\\lambda _ { r } , \\lambda _ { c } , \\lambda _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 378, + 273, + 391 + ], + "score": 1.0, + "content": ": Model hyper-parameters.", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 90, + 389, + 269, + 402 + ], + "spans": [ + { + "bbox": [ + 90, + 389, + 129, + 402 + ], + "score": 1.0, + "content": "Ensure:", + "type": "text" + }, + { + "bbox": [ + 129, + 390, + 187, + 401 + ], + "score": 0.89, + "content": "\\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 389, + 269, + 402 + ], + "score": 1.0, + "content": ": Model parameters.", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 93, + 399, + 208, + 413 + ], + "spans": [ + { + "bbox": [ + 93, + 399, + 145, + 413 + ], + "score": 1.0, + "content": "1: Initialize", + "type": "text" + }, + { + "bbox": [ + 145, + 401, + 204, + 412 + ], + "score": 0.92, + "content": "\\bar { \\theta } _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 399, + 208, + 413 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 93, + 410, + 137, + 424 + ], + "spans": [ + { + "bbox": [ + 93, + 410, + 137, + 424 + ], + "score": 1.0, + "content": "2: repeat", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 422, + 378, + 435 + ], + "spans": [ + { + "bbox": [ + 94, + 422, + 105, + 434 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 116, + 422, + 378, + 435 + ], + "score": 1.0, + "content": "Sample a mini-batch from each of source/target domain datasets.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 94, + 432, + 474, + 451 + ], + "spans": [ + { + "bbox": [ + 94, + 435, + 105, + 446 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 116, + 432, + 149, + 451 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 434, + 168, + 447 + ], + "score": 0.91, + "content": "\\{ \\theta _ { s } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 432, + 230, + 451 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 231, + 434, + 243, + 446 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 432, + 406, + 451 + ], + "score": 1.0, + "content": "in Eq.(10) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 407, + 433, + 472, + 448 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\theta _ { s } = \\theta _ { s } - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\theta _ { s } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 432, + 474, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 94, + 444, + 471, + 468 + ], + "spans": [ + { + "bbox": [ + 94, + 450, + 105, + 462 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 114, + 444, + 471, + 468 + ], + "score": 1.0, + "content": "Update {θp} by minimizing Lp in Eq.(9) through the gradient descent:θp = θp − η ∂Lp∂θp .", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 94, + 464, + 463, + 479 + ], + "spans": [ + { + "bbox": [ + 94, + 466, + 105, + 477 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 116, + 464, + 149, + 479 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 465, + 168, + 477 + ], + "score": 0.91, + "content": "\\{ \\theta _ { c } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 464, + 230, + 479 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 231, + 465, + 243, + 477 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 464, + 398, + 479 + ], + "score": 1.0, + "content": "in Eq.(8) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 398, + 464, + 463, + 479 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\theta _ { c } = \\theta _ { c } - \\eta \\frac { \\partial \\mathcal { L } _ { c } } { \\partial \\theta _ { c } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 478, + 459, + 496 + ], + "spans": [ + { + "bbox": [ + 95, + 480, + 105, + 492 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 116, + 478, + 149, + 496 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 480, + 165, + 493 + ], + "score": 0.9, + "content": "\\{ \\phi \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 478, + 227, + 496 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 228, + 480, + 241, + 493 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 478, + 397, + 496 + ], + "score": 1.0, + "content": "in Eq.(6) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 397, + 479, + 457, + 494 + ], + "score": 0.87, + "content": "\\phi = \\phi - \\eta \\frac { \\partial \\mathcal { L } _ { \\phi } } { \\partial \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 478, + 459, + 496 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 95, + 491, + 462, + 511 + ], + "spans": [ + { + "bbox": [ + 95, + 496, + 105, + 507 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 116, + 491, + 149, + 511 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 149, + 495, + 167, + 507 + ], + "score": 0.91, + "content": "\\{ \\psi \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 491, + 228, + 511 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + }, + { + "bbox": [ + 228, + 495, + 243, + 507 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 491, + 398, + 511 + ], + "score": 1.0, + "content": "in Eq.(5) through the gradient descent:", + "type": "text" + }, + { + "bbox": [ + 398, + 494, + 460, + 510 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\psi = \\psi - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\psi } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 491, + 462, + 511 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 95, + 507, + 186, + 520 + ], + "spans": [ + { + "bbox": [ + 95, + 509, + 105, + 519 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 106, + 507, + 186, + 520 + ], + "score": 1.0, + "content": "until Convergence;", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 91, + 519, + 206, + 532 + ], + "spans": [ + { + "bbox": [ + 91, + 519, + 134, + 532 + ], + "score": 1.0, + "content": "10: return", + "type": "text" + }, + { + "bbox": [ + 134, + 519, + 202, + 531 + ], + "score": 0.92, + "content": "\\{ \\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 519, + 206, + 532 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 11, + "bbox_fs": [ + 90, + 367, + 474, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 542, + 522, + 576 + ], + "lines": [ + { + "bbox": [ + 90, + 541, + 522, + 555 + ], + "spans": [ + { + "bbox": [ + 90, + 541, + 522, + 555 + ], + "score": 1.0, + "content": "achieves the best performance for all transfer tasks. Evaluations were obtained by training all models (ADDA,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 552, + 522, + 567 + ], + "spans": [ + { + "bbox": [ + 89, + 552, + 522, + 567 + ], + "score": 1.0, + "content": "DSN, and ours) from scratch on the PACS dataset. Note that the overall performance figures are low due to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 90, + 564, + 441, + 577 + ], + "spans": [ + { + "bbox": [ + 90, + 564, + 441, + 577 + ], + "score": 1.0, + "content": "the extreme difficulty of the transfer task, induced by large differences among domains.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 89, + 541, + 522, + 577 + ] + }, + { + "type": "title", + "bbox": [ + 90, + 593, + 373, + 605 + ], + "lines": [ + { + "bbox": [ + 89, + 592, + 374, + 607 + ], + "spans": [ + { + "bbox": [ + 89, + 592, + 374, + 607 + ], + "score": 1.0, + "content": "H ANALYSIS OF SHARED/PRIVATE SPACE EMBEDDING", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 91, + 618, + 521, + 651 + ], + "lines": [ + { + "bbox": [ + 88, + 616, + 522, + 632 + ], + "spans": [ + { + "bbox": [ + 88, + 616, + 522, + 632 + ], + "score": 1.0, + "content": "In the experiments conducted, we showed that our approach is able to achieve better performance than the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 88, + 628, + 522, + 643 + ], + "spans": [ + { + "bbox": [ + 88, + 628, + 522, + 643 + ], + "score": 1.0, + "content": "competing methods including the extended DSN with one private encoder (1p-DSN) which is the most similar", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 639, + 155, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 639, + 155, + 653 + ], + "score": 1.0, + "content": "method to ours.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 88, + 616, + 522, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 103, + 91, + 471, + 461 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 103, + 91, + 471, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 103, + 91, + 471, + 461 + ], + "spans": [ + { + "bbox": [ + 103, + 91, + 471, + 461 + ], + "score": 0.977, + "type": "image", + "image_path": "07d4618a501cd10a00bd625bd5df1054e718feb2cb44cd9581d0cca9092af8f0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 103, + 91, + 471, + 214.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 103, + 214.33333333333331, + 471, + 337.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 103, + 337.66666666666663, + 471, + 460.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 468, + 522, + 519 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 468, + 522, + 480 + ], + "spans": [ + { + "bbox": [ + 90, + 468, + 522, + 480 + ], + "score": 1.0, + "content": "Figure 5: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 479, + 523, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 523, + 490 + ], + "score": 1.0, + "content": "show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 489, + 522, + 500 + ], + "spans": [ + { + "bbox": [ + 89, + 489, + 522, + 500 + ], + "score": 1.0, + "content": "(c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 498, + 522, + 511 + ], + "spans": [ + { + "bbox": [ + 89, + 498, + 522, + 511 + ], + "score": 1.0, + "content": "in the right column represent points from the shared space, while the smaller ones are from the private spaces. (e),(f)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 509, + 197, + 520 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 197, + 520 + ], + "score": 1.0, + "content": "learned features for 1p-DSN.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 90, + 535, + 520, + 558 + ], + "lines": [ + { + "bbox": [ + 90, + 534, + 521, + 547 + ], + "spans": [ + { + "bbox": [ + 90, + 534, + 521, + 547 + ], + "score": 1.0, + "content": "Indeed, fig. 5 depicts the embedding of the MTDA-ITA learned private/shared features, those of 1p-DSN", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 546, + 447, + 559 + ], + "spans": [ + { + "bbox": [ + 90, + 546, + 447, + 559 + ], + "score": 1.0, + "content": "and the original features from different domains for Digit datasets (SVHN is the source).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 91, + 563, + 521, + 652 + ], + "lines": [ + { + "bbox": [ + 90, + 564, + 521, + 575 + ], + "spans": [ + { + "bbox": [ + 90, + 564, + 521, + 575 + ], + "score": 1.0, + "content": "Notice that both MTDA-ITA and 1p-DSN reduces the domain mismatch for the shared features (circle", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 574, + 522, + 586 + ], + "spans": [ + { + "bbox": [ + 88, + 574, + 522, + 586 + ], + "score": 1.0, + "content": "markers in fig. 5) and separate the shared features from private features. On the other hand, MTDA-ITA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 586, + 522, + 598 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 522, + 598 + ], + "score": 1.0, + "content": "increases the domain separation for the private features (triangle markers, pure and well-separated domain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "score": 1.0, + "content": "clusters in fig. 5c) while 1p-DSN is unable to enforce the private representation of different domains to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "score": 1.0, + "content": "be different (fig. 5e) that may result in redundancy of different private spaces. This is partially due to the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 618, + 523, + 631 + ], + "spans": [ + { + "bbox": [ + 88, + 618, + 397, + 631 + ], + "score": 1.0, + "content": "proposed multi-domain separation loss through the use of the domain classifier", + "type": "text" + }, + { + "bbox": [ + 398, + 619, + 407, + 628 + ], + "score": 0.76, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 618, + 523, + 631 + ], + "score": 1.0, + "content": ", which penalizes the domain", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 629, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 521, + 642 + ], + "score": 1.0, + "content": "mismatch for the shared features and rewards the mismatch for the private features, something the 1p-DSN", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "score": 1.0, + "content": "fails to account for. Moreover, as supported by the quantitative results in tab. 1, the class label separation in", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 294, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 670, + 296, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 103, + 91, + 471, + 461 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 103, + 91, + 471, + 461 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 103, + 91, + 471, + 461 + ], + "spans": [ + { + "bbox": [ + 103, + 91, + 471, + 461 + ], + "score": 0.977, + "type": "image", + "image_path": "07d4618a501cd10a00bd625bd5df1054e718feb2cb44cd9581d0cca9092af8f0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 103, + 91, + 471, + 214.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 103, + 214.33333333333331, + 471, + 337.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 103, + 337.66666666666663, + 471, + 460.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 468, + 522, + 519 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 468, + 522, + 480 + ], + "spans": [ + { + "bbox": [ + 90, + 468, + 522, + 480 + ], + "score": 1.0, + "content": "Figure 5: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 479, + 523, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 523, + 490 + ], + "score": 1.0, + "content": "show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 489, + 522, + 500 + ], + "spans": [ + { + "bbox": [ + 89, + 489, + 522, + 500 + ], + "score": 1.0, + "content": "(c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 498, + 522, + 511 + ], + "spans": [ + { + "bbox": [ + 89, + 498, + 522, + 511 + ], + "score": 1.0, + "content": "in the right column represent points from the shared space, while the smaller ones are from the private spaces. (e),(f)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 509, + 197, + 520 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 197, + 520 + ], + "score": 1.0, + "content": "learned features for 1p-DSN.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 90, + 535, + 520, + 558 + ], + "lines": [ + { + "bbox": [ + 90, + 534, + 521, + 547 + ], + "spans": [ + { + "bbox": [ + 90, + 534, + 521, + 547 + ], + "score": 1.0, + "content": "Indeed, fig. 5 depicts the embedding of the MTDA-ITA learned private/shared features, those of 1p-DSN", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 546, + 447, + 559 + ], + "spans": [ + { + "bbox": [ + 90, + 546, + 447, + 559 + ], + "score": 1.0, + "content": "and the original features from different domains for Digit datasets (SVHN is the source).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 90, + 534, + 521, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 91, + 563, + 521, + 652 + ], + "lines": [ + { + "bbox": [ + 90, + 564, + 521, + 575 + ], + "spans": [ + { + "bbox": [ + 90, + 564, + 521, + 575 + ], + "score": 1.0, + "content": "Notice that both MTDA-ITA and 1p-DSN reduces the domain mismatch for the shared features (circle", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 574, + 522, + 586 + ], + "spans": [ + { + "bbox": [ + 88, + 574, + 522, + 586 + ], + "score": 1.0, + "content": "markers in fig. 5) and separate the shared features from private features. On the other hand, MTDA-ITA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 586, + 522, + 598 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 522, + 598 + ], + "score": 1.0, + "content": "increases the domain separation for the private features (triangle markers, pure and well-separated domain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "spans": [ + { + "bbox": [ + 89, + 596, + 522, + 609 + ], + "score": 1.0, + "content": "clusters in fig. 5c) while 1p-DSN is unable to enforce the private representation of different domains to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "spans": [ + { + "bbox": [ + 89, + 606, + 522, + 621 + ], + "score": 1.0, + "content": "be different (fig. 5e) that may result in redundancy of different private spaces. This is partially due to the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 618, + 523, + 631 + ], + "spans": [ + { + "bbox": [ + 88, + 618, + 397, + 631 + ], + "score": 1.0, + "content": "proposed multi-domain separation loss through the use of the domain classifier", + "type": "text" + }, + { + "bbox": [ + 398, + 619, + 407, + 628 + ], + "score": 0.76, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 618, + 523, + 631 + ], + "score": 1.0, + "content": ", which penalizes the domain", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 629, + 521, + 642 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 521, + 642 + ], + "score": 1.0, + "content": "mismatch for the shared features and rewards the mismatch for the private features, something the 1p-DSN", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "spans": [ + { + "bbox": [ + 88, + 639, + 522, + 653 + ], + "score": 1.0, + "content": "fails to account for. Moreover, as supported by the quantitative results in tab. 1, the class label separation in", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 88, + 564, + 523, + 653 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 73, + 519, + 213 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 73, + 519, + 213 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 90, + 73, + 519, + 213 + ], + "spans": [ + { + "bbox": [ + 90, + 73, + 519, + 213 + ], + "score": 0.982, + "html": "
methodC08→C05C08→C09C08→C13C08→C14C09→C05C09→C08C09→C13C09→C14
Source Only33.70±0.3350.10±0.2850.80±0.3240.13±0.2633.32±0.4448.24±0.3249.24±0.3036.19±0.27
1-NN28.7535.3939.7932.1326.8235.3034.2628.41
CORAL (Sun & Saenko (2016))35.89 ± 0.4455.79±0.5060.00 ±0.2940.67 ± 0.4835.89 ± 0.4051.56±0.4350.45 ± 0.4140.67± 0.35
DANN (Ganin et al. (2016))40.20±0.5056.89 ± 0.3955.83 ±0.4043.25±0.4150.63±0.3858.40 ± 0.5155.81 ± 0.5348.90 ±0.43
ADDA(Tzeng et al. (2017))37.40±0.6858.40±0.7360.40±0.8342.10±0.4829.40±0.7053.30±0.4945.30± 0.5338.30±0.63
c-ADDA41.60±0.6439.65 ±0.7050.00 ±0.5246.25±0.5245.01±0.6352.14 ± 0.5337.43 ± 0.6043.26±0.58
DTN (Zhang et al. (2015))44.13 ± 0.4157.42 ± 0.4255.89±0.4845.76± 0.3944.53± 0.4957.34±0.3552.43±0.3851.55 ± 0.40
c-DTN45.10 ±0.4449.78±0.5047.43 ±0.4645.79 ±0.4849.80±0.4055.69 ± 0.3550.10 ± 0.3852.31 ±0.29
PixelDA (Bousmalis et al. (2017))46.45 ± 0.4544.33±0.3844.87 ± 0.4146.83±0.2945.63±0.3416.37 ± 0.2745.43 ± 0.3547.00± 0.49
UNIT (Liu et al. (2017))43.88±0.1843.99±0.2344.47±0.1944.47 ± 0.2444.47 ± 0.1743.95± 0.2144.64±0.2244.47 ±0.19
DSN (Bousmalis et al. (2016)) c-DSN46.25± 0.5347.50±0.6062.15 ± 0.5839.72± 0.5545.85±0.4856.65±0.5056.5±0.3842.87±0.43
45.82 ± 0.5344.64± 0.4245.60 ± 0.4846.32±0.5245.18 ± 0.4745.52 ± 0.5544.79 ± 0.5347.37 ± 0.48
1p-DSN44.12 ± 0.7344.14 ± 0.2045.00 ±0.3845.62 ± 0.4244.78 ±0.4745.02 ±0.6544.21 ± 0.4846.97 ± 0.38
mp-DSN42.19±0.4644.70±0.5342.47 ± 0.4840.50±0.3945.00±0.5143.80±0.5045.79±0.4842.39 ±0.49
s-MTDA-ITA44.77±0.1945.61±0.1860.00±0.2746.70±0.2849.06±0.2455.33±0.2259.90±0.3050.64± 0.26
c-MTDA-ITA MTDA-ITA44.35± 0.2742.67 ± 0.2458.90 ±0.2644.32 ± 0.2646.74 ±0.2254.11 ± 0.2156.89 ± 0.2349.64 ± 0.19
46.30 ± 0.2560.60 ± 0.1860.50±0.1950.40±0.2055.59± 0.2557.80±0.2164.20 ±0.1856.34 ± 0.20
", + "type": "table", + "image_path": "c85b4ef391c96ca0791a2ae020ce98ec467cb561c381b65c00be093640811879.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 90, + 73, + 519, + 119.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 90, + 119.66666666666666, + 519, + 166.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 90, + 166.33333333333331, + 519, + 212.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 91, + 243, + 520, + 317 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 182, + 230, + 427, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 182, + 228, + 428, + 242 + ], + "spans": [ + { + "bbox": [ + 182, + 228, + 428, + 242 + ], + "score": 1.0, + "content": "Table 6: Classification results on Multi-PIE dataset. The best (red).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 91, + 243, + 520, + 317 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 91, + 244, + 520, + 316 + ], + "spans": [ + { + "bbox": [ + 91, + 244, + 520, + 316 + ], + "score": 0.977, + "html": "
methodP→AP→CP→SA→PA→CA→S
1-NN15.2818.1625.6022.7019.7522.70
ADDA(Tzeng et al. (2017))24.35 ± 2.3720.12 ± 2.5022.45 ± 2.1132.57 ± 2.7017.68 ± 2.0418.90 ± 2.48
DSN (Bousmalis et al. (2016))28.42 ±2.1221.14 ± 2.082.04 ± 1.9029.54 ± 1.9525.89 ± 1.8824.69 ± 2.08
s-MTDA-ITA28.02 ± 1.5921.64±1.2426.24 ± 1.6031.06 ± 1.5025.09±1.4025.89±1.03
c-MTDA-ITA25.35 ± 1.8020.24 ± 1.3923.64 ± 1.6026.54 ± 1.3320.30 ± 1.2922.38 ± 1.45
MTDA-ITA31.40 ±1.5523.05 ±1.0428.24 ±1.7835.74 ± 1.5027.00 ±1.2528.90 ±1.60
", + "type": "table", + "image_path": "5010897f659b60be96fdf8d14f7bef3429234f0e1fd20936d49e7805613ab995.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 91, + 243, + 520, + 267.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 91, + 267.6666666666667, + 520, + 292.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 91, + 292.33333333333337, + 520, + 317.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 88, + 331, + 520, + 352 + ], + "lines": [ + { + "bbox": [ + 88, + 330, + 522, + 343 + ], + "spans": [ + { + "bbox": [ + 88, + 330, + 522, + 343 + ], + "score": 1.0, + "content": "Table 7: Classification results on PACS dataset classification. A:Art-painting, C:Cartoon, S:Sketch, P:Photo. The best", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 340, + 114, + 354 + ], + "spans": [ + { + "bbox": [ + 88, + 340, + 114, + 354 + ], + "score": 1.0, + "content": "(red).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 88, + 358, + 521, + 392 + ], + "lines": [ + { + "bbox": [ + 90, + 358, + 522, + 371 + ], + "spans": [ + { + "bbox": [ + 90, + 358, + 522, + 371 + ], + "score": 1.0, + "content": "the shared space for 1p-DSN, fig. 5f, is still evident but not as strong as in the MTDA-ITA, fig. 5d. This can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 369, + 521, + 382 + ], + "spans": [ + { + "bbox": [ + 90, + 369, + 521, + 382 + ], + "score": 1.0, + "content": "be attributed to the lack of redundancy in the private space that helps MTDA-ITA to learn more disentangled", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 380, + 519, + 394 + ], + "spans": [ + { + "bbox": [ + 90, + 380, + 519, + 394 + ], + "score": 1.0, + "content": "shared features and usage of the target samples during training, something the 1p-DSN fails to account for.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 284, + 671, + 295, + 680 + ], + "lines": [ + { + "bbox": [ + 282, + 669, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 282, + 669, + 297, + 684 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 73, + 519, + 213 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 90, + 73, + 519, + 213 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 90, + 73, + 519, + 213 + ], + "spans": [ + { + "bbox": [ + 90, + 73, + 519, + 213 + ], + "score": 0.982, + "html": "
methodC08→C05C08→C09C08→C13C08→C14C09→C05C09→C08C09→C13C09→C14
Source Only33.70±0.3350.10±0.2850.80±0.3240.13±0.2633.32±0.4448.24±0.3249.24±0.3036.19±0.27
1-NN28.7535.3939.7932.1326.8235.3034.2628.41
CORAL (Sun & Saenko (2016))35.89 ± 0.4455.79±0.5060.00 ±0.2940.67 ± 0.4835.89 ± 0.4051.56±0.4350.45 ± 0.4140.67± 0.35
DANN (Ganin et al. (2016))40.20±0.5056.89 ± 0.3955.83 ±0.4043.25±0.4150.63±0.3858.40 ± 0.5155.81 ± 0.5348.90 ±0.43
ADDA(Tzeng et al. (2017))37.40±0.6858.40±0.7360.40±0.8342.10±0.4829.40±0.7053.30±0.4945.30± 0.5338.30±0.63
c-ADDA41.60±0.6439.65 ±0.7050.00 ±0.5246.25±0.5245.01±0.6352.14 ± 0.5337.43 ± 0.6043.26±0.58
DTN (Zhang et al. (2015))44.13 ± 0.4157.42 ± 0.4255.89±0.4845.76± 0.3944.53± 0.4957.34±0.3552.43±0.3851.55 ± 0.40
c-DTN45.10 ±0.4449.78±0.5047.43 ±0.4645.79 ±0.4849.80±0.4055.69 ± 0.3550.10 ± 0.3852.31 ±0.29
PixelDA (Bousmalis et al. (2017))46.45 ± 0.4544.33±0.3844.87 ± 0.4146.83±0.2945.63±0.3416.37 ± 0.2745.43 ± 0.3547.00± 0.49
UNIT (Liu et al. (2017))43.88±0.1843.99±0.2344.47±0.1944.47 ± 0.2444.47 ± 0.1743.95± 0.2144.64±0.2244.47 ±0.19
DSN (Bousmalis et al. (2016)) c-DSN46.25± 0.5347.50±0.6062.15 ± 0.5839.72± 0.5545.85±0.4856.65±0.5056.5±0.3842.87±0.43
45.82 ± 0.5344.64± 0.4245.60 ± 0.4846.32±0.5245.18 ± 0.4745.52 ± 0.5544.79 ± 0.5347.37 ± 0.48
1p-DSN44.12 ± 0.7344.14 ± 0.2045.00 ±0.3845.62 ± 0.4244.78 ±0.4745.02 ±0.6544.21 ± 0.4846.97 ± 0.38
mp-DSN42.19±0.4644.70±0.5342.47 ± 0.4840.50±0.3945.00±0.5143.80±0.5045.79±0.4842.39 ±0.49
s-MTDA-ITA44.77±0.1945.61±0.1860.00±0.2746.70±0.2849.06±0.2455.33±0.2259.90±0.3050.64± 0.26
c-MTDA-ITA MTDA-ITA44.35± 0.2742.67 ± 0.2458.90 ±0.2644.32 ± 0.2646.74 ±0.2254.11 ± 0.2156.89 ± 0.2349.64 ± 0.19
46.30 ± 0.2560.60 ± 0.1860.50±0.1950.40±0.2055.59± 0.2557.80±0.2164.20 ±0.1856.34 ± 0.20
", + "type": "table", + "image_path": "c85b4ef391c96ca0791a2ae020ce98ec467cb561c381b65c00be093640811879.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 90, + 73, + 519, + 119.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 90, + 119.66666666666666, + 519, + 166.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 90, + 166.33333333333331, + 519, + 212.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 91, + 243, + 520, + 317 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 182, + 230, + 427, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 182, + 228, + 428, + 242 + ], + "spans": [ + { + "bbox": [ + 182, + 228, + 428, + 242 + ], + "score": 1.0, + "content": "Table 6: Classification results on Multi-PIE dataset. The best (red).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 91, + 243, + 520, + 317 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 91, + 244, + 520, + 316 + ], + "spans": [ + { + "bbox": [ + 91, + 244, + 520, + 316 + ], + "score": 0.977, + "html": "
methodP→AP→CP→SA→PA→CA→S
1-NN15.2818.1625.6022.7019.7522.70
ADDA(Tzeng et al. (2017))24.35 ± 2.3720.12 ± 2.5022.45 ± 2.1132.57 ± 2.7017.68 ± 2.0418.90 ± 2.48
DSN (Bousmalis et al. (2016))28.42 ±2.1221.14 ± 2.082.04 ± 1.9029.54 ± 1.9525.89 ± 1.8824.69 ± 2.08
s-MTDA-ITA28.02 ± 1.5921.64±1.2426.24 ± 1.6031.06 ± 1.5025.09±1.4025.89±1.03
c-MTDA-ITA25.35 ± 1.8020.24 ± 1.3923.64 ± 1.6026.54 ± 1.3320.30 ± 1.2922.38 ± 1.45
MTDA-ITA31.40 ±1.5523.05 ±1.0428.24 ±1.7835.74 ± 1.5027.00 ±1.2528.90 ±1.60
", + "type": "table", + "image_path": "5010897f659b60be96fdf8d14f7bef3429234f0e1fd20936d49e7805613ab995.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 91, + 243, + 520, + 267.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 91, + 267.6666666666667, + 520, + 292.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 91, + 292.33333333333337, + 520, + 317.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 88, + 331, + 520, + 352 + ], + "lines": [ + { + "bbox": [ + 88, + 330, + 522, + 343 + ], + "spans": [ + { + "bbox": [ + 88, + 330, + 522, + 343 + ], + "score": 1.0, + "content": "Table 7: Classification results on PACS dataset classification. A:Art-painting, C:Cartoon, S:Sketch, P:Photo. The best", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 340, + 114, + 354 + ], + "spans": [ + { + "bbox": [ + 88, + 340, + 114, + 354 + ], + "score": 1.0, + "content": "(red).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 88, + 330, + 522, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 88, + 358, + 521, + 392 + ], + "lines": [ + { + "bbox": [ + 90, + 358, + 522, + 371 + ], + "spans": [ + { + "bbox": [ + 90, + 358, + 522, + 371 + ], + "score": 1.0, + "content": "the shared space for 1p-DSN, fig. 5f, is still evident but not as strong as in the MTDA-ITA, fig. 5d. This can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 369, + 521, + 382 + ], + "spans": [ + { + "bbox": [ + 90, + 369, + 521, + 382 + ], + "score": 1.0, + "content": "be attributed to the lack of redundancy in the private space that helps MTDA-ITA to learn more disentangled", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 380, + 519, + 394 + ], + "spans": [ + { + "bbox": [ + 90, + 380, + 519, + 394 + ], + "score": 1.0, + "content": "shared features and usage of the target samples during training, something the 1p-DSN fails to account for.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 90, + 358, + 522, + 394 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 81, + 447, + 645 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 161, + 81, + 447, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 161, + 81, + 447, + 645 + ], + "spans": [ + { + "bbox": [ + 161, + 81, + 447, + 645 + ], + "score": 0.977, + "type": "image", + "image_path": "f7a643eee7751386eefa467035b5bfae216f67ec9df6317cb9b21d6dd8832516.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 161, + 81, + 447, + 93.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 161, + 93.0, + 447, + 105.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 161, + 105.0, + 447, + 117.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 161, + 117.0, + 447, + 129.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 161, + 129.0, + 447, + 141.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 161, + 141.0, + 447, + 153.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 161, + 153.0, + 447, + 165.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 161, + 165.0, + 447, + 177.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 177.0, + 447, + 189.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 161, + 189.0, + 447, + 201.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 161, + 201.0, + 447, + 213.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 161, + 213.0, + 447, + 225.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 161, + 225.0, + 447, + 237.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 161, + 237.0, + 447, + 249.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 161, + 249.0, + 447, + 261.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 161, + 261.0, + 447, + 273.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 161, + 273.0, + 447, + 285.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 161, + 285.0, + 447, + 297.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 161, + 297.0, + 447, + 309.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 161, + 309.0, + 447, + 321.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 161, + 321.0, + 447, + 333.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 161, + 333.0, + 447, + 345.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 161, + 345.0, + 447, + 357.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 161, + 357.0, + 447, + 369.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 161, + 369.0, + 447, + 381.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 161, + 381.0, + 447, + 393.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 161, + 393.0, + 447, + 405.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 161, + 405.0, + 447, + 417.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 161, + 417.0, + 447, + 429.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 161, + 429.0, + 447, + 441.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 161, + 441.0, + 447, + 453.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 161, + 453.0, + 447, + 465.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 161, + 465.0, + 447, + 477.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 161, + 477.0, + 447, + 489.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 161, + 489.0, + 447, + 501.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 161, + 501.0, + 447, + 513.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 161, + 513.0, + 447, + 525.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 161, + 525.0, + 447, + 537.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 161, + 537.0, + 447, + 549.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 161, + 549.0, + 447, + 561.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 161, + 561.0, + 447, + 573.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 161, + 573.0, + 447, + 585.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 161, + 585.0, + 447, + 597.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 161, + 597.0, + 447, + 609.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 161, + 609.0, + 447, + 621.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 161, + 621.0, + 447, + 633.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 161, + 633.0, + 447, + 645.0 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 649, + 523, + 719 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 649, + 522, + 660 + ], + "spans": [ + { + "bbox": [ + 90, + 649, + 522, + 660 + ], + "score": 1.0, + "content": "Figure 6: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 89, + 658, + 523, + 671 + ], + "spans": [ + { + "bbox": [ + 89, + 658, + 523, + 671 + ], + "score": 1.0, + "content": "show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 90, + 669, + 522, + 681 + ], + "spans": [ + { + "bbox": [ + 90, + 669, + 522, + 681 + ], + "score": 1.0, + "content": "(c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters21", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 89, + 679, + 522, + 691 + ], + "spans": [ + { + "bbox": [ + 89, + 679, + 522, + 691 + ], + "score": 1.0, + "content": "in the right column represent points from the shared space, while the smaller ones are from the private spaces. The", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 89, + 689, + 521, + 700 + ], + "spans": [ + { + "bbox": [ + 89, + 689, + 521, + 700 + ], + "score": 1.0, + "content": "remaining figures depict the learned features without: (e),(f) the classifier entropy loss, MTDA-woE; (g),(h) the private", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 89, + 697, + 523, + 712 + ], + "spans": [ + { + "bbox": [ + 89, + 697, + 523, + 712 + ], + "score": 1.0, + "content": "encoder, MTDA-woP; (i),(j) the reconstruction loss/decoder, MTDA-woR; and (k),(l) the multi-domain separation loss,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 89, + 708, + 142, + 720 + ], + "spans": [ + { + "bbox": [ + 89, + 708, + 142, + 720 + ], + "score": 1.0, + "content": "MTDA-woD.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50 + } + ], + "index": 36.5 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 90, + 40, + 292, + 50 + ], + "lines": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "spans": [ + { + "bbox": [ + 90, + 39, + 292, + 52 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 81, + 447, + 645 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 161, + 81, + 447, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 161, + 81, + 447, + 645 + ], + "spans": [ + { + "bbox": [ + 161, + 81, + 447, + 645 + ], + "score": 0.977, + "type": "image", + "image_path": "f7a643eee7751386eefa467035b5bfae216f67ec9df6317cb9b21d6dd8832516.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 161, + 81, + 447, + 93.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 161, + 93.0, + 447, + 105.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 161, + 105.0, + 447, + 117.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 161, + 117.0, + 447, + 129.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 161, + 129.0, + 447, + 141.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 161, + 141.0, + 447, + 153.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 161, + 153.0, + 447, + 165.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 161, + 165.0, + 447, + 177.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 177.0, + 447, + 189.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 161, + 189.0, + 447, + 201.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 161, + 201.0, + 447, + 213.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 161, + 213.0, + 447, + 225.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 161, + 225.0, + 447, + 237.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 161, + 237.0, + 447, + 249.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 161, + 249.0, + 447, + 261.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 161, + 261.0, + 447, + 273.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 161, + 273.0, + 447, + 285.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 161, + 285.0, + 447, + 297.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 161, + 297.0, + 447, + 309.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 161, + 309.0, + 447, + 321.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 161, + 321.0, + 447, + 333.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 161, + 333.0, + 447, + 345.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 161, + 345.0, + 447, + 357.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 161, + 357.0, + 447, + 369.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 161, + 369.0, + 447, + 381.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 161, + 381.0, + 447, + 393.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 161, + 393.0, + 447, + 405.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 161, + 405.0, + 447, + 417.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 161, + 417.0, + 447, + 429.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 161, + 429.0, + 447, + 441.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 161, + 441.0, + 447, + 453.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 161, + 453.0, + 447, + 465.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 161, + 465.0, + 447, + 477.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 161, + 477.0, + 447, + 489.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 161, + 489.0, + 447, + 501.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 161, + 501.0, + 447, + 513.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 161, + 513.0, + 447, + 525.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 161, + 525.0, + 447, + 537.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 161, + 537.0, + 447, + 549.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 161, + 549.0, + 447, + 561.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 161, + 561.0, + 447, + 573.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 161, + 573.0, + 447, + 585.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 161, + 585.0, + 447, + 597.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 161, + 597.0, + 447, + 609.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 161, + 609.0, + 447, + 621.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 161, + 621.0, + 447, + 633.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 161, + 633.0, + 447, + 645.0 + ], + "spans": [], + "index": 46 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 90, + 649, + 523, + 719 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 90, + 649, + 522, + 660 + ], + "spans": [ + { + "bbox": [ + 90, + 649, + 522, + 660 + ], + "score": 1.0, + "content": "Figure 6: Feature visualization for embedding of digit datasets using t-SNE algorithm. The first and the second columns", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 89, + 658, + 523, + 671 + ], + "spans": [ + { + "bbox": [ + 89, + 658, + 523, + 671 + ], + "score": 1.0, + "content": "show the domains and classes, respectively, with color indicating domain and class membership. (a),(b) Original features.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 90, + 669, + 522, + 681 + ], + "spans": [ + { + "bbox": [ + 90, + 669, + 522, + 681 + ], + "score": 1.0, + "content": "(c),(d) learned features for MTDA-ITA (triangle marker: private features, circle marker: shared features). Large clusters21", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 89, + 679, + 522, + 691 + ], + "spans": [ + { + "bbox": [ + 89, + 679, + 522, + 691 + ], + "score": 1.0, + "content": "in the right column represent points from the shared space, while the smaller ones are from the private spaces. The", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 89, + 689, + 521, + 700 + ], + "spans": [ + { + "bbox": [ + 89, + 689, + 521, + 700 + ], + "score": 1.0, + "content": "remaining figures depict the learned features without: (e),(f) the classifier entropy loss, MTDA-woE; (g),(h) the private", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 89, + 697, + 523, + 712 + ], + "spans": [ + { + "bbox": [ + 89, + 697, + 523, + 712 + ], + "score": 1.0, + "content": "encoder, MTDA-woP; (i),(j) the reconstruction loss/decoder, MTDA-woR; and (k),(l) the multi-domain separation loss,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 89, + 708, + 142, + 720 + ], + "spans": [ + { + "bbox": [ + 89, + 708, + 142, + 720 + ], + "score": 1.0, + "content": "MTDA-woD.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50 + } + ], + "index": 36.5 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/BJxLH2AcYX/BJxLH2AcYX_model.json b/parse/train/BJxLH2AcYX/BJxLH2AcYX_model.json new file mode 100644 index 0000000000000000000000000000000000000000..58e40644e22ed4d9252147079aaa03111d1dcdc7 --- /dev/null +++ b/parse/train/BJxLH2AcYX/BJxLH2AcYX_model.json @@ -0,0 +1,28518 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 354, + 621, + 1349, + 621, + 1349, + 1077, + 354, + 1077 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 253, + 1197, + 1449, + 1197, + 1449, + 1472, + 253, + 1472 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 253, + 1488, + 1447, + 1488, + 1447, + 1673, + 253, + 1673 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 254, + 1688, + 1447, + 1688, + 1447, + 1811, + 254, + 1811 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 255, + 254, + 1452, + 254, + 1452, + 361, + 255, + 361 + ], + "score": 0.962 + }, + { + "category_id": 2, + "poly": [ + 253, + 112, + 812, + 112, + 812, + 141, + 253, + 141 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 257, + 1130, + 529, + 1130, + 529, + 1165, + 257, + 1165 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 773, + 555, + 927, + 555, + 927, + 589, + 773, + 589 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 261, + 414, + 625, + 414, + 625, + 474, + 261, + 474 + ], + "score": 0.779 + }, + { + "category_id": 2, + "poly": [ + 796, + 1866, + 810, + 1866, + 810, + 1889, + 796, + 1889 + ], + "score": 0.752 + }, + { + "category_id": 15, + "poly": [ + 250.0, + 253.0, + 1457.0, + 253.0, + 1457.0, + 309.0, + 250.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 311.0, + 1121.0, + 311.0, + 1121.0, + 364.0, + 248.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 109.0, + 813.0, + 109.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1126.0, + 534.0, + 1126.0, + 534.0, + 1173.0, + 249.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 554.0, + 932.0, + 554.0, + 932.0, + 592.0, + 770.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1865.0, + 815.0, + 1865.0, + 815.0, + 1895.0, + 794.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 618.0, + 1350.0, + 618.0, + 1350.0, + 657.0, + 350.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 650.0, + 1351.0, + 650.0, + 1351.0, + 688.0, + 347.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 683.0, + 1352.0, + 683.0, + 1352.0, + 716.0, + 350.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 712.0, + 1351.0, + 712.0, + 1351.0, + 748.0, + 350.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 742.0, + 1350.0, + 742.0, + 1350.0, + 777.0, + 349.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 771.0, + 1351.0, + 771.0, + 1351.0, + 807.0, + 349.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 805.0, + 1351.0, + 805.0, + 1351.0, + 837.0, + 350.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 835.0, + 1349.0, + 835.0, + 1349.0, + 868.0, + 350.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 866.0, + 1349.0, + 866.0, + 1349.0, + 899.0, + 350.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 896.0, + 1351.0, + 896.0, + 1351.0, + 929.0, + 350.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 927.0, + 1350.0, + 927.0, + 1350.0, + 959.0, + 350.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 953.0, + 1352.0, + 953.0, + 1352.0, + 990.0, + 347.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 986.0, + 1352.0, + 986.0, + 1352.0, + 1022.0, + 348.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1017.0, + 1350.0, + 1017.0, + 1350.0, + 1053.0, + 348.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1048.0, + 846.0, + 1048.0, + 846.0, + 1081.0, + 351.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1199.0, + 1449.0, + 1199.0, + 1449.0, + 1232.0, + 250.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1225.0, + 1452.0, + 1225.0, + 1452.0, + 1267.0, + 245.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1260.0, + 1450.0, + 1260.0, + 1450.0, + 1293.0, + 250.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1285.0, + 1452.0, + 1285.0, + 1452.0, + 1328.0, + 246.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1320.0, + 1452.0, + 1320.0, + 1452.0, + 1357.0, + 249.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1347.0, + 1453.0, + 1347.0, + 1453.0, + 1387.0, + 248.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1377.0, + 1452.0, + 1377.0, + 1452.0, + 1418.0, + 248.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1406.0, + 1456.0, + 1406.0, + 1456.0, + 1450.0, + 246.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1443.0, + 874.0, + 1443.0, + 874.0, + 1478.0, + 249.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1488.0, + 1448.0, + 1488.0, + 1448.0, + 1519.0, + 250.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1518.0, + 1451.0, + 1518.0, + 1451.0, + 1554.0, + 247.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1550.0, + 1450.0, + 1550.0, + 1450.0, + 1582.0, + 250.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1582.0, + 1450.0, + 1582.0, + 1450.0, + 1614.0, + 250.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1610.0, + 1451.0, + 1610.0, + 1451.0, + 1645.0, + 249.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1639.0, + 1171.0, + 1639.0, + 1171.0, + 1677.0, + 245.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1687.0, + 1451.0, + 1687.0, + 1451.0, + 1726.0, + 248.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1716.0, + 1447.0, + 1716.0, + 1447.0, + 1751.0, + 250.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1748.0, + 1451.0, + 1748.0, + 1451.0, + 1786.0, + 247.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1777.0, + 1450.0, + 1777.0, + 1450.0, + 1816.0, + 250.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 416.0, + 505.0, + 416.0, + 505.0, + 445.0, + 262.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 256.0, + 443.0, + 629.0, + 443.0, + 629.0, + 476.0, + 256.0, + 476.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 252, + 725, + 1448, + 725, + 1448, + 1399, + 252, + 1399 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 253, + 1413, + 1450, + 1413, + 1450, + 1811, + 253, + 1811 + ], + "score": 0.981 + }, + { + "category_id": 3, + "poly": [ + 360, + 234, + 1334, + 234, + 1334, + 582, + 360, + 582 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 252, + 607, + 1448, + 607, + 1448, + 691, + 252, + 691 + ], + "score": 0.962 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 796, + 1866, + 813, + 1866, + 813, + 1889, + 796, + 1889 + ], + "score": 0.78 + }, + { + "category_id": 13, + "poly": [ + 1133, + 244, + 1157, + 244, + 1157, + 260, + 1133, + 260 + ], + "score": 0.37, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 439, + 370, + 469, + 370, + 469, + 390, + 439, + 390 + ], + "score": 0.37, + "latex": "D _ { 0 }" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 233.0, + 658.0, + 233.0, + 658.0, + 266.0, + 622.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 234.0, + 1132.0, + 234.0, + 1132.0, + 266.0, + 1128.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 234.0, + 1164.0, + 234.0, + 1164.0, + 266.0, + 1158.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 277.0, + 580.0, + 277.0, + 580.0, + 316.0, + 466.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 311.0, + 1061.0, + 311.0, + 1061.0, + 344.0, + 943.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 356.0, + 438.0, + 356.0, + 438.0, + 400.0, + 433.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 356.0, + 558.0, + 356.0, + 558.0, + 400.0, + 470.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 347.0, + 776.0, + 347.0, + 776.0, + 392.0, + 654.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 333.0, + 1280.0, + 333.0, + 1280.0, + 370.0, + 1165.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 388.0, + 548.0, + 388.0, + 548.0, + 413.0, + 508.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 390.0, + 1005.0, + 390.0, + 1005.0, + 436.0, + 870.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 461.0, + 714.0, + 461.0, + 714.0, + 499.0, + 598.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 440.0, + 1188.0, + 440.0, + 1188.0, + 484.0, + 1071.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 546.0, + 802.0, + 546.0, + 802.0, + 587.0, + 383.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 547.0, + 1311.0, + 547.0, + 1311.0, + 585.0, + 886.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 375.5, + 705.0, + 375.5, + 705.0, + 411.5, + 633.0, + 411.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 607.0, + 1448.0, + 607.0, + 1448.0, + 639.0, + 251.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 632.0, + 1451.0, + 632.0, + 1451.0, + 671.0, + 248.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 665.0, + 349.0, + 665.0, + 349.0, + 694.0, + 251.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1863.0, + 817.0, + 1863.0, + 817.0, + 1898.0, + 794.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 729.0, + 1451.0, + 729.0, + 1451.0, + 762.0, + 249.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 760.0, + 1451.0, + 760.0, + 1451.0, + 793.0, + 249.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 784.0, + 1452.0, + 784.0, + 1452.0, + 828.0, + 245.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 820.0, + 1451.0, + 820.0, + 1451.0, + 854.0, + 249.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 850.0, + 1451.0, + 850.0, + 1451.0, + 883.0, + 249.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 875.0, + 1453.0, + 875.0, + 1453.0, + 920.0, + 245.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 910.0, + 1453.0, + 910.0, + 1453.0, + 948.0, + 248.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 938.0, + 1452.0, + 938.0, + 1452.0, + 977.0, + 247.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 971.0, + 1452.0, + 971.0, + 1452.0, + 1007.0, + 247.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1003.0, + 1451.0, + 1003.0, + 1451.0, + 1036.0, + 248.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1032.0, + 1452.0, + 1032.0, + 1452.0, + 1066.0, + 249.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1061.0, + 1452.0, + 1061.0, + 1452.0, + 1098.0, + 249.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1094.0, + 1451.0, + 1094.0, + 1451.0, + 1128.0, + 249.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1121.0, + 1452.0, + 1121.0, + 1452.0, + 1160.0, + 247.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1152.0, + 1452.0, + 1152.0, + 1452.0, + 1191.0, + 247.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1184.0, + 1452.0, + 1184.0, + 1452.0, + 1221.0, + 247.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1214.0, + 1452.0, + 1214.0, + 1452.0, + 1251.0, + 248.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1245.0, + 1451.0, + 1245.0, + 1451.0, + 1282.0, + 248.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1277.0, + 1451.0, + 1277.0, + 1451.0, + 1311.0, + 249.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1307.0, + 1451.0, + 1307.0, + 1451.0, + 1340.0, + 249.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1333.0, + 1451.0, + 1333.0, + 1451.0, + 1374.0, + 248.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1367.0, + 1443.0, + 1367.0, + 1443.0, + 1405.0, + 249.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1414.0, + 1454.0, + 1414.0, + 1454.0, + 1450.0, + 249.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1443.0, + 1450.0, + 1443.0, + 1450.0, + 1481.0, + 248.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1472.0, + 1451.0, + 1472.0, + 1451.0, + 1513.0, + 248.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1504.0, + 1450.0, + 1504.0, + 1450.0, + 1541.0, + 249.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1535.0, + 1451.0, + 1535.0, + 1451.0, + 1572.0, + 248.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1564.0, + 1453.0, + 1564.0, + 1453.0, + 1603.0, + 248.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1598.0, + 1450.0, + 1598.0, + 1450.0, + 1630.0, + 249.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1628.0, + 1451.0, + 1628.0, + 1451.0, + 1664.0, + 249.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1657.0, + 1451.0, + 1657.0, + 1451.0, + 1694.0, + 249.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1689.0, + 1453.0, + 1689.0, + 1453.0, + 1722.0, + 249.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1720.0, + 1450.0, + 1720.0, + 1450.0, + 1752.0, + 250.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1748.0, + 1450.0, + 1748.0, + 1450.0, + 1784.0, + 250.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1779.0, + 1451.0, + 1779.0, + 1451.0, + 1816.0, + 249.0, + 1816.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 251, + 444, + 1451, + 444, + 1451, + 606, + 251, + 606 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 251, + 1316, + 1449, + 1316, + 1449, + 1443, + 251, + 1443 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 253, + 1686, + 1449, + 1686, + 1449, + 1816, + 253, + 1816 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 254, + 869, + 1449, + 869, + 1449, + 993, + 254, + 993 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 253, + 620, + 1447, + 620, + 1447, + 712, + 253, + 712 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 323, + 1018, + 1449, + 1018, + 1449, + 1292, + 323, + 1292 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 250, + 1537, + 1448, + 1537, + 1448, + 1614, + 250, + 1614 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 249, + 265, + 1450, + 265, + 1450, + 330, + 249, + 330 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 623, + 1630, + 1078, + 1630, + 1078, + 1672, + 623, + 1672 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 537, + 713, + 1157, + 713, + 1157, + 750, + 537, + 750 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 419, + 810, + 1277, + 810, + 1277, + 851, + 419, + 851 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 253, + 757, + 897, + 757, + 897, + 791, + 253, + 791 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 253, + 111, + 813, + 111, + 813, + 142, + 253, + 142 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 255, + 374, + 666, + 374, + 666, + 411, + 255, + 411 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 253, + 1479, + 500, + 1479, + 500, + 1512, + 253, + 1512 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1411, + 1634, + 1444, + 1634, + 1444, + 1664, + 1411, + 1664 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1411, + 715, + 1445, + 715, + 1445, + 746, + 1411, + 746 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1411, + 815, + 1445, + 815, + 1445, + 845, + 1411, + 845 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 796, + 1865, + 814, + 1865, + 814, + 1890, + 796, + 1890 + ], + "score": 0.792 + }, + { + "category_id": 13, + "poly": [ + 1002, + 621, + 1149, + 621, + 1149, + 656, + 1002, + 656 + ], + "score": 0.94, + "latex": "\\boldsymbol { z } = [ z _ { s } , z _ { p } ]" + }, + { + "category_id": 13, + "poly": [ + 1157, + 508, + 1396, + 508, + 1396, + 544, + 1157, + 544 + ], + "score": 0.93, + "latex": "\\pmb { y } _ { i } = [ y _ { i } ^ { 1 } , y _ { i } ^ { 2 } , . . . , y _ { i } ^ { K } ]" + }, + { + "category_id": 13, + "poly": [ + 446, + 475, + 799, + 475, + 799, + 512, + 446, + 512 + ], + "score": 0.93, + "latex": "( \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } ) = \\{ ( \\pmb { x } _ { i } , \\pmb { y } _ { i } , \\pmb { d } _ { i } ) \\} _ { i = 0 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 253, + 1749, + 571, + 1749, + 571, + 1783, + 253, + 1783 + ], + "score": 0.93, + "latex": "p ( \\pmb { x } | \\pmb { z } ) = p ( \\pmb { z } | \\pmb { x } ) p ( \\pmb { x } ) / p ( \\pmb { z } )" + }, + { + "category_id": 13, + "poly": [ + 253, + 901, + 339, + 901, + 339, + 934, + 253, + 934 + ], + "score": 0.92, + "latex": "I ( { \\pmb x } ; { \\pmb y } )" + }, + { + "category_id": 13, + "poly": [ + 974, + 1719, + 1082, + 1719, + 1082, + 1752, + 974, + 1752 + ], + "score": 0.92, + "latex": "q ( { \\pmb x } | { \\pmb z } ; \\phi )" + }, + { + "category_id": 13, + "poly": [ + 623, + 1379, + 727, + 1379, + 727, + 1413, + 623, + 1413 + ], + "score": 0.92, + "latex": "p ( z _ { s } , z _ { p } )" + }, + { + "category_id": 13, + "poly": [ + 253, + 542, + 485, + 542, + 485, + 577, + 253, + 577 + ], + "score": 0.92, + "latex": "d _ { i } = [ d _ { i } ^ { 1 } , d _ { i } ^ { 2 } , . . . , d _ { i } ^ { M } ]" + }, + { + "category_id": 13, + "poly": [ + 430, + 870, + 485, + 870, + 485, + 903, + 430, + 903 + ], + "score": 0.92, + "latex": "p ( d )" + }, + { + "category_id": 13, + "poly": [ + 324, + 870, + 380, + 870, + 380, + 903, + 324, + 903 + ], + "score": 0.92, + "latex": "p ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 327, + 1689, + 394, + 1689, + 394, + 1721, + 327, + 1721 + ], + "score": 0.92, + "latex": "H ( { \\pmb x } )" + }, + { + "category_id": 14, + "poly": [ + 620, + 1630, + 1079, + 1630, + 1079, + 1671, + 620, + 1671 + ], + "score": 0.92, + "latex": "I ( \\pmb { x } ; z ) \\geq H ( \\pmb { x } ) + \\mathbb { E } _ { p ( \\pmb { x } , z ) } [ \\ln q ( \\pmb { x } | z ; \\phi ) ]" + }, + { + "category_id": 13, + "poly": [ + 948, + 1409, + 1170, + 1409, + 1170, + 1444, + 948, + 1444 + ], + "score": 0.92, + "latex": "I ( d ; \\dot { z } _ { p } ) - I ( d ; z _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 351, + 1092, + 447, + 1092, + 447, + 1127, + 351, + 1127 + ], + "score": 0.92, + "latex": "I ( \\pmb { y } ; \\pmb { z } _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 254, + 682, + 447, + 682, + 447, + 717, + 254, + 717 + ], + "score": 0.91, + "latex": "p ( { \\pmb x } , { \\pmb y } , d , z _ { s } , z _ { p } )" + }, + { + "category_id": 13, + "poly": [ + 1316, + 1779, + 1448, + 1779, + 1448, + 1813, + 1316, + 1813 + ], + "score": 0.91, + "latex": "I ( d ; z _ { s } ) \\geq" + }, + { + "category_id": 13, + "poly": [ + 611, + 1568, + 880, + 1568, + 880, + 1616, + 611, + 1616 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\Sigma _ { { \\pmb x } } p ( { \\pmb x } | z ) \\ln \\frac { p ( { \\pmb x } | z ) } { q ( { \\pmb x } | z ) } \\ge 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1162, + 1688, + 1270, + 1688, + 1270, + 1721, + 1162, + 1721 + ], + "score": 0.91, + "latex": "q ( { \\pmb x } | { \\pmb z } ; { \\pmb \\phi } )" + }, + { + "category_id": 13, + "poly": [ + 756, + 1779, + 1258, + 1779, + 1258, + 1816, + 756, + 1816 + ], + "score": 0.9, + "latex": "I ( d ; z _ { p } ) \\geq H ( d ) + \\mathbb { E } _ { p ( d , z _ { p } ) } [ \\ln q ( d | z _ { p } ; \\psi ) ]" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1750, + 1136, + 1750, + 1136, + 1782, + 1079, + 1782 + ], + "score": 0.9, + "latex": "p ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 352, + 1166, + 580, + 1166, + 580, + 1201, + 352, + 1201 + ], + "score": 0.9, + "latex": "I ( d ; z _ { p } ) - I ( d ; z _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 1199, + 901, + 1271, + 901, + 1271, + 932, + 1199, + 932 + ], + "score": 0.9, + "latex": "\\lambda _ { r } , \\lambda _ { c }" + }, + { + "category_id": 14, + "poly": [ + 422, + 808, + 1275, + 808, + 1275, + 851, + 422, + 851 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\boldsymbol { \\mathcal { L } } ( \\boldsymbol { \\theta } _ { s } , \\boldsymbol { \\theta } _ { p } , \\boldsymbol { \\theta } _ { c } ; \\boldsymbol { x } , \\boldsymbol { y } , \\boldsymbol { d } ) = \\lambda _ { r } \\boldsymbol { I } ( \\boldsymbol { x } ; \\boldsymbol { z } ) + \\lambda _ { c } \\boldsymbol { I } ( \\boldsymbol { y } ; \\boldsymbol { z } _ { s } ) + \\lambda _ { d } \\big ( \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { p } ) - \\boldsymbol { I } ( \\boldsymbol { d } ; \\boldsymbol { z } _ { s } ) \\big ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1328, + 902, + 1361, + 902, + 1361, + 931, + 1328, + 931 + ], + "score": 0.89, + "latex": "\\lambda _ { d }" + }, + { + "category_id": 13, + "poly": [ + 351, + 1020, + 439, + 1020, + 439, + 1054, + 351, + 1054 + ], + "score": 0.89, + "latex": "I ( { \\pmb x } ; z )" + }, + { + "category_id": 14, + "poly": [ + 540, + 711, + 1157, + 711, + 1157, + 750, + 540, + 750 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { p ( { \\pmb x } , { \\pmb y } , { \\pmb d } , z _ { s } , z _ { p } ) = p ( { \\pmb x } ) p ( { \\pmb d } ) p ( z _ { s } | { \\pmb x } ) p ( z _ { p } | { \\pmb x } ) p ( { \\pmb y } | z _ { s } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1327, + 627, + 1358, + 627, + 1358, + 655, + 1327, + 655 + ], + "score": 0.87, + "latex": "z _ { p }" + }, + { + "category_id": 13, + "poly": [ + 908, + 1383, + 938, + 1383, + 938, + 1409, + 908, + 1409 + ], + "score": 0.87, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 784, + 516, + 815, + 516, + 815, + 541, + 784, + 541 + ], + "score": 0.86, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 1240, + 627, + 1271, + 627, + 1271, + 652, + 1240, + 652 + ], + "score": 0.86, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 252, + 1350, + 301, + 1350, + 301, + 1382, + 252, + 1382 + ], + "score": 0.86, + "latex": "( z _ { p } )" + }, + { + "category_id": 13, + "poly": [ + 302, + 511, + 386, + 511, + 386, + 539, + 302, + 539 + ], + "score": 0.86, + "latex": "M - 1" + }, + { + "category_id": 13, + "poly": [ + 989, + 1383, + 1020, + 1383, + 1020, + 1411, + 989, + 1411 + ], + "score": 0.85, + "latex": "z _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 549, + 1294, + 549, + 1294, + 574, + 1263, + 574 + ], + "score": 0.84, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1319, + 1314, + 1319, + 1314, + 1349, + 1266, + 1349 + ], + "score": 0.84, + "latex": "( z _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 1011, + 478, + 1044, + 478, + 1044, + 506, + 1011, + 506 + ], + "score": 0.83, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 576, + 1721, + 594, + 1721, + 594, + 1750, + 576, + 1750 + ], + "score": 0.83, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 839, + 448, + 866, + 448, + 866, + 474, + 839, + 474 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 852, + 658, + 872, + 658, + 872, + 679, + 852, + 679 + ], + "score": 0.78, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 816, + 628, + 838, + 628, + 838, + 649, + 816, + 649 + ], + "score": 0.77, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1078, + 906, + 1101, + 906, + 1101, + 928, + 1078, + 928 + ], + "score": 0.75, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 960, + 513, + 973, + 513, + 973, + 538, + 960, + 538 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 577, + 544, + 605, + 544, + 605, + 571, + 577, + 571 + ], + "score": 0.72, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 690, + 544, + 722, + 544, + 722, + 571, + 690, + 571 + ], + "score": 0.7, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 853, + 1056, + 875, + 1056, + 875, + 1078, + 853, + 1078 + ], + "score": 0.69, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 939, + 1057, + 958, + 1057, + 958, + 1078, + 939, + 1078 + ], + "score": 0.67, + "latex": "_ { z }" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1694, + 1149, + 1694, + 1149, + 1716, + 1127, + 1716 + ], + "score": 0.65, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 253, + 1415, + 273, + 1415, + 273, + 1437, + 253, + 1437 + ], + "score": 0.64, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1157, + 904, + 1178, + 904, + 1178, + 932, + 1157, + 932 + ], + "score": 0.64, + "latex": "\\textbf { { y } }" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 109.0, + 813.0, + 109.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 372.0, + 668.0, + 372.0, + 668.0, + 416.0, + 247.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1477.0, + 503.0, + 1477.0, + 503.0, + 1516.0, + 248.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1863.0, + 815.0, + 1863.0, + 815.0, + 1894.0, + 794.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 446.0, + 838.0, + 446.0, + 838.0, + 480.0, + 250.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 446.0, + 1451.0, + 446.0, + 1451.0, + 480.0, + 867.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 473.0, + 445.0, + 473.0, + 445.0, + 520.0, + 247.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 473.0, + 1010.0, + 473.0, + 1010.0, + 520.0, + 800.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 473.0, + 1455.0, + 473.0, + 1455.0, + 520.0, + 1045.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 509.0, + 301.0, + 509.0, + 301.0, + 547.0, + 250.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 509.0, + 783.0, + 509.0, + 783.0, + 547.0, + 387.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 509.0, + 959.0, + 509.0, + 959.0, + 547.0, + 816.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 509.0, + 1156.0, + 509.0, + 1156.0, + 547.0, + 974.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 509.0, + 1451.0, + 509.0, + 1451.0, + 547.0, + 1397.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 536.0, + 252.0, + 536.0, + 252.0, + 581.0, + 248.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 536.0, + 576.0, + 536.0, + 576.0, + 581.0, + 486.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 536.0, + 689.0, + 536.0, + 689.0, + 581.0, + 606.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 536.0, + 1262.0, + 536.0, + 1262.0, + 581.0, + 723.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 536.0, + 1455.0, + 536.0, + 1455.0, + 581.0, + 1295.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 571.0, + 997.0, + 571.0, + 997.0, + 608.0, + 248.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1317.0, + 1265.0, + 1317.0, + 1265.0, + 1353.0, + 248.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1317.0, + 1449.0, + 1317.0, + 1449.0, + 1353.0, + 1315.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1347.0, + 251.0, + 1347.0, + 251.0, + 1384.0, + 247.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1347.0, + 1450.0, + 1347.0, + 1450.0, + 1384.0, + 302.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1378.0, + 622.0, + 1378.0, + 622.0, + 1414.0, + 250.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1378.0, + 907.0, + 1378.0, + 907.0, + 1414.0, + 728.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1378.0, + 988.0, + 1378.0, + 988.0, + 1414.0, + 939.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1378.0, + 1450.0, + 1378.0, + 1450.0, + 1414.0, + 1021.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1406.0, + 252.0, + 1406.0, + 252.0, + 1448.0, + 246.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1406.0, + 947.0, + 1406.0, + 947.0, + 1448.0, + 274.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1406.0, + 1183.0, + 1406.0, + 1183.0, + 1448.0, + 1171.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1688.0, + 326.0, + 1688.0, + 326.0, + 1724.0, + 250.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1688.0, + 1126.0, + 1688.0, + 1126.0, + 1724.0, + 395.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1688.0, + 1161.0, + 1688.0, + 1161.0, + 1724.0, + 1150.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1688.0, + 1448.0, + 1688.0, + 1448.0, + 1724.0, + 1271.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1718.0, + 575.0, + 1718.0, + 575.0, + 1754.0, + 250.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1718.0, + 973.0, + 1718.0, + 973.0, + 1754.0, + 595.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1718.0, + 1449.0, + 1718.0, + 1449.0, + 1754.0, + 1083.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1747.0, + 252.0, + 1747.0, + 252.0, + 1787.0, + 249.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1747.0, + 1078.0, + 1747.0, + 1078.0, + 1787.0, + 572.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1747.0, + 1453.0, + 1747.0, + 1453.0, + 1787.0, + 1137.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1775.0, + 755.0, + 1775.0, + 755.0, + 1820.0, + 248.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1775.0, + 1315.0, + 1775.0, + 1315.0, + 1820.0, + 1259.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 867.0, + 323.0, + 867.0, + 323.0, + 907.0, + 249.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 867.0, + 429.0, + 867.0, + 429.0, + 907.0, + 381.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 867.0, + 1452.0, + 867.0, + 1452.0, + 907.0, + 486.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 899.0, + 1077.0, + 899.0, + 1077.0, + 938.0, + 340.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 899.0, + 1156.0, + 899.0, + 1156.0, + 938.0, + 1102.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 899.0, + 1198.0, + 899.0, + 1198.0, + 938.0, + 1179.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 899.0, + 1327.0, + 899.0, + 1327.0, + 938.0, + 1272.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 899.0, + 1450.0, + 899.0, + 1450.0, + 938.0, + 1362.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 927.0, + 1450.0, + 927.0, + 1450.0, + 969.0, + 249.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 960.0, + 746.0, + 960.0, + 746.0, + 997.0, + 249.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 620.0, + 815.0, + 620.0, + 815.0, + 657.0, + 249.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 620.0, + 1001.0, + 620.0, + 1001.0, + 657.0, + 839.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 620.0, + 1239.0, + 620.0, + 1239.0, + 657.0, + 1150.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 620.0, + 1326.0, + 620.0, + 1326.0, + 657.0, + 1272.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 620.0, + 1450.0, + 620.0, + 1450.0, + 657.0, + 1359.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 648.0, + 851.0, + 648.0, + 851.0, + 689.0, + 249.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 648.0, + 1450.0, + 648.0, + 1450.0, + 689.0, + 873.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 681.0, + 253.0, + 681.0, + 253.0, + 718.0, + 249.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 681.0, + 483.0, + 681.0, + 483.0, + 718.0, + 448.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1019.0, + 350.0, + 1019.0, + 350.0, + 1054.0, + 323.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1019.0, + 1448.0, + 1019.0, + 1448.0, + 1054.0, + 440.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1051.0, + 852.0, + 1051.0, + 852.0, + 1084.0, + 348.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1051.0, + 938.0, + 1051.0, + 938.0, + 1084.0, + 876.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1051.0, + 977.0, + 1051.0, + 977.0, + 1084.0, + 959.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1090.0, + 350.0, + 1090.0, + 350.0, + 1128.0, + 322.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1090.0, + 1449.0, + 1090.0, + 1449.0, + 1128.0, + 448.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1123.0, + 454.0, + 1123.0, + 454.0, + 1158.0, + 348.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1164.0, + 351.0, + 1164.0, + 351.0, + 1203.0, + 322.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1164.0, + 1451.0, + 1164.0, + 1451.0, + 1203.0, + 581.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1197.0, + 1449.0, + 1197.0, + 1449.0, + 1232.0, + 347.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1229.0, + 1449.0, + 1229.0, + 1449.0, + 1261.0, + 350.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1259.0, + 977.0, + 1259.0, + 977.0, + 1294.0, + 347.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1534.0, + 1449.0, + 1534.0, + 1449.0, + 1576.0, + 247.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 244.0, + 1569.0, + 610.0, + 1569.0, + 610.0, + 1620.0, + 244.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1569.0, + 955.0, + 1569.0, + 955.0, + 1620.0, + 881.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 265.0, + 1450.0, + 265.0, + 1450.0, + 301.0, + 249.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 296.0, + 656.0, + 296.0, + 656.0, + 331.0, + 248.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 756.0, + 897.0, + 756.0, + 897.0, + 795.0, + 248.0, + 795.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 251, + 932, + 1449, + 932, + 1449, + 1128, + 251, + 1128 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 251, + 1138, + 1450, + 1138, + 1450, + 1284, + 251, + 1284 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 252, + 1368, + 1450, + 1368, + 1450, + 1494, + 252, + 1494 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 574, + 203, + 1144, + 203, + 1144, + 638, + 574, + 638 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 251, + 659, + 1450, + 659, + 1450, + 802, + 251, + 802 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 252, + 818, + 1450, + 818, + 1450, + 921, + 252, + 921 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 524, + 1558, + 1170, + 1558, + 1170, + 1649, + 524, + 1649 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 252, + 1665, + 1445, + 1665, + 1445, + 1731, + 252, + 1731 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 249, + 1753, + 1447, + 1753, + 1447, + 1812, + 249, + 1812 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 254, + 1508, + 802, + 1508, + 802, + 1541, + 254, + 1541 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.901 + }, + { + "category_id": 8, + "poly": [ + 258, + 1299, + 1379, + 1299, + 1379, + 1338, + 258, + 1338 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1411, + 1588, + 1445, + 1588, + 1445, + 1619, + 1411, + 1619 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1412, + 1304, + 1445, + 1304, + 1445, + 1334, + 1412, + 1334 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 795, + 1865, + 813, + 1865, + 813, + 1889, + 795, + 1889 + ], + "score": 0.798 + }, + { + "category_id": 14, + "poly": [ + 524, + 1554, + 1175, + 1554, + 1175, + 1651, + 524, + 1651 + ], + "score": 0.94, + "latex": "\\hat { \\phi } = \\underset { \\phi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { F } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\pmb { x } _ { i } - F \\big ( E _ { s } ( \\pmb { x } _ { i } ) , E _ { p } ( \\pmb { x } _ { i } ) \\big ) \\| _ { 1 } ." + }, + { + "category_id": 13, + "poly": [ + 1043, + 1061, + 1152, + 1061, + 1152, + 1096, + 1043, + 1096 + ], + "score": 0.94, + "latex": "C ( z _ { s } ; \\theta _ { c } )" + }, + { + "category_id": 13, + "poly": [ + 833, + 1026, + 928, + 1026, + 928, + 1058, + 833, + 1058 + ], + "score": 0.94, + "latex": "D ( z ; \\psi )" + }, + { + "category_id": 13, + "poly": [ + 603, + 995, + 745, + 995, + 745, + 1029, + 603, + 1029 + ], + "score": 0.93, + "latex": "F ( z _ { s } , z _ { p } ; \\phi )" + }, + { + "category_id": 13, + "poly": [ + 1276, + 854, + 1401, + 854, + 1401, + 888, + 1276, + 888 + ], + "score": 0.93, + "latex": "q ( \\pmb { y } | \\pmb { z } _ { s } ; \\theta _ { c } )" + }, + { + "category_id": 13, + "poly": [ + 396, + 1666, + 537, + 1666, + 537, + 1700, + 396, + 1700 + ], + "score": 0.93, + "latex": "F ( z _ { s } , z _ { p } ; \\phi )" + }, + { + "category_id": 13, + "poly": [ + 427, + 965, + 540, + 965, + 540, + 999, + 427, + 999 + ], + "score": 0.93, + "latex": "E _ { p } ( \\pmb { x } ; \\theta _ { p } )" + }, + { + "category_id": 13, + "poly": [ + 873, + 1431, + 992, + 1431, + 992, + 1465, + 873, + 1465 + ], + "score": 0.93, + "latex": "( \\theta _ { s } , \\theta _ { p } , \\theta _ { c } )" + }, + { + "category_id": 13, + "poly": [ + 575, + 855, + 670, + 855, + 670, + 888, + 575, + 888 + ], + "score": 0.93, + "latex": "I ( \\pmb { y } ; \\pmb { z } _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 405, + 718, + 506, + 718, + 506, + 748, + 405, + 748 + ], + "score": 0.93, + "latex": "F \\big ( z _ { p } , z _ { s } \\big )" + }, + { + "category_id": 13, + "poly": [ + 700, + 820, + 821, + 820, + 821, + 854, + 700, + 854 + ], + "score": 0.93, + "latex": "q ( d | z _ { p } ; \\psi )" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1250, + 1286, + 1250, + 1286, + 1283, + 1207, + 1283 + ], + "score": 0.93, + "latex": "q ( d | z )" + }, + { + "category_id": 13, + "poly": [ + 624, + 662, + 693, + 662, + 693, + 692, + 624, + 692 + ], + "score": 0.92, + "latex": "E _ { s } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1210, + 1393, + 1210, + 1393, + 1248, + 1036, + 1248 + ], + "score": 0.92, + "latex": "p ( z _ { s } | \\pmb { x } ) = \\delta \\big ( z _ { s } - E _ { s } ( \\pmb { x } ; \\theta _ { s } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 426, + 1212, + 619, + 1212, + 619, + 1247, + 426, + 1247 + ], + "score": 0.92, + "latex": "p ( \\pmb { z } _ { s } | \\pmb { x } ) , p ( \\pmb { z } _ { p } | \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 356, + 934, + 467, + 934, + 467, + 967, + 356, + 967 + ], + "score": 0.92, + "latex": "E _ { s } ( \\pmb { x } ; \\theta _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1753, + 1403, + 1753, + 1403, + 1784, + 1293, + 1784 + ], + "score": 0.92, + "latex": "q ( d | z _ { s } , \\psi )" + }, + { + "category_id": 13, + "poly": [ + 254, + 1170, + 486, + 1170, + 486, + 1213, + 254, + 1213 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { \\bar { N } } \\delta ( { \\bf d } - { \\bf d } _ { i } ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 421, + 1139, + 608, + 1139, + 608, + 1174, + 421, + 1174 + ], + "score": 0.92, + "latex": "p ( { \\pmb d } ) , p ( { \\pmb x } ) , p ( { \\pmb y } )" + }, + { + "category_id": 13, + "poly": [ + 504, + 690, + 575, + 690, + 575, + 720, + 504, + 720 + ], + "score": 0.92, + "latex": "E _ { p } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 253, + 819, + 608, + 819, + 608, + 857, + 253, + 857 + ], + "score": 0.91, + "latex": "H ( d ) + \\mathbb { E } _ { p ( d , z _ { s } ) } [ \\ln q ( d | z _ { s } ; \\psi ) ]" + }, + { + "category_id": 13, + "poly": [ + 908, + 889, + 995, + 889, + 995, + 922, + 908, + 922 + ], + "score": 0.91, + "latex": "p ( \\pmb { y } | \\boldsymbol { z } _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 1304, + 1369, + 1355, + 1369, + 1355, + 1403, + 1304, + 1403 + ], + "score": 0.91, + "latex": "\\left. . \\right. _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 704, + 853, + 1191, + 853, + 1191, + 889, + 704, + 889 + ], + "score": 0.91, + "latex": "I ( \\pmb { y } ; \\pmb { z _ { s } } ) \\geq H ( \\pmb { y } ) + \\mathbb { E } _ { p ( \\pmb { y } , \\pmb { z _ { s } } ) } [ \\ln q ( \\pmb { y } | \\pmb { z _ { s } } ; \\pmb { \\theta _ { c } } ) ]" + }, + { + "category_id": 13, + "poly": [ + 252, + 1248, + 591, + 1248, + 591, + 1285, + 252, + 1285 + ], + "score": 0.9, + "latex": "p ( z _ { p } | \\pmb { x } ) = \\delta \\big ( z _ { p } - E _ { p } ( \\pmb { x } ; \\theta _ { p } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 252, + 1462, + 315, + 1462, + 315, + 1494, + 252, + 1494 + ], + "score": 0.89, + "latex": "( \\phi , \\psi )" + }, + { + "category_id": 13, + "poly": [ + 253, + 1781, + 364, + 1781, + 364, + 1813, + 253, + 1813 + ], + "score": 0.89, + "latex": "q ( d | z _ { p } ; \\psi )" + }, + { + "category_id": 13, + "poly": [ + 797, + 1759, + 816, + 1759, + 816, + 1783, + 797, + 1783 + ], + "score": 0.89, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1140, + 1446, + 1140, + 1446, + 1175, + 1345, + 1175 + ], + "score": 0.88, + "latex": "\\begin{array} { r l } { p ( d ) } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 293, + 1402, + 328, + 1402, + 328, + 1431, + 293, + 1431 + ], + "score": 0.88, + "latex": "L _ { 1 }" + }, + { + "category_id": 14, + "poly": [ + 273, + 1298, + 1390, + 1298, + 1390, + 1338, + 273, + 1338 + ], + "score": 0.88, + "latex": "q ( \\boldsymbol { y } | \\boldsymbol { z } _ { s } ) = \\mathrm { S o f t M a x } ( C ( \\boldsymbol { z } _ { s } ; \\boldsymbol { \\theta } _ { c } ) ) , q ( d | \\boldsymbol { z } ) = \\mathrm { S o f t M a x } ( D ( \\boldsymbol { z } ; \\boldsymbol { \\psi } ) ) , q ( \\boldsymbol { x } | \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\propto \\exp ( \\| \\boldsymbol { x } - \\boldsymbol { F } ( \\boldsymbol { z } ; \\boldsymbol { \\phi } ) \\| _ { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 756, + 496, + 778, + 496, + 778, + 513, + 756, + 513 + ], + "score": 0.88, + "latex": "E _ { s }" + }, + { + "category_id": 13, + "poly": [ + 839, + 936, + 867, + 936, + 867, + 965, + 839, + 965 + ], + "score": 0.87, + "latex": "\\theta _ { s }" + }, + { + "category_id": 13, + "poly": [ + 425, + 1031, + 457, + 1031, + 457, + 1060, + 425, + 1060 + ], + "score": 0.87, + "latex": "z _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1030, + 969, + 1061, + 969, + 1061, + 1000, + 1030, + 1000 + ], + "score": 0.87, + "latex": "z _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 1029, + 1398, + 1029, + 1398, + 1060, + 1366, + 1060 + ], + "score": 0.87, + "latex": "z _ { p }" + }, + { + "category_id": 13, + "poly": [ + 708, + 889, + 736, + 889, + 736, + 918, + 708, + 918 + ], + "score": 0.87, + "latex": "\\theta _ { c }" + }, + { + "category_id": 13, + "poly": [ + 337, + 970, + 367, + 970, + 367, + 996, + 337, + 996 + ], + "score": 0.86, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1030, + 375, + 1030, + 375, + 1057, + 345, + 1057 + ], + "score": 0.85, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 356, + 1098, + 386, + 1098, + 386, + 1124, + 356, + 1124 + ], + "score": 0.85, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 438, + 1095, + 466, + 1095, + 466, + 1121, + 438, + 1121 + ], + "score": 0.85, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1286, + 1030, + 1317, + 1030, + 1317, + 1057, + 1286, + 1057 + ], + "score": 0.85, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 579, + 1510, + 599, + 1510, + 599, + 1540, + 579, + 1540 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1082, + 415, + 1093, + 415, + 1093, + 430, + 1082, + 430 + ], + "score": 0.84, + "latex": "\\hat { x }" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1095, + 1031, + 1095, + 1031, + 1126, + 1011, + 1126 + ], + "score": 0.84, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 1146, + 1699, + 1170, + 1699, + 1170, + 1725, + 1146, + 1725 + ], + "score": 0.84, + "latex": "F" + }, + { + "category_id": 13, + "poly": [ + 1354, + 820, + 1376, + 820, + 1376, + 852, + 1354, + 852 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1083, + 280, + 1093, + 280, + 1093, + 299, + 1083, + 299 + ], + "score": 0.83, + "latex": "\\hat { d }" + }, + { + "category_id": 13, + "poly": [ + 974, + 417, + 990, + 417, + 990, + 431, + 974, + 431 + ], + "score": 0.83, + "latex": "F ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1055, + 664, + 1096, + 664, + 1096, + 690, + 1055, + 690 + ], + "score": 0.82, + "latex": "( z _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 973, + 551, + 989, + 551, + 989, + 566, + 973, + 566 + ], + "score": 0.82, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1009, + 691, + 1050, + 691, + 1050, + 719, + 1009, + 719 + ], + "score": 0.82, + "latex": "( z _ { p } )" + }, + { + "category_id": 13, + "poly": [ + 754, + 1028, + 776, + 1028, + 776, + 1053, + 754, + 1053 + ], + "score": 0.82, + "latex": "\\hat { \\pmb x }" + }, + { + "category_id": 13, + "poly": [ + 253, + 1062, + 285, + 1062, + 285, + 1090, + 253, + 1090 + ], + "score": 0.81, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 445, + 774, + 468, + 774, + 468, + 798, + 445, + 798 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 919, + 1058, + 940, + 1058, + 940, + 1090, + 919, + 1090 + ], + "score": 0.81, + "latex": "\\hat { d }" + }, + { + "category_id": 13, + "poly": [ + 777, + 1509, + 802, + 1509, + 802, + 1536, + 777, + 1536 + ], + "score": 0.81, + "latex": "F" + }, + { + "category_id": 13, + "poly": [ + 346, + 746, + 371, + 746, + 371, + 770, + 346, + 770 + ], + "score": 0.8, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 838, + 498, + 856, + 498, + 856, + 511, + 838, + 511 + ], + "score": 0.8, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1345, + 667, + 1366, + 667, + 1366, + 687, + 1345, + 687 + ], + "score": 0.76, + "latex": "_ { \\pmb { x } }" + }, + { + "category_id": 13, + "poly": [ + 971, + 1219, + 993, + 1219, + 993, + 1242, + 971, + 1242 + ], + "score": 0.75, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 977, + 970, + 999, + 970, + 999, + 992, + 977, + 992 + ], + "score": 0.75, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1425, + 1673, + 1446, + 1673, + 1446, + 1694, + 1425, + 1694 + ], + "score": 0.73, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 941, + 1124, + 941, + 1124, + 962, + 1103, + 962 + ], + "score": 0.73, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 550, + 1092, + 550, + 1092, + 564, + 1083, + 564 + ], + "score": 0.72, + "latex": "\\hat { c }" + }, + { + "category_id": 13, + "poly": [ + 755, + 332, + 779, + 332, + 779, + 354, + 755, + 354 + ], + "score": 0.71, + "latex": "E _ { p }" + }, + { + "category_id": 13, + "poly": [ + 962, + 1248, + 1050, + 1248, + 1050, + 1284, + 962, + 1284 + ], + "score": 0.58, + "latex": "q ( \\pmb { y } | \\pmb { z } _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 838, + 334, + 857, + 334, + 857, + 352, + 838, + 352 + ], + "score": 0.57, + "latex": "z _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1063, + 1249, + 1142, + 1249, + 1142, + 1284, + 1063, + 1284 + ], + "score": 0.56, + "latex": "q ( { \\pmb x } | z )" + }, + { + "category_id": 13, + "poly": [ + 973, + 284, + 990, + 284, + 990, + 298, + 973, + 298 + ], + "score": 0.55, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 627, + 418, + 639, + 418, + 639, + 428, + 627, + 428 + ], + "score": 0.5, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 421, + 1369, + 451, + 1369, + 451, + 1402, + 421, + 1402 + ], + "score": 0.39, + "latex": "( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1411, + 1178, + 1447, + 1178, + 1447, + 1207, + 1411, + 1207 + ], + "score": 0.27, + "latex": "\\mathrm { P u }" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 223.0, + 729.0, + 223.0, + 729.0, + 255.0, + 600.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 247.0, + 725.0, + 247.0, + 725.0, + 276.0, + 602.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 410.0, + 1081.0, + 410.0, + 1081.0, + 432.0, + 1078.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 410.0, + 1098.0, + 410.0, + 1098.0, + 432.0, + 1094.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 329.5, + 778.0, + 329.5, + 778.0, + 353.5, + 760.0, + 353.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 333.0, + 855.0, + 333.0, + 855.0, + 352.0, + 841.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 498.0, + 768.0, + 498.0, + 768.0, + 509.0, + 765.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 495.5, + 856.0, + 495.5, + 856.0, + 510.5, + 840.0, + 510.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 660.0, + 623.0, + 660.0, + 623.0, + 694.0, + 250.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 660.0, + 1054.0, + 660.0, + 1054.0, + 694.0, + 694.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 660.0, + 1344.0, + 660.0, + 1344.0, + 694.0, + 1097.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 660.0, + 1451.0, + 660.0, + 1451.0, + 694.0, + 1367.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 690.0, + 503.0, + 690.0, + 503.0, + 722.0, + 251.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 690.0, + 1008.0, + 690.0, + 1008.0, + 722.0, + 576.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 690.0, + 1450.0, + 690.0, + 1450.0, + 722.0, + 1051.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 718.0, + 404.0, + 718.0, + 404.0, + 749.0, + 251.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 718.0, + 1450.0, + 718.0, + 1450.0, + 749.0, + 507.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 743.0, + 345.0, + 743.0, + 345.0, + 776.0, + 250.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 743.0, + 1450.0, + 743.0, + 1450.0, + 776.0, + 372.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 772.0, + 444.0, + 772.0, + 444.0, + 804.0, + 248.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 772.0, + 1141.0, + 772.0, + 1141.0, + 804.0, + 469.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1746.0, + 796.0, + 1746.0, + 796.0, + 1788.0, + 284.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1746.0, + 1292.0, + 1746.0, + 1292.0, + 1788.0, + 817.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1746.0, + 1452.0, + 1746.0, + 1452.0, + 1788.0, + 1404.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1776.0, + 252.0, + 1776.0, + 252.0, + 1817.0, + 248.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1776.0, + 379.0, + 1776.0, + 379.0, + 1817.0, + 365.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1506.0, + 578.0, + 1506.0, + 578.0, + 1544.0, + 249.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1506.0, + 776.0, + 1506.0, + 776.0, + 1544.0, + 600.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1862.0, + 817.0, + 1862.0, + 817.0, + 1895.0, + 792.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 935.0, + 355.0, + 935.0, + 355.0, + 968.0, + 250.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 935.0, + 838.0, + 935.0, + 838.0, + 968.0, + 468.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 935.0, + 1102.0, + 935.0, + 1102.0, + 968.0, + 868.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 935.0, + 1450.0, + 935.0, + 1450.0, + 968.0, + 1125.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 965.0, + 336.0, + 965.0, + 336.0, + 998.0, + 251.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 965.0, + 426.0, + 965.0, + 426.0, + 998.0, + 368.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 965.0, + 976.0, + 965.0, + 976.0, + 998.0, + 541.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 965.0, + 1029.0, + 965.0, + 1029.0, + 998.0, + 1000.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 965.0, + 1450.0, + 965.0, + 1450.0, + 998.0, + 1062.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 993.0, + 602.0, + 993.0, + 602.0, + 1034.0, + 248.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 993.0, + 1453.0, + 993.0, + 1453.0, + 1034.0, + 746.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1025.0, + 344.0, + 1025.0, + 344.0, + 1062.0, + 250.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1025.0, + 424.0, + 1025.0, + 424.0, + 1062.0, + 376.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1025.0, + 753.0, + 1025.0, + 753.0, + 1062.0, + 458.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1025.0, + 832.0, + 1025.0, + 832.0, + 1062.0, + 777.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1025.0, + 1285.0, + 1025.0, + 1285.0, + 1062.0, + 929.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1025.0, + 1365.0, + 1025.0, + 1365.0, + 1062.0, + 1318.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1025.0, + 1453.0, + 1025.0, + 1453.0, + 1062.0, + 1399.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1059.0, + 918.0, + 1059.0, + 918.0, + 1096.0, + 286.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1059.0, + 1042.0, + 1059.0, + 1042.0, + 1096.0, + 941.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1059.0, + 1449.0, + 1059.0, + 1449.0, + 1096.0, + 1153.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1092.0, + 355.0, + 1092.0, + 355.0, + 1129.0, + 248.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1092.0, + 437.0, + 1092.0, + 437.0, + 1129.0, + 387.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1092.0, + 1010.0, + 1092.0, + 1010.0, + 1129.0, + 467.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1092.0, + 1043.0, + 1092.0, + 1043.0, + 1129.0, + 1032.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1135.0, + 420.0, + 1135.0, + 420.0, + 1175.0, + 248.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1135.0, + 1344.0, + 1135.0, + 1344.0, + 1175.0, + 609.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1152.0, + 253.0, + 1152.0, + 253.0, + 1224.0, + 235.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1152.0, + 1410.0, + 1152.0, + 1410.0, + 1224.0, + 487.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1448.0, + 1152.0, + 1463.0, + 1152.0, + 1463.0, + 1224.0, + 1448.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1210.0, + 425.0, + 1210.0, + 425.0, + 1248.0, + 246.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1210.0, + 970.0, + 1210.0, + 970.0, + 1248.0, + 620.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1210.0, + 1035.0, + 1210.0, + 1035.0, + 1248.0, + 994.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1210.0, + 1450.0, + 1210.0, + 1450.0, + 1248.0, + 1394.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1246.0, + 251.0, + 1246.0, + 251.0, + 1285.0, + 248.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1246.0, + 961.0, + 1246.0, + 961.0, + 1285.0, + 592.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1246.0, + 1062.0, + 1246.0, + 1062.0, + 1285.0, + 1051.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1246.0, + 1206.0, + 1246.0, + 1206.0, + 1285.0, + 1143.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1246.0, + 1321.0, + 1246.0, + 1321.0, + 1285.0, + 1287.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1367.0, + 420.0, + 1367.0, + 420.0, + 1407.0, + 249.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1367.0, + 1303.0, + 1367.0, + 1303.0, + 1407.0, + 452.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1367.0, + 1451.0, + 1367.0, + 1451.0, + 1407.0, + 1356.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1399.0, + 292.0, + 1399.0, + 292.0, + 1434.0, + 249.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1399.0, + 1449.0, + 1399.0, + 1449.0, + 1434.0, + 329.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1428.0, + 872.0, + 1428.0, + 872.0, + 1466.0, + 248.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1428.0, + 1450.0, + 1428.0, + 1450.0, + 1466.0, + 993.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1459.0, + 251.0, + 1459.0, + 251.0, + 1497.0, + 248.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1459.0, + 800.0, + 1459.0, + 800.0, + 1497.0, + 316.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 816.0, + 252.0, + 816.0, + 252.0, + 857.0, + 248.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 816.0, + 699.0, + 816.0, + 699.0, + 857.0, + 609.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 816.0, + 1353.0, + 816.0, + 1353.0, + 857.0, + 822.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 816.0, + 1453.0, + 816.0, + 1453.0, + 857.0, + 1377.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 850.0, + 574.0, + 850.0, + 574.0, + 894.0, + 248.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 850.0, + 703.0, + 850.0, + 703.0, + 894.0, + 671.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 850.0, + 1275.0, + 850.0, + 1275.0, + 894.0, + 1192.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 850.0, + 1451.0, + 850.0, + 1451.0, + 894.0, + 1402.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 886.0, + 707.0, + 886.0, + 707.0, + 925.0, + 248.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 886.0, + 907.0, + 886.0, + 907.0, + 925.0, + 737.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 886.0, + 1009.0, + 886.0, + 1009.0, + 925.0, + 996.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1663.0, + 395.0, + 1663.0, + 395.0, + 1702.0, + 249.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1663.0, + 1424.0, + 1663.0, + 1424.0, + 1702.0, + 538.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1447.0, + 1663.0, + 1451.0, + 1663.0, + 1451.0, + 1702.0, + 1447.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1695.0, + 1145.0, + 1695.0, + 1145.0, + 1730.0, + 249.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1695.0, + 1181.0, + 1695.0, + 1181.0, + 1730.0, + 1171.0, + 1730.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 251, + 919, + 1451, + 919, + 1451, + 1169, + 251, + 1169 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 252, + 1330, + 1446, + 1330, + 1446, + 1455, + 252, + 1455 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 251, + 414, + 1448, + 414, + 1448, + 509, + 251, + 509 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 309, + 716, + 1385, + 716, + 1385, + 912, + 309, + 912 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 307, + 1512, + 1390, + 1512, + 1390, + 1807, + 307, + 1807 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 379, + 1227, + 1315, + 1227, + 1315, + 1321, + 379, + 1321 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 415, + 312, + 1283, + 312, + 1283, + 403, + 415, + 403 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 256, + 639, + 1448, + 639, + 1448, + 705, + 256, + 705 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 560, + 567, + 1133, + 567, + 1133, + 625, + 560, + 625 + ], + "score": 0.936 + }, + { + "category_id": 2, + "poly": [ + 252, + 110, + 813, + 110, + 813, + 142, + 252, + 142 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 253, + 266, + 914, + 266, + 914, + 301, + 253, + 301 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1410, + 574, + 1446, + 574, + 1446, + 605, + 1410, + 605 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1411, + 848, + 1445, + 848, + 1445, + 878, + 1411, + 878 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1411, + 1259, + 1445, + 1259, + 1445, + 1289, + 1411, + 1289 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1410, + 342, + 1446, + 342, + 1446, + 373, + 1410, + 373 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1396, + 1741, + 1446, + 1741, + 1446, + 1772, + 1396, + 1772 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 795, + 1865, + 814, + 1865, + 814, + 1890, + 795, + 1890 + ], + "score": 0.795 + }, + { + "category_id": 0, + "poly": [ + 253, + 1181, + 905, + 1181, + 905, + 1216, + 253, + 1216 + ], + "score": 0.787 + }, + { + "category_id": 0, + "poly": [ + 252, + 1469, + 899, + 1469, + 899, + 1502, + 252, + 1502 + ], + "score": 0.65 + }, + { + "category_id": 0, + "poly": [ + 252, + 522, + 886, + 522, + 886, + 556, + 252, + 556 + ], + "score": 0.581 + }, + { + "category_id": 1, + "poly": [ + 252, + 522, + 886, + 522, + 886, + 556, + 252, + 556 + ], + "score": 0.309 + }, + { + "category_id": 1, + "poly": [ + 252, + 1469, + 899, + 1469, + 899, + 1502, + 252, + 1502 + ], + "score": 0.268 + }, + { + "category_id": 1, + "poly": [ + 253, + 1181, + 905, + 1181, + 905, + 1216, + 253, + 1216 + ], + "score": 0.12 + }, + { + "category_id": 14, + "poly": [ + 312, + 1511, + 1389, + 1511, + 1389, + 1812, + 312, + 1812 + ], + "score": 0.94, + "latex": "\\begin{array} { c } { \\displaystyle \\hat { \\theta _ { s } } = \\arg \\operatorname* { m i n } _ { \\theta _ { s } } \\mathcal { L } _ { S } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { c } } { N } \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) } \\\\ { \\displaystyle + \\frac { \\lambda _ { c } } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\frac { 1 } { N - N _ { s } } \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) \\big ) \\bigg ) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 414, + 308, + 1284, + 308, + 1284, + 404, + 414, + 404 + ], + "score": 0.94, + "latex": "\\hat { \\psi } = \\underset { \\psi } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } _ { D } = - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { s } ( \\pmb { x } _ { i } ) \\big ) - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } \\ln D \\big ( E _ { p } ( \\pmb { x } _ { i } ) \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 382, + 1226, + 1317, + 1226, + 1317, + 1322, + 382, + 1322 + ], + "score": 0.94, + "latex": "\\hat { \\theta _ { p } } = \\underset { \\theta _ { p } } { \\operatorname { a r g m i n } } \\ \\mathcal { L } _ { P } = \\frac { \\lambda _ { r } } { N } \\sum _ { i = 1 } ^ { N } \\| \\boldsymbol { x } _ { i } - F \\big ( E _ { s } ( \\boldsymbol { x } _ { i } ) , E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) \\| _ { 1 } - \\frac { \\lambda _ { d } } { N } \\sum _ { i = 1 } ^ { N } d _ { i } ^ { \\top } D \\big ( E _ { p } ( \\boldsymbol { x } _ { i } ) \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 312, + 712, + 1387, + 712, + 1387, + 915, + 312, + 915 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { C } = - \\ \\frac { 1 } { N } \\displaystyle \\sum _ { i = 1 } ^ { N _ { s } } y _ { i } ^ { T } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) - \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln C \\big ( E _ { s } ( x _ { i } ) \\big ) } \\\\ & { \\qquad + \\ \\frac { \\lambda _ { c } } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) ^ { \\top } \\ln \\bigg ( \\displaystyle \\frac { 1 } { N - N _ { s } } \\displaystyle \\sum _ { i = N _ { s } + 1 } ^ { N } C \\big ( E _ { s } ( x _ { i } ) \\big ) \\bigg ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 378, + 982, + 467, + 982, + 467, + 1016, + 378, + 1016 + ], + "score": 0.92, + "latex": "q ( \\pmb { y } | \\pmb { z } _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 720, + 640, + 787, + 640, + 787, + 674, + 720, + 674 + ], + "score": 0.92, + "latex": "H ( \\pmb { y } )" + }, + { + "category_id": 13, + "poly": [ + 485, + 1014, + 574, + 1014, + 574, + 1047, + 485, + 1047 + ], + "score": 0.92, + "latex": "p ( \\pmb { y } | \\pmb { z } _ { s } )" + }, + { + "category_id": 13, + "poly": [ + 253, + 416, + 348, + 416, + 348, + 449, + 253, + 449 + ], + "score": 0.91, + "latex": "D ( z ; \\psi )" + }, + { + "category_id": 14, + "poly": [ + 563, + 564, + 1137, + 564, + 1137, + 629, + 563, + 629 + ], + "score": 0.9, + "latex": "\\hat { \\theta } _ { c } = \\underset { \\theta _ { c } } { \\arg \\operatorname* { m i n } } \\big \\{ - H ( \\pmb { y } ) - \\mathbb { E } _ { p ( \\pmb { y } , z _ { s } ) } \\big [ \\ln q ( \\pmb { y } | z _ { s } ) \\big ] \\big \\} ." + }, + { + "category_id": 13, + "poly": [ + 1385, + 641, + 1446, + 641, + 1446, + 674, + 1385, + 674 + ], + "score": 0.9, + "latex": "H [ \\pmb { y } ]" + }, + { + "category_id": 13, + "poly": [ + 882, + 1332, + 918, + 1332, + 918, + 1365, + 882, + 1365 + ], + "score": 0.9, + "latex": "E _ { p }" + }, + { + "category_id": 13, + "poly": [ + 326, + 923, + 362, + 923, + 362, + 953, + 326, + 953 + ], + "score": 0.88, + "latex": "N _ { s }" + }, + { + "category_id": 13, + "poly": [ + 579, + 524, + 607, + 524, + 607, + 554, + 579, + 554 + ], + "score": 0.88, + "latex": "\\theta _ { c }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1183, + 597, + 1183, + 597, + 1217, + 568, + 1217 + ], + "score": 0.88, + "latex": "\\theta _ { p }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1471, + 596, + 1471, + 596, + 1500, + 568, + 1500 + ], + "score": 0.88, + "latex": "\\theta _ { s }" + }, + { + "category_id": 13, + "poly": [ + 861, + 1471, + 895, + 1471, + 895, + 1500, + 861, + 1500 + ], + "score": 0.87, + "latex": "E _ { s }" + }, + { + "category_id": 13, + "poly": [ + 959, + 1079, + 990, + 1079, + 990, + 1105, + 959, + 1105 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 579, + 270, + 601, + 270, + 601, + 300, + 579, + 300 + ], + "score": 0.85, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 867, + 1185, + 902, + 1185, + 902, + 1217, + 867, + 1217 + ], + "score": 0.85, + "latex": "E _ { p }" + }, + { + "category_id": 13, + "poly": [ + 644, + 1106, + 669, + 1106, + 669, + 1132, + 644, + 1132 + ], + "score": 0.84, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1204, + 923, + 1228, + 923, + 1228, + 949, + 1204, + 949 + ], + "score": 0.83, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 860, + 673, + 883, + 673, + 883, + 699, + 860, + 699 + ], + "score": 0.83, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 373, + 1136, + 430, + 1136, + 430, + 1164, + 373, + 1164 + ], + "score": 0.82, + "latex": "\\ln K" + }, + { + "category_id": 13, + "poly": [ + 1066, + 447, + 1092, + 447, + 1092, + 473, + 1066, + 473 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 858, + 525, + 883, + 525, + 883, + 551, + 858, + 551 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 885, + 270, + 912, + 270, + 912, + 295, + 885, + 295 + ], + "score": 0.77, + "latex": "D" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 108.0, + 813.0, + 108.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 265.0, + 578.0, + 265.0, + 578.0, + 305.0, + 250.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 265.0, + 884.0, + 265.0, + 884.0, + 305.0, + 602.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1864.0, + 816.0, + 1864.0, + 816.0, + 1895.0, + 794.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1175.0, + 567.0, + 1175.0, + 567.0, + 1225.0, + 250.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1175.0, + 866.0, + 1175.0, + 866.0, + 1225.0, + 598.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1175.0, + 907.0, + 1175.0, + 907.0, + 1225.0, + 903.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1465.0, + 567.0, + 1465.0, + 567.0, + 1508.0, + 250.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1465.0, + 860.0, + 1465.0, + 860.0, + 1508.0, + 597.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1465.0, + 900.0, + 1465.0, + 900.0, + 1508.0, + 896.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 520.0, + 578.0, + 520.0, + 578.0, + 559.0, + 251.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 520.0, + 857.0, + 520.0, + 857.0, + 559.0, + 608.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 920.0, + 325.0, + 920.0, + 325.0, + 957.0, + 250.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 920.0, + 1203.0, + 920.0, + 1203.0, + 957.0, + 363.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 920.0, + 1450.0, + 920.0, + 1450.0, + 957.0, + 1229.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 952.0, + 1451.0, + 952.0, + 1451.0, + 987.0, + 251.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 984.0, + 377.0, + 984.0, + 377.0, + 1018.0, + 251.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 984.0, + 1451.0, + 984.0, + 1451.0, + 1018.0, + 468.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1013.0, + 484.0, + 1013.0, + 484.0, + 1049.0, + 248.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1013.0, + 1452.0, + 1013.0, + 1452.0, + 1049.0, + 575.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1040.0, + 1452.0, + 1040.0, + 1452.0, + 1080.0, + 247.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1075.0, + 958.0, + 1075.0, + 958.0, + 1110.0, + 248.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1075.0, + 1451.0, + 1075.0, + 1451.0, + 1110.0, + 991.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1101.0, + 643.0, + 1101.0, + 643.0, + 1141.0, + 247.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1101.0, + 1452.0, + 1101.0, + 1452.0, + 1141.0, + 670.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1133.0, + 372.0, + 1133.0, + 372.0, + 1173.0, + 247.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1133.0, + 1031.0, + 1133.0, + 1031.0, + 1173.0, + 431.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1329.0, + 881.0, + 1329.0, + 881.0, + 1369.0, + 248.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1329.0, + 1452.0, + 1329.0, + 1452.0, + 1369.0, + 919.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1356.0, + 1450.0, + 1356.0, + 1450.0, + 1401.0, + 247.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1390.0, + 1449.0, + 1390.0, + 1449.0, + 1428.0, + 249.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1419.0, + 891.0, + 1419.0, + 891.0, + 1459.0, + 248.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 411.0, + 1449.0, + 411.0, + 1449.0, + 449.0, + 349.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 446.0, + 1065.0, + 446.0, + 1065.0, + 480.0, + 251.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 446.0, + 1447.0, + 446.0, + 1447.0, + 480.0, + 1093.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 472.0, + 878.0, + 472.0, + 878.0, + 513.0, + 247.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 636.0, + 719.0, + 636.0, + 719.0, + 679.0, + 248.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 636.0, + 1384.0, + 636.0, + 1384.0, + 679.0, + 788.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 669.0, + 859.0, + 669.0, + 859.0, + 707.0, + 249.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 669.0, + 1419.0, + 669.0, + 1419.0, + 707.0, + 884.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 520.0, + 578.0, + 520.0, + 578.0, + 559.0, + 251.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 520.0, + 857.0, + 520.0, + 857.0, + 559.0, + 608.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1465.0, + 567.0, + 1465.0, + 567.0, + 1508.0, + 250.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1465.0, + 860.0, + 1465.0, + 860.0, + 1508.0, + 597.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1465.0, + 900.0, + 1465.0, + 900.0, + 1508.0, + 896.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1175.0, + 567.0, + 1175.0, + 567.0, + 1225.0, + 250.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1175.0, + 866.0, + 1175.0, + 866.0, + 1225.0, + 598.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1175.0, + 907.0, + 1175.0, + 907.0, + 1225.0, + 903.0, + 1225.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 253, + 1213, + 1449, + 1213, + 1449, + 1672, + 253, + 1672 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 252, + 711, + 1449, + 711, + 1449, + 1197, + 252, + 1197 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 253, + 266, + 1449, + 266, + 1449, + 574, + 253, + 574 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 254, + 1688, + 1449, + 1688, + 1449, + 1811, + 254, + 1811 + ], + "score": 0.974 + }, + { + "category_id": 2, + "poly": [ + 253, + 113, + 812, + 113, + 812, + 141, + 253, + 141 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 255, + 632, + 542, + 632, + 542, + 667, + 255, + 667 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 796, + 1867, + 813, + 1867, + 813, + 1889, + 796, + 1889 + ], + "score": 0.793 + }, + { + "category_id": 13, + "poly": [ + 1048, + 512, + 1084, + 512, + 1084, + 544, + 1048, + 544 + ], + "score": 0.9, + "latex": "E _ { p }" + }, + { + "category_id": 13, + "poly": [ + 440, + 451, + 474, + 451, + 474, + 480, + 440, + 480 + ], + "score": 0.89, + "latex": "E _ { s }" + }, + { + "category_id": 13, + "poly": [ + 889, + 269, + 924, + 269, + 924, + 298, + 889, + 298 + ], + "score": 0.89, + "latex": "E _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1282, + 299, + 1317, + 299, + 1317, + 328, + 1282, + 328 + ], + "score": 0.89, + "latex": "E _ { s }" + }, + { + "category_id": 13, + "poly": [ + 781, + 512, + 815, + 512, + 815, + 541, + 781, + 541 + ], + "score": 0.88, + "latex": "E _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1357, + 362, + 1387, + 362, + 1387, + 389, + 1357, + 389 + ], + "score": 0.86, + "latex": "z _ { s }" + }, + { + "category_id": 13, + "poly": [ + 535, + 543, + 562, + 543, + 562, + 569, + 535, + 569 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1233, + 512, + 1258, + 512, + 1258, + 538, + 1233, + 538 + ], + "score": 0.8, + "latex": "F" + }, + { + "category_id": 13, + "poly": [ + 1417, + 512, + 1442, + 512, + 1442, + 538, + 1417, + 538 + ], + "score": 0.79, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 636, + 360, + 663, + 360, + 663, + 387, + 636, + 387 + ], + "score": 0.78, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1050, + 360, + 1070, + 360, + 1070, + 386, + 1050, + 386 + ], + "score": 0.69, + "latex": "^ d" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 253.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 629.0, + 548.0, + 629.0, + 548.0, + 673.0, + 247.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1864.0, + 816.0, + 1864.0, + 816.0, + 1894.0, + 795.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1215.0, + 1453.0, + 1215.0, + 1453.0, + 1249.0, + 250.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1246.0, + 1450.0, + 1246.0, + 1450.0, + 1280.0, + 250.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1276.0, + 1452.0, + 1276.0, + 1452.0, + 1309.0, + 249.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1307.0, + 1450.0, + 1307.0, + 1450.0, + 1340.0, + 249.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1338.0, + 1449.0, + 1338.0, + 1449.0, + 1371.0, + 249.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1365.0, + 1450.0, + 1365.0, + 1450.0, + 1405.0, + 246.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1397.0, + 1452.0, + 1397.0, + 1452.0, + 1434.0, + 248.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1426.0, + 1453.0, + 1426.0, + 1453.0, + 1465.0, + 246.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1456.0, + 1452.0, + 1456.0, + 1452.0, + 1495.0, + 248.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1487.0, + 1453.0, + 1487.0, + 1453.0, + 1526.0, + 248.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1519.0, + 1454.0, + 1519.0, + 1454.0, + 1555.0, + 248.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1550.0, + 1452.0, + 1550.0, + 1452.0, + 1584.0, + 250.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1580.0, + 1450.0, + 1580.0, + 1450.0, + 1617.0, + 248.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1609.0, + 1452.0, + 1609.0, + 1452.0, + 1648.0, + 246.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1644.0, + 347.0, + 1644.0, + 347.0, + 1675.0, + 249.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 712.0, + 1450.0, + 712.0, + 1450.0, + 745.0, + 251.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 739.0, + 1449.0, + 739.0, + 1449.0, + 776.0, + 248.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 772.0, + 1452.0, + 772.0, + 1452.0, + 809.0, + 249.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 801.0, + 1453.0, + 801.0, + 1453.0, + 838.0, + 248.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 832.0, + 1453.0, + 832.0, + 1453.0, + 869.0, + 248.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 861.0, + 1454.0, + 861.0, + 1454.0, + 900.0, + 247.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 895.0, + 1453.0, + 895.0, + 1453.0, + 928.0, + 249.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 924.0, + 1453.0, + 924.0, + 1453.0, + 961.0, + 248.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 950.0, + 1452.0, + 950.0, + 1452.0, + 993.0, + 245.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 984.0, + 1450.0, + 984.0, + 1450.0, + 1022.0, + 247.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1017.0, + 1450.0, + 1017.0, + 1450.0, + 1049.0, + 249.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1044.0, + 1452.0, + 1044.0, + 1452.0, + 1082.0, + 247.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1074.0, + 1452.0, + 1074.0, + 1452.0, + 1112.0, + 247.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1108.0, + 1449.0, + 1108.0, + 1449.0, + 1141.0, + 249.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1137.0, + 1450.0, + 1137.0, + 1450.0, + 1172.0, + 247.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1169.0, + 818.0, + 1169.0, + 818.0, + 1201.0, + 249.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 266.0, + 888.0, + 266.0, + 888.0, + 305.0, + 248.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 266.0, + 1452.0, + 266.0, + 1452.0, + 305.0, + 925.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 297.0, + 1281.0, + 297.0, + 1281.0, + 332.0, + 249.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 297.0, + 1452.0, + 297.0, + 1452.0, + 332.0, + 1318.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 324.0, + 1452.0, + 324.0, + 1452.0, + 366.0, + 246.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 356.0, + 635.0, + 356.0, + 635.0, + 396.0, + 248.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 356.0, + 1049.0, + 356.0, + 1049.0, + 396.0, + 664.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 356.0, + 1356.0, + 356.0, + 1356.0, + 396.0, + 1071.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1388.0, + 356.0, + 1452.0, + 356.0, + 1452.0, + 396.0, + 1388.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 386.0, + 1450.0, + 386.0, + 1450.0, + 424.0, + 248.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 419.0, + 1452.0, + 419.0, + 1452.0, + 454.0, + 249.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 450.0, + 439.0, + 450.0, + 439.0, + 485.0, + 249.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 450.0, + 1450.0, + 450.0, + 1450.0, + 485.0, + 475.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 479.0, + 1452.0, + 479.0, + 1452.0, + 515.0, + 250.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 508.0, + 780.0, + 508.0, + 780.0, + 546.0, + 248.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 508.0, + 1047.0, + 508.0, + 1047.0, + 546.0, + 816.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 508.0, + 1232.0, + 508.0, + 1232.0, + 546.0, + 1085.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 508.0, + 1416.0, + 508.0, + 1416.0, + 546.0, + 1259.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1443.0, + 508.0, + 1454.0, + 508.0, + 1454.0, + 546.0, + 1443.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 538.0, + 534.0, + 538.0, + 534.0, + 579.0, + 246.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 538.0, + 1411.0, + 538.0, + 1411.0, + 579.0, + 563.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1685.0, + 1452.0, + 1685.0, + 1452.0, + 1726.0, + 249.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1719.0, + 1452.0, + 1719.0, + 1452.0, + 1754.0, + 250.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1750.0, + 1449.0, + 1750.0, + 1449.0, + 1784.0, + 250.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1777.0, + 1450.0, + 1777.0, + 1450.0, + 1815.0, + 247.0, + 1815.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 252, + 266, + 1448, + 266, + 1448, + 542, + 252, + 542 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 253, + 914, + 1450, + 914, + 1450, + 1191, + 253, + 1191 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 253, + 654, + 1450, + 654, + 1450, + 900, + 253, + 900 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 252, + 1206, + 1449, + 1206, + 1449, + 1602, + 252, + 1602 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 256, + 1696, + 1449, + 1696, + 1449, + 1759, + 256, + 1759 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 284, + 1782, + 1033, + 1782, + 1033, + 1811, + 284, + 1811 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 253, + 113, + 812, + 113, + 812, + 141, + 253, + 141 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 255, + 1640, + 536, + 1640, + 536, + 1671, + 255, + 1671 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 257, + 586, + 668, + 586, + 668, + 622, + 257, + 622 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 796, + 1865, + 813, + 1865, + 813, + 1888, + 796, + 1888 + ], + "score": 0.788 + }, + { + "category_id": 13, + "poly": [ + 991, + 1728, + 1089, + 1728, + 1089, + 1757, + 991, + 1757 + ], + "score": 0.89, + "latex": "3 2 \\times 3 2" + }, + { + "category_id": 13, + "poly": [ + 1162, + 1128, + 1241, + 1128, + 1241, + 1162, + 1162, + 1162 + ], + "score": 0.88, + "latex": "[ - 1 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 486, + 977, + 565, + 977, + 565, + 1009, + 486, + 1009 + ], + "score": 0.81, + "latex": "[ - 1 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 797, + 1130, + 822, + 1130, + 822, + 1156, + 797, + 1156 + ], + "score": 0.8, + "latex": "F" + }, + { + "category_id": 13, + "poly": [ + 1324, + 977, + 1443, + 977, + 1443, + 1008, + 1324, + 1008 + ], + "score": 0.72, + "latex": "\\lambda _ { d } = 0 . 2 0" + }, + { + "category_id": 13, + "poly": [ + 1082, + 977, + 1315, + 977, + 1315, + 1008, + 1082, + 1008 + ], + "score": 0.71, + "latex": "\\lambda _ { r } = 1 . 0 , \\lambda _ { c } = 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 426, + 388, + 494, + 388, + 494, + 418, + 426, + 418 + ], + "score": 0.51, + "latex": "6 \\& 9" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1777.0, + 1034.0, + 1777.0, + 1034.0, + 1816.0, + 287.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 253.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1637.0, + 540.0, + 1637.0, + 540.0, + 1676.0, + 250.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 585.0, + 672.0, + 585.0, + 672.0, + 625.0, + 249.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1862.0, + 816.0, + 1862.0, + 816.0, + 1897.0, + 794.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 267.0, + 1448.0, + 267.0, + 1448.0, + 300.0, + 249.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 298.0, + 1449.0, + 298.0, + 1449.0, + 331.0, + 249.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 327.0, + 1448.0, + 327.0, + 1448.0, + 362.0, + 248.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 360.0, + 1451.0, + 360.0, + 1451.0, + 392.0, + 251.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 387.0, + 425.0, + 387.0, + 425.0, + 425.0, + 247.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 387.0, + 1452.0, + 387.0, + 1452.0, + 425.0, + 495.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 417.0, + 1449.0, + 417.0, + 1449.0, + 456.0, + 247.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 451.0, + 1449.0, + 451.0, + 1449.0, + 484.0, + 249.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 480.0, + 1448.0, + 480.0, + 1448.0, + 515.0, + 248.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 511.0, + 1376.0, + 511.0, + 1376.0, + 545.0, + 247.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 914.0, + 1450.0, + 914.0, + 1450.0, + 949.0, + 249.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 944.0, + 1457.0, + 944.0, + 1457.0, + 983.0, + 246.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 973.0, + 485.0, + 973.0, + 485.0, + 1015.0, + 245.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 973.0, + 1081.0, + 973.0, + 1081.0, + 1015.0, + 566.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 973.0, + 1323.0, + 973.0, + 1323.0, + 1015.0, + 1316.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1444.0, + 973.0, + 1455.0, + 973.0, + 1455.0, + 1015.0, + 1444.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1008.0, + 1450.0, + 1008.0, + 1450.0, + 1041.0, + 250.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1039.0, + 1449.0, + 1039.0, + 1449.0, + 1068.0, + 250.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1068.0, + 1450.0, + 1068.0, + 1450.0, + 1100.0, + 248.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1101.0, + 1447.0, + 1101.0, + 1447.0, + 1130.0, + 250.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1130.0, + 796.0, + 1130.0, + 796.0, + 1162.0, + 249.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1130.0, + 1161.0, + 1130.0, + 1161.0, + 1162.0, + 823.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1130.0, + 1449.0, + 1130.0, + 1449.0, + 1162.0, + 1242.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1160.0, + 645.0, + 1160.0, + 645.0, + 1194.0, + 247.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 654.0, + 1449.0, + 654.0, + 1449.0, + 688.0, + 252.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 686.0, + 1451.0, + 686.0, + 1451.0, + 720.0, + 250.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 715.0, + 1451.0, + 715.0, + 1451.0, + 753.0, + 248.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 743.0, + 1454.0, + 743.0, + 1454.0, + 782.0, + 246.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 778.0, + 1450.0, + 778.0, + 1450.0, + 812.0, + 250.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 808.0, + 1449.0, + 808.0, + 1449.0, + 842.0, + 249.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 838.0, + 1447.0, + 838.0, + 1447.0, + 872.0, + 250.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 866.0, + 542.0, + 866.0, + 542.0, + 906.0, + 248.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1207.0, + 1450.0, + 1207.0, + 1450.0, + 1242.0, + 249.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1236.0, + 1452.0, + 1236.0, + 1452.0, + 1272.0, + 251.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1267.0, + 1449.0, + 1267.0, + 1449.0, + 1303.0, + 249.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1296.0, + 1454.0, + 1296.0, + 1454.0, + 1335.0, + 248.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1328.0, + 1450.0, + 1328.0, + 1450.0, + 1363.0, + 249.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1358.0, + 1449.0, + 1358.0, + 1449.0, + 1393.0, + 248.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1387.0, + 1452.0, + 1387.0, + 1452.0, + 1427.0, + 248.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1419.0, + 1452.0, + 1419.0, + 1452.0, + 1455.0, + 249.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1448.0, + 1453.0, + 1448.0, + 1453.0, + 1487.0, + 248.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1479.0, + 1449.0, + 1479.0, + 1449.0, + 1518.0, + 248.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1510.0, + 1449.0, + 1510.0, + 1449.0, + 1545.0, + 248.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1542.0, + 1449.0, + 1542.0, + 1449.0, + 1577.0, + 251.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1570.0, + 1134.0, + 1570.0, + 1134.0, + 1608.0, + 248.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1696.0, + 1453.0, + 1696.0, + 1453.0, + 1731.0, + 250.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1728.0, + 990.0, + 1728.0, + 990.0, + 1761.0, + 249.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1728.0, + 1453.0, + 1728.0, + 1453.0, + 1761.0, + 1090.0, + 1761.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 252, + 1473, + 1452, + 1473, + 1452, + 1811, + 252, + 1811 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 252, + 690, + 1448, + 690, + 1448, + 1273, + 252, + 1273 + ], + "score": 0.982 + }, + { + "category_id": 3, + "poly": [ + 368, + 202, + 1324, + 202, + 1324, + 520, + 368, + 520 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 255, + 528, + 1447, + 528, + 1447, + 640, + 255, + 640 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 255, + 1387, + 569, + 1387, + 569, + 1418, + 255, + 1418 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 796, + 1866, + 813, + 1866, + 813, + 1889, + 796, + 1889 + ], + "score": 0.786 + }, + { + "category_id": 13, + "poly": [ + 1139, + 1567, + 1278, + 1567, + 1278, + 1596, + 1139, + 1596 + ], + "score": 0.92, + "latex": "6 4 \\times 6 4 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 747, + 1658, + 801, + 1658, + 801, + 1686, + 747, + 1686 + ], + "score": 0.85, + "latex": "C 0 9" + }, + { + "category_id": 13, + "poly": [ + 1279, + 1628, + 1333, + 1628, + 1333, + 1656, + 1279, + 1656 + ], + "score": 0.84, + "latex": "C 1 4" + }, + { + "category_id": 13, + "poly": [ + 487, + 1536, + 540, + 1536, + 540, + 1564, + 487, + 1564 + ], + "score": 0.83, + "latex": "C 1 4" + }, + { + "category_id": 13, + "poly": [ + 1173, + 1628, + 1226, + 1628, + 1226, + 1657, + 1173, + 1657 + ], + "score": 0.81, + "latex": "C 1 3" + }, + { + "category_id": 13, + "poly": [ + 1390, + 1506, + 1442, + 1506, + 1442, + 1535, + 1390, + 1535 + ], + "score": 0.81, + "latex": "C 0 5" + }, + { + "category_id": 13, + "poly": [ + 644, + 1658, + 698, + 1658, + 698, + 1687, + 644, + 1687 + ], + "score": 0.81, + "latex": "C 0 8" + }, + { + "category_id": 13, + "poly": [ + 582, + 1658, + 635, + 1658, + 635, + 1687, + 582, + 1687 + ], + "score": 0.8, + "latex": "C 0 5" + }, + { + "category_id": 13, + "poly": [ + 972, + 999, + 997, + 999, + 997, + 1026, + 972, + 1026 + ], + "score": 0.73, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 315, + 1536, + 368, + 1536, + 368, + 1565, + 315, + 1565 + ], + "score": 0.67, + "latex": "C 0 9" + }, + { + "category_id": 13, + "poly": [ + 252, + 1536, + 305, + 1536, + 305, + 1565, + 252, + 1565 + ], + "score": 0.64, + "latex": "C 0 8" + }, + { + "category_id": 13, + "poly": [ + 378, + 1536, + 431, + 1536, + 431, + 1565, + 378, + 1565 + ], + "score": 0.59, + "latex": "C 1 3" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 203.0, + 651.0, + 203.0, + 651.0, + 262.0, + 365.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 212.0, + 1323.0, + 212.0, + 1323.0, + 268.0, + 1051.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 256.0, + 506.0, + 256.0, + 506.0, + 292.0, + 410.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 262.0, + 641.0, + 262.0, + 641.0, + 294.0, + 615.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 267.0, + 1323.0, + 267.0, + 1323.0, + 323.0, + 1053.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 304.0, + 446.0, + 304.0, + 446.0, + 339.0, + 406.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 299.0, + 608.0, + 299.0, + 608.0, + 345.0, + 447.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 348.0, + 506.0, + 348.0, + 506.0, + 387.0, + 472.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 350.0, + 436.0, + 350.0, + 436.0, + 386.0, + 407.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 396.0, + 506.0, + 396.0, + 506.0, + 427.0, + 473.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 400.0, + 435.0, + 400.0, + 435.0, + 423.0, + 411.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 426.0, + 953.0, + 426.0, + 953.0, + 439.0, + 940.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 433.0, + 519.0, + 433.0, + 519.0, + 486.0, + 368.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 437.0, + 887.0, + 437.0, + 887.0, + 446.0, + 876.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 446.0, + 968.0, + 446.0, + 968.0, + 459.0, + 953.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 450.0, + 898.0, + 450.0, + 898.0, + 462.0, + 882.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 455.0, + 879.0, + 455.0, + 879.0, + 464.0, + 869.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 457.0, + 900.0, + 457.0, + 900.0, + 479.0, + 866.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 490.0, + 603.0, + 490.0, + 603.0, + 525.0, + 419.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 491.0, + 941.0, + 491.0, + 941.0, + 523.0, + 759.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 491.0, + 1299.0, + 491.0, + 1299.0, + 525.0, + 1077.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 524.0, + 1453.0, + 524.0, + 1453.0, + 564.0, + 248.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 554.0, + 1449.0, + 554.0, + 1449.0, + 587.0, + 249.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 583.0, + 1451.0, + 583.0, + 1451.0, + 617.0, + 248.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 611.0, + 354.0, + 611.0, + 354.0, + 643.0, + 246.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1382.0, + 575.0, + 1382.0, + 575.0, + 1423.0, + 249.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1863.0, + 815.0, + 1863.0, + 815.0, + 1893.0, + 794.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1474.0, + 1453.0, + 1474.0, + 1453.0, + 1511.0, + 248.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1504.0, + 1389.0, + 1504.0, + 1389.0, + 1541.0, + 248.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1443.0, + 1504.0, + 1452.0, + 1504.0, + 1452.0, + 1541.0, + 1443.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1532.0, + 251.0, + 1532.0, + 251.0, + 1572.0, + 248.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1532.0, + 314.0, + 1532.0, + 314.0, + 1572.0, + 306.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1532.0, + 377.0, + 1532.0, + 377.0, + 1572.0, + 369.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1532.0, + 486.0, + 1532.0, + 486.0, + 1572.0, + 432.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1532.0, + 1453.0, + 1532.0, + 1453.0, + 1572.0, + 541.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1567.0, + 1138.0, + 1567.0, + 1138.0, + 1601.0, + 248.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1567.0, + 1451.0, + 1567.0, + 1451.0, + 1601.0, + 1279.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1597.0, + 1451.0, + 1597.0, + 1451.0, + 1631.0, + 249.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1628.0, + 1172.0, + 1628.0, + 1172.0, + 1663.0, + 251.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1628.0, + 1278.0, + 1628.0, + 1278.0, + 1663.0, + 1227.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 1628.0, + 1451.0, + 1628.0, + 1451.0, + 1663.0, + 1334.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1654.0, + 581.0, + 1654.0, + 581.0, + 1694.0, + 247.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1654.0, + 643.0, + 1654.0, + 643.0, + 1694.0, + 636.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1654.0, + 746.0, + 1654.0, + 746.0, + 1694.0, + 699.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 1654.0, + 1451.0, + 1654.0, + 1451.0, + 1694.0, + 802.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1686.0, + 1451.0, + 1686.0, + 1451.0, + 1726.0, + 248.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1717.0, + 1451.0, + 1717.0, + 1451.0, + 1754.0, + 248.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1750.0, + 1451.0, + 1750.0, + 1451.0, + 1784.0, + 251.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1779.0, + 1014.0, + 1779.0, + 1014.0, + 1816.0, + 247.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 691.0, + 1451.0, + 691.0, + 1451.0, + 729.0, + 247.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 721.0, + 1452.0, + 721.0, + 1452.0, + 760.0, + 247.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 753.0, + 1455.0, + 753.0, + 1455.0, + 789.0, + 247.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 783.0, + 1452.0, + 783.0, + 1452.0, + 820.0, + 248.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 813.0, + 1452.0, + 813.0, + 1452.0, + 849.0, + 247.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 844.0, + 1451.0, + 844.0, + 1451.0, + 880.0, + 248.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 875.0, + 1452.0, + 875.0, + 1452.0, + 909.0, + 247.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 907.0, + 1451.0, + 907.0, + 1451.0, + 940.0, + 248.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 935.0, + 1452.0, + 935.0, + 1452.0, + 971.0, + 247.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 963.0, + 1451.0, + 963.0, + 1451.0, + 1006.0, + 248.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 999.0, + 971.0, + 999.0, + 971.0, + 1032.0, + 249.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 999.0, + 1449.0, + 999.0, + 1449.0, + 1032.0, + 998.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1027.0, + 1453.0, + 1027.0, + 1453.0, + 1064.0, + 248.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1056.0, + 1451.0, + 1056.0, + 1451.0, + 1094.0, + 248.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1088.0, + 1451.0, + 1088.0, + 1451.0, + 1126.0, + 248.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1118.0, + 1449.0, + 1118.0, + 1449.0, + 1156.0, + 247.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1148.0, + 1451.0, + 1148.0, + 1451.0, + 1184.0, + 249.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1182.0, + 1449.0, + 1182.0, + 1449.0, + 1215.0, + 249.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1210.0, + 1451.0, + 1210.0, + 1451.0, + 1247.0, + 248.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1240.0, + 1034.0, + 1240.0, + 1034.0, + 1276.0, + 247.0, + 1276.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 250, + 259, + 1445, + 259, + 1445, + 692, + 250, + 692 + ], + "score": 0.981, + "html": "
methodS→MS→MMS→UM→SM→MMM→UAve.ranking
Source Only62.10±0.6040.43±0.7039.90± 0.6030.29± 0.5955.98 ± 0.4878.30±0.3814.00
1-NN35.8618.2129.3128.0112.5841.2215.00
CORAL (Sun & Saenko (2016))63.10 ± 0.6154.37 ± 0.5350.15 ± 0.6333.40 ±0.7457.70 ± 0.6981.05 ± 0.8011.33
DANN (Ganin et al. (2016))73.80 ±0.4961.05±0.8062.54 ± 0.9135.50 ± 0.6577.40±0.7381.60 ± 0.608.75
ADDA(Tzeng et al. (2017))77.68 ± 0.9264.23±0.7064.10±0.7930.04± 0.9891.47 ±1.090.51 ± 0.806.43
C-ADDA80.10 ± 0.6956.80 ± 0.7964.80±0.8827.50 ± 0.8683.30 ±0.9084.10 ±0.988.95
DTN (Zhang et al. (2015))81.40± 0.4263.70±0.3960.12 ± 0.5240.40±0.5085.70±0.3985.80 ± 0.466.04
c-DTN PixelDA (Bousmalis et al. (2017))82.10 ±0.6259.30 ± 0.5956.87 ± 0.6538.32 ± 0.5080.90±0.8079.31 ±0.787.96
UNIT (Liu et al. (2017))98.10*94.10*1
DSN (Bousmalis et al. (2016))90.6*92.901
c-DSN82.70±0.37 83.10 ±0.2064.80± 0.4065.30± 0.2849.30±0.3083.20±0.3091.65 ± 0.402.85
1p-DSN81.00 ± 0.4760.56 ± 0.36 58.22 ± 0.6860.35 ± 0.5946.80 ± 0.4580.49±0.4088.21 ±0.384.84
mp-DSN83.40±0.3061.00±0.5058.06 ±0.48 58.10 ±0.6445.11 ± 0.3377.33 ± 0.52 79.30±0.5985.16 ± 0.634.90
s-MTDA-ITA82.90±0.1363.10±0.2863.54±0.3047.35 ± 0.4082.42± 0.1986.45 ± 0.715.33
c-MTDA-ITA79.20±0.2859.90 ± 0.3063.70±0.2649.60± 0.25 45.30 ±0.3077.12 ± 0.2289.21±0.282.88
MTDA-ITA84.60±0.2465.30 ± 0.1570.03 ±0.2052.01 ± 0.2185.50 ±0.1887.47 ± 0.25 94.20±0.204.25
1.16
" + }, + { + "category_id": 5, + "poly": [ + 251, + 942, + 1445, + 942, + 1445, + 1345, + 251, + 1345 + ], + "score": 0.978, + "html": "
methodC13→C05C13→C08C13→C09C13→C14C14→C05C14→C08C14→C09C14→C13Ave.ranking
Source Only50.79± 0.3345.90± 0.5040.04±0.4059.68± 0.2960.03±0.5536.80± 0.6140.11 ±0.5060.57± 0.3616.08
1-NN CORAL33.21 54.89 ± 0.5237.0134.4548.79 68.90± 0.3547.4428.2430.8644.8617.00
DANN57.86± 0.4148.90±0.48 50.30 ± 0.4340.30±0.53 45.30± 0.5070.68 ± 0.3559.98 ± 0.45 57.20± 0.4540.63±0.55 40.22±0.5540.80 ±0.53 40.77 ± 0.4565.11 ± 0.45 70.50± 0.5511.95 9.92
ADDA64.83±0.6963.20±0.4555.48 ±0.6574.25±0.5573.62 ± 0.7543.56±0.9538.68 ±0.9572.84±0.759.33
C-ADDA59.20±0.2530.70±0.6353.20±0.4068.33±0.3565.88±0.3830.60±0.6145.34±0.4864.30±0.4011.50
DTN63.78±0.2960.45±0.3560.55±0.3572.60±0.2570.67±0.3041.55±0.6541.45 ± 0.4570.67±0.45
c-DTN57.53 ± 0.4255.24 ± 0.4557.14 ± 0.3965.16 ± 0.3563.80 ±0.4238.97 ± 0.7139.80 ± 0.6562.10 ± 0.458.75 10.92
PixelDA45.68± 0.5244.95± 0.4244.45±0.5590.50± 0.2546.28±0.6045.89±0.6144.45± 0.5169.15 ± 0.459.95
UNIT44.14±0.1044.47 ± 0.1144.21 ±0.1244.47 ± 0.1143.03 ±0.144.44 ±0.1544.47 ± 0.1544.47 ± 0.0511.07
DSN64.15± 0.3057.70±0.3849.15± 0.4580.75±0.2782.20±0.2838.75±0.5345.00± 0.2580.50± 0.35
c-DSN57.34±0.4531.63 ±0.6051.17 ±0.4074.52 ±0.3782.01 ± 0.3534.25±0.5842.63 ± 0.5579.42 ±0.355.15 8.20
1p-DSN55.84± 0.5030.03 ± 0.5049.06 ± 0.3872.11 ± 0.5081.22 ±0.4533.33 ± 0.5842.03 ±0.2478.78±0.578.63
mp-DSN55.20±0.4630.40±0.5047.80±0.3575.30 ± 0.2580.75±0.2030.20 ±0.5543.00±0.3579.02±0.408.88
s-MTDA-ITA70.10± 0.2758.90±0.2558.10±0.2780.12±0.1582.05±0.1845.90±0.3052.67±0.3081.60±0.24
c-MTDA-ITA60.34 ± 0.1755.67 ± 0.2157.10 ± 0.2373.50 ± 0.2076.80 ±0.1043.10 ±0.1248.10 ± 0.1480.90 ± 0.113.65
MTDA-ITA78.40 ± 0.266.70 ± 0.1770.30± 0.1485.49 ± 0.1187.20±0.1061.40±0.1460.05 ±0.1386.70 ± 0.105.01 1.20
" + }, + { + "category_id": 1, + "poly": [ + 253, + 1535, + 1448, + 1535, + 1448, + 1689, + 253, + 1689 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 254, + 1690, + 1447, + 1690, + 1447, + 1812, + 254, + 1812 + ], + "score": 0.957 + }, + { + "category_id": 6, + "poly": [ + 251, + 733, + 1452, + 733, + 1452, + 903, + 251, + 903 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 252, + 112, + 812, + 112, + 812, + 141, + 252, + 141 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 255, + 1472, + 558, + 1472, + 558, + 1503, + 255, + 1503 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 795, + 1865, + 814, + 1865, + 814, + 1889, + 795, + 1889 + ], + "score": 0.801 + }, + { + "category_id": 1, + "poly": [ + 249, + 1389, + 1446, + 1389, + 1446, + 1445, + 249, + 1445 + ], + "score": 0.469 + }, + { + "category_id": 6, + "poly": [ + 249, + 1389, + 1446, + 1389, + 1446, + 1445, + 249, + 1445 + ], + "score": 0.365 + }, + { + "category_id": 13, + "poly": [ + 1262, + 1597, + 1344, + 1597, + 1344, + 1628, + 1262, + 1628 + ], + "score": 0.92, + "latex": "\\lambda _ { r } = 0" + }, + { + "category_id": 13, + "poly": [ + 1054, + 1629, + 1136, + 1629, + 1136, + 1658, + 1054, + 1658 + ], + "score": 0.91, + "latex": "\\lambda _ { c } = 0" + }, + { + "category_id": 13, + "poly": [ + 924, + 1659, + 1007, + 1659, + 1007, + 1688, + 924, + 1688 + ], + "score": 0.91, + "latex": "\\lambda _ { d } = 0" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1749, + 1291, + 1749, + 1291, + 1779, + 1203, + 1779 + ], + "score": 0.89, + "latex": "\\approx 2 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1058, + 1718, + 1157, + 1718, + 1157, + 1748, + 1058, + 1748 + ], + "score": 0.87, + "latex": "\\approx 3 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 711, + 846, + 798, + 846, + 798, + 873, + 711, + 873 + ], + "score": 0.85, + "latex": "> 5 0 0 \\mathrm { K }" + }, + { + "category_id": 13, + "poly": [ + 1172, + 846, + 1280, + 846, + 1280, + 875, + 1172, + 875 + ], + "score": 0.85, + "latex": "( \\approx 1 , 0 0 0 )" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1750, + 1152, + 1750, + 1152, + 1779, + 1056, + 1779 + ], + "score": 0.82, + "latex": "\\approx 4 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 591, + 264, + 655, + 264, + 655, + 283, + 591, + 283 + ], + "score": 0.73, + "latex": "\\overline { { S \\to \\mathbf { M } } }" + }, + { + "category_id": 13, + "poly": [ + 391, + 967, + 493, + 967, + 493, + 988, + 391, + 988 + ], + "score": 0.7, + "latex": "\\overline { { 5 0 . 7 9 \\pm 0 . 3 3 } }" + }, + { + "category_id": 13, + "poly": [ + 509, + 967, + 610, + 967, + 610, + 988, + 509, + 988 + ], + "score": 0.69, + "latex": "\\overline { { 4 5 . 9 0 \\pm 0 . 5 0 } }" + }, + { + "category_id": 13, + "poly": [ + 968, + 267, + 1031, + 267, + 1031, + 283, + 968, + 283 + ], + "score": 0.66, + "latex": "\\mathbf { M } \\to \\mathbf { S }" + }, + { + "category_id": 13, + "poly": [ + 844, + 267, + 904, + 267, + 904, + 283, + 844, + 283 + ], + "score": 0.65, + "latex": "\\mathbf { S } \\to \\mathbf { U }" + }, + { + "category_id": 13, + "poly": [ + 626, + 967, + 728, + 967, + 728, + 988, + 626, + 988 + ], + "score": 0.63, + "latex": "\\overline { { 4 0 . 0 4 \\pm 0 . 4 0 } }" + }, + { + "category_id": 13, + "poly": [ + 507, + 1012, + 610, + 1012, + 610, + 1031, + 507, + 1031 + ], + "score": 0.63, + "latex": "4 8 . 9 0 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 1075, + 384, + 1176, + 384, + 1176, + 402, + 1075, + 402 + ], + "score": 0.6, + "latex": "9 1 . 4 7 \\pm 1 . 0" + }, + { + "category_id": 13, + "poly": [ + 1071, + 455, + 1180, + 455, + 1180, + 473, + 1071, + 473 + ], + "score": 0.6, + "latex": "8 0 . 9 0 \\pm 0 . 8 0" + }, + { + "category_id": 13, + "poly": [ + 1071, + 408, + 1180, + 408, + 1180, + 426, + 1071, + 426 + ], + "score": 0.59, + "latex": "8 3 . 3 0 \\pm 0 . 9 0" + }, + { + "category_id": 13, + "poly": [ + 507, + 1274, + 611, + 1274, + 611, + 1294, + 507, + 1294 + ], + "score": 0.59, + "latex": "\\overline { { 5 8 . 9 0 \\pm 0 . 2 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1012, + 1198, + 1012, + 1198, + 1031, + 1095, + 1031 + ], + "score": 0.59, + "latex": "4 0 . 8 0 \\pm 0 . 5 3" + }, + { + "category_id": 13, + "poly": [ + 508, + 1033, + 610, + 1033, + 610, + 1053, + 508, + 1053 + ], + "score": 0.58, + "latex": "5 0 . 3 0 \\pm 0 . 4 3" + }, + { + "category_id": 13, + "poly": [ + 977, + 1012, + 1080, + 1012, + 1080, + 1031, + 977, + 1031 + ], + "score": 0.58, + "latex": "4 0 . 6 3 \\pm 0 . 5 5" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1274, + 1199, + 1274, + 1199, + 1294, + 1095, + 1294 + ], + "score": 0.58, + "latex": "\\overline { { 5 2 . 6 7 \\pm 0 . 3 0 } }" + }, + { + "category_id": 13, + "poly": [ + 626, + 1012, + 727, + 1012, + 727, + 1031, + 626, + 1031 + ], + "score": 0.58, + "latex": "4 0 . 3 0 \\pm 0 . 5 3" + }, + { + "category_id": 13, + "poly": [ + 742, + 1012, + 845, + 1012, + 845, + 1031, + 742, + 1031 + ], + "score": 0.57, + "latex": "6 8 . 9 0 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 979, + 967, + 1079, + 967, + 1079, + 988, + 979, + 988 + ], + "score": 0.57, + "latex": "3 6 . 8 0 \\pm 0 . 6 1" + }, + { + "category_id": 13, + "poly": [ + 861, + 967, + 962, + 967, + 962, + 988, + 861, + 988 + ], + "score": 0.57, + "latex": "\\overline { { 6 0 . 0 3 \\pm 0 . 5 5 } }" + }, + { + "category_id": 13, + "poly": [ + 390, + 1011, + 493, + 1011, + 493, + 1031, + 390, + 1031 + ], + "score": 0.56, + "latex": "5 4 . 8 9 \\pm 0 . 5 2" + }, + { + "category_id": 13, + "poly": [ + 1070, + 361, + 1179, + 361, + 1179, + 378, + 1070, + 378 + ], + "score": 0.56, + "latex": "7 7 . 4 0 \\pm 0 . 7 3" + }, + { + "category_id": 13, + "poly": [ + 694, + 455, + 802, + 455, + 802, + 473, + 694, + 473 + ], + "score": 0.55, + "latex": "5 9 . 3 0 \\pm 0 . 5 9" + }, + { + "category_id": 13, + "poly": [ + 626, + 1274, + 727, + 1274, + 727, + 1294, + 626, + 1294 + ], + "score": 0.54, + "latex": "\\overline { { 5 8 . 1 0 \\pm 0 . 2 7 } }" + }, + { + "category_id": 13, + "poly": [ + 821, + 383, + 928, + 383, + 928, + 402, + 821, + 402 + ], + "score": 0.53, + "latex": "6 4 . 1 0 \\pm 0 . 7 9" + }, + { + "category_id": 13, + "poly": [ + 1199, + 290, + 1304, + 290, + 1304, + 308, + 1199, + 308 + ], + "score": 0.53, + "latex": "7 8 . 3 0 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 391, + 1033, + 493, + 1033, + 493, + 1053, + 391, + 1053 + ], + "score": 0.52, + "latex": "5 7 . 8 6 \\pm 0 . 4 1" + }, + { + "category_id": 13, + "poly": [ + 743, + 968, + 845, + 968, + 845, + 988, + 743, + 988 + ], + "score": 0.52, + "latex": "\\overline { { 5 9 . 6 8 \\pm 0 . 2 9 } }" + }, + { + "category_id": 13, + "poly": [ + 820, + 408, + 928, + 408, + 928, + 426, + 820, + 426 + ], + "score": 0.52, + "latex": "6 4 . 8 0 \\pm 0 . 8 8" + }, + { + "category_id": 13, + "poly": [ + 861, + 1012, + 963, + 1012, + 963, + 1031, + 861, + 1031 + ], + "score": 0.52, + "latex": "5 9 . 9 8 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 626, + 1033, + 728, + 1033, + 728, + 1053, + 626, + 1053 + ], + "score": 0.52, + "latex": "4 5 . 3 0 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 626, + 1077, + 728, + 1077, + 728, + 1097, + 626, + 1097 + ], + "score": 0.51, + "latex": "5 3 . 2 0 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 1072, + 290, + 1179, + 290, + 1179, + 308, + 1072, + 308 + ], + "score": 0.51, + "latex": "5 5 . 9 8 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 508, + 1252, + 611, + 1252, + 611, + 1273, + 508, + 1273 + ], + "score": 0.51, + "latex": "3 0 . 4 0 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1252, + 1198, + 1252, + 1198, + 1273, + 1095, + 1273 + ], + "score": 0.51, + "latex": "4 3 . 0 0 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 393, + 943, + 491, + 943, + 491, + 965, + 393, + 965 + ], + "score": 0.51, + "latex": "\\overline { { \\mathbf { C } 1 3 \\to \\mathbf { C } 0 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1124, + 1198, + 1124, + 1198, + 1141, + 1096, + 1141 + ], + "score": 0.5, + "latex": "3 9 . 8 0 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 391, + 1274, + 493, + 1274, + 493, + 1294, + 391, + 1294 + ], + "score": 0.5, + "latex": "\\overline { { 7 0 . 1 0 \\pm 0 . 2 7 } }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 967, + 1315, + 967, + 1315, + 988, + 1214, + 988 + ], + "score": 0.49, + "latex": "\\overline { { 6 0 . 5 7 \\pm 0 . 3 6 } }" + }, + { + "category_id": 13, + "poly": [ + 947, + 455, + 1055, + 455, + 1055, + 473, + 947, + 473 + ], + "score": 0.49, + "latex": "3 8 . 3 2 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 946, + 360, + 1054, + 360, + 1054, + 378, + 946, + 378 + ], + "score": 0.48, + "latex": "3 5 . 5 0 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 862, + 1033, + 963, + 1033, + 963, + 1053, + 862, + 1053 + ], + "score": 0.48, + "latex": "5 7 . 2 0 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 946, + 384, + 1054, + 384, + 1054, + 402, + 946, + 402 + ], + "score": 0.47, + "latex": "3 0 . 0 4 \\pm 0 . 9 8" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1033, + 1198, + 1033, + 1198, + 1053, + 1095, + 1053 + ], + "score": 0.47, + "latex": "4 0 . 7 7 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 978, + 1077, + 1080, + 1077, + 1080, + 1097, + 978, + 1097 + ], + "score": 0.46, + "latex": "3 0 . 6 0 \\pm 0 . 6 1" + }, + { + "category_id": 13, + "poly": [ + 391, + 1077, + 493, + 1077, + 493, + 1097, + 391, + 1097 + ], + "score": 0.46, + "latex": "5 9 . 2 0 \\pm 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 625, + 1252, + 728, + 1252, + 728, + 1273, + 625, + 1273 + ], + "score": 0.46, + "latex": "4 7 . 8 0 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1077, + 1198, + 1077, + 1198, + 1097, + 1095, + 1097 + ], + "score": 0.45, + "latex": "4 5 . 3 4 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1056, + 1197, + 1056, + 1197, + 1075, + 1095, + 1075 + ], + "score": 0.44, + "latex": "3 8 . 6 8 \\pm 0 . 9 5" + }, + { + "category_id": 13, + "poly": [ + 1095, + 968, + 1198, + 968, + 1198, + 989, + 1095, + 989 + ], + "score": 0.44, + "latex": "\\overline { { 4 0 . 1 1 \\pm 0 . 5 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1033, + 1316, + 1033, + 1316, + 1053, + 1214, + 1053 + ], + "score": 0.44, + "latex": "7 0 . 5 0 \\pm 0 . 5 5" + }, + { + "category_id": 13, + "poly": [ + 978, + 1033, + 1080, + 1033, + 1080, + 1053, + 978, + 1053 + ], + "score": 0.43, + "latex": "4 0 . 2 2 \\pm 0 . 5 5" + }, + { + "category_id": 13, + "poly": [ + 1198, + 455, + 1305, + 455, + 1305, + 473, + 1198, + 473 + ], + "score": 0.43, + "latex": "7 9 . 3 1 \\pm 0 . 7 8" + }, + { + "category_id": 13, + "poly": [ + 507, + 1056, + 610, + 1056, + 610, + 1075, + 507, + 1075 + ], + "score": 0.43, + "latex": "\\overline { { 6 3 . 2 0 \\pm 0 . 4 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 338, + 1180, + 338, + 1180, + 356, + 1070, + 356 + ], + "score": 0.43, + "latex": "5 7 . 7 0 \\pm 0 . 6 9" + }, + { + "category_id": 13, + "poly": [ + 820, + 455, + 928, + 455, + 928, + 472, + 820, + 472 + ], + "score": 0.43, + "latex": "5 6 . 8 7 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 708, + 266, + 790, + 266, + 790, + 283, + 708, + 283 + ], + "score": 0.42, + "latex": "\\mathbf S \\to \\mathbf M \\mathbf M" + }, + { + "category_id": 13, + "poly": [ + 743, + 1033, + 845, + 1033, + 845, + 1053, + 743, + 1053 + ], + "score": 0.42, + "latex": "7 0 . 6 8 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 992, + 508, + 1007, + 508, + 1007, + 520, + 992, + 520 + ], + "score": 0.42, + "latex": "-" + }, + { + "category_id": 13, + "poly": [ + 694, + 408, + 802, + 408, + 802, + 425, + 694, + 425 + ], + "score": 0.42, + "latex": "5 6 . 8 0 \\pm 0 . 7 9" + }, + { + "category_id": 13, + "poly": [ + 508, + 1078, + 610, + 1078, + 610, + 1097, + 508, + 1097 + ], + "score": 0.42, + "latex": "3 0 . 7 0 \\pm 0 . 6 3" + }, + { + "category_id": 13, + "poly": [ + 507, + 1123, + 610, + 1123, + 610, + 1141, + 507, + 1141 + ], + "score": 0.41, + "latex": "5 5 . 2 4 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 391, + 1252, + 495, + 1252, + 495, + 1273, + 391, + 1273 + ], + "score": 0.41, + "latex": "5 5 . 2 0 \\pm 0 . 4 6" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1230, + 1198, + 1230, + 1198, + 1250, + 1095, + 1250 + ], + "score": 0.4, + "latex": "4 2 . 0 3 \\pm 0 . 2 4" + }, + { + "category_id": 13, + "poly": [ + 978, + 1123, + 1080, + 1123, + 1080, + 1141, + 978, + 1141 + ], + "score": 0.4, + "latex": "3 8 . 9 7 \\pm 0 . 7 1" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1101, + 1315, + 1101, + 1315, + 1119, + 1214, + 1119 + ], + "score": 0.4, + "latex": "7 0 . 6 7 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1011, + 1316, + 1011, + 1316, + 1031, + 1212, + 1031 + ], + "score": 0.4, + "latex": "6 5 . 1 1 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1198, + 408, + 1305, + 408, + 1305, + 425, + 1198, + 425 + ], + "score": 0.4, + "latex": "8 4 . 1 0 \\pm 0 . 9 8" + }, + { + "category_id": 13, + "poly": [ + 946, + 408, + 1055, + 408, + 1055, + 425, + 946, + 425 + ], + "score": 0.4, + "latex": "2 7 . 5 0 \\pm 0 . 8 6" + }, + { + "category_id": 13, + "poly": [ + 508, + 943, + 610, + 943, + 610, + 966, + 508, + 966 + ], + "score": 0.4, + "latex": "\\overline { { \\mathbf { C } 1 3 \\to \\mathbf { C } 0 8 } }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1077, + 1315, + 1077, + 1315, + 1097, + 1213, + 1097 + ], + "score": 0.39, + "latex": "6 4 . 3 0 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 861, + 1124, + 962, + 1124, + 962, + 1141, + 861, + 1141 + ], + "score": 0.39, + "latex": "6 3 . 8 0 \\pm 0 . 4 2" + }, + { + "category_id": 13, + "poly": [ + 743, + 1275, + 845, + 1275, + 845, + 1294, + 743, + 1294 + ], + "score": 0.39, + "latex": "\\overline { { 8 0 . 1 2 \\pm 0 . 1 5 } }" + }, + { + "category_id": 13, + "poly": [ + 391, + 1055, + 493, + 1055, + 493, + 1075, + 391, + 1075 + ], + "score": 0.38, + "latex": "\\overline { { 6 4 . 8 3 \\pm 0 . 6 9 } }" + }, + { + "category_id": 13, + "poly": [ + 626, + 1056, + 727, + 1056, + 727, + 1075, + 626, + 1075 + ], + "score": 0.38, + "latex": "\\overline { { 5 5 . 4 8 \\pm 0 . 6 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1124, + 1315, + 1124, + 1315, + 1141, + 1213, + 1141 + ], + "score": 0.38, + "latex": "6 2 . 1 0 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 509, + 1230, + 611, + 1230, + 611, + 1251, + 509, + 1251 + ], + "score": 0.37, + "latex": "3 0 . 0 3 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 861, + 1078, + 963, + 1078, + 963, + 1097, + 861, + 1097 + ], + "score": 0.37, + "latex": "6 5 . 8 8 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 820, + 360, + 928, + 360, + 928, + 378, + 820, + 378 + ], + "score": 0.37, + "latex": "6 2 . 5 4 \\pm 0 . 9 1" + }, + { + "category_id": 13, + "poly": [ + 391, + 1101, + 493, + 1101, + 493, + 1119, + 391, + 1119 + ], + "score": 0.37, + "latex": "6 3 . 7 8 \\pm 0 . 2 9" + }, + { + "category_id": 13, + "poly": [ + 308, + 846, + 385, + 846, + 385, + 873, + 308, + 873 + ], + "score": 0.37, + "latex": "\\bf \\Pi ^ { * } U N I T" + }, + { + "category_id": 13, + "poly": [ + 1071, + 431, + 1180, + 431, + 1180, + 448, + 1071, + 448 + ], + "score": 0.37, + "latex": "8 5 . 7 0 \\pm 0 . 3 9" + }, + { + "category_id": 13, + "poly": [ + 744, + 1056, + 845, + 1056, + 845, + 1075, + 744, + 1075 + ], + "score": 0.37, + "latex": "\\overline { { 7 4 . 2 5 \\pm 0 . 5 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1198, + 360, + 1306, + 360, + 1306, + 378, + 1198, + 378 + ], + "score": 0.37, + "latex": "8 1 . 6 0 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 992, + 483, + 1007, + 483, + 1007, + 497, + 992, + 497 + ], + "score": 0.37, + "latex": "-" + }, + { + "category_id": 13, + "poly": [ + 743, + 1101, + 845, + 1101, + 845, + 1119, + 743, + 1119 + ], + "score": 0.36, + "latex": "7 2 . 6 0 \\pm 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 567, + 638, + 677, + 638, + 677, + 660, + 567, + 660 + ], + "score": 0.36, + "latex": "7 9 . 2 0 \\pm 0 . 2 8" + }, + { + "category_id": 13, + "poly": [ + 1069, + 593, + 1180, + 593, + 1180, + 613, + 1069, + 613 + ], + "score": 0.36, + "latex": "7 9 . 3 0 \\pm 0 . 5 9" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1209, + 1198, + 1209, + 1198, + 1228, + 1095, + 1228 + ], + "score": 0.35, + "latex": "4 2 . 6 3 \\pm 0 . 5 5" + }, + { + "category_id": 13, + "poly": [ + 628, + 943, + 727, + 943, + 727, + 965, + 628, + 965 + ], + "score": 0.35, + "latex": "\\overline { { \\mathbf { C } 1 3 \\to \\mathbf { C } 0 9 } }" + }, + { + "category_id": 13, + "poly": [ + 743, + 1124, + 845, + 1124, + 845, + 1141, + 743, + 1141 + ], + "score": 0.35, + "latex": "6 5 . 1 6 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1081, + 267, + 1171, + 267, + 1171, + 283, + 1081, + 283 + ], + "score": 0.35, + "latex": "\\mathbf { M } \\to \\mathbf { M M }" + }, + { + "category_id": 13, + "poly": [ + 977, + 1274, + 1080, + 1274, + 1080, + 1294, + 977, + 1294 + ], + "score": 0.34, + "latex": "\\overline { { 4 5 . 9 0 \\pm 0 . 3 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1197, + 384, + 1306, + 384, + 1306, + 402, + 1197, + 402 + ], + "score": 0.34, + "latex": "9 0 . 5 1 \\pm 0 . 8 0" + }, + { + "category_id": 13, + "poly": [ + 627, + 1101, + 727, + 1101, + 727, + 1119, + 627, + 1119 + ], + "score": 0.34, + "latex": "6 0 . 5 5 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1198, + 338, + 1306, + 338, + 1306, + 355, + 1198, + 355 + ], + "score": 0.34, + "latex": "8 1 . 0 5 \\pm 0 . 8 0" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1253, + 1316, + 1253, + 1316, + 1273, + 1212, + 1273 + ], + "score": 0.34, + "latex": "7 9 . 0 2 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 693, + 383, + 802, + 383, + 802, + 402, + 693, + 402 + ], + "score": 0.34, + "latex": "6 4 . 2 3 \\pm 0 . 7 0" + }, + { + "category_id": 13, + "poly": [ + 945, + 548, + 1054, + 548, + 1054, + 567, + 945, + 567 + ], + "score": 0.34, + "latex": "4 6 . 8 0 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 821, + 290, + 926, + 290, + 926, + 309, + 821, + 309 + ], + "score": 0.33, + "latex": "3 9 . 9 0 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 820, + 430, + 928, + 430, + 928, + 448, + 820, + 448 + ], + "score": 0.33, + "latex": "6 0 . 1 2 \\pm 0 . 5 2" + }, + { + "category_id": 13, + "poly": [ + 946, + 338, + 1054, + 338, + 1054, + 355, + 946, + 355 + ], + "score": 0.33, + "latex": "3 3 . 4 0 \\pm 0 . 7 4" + }, + { + "category_id": 13, + "poly": [ + 740, + 508, + 756, + 508, + 756, + 519, + 740, + 519 + ], + "score": 0.33, + "latex": "^ -" + }, + { + "category_id": 13, + "poly": [ + 741, + 483, + 756, + 483, + 756, + 497, + 741, + 497 + ], + "score": 0.33, + "latex": "^ -" + }, + { + "category_id": 13, + "poly": [ + 625, + 1230, + 729, + 1230, + 729, + 1251, + 625, + 1251 + ], + "score": 0.33, + "latex": "4 9 . 0 6 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 693, + 592, + 803, + 592, + 803, + 613, + 693, + 613 + ], + "score": 0.32, + "latex": "6 1 . 0 0 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 861, + 1056, + 962, + 1056, + 962, + 1075, + 861, + 1075 + ], + "score": 0.32, + "latex": "\\overline { { 7 3 . 6 2 \\pm 0 . 7 5 } }" + }, + { + "category_id": 13, + "poly": [ + 819, + 592, + 929, + 592, + 929, + 613, + 819, + 613 + ], + "score": 0.32, + "latex": "5 8 . 1 0 \\pm 0 . 6 4" + }, + { + "category_id": 13, + "poly": [ + 1068, + 548, + 1180, + 548, + 1180, + 567, + 1068, + 567 + ], + "score": 0.32, + "latex": "8 0 . 4 9 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 508, + 1209, + 611, + 1209, + 611, + 1228, + 508, + 1228 + ], + "score": 0.32, + "latex": "3 1 . 6 3 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 864, + 1168, + 959, + 1168, + 959, + 1184, + 864, + 1184 + ], + "score": 0.32, + "latex": "4 3 . 0 3 \\pm 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 391, + 1230, + 494, + 1230, + 494, + 1250, + 391, + 1250 + ], + "score": 0.32, + "latex": "5 5 . 8 4 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1102, + 1198, + 1102, + 1198, + 1119, + 1096, + 1119 + ], + "score": 0.32, + "latex": "4 1 . 4 5 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 743, + 1252, + 845, + 1252, + 845, + 1273, + 743, + 1273 + ], + "score": 0.31, + "latex": "7 5 . 3 0 \\pm 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 1071, + 527, + 1180, + 527, + 1180, + 544, + 1071, + 544 + ], + "score": 0.31, + "latex": "8 3 . 2 0 \\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 625, + 1209, + 728, + 1209, + 728, + 1228, + 625, + 1228 + ], + "score": 0.31, + "latex": "5 1 . 1 7 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 392, + 1123, + 493, + 1123, + 493, + 1141, + 392, + 1141 + ], + "score": 0.31, + "latex": "5 7 . 5 3 \\pm 0 . 4 2" + }, + { + "category_id": 13, + "poly": [ + 861, + 1275, + 963, + 1275, + 963, + 1294, + 861, + 1294 + ], + "score": 0.31, + "latex": "\\overline { { 8 2 . 0 5 \\pm 0 . 1 8 } }" + }, + { + "category_id": 13, + "poly": [ + 981, + 943, + 1079, + 943, + 1079, + 965, + 981, + 965 + ], + "score": 0.31, + "latex": "\\overline { { \\mathbf { C } 1 4 \\to \\mathbf { C } 0 8 } }" + }, + { + "category_id": 13, + "poly": [ + 568, + 289, + 676, + 289, + 676, + 308, + 568, + 308 + ], + "score": 0.3, + "latex": "6 2 . 1 0 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 743, + 1078, + 845, + 1078, + 845, + 1097, + 743, + 1097 + ], + "score": 0.3, + "latex": "6 8 . 3 3 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 508, + 1101, + 610, + 1101, + 610, + 1119, + 508, + 1119 + ], + "score": 0.3, + "latex": "6 0 . 4 5 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 945, + 591, + 1055, + 591, + 1055, + 613, + 945, + 613 + ], + "score": 0.29, + "latex": "4 7 . 3 5 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 626, + 1123, + 728, + 1123, + 728, + 1141, + 626, + 1141 + ], + "score": 0.29, + "latex": "5 7 . 1 4 \\pm 0 . 3 9" + }, + { + "category_id": 13, + "poly": [ + 1068, + 638, + 1180, + 638, + 1180, + 660, + 1068, + 660 + ], + "score": 0.29, + "latex": "7 7 . 1 2 \\pm 0 . 2 2" + }, + { + "category_id": 13, + "poly": [ + 862, + 1101, + 963, + 1101, + 963, + 1119, + 862, + 1119 + ], + "score": 0.29, + "latex": "7 0 . 6 7 \\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 977, + 1056, + 1080, + 1056, + 1080, + 1075, + 977, + 1075 + ], + "score": 0.29, + "latex": "\\overline { { 4 3 . 5 6 \\pm 0 . 9 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 943, + 1196, + 943, + 1196, + 966, + 1103, + 966 + ], + "score": 0.28, + "latex": "\\overline { { \\mathbf { C } 1 4 \\to \\mathbf { C } 0 9 } }" + }, + { + "category_id": 13, + "poly": [ + 977, + 1252, + 1081, + 1252, + 1081, + 1272, + 977, + 1272 + ], + "score": 0.28, + "latex": "3 0 . 2 0 \\pm 0 . 5 5" + }, + { + "category_id": 13, + "poly": [ + 391, + 1295, + 495, + 1295, + 495, + 1316, + 391, + 1316 + ], + "score": 0.28, + "latex": "6 0 . 3 4 \\pm 0 . 1 7" + }, + { + "category_id": 13, + "poly": [ + 947, + 528, + 1054, + 528, + 1054, + 544, + 947, + 544 + ], + "score": 0.28, + "latex": "4 9 . 3 0 \\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 568, + 450, + 677, + 450, + 677, + 474, + 568, + 474 + ], + "score": 0.28, + "latex": "8 2 . 1 0 \\pm 0 . 6 2" + }, + { + "category_id": 13, + "poly": [ + 977, + 1167, + 1080, + 1167, + 1080, + 1185, + 977, + 1185 + ], + "score": 0.28, + "latex": "4 4 . 4 4 \\pm 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1147, + 1315, + 1147, + 1315, + 1163, + 1213, + 1163 + ], + "score": 0.27, + "latex": "6 9 . 1 5 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1188, + 1198, + 1188, + 1198, + 1207, + 1095, + 1207 + ], + "score": 0.27, + "latex": "\\overline { { 4 5 . 0 0 \\pm 0 . 2 5 } }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1275, + 1315, + 1275, + 1315, + 1294, + 1213, + 1294 + ], + "score": 0.27, + "latex": "\\overline { { 8 1 . 6 0 \\pm 0 . 2 4 } }" + }, + { + "category_id": 13, + "poly": [ + 1215, + 1230, + 1314, + 1230, + 1314, + 1250, + 1215, + 1250 + ], + "score": 0.27, + "latex": "7 8 . 7 8 \\pm 0 . 5 7" + }, + { + "category_id": 13, + "poly": [ + 819, + 547, + 928, + 547, + 928, + 567, + 819, + 567 + ], + "score": 0.27, + "latex": "6 0 . 3 5 \\pm 0 . 5 9" + }, + { + "category_id": 13, + "poly": [ + 867, + 483, + 882, + 483, + 882, + 497, + 867, + 497 + ], + "score": 0.26, + "latex": "^ -" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1055, + 1316, + 1055, + 1316, + 1075, + 1214, + 1075 + ], + "score": 0.26, + "latex": "\\overline { { 7 2 . 8 4 \\pm 0 . 7 5 } }" + }, + { + "category_id": 13, + "poly": [ + 891, + 944, + 960, + 944, + 960, + 965, + 891, + 965 + ], + "score": 0.26, + "latex": "\\overline { { 4 \\mathrm { C } 0 5 } }" + }, + { + "category_id": 13, + "poly": [ + 742, + 1231, + 845, + 1231, + 845, + 1250, + 742, + 1250 + ], + "score": 0.26, + "latex": "7 2 . 1 1 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 860, + 1253, + 962, + 1253, + 962, + 1273, + 860, + 1273 + ], + "score": 0.26, + "latex": "8 0 . 7 5 \\pm 0 . 2 0" + }, + { + "category_id": 13, + "poly": [ + 686, + 638, + 1175, + 638, + 1175, + 658, + 686, + 658 + ], + "score": 0.26, + "latex": "| 5 9 . 9 0 \\pm 0 . 3 0 | | 6 3 . 7 0 \\pm 0 . 2 6 | | 4 5 . 3 0 \\pm 0 . 3 0 | | 7 7 . 1 2 \\pm 0 . 2 2" + }, + { + "category_id": 13, + "poly": [ + 978, + 1101, + 1080, + 1101, + 1080, + 1119, + 978, + 1119 + ], + "score": 0.26, + "latex": "4 1 . 5 5 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 391, + 1208, + 494, + 1208, + 494, + 1228, + 391, + 1228 + ], + "score": 0.25, + "latex": "5 7 . 3 4 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1101, + 1147, + 1198, + 1147, + 1198, + 1163, + 1101, + 1163 + ], + "score": 0.25, + "latex": "\\ d ^ { 4 4 . 4 5 } \\pm 0 . 5 1" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 731.0, + 1452.0, + 731.0, + 1452.0, + 767.0, + 248.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 760.0, + 1453.0, + 760.0, + 1453.0, + 794.0, + 248.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 787.0, + 1453.0, + 787.0, + 1453.0, + 822.0, + 248.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 818.0, + 1448.0, + 818.0, + 1448.0, + 849.0, + 250.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 845.0, + 307.0, + 845.0, + 307.0, + 878.0, + 247.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 845.0, + 710.0, + 845.0, + 710.0, + 878.0, + 386.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 845.0, + 1171.0, + 845.0, + 1171.0, + 878.0, + 799.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 845.0, + 1450.0, + 845.0, + 1450.0, + 878.0, + 1281.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 871.0, + 894.0, + 871.0, + 894.0, + 905.0, + 249.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 109.0, + 813.0, + 109.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1469.0, + 562.0, + 1469.0, + 562.0, + 1508.0, + 250.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1863.0, + 816.0, + 1863.0, + 816.0, + 1894.0, + 794.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1389.0, + 1447.0, + 1389.0, + 1447.0, + 1419.0, + 250.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1416.0, + 679.0, + 1416.0, + 679.0, + 1446.0, + 249.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1533.0, + 1449.0, + 1533.0, + 1449.0, + 1570.0, + 250.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1564.0, + 1448.0, + 1564.0, + 1448.0, + 1601.0, + 249.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1593.0, + 1261.0, + 1593.0, + 1261.0, + 1635.0, + 248.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1593.0, + 1449.0, + 1593.0, + 1449.0, + 1635.0, + 1345.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1627.0, + 1053.0, + 1627.0, + 1053.0, + 1664.0, + 249.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1627.0, + 1452.0, + 1627.0, + 1452.0, + 1664.0, + 1137.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1657.0, + 923.0, + 1657.0, + 923.0, + 1694.0, + 249.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1657.0, + 1019.0, + 1657.0, + 1019.0, + 1694.0, + 1008.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1689.0, + 1450.0, + 1689.0, + 1450.0, + 1723.0, + 250.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1716.0, + 1057.0, + 1716.0, + 1057.0, + 1755.0, + 248.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1716.0, + 1449.0, + 1716.0, + 1449.0, + 1755.0, + 1158.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1743.0, + 1055.0, + 1743.0, + 1055.0, + 1788.0, + 246.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1743.0, + 1202.0, + 1743.0, + 1202.0, + 1788.0, + 1153.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1743.0, + 1453.0, + 1743.0, + 1453.0, + 1788.0, + 1292.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1776.0, + 1451.0, + 1776.0, + 1451.0, + 1818.0, + 247.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1389.0, + 1447.0, + 1389.0, + 1447.0, + 1419.0, + 250.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1416.0, + 679.0, + 1416.0, + 679.0, + 1446.0, + 249.0, + 1446.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 253, + 758, + 1449, + 758, + 1449, + 1126, + 253, + 1126 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 253, + 1140, + 1450, + 1140, + 1450, + 1416, + 253, + 1416 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 253, + 1565, + 1449, + 1565, + 1449, + 1811, + 253, + 1811 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 253, + 451, + 1448, + 451, + 1448, + 633, + 253, + 633 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 254, + 267, + 1448, + 267, + 1448, + 449, + 254, + 449 + ], + "score": 0.972 + }, + { + "category_id": 2, + "poly": [ + 253, + 113, + 812, + 113, + 812, + 141, + 253, + 141 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 254, + 1483, + 499, + 1483, + 499, + 1518, + 254, + 1518 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 791, + 1866, + 819, + 1866, + 819, + 1889, + 791, + 1889 + ], + "score": 0.841 + }, + { + "category_id": 0, + "poly": [ + 256, + 693, + 628, + 693, + 628, + 724, + 256, + 724 + ], + "score": 0.739 + }, + { + "category_id": 0, + "poly": [ + 254, + 693, + 627, + 693, + 627, + 724, + 254, + 724 + ], + "score": 0.457 + }, + { + "category_id": 13, + "poly": [ + 1091, + 1173, + 1127, + 1173, + 1127, + 1206, + 1091, + 1206 + ], + "score": 0.9, + "latex": "E _ { p }" + }, + { + "category_id": 13, + "poly": [ + 879, + 510, + 950, + 510, + 950, + 541, + 879, + 541 + ], + "score": 0.88, + "latex": "\\approx 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1338, + 913, + 1364, + 913, + 1364, + 940, + 1338, + 940 + ], + "score": 0.77, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 899, + 1204, + 932, + 1204, + 932, + 1235, + 899, + 1235 + ], + "score": 0.35, + "latex": "6 \\mathrm { g }" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 253.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1479.0, + 505.0, + 1479.0, + 505.0, + 1527.0, + 246.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 824.0, + 1862.0, + 824.0, + 1901.0, + 786.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 688.0, + 633.0, + 688.0, + 633.0, + 729.0, + 248.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 690.0, + 631.0, + 690.0, + 631.0, + 728.0, + 249.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 760.0, + 1449.0, + 760.0, + 1449.0, + 793.0, + 250.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 788.0, + 1449.0, + 788.0, + 1449.0, + 824.0, + 249.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 817.0, + 1450.0, + 817.0, + 1450.0, + 856.0, + 248.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 848.0, + 1452.0, + 848.0, + 1452.0, + 888.0, + 248.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 881.0, + 1453.0, + 881.0, + 1453.0, + 918.0, + 248.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 911.0, + 1337.0, + 911.0, + 1337.0, + 946.0, + 246.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 911.0, + 1450.0, + 911.0, + 1450.0, + 946.0, + 1365.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 941.0, + 1456.0, + 941.0, + 1456.0, + 979.0, + 245.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 974.0, + 1450.0, + 974.0, + 1450.0, + 1007.0, + 250.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1000.0, + 1454.0, + 1000.0, + 1454.0, + 1040.0, + 248.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1032.0, + 1449.0, + 1032.0, + 1449.0, + 1070.0, + 249.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1063.0, + 1452.0, + 1063.0, + 1452.0, + 1100.0, + 249.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1095.0, + 1008.0, + 1095.0, + 1008.0, + 1132.0, + 249.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1141.0, + 1454.0, + 1141.0, + 1454.0, + 1174.0, + 249.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1171.0, + 1090.0, + 1171.0, + 1090.0, + 1204.0, + 250.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1171.0, + 1450.0, + 1171.0, + 1450.0, + 1204.0, + 1128.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1204.0, + 898.0, + 1204.0, + 898.0, + 1236.0, + 250.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1204.0, + 1450.0, + 1204.0, + 1450.0, + 1236.0, + 933.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1228.0, + 1454.0, + 1228.0, + 1454.0, + 1273.0, + 246.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1264.0, + 1451.0, + 1264.0, + 1451.0, + 1300.0, + 248.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1291.0, + 1451.0, + 1291.0, + 1451.0, + 1329.0, + 248.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1324.0, + 1450.0, + 1324.0, + 1450.0, + 1360.0, + 249.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1351.0, + 1455.0, + 1351.0, + 1455.0, + 1393.0, + 245.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1384.0, + 1380.0, + 1384.0, + 1380.0, + 1420.0, + 248.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1565.0, + 1452.0, + 1565.0, + 1452.0, + 1603.0, + 249.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1597.0, + 1449.0, + 1597.0, + 1449.0, + 1631.0, + 249.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1627.0, + 1454.0, + 1627.0, + 1454.0, + 1664.0, + 245.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1658.0, + 1450.0, + 1658.0, + 1450.0, + 1692.0, + 249.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1689.0, + 1450.0, + 1689.0, + 1450.0, + 1723.0, + 249.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1718.0, + 1450.0, + 1718.0, + 1450.0, + 1756.0, + 248.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1749.0, + 1450.0, + 1749.0, + 1450.0, + 1787.0, + 249.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1779.0, + 1286.0, + 1779.0, + 1286.0, + 1815.0, + 249.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 450.0, + 1448.0, + 450.0, + 1448.0, + 482.0, + 250.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 478.0, + 1449.0, + 478.0, + 1449.0, + 514.0, + 248.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 511.0, + 878.0, + 511.0, + 878.0, + 546.0, + 249.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 511.0, + 1453.0, + 511.0, + 1453.0, + 546.0, + 951.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 539.0, + 1451.0, + 539.0, + 1451.0, + 576.0, + 248.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 573.0, + 1449.0, + 573.0, + 1449.0, + 605.0, + 250.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 602.0, + 1080.0, + 602.0, + 1080.0, + 638.0, + 250.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 266.0, + 1449.0, + 266.0, + 1449.0, + 301.0, + 250.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 299.0, + 1449.0, + 299.0, + 1449.0, + 330.0, + 250.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 327.0, + 1453.0, + 327.0, + 1453.0, + 364.0, + 247.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 359.0, + 1451.0, + 359.0, + 1451.0, + 391.0, + 250.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 388.0, + 1451.0, + 388.0, + 1451.0, + 423.0, + 248.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 418.0, + 900.0, + 418.0, + 900.0, + 454.0, + 250.0, + 454.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 253, + 112, + 813, + 112, + 813, + 142, + 253, + 142 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 790, + 1865, + 817, + 1865, + 817, + 1890, + 790, + 1890 + ], + "score": 0.8 + }, + { + "category_id": 0, + "poly": [ + 254, + 264, + 445, + 264, + 445, + 299, + 254, + 299 + ], + "score": 0.769 + }, + { + "category_id": 1, + "poly": [ + 249, + 957, + 1440, + 957, + 1440, + 1023, + 249, + 1023 + ], + "score": 0.574 + }, + { + "category_id": 1, + "poly": [ + 248, + 316, + 1453, + 316, + 1453, + 410, + 248, + 410 + ], + "score": 0.555 + }, + { + "category_id": 1, + "poly": [ + 249, + 1748, + 1447, + 1748, + 1447, + 1811, + 249, + 1811 + ], + "score": 0.524 + }, + { + "category_id": 1, + "poly": [ + 249, + 782, + 1448, + 782, + 1448, + 849, + 249, + 849 + ], + "score": 0.521 + }, + { + "category_id": 1, + "poly": [ + 246, + 433, + 1440, + 433, + 1440, + 498, + 246, + 498 + ], + "score": 0.519 + }, + { + "category_id": 1, + "poly": [ + 251, + 870, + 1446, + 870, + 1446, + 936, + 251, + 936 + ], + "score": 0.504 + }, + { + "category_id": 1, + "poly": [ + 253, + 1043, + 1445, + 1043, + 1445, + 1111, + 253, + 1111 + ], + "score": 0.496 + }, + { + "category_id": 1, + "poly": [ + 254, + 1131, + 1443, + 1131, + 1443, + 1228, + 254, + 1228 + ], + "score": 0.494 + }, + { + "category_id": 1, + "poly": [ + 249, + 695, + 1449, + 695, + 1449, + 760, + 249, + 760 + ], + "score": 0.486 + }, + { + "category_id": 1, + "poly": [ + 254, + 1512, + 1444, + 1512, + 1444, + 1607, + 254, + 1607 + ], + "score": 0.463 + }, + { + "category_id": 1, + "poly": [ + 256, + 519, + 1445, + 519, + 1445, + 587, + 256, + 587 + ], + "score": 0.459 + }, + { + "category_id": 1, + "poly": [ + 250, + 1336, + 1441, + 1336, + 1441, + 1404, + 250, + 1404 + ], + "score": 0.457 + }, + { + "category_id": 1, + "poly": [ + 250, + 1629, + 1446, + 1629, + 1446, + 1724, + 250, + 1724 + ], + "score": 0.453 + }, + { + "category_id": 1, + "poly": [ + 255, + 608, + 1447, + 608, + 1447, + 674, + 255, + 674 + ], + "score": 0.452 + }, + { + "category_id": 1, + "poly": [ + 246, + 1248, + 1445, + 1248, + 1445, + 1316, + 246, + 1316 + ], + "score": 0.421 + }, + { + "category_id": 1, + "poly": [ + 248, + 1423, + 1438, + 1423, + 1438, + 1490, + 248, + 1490 + ], + "score": 0.356 + }, + { + "category_id": 1, + "poly": [ + 254, + 264, + 445, + 264, + 445, + 299, + 254, + 299 + ], + "score": 0.092 + }, + { + "category_id": 15, + "poly": [ + 251.0, + 107.0, + 814.0, + 107.0, + 814.0, + 146.0, + 251.0, + 146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 822.0, + 1862.0, + 822.0, + 1902.0, + 786.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 267.0, + 445.0, + 267.0, + 445.0, + 300.0, + 251.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 958.0, + 1446.0, + 958.0, + 1446.0, + 994.0, + 250.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 987.0, + 816.0, + 987.0, + 816.0, + 1022.0, + 276.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 314.0, + 1451.0, + 314.0, + 1451.0, + 354.0, + 248.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 345.0, + 1453.0, + 345.0, + 1453.0, + 389.0, + 274.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 378.0, + 532.0, + 378.0, + 532.0, + 412.0, + 278.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1743.0, + 1452.0, + 1743.0, + 1452.0, + 1788.0, + 246.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1777.0, + 353.0, + 1777.0, + 353.0, + 1813.0, + 276.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 779.0, + 1449.0, + 779.0, + 1449.0, + 822.0, + 249.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 816.0, + 889.0, + 816.0, + 889.0, + 848.0, + 279.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 432.0, + 1446.0, + 432.0, + 1446.0, + 472.0, + 250.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 466.0, + 1239.0, + 466.0, + 1239.0, + 499.0, + 277.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 873.0, + 1450.0, + 873.0, + 1450.0, + 906.0, + 251.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 903.0, + 1307.0, + 903.0, + 1307.0, + 935.0, + 272.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1047.0, + 1449.0, + 1047.0, + 1449.0, + 1080.0, + 251.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1077.0, + 1449.0, + 1077.0, + 1449.0, + 1111.0, + 277.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1131.0, + 1450.0, + 1131.0, + 1450.0, + 1170.0, + 250.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1163.0, + 1450.0, + 1163.0, + 1450.0, + 1202.0, + 278.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 1194.0, + 994.0, + 1194.0, + 994.0, + 1229.0, + 279.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 695.0, + 1453.0, + 695.0, + 1453.0, + 731.0, + 250.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 728.0, + 742.0, + 728.0, + 742.0, + 760.0, + 278.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1512.0, + 1448.0, + 1512.0, + 1448.0, + 1550.0, + 251.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 1541.0, + 1451.0, + 1541.0, + 1451.0, + 1581.0, + 275.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1573.0, + 990.0, + 1573.0, + 990.0, + 1611.0, + 278.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 519.0, + 1449.0, + 519.0, + 1449.0, + 558.0, + 249.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 552.0, + 1149.0, + 552.0, + 1149.0, + 587.0, + 276.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1336.0, + 1447.0, + 1336.0, + 1447.0, + 1374.0, + 248.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1367.0, + 1175.0, + 1367.0, + 1175.0, + 1404.0, + 278.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1628.0, + 1450.0, + 1628.0, + 1450.0, + 1667.0, + 249.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1657.0, + 1451.0, + 1657.0, + 1451.0, + 1701.0, + 274.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1689.0, + 488.0, + 1689.0, + 488.0, + 1727.0, + 276.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 609.0, + 1451.0, + 609.0, + 1451.0, + 645.0, + 251.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 640.0, + 1045.0, + 640.0, + 1045.0, + 676.0, + 276.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1248.0, + 1449.0, + 1248.0, + 1449.0, + 1287.0, + 249.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1281.0, + 1261.0, + 1281.0, + 1261.0, + 1317.0, + 277.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1426.0, + 1442.0, + 1426.0, + 1442.0, + 1459.0, + 250.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 1455.0, + 1368.0, + 1455.0, + 1368.0, + 1492.0, + 279.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 267.0, + 445.0, + 267.0, + 445.0, + 300.0, + 251.0, + 300.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 253, + 112, + 813, + 112, + 813, + 142, + 253, + 142 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 790, + 1865, + 819, + 1865, + 819, + 1890, + 790, + 1890 + ], + "score": 0.814 + }, + { + "category_id": 1, + "poly": [ + 250, + 883, + 1445, + 883, + 1445, + 947, + 250, + 947 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 254, + 1325, + 1449, + 1325, + 1449, + 1421, + 254, + 1421 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 249, + 1242, + 1448, + 1242, + 1448, + 1308, + 249, + 1308 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 251, + 378, + 1443, + 378, + 1443, + 443, + 251, + 443 + ], + "score": 0.674 + }, + { + "category_id": 1, + "poly": [ + 252, + 266, + 1452, + 266, + 1452, + 360, + 252, + 360 + ], + "score": 0.673 + }, + { + "category_id": 1, + "poly": [ + 256, + 686, + 1451, + 686, + 1451, + 753, + 256, + 753 + ], + "score": 0.666 + }, + { + "category_id": 1, + "poly": [ + 255, + 1047, + 1449, + 1047, + 1449, + 1113, + 255, + 1113 + ], + "score": 0.665 + }, + { + "category_id": 1, + "poly": [ + 259, + 964, + 1447, + 964, + 1447, + 1031, + 259, + 1031 + ], + "score": 0.646 + }, + { + "category_id": 1, + "poly": [ + 248, + 1521, + 1447, + 1521, + 1447, + 1615, + 248, + 1615 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 244, + 1438, + 1448, + 1438, + 1448, + 1505, + 244, + 1505 + ], + "score": 0.641 + }, + { + "category_id": 1, + "poly": [ + 254, + 460, + 1451, + 460, + 1451, + 556, + 254, + 556 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 251, + 1717, + 1453, + 1717, + 1453, + 1811, + 251, + 1811 + ], + "score": 0.619 + }, + { + "category_id": 1, + "poly": [ + 246, + 1634, + 1448, + 1634, + 1448, + 1699, + 246, + 1699 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 253, + 573, + 1451, + 573, + 1451, + 668, + 253, + 668 + ], + "score": 0.611 + }, + { + "category_id": 1, + "poly": [ + 250, + 769, + 1453, + 769, + 1453, + 863, + 250, + 863 + ], + "score": 0.504 + }, + { + "category_id": 1, + "poly": [ + 251, + 1129, + 1451, + 1129, + 1451, + 1224, + 251, + 1224 + ], + "score": 0.474 + }, + { + "category_id": 15, + "poly": [ + 251.0, + 107.0, + 814.0, + 107.0, + 814.0, + 146.0, + 251.0, + 146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 824.0, + 1862.0, + 824.0, + 1901.0, + 786.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 879.0, + 1451.0, + 879.0, + 1451.0, + 921.0, + 247.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 915.0, + 758.0, + 915.0, + 758.0, + 947.0, + 277.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1326.0, + 1454.0, + 1326.0, + 1454.0, + 1361.0, + 251.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 273.0, + 1353.0, + 1453.0, + 1353.0, + 1453.0, + 1396.0, + 273.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1389.0, + 730.0, + 1389.0, + 730.0, + 1420.0, + 277.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1239.0, + 1455.0, + 1239.0, + 1455.0, + 1283.0, + 246.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1275.0, + 528.0, + 1275.0, + 528.0, + 1309.0, + 277.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 377.0, + 1449.0, + 377.0, + 1449.0, + 416.0, + 251.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 409.0, + 828.0, + 409.0, + 828.0, + 444.0, + 276.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 264.0, + 1451.0, + 264.0, + 1451.0, + 304.0, + 249.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 292.0, + 1453.0, + 292.0, + 1453.0, + 339.0, + 274.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 325.0, + 488.0, + 325.0, + 488.0, + 363.0, + 278.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 684.0, + 1451.0, + 684.0, + 1451.0, + 726.0, + 248.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 718.0, + 998.0, + 718.0, + 998.0, + 753.0, + 276.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1046.0, + 1449.0, + 1046.0, + 1449.0, + 1085.0, + 250.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1080.0, + 997.0, + 1080.0, + 997.0, + 1111.0, + 276.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 965.0, + 1451.0, + 965.0, + 1451.0, + 1001.0, + 253.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 996.0, + 1343.0, + 996.0, + 1343.0, + 1033.0, + 276.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1518.0, + 1451.0, + 1518.0, + 1451.0, + 1562.0, + 248.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1551.0, + 1450.0, + 1551.0, + 1450.0, + 1589.0, + 278.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 1582.0, + 628.0, + 1582.0, + 628.0, + 1620.0, + 275.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1439.0, + 1450.0, + 1439.0, + 1450.0, + 1475.0, + 250.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 271.0, + 1469.0, + 1179.0, + 1469.0, + 1179.0, + 1504.0, + 271.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 460.0, + 1454.0, + 460.0, + 1454.0, + 498.0, + 251.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 491.0, + 1452.0, + 491.0, + 1452.0, + 529.0, + 276.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 524.0, + 882.0, + 524.0, + 882.0, + 556.0, + 274.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1719.0, + 1449.0, + 1719.0, + 1449.0, + 1752.0, + 251.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1746.0, + 1454.0, + 1746.0, + 1454.0, + 1790.0, + 274.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1779.0, + 555.0, + 1779.0, + 555.0, + 1813.0, + 278.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1632.0, + 1452.0, + 1632.0, + 1452.0, + 1674.0, + 248.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1665.0, + 1221.0, + 1665.0, + 1221.0, + 1700.0, + 277.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 573.0, + 1451.0, + 573.0, + 1451.0, + 611.0, + 250.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 605.0, + 1454.0, + 605.0, + 1454.0, + 643.0, + 276.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 634.0, + 349.0, + 634.0, + 349.0, + 671.0, + 277.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 767.0, + 1453.0, + 767.0, + 1453.0, + 808.0, + 249.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 799.0, + 1451.0, + 799.0, + 1451.0, + 837.0, + 277.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 826.0, + 351.0, + 826.0, + 351.0, + 867.0, + 276.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1128.0, + 1448.0, + 1128.0, + 1448.0, + 1166.0, + 250.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1160.0, + 1454.0, + 1160.0, + 1454.0, + 1199.0, + 274.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1190.0, + 350.0, + 1190.0, + 350.0, + 1225.0, + 277.0, + 1225.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 253, + 112, + 813, + 112, + 813, + 142, + 253, + 142 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 790, + 1865, + 819, + 1865, + 819, + 1890, + 790, + 1890 + ], + "score": 0.807 + }, + { + "category_id": 1, + "poly": [ + 248, + 265, + 1446, + 265, + 1446, + 330, + 248, + 330 + ], + "score": 0.718 + }, + { + "category_id": 1, + "poly": [ + 247, + 1187, + 1449, + 1187, + 1449, + 1253, + 247, + 1253 + ], + "score": 0.681 + }, + { + "category_id": 1, + "poly": [ + 252, + 1104, + 1447, + 1104, + 1447, + 1169, + 252, + 1169 + ], + "score": 0.657 + }, + { + "category_id": 1, + "poly": [ + 247, + 1269, + 1449, + 1269, + 1449, + 1337, + 247, + 1337 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 253, + 1022, + 1449, + 1022, + 1449, + 1087, + 253, + 1087 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 252, + 938, + 1444, + 938, + 1444, + 1004, + 252, + 1004 + ], + "score": 0.593 + }, + { + "category_id": 1, + "poly": [ + 253, + 771, + 1448, + 771, + 1448, + 838, + 253, + 838 + ], + "score": 0.57 + }, + { + "category_id": 1, + "poly": [ + 252, + 461, + 1444, + 461, + 1444, + 527, + 252, + 527 + ], + "score": 0.569 + }, + { + "category_id": 1, + "poly": [ + 254, + 1353, + 1450, + 1353, + 1450, + 1450, + 254, + 1450 + ], + "score": 0.567 + }, + { + "category_id": 1, + "poly": [ + 257, + 855, + 1449, + 855, + 1449, + 921, + 257, + 921 + ], + "score": 0.565 + }, + { + "category_id": 1, + "poly": [ + 248, + 1467, + 1449, + 1467, + 1449, + 1533, + 248, + 1533 + ], + "score": 0.55 + }, + { + "category_id": 1, + "poly": [ + 249, + 544, + 1446, + 544, + 1446, + 640, + 249, + 640 + ], + "score": 0.537 + }, + { + "category_id": 1, + "poly": [ + 251, + 348, + 1451, + 348, + 1451, + 443, + 251, + 443 + ], + "score": 0.514 + }, + { + "category_id": 1, + "poly": [ + 247, + 1663, + 1447, + 1663, + 1447, + 1729, + 247, + 1729 + ], + "score": 0.499 + }, + { + "category_id": 1, + "poly": [ + 258, + 1550, + 1446, + 1550, + 1446, + 1646, + 258, + 1646 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 251, + 1748, + 1443, + 1748, + 1443, + 1812, + 251, + 1812 + ], + "score": 0.479 + }, + { + "category_id": 1, + "poly": [ + 252, + 658, + 1452, + 658, + 1452, + 754, + 252, + 754 + ], + "score": 0.278 + }, + { + "category_id": 15, + "poly": [ + 251.0, + 107.0, + 814.0, + 107.0, + 814.0, + 146.0, + 251.0, + 146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 823.0, + 1862.0, + 823.0, + 1900.0, + 786.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 261.0, + 1451.0, + 261.0, + 1451.0, + 305.0, + 247.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 298.0, + 689.0, + 298.0, + 689.0, + 327.0, + 275.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1188.0, + 1450.0, + 1188.0, + 1450.0, + 1224.0, + 249.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1219.0, + 941.0, + 1219.0, + 941.0, + 1252.0, + 276.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1106.0, + 1451.0, + 1106.0, + 1451.0, + 1142.0, + 249.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 1137.0, + 764.0, + 1137.0, + 764.0, + 1169.0, + 279.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1273.0, + 1453.0, + 1273.0, + 1453.0, + 1306.0, + 251.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1302.0, + 1378.0, + 1302.0, + 1378.0, + 1337.0, + 276.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1021.0, + 1449.0, + 1021.0, + 1449.0, + 1060.0, + 249.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 1052.0, + 1074.0, + 1052.0, + 1074.0, + 1088.0, + 279.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 937.0, + 1450.0, + 937.0, + 1450.0, + 976.0, + 248.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 970.0, + 1381.0, + 970.0, + 1381.0, + 1005.0, + 276.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 774.0, + 1449.0, + 774.0, + 1449.0, + 807.0, + 252.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 803.0, + 1319.0, + 803.0, + 1319.0, + 838.0, + 276.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 464.0, + 1447.0, + 464.0, + 1447.0, + 497.0, + 253.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 494.0, + 1370.0, + 494.0, + 1370.0, + 528.0, + 277.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1353.0, + 1451.0, + 1353.0, + 1451.0, + 1392.0, + 250.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1383.0, + 1450.0, + 1383.0, + 1450.0, + 1424.0, + 276.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1413.0, + 974.0, + 1413.0, + 974.0, + 1453.0, + 277.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 852.0, + 1449.0, + 852.0, + 1449.0, + 896.0, + 249.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 887.0, + 1115.0, + 887.0, + 1115.0, + 923.0, + 278.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1468.0, + 1454.0, + 1468.0, + 1454.0, + 1503.0, + 249.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1499.0, + 1111.0, + 1499.0, + 1111.0, + 1535.0, + 277.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 543.0, + 1451.0, + 543.0, + 1451.0, + 582.0, + 250.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 577.0, + 1451.0, + 577.0, + 1451.0, + 612.0, + 276.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 608.0, + 776.0, + 608.0, + 776.0, + 642.0, + 275.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 344.0, + 1451.0, + 344.0, + 1451.0, + 389.0, + 247.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 376.0, + 1454.0, + 376.0, + 1454.0, + 420.0, + 274.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 408.0, + 488.0, + 408.0, + 488.0, + 446.0, + 277.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1666.0, + 1450.0, + 1666.0, + 1450.0, + 1699.0, + 251.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 277.0, + 1696.0, + 1371.0, + 1696.0, + 1371.0, + 1730.0, + 277.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1550.0, + 1452.0, + 1550.0, + 1452.0, + 1588.0, + 251.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1583.0, + 1449.0, + 1583.0, + 1449.0, + 1618.0, + 278.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1614.0, + 758.0, + 1614.0, + 758.0, + 1648.0, + 278.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1748.0, + 1449.0, + 1748.0, + 1449.0, + 1784.0, + 251.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 1779.0, + 953.0, + 1779.0, + 953.0, + 1812.0, + 276.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 657.0, + 1453.0, + 657.0, + 1453.0, + 697.0, + 252.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 688.0, + 1451.0, + 688.0, + 1451.0, + 727.0, + 275.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 720.0, + 619.0, + 720.0, + 619.0, + 758.0, + 275.0, + 758.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 254, + 694, + 1448, + 694, + 1448, + 909, + 254, + 909 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 253, + 1396, + 1448, + 1396, + 1448, + 1673, + 253, + 1673 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 253, + 1093, + 1449, + 1093, + 1449, + 1276, + 253, + 1276 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 253, + 494, + 1448, + 494, + 1448, + 680, + 253, + 680 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 253, + 924, + 1447, + 924, + 1447, + 1077, + 253, + 1077 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 256, + 1688, + 1447, + 1688, + 1447, + 1811, + 256, + 1811 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 254, + 265, + 1450, + 265, + 1450, + 360, + 254, + 360 + ], + "score": 0.964 + }, + { + "category_id": 2, + "poly": [ + 253, + 112, + 812, + 112, + 812, + 141, + 253, + 141 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 253, + 1324, + 1175, + 1324, + 1175, + 1361, + 253, + 1361 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 791, + 1866, + 818, + 1866, + 818, + 1889, + 791, + 1889 + ], + "score": 0.822 + }, + { + "category_id": 0, + "poly": [ + 255, + 423, + 1364, + 423, + 1364, + 460, + 255, + 460 + ], + "score": 0.775 + }, + { + "category_id": 1, + "poly": [ + 255, + 423, + 1364, + 423, + 1364, + 460, + 255, + 460 + ], + "score": 0.14 + }, + { + "category_id": 13, + "poly": [ + 726, + 1610, + 889, + 1610, + 889, + 1644, + 726, + 1644 + ], + "score": 0.91, + "latex": "G ( \\pmb { x } , \\pmb { d } ) \\pmb { x } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1416, + 1581, + 1446, + 1581, + 1446, + 1608, + 1416, + 1608 + ], + "score": 0.85, + "latex": "\\mathbf { x } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 721, + 1581, + 746, + 1581, + 746, + 1608, + 721, + 1608 + ], + "score": 0.82, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1521, + 1103, + 1521, + 1103, + 1548, + 1078, + 1548 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1357, + 1612, + 1383, + 1612, + 1383, + 1638, + 1357, + 1638 + ], + "score": 0.79, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1157, + 1586, + 1179, + 1586, + 1179, + 1608, + 1157, + 1608 + ], + "score": 0.74, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 695, + 1612, + 715, + 1612, + 715, + 1638, + 695, + 1638 + ], + "score": 0.48, + "latex": "^ d" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 109.0, + 813.0, + 109.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1321.0, + 1180.0, + 1321.0, + 1180.0, + 1367.0, + 246.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1862.0, + 823.0, + 1862.0, + 823.0, + 1901.0, + 787.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 422.0, + 1366.0, + 422.0, + 1366.0, + 461.0, + 250.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 695.0, + 1451.0, + 695.0, + 1451.0, + 730.0, + 257.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 726.0, + 1451.0, + 726.0, + 1451.0, + 761.0, + 250.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 752.0, + 1451.0, + 752.0, + 1451.0, + 792.0, + 248.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 787.0, + 1451.0, + 787.0, + 1451.0, + 823.0, + 250.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 814.0, + 1452.0, + 814.0, + 1452.0, + 852.0, + 247.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 845.0, + 1452.0, + 845.0, + 1452.0, + 884.0, + 247.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 877.0, + 1285.0, + 877.0, + 1285.0, + 913.0, + 250.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1398.0, + 1452.0, + 1398.0, + 1452.0, + 1431.0, + 250.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1426.0, + 1451.0, + 1426.0, + 1451.0, + 1465.0, + 248.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1460.0, + 1451.0, + 1460.0, + 1451.0, + 1493.0, + 250.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1485.0, + 1451.0, + 1485.0, + 1451.0, + 1528.0, + 246.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1517.0, + 1077.0, + 1517.0, + 1077.0, + 1557.0, + 248.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1517.0, + 1451.0, + 1517.0, + 1451.0, + 1557.0, + 1104.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1545.0, + 1451.0, + 1545.0, + 1451.0, + 1590.0, + 245.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1579.0, + 720.0, + 1579.0, + 720.0, + 1616.0, + 249.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1579.0, + 1156.0, + 1579.0, + 1156.0, + 1616.0, + 747.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1579.0, + 1415.0, + 1579.0, + 1415.0, + 1616.0, + 1180.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1606.0, + 694.0, + 1606.0, + 694.0, + 1650.0, + 248.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1606.0, + 725.0, + 1606.0, + 725.0, + 1650.0, + 716.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1606.0, + 1356.0, + 1606.0, + 1356.0, + 1650.0, + 890.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1384.0, + 1606.0, + 1451.0, + 1606.0, + 1451.0, + 1650.0, + 1384.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1643.0, + 1298.0, + 1643.0, + 1298.0, + 1676.0, + 250.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1092.0, + 1450.0, + 1092.0, + 1450.0, + 1126.0, + 248.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1123.0, + 1449.0, + 1123.0, + 1449.0, + 1159.0, + 248.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1153.0, + 1453.0, + 1153.0, + 1453.0, + 1191.0, + 246.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1186.0, + 1453.0, + 1186.0, + 1453.0, + 1221.0, + 249.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1217.0, + 1450.0, + 1217.0, + 1450.0, + 1249.0, + 250.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 1248.0, + 761.0, + 1248.0, + 761.0, + 1279.0, + 252.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 491.0, + 1451.0, + 491.0, + 1451.0, + 532.0, + 246.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 524.0, + 1451.0, + 524.0, + 1451.0, + 561.0, + 248.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 558.0, + 1449.0, + 558.0, + 1449.0, + 590.0, + 250.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 587.0, + 1452.0, + 587.0, + 1452.0, + 623.0, + 248.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 612.0, + 1451.0, + 612.0, + 1451.0, + 656.0, + 246.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 647.0, + 1080.0, + 647.0, + 1080.0, + 684.0, + 248.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 923.0, + 1450.0, + 923.0, + 1450.0, + 957.0, + 249.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 956.0, + 1450.0, + 956.0, + 1450.0, + 989.0, + 252.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 986.0, + 1450.0, + 986.0, + 1450.0, + 1019.0, + 250.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1018.0, + 1450.0, + 1018.0, + 1450.0, + 1051.0, + 250.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1048.0, + 1136.0, + 1048.0, + 1136.0, + 1081.0, + 250.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 1688.0, + 1450.0, + 1688.0, + 1450.0, + 1723.0, + 258.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1716.0, + 1450.0, + 1716.0, + 1450.0, + 1751.0, + 250.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1748.0, + 1450.0, + 1748.0, + 1450.0, + 1783.0, + 249.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1776.0, + 1451.0, + 1776.0, + 1451.0, + 1818.0, + 248.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 265.0, + 1453.0, + 265.0, + 1453.0, + 303.0, + 251.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 293.0, + 1450.0, + 293.0, + 1450.0, + 337.0, + 274.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 276.0, + 326.0, + 919.0, + 326.0, + 919.0, + 364.0, + 276.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 422.0, + 1366.0, + 422.0, + 1366.0, + 461.0, + 250.0, + 461.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 253, + 848, + 1449, + 848, + 1449, + 1275, + 253, + 1275 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 253, + 1290, + 1451, + 1290, + 1451, + 1567, + 253, + 1567 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 253, + 375, + 1448, + 375, + 1448, + 557, + 253, + 557 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 254, + 678, + 1448, + 678, + 1448, + 832, + 254, + 832 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 253, + 266, + 1446, + 266, + 1446, + 359, + 253, + 359 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 255, + 1687, + 1449, + 1687, + 1449, + 1811, + 255, + 1811 + ], + "score": 0.971 + }, + { + "category_id": 2, + "poly": [ + 253, + 113, + 812, + 113, + 812, + 141, + 253, + 141 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 253, + 606, + 1045, + 606, + 1045, + 643, + 253, + 643 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 256, + 1615, + 697, + 1615, + 697, + 1650, + 256, + 1650 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 791, + 1866, + 818, + 1866, + 818, + 1889, + 791, + 1889 + ], + "score": 0.819 + }, + { + "category_id": 13, + "poly": [ + 1317, + 1290, + 1443, + 1290, + 1443, + 1326, + 1317, + 1326 + ], + "score": 0.92, + "latex": "\\| z _ { s } - z _ { p } \\| ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1092, + 1032, + 1173, + 1032, + 1173, + 1061, + 1092, + 1061 + ], + "score": 0.9, + "latex": "M + 1" + }, + { + "category_id": 13, + "poly": [ + 333, + 1384, + 378, + 1384, + 378, + 1419, + 333, + 1419 + ], + "score": 0.9, + "latex": "E _ { \\theta _ { p } }" + }, + { + "category_id": 13, + "poly": [ + 1018, + 1475, + 1060, + 1475, + 1060, + 1508, + 1018, + 1508 + ], + "score": 0.9, + "latex": "D _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 824, + 1063, + 902, + 1063, + 902, + 1092, + 824, + 1092 + ], + "score": 0.89, + "latex": "M + 3" + }, + { + "category_id": 13, + "poly": [ + 817, + 1444, + 859, + 1444, + 859, + 1479, + 817, + 1479 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 252, + 297, + 801, + 297, + 801, + 331, + 252, + 331 + ], + "score": 0.88, + "latex": "\\pmb { x } ^ { \\prime } = G _ { \\mathbf { X } ^ { \\prime } , \\mathbf { X } } ( \\pmb { x } ) = D e c o d e r _ { X ^ { \\prime } } ( E n c o d e r _ { X } ( \\pmb { x } ) )" + }, + { + "category_id": 13, + "poly": [ + 988, + 268, + 1018, + 268, + 1018, + 295, + 988, + 295 + ], + "score": 0.87, + "latex": "\\mathbf { x } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1171, + 267, + 1206, + 267, + 1206, + 296, + 1171, + 296 + ], + "score": 0.85, + "latex": "\\mathbf { X } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 619, + 274, + 640, + 274, + 640, + 295, + 619, + 295 + ], + "score": 0.75, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1719, + 1330, + 1719, + 1330, + 1747, + 1292, + 1747 + ], + "score": 0.73, + "latex": "\\mathbf { C } =" + }, + { + "category_id": 13, + "poly": [ + 251, + 1031, + 303, + 1031, + 303, + 1064, + 251, + 1064 + ], + "score": 0.71, + "latex": "\\bar { ( \\cal M ) }" + }, + { + "category_id": 13, + "poly": [ + 976, + 298, + 1143, + 298, + 1143, + 331, + 976, + 331 + ], + "score": 0.65, + "latex": "E n c o d e r _ { X } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 675, + 1719, + 715, + 1719, + 715, + 1747, + 675, + 1747 + ], + "score": 0.61, + "latex": "\\mathbf { K } { = }" + }, + { + "category_id": 13, + "poly": [ + 929, + 268, + 957, + 268, + 957, + 296, + 929, + 296 + ], + "score": 0.52, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1720, + 1053, + 1720, + 1053, + 1747, + 1012, + 1747 + ], + "score": 0.5, + "latex": "\\begin{array} { r } { \\boldsymbol { \\mathrm { D } } \\boldsymbol { = } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 529, + 1719, + 570, + 1719, + 570, + 1748, + 529, + 1748 + ], + "score": 0.48, + "latex": "{ \\bf N } =" + }, + { + "category_id": 13, + "poly": [ + 851, + 1720, + 887, + 1720, + 887, + 1747, + 851, + 1747 + ], + "score": 0.38, + "latex": "{ \\mathrm { S } } { = } { }" + }, + { + "category_id": 13, + "poly": [ + 1080, + 298, + 1143, + 298, + 1143, + 331, + 1080, + 331 + ], + "score": 0.3, + "latex": "\\ b _ { X } ( \\ b { x } )" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 253.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 603.0, + 1051.0, + 603.0, + 1051.0, + 649.0, + 249.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1612.0, + 702.0, + 1612.0, + 702.0, + 1656.0, + 249.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1862.0, + 822.0, + 1862.0, + 822.0, + 1900.0, + 787.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 845.0, + 1453.0, + 845.0, + 1453.0, + 886.0, + 248.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 878.0, + 1450.0, + 878.0, + 1450.0, + 912.0, + 249.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 909.0, + 1453.0, + 909.0, + 1453.0, + 943.0, + 250.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 939.0, + 1450.0, + 939.0, + 1450.0, + 978.0, + 248.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 971.0, + 1452.0, + 971.0, + 1452.0, + 1007.0, + 246.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1000.0, + 1450.0, + 1000.0, + 1450.0, + 1038.0, + 246.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1030.0, + 1091.0, + 1030.0, + 1091.0, + 1067.0, + 304.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1030.0, + 1450.0, + 1030.0, + 1450.0, + 1067.0, + 1174.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1063.0, + 823.0, + 1063.0, + 823.0, + 1097.0, + 249.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1063.0, + 1450.0, + 1063.0, + 1450.0, + 1097.0, + 903.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1093.0, + 1449.0, + 1093.0, + 1449.0, + 1127.0, + 249.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1122.0, + 1452.0, + 1122.0, + 1452.0, + 1160.0, + 249.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1152.0, + 1452.0, + 1152.0, + 1452.0, + 1189.0, + 246.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1182.0, + 1452.0, + 1182.0, + 1452.0, + 1217.0, + 249.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1215.0, + 1449.0, + 1215.0, + 1449.0, + 1250.0, + 250.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1243.0, + 688.0, + 1243.0, + 688.0, + 1283.0, + 246.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1287.0, + 1316.0, + 1287.0, + 1316.0, + 1328.0, + 246.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1444.0, + 1287.0, + 1451.0, + 1287.0, + 1451.0, + 1328.0, + 1444.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1320.0, + 1454.0, + 1320.0, + 1454.0, + 1359.0, + 246.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1350.0, + 1452.0, + 1350.0, + 1452.0, + 1388.0, + 248.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1384.0, + 332.0, + 1384.0, + 332.0, + 1417.0, + 250.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1384.0, + 1451.0, + 1384.0, + 1451.0, + 1417.0, + 379.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1413.0, + 1455.0, + 1413.0, + 1455.0, + 1448.0, + 248.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1443.0, + 816.0, + 1443.0, + 816.0, + 1480.0, + 249.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1443.0, + 1452.0, + 1443.0, + 1452.0, + 1480.0, + 860.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1476.0, + 1017.0, + 1476.0, + 1017.0, + 1508.0, + 250.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1476.0, + 1451.0, + 1476.0, + 1451.0, + 1508.0, + 1061.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1503.0, + 1452.0, + 1503.0, + 1452.0, + 1540.0, + 249.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1536.0, + 1454.0, + 1536.0, + 1454.0, + 1569.0, + 250.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 375.0, + 1449.0, + 375.0, + 1449.0, + 407.0, + 252.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 406.0, + 1448.0, + 406.0, + 1448.0, + 437.0, + 252.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 435.0, + 1451.0, + 435.0, + 1451.0, + 470.0, + 247.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 463.0, + 1452.0, + 463.0, + 1452.0, + 501.0, + 248.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 492.0, + 1454.0, + 492.0, + 1454.0, + 533.0, + 246.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 528.0, + 361.0, + 528.0, + 361.0, + 560.0, + 249.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 681.0, + 1449.0, + 681.0, + 1449.0, + 714.0, + 251.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 711.0, + 1451.0, + 711.0, + 1451.0, + 744.0, + 250.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 738.0, + 1451.0, + 738.0, + 1451.0, + 778.0, + 247.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 770.0, + 1451.0, + 770.0, + 1451.0, + 808.0, + 247.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 803.0, + 545.0, + 803.0, + 545.0, + 836.0, + 250.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 261.0, + 618.0, + 261.0, + 618.0, + 305.0, + 247.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 261.0, + 928.0, + 261.0, + 928.0, + 305.0, + 641.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 261.0, + 987.0, + 261.0, + 987.0, + 305.0, + 958.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 261.0, + 1170.0, + 261.0, + 1170.0, + 305.0, + 1019.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 261.0, + 1450.0, + 261.0, + 1450.0, + 305.0, + 1207.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 296.0, + 975.0, + 296.0, + 975.0, + 334.0, + 802.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 296.0, + 1450.0, + 296.0, + 1450.0, + 334.0, + 1144.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 326.0, + 895.0, + 326.0, + 895.0, + 365.0, + 249.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1686.0, + 1450.0, + 1686.0, + 1450.0, + 1724.0, + 248.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1718.0, + 528.0, + 1718.0, + 528.0, + 1753.0, + 249.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1718.0, + 674.0, + 1718.0, + 674.0, + 1753.0, + 571.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1718.0, + 850.0, + 1718.0, + 850.0, + 1753.0, + 716.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1718.0, + 1011.0, + 1718.0, + 1011.0, + 1753.0, + 888.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1718.0, + 1291.0, + 1718.0, + 1291.0, + 1753.0, + 1054.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1718.0, + 1452.0, + 1718.0, + 1452.0, + 1753.0, + 1331.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1749.0, + 1449.0, + 1749.0, + 1449.0, + 1784.0, + 251.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1780.0, + 372.0, + 1780.0, + 372.0, + 1812.0, + 249.0, + 1812.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 477, + 262, + 1221, + 262, + 1221, + 1210, + 477, + 1210 + ], + "score": 0.972, + "html": "
LayerEncoders (shared,private)
1CONV-(N16,K7,S1), ReLU
2CONV-(N32,K3,S2), ReLU
3CONV-(N64,K3,S2), ReLU
4RESBLK-(N64,K3,S1)
5RESBLK-(N64,K3,S1)
6RESBLK-(N64,K3,S1)
7RESBLK-(N64,K3,S1)
LayerDecoder
1RESBLK-(N64,K3,S1)
2RESBLK-(N64,K3,S1)
3RESBLK-(N64,K3,S1)
4RESBLK-(N64,K3,S1)
5DCONV-(N32,K3,S2), ReLU
6DCONV-(N16,K3,S2), ReLU
7DCONV-(N1,K1,S1), Tanh
LayerDiscriminator
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1),ReLU
3CONV-(N16,K3,S1),ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(ND), Softmax
LayerClassifier
1CONV-(N4,K3,S1),ReLU
2CONV-(N8,K3,S1), ReLU
3CONV-(N16,K3,S1), ReLU
4CONV-(N32,K3,S1),ReLU
5CONV-(N1,K3,S1),ReLU
6DENSE-(NC), Softmax
" + }, + { + "category_id": 1, + "poly": [ + 253, + 1717, + 1451, + 1717, + 1451, + 1811, + 253, + 1811 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 253, + 1502, + 1446, + 1502, + 1446, + 1596, + 253, + 1596 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 254, + 1350, + 1165, + 1350, + 1165, + 1384, + 254, + 1384 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 255, + 1645, + 534, + 1645, + 534, + 1682, + 255, + 1682 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.896 + }, + { + "category_id": 6, + "poly": [ + 593, + 1232, + 1104, + 1232, + 1104, + 1262, + 593, + 1262 + ], + "score": 0.87 + }, + { + "category_id": 0, + "poly": [ + 255, + 1278, + 789, + 1278, + 789, + 1315, + 255, + 1315 + ], + "score": 0.856 + }, + { + "category_id": 0, + "poly": [ + 251, + 1430, + 1259, + 1430, + 1259, + 1469, + 251, + 1469 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 790, + 1866, + 819, + 1866, + 819, + 1889, + 790, + 1889 + ], + "score": 0.829 + }, + { + "category_id": 13, + "poly": [ + 673, + 1719, + 856, + 1719, + 856, + 1748, + 673, + 1748 + ], + "score": 0.88, + "latex": "2 2 7 \\times 2 2 7 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1535, + 1253, + 1535, + 1253, + 1563, + 1200, + 1563 + ], + "score": 0.88, + "latex": "C 0 9" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1535, + 1089, + 1535, + 1089, + 1564, + 1036, + 1564 + ], + "score": 0.78, + "latex": "C 0 5" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1535, + 1152, + 1535, + 1152, + 1564, + 1099, + 1564 + ], + "score": 0.73, + "latex": "C 0 8" + }, + { + "category_id": 13, + "poly": [ + 470, + 1565, + 537, + 1565, + 537, + 1595, + 470, + 1595 + ], + "score": 0.63, + "latex": "5 \\& 6" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1641.0, + 540.0, + 1641.0, + 540.0, + 1688.0, + 249.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1231.0, + 1108.0, + 1231.0, + 1108.0, + 1265.0, + 590.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1273.0, + 792.0, + 1273.0, + 792.0, + 1320.0, + 247.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1430.0, + 1264.0, + 1430.0, + 1264.0, + 1472.0, + 246.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 824.0, + 1862.0, + 824.0, + 1901.0, + 786.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1712.0, + 672.0, + 1712.0, + 672.0, + 1755.0, + 248.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1712.0, + 1452.0, + 1712.0, + 1452.0, + 1755.0, + 857.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1746.0, + 1452.0, + 1746.0, + 1452.0, + 1787.0, + 248.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 1775.0, + 1452.0, + 1775.0, + 1452.0, + 1820.0, + 245.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1502.0, + 1450.0, + 1502.0, + 1450.0, + 1538.0, + 249.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1532.0, + 1035.0, + 1532.0, + 1035.0, + 1569.0, + 247.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1532.0, + 1098.0, + 1532.0, + 1098.0, + 1569.0, + 1090.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1532.0, + 1199.0, + 1532.0, + 1199.0, + 1569.0, + 1153.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1532.0, + 1450.0, + 1532.0, + 1450.0, + 1569.0, + 1254.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1567.0, + 469.0, + 1567.0, + 469.0, + 1596.0, + 250.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1567.0, + 546.0, + 1567.0, + 546.0, + 1596.0, + 538.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1348.0, + 1170.0, + 1348.0, + 1170.0, + 1388.0, + 249.0, + 1388.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 251, + 1018, + 1444, + 1018, + 1444, + 1481, + 251, + 1481 + ], + "score": 0.98, + "html": "
methodMM→SMM→MMM→UU→SU→MU→MM
Source Only40.00±0.6184.46 ± 0.2980.43± 0.5023.41± 0.5250.64 ± 0.3741.45±0.38
1-NN21.4582.1336.9015.3438.4518.54
CORAL (Sun & Saenko (2016))40.20 ±0.6084.90 ± 0.7087.54 ± 0.4438.90 ± 0.9685.01 ± 0.6160.45 ± 0.70
DANN (Ganin et al. (2016))51.80 ± 0.9161.05 ± 0.7185.34 ± 0.6435.50± 0.8477.40 ± 0.6461.60 ± 0.64
ADDA (Tzeng et al. (2017))40.64 ± 0.8692.82 ± 0.4880.70±0.4841.23 ± 0.7890.10± 0.5856.21 ± 0.79
c-ADDA35.43 ± 0.9488.47 ± 0.6174.19 ± 0.5839.36 ± 0.9984.67 ± 0.9452.54 ± 0.88
DTN (Zhang et al. (2015)) c-DTN48.80± 0.6688.80±0.3890.68 ± 0.3542.43 ± 0.6189.04 ± 0.3655.78 ± 0.40
UNIT (Liu et al. (2017))44.21 ± 0.6183.60 ± 0.5484.98 ± 0.4139.75 ± 0.6485.04 ± 0.4548.86 ± 0.54
DSN (Bousmalis et al. (2016))1 51.50 ± 0.64/90.60
c-DSN47.10 ± 0.5090.20± 0.3189.95± 0.2948.20 ± 0.5991.40 ± 0.3060.45 ± 0.35
1p-DSN45.00 ± 0.6084.60 ± 0.4084.80 ±0.3940.50 ± 0.6186.05 ± 0.4656.25 ± 0.50
mp-DSN47.15 ± 0.6481.96 ± 0.6083.03 ±0.4939.30 ± 0.5184.55 ± 0.5655.03 ± 0.60
s-MTDA-ITA85.51 ± 0.5483.24± 0.2438.30 ± 0.7487.40 ± 0.3555.47 ± 0.44
c-MTDA-ITA50.55± 0.18 47.32±0.1994.82±0.2189.05 ±0.2840.13±0.3087.10 ± 0.2561.01 ± 0.24
MTDA-ITA53.50 ± 0.2290.20 ±0.30 98.20 ± 0.1090.01 ± 0.24 94.10 ± 0.1141.10 ± 0.35 46.00 ± 0.4885.35 ± 0.28 91.50 ± 0.2360.31± 0.34 67.30 ± 0.15
" + }, + { + "category_id": 1, + "poly": [ + 253, + 1688, + 1448, + 1688, + 1448, + 1811, + 253, + 1811 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 398, + 225, + 1292, + 225, + 1292, + 887, + 398, + 887 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 246, + 923, + 1445, + 923, + 1445, + 982, + 246, + 982 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 790, + 1865, + 818, + 1865, + 818, + 1889, + 790, + 1889 + ], + "score": 0.83 + }, + { + "category_id": 6, + "poly": [ + 252, + 1521, + 1449, + 1521, + 1449, + 1663, + 252, + 1663 + ], + "score": 0.465 + }, + { + "category_id": 13, + "poly": [ + 991, + 1607, + 1098, + 1607, + 1098, + 1636, + 991, + 1636 + ], + "score": 0.83, + "latex": "( \\approx 1 , 0 0 0 )" + }, + { + "category_id": 13, + "poly": [ + 520, + 1607, + 609, + 1607, + 609, + 1634, + 520, + 1634 + ], + "score": 0.82, + "latex": "> 5 0 0 \\mathrm { K }" + }, + { + "category_id": 13, + "poly": [ + 748, + 1022, + 851, + 1022, + 851, + 1046, + 748, + 1046 + ], + "score": 0.75, + "latex": "\\mathbf { \\overline { { M M } } } \\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 882, + 1293, + 1004, + 1293, + 1004, + 1314, + 882, + 1314 + ], + "score": 0.67, + "latex": "8 9 . 9 5 \\pm 0 . 2 9" + }, + { + "category_id": 13, + "poly": [ + 893, + 1023, + 993, + 1023, + 993, + 1046, + 893, + 1046 + ], + "score": 0.66, + "latex": "\\mathbf { M M } \\to \\mathbf { U }" + }, + { + "category_id": 13, + "poly": [ + 882, + 1320, + 1003, + 1320, + 1003, + 1340, + 882, + 1340 + ], + "score": 0.64, + "latex": "8 4 . 8 0 \\pm 0 . 3 9" + }, + { + "category_id": 13, + "poly": [ + 739, + 1240, + 862, + 1240, + 862, + 1260, + 739, + 1260 + ], + "score": 0.63, + "latex": "8 3 . 6 0 \\pm 0 . 5 4" + }, + { + "category_id": 13, + "poly": [ + 882, + 1240, + 1004, + 1240, + 1004, + 1260, + 882, + 1260 + ], + "score": 0.63, + "latex": "8 4 . 9 8 \\pm 0 . 4 1" + }, + { + "category_id": 13, + "poly": [ + 1307, + 1579, + 1387, + 1579, + 1387, + 1606, + 1307, + 1606 + ], + "score": 0.62, + "latex": "\\bf \\Pi ^ { * } U N I T" + }, + { + "category_id": 13, + "poly": [ + 739, + 1369, + 862, + 1369, + 862, + 1392, + 739, + 1392 + ], + "score": 0.62, + "latex": "8 5 . 5 1 \\pm 0 . 5 4" + }, + { + "category_id": 13, + "poly": [ + 740, + 1294, + 861, + 1294, + 861, + 1313, + 740, + 1313 + ], + "score": 0.62, + "latex": "9 0 . 2 0 \\pm 0 . 3 1" + }, + { + "category_id": 13, + "poly": [ + 1190, + 1023, + 1266, + 1023, + 1266, + 1046, + 1190, + 1046 + ], + "score": 0.61, + "latex": "\\overline { { \\mathrm { U } \\to \\mathbf { M } } }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1023, + 1120, + 1023, + 1120, + 1046, + 1051, + 1046 + ], + "score": 0.61, + "latex": "\\overline { { \\mathrm { ~ U ~ } \\to \\mathrm { ~ S ~ } } }" + }, + { + "category_id": 13, + "poly": [ + 739, + 1396, + 861, + 1396, + 861, + 1420, + 739, + 1420 + ], + "score": 0.59, + "latex": "\\overline { { 9 4 . 8 2 \\pm 0 . 2 1 } }" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1370, + 1148, + 1370, + 1148, + 1392, + 1024, + 1392 + ], + "score": 0.56, + "latex": "3 8 . 3 0 \\pm 0 . 7 4" + }, + { + "category_id": 13, + "poly": [ + 881, + 1422, + 1004, + 1422, + 1004, + 1446, + 881, + 1446 + ], + "score": 0.56, + "latex": "9 0 . 0 1 \\pm 0 . 2 4" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1295, + 1147, + 1295, + 1147, + 1313, + 1024, + 1313 + ], + "score": 0.56, + "latex": "4 8 . 2 0 \\pm 0 . 5 9" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1398, + 1148, + 1398, + 1148, + 1420, + 1024, + 1420 + ], + "score": 0.56, + "latex": "4 0 . 1 3 \\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1320, + 1146, + 1320, + 1146, + 1340, + 1024, + 1340 + ], + "score": 0.54, + "latex": "4 0 . 5 0 \\pm 0 . 6 1" + }, + { + "category_id": 13, + "poly": [ + 882, + 1369, + 1004, + 1369, + 1004, + 1392, + 882, + 1392 + ], + "score": 0.54, + "latex": "8 3 . 2 4 \\pm 0 . 2 4" + }, + { + "category_id": 13, + "poly": [ + 881, + 1397, + 1004, + 1397, + 1004, + 1420, + 881, + 1420 + ], + "score": 0.53, + "latex": "8 9 . 0 5 \\pm 0 . 2 8" + }, + { + "category_id": 13, + "poly": [ + 739, + 1183, + 861, + 1183, + 861, + 1206, + 739, + 1206 + ], + "score": 0.51, + "latex": "8 8 . 4 7 \\pm 0 . 6 1" + }, + { + "category_id": 13, + "poly": [ + 739, + 1319, + 862, + 1319, + 862, + 1340, + 739, + 1340 + ], + "score": 0.51, + "latex": "8 4 . 6 0 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 609, + 1022, + 705, + 1022, + 705, + 1046, + 609, + 1046 + ], + "score": 0.5, + "latex": "\\overline { { { \\bf { M M } } \\to { \\bf { S } } } }" + }, + { + "category_id": 13, + "poly": [ + 935, + 1273, + 951, + 1273, + 951, + 1285, + 935, + 1285 + ], + "score": 0.49, + "latex": "-" + }, + { + "category_id": 13, + "poly": [ + 741, + 1049, + 862, + 1049, + 862, + 1074, + 741, + 1074 + ], + "score": 0.47, + "latex": "8 4 . 4 6 \\pm 0 . 2 9" + }, + { + "category_id": 13, + "poly": [ + 740, + 1422, + 863, + 1422, + 863, + 1446, + 740, + 1446 + ], + "score": 0.47, + "latex": "9 0 . 2 0 \\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 738, + 1129, + 861, + 1129, + 861, + 1153, + 738, + 1153 + ], + "score": 0.47, + "latex": "6 1 . 0 5 \\pm 0 . 7 1" + }, + { + "category_id": 13, + "poly": [ + 882, + 1049, + 1004, + 1049, + 1004, + 1074, + 882, + 1074 + ], + "score": 0.44, + "latex": "\\overline { { 8 0 . 4 3 \\pm 0 . 5 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1025, + 1423, + 1147, + 1423, + 1147, + 1446, + 1025, + 1446 + ], + "score": 0.43, + "latex": "4 1 . 1 0 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1241, + 1147, + 1241, + 1147, + 1259, + 1024, + 1259 + ], + "score": 0.43, + "latex": "3 9 . 7 5 \\pm 0 . 6 4" + }, + { + "category_id": 13, + "poly": [ + 739, + 1211, + 862, + 1211, + 862, + 1233, + 739, + 1233 + ], + "score": 0.43, + "latex": "8 8 . 8 0 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1370, + 1290, + 1370, + 1290, + 1393, + 1166, + 1393 + ], + "score": 0.43, + "latex": "8 7 . 4 0 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1295, + 1290, + 1295, + 1290, + 1313, + 1166, + 1313 + ], + "score": 0.42, + "latex": "\\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 882, + 1183, + 1004, + 1183, + 1004, + 1206, + 882, + 1206 + ], + "score": 0.42, + "latex": "7 4 . 1 9 \\pm 0 . 5 8" + }, + { + "category_id": 13, + "poly": [ + 881, + 1103, + 1003, + 1103, + 1003, + 1127, + 881, + 1127 + ], + "score": 0.42, + "latex": "8 7 . 5 4 \\pm 0 . 4 4" + }, + { + "category_id": 13, + "poly": [ + 739, + 1344, + 862, + 1344, + 862, + 1366, + 739, + 1366 + ], + "score": 0.4, + "latex": "8 1 . 9 6 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 739, + 1157, + 861, + 1157, + 861, + 1180, + 739, + 1180 + ], + "score": 0.4, + "latex": "9 2 . 8 2 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 881, + 1211, + 1003, + 1211, + 1003, + 1233, + 881, + 1233 + ], + "score": 0.39, + "latex": "9 0 . 6 8 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 833, + 1608, + 848, + 1608, + 848, + 1631, + 833, + 1631 + ], + "score": 0.39, + "latex": "^ *" + }, + { + "category_id": 13, + "poly": [ + 882, + 1157, + 1004, + 1157, + 1004, + 1180, + 882, + 1180 + ], + "score": 0.38, + "latex": "8 0 . 7 0 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 882, + 1345, + 1004, + 1345, + 1004, + 1366, + 882, + 1366 + ], + "score": 0.37, + "latex": "8 3 . 0 3 \\pm 0 . 4 9" + }, + { + "category_id": 13, + "poly": [ + 577, + 1446, + 1433, + 1446, + 1433, + 1478, + 577, + 1478 + ], + "score": 0.37, + "latex": "\\left| \\left| 5 3 . 5 0 \\pm 0 . 2 2 \\right| \\left| 9 8 . 2 0 \\pm 0 . 1 0 \\right| \\left| 9 4 . 1 0 \\pm 0 . 1 1 \\right| \\left| 4 6 . 0 0 \\pm 0 . 4 8 \\right| \\left| 9 1 . 5 0 \\pm 0 . 2 3 \\right| \\left| 6 7 . 3 0 \\pm 0 . 1 5 1 \\pm 0 . 0 0 0 \\right| \\right| ." + }, + { + "category_id": 13, + "poly": [ + 1024, + 1183, + 1148, + 1183, + 1148, + 1206, + 1024, + 1206 + ], + "score": 0.37, + "latex": "3 9 . 3 6 \\pm 0 . 9 9" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1345, + 1148, + 1345, + 1148, + 1366, + 1024, + 1366 + ], + "score": 0.37, + "latex": "3 9 . 3 0 \\pm 0 . 5 1" + }, + { + "category_id": 13, + "poly": [ + 793, + 1273, + 809, + 1273, + 809, + 1285, + 793, + 1285 + ], + "score": 0.36, + "latex": "-" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1397, + 1290, + 1397, + 1290, + 1420, + 1166, + 1420 + ], + "score": 0.36, + "latex": "8 7 . 1 0 \\pm 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1211, + 1147, + 1211, + 1147, + 1233, + 1024, + 1233 + ], + "score": 0.35, + "latex": "4 2 . 4 3 \\pm 0 . 6 1" + }, + { + "category_id": 13, + "poly": [ + 739, + 1103, + 862, + 1103, + 862, + 1127, + 739, + 1127 + ], + "score": 0.33, + "latex": "8 4 . 9 0 \\pm 0 . 7 0" + }, + { + "category_id": 13, + "poly": [ + 596, + 1423, + 720, + 1423, + 720, + 1446, + 596, + 1446 + ], + "score": 0.32, + "latex": "4 7 . 3 2 \\pm 0 . 1 9" + }, + { + "category_id": 13, + "poly": [ + 1167, + 1240, + 1289, + 1240, + 1289, + 1260, + 1167, + 1260 + ], + "score": 0.32, + "latex": "8 5 . 0 4 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 882, + 1129, + 1004, + 1129, + 1004, + 1154, + 882, + 1154 + ], + "score": 0.31, + "latex": "8 5 . 3 4 \\pm 0 . 6 4" + }, + { + "category_id": 13, + "poly": [ + 1321, + 1024, + 1419, + 1024, + 1419, + 1046, + 1321, + 1046 + ], + "score": 0.3, + "latex": "\\mathbf { \\overline { { U } } } \\mathbf { } \\mathbf { M } \\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1049, + 1289, + 1049, + 1289, + 1074, + 1168, + 1074 + ], + "score": 0.27, + "latex": "\\overline { { 5 0 . 6 4 \\pm 0 . 3 7 } }" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 296.0, + 943.0, + 296.0, + 943.0, + 503.0, + 913.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 306.0, + 645.0, + 306.0, + 645.0, + 326.0, + 565.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 308.0, + 424.0, + 308.0, + 424.0, + 324.0, + 404.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 314.0, + 426.0, + 314.0, + 426.0, + 385.0, + 398.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 322.0, + 650.0, + 322.0, + 650.0, + 342.0, + 564.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 335.0, + 651.0, + 335.0, + 651.0, + 355.0, + 563.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 350.0, + 652.0, + 350.0, + 652.0, + 371.0, + 564.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 364.0, + 652.0, + 364.0, + 652.0, + 384.0, + 564.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 378.0, + 424.0, + 378.0, + 424.0, + 396.0, + 404.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 391.0, + 424.0, + 391.0, + 424.0, + 411.0, + 404.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 407.0, + 423.0, + 407.0, + 423.0, + 425.0, + 404.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 421.0, + 424.0, + 421.0, + 424.0, + 441.0, + 404.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 436.0, + 423.0, + 436.0, + 423.0, + 452.0, + 404.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 450.0, + 423.0, + 450.0, + 423.0, + 467.0, + 404.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 464.0, + 423.0, + 464.0, + 423.0, + 481.0, + 404.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 477.0, + 425.0, + 477.0, + 425.0, + 499.0, + 403.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 493.0, + 424.0, + 493.0, + 424.0, + 508.0, + 405.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 497.0, + 934.0, + 497.0, + 934.0, + 505.0, + 925.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 502.0, + 498.0, + 502.0, + 498.0, + 519.0, + 456.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 502.0, + 625.0, + 502.0, + 625.0, + 519.0, + 573.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 503.0, + 741.0, + 503.0, + 741.0, + 519.0, + 701.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 502.0, + 1007.0, + 502.0, + 1007.0, + 520.0, + 970.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 503.0, + 1129.0, + 503.0, + 1129.0, + 518.0, + 1092.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 503.0, + 1251.0, + 503.0, + 1251.0, + 518.0, + 1215.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 533.0, + 730.0, + 533.0, + 730.0, + 565.0, + 458.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 534.0, + 1263.0, + 534.0, + 1263.0, + 565.0, + 948.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 616.0, + 436.0, + 616.0, + 436.0, + 799.0, + 403.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 617.0, + 944.0, + 617.0, + 944.0, + 832.0, + 914.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 822.0, + 491.0, + 822.0, + 491.0, + 837.0, + 464.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 822.0, + 616.0, + 822.0, + 616.0, + 837.0, + 582.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 822.0, + 735.0, + 822.0, + 735.0, + 837.0, + 707.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 822.0, + 1001.0, + 822.0, + 1001.0, + 838.0, + 975.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 822.0, + 1125.0, + 822.0, + 1125.0, + 837.0, + 1097.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 822.0, + 1249.0, + 822.0, + 1249.0, + 836.0, + 1216.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 854.0, + 738.0, + 854.0, + 738.0, + 885.0, + 448.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 854.0, + 1241.0, + 854.0, + 1241.0, + 885.0, + 969.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 790.0, + 427.0, + 790.0, + 427.0, + 803.5, + 410.0, + 803.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 924.0, + 1450.0, + 924.0, + 1450.0, + 956.0, + 249.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 953.0, + 1269.0, + 953.0, + 1269.0, + 984.0, + 251.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 823.0, + 1862.0, + 823.0, + 1901.0, + 786.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1521.0, + 1450.0, + 1521.0, + 1450.0, + 1554.0, + 248.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1550.0, + 1449.0, + 1550.0, + 1449.0, + 1582.0, + 249.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1577.0, + 1306.0, + 1577.0, + 1306.0, + 1612.0, + 248.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1388.0, + 1577.0, + 1450.0, + 1577.0, + 1450.0, + 1612.0, + 1388.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1603.0, + 519.0, + 1603.0, + 519.0, + 1641.0, + 247.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1603.0, + 832.0, + 1603.0, + 832.0, + 1641.0, + 610.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1603.0, + 990.0, + 1603.0, + 990.0, + 1641.0, + 849.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1603.0, + 1453.0, + 1603.0, + 1453.0, + 1641.0, + 1099.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1632.0, + 721.0, + 1632.0, + 721.0, + 1668.0, + 248.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1688.0, + 1451.0, + 1688.0, + 1451.0, + 1723.0, + 250.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1717.0, + 1449.0, + 1717.0, + 1449.0, + 1752.0, + 249.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1747.0, + 1451.0, + 1747.0, + 1451.0, + 1786.0, + 249.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1777.0, + 1449.0, + 1777.0, + 1449.0, + 1816.0, + 249.0, + 1816.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 251, + 258, + 1442, + 258, + 1442, + 889, + 251, + 889 + ], + "score": 0.982, + "html": "
methodC05 →C08C05→C09C05→C13C05→C14
Source Only31.56 ± 0.4040.67 ± 0.3639.89 ± 0.2254.70± 0.25
1-NN CORAL (Sun & Saenko (2016))27.28 36.55 ± 0.6631.22 38.60 ± 0.6733.6647.04 55.29 ± 0.47
40.60 ± 0.58
DANN (citeganin2016domain)40.30 ± 0.6041.20 ± 0.6540.12 ± 0.6058.90 ± 0.38
ADDA (Tzeng et al. (2017)) C-ADDA33.21 ± 0.8130.86 ± 0.9052.44± 0.8070.18 ± 0.60
46.88 ± 0.6536.38 ± 0.8839.14 ± 0.8565.41 ± 0.69
DTN (Zhang et al. (2015))38.50 ± 0.5130.56 ± 0.4655.78 ± 0.3668.90 ± 0.31
c-DTN41.70 ± 0.4231.10 ± 0.4850.19 ± 0.4560.34 ± 0.35
PixelDA (Bousmalis et al. (2017))44.93 ±0.4244.75 ± 0.4545.18 ± 0.4546.88 ± 0.49
UNIT (Liu et al. (2017))44.47 ± 0.2144.47 ± 0.2144.47 ± 0.2044.51 ± 0.28
DSN (Bousmalis et al. (2016))45.12 ± 0.4644.35± 0.4948.12 ± 0.5375.00 ± 0.39
c-DSN42.52 ± 0.4838.54 ± 0.6434.15 ± 0.6469.45 ± 0.55
1p-DSN41.64 ± 0.5837.84 ± 0.6334.65 ± 0.4468.75 ± 0.85
mp-DSN41.30 ± 0.2835.14 ± 0.3534.40 ± 0.3565.70 ± 0.27
s-MTDA-ITA44.40 ± 0.2344.60 ± 0.2547.65 ± 0.2780.20 ± 0.13
c-MTDA-ITA40.49 ± 0.2540.70 ± 0.2542.80 ± 0.2571.60 ± 0.10
MTDA-ITA49.01 ± 0.20
48.20± 0.2753.13 ±0.2284.29 9±0.10
" + }, + { + "category_id": 1, + "poly": [ + 254, + 1718, + 1449, + 1718, + 1449, + 1811, + 254, + 1811 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 252, + 112, + 812, + 112, + 812, + 141, + 252, + 141 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 252, + 1648, + 1037, + 1648, + 1037, + 1683, + 252, + 1683 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 253, + 1507, + 1450, + 1507, + 1450, + 1601, + 253, + 1601 + ], + "score": 0.857 + }, + { + "category_id": 6, + "poly": [ + 459, + 928, + 1238, + 928, + 1238, + 958, + 459, + 958 + ], + "score": 0.845 + }, + { + "category_id": 2, + "poly": [ + 790, + 1865, + 819, + 1865, + 819, + 1890, + 790, + 1890 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 256, + 1023, + 1333, + 1023, + 1333, + 1486, + 256, + 1486 + ], + "score": 0.787 + }, + { + "category_id": 0, + "poly": [ + 254, + 982, + 662, + 982, + 662, + 1016, + 254, + 1016 + ], + "score": 0.73 + }, + { + "category_id": 7, + "poly": [ + 253, + 1507, + 1450, + 1507, + 1450, + 1601, + 253, + 1601 + ], + "score": 0.266 + }, + { + "category_id": 13, + "poly": [ + 374, + 1442, + 563, + 1442, + 563, + 1477, + 374, + 1477 + ], + "score": 0.92, + "latex": "\\{ \\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi \\}" + }, + { + "category_id": 13, + "poly": [ + 405, + 1115, + 567, + 1115, + 567, + 1147, + 405, + 1147 + ], + "score": 0.92, + "latex": "\\bar { \\theta } _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi" + }, + { + "category_id": 13, + "poly": [ + 415, + 1207, + 469, + 1207, + 469, + 1242, + 415, + 1242 + ], + "score": 0.91, + "latex": "\\{ \\theta _ { s } \\}" + }, + { + "category_id": 13, + "poly": [ + 415, + 1292, + 469, + 1292, + 469, + 1327, + 415, + 1327 + ], + "score": 0.91, + "latex": "\\{ \\theta _ { c } \\}" + }, + { + "category_id": 13, + "poly": [ + 415, + 1375, + 464, + 1375, + 464, + 1410, + 415, + 1410 + ], + "score": 0.91, + "latex": "\\{ \\psi \\}" + }, + { + "category_id": 13, + "poly": [ + 415, + 1249, + 470, + 1249, + 470, + 1285, + 415, + 1285 + ], + "score": 0.91, + "latex": "\\{ \\theta _ { p } \\}" + }, + { + "category_id": 13, + "poly": [ + 1108, + 1374, + 1278, + 1374, + 1278, + 1417, + 1108, + 1417 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\psi = \\psi - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\psi } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 415, + 1334, + 461, + 1334, + 461, + 1370, + 415, + 1370 + ], + "score": 0.9, + "latex": "\\{ \\phi \\}" + }, + { + "category_id": 13, + "poly": [ + 371, + 1021, + 533, + 1021, + 533, + 1055, + 371, + 1055 + ], + "score": 0.9, + "latex": "\\{ \\mathbf { X } , \\mathbf { Y } , \\mathbf { D } \\} { : \\mathrm { M } }" + }, + { + "category_id": 13, + "poly": [ + 636, + 1377, + 675, + 1377, + 675, + 1410, + 636, + 1410 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 634, + 1336, + 671, + 1336, + 671, + 1370, + 634, + 1370 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 1131, + 1205, + 1313, + 1205, + 1313, + 1246, + 1131, + 1246 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\theta _ { s } = \\theta _ { s } - \\eta \\frac { \\partial \\mathcal { L } _ { s } } { \\partial \\theta _ { s } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 643, + 1251, + 678, + 1251, + 678, + 1285, + 643, + 1285 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 360, + 1085, + 522, + 1085, + 522, + 1116, + 360, + 1116 + ], + "score": 0.89, + "latex": "\\theta _ { s } , \\theta _ { p } , \\theta _ { c } , \\phi , \\psi" + }, + { + "category_id": 13, + "poly": [ + 642, + 1208, + 676, + 1208, + 676, + 1239, + 642, + 1239 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { s }" + }, + { + "category_id": 13, + "poly": [ + 642, + 1294, + 676, + 1294, + 676, + 1325, + 642, + 1325 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 348, + 1055, + 461, + 1055, + 461, + 1084, + 348, + 1084 + ], + "score": 0.88, + "latex": "\\lambda _ { r } , \\lambda _ { c } , \\lambda _ { d }" + }, + { + "category_id": 13, + "poly": [ + 1104, + 1331, + 1271, + 1331, + 1271, + 1374, + 1104, + 1374 + ], + "score": 0.87, + "latex": "\\phi = \\phi - \\eta \\frac { \\partial \\mathcal { L } _ { \\phi } } { \\partial \\phi }" + }, + { + "category_id": 13, + "poly": [ + 904, + 269, + 1055, + 269, + 1055, + 291, + 904, + 291 + ], + "score": 0.87, + "latex": "\\mathrm { C } 0 5 \\mathrm { C } 0 9" + }, + { + "category_id": 13, + "poly": [ + 1089, + 269, + 1239, + 269, + 1239, + 292, + 1089, + 292 + ], + "score": 0.85, + "latex": "{ \\bf C } 0 5 { \\bf C } 1 3" + }, + { + "category_id": 13, + "poly": [ + 1110, + 1246, + 1296, + 1246, + 1296, + 1290, + 1110, + 1290 + ], + "score": 0.85, + "latex": "\\begin{array} { r } { \\theta _ { p } = \\theta _ { p } - \\eta \\frac { \\partial \\mathcal { L } _ { p } } { \\partial \\theta _ { p } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1291, + 1288, + 1291, + 1288, + 1331, + 1107, + 1331 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { \\theta _ { c } = \\theta _ { c } - \\eta \\frac { \\partial \\mathcal { L } _ { c } } { \\partial \\theta _ { c } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1269, + 440, + 1430, + 440, + 1430, + 466, + 1269, + 466 + ], + "score": 0.82, + "latex": "7 0 . 1 8 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 1273, + 269, + 1423, + 269, + 1423, + 292, + 1273, + 292 + ], + "score": 0.79, + "latex": "{ \\mathrm { C } } 0 5 \\to { \\mathrm { C } } 1 4" + }, + { + "category_id": 13, + "poly": [ + 900, + 441, + 1058, + 441, + 1058, + 466, + 900, + 466 + ], + "score": 0.79, + "latex": "3 0 . 8 6 \\pm 0 . 9 0" + }, + { + "category_id": 13, + "poly": [ + 1084, + 581, + 1243, + 581, + 1243, + 605, + 1084, + 605 + ], + "score": 0.73, + "latex": "4 5 . 1 8 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 900, + 406, + 1058, + 406, + 1058, + 432, + 900, + 432 + ], + "score": 0.73, + "latex": "4 1 . 2 0 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 900, + 510, + 1058, + 510, + 1058, + 535, + 900, + 535 + ], + "score": 0.72, + "latex": "3 0 . 5 6 \\pm 0 . 4 6" + }, + { + "category_id": 13, + "poly": [ + 1085, + 405, + 1243, + 405, + 1243, + 432, + 1085, + 432 + ], + "score": 0.72, + "latex": "4 0 . 1 2 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 718, + 264, + 870, + 264, + 870, + 293, + 718, + 293 + ], + "score": 0.71, + "latex": "\\overline { { \\mathbf { C 0 5 } \\to \\mathbf { C 0 8 } } }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 406, + 1428, + 406, + 1428, + 431, + 1269, + 431 + ], + "score": 0.71, + "latex": "5 8 . 9 0 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 1269, + 509, + 1428, + 509, + 1428, + 535, + 1269, + 535 + ], + "score": 0.71, + "latex": "6 8 . 9 0 \\pm 0 . 3 1" + }, + { + "category_id": 13, + "poly": [ + 900, + 649, + 1058, + 649, + 1058, + 674, + 900, + 674 + ], + "score": 0.71, + "latex": "4 4 . 3 5 \\pm 0 . 4 9" + }, + { + "category_id": 13, + "poly": [ + 900, + 580, + 1058, + 580, + 1058, + 604, + 900, + 604 + ], + "score": 0.7, + "latex": "4 4 . 7 5 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1085, + 440, + 1243, + 440, + 1243, + 466, + 1085, + 466 + ], + "score": 0.69, + "latex": "5 2 . 4 4 \\pm 0 . 8 0" + }, + { + "category_id": 13, + "poly": [ + 1084, + 649, + 1242, + 649, + 1242, + 674, + 1084, + 674 + ], + "score": 0.69, + "latex": "4 8 . 1 2 \\pm 0 . 5 3" + }, + { + "category_id": 13, + "poly": [ + 900, + 372, + 1058, + 372, + 1058, + 398, + 900, + 398 + ], + "score": 0.67, + "latex": "3 8 . 6 0 \\pm 0 . 6 7" + }, + { + "category_id": 13, + "poly": [ + 1269, + 649, + 1428, + 649, + 1428, + 674, + 1269, + 674 + ], + "score": 0.67, + "latex": "7 5 . 0 0 \\pm 0 . 3 9" + }, + { + "category_id": 13, + "poly": [ + 1269, + 372, + 1428, + 372, + 1428, + 397, + 1269, + 397 + ], + "score": 0.67, + "latex": "5 5 . 2 9 \\pm 0 . 4 7" + }, + { + "category_id": 13, + "poly": [ + 900, + 475, + 1058, + 475, + 1058, + 500, + 900, + 500 + ], + "score": 0.67, + "latex": "3 6 . 3 8 \\pm 0 . 8 8" + }, + { + "category_id": 13, + "poly": [ + 1084, + 544, + 1243, + 544, + 1243, + 570, + 1084, + 570 + ], + "score": 0.66, + "latex": "5 0 . 1 9 \\pm 0 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1084, + 372, + 1243, + 372, + 1243, + 398, + 1084, + 398 + ], + "score": 0.66, + "latex": "4 0 . 6 0 \\pm 0 . 5 8" + }, + { + "category_id": 13, + "poly": [ + 1084, + 475, + 1243, + 475, + 1243, + 500, + 1084, + 500 + ], + "score": 0.66, + "latex": "3 9 . 1 4 \\pm 0 . 8 5" + }, + { + "category_id": 13, + "poly": [ + 1269, + 475, + 1428, + 475, + 1428, + 500, + 1269, + 500 + ], + "score": 0.63, + "latex": "6 5 . 4 1 \\pm 0 . 6 9" + }, + { + "category_id": 13, + "poly": [ + 900, + 614, + 1058, + 614, + 1058, + 639, + 900, + 639 + ], + "score": 0.62, + "latex": "4 4 . 4 7 \\pm 0 . 2 1" + }, + { + "category_id": 13, + "poly": [ + 1269, + 580, + 1429, + 580, + 1429, + 604, + 1269, + 604 + ], + "score": 0.62, + "latex": "4 6 . 8 8 \\pm 0 . 4 9" + }, + { + "category_id": 13, + "poly": [ + 901, + 543, + 1058, + 543, + 1058, + 570, + 901, + 570 + ], + "score": 0.62, + "latex": "3 1 . 1 0 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 1085, + 614, + 1243, + 614, + 1243, + 639, + 1085, + 639 + ], + "score": 0.59, + "latex": "4 4 . 4 7 \\pm 0 . 2 0" + }, + { + "category_id": 13, + "poly": [ + 1269, + 543, + 1428, + 543, + 1428, + 569, + 1269, + 569 + ], + "score": 0.58, + "latex": "6 0 . 3 4 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 715, + 441, + 873, + 441, + 873, + 466, + 715, + 466 + ], + "score": 0.57, + "latex": "3 3 . 2 1 \\pm 0 . 8 1" + }, + { + "category_id": 13, + "poly": [ + 900, + 682, + 1058, + 682, + 1058, + 709, + 900, + 709 + ], + "score": 0.56, + "latex": "3 8 . 5 4 \\pm 0 . 6 4" + }, + { + "category_id": 13, + "poly": [ + 715, + 405, + 874, + 405, + 874, + 432, + 715, + 432 + ], + "score": 0.56, + "latex": "4 0 . 3 0 \\pm 0 . 6 0" + }, + { + "category_id": 13, + "poly": [ + 1085, + 510, + 1243, + 510, + 1243, + 535, + 1085, + 535 + ], + "score": 0.54, + "latex": "5 5 . 7 8 \\pm 0 . 3 6" + }, + { + "category_id": 13, + "poly": [ + 715, + 510, + 872, + 510, + 872, + 535, + 715, + 535 + ], + "score": 0.5, + "latex": "3 8 . 5 0 \\pm 0 . 5 1" + }, + { + "category_id": 13, + "poly": [ + 1085, + 682, + 1243, + 682, + 1243, + 709, + 1085, + 709 + ], + "score": 0.5, + "latex": "3 4 . 1 5 \\pm 0 . 6 4" + }, + { + "category_id": 13, + "poly": [ + 1269, + 614, + 1428, + 614, + 1428, + 639, + 1269, + 639 + ], + "score": 0.48, + "latex": "4 4 . 5 1 \\pm 0 . 2 8" + }, + { + "category_id": 13, + "poly": [ + 901, + 304, + 1058, + 304, + 1058, + 329, + 901, + 329 + ], + "score": 0.45, + "latex": "4 0 . 6 7 \\pm 0 . 3 6" + }, + { + "category_id": 13, + "poly": [ + 715, + 474, + 873, + 474, + 873, + 502, + 715, + 502 + ], + "score": 0.42, + "latex": "4 6 . 8 8 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 1270, + 683, + 1428, + 683, + 1428, + 708, + 1270, + 708 + ], + "score": 0.4, + "latex": "6 9 . 4 5 \\pm 0 . 5 5" + }, + { + "category_id": 13, + "poly": [ + 715, + 372, + 874, + 372, + 874, + 398, + 715, + 398 + ], + "score": 0.39, + "latex": "3 6 . 5 5 \\pm 0 . 6 6" + }, + { + "category_id": 13, + "poly": [ + 715, + 649, + 874, + 649, + 874, + 674, + 715, + 674 + ], + "score": 0.38, + "latex": "4 5 . 1 2 \\pm 0 . 4 6" + }, + { + "category_id": 13, + "poly": [ + 1083, + 782, + 1243, + 782, + 1243, + 812, + 1083, + 812 + ], + "score": 0.35, + "latex": "4 7 . 6 5 \\pm 0 . 2 7" + }, + { + "category_id": 13, + "poly": [ + 715, + 543, + 873, + 543, + 873, + 570, + 715, + 570 + ], + "score": 0.3, + "latex": "4 1 . 7 0 \\pm 0 . 4 2" + }, + { + "category_id": 13, + "poly": [ + 715, + 580, + 873, + 580, + 873, + 604, + 715, + 604 + ], + "score": 0.3, + "latex": "4 4 . 9 3 \\pm 0 . 4 2" + }, + { + "category_id": 13, + "poly": [ + 899, + 782, + 1058, + 782, + 1058, + 812, + 899, + 812 + ], + "score": 0.29, + "latex": "4 4 . 6 0 \\pm 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 1085, + 714, + 1244, + 714, + 1244, + 743, + 1085, + 743 + ], + "score": 0.28, + "latex": "3 4 . 6 5 \\pm 0 . 4 4" + }, + { + "category_id": 13, + "poly": [ + 1084, + 746, + 1243, + 746, + 1243, + 778, + 1084, + 778 + ], + "score": 0.28, + "latex": "3 4 . 4 0 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1269, + 304, + 1431, + 304, + 1431, + 329, + 1269, + 329 + ], + "score": 0.27, + "latex": "5 4 . 7 0 \\pm 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 715, + 613, + 873, + 613, + 873, + 639, + 715, + 639 + ], + "score": 0.27, + "latex": "4 4 . 4 7 \\pm 0 . 2 1" + }, + { + "category_id": 13, + "poly": [ + 899, + 713, + 1058, + 713, + 1058, + 743, + 899, + 743 + ], + "score": 0.26, + "latex": "3 7 . 8 4 \\pm 0 . 6 3" + }, + { + "category_id": 13, + "poly": [ + 1086, + 304, + 1242, + 304, + 1242, + 329, + 1086, + 329 + ], + "score": 0.25, + "latex": "3 9 . 8 9 \\pm 0 . 2 2" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 109.0, + 813.0, + 109.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1647.0, + 1041.0, + 1647.0, + 1041.0, + 1687.0, + 249.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 926.0, + 1240.0, + 926.0, + 1240.0, + 959.0, + 457.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1862.0, + 824.0, + 1862.0, + 824.0, + 1899.0, + 786.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 979.0, + 661.0, + 979.0, + 661.0, + 1022.0, + 250.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1504.0, + 1450.0, + 1504.0, + 1450.0, + 1542.0, + 250.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1534.0, + 1450.0, + 1534.0, + 1450.0, + 1575.0, + 249.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1569.0, + 1226.0, + 1569.0, + 1226.0, + 1603.0, + 250.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1713.0, + 1450.0, + 1713.0, + 1450.0, + 1757.0, + 247.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1746.0, + 1452.0, + 1746.0, + 1452.0, + 1787.0, + 246.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1776.0, + 432.0, + 1776.0, + 432.0, + 1814.0, + 249.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1504.0, + 1450.0, + 1504.0, + 1450.0, + 1542.0, + 250.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1534.0, + 1450.0, + 1534.0, + 1450.0, + 1575.0, + 249.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1569.0, + 1226.0, + 1569.0, + 1226.0, + 1603.0, + 250.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1021.0, + 370.0, + 1021.0, + 370.0, + 1056.0, + 251.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1021.0, + 725.0, + 1021.0, + 725.0, + 1056.0, + 534.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1052.0, + 347.0, + 1052.0, + 347.0, + 1087.0, + 343.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1052.0, + 759.0, + 1052.0, + 759.0, + 1087.0, + 462.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1082.0, + 359.0, + 1082.0, + 359.0, + 1118.0, + 251.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1082.0, + 749.0, + 1082.0, + 749.0, + 1118.0, + 523.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 260.0, + 1111.0, + 404.0, + 1111.0, + 404.0, + 1149.0, + 260.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1111.0, + 579.0, + 1111.0, + 579.0, + 1149.0, + 568.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 261.0, + 1140.0, + 383.0, + 1140.0, + 383.0, + 1179.0, + 261.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 1174.0, + 293.0, + 1174.0, + 293.0, + 1207.0, + 262.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1173.0, + 1052.0, + 1173.0, + 1052.0, + 1209.0, + 323.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 1209.0, + 293.0, + 1209.0, + 293.0, + 1241.0, + 262.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1200.0, + 414.0, + 1200.0, + 414.0, + 1254.0, + 324.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1200.0, + 641.0, + 1200.0, + 641.0, + 1254.0, + 470.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1200.0, + 1130.0, + 1200.0, + 1130.0, + 1254.0, + 677.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1200.0, + 1319.0, + 1200.0, + 1319.0, + 1254.0, + 1314.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 1251.0, + 293.0, + 1251.0, + 293.0, + 1285.0, + 262.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 1295.0, + 293.0, + 1295.0, + 293.0, + 1327.0, + 262.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1290.0, + 414.0, + 1290.0, + 414.0, + 1331.0, + 324.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1290.0, + 641.0, + 1290.0, + 641.0, + 1331.0, + 470.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1290.0, + 1106.0, + 1290.0, + 1106.0, + 1331.0, + 677.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 264.0, + 1336.0, + 293.0, + 1336.0, + 293.0, + 1369.0, + 264.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1330.0, + 414.0, + 1330.0, + 414.0, + 1380.0, + 324.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1330.0, + 633.0, + 1330.0, + 633.0, + 1380.0, + 462.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1330.0, + 1103.0, + 1330.0, + 1103.0, + 1380.0, + 672.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1330.0, + 1276.0, + 1330.0, + 1276.0, + 1380.0, + 1272.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 264.0, + 1378.0, + 293.0, + 1378.0, + 293.0, + 1409.0, + 264.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1366.0, + 414.0, + 1366.0, + 414.0, + 1420.0, + 324.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1366.0, + 635.0, + 1366.0, + 635.0, + 1420.0, + 465.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1366.0, + 1107.0, + 1366.0, + 1107.0, + 1420.0, + 676.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1366.0, + 1284.0, + 1366.0, + 1284.0, + 1420.0, + 1279.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1415.0, + 292.0, + 1415.0, + 292.0, + 1442.0, + 265.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1409.0, + 519.0, + 1409.0, + 519.0, + 1446.0, + 297.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 1443.0, + 373.0, + 1443.0, + 373.0, + 1478.0, + 253.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1443.0, + 573.0, + 1443.0, + 573.0, + 1478.0, + 564.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1236.0, + 1309.0, + 1236.0, + 1309.0, + 1300.0, + 317.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.25, + 1237.5, + 1298.25, + 1237.5, + 1298.25, + 1272.5, + 1247.25, + 1272.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.25, + 1283.5, + 1290.25, + 1283.5, + 1290.25, + 1315.0, + 1242.25, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1303.0, + 1289.0, + 1303.0, + 1289.0, + 1333.0, + 1221.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1322.0, + 1273.0, + 1322.0, + 1273.0, + 1357.5, + 1221.0, + 1357.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 253, + 1565, + 1448, + 1565, + 1448, + 1812, + 253, + 1812 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 287, + 254, + 1310, + 254, + 1310, + 1281, + 287, + 1281 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 252, + 1302, + 1450, + 1302, + 1450, + 1443, + 252, + 1443 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 251, + 1487, + 1445, + 1487, + 1445, + 1551, + 251, + 1551 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 791, + 1866, + 819, + 1866, + 819, + 1889, + 791, + 1889 + ], + "score": 0.835 + }, + { + "category_id": 13, + "poly": [ + 1106, + 1720, + 1132, + 1720, + 1132, + 1746, + 1106, + 1746 + ], + "score": 0.76, + "latex": "D" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 258.0, + 656.0, + 258.0, + 656.0, + 295.0, + 545.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 256.0, + 1189.0, + 256.0, + 1189.0, + 296.0, + 1094.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 324.0, + 751.0, + 324.0, + 751.0, + 340.0, + 719.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 337.0, + 755.0, + 337.0, + 755.0, + 357.0, + 718.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 352.0, + 768.0, + 352.0, + 768.0, + 372.0, + 718.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 367.0, + 330.0, + 367.0, + 330.0, + 539.0, + 291.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 370.0, + 748.0, + 370.0, + 748.0, + 386.0, + 719.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 591.0, + 620.0, + 591.0, + 620.0, + 623.0, + 580.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 589.0, + 1162.0, + 589.0, + 1162.0, + 624.0, + 1119.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 660.0, + 568.0, + 660.0, + 568.0, + 685.0, + 506.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 659.0, + 722.0, + 659.0, + 722.0, + 695.0, + 676.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 702.0, + 827.0, + 702.0, + 827.0, + 733.0, + 749.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 704.0, + 332.0, + 704.0, + 332.0, + 851.0, + 290.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 733.0, + 815.0, + 733.0, + 815.0, + 762.0, + 751.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 841.0, + 539.0, + 841.0, + 539.0, + 860.0, + 469.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 856.0, + 550.0, + 856.0, + 550.0, + 875.0, + 466.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 870.0, + 565.0, + 870.0, + 565.0, + 892.0, + 466.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 885.0, + 547.0, + 885.0, + 547.0, + 907.0, + 466.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 919.0, + 620.0, + 919.0, + 620.0, + 952.0, + 580.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 917.0, + 1162.0, + 917.0, + 1162.0, + 954.0, + 1118.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 970.0, + 463.0, + 970.0, + 463.0, + 1005.0, + 388.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 979.0, + 742.0, + 979.0, + 742.0, + 998.0, + 672.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 992.0, + 755.0, + 992.0, + 755.0, + 1014.0, + 671.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1004.0, + 450.0, + 1004.0, + 450.0, + 1032.0, + 385.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1007.0, + 770.0, + 1007.0, + 770.0, + 1029.0, + 671.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1022.0, + 751.0, + 1022.0, + 751.0, + 1045.0, + 671.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1056.0, + 326.0, + 1056.0, + 326.0, + 1163.0, + 283.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1176.0, + 837.0, + 1176.0, + 837.0, + 1208.0, + 763.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1209.0, + 825.0, + 1209.0, + 825.0, + 1237.0, + 765.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1249.0, + 620.0, + 1249.0, + 620.0, + 1281.0, + 581.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1249.0, + 1156.0, + 1249.0, + 1156.0, + 1278.0, + 1124.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1302.0, + 1450.0, + 1302.0, + 1450.0, + 1334.0, + 251.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1331.0, + 1453.0, + 1331.0, + 1453.0, + 1362.0, + 249.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1359.0, + 1450.0, + 1359.0, + 1450.0, + 1390.0, + 249.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1384.0, + 1450.0, + 1384.0, + 1450.0, + 1420.0, + 248.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1414.0, + 549.0, + 1414.0, + 549.0, + 1445.0, + 249.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1863.0, + 823.0, + 1863.0, + 823.0, + 1902.0, + 786.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 1568.0, + 1449.0, + 1568.0, + 1449.0, + 1598.0, + 252.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1595.0, + 1451.0, + 1595.0, + 1451.0, + 1630.0, + 247.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1629.0, + 1451.0, + 1629.0, + 1451.0, + 1663.0, + 249.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1656.0, + 1452.0, + 1656.0, + 1452.0, + 1694.0, + 249.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1686.0, + 1452.0, + 1686.0, + 1452.0, + 1725.0, + 248.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1717.0, + 1105.0, + 1717.0, + 1105.0, + 1755.0, + 246.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1717.0, + 1453.0, + 1717.0, + 1453.0, + 1755.0, + 1133.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1748.0, + 1449.0, + 1748.0, + 1449.0, + 1786.0, + 248.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1777.0, + 1452.0, + 1777.0, + 1452.0, + 1816.0, + 246.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1486.0, + 1448.0, + 1486.0, + 1448.0, + 1522.0, + 251.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1519.0, + 1243.0, + 1519.0, + 1243.0, + 1555.0, + 251.0, + 1555.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 250, + 204, + 1444, + 204, + 1444, + 594, + 250, + 594 + ], + "score": 0.982, + "html": "
methodC08→C05C08→C09C08→C13C08→C14C09→C05C09→C08C09→C13C09→C14
Source Only33.70±0.3350.10±0.2850.80±0.3240.13±0.2633.32±0.4448.24±0.3249.24±0.3036.19±0.27
1-NN28.7535.3939.7932.1326.8235.3034.2628.41
CORAL (Sun & Saenko (2016))35.89 ± 0.4455.79±0.5060.00 ±0.2940.67 ± 0.4835.89 ± 0.4051.56±0.4350.45 ± 0.4140.67± 0.35
DANN (Ganin et al. (2016))40.20±0.5056.89 ± 0.3955.83 ±0.4043.25±0.4150.63±0.3858.40 ± 0.5155.81 ± 0.5348.90 ±0.43
ADDA(Tzeng et al. (2017))37.40±0.6858.40±0.7360.40±0.8342.10±0.4829.40±0.7053.30±0.4945.30± 0.5338.30±0.63
c-ADDA41.60±0.6439.65 ±0.7050.00 ±0.5246.25±0.5245.01±0.6352.14 ± 0.5337.43 ± 0.6043.26±0.58
DTN (Zhang et al. (2015))44.13 ± 0.4157.42 ± 0.4255.89±0.4845.76± 0.3944.53± 0.4957.34±0.3552.43±0.3851.55 ± 0.40
c-DTN45.10 ±0.4449.78±0.5047.43 ±0.4645.79 ±0.4849.80±0.4055.69 ± 0.3550.10 ± 0.3852.31 ±0.29
PixelDA (Bousmalis et al. (2017))46.45 ± 0.4544.33±0.3844.87 ± 0.4146.83±0.2945.63±0.3416.37 ± 0.2745.43 ± 0.3547.00± 0.49
UNIT (Liu et al. (2017))43.88±0.1843.99±0.2344.47±0.1944.47 ± 0.2444.47 ± 0.1743.95± 0.2144.64±0.2244.47 ±0.19
DSN (Bousmalis et al. (2016)) c-DSN46.25± 0.5347.50±0.6062.15 ± 0.5839.72± 0.5545.85±0.4856.65±0.5056.5±0.3842.87±0.43
45.82 ± 0.5344.64± 0.4245.60 ± 0.4846.32±0.5245.18 ± 0.4745.52 ± 0.5544.79 ± 0.5347.37 ± 0.48
1p-DSN44.12 ± 0.7344.14 ± 0.2045.00 ±0.3845.62 ± 0.4244.78 ±0.4745.02 ±0.6544.21 ± 0.4846.97 ± 0.38
mp-DSN42.19±0.4644.70±0.5342.47 ± 0.4840.50±0.3945.00±0.5143.80±0.5045.79±0.4842.39 ±0.49
s-MTDA-ITA44.77±0.1945.61±0.1860.00±0.2746.70±0.2849.06±0.2455.33±0.2259.90±0.3050.64± 0.26
c-MTDA-ITA MTDA-ITA44.35± 0.2742.67 ± 0.2458.90 ±0.2644.32 ± 0.2646.74 ±0.2254.11 ± 0.2156.89 ± 0.2349.64 ± 0.19
46.30 ± 0.2560.60 ± 0.1860.50±0.1950.40±0.2055.59± 0.2557.80±0.2164.20 ±0.1856.34 ± 0.20
" + }, + { + "category_id": 5, + "poly": [ + 253, + 680, + 1447, + 680, + 1447, + 880, + 253, + 880 + ], + "score": 0.977, + "html": "
methodP→AP→CP→SA→PA→CA→S
1-NN15.2818.1625.6022.7019.7522.70
ADDA(Tzeng et al. (2017))24.35 ± 2.3720.12 ± 2.5022.45 ± 2.1132.57 ± 2.7017.68 ± 2.0418.90 ± 2.48
DSN (Bousmalis et al. (2016))28.42 ±2.1221.14 ± 2.082.04 ± 1.9029.54 ± 1.9525.89 ± 1.8824.69 ± 2.08
s-MTDA-ITA28.02 ± 1.5921.64±1.2426.24 ± 1.6031.06 ± 1.5025.09±1.4025.89±1.03
c-MTDA-ITA25.35 ± 1.8020.24 ± 1.3923.64 ± 1.6026.54 ± 1.3320.30 ± 1.2922.38 ± 1.45
MTDA-ITA31.40 ±1.5523.05 ±1.0428.24 ±1.7835.74 ± 1.5027.00 ±1.2528.90 ±1.60
" + }, + { + "category_id": 1, + "poly": [ + 245, + 996, + 1449, + 996, + 1449, + 1091, + 245, + 1091 + ], + "score": 0.957 + }, + { + "category_id": 2, + "poly": [ + 252, + 112, + 812, + 112, + 812, + 141, + 252, + 141 + ], + "score": 0.895 + }, + { + "category_id": 6, + "poly": [ + 508, + 639, + 1187, + 639, + 1187, + 668, + 508, + 668 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 790, + 1865, + 821, + 1865, + 821, + 1889, + 790, + 1889 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 246, + 921, + 1446, + 921, + 1446, + 978, + 246, + 978 + ], + "score": 0.542 + }, + { + "category_id": 6, + "poly": [ + 246, + 921, + 1446, + 921, + 1446, + 978, + 246, + 978 + ], + "score": 0.1 + }, + { + "category_id": 13, + "poly": [ + 1340, + 687, + 1409, + 687, + 1409, + 706, + 1340, + 706 + ], + "score": 0.73, + "latex": "\\mathbf A \\to \\mathbf S" + }, + { + "category_id": 13, + "poly": [ + 653, + 206, + 748, + 206, + 748, + 227, + 653, + 227 + ], + "score": 0.7, + "latex": "\\overline { { \\mathbf { C 0 8 } \\to \\mathbf { C 0 9 } } }" + }, + { + "category_id": 13, + "poly": [ + 1105, + 275, + 1204, + 275, + 1204, + 291, + 1105, + 291 + ], + "score": 0.68, + "latex": "5 1 . 5 6 \\pm 0 . 4 3" + }, + { + "category_id": 13, + "poly": [ + 539, + 206, + 634, + 206, + 634, + 227, + 539, + 227 + ], + "score": 0.63, + "latex": "\\mathbf { \\overline { { C O 8 \\to C O 5 } } }" + }, + { + "category_id": 13, + "poly": [ + 994, + 233, + 1090, + 233, + 1090, + 249, + 994, + 249 + ], + "score": 0.6, + "latex": "3 3 . 3 2 \\pm 0 . 4 4" + }, + { + "category_id": 13, + "poly": [ + 1107, + 210, + 1202, + 210, + 1202, + 226, + 1107, + 226 + ], + "score": 0.58, + "latex": "\\mathbf { C 0 9 } \\to \\mathbf { C 0 8 }" + }, + { + "category_id": 13, + "poly": [ + 992, + 275, + 1090, + 275, + 1090, + 291, + 992, + 291 + ], + "score": 0.56, + "latex": "3 5 . 8 9 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 992, + 380, + 1091, + 380, + 1091, + 397, + 992, + 397 + ], + "score": 0.56, + "latex": "4 9 . 8 0 \\pm 0 . 4 0" + }, + { + "category_id": 13, + "poly": [ + 1107, + 234, + 1203, + 234, + 1203, + 249, + 1107, + 249 + ], + "score": 0.52, + "latex": "4 8 . 2 4 \\pm 0 . 3 2" + }, + { + "category_id": 13, + "poly": [ + 1222, + 210, + 1315, + 210, + 1315, + 226, + 1222, + 226 + ], + "score": 0.49, + "latex": "{ \\bf C } 0 9 { \\bf C } 1 3" + }, + { + "category_id": 13, + "poly": [ + 651, + 380, + 750, + 380, + 750, + 397, + 651, + 397 + ], + "score": 0.48, + "latex": "4 9 . 7 8 \\pm 0 . 5 0" + }, + { + "category_id": 13, + "poly": [ + 764, + 421, + 863, + 421, + 863, + 440, + 764, + 440 + ], + "score": 0.47, + "latex": "\\overline { { 4 4 . 4 7 \\pm 0 . 1 9 } }" + }, + { + "category_id": 13, + "poly": [ + 1219, + 275, + 1317, + 275, + 1317, + 291, + 1219, + 291 + ], + "score": 0.46, + "latex": "5 0 . 4 5 \\pm 0 . 4 1" + }, + { + "category_id": 13, + "poly": [ + 994, + 210, + 1088, + 210, + 1088, + 226, + 994, + 226 + ], + "score": 0.45, + "latex": "{ \\bf C } 0 9 \\to { \\bf C } 0 5" + }, + { + "category_id": 13, + "poly": [ + 877, + 276, + 976, + 276, + 976, + 291, + 877, + 291 + ], + "score": 0.44, + "latex": "4 0 . 6 / \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 766, + 206, + 863, + 206, + 863, + 227, + 766, + 227 + ], + "score": 0.44, + "latex": "\\overline { { \\mathbf { C 0 8 } \\to \\mathbf { C 1 3 } } }" + }, + { + "category_id": 13, + "poly": [ + 764, + 527, + 863, + 527, + 863, + 546, + 764, + 546 + ], + "score": 0.43, + "latex": "\\overline { { 6 0 . 0 0 \\pm 0 . 2 7 } }" + }, + { + "category_id": 13, + "poly": [ + 878, + 505, + 977, + 505, + 977, + 524, + 878, + 524 + ], + "score": 0.43, + "latex": "4 0 . 5 0 \\pm 0 . 3 9" + }, + { + "category_id": 13, + "poly": [ + 992, + 359, + 1090, + 359, + 1090, + 375, + 992, + 375 + ], + "score": 0.41, + "latex": "4 4 . 5 3 \\pm 0 . 4 9" + }, + { + "category_id": 13, + "poly": [ + 1106, + 381, + 1204, + 381, + 1204, + 397, + 1106, + 397 + ], + "score": 0.41, + "latex": "5 5 . 6 9 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 763, + 380, + 863, + 380, + 863, + 397, + 763, + 397 + ], + "score": 0.39, + "latex": "4 7 . 4 3 \\pm 0 . 4 6" + }, + { + "category_id": 13, + "poly": [ + 651, + 230, + 749, + 230, + 749, + 249, + 651, + 249 + ], + "score": 0.39, + "latex": "\\overline { { 5 0 . 1 0 \\pm 0 . 2 8 } }" + }, + { + "category_id": 13, + "poly": [ + 992, + 338, + 1090, + 338, + 1090, + 354, + 992, + 354 + ], + "score": 0.38, + "latex": "4 5 . 0 1 \\pm 0 . 6 3" + }, + { + "category_id": 13, + "poly": [ + 1220, + 233, + 1317, + 233, + 1317, + 249, + 1220, + 249 + ], + "score": 0.37, + "latex": "4 9 . 2 4 \\pm 0 . 3 0" + }, + { + "category_id": 13, + "poly": [ + 651, + 401, + 750, + 401, + 750, + 418, + 651, + 418 + ], + "score": 0.37, + "latex": "4 4 . 3 3 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 878, + 381, + 976, + 381, + 976, + 397, + 878, + 397 + ], + "score": 0.37, + "latex": "4 5 . 7 9 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 764, + 484, + 863, + 484, + 863, + 503, + 764, + 503 + ], + "score": 0.37, + "latex": "4 5 . 0 0 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 878, + 234, + 976, + 234, + 976, + 249, + 878, + 249 + ], + "score": 0.36, + "latex": "4 0 . 1 3 \\pm 0 . 2 6" + }, + { + "category_id": 13, + "poly": [ + 992, + 401, + 1090, + 401, + 1090, + 418, + 992, + 418 + ], + "score": 0.36, + "latex": "4 5 . 6 3 \\pm 0 . 3 4" + }, + { + "category_id": 13, + "poly": [ + 650, + 421, + 750, + 421, + 750, + 440, + 650, + 440 + ], + "score": 0.36, + "latex": "\\overline { { 4 3 . 9 9 \\pm 0 . 2 3 } }" + }, + { + "category_id": 13, + "poly": [ + 878, + 527, + 977, + 527, + 977, + 546, + 878, + 546 + ], + "score": 0.36, + "latex": "\\overline { { 4 6 . 7 0 \\pm 0 . 2 8 } }" + }, + { + "category_id": 13, + "poly": [ + 992, + 527, + 1091, + 527, + 1091, + 546, + 992, + 546 + ], + "score": 0.36, + "latex": "\\overline { { 4 9 . 0 6 \\pm 0 . 2 4 } }" + }, + { + "category_id": 13, + "poly": [ + 1219, + 380, + 1317, + 380, + 1317, + 397, + 1219, + 397 + ], + "score": 0.35, + "latex": "5 0 . 1 0 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 764, + 505, + 863, + 505, + 863, + 525, + 764, + 525 + ], + "score": 0.34, + "latex": "4 2 . 4 7 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 1106, + 359, + 1203, + 359, + 1203, + 375, + 1106, + 375 + ], + "score": 0.34, + "latex": "5 7 . 3 4 \\pm 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 651, + 527, + 749, + 527, + 749, + 546, + 651, + 546 + ], + "score": 0.34, + "latex": "\\overline { { 4 5 . 6 1 \\pm 0 . 1 8 } }" + }, + { + "category_id": 13, + "poly": [ + 878, + 401, + 977, + 401, + 977, + 418, + 878, + 418 + ], + "score": 0.34, + "latex": "4 6 . 8 3 \\pm 0 . 2 9" + }, + { + "category_id": 13, + "poly": [ + 877, + 464, + 977, + 464, + 977, + 482, + 877, + 482 + ], + "score": 0.33, + "latex": "4 6 . 3 2 \\pm 0 . 5 2" + }, + { + "category_id": 13, + "poly": [ + 878, + 484, + 977, + 484, + 977, + 503, + 878, + 503 + ], + "score": 0.33, + "latex": "4 5 . 6 2 \\pm 0 . 4 2" + }, + { + "category_id": 13, + "poly": [ + 877, + 359, + 977, + 359, + 977, + 375, + 877, + 375 + ], + "score": 0.32, + "latex": "4 5 . 7 6 \\pm 0 . 3 9" + }, + { + "category_id": 13, + "poly": [ + 764, + 401, + 863, + 401, + 863, + 418, + 764, + 418 + ], + "score": 0.32, + "latex": "4 4 . 8 7 \\pm 0 . 4 1" + }, + { + "category_id": 13, + "poly": [ + 1106, + 338, + 1204, + 338, + 1204, + 354, + 1106, + 354 + ], + "score": 0.32, + "latex": "5 2 . 1 4 \\pm 0 . 5 3" + }, + { + "category_id": 13, + "poly": [ + 765, + 230, + 863, + 230, + 863, + 249, + 765, + 249 + ], + "score": 0.32, + "latex": "\\overline { { 5 0 . 8 0 \\pm 0 . 3 2 } }" + }, + { + "category_id": 13, + "poly": [ + 992, + 464, + 1091, + 464, + 1091, + 482, + 992, + 482 + ], + "score": 0.31, + "latex": "4 5 . 1 8 \\pm 0 . 4 7" + }, + { + "category_id": 13, + "poly": [ + 992, + 505, + 1090, + 505, + 1090, + 524, + 992, + 524 + ], + "score": 0.31, + "latex": "4 5 . 0 0 \\pm 0 . 5 1" + }, + { + "category_id": 13, + "poly": [ + 650, + 505, + 749, + 505, + 749, + 525, + 650, + 525 + ], + "score": 0.3, + "latex": "4 4 . 7 0 \\pm 0 . 5 3" + }, + { + "category_id": 13, + "poly": [ + 1336, + 210, + 1433, + 210, + 1433, + 226, + 1336, + 226 + ], + "score": 0.3, + "latex": "{ \\bf C } 0 9 { \\bf C } 1 4" + }, + { + "category_id": 13, + "poly": [ + 1106, + 401, + 1203, + 401, + 1203, + 419, + 1106, + 419 + ], + "score": 0.3, + "latex": "1 6 . 3 7 \\pm 0 . 2 7" + }, + { + "category_id": 13, + "poly": [ + 1105, + 484, + 1204, + 484, + 1204, + 503, + 1105, + 503 + ], + "score": 0.29, + "latex": "4 5 . 0 2 \\pm 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 991, + 421, + 1091, + 421, + 1091, + 439, + 991, + 439 + ], + "score": 0.29, + "latex": "\\overline { { 4 4 . 4 7 \\pm 0 . 1 7 } }" + }, + { + "category_id": 13, + "poly": [ + 1218, + 358, + 1316, + 358, + 1316, + 375, + 1218, + 375 + ], + "score": 0.29, + "latex": "5 2 . 4 3 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 1332, + 487, + 1435, + 487, + 1435, + 503, + 1332, + 503 + ], + "score": 0.29, + "latex": "4 6 . 9 7 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 877, + 421, + 977, + 421, + 977, + 439, + 877, + 439 + ], + "score": 0.28, + "latex": "\\overline { { 4 4 . 4 7 \\pm 0 . 2 4 } }" + }, + { + "category_id": 13, + "poly": [ + 1331, + 467, + 1435, + 467, + 1435, + 482, + 1331, + 482 + ], + "score": 0.28, + "latex": "4 7 . 3 7 \\pm 0 . 4 8" + }, + { + "category_id": 13, + "poly": [ + 1104, + 421, + 1204, + 421, + 1204, + 440, + 1104, + 440 + ], + "score": 0.27, + "latex": "\\overline { { 4 3 . 9 5 \\pm 0 . 2 1 } }" + }, + { + "category_id": 13, + "poly": [ + 992, + 296, + 1091, + 296, + 1091, + 312, + 992, + 312 + ], + "score": 0.27, + "latex": "5 0 . 6 3 \\pm 0 . 3 8" + }, + { + "category_id": 13, + "poly": [ + 763, + 464, + 864, + 464, + 864, + 482, + 763, + 482 + ], + "score": 0.27, + "latex": "4 5 . 6 0 \\pm 0 . 4 8" + }, + { + "category_id": 14, + "poly": [ + 276, + 675, + 1389, + 675, + 1389, + 881, + 276, + 881 + ], + "score": 0.27, + "latex": "\\begin{array} { r l } { \\frac { \\mathrm { m e t h o d } } { \\mathrm { A D B A } ( \\mathrm { T r } ) } \\frac { | \\mathrm { P } \\mathrm { A } | } { | \\mathrm { A D V } | } \\frac { \\mathrm { P } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 1 6 } } \\frac { \\mathrm { P } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 5 } 6 0 } \\frac { \\mathrm { P } \\mathrm { S } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { P } } { | \\mathrm { B } \\mathrm { 2 5 } 6 } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } \\mathrm { A } } \\frac { \\mathrm { A } \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 2 } . \\mathrm { 7 0 } \\mathrm { B } \\mathrm { 2 } \\mathrm { B } } } & = \\mathrm { 8 } \\frac { \\mathrm { A } } { \\mathrm { 8 } } \\frac { \\mathrm { C } } { | \\mathrm { B } \\mathrm { 2 } . \\mathrm { 3 } 6 0 } \\frac { \\mathrm { A } } { \\mathrm { 8 } } \\frac { \\mathrm { C } } | \\mathrm { B } \\mathrm 2 \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 650, + 484, + 750, + 484, + 750, + 503, + 650, + 503 + ], + "score": 0.27, + "latex": "4 4 . 1 4 \\pm 0 . 2 0" + }, + { + "category_id": 13, + "poly": [ + 992, + 484, + 1091, + 484, + 1091, + 503, + 992, + 503 + ], + "score": 0.25, + "latex": "4 4 . 7 8 \\pm 0 . 4 7" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 109.0, + 813.0, + 109.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 635.0, + 1189.0, + 635.0, + 1189.0, + 673.0, + 508.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1861.0, + 827.0, + 1861.0, + 827.0, + 1900.0, + 784.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 919.0, + 1451.0, + 919.0, + 1451.0, + 953.0, + 247.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 946.0, + 318.0, + 946.0, + 318.0, + 984.0, + 245.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 995.0, + 1450.0, + 995.0, + 1450.0, + 1033.0, + 251.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1025.0, + 1449.0, + 1025.0, + 1449.0, + 1063.0, + 251.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 1057.0, + 1443.0, + 1057.0, + 1443.0, + 1095.0, + 251.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 919.0, + 1451.0, + 919.0, + 1451.0, + 953.0, + 247.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 245.0, + 946.0, + 318.0, + 946.0, + 318.0, + 984.0, + 245.0, + 984.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 448, + 225, + 1243, + 225, + 1243, + 1794, + 448, + 1794 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 250, + 1803, + 1453, + 1803, + 1453, + 1999, + 250, + 1999 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 252, + 113, + 812, + 113, + 812, + 141, + 252, + 141 + ], + "score": 0.891 + }, + { + "category_id": 15, + "poly": [ + 627.0, + 218.0, + 732.0, + 218.0, + 732.0, + 236.0, + 627.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 218.0, + 1106.0, + 218.0, + 1106.0, + 236.0, + 1019.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 270.0, + 491.0, + 270.0, + 491.0, + 452.0, + 446.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 274.0, + 825.0, + 274.0, + 825.0, + 302.0, + 805.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 276.0, + 704.0, + 276.0, + 704.0, + 290.0, + 692.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 302.0, + 662.0, + 302.0, + 662.0, + 352.0, + 632.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 307.0, + 722.0, + 307.0, + 722.0, + 342.0, + 688.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 309.0, + 954.0, + 309.0, + 954.0, + 337.0, + 940.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 304.0, + 1036.0, + 304.0, + 1036.0, + 326.0, + 1015.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 312.0, + 1064.0, + 312.0, + 1064.0, + 335.0, + 1041.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 319.0, + 1085.0, + 319.0, + 1085.0, + 345.0, + 1073.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 323.0, + 1162.0, + 323.0, + 1162.0, + 338.0, + 1150.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 344.0, + 921.0, + 344.0, + 921.0, + 371.0, + 903.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 366.0, + 1062.0, + 366.0, + 1062.0, + 378.0, + 1048.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 368.0, + 1026.0, + 368.0, + 1026.0, + 384.0, + 1010.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 368.0, + 1097.0, + 368.0, + 1097.0, + 389.0, + 1076.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 384.0, + 707.0, + 384.0, + 707.0, + 401.0, + 681.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 382.0, + 1090.0, + 382.0, + 1090.0, + 403.0, + 1061.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 408.0, + 669.0, + 408.0, + 669.0, + 431.0, + 650.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 410.0, + 688.0, + 410.0, + 688.0, + 432.0, + 674.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 417.0, + 706.0, + 417.0, + 706.0, + 429.0, + 690.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 411.0, + 1022.0, + 411.0, + 1022.0, + 448.0, + 987.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 404.0, + 1095.0, + 404.0, + 1095.0, + 441.0, + 1069.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 460.0, + 704.0, + 460.0, + 704.0, + 500.0, + 655.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 460.0, + 1087.0, + 460.0, + 1087.0, + 498.0, + 1038.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 533.0, + 830.0, + 533.0, + 830.0, + 563.0, + 811.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 547.0, + 485.0, + 547.0, + 485.0, + 688.0, + 450.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 554.0, + 767.0, + 554.0, + 767.0, + 606.0, + 653.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 554.0, + 954.0, + 554.0, + 954.0, + 570.0, + 940.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 559.0, + 1136.0, + 559.0, + 1136.0, + 599.0, + 1041.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 603.0, + 1176.0, + 603.0, + 1176.0, + 617.0, + 1153.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 637.0, + 660.0, + 637.0, + 660.0, + 651.0, + 646.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 655.0, + 623.0, + 655.0, + 623.0, + 664.0, + 615.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 667.0, + 849.0, + 667.0, + 849.0, + 713.0, + 757.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 719.0, + 699.0, + 719.0, + 699.0, + 754.0, + 660.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 719.0, + 1083.0, + 719.0, + 1083.0, + 754.0, + 1040.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 791.0, + 489.0, + 791.0, + 489.0, + 951.0, + 443.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 863.0, + 849.0, + 863.0, + 849.0, + 909.0, + 783.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 851.0, + 949.0, + 851.0, + 949.0, + 891.0, + 893.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 865.0, + 1230.0, + 865.0, + 1230.0, + 910.0, + 1169.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 909.0, + 1027.0, + 909.0, + 1027.0, + 942.0, + 980.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 938.0, + 739.0, + 938.0, + 739.0, + 959.0, + 714.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 936.0, + 1122.0, + 936.0, + 1122.0, + 963.0, + 1078.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 975.0, + 702.0, + 975.0, + 702.0, + 1015.0, + 655.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 975.0, + 1081.0, + 975.0, + 1081.0, + 1011.0, + 1041.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1048.0, + 485.0, + 1048.0, + 485.0, + 1204.0, + 445.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1089.0, + 529.0, + 1089.0, + 529.0, + 1123.0, + 510.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1098.0, + 737.0, + 1098.0, + 737.0, + 1208.0, + 667.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1110.0, + 779.0, + 1110.0, + 779.0, + 1131.0, + 765.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1114.0, + 1164.0, + 1114.0, + 1164.0, + 1138.0, + 1120.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1150.0, + 641.0, + 1150.0, + 641.0, + 1168.0, + 627.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1147.0, + 665.0, + 1147.0, + 665.0, + 1162.0, + 653.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1152.0, + 725.0, + 1152.0, + 725.0, + 1171.0, + 713.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1135.0, + 753.0, + 1135.0, + 753.0, + 1166.0, + 734.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1143.0, + 769.0, + 1143.0, + 769.0, + 1162.0, + 755.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1152.0, + 1024.0, + 1152.0, + 1024.0, + 1168.0, + 1008.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1135.0, + 1134.0, + 1135.0, + 1134.0, + 1166.0, + 1109.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1145.0, + 1150.0, + 1145.0, + 1150.0, + 1164.0, + 1136.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1164.0, + 669.0, + 1164.0, + 669.0, + 1204.0, + 643.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1168.0, + 751.0, + 1168.0, + 751.0, + 1180.0, + 735.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1159.0, + 1144.0, + 1159.0, + 1144.0, + 1187.0, + 1116.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1189.0, + 630.0, + 1189.0, + 630.0, + 1206.0, + 617.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1230.0, + 704.0, + 1230.0, + 704.0, + 1274.0, + 655.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1230.0, + 1087.0, + 1230.0, + 1087.0, + 1270.0, + 1038.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1298.0, + 1090.0, + 1298.0, + 1090.0, + 1342.0, + 1015.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1309.0, + 487.0, + 1309.0, + 487.0, + 1465.0, + 447.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1355.0, + 716.0, + 1355.0, + 716.0, + 1371.0, + 702.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1376.0, + 651.0, + 1376.0, + 651.0, + 1394.0, + 639.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 1380.0, + 1092.0, + 1380.0, + 1092.0, + 1395.0, + 1076.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1389.0, + 667.0, + 1389.0, + 667.0, + 1425.0, + 650.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1395.0, + 746.0, + 1395.0, + 746.0, + 1416.0, + 732.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1390.0, + 765.0, + 1390.0, + 765.0, + 1422.0, + 748.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1397.0, + 1099.0, + 1397.0, + 1099.0, + 1416.0, + 1074.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1397.0, + 1127.0, + 1397.0, + 1127.0, + 1415.0, + 1115.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1392.0, + 1143.0, + 1392.0, + 1143.0, + 1408.0, + 1130.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1409.0, + 1062.0, + 1409.0, + 1062.0, + 1423.0, + 1047.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1411.0, + 1197.0, + 1411.0, + 1197.0, + 1423.0, + 1181.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1425.0, + 554.0, + 1425.0, + 554.0, + 1444.0, + 529.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1416.0, + 1099.0, + 1416.0, + 1099.0, + 1448.0, + 1047.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1423.0, + 1153.0, + 1423.0, + 1153.0, + 1437.0, + 1136.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1434.0, + 1178.0, + 1434.0, + 1178.0, + 1449.0, + 1164.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1451.0, + 723.0, + 1451.0, + 723.0, + 1470.0, + 690.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1448.0, + 1104.0, + 1448.0, + 1104.0, + 1472.0, + 1055.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1489.0, + 699.0, + 1489.0, + 699.0, + 1526.0, + 660.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1491.0, + 1081.0, + 1491.0, + 1081.0, + 1528.0, + 1041.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1564.0, + 487.0, + 1564.0, + 487.0, + 1721.0, + 449.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1745.0, + 704.0, + 1745.0, + 704.0, + 1783.0, + 655.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1747.0, + 1081.0, + 1747.0, + 1081.0, + 1783.0, + 1041.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 341.0, + 787.0, + 341.0, + 787.0, + 356.5, + 782.0, + 356.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.25, + 377.5, + 1119.25, + 377.5, + 1119.25, + 433.5, + 1115.25, + 433.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 396.5, + 1078.0, + 396.5, + 1078.0, + 449.0, + 994.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 403.0, + 1097.0, + 403.0, + 1097.0, + 459.0, + 1034.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 531.5, + 841.0, + 531.5, + 841.0, + 562.5, + 836.0, + 562.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 602.5, + 781.0, + 602.5, + 781.0, + 619.5, + 778.0, + 619.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.25, + 849.5, + 560.25, + 849.5, + 560.25, + 887.5, + 510.25, + 887.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 915.5, + 639.0, + 915.5, + 639.0, + 933.0, + 606.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1187.5, + 1013.0, + 1187.5, + 1013.0, + 1205.0, + 994.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1803.0, + 1451.0, + 1803.0, + 1451.0, + 1836.0, + 250.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1830.0, + 1453.0, + 1830.0, + 1453.0, + 1865.0, + 248.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1859.0, + 1451.0, + 1859.0, + 1451.0, + 1892.0, + 250.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1888.0, + 1451.0, + 1888.0, + 1451.0, + 1921.0, + 249.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1914.0, + 1449.0, + 1914.0, + 1449.0, + 1947.0, + 248.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1938.0, + 1453.0, + 1938.0, + 1453.0, + 1978.0, + 248.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1968.0, + 396.0, + 1968.0, + 396.0, + 2002.0, + 248.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 110.0, + 813.0, + 110.0, + 813.0, + 145.0, + 252.0, + 145.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/BkSmc8qll/BkSmc8qll.md b/parse/train/BkSmc8qll/BkSmc8qll.md new file mode 100644 index 0000000000000000000000000000000000000000..0f1ad19c093df6a31b22b2cef53b080d5e587538 --- /dev/null +++ b/parse/train/BkSmc8qll/BkSmc8qll.md @@ -0,0 +1,451 @@ +# DYNAMIC NEURAL TURING MACHINE WITH CONTINUOUS AND DISCRETE ADDRESSING SCHEMES + +Caglar Gulcehre∗, Sarath Chandar∗, Kyunghyun Cho†, Yoshua Bengio∗ + +∗ University of Montreal, name.lastname@umontreal.ca † New York University, name.lastname@nyu.edu + +# ABSTRACT + +In this paper, we extend neural Turing machine (NTM) into a dynamic neural Turing machine (D-NTM) by introducing a trainable memory addressing scheme. This addressing scheme maintains for each memory cell two separate vectors, content and address vectors. This allows the D-NTM to learn a wide variety of location-based addressing strategies including both linear and nonlinear ones. We implement the D-NTM with both continuous, differentiable and discrete, non-differentiable read/write mechanisms. We investigate the mechanisms and effects for learning to read and write to a memory through experiments on Facebook bAbI tasks using both a feedforward and GRU-controller. The D-NTM is evaluated on a set of Facebook bAbI tasks and shown to outperform NTM and LSTM baselines. We also provide further experimental results on sequential MNIST, associative recall and copy tasks. + +# 1 INTRODUCTION + +Designing general-purpose learning algorithms is one of the long-standing goals of artificial intelligence. Despite the success of deep learning in this area (see, e.g., (Goodfellow et al., 2016)) there are still a set of complex tasks that are not well addressed by conventional neural networks. Those tasks often require a neural network to be equipped with an explicit, external memory in which a larger, potentially unbounded, set of facts need to be stored. They include, but are not limited to, episodic question-answering (Weston et al., 2015b; Hermann et al., 2015; Hill et al., 2015), compact algorithms (Zaremba et al., 2015), dialogue (Serban et al., 2016; Vinyals & Le, 2015) and video caption generation (Yao et al., 2015). + +Recently two promising approaches based on neural networks to this type of tasks have been proposed. Memory networks (Weston et al., 2015b) explicitly store all the facts, or information, available for each episode in an external memory (as continuous vectors) and use the attention-based mechanism to index them when returning an output. On the other hand, neural Turing machines (NTM, (Graves et al., 2014)) read each fact in an episode and decides whether to read, write the fact or do both to the external, differentiable memory. + +A crucial difference between these two models is that the memory network does not have a mechanism to modify the content of the external memory, while the NTM does. In practice, this leads to easier learning in the memory network, which in turn resulted in it being used more in real tasks (Bordes et al., 2015; Dodge et al., 2015). On the contrary, the NTM has mainly been tested on a series of small-scale, carefully-crafted tasks such as copy and associative recall. The NTM, however is more expressive, precisely because it can store and modify the internal state of the network as it processes an episode. + +The original NTM supports two modes of addressing (which can be used simultaneously.) They are content-based and location-based addressing. We notice that the location-based strategy is based on linear addressing. The distance between each pair of consecutive memory cells is fixed to a constant. We address this limitation, in this paper, by introducing a learnable address vector for each memory cell of the NTM with least recently used memory addressing mechanism, and we call this variant a dynamic neural Turing machine (D-NTM). + +We evaluate the proposed D-NTM on the full set of Facebook bAbI task (Weston et al., 2015b) using either continuous, differentiable attention or discrete, non-differentiable attention (Zaremba & Sutskever, 2015) as an addressing strategy. Our experiments reveal that it is possible to use the discrete, non-differentiable attention mechanism, and in fact, the D-NTM with the discrete attention and GRU controller outperforms the one with the continuous attention. After we published our paper on arXiv, a new extension of NTM called DNC (Graves et al., 2016) has also provided results on bAbI task as well. + +We also provide results on sequential-MNIST and algorithmic tasks proposed by (Graves et al., 2014) in order to investigate the ability of our model when dealing with long-term dependencies. + +# Our Contributions + +1. We propose a generalization of Neural Turing Machine called a dynamic neural Turing machine (D-NTM) which employs a learnable and location-based addressing. +2. We demonstrate the application of neural Turing machines on a more natural and less toyish task: episodic question-answering besides the toy tasks. We provide detailed analysis of our model on this task. +3. We propose to use the discrete attention mechanism and empirically show that, it can outperform the continuous attention based addressing for episodic QA task. +4. We propose a curriculum strategy for our model with the feedforward controller and discrete attention that improves our results significantly. + +# 2 DYNAMIC NEURAL TURING MACHINE + +The proposed dynamic neural Turing machine (D-NTM) extends the neural Turing machine (NTM, (Graves et al., 2014)) which has a modular design. The NTM consists of two main modules, a controller and, a memory. The controller, which is often implemented as a recurrent neural network, issues a command to the memory so as to read, write to and erase a subset of memory cells. Although the memory was originally envisioned as an integrated module, it is not necessary, and the memory may be an external, black box (Zaremba & Sutskever, 2015). + +# 2.1 CONTROLLER + +At each time step $t$ , the controller (1) receives an input value $\mathbf { x } ^ { t }$ , (2) addresses and reads the memory and creates the content vector $\phi ^ { t }$ , (3) erases/writes a portion of the memory, (4) updates its own hidden state $\mathbf { h } _ { t }$ , and (5) outputs a value $\mathbf { y } ^ { t }$ (if needed.) In this paper, we use both a gated recurrent unit (GRU, (Cho et al., 2014)) and a feedforward-controller to implement the controller such that for a GRU controller + +$$ +\mathbf { h } ^ { t } = \mathbf { G } \mathbf { R } \mathbf { U } ( \mathbf { x } ^ { t } , \mathbf { h } ^ { t - 1 } , \boldsymbol { \phi } ^ { t } ) +$$ + +or for a feedforward-controller + +$$ +\mathbf { h } ^ { t } = \sigma ( \mathbf { x } ^ { t } , \phi ^ { t } ) . +$$ + +# 2.2 MEMORY + +We use a rectangular matrix $\mathbf { M } \in \mathbb { R } ^ { N \times ( d _ { c } + d _ { a } ) }$ to denote $N$ memory cells. Unlike the original NTM, we partition each memory cell vector into two parts: + +$$ +\mathbf { M } = \left[ \mathbf { A } ; \mathbf { C } \right] . +$$ + +The first part $\mathbf { A } \in \mathbb { R } ^ { N \times d _ { a } }$ is a learnable address matrix, and the second $\mathbf { C } \in \mathbb { R } ^ { N \times d _ { c } }$ a content matrix. In other words, each memory cell $\mathbf { m } _ { i }$ is now + +$$ +\mathbf { m } _ { i } = \left[ \mathbf { a } _ { i } ; \mathbf { c } _ { i } \right] . +$$ + +The address part ${ \bf a } _ { i }$ is considered a model parameter that is updated during training. During inference, the address part is not overwritten by the controller and remains constant. On the other hand, the content part $\mathbf { c } _ { i }$ is both read and written by the controller both during training and inference. At the beginning of each episode, the content part of the memory is refreshed to be an all-zero matrix, $\mathbf { C } ^ { 0 } = \mathbf { 0 }$ . This introduction of the learnable address portion for each memory cell allows the model to learn sophisticated location-based addressing strategies. A similar addressing mechanism is also explored in (Reed & de Freitas, 2015) in the context of learning program traces. + +# 2.3 MEMORY ADDRESSING + +Memory addressing in the D-NTM is equivalent to computing an $N$ -dimensional address vector. The DNTM computes three such vectors for respectively reading $\mathbf { w } ^ { t } \in \mathbb { R } ^ { N }$ , erasing $\mathbf { e } ^ { t } \in \mathbb { R } ^ { d _ { c } }$ and writing $\mathbf { u } ^ { t } \in$ $\mathbb { R } ^ { N }$ . Specifically for writing, the controller further computes a candidate memory content vector $\bar { \mathbf { c } } ^ { t } \in$ + +![](images/e74c8fe56f94d776f121ec40a19574de3775bfc3a8c2b406e74e9fa626fad5ca.jpg) +Figure 1: A graphical illustration of the proposed dynamic neural Turing machine with the recurrent-controller. The controller receives the fact as a continuous vector encoded by a recurrent neural network, computes the read and write weights for addressing the memory. If the D-NTM automatically detects that a query has been received, it returns an answer and terminates. + +$\mathbb { R } ^ { d _ { c } }$ based on its current hidden state of the controller $\mathbf { h } ^ { t } \in \mathbb { R } ^ { d _ { h } }$ and the input of the controller scaled with a scalar gate $\alpha ^ { t }$ which is a function of the hidden state and the input of the controller as well, see Eqn 4. + +$$ +\begin{array} { l } { { \boldsymbol { \alpha } } ^ { t } = \mathbf { f } ( \mathbf { h } ^ { t } , \mathbf { x } ^ { t } ) , } \\ { { \bar { \mathbf { c } } } ^ { t } = \mathrm { R e L U } ( \mathbf { W } _ { m } \mathbf { h } ^ { t } + { \boldsymbol { \alpha } } ^ { t } \mathbf { W } _ { x } \mathbf { x } ^ { t } + \mathbf { b _ { m } } ) . } \end{array} +$$ + +Reading With the read vector $\mathbf { w } ^ { t }$ , the content vector read from the memory $\phi ^ { t } \in \mathbb { R } ^ { d _ { a } + d _ { c } }$ is retrieved by + +$$ +\begin{array} { r } { \phi ^ { t } = ( \mathbf { w } ^ { t } ) ^ { \top } \mathbf { M } ^ { t - 1 } , } \end{array} +$$ + +where $\mathbf { w } ^ { t }$ is a row vector. + +Erasing and Writing Given the erase, write and candidate memory content vectors $( \mathbf { e } ^ { t } , u _ { j } ^ { t }$ , and $\bar { \mathbf { c } } ^ { t }$ respectively) generated by a simple MLP conditioned on the hidden state of the controller $\mathbf { h } ^ { t }$ , the memory matrix is updated by, + +$$ +\mathbf { C } ^ { t } [ j ] = ( 1 - \mathbf { e } ^ { t } u _ { j } ^ { t } ) \odot \mathbf { C } ^ { t - 1 } [ j ] + u _ { j } ^ { t } \bar { \mathbf { c } } ^ { t } . +$$ + +where the subscript $j$ in $\mathbf { C } ^ { t } [ j ]$ denotes the $j$ -th row of the content part $\mathbf { C } ^ { t }$ of the memory matrix $\mathbf { M } ^ { t }$ . + +No Operation (NOP) As found in (Joulin & Mikolov, 2015), an additional NOP action might be beneficial for the controller not to access the memory once in a while. We model this situation by designating one memory cell as a NOP cell. Reading or writing from this memory cell is ignored. + +# 2.4 LEARNING + +Once the proposed D-NTM is executed, it returns the output distribution $p ( \mathbf { y } | \mathbf { x } _ { 1 } , \ldots , \mathbf { x } _ { T } )$ . As a result, we define a cost function as the negative log-likelihood: + +$$ +C ( \boldsymbol { \theta } ) = \frac { 1 } { N } \sum _ { n = 1 } ^ { N } - \log p ( \mathbf { y } ^ { n } | \mathbf { x } _ { 1 } ^ { n } , \ldots , \mathbf { x } _ { T } ^ { n } ) , +$$ + +where $\theta$ is a set of all the parameters. As the proposed D-NTM, just like the original NTM, is fully end-to-end differentiable, we can compute the gradient of this cost function by using backpropagation and learn the parameters of the model with a gradient-based optimization algorithm, such as stochastic gradient descent, to train it end-to-end. + +# 3 ADDRESSING MECHANISM + +# 3.1 ADDRESS VECTORS + +Each of the address vectors (both read and write) is computed in the same way. The way they are computed are very similar to the content based addressing in (Graves et al., 2014). First, the controller computes a key vector: + +$$ +\mathbf { k } ^ { t } = \mathbf { W } _ { k } ^ { \top } \mathbf { h } ^ { t } + \mathbf { b } _ { k } , +$$ + +where $\mathbf { W } _ { k } \ \in \ \mathbb { R } ^ { N \times ( d _ { a } + d _ { c } ) }$ and $\mathbf { b } _ { k } \ \in \ \mathbb { R } ^ { d _ { a } + d _ { c } }$ if the read head is being computed, otherwise $\mathbf { W } _ { k } \in \mathbb { R } ^ { N \times d _ { c } }$ and $\mathbf { b } _ { k } \in \mathbb { R } ^ { d _ { c } }$ if the write head weights are being computed. They can be the parameters for a specific head (either read or write.) Also, the sharpening factor $\beta _ { t } \in \mathbb { R } ^ { \ge 1 }$ is computed as: + +$$ +\begin{array} { c } { { \mathrm { s o f t p l u s } ( x ) = \log ( \exp ( x ) + 1 ) } } \\ { { \beta _ { t } = \mathrm { s o f t p l u s } ( \mathbf { u } _ { \beta } ^ { \top } \mathbf { h } ^ { t } + b _ { \beta } ) + 1 . } } \end{array} +$$ + +$\mathbf { u } _ { \beta }$ and $b _ { \beta }$ are the parameters of the sharpening $\beta _ { t }$ . + +The address vector is then computed by, + +$$ +\begin{array} { c } { { z _ { i } ^ { t } = \beta ^ { t } S \left( \mathbf { k } ^ { t } , \mathbf { m } _ { i } ^ { t } \right) } } \\ { { w _ { i } ^ { t } = \displaystyle \frac { \exp ( z _ { i } ^ { t } ) } { \sum _ { j } \exp ( z _ { j } ^ { t } ) } , } } \end{array} +$$ + +where the similarity function $S \in \mathbb { R } ^ { \geq 0 }$ is defined as + +$$ +S ( \mathbf { x } , \mathbf { y } ) = \frac { \mathbf { x } \cdot \mathbf { y } } { ( | | \mathbf { x } | | | \mathbf { y } | | + \epsilon ) } . +$$ + +# 3.2 MULTI-STEP ADDRESSING + +At each time-step, controller may require more than one-step for accessing to the memory. The original NTM addresses this by implementing multiple sets of read, erase and write heads. In this paper, we explore an option of allowing each head to operate more than once at each time step, similar to the multi-hop mechanism from the end-to-end memory network (Sukhbaatar et al., 2015). + +# 3.3 DYNAMIC LEAST RECENTLY USED ADDRESSING + +We introduce a memory addressing schema that can learn to put more emphasis on the least recently used (LRU) memory locations. As observed in (Santoro et al., 2016; Rae et al., 2016), we find it easier to learn the write operations with the use of LRU addressing. + +To learn a LRU based addressing, first we compute the exponentially moving averages of the logits $\left( \mathbf { z } _ { t } \right)$ as $\mathbf { v } _ { t }$ , $\mathbf { v } _ { t } ~ = ~ 0 . 1 \mathbf { v } _ { t - 1 } + 0 . 9 \mathbf { z } _ { t }$ . We rescale the accumulated $\mathbf { v } _ { t }$ with $\gamma _ { t }$ , such that the controller adjusts the influence of how much previously written memory locations should effect the attention weights of a particular time-step. Next, we subtract $\mathbf { v } _ { t }$ from $\mathbf { z } _ { t }$ in order to reduce the weights of previously read or written memory locations. $\gamma _ { t }$ is a shallow MLP with a scalar output and it is conditioned on the hidden state of the controller. $\gamma _ { t }$ is parametrized with the parameters $\mathbf { u } _ { \gamma }$ and ${ \bf { b } } _ { \gamma }$ , + +$$ +\begin{array} { r } { \gamma _ { t } = \operatorname { s i g m o i d } ( \mathbf { u } _ { \gamma } ^ { \top } \mathbf { h } _ { t } + \mathbf { b } _ { \gamma } ) , } \\ { \mathbf { w } _ { t } = \operatorname { s o f t m a x } \bigl ( \mathbf { z } _ { t } - \gamma _ { t } \mathbf { v } _ { t - 1 } \bigr ) . } \end{array} +$$ + +This addressing method increases the weights of the least recently used rows of the memory. The magnitude of the influence of the least-recently used memory locations is being learned and adjusted with $\gamma _ { t }$ . Our LRU addressing is dynamic due to the model’s ability to switch between pure content-based addressing and LRU. During the training, we do not backpropagate through $\mathbf { v } _ { t }$ . Due to the dynamic nature of this addressing mechanism, it can be used for both read and write operations. If needed, the model will automatically learn to disable LRU while reading from the memory. + +# 4 GENERATING DISCRETE ADDRESS VECTORS + +In this section, we describe the discrete attention based addressing strategy. + +Discrete Addressing Let us use w to denote an address vector (either read, write or erase) at time $t$ . By definition in Eq. (10), every element in this address vector is positive and sums up to one. In other words, we can treat this vector as the probabilities of a categorical distribution $\mathcal C ( \mathbf w )$ with $\dim ( \mathbf { w } )$ choices: + +$$ +p ( j ) = w _ { j } , +$$ + +where $w _ { j }$ is the $j$ -th element of w. We can readily sample from this categorical distribution and form an one-hot vector w˜ such that + +$$ +\tilde { w } _ { k } = I ( k = j ) , +$$ + +where $j \sim \mathcal { C } ( \mathbf { w } )$ , and $I$ is an indicator function. + +Training We use this sampling-based strategy for all the heads during training. This clearly makes the use of backpropagation infeasible to compute the gradient, as the sampling procedure is not differentiable. Thus, we use REINFORCE (Williams, 1992) together with the three variance reduction techniques–global baseline, input-dependent baseline and variance normalization– suggested in (Mnih & Gregor, 2014). + +Let us define $R ( { \bf x } ) = \log p ( { \bf y } | { \bf x } _ { 1 } , \ldots , { \bf x } _ { T } )$ as a reward. We first center and re-scale the reward by + +$$ +\tilde { R } ( { \bf x } ) = \frac { R ( { \bf x } ) - b } { \sqrt { \sigma ^ { 2 } + \epsilon } } , +$$ + +where $b$ and $\sigma$ is running average and standard deviation of $R$ . We can further center it for each input $\mathbf { x }$ separately, i.e., + +$$ +\tilde { R } ( \mathbf { x } ) \gets \tilde { R } ( \mathbf { x } ) - b ( \mathbf { x } ) , +$$ + +where $b ( \mathbf x )$ is computed by a baseline network which takes as input $\mathbf { x }$ and predicts its estimated reward. The baseline network is trained to minimize the Huber loss (Huber, 1964) between the true reward $\tilde { R } ( { \bf x } ) ^ { * }$ and the predicted reward $b ( \mathbf { x } )$ . We use the Huber loss, which is defined by + +$$ +H _ { \delta } ( x ) = { \left\{ \begin{array} { l l } { x ^ { 2 } } & { { \mathrm { f o r ~ } } | x | \leq \delta , } \\ { \delta ( 2 | x | - \delta ) , } & { { \mathrm { o t h e r w i s e } } , } \end{array} \right. } +$$ + +due to its robustness. As a further measure to reduce the variance, we regularize the negative entropy of all those category distributions to facilitate a better exploration during training (Xu et al., 2015). + +Then, the cost function for each training example is approximated as + +$$ +\begin{array} { r l } & { C ^ { n } ( \boldsymbol { \theta } ) = - \log p ( \mathbf { y } | \mathbf { x } _ { 1 : T } , \tilde { w } _ { 1 : J } , \tilde { u } _ { 1 : J } , \tilde { e } _ { 1 : J } ) } \\ & { \qquad - \displaystyle \sum _ { j = 1 } ^ { J } \tilde { R } ( \mathbf { x } ^ { n } ) ( \log p ( \tilde { w } _ { j } | \mathbf { x } _ { 1 : T } ) + \log p ( \tilde { u } _ { j } | \mathbf { x } _ { 1 : T } ) + \log p ( \tilde { e } _ { j } | \mathbf { x } _ { 1 : T } ) ) } \\ & { \qquad - \lambda _ { H } \displaystyle \sum _ { j = 1 } ^ { J } ( \mathcal { H } ( w _ { j } | \mathbf { x } _ { 1 : T } ) + \mathcal { H } ( u _ { j } | \mathbf { x } _ { 1 : T } ) + \mathcal { H } ( e _ { j } | \mathbf { x } _ { 1 : T } ) ) . } \end{array} +$$ + +where $J$ is the number of addressing steps, $\lambda _ { H }$ is the entropy regularization coefficient, and $\mathcal { H }$ denotes the entropy. + +Inference Once training is over, we switch to a deterministic strategy. We simply choose an element of w with the largest value to be the index of the target memory cell, such that + +$$ +\tilde { w } _ { k } = \mathbf { I } ( k = \operatorname { a r g m a x } ( \mathbf { w } ) ) . +$$ + +Curriculum Learning for the Discrete Attention Training discrete attention with feed-forward controller and REINFORCE is challenging. We propose to use a curriculum strategy for training with the discrete attention in order to tackle this problem. For each minibatch, we sample $\pi$ from a binomial distribution with the probability $p ^ { t }$ , $\pi ^ { t } \sim \mathrm { B i n } ( p ^ { t } )$ . The model will either use the discrete or the continuous-attention based on the $\pi ^ { t }$ . We start the training procedure with $p ^ { 0 } = 1$ and during the training $p ^ { t }$ is annealed to 0 by setting $\begin{array} { r } { p ^ { t } = \frac { p ^ { 0 } } { \sqrt { 1 + t } } } \end{array}$ . + +We can rewrite the weights $\mathbf { w } _ { t }$ as in Equation 14, where it is expressed as the combination of continuous attention weights $\hat { \mathbf { w } } ^ { t }$ and discrete attention weights $\tilde { \mathbf { w } } ^ { t }$ with $\mathit { \Pi } _ { \pi ^ { t } }$ being a binary variable that chooses to use one of them, + +$$ +\mathbf { w } ^ { t } \pi ^ { t } \bar { \mathbf { w } } ^ { t } + ( 1 - \pi ^ { t } ) \tilde { \mathbf { w } } ^ { t } . +$$ + +By using this curriculum learning strategy, at the beginning of the training, the model learns to use the memory mainly with the continuous attention. As we anneal the $p ^ { t }$ , the model will rely more on the discrete attention. + +# 5 REGULARIZING DYNAMIC NEURAL TURING MACHINES + +When the controller of D-NTM is a powerful recurrent neural network, it is important to regularize training of the D-NTM so as to avoid suboptimal solutions in which the D-NTM ignores the memory and works as a simple recurrent neural network. + +Read-Write Consistency Regularizer One such suboptimal solution we have observed in our preliminary experiments with the proposed D-NTM is that the D-NTM uses the address part A of the memory matrix simply as an additional weight matrix, rather than as a means to accessing the content part C. We found that this pathological case can be effectively avoided by encouraging the read head to point to a memory cell which has also been pointed by the write head. This can be implemented as the following regularization term: + +$$ +R _ { \mathrm { r w } } ( \mathbf { w } , \mathbf { u } ) = \lambda \sum _ { t ^ { \prime } = 1 } ^ { T } | | 1 - ( \frac { 1 } { t ^ { \prime } } \sum _ { t = 1 } ^ { t ^ { \prime } } \mathbf { u } _ { t } ) ^ { \top } \mathbf { w } _ { t ^ { \prime } } | | _ { 2 } ^ { 2 } +$$ + +In the equations above, $\mathbf { u } _ { t }$ is the write and $\mathbf { w } _ { t }$ is the read weights. + +Next Input Prediction as Regularization Temporal structure is a strong signal that should be exploited by the controller based on a recurrent neural network. We exploit this structure by letting the controller predict the input in the future. We maximize the predictability of the next input by the controller during training. This is equivalent to minimizing the following regularizer: + +$$ +R _ { \mathrm { p r e d } } ( \mathbf { W } ) = - \log p ( \mathbf { f } _ { t + 1 } | \mathbf { f } _ { t } , \mathbf { w } _ { t } , \mathbf { u } _ { t } , \mathbf { M } _ { t } ; \mathbf { W } ) ) +$$ + +where $f _ { t }$ is the current input and $f _ { t + 1 }$ is the input at next timestep. We found this regularizer to be effective in our preliminary experiments and use it for bAbI tasks. + +# 6 RELATED WORK + +A recurrent neural network (RNN), which is used as a controller in the proposed D-NTM, has an implicit memory in the form of recurring hidden states. Even with this implicit memory, a vanilla RNN is however known to have difficulties in storing information for long time-spans (Bengio et al., 1994; Hochreiter, 1991). Long short-term memory (LSTM, (Hochreiter & Schmidhuber, 1997)) and gated recurrent units (GRU, (Cho et al., 2014)) have been found to address this issue. However all these models based solely on RNNs have been found to be limited when they are used to solve, e.g., algorithmic tasks and episodic question-answering. + +In addition to the finite random access memory of the neural Turing machine, based on which the D-NTM is designed, other data structures have been proposed as external memory for neural networks. In (Sun et al., 1997; Grefenstette et al., 2015; Joulin & Mikolov, 2015), a continuous, differentiable stack was proposed. In (Zaremba et al., 2015; Zaremba & Sutskever, 2015), grid and tape storages are used. These approaches differ from the NTM in that their memory is unbounded and can grow indefinitely. On the other hand, they are often not randomly accessible. + +Memory networks (Weston et al., 2015b) form another family of neural networks with external memory. In this class of neural networks, information is stored explicitly as it is (in the form of its continuous representation) in the memory, without being erased or modified during an episode. Memory networks and their variants have been applied to various tasks successfully (Sukhbaatar et al., 2015; Bordes et al., 2015; Dodge et al., 2015; Xiong et al., 2016). Miller et al. (2016) have also independently proposed the idea of having separate key and value vectors for memory networks. + +Another related family of models is the attention-based neural networks. Neural networks with continuous or discrete attention over an input have shown promising results on a variety of challenging tasks, including machine translation (Bahdanau et al., 2015; Luong et al., 2015), speech recognition (Chorowski et al., 2015), machine reading comprehension (Hermann et al., 2015) and image caption generation (Xu et al., 2015). + +The latter two, the memory network and attention-based networks, are however clearly distinguishable from the D-NTM by the fact that they do not modify the content of the memory. + +# 7 EXPERIMENTS + +We provide experimental results to demonstrate the abilities of our model, first on Facebook bAbI task (Weston et al., 2015a). We give detailed analysis and experimental results on this task. We also compare different variations of NTM on bAbI tasks. We have performed experiments on sequential permuted MNIST (Le et al., 2015) and on toy tasks to compare other published models on these tasks with a recurrent controller. The details of our experiments are provided in the supplementary material. + +# 7.1 EPISODIC QUESTION-ANSWERING: BABI TASKS + +In this section, we evaluate the proposed D-NTM on the recently proposed episodic question-answering task called Facebook bAbI. We use the dataset with 10k training examples per sub-task provided by Facebook.1 For each episode, the D-NTM reads a sequence of factual sentences followed by a question, all of which are given as natural language sentences. The D-NTM is expected to store and retrieve relevant information in the memory in order to answer the question based on the presented facts. Exact implementation details and hyper-parameter settings are provided in the appendix. + +# 7.1.1 GOALS + +The goal of this experiment is three-fold. First, we present for the first time the performance of a memory-based network that can both read and write dynamically on the Facebook bAbI tasks2. We aim to understand whether a model that has to learn to write an incoming fact to the memory, rather than storing it as it is, is able to work well, and to do so, we compare both the original NTM and proposed D-NTM against an LSTM-RNN. + +Second, we investigate the effect of having to learn how to write. The fact that the NTM needs to learn to write likely has adverse effect on the overall performance, when compared to, for instance, end-to-end memory networks (MemN2N, (Sukhbaatar et al., 2015)) and dynamic memory network $\mathrm { ( D M N + }$ , (Xiong et al., 2016)) both of which simply store the incoming facts as they are. We quantify this effect in this experiment. Lastly, we show the effect of the proposed learnable addressing scheme. + +We further explore the effect of using a feedforward controller instead of the GRU controller. In addition to the explicit memory, the GRU controller can use its own internal hidden state as the memory. On the other hand, the feedforward controller must solely rely on the explicit memory, as it is the only memory available. + +# 7.1.2 RESULTS AND ANALYSIS + +In Table 1, we first observe that the NTMs are indeed capable of solving this type of episodic question-answering better than the vanilla LSTM-RNN. Although the availability of explicit memory in the NTM has already suggested this result, we note that this is the first time neural Turing machines have been used in this specific task. + +All the variants of NTM with the GRU controller outperform the vanilla LSTM-RNN. However, not all of them perform equally well. First, it is clear that the proposed dynamic NTM (D-NTM) using the GRU controller outperforms the original NTM with the GRU controller (NTM, CBA only NTM vs. continuous D-NTM, Discrete D-NTM). As discussed earlier, the learnable addressing scheme of the D-NTM allows the controller to access the memory slots by location in a potentially nonlinear way. We expect it to help with tasks that have non-trivial access patterns, and as anticipated, we see a large gain with the D-NTM over the original NTM in the tasks of, for instance, 12 - Conjunction and 17 - Positional Reasoning. + +Among the recurrent variants of the proposed D-NTM, we notice significant improvements by using discrete addressing over using continuous addressing. We conjecture that this is due to certain types of tasks that require precise/sharp retrieval of a stored fact, in which case continuous addressing is in disadvantage over discrete addressing. This is evident from the observation that the D-NTM with discrete addressing significantly outperforms that with continuous addressing in the tasks of 8 - + +
DMN+L1-step CBA1-step Soft D-NTM1-step Discrete D-NTM3-steps LBA* NTM3-steps CBA NTM3-steps Soft D-NTM3-steps Discrete D-NTM
Task 1LSTM 0.00MemN2N 0.000.00NTM 16.30NTM 16.885.416.660.000.000.000.00
281.900.300.3057.0855.7058.5456.0461.6759.3846.6662.29
383.102.101.1074.1655.0074.5872.0883.5465.2147.0841.45
40.200.000.000.000.000.000.000.000.000.000.00
51.200.800.501.4620.411.661.040.831.461.251.45
651.800.100.0023.3321.0440.2044.7948.1354.8020.6211.04
724.902.002.4021.6721.6719.1619.587.9237.707.295.62
834.100.900.0025.7621.0512.5818.4625.388.8211.020.74
920.200.300.0024.7924.1736.6634.3737.800.0039.3732.50
1030.100.000.0041.4633.1352.2950.8356.2523.7520.0020.83
1110.300.100.0018.9631.8831.454.163.960.2830.6216.87
1223.400.000.0025.8330.007.706.6628.7523.755.414.58
136.100.000.006.675.635.622.295.8383.137.915.00
1481.000.100.2058.5459.1760.0063.7561.8857.7158.1260.20
1578.700.000.0036.4642.3036.8739.2735.6221.8836.0440.26
1651.9051.8045.3071.1571.1549.1651.3546.1550.0046.0445.41
1750.1018.604.2043.7543.7517.9116.0443.7556.2521.259.16
186.805.302.103.9647.503.953.5447.5047.506.871.66
1990.302.300.0075.8971.5173.7464.6361.5663.6575.8876.66
202.100.000.001.250.002.703.120.400.003.330.00
Avg.Err.36.414.242.8131.4233.6029.5127.9332.8532.7624.2421.79
+ +Table 1: Test error rates $( \% )$ on the 20 bAbI QA tasks for models using 10k training examples with the GRU and feedforward controller. FF stands for the experiments that are conducted with feedforward controller. Let us, note that $\mathrm { L B A ^ { * } }$ refers to NTM that uses both LBA and CBA. In this table, we compare multi-step vs single-step addressing, original NTM with location based+content based addressing vs only content based addressing, and discrete vs continuous addressing on bAbI. + +Lists/Sets and 11 - Basic Coreference. Furthermore, this is in line with an earlier observation in (Xu et al., 2015), where discrete addressing was found to generalize better in the task of image caption generation. + +In Table 2, we also observe that the D-NTM with the feedforward controller and discrete attention performs worse than LSTM and D-NTM with continuous-attention. However, when the proposed curriculum strategy from Sec. 4 is used, the average test error drops from 68.30 to 37.79. + +We empirically found training of the feedforward controller more difficult than that of the recurrent controller. We train our feedforward controller based models four times longer (in terms of the number of updates) than the recurrent controller based ones in order to ensure that they are converged for most of the tasks. On the other hand, the models trained with the GRU controller overfit on bAbI tasks very quickly. For example, on tasks 3 and 16 the feedforward controller based model underfits (i.e., high training loss) at the end of the training, whereas with the same number of units the model with the GRU controller can overfit on those tasks after 3,000 updates only. + +When our results are compared to the variants of the memory network Weston et al. (2015b) (MemN2N and $\mathrm { D M N + }$ ), we notice a significant performance gap. We attribute this gap to the difficulty in learning to manipulate and store a complex input. + +Table 2: Test error rates $( \% )$ on the 20 bAbI QA tasks for models using 10k training examples with feedforward controller. + +
TaskFF Soft D-NTMFF Discrete D-NTMFF Discrete* D-NTM
14.3881.6714.79
227.576.6776.67
371.2579.3870.83
40.0078.6544.06
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.62
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.92
1939.4771.4671.56
200.076.569.79
Avg.Err.12.8168.3037.79
+ +We also provide further experiments investigating different extensions on D-NTM in the appendix. + +# 7.2 SEQUENTIAL pMNIST + +In sequential MNIST task, the pixels of the MNIST digits are provided to the model in scan line order, left to right and top to bottom (Le et al., 2015). At the end of sequence of pixels, the model predicts the label of the digit in the sequence of pixels. We experiment D-NTM on the variation of sequential MNIST where the order of the pixels is randomly shuffled, we call this task as permuted MNIST $( p \mathrm { M N I S T } )$ . An important contribution of this task to our paper, in particular, is to measure the model’s ability to perform well when dealing with long-term dependencies. We report our results in Table $3 ^ { 3 }$ , we observe improvements over other models that we compare against. In Table 3, ”discrete addressing with MAB” refers to D-NTM model using REINFORCE with baseline computed from moving averages of the reward. Discrete addressing with IB refers to D-NTM using REINFORCE with input-based baseline. + +# 7.3 NTM TOY TASKS + +We explore the possibility of using D-NTM to solve algorithmic tasks such as copy and associative recall tasks. We train our model on the same lengths of sequences that is experimented in (Graves et al., 2014). We report our results in Table 4. We find out that D-NTM using continuous-attention can successfully learn the ”Copy” and ”Associative Recall” tasks. + +In Table 4, we train our model on sequences of the same length as the experiments in (Graves et al., 2014) and test the model on the sequences of the maximum length seen during the training. We consider model to be successful on copy or associative recall if its validation cost (binary cross-entropy) is lower than 0.02 over the sequences of maximum length seen during the training. We set the threshold to 0.02 to determine whether a model is successful on a task. Because empirically we observe that the models have higher validation costs perform badly in terms of generalization over the longer sequences. ”D-NTM discrete” model in this table is trained with REINFORCE using moving averages to estimate the baseline. + +Table 3: Sequential $p { \mathrm { M N I S T } } .$ . + +
Test Acc
D-NTMdiscrete MAB89.6
D-NTMdiscrete IB92.3
Soft D-NTM93.4
NTM90.9
I-RNN (Le et al., 2015)82.0
Zoneout (Krueger et al.,2016)93.1
LSTM (Krueger et al., 2016)89.8
Unitary-RNN (Arjovsky et al., 2015)91.4
Recurrent Dropout (Krueger et al., 2016)92.5
+ +Table 4: NTM Toy Tasks. + +
Copy TasksAssociative Recall
Soft D-NTMSuccessSuccess
D-NTMdiscreteSuccessFailure
NTMSuccessSuccess
+ +# 8 CONCLUSION AND FUTURE WORK + +In this paper we extend neural Turing machines (NTM) by introducing a learnable addressing scheme which allows the NTM to be capable of performing highly nonlinear location-based addressing. This extension, to which we refer by dynamic NTM (D-NTM), is extensively tested with various configurations, including different addressing mechanisms (continuous vs. discrete) and different number of addressing steps, on the Facebook bAbI tasks. This is the first time an NTM-type model was tested on this task, and we observe that the NTM, especially the proposed D-NTM, performs better than vanilla LSTM-RNN. Furthermore, the experiments revealed that the discrete, discrete addressing works better than the continuous addressing with the GRU controller, and our analysis reveals that this is the case when the task requires precise retrieval of memory content. + +Our experiments show that the NTM-based models can be weaker than other variants of memory networks which do not learn but have an explicit mechanism of storing incoming facts as they are. We conjecture that this is due to the difficulty in learning how to write, manipulate and delete the content of memory. Despite this difficulty, we find the NTM-based approach, such as the proposed D-NTM, to be a better, future-proof approach, because it can scale to a much longer horizon (where it becomes impossible to explicitly store all the experiences.) + +On $p \mathrm { M N I S T }$ task, we show that our model can outperform other similar type of approaches proposed to deal with the long-term dependencies. On copy and associative recall tasks, we show that our model can solve the algorithmic problems that are proposed to solve with NTM type of models. + +The success of both the learnable address and the discrete addressing scheme suggests two future research directions. First, we should try both of these schemes in a wider array of memory-based models, as they are not specific to the neural Turing machines. Second, the proposed D-NTM needs to be evaluated on a diverse set of applications, such as text summarization (Rush et al., 2015), visual questionanswering (Antol et al., 2015) and machine translation, in order to make a more concrete conclusion. + +# REFERENCES + +Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pp. 2425–2433, 2015. +Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. arXiv preprint arXiv:1511.06464, 2015. +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In Proceedings Of The International Conference on Representation Learning (ICLR 2015), 2015. +Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994. +Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question answering with memory networks. arXiv preprint arXiv:1506.02075, 2015. +Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. +Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. Attention-based models for speech recognition. arXiv preprint arXiv:1506.07503, 2015. +Tim Cooijmans, Nicolas Ballas, Cesar Laurent, and Aaron Courville. Recurrent batch normalization. ´ arXiv preprint arXiv:1603.09025, 2016. +Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur Szlam, and Jason Weston. Evaluating prerequisite qualities for learning end-to-end dialog systems. CoRR, abs/1511.06931, 2015. +Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. Book in preparation for MIT Press, 2016. URL http://www.deeplearningbook.org. +Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. +Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. ´ Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626): 471–476, 2016. +Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp. 1819–1827, 2015. +Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions. arXiv preprint arXiv:1603.00391, 2016. +Karl Moritz Hermann, Toma´s Ko ˇ cisk ˇ y, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa \` Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. arXiv preprint arXiv:1506.03340, 2015. + +Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s books with explicit memory representations. arXiv preprint arXiv:1511.02301, 2015. + +Sepp Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Diploma, Technische Universitat¨ Munchen ¨ , pp. 91, 1991. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural Computation, 9(8): 1735–1780, 1997. + +Peter J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73–101, 03 1964. + +Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In Advances in Neural Information Processing Systems, pp. 190–198, 2015. + +Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. + +David Krueger, Tegan Maharaj, Janos Kram ´ ar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary ´ Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016. + +Quoc V Le, Navdeep Jaitly, and Geoffrey E Hinton. A simple way to initialize recurrent networks of rectified linear units. arXiv preprint arXiv:1504.00941, 2015. + +Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In Proceedings Of The Conference on Empirical Methods for Natural Language Processing (EMNLP 2015), 2015. + +Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. Key-value memory networks for directly reading documents. CoRR, abs/1606.03126, 2016. URL http://arxiv.org/abs/1606.03126. + +Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. arXiv preprint arXiv:1402.0030, 2014. + +Jack W Rae, Jonathan J Hunt, Tim Harley, Ivo Danihelka, Andrew Senior, Greg Wayne, Alex Graves, and Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes. In Advances in NIPS. 2016. + +Scott Reed and Nando de Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. + +Alexander M. Rush, Sumit Chopra, and Jason Weston. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pp. 379–389, 2015. + +Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-shot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016. + +Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. Building end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI-16), 2016. + +Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. arXiv preprint arXiv:1503.08895, 2015. + +Guo-Zheng Sun, C. Lee Giles, and Hsing-Hen Chen. The neural network pushdown automaton: Architecture, dynamics and training. In Adaptive Processing of Sequences and Data Structures, International Summer School on Neural Networks, pp. 296–345, 1997. + +Oriol Vinyals and Quoc Le. A neural conversational model. arXiv preprint arXiv:1506.05869, 2015. + +Jason Weston, Antoine Bordes, Sumit Chopra, and Tomas Mikolov. Towards ai-complete question answering: a set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015a. + +Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Proceedings Of The International Conference on Representation Learning (ICLR 2015), 2015b. In Press. + +Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8:229–256, 1992. +Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. CoRR, abs/1603.01417, 2016. +Kelvin Xu, Jimmy Ba, Ryan Kiros, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings Of The International Conference on Representation Learning (ICLR 2015), 2015. +Li Yao, Atousa Torabi, Kyunghyun Cho, Nicolas Ballas, Christopher Pal, Hugo Larochelle, and Aaron Courville. Describing videos by exploiting temporal structure. In Computer Vision (ICCV), 2015 IEEE International Conference on. IEEE, 2015. +Wojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines. CoRR, abs/1505.00521, 2015. +Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms from examples. arXiv preprint arXiv:1511.07275, 2015. + +# A EXPERIMENTAL DETAILS + +A.1 MODEL AND TRAINING DETAILS FOR BABI + +We use the same hyperparameters for all the tasks for a given model. + +# A.1.1 FACT REPRESENTATION + +We use a recurrent neural network with GRU units to encode a variable-length fact into a fixed-size vector representation. This allows the D-NTM to exploit the word ordering in each fact, unlike when facts are encoded as bag-of-words vectors. + +# A.1.2 CONTROLLER + +We experiment with both a recurrent and feedforward neural network as the controller that generates the read and write weights. The controller has 180 units. We train our feed-forward controller using noisy-tanh activation function (Gulcehre et al., 2016) since we were experiencing training difficulties with sigmoid and tanh activation functions. We use both single-step and three-steps addressing with our GRU controller. + +# A.1.3 MEMORY + +The memory contains 120 memory cells. Each memory cell consists of a 16-dimensional address part and 28-dimensional content part. + +# A.1.4 TRAINING DETAILS + +We set aside a random $1 0 \%$ of the training examples as a validation set for each sub-task and use it for early-stopping and hyperparameter search. We train one D-NTM for each sub-task, using Adam (Kingma & Ba, 2014) with its learning rate set to 0.003 and 0.007 respectively for GRU and Feedforward controller. The size of each minibatch is 160, and each minibatch is constructed uniform-randomly from the training set. + +# A.2 MODEL AND TRAINING DETAILS FOR SEQUENTIAL MNIST + +On sequential MNIST task we try to keep the capacity of our model to be close to our baselines. We use 100 GRU units in the controller and each content vector of size 8 and with address vectors of size 8. We use a learning rate of $1 e - 3$ and trained the model with adam optimizer. We did not use the read and write consistency regularization in any of our models. + +# A.3 MODEL AND TRAINING DETAILS FOR TOY TASKS + +On both copy and associative recall tasks, we try to keep the capacity of our model to be close to our baselines. We use 100 GRU units in the controller and each content vector of has a size of 8 and using address vector of size 8. We use a learning rate of $1 e - 3$ and trained the model with adam optimizer. We did not use the read and write consistency regularization in any of our models. For the model with the discrete attention we use REINFORCE with baseline computed using moving averages. + +# B VISUALIZATION OF DISCRETE ATTENTION + +We visualize the attention of D-NTM with GRU controller with discrete attention in Figure 2. From this example, we can see that D-NTM has learned to find the correct supporting fact even without any supervision for the particular story in the visualization. + +# C LEARNING CURVES FOR THE RECURRENT CONTROLLER + +In Figure 3, we compare the learning curves of the continuous and discrete attention D-NTM model with recurrent controller on Task 1. Surprisingly, the discrete attention D-NTM converges faster than the continuous-attention model. The main difficulty of learning continuous-attention is due to the fact that learning to write with continuous-attention can be challenging. + +![](images/13f6be6816a8ab4ad76006098fb07e79c29ffd5f83dd209f067809e9e7421ce0.jpg) +Figure 2: An example view of the discrete attention over the memory slots for both read (left) and write heads(right). $\mathbf { X }$ -axis the denotes the memory locations that are being accessed and y-axis corresponds to the content in the particular memory location. In this figure, we visualize the discrete-attention model with 3-reading steps and on task-20. It is easy to see that the NTM with discrete-attention accesses to the relevant part of the memory. We only visualize the last-step of the 3-steps writing. Because with discrete attention usually the model just reads the empty slots of the memory. + +![](images/750ba06aa86f2fcf7b1b8477065761e178ba6d5f4dd1a6aa03704d44c0632dbd.jpg) +Figure 3: A visualization for the learning curves of continuous and discrete D-NTM models trained on Task 1 using 3 steps. In most tasks, we observe that the discrete attention model with GRU controller does converge faster than the continuous-attention model. + +# D A COMPARISON BETWEEN THE LEARNINGCURVES OF INPUT BASED BASELINE AND REGULAR BASELINE ON $p$ MNIST + +In Figure 4, we show the learning curves of input-based-baseline (ibb) and regular REINFORCE with moving averages baseline (mab) on the $p \mathrm { M N I S T }$ task. We observe that input-based-baseline in general is much easier to optimize and converges faster as well. But it can quickly overfit to the task as well. + +# E TRAINING WITH CONTINUOUS-ATTENTION AND TESTING WITH DISCRETE-ATTENTION + +In Table 5, we provide results investigating the effects of using discrete attention model at the test-time for a model trained with feed-forward controller and continuous attention. Discrete∗ D-NTM model bootstraps the discrete attention with the continuous attention, using the curriculum method that we have introduced in Section ”Curriculum Learning for the Discrete Attention”. Discrete† D-NTM model is the continuous-attention model which uses discrete-attention at the test time. We observe that the Discrete† D-NTM model which is trained with continuous-attention outperforms Discrete D-NTM model. + +![](images/858347b800db4692b6c99c985fab4630eac6e5ffb7ee526a7c12130eaa865972.jpg) +Figure 4: We compare the learning curves of our D-NTM model using discrete attention on $p \mathrm { M N I S T }$ task with input-based baseline and regular REINFORCE baseline. The $\mathbf { X }$ -axis is the loss and y-axis is the number of epochs. + +
TaskcontinuousD-NTMDiscreteD-NTMDiscrete*D-NTMDiscrete†D-NTM
14.3881.6714.7972.2881.6778.9579.6968.5431.6749.17
23427.571.2576.6779.3878.6576.6770.8344.06
0.00
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.6279.3237.7125.6382.0874.3847.0877.0873.9653.0230.4211.4676.0513.96
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.9271.569.79
192039.4771.4676.56
0.0
Avg12.8168.3037.7957.21
+ +Table 5: Test error rates $( \% )$ on the 20 bAbI QA tasks for models using 10k training examples with the feedforward controller. Discrete∗ D-NTM model bootstraps the discrete attention with the continuous attention, using the curriculum method that we have introduced in Section 4. Discrete† D-NTM model is the continuous-attention model which uses discrete-attention at the test time. + +# F D-NTM WITH BOW FACT REPRESENTATION + +In Table 6, we provide results for D-NTM using BoW with positional encoding (PE) Sukhbaatar et al. (2015) as the representation of the input facts. The facts representations are provided as an input to the GRU controller. In agreement to our results with the GRU fact representation, with the BoW fact representation we observe improvements with multi-step of addressing over single-step and discrete addressing over continuous addressing. + +Table 6: Test error rates $( \% )$ on the 20 bAbI QA tasks for models using 10k training examples with the GRU controller and representations of facts are obtained with $\mathrm { B o W }$ using positional encoding. + +
TaskSoftD-NTM(1-step)DiscreteD-NTM(1-step)SoftD-NTM(3-steps)DiscreteD-NTM(3-steps)
10.000.000.000.00
23461.0459.3756.8755.62
55.6257.562.557.5
27.2924.8926.4527.08
513.5512.0815.8314.78
613.5414.3721.8713.33
78.546.258.7514.58
81.691.363.013.02
917.716.6637.7017.08
1026.0427.0826.8723.952.294.16
1120.413.952.5
120.410.830.20
133.121.044.795.83
1462.0858.3361.2560.620.05
1531.6626.250.62
1654.4748.5448.9548.9530.6236.04
1743.7531.8743.75
1833.7539.3736.66
192064.6369.2167.2365.46
1.250.001.450.00
Avg27.0224.9826.3624.05
\ No newline at end of file diff --git a/parse/train/BkSmc8qll/BkSmc8qll_content_list.json b/parse/train/BkSmc8qll/BkSmc8qll_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5fbc68d29e00b3954144bbecdbe0820ead7e213 --- /dev/null +++ b/parse/train/BkSmc8qll/BkSmc8qll_content_list.json @@ -0,0 +1,2194 @@ +[ + { + "type": "text", + "text": "DYNAMIC NEURAL TURING MACHINE WITH CONTINUOUS AND DISCRETE ADDRESSING SCHEMES ", + "text_level": 1, + "bbox": [ + 176, + 98, + 825, + 145 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Caglar Gulcehre∗, Sarath Chandar∗, Kyunghyun Cho†, Yoshua Bengio∗ ", + "bbox": [ + 189, + 167, + 686, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "∗ University of Montreal, name.lastname@umontreal.ca † New York University, name.lastname@nyu.edu ", + "bbox": [ + 186, + 185, + 583, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 247, + 544, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we extend neural Turing machine (NTM) into a dynamic neural Turing machine (D-NTM) by introducing a trainable memory addressing scheme. This addressing scheme maintains for each memory cell two separate vectors, content and address vectors. This allows the D-NTM to learn a wide variety of location-based addressing strategies including both linear and nonlinear ones. We implement the D-NTM with both continuous, differentiable and discrete, non-differentiable read/write mechanisms. We investigate the mechanisms and effects for learning to read and write to a memory through experiments on Facebook bAbI tasks using both a feedforward and GRU-controller. The D-NTM is evaluated on a set of Facebook bAbI tasks and shown to outperform NTM and LSTM baselines. We also provide further experimental results on sequential MNIST, associative recall and copy tasks. ", + "bbox": [ + 233, + 279, + 764, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 453, + 336, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Designing general-purpose learning algorithms is one of the long-standing goals of artificial intelligence. Despite the success of deep learning in this area (see, e.g., (Goodfellow et al., 2016)) there are still a set of complex tasks that are not well addressed by conventional neural networks. Those tasks often require a neural network to be equipped with an explicit, external memory in which a larger, potentially unbounded, set of facts need to be stored. They include, but are not limited to, episodic question-answering (Weston et al., 2015b; Hermann et al., 2015; Hill et al., 2015), compact algorithms (Zaremba et al., 2015), dialogue (Serban et al., 2016; Vinyals & Le, 2015) and video caption generation (Yao et al., 2015). ", + "bbox": [ + 174, + 484, + 825, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently two promising approaches based on neural networks to this type of tasks have been proposed. Memory networks (Weston et al., 2015b) explicitly store all the facts, or information, available for each episode in an external memory (as continuous vectors) and use the attention-based mechanism to index them when returning an output. On the other hand, neural Turing machines (NTM, (Graves et al., 2014)) read each fact in an episode and decides whether to read, write the fact or do both to the external, differentiable memory. ", + "bbox": [ + 174, + 584, + 825, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A crucial difference between these two models is that the memory network does not have a mechanism to modify the content of the external memory, while the NTM does. In practice, this leads to easier learning in the memory network, which in turn resulted in it being used more in real tasks (Bordes et al., 2015; Dodge et al., 2015). On the contrary, the NTM has mainly been tested on a series of small-scale, carefully-crafted tasks such as copy and associative recall. The NTM, however is more expressive, precisely because it can store and modify the internal state of the network as it processes an episode. ", + "bbox": [ + 174, + 671, + 825, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The original NTM supports two modes of addressing (which can be used simultaneously.) They are content-based and location-based addressing. We notice that the location-based strategy is based on linear addressing. The distance between each pair of consecutive memory cells is fixed to a constant. We address this limitation, in this paper, by introducing a learnable address vector for each memory cell of the NTM with least recently used memory addressing mechanism, and we call this variant a dynamic neural Turing machine (D-NTM). ", + "bbox": [ + 174, + 757, + 825, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We evaluate the proposed D-NTM on the full set of Facebook bAbI task (Weston et al., 2015b) using either continuous, differentiable attention or discrete, non-differentiable attention (Zaremba & Sutskever, 2015) as an addressing strategy. Our experiments reveal that it is possible to use the discrete, non-differentiable attention mechanism, and in fact, the D-NTM with the discrete attention and GRU controller outperforms the one with the continuous attention. After we published our paper on arXiv, a new extension of NTM called DNC (Graves et al., 2016) has also provided results on bAbI task as well. ", + "bbox": [ + 174, + 843, + 825, + 922 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We also provide results on sequential-MNIST and algorithmic tasks proposed by (Graves et al., 2014) in order to investigate the ability of our model when dealing with long-term dependencies. ", + "bbox": [ + 169, + 103, + 825, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Contributions ", + "text_level": 1, + "bbox": [ + 174, + 145, + 307, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We propose a generalization of Neural Turing Machine called a dynamic neural Turing machine (D-NTM) which employs a learnable and location-based addressing. \n2. We demonstrate the application of neural Turing machines on a more natural and less toyish task: episodic question-answering besides the toy tasks. We provide detailed analysis of our model on this task. \n3. We propose to use the discrete attention mechanism and empirically show that, it can outperform the continuous attention based addressing for episodic QA task. \n4. We propose a curriculum strategy for our model with the feedforward controller and discrete attention that improves our results significantly. ", + "bbox": [ + 173, + 166, + 826, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 DYNAMIC NEURAL TURING MACHINE ", + "text_level": 1, + "bbox": [ + 174, + 305, + 526, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed dynamic neural Turing machine (D-NTM) extends the neural Turing machine (NTM, (Graves et al., 2014)) which has a modular design. The NTM consists of two main modules, a controller and, a memory. The controller, which is often implemented as a recurrent neural network, issues a command to the memory so as to read, write to and erase a subset of memory cells. Although the memory was originally envisioned as an integrated module, it is not necessary, and the memory may be an external, black box (Zaremba & Sutskever, 2015). ", + "bbox": [ + 174, + 337, + 825, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 CONTROLLER ", + "text_level": 1, + "bbox": [ + 174, + 433, + 310, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At each time step $t$ , the controller (1) receives an input value $\\mathbf { x } ^ { t }$ , (2) addresses and reads the memory and creates the content vector $\\phi ^ { t }$ , (3) erases/writes a portion of the memory, (4) updates its own hidden state $\\mathbf { h } _ { t }$ , and (5) outputs a value $\\mathbf { y } ^ { t }$ (if needed.) In this paper, we use both a gated recurrent unit (GRU, (Cho et al., 2014)) and a feedforward-controller to implement the controller such that for a GRU controller ", + "bbox": [ + 174, + 458, + 823, + 512 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/4cf31fbf6dbc86933cb0ad73eeb8188150e039682e5c9e701f3ad20f27d6c9af.jpg", + "text": "$$\n\\mathbf { h } ^ { t } = \\mathbf { G } \\mathbf { R } \\mathbf { U } ( \\mathbf { x } ^ { t } , \\mathbf { h } ^ { t - 1 } , \\boldsymbol { \\phi } ^ { t } )\n$$", + "text_format": "latex", + "bbox": [ + 413, + 518, + 583, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or for a feedforward-controller ", + "bbox": [ + 174, + 544, + 377, + 558 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/5de065f223a91cdf87a45c3509c4aedeb7249eabc03bdc16c64b06a19366a234.jpg", + "text": "$$\n\\mathbf { h } ^ { t } = \\sigma ( \\mathbf { x } ^ { t } , \\phi ^ { t } ) .\n$$", + "text_format": "latex", + "bbox": [ + 444, + 563, + 553, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 MEMORY ", + "text_level": 1, + "bbox": [ + 174, + 597, + 281, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use a rectangular matrix $\\mathbf { M } \\in \\mathbb { R } ^ { N \\times ( d _ { c } + d _ { a } ) }$ to denote $N$ memory cells. Unlike the original NTM, we partition each memory cell vector into two parts: ", + "bbox": [ + 176, + 619, + 823, + 648 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/befc9cc2f17dd794ef5948d1e6e6deee3276d9b7c5d860e7b70f86a8f3d13223.jpg", + "text": "$$\n\\mathbf { M } = \\left[ \\mathbf { A } ; \\mathbf { C } \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 450, + 655, + 547, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first part $\\mathbf { A } \\in \\mathbb { R } ^ { N \\times d _ { a } }$ is a learnable address matrix, and the second $\\mathbf { C } \\in \\mathbb { R } ^ { N \\times d _ { c } }$ a content matrix. In other words, each memory cell $\\mathbf { m } _ { i }$ is now ", + "bbox": [ + 169, + 679, + 825, + 708 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/21127d76e3d09eb46b3c25f154ce493fbf16c81d5f1b19c4e682ad1bb66b36b8.jpg", + "text": "$$\n\\mathbf { m } _ { i } = \\left[ \\mathbf { a } _ { i } ; \\mathbf { c } _ { i } \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 449, + 714, + 547, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The address part ${ \\bf a } _ { i }$ is considered a model parameter that is updated during training. During inference, the address part is not overwritten by the controller and remains constant. On the other hand, the content part $\\mathbf { c } _ { i }$ is both read and written by the controller both during training and inference. At the beginning of each episode, the content part of the memory is refreshed to be an all-zero matrix, $\\mathbf { C } ^ { 0 } = \\mathbf { 0 }$ . This introduction of the learnable address portion for each memory cell allows the model to learn sophisticated location-based addressing strategies. A similar addressing mechanism is also explored in (Reed & de Freitas, 2015) in the context of learning program traces. ", + "bbox": [ + 174, + 744, + 826, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3 MEMORY ADDRESSING ", + "text_level": 1, + "bbox": [ + 176, + 854, + 377, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memory addressing in the D-NTM is equivalent to computing an $N$ -dimensional address vector. The DNTM computes three such vectors for respectively reading $\\mathbf { w } ^ { t } \\in \\mathbb { R } ^ { N }$ , erasing $\\mathbf { e } ^ { t } \\in \\mathbb { R } ^ { d _ { c } }$ and writing $\\mathbf { u } ^ { t } \\in$ $\\mathbb { R } ^ { N }$ . Specifically for writing, the controller further computes a candidate memory content vector $\\bar { \\mathbf { c } } ^ { t } \\in$ ", + "bbox": [ + 176, + 880, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e74c8fe56f94d776f121ec40a19574de3775bfc3a8c2b406e74e9fa626fad5ca.jpg", + "image_caption": [ + "Figure 1: A graphical illustration of the proposed dynamic neural Turing machine with the recurrent-controller. The controller receives the fact as a continuous vector encoded by a recurrent neural network, computes the read and write weights for addressing the memory. If the D-NTM automatically detects that a query has been received, it returns an answer and terminates. " + ], + "image_footnote": [], + "bbox": [ + 186, + 108, + 815, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbb { R } ^ { d _ { c } }$ based on its current hidden state of the controller $\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d _ { h } }$ and the input of the controller scaled with a scalar gate $\\alpha ^ { t }$ which is a function of the hidden state and the input of the controller as well, see Eqn 4. ", + "bbox": [ + 171, + 428, + 823, + 458 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/4cbcbe6a8a47923e844608152db23b8e222b75dd7af8100931817e3aa59284c0.jpg", + "text": "$$\n\\begin{array} { l } { { \\boldsymbol { \\alpha } } ^ { t } = \\mathbf { f } ( \\mathbf { h } ^ { t } , \\mathbf { x } ^ { t } ) , } \\\\ { { \\bar { \\mathbf { c } } } ^ { t } = \\mathrm { R e L U } ( \\mathbf { W } _ { m } \\mathbf { h } ^ { t } + { \\boldsymbol { \\alpha } } ^ { t } \\mathbf { W } _ { x } \\mathbf { x } ^ { t } + \\mathbf { b _ { m } } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 362, + 462, + 633, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reading With the read vector $\\mathbf { w } ^ { t }$ , the content vector read from the memory $\\phi ^ { t } \\in \\mathbb { R } ^ { d _ { a } + d _ { c } }$ is retrieved by ", + "bbox": [ + 169, + 513, + 823, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/89a16942bf9824a4e698cc642095618bfc16c35296ad094137f40ca6dc120238.jpg", + "text": "$$\n\\begin{array} { r } { \\phi ^ { t } = ( \\mathbf { w } ^ { t } ) ^ { \\top } \\mathbf { M } ^ { t - 1 } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 434, + 547, + 562, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf { w } ^ { t }$ is a row vector. ", + "bbox": [ + 174, + 574, + 341, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Erasing and Writing Given the erase, write and candidate memory content vectors $( \\mathbf { e } ^ { t } , u _ { j } ^ { t }$ , and $\\bar { \\mathbf { c } } ^ { t }$ respectively) generated by a simple MLP conditioned on the hidden state of the controller $\\mathbf { h } ^ { t }$ , the memory matrix is updated by, ", + "bbox": [ + 173, + 602, + 825, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/15dad782aa04c6aaaed7e3cce13a1a57276f398970a2dd95e9780ae3088c345a.jpg", + "text": "$$\n\\mathbf { C } ^ { t } [ j ] = ( 1 - \\mathbf { e } ^ { t } u _ { j } ^ { t } ) \\odot \\mathbf { C } ^ { t - 1 } [ j ] + u _ { j } ^ { t } \\bar { \\mathbf { c } } ^ { t } .\n$$", + "text_format": "latex", + "bbox": [ + 367, + 651, + 629, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the subscript $j$ in $\\mathbf { C } ^ { t } [ j ]$ denotes the $j$ -th row of the content part $\\mathbf { C } ^ { t }$ of the memory matrix $\\mathbf { M } ^ { t }$ . ", + "bbox": [ + 173, + 676, + 825, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "No Operation (NOP) As found in (Joulin & Mikolov, 2015), an additional NOP action might be beneficial for the controller not to access the memory once in a while. We model this situation by designating one memory cell as a NOP cell. Reading or writing from this memory cell is ignored. ", + "bbox": [ + 174, + 705, + 825, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4 LEARNING ", + "text_level": 1, + "bbox": [ + 174, + 763, + 290, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the proposed D-NTM is executed, it returns the output distribution $p ( \\mathbf { y } | \\mathbf { x } _ { 1 } , \\ldots , \\mathbf { x } _ { T } )$ . As a result, we define a cost function as the negative log-likelihood: ", + "bbox": [ + 174, + 789, + 825, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/31840458fce11743d0141f914540b793f1f0e2db154629e1af8bd93b447893bd.jpg", + "text": "$$\nC ( \\boldsymbol { \\theta } ) = \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } - \\log p ( \\mathbf { y } ^ { n } | \\mathbf { x } _ { 1 } ^ { n } , \\ldots , \\mathbf { x } _ { T } ^ { n } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 364, + 821, + 633, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ is a set of all the parameters. As the proposed D-NTM, just like the original NTM, is fully end-to-end differentiable, we can compute the gradient of this cost function by using backpropagation and learn the parameters of the model with a gradient-based optimization algorithm, such as stochastic gradient descent, to train it end-to-end. ", + "bbox": [ + 174, + 869, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 ADDRESSING MECHANISM ", + "text_level": 1, + "bbox": [ + 176, + 102, + 429, + 118 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 ADDRESS VECTORS ", + "text_level": 1, + "bbox": [ + 174, + 132, + 354, + 146 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each of the address vectors (both read and write) is computed in the same way. The way they are computed are very similar to the content based addressing in (Graves et al., 2014). First, the controller computes a key vector: ", + "bbox": [ + 176, + 157, + 825, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/a6ae921a61326ce67cb1bb7a9f328bc76b1cb77437c57f434b7bc8ff830a4043.jpg", + "text": "$$\n\\mathbf { k } ^ { t } = \\mathbf { W } _ { k } ^ { \\top } \\mathbf { h } ^ { t } + \\mathbf { b } _ { k } ,\n$$", + "text_format": "latex", + "bbox": [ + 431, + 195, + 562, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf { W } _ { k } \\ \\in \\ \\mathbb { R } ^ { N \\times ( d _ { a } + d _ { c } ) }$ and $\\mathbf { b } _ { k } \\ \\in \\ \\mathbb { R } ^ { d _ { a } + d _ { c } }$ if the read head is being computed, otherwise $\\mathbf { W } _ { k } \\in \\mathbb { R } ^ { N \\times d _ { c } }$ and $\\mathbf { b } _ { k } \\in \\mathbb { R } ^ { d _ { c } }$ if the write head weights are being computed. They can be the parameters for a specific head (either read or write.) Also, the sharpening factor $\\beta _ { t } \\in \\mathbb { R } ^ { \\ge 1 }$ is computed as: ", + "bbox": [ + 173, + 217, + 825, + 262 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/6e1477f491f90bacd403cb5066f8a4f9466443d67b7569e458d77982da6a2ae1.jpg", + "text": "$$\n\\begin{array} { c } { { \\mathrm { s o f t p l u s } ( x ) = \\log ( \\exp ( x ) + 1 ) } } \\\\ { { \\beta _ { t } = \\mathrm { s o f t p l u s } ( \\mathbf { u } _ { \\beta } ^ { \\top } \\mathbf { h } ^ { t } + b _ { \\beta } ) + 1 . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 362, + 265, + 633, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf { u } _ { \\beta }$ and $b _ { \\beta }$ are the parameters of the sharpening $\\beta _ { t }$ . ", + "bbox": [ + 173, + 309, + 506, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The address vector is then computed by, ", + "bbox": [ + 174, + 330, + 437, + 344 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/6581f2869a23525f78855075864ae35905f703aaccd08b2b2b4ad6fd12c3d9e4.jpg", + "text": "$$\n\\begin{array} { c } { { z _ { i } ^ { t } = \\beta ^ { t } S \\left( \\mathbf { k } ^ { t } , \\mathbf { m } _ { i } ^ { t } \\right) } } \\\\ { { w _ { i } ^ { t } = \\displaystyle \\frac { \\exp ( z _ { i } ^ { t } ) } { \\sum _ { j } \\exp ( z _ { j } ^ { t } ) } , } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 433, + 348, + 563, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the similarity function $S \\in \\mathbb { R } ^ { \\geq 0 }$ is defined as ", + "bbox": [ + 173, + 411, + 514, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/b7e486ae71fe152885ee338690d3d901339abd8eabfc9bf45b88561c8d7f3f13.jpg", + "text": "$$\nS ( \\mathbf { x } , \\mathbf { y } ) = \\frac { \\mathbf { x } \\cdot \\mathbf { y } } { ( | | \\mathbf { x } | | | \\mathbf { y } | | + \\epsilon ) } .\n$$", + "text_format": "latex", + "bbox": [ + 406, + 431, + 589, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 MULTI-STEP ADDRESSING ", + "text_level": 1, + "bbox": [ + 176, + 477, + 398, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each time-step, controller may require more than one-step for accessing to the memory. The original NTM addresses this by implementing multiple sets of read, erase and write heads. In this paper, we explore an option of allowing each head to operate more than once at each time step, similar to the multi-hop mechanism from the end-to-end memory network (Sukhbaatar et al., 2015). ", + "bbox": [ + 173, + 502, + 825, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 DYNAMIC LEAST RECENTLY USED ADDRESSING ", + "text_level": 1, + "bbox": [ + 174, + 571, + 555, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a memory addressing schema that can learn to put more emphasis on the least recently used (LRU) memory locations. As observed in (Santoro et al., 2016; Rae et al., 2016), we find it easier to learn the write operations with the use of LRU addressing. ", + "bbox": [ + 174, + 597, + 825, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn a LRU based addressing, first we compute the exponentially moving averages of the logits $\\left( \\mathbf { z } _ { t } \\right)$ as $\\mathbf { v } _ { t }$ , $\\mathbf { v } _ { t } ~ = ~ 0 . 1 \\mathbf { v } _ { t - 1 } + 0 . 9 \\mathbf { z } _ { t }$ . We rescale the accumulated $\\mathbf { v } _ { t }$ with $\\gamma _ { t }$ , such that the controller adjusts the influence of how much previously written memory locations should effect the attention weights of a particular time-step. Next, we subtract $\\mathbf { v } _ { t }$ from $\\mathbf { z } _ { t }$ in order to reduce the weights of previously read or written memory locations. $\\gamma _ { t }$ is a shallow MLP with a scalar output and it is conditioned on the hidden state of the controller. $\\gamma _ { t }$ is parametrized with the parameters $\\mathbf { u } _ { \\gamma }$ and ${ \\bf { b } } _ { \\gamma }$ , ", + "bbox": [ + 173, + 643, + 825, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/2fca1f6902ab2cd18ff1732cca4ab7b64bfee10e70afc5c4af35f34445161e14.jpg", + "text": "$$\n\\begin{array} { r } { \\gamma _ { t } = \\operatorname { s i g m o i d } ( \\mathbf { u } _ { \\gamma } ^ { \\top } \\mathbf { h } _ { t } + \\mathbf { b } _ { \\gamma } ) , } \\\\ { \\mathbf { w } _ { t } = \\operatorname { s o f t m a x } \\bigl ( \\mathbf { z } _ { t } - \\gamma _ { t } \\mathbf { v } _ { t - 1 } \\bigr ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 400, + 728, + 596, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This addressing method increases the weights of the least recently used rows of the memory. The magnitude of the influence of the least-recently used memory locations is being learned and adjusted with $\\gamma _ { t }$ . Our LRU addressing is dynamic due to the model’s ability to switch between pure content-based addressing and LRU. During the training, we do not backpropagate through $\\mathbf { v } _ { t }$ . Due to the dynamic nature of this addressing mechanism, it can be used for both read and write operations. If needed, the model will automatically learn to disable LRU while reading from the memory. ", + "bbox": [ + 173, + 779, + 825, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 GENERATING DISCRETE ADDRESS VECTORS ", + "text_level": 1, + "bbox": [ + 174, + 878, + 580, + 895 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe the discrete attention based addressing strategy. ", + "bbox": [ + 173, + 909, + 666, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discrete Addressing Let us use w to denote an address vector (either read, write or erase) at time $t$ . By definition in Eq. (10), every element in this address vector is positive and sums up to one. In other words, we can treat this vector as the probabilities of a categorical distribution $\\mathcal C ( \\mathbf w )$ with $\\dim ( \\mathbf { w } )$ choices: ", + "bbox": [ + 173, + 103, + 825, + 156 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a48b43c75935fa5ebf2a39183e9b61d2e7b31453a88848fa1484d5c7e26baacf.jpg", + "text": "$$\np ( j ) = w _ { j } ,\n$$", + "text_format": "latex", + "bbox": [ + 459, + 155, + 535, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w _ { j }$ is the $j$ -th element of w. We can readily sample from this categorical distribution and form an one-hot vector w˜ such that ", + "bbox": [ + 174, + 172, + 823, + 199 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/967cc989c6c98b115376ccbbce7e228f0ce69961352ea3475e7e07f47aa4bb10.jpg", + "text": "$$\n\\tilde { w } _ { k } = I ( k = j ) ,\n$$", + "text_format": "latex", + "bbox": [ + 444, + 196, + 552, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j \\sim \\mathcal { C } ( \\mathbf { w } )$ , and $I$ is an indicator function. ", + "bbox": [ + 174, + 214, + 486, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training We use this sampling-based strategy for all the heads during training. This clearly makes the use of backpropagation infeasible to compute the gradient, as the sampling procedure is not differentiable. Thus, we use REINFORCE (Williams, 1992) together with the three variance reduction techniques–global baseline, input-dependent baseline and variance normalization– suggested in (Mnih & Gregor, 2014). ", + "bbox": [ + 173, + 243, + 825, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us define $R ( { \\bf x } ) = \\log p ( { \\bf y } | { \\bf x } _ { 1 } , \\ldots , { \\bf x } _ { T } )$ as a reward. We first center and re-scale the reward by ", + "bbox": [ + 169, + 315, + 812, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/24dcd1de042ab3bc1efc7f9256856a07e0724973f8589335669d4d208f1218c1.jpg", + "text": "$$\n\\tilde { R } ( { \\bf x } ) = \\frac { R ( { \\bf x } ) - b } { \\sqrt { \\sigma ^ { 2 } + \\epsilon } } ,\n$$", + "text_format": "latex", + "bbox": [ + 433, + 334, + 563, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b$ and $\\sigma$ is running average and standard deviation of $R$ . We can further center it for each input $\\mathbf { x }$ separately, i.e., ", + "bbox": [ + 173, + 371, + 825, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/73ad2c0af30c26dbbee2a055bafa03cbc552876676f97e051deb52d5f0acf207.jpg", + "text": "$$\n\\tilde { R } ( \\mathbf { x } ) \\gets \\tilde { R } ( \\mathbf { x } ) - b ( \\mathbf { x } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 421, + 397, + 573, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b ( \\mathbf x )$ is computed by a baseline network which takes as input $\\mathbf { x }$ and predicts its estimated reward. The baseline network is trained to minimize the Huber loss (Huber, 1964) between the true reward $\\tilde { R } ( { \\bf x } ) ^ { * }$ and the predicted reward $b ( \\mathbf { x } )$ . We use the Huber loss, which is defined by ", + "bbox": [ + 173, + 416, + 825, + 459 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/e4f5a1c3730e593305a0c56f0c266c5f41401208bfe1b3dfdf05bc020f65d5dd.jpg", + "text": "$$\nH _ { \\delta } ( x ) = { \\left\\{ \\begin{array} { l l } { x ^ { 2 } } & { { \\mathrm { f o r ~ } } | x | \\leq \\delta , } \\\\ { \\delta ( 2 | x | - \\delta ) , } & { { \\mathrm { o t h e r w i s e } } , } \\end{array} \\right. }\n$$", + "text_format": "latex", + "bbox": [ + 372, + 463, + 622, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "due to its robustness. As a further measure to reduce the variance, we regularize the negative entropy of all those category distributions to facilitate a better exploration during training (Xu et al., 2015). ", + "bbox": [ + 171, + 502, + 825, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the cost function for each training example is approximated as ", + "bbox": [ + 174, + 535, + 625, + 549 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/bffd054bc17ac384c9d9037ee3996704175726c7a06e6adc5c2450df911bf792.jpg", + "text": "$$\n\\begin{array} { r l } & { C ^ { n } ( \\boldsymbol { \\theta } ) = - \\log p ( \\mathbf { y } | \\mathbf { x } _ { 1 : T } , \\tilde { w } _ { 1 : J } , \\tilde { u } _ { 1 : J } , \\tilde { e } _ { 1 : J } ) } \\\\ & { \\qquad - \\displaystyle \\sum _ { j = 1 } ^ { J } \\tilde { R } ( \\mathbf { x } ^ { n } ) ( \\log p ( \\tilde { w } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { u } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { e } _ { j } | \\mathbf { x } _ { 1 : T } ) ) } \\\\ & { \\qquad - \\lambda _ { H } \\displaystyle \\sum _ { j = 1 } ^ { J } ( \\mathcal { H } ( w _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( u _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( e _ { j } | \\mathbf { x } _ { 1 : T } ) ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 289, + 551, + 784, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $J$ is the number of addressing steps, $\\lambda _ { H }$ is the entropy regularization coefficient, and $\\mathcal { H }$ denotes the entropy. ", + "bbox": [ + 171, + 665, + 823, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference Once training is over, we switch to a deterministic strategy. We simply choose an element of w with the largest value to be the index of the target memory cell, such that ", + "bbox": [ + 169, + 707, + 823, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/92541e0be0f429b20ffa6a34399d1519d61c50f5dd696ad4ec1c1e4ef7a3cc0d.jpg", + "text": "$$\n\\tilde { w } _ { k } = \\mathbf { I } ( k = \\operatorname { a r g m a x } ( \\mathbf { w } ) ) .\n$$", + "text_format": "latex", + "bbox": [ + 411, + 737, + 584, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Curriculum Learning for the Discrete Attention Training discrete attention with feed-forward controller and REINFORCE is challenging. We propose to use a curriculum strategy for training with the discrete attention in order to tackle this problem. For each minibatch, we sample $\\pi$ from a binomial distribution with the probability $p ^ { t }$ , $\\pi ^ { t } \\sim \\mathrm { B i n } ( p ^ { t } )$ . The model will either use the discrete or the continuous-attention based on the $\\pi ^ { t }$ . We start the training procedure with $p ^ { 0 } = 1$ and during the training $p ^ { t }$ is annealed to 0 by setting $\\begin{array} { r } { p ^ { t } = \\frac { p ^ { 0 } } { \\sqrt { 1 + t } } } \\end{array}$ . ", + "bbox": [ + 173, + 763, + 825, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can rewrite the weights $\\mathbf { w } _ { t }$ as in Equation 14, where it is expressed as the combination of continuous attention weights $\\hat { \\mathbf { w } } ^ { t }$ and discrete attention weights $\\tilde { \\mathbf { w } } ^ { t }$ with $\\mathit { \\Pi } _ { \\pi ^ { t } }$ being a binary variable that chooses to use one of them, ", + "bbox": [ + 174, + 859, + 825, + 901 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/40eea2ba87dd6930595aeecbe70dc2633c2eb8c53228b730ad911bb606067a39.jpg", + "text": "$$\n\\mathbf { w } ^ { t } \\pi ^ { t } \\bar { \\mathbf { w } } ^ { t } + ( 1 - \\pi ^ { t } ) \\tilde { \\mathbf { w } } ^ { t } .\n$$", + "text_format": "latex", + "bbox": [ + 406, + 902, + 591, + 921 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By using this curriculum learning strategy, at the beginning of the training, the model learns to use the memory mainly with the continuous attention. As we anneal the $p ^ { t }$ , the model will rely more on the discrete attention. ", + "bbox": [ + 174, + 103, + 823, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 REGULARIZING DYNAMIC NEURAL TURING MACHINES ", + "text_level": 1, + "bbox": [ + 174, + 162, + 671, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When the controller of D-NTM is a powerful recurrent neural network, it is important to regularize training of the D-NTM so as to avoid suboptimal solutions in which the D-NTM ignores the memory and works as a simple recurrent neural network. ", + "bbox": [ + 174, + 194, + 825, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Read-Write Consistency Regularizer One such suboptimal solution we have observed in our preliminary experiments with the proposed D-NTM is that the D-NTM uses the address part A of the memory matrix simply as an additional weight matrix, rather than as a means to accessing the content part C. We found that this pathological case can be effectively avoided by encouraging the read head to point to a memory cell which has also been pointed by the write head. This can be implemented as the following regularization term: ", + "bbox": [ + 174, + 250, + 825, + 330 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/eea88a11afc5bbd2611efe1c1a4ac0e2a2be4fba9a0874199647da45aad7a8d5.jpg", + "text": "$$\nR _ { \\mathrm { r w } } ( \\mathbf { w } , \\mathbf { u } ) = \\lambda \\sum _ { t ^ { \\prime } = 1 } ^ { T } | | 1 - ( \\frac { 1 } { t ^ { \\prime } } \\sum _ { t = 1 } ^ { t ^ { \\prime } } \\mathbf { u } _ { t } ) ^ { \\top } \\mathbf { w } _ { t ^ { \\prime } } | | _ { 2 } ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 349, + 337, + 648, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the equations above, $\\mathbf { u } _ { t }$ is the write and $\\mathbf { w } _ { t }$ is the read weights. ", + "bbox": [ + 173, + 393, + 604, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next Input Prediction as Regularization Temporal structure is a strong signal that should be exploited by the controller based on a recurrent neural network. We exploit this structure by letting the controller predict the input in the future. We maximize the predictability of the next input by the controller during training. This is equivalent to minimizing the following regularizer: ", + "bbox": [ + 173, + 422, + 825, + 478 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/daa07f4d77ca80eb4471c5abe2b0db80830859f8307bfb89e9b482206c4def7c.jpg", + "text": "$$\nR _ { \\mathrm { p r e d } } ( \\mathbf { W } ) = - \\log p ( \\mathbf { f } _ { t + 1 } | \\mathbf { f } _ { t } , \\mathbf { w } _ { t } , \\mathbf { u } _ { t } , \\mathbf { M } _ { t } ; \\mathbf { W } ) )\n$$", + "text_format": "latex", + "bbox": [ + 343, + 484, + 651, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f _ { t }$ is the current input and $f _ { t + 1 }$ is the input at next timestep. We found this regularizer to be effective in our preliminary experiments and use it for bAbI tasks. ", + "bbox": [ + 171, + 506, + 821, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 553, + 344, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A recurrent neural network (RNN), which is used as a controller in the proposed D-NTM, has an implicit memory in the form of recurring hidden states. Even with this implicit memory, a vanilla RNN is however known to have difficulties in storing information for long time-spans (Bengio et al., 1994; Hochreiter, 1991). Long short-term memory (LSTM, (Hochreiter & Schmidhuber, 1997)) and gated recurrent units (GRU, (Cho et al., 2014)) have been found to address this issue. However all these models based solely on RNNs have been found to be limited when they are used to solve, e.g., algorithmic tasks and episodic question-answering. ", + "bbox": [ + 173, + 583, + 825, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the finite random access memory of the neural Turing machine, based on which the D-NTM is designed, other data structures have been proposed as external memory for neural networks. In (Sun et al., 1997; Grefenstette et al., 2015; Joulin & Mikolov, 2015), a continuous, differentiable stack was proposed. In (Zaremba et al., 2015; Zaremba & Sutskever, 2015), grid and tape storages are used. These approaches differ from the NTM in that their memory is unbounded and can grow indefinitely. On the other hand, they are often not randomly accessible. ", + "bbox": [ + 174, + 684, + 825, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory networks (Weston et al., 2015b) form another family of neural networks with external memory. In this class of neural networks, information is stored explicitly as it is (in the form of its continuous representation) in the memory, without being erased or modified during an episode. Memory networks and their variants have been applied to various tasks successfully (Sukhbaatar et al., 2015; Bordes et al., 2015; Dodge et al., 2015; Xiong et al., 2016). Miller et al. (2016) have also independently proposed the idea of having separate key and value vectors for memory networks. ", + "bbox": [ + 174, + 770, + 825, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another related family of models is the attention-based neural networks. Neural networks with continuous or discrete attention over an input have shown promising results on a variety of challenging tasks, including machine translation (Bahdanau et al., 2015; Luong et al., 2015), speech recognition (Chorowski et al., 2015), machine reading comprehension (Hermann et al., 2015) and image caption generation (Xu et al., 2015). ", + "bbox": [ + 174, + 857, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The latter two, the memory network and attention-based networks, are however clearly distinguishable from the D-NTM by the fact that they do not modify the content of the memory. ", + "bbox": [ + 171, + 103, + 823, + 131 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 150, + 326, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide experimental results to demonstrate the abilities of our model, first on Facebook bAbI task (Weston et al., 2015a). We give detailed analysis and experimental results on this task. We also compare different variations of NTM on bAbI tasks. We have performed experiments on sequential permuted MNIST (Le et al., 2015) and on toy tasks to compare other published models on these tasks with a recurrent controller. The details of our experiments are provided in the supplementary material. ", + "bbox": [ + 174, + 181, + 825, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1 EPISODIC QUESTION-ANSWERING: BABI TASKS ", + "text_level": 1, + "bbox": [ + 174, + 263, + 552, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed D-NTM on the recently proposed episodic question-answering task called Facebook bAbI. We use the dataset with 10k training examples per sub-task provided by Facebook.1 For each episode, the D-NTM reads a sequence of factual sentences followed by a question, all of which are given as natural language sentences. The D-NTM is expected to store and retrieve relevant information in the memory in order to answer the question based on the presented facts. Exact implementation details and hyper-parameter settings are provided in the appendix. ", + "bbox": [ + 174, + 289, + 825, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1.1 GOALS", + "text_level": 1, + "bbox": [ + 176, + 385, + 277, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The goal of this experiment is three-fold. First, we present for the first time the performance of a memory-based network that can both read and write dynamically on the Facebook bAbI tasks2. We aim to understand whether a model that has to learn to write an incoming fact to the memory, rather than storing it as it is, is able to work well, and to do so, we compare both the original NTM and proposed D-NTM against an LSTM-RNN. ", + "bbox": [ + 174, + 409, + 825, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, we investigate the effect of having to learn how to write. The fact that the NTM needs to learn to write likely has adverse effect on the overall performance, when compared to, for instance, end-to-end memory networks (MemN2N, (Sukhbaatar et al., 2015)) and dynamic memory network $\\mathrm { ( D M N + }$ , (Xiong et al., 2016)) both of which simply store the incoming facts as they are. We quantify this effect in this experiment. Lastly, we show the effect of the proposed learnable addressing scheme. ", + "bbox": [ + 174, + 483, + 825, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further explore the effect of using a feedforward controller instead of the GRU controller. In addition to the explicit memory, the GRU controller can use its own internal hidden state as the memory. On the other hand, the feedforward controller must solely rely on the explicit memory, as it is the only memory available. ", + "bbox": [ + 174, + 556, + 825, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1.2 RESULTS AND ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 625, + 401, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we first observe that the NTMs are indeed capable of solving this type of episodic question-answering better than the vanilla LSTM-RNN. Although the availability of explicit memory in the NTM has already suggested this result, we note that this is the first time neural Turing machines have been used in this specific task. ", + "bbox": [ + 174, + 648, + 825, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All the variants of NTM with the GRU controller outperform the vanilla LSTM-RNN. However, not all of them perform equally well. First, it is clear that the proposed dynamic NTM (D-NTM) using the GRU controller outperforms the original NTM with the GRU controller (NTM, CBA only NTM vs. continuous D-NTM, Discrete D-NTM). As discussed earlier, the learnable addressing scheme of the D-NTM allows the controller to access the memory slots by location in a potentially nonlinear way. We expect it to help with tasks that have non-trivial access patterns, and as anticipated, we see a large gain with the D-NTM over the original NTM in the tasks of, for instance, 12 - Conjunction and 17 - Positional Reasoning. ", + "bbox": [ + 174, + 708, + 825, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among the recurrent variants of the proposed D-NTM, we notice significant improvements by using discrete addressing over using continuous addressing. We conjecture that this is due to certain types of tasks that require precise/sharp retrieval of a stored fact, in which case continuous addressing is in disadvantage over discrete addressing. This is evident from the observation that the D-NTM with discrete addressing significantly outperforms that with continuous addressing in the tasks of 8 - ", + "bbox": [ + 174, + 808, + 825, + 876 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d8d524f20ba7886f2d7b3b5ca3db3d355229f453ab7efd976a9be7efb577b54.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DMN+L1-step CBA1-step Soft D-NTM1-step Discrete D-NTM3-steps LBA* NTM3-steps CBA NTM3-steps Soft D-NTM3-steps Discrete D-NTM
Task 1LSTM 0.00MemN2N 0.000.00NTM 16.30NTM 16.885.416.660.000.000.000.00
281.900.300.3057.0855.7058.5456.0461.6759.3846.6662.29
383.102.101.1074.1655.0074.5872.0883.5465.2147.0841.45
40.200.000.000.000.000.000.000.000.000.000.00
51.200.800.501.4620.411.661.040.831.461.251.45
651.800.100.0023.3321.0440.2044.7948.1354.8020.6211.04
724.902.002.4021.6721.6719.1619.587.9237.707.295.62
834.100.900.0025.7621.0512.5818.4625.388.8211.020.74
920.200.300.0024.7924.1736.6634.3737.800.0039.3732.50
1030.100.000.0041.4633.1352.2950.8356.2523.7520.0020.83
1110.300.100.0018.9631.8831.454.163.960.2830.6216.87
1223.400.000.0025.8330.007.706.6628.7523.755.414.58
136.100.000.006.675.635.622.295.8383.137.915.00
1481.000.100.2058.5459.1760.0063.7561.8857.7158.1260.20
1578.700.000.0036.4642.3036.8739.2735.6221.8836.0440.26
1651.9051.8045.3071.1571.1549.1651.3546.1550.0046.0445.41
1750.1018.604.2043.7543.7517.9116.0443.7556.2521.259.16
186.805.302.103.9647.503.953.5447.5047.506.871.66
1990.302.300.0075.8971.5173.7464.6361.5663.6575.8876.66
202.100.000.001.250.002.703.120.400.003.330.00
Avg.Err.36.414.242.8131.4233.6029.5127.9332.8532.7624.2421.79
", + "bbox": [ + 194, + 94, + 803, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: Test error rates $( \\% )$ on the 20 bAbI QA tasks for models using 10k training examples with the GRU and feedforward controller. FF stands for the experiments that are conducted with feedforward controller. Let us, note that $\\mathrm { L B A ^ { * } }$ refers to NTM that uses both LBA and CBA. In this table, we compare multi-step vs single-step addressing, original NTM with location based+content based addressing vs only content based addressing, and discrete vs continuous addressing on bAbI. ", + "bbox": [ + 173, + 309, + 825, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lists/Sets and 11 - Basic Coreference. Furthermore, this is in line with an earlier observation in (Xu et al., 2015), where discrete addressing was found to generalize better in the task of image caption generation. ", + "bbox": [ + 174, + 410, + 823, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 2, we also observe that the D-NTM with the feedforward controller and discrete attention performs worse than LSTM and D-NTM with continuous-attention. However, when the proposed curriculum strategy from Sec. 4 is used, the average test error drops from 68.30 to 37.79. ", + "bbox": [ + 174, + 444, + 823, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We empirically found training of the feedforward controller more difficult than that of the recurrent controller. We train our feedforward controller based models four times longer (in terms of the number of updates) than the recurrent controller based ones in order to ensure that they are converged for most of the tasks. On the other hand, the models trained with the GRU controller overfit on bAbI tasks very quickly. For example, on tasks 3 and 16 the feedforward controller based model underfits (i.e., high training loss) at the end of the training, whereas with the same number of units the model with the GRU controller can overfit on those tasks after 3,000 updates only. ", + "bbox": [ + 173, + 491, + 825, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When our results are compared to the variants of the memory network Weston et al. (2015b) (MemN2N and $\\mathrm { D M N + }$ ), we notice a significant performance gap. We attribute this gap to the difficulty in learning to manipulate and store a complex input. ", + "bbox": [ + 174, + 590, + 821, + 632 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/15be19b050d013aa2678635d887fe055e193f11fbac6d6b0f3bc9e3da80e1f5c.jpg", + "table_caption": [ + "Table 2: Test error rates $( \\% )$ on the 20 bAbI QA tasks for models using 10k training examples with feedforward controller. " + ], + "table_footnote": [], + "table_body": "
TaskFF Soft D-NTMFF Discrete D-NTMFF Discrete* D-NTM
14.3881.6714.79
227.576.6776.67
371.2579.3870.83
40.0078.6544.06
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.62
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.92
1939.4771.4671.56
200.076.569.79
Avg.Err.12.8168.3037.79
", + "bbox": [ + 388, + 645, + 607, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also provide further experiments investigating different extensions on D-NTM in the appendix. ", + "bbox": [ + 171, + 909, + 818, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2 SEQUENTIAL pMNIST ", + "text_level": 1, + "bbox": [ + 176, + 103, + 372, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In sequential MNIST task, the pixels of the MNIST digits are provided to the model in scan line order, left to right and top to bottom (Le et al., 2015). At the end of sequence of pixels, the model predicts the label of the digit in the sequence of pixels. We experiment D-NTM on the variation of sequential MNIST where the order of the pixels is randomly shuffled, we call this task as permuted MNIST $( p \\mathrm { M N I S T } )$ . An important contribution of this task to our paper, in particular, is to measure the model’s ability to perform well when dealing with long-term dependencies. We report our results in Table $3 ^ { 3 }$ , we observe improvements over other models that we compare against. In Table 3, ”discrete addressing with MAB” refers to D-NTM model using REINFORCE with baseline computed from moving averages of the reward. Discrete addressing with IB refers to D-NTM using REINFORCE with input-based baseline. ", + "bbox": [ + 174, + 130, + 825, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7.3 NTM TOY TASKS ", + "text_level": 1, + "bbox": [ + 176, + 268, + 338, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We explore the possibility of using D-NTM to solve algorithmic tasks such as copy and associative recall tasks. We train our model on the same lengths of sequences that is experimented in (Graves et al., 2014). We report our results in Table 4. We find out that D-NTM using continuous-attention can successfully learn the ”Copy” and ”Associative Recall” tasks. ", + "bbox": [ + 174, + 295, + 825, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Table 4, we train our model on sequences of the same length as the experiments in (Graves et al., 2014) and test the model on the sequences of the maximum length seen during the training. We consider model to be successful on copy or associative recall if its validation cost (binary cross-entropy) is lower than 0.02 over the sequences of maximum length seen during the training. We set the threshold to 0.02 to determine whether a model is successful on a task. Because empirically we observe that the models have higher validation costs perform badly in terms of generalization over the longer sequences. ”D-NTM discrete” model in this table is trained with REINFORCE using moving averages to estimate the baseline. ", + "bbox": [ + 173, + 354, + 825, + 448 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/14592f49a2c7ebec424d694080a5007ed1bdca9bbff954eff64d83cf13a9c5d1.jpg", + "table_caption": [ + "Table 3: Sequential $p { \\mathrm { M N I S T } } .$ . " + ], + "table_footnote": [], + "table_body": "
Test Acc
D-NTMdiscrete MAB89.6
D-NTMdiscrete IB92.3
Soft D-NTM93.4
NTM90.9
I-RNN (Le et al., 2015)82.0
Zoneout (Krueger et al.,2016)93.1
LSTM (Krueger et al., 2016)89.8
Unitary-RNN (Arjovsky et al., 2015)91.4
Recurrent Dropout (Krueger et al., 2016)92.5
", + "bbox": [ + 179, + 463, + 495, + 604 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d69ed6640d1288038d9eaeb3dd7f3152bddf9f7e0e4fa3cb18154291caeb585f.jpg", + "table_caption": [ + "Table 4: NTM Toy Tasks. " + ], + "table_footnote": [], + "table_body": "
Copy TasksAssociative Recall
Soft D-NTMSuccessSuccess
D-NTMdiscreteSuccessFailure
NTMSuccessSuccess
", + "bbox": [ + 500, + 502, + 818, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8 CONCLUSION AND FUTURE WORK ", + "text_level": 1, + "bbox": [ + 174, + 661, + 495, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper we extend neural Turing machines (NTM) by introducing a learnable addressing scheme which allows the NTM to be capable of performing highly nonlinear location-based addressing. This extension, to which we refer by dynamic NTM (D-NTM), is extensively tested with various configurations, including different addressing mechanisms (continuous vs. discrete) and different number of addressing steps, on the Facebook bAbI tasks. This is the first time an NTM-type model was tested on this task, and we observe that the NTM, especially the proposed D-NTM, performs better than vanilla LSTM-RNN. Furthermore, the experiments revealed that the discrete, discrete addressing works better than the continuous addressing with the GRU controller, and our analysis reveals that this is the case when the task requires precise retrieval of memory content. ", + "bbox": [ + 173, + 693, + 825, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our experiments show that the NTM-based models can be weaker than other variants of memory networks which do not learn but have an explicit mechanism of storing incoming facts as they are. We conjecture that this is due to the difficulty in learning how to write, manipulate and delete the content of memory. Despite this difficulty, we find the NTM-based approach, such as the proposed D-NTM, to be a better, future-proof approach, because it can scale to a much longer horizon (where it becomes impossible to explicitly store all the experiences.) ", + "bbox": [ + 174, + 819, + 825, + 872 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 131 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "On $p \\mathrm { M N I S T }$ task, we show that our model can outperform other similar type of approaches proposed to deal with the long-term dependencies. On copy and associative recall tasks, we show that our model can solve the algorithmic problems that are proposed to solve with NTM type of models. ", + "bbox": [ + 174, + 137, + 825, + 178 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The success of both the learnable address and the discrete addressing scheme suggests two future research directions. First, we should try both of these schemes in a wider array of memory-based models, as they are not specific to the neural Turing machines. Second, the proposed D-NTM needs to be evaluated on a diverse set of applications, such as text summarization (Rush et al., 2015), visual questionanswering (Antol et al., 2015) and machine translation, in order to make a more concrete conclusion. ", + "bbox": [ + 174, + 184, + 825, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 271, + 285, + 286 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pp. 2425–2433, 2015. \nMartin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. arXiv preprint arXiv:1511.06464, 2015. \nDzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In Proceedings Of The International Conference on Representation Learning (ICLR 2015), 2015. \nYoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994. \nAntoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question answering with memory networks. arXiv preprint arXiv:1506.02075, 2015. \nKyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. \nJan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio. Attention-based models for speech recognition. arXiv preprint arXiv:1506.07503, 2015. \nTim Cooijmans, Nicolas Ballas, Cesar Laurent, and Aaron Courville. Recurrent batch normalization. ´ arXiv preprint arXiv:1603.09025, 2016. \nJesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur Szlam, and Jason Weston. Evaluating prerequisite qualities for learning end-to-end dialog systems. CoRR, abs/1511.06931, 2015. \nIan Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. Book in preparation for MIT Press, 2016. URL http://www.deeplearningbook.org. \nAlex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. \nAlex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. ´ Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626): 471–476, 2016. \nEdward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp. 1819–1827, 2015. \nCaglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions. arXiv preprint arXiv:1603.00391, 2016. \nKarl Moritz Hermann, Toma´s Ko ˇ cisk ˇ y, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa \\` Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. arXiv preprint arXiv:1506.03340, 2015. ", + "bbox": [ + 171, + 291, + 828, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s books with explicit memory representations. arXiv preprint arXiv:1511.02301, 2015. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sepp Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Diploma, Technische Universitat¨ Munchen ¨ , pp. 91, 1991. ", + "bbox": [ + 173, + 140, + 823, + 167 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural Computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 173, + 176, + 825, + 204 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Peter J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73–101, 03 1964. ", + "bbox": [ + 173, + 212, + 825, + 228 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In Advances in Neural Information Processing Systems, pp. 190–198, 2015. ", + "bbox": [ + 173, + 236, + 823, + 265 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. ", + "bbox": [ + 173, + 272, + 823, + 300 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Krueger, Tegan Maharaj, Janos Kram ´ ar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary ´ Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016. ", + "bbox": [ + 174, + 308, + 825, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quoc V Le, Navdeep Jaitly, and Geoffrey E Hinton. A simple way to initialize recurrent networks of rectified linear units. arXiv preprint arXiv:1504.00941, 2015. ", + "bbox": [ + 173, + 357, + 823, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In Proceedings Of The Conference on Empirical Methods for Natural Language Processing (EMNLP 2015), 2015. ", + "bbox": [ + 174, + 393, + 823, + 435 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. Key-value memory networks for directly reading documents. CoRR, abs/1606.03126, 2016. URL http://arxiv.org/abs/1606.03126. ", + "bbox": [ + 174, + 443, + 823, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. arXiv preprint arXiv:1402.0030, 2014. ", + "bbox": [ + 169, + 492, + 825, + 521 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jack W Rae, Jonathan J Hunt, Tim Harley, Ivo Danihelka, Andrew Senior, Greg Wayne, Alex Graves, and Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes. In Advances in NIPS. 2016. ", + "bbox": [ + 173, + 529, + 825, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scott Reed and Nando de Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279, 2015. ", + "bbox": [ + 171, + 578, + 825, + 608 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexander M. Rush, Sumit Chopra, and Jason Weston. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pp. 379–389, 2015. ", + "bbox": [ + 173, + 614, + 823, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-shot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016. ", + "bbox": [ + 169, + 664, + 823, + 694 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. Building end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI-16), 2016. ", + "bbox": [ + 173, + 700, + 825, + 742 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. arXiv preprint arXiv:1503.08895, 2015. ", + "bbox": [ + 173, + 751, + 825, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Guo-Zheng Sun, C. Lee Giles, and Hsing-Hen Chen. The neural network pushdown automaton: Architecture, dynamics and training. In Adaptive Processing of Sequences and Data Structures, International Summer School on Neural Networks, pp. 296–345, 1997. ", + "bbox": [ + 173, + 786, + 826, + 828 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Oriol Vinyals and Quoc Le. A neural conversational model. arXiv preprint arXiv:1506.05869, 2015. ", + "bbox": [ + 171, + 837, + 825, + 852 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jason Weston, Antoine Bordes, Sumit Chopra, and Tomas Mikolov. Towards ai-complete question answering: a set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015a. ", + "bbox": [ + 173, + 859, + 825, + 888 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Proceedings Of The International Conference on Representation Learning (ICLR 2015), 2015b. In Press. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8:229–256, 1992. \nCaiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. CoRR, abs/1603.01417, 2016. \nKelvin Xu, Jimmy Ba, Ryan Kiros, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings Of The International Conference on Representation Learning (ICLR 2015), 2015. \nLi Yao, Atousa Torabi, Kyunghyun Cho, Nicolas Ballas, Christopher Pal, Hugo Larochelle, and Aaron Courville. Describing videos by exploiting temporal structure. In Computer Vision (ICCV), 2015 IEEE International Conference on. IEEE, 2015. \nWojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines. CoRR, abs/1505.00521, 2015. \nWojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms from examples. arXiv preprint arXiv:1511.07275, 2015. ", + "bbox": [ + 171, + 103, + 826, + 340 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A EXPERIMENTAL DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 421, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1 MODEL AND TRAINING DETAILS FOR BABI ", + "bbox": [ + 174, + 132, + 522, + 147 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We use the same hyperparameters for all the tasks for a given model. ", + "bbox": [ + 174, + 157, + 622, + 172 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1.1 FACT REPRESENTATION ", + "text_level": 1, + "bbox": [ + 176, + 185, + 398, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We use a recurrent neural network with GRU units to encode a variable-length fact into a fixed-size vector representation. This allows the D-NTM to exploit the word ordering in each fact, unlike when facts are encoded as bag-of-words vectors. ", + "bbox": [ + 174, + 210, + 825, + 251 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1.2 CONTROLLER ", + "text_level": 1, + "bbox": [ + 176, + 265, + 328, + 280 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We experiment with both a recurrent and feedforward neural network as the controller that generates the read and write weights. The controller has 180 units. We train our feed-forward controller using noisy-tanh activation function (Gulcehre et al., 2016) since we were experiencing training difficulties with sigmoid and tanh activation functions. We use both single-step and three-steps addressing with our GRU controller. ", + "bbox": [ + 174, + 289, + 825, + 356 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1.3 MEMORY ", + "text_level": 1, + "bbox": [ + 176, + 369, + 297, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The memory contains 120 memory cells. Each memory cell consists of a 16-dimensional address part and 28-dimensional content part. ", + "bbox": [ + 176, + 393, + 823, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1.4 TRAINING DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 435, + 366, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We set aside a random $1 0 \\%$ of the training examples as a validation set for each sub-task and use it for early-stopping and hyperparameter search. We train one D-NTM for each sub-task, using Adam (Kingma & Ba, 2014) with its learning rate set to 0.003 and 0.007 respectively for GRU and Feedforward controller. The size of each minibatch is 160, and each minibatch is constructed uniform-randomly from the training set. ", + "bbox": [ + 174, + 459, + 825, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.2 MODEL AND TRAINING DETAILS FOR SEQUENTIAL MNIST ", + "text_level": 1, + "bbox": [ + 174, + 542, + 632, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "On sequential MNIST task we try to keep the capacity of our model to be close to our baselines. We use 100 GRU units in the controller and each content vector of size 8 and with address vectors of size 8. We use a learning rate of $1 e - 3$ and trained the model with adam optimizer. We did not use the read and write consistency regularization in any of our models. ", + "bbox": [ + 174, + 568, + 825, + 621 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.3 MODEL AND TRAINING DETAILS FOR TOY TASKS ", + "text_level": 1, + "bbox": [ + 174, + 637, + 563, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "On both copy and associative recall tasks, we try to keep the capacity of our model to be close to our baselines. We use 100 GRU units in the controller and each content vector of has a size of 8 and using address vector of size 8. We use a learning rate of $1 e - 3$ and trained the model with adam optimizer. We did not use the read and write consistency regularization in any of our models. For the model with the discrete attention we use REINFORCE with baseline computed using moving averages. ", + "bbox": [ + 174, + 662, + 825, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B VISUALIZATION OF DISCRETE ATTENTION ", + "text_level": 1, + "bbox": [ + 174, + 748, + 563, + 765 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We visualize the attention of D-NTM with GRU controller with discrete attention in Figure 2. From this example, we can see that D-NTM has learned to find the correct supporting fact even without any supervision for the particular story in the visualization. ", + "bbox": [ + 174, + 779, + 825, + 820 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C LEARNING CURVES FOR THE RECURRENT CONTROLLER", + "text_level": 1, + "bbox": [ + 174, + 839, + 676, + 854 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 3, we compare the learning curves of the continuous and discrete attention D-NTM model with recurrent controller on Task 1. Surprisingly, the discrete attention D-NTM converges faster than the continuous-attention model. The main difficulty of learning continuous-attention is due to the fact that learning to write with continuous-attention can be challenging. ", + "bbox": [ + 176, + 869, + 823, + 924 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/13f6be6816a8ab4ad76006098fb07e79c29ffd5f83dd209f067809e9e7421ce0.jpg", + "image_caption": [ + "Figure 2: An example view of the discrete attention over the memory slots for both read (left) and write heads(right). $\\mathbf { X }$ -axis the denotes the memory locations that are being accessed and y-axis corresponds to the content in the particular memory location. In this figure, we visualize the discrete-attention model with 3-reading steps and on task-20. It is easy to see that the NTM with discrete-attention accesses to the relevant part of the memory. We only visualize the last-step of the 3-steps writing. Because with discrete attention usually the model just reads the empty slots of the memory. " + ], + "image_footnote": [], + "bbox": [ + 225, + 98, + 696, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/750ba06aa86f2fcf7b1b8477065761e178ba6d5f4dd1a6aa03704d44c0632dbd.jpg", + "image_caption": [ + "Figure 3: A visualization for the learning curves of continuous and discrete D-NTM models trained on Task 1 using 3 steps. In most tasks, we observe that the discrete attention model with GRU controller does converge faster than the continuous-attention model. " + ], + "image_footnote": [], + "bbox": [ + 281, + 400, + 709, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D A COMPARISON BETWEEN THE LEARNINGCURVES OF INPUT BASED BASELINE AND REGULAR BASELINE ON $p$ MNIST", + "text_level": 1, + "bbox": [ + 179, + 727, + 821, + 760 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Figure 4, we show the learning curves of input-based-baseline (ibb) and regular REINFORCE with moving averages baseline (mab) on the $p \\mathrm { M N I S T }$ task. We observe that input-based-baseline in general is much easier to optimize and converges faster as well. But it can quickly overfit to the task as well. ", + "bbox": [ + 173, + 773, + 826, + 815 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "E TRAINING WITH CONTINUOUS-ATTENTION AND TESTING WITH DISCRETE-ATTENTION ", + "text_level": 1, + "bbox": [ + 174, + 835, + 821, + 868 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Table 5, we provide results investigating the effects of using discrete attention model at the test-time for a model trained with feed-forward controller and continuous attention. Discrete∗ D-NTM model bootstraps the discrete attention with the continuous attention, using the curriculum method that we have introduced in Section ”Curriculum Learning for the Discrete Attention”. Discrete† D-NTM model is the continuous-attention model which uses discrete-attention at the test time. We observe that the Discrete† D-NTM model which is trained with continuous-attention outperforms Discrete D-NTM model. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/858347b800db4692b6c99c985fab4630eac6e5ffb7ee526a7c12130eaa865972.jpg", + "image_caption": [ + "Figure 4: We compare the learning curves of our D-NTM model using discrete attention on $p \\mathrm { M N I S T }$ task with input-based baseline and regular REINFORCE baseline. The $\\mathbf { X }$ -axis is the loss and y-axis is the number of epochs. " + ], + "image_footnote": [], + "bbox": [ + 281, + 123, + 710, + 366 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 476, + 825, + 520 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/d8842f21cc4b747e67123ff145e3356ea1f79cdfc3c8d7e3c47994ff0c97625c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskcontinuousD-NTMDiscreteD-NTMDiscrete*D-NTMDiscrete†D-NTM
14.3881.6714.7972.2881.6778.9579.6968.5431.6749.17
23427.571.2576.6779.3878.6576.6770.8344.06
0.00
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.6279.3237.7125.6382.0874.3847.0877.0873.9653.0230.4211.4676.0513.96
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.9271.569.79
192039.4771.4676.56
0.0
Avg12.8168.3037.7957.21
", + "bbox": [ + 320, + 545, + 678, + 827 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 5: Test error rates $( \\% )$ on the 20 bAbI QA tasks for models using 10k training examples with the feedforward controller. Discrete∗ D-NTM model bootstraps the discrete attention with the continuous attention, using the curriculum method that we have introduced in Section 4. Discrete† D-NTM model is the continuous-attention model which uses discrete-attention at the test time. ", + "bbox": [ + 173, + 835, + 825, + 892 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "F D-NTM WITH BOW FACT REPRESENTATION ", + "text_level": 1, + "bbox": [ + 173, + 102, + 578, + 118 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In Table 6, we provide results for D-NTM using BoW with positional encoding (PE) Sukhbaatar et al. (2015) as the representation of the input facts. The facts representations are provided as an input to the GRU controller. In agreement to our results with the GRU fact representation, with the BoW fact representation we observe improvements with multi-step of addressing over single-step and discrete addressing over continuous addressing. ", + "bbox": [ + 174, + 132, + 825, + 200 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/0a6c5ddf5f16fcf35d62d0c38760c7d4999bd44c3759192a955461176c3e8610.jpg", + "table_caption": [ + "Table 6: Test error rates $( \\% )$ on the 20 bAbI QA tasks for models using 10k training examples with the GRU controller and representations of facts are obtained with $\\mathrm { B o W }$ using positional encoding. " + ], + "table_footnote": [], + "table_body": "
TaskSoftD-NTM(1-step)DiscreteD-NTM(1-step)SoftD-NTM(3-steps)DiscreteD-NTM(3-steps)
10.000.000.000.00
23461.0459.3756.8755.62
55.6257.562.557.5
27.2924.8926.4527.08
513.5512.0815.8314.78
613.5414.3721.8713.33
78.546.258.7514.58
81.691.363.013.02
917.716.6637.7017.08
1026.0427.0826.8723.952.294.16
1120.413.952.5
120.410.830.20
133.121.044.795.83
1462.0858.3361.2560.620.05
1531.6626.250.62
1654.4748.5448.9548.9530.6236.04
1743.7531.8743.75
1833.7539.3736.66
192064.6369.2167.2365.46
1.250.001.450.00
Avg27.0224.9826.3624.05
", + "bbox": [ + 240, + 205, + 758, + 484 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/parse/train/BkSmc8qll/BkSmc8qll_middle.json b/parse/train/BkSmc8qll/BkSmc8qll_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..63ebd61e7224e09e6e1735c92b59fb98383b9ab6 --- /dev/null +++ b/parse/train/BkSmc8qll/BkSmc8qll_middle.json @@ -0,0 +1,37973 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 505, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "score": 1.0, + "content": "DYNAMIC NEURAL TURING MACHINE WITH CONTIN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 444, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 444, + 116 + ], + "score": 1.0, + "content": "UOUS AND DISCRETE ADDRESSING SCHEMES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 116, + 133, + 420, + 146 + ], + "lines": [ + { + "bbox": [ + 114, + 132, + 422, + 149 + ], + "spans": [ + { + "bbox": [ + 114, + 132, + 422, + 149 + ], + "score": 1.0, + "content": "Caglar Gulcehre∗, Sarath Chandar∗, Kyunghyun Cho†, Yoshua Bengio∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 114, + 147, + 357, + 167 + ], + "lines": [ + { + "bbox": [ + 112, + 144, + 359, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 144, + 359, + 158 + ], + "score": 1.0, + "content": "∗ University of Montreal, name.lastname@umontreal.ca", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 155, + 324, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 155, + 324, + 169 + ], + "score": 1.0, + "content": "† New York University, name.lastname@nyu.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 278, + 196, + 333, + 208 + ], + "lines": [ + { + "bbox": [ + 276, + 195, + 336, + 210 + ], + "spans": [ + { + "bbox": [ + 276, + 195, + 336, + 210 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 221, + 468, + 336 + ], + "lines": [ + { + "bbox": [ + 141, + 220, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 220, + 470, + 235 + ], + "score": 1.0, + "content": "In this paper, we extend neural Turing machine (NTM) into a dynamic neural Turing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "score": 1.0, + "content": "machine (D-NTM) by introducing a trainable memory addressing scheme. This", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 242, + 470, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 470, + 254 + ], + "score": 1.0, + "content": "addressing scheme maintains for each memory cell two separate vectors, content and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "score": 1.0, + "content": "address vectors. This allows the D-NTM to learn a wide variety of location-based", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "addressing strategies including both linear and nonlinear ones. We implement", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 273, + 470, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 470, + 285 + ], + "score": 1.0, + "content": "the D-NTM with both continuous, differentiable and discrete, non-differentiable", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 284, + 469, + 296 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 469, + 296 + ], + "score": 1.0, + "content": "read/write mechanisms. We investigate the mechanisms and effects for learning to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 294, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 470, + 306 + ], + "score": 1.0, + "content": "read and write to a memory through experiments on Facebook bAbI tasks using both", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 304, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 469, + 316 + ], + "score": 1.0, + "content": "a feedforward and GRU-controller. The D-NTM is evaluated on a set of Facebook", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "score": 1.0, + "content": "bAbI tasks and shown to outperform NTM and LSTM baselines. We also provide", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 325, + 470, + 338 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 338 + ], + "score": 1.0, + "content": "further experimental results on sequential MNIST, associative recall and copy tasks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 359, + 206, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 208, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 208, + 375 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "Designing general-purpose learning algorithms is one of the long-standing goals of artificial intelligence.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "Despite the success of deep learning in this area (see, e.g., (Goodfellow et al., 2016)) there are still a set", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 405, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 506, + 417 + ], + "score": 1.0, + "content": "of complex tasks that are not well addressed by conventional neural networks. Those tasks often require a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 416, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 506, + 428 + ], + "score": 1.0, + "content": "neural network to be equipped with an explicit, external memory in which a larger, potentially unbounded,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "set of facts need to be stored. They include, but are not limited to, episodic question-answering (Weston", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "et al., 2015b; Hermann et al., 2015; Hill et al., 2015), compact algorithms (Zaremba et al., 2015),", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 504, + 460 + ], + "score": 1.0, + "content": "dialogue (Serban et al., 2016; Vinyals & Le, 2015) and video caption generation (Yao et al., 2015).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "Recently two promising approaches based on neural networks to this type of tasks have been proposed.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "Memory networks (Weston et al., 2015b) explicitly store all the facts, or information, available for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "each episode in an external memory (as continuous vectors) and use the attention-based mechanism", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "to index them when returning an output. On the other hand, neural Turing machines (NTM, (Graves", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 505, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 517 + ], + "score": 1.0, + "content": "et al., 2014)) read each fact in an episode and decides whether to read, write the fact or do both to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 514, + 251, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 251, + 529 + ], + "score": 1.0, + "content": "the external, differentiable memory.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "A crucial difference between these two models is that the memory network does not have a mechanism", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 542, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 506, + 554 + ], + "score": 1.0, + "content": "to modify the content of the external memory, while the NTM does. In practice, this leads to easier", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "learning in the memory network, which in turn resulted in it being used more in real tasks (Bordes et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "2015; Dodge et al., 2015). On the contrary, the NTM has mainly been tested on a series of small-scale,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "carefully-crafted tasks such as copy and associative recall. The NTM, however is more expressive,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 507, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 507, + 597 + ], + "score": 1.0, + "content": "precisely because it can store and modify the internal state of the network as it processes an episode.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "The original NTM supports two modes of addressing (which can be used simultaneously.) They are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "content-based and location-based addressing. We notice that the location-based strategy is based on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 622, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 506, + 633 + ], + "score": 1.0, + "content": "linear addressing. The distance between each pair of consecutive memory cells is fixed to a constant.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "We address this limitation, in this paper, by introducing a learnable address vector for each memory", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 642, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 654 + ], + "score": 1.0, + "content": "cell of the NTM with least recently used memory addressing mechanism, and we call this variant a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 652, + 280, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 280, + 665 + ], + "score": 1.0, + "content": "dynamic neural Turing machine (D-NTM).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "We evaluate the proposed D-NTM on the full set of Facebook bAbI task (Weston et al., 2015b)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "score": 1.0, + "content": "using either continuous, differentiable attention or discrete, non-differentiable attention (Zaremba &", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 507, + 703 + ], + "score": 1.0, + "content": "Sutskever, 2015) as an addressing strategy. Our experiments reveal that it is possible to use the discrete,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "non-differentiable attention mechanism, and in fact, the D-NTM with the discrete attention and GRU", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "controller outperforms the one with the continuous attention. After we published our paper on arXiv, a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "new extension of NTM called DNC (Graves et al., 2016) has also provided results on bAbI task as well.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 505, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 507, + 97 + ], + "score": 1.0, + "content": "DYNAMIC NEURAL TURING MACHINE WITH CONTIN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 444, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 444, + 116 + ], + "score": 1.0, + "content": "UOUS AND DISCRETE ADDRESSING SCHEMES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 116, + 133, + 420, + 146 + ], + "lines": [ + { + "bbox": [ + 114, + 132, + 422, + 149 + ], + "spans": [ + { + "bbox": [ + 114, + 132, + 422, + 149 + ], + "score": 1.0, + "content": "Caglar Gulcehre∗, Sarath Chandar∗, Kyunghyun Cho†, Yoshua Bengio∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 114, + 132, + 422, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 114, + 147, + 357, + 167 + ], + "lines": [ + { + "bbox": [ + 112, + 144, + 359, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 144, + 359, + 158 + ], + "score": 1.0, + "content": "∗ University of Montreal, name.lastname@umontreal.ca", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 155, + 324, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 155, + 324, + 169 + ], + "score": 1.0, + "content": "† New York University, name.lastname@nyu.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 112, + 144, + 359, + 169 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 196, + 333, + 208 + ], + "lines": [ + { + "bbox": [ + 276, + 195, + 336, + 210 + ], + "spans": [ + { + "bbox": [ + 276, + 195, + 336, + 210 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 221, + 468, + 336 + ], + "lines": [ + { + "bbox": [ + 141, + 220, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 220, + 470, + 235 + ], + "score": 1.0, + "content": "In this paper, we extend neural Turing machine (NTM) into a dynamic neural Turing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 231, + 470, + 244 + ], + "score": 1.0, + "content": "machine (D-NTM) by introducing a trainable memory addressing scheme. This", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 242, + 470, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 470, + 254 + ], + "score": 1.0, + "content": "addressing scheme maintains for each memory cell two separate vectors, content and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "score": 1.0, + "content": "address vectors. This allows the D-NTM to learn a wide variety of location-based", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 275 + ], + "score": 1.0, + "content": "addressing strategies including both linear and nonlinear ones. We implement", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 273, + 470, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 470, + 285 + ], + "score": 1.0, + "content": "the D-NTM with both continuous, differentiable and discrete, non-differentiable", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 284, + 469, + 296 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 469, + 296 + ], + "score": 1.0, + "content": "read/write mechanisms. We investigate the mechanisms and effects for learning to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 294, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 470, + 306 + ], + "score": 1.0, + "content": "read and write to a memory through experiments on Facebook bAbI tasks using both", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 304, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 469, + 316 + ], + "score": 1.0, + "content": "a feedforward and GRU-controller. The D-NTM is evaluated on a set of Facebook", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "score": 1.0, + "content": "bAbI tasks and shown to outperform NTM and LSTM baselines. We also provide", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 325, + 470, + 338 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 338 + ], + "score": 1.0, + "content": "further experimental results on sequential MNIST, associative recall and copy tasks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 220, + 470, + 338 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 359, + 206, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 208, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 208, + 375 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "Designing general-purpose learning algorithms is one of the long-standing goals of artificial intelligence.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "Despite the success of deep learning in this area (see, e.g., (Goodfellow et al., 2016)) there are still a set", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 405, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 506, + 417 + ], + "score": 1.0, + "content": "of complex tasks that are not well addressed by conventional neural networks. Those tasks often require a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 416, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 506, + 428 + ], + "score": 1.0, + "content": "neural network to be equipped with an explicit, external memory in which a larger, potentially unbounded,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "set of facts need to be stored. They include, but are not limited to, episodic question-answering (Weston", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "et al., 2015b; Hermann et al., 2015; Hill et al., 2015), compact algorithms (Zaremba et al., 2015),", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 504, + 460 + ], + "score": 1.0, + "content": "dialogue (Serban et al., 2016; Vinyals & Le, 2015) and video caption generation (Yao et al., 2015).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 384, + 506, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "Recently two promising approaches based on neural networks to this type of tasks have been proposed.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "Memory networks (Weston et al., 2015b) explicitly store all the facts, or information, available for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "each episode in an external memory (as continuous vectors) and use the attention-based mechanism", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "to index them when returning an output. On the other hand, neural Turing machines (NTM, (Graves", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 505, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 517 + ], + "score": 1.0, + "content": "et al., 2014)) read each fact in an episode and decides whether to read, write the fact or do both to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 514, + 251, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 251, + 529 + ], + "score": 1.0, + "content": "the external, differentiable memory.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 463, + 506, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "A crucial difference between these two models is that the memory network does not have a mechanism", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 542, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 506, + 554 + ], + "score": 1.0, + "content": "to modify the content of the external memory, while the NTM does. In practice, this leads to easier", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "learning in the memory network, which in turn resulted in it being used more in real tasks (Bordes et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "2015; Dodge et al., 2015). On the contrary, the NTM has mainly been tested on a series of small-scale,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "carefully-crafted tasks such as copy and associative recall. The NTM, however is more expressive,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 507, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 507, + 597 + ], + "score": 1.0, + "content": "precisely because it can store and modify the internal state of the network as it processes an episode.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 531, + 507, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "The original NTM supports two modes of addressing (which can be used simultaneously.) They are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "content-based and location-based addressing. We notice that the location-based strategy is based on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 622, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 506, + 633 + ], + "score": 1.0, + "content": "linear addressing. The distance between each pair of consecutive memory cells is fixed to a constant.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "We address this limitation, in this paper, by introducing a learnable address vector for each memory", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 642, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 654 + ], + "score": 1.0, + "content": "cell of the NTM with least recently used memory addressing mechanism, and we call this variant a", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 652, + 280, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 280, + 665 + ], + "score": 1.0, + "content": "dynamic neural Turing machine (D-NTM).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 599, + 506, + 665 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "We evaluate the proposed D-NTM on the full set of Facebook bAbI task (Weston et al., 2015b)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 691 + ], + "score": 1.0, + "content": "using either continuous, differentiable attention or discrete, non-differentiable attention (Zaremba &", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 507, + 703 + ], + "score": 1.0, + "content": "Sutskever, 2015) as an addressing strategy. Our experiments reveal that it is possible to use the discrete,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "non-differentiable attention mechanism, and in fact, the D-NTM with the discrete attention and GRU", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "controller outperforms the one with the continuous attention. After we published our paper on arXiv, a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "new extension of NTM called DNC (Graves et al., 2016) has also provided results on bAbI task as well.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 668, + 507, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "We also provide results on sequential-MNIST and algorithmic tasks proposed by (Graves et al., 2014)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 468, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 468, + 106 + ], + "score": 1.0, + "content": "in order to investigate the ability of our model when dealing with long-term dependencies.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 115, + 188, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 115, + 189, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 189, + 129 + ], + "score": 1.0, + "content": "Our Contributions", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 506, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "1. We propose a generalization of Neural Turing Machine called a dynamic neural Turing machine", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 141, + 394, + 157 + ], + "spans": [ + { + "bbox": [ + 117, + 141, + 394, + 157 + ], + "score": 1.0, + "content": "(D-NTM) which employs a learnable and location-based addressing.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 154, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 117, + 164 + ], + "score": 1.0, + "content": "2.", + "type": "text" + }, + { + "bbox": [ + 118, + 154, + 506, + 166 + ], + "score": 1.0, + "content": "We demonstrate the application of neural Turing machines on a more natural and less toyish task:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 118, + 164, + 505, + 176 + ], + "score": 1.0, + "content": "episodic question-answering besides the toy tasks. We provide detailed analysis of our model on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 174, + 156, + 186 + ], + "spans": [ + { + "bbox": [ + 118, + 174, + 156, + 186 + ], + "score": 1.0, + "content": "this task.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 504, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 504, + 197 + ], + "score": 1.0, + "content": "3. We propose to use the discrete attention mechanism and empirically show that, it can outperform", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 195, + 375, + 207 + ], + "spans": [ + { + "bbox": [ + 118, + 195, + 375, + 207 + ], + "score": 1.0, + "content": "the continuous attention based addressing for episodic QA task.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 117, + 216 + ], + "score": 1.0, + "content": "4.", + "type": "text" + }, + { + "bbox": [ + 118, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "We propose a curriculum strategy for our model with the feedforward controller and discrete", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 216, + 311, + 229 + ], + "spans": [ + { + "bbox": [ + 118, + 216, + 311, + 229 + ], + "score": 1.0, + "content": "attention that improves our results significantly.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 242, + 322, + 255 + ], + "lines": [ + { + "bbox": [ + 104, + 241, + 323, + 257 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 323, + 257 + ], + "score": 1.0, + "content": "2 DYNAMIC NEURAL TURING MACHINE", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "score": 1.0, + "content": "The proposed dynamic neural Turing machine (D-NTM) extends the neural Turing machine (NTM,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "score": 1.0, + "content": "(Graves et al., 2014)) which has a modular design. The NTM consists of two main modules, a controller", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 288, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 506, + 300 + ], + "score": 1.0, + "content": "and, a memory. The controller, which is often implemented as a recurrent neural network, issues a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "command to the memory so as to read, write to and erase a subset of memory cells. Although the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "memory was originally envisioned as an integrated module, it is not necessary, and the memory may", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 320, + 331, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 331, + 331 + ], + "score": 1.0, + "content": "be an external, black box (Zaremba & Sutskever, 2015).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 107, + 343, + 190, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 192, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 192, + 356 + ], + "score": 1.0, + "content": "2.1 CONTROLLER", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 504, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 174, + 376 + ], + "score": 1.0, + "content": "At each time step", + "type": "text" + }, + { + "bbox": [ + 174, + 365, + 179, + 374 + ], + "score": 0.61, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 363, + 336, + 376 + ], + "score": 1.0, + "content": ", the controller (1) receives an input value", + "type": "text" + }, + { + "bbox": [ + 337, + 363, + 347, + 374 + ], + "score": 0.85, + "content": "\\mathbf { x } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 363, + 505, + 376 + ], + "score": 1.0, + "content": ", (2) addresses and reads the memory and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 205, + 387 + ], + "score": 1.0, + "content": "creates the content vector", + "type": "text" + }, + { + "bbox": [ + 206, + 374, + 216, + 385 + ], + "score": 0.85, + "content": "\\phi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 373, + 506, + 387 + ], + "score": 1.0, + "content": ", (3) erases/writes a portion of the memory, (4) updates its own hidden state", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 117, + 396 + ], + "score": 0.86, + "content": "\\mathbf { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 384, + 211, + 396 + ], + "score": 1.0, + "content": ", and (5) outputs a value", + "type": "text" + }, + { + "bbox": [ + 212, + 385, + 222, + 396 + ], + "score": 0.87, + "content": "\\mathbf { y } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "(if needed.) In this paper, we use both a gated recurrent unit (GRU, (Cho", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "et al., 2014)) and a feedforward-controller to implement the controller such that for a GRU controller", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 411, + 357, + 426 + ], + "lines": [ + { + "bbox": [ + 253, + 411, + 357, + 426 + ], + "spans": [ + { + "bbox": [ + 253, + 411, + 357, + 426 + ], + "score": 0.93, + "content": "\\mathbf { h } ^ { t } = \\mathbf { G } \\mathbf { R } \\mathbf { U } ( \\mathbf { x } ^ { t } , \\mathbf { h } ^ { t - 1 } , \\boldsymbol { \\phi } ^ { t } )", + "type": "interline_equation", + "image_path": "4cf31fbf6dbc86933cb0ad73eeb8188150e039682e5c9e701f3ad20f27d6c9af.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 253, + 411, + 357, + 426 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 231, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 231, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 231, + 443 + ], + "score": 1.0, + "content": "or for a feedforward-controller", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 272, + 446, + 339, + 461 + ], + "lines": [ + { + "bbox": [ + 272, + 446, + 339, + 461 + ], + "spans": [ + { + "bbox": [ + 272, + 446, + 339, + 461 + ], + "score": 0.92, + "content": "\\mathbf { h } ^ { t } = \\sigma ( \\mathbf { x } ^ { t } , \\phi ^ { t } ) .", + "type": "interline_equation", + "image_path": "5de065f223a91cdf87a45c3509c4aedeb7249eabc03bdc16c64b06a19366a234.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 272, + 446, + 339, + 461 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 473, + 172, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 173, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 173, + 486 + ], + "score": 1.0, + "content": "2.2 MEMORY", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 491, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 218, + 506 + ], + "score": 1.0, + "content": "We use a rectangular matrix", + "type": "text" + }, + { + "bbox": [ + 218, + 491, + 292, + 503 + ], + "score": 0.92, + "content": "\\mathbf { M } \\in \\mathbb { R } ^ { N \\times ( d _ { c } + d _ { a } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 489, + 333, + 506 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + }, + { + "bbox": [ + 333, + 493, + 343, + 502 + ], + "score": 0.84, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 489, + 506, + 506 + ], + "score": 1.0, + "content": "memory cells. Unlike the original NTM,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 502, + 317, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 317, + 516 + ], + "score": 1.0, + "content": "we partition each memory cell vector into two parts:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "interline_equation", + "bbox": [ + 276, + 519, + 335, + 533 + ], + "lines": [ + { + "bbox": [ + 276, + 519, + 335, + 533 + ], + "spans": [ + { + "bbox": [ + 276, + 519, + 335, + 533 + ], + "score": 0.92, + "content": "\\mathbf { M } = \\left[ \\mathbf { A } ; \\mathbf { C } \\right] .", + "type": "interline_equation", + "image_path": "befc9cc2f17dd794ef5948d1e6e6deee3276d9b7c5d860e7b70f86a8f3d13223.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 276, + 519, + 335, + 533 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 538, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 104, + 536, + 508, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 159, + 553 + ], + "score": 1.0, + "content": "The first part", + "type": "text" + }, + { + "bbox": [ + 159, + 538, + 210, + 550 + ], + "score": 0.92, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { N \\times d _ { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 536, + 387, + 553 + ], + "score": 1.0, + "content": "is a learnable address matrix, and the second", + "type": "text" + }, + { + "bbox": [ + 387, + 538, + 437, + 550 + ], + "score": 0.92, + "content": "\\mathbf { C } \\in \\mathbb { R } ^ { N \\times d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 536, + 508, + 553 + ], + "score": 1.0, + "content": "a content matrix.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 549, + 286, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 243, + 562 + ], + "score": 1.0, + "content": "In other words, each memory cell", + "type": "text" + }, + { + "bbox": [ + 243, + 551, + 257, + 561 + ], + "score": 0.88, + "content": "\\mathbf { m } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 549, + 286, + 562 + ], + "score": 1.0, + "content": "is now", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 566, + 335, + 579 + ], + "lines": [ + { + "bbox": [ + 275, + 566, + 335, + 579 + ], + "spans": [ + { + "bbox": [ + 275, + 566, + 335, + 579 + ], + "score": 0.91, + "content": "\\mathbf { m } _ { i } = \\left[ \\mathbf { a } _ { i } ; \\mathbf { c } _ { i } \\right] .", + "type": "interline_equation", + "image_path": "21127d76e3d09eb46b3c25f154ce493fbf16c81d5f1b19c4e682ad1bb66b36b8.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 275, + 566, + 335, + 579 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 506, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 172, + 604 + ], + "score": 1.0, + "content": "The address part", + "type": "text" + }, + { + "bbox": [ + 173, + 592, + 183, + 601 + ], + "score": 0.87, + "content": "{ \\bf a } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 589, + 507, + 604 + ], + "score": 1.0, + "content": "is considered a model parameter that is updated during training. During inference,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "the address part is not overwritten by the controller and remains constant. On the other hand, the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 156, + 623 + ], + "score": 1.0, + "content": "content part", + "type": "text" + }, + { + "bbox": [ + 157, + 613, + 167, + 622 + ], + "score": 0.86, + "content": "\\mathbf { c } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "is both read and written by the controller both during training and inference. At the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 470, + 633 + ], + "score": 1.0, + "content": "beginning of each episode, the content part of the memory is refreshed to be an all-zero matrix,", + "type": "text" + }, + { + "bbox": [ + 470, + 622, + 503, + 632 + ], + "score": 0.9, + "content": "\\mathbf { C } ^ { 0 } = \\mathbf { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 504, + 622, + 506, + 633 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "This introduction of the learnable address portion for each memory cell allows the model to learn", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "sophisticated location-based addressing strategies. A similar addressing mechanism is also explored", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 653, + 390, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 390, + 666 + ], + "score": 1.0, + "content": "in (Reed & de Freitas, 2015) in the context of learning program traces.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 231, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 232, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 232, + 690 + ], + "score": 1.0, + "content": "2.3 MEMORY ADDRESSING", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 697, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 507, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 357, + 711 + ], + "score": 1.0, + "content": "Memory addressing in the D-NTM is equivalent to computing an", + "type": "text" + }, + { + "bbox": [ + 357, + 698, + 367, + 707 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 696, + 507, + 711 + ], + "score": 1.0, + "content": "-dimensional address vector. The D-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 707, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 104, + 707, + 327, + 722 + ], + "score": 1.0, + "content": "NTM computes three such vectors for respectively reading", + "type": "text" + }, + { + "bbox": [ + 328, + 708, + 367, + 719 + ], + "score": 0.9, + "content": "\\mathbf { w } ^ { t } \\in \\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 707, + 399, + 722 + ], + "score": 1.0, + "content": ", erasing", + "type": "text" + }, + { + "bbox": [ + 400, + 708, + 437, + 720 + ], + "score": 0.92, + "content": "\\mathbf { e } ^ { t } \\in \\mathbb { R } ^ { d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 707, + 483, + 722 + ], + "score": 1.0, + "content": "and writing", + "type": "text" + }, + { + "bbox": [ + 483, + 708, + 505, + 720 + ], + "score": 0.88, + "content": "\\mathbf { u } ^ { t } \\in", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 122, + 730 + ], + "score": 0.87, + "content": "\\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 719, + 484, + 734 + ], + "score": 1.0, + "content": ". Specifically for writing, the controller further computes a candidate memory content vector", + "type": "text" + }, + { + "bbox": [ + 485, + 720, + 505, + 731 + ], + "score": 0.86, + "content": "\\bar { \\mathbf { c } } ^ { t } \\in", + "type": "inline_equation" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "We also provide results on sequential-MNIST and algorithmic tasks proposed by (Graves et al., 2014)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 468, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 468, + 106 + ], + "score": 1.0, + "content": "in order to investigate the ability of our model when dealing with long-term dependencies.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 506, + 106 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 115, + 188, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 115, + 189, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 189, + 129 + ], + "score": 1.0, + "content": "Our Contributions", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "list", + "bbox": [ + 106, + 132, + 506, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "1. We propose a generalization of Neural Turing Machine called a dynamic neural Turing machine", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 141, + 394, + 157 + ], + "spans": [ + { + "bbox": [ + 117, + 141, + 394, + 157 + ], + "score": 1.0, + "content": "(D-NTM) which employs a learnable and location-based addressing.", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 154, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 117, + 164 + ], + "score": 1.0, + "content": "2.", + "type": "text" + }, + { + "bbox": [ + 118, + 154, + 506, + 166 + ], + "score": 1.0, + "content": "We demonstrate the application of neural Turing machines on a more natural and less toyish task:", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 118, + 164, + 505, + 176 + ], + "score": 1.0, + "content": "episodic question-answering besides the toy tasks. We provide detailed analysis of our model on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 174, + 156, + 186 + ], + "spans": [ + { + "bbox": [ + 118, + 174, + 156, + 186 + ], + "score": 1.0, + "content": "this task.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 184, + 504, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 504, + 197 + ], + "score": 1.0, + "content": "3. We propose to use the discrete attention mechanism and empirically show that, it can outperform", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 195, + 375, + 207 + ], + "spans": [ + { + "bbox": [ + 118, + 195, + 375, + 207 + ], + "score": 1.0, + "content": "the continuous attention based addressing for episodic QA task.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 117, + 216 + ], + "score": 1.0, + "content": "4.", + "type": "text" + }, + { + "bbox": [ + 118, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "We propose a curriculum strategy for our model with the feedforward controller and discrete", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 216, + 311, + 229 + ], + "spans": [ + { + "bbox": [ + 118, + 216, + 311, + 229 + ], + "score": 1.0, + "content": "attention that improves our results significantly.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + } + ], + "index": 7, + "bbox_fs": [ + 106, + 132, + 506, + 229 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 242, + 322, + 255 + ], + "lines": [ + { + "bbox": [ + 104, + 241, + 323, + 257 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 323, + 257 + ], + "score": 1.0, + "content": "2 DYNAMIC NEURAL TURING MACHINE", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "score": 1.0, + "content": "The proposed dynamic neural Turing machine (D-NTM) extends the neural Turing machine (NTM,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 290 + ], + "score": 1.0, + "content": "(Graves et al., 2014)) which has a modular design. The NTM consists of two main modules, a controller", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 288, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 506, + 300 + ], + "score": 1.0, + "content": "and, a memory. The controller, which is often implemented as a recurrent neural network, issues a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "command to the memory so as to read, write to and erase a subset of memory cells. Although the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "memory was originally envisioned as an integrated module, it is not necessary, and the memory may", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 320, + 331, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 331, + 331 + ], + "score": 1.0, + "content": "be an external, black box (Zaremba & Sutskever, 2015).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 267, + 506, + 331 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 343, + 190, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 192, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 192, + 356 + ], + "score": 1.0, + "content": "2.1 CONTROLLER", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 504, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 174, + 376 + ], + "score": 1.0, + "content": "At each time step", + "type": "text" + }, + { + "bbox": [ + 174, + 365, + 179, + 374 + ], + "score": 0.61, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 363, + 336, + 376 + ], + "score": 1.0, + "content": ", the controller (1) receives an input value", + "type": "text" + }, + { + "bbox": [ + 337, + 363, + 347, + 374 + ], + "score": 0.85, + "content": "\\mathbf { x } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 363, + 505, + 376 + ], + "score": 1.0, + "content": ", (2) addresses and reads the memory and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 205, + 387 + ], + "score": 1.0, + "content": "creates the content vector", + "type": "text" + }, + { + "bbox": [ + 206, + 374, + 216, + 385 + ], + "score": 0.85, + "content": "\\phi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 373, + 506, + 387 + ], + "score": 1.0, + "content": ", (3) erases/writes a portion of the memory, (4) updates its own hidden state", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 117, + 396 + ], + "score": 0.86, + "content": "\\mathbf { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 384, + 211, + 396 + ], + "score": 1.0, + "content": ", and (5) outputs a value", + "type": "text" + }, + { + "bbox": [ + 212, + 385, + 222, + 396 + ], + "score": 0.87, + "content": "\\mathbf { y } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "(if needed.) In this paper, we use both a gated recurrent unit (GRU, (Cho", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "et al., 2014)) and a feedforward-controller to implement the controller such that for a GRU controller", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 363, + 506, + 407 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 411, + 357, + 426 + ], + "lines": [ + { + "bbox": [ + 253, + 411, + 357, + 426 + ], + "spans": [ + { + "bbox": [ + 253, + 411, + 357, + 426 + ], + "score": 0.93, + "content": "\\mathbf { h } ^ { t } = \\mathbf { G } \\mathbf { R } \\mathbf { U } ( \\mathbf { x } ^ { t } , \\mathbf { h } ^ { t - 1 } , \\boldsymbol { \\phi } ^ { t } )", + "type": "interline_equation", + "image_path": "4cf31fbf6dbc86933cb0ad73eeb8188150e039682e5c9e701f3ad20f27d6c9af.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 253, + 411, + 357, + 426 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 231, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 231, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 231, + 443 + ], + "score": 1.0, + "content": "or for a feedforward-controller", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 430, + 231, + 443 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 272, + 446, + 339, + 461 + ], + "lines": [ + { + "bbox": [ + 272, + 446, + 339, + 461 + ], + "spans": [ + { + "bbox": [ + 272, + 446, + 339, + 461 + ], + "score": 0.92, + "content": "\\mathbf { h } ^ { t } = \\sigma ( \\mathbf { x } ^ { t } , \\phi ^ { t } ) .", + "type": "interline_equation", + "image_path": "5de065f223a91cdf87a45c3509c4aedeb7249eabc03bdc16c64b06a19366a234.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 272, + 446, + 339, + 461 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 473, + 172, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 173, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 173, + 486 + ], + "score": 1.0, + "content": "2.2 MEMORY", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 491, + 504, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 218, + 506 + ], + "score": 1.0, + "content": "We use a rectangular matrix", + "type": "text" + }, + { + "bbox": [ + 218, + 491, + 292, + 503 + ], + "score": 0.92, + "content": "\\mathbf { M } \\in \\mathbb { R } ^ { N \\times ( d _ { c } + d _ { a } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 489, + 333, + 506 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + }, + { + "bbox": [ + 333, + 493, + 343, + 502 + ], + "score": 0.84, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 489, + 506, + 506 + ], + "score": 1.0, + "content": "memory cells. Unlike the original NTM,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 502, + 317, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 317, + 516 + ], + "score": 1.0, + "content": "we partition each memory cell vector into two parts:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 489, + 506, + 516 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 276, + 519, + 335, + 533 + ], + "lines": [ + { + "bbox": [ + 276, + 519, + 335, + 533 + ], + "spans": [ + { + "bbox": [ + 276, + 519, + 335, + 533 + ], + "score": 0.92, + "content": "\\mathbf { M } = \\left[ \\mathbf { A } ; \\mathbf { C } \\right] .", + "type": "interline_equation", + "image_path": "befc9cc2f17dd794ef5948d1e6e6deee3276d9b7c5d860e7b70f86a8f3d13223.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 276, + 519, + 335, + 533 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 538, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 104, + 536, + 508, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 159, + 553 + ], + "score": 1.0, + "content": "The first part", + "type": "text" + }, + { + "bbox": [ + 159, + 538, + 210, + 550 + ], + "score": 0.92, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { N \\times d _ { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 536, + 387, + 553 + ], + "score": 1.0, + "content": "is a learnable address matrix, and the second", + "type": "text" + }, + { + "bbox": [ + 387, + 538, + 437, + 550 + ], + "score": 0.92, + "content": "\\mathbf { C } \\in \\mathbb { R } ^ { N \\times d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 536, + 508, + 553 + ], + "score": 1.0, + "content": "a content matrix.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 549, + 286, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 243, + 562 + ], + "score": 1.0, + "content": "In other words, each memory cell", + "type": "text" + }, + { + "bbox": [ + 243, + 551, + 257, + 561 + ], + "score": 0.88, + "content": "\\mathbf { m } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 549, + 286, + 562 + ], + "score": 1.0, + "content": "is now", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 536, + 508, + 562 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 566, + 335, + 579 + ], + "lines": [ + { + "bbox": [ + 275, + 566, + 335, + 579 + ], + "spans": [ + { + "bbox": [ + 275, + 566, + 335, + 579 + ], + "score": 0.91, + "content": "\\mathbf { m } _ { i } = \\left[ \\mathbf { a } _ { i } ; \\mathbf { c } _ { i } \\right] .", + "type": "interline_equation", + "image_path": "21127d76e3d09eb46b3c25f154ce493fbf16c81d5f1b19c4e682ad1bb66b36b8.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 275, + 566, + 335, + 579 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 506, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 172, + 604 + ], + "score": 1.0, + "content": "The address part", + "type": "text" + }, + { + "bbox": [ + 173, + 592, + 183, + 601 + ], + "score": 0.87, + "content": "{ \\bf a } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 589, + 507, + 604 + ], + "score": 1.0, + "content": "is considered a model parameter that is updated during training. During inference,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "the address part is not overwritten by the controller and remains constant. On the other hand, the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 156, + 623 + ], + "score": 1.0, + "content": "content part", + "type": "text" + }, + { + "bbox": [ + 157, + 613, + 167, + 622 + ], + "score": 0.86, + "content": "\\mathbf { c } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "is both read and written by the controller both during training and inference. At the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 470, + 633 + ], + "score": 1.0, + "content": "beginning of each episode, the content part of the memory is refreshed to be an all-zero matrix,", + "type": "text" + }, + { + "bbox": [ + 470, + 622, + 503, + 632 + ], + "score": 0.9, + "content": "\\mathbf { C } ^ { 0 } = \\mathbf { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 504, + 622, + 506, + 633 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "This introduction of the learnable address portion for each memory cell allows the model to learn", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "sophisticated location-based addressing strategies. A similar addressing mechanism is also explored", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 653, + 390, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 390, + 666 + ], + "score": 1.0, + "content": "in (Reed & de Freitas, 2015) in the context of learning program traces.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 589, + 507, + 666 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 231, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 232, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 232, + 690 + ], + "score": 1.0, + "content": "2.3 MEMORY ADDRESSING", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 697, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 507, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 357, + 711 + ], + "score": 1.0, + "content": "Memory addressing in the D-NTM is equivalent to computing an", + "type": "text" + }, + { + "bbox": [ + 357, + 698, + 367, + 707 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 696, + 507, + 711 + ], + "score": 1.0, + "content": "-dimensional address vector. The D-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 707, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 104, + 707, + 327, + 722 + ], + "score": 1.0, + "content": "NTM computes three such vectors for respectively reading", + "type": "text" + }, + { + "bbox": [ + 328, + 708, + 367, + 719 + ], + "score": 0.9, + "content": "\\mathbf { w } ^ { t } \\in \\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 707, + 399, + 722 + ], + "score": 1.0, + "content": ", erasing", + "type": "text" + }, + { + "bbox": [ + 400, + 708, + 437, + 720 + ], + "score": 0.92, + "content": "\\mathbf { e } ^ { t } \\in \\mathbb { R } ^ { d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 707, + 483, + 722 + ], + "score": 1.0, + "content": "and writing", + "type": "text" + }, + { + "bbox": [ + 483, + 708, + 505, + 720 + ], + "score": 0.88, + "content": "\\mathbf { u } ^ { t } \\in", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 122, + 730 + ], + "score": 0.87, + "content": "\\mathbb { R } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 719, + 484, + 734 + ], + "score": 1.0, + "content": ". Specifically for writing, the controller further computes a candidate memory content vector", + "type": "text" + }, + { + "bbox": [ + 485, + 720, + 505, + 731 + ], + "score": 0.86, + "content": "\\bar { \\mathbf { c } } ^ { t } \\in", + "type": "inline_equation" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 104, + 696, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 86, + 499, + 254 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 86, + 499, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 86, + 499, + 254 + ], + "spans": [ + { + "bbox": [ + 114, + 86, + 499, + 254 + ], + "score": 0.962, + "type": "image", + "image_path": "e74c8fe56f94d776f121ec40a19574de3775bfc3a8c2b406e74e9fa626fad5ca.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 86, + 499, + 142.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 142.0, + 499, + 198.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 198.0, + 499, + 254.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 277, + 505, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "Figure 1: A graphical illustration of the proposed dynamic neural Turing machine with the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "recurrent-controller. The controller receives the fact as a continuous vector encoded by a recurrent", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "neural network, computes the read and write weights for addressing the memory. If the D-NTM", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 308, + 462, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 462, + 321 + ], + "score": 1.0, + "content": "automatically detects that a query has been received, it returns an answer and terminates.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 105, + 339, + 504, + 363 + ], + "lines": [ + { + "bbox": [ + 107, + 338, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 107, + 340, + 123, + 350 + ], + "score": 0.88, + "content": "\\mathbb { R } ^ { d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 338, + 307, + 353 + ], + "score": 1.0, + "content": "based on its current hidden state of the controller", + "type": "text" + }, + { + "bbox": [ + 307, + 340, + 346, + 351 + ], + "score": 0.92, + "content": "\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 338, + 505, + 353 + ], + "score": 1.0, + "content": "and the input of the controller scaled with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 349, + 507, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 156, + 363 + ], + "score": 1.0, + "content": "a scalar gate", + "type": "text" + }, + { + "bbox": [ + 156, + 351, + 167, + 361 + ], + "score": 0.88, + "content": "\\alpha ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 349, + 507, + 363 + ], + "score": 1.0, + "content": "which is a function of the hidden state and the input of the controller as well, see Eqn 4.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 366, + 388, + 397 + ], + "lines": [ + { + "bbox": [ + 222, + 366, + 388, + 397 + ], + "spans": [ + { + "bbox": [ + 222, + 366, + 388, + 397 + ], + "score": 0.92, + "content": "\\begin{array} { l } { { \\boldsymbol { \\alpha } } ^ { t } = \\mathbf { f } ( \\mathbf { h } ^ { t } , \\mathbf { x } ^ { t } ) , } \\\\ { { \\bar { \\mathbf { c } } } ^ { t } = \\mathrm { R e L U } ( \\mathbf { W } _ { m } \\mathbf { h } ^ { t } + { \\boldsymbol { \\alpha } } ^ { t } \\mathbf { W } _ { x } \\mathbf { x } ^ { t } + \\mathbf { b _ { m } } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "4cbcbe6a8a47923e844608152db23b8e222b75dd7af8100931817e3aa59284c0.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 366, + 388, + 381.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 222, + 381.5, + 388, + 397.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 407, + 504, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 231, + 420 + ], + "score": 1.0, + "content": "Reading With the read vector", + "type": "text" + }, + { + "bbox": [ + 232, + 408, + 244, + 419 + ], + "score": 0.88, + "content": "\\mathbf { w } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 406, + 406, + 420 + ], + "score": 1.0, + "content": ", the content vector read from the memory", + "type": "text" + }, + { + "bbox": [ + 407, + 407, + 459, + 420 + ], + "score": 0.93, + "content": "\\phi ^ { t } \\in \\mathbb { R } ^ { d _ { a } + d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 406, + 506, + 420 + ], + "score": 1.0, + "content": "is retrieved", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 418, + 120, + 433 + ], + "spans": [ + { + "bbox": [ + 104, + 418, + 120, + 433 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 434, + 344, + 450 + ], + "lines": [ + { + "bbox": [ + 266, + 434, + 344, + 450 + ], + "spans": [ + { + "bbox": [ + 266, + 434, + 344, + 450 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\phi ^ { t } = ( \\mathbf { w } ^ { t } ) ^ { \\top } \\mathbf { M } ^ { t - 1 } , } \\end{array}", + "type": "interline_equation", + "image_path": "89a16942bf9824a4e698cc642095618bfc16c35296ad094137f40ca6dc120238.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 266, + 434, + 344, + 450 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 209, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 211, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 133, + 468 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 455, + 146, + 465 + ], + "score": 0.86, + "content": "\\mathbf { w } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 452, + 211, + 468 + ], + "score": 1.0, + "content": "is a row vector.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 455, + 491 + ], + "score": 1.0, + "content": "Erasing and Writing Given the erase, write and candidate memory content vectors", + "type": "text" + }, + { + "bbox": [ + 455, + 477, + 484, + 491 + ], + "score": 0.86, + "content": "( \\mathbf { e } ^ { t } , u _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 477, + 506, + 491 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 489, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 116, + 501 + ], + "score": 0.83, + "content": "\\bar { \\mathbf { c } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 489, + 475, + 503 + ], + "score": 1.0, + "content": "respectively) generated by a simple MLP conditioned on the hidden state of the controller", + "type": "text" + }, + { + "bbox": [ + 476, + 491, + 487, + 500 + ], + "score": 0.85, + "content": "\\mathbf { h } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 489, + 505, + 503 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 501, + 228, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 228, + 514 + ], + "score": 1.0, + "content": "memory matrix is updated by,", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 516, + 385, + 532 + ], + "lines": [ + { + "bbox": [ + 225, + 516, + 385, + 532 + ], + "spans": [ + { + "bbox": [ + 225, + 516, + 385, + 532 + ], + "score": 0.93, + "content": "\\mathbf { C } ^ { t } [ j ] = ( 1 - \\mathbf { e } ^ { t } u _ { j } ^ { t } ) \\odot \\mathbf { C } ^ { t - 1 } [ j ] + u _ { j } ^ { t } \\bar { \\mathbf { c } } ^ { t } .", + "type": "interline_equation", + "image_path": "15dad782aa04c6aaaed7e3cce13a1a57276f398970a2dd95e9780ae3088c345a.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 225, + 516, + 385, + 532 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 185, + 550 + ], + "score": 1.0, + "content": "where the subscript", + "type": "text" + }, + { + "bbox": [ + 185, + 538, + 191, + 549 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 536, + 203, + 550 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 203, + 537, + 226, + 549 + ], + "score": 0.92, + "content": "\\mathbf { C } ^ { t } [ j ]", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 536, + 274, + 550 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 275, + 538, + 280, + 549 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 536, + 384, + 550 + ], + "score": 1.0, + "content": "-th row of the content part", + "type": "text" + }, + { + "bbox": [ + 385, + 537, + 397, + 547 + ], + "score": 0.88, + "content": "\\mathbf { C } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 536, + 487, + 550 + ], + "score": 1.0, + "content": "of the memory matrix", + "type": "text" + }, + { + "bbox": [ + 487, + 537, + 503, + 547 + ], + "score": 0.88, + "content": "\\mathbf { M } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 536, + 506, + 550 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 559, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "score": 1.0, + "content": "No Operation (NOP) As found in (Joulin & Mikolov, 2015), an additional NOP action might be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "beneficial for the controller not to access the memory once in a while. We model this situation by", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 581, + 498, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 498, + 594 + ], + "score": 1.0, + "content": "designating one memory cell as a NOP cell. Reading or writing from this memory cell is ignored.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 605, + 178, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 178, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 178, + 617 + ], + "score": 1.0, + "content": "2.4 LEARNING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 388, + 638 + ], + "score": 1.0, + "content": "Once the proposed D-NTM is executed, it returns the output distribution", + "type": "text" + }, + { + "bbox": [ + 388, + 626, + 455, + 637 + ], + "score": 0.91, + "content": "p ( \\mathbf { y } | \\mathbf { x } _ { 1 } , \\ldots , \\mathbf { x } _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 625, + 505, + 638 + ], + "score": 1.0, + "content": ". As a result,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 635, + 331, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 331, + 648 + ], + "score": 1.0, + "content": "we define a cost function as the negative log-likelihood:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 651, + 388, + 686 + ], + "lines": [ + { + "bbox": [ + 223, + 651, + 388, + 686 + ], + "spans": [ + { + "bbox": [ + 223, + 651, + 388, + 686 + ], + "score": 0.95, + "content": "C ( \\boldsymbol { \\theta } ) = \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } - \\log p ( \\mathbf { y } ^ { n } | \\mathbf { x } _ { 1 } ^ { n } , \\ldots , \\mathbf { x } _ { T } ^ { n } ) ,", + "type": "interline_equation", + "image_path": "31840458fce11743d0141f914540b793f1f0e2db154629e1af8bd93b447893bd.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 651, + 388, + 668.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 223, + 668.5, + 388, + 686.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 133, + 703 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 690, + 140, + 699 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 689, + 505, + 703 + ], + "score": 1.0, + "content": "is a set of all the parameters. As the proposed D-NTM, just like the original NTM, is fully", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "end-to-end differentiable, we can compute the gradient of this cost function by using backpropagation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and learn the parameters of the model with a gradient-based optimization algorithm, such as stochastic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 721, + 263, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 263, + 732 + ], + "score": 1.0, + "content": "gradient descent, to train it end-to-end.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 86, + 499, + 254 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 86, + 499, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 86, + 499, + 254 + ], + "spans": [ + { + "bbox": [ + 114, + 86, + 499, + 254 + ], + "score": 0.962, + "type": "image", + "image_path": "e74c8fe56f94d776f121ec40a19574de3775bfc3a8c2b406e74e9fa626fad5ca.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 114, + 86, + 499, + 142.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 142.0, + 499, + 198.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 198.0, + 499, + 254.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 277, + 505, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "Figure 1: A graphical illustration of the proposed dynamic neural Turing machine with the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "recurrent-controller. The controller receives the fact as a continuous vector encoded by a recurrent", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "neural network, computes the read and write weights for addressing the memory. If the D-NTM", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 308, + 462, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 462, + 321 + ], + "score": 1.0, + "content": "automatically detects that a query has been received, it returns an answer and terminates.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 105, + 339, + 504, + 363 + ], + "lines": [ + { + "bbox": [ + 107, + 338, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 107, + 340, + 123, + 350 + ], + "score": 0.88, + "content": "\\mathbb { R } ^ { d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 338, + 307, + 353 + ], + "score": 1.0, + "content": "based on its current hidden state of the controller", + "type": "text" + }, + { + "bbox": [ + 307, + 340, + 346, + 351 + ], + "score": 0.92, + "content": "\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 338, + 505, + 353 + ], + "score": 1.0, + "content": "and the input of the controller scaled with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 349, + 507, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 156, + 363 + ], + "score": 1.0, + "content": "a scalar gate", + "type": "text" + }, + { + "bbox": [ + 156, + 351, + 167, + 361 + ], + "score": 0.88, + "content": "\\alpha ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 349, + 507, + 363 + ], + "score": 1.0, + "content": "which is a function of the hidden state and the input of the controller as well, see Eqn 4.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 338, + 507, + 363 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 366, + 388, + 397 + ], + "lines": [ + { + "bbox": [ + 222, + 366, + 388, + 397 + ], + "spans": [ + { + "bbox": [ + 222, + 366, + 388, + 397 + ], + "score": 0.92, + "content": "\\begin{array} { l } { { \\boldsymbol { \\alpha } } ^ { t } = \\mathbf { f } ( \\mathbf { h } ^ { t } , \\mathbf { x } ^ { t } ) , } \\\\ { { \\bar { \\mathbf { c } } } ^ { t } = \\mathrm { R e L U } ( \\mathbf { W } _ { m } \\mathbf { h } ^ { t } + { \\boldsymbol { \\alpha } } ^ { t } \\mathbf { W } _ { x } \\mathbf { x } ^ { t } + \\mathbf { b _ { m } } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "4cbcbe6a8a47923e844608152db23b8e222b75dd7af8100931817e3aa59284c0.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 366, + 388, + 381.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 222, + 381.5, + 388, + 397.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 407, + 504, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 231, + 420 + ], + "score": 1.0, + "content": "Reading With the read vector", + "type": "text" + }, + { + "bbox": [ + 232, + 408, + 244, + 419 + ], + "score": 0.88, + "content": "\\mathbf { w } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 406, + 406, + 420 + ], + "score": 1.0, + "content": ", the content vector read from the memory", + "type": "text" + }, + { + "bbox": [ + 407, + 407, + 459, + 420 + ], + "score": 0.93, + "content": "\\phi ^ { t } \\in \\mathbb { R } ^ { d _ { a } + d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 406, + 506, + 420 + ], + "score": 1.0, + "content": "is retrieved", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 418, + 120, + 433 + ], + "spans": [ + { + "bbox": [ + 104, + 418, + 120, + 433 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 406, + 506, + 433 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 434, + 344, + 450 + ], + "lines": [ + { + "bbox": [ + 266, + 434, + 344, + 450 + ], + "spans": [ + { + "bbox": [ + 266, + 434, + 344, + 450 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\phi ^ { t } = ( \\mathbf { w } ^ { t } ) ^ { \\top } \\mathbf { M } ^ { t - 1 } , } \\end{array}", + "type": "interline_equation", + "image_path": "89a16942bf9824a4e698cc642095618bfc16c35296ad094137f40ca6dc120238.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 266, + 434, + 344, + 450 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 209, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 211, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 133, + 468 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 455, + 146, + 465 + ], + "score": 0.86, + "content": "\\mathbf { w } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 452, + 211, + 468 + ], + "score": 1.0, + "content": "is a row vector.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 452, + 211, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 455, + 491 + ], + "score": 1.0, + "content": "Erasing and Writing Given the erase, write and candidate memory content vectors", + "type": "text" + }, + { + "bbox": [ + 455, + 477, + 484, + 491 + ], + "score": 0.86, + "content": "( \\mathbf { e } ^ { t } , u _ { j } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 477, + 506, + 491 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 489, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 116, + 501 + ], + "score": 0.83, + "content": "\\bar { \\mathbf { c } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 489, + 475, + 503 + ], + "score": 1.0, + "content": "respectively) generated by a simple MLP conditioned on the hidden state of the controller", + "type": "text" + }, + { + "bbox": [ + 476, + 491, + 487, + 500 + ], + "score": 0.85, + "content": "\\mathbf { h } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 489, + 505, + 503 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 501, + 228, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 228, + 514 + ], + "score": 1.0, + "content": "memory matrix is updated by,", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 477, + 506, + 514 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 516, + 385, + 532 + ], + "lines": [ + { + "bbox": [ + 225, + 516, + 385, + 532 + ], + "spans": [ + { + "bbox": [ + 225, + 516, + 385, + 532 + ], + "score": 0.93, + "content": "\\mathbf { C } ^ { t } [ j ] = ( 1 - \\mathbf { e } ^ { t } u _ { j } ^ { t } ) \\odot \\mathbf { C } ^ { t - 1 } [ j ] + u _ { j } ^ { t } \\bar { \\mathbf { c } } ^ { t } .", + "type": "interline_equation", + "image_path": "15dad782aa04c6aaaed7e3cce13a1a57276f398970a2dd95e9780ae3088c345a.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 225, + 516, + 385, + 532 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 185, + 550 + ], + "score": 1.0, + "content": "where the subscript", + "type": "text" + }, + { + "bbox": [ + 185, + 538, + 191, + 549 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 536, + 203, + 550 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 203, + 537, + 226, + 549 + ], + "score": 0.92, + "content": "\\mathbf { C } ^ { t } [ j ]", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 536, + 274, + 550 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 275, + 538, + 280, + 549 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 536, + 384, + 550 + ], + "score": 1.0, + "content": "-th row of the content part", + "type": "text" + }, + { + "bbox": [ + 385, + 537, + 397, + 547 + ], + "score": 0.88, + "content": "\\mathbf { C } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 536, + 487, + 550 + ], + "score": 1.0, + "content": "of the memory matrix", + "type": "text" + }, + { + "bbox": [ + 487, + 537, + 503, + 547 + ], + "score": 0.88, + "content": "\\mathbf { M } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 536, + 506, + 550 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 536, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 559, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 573 + ], + "score": 1.0, + "content": "No Operation (NOP) As found in (Joulin & Mikolov, 2015), an additional NOP action might be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "beneficial for the controller not to access the memory once in a while. We model this situation by", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 581, + 498, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 498, + 594 + ], + "score": 1.0, + "content": "designating one memory cell as a NOP cell. Reading or writing from this memory cell is ignored.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 559, + 506, + 594 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 605, + 178, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 178, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 178, + 617 + ], + "score": 1.0, + "content": "2.4 LEARNING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 388, + 638 + ], + "score": 1.0, + "content": "Once the proposed D-NTM is executed, it returns the output distribution", + "type": "text" + }, + { + "bbox": [ + 388, + 626, + 455, + 637 + ], + "score": 0.91, + "content": "p ( \\mathbf { y } | \\mathbf { x } _ { 1 } , \\ldots , \\mathbf { x } _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 625, + 505, + 638 + ], + "score": 1.0, + "content": ". As a result,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 635, + 331, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 331, + 648 + ], + "score": 1.0, + "content": "we define a cost function as the negative log-likelihood:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 625, + 505, + 648 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 651, + 388, + 686 + ], + "lines": [ + { + "bbox": [ + 223, + 651, + 388, + 686 + ], + "spans": [ + { + "bbox": [ + 223, + 651, + 388, + 686 + ], + "score": 0.95, + "content": "C ( \\boldsymbol { \\theta } ) = \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } - \\log p ( \\mathbf { y } ^ { n } | \\mathbf { x } _ { 1 } ^ { n } , \\ldots , \\mathbf { x } _ { T } ^ { n } ) ,", + "type": "interline_equation", + "image_path": "31840458fce11743d0141f914540b793f1f0e2db154629e1af8bd93b447893bd.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 651, + 388, + 668.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 223, + 668.5, + 388, + 686.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 133, + 703 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 690, + 140, + 699 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 689, + 505, + 703 + ], + "score": 1.0, + "content": "is a set of all the parameters. As the proposed D-NTM, just like the original NTM, is fully", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "end-to-end differentiable, we can compute the gradient of this cost function by using backpropagation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and learn the parameters of the model with a gradient-based optimization algorithm, such as stochastic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 721, + 263, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 263, + 732 + ], + "score": 1.0, + "content": "gradient descent, to train it end-to-end.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 689, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 263, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 265, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 265, + 96 + ], + "score": 1.0, + "content": "3 ADDRESSING MECHANISM", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 217, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 217, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 217, + 118 + ], + "score": 1.0, + "content": "3.1 ADDRESS VECTORS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 125, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "score": 1.0, + "content": "Each of the address vectors (both read and write) is computed in the same way. The way they are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 136, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 147 + ], + "score": 1.0, + "content": "computed are very similar to the content based addressing in (Graves et al., 2014). First, the controller", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 147, + 200, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 200, + 157 + ], + "score": 1.0, + "content": "computes a key vector:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 155, + 344, + 171 + ], + "lines": [ + { + "bbox": [ + 264, + 155, + 344, + 171 + ], + "spans": [ + { + "bbox": [ + 264, + 155, + 344, + 171 + ], + "score": 0.92, + "content": "\\mathbf { k } ^ { t } = \\mathbf { W } _ { k } ^ { \\top } \\mathbf { h } ^ { t } + \\mathbf { b } _ { k } ,", + "type": "interline_equation", + "image_path": "a6ae921a61326ce67cb1bb7a9f328bc76b1cb77437c57f434b7bc8ff830a4043.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 264, + 155, + 344, + 171 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 104, + 169, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 135, + 187 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 172, + 221, + 184 + ], + "score": 0.92, + "content": "\\mathbf { W } _ { k } \\ \\in \\ \\mathbb { R } ^ { N \\times ( d _ { a } + d _ { c } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 169, + 243, + 187 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 172, + 305, + 185 + ], + "score": 0.92, + "content": "\\mathbf { b } _ { k } \\ \\in \\ \\mathbb { R } ^ { d _ { a } + d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 169, + 506, + 187 + ], + "score": 1.0, + "content": "if the read head is being computed, otherwise", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 182, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 164, + 196 + ], + "score": 0.92, + "content": "\\mathbf { W } _ { k } \\in \\mathbb { R } ^ { N \\times d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 182, + 182, + 198 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 182, + 184, + 222, + 196 + ], + "score": 0.9, + "content": "\\mathbf { b } _ { k } \\in \\mathbb { R } ^ { d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 182, + 506, + 198 + ], + "score": 1.0, + "content": "if the write head weights are being computed. They can be the parameters", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 487, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 380, + 209 + ], + "score": 1.0, + "content": "for a specific head (either read or write.) Also, the sharpening factor", + "type": "text" + }, + { + "bbox": [ + 380, + 196, + 420, + 208 + ], + "score": 0.93, + "content": "\\beta _ { t } \\in \\mathbb { R } ^ { \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 195, + 487, + 209 + ], + "score": 1.0, + "content": "is computed as:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 210, + 388, + 243 + ], + "lines": [ + { + "bbox": [ + 222, + 210, + 388, + 243 + ], + "spans": [ + { + "bbox": [ + 222, + 210, + 388, + 243 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\mathrm { s o f t p l u s } ( x ) = \\log ( \\exp ( x ) + 1 ) } } \\\\ { { \\beta _ { t } = \\mathrm { s o f t p l u s } ( \\mathbf { u } _ { \\beta } ^ { \\top } \\mathbf { h } ^ { t } + b _ { \\beta } ) + 1 . } } \\end{array}", + "type": "interline_equation", + "image_path": "6e1477f491f90bacd403cb5066f8a4f9466443d67b7569e458d77982da6a2ae1.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 210, + 388, + 226.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 222, + 226.5, + 388, + 243.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 310, + 258 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 311, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 119, + 258 + ], + "score": 0.84, + "content": "\\mathbf { u } _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 245, + 137, + 260 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 138, + 246, + 148, + 258 + ], + "score": 0.89, + "content": "b _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 245, + 297, + 260 + ], + "score": 1.0, + "content": "are the parameters of the sharpening", + "type": "text" + }, + { + "bbox": [ + 297, + 246, + 307, + 258 + ], + "score": 0.88, + "content": "\\beta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 245, + 311, + 260 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 262, + 268, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 269, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 269, + 277 + ], + "score": 1.0, + "content": "The address vector is then computed by,", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 276, + 345, + 324 + ], + "lines": [ + { + "bbox": [ + 265, + 276, + 345, + 324 + ], + "spans": [ + { + "bbox": [ + 265, + 276, + 345, + 324 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { z _ { i } ^ { t } = \\beta ^ { t } S \\left( \\mathbf { k } ^ { t } , \\mathbf { m } _ { i } ^ { t } \\right) } } \\\\ { { w _ { i } ^ { t } = \\displaystyle \\frac { \\exp ( z _ { i } ^ { t } ) } { \\sum _ { j } \\exp ( z _ { j } ^ { t } ) } , } } \\end{array}", + "type": "interline_equation", + "image_path": "6581f2869a23525f78855075864ae35905f703aaccd08b2b2b4ad6fd12c3d9e4.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 265, + 276, + 345, + 300.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 265, + 300.0, + 345, + 324.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 315, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 316, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 224, + 340 + ], + "score": 1.0, + "content": "where the similarity function", + "type": "text" + }, + { + "bbox": [ + 225, + 326, + 262, + 338 + ], + "score": 0.93, + "content": "S \\in \\mathbb { R } ^ { \\geq 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 326, + 316, + 340 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 342, + 361, + 367 + ], + "lines": [ + { + "bbox": [ + 249, + 342, + 361, + 367 + ], + "spans": [ + { + "bbox": [ + 249, + 342, + 361, + 367 + ], + "score": 0.94, + "content": "S ( \\mathbf { x } , \\mathbf { y } ) = \\frac { \\mathbf { x } \\cdot \\mathbf { y } } { ( | | \\mathbf { x } | | | \\mathbf { y } | | + \\epsilon ) } .", + "type": "interline_equation", + "image_path": "b7e486ae71fe152885ee338690d3d901339abd8eabfc9bf45b88561c8d7f3f13.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 249, + 342, + 361, + 367 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 378, + 244, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 245, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 245, + 389 + ], + "score": 1.0, + "content": "3.2 MULTI-STEP ADDRESSING", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 398, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "At each time-step, controller may require more than one-step for accessing to the memory. The original", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "NTM addresses this by implementing multiple sets of read, erase and write heads. In this paper, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "score": 1.0, + "content": "explore an option of allowing each head to operate more than once at each time step, similar to the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 429, + 451, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 451, + 441 + ], + "score": 1.0, + "content": "multi-hop mechanism from the end-to-end memory network (Sukhbaatar et al., 2015).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 453, + 340, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 340, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 340, + 466 + ], + "score": 1.0, + "content": "3.3 DYNAMIC LEAST RECENTLY USED ADDRESSING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "We introduce a memory addressing schema that can learn to put more emphasis on the least recently", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 484, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 496 + ], + "score": 1.0, + "content": "used (LRU) memory locations. As observed in (Santoro et al., 2016; Rae et al., 2016), we find it easier", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 493, + 352, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 352, + 507 + ], + "score": 1.0, + "content": "to learn the write operations with the use of LRU addressing.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 488, + 523 + ], + "score": 1.0, + "content": "To learn a LRU based addressing, first we compute the exponentially moving averages of the logits", + "type": "text" + }, + { + "bbox": [ + 488, + 511, + 505, + 522 + ], + "score": 0.83, + "content": "\\left( \\mathbf { z } _ { t } \\right)", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 117, + 533 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 117, + 523, + 127, + 532 + ], + "score": 0.69, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 521, + 132, + 533 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 132, + 522, + 226, + 532 + ], + "score": 0.9, + "content": "\\mathbf { v } _ { t } ~ = ~ 0 . 1 \\mathbf { v } _ { t - 1 } + 0 . 9 \\mathbf { z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 521, + 339, + 533 + ], + "score": 1.0, + "content": ". We rescale the accumulated", + "type": "text" + }, + { + "bbox": [ + 340, + 522, + 351, + 532 + ], + "score": 0.87, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 521, + 371, + 533 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 371, + 523, + 381, + 532 + ], + "score": 0.86, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 521, + 505, + 533 + ], + "score": 1.0, + "content": ", such that the controller adjusts", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "the influence of how much previously written memory locations should effect the attention weights", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 281, + 554 + ], + "score": 1.0, + "content": "of a particular time-step. Next, we subtract", + "type": "text" + }, + { + "bbox": [ + 282, + 543, + 293, + 553 + ], + "score": 0.86, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 542, + 316, + 554 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 316, + 543, + 326, + 552 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "in order to reduce the weights of previously", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 244, + 564 + ], + "score": 1.0, + "content": "read or written memory locations.", + "type": "text" + }, + { + "bbox": [ + 245, + 554, + 254, + 563 + ], + "score": 0.85, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "is a shallow MLP with a scalar output and it is conditioned on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 561, + 444, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 240, + 577 + ], + "score": 1.0, + "content": "the hidden state of the controller.", + "type": "text" + }, + { + "bbox": [ + 241, + 564, + 250, + 574 + ], + "score": 0.86, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 561, + 396, + 577 + ], + "score": 1.0, + "content": "is parametrized with the parameters", + "type": "text" + }, + { + "bbox": [ + 397, + 564, + 409, + 575 + ], + "score": 0.88, + "content": "\\mathbf { u } _ { \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 561, + 427, + 577 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 427, + 563, + 440, + 575 + ], + "score": 0.89, + "content": "{ \\bf { b } } _ { \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 561, + 444, + 577 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 577, + 365, + 610 + ], + "lines": [ + { + "bbox": [ + 245, + 577, + 365, + 610 + ], + "spans": [ + { + "bbox": [ + 245, + 577, + 365, + 610 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\gamma _ { t } = \\operatorname { s i g m o i d } ( \\mathbf { u } _ { \\gamma } ^ { \\top } \\mathbf { h } _ { t } + \\mathbf { b } _ { \\gamma } ) , } \\\\ { \\mathbf { w } _ { t } = \\operatorname { s o f t m a x } \\bigl ( \\mathbf { z } _ { t } - \\gamma _ { t } \\mathbf { v } _ { t - 1 } \\bigr ) . } \\end{array}", + "type": "interline_equation", + "image_path": "2fca1f6902ab2cd18ff1732cca4ab7b64bfee10e70afc5c4af35f34445161e14.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 245, + 577, + 365, + 593.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 245, + 593.5, + 365, + 610.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "This addressing method increases the weights of the least recently used rows of the memory. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "magnitude of the influence of the least-recently used memory locations is being learned and adjusted", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 125, + 651 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 125, + 640, + 135, + 650 + ], + "score": 0.85, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 639, + 505, + 651 + ], + "score": 1.0, + "content": ". Our LRU addressing is dynamic due to the model’s ability to switch between pure content-based", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 649, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 409, + 663 + ], + "score": 1.0, + "content": "addressing and LRU. During the training, we do not backpropagate through", + "type": "text" + }, + { + "bbox": [ + 409, + 650, + 420, + 660 + ], + "score": 0.85, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 649, + 506, + 663 + ], + "score": 1.0, + "content": ". Due to the dynamic", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "nature of this addressing mechanism, it can be used for both read and write operations. If needed,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 669, + 440, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 440, + 683 + ], + "score": 1.0, + "content": "the model will automatically learn to disable LRU while reading from the memory.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 107, + 696, + 355, + 709 + ], + "lines": [ + { + "bbox": [ + 104, + 695, + 356, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 695, + 356, + 712 + ], + "score": 1.0, + "content": "4 GENERATING DISCRETE ADDRESS VECTORS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 408, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 409, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 409, + 735 + ], + "score": 1.0, + "content": "In this section, we describe the discrete attention based addressing strategy.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 263, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 265, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 265, + 96 + ], + "score": 1.0, + "content": "3 ADDRESSING MECHANISM", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 105, + 217, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 217, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 217, + 118 + ], + "score": 1.0, + "content": "3.1 ADDRESS VECTORS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 125, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "score": 1.0, + "content": "Each of the address vectors (both read and write) is computed in the same way. The way they are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 136, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 147 + ], + "score": 1.0, + "content": "computed are very similar to the content based addressing in (Graves et al., 2014). First, the controller", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 147, + 200, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 200, + 157 + ], + "score": 1.0, + "content": "computes a key vector:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 124, + 505, + 157 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 155, + 344, + 171 + ], + "lines": [ + { + "bbox": [ + 264, + 155, + 344, + 171 + ], + "spans": [ + { + "bbox": [ + 264, + 155, + 344, + 171 + ], + "score": 0.92, + "content": "\\mathbf { k } ^ { t } = \\mathbf { W } _ { k } ^ { \\top } \\mathbf { h } ^ { t } + \\mathbf { b } _ { k } ,", + "type": "interline_equation", + "image_path": "a6ae921a61326ce67cb1bb7a9f328bc76b1cb77437c57f434b7bc8ff830a4043.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 264, + 155, + 344, + 171 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 104, + 169, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 135, + 187 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 172, + 221, + 184 + ], + "score": 0.92, + "content": "\\mathbf { W } _ { k } \\ \\in \\ \\mathbb { R } ^ { N \\times ( d _ { a } + d _ { c } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 169, + 243, + 187 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 172, + 305, + 185 + ], + "score": 0.92, + "content": "\\mathbf { b } _ { k } \\ \\in \\ \\mathbb { R } ^ { d _ { a } + d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 169, + 506, + 187 + ], + "score": 1.0, + "content": "if the read head is being computed, otherwise", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 182, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 164, + 196 + ], + "score": 0.92, + "content": "\\mathbf { W } _ { k } \\in \\mathbb { R } ^ { N \\times d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 182, + 182, + 198 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 182, + 184, + 222, + 196 + ], + "score": 0.9, + "content": "\\mathbf { b } _ { k } \\in \\mathbb { R } ^ { d _ { c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 182, + 506, + 198 + ], + "score": 1.0, + "content": "if the write head weights are being computed. They can be the parameters", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 487, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 380, + 209 + ], + "score": 1.0, + "content": "for a specific head (either read or write.) Also, the sharpening factor", + "type": "text" + }, + { + "bbox": [ + 380, + 196, + 420, + 208 + ], + "score": 0.93, + "content": "\\beta _ { t } \\in \\mathbb { R } ^ { \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 195, + 487, + 209 + ], + "score": 1.0, + "content": "is computed as:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 104, + 169, + 506, + 209 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 222, + 210, + 388, + 243 + ], + "lines": [ + { + "bbox": [ + 222, + 210, + 388, + 243 + ], + "spans": [ + { + "bbox": [ + 222, + 210, + 388, + 243 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\mathrm { s o f t p l u s } ( x ) = \\log ( \\exp ( x ) + 1 ) } } \\\\ { { \\beta _ { t } = \\mathrm { s o f t p l u s } ( \\mathbf { u } _ { \\beta } ^ { \\top } \\mathbf { h } ^ { t } + b _ { \\beta } ) + 1 . } } \\end{array}", + "type": "interline_equation", + "image_path": "6e1477f491f90bacd403cb5066f8a4f9466443d67b7569e458d77982da6a2ae1.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 222, + 210, + 388, + 226.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 222, + 226.5, + 388, + 243.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 310, + 258 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 311, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 119, + 258 + ], + "score": 0.84, + "content": "\\mathbf { u } _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 245, + 137, + 260 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 138, + 246, + 148, + 258 + ], + "score": 0.89, + "content": "b _ { \\beta }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 245, + 297, + 260 + ], + "score": 1.0, + "content": "are the parameters of the sharpening", + "type": "text" + }, + { + "bbox": [ + 297, + 246, + 307, + 258 + ], + "score": 0.88, + "content": "\\beta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 245, + 311, + 260 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 245, + 311, + 260 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 262, + 268, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 269, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 269, + 277 + ], + "score": 1.0, + "content": "The address vector is then computed by,", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 259, + 269, + 277 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 276, + 345, + 324 + ], + "lines": [ + { + "bbox": [ + 265, + 276, + 345, + 324 + ], + "spans": [ + { + "bbox": [ + 265, + 276, + 345, + 324 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { z _ { i } ^ { t } = \\beta ^ { t } S \\left( \\mathbf { k } ^ { t } , \\mathbf { m } _ { i } ^ { t } \\right) } } \\\\ { { w _ { i } ^ { t } = \\displaystyle \\frac { \\exp ( z _ { i } ^ { t } ) } { \\sum _ { j } \\exp ( z _ { j } ^ { t } ) } , } } \\end{array}", + "type": "interline_equation", + "image_path": "6581f2869a23525f78855075864ae35905f703aaccd08b2b2b4ad6fd12c3d9e4.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 265, + 276, + 345, + 300.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 265, + 300.0, + 345, + 324.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 315, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 316, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 224, + 340 + ], + "score": 1.0, + "content": "where the similarity function", + "type": "text" + }, + { + "bbox": [ + 225, + 326, + 262, + 338 + ], + "score": 0.93, + "content": "S \\in \\mathbb { R } ^ { \\geq 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 326, + 316, + 340 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 326, + 316, + 340 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 342, + 361, + 367 + ], + "lines": [ + { + "bbox": [ + 249, + 342, + 361, + 367 + ], + "spans": [ + { + "bbox": [ + 249, + 342, + 361, + 367 + ], + "score": 0.94, + "content": "S ( \\mathbf { x } , \\mathbf { y } ) = \\frac { \\mathbf { x } \\cdot \\mathbf { y } } { ( | | \\mathbf { x } | | | \\mathbf { y } | | + \\epsilon ) } .", + "type": "interline_equation", + "image_path": "b7e486ae71fe152885ee338690d3d901339abd8eabfc9bf45b88561c8d7f3f13.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 249, + 342, + 361, + 367 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 378, + 244, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 245, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 245, + 389 + ], + "score": 1.0, + "content": "3.2 MULTI-STEP ADDRESSING", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 398, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "At each time-step, controller may require more than one-step for accessing to the memory. The original", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "NTM addresses this by implementing multiple sets of read, erase and write heads. In this paper, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "score": 1.0, + "content": "explore an option of allowing each head to operate more than once at each time step, similar to the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 429, + 451, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 451, + 441 + ], + "score": 1.0, + "content": "multi-hop mechanism from the end-to-end memory network (Sukhbaatar et al., 2015).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 398, + 506, + 441 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 453, + 340, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 340, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 340, + 466 + ], + "score": 1.0, + "content": "3.3 DYNAMIC LEAST RECENTLY USED ADDRESSING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "We introduce a memory addressing schema that can learn to put more emphasis on the least recently", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 484, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 496 + ], + "score": 1.0, + "content": "used (LRU) memory locations. As observed in (Santoro et al., 2016; Rae et al., 2016), we find it easier", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 493, + 352, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 352, + 507 + ], + "score": 1.0, + "content": "to learn the write operations with the use of LRU addressing.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 473, + 506, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 488, + 523 + ], + "score": 1.0, + "content": "To learn a LRU based addressing, first we compute the exponentially moving averages of the logits", + "type": "text" + }, + { + "bbox": [ + 488, + 511, + 505, + 522 + ], + "score": 0.83, + "content": "\\left( \\mathbf { z } _ { t } \\right)", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 117, + 533 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 117, + 523, + 127, + 532 + ], + "score": 0.69, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 521, + 132, + 533 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 132, + 522, + 226, + 532 + ], + "score": 0.9, + "content": "\\mathbf { v } _ { t } ~ = ~ 0 . 1 \\mathbf { v } _ { t - 1 } + 0 . 9 \\mathbf { z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 521, + 339, + 533 + ], + "score": 1.0, + "content": ". We rescale the accumulated", + "type": "text" + }, + { + "bbox": [ + 340, + 522, + 351, + 532 + ], + "score": 0.87, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 521, + 371, + 533 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 371, + 523, + 381, + 532 + ], + "score": 0.86, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 521, + 505, + 533 + ], + "score": 1.0, + "content": ", such that the controller adjusts", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "the influence of how much previously written memory locations should effect the attention weights", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 281, + 554 + ], + "score": 1.0, + "content": "of a particular time-step. Next, we subtract", + "type": "text" + }, + { + "bbox": [ + 282, + 543, + 293, + 553 + ], + "score": 0.86, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 542, + 316, + 554 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 316, + 543, + 326, + 552 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "in order to reduce the weights of previously", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 244, + 564 + ], + "score": 1.0, + "content": "read or written memory locations.", + "type": "text" + }, + { + "bbox": [ + 245, + 554, + 254, + 563 + ], + "score": 0.85, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "is a shallow MLP with a scalar output and it is conditioned on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 561, + 444, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 240, + 577 + ], + "score": 1.0, + "content": "the hidden state of the controller.", + "type": "text" + }, + { + "bbox": [ + 241, + 564, + 250, + 574 + ], + "score": 0.86, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 561, + 396, + 577 + ], + "score": 1.0, + "content": "is parametrized with the parameters", + "type": "text" + }, + { + "bbox": [ + 397, + 564, + 409, + 575 + ], + "score": 0.88, + "content": "\\mathbf { u } _ { \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 561, + 427, + 577 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 427, + 563, + 440, + 575 + ], + "score": 0.89, + "content": "{ \\bf { b } } _ { \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 561, + 444, + 577 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 510, + 505, + 577 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 577, + 365, + 610 + ], + "lines": [ + { + "bbox": [ + 245, + 577, + 365, + 610 + ], + "spans": [ + { + "bbox": [ + 245, + 577, + 365, + 610 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\gamma _ { t } = \\operatorname { s i g m o i d } ( \\mathbf { u } _ { \\gamma } ^ { \\top } \\mathbf { h } _ { t } + \\mathbf { b } _ { \\gamma } ) , } \\\\ { \\mathbf { w } _ { t } = \\operatorname { s o f t m a x } \\bigl ( \\mathbf { z } _ { t } - \\gamma _ { t } \\mathbf { v } _ { t - 1 } \\bigr ) . } \\end{array}", + "type": "interline_equation", + "image_path": "2fca1f6902ab2cd18ff1732cca4ab7b64bfee10e70afc5c4af35f34445161e14.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 245, + 577, + 365, + 593.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 245, + 593.5, + 365, + 610.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "This addressing method increases the weights of the least recently used rows of the memory. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "magnitude of the influence of the least-recently used memory locations is being learned and adjusted", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 125, + 651 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 125, + 640, + 135, + 650 + ], + "score": 0.85, + "content": "\\gamma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 639, + 505, + 651 + ], + "score": 1.0, + "content": ". Our LRU addressing is dynamic due to the model’s ability to switch between pure content-based", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 649, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 409, + 663 + ], + "score": 1.0, + "content": "addressing and LRU. During the training, we do not backpropagate through", + "type": "text" + }, + { + "bbox": [ + 409, + 650, + 420, + 660 + ], + "score": 0.85, + "content": "\\mathbf { v } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 649, + 506, + 663 + ], + "score": 1.0, + "content": ". Due to the dynamic", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "nature of this addressing mechanism, it can be used for both read and write operations. If needed,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 669, + 440, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 440, + 683 + ], + "score": 1.0, + "content": "the model will automatically learn to disable LRU while reading from the memory.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 617, + 506, + 683 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 696, + 355, + 709 + ], + "lines": [ + { + "bbox": [ + 104, + 695, + 356, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 695, + 356, + 712 + ], + "score": 1.0, + "content": "4 GENERATING DISCRETE ADDRESS VECTORS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 408, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 409, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 409, + 735 + ], + "score": 1.0, + "content": "In this section, we describe the discrete attention based addressing strategy.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 718, + 409, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Discrete Addressing Let us use w to denote an address vector (either read, write or erase) at time", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 111, + 103 + ], + "score": 0.5, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 93, + 505, + 105 + ], + "score": 1.0, + "content": ". By definition in Eq. (10), every element in this address vector is positive and sums up to one. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 427, + 117 + ], + "score": 1.0, + "content": "other words, we can treat this vector as the probabilities of a categorical distribution", + "type": "text" + }, + { + "bbox": [ + 427, + 104, + 450, + 115 + ], + "score": 0.9, + "content": "\\mathcal C ( \\mathbf w )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 102, + 471, + 117 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 471, + 104, + 505, + 115 + ], + "score": 0.78, + "content": "\\dim ( \\mathbf { w } )", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 142, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 142, + 126 + ], + "score": 1.0, + "content": "choices:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "interline_equation", + "bbox": [ + 281, + 123, + 328, + 137 + ], + "lines": [ + { + "bbox": [ + 281, + 123, + 328, + 137 + ], + "spans": [ + { + "bbox": [ + 281, + 123, + 328, + 137 + ], + "score": 0.91, + "content": "p ( j ) = w _ { j } ,", + "type": "interline_equation", + "image_path": "a48b43c75935fa5ebf2a39183e9b61d2e7b31453a88848fa1484d5c7e26baacf.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 281, + 123, + 328, + 137 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 137, + 504, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 132, + 149 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 138, + 145, + 149 + ], + "score": 0.87, + "content": "w _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 137, + 169, + 149 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 170, + 138, + 176, + 149 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "-th element of w. We can readily sample from this categorical distribution and form", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 228, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 228, + 159 + ], + "score": 1.0, + "content": "an one-hot vector w˜ such that", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 272, + 156, + 338, + 170 + ], + "lines": [ + { + "bbox": [ + 272, + 156, + 338, + 170 + ], + "spans": [ + { + "bbox": [ + 272, + 156, + 338, + 170 + ], + "score": 0.91, + "content": "\\tilde { w } _ { k } = I ( k = j ) ,", + "type": "interline_equation", + "image_path": "967cc989c6c98b115376ccbbce7e228f0ce69961352ea3475e7e07f47aa4bb10.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 272, + 156, + 338, + 170 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 170, + 298, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 299, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 133, + 183 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 171, + 174, + 182 + ], + "score": 0.92, + "content": "j \\sim \\mathcal { C } ( \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 170, + 194, + 183 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 195, + 171, + 201, + 181 + ], + "score": 0.79, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 170, + 299, + 183 + ], + "score": 1.0, + "content": "is an indicator function.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "Training We use this sampling-based strategy for all the heads during training. This clearly makes", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 203, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 215 + ], + "score": 1.0, + "content": "the use of backpropagation infeasible to compute the gradient, as the sampling procedure is not", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "differentiable. Thus, we use REINFORCE (Williams, 1992) together with the three variance reduction", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 224, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 505, + 237 + ], + "score": 1.0, + "content": "techniques–global baseline, input-dependent baseline and variance normalization– suggested in (Mnih", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 234, + 178, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 178, + 248 + ], + "score": 1.0, + "content": "& Gregor, 2014).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 104, + 250, + 497, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 249, + 497, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 160, + 266 + ], + "score": 1.0, + "content": "Let us define", + "type": "text" + }, + { + "bbox": [ + 160, + 251, + 277, + 263 + ], + "score": 0.93, + "content": "R ( { \\bf x } ) = \\log p ( { \\bf y } | { \\bf x } _ { 1 } , \\ldots , { \\bf x } _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 249, + 497, + 266 + ], + "score": 1.0, + "content": "as a reward. We first center and re-scale the reward by", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 265, + 345, + 293 + ], + "lines": [ + { + "bbox": [ + 265, + 265, + 345, + 293 + ], + "spans": [ + { + "bbox": [ + 265, + 265, + 345, + 293 + ], + "score": 0.95, + "content": "\\tilde { R } ( { \\bf x } ) = \\frac { R ( { \\bf x } ) - b } { \\sqrt { \\sigma ^ { 2 } + \\epsilon } } ,", + "type": "interline_equation", + "image_path": "24dcd1de042ab3bc1efc7f9256856a07e0724973f8589335669d4d208f1218c1.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 265, + 265, + 345, + 279.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 265, + 279.0, + 345, + 293.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 132, + 308 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 295, + 138, + 305 + ], + "score": 0.72, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 294, + 155, + 308 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 297, + 163, + 305 + ], + "score": 0.8, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 294, + 340, + 308 + ], + "score": 1.0, + "content": "is running average and standard deviation of", + "type": "text" + }, + { + "bbox": [ + 341, + 295, + 349, + 305 + ], + "score": 0.82, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 294, + 505, + 308 + ], + "score": 1.0, + "content": ". We can further center it for each input", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 305, + 177, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 114, + 315 + ], + "score": 0.53, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 305, + 177, + 318 + ], + "score": 1.0, + "content": "separately, i.e.,", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 315, + 351, + 330 + ], + "lines": [ + { + "bbox": [ + 258, + 315, + 351, + 330 + ], + "spans": [ + { + "bbox": [ + 258, + 315, + 351, + 330 + ], + "score": 0.92, + "content": "\\tilde { R } ( \\mathbf { x } ) \\gets \\tilde { R } ( \\mathbf { x } ) - b ( \\mathbf { x } ) ,", + "type": "interline_equation", + "image_path": "73ad2c0af30c26dbbee2a055bafa03cbc552876676f97e051deb52d5f0acf207.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 258, + 315, + 351, + 330 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 132, + 342 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 330, + 151, + 342 + ], + "score": 0.91, + "content": "b ( \\mathbf x )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 330, + 367, + 342 + ], + "score": 1.0, + "content": "is computed by a baseline network which takes as input", + "type": "text" + }, + { + "bbox": [ + 368, + 333, + 375, + 340 + ], + "score": 0.65, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 330, + 506, + 342 + ], + "score": 1.0, + "content": "and predicts its estimated reward.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "The baseline network is trained to minimize the Huber loss (Huber, 1964) between the true reward", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 351, + 434, + 366 + ], + "spans": [ + { + "bbox": [ + 107, + 351, + 133, + 365 + ], + "score": 0.93, + "content": "\\tilde { R } ( { \\bf x } ) ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 351, + 236, + 366 + ], + "score": 1.0, + "content": "and the predicted reward", + "type": "text" + }, + { + "bbox": [ + 236, + 352, + 255, + 365 + ], + "score": 0.9, + "content": "b ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 351, + 434, + 366 + ], + "score": 1.0, + "content": ". We use the Huber loss, which is defined by", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 367, + 381, + 395 + ], + "lines": [ + { + "bbox": [ + 228, + 367, + 381, + 395 + ], + "spans": [ + { + "bbox": [ + 228, + 367, + 381, + 395 + ], + "score": 0.94, + "content": "H _ { \\delta } ( x ) = { \\left\\{ \\begin{array} { l l } { x ^ { 2 } } & { { \\mathrm { f o r ~ } } | x | \\leq \\delta , } \\\\ { \\delta ( 2 | x | - \\delta ) , } & { { \\mathrm { o t h e r w i s e } } , } \\end{array} \\right. }", + "type": "interline_equation", + "image_path": "e4f5a1c3730e593305a0c56f0c266c5f41401208bfe1b3dfdf05bc020f65d5dd.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 367, + 381, + 381.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 228, + 381.0, + 381, + 395.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 398, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "score": 1.0, + "content": "due to its robustness. As a further measure to reduce the variance, we regularize the negative entropy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 408, + 502, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 502, + 420 + ], + "score": 1.0, + "content": "of all those category distributions to facilitate a better exploration during training (Xu et al., 2015).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 383, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 383, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 383, + 439 + ], + "score": 1.0, + "content": "Then, the cost function for each training example is approximated as", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 437, + 480, + 524 + ], + "lines": [ + { + "bbox": [ + 177, + 437, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 177, + 437, + 480, + 524 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { C ^ { n } ( \\boldsymbol { \\theta } ) = - \\log p ( \\mathbf { y } | \\mathbf { x } _ { 1 : T } , \\tilde { w } _ { 1 : J } , \\tilde { u } _ { 1 : J } , \\tilde { e } _ { 1 : J } ) } \\\\ & { \\qquad - \\displaystyle \\sum _ { j = 1 } ^ { J } \\tilde { R } ( \\mathbf { x } ^ { n } ) ( \\log p ( \\tilde { w } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { u } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { e } _ { j } | \\mathbf { x } _ { 1 : T } ) ) } \\\\ & { \\qquad - \\lambda _ { H } \\displaystyle \\sum _ { j = 1 } ^ { J } ( \\mathcal { H } ( w _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( u _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( e _ { j } | \\mathbf { x } _ { 1 : T } ) ) . } \\end{array}", + "type": "interline_equation", + "image_path": "bffd054bc17ac384c9d9037ee3996704175726c7a06e6adc5c2450df911bf792.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 177, + 437, + 480, + 466.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 177, + 466.0, + 480, + 495.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 177, + 495.0, + 480, + 524.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 527, + 504, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 132, + 540 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 528, + 140, + 537 + ], + "score": 0.85, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 525, + 274, + 540 + ], + "score": 1.0, + "content": "is the number of addressing steps,", + "type": "text" + }, + { + "bbox": [ + 274, + 528, + 289, + 538 + ], + "score": 0.89, + "content": "\\lambda _ { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 525, + 462, + 540 + ], + "score": 1.0, + "content": "is the entropy regularization coefficient, and", + "type": "text" + }, + { + "bbox": [ + 462, + 528, + 472, + 537 + ], + "score": 0.82, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 156, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 156, + 551 + ], + "score": 1.0, + "content": "the entropy.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 104, + 560, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "Inference Once training is over, we switch to a deterministic strategy. We simply choose an element", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 570, + 421, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 421, + 583 + ], + "score": 1.0, + "content": "of w with the largest value to be the index of the target memory cell, such that", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 584, + 358, + 598 + ], + "lines": [ + { + "bbox": [ + 252, + 584, + 358, + 598 + ], + "spans": [ + { + "bbox": [ + 252, + 584, + 358, + 598 + ], + "score": 0.91, + "content": "\\tilde { w } _ { k } = \\mathbf { I } ( k = \\operatorname { a r g m a x } ( \\mathbf { w } ) ) .", + "type": "interline_equation", + "image_path": "92541e0be0f429b20ffa6a34399d1519d61c50f5dd696ad4ec1c1e4ef7a3cc0d.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 252, + 584, + 358, + 598 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "Curriculum Learning for the Discrete Attention Training discrete attention with feed-forward", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "controller and REINFORCE is challenging. We propose to use a curriculum strategy for training", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 467, + 639 + ], + "score": 1.0, + "content": "with the discrete attention in order to tackle this problem. For each minibatch, we sample", + "type": "text" + }, + { + "bbox": [ + 467, + 629, + 475, + 637 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "from a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 275, + 650 + ], + "score": 1.0, + "content": "binomial distribution with the probability", + "type": "text" + }, + { + "bbox": [ + 275, + 637, + 285, + 649 + ], + "score": 0.59, + "content": "p ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 636, + 288, + 650 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 288, + 637, + 345, + 649 + ], + "score": 0.89, + "content": "\\pi ^ { t } \\sim \\mathrm { B i n } ( p ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 636, + 506, + 650 + ], + "score": 1.0, + "content": ". The model will either use the discrete", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 647, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 647, + 268, + 663 + ], + "score": 1.0, + "content": "or the continuous-attention based on the", + "type": "text" + }, + { + "bbox": [ + 269, + 649, + 279, + 659 + ], + "score": 0.87, + "content": "\\pi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 647, + 429, + 663 + ], + "score": 1.0, + "content": ". We start the training procedure with", + "type": "text" + }, + { + "bbox": [ + 429, + 648, + 459, + 660 + ], + "score": 0.92, + "content": "p ^ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 647, + 506, + 663 + ], + "score": 1.0, + "content": "and during", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 659, + 320, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 154, + 678 + ], + "score": 1.0, + "content": "the training", + "type": "text" + }, + { + "bbox": [ + 155, + 662, + 164, + 675 + ], + "score": 0.87, + "content": "p ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 661, + 272, + 678 + ], + "score": 1.0, + "content": "is annealed to 0 by setting", + "type": "text" + }, + { + "bbox": [ + 272, + 659, + 317, + 678 + ], + "score": 0.93, + "content": "\\begin{array} { r } { p ^ { t } = \\frac { p ^ { 0 } } { \\sqrt { 1 + t } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 661, + 320, + 678 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 210, + 694 + ], + "score": 1.0, + "content": "We can rewrite the weights", + "type": "text" + }, + { + "bbox": [ + 211, + 684, + 224, + 693 + ], + "score": 0.87, + "content": "\\mathbf { w } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 682, + 505, + 694 + ], + "score": 1.0, + "content": "as in Equation 14, where it is expressed as the combination of continuous", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 177, + 705 + ], + "score": 1.0, + "content": "attention weights", + "type": "text" + }, + { + "bbox": [ + 177, + 693, + 190, + 702 + ], + "score": 0.88, + "content": "\\hat { \\mathbf { w } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 692, + 313, + 705 + ], + "score": 1.0, + "content": "and discrete attention weights", + "type": "text" + }, + { + "bbox": [ + 313, + 693, + 326, + 702 + ], + "score": 0.88, + "content": "\\tilde { \\mathbf { w } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 692, + 348, + 705 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 348, + 693, + 358, + 702 + ], + "score": 0.88, + "content": "\\mathit { \\Pi } _ { \\pi ^ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "being a binary variable that chooses", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 703, + 185, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 185, + 714 + ], + "score": 1.0, + "content": "to use one of them,", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 715, + 362, + 730 + ], + "lines": [ + { + "bbox": [ + 249, + 715, + 362, + 730 + ], + "spans": [ + { + "bbox": [ + 249, + 715, + 362, + 730 + ], + "score": 0.92, + "content": "\\mathbf { w } ^ { t } \\pi ^ { t } \\bar { \\mathbf { w } } ^ { t } + ( 1 - \\pi ^ { t } ) \\tilde { \\mathbf { w } } ^ { t } .", + "type": "interline_equation", + "image_path": "40eea2ba87dd6930595aeecbe70dc2633c2eb8c53228b730ad911bb606067a39.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 249, + 715, + 362, + 730 + ], + "spans": [], + "index": 45 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Discrete Addressing Let us use w to denote an address vector (either read, write or erase) at time", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 111, + 103 + ], + "score": 0.5, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 93, + 505, + 105 + ], + "score": 1.0, + "content": ". By definition in Eq. (10), every element in this address vector is positive and sums up to one. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 427, + 117 + ], + "score": 1.0, + "content": "other words, we can treat this vector as the probabilities of a categorical distribution", + "type": "text" + }, + { + "bbox": [ + 427, + 104, + 450, + 115 + ], + "score": 0.9, + "content": "\\mathcal C ( \\mathbf w )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 102, + 471, + 117 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 471, + 104, + 505, + 115 + ], + "score": 0.78, + "content": "\\dim ( \\mathbf { w } )", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 142, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 142, + 126 + ], + "score": 1.0, + "content": "choices:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 81, + 506, + 126 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 281, + 123, + 328, + 137 + ], + "lines": [ + { + "bbox": [ + 281, + 123, + 328, + 137 + ], + "spans": [ + { + "bbox": [ + 281, + 123, + 328, + 137 + ], + "score": 0.91, + "content": "p ( j ) = w _ { j } ,", + "type": "interline_equation", + "image_path": "a48b43c75935fa5ebf2a39183e9b61d2e7b31453a88848fa1484d5c7e26baacf.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 281, + 123, + 328, + 137 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 137, + 504, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 132, + 149 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 138, + 145, + 149 + ], + "score": 0.87, + "content": "w _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 137, + 169, + 149 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 170, + 138, + 176, + 149 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "-th element of w. We can readily sample from this categorical distribution and form", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 228, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 228, + 159 + ], + "score": 1.0, + "content": "an one-hot vector w˜ such that", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 137, + 505, + 159 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 272, + 156, + 338, + 170 + ], + "lines": [ + { + "bbox": [ + 272, + 156, + 338, + 170 + ], + "spans": [ + { + "bbox": [ + 272, + 156, + 338, + 170 + ], + "score": 0.91, + "content": "\\tilde { w } _ { k } = I ( k = j ) ,", + "type": "interline_equation", + "image_path": "967cc989c6c98b115376ccbbce7e228f0ce69961352ea3475e7e07f47aa4bb10.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 272, + 156, + 338, + 170 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 170, + 298, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 299, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 133, + 183 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 171, + 174, + 182 + ], + "score": 0.92, + "content": "j \\sim \\mathcal { C } ( \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 170, + 194, + 183 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 195, + 171, + 201, + 181 + ], + "score": 0.79, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 170, + 299, + 183 + ], + "score": 1.0, + "content": "is an indicator function.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 170, + 299, + 183 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "Training We use this sampling-based strategy for all the heads during training. This clearly makes", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 203, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 215 + ], + "score": 1.0, + "content": "the use of backpropagation infeasible to compute the gradient, as the sampling procedure is not", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "differentiable. Thus, we use REINFORCE (Williams, 1992) together with the three variance reduction", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 224, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 505, + 237 + ], + "score": 1.0, + "content": "techniques–global baseline, input-dependent baseline and variance normalization– suggested in (Mnih", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 234, + 178, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 178, + 248 + ], + "score": 1.0, + "content": "& Gregor, 2014).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 192, + 505, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 250, + 497, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 249, + 497, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 160, + 266 + ], + "score": 1.0, + "content": "Let us define", + "type": "text" + }, + { + "bbox": [ + 160, + 251, + 277, + 263 + ], + "score": 0.93, + "content": "R ( { \\bf x } ) = \\log p ( { \\bf y } | { \\bf x } _ { 1 } , \\ldots , { \\bf x } _ { T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 249, + 497, + 266 + ], + "score": 1.0, + "content": "as a reward. We first center and re-scale the reward by", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 249, + 497, + 266 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 265, + 345, + 293 + ], + "lines": [ + { + "bbox": [ + 265, + 265, + 345, + 293 + ], + "spans": [ + { + "bbox": [ + 265, + 265, + 345, + 293 + ], + "score": 0.95, + "content": "\\tilde { R } ( { \\bf x } ) = \\frac { R ( { \\bf x } ) - b } { \\sqrt { \\sigma ^ { 2 } + \\epsilon } } ,", + "type": "interline_equation", + "image_path": "24dcd1de042ab3bc1efc7f9256856a07e0724973f8589335669d4d208f1218c1.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 265, + 265, + 345, + 279.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 265, + 279.0, + 345, + 293.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 132, + 308 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 295, + 138, + 305 + ], + "score": 0.72, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 294, + 155, + 308 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 297, + 163, + 305 + ], + "score": 0.8, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 294, + 340, + 308 + ], + "score": 1.0, + "content": "is running average and standard deviation of", + "type": "text" + }, + { + "bbox": [ + 341, + 295, + 349, + 305 + ], + "score": 0.82, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 294, + 505, + 308 + ], + "score": 1.0, + "content": ". We can further center it for each input", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 305, + 177, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 114, + 315 + ], + "score": 0.53, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 305, + 177, + 318 + ], + "score": 1.0, + "content": "separately, i.e.,", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 294, + 505, + 318 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 315, + 351, + 330 + ], + "lines": [ + { + "bbox": [ + 258, + 315, + 351, + 330 + ], + "spans": [ + { + "bbox": [ + 258, + 315, + 351, + 330 + ], + "score": 0.92, + "content": "\\tilde { R } ( \\mathbf { x } ) \\gets \\tilde { R } ( \\mathbf { x } ) - b ( \\mathbf { x } ) ,", + "type": "interline_equation", + "image_path": "73ad2c0af30c26dbbee2a055bafa03cbc552876676f97e051deb52d5f0acf207.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 258, + 315, + 351, + 330 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 505, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 132, + 342 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 330, + 151, + 342 + ], + "score": 0.91, + "content": "b ( \\mathbf x )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 330, + 367, + 342 + ], + "score": 1.0, + "content": "is computed by a baseline network which takes as input", + "type": "text" + }, + { + "bbox": [ + 368, + 333, + 375, + 340 + ], + "score": 0.65, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 330, + 506, + 342 + ], + "score": 1.0, + "content": "and predicts its estimated reward.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 353 + ], + "score": 1.0, + "content": "The baseline network is trained to minimize the Huber loss (Huber, 1964) between the true reward", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 351, + 434, + 366 + ], + "spans": [ + { + "bbox": [ + 107, + 351, + 133, + 365 + ], + "score": 0.93, + "content": "\\tilde { R } ( { \\bf x } ) ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 351, + 236, + 366 + ], + "score": 1.0, + "content": "and the predicted reward", + "type": "text" + }, + { + "bbox": [ + 236, + 352, + 255, + 365 + ], + "score": 0.9, + "content": "b ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 351, + 434, + 366 + ], + "score": 1.0, + "content": ". We use the Huber loss, which is defined by", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 330, + 506, + 366 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 367, + 381, + 395 + ], + "lines": [ + { + "bbox": [ + 228, + 367, + 381, + 395 + ], + "spans": [ + { + "bbox": [ + 228, + 367, + 381, + 395 + ], + "score": 0.94, + "content": "H _ { \\delta } ( x ) = { \\left\\{ \\begin{array} { l l } { x ^ { 2 } } & { { \\mathrm { f o r ~ } } | x | \\leq \\delta , } \\\\ { \\delta ( 2 | x | - \\delta ) , } & { { \\mathrm { o t h e r w i s e } } , } \\end{array} \\right. }", + "type": "interline_equation", + "image_path": "e4f5a1c3730e593305a0c56f0c266c5f41401208bfe1b3dfdf05bc020f65d5dd.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 367, + 381, + 381.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 228, + 381.0, + 381, + 395.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 398, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "score": 1.0, + "content": "due to its robustness. As a further measure to reduce the variance, we regularize the negative entropy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 408, + 502, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 502, + 420 + ], + "score": 1.0, + "content": "of all those category distributions to facilitate a better exploration during training (Xu et al., 2015).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 396, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 383, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 383, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 383, + 439 + ], + "score": 1.0, + "content": "Then, the cost function for each training example is approximated as", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 421, + 383, + 439 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 437, + 480, + 524 + ], + "lines": [ + { + "bbox": [ + 177, + 437, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 177, + 437, + 480, + 524 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { C ^ { n } ( \\boldsymbol { \\theta } ) = - \\log p ( \\mathbf { y } | \\mathbf { x } _ { 1 : T } , \\tilde { w } _ { 1 : J } , \\tilde { u } _ { 1 : J } , \\tilde { e } _ { 1 : J } ) } \\\\ & { \\qquad - \\displaystyle \\sum _ { j = 1 } ^ { J } \\tilde { R } ( \\mathbf { x } ^ { n } ) ( \\log p ( \\tilde { w } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { u } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { e } _ { j } | \\mathbf { x } _ { 1 : T } ) ) } \\\\ & { \\qquad - \\lambda _ { H } \\displaystyle \\sum _ { j = 1 } ^ { J } ( \\mathcal { H } ( w _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( u _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( e _ { j } | \\mathbf { x } _ { 1 : T } ) ) . } \\end{array}", + "type": "interline_equation", + "image_path": "bffd054bc17ac384c9d9037ee3996704175726c7a06e6adc5c2450df911bf792.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 177, + 437, + 480, + 466.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 177, + 466.0, + 480, + 495.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 177, + 495.0, + 480, + 524.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 527, + 504, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 132, + 540 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 528, + 140, + 537 + ], + "score": 0.85, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 525, + 274, + 540 + ], + "score": 1.0, + "content": "is the number of addressing steps,", + "type": "text" + }, + { + "bbox": [ + 274, + 528, + 289, + 538 + ], + "score": 0.89, + "content": "\\lambda _ { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 525, + 462, + 540 + ], + "score": 1.0, + "content": "is the entropy regularization coefficient, and", + "type": "text" + }, + { + "bbox": [ + 462, + 528, + 472, + 537 + ], + "score": 0.82, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 156, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 156, + 551 + ], + "score": 1.0, + "content": "the entropy.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 525, + 506, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 560, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "Inference Once training is over, we switch to a deterministic strategy. We simply choose an element", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 570, + 421, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 421, + 583 + ], + "score": 1.0, + "content": "of w with the largest value to be the index of the target memory cell, such that", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 559, + 505, + 583 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 584, + 358, + 598 + ], + "lines": [ + { + "bbox": [ + 252, + 584, + 358, + 598 + ], + "spans": [ + { + "bbox": [ + 252, + 584, + 358, + 598 + ], + "score": 0.91, + "content": "\\tilde { w } _ { k } = \\mathbf { I } ( k = \\operatorname { a r g m a x } ( \\mathbf { w } ) ) .", + "type": "interline_equation", + "image_path": "92541e0be0f429b20ffa6a34399d1519d61c50f5dd696ad4ec1c1e4ef7a3cc0d.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 252, + 584, + 358, + 598 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "Curriculum Learning for the Discrete Attention Training discrete attention with feed-forward", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 630 + ], + "score": 1.0, + "content": "controller and REINFORCE is challenging. We propose to use a curriculum strategy for training", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 467, + 639 + ], + "score": 1.0, + "content": "with the discrete attention in order to tackle this problem. For each minibatch, we sample", + "type": "text" + }, + { + "bbox": [ + 467, + 629, + 475, + 637 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "from a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 275, + 650 + ], + "score": 1.0, + "content": "binomial distribution with the probability", + "type": "text" + }, + { + "bbox": [ + 275, + 637, + 285, + 649 + ], + "score": 0.59, + "content": "p ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 636, + 288, + 650 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 288, + 637, + 345, + 649 + ], + "score": 0.89, + "content": "\\pi ^ { t } \\sim \\mathrm { B i n } ( p ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 636, + 506, + 650 + ], + "score": 1.0, + "content": ". The model will either use the discrete", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 647, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 647, + 268, + 663 + ], + "score": 1.0, + "content": "or the continuous-attention based on the", + "type": "text" + }, + { + "bbox": [ + 269, + 649, + 279, + 659 + ], + "score": 0.87, + "content": "\\pi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 647, + 429, + 663 + ], + "score": 1.0, + "content": ". We start the training procedure with", + "type": "text" + }, + { + "bbox": [ + 429, + 648, + 459, + 660 + ], + "score": 0.92, + "content": "p ^ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 647, + 506, + 663 + ], + "score": 1.0, + "content": "and during", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 659, + 320, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 154, + 678 + ], + "score": 1.0, + "content": "the training", + "type": "text" + }, + { + "bbox": [ + 155, + 662, + 164, + 675 + ], + "score": 0.87, + "content": "p ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 661, + 272, + 678 + ], + "score": 1.0, + "content": "is annealed to 0 by setting", + "type": "text" + }, + { + "bbox": [ + 272, + 659, + 317, + 678 + ], + "score": 0.93, + "content": "\\begin{array} { r } { p ^ { t } = \\frac { p ^ { 0 } } { \\sqrt { 1 + t } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 661, + 320, + 678 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5, + "bbox_fs": [ + 104, + 605, + 506, + 678 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 210, + 694 + ], + "score": 1.0, + "content": "We can rewrite the weights", + "type": "text" + }, + { + "bbox": [ + 211, + 684, + 224, + 693 + ], + "score": 0.87, + "content": "\\mathbf { w } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 682, + 505, + 694 + ], + "score": 1.0, + "content": "as in Equation 14, where it is expressed as the combination of continuous", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 177, + 705 + ], + "score": 1.0, + "content": "attention weights", + "type": "text" + }, + { + "bbox": [ + 177, + 693, + 190, + 702 + ], + "score": 0.88, + "content": "\\hat { \\mathbf { w } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 692, + 313, + 705 + ], + "score": 1.0, + "content": "and discrete attention weights", + "type": "text" + }, + { + "bbox": [ + 313, + 693, + 326, + 702 + ], + "score": 0.88, + "content": "\\tilde { \\mathbf { w } } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 692, + 348, + 705 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 348, + 693, + 358, + 702 + ], + "score": 0.88, + "content": "\\mathit { \\Pi } _ { \\pi ^ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "being a binary variable that chooses", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 703, + 185, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 185, + 714 + ], + "score": 1.0, + "content": "to use one of them,", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 682, + 505, + 714 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 715, + 362, + 730 + ], + "lines": [ + { + "bbox": [ + 249, + 715, + 362, + 730 + ], + "spans": [ + { + "bbox": [ + 249, + 715, + 362, + 730 + ], + "score": 0.92, + "content": "\\mathbf { w } ^ { t } \\pi ^ { t } \\bar { \\mathbf { w } } ^ { t } + ( 1 - \\pi ^ { t } ) \\tilde { \\mathbf { w } } ^ { t } .", + "type": "interline_equation", + "image_path": "40eea2ba87dd6930595aeecbe70dc2633c2eb8c53228b730ad911bb606067a39.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 249, + 715, + 362, + 730 + ], + "spans": [], + "index": 45 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "By using this curriculum learning strategy, at the beginning of the training, the model learns to use", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 378, + 105 + ], + "score": 1.0, + "content": "the memory mainly with the continuous attention. As we anneal the", + "type": "text" + }, + { + "bbox": [ + 378, + 93, + 388, + 104 + ], + "score": 0.87, + "content": "p ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 93, + 505, + 105 + ], + "score": 1.0, + "content": ", the model will rely more on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 194, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 194, + 115 + ], + "score": 1.0, + "content": "the discrete attention.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 107, + 129, + 411, + 143 + ], + "lines": [ + { + "bbox": [ + 104, + 129, + 412, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 412, + 145 + ], + "score": 1.0, + "content": "5 REGULARIZING DYNAMIC NEURAL TURING MACHINES", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "When the controller of D-NTM is a powerful recurrent neural network, it is important to regularize", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "training of the D-NTM so as to avoid suboptimal solutions in which the D-NTM ignores the memory", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 300, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 300, + 187 + ], + "score": 1.0, + "content": "and works as a simple recurrent neural network.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Read-Write Consistency Regularizer One such suboptimal solution we have observed in our", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "preliminary experiments with the proposed D-NTM is that the D-NTM uses the address part A of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "the memory matrix simply as an additional weight matrix, rather than as a means to accessing the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "content part C. We found that this pathological case can be effectively avoided by encouraging the read", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "head to point to a memory cell which has also been pointed by the write head. This can be implemented", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 254, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 254, + 265 + ], + "score": 1.0, + "content": "as the following regularization term:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 267, + 397, + 303 + ], + "lines": [ + { + "bbox": [ + 214, + 267, + 397, + 303 + ], + "spans": [ + { + "bbox": [ + 214, + 267, + 397, + 303 + ], + "score": 0.94, + "content": "R _ { \\mathrm { r w } } ( \\mathbf { w } , \\mathbf { u } ) = \\lambda \\sum _ { t ^ { \\prime } = 1 } ^ { T } | | 1 - ( \\frac { 1 } { t ^ { \\prime } } \\sum _ { t = 1 } ^ { t ^ { \\prime } } \\mathbf { u } _ { t } ) ^ { \\top } \\mathbf { w } _ { t ^ { \\prime } } | | _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "eea88a11afc5bbd2611efe1c1a4ac0e2a2be4fba9a0874199647da45aad7a8d5.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 267, + 397, + 285.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 214, + 285.0, + 397, + 303.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 370, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 370, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 201, + 326 + ], + "score": 1.0, + "content": "In the equations above,", + "type": "text" + }, + { + "bbox": [ + 201, + 315, + 212, + 324 + ], + "score": 0.85, + "content": "\\mathbf { u } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 311, + 277, + 326 + ], + "score": 1.0, + "content": "is the write and", + "type": "text" + }, + { + "bbox": [ + 277, + 315, + 290, + 324 + ], + "score": 0.88, + "content": "\\mathbf { w } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 311, + 370, + 326 + ], + "score": 1.0, + "content": "is the read weights.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "Next Input Prediction as Regularization Temporal structure is a strong signal that should be", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 345, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 506, + 361 + ], + "score": 1.0, + "content": "exploited by the controller based on a recurrent neural network. We exploit this structure by letting", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "the controller predict the input in the future. We maximize the predictability of the next input by the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 448, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 448, + 380 + ], + "score": 1.0, + "content": "controller during training. This is equivalent to minimizing the following regularizer:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 384, + 399, + 397 + ], + "lines": [ + { + "bbox": [ + 210, + 384, + 399, + 397 + ], + "spans": [ + { + "bbox": [ + 210, + 384, + 399, + 397 + ], + "score": 0.92, + "content": "R _ { \\mathrm { p r e d } } ( \\mathbf { W } ) = - \\log p ( \\mathbf { f } _ { t + 1 } | \\mathbf { f } _ { t } , \\mathbf { w } _ { t } , \\mathbf { u } _ { t } , \\mathbf { M } _ { t } ; \\mathbf { W } ) )", + "type": "interline_equation", + "image_path": "daa07f4d77ca80eb4471c5abe2b0db80830859f8307bfb89e9b482206c4def7c.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 210, + 384, + 399, + 397 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 401, + 503, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 133, + 415 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 402, + 143, + 413 + ], + "score": 0.87, + "content": "f _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 400, + 240, + 415 + ], + "score": 1.0, + "content": "is the current input and", + "type": "text" + }, + { + "bbox": [ + 240, + 402, + 259, + 413 + ], + "score": 0.92, + "content": "f _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 400, + 505, + 415 + ], + "score": 1.0, + "content": "is the input at next timestep. We found this regularizer to be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 412, + 372, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 372, + 424 + ], + "score": 1.0, + "content": "effective in our preliminary experiments and use it for bAbI tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 438, + 211, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 213, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 213, + 453 + ], + "score": 1.0, + "content": "6 RELATED WORK", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "A recurrent neural network (RNN), which is used as a controller in the proposed D-NTM, has an", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "score": 1.0, + "content": "implicit memory in the form of recurring hidden states. Even with this implicit memory, a vanilla", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "RNN is however known to have difficulties in storing information for long time-spans (Bengio et al.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 494, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 506 + ], + "score": 1.0, + "content": "1994; Hochreiter, 1991). Long short-term memory (LSTM, (Hochreiter & Schmidhuber, 1997)) and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 505, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 516 + ], + "score": 1.0, + "content": "gated recurrent units (GRU, (Cho et al., 2014)) have been found to address this issue. However all", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 514, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 507, + 529 + ], + "score": 1.0, + "content": "these models based solely on RNNs have been found to be limited when they are used to solve, e.g.,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 313, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 313, + 540 + ], + "score": 1.0, + "content": "algorithmic tasks and episodic question-answering.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "In addition to the finite random access memory of the neural Turing machine, based on which the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 552, + 507, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 507, + 565 + ], + "score": 1.0, + "content": "D-NTM is designed, other data structures have been proposed as external memory for neural networks.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 563, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 575 + ], + "score": 1.0, + "content": "In (Sun et al., 1997; Grefenstette et al., 2015; Joulin & Mikolov, 2015), a continuous, differentiable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "score": 1.0, + "content": "stack was proposed. In (Zaremba et al., 2015; Zaremba & Sutskever, 2015), grid and tape storages", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "score": 1.0, + "content": "are used. These approaches differ from the NTM in that their memory is unbounded and can grow", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "score": 1.0, + "content": "indefinitely. On the other hand, they are often not randomly accessible.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 507, + 624 + ], + "score": 1.0, + "content": "Memory networks (Weston et al., 2015b) form another family of neural networks with external memory.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "In this class of neural networks, information is stored explicitly as it is (in the form of its continuous", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 631, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 506, + 644 + ], + "score": 1.0, + "content": "representation) in the memory, without being erased or modified during an episode. Memory networks", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "and their variants have been applied to various tasks successfully (Sukhbaatar et al., 2015; Bordes et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "score": 1.0, + "content": "2015; Dodge et al., 2015; Xiong et al., 2016). Miller et al. (2016) have also independently proposed", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 662, + 395, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 395, + 675 + ], + "score": 1.0, + "content": "the idea of having separate key and value vectors for memory networks.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 690 + ], + "score": 1.0, + "content": "Another related family of models is the attention-based neural networks. Neural networks with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "continuous or discrete attention over an input have shown promising results on a variety of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "challenging tasks, including machine translation (Bahdanau et al., 2015; Luong et al., 2015), speech", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "recognition (Chorowski et al., 2015), machine reading comprehension (Hermann et al., 2015) and", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 280, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 280, + 732 + ], + "score": 1.0, + "content": "image caption generation (Xu et al., 2015).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "By using this curriculum learning strategy, at the beginning of the training, the model learns to use", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 378, + 105 + ], + "score": 1.0, + "content": "the memory mainly with the continuous attention. As we anneal the", + "type": "text" + }, + { + "bbox": [ + 378, + 93, + 388, + 104 + ], + "score": 0.87, + "content": "p ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 93, + 505, + 105 + ], + "score": 1.0, + "content": ", the model will rely more on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 194, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 194, + 115 + ], + "score": 1.0, + "content": "the discrete attention.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 115 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 129, + 411, + 143 + ], + "lines": [ + { + "bbox": [ + 104, + 129, + 412, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 412, + 145 + ], + "score": 1.0, + "content": "5 REGULARIZING DYNAMIC NEURAL TURING MACHINES", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "When the controller of D-NTM is a powerful recurrent neural network, it is important to regularize", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "training of the D-NTM so as to avoid suboptimal solutions in which the D-NTM ignores the memory", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 300, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 300, + 187 + ], + "score": 1.0, + "content": "and works as a simple recurrent neural network.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 154, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Read-Write Consistency Regularizer One such suboptimal solution we have observed in our", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "preliminary experiments with the proposed D-NTM is that the D-NTM uses the address part A of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "the memory matrix simply as an additional weight matrix, rather than as a means to accessing the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "content part C. We found that this pathological case can be effectively avoided by encouraging the read", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 253 + ], + "score": 1.0, + "content": "head to point to a memory cell which has also been pointed by the write head. This can be implemented", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 254, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 254, + 265 + ], + "score": 1.0, + "content": "as the following regularization term:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 198, + 506, + 265 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 267, + 397, + 303 + ], + "lines": [ + { + "bbox": [ + 214, + 267, + 397, + 303 + ], + "spans": [ + { + "bbox": [ + 214, + 267, + 397, + 303 + ], + "score": 0.94, + "content": "R _ { \\mathrm { r w } } ( \\mathbf { w } , \\mathbf { u } ) = \\lambda \\sum _ { t ^ { \\prime } = 1 } ^ { T } | | 1 - ( \\frac { 1 } { t ^ { \\prime } } \\sum _ { t = 1 } ^ { t ^ { \\prime } } \\mathbf { u } _ { t } ) ^ { \\top } \\mathbf { w } _ { t ^ { \\prime } } | | _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "eea88a11afc5bbd2611efe1c1a4ac0e2a2be4fba9a0874199647da45aad7a8d5.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 267, + 397, + 285.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 214, + 285.0, + 397, + 303.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 370, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 370, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 201, + 326 + ], + "score": 1.0, + "content": "In the equations above,", + "type": "text" + }, + { + "bbox": [ + 201, + 315, + 212, + 324 + ], + "score": 0.85, + "content": "\\mathbf { u } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 311, + 277, + 326 + ], + "score": 1.0, + "content": "is the write and", + "type": "text" + }, + { + "bbox": [ + 277, + 315, + 290, + 324 + ], + "score": 0.88, + "content": "\\mathbf { w } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 311, + 370, + 326 + ], + "score": 1.0, + "content": "is the read weights.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 311, + 370, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "Next Input Prediction as Regularization Temporal structure is a strong signal that should be", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 345, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 506, + 361 + ], + "score": 1.0, + "content": "exploited by the controller based on a recurrent neural network. We exploit this structure by letting", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "the controller predict the input in the future. We maximize the predictability of the next input by the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 448, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 448, + 380 + ], + "score": 1.0, + "content": "controller during training. This is equivalent to minimizing the following regularizer:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 335, + 506, + 380 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 384, + 399, + 397 + ], + "lines": [ + { + "bbox": [ + 210, + 384, + 399, + 397 + ], + "spans": [ + { + "bbox": [ + 210, + 384, + 399, + 397 + ], + "score": 0.92, + "content": "R _ { \\mathrm { p r e d } } ( \\mathbf { W } ) = - \\log p ( \\mathbf { f } _ { t + 1 } | \\mathbf { f } _ { t } , \\mathbf { w } _ { t } , \\mathbf { u } _ { t } , \\mathbf { M } _ { t } ; \\mathbf { W } ) )", + "type": "interline_equation", + "image_path": "daa07f4d77ca80eb4471c5abe2b0db80830859f8307bfb89e9b482206c4def7c.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 210, + 384, + 399, + 397 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 401, + 503, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 133, + 415 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 402, + 143, + 413 + ], + "score": 0.87, + "content": "f _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 400, + 240, + 415 + ], + "score": 1.0, + "content": "is the current input and", + "type": "text" + }, + { + "bbox": [ + 240, + 402, + 259, + 413 + ], + "score": 0.92, + "content": "f _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 400, + 505, + 415 + ], + "score": 1.0, + "content": "is the input at next timestep. We found this regularizer to be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 412, + 372, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 372, + 424 + ], + "score": 1.0, + "content": "effective in our preliminary experiments and use it for bAbI tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 400, + 505, + 424 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 438, + 211, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 213, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 213, + 453 + ], + "score": 1.0, + "content": "6 RELATED WORK", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "A recurrent neural network (RNN), which is used as a controller in the proposed D-NTM, has an", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "score": 1.0, + "content": "implicit memory in the form of recurring hidden states. Even with this implicit memory, a vanilla", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "RNN is however known to have difficulties in storing information for long time-spans (Bengio et al.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 494, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 506 + ], + "score": 1.0, + "content": "1994; Hochreiter, 1991). Long short-term memory (LSTM, (Hochreiter & Schmidhuber, 1997)) and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 505, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 516 + ], + "score": 1.0, + "content": "gated recurrent units (GRU, (Cho et al., 2014)) have been found to address this issue. However all", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 514, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 507, + 529 + ], + "score": 1.0, + "content": "these models based solely on RNNs have been found to be limited when they are used to solve, e.g.,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 313, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 313, + 540 + ], + "score": 1.0, + "content": "algorithmic tasks and episodic question-answering.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 463, + 507, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "In addition to the finite random access memory of the neural Turing machine, based on which the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 552, + 507, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 507, + 565 + ], + "score": 1.0, + "content": "D-NTM is designed, other data structures have been proposed as external memory for neural networks.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 563, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 575 + ], + "score": 1.0, + "content": "In (Sun et al., 1997; Grefenstette et al., 2015; Joulin & Mikolov, 2015), a continuous, differentiable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "score": 1.0, + "content": "stack was proposed. In (Zaremba et al., 2015; Zaremba & Sutskever, 2015), grid and tape storages", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 596 + ], + "score": 1.0, + "content": "are used. These approaches differ from the NTM in that their memory is unbounded and can grow", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 392, + 606 + ], + "score": 1.0, + "content": "indefinitely. On the other hand, they are often not randomly accessible.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 541, + 507, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 507, + 624 + ], + "score": 1.0, + "content": "Memory networks (Weston et al., 2015b) form another family of neural networks with external memory.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "In this class of neural networks, information is stored explicitly as it is (in the form of its continuous", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 631, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 506, + 644 + ], + "score": 1.0, + "content": "representation) in the memory, without being erased or modified during an episode. Memory networks", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "and their variants have been applied to various tasks successfully (Sukhbaatar et al., 2015; Bordes et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "score": 1.0, + "content": "2015; Dodge et al., 2015; Xiong et al., 2016). Miller et al. (2016) have also independently proposed", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 662, + 395, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 395, + 675 + ], + "score": 1.0, + "content": "the idea of having separate key and value vectors for memory networks.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 609, + 507, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 690 + ], + "score": 1.0, + "content": "Another related family of models is the attention-based neural networks. Neural networks with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "continuous or discrete attention over an input have shown promising results on a variety of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "challenging tasks, including machine translation (Bahdanau et al., 2015; Luong et al., 2015), speech", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "recognition (Chorowski et al., 2015), machine reading comprehension (Hermann et al., 2015) and", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 280, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 280, + 732 + ], + "score": 1.0, + "content": "image caption generation (Xu et al., 2015).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 679, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The latter two, the memory network and attention-based networks, are however clearly distinguishable", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 428, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 428, + 106 + ], + "score": 1.0, + "content": "from the D-NTM by the fact that they do not modify the content of the memory.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 119, + 200, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 118, + 201, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 201, + 134 + ], + "score": 1.0, + "content": "7 EXPERIMENTS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "We provide experimental results to demonstrate the abilities of our model, first on Facebook bAbI", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "task (Weston et al., 2015a). We give detailed analysis and experimental results on this task. We also", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "compare different variations of NTM on bAbI tasks. We have performed experiments on sequential", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 187 + ], + "score": 1.0, + "content": "permuted MNIST (Le et al., 2015) and on toy tasks to compare other published models on these tasks", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 185, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 507, + 198 + ], + "score": 1.0, + "content": "with a recurrent controller. The details of our experiments are provided in the supplementary material.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 209, + 338, + 221 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 339, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 339, + 223 + ], + "score": 1.0, + "content": "7.1 EPISODIC QUESTION-ANSWERING: BABI TASKS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 229, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 243 + ], + "score": 1.0, + "content": "In this section, we evaluate the proposed D-NTM on the recently proposed episodic question-answering", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "score": 1.0, + "content": "task called Facebook bAbI. We use the dataset with 10k training examples per sub-task provided by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Facebook.1 For each episode, the D-NTM reads a sequence of factual sentences followed by a question,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "all of which are given as natural language sentences. The D-NTM is expected to store and retrieve", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "relevant information in the memory in order to answer the question based on the presented facts. Exact", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 282, + 438, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 438, + 296 + ], + "score": 1.0, + "content": "implementation details and hyper-parameter settings are provided in the appendix.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 305, + 170, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 172, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 172, + 319 + ], + "score": 1.0, + "content": "7.1.1 GOALS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "The goal of this experiment is three-fold. First, we present for the first time the performance of a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "memory-based network that can both read and write dynamically on the Facebook bAbI tasks2. We aim", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "to understand whether a model that has to learn to write an incoming fact to the memory, rather than", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "storing it as it is, is able to work well, and to do so, we compare both the original NTM and proposed", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 367, + 240, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 240, + 379 + ], + "score": 1.0, + "content": "D-NTM against an LSTM-RNN.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "Second, we investigate the effect of having to learn how to write. The fact that the NTM needs to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 392, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 407 + ], + "score": 1.0, + "content": "learn to write likely has adverse effect on the overall performance, when compared to, for instance,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "end-to-end memory networks (MemN2N, (Sukhbaatar et al., 2015)) and dynamic memory network", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 107, + 415, + 139, + 425 + ], + "score": 0.67, + "content": "\\mathrm { ( D M N + }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 414, + 505, + 427 + ], + "score": 1.0, + "content": ", (Xiong et al., 2016)) both of which simply store the incoming facts as they are. We quantify", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "this effect in this experiment. Lastly, we show the effect of the proposed learnable addressing scheme.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "We further explore the effect of using a feedforward controller instead of the GRU controller. In addition", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "to the explicit memory, the GRU controller can use its own internal hidden state as the memory. On", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "the other hand, the feedforward controller must solely rely on the explicit memory, as it is the only", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 183, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 183, + 485 + ], + "score": 1.0, + "content": "memory available.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 108, + 495, + 246, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 248, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 248, + 508 + ], + "score": 1.0, + "content": "7.1.2 RESULTS AND ANALYSIS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "In Table 1, we first observe that the NTMs are indeed capable of solving this type of episodic", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "score": 1.0, + "content": "question-answering better than the vanilla LSTM-RNN. Although the availability of explicit memory", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "in the NTM has already suggested this result, we note that this is the first time neural Turing machines", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 545, + 250, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 250, + 558 + ], + "score": 1.0, + "content": "have been used in this specific task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "All the variants of NTM with the GRU controller outperform the vanilla LSTM-RNN. However, not all", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "of them perform equally well. First, it is clear that the proposed dynamic NTM (D-NTM) using the GRU", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "controller outperforms the original NTM with the GRU controller (NTM, CBA only NTM vs. continuous", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "D-NTM, Discrete D-NTM). As discussed earlier, the learnable addressing scheme of the D-NTM allows", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "the controller to access the memory slots by location in a potentially nonlinear way. We expect it to help", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "score": 1.0, + "content": "with tasks that have non-trivial access patterns, and as anticipated, we see a large gain with the D-NTM", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "score": 1.0, + "content": "over the original NTM in the tasks of, for instance, 12 - Conjunction and 17 - Positional Reasoning.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Among the recurrent variants of the proposed D-NTM, we notice significant improvements by using", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "discrete addressing over using continuous addressing. We conjecture that this is due to certain types", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 660, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 506, + 675 + ], + "score": 1.0, + "content": "of tasks that require precise/sharp retrieval of a stored fact, in which case continuous addressing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "is in disadvantage over discrete addressing. This is evident from the observation that the D-NTM", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 682, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 506, + 694 + ], + "score": 1.0, + "content": "with discrete addressing significantly outperforms that with continuous addressing in the tasks of 8 -", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 702, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 440, + 713 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 440, + 713 + ], + "score": 1.0, + "content": "1 https://research.facebook.com/researchers/1543934539189348", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 710, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 710, + 506, + 724 + ], + "score": 1.0, + "content": "2Similar experiments were done in the recently published (Graves et al., 2016), but D-NTM results for bAbI", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 286, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 286, + 732 + ], + "score": 1.0, + "content": "tasks were already available in arxiv by that time.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The latter two, the memory network and attention-based networks, are however clearly distinguishable", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 428, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 428, + 106 + ], + "score": 1.0, + "content": "from the D-NTM by the fact that they do not modify the content of the memory.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 119, + 200, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 118, + 201, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 201, + 134 + ], + "score": 1.0, + "content": "7 EXPERIMENTS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "We provide experimental results to demonstrate the abilities of our model, first on Facebook bAbI", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "task (Weston et al., 2015a). We give detailed analysis and experimental results on this task. We also", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "compare different variations of NTM on bAbI tasks. We have performed experiments on sequential", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 187 + ], + "score": 1.0, + "content": "permuted MNIST (Le et al., 2015) and on toy tasks to compare other published models on these tasks", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 185, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 507, + 198 + ], + "score": 1.0, + "content": "with a recurrent controller. The details of our experiments are provided in the supplementary material.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 143, + 507, + 198 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 209, + 338, + 221 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 339, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 339, + 223 + ], + "score": 1.0, + "content": "7.1 EPISODIC QUESTION-ANSWERING: BABI TASKS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 229, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 243 + ], + "score": 1.0, + "content": "In this section, we evaluate the proposed D-NTM on the recently proposed episodic question-answering", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "score": 1.0, + "content": "task called Facebook bAbI. We use the dataset with 10k training examples per sub-task provided by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Facebook.1 For each episode, the D-NTM reads a sequence of factual sentences followed by a question,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "all of which are given as natural language sentences. The D-NTM is expected to store and retrieve", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "relevant information in the memory in order to answer the question based on the presented facts. Exact", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 282, + 438, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 438, + 296 + ], + "score": 1.0, + "content": "implementation details and hyper-parameter settings are provided in the appendix.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 227, + 506, + 296 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 305, + 170, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 172, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 172, + 319 + ], + "score": 1.0, + "content": "7.1.1 GOALS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "The goal of this experiment is three-fold. First, we present for the first time the performance of a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "memory-based network that can both read and write dynamically on the Facebook bAbI tasks2. We aim", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "to understand whether a model that has to learn to write an incoming fact to the memory, rather than", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "storing it as it is, is able to work well, and to do so, we compare both the original NTM and proposed", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 367, + 240, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 240, + 379 + ], + "score": 1.0, + "content": "D-NTM against an LSTM-RNN.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 324, + 506, + 379 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "Second, we investigate the effect of having to learn how to write. The fact that the NTM needs to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 392, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 407 + ], + "score": 1.0, + "content": "learn to write likely has adverse effect on the overall performance, when compared to, for instance,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "end-to-end memory networks (MemN2N, (Sukhbaatar et al., 2015)) and dynamic memory network", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 107, + 415, + 139, + 425 + ], + "score": 0.67, + "content": "\\mathrm { ( D M N + }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 414, + 505, + 427 + ], + "score": 1.0, + "content": ", (Xiong et al., 2016)) both of which simply store the incoming facts as they are. We quantify", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "this effect in this experiment. Lastly, we show the effect of the proposed learnable addressing scheme.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 382, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "We further explore the effect of using a feedforward controller instead of the GRU controller. In addition", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "to the explicit memory, the GRU controller can use its own internal hidden state as the memory. On", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "the other hand, the feedforward controller must solely rely on the explicit memory, as it is the only", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 183, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 183, + 485 + ], + "score": 1.0, + "content": "memory available.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 441, + 506, + 485 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 495, + 246, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 248, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 248, + 508 + ], + "score": 1.0, + "content": "7.1.2 RESULTS AND ANALYSIS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "In Table 1, we first observe that the NTMs are indeed capable of solving this type of episodic", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "score": 1.0, + "content": "question-answering better than the vanilla LSTM-RNN. Although the availability of explicit memory", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "in the NTM has already suggested this result, we note that this is the first time neural Turing machines", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 545, + 250, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 250, + 558 + ], + "score": 1.0, + "content": "have been used in this specific task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 513, + 506, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "All the variants of NTM with the GRU controller outperform the vanilla LSTM-RNN. However, not all", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "of them perform equally well. First, it is clear that the proposed dynamic NTM (D-NTM) using the GRU", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "controller outperforms the original NTM with the GRU controller (NTM, CBA only NTM vs. continuous", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "D-NTM, Discrete D-NTM). As discussed earlier, the learnable addressing scheme of the D-NTM allows", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "the controller to access the memory slots by location in a potentially nonlinear way. We expect it to help", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "score": 1.0, + "content": "with tasks that have non-trivial access patterns, and as anticipated, we see a large gain with the D-NTM", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "score": 1.0, + "content": "over the original NTM in the tasks of, for instance, 12 - Conjunction and 17 - Positional Reasoning.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 561, + 506, + 637 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Among the recurrent variants of the proposed D-NTM, we notice significant improvements by using", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "discrete addressing over using continuous addressing. We conjecture that this is due to certain types", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 660, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 506, + 675 + ], + "score": 1.0, + "content": "of tasks that require precise/sharp retrieval of a stored fact, in which case continuous addressing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "is in disadvantage over discrete addressing. This is evident from the observation that the D-NTM", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 682, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 506, + 694 + ], + "score": 1.0, + "content": "with discrete addressing significantly outperforms that with continuous addressing in the tasks of 8 -", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 104, + 640, + 506, + 694 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 119, + 75, + 492, + 237 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 119, + 75, + 492, + 237 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 75, + 492, + 237 + ], + "spans": [ + { + "bbox": [ + 119, + 75, + 492, + 237 + ], + "score": 0.982, + "html": "
DMN+L1-step CBA1-step Soft D-NTM1-step Discrete D-NTM3-steps LBA* NTM3-steps CBA NTM3-steps Soft D-NTM3-steps Discrete D-NTM
Task 1LSTM 0.00MemN2N 0.000.00NTM 16.30NTM 16.885.416.660.000.000.000.00
281.900.300.3057.0855.7058.5456.0461.6759.3846.6662.29
383.102.101.1074.1655.0074.5872.0883.5465.2147.0841.45
40.200.000.000.000.000.000.000.000.000.000.00
51.200.800.501.4620.411.661.040.831.461.251.45
651.800.100.0023.3321.0440.2044.7948.1354.8020.6211.04
724.902.002.4021.6721.6719.1619.587.9237.707.295.62
834.100.900.0025.7621.0512.5818.4625.388.8211.020.74
920.200.300.0024.7924.1736.6634.3737.800.0039.3732.50
1030.100.000.0041.4633.1352.2950.8356.2523.7520.0020.83
1110.300.100.0018.9631.8831.454.163.960.2830.6216.87
1223.400.000.0025.8330.007.706.6628.7523.755.414.58
136.100.000.006.675.635.622.295.8383.137.915.00
1481.000.100.2058.5459.1760.0063.7561.8857.7158.1260.20
1578.700.000.0036.4642.3036.8739.2735.6221.8836.0440.26
1651.9051.8045.3071.1571.1549.1651.3546.1550.0046.0445.41
1750.1018.604.2043.7543.7517.9116.0443.7556.2521.259.16
186.805.302.103.9647.503.953.5447.5047.506.871.66
1990.302.300.0075.8971.5173.7464.6361.5663.6575.8876.66
202.100.000.001.250.002.703.120.400.003.330.00
Avg.Err.36.414.242.8131.4233.6029.5127.9332.8532.7624.2421.79
", + "type": "table", + "image_path": "2d8d524f20ba7886f2d7b3b5ca3db3d355229f453ab7efd976a9be7efb577b54.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 75, + 492, + 129.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 129.0, + 492, + 183.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 183.0, + 492, + 237.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 504, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 204, + 257 + ], + "score": 1.0, + "content": "Table 1: Test error rates", + "type": "text" + }, + { + "bbox": [ + 204, + 245, + 220, + 256 + ], + "score": 0.8, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 245, + 504, + 257 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "the GRU and feedforward controller. FF stands for the experiments that are conducted with feedforward", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 221, + 277 + ], + "score": 1.0, + "content": "controller. Let us, note that", + "type": "text" + }, + { + "bbox": [ + 222, + 266, + 247, + 276 + ], + "score": 0.67, + "content": "\\mathrm { L B A ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 266, + 505, + 277 + ], + "score": 1.0, + "content": "refers to NTM that uses both LBA and CBA. In this table, we", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "score": 1.0, + "content": "compare multi-step vs single-step addressing, original NTM with location based+content based", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 287, + 478, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 478, + 299 + ], + "score": 1.0, + "content": "addressing vs only content based addressing, and discrete vs continuous addressing on bAbI.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "Lists/Sets and 11 - Basic Coreference. Furthermore, this is in line with an earlier observation in (Xu et al.,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "2015), where discrete addressing was found to generalize better in the task of image caption generation.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 504, + 384 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "In Table 2, we also observe that the D-NTM with the feedforward controller and discrete attention", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "performs worse than LSTM and D-NTM with continuous-attention. However, when the proposed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 373, + 462, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 462, + 385 + ], + "score": 1.0, + "content": "curriculum strategy from Sec. 4 is used, the average test error drops from 68.30 to 37.79.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "score": 1.0, + "content": "We empirically found training of the feedforward controller more difficult than that of the recurrent", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 400, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 411 + ], + "score": 1.0, + "content": "controller. We train our feedforward controller based models four times longer (in terms of the number", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 411, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 421 + ], + "score": 1.0, + "content": "of updates) than the recurrent controller based ones in order to ensure that they are converged for most", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 421, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 432 + ], + "score": 1.0, + "content": "of the tasks. On the other hand, the models trained with the GRU controller overfit on bAbI tasks", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "very quickly. For example, on tasks 3 and 16 the feedforward controller based model underfits (i.e.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 442, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 453 + ], + "score": 1.0, + "content": "high training loss) at the end of the training, whereas with the same number of units the model with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 451, + 389, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 389, + 464 + ], + "score": 1.0, + "content": "the GRU controller can overfit on those tasks after 3,000 updates only.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 503, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "When our results are compared to the variants of the memory network Weston et al. (2015b) (MemN2N", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 477, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 123, + 492 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 479, + 153, + 489 + ], + "score": 0.58, + "content": "\\mathrm { D M N + }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 477, + 505, + 492 + ], + "score": 1.0, + "content": "), we notice a significant performance gap. We attribute this gap to the difficulty in learning", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 489, + 271, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 271, + 501 + ], + "score": 1.0, + "content": "to manipulate and store a complex input.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 238, + 511, + 372, + 673 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 238, + 511, + 372, + 673 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 238, + 511, + 372, + 673 + ], + "spans": [ + { + "bbox": [ + 238, + 511, + 372, + 673 + ], + "score": 0.96, + "html": "
TaskFF Soft D-NTMFF Discrete D-NTMFF Discrete* D-NTM
14.3881.6714.79
227.576.6776.67
371.2579.3870.83
40.0078.6544.06
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.62
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.92
1939.4771.4671.56
200.076.569.79
Avg.Err.12.8168.3037.79
", + "type": "table", + "image_path": "15be19b050d013aa2678635d887fe055e193f11fbac6d6b0f3bc9e3da80e1f5c.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 238, + 511, + 372, + 673 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 105, + 681, + 503, + 703 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 204, + 694 + ], + "score": 1.0, + "content": "Table 2: Test error rates", + "type": "text" + }, + { + "bbox": [ + 204, + 682, + 220, + 692 + ], + "score": 0.78, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 691, + 201, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 201, + 703 + ], + "score": 1.0, + "content": "feedforward controller.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + } + ], + "index": 23.75 + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 501, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 720, + 501, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 501, + 734 + ], + "score": 1.0, + "content": "We also provide further experiments investigating different extensions on D-NTM in the appendix.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 119, + 75, + 492, + 237 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 119, + 75, + 492, + 237 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 75, + 492, + 237 + ], + "spans": [ + { + "bbox": [ + 119, + 75, + 492, + 237 + ], + "score": 0.982, + "html": "
DMN+L1-step CBA1-step Soft D-NTM1-step Discrete D-NTM3-steps LBA* NTM3-steps CBA NTM3-steps Soft D-NTM3-steps Discrete D-NTM
Task 1LSTM 0.00MemN2N 0.000.00NTM 16.30NTM 16.885.416.660.000.000.000.00
281.900.300.3057.0855.7058.5456.0461.6759.3846.6662.29
383.102.101.1074.1655.0074.5872.0883.5465.2147.0841.45
40.200.000.000.000.000.000.000.000.000.000.00
51.200.800.501.4620.411.661.040.831.461.251.45
651.800.100.0023.3321.0440.2044.7948.1354.8020.6211.04
724.902.002.4021.6721.6719.1619.587.9237.707.295.62
834.100.900.0025.7621.0512.5818.4625.388.8211.020.74
920.200.300.0024.7924.1736.6634.3737.800.0039.3732.50
1030.100.000.0041.4633.1352.2950.8356.2523.7520.0020.83
1110.300.100.0018.9631.8831.454.163.960.2830.6216.87
1223.400.000.0025.8330.007.706.6628.7523.755.414.58
136.100.000.006.675.635.622.295.8383.137.915.00
1481.000.100.2058.5459.1760.0063.7561.8857.7158.1260.20
1578.700.000.0036.4642.3036.8739.2735.6221.8836.0440.26
1651.9051.8045.3071.1571.1549.1651.3546.1550.0046.0445.41
1750.1018.604.2043.7543.7517.9116.0443.7556.2521.259.16
186.805.302.103.9647.503.953.5447.5047.506.871.66
1990.302.300.0075.8971.5173.7464.6361.5663.6575.8876.66
202.100.000.001.250.002.703.120.400.003.330.00
Avg.Err.36.414.242.8131.4233.6029.5127.9332.8532.7624.2421.79
", + "type": "table", + "image_path": "2d8d524f20ba7886f2d7b3b5ca3db3d355229f453ab7efd976a9be7efb577b54.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 75, + 492, + 129.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 129.0, + 492, + 183.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 183.0, + 492, + 237.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 504, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 204, + 257 + ], + "score": 1.0, + "content": "Table 1: Test error rates", + "type": "text" + }, + { + "bbox": [ + 204, + 245, + 220, + 256 + ], + "score": 0.8, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 245, + 504, + 257 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "the GRU and feedforward controller. FF stands for the experiments that are conducted with feedforward", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 221, + 277 + ], + "score": 1.0, + "content": "controller. Let us, note that", + "type": "text" + }, + { + "bbox": [ + 222, + 266, + 247, + 276 + ], + "score": 0.67, + "content": "\\mathrm { L B A ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 266, + 505, + 277 + ], + "score": 1.0, + "content": "refers to NTM that uses both LBA and CBA. In this table, we", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "score": 1.0, + "content": "compare multi-step vs single-step addressing, original NTM with location based+content based", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 287, + 478, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 478, + 299 + ], + "score": 1.0, + "content": "addressing vs only content based addressing, and discrete vs continuous addressing on bAbI.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 245, + 506, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "Lists/Sets and 11 - Basic Coreference. Furthermore, this is in line with an earlier observation in (Xu et al.,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "2015), where discrete addressing was found to generalize better in the task of image caption generation.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 324, + 506, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 504, + 384 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "In Table 2, we also observe that the D-NTM with the feedforward controller and discrete attention", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "performs worse than LSTM and D-NTM with continuous-attention. However, when the proposed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 373, + 462, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 462, + 385 + ], + "score": 1.0, + "content": "curriculum strategy from Sec. 4 is used, the average test error drops from 68.30 to 37.79.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 351, + 506, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "score": 1.0, + "content": "We empirically found training of the feedforward controller more difficult than that of the recurrent", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 400, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 411 + ], + "score": 1.0, + "content": "controller. We train our feedforward controller based models four times longer (in terms of the number", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 411, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 421 + ], + "score": 1.0, + "content": "of updates) than the recurrent controller based ones in order to ensure that they are converged for most", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 421, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 432 + ], + "score": 1.0, + "content": "of the tasks. On the other hand, the models trained with the GRU controller overfit on bAbI tasks", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "very quickly. For example, on tasks 3 and 16 the feedforward controller based model underfits (i.e.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 442, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 453 + ], + "score": 1.0, + "content": "high training loss) at the end of the training, whereas with the same number of units the model with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 451, + 389, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 389, + 464 + ], + "score": 1.0, + "content": "the GRU controller can overfit on those tasks after 3,000 updates only.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 389, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 503, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "When our results are compared to the variants of the memory network Weston et al. (2015b) (MemN2N", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 477, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 123, + 492 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 479, + 153, + 489 + ], + "score": 0.58, + "content": "\\mathrm { D M N + }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 477, + 505, + 492 + ], + "score": 1.0, + "content": "), we notice a significant performance gap. We attribute this gap to the difficulty in learning", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 489, + 271, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 271, + 501 + ], + "score": 1.0, + "content": "to manipulate and store a complex input.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 468, + 505, + 501 + ] + }, + { + "type": "table", + "bbox": [ + 238, + 511, + 372, + 673 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 238, + 511, + 372, + 673 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 238, + 511, + 372, + 673 + ], + "spans": [ + { + "bbox": [ + 238, + 511, + 372, + 673 + ], + "score": 0.96, + "html": "
TaskFF Soft D-NTMFF Discrete D-NTMFF Discrete* D-NTM
14.3881.6714.79
227.576.6776.67
371.2579.3870.83
40.0078.6544.06
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.62
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.92
1939.4771.4671.56
200.076.569.79
Avg.Err.12.8168.3037.79
", + "type": "table", + "image_path": "15be19b050d013aa2678635d887fe055e193f11fbac6d6b0f3bc9e3da80e1f5c.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 238, + 511, + 372, + 673 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 105, + 681, + 503, + 703 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 204, + 694 + ], + "score": 1.0, + "content": "Table 2: Test error rates", + "type": "text" + }, + { + "bbox": [ + 204, + 682, + 220, + 692 + ], + "score": 0.78, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 691, + 201, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 201, + 703 + ], + "score": 1.0, + "content": "feedforward controller.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + } + ], + "index": 23.75 + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 501, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 720, + 501, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 501, + 734 + ], + "score": 1.0, + "content": "We also provide further experiments investigating different extensions on D-NTM in the appendix.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 720, + 501, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 228, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 230, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 230, + 96 + ], + "score": 1.0, + "content": "7.2 SEQUENTIAL pMNIST", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "In sequential MNIST task, the pixels of the MNIST digits are provided to the model in scan line order,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 125 + ], + "score": 1.0, + "content": "left to right and top to bottom (Le et al., 2015). At the end of sequence of pixels, the model predicts", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 1.0, + "content": "the label of the digit in the sequence of pixels. We experiment D-NTM on the variation of sequential", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "MNIST where the order of the pixels is randomly shuffled, we call this task as permuted MNIST", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 145, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 109, + 145, + 147, + 156 + ], + "score": 0.47, + "content": "( p \\mathrm { M N I S T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 145, + 506, + 157 + ], + "score": 1.0, + "content": ". An important contribution of this task to our paper, in particular, is to measure the model’s", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 155, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 477, + 169 + ], + "score": 1.0, + "content": "ability to perform well when dealing with long-term dependencies. We report our results in Table", + "type": "text" + }, + { + "bbox": [ + 478, + 156, + 488, + 167 + ], + "score": 0.6, + "content": "3 ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 155, + 506, + 169 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "observe improvements over other models that we compare against. In Table 3, ”discrete addressing with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "MAB” refers to D-NTM model using REINFORCE with baseline computed from moving averages of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "the reward. Discrete addressing with IB refers to D-NTM using REINFORCE with input-based baseline.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 213, + 207, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 208, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 208, + 227 + ], + "score": 1.0, + "content": "7.3 NTM TOY TASKS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 505, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "score": 1.0, + "content": "We explore the possibility of using D-NTM to solve algorithmic tasks such as copy and associative", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "recall tasks. We train our model on the same lengths of sequences that is experimented in (Graves", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "score": 1.0, + "content": "et al., 2014). We report our results in Table 4. We find out that D-NTM using continuous-attention", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 264, + 372, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 264, + 372, + 277 + ], + "score": 1.0, + "content": "can successfully learn the ”Copy” and ”Associative Recall” tasks.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "In Table 4, we train our model on sequences of the same length as the experiments in (Graves et al., 2014)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "and test the model on the sequences of the maximum length seen during the training. We consider model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "score": 1.0, + "content": "to be successful on copy or associative recall if its validation cost (binary cross-entropy) is lower than", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 313, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 324 + ], + "score": 1.0, + "content": "0.02 over the sequences of maximum length seen during the training. We set the threshold to 0.02 to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "determine whether a model is successful on a task. Because empirically we observe that the models have", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "higher validation costs perform badly in terms of generalization over the longer sequences. ”D-NTM", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "discrete” model in this table is trained with REINFORCE using moving averages to estimate the baseline.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 110, + 367, + 303, + 479 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 110, + 367, + 303, + 479 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 367, + 303, + 479 + ], + "spans": [ + { + "bbox": [ + 110, + 367, + 303, + 479 + ], + "score": 0.978, + "html": "
Test Acc
D-NTMdiscrete MAB89.6
D-NTMdiscrete IB92.3
Soft D-NTM93.4
NTM90.9
I-RNN (Le et al., 2015)82.0
Zoneout (Krueger et al.,2016)93.1
LSTM (Krueger et al., 2016)89.8
Unitary-RNN (Arjovsky et al., 2015)91.4
Recurrent Dropout (Krueger et al., 2016)92.5
", + "type": "table", + "image_path": "14592f49a2c7ebec424d694080a5007ed1bdca9bbff954eff64d83cf13a9c5d1.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 110, + 367, + 303, + 381.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 110, + 381.0, + 303, + 395.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 110, + 395.0, + 303, + 409.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 110, + 409.0, + 303, + 423.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 110, + 423.0, + 303, + 437.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 110, + 437.0, + 303, + 451.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 110, + 451.0, + 303, + 465.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 110, + 465.0, + 303, + 479.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 147, + 488, + 266, + 499 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 487, + 266, + 501 + ], + "spans": [ + { + "bbox": [ + 146, + 487, + 226, + 501 + ], + "score": 1.0, + "content": "Table 3: Sequential", + "type": "text" + }, + { + "bbox": [ + 226, + 488, + 264, + 499 + ], + "score": 0.39, + "content": "p { \\mathrm { M N I S T } } .", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 487, + 266, + 501 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + } + ], + "index": 29.75 + }, + { + "type": "table", + "bbox": [ + 306, + 398, + 501, + 449 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 306, + 398, + 501, + 449 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 306, + 398, + 501, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 501, + 449 + ], + "score": 0.971, + "html": "
Copy TasksAssociative Recall
Soft D-NTMSuccessSuccess
D-NTMdiscreteSuccessFailure
NTMSuccessSuccess
", + "type": "table", + "image_path": "d69ed6640d1288038d9eaeb3dd7f3152bddf9f7e0e4fa3cb18154291caeb585f.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 306, + 398, + 501, + 415.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 306, + 415.0, + 501, + 432.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 306, + 432.0, + 501, + 449.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 335, + 457, + 438, + 468 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 455, + 439, + 470 + ], + "spans": [ + { + "bbox": [ + 334, + 455, + 439, + 470 + ], + "score": 1.0, + "content": "Table 4: NTM Toy Tasks.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32.0 + }, + { + "type": "title", + "bbox": [ + 107, + 524, + 303, + 536 + ], + "lines": [ + { + "bbox": [ + 104, + 522, + 304, + 539 + ], + "spans": [ + { + "bbox": [ + 104, + 522, + 304, + 539 + ], + "score": 1.0, + "content": "8 CONCLUSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "In this paper we extend neural Turing machines (NTM) by introducing a learnable addressing scheme", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 558, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 507, + 573 + ], + "score": 1.0, + "content": "which allows the NTM to be capable of performing highly nonlinear location-based addressing.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "score": 1.0, + "content": "This extension, to which we refer by dynamic NTM (D-NTM), is extensively tested with various", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "configurations, including different addressing mechanisms (continuous vs. discrete) and different", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "number of addressing steps, on the Facebook bAbI tasks. This is the first time an NTM-type model", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 601, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 614 + ], + "score": 1.0, + "content": "was tested on this task, and we observe that the NTM, especially the proposed D-NTM, performs better", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 610, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 625 + ], + "score": 1.0, + "content": "than vanilla LSTM-RNN. Furthermore, the experiments revealed that the discrete, discrete addressing", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "score": 1.0, + "content": "works better than the continuous addressing with the GRU controller, and our analysis reveals that", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 632, + 405, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 405, + 645 + ], + "score": 1.0, + "content": "this is the case when the task requires precise retrieval of memory content.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Our experiments show that the NTM-based models can be weaker than other variants of memory", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 659, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 671 + ], + "score": 1.0, + "content": "networks which do not learn but have an explicit mechanism of storing incoming facts as they are. We", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "conjecture that this is due to the difficulty in learning how to write, manipulate and delete the content", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "score": 1.0, + "content": "of memory. Despite this difficulty, we find the NTM-based approach, such as the proposed D-NTM,", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 703, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 117, + 700, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 117, + 700, + 506, + 715 + ], + "score": 1.0, + "content": "3Let us note that, the current state of art on this task is recurrent batch normalization with LSTM (Cooijmans", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 168, + 724 + ], + "score": 1.0, + "content": "et al., 2016) with", + "type": "text" + }, + { + "bbox": [ + 168, + 712, + 192, + 722 + ], + "score": 0.86, + "content": "9 5 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "accuracy. It is possible to use recurrent batch normalization in our model and potentially", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 251, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 251, + 732 + ], + "score": 1.0, + "content": "improve our results on this task as well.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 228, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 230, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 230, + 96 + ], + "score": 1.0, + "content": "7.2 SEQUENTIAL pMNIST", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "In sequential MNIST task, the pixels of the MNIST digits are provided to the model in scan line order,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 125 + ], + "score": 1.0, + "content": "left to right and top to bottom (Le et al., 2015). At the end of sequence of pixels, the model predicts", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 1.0, + "content": "the label of the digit in the sequence of pixels. We experiment D-NTM on the variation of sequential", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "MNIST where the order of the pixels is randomly shuffled, we call this task as permuted MNIST", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 145, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 109, + 145, + 147, + 156 + ], + "score": 0.47, + "content": "( p \\mathrm { M N I S T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 145, + 506, + 157 + ], + "score": 1.0, + "content": ". An important contribution of this task to our paper, in particular, is to measure the model’s", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 155, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 477, + 169 + ], + "score": 1.0, + "content": "ability to perform well when dealing with long-term dependencies. We report our results in Table", + "type": "text" + }, + { + "bbox": [ + 478, + 156, + 488, + 167 + ], + "score": 0.6, + "content": "3 ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 155, + 506, + 169 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "observe improvements over other models that we compare against. In Table 3, ”discrete addressing with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "score": 1.0, + "content": "MAB” refers to D-NTM model using REINFORCE with baseline computed from moving averages of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "the reward. Discrete addressing with IB refers to D-NTM using REINFORCE with input-based baseline.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 103, + 506, + 200 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 213, + 207, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 208, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 208, + 227 + ], + "score": 1.0, + "content": "7.3 NTM TOY TASKS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 505, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "score": 1.0, + "content": "We explore the possibility of using D-NTM to solve algorithmic tasks such as copy and associative", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "recall tasks. We train our model on the same lengths of sequences that is experimented in (Graves", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "score": 1.0, + "content": "et al., 2014). We report our results in Table 4. We find out that D-NTM using continuous-attention", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 264, + 372, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 264, + 372, + 277 + ], + "score": 1.0, + "content": "can successfully learn the ”Copy” and ”Associative Recall” tasks.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 234, + 505, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 281, + 505, + 355 + ], + "lines": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "In Table 4, we train our model on sequences of the same length as the experiments in (Graves et al., 2014)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "and test the model on the sequences of the maximum length seen during the training. We consider model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 314 + ], + "score": 1.0, + "content": "to be successful on copy or associative recall if its validation cost (binary cross-entropy) is lower than", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 313, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 324 + ], + "score": 1.0, + "content": "0.02 over the sequences of maximum length seen during the training. We set the threshold to 0.02 to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "determine whether a model is successful on a task. Because empirically we observe that the models have", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "higher validation costs perform badly in terms of generalization over the longer sequences. ”D-NTM", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "discrete” model in this table is trained with REINFORCE using moving averages to estimate the baseline.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 281, + 506, + 356 + ] + }, + { + "type": "table", + "bbox": [ + 110, + 367, + 303, + 479 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 110, + 367, + 303, + 479 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 367, + 303, + 479 + ], + "spans": [ + { + "bbox": [ + 110, + 367, + 303, + 479 + ], + "score": 0.978, + "html": "
Test Acc
D-NTMdiscrete MAB89.6
D-NTMdiscrete IB92.3
Soft D-NTM93.4
NTM90.9
I-RNN (Le et al., 2015)82.0
Zoneout (Krueger et al.,2016)93.1
LSTM (Krueger et al., 2016)89.8
Unitary-RNN (Arjovsky et al., 2015)91.4
Recurrent Dropout (Krueger et al., 2016)92.5
", + "type": "table", + "image_path": "14592f49a2c7ebec424d694080a5007ed1bdca9bbff954eff64d83cf13a9c5d1.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 110, + 367, + 303, + 381.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 110, + 381.0, + 303, + 395.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 110, + 395.0, + 303, + 409.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 110, + 409.0, + 303, + 423.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 110, + 423.0, + 303, + 437.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 110, + 437.0, + 303, + 451.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 110, + 451.0, + 303, + 465.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 110, + 465.0, + 303, + 479.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 147, + 488, + 266, + 499 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 487, + 266, + 501 + ], + "spans": [ + { + "bbox": [ + 146, + 487, + 226, + 501 + ], + "score": 1.0, + "content": "Table 3: Sequential", + "type": "text" + }, + { + "bbox": [ + 226, + 488, + 264, + 499 + ], + "score": 0.39, + "content": "p { \\mathrm { M N I S T } } .", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 487, + 266, + 501 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + } + ], + "index": 29.75 + }, + { + "type": "table", + "bbox": [ + 306, + 398, + 501, + 449 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 306, + 398, + 501, + 449 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 306, + 398, + 501, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 501, + 449 + ], + "score": 0.971, + "html": "
Copy TasksAssociative Recall
Soft D-NTMSuccessSuccess
D-NTMdiscreteSuccessFailure
NTMSuccessSuccess
", + "type": "table", + "image_path": "d69ed6640d1288038d9eaeb3dd7f3152bddf9f7e0e4fa3cb18154291caeb585f.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 306, + 398, + 501, + 415.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 306, + 415.0, + 501, + 432.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 306, + 432.0, + 501, + 449.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 335, + 457, + 438, + 468 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 455, + 439, + 470 + ], + "spans": [ + { + "bbox": [ + 334, + 455, + 439, + 470 + ], + "score": 1.0, + "content": "Table 4: NTM Toy Tasks.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32.0 + }, + { + "type": "title", + "bbox": [ + 107, + 524, + 303, + 536 + ], + "lines": [ + { + "bbox": [ + 104, + 522, + 304, + 539 + ], + "spans": [ + { + "bbox": [ + 104, + 522, + 304, + 539 + ], + "score": 1.0, + "content": "8 CONCLUSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "In this paper we extend neural Turing machines (NTM) by introducing a learnable addressing scheme", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 558, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 507, + 573 + ], + "score": 1.0, + "content": "which allows the NTM to be capable of performing highly nonlinear location-based addressing.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 505, + 582 + ], + "score": 1.0, + "content": "This extension, to which we refer by dynamic NTM (D-NTM), is extensively tested with various", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "configurations, including different addressing mechanisms (continuous vs. discrete) and different", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "number of addressing steps, on the Facebook bAbI tasks. This is the first time an NTM-type model", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 601, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 614 + ], + "score": 1.0, + "content": "was tested on this task, and we observe that the NTM, especially the proposed D-NTM, performs better", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 610, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 625 + ], + "score": 1.0, + "content": "than vanilla LSTM-RNN. Furthermore, the experiments revealed that the discrete, discrete addressing", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "score": 1.0, + "content": "works better than the continuous addressing with the GRU controller, and our analysis reveals that", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 632, + 405, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 405, + 645 + ], + "score": 1.0, + "content": "this is the case when the task requires precise retrieval of memory content.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 549, + 507, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Our experiments show that the NTM-based models can be weaker than other variants of memory", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 659, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 671 + ], + "score": 1.0, + "content": "networks which do not learn but have an explicit mechanism of storing incoming facts as they are. We", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "conjecture that this is due to the difficulty in learning how to write, manipulate and delete the content", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "score": 1.0, + "content": "of memory. Despite this difficulty, we find the NTM-based approach, such as the proposed D-NTM,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "to be a better, future-proof approach, because it can scale to a much longer horizon (where it becomes", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 307, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 307, + 105 + ], + "score": 1.0, + "content": "impossible to explicitly store all the experiences.)", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 648, + 507, + 693 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "to be a better, future-proof approach, because it can scale to a much longer horizon (where it becomes", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 307, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 307, + 105 + ], + "score": 1.0, + "content": "impossible to explicitly store all the experiences.)", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 109, + 505, + 141 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 120, + 122 + ], + "score": 1.0, + "content": "On", + "type": "text" + }, + { + "bbox": [ + 121, + 110, + 158, + 121 + ], + "score": 0.38, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 109, + 505, + 122 + ], + "score": 1.0, + "content": "task, we show that our model can outperform other similar type of approaches proposed", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "score": 1.0, + "content": "to deal with the long-term dependencies. On copy and associative recall tasks, we show that our model", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 130, + 463, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 130, + 463, + 142 + ], + "score": 1.0, + "content": "can solve the algorithmic problems that are proposed to solve with NTM type of models.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 146, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 106, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "The success of both the learnable address and the discrete addressing scheme suggests two future", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 507, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 507, + 170 + ], + "score": 1.0, + "content": "research directions. First, we should try both of these schemes in a wider array of memory-based models,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "as they are not specific to the neural Turing machines. Second, the proposed D-NTM needs to be", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 178, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 506, + 190 + ], + "score": 1.0, + "content": "evaluated on a diverse set of applications, such as text summarization (Rush et al., 2015), visual question-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 189, + 507, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 507, + 200 + ], + "score": 1.0, + "content": "answering (Antol et al., 2015) and machine translation, in order to make a more concrete conclusion.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 108, + 215, + 175, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 176, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 176, + 228 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 105, + 231, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 254, + 486, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 486, + 266 + ], + "score": 1.0, + "content": "Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pp. 2425–2433, 2015.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 270, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 270, + 506, + 286 + ], + "score": 1.0, + "content": "Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. arXiv", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 113, + 282, + 254, + 294 + ], + "spans": [ + { + "bbox": [ + 113, + 282, + 254, + 294 + ], + "score": 1.0, + "content": "preprint arXiv:1511.06464, 2015.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 298, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 104, + 298, + 505, + 313 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 311, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 506, + 323 + ], + "score": 1.0, + "content": "learning to align and translate. In Proceedings Of The International Conference on Representation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 320, + 236, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 236, + 334 + ], + "score": 1.0, + "content": "Learning (ICLR 2015), 2015.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 348, + 446, + 362 + ], + "spans": [ + { + "bbox": [ + 114, + 348, + 446, + 362 + ], + "score": 1.0, + "content": "descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 376, + 421, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 376, + 421, + 390 + ], + "score": 1.0, + "content": "answering with memory networks. arXiv preprint arXiv:1506.02075, 2015.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 405, + 507, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 405, + 507, + 419 + ], + "score": 1.0, + "content": "Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 415, + 274, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 274, + 428 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 444, + 471, + 456 + ], + "spans": [ + { + "bbox": [ + 116, + 444, + 471, + 456 + ], + "score": 1.0, + "content": "Attention-based models for speech recognition. arXiv preprint arXiv:1506.07503, 2015.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 507, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 507, + 475 + ], + "score": 1.0, + "content": "Tim Cooijmans, Nicolas Ballas, Cesar Laurent, and Aaron Courville. Recurrent batch normalization. ´", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 472, + 279, + 484 + ], + "spans": [ + { + "bbox": [ + 115, + 472, + 279, + 484 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1603.09025, 2016.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 488, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 104, + 488, + 506, + 503 + ], + "score": 1.0, + "content": "Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 499, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 499, + 507, + 514 + ], + "score": 1.0, + "content": "Szlam, and Jason Weston. Evaluating prerequisite qualities for learning end-to-end dialog systems.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 510, + 239, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 510, + 239, + 523 + ], + "score": 1.0, + "content": "CoRR, abs/1511.06931, 2015.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. Book in preparation for MIT", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 538, + 382, + 552 + ], + "spans": [ + { + "bbox": [ + 115, + 538, + 382, + 552 + ], + "score": 1.0, + "content": "Press, 2016. URL http://www.deeplearningbook.org.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 555, + 507, + 569 + ], + "spans": [ + { + "bbox": [ + 104, + 555, + 507, + 569 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 566, + 143, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 143, + 579 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 595, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 116, + 595, + 507, + 607 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. ´", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 605, + 507, + 619 + ], + "spans": [ + { + "bbox": [ + 115, + 605, + 507, + 619 + ], + "score": 1.0, + "content": "Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 615, + 181, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 615, + 181, + 627 + ], + "score": 1.0, + "content": "471–476, 2016.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 642, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 114, + 642, + 507, + 658 + ], + "score": 1.0, + "content": "transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 653, + 191, + 666 + ], + "spans": [ + { + "bbox": [ + 115, + 653, + 191, + 666 + ], + "score": 1.0, + "content": "1819–1827, 2015.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 671, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 671, + 506, + 685 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 682, + 279, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 682, + 279, + 695 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1603.00391, 2016.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "Karl Moritz Hermann, Toma´s Ko ˇ cisk ˇ y, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa `", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. arXiv preprint", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 219, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 219, + 732 + ], + "score": 1.0, + "content": "arXiv:1506.03340, 2015.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 29.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 104 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 105 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 109, + 505, + 141 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 120, + 122 + ], + "score": 1.0, + "content": "On", + "type": "text" + }, + { + "bbox": [ + 121, + 110, + 158, + 121 + ], + "score": 0.38, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 109, + 505, + 122 + ], + "score": 1.0, + "content": "task, we show that our model can outperform other similar type of approaches proposed", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "score": 1.0, + "content": "to deal with the long-term dependencies. On copy and associative recall tasks, we show that our model", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 130, + 463, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 130, + 463, + 142 + ], + "score": 1.0, + "content": "can solve the algorithmic problems that are proposed to solve with NTM type of models.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 109, + 505, + 142 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 146, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 106, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "The success of both the learnable address and the discrete addressing scheme suggests two future", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 507, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 507, + 170 + ], + "score": 1.0, + "content": "research directions. First, we should try both of these schemes in a wider array of memory-based models,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "as they are not specific to the neural Turing machines. Second, the proposed D-NTM needs to be", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 178, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 506, + 190 + ], + "score": 1.0, + "content": "evaluated on a diverse set of applications, such as text summarization (Rush et al., 2015), visual question-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 189, + 507, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 507, + 200 + ], + "score": 1.0, + "content": "answering (Antol et al., 2015) and machine translation, in order to make a more concrete conclusion.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 146, + 507, + 200 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 215, + 175, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 176, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 176, + 228 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "list", + "bbox": [ + 105, + 231, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick,", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 254, + 486, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 486, + 266 + ], + "score": 1.0, + "content": "Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pp. 2425–2433, 2015.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 270, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 270, + 506, + 286 + ], + "score": 1.0, + "content": "Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent neural networks. arXiv", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 282, + 254, + 294 + ], + "spans": [ + { + "bbox": [ + 113, + 282, + 254, + 294 + ], + "score": 1.0, + "content": "preprint arXiv:1511.06464, 2015.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 298, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 104, + 298, + 505, + 313 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 311, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 506, + 323 + ], + "score": 1.0, + "content": "learning to align and translate. In Proceedings Of The International Conference on Representation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 320, + 236, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 236, + 334 + ], + "score": 1.0, + "content": "Learning (ICLR 2015), 2015.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 348, + 446, + 362 + ], + "spans": [ + { + "bbox": [ + 114, + 348, + 446, + 362 + ], + "score": 1.0, + "content": "descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 376, + 421, + 390 + ], + "spans": [ + { + "bbox": [ + 115, + 376, + 421, + 390 + ], + "score": 1.0, + "content": "answering with memory networks. arXiv preprint arXiv:1506.02075, 2015.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 405, + 507, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 405, + 507, + 419 + ], + "score": 1.0, + "content": "Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 415, + 274, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 274, + 428 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio.", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 444, + 471, + 456 + ], + "spans": [ + { + "bbox": [ + 116, + 444, + 471, + 456 + ], + "score": 1.0, + "content": "Attention-based models for speech recognition. arXiv preprint arXiv:1506.07503, 2015.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 460, + 507, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 507, + 475 + ], + "score": 1.0, + "content": "Tim Cooijmans, Nicolas Ballas, Cesar Laurent, and Aaron Courville. Recurrent batch normalization. ´", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 472, + 279, + 484 + ], + "spans": [ + { + "bbox": [ + 115, + 472, + 279, + 484 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1603.09025, 2016.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 488, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 104, + 488, + 506, + 503 + ], + "score": 1.0, + "content": "Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 499, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 499, + 507, + 514 + ], + "score": 1.0, + "content": "Szlam, and Jason Weston. Evaluating prerequisite qualities for learning end-to-end dialog systems.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 510, + 239, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 510, + 239, + 523 + ], + "score": 1.0, + "content": "CoRR, abs/1511.06931, 2015.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. Book in preparation for MIT", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 538, + 382, + 552 + ], + "spans": [ + { + "bbox": [ + 115, + 538, + 382, + 552 + ], + "score": 1.0, + "content": "Press, 2016. URL http://www.deeplearningbook.org.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 555, + 507, + 569 + ], + "spans": [ + { + "bbox": [ + 104, + 555, + 507, + 569 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401,", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 566, + 143, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 143, + 579 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 595, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 116, + 595, + 507, + 607 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. ´", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 605, + 507, + 619 + ], + "spans": [ + { + "bbox": [ + 115, + 605, + 507, + 619 + ], + "score": 1.0, + "content": "Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626):", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 615, + 181, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 615, + 181, + 627 + ], + "score": 1.0, + "content": "471–476, 2016.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 642, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 114, + 642, + 507, + 658 + ], + "score": 1.0, + "content": "transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 653, + 191, + 666 + ], + "spans": [ + { + "bbox": [ + 115, + 653, + 191, + 666 + ], + "score": 1.0, + "content": "1819–1827, 2015.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 671, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 671, + 506, + 685 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions.", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 682, + 279, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 682, + 279, + 695 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1603.00391, 2016.", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "Karl Moritz Hermann, Toma´s Ko ˇ cisk ˇ y, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa `", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. arXiv preprint", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 219, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 219, + 732 + ], + "score": 1.0, + "content": "arXiv:1506.03340, 2015.", + "type": "text" + } + ], + "index": 48, + "is_list_end_line": true + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 234, + 507, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 92, + 502, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 92, + 502, + 105 + ], + "score": 1.0, + "content": "children’s books with explicit memory representations. arXiv preprint arXiv:1511.02301, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "Sepp Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Diploma, Technische Universitat¨", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 216, + 134 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 216, + 134 + ], + "score": 1.0, + "content": "Munchen ¨ , pp. 91, 1991.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural Computation, 9(8):", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 150, + 191, + 162 + ], + "spans": [ + { + "bbox": [ + 114, + 150, + 191, + 162 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "score": 1.0, + "content": "Peter J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73–101, 03 1964.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 187, + 504, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 197, + 443, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 443, + 210 + ], + "score": 1.0, + "content": "nets. In Advances in Neural Information Processing Systems, pp. 190–198, 2015.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 225, + 143, + 239 + ], + "spans": [ + { + "bbox": [ + 114, + 225, + 143, + 239 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 244, + 505, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 258 + ], + "score": 1.0, + "content": "David Krueger, Tegan Maharaj, Janos Kram ´ ar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary ´", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "score": 1.0, + "content": "Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 475, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 475, + 278 + ], + "score": 1.0, + "content": "rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 504, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "Quoc V Le, Navdeep Jaitly, and Geoffrey E Hinton. A simple way to initialize recurrent networks", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 376, + 306 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 376, + 306 + ], + "score": 1.0, + "content": "of rectified linear units. arXiv preprint arXiv:1504.00941, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 504, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 324, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 506, + 335 + ], + "score": 1.0, + "content": "neural machine translation. In Proceedings Of The Conference on Empirical Methods for Natural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 334, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 116, + 334, + 297, + 346 + ], + "score": 1.0, + "content": "Language Processing (EMNLP 2015), 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 504, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "Key-value memory networks for directly reading documents. CoRR, abs/1606.03126, 2016. URL", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 372, + 308, + 385 + ], + "spans": [ + { + "bbox": [ + 116, + 372, + 308, + 385 + ], + "score": 1.0, + "content": "http://arxiv.org/abs/1606.03126.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 104, + 390, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 402, + 249, + 413 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 249, + 413 + ], + "score": 1.0, + "content": "preprint arXiv:1402.0030, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 507, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 507, + 433 + ], + "score": 1.0, + "content": "Jack W Rae, Jonathan J Hunt, Tim Harley, Ivo Danihelka, Andrew Senior, Greg Wayne, Alex Graves,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 430, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 507, + 443 + ], + "score": 1.0, + "content": "and Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 440, + 229, + 452 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 229, + 452 + ], + "score": 1.0, + "content": "In Advances in NIPS. 2016.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 458, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Scott Reed and Nando de Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 468, + 144, + 482 + ], + "spans": [ + { + "bbox": [ + 114, + 468, + 144, + 482 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "Alexander M. Rush, Sumit Chopra, and Jason Weston. A neural attention model for abstractive sentence", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 497, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 115, + 497, + 506, + 512 + ], + "score": 1.0, + "content": "summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 508, + 474, + 521 + ], + "spans": [ + { + "bbox": [ + 115, + 508, + 474, + 521 + ], + "score": 1.0, + "content": "Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pp. 379–389, 2015.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 104, + 526, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-shot", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 537, + 487, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 537, + 487, + 550 + ], + "score": 1.0, + "content": "learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. Building", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 577, + 405, + 589 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 405, + 589 + ], + "score": 1.0, + "content": "of the 30th AAAI Conference on Artificial Intelligence (AAAI-16), 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 605, + 279, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 279, + 617 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1503.08895, 2015.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 506, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "Guo-Zheng Sun, C. Lee Giles, and Hsing-Hen Chen. The neural network pushdown automaton:", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 115, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "Architecture, dynamics and training. In Adaptive Processing of Sequences and Data Structures,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 644, + 402, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 644, + 402, + 657 + ], + "score": 1.0, + "content": "International Summer School on Neural Networks, pp. 296–345, 1997.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 105, + 663, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "Oriol Vinyals and Quoc Le. A neural conversational model. arXiv preprint arXiv:1506.05869, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "score": 1.0, + "content": "Jason Weston, Antoine Bordes, Sumit Chopra, and Tomas Mikolov. Towards ai-complete question", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 691, + 452, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 452, + 704 + ], + "score": 1.0, + "content": "answering: a set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015a.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Proceedings Of The", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 457, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 457, + 733 + ], + "score": 1.0, + "content": "International Conference on Representation Learning (ICLR 2015), 2015b. In Press.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 92, + 502, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 92, + 502, + 105 + ], + "score": 1.0, + "content": "children’s books with explicit memory representations. arXiv preprint arXiv:1511.02301, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "Sepp Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Diploma, Technische Universitat¨", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 216, + 134 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 216, + 134 + ], + "score": 1.0, + "content": "Munchen ¨ , pp. 91, 1991.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 110, + 506, + 134 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural Computation, 9(8):", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 150, + 191, + 162 + ], + "spans": [ + { + "bbox": [ + 114, + 150, + 191, + 162 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 139, + 506, + 162 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "score": 1.0, + "content": "Peter J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73–101, 03 1964.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 168, + 505, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 187, + 504, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 197, + 443, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 443, + 210 + ], + "score": 1.0, + "content": "nets. In Advances in Neural Information Processing Systems, pp. 190–198, 2015.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 186, + 505, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 225, + 143, + 239 + ], + "spans": [ + { + "bbox": [ + 114, + 225, + 143, + 239 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 215, + 506, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 244, + 505, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 258 + ], + "score": 1.0, + "content": "David Krueger, Tegan Maharaj, Janos Kram ´ ar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary ´", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "score": 1.0, + "content": "Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Aaron Courville, et al. Zoneout: Regularizing", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 475, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 475, + 278 + ], + "score": 1.0, + "content": "rnns by randomly preserving hidden activations. arXiv preprint arXiv:1606.01305, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 243, + 506, + 278 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 504, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "Quoc V Le, Navdeep Jaitly, and Geoffrey E Hinton. A simple way to initialize recurrent networks", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 376, + 306 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 376, + 306 + ], + "score": 1.0, + "content": "of rectified linear units. arXiv preprint arXiv:1504.00941, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 284, + 505, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 504, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 324, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 506, + 335 + ], + "score": 1.0, + "content": "neural machine translation. In Proceedings Of The Conference on Empirical Methods for Natural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 334, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 116, + 334, + 297, + 346 + ], + "score": 1.0, + "content": "Language Processing (EMNLP 2015), 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 313, + 506, + 346 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 504, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "Key-value memory networks for directly reading documents. CoRR, abs/1606.03126, 2016. URL", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 372, + 308, + 385 + ], + "spans": [ + { + "bbox": [ + 116, + 372, + 308, + 385 + ], + "score": 1.0, + "content": "http://arxiv.org/abs/1606.03126.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 351, + 506, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 390, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 402, + 249, + 413 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 249, + 413 + ], + "score": 1.0, + "content": "preprint arXiv:1402.0030, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 390, + 505, + 413 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 507, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 507, + 433 + ], + "score": 1.0, + "content": "Jack W Rae, Jonathan J Hunt, Tim Harley, Ivo Danihelka, Andrew Senior, Greg Wayne, Alex Graves,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 430, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 507, + 443 + ], + "score": 1.0, + "content": "and Timothy P Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 440, + 229, + 452 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 229, + 452 + ], + "score": 1.0, + "content": "In Advances in NIPS. 2016.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 418, + 507, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 458, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Scott Reed and Nando de Freitas. Neural programmer-interpreters. arXiv preprint arXiv:1511.06279,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 468, + 144, + 482 + ], + "spans": [ + { + "bbox": [ + 114, + 468, + 144, + 482 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 458, + 506, + 482 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "Alexander M. Rush, Sumit Chopra, and Jason Weston. A neural attention model for abstractive sentence", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 497, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 115, + 497, + 506, + 512 + ], + "score": 1.0, + "content": "summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 508, + 474, + 521 + ], + "spans": [ + { + "bbox": [ + 115, + 508, + 474, + 521 + ], + "score": 1.0, + "content": "Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pp. 379–389, 2015.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 488, + 506, + 521 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 526, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-shot", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 537, + 487, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 537, + 487, + 550 + ], + "score": 1.0, + "content": "learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 526, + 505, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. Building", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 577, + 405, + 589 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 405, + 589 + ], + "score": 1.0, + "content": "of the 30th AAAI Conference on Artificial Intelligence (AAAI-16), 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 555, + 505, + 589 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 605, + 279, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 279, + 617 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1503.08895, 2015.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 594, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 506, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "Guo-Zheng Sun, C. Lee Giles, and Hsing-Hen Chen. The neural network pushdown automaton:", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 115, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "Architecture, dynamics and training. In Adaptive Processing of Sequences and Data Structures,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 644, + 402, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 644, + 402, + 657 + ], + "score": 1.0, + "content": "International Summer School on Neural Networks, pp. 296–345, 1997.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 623, + 506, + 657 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 663, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "Oriol Vinyals and Quoc Le. A neural conversational model. arXiv preprint arXiv:1506.05869, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 662, + 506, + 676 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "score": 1.0, + "content": "Jason Weston, Antoine Bordes, Sumit Chopra, and Tomas Mikolov. Towards ai-complete question", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 691, + 452, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 452, + 704 + ], + "score": 1.0, + "content": "answering: a set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015a.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 681, + 504, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Proceedings Of The", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 457, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 457, + 733 + ], + "score": 1.0, + "content": "International Conference on Representation Learning (ICLR 2015), 2015b. In Press.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 710, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 306, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 306, + 105 + ], + "score": 1.0, + "content": "learning. Machine Learning, 8:229–256, 1992.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "score": 1.0, + "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "score": 1.0, + "content": "question answering. CoRR, abs/1603.01417, 2016.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 139, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 154 + ], + "score": 1.0, + "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 161, + 443, + 174 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 443, + 174 + ], + "score": 1.0, + "content": "Of The International Conference on Representation Learning (ICLR 2015), 2015.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "Li Yao, Atousa Torabi, Kyunghyun Cho, Nicolas Ballas, Christopher Pal, Hugo Larochelle, and Aaron", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 115, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "Courville. Describing videos by exploiting temporal structure. In Computer Vision (ICCV), 2015", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 200, + 310, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 200, + 310, + 213 + ], + "score": 1.0, + "content": "IEEE International Conference on. IEEE, 2015.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "Wojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines. CoRR,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 230, + 210, + 241 + ], + "spans": [ + { + "bbox": [ + 116, + 230, + 210, + 241 + ], + "score": 1.0, + "content": "abs/1505.00521, 2015.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 247, + 504, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 504, + 261 + ], + "score": 1.0, + "content": "Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 258, + 345, + 271 + ], + "spans": [ + { + "bbox": [ + 115, + 258, + 345, + 271 + ], + "score": 1.0, + "content": "from examples. arXiv preprint arXiv:1511.07275, 2015.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 82, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 94, + 306, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 306, + 105 + ], + "score": 1.0, + "content": "learning. Machine Learning, 8:229–256, 1992.", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "score": 1.0, + "content": "Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 322, + 135 + ], + "score": 1.0, + "content": "question answering. CoRR, abs/1603.01417, 2016.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 139, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 154 + ], + "score": 1.0, + "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "Bengio. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 161, + 443, + 174 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 443, + 174 + ], + "score": 1.0, + "content": "Of The International Conference on Representation Learning (ICLR 2015), 2015.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "Li Yao, Atousa Torabi, Kyunghyun Cho, Nicolas Ballas, Christopher Pal, Hugo Larochelle, and Aaron", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 115, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "Courville. Describing videos by exploiting temporal structure. In Computer Vision (ICCV), 2015", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 200, + 310, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 200, + 310, + 213 + ], + "score": 1.0, + "content": "IEEE International Conference on. IEEE, 2015.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "Wojciech Zaremba and Ilya Sutskever. Reinforcement learning neural turing machines. CoRR,", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 230, + 210, + 241 + ], + "spans": [ + { + "bbox": [ + 116, + 230, + 210, + 241 + ], + "score": 1.0, + "content": "abs/1505.00521, 2015.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 247, + 504, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 504, + 261 + ], + "score": 1.0, + "content": "Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 258, + 345, + 271 + ], + "spans": [ + { + "bbox": [ + 115, + 258, + 345, + 271 + ], + "score": 1.0, + "content": "from examples. arXiv preprint arXiv:1511.07275, 2015.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 83, + 506, + 271 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 258, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 259, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 259, + 95 + ], + "score": 1.0, + "content": "A EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 320, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 320, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 320, + 118 + ], + "score": 1.0, + "content": "A.1 MODEL AND TRAINING DETAILS FOR BABI", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 125, + 381, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 125, + 382, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 382, + 138 + ], + "score": 1.0, + "content": "We use the same hyperparameters for all the tasks for a given model.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 147, + 244, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 245, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 245, + 160 + ], + "score": 1.0, + "content": "A.1.1 FACT REPRESENTATION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 179 + ], + "score": 1.0, + "content": "We use a recurrent neural network with GRU units to encode a variable-length fact into a fixed-size", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "vector representation. This allows the D-NTM to exploit the word ordering in each fact, unlike when", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 188, + 279, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 279, + 200 + ], + "score": 1.0, + "content": "facts are encoded as bag-of-words vectors.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 210, + 201, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 203, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 203, + 224 + ], + "score": 1.0, + "content": "A.1.2 CONTROLLER", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 229, + 505, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "We experiment with both a recurrent and feedforward neural network as the controller that generates", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 237, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 253 + ], + "score": 1.0, + "content": "the read and write weights. The controller has 180 units. We train our feed-forward controller using", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "noisy-tanh activation function (Gulcehre et al., 2016) since we were experiencing training difficulties", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "with sigmoid and tanh activation functions. We use both single-step and three-steps addressing with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 271, + 188, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 188, + 282 + ], + "score": 1.0, + "content": "our GRU controller.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 108, + 293, + 182, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 184, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 184, + 307 + ], + "score": 1.0, + "content": "A.1.3 MEMORY", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 108, + 312, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "The memory contains 120 memory cells. Each memory cell consists of a 16-dimensional address", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 322, + 258, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 258, + 335 + ], + "score": 1.0, + "content": "part and 28-dimensional content part.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 224, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 226, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 226, + 358 + ], + "score": 1.0, + "content": "A.1.4 TRAINING DETAILS", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 505, + 417 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 200, + 376 + ], + "score": 1.0, + "content": "We set aside a random", + "type": "text" + }, + { + "bbox": [ + 201, + 364, + 221, + 375 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "of the training examples as a validation set for each sub-task and use", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 104, + 373, + 506, + 388 + ], + "score": 1.0, + "content": "it for early-stopping and hyperparameter search. We train one D-NTM for each sub-task, using", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Adam (Kingma & Ba, 2014) with its learning rate set to 0.003 and 0.007 respectively for GRU", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "and Feedforward controller. The size of each minibatch is 160, and each minibatch is constructed", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 406, + 269, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 269, + 419 + ], + "score": 1.0, + "content": "uniform-randomly from the training set.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 107, + 430, + 387, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 388, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 388, + 443 + ], + "score": 1.0, + "content": "A.2 MODEL AND TRAINING DETAILS FOR SEQUENTIAL MNIST", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "On sequential MNIST task we try to keep the capacity of our model to be close to our baselines. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "use 100 GRU units in the controller and each content vector of size 8 and with address vectors of size", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 220, + 483 + ], + "score": 1.0, + "content": "8. We use a learning rate of", + "type": "text" + }, + { + "bbox": [ + 220, + 471, + 249, + 481 + ], + "score": 0.82, + "content": "1 e - 3", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "and trained the model with adam optimizer. We did not use the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 481, + 359, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 359, + 493 + ], + "score": 1.0, + "content": "read and write consistency regularization in any of our models.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 505, + 345, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 346, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 346, + 518 + ], + "score": 1.0, + "content": "A.3 MODEL AND TRAINING DETAILS FOR TOY TASKS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "score": 1.0, + "content": "On both copy and associative recall tasks, we try to keep the capacity of our model to be close to our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "baselines. We use 100 GRU units in the controller and each content vector of has a size of 8 and using", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 546, + 507, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 304, + 559 + ], + "score": 1.0, + "content": "address vector of size 8. We use a learning rate of", + "type": "text" + }, + { + "bbox": [ + 304, + 547, + 333, + 557 + ], + "score": 0.84, + "content": "1 e - 3", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 546, + 507, + 559 + ], + "score": 1.0, + "content": "and trained the model with adam optimizer.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "We did not use the read and write consistency regularization in any of our models. For the model with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 565, + 474, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 474, + 581 + ], + "score": 1.0, + "content": "the discrete attention we use REINFORCE with baseline computed using moving averages.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 593, + 345, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 347, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 347, + 608 + ], + "score": 1.0, + "content": "B VISUALIZATION OF DISCRETE ATTENTION", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "score": 1.0, + "content": "We visualize the attention of D-NTM with GRU controller with discrete attention in Figure 2. From", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "this example, we can see that D-NTM has learned to find the correct supporting fact even without", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 639, + 344, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 344, + 652 + ], + "score": 1.0, + "content": "any supervision for the particular story in the visualization.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 107, + 665, + 414, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 663, + 416, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 416, + 680 + ], + "score": 1.0, + "content": "C LEARNING CURVES FOR THE RECURRENT CONTROLLER", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "In Figure 3, we compare the learning curves of the continuous and discrete attention D-NTM model", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "with recurrent controller on Task 1. Surprisingly, the discrete attention D-NTM converges faster than", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "the continuous-attention model. The main difficulty of learning continuous-attention is due to the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 718, + 394, + 735 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 394, + 735 + ], + "score": 1.0, + "content": "fact that learning to write with continuous-attention can be challenging.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 258, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 259, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 259, + 95 + ], + "score": 1.0, + "content": "A EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 320, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 320, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 320, + 118 + ], + "score": 1.0, + "content": "A.1 MODEL AND TRAINING DETAILS FOR BABI", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 105, + 320, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 125, + 381, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 125, + 382, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 382, + 138 + ], + "score": 1.0, + "content": "We use the same hyperparameters for all the tasks for a given model.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 125, + 382, + 138 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 147, + 244, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 245, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 245, + 160 + ], + "score": 1.0, + "content": "A.1.1 FACT REPRESENTATION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 179 + ], + "score": 1.0, + "content": "We use a recurrent neural network with GRU units to encode a variable-length fact into a fixed-size", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "vector representation. This allows the D-NTM to exploit the word ordering in each fact, unlike when", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 188, + 279, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 279, + 200 + ], + "score": 1.0, + "content": "facts are encoded as bag-of-words vectors.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 165, + 505, + 200 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 210, + 201, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 203, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 203, + 224 + ], + "score": 1.0, + "content": "A.1.2 CONTROLLER", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 229, + 505, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "We experiment with both a recurrent and feedforward neural network as the controller that generates", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 237, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 253 + ], + "score": 1.0, + "content": "the read and write weights. The controller has 180 units. We train our feed-forward controller using", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "noisy-tanh activation function (Gulcehre et al., 2016) since we were experiencing training difficulties", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "with sigmoid and tanh activation functions. We use both single-step and three-steps addressing with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 271, + 188, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 188, + 282 + ], + "score": 1.0, + "content": "our GRU controller.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 229, + 506, + 282 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 293, + 182, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 184, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 184, + 307 + ], + "score": 1.0, + "content": "A.1.3 MEMORY", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 108, + 312, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "The memory contains 120 memory cells. Each memory cell consists of a 16-dimensional address", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 322, + 258, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 258, + 335 + ], + "score": 1.0, + "content": "part and 28-dimensional content part.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 312, + 505, + 335 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 224, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 226, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 226, + 358 + ], + "score": 1.0, + "content": "A.1.4 TRAINING DETAILS", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 505, + 417 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 200, + 376 + ], + "score": 1.0, + "content": "We set aside a random", + "type": "text" + }, + { + "bbox": [ + 201, + 364, + 221, + 375 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "of the training examples as a validation set for each sub-task and use", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 104, + 373, + 506, + 388 + ], + "score": 1.0, + "content": "it for early-stopping and hyperparameter search. We train one D-NTM for each sub-task, using", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Adam (Kingma & Ba, 2014) with its learning rate set to 0.003 and 0.007 respectively for GRU", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "and Feedforward controller. The size of each minibatch is 160, and each minibatch is constructed", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 406, + 269, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 269, + 419 + ], + "score": 1.0, + "content": "uniform-randomly from the training set.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 363, + 506, + 419 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 430, + 387, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 388, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 388, + 443 + ], + "score": 1.0, + "content": "A.2 MODEL AND TRAINING DETAILS FOR SEQUENTIAL MNIST", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "On sequential MNIST task we try to keep the capacity of our model to be close to our baselines. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "use 100 GRU units in the controller and each content vector of size 8 and with address vectors of size", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 220, + 483 + ], + "score": 1.0, + "content": "8. We use a learning rate of", + "type": "text" + }, + { + "bbox": [ + 220, + 471, + 249, + 481 + ], + "score": 0.82, + "content": "1 e - 3", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "and trained the model with adam optimizer. We did not use the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 481, + 359, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 359, + 493 + ], + "score": 1.0, + "content": "read and write consistency regularization in any of our models.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 450, + 505, + 493 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 505, + 345, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 346, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 346, + 518 + ], + "score": 1.0, + "content": "A.3 MODEL AND TRAINING DETAILS FOR TOY TASKS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 537 + ], + "score": 1.0, + "content": "On both copy and associative recall tasks, we try to keep the capacity of our model to be close to our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "baselines. We use 100 GRU units in the controller and each content vector of has a size of 8 and using", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 546, + 507, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 304, + 559 + ], + "score": 1.0, + "content": "address vector of size 8. We use a learning rate of", + "type": "text" + }, + { + "bbox": [ + 304, + 547, + 333, + 557 + ], + "score": 0.84, + "content": "1 e - 3", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 546, + 507, + 559 + ], + "score": 1.0, + "content": "and trained the model with adam optimizer.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "We did not use the read and write consistency regularization in any of our models. For the model with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 565, + 474, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 474, + 581 + ], + "score": 1.0, + "content": "the discrete attention we use REINFORCE with baseline computed using moving averages.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 525, + 507, + 581 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 593, + 345, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 347, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 347, + 608 + ], + "score": 1.0, + "content": "B VISUALIZATION OF DISCRETE ATTENTION", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "score": 1.0, + "content": "We visualize the attention of D-NTM with GRU controller with discrete attention in Figure 2. From", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "this example, we can see that D-NTM has learned to find the correct supporting fact even without", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 639, + 344, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 344, + 652 + ], + "score": 1.0, + "content": "any supervision for the particular story in the visualization.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 616, + 505, + 652 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 665, + 414, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 663, + 416, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 416, + 680 + ], + "score": 1.0, + "content": "C LEARNING CURVES FOR THE RECURRENT CONTROLLER", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "In Figure 3, we compare the learning curves of the continuous and discrete attention D-NTM model", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "with recurrent controller on Task 1. Surprisingly, the discrete attention D-NTM converges faster than", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "the continuous-attention model. The main difficulty of learning continuous-attention is due to the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 718, + 394, + 735 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 394, + 735 + ], + "score": 1.0, + "content": "fact that learning to write with continuous-attention can be challenging.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 689, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 78, + 426, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 138, + 78, + 426, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 78, + 426, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 78, + 426, + 213 + ], + "score": 0.965, + "type": "image", + "image_path": "13f6be6816a8ab4ad76006098fb07e79c29ffd5f83dd209f067809e9e7421ce0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 138, + 78, + 426, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 138, + 123.0, + 426, + 168.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 138, + 168.0, + 426, + 213.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 220, + 506, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "Figure 2: An example view of the discrete attention over the memory slots for both read (left) and write", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 232, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 159, + 243 + ], + "score": 1.0, + "content": "heads(right).", + "type": "text" + }, + { + "bbox": [ + 159, + 233, + 165, + 241 + ], + "score": 0.25, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 232, + 505, + 243 + ], + "score": 1.0, + "content": "-axis the denotes the memory locations that are being accessed and y-axis corresponds", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "to the content in the particular memory location. In this figure, we visualize the discrete-attention model", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "with 3-reading steps and on task-20. It is easy to see that the NTM with discrete-attention accesses", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 262, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 506, + 275 + ], + "score": 1.0, + "content": "to the relevant part of the memory. We only visualize the last-step of the 3-steps writing. Because", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 272, + 437, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 437, + 286 + ], + "score": 1.0, + "content": "with discrete attention usually the model just reads the empty slots of the memory.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 172, + 317, + 434, + 506 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 172, + 317, + 434, + 506 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 172, + 317, + 434, + 506 + ], + "spans": [ + { + "bbox": [ + 172, + 317, + 434, + 506 + ], + "score": 0.969, + "type": "image", + "image_path": "750ba06aa86f2fcf7b1b8477065761e178ba6d5f4dd1a6aa03704d44c0632dbd.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 172, + 317, + 434, + 380.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 172, + 380.0, + 434, + 443.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 172, + 443.0, + 434, + 506.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 523, + 505, + 555 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "Figure 3: A visualization for the learning curves of continuous and discrete D-NTM models trained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 534, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 545 + ], + "score": 1.0, + "content": "on Task 1 using 3 steps. In most tasks, we observe that the discrete attention model with GRU controller", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 544, + 339, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 339, + 556 + ], + "score": 1.0, + "content": "does converge faster than the continuous-attention model.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 110, + 576, + 503, + 602 + ], + "lines": [ + { + "bbox": [ + 107, + 574, + 346, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 346, + 590 + ], + "score": 1.0, + "content": "D A COMPARISON BETWEEN THE LEARNING", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 587, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 127, + 587, + 457, + 604 + ], + "score": 1.0, + "content": "CURVES OF INPUT BASED BASELINE AND REGULAR BASELINE ON", + "type": "text" + }, + { + "bbox": [ + 458, + 591, + 465, + 603 + ], + "score": 0.28, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 587, + 505, + 604 + ], + "score": 1.0, + "content": "MNIST", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 506, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "score": 1.0, + "content": "In Figure 4, we show the learning curves of input-based-baseline (ibb) and regular REINFORCE with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 259, + 636 + ], + "score": 1.0, + "content": "moving averages baseline (mab) on the", + "type": "text" + }, + { + "bbox": [ + 259, + 625, + 296, + 636 + ], + "score": 0.47, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "task. We observe that input-based-baseline in general", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 635, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 104, + 635, + 506, + 647 + ], + "score": 1.0, + "content": "is much easier to optimize and converges faster as well. But it can quickly overfit to the task as well.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 662, + 503, + 688 + ], + "lines": [ + { + "bbox": [ + 104, + 660, + 182, + 676 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 182, + 676 + ], + "score": 1.0, + "content": "E TRAINING", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 675, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 125, + 675, + 505, + 689 + ], + "score": 1.0, + "content": "WITH CONTINUOUS-ATTENTION AND TESTING WITH DISCRETE-ATTENTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "In Table 5, we provide results investigating the effects of using discrete attention model at the test-time", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "for a model trained with feed-forward controller and continuous attention. Discrete∗ D-NTM model", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "bootstraps the discrete attention with the continuous attention, using the curriculum method that we have", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 78, + 426, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 138, + 78, + 426, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 78, + 426, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 78, + 426, + 213 + ], + "score": 0.965, + "type": "image", + "image_path": "13f6be6816a8ab4ad76006098fb07e79c29ffd5f83dd209f067809e9e7421ce0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 138, + 78, + 426, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 138, + 123.0, + 426, + 168.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 138, + 168.0, + 426, + 213.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 220, + 506, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "Figure 2: An example view of the discrete attention over the memory slots for both read (left) and write", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 232, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 159, + 243 + ], + "score": 1.0, + "content": "heads(right).", + "type": "text" + }, + { + "bbox": [ + 159, + 233, + 165, + 241 + ], + "score": 0.25, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 232, + 505, + 243 + ], + "score": 1.0, + "content": "-axis the denotes the memory locations that are being accessed and y-axis corresponds", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "to the content in the particular memory location. In this figure, we visualize the discrete-attention model", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "with 3-reading steps and on task-20. It is easy to see that the NTM with discrete-attention accesses", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 262, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 506, + 275 + ], + "score": 1.0, + "content": "to the relevant part of the memory. We only visualize the last-step of the 3-steps writing. Because", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 272, + 437, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 437, + 286 + ], + "score": 1.0, + "content": "with discrete attention usually the model just reads the empty slots of the memory.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 172, + 317, + 434, + 506 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 172, + 317, + 434, + 506 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 172, + 317, + 434, + 506 + ], + "spans": [ + { + "bbox": [ + 172, + 317, + 434, + 506 + ], + "score": 0.969, + "type": "image", + "image_path": "750ba06aa86f2fcf7b1b8477065761e178ba6d5f4dd1a6aa03704d44c0632dbd.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 172, + 317, + 434, + 380.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 172, + 380.0, + 434, + 443.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 172, + 443.0, + 434, + 506.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 523, + 505, + 555 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "Figure 3: A visualization for the learning curves of continuous and discrete D-NTM models trained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 534, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 545 + ], + "score": 1.0, + "content": "on Task 1 using 3 steps. In most tasks, we observe that the discrete attention model with GRU controller", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 544, + 339, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 339, + 556 + ], + "score": 1.0, + "content": "does converge faster than the continuous-attention model.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 110, + 576, + 503, + 602 + ], + "lines": [ + { + "bbox": [ + 107, + 574, + 346, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 346, + 590 + ], + "score": 1.0, + "content": "D A COMPARISON BETWEEN THE LEARNING", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 587, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 127, + 587, + 457, + 604 + ], + "score": 1.0, + "content": "CURVES OF INPUT BASED BASELINE AND REGULAR BASELINE ON", + "type": "text" + }, + { + "bbox": [ + 458, + 591, + 465, + 603 + ], + "score": 0.28, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 587, + 505, + 604 + ], + "score": 1.0, + "content": "MNIST", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 506, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "score": 1.0, + "content": "In Figure 4, we show the learning curves of input-based-baseline (ibb) and regular REINFORCE with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 259, + 636 + ], + "score": 1.0, + "content": "moving averages baseline (mab) on the", + "type": "text" + }, + { + "bbox": [ + 259, + 625, + 296, + 636 + ], + "score": 0.47, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "task. We observe that input-based-baseline in general", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 635, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 104, + 635, + 506, + 647 + ], + "score": 1.0, + "content": "is much easier to optimize and converges faster as well. But it can quickly overfit to the task as well.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 614, + 506, + 647 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 662, + 503, + 688 + ], + "lines": [ + { + "bbox": [ + 104, + 660, + 182, + 676 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 182, + 676 + ], + "score": 1.0, + "content": "E TRAINING", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 675, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 125, + 675, + 505, + 689 + ], + "score": 1.0, + "content": "WITH CONTINUOUS-ATTENTION AND TESTING WITH DISCRETE-ATTENTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "In Table 5, we provide results investigating the effects of using discrete attention model at the test-time", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "for a model trained with feed-forward controller and continuous attention. Discrete∗ D-NTM model", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "bootstraps the discrete attention with the continuous attention, using the curriculum method that we have", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "introduced in Section ”Curriculum Learning for the Discrete Attention”. Discrete† D-NTM model is the", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "continuous-attention model which uses discrete-attention at the test time. We observe that the Discrete†", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 401, + 491, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 491, + 412 + ], + "score": 1.0, + "content": "D-NTM model which is trained with continuous-attention outperforms Discrete D-NTM model.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 700, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 172, + 98, + 435, + 290 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 172, + 98, + 435, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 172, + 98, + 435, + 290 + ], + "spans": [ + { + "bbox": [ + 172, + 98, + 435, + 290 + ], + "score": 0.969, + "type": "image", + "image_path": "858347b800db4692b6c99c985fab4630eac6e5ffb7ee526a7c12130eaa865972.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 172, + 98, + 435, + 162.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 172, + 162.0, + 435, + 226.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 172, + 226.0, + 435, + 290.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 307, + 505, + 340 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 467, + 320 + ], + "score": 1.0, + "content": "Figure 4: We compare the learning curves of our D-NTM model using discrete attention on", + "type": "text" + }, + { + "bbox": [ + 468, + 308, + 505, + 319 + ], + "score": 0.64, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 392, + 330 + ], + "score": 1.0, + "content": "task with input-based baseline and regular REINFORCE baseline. The", + "type": "text" + }, + { + "bbox": [ + 392, + 320, + 398, + 328 + ], + "score": 0.41, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "-axis is the loss and y-axis", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 329, + 207, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 207, + 341 + ], + "score": 1.0, + "content": "is the number of epochs.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "introduced in Section ”Curriculum Learning for the Discrete Attention”. Discrete† D-NTM model is the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "continuous-attention model which uses discrete-attention at the test time. We observe that the Discrete†", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 401, + 491, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 491, + 412 + ], + "score": 1.0, + "content": "D-NTM model which is trained with continuous-attention outperforms Discrete D-NTM model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 196, + 432, + 415, + 655 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 196, + 432, + 415, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 432, + 415, + 655 + ], + "spans": [ + { + "bbox": [ + 196, + 432, + 415, + 655 + ], + "score": 0.957, + "html": "
TaskcontinuousD-NTMDiscreteD-NTMDiscrete*D-NTMDiscrete†D-NTM
14.3881.6714.7972.2881.6778.9579.6968.5431.6749.17
23427.571.2576.6779.3878.6576.6770.8344.06
0.00
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.6279.3237.7125.6382.0874.3847.0877.0873.9653.0230.4211.4676.0513.96
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.9271.569.79
192039.4771.4676.56
0.0
Avg12.8168.3037.7957.21
", + "type": "table", + "image_path": "d8842f21cc4b747e67123ff145e3356ea1f79cdfc3c8d7e3c47994ff0c97625c.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 432, + 415, + 444.3888888888889 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 196, + 444.3888888888889, + 415, + 456.7777777777778 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 196, + 456.7777777777778, + 415, + 469.16666666666674 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 196, + 469.16666666666674, + 415, + 481.55555555555566 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 196, + 481.55555555555566, + 415, + 493.94444444444457 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 196, + 493.94444444444457, + 415, + 506.3333333333335 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 196, + 506.3333333333335, + 415, + 518.7222222222224 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 196, + 518.7222222222224, + 415, + 531.1111111111113 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 196, + 531.1111111111113, + 415, + 543.5000000000002 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 196, + 543.5000000000002, + 415, + 555.8888888888891 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 196, + 555.8888888888891, + 415, + 568.277777777778 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 196, + 568.277777777778, + 415, + 580.666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 196, + 580.666666666667, + 415, + 593.0555555555559 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 196, + 593.0555555555559, + 415, + 605.4444444444448 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 196, + 605.4444444444448, + 415, + 617.8333333333337 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 196, + 617.8333333333337, + 415, + 630.2222222222226 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 196, + 630.2222222222226, + 415, + 642.6111111111115 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 196, + 642.6111111111115, + 415, + 655.0000000000005 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 662, + 505, + 707 + ], + "lines": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 200, + 674 + ], + "score": 1.0, + "content": "Table 5: Test error rates", + "type": "text" + }, + { + "bbox": [ + 201, + 663, + 216, + 673 + ], + "score": 0.8, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "feedforward controller. Discrete∗ D-NTM model bootstraps the discrete attention with the continuous", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "score": 1.0, + "content": "attention, using the curriculum method that we have introduced in Section 4. Discrete† D-NTM model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 696, + 423, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 423, + 707 + ], + "score": 1.0, + "content": "is the continuous-attention model which uses discrete-attention at the test time.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 172, + 98, + 435, + 290 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 172, + 98, + 435, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 172, + 98, + 435, + 290 + ], + "spans": [ + { + "bbox": [ + 172, + 98, + 435, + 290 + ], + "score": 0.969, + "type": "image", + "image_path": "858347b800db4692b6c99c985fab4630eac6e5ffb7ee526a7c12130eaa865972.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 172, + 98, + 435, + 162.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 172, + 162.0, + 435, + 226.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 172, + 226.0, + 435, + 290.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 307, + 505, + 340 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 467, + 320 + ], + "score": 1.0, + "content": "Figure 4: We compare the learning curves of our D-NTM model using discrete attention on", + "type": "text" + }, + { + "bbox": [ + 468, + 308, + 505, + 319 + ], + "score": 0.64, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 392, + 330 + ], + "score": 1.0, + "content": "task with input-based baseline and regular REINFORCE baseline. The", + "type": "text" + }, + { + "bbox": [ + 392, + 320, + 398, + 328 + ], + "score": 0.41, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "-axis is the loss and y-axis", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 329, + 207, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 207, + 341 + ], + "score": 1.0, + "content": "is the number of epochs.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 505, + 412 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 105, + 377, + 505, + 412 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 196, + 432, + 415, + 655 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 196, + 432, + 415, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 432, + 415, + 655 + ], + "spans": [ + { + "bbox": [ + 196, + 432, + 415, + 655 + ], + "score": 0.957, + "html": "
TaskcontinuousD-NTMDiscreteD-NTMDiscrete*D-NTMDiscrete†D-NTM
14.3881.6714.7972.2881.6778.9579.6968.5431.6749.17
23427.571.2576.6779.3878.6576.6770.8344.06
0.00
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.6279.3237.7125.6382.0874.3847.0877.0873.9653.0230.4211.4676.0513.96
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.9271.569.79
192039.4771.4676.56
0.0
Avg12.8168.3037.7957.21
", + "type": "table", + "image_path": "d8842f21cc4b747e67123ff145e3356ea1f79cdfc3c8d7e3c47994ff0c97625c.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 432, + 415, + 444.3888888888889 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 196, + 444.3888888888889, + 415, + 456.7777777777778 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 196, + 456.7777777777778, + 415, + 469.16666666666674 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 196, + 469.16666666666674, + 415, + 481.55555555555566 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 196, + 481.55555555555566, + 415, + 493.94444444444457 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 196, + 493.94444444444457, + 415, + 506.3333333333335 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 196, + 506.3333333333335, + 415, + 518.7222222222224 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 196, + 518.7222222222224, + 415, + 531.1111111111113 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 196, + 531.1111111111113, + 415, + 543.5000000000002 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 196, + 543.5000000000002, + 415, + 555.8888888888891 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 196, + 555.8888888888891, + 415, + 568.277777777778 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 196, + 568.277777777778, + 415, + 580.666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 196, + 580.666666666667, + 415, + 593.0555555555559 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 196, + 593.0555555555559, + 415, + 605.4444444444448 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 196, + 605.4444444444448, + 415, + 617.8333333333337 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 196, + 617.8333333333337, + 415, + 630.2222222222226 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 196, + 630.2222222222226, + 415, + 642.6111111111115 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 196, + 642.6111111111115, + 415, + 655.0000000000005 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 662, + 505, + 707 + ], + "lines": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 200, + 674 + ], + "score": 1.0, + "content": "Table 5: Test error rates", + "type": "text" + }, + { + "bbox": [ + 201, + 663, + 216, + 673 + ], + "score": 0.8, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "feedforward controller. Discrete∗ D-NTM model bootstraps the discrete attention with the continuous", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 696 + ], + "score": 1.0, + "content": "attention, using the curriculum method that we have introduced in Section 4. Discrete† D-NTM model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 696, + 423, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 423, + 707 + ], + "score": 1.0, + "content": "is the continuous-attention model which uses discrete-attention at the test time.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 662, + 505, + 707 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 354, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 356, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 356, + 96 + ], + "score": 1.0, + "content": "F D-NTM WITH BOW FACT REPRESENTATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 507, + 118 + ], + "score": 1.0, + "content": "In Table 6, we provide results for D-NTM using BoW with positional encoding (PE) Sukhbaatar et al.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "(2015) as the representation of the input facts. The facts representations are provided as an input to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "the GRU controller. In agreement to our results with the GRU fact representation, with the BoW fact", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "representation we observe improvements with multi-step of addressing over single-step and discrete", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 265, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 265, + 161 + ], + "score": 1.0, + "content": "addressing over continuous addressing.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 147, + 163, + 464, + 384 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 147, + 163, + 464, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 163, + 464, + 384 + ], + "spans": [ + { + "bbox": [ + 147, + 163, + 464, + 384 + ], + "score": 0.98, + "html": "
TaskSoftD-NTM(1-step)DiscreteD-NTM(1-step)SoftD-NTM(3-steps)DiscreteD-NTM(3-steps)
10.000.000.000.00
23461.0459.3756.8755.62
55.6257.562.557.5
27.2924.8926.4527.08
513.5512.0815.8314.78
613.5414.3721.8713.33
78.546.258.7514.58
81.691.363.013.02
917.716.6637.7017.08
1026.0427.0826.8723.952.294.16
1120.413.952.5
120.410.830.20
133.121.044.795.83
1462.0858.3361.2560.620.05
1531.6626.250.62
1654.4748.5448.9548.9530.6236.04
1743.7531.8743.75
1833.7539.3736.66
192064.6369.2167.2365.46
1.250.001.450.00
Avg27.0224.9826.3624.05
", + "type": "table", + "image_path": "0a6c5ddf5f16fcf35d62d0c38760c7d4999bd44c3759192a955461176c3e8610.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 147, + 163, + 464, + 236.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 147, + 236.66666666666669, + 464, + 310.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 147, + 310.33333333333337, + 464, + 384.00000000000006 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 392, + 504, + 414 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 204, + 405 + ], + "score": 1.0, + "content": "Table 6: Test error rates", + "type": "text" + }, + { + "bbox": [ + 204, + 393, + 220, + 403 + ], + "score": 0.8, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 401, + 500, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 368, + 416 + ], + "score": 1.0, + "content": "the GRU controller and representations of facts are obtained with", + "type": "text" + }, + { + "bbox": [ + 369, + 403, + 392, + 413 + ], + "score": 0.46, + "content": "\\mathrm { B o W }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 401, + 500, + 416 + ], + "score": 1.0, + "content": "using positional encoding.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + } + ], + "index": 8.25 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 354, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 356, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 356, + 96 + ], + "score": 1.0, + "content": "F D-NTM WITH BOW FACT REPRESENTATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 507, + 118 + ], + "score": 1.0, + "content": "In Table 6, we provide results for D-NTM using BoW with positional encoding (PE) Sukhbaatar et al.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "(2015) as the representation of the input facts. The facts representations are provided as an input to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "the GRU controller. In agreement to our results with the GRU fact representation, with the BoW fact", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "representation we observe improvements with multi-step of addressing over single-step and discrete", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 265, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 265, + 161 + ], + "score": 1.0, + "content": "addressing over continuous addressing.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 105, + 507, + 161 + ] + }, + { + "type": "table", + "bbox": [ + 147, + 163, + 464, + 384 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 147, + 163, + 464, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 163, + 464, + 384 + ], + "spans": [ + { + "bbox": [ + 147, + 163, + 464, + 384 + ], + "score": 0.98, + "html": "
TaskSoftD-NTM(1-step)DiscreteD-NTM(1-step)SoftD-NTM(3-steps)DiscreteD-NTM(3-steps)
10.000.000.000.00
23461.0459.3756.8755.62
55.6257.562.557.5
27.2924.8926.4527.08
513.5512.0815.8314.78
613.5414.3721.8713.33
78.546.258.7514.58
81.691.363.013.02
917.716.6637.7017.08
1026.0427.0826.8723.952.294.16
1120.413.952.5
120.410.830.20
133.121.044.795.83
1462.0858.3361.2560.620.05
1531.6626.250.62
1654.4748.5448.9548.9530.6236.04
1743.7531.8743.75
1833.7539.3736.66
192064.6369.2167.2365.46
1.250.001.450.00
Avg27.0224.9826.3624.05
", + "type": "table", + "image_path": "0a6c5ddf5f16fcf35d62d0c38760c7d4999bd44c3759192a955461176c3e8610.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 147, + 163, + 464, + 236.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 147, + 236.66666666666669, + 464, + 310.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 147, + 310.33333333333337, + 464, + 384.00000000000006 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 392, + 504, + 414 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 204, + 405 + ], + "score": 1.0, + "content": "Table 6: Test error rates", + "type": "text" + }, + { + "bbox": [ + 204, + 393, + 220, + 403 + ], + "score": 0.8, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "on the 20 bAbI QA tasks for models using 10k training examples with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 401, + 500, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 368, + 416 + ], + "score": 1.0, + "content": "the GRU controller and representations of facts are obtained with", + "type": "text" + }, + { + "bbox": [ + 369, + 403, + 392, + 413 + ], + "score": 0.46, + "content": "\\mathrm { B o W }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 401, + 500, + 416 + ], + "score": 1.0, + "content": "using positional encoding.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + } + ], + "index": 8.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/BkSmc8qll/BkSmc8qll_model.json b/parse/train/BkSmc8qll/BkSmc8qll_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0f7b35a182cbf40b7db38ce53efcec480f638ab --- /dev/null +++ b/parse/train/BkSmc8qll/BkSmc8qll_model.json @@ -0,0 +1,17617 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 615, + 1302, + 615, + 1302, + 936, + 398, + 936 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1068, + 1405, + 1068, + 1405, + 1273, + 298, + 1273 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1478, + 1404, + 1478, + 1404, + 1653, + 298, + 1653 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1288, + 1403, + 1288, + 1403, + 1463, + 299, + 1463 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1858, + 1404, + 1858, + 1404, + 2033, + 299, + 2033 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1668, + 1403, + 1668, + 1403, + 1843, + 299, + 1843 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 300, + 218, + 1405, + 218, + 1405, + 321, + 300, + 321 + ], + "score": 0.962 + }, + { + "category_id": 0, + "poly": [ + 302, + 998, + 573, + 998, + 573, + 1033, + 302, + 1033 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 773, + 546, + 927, + 546, + 927, + 579, + 773, + 579 + ], + "score": 0.872 + }, + { + "category_id": 1, + "poly": [ + 323, + 372, + 1167, + 372, + 1167, + 406, + 323, + 406 + ], + "score": 0.813 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 319, + 410, + 993, + 410, + 993, + 465, + 319, + 465 + ], + "score": 0.693 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 219.0, + 1411.0, + 219.0, + 1411.0, + 271.0, + 296.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 272.0, + 1235.0, + 272.0, + 1235.0, + 324.0, + 294.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 995.0, + 579.0, + 995.0, + 579.0, + 1042.0, + 294.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 542.0, + 934.0, + 542.0, + 934.0, + 585.0, + 768.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 613.0, + 1306.0, + 613.0, + 1306.0, + 653.0, + 392.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 643.0, + 1306.0, + 643.0, + 1306.0, + 680.0, + 392.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 673.0, + 1306.0, + 673.0, + 1306.0, + 708.0, + 394.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 703.0, + 1305.0, + 703.0, + 1305.0, + 734.0, + 394.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 733.0, + 1306.0, + 733.0, + 1306.0, + 764.0, + 394.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 759.0, + 1306.0, + 759.0, + 1306.0, + 793.0, + 393.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 789.0, + 1305.0, + 789.0, + 1305.0, + 823.0, + 393.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 818.0, + 1306.0, + 818.0, + 1306.0, + 851.0, + 392.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 847.0, + 1305.0, + 847.0, + 1305.0, + 879.0, + 394.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 876.0, + 1305.0, + 876.0, + 1305.0, + 909.0, + 393.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 905.0, + 1308.0, + 905.0, + 1308.0, + 940.0, + 393.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1067.0, + 1408.0, + 1067.0, + 1408.0, + 1104.0, + 293.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 1407.0, + 1097.0, + 1407.0, + 1134.0, + 293.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1126.0, + 1407.0, + 1126.0, + 1407.0, + 1160.0, + 295.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1156.0, + 1407.0, + 1156.0, + 1407.0, + 1190.0, + 295.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1185.0, + 1406.0, + 1185.0, + 1406.0, + 1218.0, + 295.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1211.0, + 1408.0, + 1211.0, + 1408.0, + 1249.0, + 291.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1400.0, + 1240.0, + 1400.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1403.0, + 1477.0, + 1403.0, + 1510.0, + 293.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1507.0, + 1406.0, + 1507.0, + 1406.0, + 1540.0, + 292.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1568.0, + 296.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1562.0, + 1408.0, + 1562.0, + 1408.0, + 1599.0, + 293.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1594.0, + 1408.0, + 1594.0, + 1408.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1622.0, + 1409.0, + 1622.0, + 1409.0, + 1659.0, + 292.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1287.0, + 1408.0, + 1287.0, + 1408.0, + 1321.0, + 295.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1315.0, + 1405.0, + 1315.0, + 1405.0, + 1350.0, + 294.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1345.0, + 1405.0, + 1345.0, + 1405.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1409.0, + 293.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1404.0, + 1407.0, + 1404.0, + 1407.0, + 1438.0, + 293.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 699.0, + 1429.0, + 699.0, + 1470.0, + 293.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1857.0, + 1406.0, + 1857.0, + 1406.0, + 1891.0, + 295.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1888.0, + 1408.0, + 1888.0, + 1408.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1912.0, + 1409.0, + 1912.0, + 1409.0, + 1953.0, + 292.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1408.0, + 2002.0, + 1408.0, + 2035.0, + 293.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1665.0, + 1405.0, + 1665.0, + 1405.0, + 1705.0, + 293.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1732.0, + 294.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1728.0, + 1407.0, + 1728.0, + 1407.0, + 1759.0, + 295.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1753.0, + 1405.0, + 1753.0, + 1405.0, + 1790.0, + 293.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1784.0, + 1408.0, + 1784.0, + 1408.0, + 1818.0, + 294.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1813.0, + 780.0, + 1813.0, + 780.0, + 1848.0, + 295.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 368.0, + 1173.0, + 368.0, + 1173.0, + 415.0, + 317.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 401.0, + 998.0, + 401.0, + 998.0, + 439.0, + 312.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 432.0, + 901.0, + 432.0, + 901.0, + 471.0, + 313.0, + 471.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 743, + 1405, + 743, + 1405, + 919, + 298, + 919 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1640, + 1406, + 1640, + 1406, + 1848, + 298, + 1848 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1010, + 1402, + 1010, + 1402, + 1129, + 298, + 1129 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 300, + 1938, + 1403, + 1938, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1365, + 1402, + 1365, + 1402, + 1430, + 300, + 1430 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 290, + 1497, + 1405, + 1497, + 1405, + 1559, + 290, + 1559 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 297, + 369, + 1406, + 369, + 1406, + 633, + 297, + 633 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 291, + 229, + 1405, + 229, + 1405, + 291, + 291, + 291 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 769, + 1444, + 926, + 1444, + 926, + 1482, + 769, + 1482 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 766, + 1574, + 929, + 1574, + 929, + 1613, + 766, + 1613 + ], + "score": 0.934 + }, + { + "category_id": 8, + "poly": [ + 708, + 1145, + 992, + 1145, + 992, + 1184, + 708, + 1184 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 1198, + 643, + 1198, + 643, + 1229, + 298, + 1229 + ], + "score": 0.929 + }, + { + "category_id": 8, + "poly": [ + 757, + 1242, + 938, + 1242, + 938, + 1281, + 757, + 1281 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 299, + 674, + 896, + 674, + 896, + 710, + 299, + 710 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 298, + 322, + 523, + 322, + 523, + 354, + 298, + 354 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 1314, + 478, + 1314, + 478, + 1344, + 299, + 1344 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 300, + 1883, + 644, + 1883, + 644, + 1914, + 300, + 1914 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1248, + 1400, + 1248, + 1400, + 1278, + 1366, + 1278 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1151, + 1400, + 1151, + 1400, + 1180, + 1366, + 1180 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 299, + 955, + 529, + 955, + 529, + 985, + 299, + 985 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.695 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 105, + 298, + 105 + ], + "score": 0.673 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 105, + 298, + 105 + ], + "score": 0.255 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.206 + }, + { + "category_id": 14, + "poly": [ + 705, + 1144, + 992, + 1144, + 992, + 1184, + 705, + 1184 + ], + "score": 0.93, + "latex": "\\mathbf { h } ^ { t } = \\mathbf { G } \\mathbf { R } \\mathbf { U } ( \\mathbf { x } ^ { t } , \\mathbf { h } ^ { t - 1 } , \\boldsymbol { \\phi } ^ { t } )" + }, + { + "category_id": 14, + "poly": [ + 757, + 1241, + 942, + 1241, + 942, + 1282, + 757, + 1282 + ], + "score": 0.92, + "latex": "\\mathbf { h } ^ { t } = \\sigma ( \\mathbf { x } ^ { t } , \\phi ^ { t } ) ." + }, + { + "category_id": 13, + "poly": [ + 1077, + 1497, + 1216, + 1497, + 1216, + 1528, + 1077, + 1528 + ], + "score": 0.92, + "latex": "\\mathbf { C } \\in \\mathbb { R } ^ { N \\times d _ { c } }" + }, + { + "category_id": 13, + "poly": [ + 443, + 1497, + 584, + 1497, + 584, + 1528, + 443, + 1528 + ], + "score": 0.92, + "latex": "\\mathbf { A } \\in \\mathbb { R } ^ { N \\times d _ { a } }" + }, + { + "category_id": 14, + "poly": [ + 768, + 1443, + 931, + 1443, + 931, + 1482, + 768, + 1482 + ], + "score": 0.92, + "latex": "\\mathbf { M } = \\left[ \\mathbf { A } ; \\mathbf { C } \\right] ." + }, + { + "category_id": 13, + "poly": [ + 608, + 1366, + 813, + 1366, + 813, + 1398, + 608, + 1398 + ], + "score": 0.92, + "latex": "\\mathbf { M } \\in \\mathbb { R } ^ { N \\times ( d _ { c } + d _ { a } ) }" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1968, + 1216, + 1968, + 1216, + 2000, + 1112, + 2000 + ], + "score": 0.92, + "latex": "\\mathbf { e } ^ { t } \\in \\mathbb { R } ^ { d _ { c } }" + }, + { + "category_id": 14, + "poly": [ + 765, + 1573, + 933, + 1573, + 933, + 1611, + 765, + 1611 + ], + "score": 0.91, + "latex": "\\mathbf { m } _ { i } = \\left[ \\mathbf { a } _ { i } ; \\mathbf { c } _ { i } \\right] ." + }, + { + "category_id": 13, + "poly": [ + 1308, + 1728, + 1399, + 1728, + 1399, + 1757, + 1308, + 1757 + ], + "score": 0.9, + "latex": "\\mathbf { C } ^ { 0 } = \\mathbf { 0 }" + }, + { + "category_id": 13, + "poly": [ + 912, + 1969, + 1021, + 1969, + 1021, + 1999, + 912, + 1999 + ], + "score": 0.9, + "latex": "\\mathbf { w } ^ { t } \\in \\mathbb { R } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1344, + 1969, + 1405, + 1969, + 1405, + 2000, + 1344, + 2000 + ], + "score": 0.88, + "latex": "\\mathbf { u } ^ { t } \\in" + }, + { + "category_id": 13, + "poly": [ + 676, + 1532, + 714, + 1532, + 714, + 1559, + 676, + 1559 + ], + "score": 0.88, + "latex": "\\mathbf { m } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 589, + 1070, + 619, + 1070, + 619, + 1100, + 589, + 1100 + ], + "score": 0.87, + "latex": "\\mathbf { y } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 2001, + 340, + 2001, + 340, + 2030, + 298, + 2030 + ], + "score": 0.87, + "latex": "\\mathbb { R } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 481, + 1647, + 509, + 1647, + 509, + 1672, + 481, + 1672 + ], + "score": 0.87, + "latex": "{ \\bf a } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1348, + 2002, + 1404, + 2002, + 1404, + 2033, + 1348, + 2033 + ], + "score": 0.86, + "latex": "\\bar { \\mathbf { c } } ^ { t } \\in" + }, + { + "category_id": 13, + "poly": [ + 297, + 1070, + 327, + 1070, + 327, + 1100, + 297, + 1100 + ], + "score": 0.86, + "latex": "\\mathbf { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 437, + 1704, + 464, + 1704, + 464, + 1730, + 437, + 1730 + ], + "score": 0.86, + "latex": "\\mathbf { c } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 937, + 1011, + 966, + 1011, + 966, + 1039, + 937, + 1039 + ], + "score": 0.85, + "latex": "\\mathbf { x } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 573, + 1039, + 601, + 1039, + 601, + 1071, + 573, + 1071 + ], + "score": 0.85, + "latex": "\\phi ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 926, + 1371, + 954, + 1371, + 954, + 1397, + 926, + 1397 + ], + "score": 0.84, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 993, + 1941, + 1020, + 1941, + 1020, + 1965, + 993, + 1965 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 485, + 1015, + 498, + 1015, + 498, + 1039, + 485, + 1039 + ], + "score": 0.61, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 670.0, + 899.0, + 670.0, + 899.0, + 715.0, + 291.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 320.0, + 526.0, + 320.0, + 526.0, + 360.0, + 296.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 483.0, + 1309.0, + 483.0, + 1350.0, + 293.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 647.0, + 1880.0, + 647.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 535.0, + 951.0, + 535.0, + 991.0, + 293.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 744.0, + 1407.0, + 744.0, + 1407.0, + 778.0, + 295.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 772.0, + 1406.0, + 772.0, + 1406.0, + 806.0, + 295.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 800.0, + 1407.0, + 800.0, + 1407.0, + 835.0, + 295.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 830.0, + 1405.0, + 830.0, + 1405.0, + 864.0, + 295.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 856.0, + 1406.0, + 856.0, + 1406.0, + 896.0, + 292.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 889.0, + 922.0, + 889.0, + 922.0, + 920.0, + 296.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1637.0, + 480.0, + 1637.0, + 480.0, + 1678.0, + 292.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1637.0, + 1409.0, + 1637.0, + 1409.0, + 1678.0, + 510.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1671.0, + 1406.0, + 1671.0, + 1406.0, + 1704.0, + 295.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1699.0, + 436.0, + 1699.0, + 436.0, + 1732.0, + 292.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1699.0, + 1406.0, + 1699.0, + 1406.0, + 1732.0, + 465.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1728.0, + 1307.0, + 1728.0, + 1307.0, + 1761.0, + 292.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 1728.0, + 1408.0, + 1728.0, + 1408.0, + 1761.0, + 1400.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1757.0, + 1406.0, + 1757.0, + 1406.0, + 1791.0, + 293.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1786.0, + 1404.0, + 1786.0, + 1404.0, + 1820.0, + 295.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1814.0, + 1086.0, + 1814.0, + 1086.0, + 1851.0, + 293.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1011.0, + 484.0, + 1011.0, + 484.0, + 1045.0, + 296.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1011.0, + 936.0, + 1011.0, + 936.0, + 1045.0, + 499.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1011.0, + 1404.0, + 1011.0, + 1404.0, + 1045.0, + 967.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1038.0, + 572.0, + 1038.0, + 572.0, + 1076.0, + 293.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1038.0, + 1406.0, + 1038.0, + 1406.0, + 1076.0, + 602.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1068.0, + 588.0, + 1068.0, + 588.0, + 1102.0, + 328.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1068.0, + 1404.0, + 1068.0, + 1404.0, + 1102.0, + 620.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 1406.0, + 1097.0, + 1406.0, + 1131.0, + 293.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1935.0, + 992.0, + 1935.0, + 992.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1935.0, + 1409.0, + 1935.0, + 1409.0, + 1976.0, + 1021.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1965.0, + 911.0, + 1965.0, + 911.0, + 2007.0, + 291.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1965.0, + 1111.0, + 1965.0, + 1111.0, + 2007.0, + 1022.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1965.0, + 1343.0, + 1965.0, + 1343.0, + 2007.0, + 1217.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1998.0, + 297.0, + 1998.0, + 297.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1998.0, + 1347.0, + 1998.0, + 1347.0, + 2039.0, + 341.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1360.0, + 607.0, + 1360.0, + 607.0, + 1407.0, + 293.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1360.0, + 925.0, + 1360.0, + 925.0, + 1407.0, + 814.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1360.0, + 1408.0, + 1360.0, + 1408.0, + 1407.0, + 955.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1397.0, + 881.0, + 1397.0, + 881.0, + 1434.0, + 295.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1489.0, + 442.0, + 1489.0, + 442.0, + 1538.0, + 289.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1489.0, + 1076.0, + 1489.0, + 1076.0, + 1538.0, + 585.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1489.0, + 1412.0, + 1489.0, + 1412.0, + 1538.0, + 1217.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1527.0, + 675.0, + 1527.0, + 675.0, + 1563.0, + 294.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1527.0, + 797.0, + 1527.0, + 797.0, + 1563.0, + 715.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 1404.0, + 367.0, + 1404.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 393.0, + 1097.0, + 393.0, + 1097.0, + 438.0, + 326.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 431.0, + 326.0, + 431.0, + 326.0, + 457.0, + 297.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 428.0, + 1407.0, + 428.0, + 1407.0, + 463.0, + 329.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 456.0, + 1404.0, + 456.0, + 1404.0, + 491.0, + 329.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 485.0, + 436.0, + 485.0, + 436.0, + 517.0, + 329.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 512.0, + 1402.0, + 512.0, + 1402.0, + 548.0, + 297.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 544.0, + 1043.0, + 544.0, + 1043.0, + 576.0, + 329.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 576.0, + 325.0, + 576.0, + 325.0, + 600.0, + 297.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 572.0, + 1404.0, + 572.0, + 1404.0, + 607.0, + 329.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 602.0, + 864.0, + 602.0, + 864.0, + 637.0, + 329.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1407.0, + 230.0, + 1407.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 257.0, + 1302.0, + 257.0, + 1302.0, + 295.0, + 293.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1197.0, + 644.0, + 1197.0, + 644.0, + 1231.0, + 296.0, + 1231.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1555, + 1405, + 1555, + 1405, + 1648, + 298, + 1648 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1327, + 1405, + 1327, + 1405, + 1425, + 297, + 1425 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1915, + 1403, + 1915, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 317, + 241, + 1388, + 241, + 1388, + 708, + 317, + 708 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 618, + 1815, + 1080, + 1815, + 1080, + 1905, + 618, + 1905 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 299, + 1737, + 1403, + 1737, + 1403, + 1799, + 299, + 1799 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 293, + 944, + 1402, + 944, + 1402, + 1009, + 293, + 1009 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 295, + 770, + 1405, + 770, + 1405, + 891, + 295, + 891 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 289, + 1133, + 1401, + 1133, + 1401, + 1198, + 289, + 1198 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 630, + 1436, + 1068, + 1436, + 1068, + 1479, + 630, + 1479 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 739, + 1208, + 958, + 1208, + 958, + 1250, + 739, + 1250 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 298, + 1264, + 583, + 1264, + 583, + 1296, + 298, + 1296 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 296, + 1491, + 1404, + 1491, + 1404, + 1527, + 296, + 1527 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 298, + 1681, + 496, + 1681, + 496, + 1712, + 298, + 1712 + ], + "score": 0.912 + }, + { + "category_id": 8, + "poly": [ + 621, + 1020, + 794, + 1020, + 794, + 1057, + 621, + 1057 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1843, + 1400, + 1843, + 1400, + 1873, + 1366, + 1873 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1443, + 1400, + 1443, + 1400, + 1472, + 1366, + 1472 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1026, + 1399, + 1026, + 1399, + 1055, + 1366, + 1055 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1070, + 1400, + 1070, + 1400, + 1098, + 1366, + 1098 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1216, + 1400, + 1216, + 1400, + 1245, + 1366, + 1245 + ], + "score": 0.874 + }, + { + "category_id": 8, + "poly": [ + 623, + 1064, + 1076, + 1064, + 1076, + 1102, + 623, + 1102 + ], + "score": 0.814 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.699 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.373 + }, + { + "category_id": 8, + "poly": [ + 621, + 1022, + 1076, + 1022, + 1076, + 1103, + 621, + 1103 + ], + "score": 0.192 + }, + { + "category_id": 14, + "poly": [ + 620, + 1811, + 1078, + 1811, + 1078, + 1906, + 620, + 1906 + ], + "score": 0.95, + "latex": "C ( \\boldsymbol { \\theta } ) = \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } - \\log p ( \\mathbf { y } ^ { n } | \\mathbf { x } _ { 1 } ^ { n } , \\ldots , \\mathbf { x } _ { T } ^ { n } ) ," + }, + { + "category_id": 14, + "poly": [ + 627, + 1435, + 1071, + 1435, + 1071, + 1480, + 627, + 1480 + ], + "score": 0.93, + "latex": "\\mathbf { C } ^ { t } [ j ] = ( 1 - \\mathbf { e } ^ { t } u _ { j } ^ { t } ) \\odot \\mathbf { C } ^ { t - 1 } [ j ] + u _ { j } ^ { t } \\bar { \\mathbf { c } } ^ { t } ." + }, + { + "category_id": 13, + "poly": [ + 1131, + 1133, + 1276, + 1133, + 1276, + 1168, + 1131, + 1168 + ], + "score": 0.93, + "latex": "\\phi ^ { t } \\in \\mathbb { R } ^ { d _ { a } + d _ { c } }" + }, + { + "category_id": 14, + "poly": [ + 739, + 1207, + 958, + 1207, + 958, + 1250, + 739, + 1250 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\phi ^ { t } = ( \\mathbf { w } ^ { t } ) ^ { \\top } \\mathbf { M } ^ { t - 1 } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 855, + 945, + 963, + 945, + 963, + 976, + 855, + 976 + ], + "score": 0.92, + "latex": "\\mathbf { h } ^ { t } \\in \\mathbb { R } ^ { d _ { h } }" + }, + { + "category_id": 13, + "poly": [ + 565, + 1492, + 629, + 1492, + 629, + 1527, + 565, + 1527 + ], + "score": 0.92, + "latex": "\\mathbf { C } ^ { t } [ j ]" + }, + { + "category_id": 14, + "poly": [ + 619, + 1017, + 1079, + 1017, + 1079, + 1105, + 619, + 1105 + ], + "score": 0.92, + "latex": "\\begin{array} { l } { { \\boldsymbol { \\alpha } } ^ { t } = \\mathbf { f } ( \\mathbf { h } ^ { t } , \\mathbf { x } ^ { t } ) , } \\\\ { { \\bar { \\mathbf { c } } } ^ { t } = \\mathrm { R e L U } ( \\mathbf { W } _ { m } \\mathbf { h } ^ { t } + { \\boldsymbol { \\alpha } } ^ { t } \\mathbf { W } _ { x } \\mathbf { x } ^ { t } + \\mathbf { b _ { m } } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1739, + 1266, + 1739, + 1266, + 1772, + 1080, + 1772 + ], + "score": 0.91, + "latex": "p ( \\mathbf { y } | \\mathbf { x } _ { 1 } , \\ldots , \\mathbf { x } _ { T } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 946, + 342, + 946, + 342, + 974, + 298, + 974 + ], + "score": 0.88, + "latex": "\\mathbb { R } ^ { d _ { c } }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 1493, + 1105, + 1493, + 1105, + 1522, + 1070, + 1522 + ], + "score": 0.88, + "latex": "\\mathbf { C } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 645, + 1135, + 680, + 1135, + 680, + 1164, + 645, + 1164 + ], + "score": 0.88, + "latex": "\\mathbf { w } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1355, + 1493, + 1398, + 1493, + 1398, + 1522, + 1355, + 1522 + ], + "score": 0.88, + "latex": "\\mathbf { M } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 436, + 977, + 466, + 977, + 466, + 1004, + 436, + 1004 + ], + "score": 0.88, + "latex": "\\alpha ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1264, + 407, + 1264, + 407, + 1293, + 371, + 1293 + ], + "score": 0.86, + "latex": "\\mathbf { w } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1327, + 1345, + 1327, + 1345, + 1364, + 1266, + 1364 + ], + "score": 0.86, + "latex": "( \\mathbf { e } ^ { t } , u _ { j } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1323, + 1364, + 1353, + 1364, + 1353, + 1391, + 1323, + 1391 + ], + "score": 0.85, + "latex": "\\mathbf { h } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1362, + 324, + 1362, + 324, + 1392, + 297, + 1392 + ], + "score": 0.83, + "latex": "\\bar { \\mathbf { c } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 764, + 1496, + 779, + 1496, + 779, + 1526, + 764, + 1526 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 516, + 1496, + 532, + 1496, + 532, + 1526, + 516, + 1526 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 371, + 1918, + 389, + 1918, + 389, + 1944, + 371, + 1944 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 243.0, + 571.0, + 243.0, + 571.0, + 289.0, + 489.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 247.0, + 826.0, + 247.0, + 826.0, + 283.0, + 684.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 307.0, + 1199.0, + 307.0, + 1199.0, + 336.0, + 1125.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 310.0, + 1332.0, + 310.0, + 1332.0, + 335.0, + 1266.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 355.0, + 414.0, + 355.0, + 414.0, + 386.0, + 320.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 367.0, + 752.0, + 367.0, + 752.0, + 380.0, + 743.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 347.0, + 1200.0, + 347.0, + 1200.0, + 371.0, + 1126.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 348.0, + 1329.0, + 348.0, + 1329.0, + 369.0, + 1267.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 381.0, + 1199.0, + 381.0, + 1199.0, + 405.0, + 1125.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 383.0, + 1331.0, + 383.0, + 1331.0, + 405.0, + 1267.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 425.0, + 406.0, + 425.0, + 406.0, + 459.0, + 319.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 416.0, + 1200.0, + 416.0, + 1200.0, + 440.0, + 1126.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 419.0, + 1329.0, + 419.0, + 1329.0, + 440.0, + 1267.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 456.0, + 994.0, + 456.0, + 994.0, + 490.0, + 916.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 453.0, + 1200.0, + 453.0, + 1200.0, + 478.0, + 1126.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 453.0, + 1332.0, + 453.0, + 1332.0, + 478.0, + 1266.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 496.0, + 424.0, + 496.0, + 424.0, + 529.0, + 318.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 500.0, + 817.0, + 500.0, + 817.0, + 520.0, + 794.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 492.0, + 1009.0, + 492.0, + 1009.0, + 550.0, + 912.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 490.0, + 1200.0, + 490.0, + 1200.0, + 514.0, + 1125.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 490.0, + 1332.0, + 490.0, + 1332.0, + 515.0, + 1266.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 526.0, + 1329.0, + 526.0, + 1329.0, + 548.0, + 1267.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 540.0, + 1006.0, + 540.0, + 1006.0, + 573.0, + 916.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 577.0, + 756.0, + 577.0, + 756.0, + 587.0, + 743.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 638.0, + 1010.0, + 638.0, + 1010.0, + 667.0, + 921.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 650.0, + 794.0, + 650.0, + 794.0, + 686.0, + 706.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 247.0, + 1272.0, + 247.0, + 1272.0, + 285.0, + 1147.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 525.0, + 1201.0, + 525.0, + 1201.0, + 550.5, + 1127.0, + 550.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 771.0, + 1403.0, + 771.0, + 1403.0, + 806.0, + 295.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 799.0, + 1403.0, + 799.0, + 1403.0, + 833.0, + 294.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 858.0, + 1284.0, + 858.0, + 1284.0, + 893.0, + 297.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1680.0, + 497.0, + 1680.0, + 497.0, + 1716.0, + 295.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1554.0, + 1406.0, + 1554.0, + 1406.0, + 1592.0, + 293.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1583.0, + 1405.0, + 1583.0, + 1405.0, + 1622.0, + 292.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1615.0, + 1386.0, + 1615.0, + 1386.0, + 1652.0, + 293.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1325.0, + 1265.0, + 1325.0, + 1265.0, + 1364.0, + 292.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1325.0, + 1406.0, + 1325.0, + 1406.0, + 1364.0, + 1346.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1360.0, + 1322.0, + 1360.0, + 1322.0, + 1398.0, + 325.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1360.0, + 1403.0, + 1360.0, + 1403.0, + 1398.0, + 1354.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1393.0, + 635.0, + 1393.0, + 635.0, + 1429.0, + 297.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1914.0, + 370.0, + 1914.0, + 370.0, + 1953.0, + 294.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1914.0, + 1405.0, + 1914.0, + 1405.0, + 1953.0, + 390.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1981.0, + 293.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 732.0, + 2003.0, + 732.0, + 2035.0, + 294.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1738.0, + 1079.0, + 1738.0, + 1079.0, + 1773.0, + 295.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1738.0, + 1405.0, + 1738.0, + 1405.0, + 1773.0, + 1267.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1765.0, + 921.0, + 1765.0, + 921.0, + 1801.0, + 294.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 940.0, + 297.0, + 940.0, + 297.0, + 983.0, + 293.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 940.0, + 854.0, + 940.0, + 854.0, + 983.0, + 343.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 940.0, + 1405.0, + 940.0, + 1405.0, + 983.0, + 964.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 971.0, + 435.0, + 971.0, + 435.0, + 1011.0, + 293.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 971.0, + 1409.0, + 971.0, + 1409.0, + 1011.0, + 467.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1129.0, + 644.0, + 1129.0, + 644.0, + 1169.0, + 292.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1129.0, + 1130.0, + 1129.0, + 1130.0, + 1169.0, + 681.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1129.0, + 1406.0, + 1129.0, + 1406.0, + 1169.0, + 1277.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1162.0, + 335.0, + 1162.0, + 335.0, + 1204.0, + 291.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1258.0, + 370.0, + 1258.0, + 370.0, + 1300.0, + 293.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1258.0, + 587.0, + 1258.0, + 587.0, + 1300.0, + 408.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1489.0, + 515.0, + 1489.0, + 515.0, + 1529.0, + 294.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1489.0, + 564.0, + 1489.0, + 564.0, + 1529.0, + 533.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1489.0, + 763.0, + 1489.0, + 763.0, + 1529.0, + 630.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1489.0, + 1069.0, + 1489.0, + 1069.0, + 1529.0, + 780.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1489.0, + 1354.0, + 1489.0, + 1354.0, + 1529.0, + 1106.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1489.0, + 1407.0, + 1489.0, + 1407.0, + 1529.0, + 1399.0, + 1529.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1418, + 1404, + 1418, + 1404, + 1596, + 297, + 1596 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1715, + 1404, + 1715, + 1404, + 1894, + 297, + 1894 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1106, + 1404, + 1106, + 1404, + 1225, + 297, + 1225 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1314, + 1404, + 1314, + 1404, + 1405, + 298, + 1405 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 480, + 1404, + 480, + 1404, + 579, + 297, + 579 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 349, + 1403, + 349, + 1403, + 436, + 300, + 436 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 696, + 953, + 1002, + 953, + 1002, + 1022, + 696, + 1022 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 738, + 820, + 958, + 820, + 958, + 896, + 738, + 896 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 737, + 436, + 958, + 436, + 958, + 474, + 737, + 474 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 297, + 908, + 876, + 908, + 876, + 942, + 297, + 942 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 299, + 1935, + 987, + 1935, + 987, + 1971, + 299, + 1971 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 297, + 2001, + 1134, + 2001, + 1134, + 2035, + 297, + 2035 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 296, + 683, + 863, + 683, + 863, + 717, + 296, + 717 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 744, + 772, + 959, + 772, + 959, + 812, + 744, + 812 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 301, + 225, + 733, + 225, + 733, + 262, + 301, + 262 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 298, + 729, + 745, + 729, + 745, + 761, + 298, + 761 + ], + "score": 0.918 + }, + { + "category_id": 8, + "poly": [ + 721, + 633, + 1076, + 633, + 1076, + 671, + 721, + 671 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 298, + 1259, + 946, + 1259, + 946, + 1291, + 298, + 1291 + ], + "score": 0.915 + }, + { + "category_id": 8, + "poly": [ + 622, + 590, + 971, + 590, + 971, + 626, + 622, + 626 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 300, + 1050, + 680, + 1050, + 680, + 1081, + 300, + 1081 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 857, + 74, + 857, + 105, + 298, + 105 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 299, + 292, + 603, + 292, + 603, + 324, + 299, + 324 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1352, + 777, + 1400, + 777, + 1400, + 808, + 1352, + 808 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1352, + 839, + 1400, + 839, + 1400, + 870, + 1352, + 870 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1615, + 1399, + 1615, + 1399, + 1645, + 1352, + 1645 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1658, + 1400, + 1658, + 1400, + 1687, + 1353, + 1687 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 593, + 1399, + 593, + 1399, + 622, + 1366, + 622 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1366, + 638, + 1400, + 638, + 1400, + 667, + 1366, + 667 + ], + "score": 0.865 + }, + { + "category_id": 8, + "poly": [ + 692, + 1610, + 1007, + 1610, + 1007, + 1650, + 692, + 1650 + ], + "score": 0.8 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2111, + 841, + 2111 + ], + "score": 0.787 + }, + { + "category_id": 8, + "poly": [ + 686, + 1657, + 1010, + 1657, + 1010, + 1691, + 686, + 1691 + ], + "score": 0.756 + }, + { + "category_id": 8, + "poly": [ + 684, + 1610, + 1012, + 1610, + 1012, + 1691, + 684, + 1691 + ], + "score": 0.161 + }, + { + "category_id": 14, + "poly": [ + 694, + 950, + 1004, + 950, + 1004, + 1020, + 694, + 1020 + ], + "score": 0.94, + "latex": "S ( \\mathbf { x } , \\mathbf { y } ) = \\frac { \\mathbf { x } \\cdot \\mathbf { y } } { ( | | \\mathbf { x } | | | \\mathbf { y } | | + \\epsilon ) } ." + }, + { + "category_id": 13, + "poly": [ + 1058, + 545, + 1169, + 545, + 1169, + 579, + 1058, + 579 + ], + "score": 0.93, + "latex": "\\beta _ { t } \\in \\mathbb { R } ^ { \\ge 1 }" + }, + { + "category_id": 13, + "poly": [ + 625, + 908, + 729, + 908, + 729, + 939, + 625, + 939 + ], + "score": 0.93, + "latex": "S \\in \\mathbb { R } ^ { \\geq 0 }" + }, + { + "category_id": 14, + "poly": [ + 735, + 433, + 958, + 433, + 958, + 475, + 735, + 475 + ], + "score": 0.92, + "latex": "\\mathbf { k } ^ { t } = \\mathbf { W } _ { k } ^ { \\top } \\mathbf { h } ^ { t } + \\mathbf { b } _ { k } ," + }, + { + "category_id": 13, + "poly": [ + 676, + 480, + 849, + 480, + 849, + 514, + 676, + 514 + ], + "score": 0.92, + "latex": "\\mathbf { b } _ { k } \\ \\in \\ \\mathbb { R } ^ { d _ { a } + d _ { c } }" + }, + { + "category_id": 13, + "poly": [ + 376, + 479, + 616, + 479, + 616, + 513, + 376, + 513 + ], + "score": 0.92, + "latex": "\\mathbf { W } _ { k } \\ \\in \\ \\mathbb { R } ^ { N \\times ( d _ { a } + d _ { c } ) }" + }, + { + "category_id": 13, + "poly": [ + 296, + 514, + 458, + 514, + 458, + 546, + 296, + 546 + ], + "score": 0.92, + "latex": "\\mathbf { W } _ { k } \\in \\mathbb { R } ^ { N \\times d _ { c } }" + }, + { + "category_id": 14, + "poly": [ + 683, + 1605, + 1014, + 1605, + 1014, + 1696, + 683, + 1696 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\gamma _ { t } = \\operatorname { s i g m o i d } ( \\mathbf { u } _ { \\gamma } ^ { \\top } \\mathbf { h } _ { t } + \\mathbf { b } _ { \\gamma } ) , } \\\\ { \\mathbf { w } _ { t } = \\operatorname { s o f t m a x } \\bigl ( \\mathbf { z } _ { t } - \\gamma _ { t } \\mathbf { v } _ { t - 1 } \\bigr ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 507, + 513, + 617, + 513, + 617, + 546, + 507, + 546 + ], + "score": 0.9, + "latex": "\\mathbf { b } _ { k } \\in \\mathbb { R } ^ { d _ { c } }" + }, + { + "category_id": 13, + "poly": [ + 368, + 1450, + 630, + 1450, + 630, + 1480, + 368, + 1480 + ], + "score": 0.9, + "latex": "\\mathbf { v } _ { t } ~ = ~ 0 . 1 \\mathbf { v } _ { t - 1 } + 0 . 9 \\mathbf { z } _ { t }" + }, + { + "category_id": 14, + "poly": [ + 619, + 586, + 1079, + 586, + 1079, + 676, + 619, + 676 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { \\mathrm { s o f t p l u s } ( x ) = \\log ( \\exp ( x ) + 1 ) } } \\\\ { { \\beta _ { t } = \\mathrm { s o f t p l u s } ( \\mathbf { u } _ { \\beta } ^ { \\top } \\mathbf { h } ^ { t } + b _ { \\beta } ) + 1 . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 737, + 769, + 961, + 769, + 961, + 900, + 737, + 900 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { z _ { i } ^ { t } = \\beta ^ { t } S \\left( \\mathbf { k } ^ { t } , \\mathbf { m } _ { i } ^ { t } \\right) } } \\\\ { { w _ { i } ^ { t } = \\displaystyle \\frac { \\exp ( z _ { i } ^ { t } ) } { \\sum _ { j } \\exp ( z _ { j } ^ { t } ) } , } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 384, + 686, + 413, + 686, + 413, + 718, + 384, + 718 + ], + "score": 0.89, + "latex": "b _ { \\beta }" + }, + { + "category_id": 13, + "poly": [ + 1188, + 1565, + 1223, + 1565, + 1223, + 1598, + 1188, + 1598 + ], + "score": 0.89, + "latex": "{ \\bf { b } } _ { \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 1568, + 1138, + 1568, + 1138, + 1598, + 1103, + 1598 + ], + "score": 0.88, + "latex": "\\mathbf { u } _ { \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 826, + 685, + 854, + 685, + 854, + 717, + 826, + 717 + ], + "score": 0.88, + "latex": "\\beta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 945, + 1452, + 975, + 1452, + 975, + 1479, + 945, + 1479 + ], + "score": 0.87, + "latex": "\\mathbf { v } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 784, + 1510, + 814, + 1510, + 814, + 1537, + 784, + 1537 + ], + "score": 0.86, + "latex": "\\mathbf { v } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1453, + 1060, + 1453, + 1060, + 1480, + 1032, + 1480 + ], + "score": 0.86, + "latex": "\\gamma _ { t }" + }, + { + "category_id": 13, + "poly": [ + 670, + 1569, + 697, + 1569, + 697, + 1595, + 670, + 1595 + ], + "score": 0.86, + "latex": "\\gamma _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1808, + 1167, + 1808, + 1167, + 1835, + 1138, + 1835 + ], + "score": 0.85, + "latex": "\\mathbf { v } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 349, + 1779, + 376, + 1779, + 376, + 1808, + 349, + 1808 + ], + "score": 0.85, + "latex": "\\gamma _ { t }" + }, + { + "category_id": 13, + "poly": [ + 681, + 1539, + 708, + 1539, + 708, + 1566, + 681, + 1566 + ], + "score": 0.85, + "latex": "\\gamma _ { t }" + }, + { + "category_id": 13, + "poly": [ + 879, + 1510, + 907, + 1510, + 907, + 1536, + 879, + 1536 + ], + "score": 0.85, + "latex": "\\mathbf { z } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 689, + 333, + 689, + 333, + 719, + 297, + 719 + ], + "score": 0.84, + "latex": "\\mathbf { u } _ { \\beta }" + }, + { + "category_id": 13, + "poly": [ + 1358, + 1422, + 1403, + 1422, + 1403, + 1451, + 1358, + 1451 + ], + "score": 0.83, + "latex": "\\left( \\mathbf { z } _ { t } \\right)" + }, + { + "category_id": 13, + "poly": [ + 326, + 1453, + 355, + 1453, + 355, + 1479, + 326, + 1479 + ], + "score": 0.69, + "latex": "\\mathbf { v } _ { t }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1931.0, + 989.0, + 1931.0, + 989.0, + 1978.0, + 291.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 738.0, + 221.0, + 738.0, + 268.0, + 291.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1259.0, + 946.0, + 1259.0, + 946.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1050.0, + 681.0, + 1050.0, + 681.0, + 1083.0, + 296.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 605.0, + 292.0, + 605.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 837.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 1357.0, + 1418.0, + 1357.0, + 1455.0, + 292.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1449.0, + 325.0, + 1449.0, + 325.0, + 1483.0, + 294.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1449.0, + 367.0, + 1449.0, + 367.0, + 1483.0, + 356.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1449.0, + 944.0, + 1449.0, + 944.0, + 1483.0, + 631.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1449.0, + 1031.0, + 1449.0, + 1031.0, + 1483.0, + 976.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1449.0, + 1405.0, + 1449.0, + 1405.0, + 1483.0, + 1061.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1477.0, + 1405.0, + 1477.0, + 1405.0, + 1512.0, + 295.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1506.0, + 783.0, + 1506.0, + 783.0, + 1541.0, + 295.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1506.0, + 878.0, + 1506.0, + 878.0, + 1541.0, + 815.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1506.0, + 1404.0, + 1506.0, + 1404.0, + 1541.0, + 908.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1537.0, + 680.0, + 1537.0, + 680.0, + 1568.0, + 295.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1568.0, + 709.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1560.0, + 669.0, + 1560.0, + 669.0, + 1603.0, + 292.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1560.0, + 1102.0, + 1560.0, + 1102.0, + 1603.0, + 698.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1560.0, + 1187.0, + 1560.0, + 1187.0, + 1603.0, + 1139.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1560.0, + 1235.0, + 1560.0, + 1235.0, + 1603.0, + 1224.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1715.0, + 1404.0, + 1715.0, + 1404.0, + 1751.0, + 294.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1746.0, + 1405.0, + 1746.0, + 1405.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1775.0, + 348.0, + 1775.0, + 348.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1775.0, + 1405.0, + 1775.0, + 1405.0, + 1809.0, + 377.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1137.0, + 1803.0, + 1137.0, + 1842.0, + 294.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1803.0, + 1406.0, + 1803.0, + 1406.0, + 1842.0, + 1168.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 1407.0, + 1833.0, + 1407.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1859.0, + 1223.0, + 1859.0, + 1223.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1107.0, + 1404.0, + 1107.0, + 1404.0, + 1141.0, + 295.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1131.0, + 1406.0, + 1131.0, + 1406.0, + 1172.0, + 292.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1163.0, + 1406.0, + 1163.0, + 1406.0, + 1198.0, + 293.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1193.0, + 1255.0, + 1193.0, + 1255.0, + 1227.0, + 295.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1315.0, + 1404.0, + 1315.0, + 1404.0, + 1349.0, + 294.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1345.0, + 1406.0, + 1345.0, + 1406.0, + 1379.0, + 294.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1370.0, + 978.0, + 1370.0, + 978.0, + 1411.0, + 292.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 472.0, + 375.0, + 472.0, + 375.0, + 522.0, + 290.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 472.0, + 675.0, + 472.0, + 675.0, + 522.0, + 617.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 472.0, + 1406.0, + 472.0, + 1406.0, + 522.0, + 850.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 507.0, + 295.0, + 507.0, + 295.0, + 552.0, + 292.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 507.0, + 506.0, + 507.0, + 506.0, + 552.0, + 459.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 507.0, + 1406.0, + 507.0, + 1406.0, + 552.0, + 618.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 544.0, + 1057.0, + 544.0, + 1057.0, + 583.0, + 292.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 544.0, + 1353.0, + 544.0, + 1353.0, + 583.0, + 1170.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 347.0, + 1404.0, + 347.0, + 1404.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 380.0, + 1405.0, + 380.0, + 1405.0, + 410.0, + 294.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 410.0, + 558.0, + 410.0, + 558.0, + 438.0, + 295.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 906.0, + 624.0, + 906.0, + 624.0, + 946.0, + 294.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 906.0, + 879.0, + 906.0, + 879.0, + 946.0, + 730.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1995.0, + 1137.0, + 1995.0, + 1137.0, + 2043.0, + 293.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 681.0, + 383.0, + 681.0, + 383.0, + 724.0, + 334.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 681.0, + 825.0, + 681.0, + 825.0, + 724.0, + 414.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 681.0, + 865.0, + 681.0, + 865.0, + 724.0, + 855.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 722.0, + 748.0, + 722.0, + 748.0, + 770.0, + 293.0, + 770.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1682, + 1405, + 1682, + 1405, + 1884, + 296, + 1884 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 537, + 1404, + 537, + 1404, + 686, + 297, + 686 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1404, + 229, + 1404, + 346, + 296, + 346 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 497, + 1219, + 1338, + 1219, + 1338, + 1457, + 497, + 1457 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 1894, + 1403, + 1894, + 1403, + 1985, + 299, + 1985 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 919, + 1403, + 919, + 1403, + 1013, + 297, + 1013 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 739, + 740, + 961, + 740, + 961, + 813, + 739, + 813 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 635, + 1023, + 1060, + 1023, + 1060, + 1097, + 635, + 1097 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 289, + 1556, + 1400, + 1556, + 1400, + 1618, + 289, + 1618 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 819, + 1404, + 819, + 1404, + 878, + 297, + 878 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 694, + 1989, + 1004, + 1989, + 1004, + 2029, + 694, + 2029 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 293, + 1106, + 1403, + 1106, + 1403, + 1167, + 293, + 1167 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 721, + 879, + 978, + 879, + 978, + 917, + 721, + 917 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 293, + 1465, + 1401, + 1465, + 1401, + 1527, + 293, + 1527 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 756, + 439, + 941, + 439, + 941, + 473, + 756, + 473 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 701, + 1626, + 995, + 1626, + 995, + 1662, + 701, + 1662 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 298, + 382, + 1401, + 382, + 1401, + 439, + 298, + 439 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 783, + 346, + 913, + 346, + 913, + 380, + 783, + 380 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 299, + 474, + 830, + 474, + 830, + 508, + 299, + 508 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 291, + 697, + 1383, + 697, + 1383, + 733, + 291, + 733 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 857, + 74, + 857, + 105, + 297, + 105 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 299, + 1178, + 1064, + 1178, + 1064, + 1211, + 299, + 1211 + ], + "score": 0.92 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1995, + 1400, + 1995, + 1400, + 2025, + 1352, + 2025 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.762 + }, + { + "category_id": 14, + "poly": [ + 738, + 738, + 959, + 738, + 959, + 815, + 738, + 815 + ], + "score": 0.95, + "latex": "\\tilde { R } ( { \\bf x } ) = \\frac { R ( { \\bf x } ) - b } { \\sqrt { \\sigma ^ { 2 } + \\epsilon } } ," + }, + { + "category_id": 14, + "poly": [ + 634, + 1021, + 1060, + 1021, + 1060, + 1099, + 634, + 1099 + ], + "score": 0.94, + "latex": "H _ { \\delta } ( x ) = { \\left\\{ \\begin{array} { l l } { x ^ { 2 } } & { { \\mathrm { f o r ~ } } | x | \\leq \\delta , } \\\\ { \\delta ( 2 | x | - \\delta ) , } & { { \\mathrm { o t h e r w i s e } } , } \\end{array} \\right. }" + }, + { + "category_id": 14, + "poly": [ + 494, + 1215, + 1336, + 1215, + 1336, + 1458, + 494, + 1458 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { C ^ { n } ( \\boldsymbol { \\theta } ) = - \\log p ( \\mathbf { y } | \\mathbf { x } _ { 1 : T } , \\tilde { w } _ { 1 : J } , \\tilde { u } _ { 1 : J } , \\tilde { e } _ { 1 : J } ) } \\\\ & { \\qquad - \\displaystyle \\sum _ { j = 1 } ^ { J } \\tilde { R } ( \\mathbf { x } ^ { n } ) ( \\log p ( \\tilde { w } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { u } _ { j } | \\mathbf { x } _ { 1 : T } ) + \\log p ( \\tilde { e } _ { j } | \\mathbf { x } _ { 1 : T } ) ) } \\\\ & { \\qquad - \\lambda _ { H } \\displaystyle \\sum _ { j = 1 } ^ { J } ( \\mathcal { H } ( w _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( u _ { j } | \\mathbf { x } _ { 1 : T } ) + \\mathcal { H } ( e _ { j } | \\mathbf { x } _ { 1 : T } ) ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 976, + 372, + 976, + 372, + 1014, + 298, + 1014 + ], + "score": 0.93, + "latex": "\\tilde { R } ( { \\bf x } ) ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 757, + 1833, + 881, + 1833, + 881, + 1884, + 757, + 1884 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { p ^ { t } = \\frac { p ^ { 0 } } { \\sqrt { 1 + t } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 447, + 699, + 772, + 699, + 772, + 733, + 447, + 733 + ], + "score": 0.93, + "latex": "R ( { \\bf x } ) = \\log p ( { \\bf y } | { \\bf x } _ { 1 } , \\ldots , { \\bf x } _ { T } )" + }, + { + "category_id": 13, + "poly": [ + 371, + 475, + 484, + 475, + 484, + 508, + 371, + 508 + ], + "score": 0.92, + "latex": "j \\sim \\mathcal { C } ( \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 1194, + 1802, + 1275, + 1802, + 1275, + 1835, + 1194, + 1835 + ], + "score": 0.92, + "latex": "p ^ { 0 } = 1" + }, + { + "category_id": 14, + "poly": [ + 718, + 876, + 977, + 876, + 977, + 918, + 718, + 918 + ], + "score": 0.92, + "latex": "\\tilde { R } ( \\mathbf { x } ) \\gets \\tilde { R } ( \\mathbf { x } ) - b ( \\mathbf { x } ) ," + }, + { + "category_id": 14, + "poly": [ + 692, + 1988, + 1006, + 1988, + 1006, + 2029, + 692, + 2029 + ], + "score": 0.92, + "latex": "\\mathbf { w } ^ { t } \\pi ^ { t } \\bar { \\mathbf { w } } ^ { t } + ( 1 - \\pi ^ { t } ) \\tilde { \\mathbf { w } } ^ { t } ." + }, + { + "category_id": 14, + "poly": [ + 783, + 343, + 913, + 343, + 913, + 381, + 783, + 381 + ], + "score": 0.91, + "latex": "p ( j ) = w _ { j } ," + }, + { + "category_id": 13, + "poly": [ + 368, + 919, + 421, + 919, + 421, + 951, + 368, + 951 + ], + "score": 0.91, + "latex": "b ( \\mathbf x )" + }, + { + "category_id": 14, + "poly": [ + 756, + 436, + 941, + 436, + 941, + 474, + 756, + 474 + ], + "score": 0.91, + "latex": "\\tilde { w } _ { k } = I ( k = j ) ," + }, + { + "category_id": 14, + "poly": [ + 701, + 1624, + 997, + 1624, + 997, + 1662, + 701, + 1662 + ], + "score": 0.91, + "latex": "\\tilde { w } _ { k } = \\mathbf { I } ( k = \\operatorname { a r g m a x } ( \\mathbf { w } ) ) ." + }, + { + "category_id": 13, + "poly": [ + 1188, + 289, + 1252, + 289, + 1252, + 322, + 1188, + 322 + ], + "score": 0.9, + "latex": "\\mathcal C ( \\mathbf w )" + }, + { + "category_id": 13, + "poly": [ + 657, + 980, + 709, + 980, + 709, + 1014, + 657, + 1014 + ], + "score": 0.9, + "latex": "b ( \\mathbf { x } )" + }, + { + "category_id": 13, + "poly": [ + 763, + 1467, + 803, + 1467, + 803, + 1496, + 763, + 1496 + ], + "score": 0.89, + "latex": "\\lambda _ { H }" + }, + { + "category_id": 13, + "poly": [ + 802, + 1771, + 959, + 1771, + 959, + 1805, + 802, + 1805 + ], + "score": 0.89, + "latex": "\\pi ^ { t } \\sim \\mathrm { B i n } ( p ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 494, + 1925, + 530, + 1925, + 530, + 1952, + 494, + 1952 + ], + "score": 0.88, + "latex": "\\hat { \\mathbf { w } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 968, + 1925, + 996, + 1925, + 996, + 1952, + 968, + 1952 + ], + "score": 0.88, + "latex": "\\mathit { \\Pi } _ { \\pi ^ { t } }" + }, + { + "category_id": 13, + "poly": [ + 872, + 1925, + 907, + 1925, + 907, + 1952, + 872, + 1952 + ], + "score": 0.88, + "latex": "\\tilde { \\mathbf { w } } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 431, + 1841, + 458, + 1841, + 458, + 1875, + 431, + 1875 + ], + "score": 0.87, + "latex": "p ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 748, + 1804, + 776, + 1804, + 776, + 1831, + 748, + 1831 + ], + "score": 0.87, + "latex": "\\pi ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 587, + 1900, + 623, + 1900, + 623, + 1926, + 587, + 1926 + ], + "score": 0.87, + "latex": "\\mathbf { w } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 369, + 386, + 404, + 386, + 404, + 415, + 369, + 415 + ], + "score": 0.87, + "latex": "w _ { j }" + }, + { + "category_id": 13, + "poly": [ + 369, + 1467, + 390, + 1467, + 390, + 1493, + 369, + 1493 + ], + "score": 0.85, + "latex": "J" + }, + { + "category_id": 13, + "poly": [ + 473, + 384, + 489, + 384, + 489, + 414, + 473, + 414 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1286, + 1467, + 1312, + 1467, + 1312, + 1493, + 1286, + 1493 + ], + "score": 0.82, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 948, + 822, + 971, + 822, + 971, + 848, + 948, + 848 + ], + "score": 0.82, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 434, + 827, + 455, + 827, + 455, + 848, + 434, + 848 + ], + "score": 0.8, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 542, + 477, + 560, + 477, + 560, + 503, + 542, + 503 + ], + "score": 0.79, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 1310, + 289, + 1403, + 289, + 1403, + 322, + 1310, + 322 + ], + "score": 0.78, + "latex": "\\dim ( \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 1299, + 1748, + 1320, + 1748, + 1320, + 1770, + 1299, + 1770 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 369, + 822, + 385, + 822, + 385, + 848, + 369, + 848 + ], + "score": 0.72, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 1023, + 925, + 1044, + 925, + 1044, + 946, + 1023, + 946 + ], + "score": 0.65, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 766, + 1771, + 793, + 1771, + 793, + 1803, + 766, + 1803 + ], + "score": 0.59, + "latex": "p ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 855, + 317, + 855, + 317, + 876, + 297, + 876 + ], + "score": 0.53, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 297, + 262, + 311, + 262, + 311, + 288, + 297, + 288 + ], + "score": 0.5, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1683.0, + 1405.0, + 1683.0, + 1405.0, + 1717.0, + 296.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1406.0, + 1710.0, + 1406.0, + 1751.0, + 293.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1298.0, + 1741.0, + 1298.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 1741.0, + 1407.0, + 1741.0, + 1407.0, + 1776.0, + 1321.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 765.0, + 1768.0, + 765.0, + 1806.0, + 294.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1768.0, + 801.0, + 1768.0, + 801.0, + 1806.0, + 794.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1768.0, + 1406.0, + 1768.0, + 1406.0, + 1806.0, + 960.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1798.0, + 747.0, + 1798.0, + 747.0, + 1842.0, + 291.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1798.0, + 1193.0, + 1798.0, + 1193.0, + 1842.0, + 777.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1798.0, + 1406.0, + 1798.0, + 1406.0, + 1842.0, + 1276.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1838.0, + 430.0, + 1838.0, + 430.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1838.0, + 756.0, + 1838.0, + 756.0, + 1885.0, + 459.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1838.0, + 891.0, + 1838.0, + 891.0, + 1885.0, + 882.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.25, + 1831.5, + 872.25, + 1831.5, + 872.25, + 1861.0, + 826.25, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 535.0, + 1405.0, + 535.0, + 1405.0, + 574.0, + 295.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 566.0, + 1404.0, + 566.0, + 1404.0, + 599.0, + 296.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 596.0, + 1405.0, + 596.0, + 1405.0, + 628.0, + 296.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 623.0, + 1405.0, + 623.0, + 1405.0, + 659.0, + 295.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 652.0, + 495.0, + 652.0, + 495.0, + 689.0, + 294.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1406.0, + 226.0, + 1406.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 293.0, + 312.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 285.0, + 1187.0, + 285.0, + 1187.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 285.0, + 1309.0, + 285.0, + 1309.0, + 326.0, + 1253.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 314.0, + 396.0, + 314.0, + 396.0, + 350.0, + 294.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1895.0, + 586.0, + 1895.0, + 586.0, + 1929.0, + 295.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1895.0, + 1403.0, + 1895.0, + 1403.0, + 1929.0, + 624.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1923.0, + 493.0, + 1923.0, + 493.0, + 1959.0, + 294.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1923.0, + 871.0, + 1923.0, + 871.0, + 1959.0, + 531.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1923.0, + 967.0, + 1923.0, + 967.0, + 1959.0, + 908.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1923.0, + 1405.0, + 1923.0, + 1405.0, + 1959.0, + 997.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1955.0, + 516.0, + 1955.0, + 516.0, + 1986.0, + 297.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 367.0, + 917.0, + 367.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 917.0, + 1022.0, + 917.0, + 1022.0, + 951.0, + 422.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 917.0, + 1407.0, + 917.0, + 1407.0, + 951.0, + 1045.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 947.0, + 1405.0, + 947.0, + 1405.0, + 981.0, + 295.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 977.0, + 656.0, + 977.0, + 656.0, + 1017.0, + 373.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 977.0, + 1206.0, + 977.0, + 1206.0, + 1017.0, + 710.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1555.0, + 1403.0, + 1555.0, + 1403.0, + 1592.0, + 293.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1586.0, + 1172.0, + 1586.0, + 1172.0, + 1621.0, + 295.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 817.0, + 368.0, + 817.0, + 368.0, + 856.0, + 294.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 817.0, + 433.0, + 817.0, + 433.0, + 856.0, + 386.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 817.0, + 947.0, + 817.0, + 947.0, + 856.0, + 456.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 817.0, + 1405.0, + 817.0, + 1405.0, + 856.0, + 972.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 849.0, + 296.0, + 849.0, + 296.0, + 885.0, + 292.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 849.0, + 493.0, + 849.0, + 493.0, + 885.0, + 318.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1144.0, + 294.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1134.0, + 1396.0, + 1134.0, + 1396.0, + 1169.0, + 295.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1461.0, + 368.0, + 1461.0, + 368.0, + 1501.0, + 292.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1461.0, + 762.0, + 1461.0, + 762.0, + 1501.0, + 391.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1461.0, + 1285.0, + 1461.0, + 1285.0, + 1501.0, + 804.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1461.0, + 1406.0, + 1461.0, + 1406.0, + 1501.0, + 1313.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 436.0, + 1490.0, + 436.0, + 1533.0, + 293.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 381.0, + 368.0, + 381.0, + 368.0, + 415.0, + 296.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 381.0, + 472.0, + 381.0, + 472.0, + 415.0, + 405.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 381.0, + 1403.0, + 381.0, + 1403.0, + 415.0, + 490.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 412.0, + 635.0, + 412.0, + 635.0, + 442.0, + 294.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 473.0, + 370.0, + 473.0, + 370.0, + 510.0, + 296.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 473.0, + 541.0, + 473.0, + 541.0, + 510.0, + 485.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 473.0, + 832.0, + 473.0, + 832.0, + 510.0, + 561.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 692.0, + 446.0, + 692.0, + 446.0, + 739.0, + 292.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 692.0, + 1382.0, + 692.0, + 1382.0, + 739.0, + 773.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1172.0, + 1064.0, + 1172.0, + 1064.0, + 1220.0, + 292.0, + 1220.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 550, + 1404, + 550, + 1404, + 729, + 298, + 729 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1286, + 1405, + 1286, + 1405, + 1493, + 297, + 1493 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1506, + 1404, + 1506, + 1404, + 1683, + 298, + 1683 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1696, + 1405, + 1696, + 1405, + 1873, + 298, + 1873 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 933, + 1404, + 933, + 1404, + 1054, + 296, + 1054 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1887, + 1402, + 1887, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 429, + 1405, + 429, + 1405, + 519, + 299, + 519 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 599, + 748, + 1101, + 748, + 1101, + 842, + 599, + 842 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1400, + 229, + 1400, + 320, + 299, + 320 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 293, + 1116, + 1399, + 1116, + 1399, + 1177, + 293, + 1177 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 590, + 1067, + 1112, + 1067, + 1112, + 1105, + 590, + 1105 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 296, + 869, + 1028, + 869, + 1028, + 903, + 296, + 903 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 298, + 361, + 1142, + 361, + 1142, + 398, + 298, + 398 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 300, + 1218, + 587, + 1218, + 587, + 1254, + 300, + 1254 + ], + "score": 0.919 + }, + { + "category_id": 9, + "poly": [ + 1352, + 778, + 1400, + 778, + 1400, + 809, + 1352, + 809 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 105, + 298, + 105 + ], + "score": 0.81 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 105, + 298, + 105 + ], + "score": 0.129 + }, + { + "category_id": 14, + "poly": [ + 597, + 743, + 1103, + 743, + 1103, + 843, + 597, + 843 + ], + "score": 0.94, + "latex": "R _ { \\mathrm { r w } } ( \\mathbf { w } , \\mathbf { u } ) = \\lambda \\sum _ { t ^ { \\prime } = 1 } ^ { T } | | 1 - ( \\frac { 1 } { t ^ { \\prime } } \\sum _ { t = 1 } ^ { t ^ { \\prime } } \\mathbf { u } _ { t } ) ^ { \\top } \\mathbf { w } _ { t ^ { \\prime } } | | _ { 2 } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 668, + 1117, + 722, + 1117, + 722, + 1149, + 668, + 1149 + ], + "score": 0.92, + "latex": "f _ { t + 1 }" + }, + { + "category_id": 14, + "poly": [ + 586, + 1067, + 1111, + 1067, + 1111, + 1105, + 586, + 1105 + ], + "score": 0.92, + "latex": "R _ { \\mathrm { p r e d } } ( \\mathbf { W } ) = - \\log p ( \\mathbf { f } _ { t + 1 } | \\mathbf { f } _ { t } , \\mathbf { w } _ { t } , \\mathbf { u } _ { t } , \\mathbf { M } _ { t } ; \\mathbf { W } ) )" + }, + { + "category_id": 13, + "poly": [ + 772, + 875, + 807, + 875, + 807, + 901, + 772, + 901 + ], + "score": 0.88, + "latex": "\\mathbf { w } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1052, + 259, + 1079, + 259, + 1079, + 291, + 1052, + 291 + ], + "score": 0.87, + "latex": "p ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1118, + 398, + 1118, + 398, + 1148, + 372, + 1148 + ], + "score": 0.87, + "latex": "f _ { t }" + }, + { + "category_id": 13, + "poly": [ + 560, + 875, + 591, + 875, + 591, + 901, + 560, + 901 + ], + "score": 0.85, + "latex": "\\mathbf { u } _ { t }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 359.0, + 1145.0, + 359.0, + 1145.0, + 404.0, + 291.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1216.0, + 593.0, + 1216.0, + 593.0, + 1259.0, + 293.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1406.0, + 552.0, + 1406.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 581.0, + 1406.0, + 581.0, + 1406.0, + 616.0, + 293.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 1405.0, + 640.0, + 1405.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 668.0, + 1406.0, + 668.0, + 1406.0, + 703.0, + 294.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 694.0, + 708.0, + 694.0, + 708.0, + 737.0, + 293.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1288.0, + 1405.0, + 1288.0, + 1405.0, + 1321.0, + 294.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1319.0, + 1403.0, + 1319.0, + 1403.0, + 1349.0, + 295.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1344.0, + 1407.0, + 1344.0, + 1407.0, + 1380.0, + 292.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1373.0, + 1407.0, + 1373.0, + 1407.0, + 1408.0, + 292.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 1406.0, + 1404.0, + 1406.0, + 1436.0, + 294.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1428.0, + 1410.0, + 1428.0, + 1410.0, + 1471.0, + 292.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1460.0, + 870.0, + 1460.0, + 870.0, + 1500.0, + 294.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1505.0, + 1405.0, + 1505.0, + 1405.0, + 1540.0, + 294.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1534.0, + 1409.0, + 1534.0, + 1409.0, + 1570.0, + 292.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1564.0, + 1406.0, + 1564.0, + 1406.0, + 1598.0, + 294.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1629.0, + 293.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1623.0, + 1405.0, + 1623.0, + 1405.0, + 1658.0, + 294.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1091.0, + 1651.0, + 1091.0, + 1686.0, + 294.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1693.0, + 1410.0, + 1693.0, + 1410.0, + 1735.0, + 292.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1725.0, + 1405.0, + 1725.0, + 1405.0, + 1761.0, + 293.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1754.0, + 1406.0, + 1754.0, + 1406.0, + 1789.0, + 291.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1783.0, + 1407.0, + 1783.0, + 1407.0, + 1818.0, + 295.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1810.0, + 1407.0, + 1810.0, + 1407.0, + 1848.0, + 293.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1840.0, + 1099.0, + 1840.0, + 1099.0, + 1877.0, + 293.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 933.0, + 1406.0, + 933.0, + 1406.0, + 971.0, + 294.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 1003.0, + 291.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 989.0, + 1405.0, + 989.0, + 1405.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1020.0, + 1246.0, + 1020.0, + 1246.0, + 1058.0, + 294.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1887.0, + 1403.0, + 1887.0, + 1403.0, + 1919.0, + 297.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1915.0, + 1407.0, + 1915.0, + 1407.0, + 1950.0, + 293.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1945.0, + 1403.0, + 1945.0, + 1403.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 779.0, + 2003.0, + 779.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 430.0, + 1403.0, + 430.0, + 1403.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 456.0, + 1405.0, + 456.0, + 1405.0, + 497.0, + 293.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 491.0, + 834.0, + 491.0, + 834.0, + 520.0, + 297.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 260.0, + 1051.0, + 260.0, + 1051.0, + 294.0, + 295.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 294.0, + 1080.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 288.0, + 541.0, + 288.0, + 541.0, + 321.0, + 296.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1113.0, + 371.0, + 1113.0, + 371.0, + 1153.0, + 292.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1113.0, + 667.0, + 1113.0, + 667.0, + 1153.0, + 399.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1153.0, + 723.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1145.0, + 1036.0, + 1145.0, + 1036.0, + 1180.0, + 295.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 866.0, + 559.0, + 866.0, + 559.0, + 907.0, + 293.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 866.0, + 771.0, + 866.0, + 771.0, + 907.0, + 592.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 866.0, + 1029.0, + 866.0, + 1029.0, + 907.0, + 808.0, + 907.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 637, + 1404, + 637, + 1404, + 818, + 298, + 818 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 901, + 1403, + 901, + 1403, + 1050, + 299, + 1050 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1065, + 1404, + 1065, + 1404, + 1213, + 299, + 1213 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1561, + 1404, + 1561, + 1404, + 1766, + 298, + 1766 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 400, + 1404, + 400, + 1404, + 547, + 298, + 547 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1226, + 1403, + 1226, + 1403, + 1344, + 299, + 1344 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1428, + 1403, + 1428, + 1403, + 1546, + 298, + 1546 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1780, + 1404, + 1780, + 1404, + 1928, + 298, + 1928 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 294, + 230, + 1400, + 230, + 1400, + 291, + 294, + 291 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 298, + 1950, + 1402, + 1950, + 1402, + 2033, + 298, + 2033 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 301, + 1377, + 684, + 1377, + 684, + 1408, + 301, + 1408 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 300, + 332, + 558, + 332, + 558, + 367, + 300, + 367 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 300, + 848, + 474, + 848, + 474, + 880, + 300, + 880 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 298, + 582, + 940, + 582, + 940, + 614, + 298, + 614 + ], + "score": 0.813 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.767 + }, + { + "category_id": 13, + "poly": [ + 299, + 1153, + 388, + 1153, + 388, + 1182, + 299, + 1182 + ], + "score": 0.67, + "latex": "\\mathrm { ( D M N + }" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1943.0, + 1224.0, + 1943.0, + 1224.0, + 1982.0, + 328.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1973.0, + 1407.0, + 1973.0, + 1407.0, + 2012.0, + 328.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 796.0, + 2004.0, + 796.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1372.0, + 689.0, + 1372.0, + 689.0, + 1413.0, + 294.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 328.0, + 561.0, + 328.0, + 561.0, + 373.0, + 293.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 841.0, + 480.0, + 841.0, + 480.0, + 888.0, + 293.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 944.0, + 580.0, + 944.0, + 620.0, + 294.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 632.0, + 1406.0, + 632.0, + 1406.0, + 677.0, + 292.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 663.0, + 1404.0, + 663.0, + 1404.0, + 705.0, + 293.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 696.0, + 1408.0, + 696.0, + 1408.0, + 735.0, + 293.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 726.0, + 1405.0, + 726.0, + 1405.0, + 764.0, + 293.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 758.0, + 1406.0, + 758.0, + 1406.0, + 793.0, + 293.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 785.0, + 1217.0, + 785.0, + 1217.0, + 824.0, + 293.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 902.0, + 1405.0, + 902.0, + 1405.0, + 934.0, + 297.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 931.0, + 1407.0, + 931.0, + 1407.0, + 966.0, + 293.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 964.0, + 1404.0, + 964.0, + 1404.0, + 996.0, + 295.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 1403.0, + 992.0, + 1403.0, + 1025.0, + 295.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1021.0, + 668.0, + 1021.0, + 668.0, + 1053.0, + 297.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1063.0, + 1406.0, + 1063.0, + 1406.0, + 1100.0, + 294.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1090.0, + 1408.0, + 1090.0, + 1408.0, + 1131.0, + 293.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1121.0, + 1406.0, + 1121.0, + 1406.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 298.0, + 1152.0, + 298.0, + 1188.0, + 295.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1152.0, + 1404.0, + 1152.0, + 1404.0, + 1188.0, + 389.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1183.0, + 1408.0, + 1183.0, + 1408.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1617.0, + 1408.0, + 1617.0, + 1408.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1646.0, + 1406.0, + 1646.0, + 1406.0, + 1681.0, + 293.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1674.0, + 1405.0, + 1674.0, + 1405.0, + 1711.0, + 293.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1703.0, + 1407.0, + 1703.0, + 1407.0, + 1742.0, + 292.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1731.0, + 1408.0, + 1731.0, + 1408.0, + 1771.0, + 293.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 399.0, + 1406.0, + 399.0, + 1406.0, + 435.0, + 296.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 430.0, + 1405.0, + 430.0, + 1405.0, + 462.0, + 296.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 457.0, + 1405.0, + 457.0, + 1405.0, + 493.0, + 294.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 485.0, + 1406.0, + 485.0, + 1406.0, + 522.0, + 292.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 516.0, + 1409.0, + 516.0, + 1409.0, + 552.0, + 294.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1225.0, + 1403.0, + 1225.0, + 1403.0, + 1259.0, + 295.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1407.0, + 1252.0, + 1407.0, + 1292.0, + 293.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1314.0, + 509.0, + 1314.0, + 509.0, + 1349.0, + 294.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1427.0, + 1403.0, + 1427.0, + 1403.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1454.0, + 1407.0, + 1454.0, + 1407.0, + 1497.0, + 292.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1484.0, + 1404.0, + 1484.0, + 1404.0, + 1521.0, + 293.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1516.0, + 696.0, + 1516.0, + 696.0, + 1550.0, + 296.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1778.0, + 1404.0, + 1778.0, + 1404.0, + 1817.0, + 293.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1806.0, + 1405.0, + 1806.0, + 1405.0, + 1846.0, + 293.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1835.0, + 1408.0, + 1835.0, + 1408.0, + 1877.0, + 291.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1866.0, + 1406.0, + 1866.0, + 1406.0, + 1899.0, + 292.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1897.0, + 1408.0, + 1897.0, + 1408.0, + 1929.0, + 296.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 257.0, + 1190.0, + 257.0, + 1190.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 331, + 210, + 1367, + 210, + 1367, + 659, + 331, + 659 + ], + "score": 0.982, + "html": "
DMN+L1-step CBA1-step Soft D-NTM1-step Discrete D-NTM3-steps LBA* NTM3-steps CBA NTM3-steps Soft D-NTM3-steps Discrete D-NTM
Task 1LSTM 0.00MemN2N 0.000.00NTM 16.30NTM 16.885.416.660.000.000.000.00
281.900.300.3057.0855.7058.5456.0461.6759.3846.6662.29
383.102.101.1074.1655.0074.5872.0883.5465.2147.0841.45
40.200.000.000.000.000.000.000.000.000.000.00
51.200.800.501.4620.411.661.040.831.461.251.45
651.800.100.0023.3321.0440.2044.7948.1354.8020.6211.04
724.902.002.4021.6721.6719.1619.587.9237.707.295.62
834.100.900.0025.7621.0512.5818.4625.388.8211.020.74
920.200.300.0024.7924.1736.6634.3737.800.0039.3732.50
1030.100.000.0041.4633.1352.2950.8356.2523.7520.0020.83
1110.300.100.0018.9631.8831.454.163.960.2830.6216.87
1223.400.000.0025.8330.007.706.6628.7523.755.414.58
136.100.000.006.675.635.622.295.8383.137.915.00
1481.000.100.2058.5459.1760.0063.7561.8857.7158.1260.20
1578.700.000.0036.4642.3036.8739.2735.6221.8836.0440.26
1651.9051.8045.3071.1571.1549.1651.3546.1550.0046.0445.41
1750.1018.604.2043.7543.7517.9116.0443.7556.2521.259.16
186.805.302.103.9647.503.953.5447.5047.506.871.66
1990.302.300.0075.8971.5173.7464.6361.5663.6575.8876.66
202.100.000.001.250.002.703.120.400.003.330.00
Avg.Err.36.414.242.8131.4233.6029.5127.9332.8532.7624.2421.79
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1082, + 1404, + 1082, + 1404, + 1288, + 297, + 1288 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 663, + 1421, + 1035, + 1421, + 1035, + 1870, + 663, + 1870 + ], + "score": 0.96, + "html": "
TaskFF Soft D-NTMFF Discrete D-NTMFF Discrete* D-NTM
14.3881.6714.79
227.576.6776.67
371.2579.3870.83
40.0078.6544.06
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.62
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.92
1939.4771.4671.56
200.076.569.79
Avg.Err.12.8168.3037.79
" + }, + { + "category_id": 1, + "poly": [ + 299, + 978, + 1400, + 978, + 1400, + 1069, + 299, + 1069 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 298, + 903, + 1401, + 903, + 1401, + 965, + 298, + 965 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 292, + 2001, + 1392, + 2001, + 1392, + 2035, + 292, + 2035 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 299, + 1301, + 1399, + 1301, + 1399, + 1392, + 299, + 1392 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.81 + }, + { + "category_id": 1, + "poly": [ + 297, + 681, + 1404, + 681, + 1404, + 830, + 297, + 830 + ], + "score": 0.795 + }, + { + "category_id": 6, + "poly": [ + 294, + 1893, + 1398, + 1893, + 1398, + 1954, + 294, + 1954 + ], + "score": 0.561 + }, + { + "category_id": 1, + "poly": [ + 294, + 1893, + 1398, + 1893, + 1398, + 1954, + 294, + 1954 + ], + "score": 0.297 + }, + { + "category_id": 6, + "poly": [ + 297, + 681, + 1404, + 681, + 1404, + 830, + 297, + 830 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 569, + 683, + 613, + 683, + 613, + 712, + 569, + 712 + ], + "score": 0.8, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 569, + 1895, + 613, + 1895, + 613, + 1924, + 569, + 1924 + ], + "score": 0.78, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 617, + 740, + 688, + 740, + 688, + 769, + 617, + 769 + ], + "score": 0.67, + "latex": "\\mathrm { L B A ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 343, + 1331, + 425, + 1331, + 425, + 1361, + 343, + 1361 + ], + "score": 0.58, + "latex": "\\mathrm { D M N + }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 568.0, + 1888.0, + 568.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1888.0, + 1403.0, + 1888.0, + 1403.0, + 1929.0, + 614.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1920.0, + 560.0, + 1920.0, + 560.0, + 1955.0, + 294.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 682.0, + 568.0, + 682.0, + 568.0, + 714.0, + 296.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 682.0, + 1402.0, + 682.0, + 1402.0, + 714.0, + 614.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 709.0, + 1404.0, + 709.0, + 1404.0, + 743.0, + 295.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 740.0, + 616.0, + 740.0, + 616.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 772.0, + 689.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 768.0, + 1406.0, + 768.0, + 1406.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1330.0, + 798.0, + 1330.0, + 831.0, + 295.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1083.0, + 1404.0, + 1083.0, + 1404.0, + 1113.0, + 296.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1143.0, + 295.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1142.0, + 1405.0, + 1142.0, + 1405.0, + 1172.0, + 296.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1171.0, + 1404.0, + 1171.0, + 1404.0, + 1201.0, + 296.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1197.0, + 1407.0, + 1197.0, + 1407.0, + 1233.0, + 292.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 1406.0, + 1228.0, + 1406.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1254.0, + 1081.0, + 1254.0, + 1081.0, + 1291.0, + 294.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 977.0, + 1404.0, + 977.0, + 1404.0, + 1013.0, + 293.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1007.0, + 1406.0, + 1007.0, + 1406.0, + 1044.0, + 293.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1038.0, + 1286.0, + 1038.0, + 1286.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 901.0, + 1406.0, + 901.0, + 1406.0, + 939.0, + 293.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 931.0, + 1406.0, + 931.0, + 1406.0, + 969.0, + 294.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1393.0, + 2000.0, + 1393.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1301.0, + 1404.0, + 1301.0, + 1404.0, + 1335.0, + 295.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1327.0, + 342.0, + 1327.0, + 342.0, + 1367.0, + 293.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1327.0, + 1405.0, + 1327.0, + 1405.0, + 1367.0, + 426.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1361.0, + 754.0, + 1361.0, + 754.0, + 1394.0, + 296.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 682.0, + 568.0, + 682.0, + 568.0, + 714.0, + 296.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 682.0, + 1402.0, + 682.0, + 1402.0, + 714.0, + 614.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 709.0, + 1404.0, + 709.0, + 1404.0, + 743.0, + 295.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 740.0, + 616.0, + 740.0, + 616.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 772.0, + 689.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 768.0, + 1406.0, + 768.0, + 1406.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1330.0, + 798.0, + 1330.0, + 831.0, + 295.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 568.0, + 1888.0, + 568.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1888.0, + 1403.0, + 1888.0, + 1403.0, + 1929.0, + 614.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1920.0, + 560.0, + 1920.0, + 560.0, + 1955.0, + 294.0, + 1955.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1525, + 1404, + 1525, + 1404, + 1790, + 297, + 1790 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 287, + 1405, + 287, + 1405, + 554, + 298, + 554 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 306, + 1021, + 842, + 1021, + 842, + 1333, + 306, + 1333 + ], + "score": 0.978, + "html": "
Test Acc
D-NTMdiscrete MAB89.6
D-NTMdiscrete IB92.3
Soft D-NTM93.4
NTM90.9
I-RNN (Le et al., 2015)82.0
Zoneout (Krueger et al.,2016)93.1
LSTM (Krueger et al., 2016)89.8
Unitary-RNN (Arjovsky et al., 2015)91.4
Recurrent Dropout (Krueger et al., 2016)92.5
" + }, + { + "category_id": 1, + "poly": [ + 298, + 651, + 1403, + 651, + 1403, + 768, + 298, + 768 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 783, + 1404, + 783, + 1404, + 988, + 297, + 988 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1804, + 1404, + 1804, + 1404, + 1922, + 298, + 1922 + ], + "score": 0.971 + }, + { + "category_id": 5, + "poly": [ + 852, + 1106, + 1392, + 1106, + 1392, + 1248, + 852, + 1248 + ], + "score": 0.971, + "html": "
Copy TasksAssociative Recall
Soft D-NTMSuccessSuccess
D-NTMdiscreteSuccessFailure
NTMSuccessSuccess
" + }, + { + "category_id": 2, + "poly": [ + 297, + 1953, + 1402, + 1953, + 1402, + 2032, + 297, + 2032 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 298, + 1456, + 842, + 1456, + 842, + 1491, + 298, + 1491 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 300, + 592, + 576, + 592, + 576, + 624, + 300, + 624 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 301, + 229, + 634, + 229, + 634, + 262, + 301, + 262 + ], + "score": 0.896 + }, + { + "category_id": 6, + "poly": [ + 411, + 1356, + 739, + 1356, + 739, + 1387, + 411, + 1387 + ], + "score": 0.886 + }, + { + "category_id": 6, + "poly": [ + 932, + 1270, + 1219, + 1270, + 1219, + 1301, + 932, + 1301 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.8 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 856, + 75, + 856, + 104, + 299, + 104 + ], + "score": 0.663 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.443 + }, + { + "category_id": 13, + "poly": [ + 469, + 1980, + 535, + 1980, + 535, + 2006, + 469, + 2006 + ], + "score": 0.86, + "latex": "9 5 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1328, + 434, + 1356, + 434, + 1356, + 464, + 1328, + 464 + ], + "score": 0.6, + "latex": "3 ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 303, + 404, + 411, + 404, + 411, + 435, + 303, + 435 + ], + "score": 0.47, + "latex": "( p \\mathrm { M N I S T } )" + }, + { + "category_id": 13, + "poly": [ + 630, + 1356, + 735, + 1356, + 735, + 1388, + 630, + 1388 + ], + "score": 0.39, + "latex": "p { \\mathrm { M N I S T } } ." + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1945.0, + 1407.0, + 1945.0, + 1407.0, + 1987.0, + 327.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1976.0, + 468.0, + 1976.0, + 468.0, + 2012.0, + 292.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1976.0, + 1403.0, + 1976.0, + 1403.0, + 2012.0, + 536.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 698.0, + 2004.0, + 698.0, + 2034.0, + 295.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1451.0, + 847.0, + 1451.0, + 847.0, + 1498.0, + 291.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 586.0, + 580.0, + 586.0, + 580.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 640.0, + 227.0, + 640.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1353.0, + 629.0, + 1353.0, + 629.0, + 1392.0, + 407.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1353.0, + 741.0, + 1353.0, + 741.0, + 1392.0, + 736.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1265.0, + 1222.0, + 1265.0, + 1222.0, + 1306.0, + 929.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1526.0, + 1406.0, + 1526.0, + 1406.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1551.0, + 1409.0, + 1551.0, + 1409.0, + 1594.0, + 291.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1587.0, + 1405.0, + 1587.0, + 1405.0, + 1618.0, + 296.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1613.0, + 1405.0, + 1613.0, + 1405.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1678.0, + 294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 1406.0, + 1671.0, + 1406.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1697.0, + 1404.0, + 1697.0, + 1404.0, + 1737.0, + 292.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1731.0, + 1405.0, + 1731.0, + 1405.0, + 1763.0, + 295.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1756.0, + 1126.0, + 1756.0, + 1126.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 287.0, + 1408.0, + 287.0, + 1408.0, + 322.0, + 293.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1405.0, + 317.0, + 1405.0, + 349.0, + 295.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 345.0, + 1406.0, + 345.0, + 1406.0, + 381.0, + 293.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 373.0, + 1406.0, + 373.0, + 1406.0, + 409.0, + 293.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 403.0, + 302.0, + 403.0, + 302.0, + 438.0, + 293.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 403.0, + 1406.0, + 403.0, + 1406.0, + 438.0, + 412.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 433.0, + 1327.0, + 433.0, + 1327.0, + 470.0, + 292.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 433.0, + 1406.0, + 433.0, + 1406.0, + 470.0, + 1357.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 464.0, + 1405.0, + 464.0, + 1405.0, + 500.0, + 295.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 491.0, + 1407.0, + 491.0, + 1407.0, + 529.0, + 293.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 522.0, + 1408.0, + 522.0, + 1408.0, + 558.0, + 295.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 1405.0, + 650.0, + 1405.0, + 688.0, + 294.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 678.0, + 1404.0, + 678.0, + 1404.0, + 715.0, + 294.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 705.0, + 1405.0, + 705.0, + 1405.0, + 745.0, + 292.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 735.0, + 1035.0, + 735.0, + 1035.0, + 772.0, + 291.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 782.0, + 1407.0, + 782.0, + 1407.0, + 819.0, + 292.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 844.0, + 1404.0, + 844.0, + 1404.0, + 874.0, + 296.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 872.0, + 1405.0, + 872.0, + 1405.0, + 902.0, + 295.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 897.0, + 1406.0, + 897.0, + 1406.0, + 932.0, + 294.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 926.0, + 1407.0, + 926.0, + 1407.0, + 962.0, + 291.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 955.0, + 1407.0, + 955.0, + 1407.0, + 990.0, + 294.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1800.0, + 1405.0, + 1800.0, + 1405.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1833.0, + 1405.0, + 1833.0, + 1405.0, + 1866.0, + 293.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1862.0, + 1405.0, + 1862.0, + 1405.0, + 1896.0, + 296.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1889.0, + 1409.0, + 1889.0, + 1409.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 408, + 1405, + 408, + 1405, + 556, + 299, + 556 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 299, + 305, + 1403, + 305, + 1403, + 393, + 299, + 393 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 294, + 230, + 1401, + 230, + 1401, + 290, + 294, + 290 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 300, + 599, + 488, + 599, + 488, + 631, + 300, + 631 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 292, + 642, + 1410, + 642, + 1410, + 2035, + 292, + 2035 + ], + "score": 0.718 + }, + { + "category_id": 13, + "poly": [ + 337, + 306, + 440, + 306, + 440, + 337, + 337, + 337 + ], + "score": 0.38, + "latex": "p \\mathrm { M N I S T }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 599.0, + 490.0, + 599.0, + 490.0, + 635.0, + 296.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 406.0, + 1405.0, + 406.0, + 1405.0, + 443.0, + 296.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 437.0, + 1409.0, + 437.0, + 1409.0, + 473.0, + 294.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 465.0, + 1405.0, + 465.0, + 1405.0, + 501.0, + 294.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 496.0, + 1406.0, + 496.0, + 1406.0, + 528.0, + 295.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 526.0, + 1409.0, + 526.0, + 1409.0, + 558.0, + 297.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 304.0, + 336.0, + 304.0, + 336.0, + 341.0, + 294.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 304.0, + 1405.0, + 304.0, + 1405.0, + 341.0, + 441.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 333.0, + 1405.0, + 333.0, + 1405.0, + 368.0, + 293.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 363.0, + 1287.0, + 363.0, + 1287.0, + 397.0, + 293.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 259.0, + 854.0, + 259.0, + 854.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 1406.0, + 650.0, + 1406.0, + 684.0, + 294.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 676.0, + 1408.0, + 676.0, + 1408.0, + 716.0, + 320.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 707.0, + 1351.0, + 707.0, + 1351.0, + 741.0, + 323.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 751.0, + 1406.0, + 751.0, + 1406.0, + 796.0, + 291.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 785.0, + 708.0, + 785.0, + 708.0, + 819.0, + 316.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 830.0, + 1405.0, + 830.0, + 1405.0, + 872.0, + 291.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 864.0, + 1406.0, + 864.0, + 1406.0, + 898.0, + 321.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 891.0, + 658.0, + 891.0, + 658.0, + 929.0, + 320.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 937.0, + 1406.0, + 937.0, + 1406.0, + 978.0, + 293.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 968.0, + 1240.0, + 968.0, + 1240.0, + 1006.0, + 318.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1406.0, + 1017.0, + 1406.0, + 1055.0, + 294.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1046.0, + 1170.0, + 1046.0, + 1170.0, + 1085.0, + 320.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1097.0, + 1406.0, + 1097.0, + 1406.0, + 1132.0, + 296.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1125.0, + 1411.0, + 1125.0, + 1411.0, + 1164.0, + 321.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1155.0, + 762.0, + 1155.0, + 762.0, + 1189.0, + 321.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1202.0, + 1411.0, + 1202.0, + 1411.0, + 1242.0, + 291.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1234.0, + 1309.0, + 1234.0, + 1309.0, + 1269.0, + 323.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1279.0, + 1409.0, + 1279.0, + 1409.0, + 1320.0, + 293.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1312.0, + 776.0, + 1312.0, + 776.0, + 1346.0, + 320.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1357.0, + 1408.0, + 1357.0, + 1408.0, + 1398.0, + 291.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1387.0, + 1411.0, + 1387.0, + 1411.0, + 1429.0, + 320.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1419.0, + 664.0, + 1419.0, + 664.0, + 1454.0, + 323.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1469.0, + 1404.0, + 1469.0, + 1404.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1496.0, + 1062.0, + 1496.0, + 1062.0, + 1535.0, + 320.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1542.0, + 1409.0, + 1542.0, + 1409.0, + 1583.0, + 291.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1573.0, + 399.0, + 1573.0, + 399.0, + 1611.0, + 320.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1625.0, + 1408.0, + 1625.0, + 1408.0, + 1659.0, + 294.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1654.0, + 1409.0, + 1654.0, + 1409.0, + 1688.0, + 323.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1681.0, + 1409.0, + 1681.0, + 1409.0, + 1720.0, + 321.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1710.0, + 504.0, + 1710.0, + 504.0, + 1744.0, + 321.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1758.0, + 1406.0, + 1758.0, + 1406.0, + 1796.0, + 294.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1786.0, + 1409.0, + 1786.0, + 1409.0, + 1828.0, + 318.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1814.0, + 531.0, + 1814.0, + 531.0, + 1852.0, + 320.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1864.0, + 1408.0, + 1864.0, + 1408.0, + 1905.0, + 291.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1895.0, + 776.0, + 1895.0, + 776.0, + 1933.0, + 320.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1947.0, + 1406.0, + 1947.0, + 1406.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2012.0, + 321.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 611.0, + 2003.0, + 611.0, + 2035.0, + 321.0, + 2035.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 862, + 2088, + 862, + 2113, + 835, + 2113 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 296, + 520, + 1402, + 520, + 1402, + 584, + 296, + 584 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 295, + 600, + 1401, + 600, + 1401, + 662, + 295, + 662 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 297, + 1653, + 1404, + 1653, + 1404, + 1716, + 297, + 1716 + ], + "score": 0.673 + }, + { + "category_id": 1, + "poly": [ + 297, + 1165, + 1403, + 1165, + 1403, + 1257, + 297, + 1257 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 289, + 1085, + 1403, + 1085, + 1403, + 1149, + 289, + 1149 + ], + "score": 0.653 + }, + { + "category_id": 1, + "poly": [ + 297, + 1354, + 1401, + 1354, + 1401, + 1448, + 297, + 1448 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 293, + 1274, + 1405, + 1274, + 1405, + 1339, + 293, + 1339 + ], + "score": 0.645 + }, + { + "category_id": 1, + "poly": [ + 295, + 389, + 1404, + 389, + 1404, + 452, + 295, + 452 + ], + "score": 0.645 + }, + { + "category_id": 1, + "poly": [ + 297, + 788, + 1400, + 788, + 1400, + 852, + 297, + 852 + ], + "score": 0.637 + }, + { + "category_id": 1, + "poly": [ + 298, + 977, + 1402, + 977, + 1402, + 1069, + 298, + 1069 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 294, + 1842, + 1403, + 1842, + 1403, + 1876, + 294, + 1876 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 298, + 679, + 1404, + 679, + 1404, + 773, + 298, + 773 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1403, + 229, + 1403, + 293, + 293, + 293 + ], + "score": 0.625 + }, + { + "category_id": 1, + "poly": [ + 296, + 1732, + 1406, + 1732, + 1406, + 1824, + 296, + 1824 + ], + "score": 0.623 + }, + { + "category_id": 1, + "poly": [ + 297, + 1543, + 1405, + 1543, + 1405, + 1636, + 297, + 1636 + ], + "score": 0.619 + }, + { + "category_id": 1, + "poly": [ + 299, + 868, + 1402, + 868, + 1402, + 960, + 299, + 960 + ], + "score": 0.619 + }, + { + "category_id": 1, + "poly": [ + 290, + 1463, + 1402, + 1463, + 1402, + 1528, + 290, + 1528 + ], + "score": 0.605 + }, + { + "category_id": 1, + "poly": [ + 295, + 1893, + 1404, + 1893, + 1404, + 1956, + 295, + 1956 + ], + "score": 0.592 + }, + { + "category_id": 1, + "poly": [ + 297, + 309, + 1401, + 309, + 1401, + 372, + 297, + 372 + ], + "score": 0.59 + }, + { + "category_id": 1, + "poly": [ + 296, + 1972, + 1400, + 1972, + 1400, + 2035, + 296, + 2035 + ], + "score": 0.584 + }, + { + "category_id": 1, + "poly": [ + 295, + 469, + 1403, + 469, + 1403, + 503, + 295, + 503 + ], + "score": 0.579 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 519.0, + 1404.0, + 519.0, + 1404.0, + 556.0, + 295.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 549.0, + 1233.0, + 549.0, + 1233.0, + 584.0, + 321.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 599.0, + 1407.0, + 599.0, + 1407.0, + 635.0, + 294.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 627.0, + 398.0, + 627.0, + 398.0, + 665.0, + 319.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1650.0, + 1407.0, + 1650.0, + 1407.0, + 1691.0, + 295.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1683.0, + 775.0, + 1683.0, + 775.0, + 1716.0, + 324.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1162.0, + 1409.0, + 1162.0, + 1409.0, + 1204.0, + 292.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1195.0, + 1410.0, + 1195.0, + 1410.0, + 1231.0, + 321.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1224.0, + 638.0, + 1224.0, + 638.0, + 1258.0, + 324.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1085.0, + 1405.0, + 1085.0, + 1405.0, + 1121.0, + 296.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1117.0, + 694.0, + 1117.0, + 694.0, + 1148.0, + 320.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1356.0, + 1405.0, + 1356.0, + 1405.0, + 1390.0, + 296.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1383.0, + 1406.0, + 1383.0, + 1406.0, + 1423.0, + 320.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1413.0, + 1318.0, + 1413.0, + 1318.0, + 1449.0, + 322.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1274.0, + 1407.0, + 1274.0, + 1407.0, + 1310.0, + 295.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1301.0, + 400.0, + 1301.0, + 400.0, + 1340.0, + 319.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 387.0, + 1407.0, + 387.0, + 1407.0, + 427.0, + 294.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 418.0, + 532.0, + 418.0, + 532.0, + 451.0, + 319.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 789.0, + 1404.0, + 789.0, + 1404.0, + 825.0, + 297.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 817.0, + 1045.0, + 817.0, + 1045.0, + 852.0, + 321.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 977.0, + 1406.0, + 977.0, + 1406.0, + 1011.0, + 296.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1007.0, + 1407.0, + 1007.0, + 1407.0, + 1040.0, + 323.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1036.0, + 856.0, + 1036.0, + 856.0, + 1071.0, + 323.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1841.0, + 1408.0, + 1841.0, + 1408.0, + 1880.0, + 296.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 677.0, + 1405.0, + 677.0, + 1405.0, + 719.0, + 293.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 708.0, + 1406.0, + 708.0, + 1406.0, + 748.0, + 322.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 738.0, + 1320.0, + 738.0, + 1320.0, + 774.0, + 322.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 225.0, + 1405.0, + 225.0, + 1405.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 258.0, + 1396.0, + 258.0, + 1396.0, + 294.0, + 324.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1732.0, + 1408.0, + 1732.0, + 1408.0, + 1768.0, + 295.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1762.0, + 1408.0, + 1762.0, + 1408.0, + 1798.0, + 322.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1790.0, + 1117.0, + 1790.0, + 1117.0, + 1827.0, + 322.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1542.0, + 1405.0, + 1542.0, + 1405.0, + 1581.0, + 292.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1573.0, + 1405.0, + 1573.0, + 1405.0, + 1611.0, + 320.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1603.0, + 1126.0, + 1603.0, + 1126.0, + 1638.0, + 321.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 870.0, + 1404.0, + 870.0, + 1404.0, + 903.0, + 295.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 900.0, + 1406.0, + 900.0, + 1406.0, + 933.0, + 323.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 928.0, + 825.0, + 928.0, + 825.0, + 962.0, + 323.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1463.0, + 1405.0, + 1463.0, + 1405.0, + 1501.0, + 296.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1494.0, + 1354.0, + 1494.0, + 1354.0, + 1530.0, + 322.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1894.0, + 1402.0, + 1894.0, + 1402.0, + 1930.0, + 295.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1922.0, + 1256.0, + 1922.0, + 1256.0, + 1958.0, + 322.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 307.0, + 1406.0, + 307.0, + 1406.0, + 347.0, + 295.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 338.0, + 602.0, + 338.0, + 602.0, + 374.0, + 321.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2002.0, + 1272.0, + 2002.0, + 1272.0, + 2038.0, + 323.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 468.0, + 1405.0, + 468.0, + 1405.0, + 503.0, + 297.0, + 503.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1407, + 229, + 1407, + 751, + 294, + 751 + ], + "score": 0.634 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 262.0, + 852.0, + 262.0, + 852.0, + 294.0, + 324.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 310.0, + 1406.0, + 310.0, + 1406.0, + 350.0, + 293.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 342.0, + 896.0, + 342.0, + 896.0, + 375.0, + 321.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 388.0, + 1406.0, + 388.0, + 1406.0, + 429.0, + 293.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 417.0, + 1405.0, + 417.0, + 1405.0, + 458.0, + 321.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 449.0, + 1233.0, + 449.0, + 1233.0, + 485.0, + 322.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 498.0, + 1405.0, + 498.0, + 1405.0, + 537.0, + 293.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 528.0, + 1406.0, + 528.0, + 1406.0, + 568.0, + 321.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 557.0, + 863.0, + 557.0, + 863.0, + 594.0, + 320.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 607.0, + 1408.0, + 607.0, + 1408.0, + 644.0, + 295.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 639.0, + 584.0, + 639.0, + 584.0, + 671.0, + 324.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 687.0, + 1401.0, + 687.0, + 1401.0, + 725.0, + 295.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 718.0, + 959.0, + 718.0, + 959.0, + 754.0, + 322.0, + 754.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 637, + 1403, + 637, + 1403, + 784, + 299, + 784 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1460, + 1404, + 1460, + 1404, + 1608, + 298, + 1608 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1012, + 1403, + 1012, + 1403, + 1160, + 299, + 1160 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1251, + 1404, + 1251, + 1404, + 1368, + 298, + 1368 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1916, + 1402, + 1916, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 464, + 1404, + 464, + 1404, + 553, + 299, + 553 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1716, + 1403, + 1716, + 1403, + 1807, + 298, + 1807 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 868, + 1400, + 868, + 1400, + 929, + 300, + 929 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 349, + 1060, + 349, + 1060, + 381, + 298, + 381 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 298, + 1848, + 1152, + 1848, + 1152, + 1883, + 298, + 1883 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 299, + 1404, + 960, + 1404, + 960, + 1435, + 299, + 1435 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 815, + 508, + 815, + 508, + 846, + 300, + 846 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 300, + 960, + 624, + 960, + 624, + 992, + 300, + 992 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 299, + 1195, + 1077, + 1195, + 1077, + 1226, + 299, + 1226 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 301, + 585, + 560, + 585, + 560, + 617, + 301, + 617 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.858 + }, + { + "category_id": 0, + "poly": [ + 301, + 411, + 678, + 411, + 678, + 443, + 301, + 443 + ], + "score": 0.844 + }, + { + "category_id": 0, + "poly": [ + 299, + 1649, + 959, + 1649, + 959, + 1684, + 299, + 1684 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.825 + }, + { + "category_id": 0, + "poly": [ + 302, + 226, + 718, + 226, + 718, + 262, + 302, + 262 + ], + "score": 0.758 + }, + { + "category_id": 1, + "poly": [ + 299, + 293, + 889, + 293, + 889, + 326, + 299, + 326 + ], + "score": 0.51 + }, + { + "category_id": 0, + "poly": [ + 299, + 293, + 889, + 293, + 889, + 326, + 299, + 326 + ], + "score": 0.502 + }, + { + "category_id": 1, + "poly": [ + 302, + 226, + 718, + 226, + 718, + 262, + 302, + 262 + ], + "score": 0.161 + }, + { + "category_id": 13, + "poly": [ + 559, + 1012, + 614, + 1012, + 614, + 1042, + 559, + 1042 + ], + "score": 0.88, + "latex": "1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 847, + 1520, + 925, + 1520, + 925, + 1548, + 847, + 1548 + ], + "score": 0.84, + "latex": "1 e - 3" + }, + { + "category_id": 13, + "poly": [ + 613, + 1310, + 692, + 1310, + 692, + 1338, + 613, + 1338 + ], + "score": 0.82, + "latex": "1 e - 3" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1844.0, + 1158.0, + 1844.0, + 1158.0, + 1890.0, + 295.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1403.0, + 962.0, + 1403.0, + 962.0, + 1439.0, + 297.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 811.0, + 512.0, + 811.0, + 512.0, + 853.0, + 295.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 962.0, + 628.0, + 962.0, + 628.0, + 995.0, + 297.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1193.0, + 1080.0, + 1193.0, + 1080.0, + 1231.0, + 295.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 581.0, + 564.0, + 581.0, + 564.0, + 623.0, + 294.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 411.0, + 681.0, + 411.0, + 681.0, + 447.0, + 297.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1645.0, + 964.0, + 1645.0, + 964.0, + 1691.0, + 292.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 867.0, + 2084.0, + 867.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 721.0, + 224.0, + 721.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 890.0, + 293.0, + 890.0, + 328.0, + 296.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 637.0, + 1405.0, + 637.0, + 1405.0, + 671.0, + 295.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 659.0, + 1407.0, + 659.0, + 1407.0, + 705.0, + 292.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 694.0, + 1405.0, + 694.0, + 1405.0, + 730.0, + 292.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 723.0, + 1405.0, + 723.0, + 1405.0, + 759.0, + 294.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 754.0, + 524.0, + 754.0, + 524.0, + 786.0, + 295.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1460.0, + 1406.0, + 1460.0, + 1406.0, + 1493.0, + 296.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1486.0, + 1406.0, + 1486.0, + 1406.0, + 1525.0, + 293.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1517.0, + 846.0, + 1517.0, + 846.0, + 1553.0, + 293.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1517.0, + 1410.0, + 1517.0, + 1410.0, + 1553.0, + 926.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1571.0, + 1317.0, + 1571.0, + 1317.0, + 1614.0, + 293.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 558.0, + 1011.0, + 558.0, + 1046.0, + 294.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1011.0, + 1405.0, + 1011.0, + 1405.0, + 1046.0, + 615.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1037.0, + 1407.0, + 1037.0, + 1407.0, + 1080.0, + 290.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1069.0, + 1404.0, + 1069.0, + 1404.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1099.0, + 1405.0, + 1099.0, + 1405.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1129.0, + 748.0, + 1129.0, + 748.0, + 1165.0, + 295.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1288.0, + 296.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1278.0, + 1404.0, + 1278.0, + 1404.0, + 1316.0, + 294.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1308.0, + 612.0, + 1308.0, + 612.0, + 1342.0, + 294.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1308.0, + 1404.0, + 1308.0, + 1404.0, + 1342.0, + 693.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1338.0, + 999.0, + 1338.0, + 999.0, + 1372.0, + 293.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1914.0, + 1406.0, + 1914.0, + 1406.0, + 1950.0, + 294.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1406.0, + 1944.0, + 1406.0, + 1982.0, + 294.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1997.0, + 1095.0, + 1997.0, + 1095.0, + 2043.0, + 291.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1404.0, + 461.0, + 1404.0, + 499.0, + 294.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 491.0, + 1404.0, + 491.0, + 1404.0, + 527.0, + 293.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 523.0, + 775.0, + 523.0, + 775.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1713.0, + 1403.0, + 1713.0, + 1403.0, + 1753.0, + 293.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1781.0, + 294.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1776.0, + 957.0, + 1776.0, + 957.0, + 1812.0, + 293.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 867.0, + 1404.0, + 867.0, + 1404.0, + 902.0, + 296.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 897.0, + 719.0, + 897.0, + 719.0, + 932.0, + 294.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 349.0, + 1062.0, + 349.0, + 1062.0, + 385.0, + 296.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 890.0, + 293.0, + 890.0, + 328.0, + 296.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 721.0, + 224.0, + 721.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1705, + 1406, + 1705, + 1406, + 1796, + 297, + 1796 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 479, + 882, + 1208, + 882, + 1208, + 1408, + 479, + 1408 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 300, + 1944, + 1402, + 1944, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 384, + 218, + 1185, + 218, + 1185, + 592, + 384, + 592 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 296, + 613, + 1406, + 613, + 1406, + 792, + 296, + 792 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 296, + 1453, + 1404, + 1453, + 1404, + 1544, + 296, + 1544 + ], + "score": 0.944 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 0, + "poly": [ + 308, + 1600, + 1399, + 1600, + 1399, + 1673, + 308, + 1673 + ], + "score": 0.766 + }, + { + "category_id": 0, + "poly": [ + 299, + 1840, + 1399, + 1840, + 1399, + 1913, + 299, + 1913 + ], + "score": 0.76 + }, + { + "category_id": 1, + "poly": [ + 308, + 1600, + 1399, + 1600, + 1399, + 1673, + 308, + 1673 + ], + "score": 0.216 + }, + { + "category_id": 13, + "poly": [ + 722, + 1737, + 824, + 1737, + 824, + 1768, + 722, + 1768 + ], + "score": 0.47, + "latex": "p \\mathrm { M N I S T }" + }, + { + "category_id": 13, + "poly": [ + 1273, + 1643, + 1293, + 1643, + 1293, + 1675, + 1273, + 1675 + ], + "score": 0.28, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 443, + 649, + 461, + 649, + 461, + 671, + 443, + 671 + ], + "score": 0.25, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 899.0, + 520.0, + 899.0, + 520.0, + 930.0, + 481.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 902.0, + 1187.0, + 902.0, + 1187.0, + 940.0, + 820.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 938.0, + 1178.0, + 938.0, + 1178.0, + 981.0, + 816.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 975.0, + 520.0, + 975.0, + 520.0, + 1005.0, + 481.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1051.0, + 520.0, + 1051.0, + 520.0, + 1081.0, + 481.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1128.0, + 520.0, + 1128.0, + 520.0, + 1155.0, + 484.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1205.0, + 519.0, + 1205.0, + 519.0, + 1230.0, + 485.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1280.0, + 519.0, + 1280.0, + 519.0, + 1305.0, + 483.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1355.0, + 519.0, + 1355.0, + 519.0, + 1381.0, + 483.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 1377.0, + 529.0, + 1377.0, + 529.0, + 1395.0, + 513.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1374.0, + 641.0, + 1374.0, + 641.0, + 1397.0, + 613.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1371.0, + 755.0, + 1371.0, + 755.0, + 1400.0, + 715.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1371.0, + 861.0, + 1371.0, + 861.0, + 1400.0, + 822.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1371.0, + 967.0, + 1371.0, + 967.0, + 1400.0, + 927.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1372.0, + 1075.0, + 1372.0, + 1075.0, + 1400.0, + 1033.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1371.0, + 1180.0, + 1371.0, + 1180.0, + 1400.0, + 1139.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.25, + 1091.0, + 651.25, + 1091.0, + 651.25, + 1109.0, + 629.25, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1148.5, + 772.0, + 1148.5, + 772.0, + 1172.5, + 720.0, + 1172.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 231.0, + 768.0, + 231.0, + 768.0, + 247.0, + 752.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 230.0, + 819.0, + 230.0, + 819.0, + 248.0, + 790.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 228.0, + 945.0, + 228.0, + 945.0, + 250.0, + 830.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 229.0, + 1032.0, + 229.0, + 1032.0, + 249.0, + 958.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 231.0, + 1163.0, + 231.0, + 1163.0, + 246.0, + 1143.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 256.0, + 600.0, + 256.0, + 600.0, + 275.0, + 488.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 296.0, + 603.0, + 296.0, + 603.0, + 321.0, + 491.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 343.0, + 600.0, + 343.0, + 600.0, + 362.0, + 407.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 384.0, + 601.0, + 384.0, + 601.0, + 406.0, + 395.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 405.0, + 723.0, + 405.0, + 723.0, + 431.0, + 673.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 403.0, + 1122.0, + 403.0, + 1122.0, + 430.0, + 1071.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 429.0, + 602.0, + 429.0, + 602.0, + 447.0, + 433.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 469.0, + 605.0, + 469.0, + 605.0, + 491.0, + 513.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 514.0, + 600.0, + 514.0, + 600.0, + 532.0, + 396.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 555.0, + 603.0, + 555.0, + 603.0, + 577.0, + 383.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 615.0, + 1403.0, + 615.0, + 1403.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 645.0, + 442.0, + 645.0, + 442.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 645.0, + 1403.0, + 645.0, + 1403.0, + 676.0, + 462.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1406.0, + 673.0, + 1406.0, + 707.0, + 294.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 700.0, + 1404.0, + 700.0, + 1404.0, + 736.0, + 293.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 730.0, + 1406.0, + 730.0, + 1406.0, + 766.0, + 293.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 757.0, + 1215.0, + 757.0, + 1215.0, + 796.0, + 293.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1454.0, + 1405.0, + 1454.0, + 1405.0, + 1488.0, + 295.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1484.0, + 1405.0, + 1484.0, + 1405.0, + 1514.0, + 295.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1513.0, + 942.0, + 1513.0, + 942.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1597.0, + 963.0, + 1597.0, + 963.0, + 1639.0, + 299.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1632.0, + 1272.0, + 1632.0, + 1272.0, + 1678.0, + 354.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1678.0, + 1294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1836.0, + 507.0, + 1836.0, + 507.0, + 1879.0, + 291.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1876.0, + 1403.0, + 1876.0, + 1403.0, + 1914.0, + 349.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1706.0, + 1404.0, + 1706.0, + 1404.0, + 1740.0, + 295.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1736.0, + 721.0, + 1736.0, + 721.0, + 1768.0, + 292.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1736.0, + 1404.0, + 1736.0, + 1404.0, + 1768.0, + 825.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1764.0, + 1408.0, + 1764.0, + 1408.0, + 1798.0, + 291.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1975.0, + 1404.0, + 1975.0, + 1404.0, + 2004.0, + 295.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 1403.0, + 2004.0, + 1403.0, + 2034.0, + 296.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1597.0, + 963.0, + 1597.0, + 963.0, + 1639.0, + 299.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1632.0, + 1272.0, + 1632.0, + 1272.0, + 1678.0, + 354.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1678.0, + 1294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 478, + 273, + 1209, + 273, + 1209, + 807, + 478, + 807 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 1048, + 1403, + 1048, + 1403, + 1145, + 297, + 1145 + ], + "score": 0.965 + }, + { + "category_id": 5, + "poly": [ + 547, + 1201, + 1153, + 1201, + 1153, + 1820, + 547, + 1820 + ], + "score": 0.957, + "html": "
TaskcontinuousD-NTMDiscreteD-NTMDiscrete*D-NTMDiscrete†D-NTM
14.3881.6714.7972.2881.6778.9579.6968.5431.6749.17
23427.571.2576.6779.3878.6576.6770.8344.06
0.00
51.6783.1317.71
61.4648.7648.13
76.0454.7923.54
81.7069.7535.6279.3237.7125.6382.0874.3847.0877.0873.9653.0230.4211.4676.0513.96
90.6339.1714.38
1019.8056.2556.25
110.0078.9639.58
126.2582.532.08
137.575.018.54
1417.578.7524.79
150.071.4239.73
1649.6571.4671.15
171.2543.7543.75
180.2448.132.9271.569.79
192039.4771.4676.56
0.0
Avg12.8168.3037.7957.21
" + }, + { + "category_id": 4, + "poly": [ + 297, + 854, + 1404, + 854, + 1404, + 945, + 297, + 945 + ], + "score": 0.95 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2113, + 835, + 2113 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 297, + 1841, + 1404, + 1841, + 1404, + 1964, + 297, + 1964 + ], + "score": 0.657 + }, + { + "category_id": 6, + "poly": [ + 297, + 1841, + 1404, + 1841, + 1404, + 1964, + 297, + 1964 + ], + "score": 0.266 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 854, + 76, + 854, + 104, + 299, + 104 + ], + "score": 0.095 + }, + { + "category_id": 13, + "poly": [ + 559, + 1843, + 601, + 1843, + 601, + 1872, + 559, + 1872 + ], + "score": 0.8, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1300, + 856, + 1403, + 856, + 1403, + 887, + 1300, + 887 + ], + "score": 0.64, + "latex": "p \\mathrm { M N I S T }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 890, + 1108, + 890, + 1108, + 912, + 1090, + 912 + ], + "score": 0.41, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 273.0, + 517.0, + 273.0, + 517.0, + 298.0, + 482.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 305.0, + 1163.0, + 305.0, + 1163.0, + 349.0, + 736.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 368.0, + 517.0, + 368.0, + 517.0, + 397.0, + 480.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 350.0, + 1181.0, + 350.0, + 1181.0, + 390.0, + 735.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 393.0, + 1135.0, + 393.0, + 1135.0, + 437.0, + 735.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 438.0, + 1152.0, + 438.0, + 1152.0, + 480.0, + 735.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 464.0, + 518.0, + 464.0, + 518.0, + 494.0, + 481.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 563.0, + 518.0, + 563.0, + 518.0, + 590.0, + 482.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 659.0, + 518.0, + 659.0, + 518.0, + 686.0, + 481.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 756.0, + 518.0, + 756.0, + 518.0, + 783.0, + 482.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 778.0, + 528.0, + 778.0, + 528.0, + 797.0, + 512.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 775.0, + 671.0, + 775.0, + 671.0, + 800.0, + 643.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 774.0, + 809.0, + 774.0, + 809.0, + 801.0, + 778.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 774.0, + 944.0, + 774.0, + 944.0, + 801.0, + 913.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 773.0, + 1081.0, + 773.0, + 1081.0, + 801.0, + 1050.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 774.0, + 1217.0, + 774.0, + 1217.0, + 801.0, + 1184.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 613.5, + 614.0, + 613.5, + 614.0, + 633.5, + 581.0, + 633.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.75, + 657.5, + 668.75, + 657.5, + 668.75, + 672.0, + 656.75, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 855.0, + 1299.0, + 855.0, + 1299.0, + 889.0, + 295.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 883.0, + 1089.0, + 883.0, + 1089.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 883.0, + 1404.0, + 883.0, + 1404.0, + 919.0, + 1109.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 914.0, + 575.0, + 914.0, + 575.0, + 948.0, + 294.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1841.0, + 558.0, + 1841.0, + 558.0, + 1873.0, + 294.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1841.0, + 1404.0, + 1841.0, + 1404.0, + 1873.0, + 602.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1869.0, + 1405.0, + 1869.0, + 1405.0, + 1904.0, + 294.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1904.0, + 1405.0, + 1904.0, + 1405.0, + 1935.0, + 293.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1935.0, + 1177.0, + 1935.0, + 1177.0, + 1964.0, + 295.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1048.0, + 1404.0, + 1048.0, + 1404.0, + 1083.0, + 296.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1083.0, + 1403.0, + 1083.0, + 1403.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1114.0, + 1365.0, + 1114.0, + 1365.0, + 1145.0, + 296.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1841.0, + 558.0, + 1841.0, + 558.0, + 1873.0, + 294.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1841.0, + 1404.0, + 1841.0, + 1404.0, + 1873.0, + 602.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1869.0, + 1405.0, + 1869.0, + 1405.0, + 1904.0, + 294.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1904.0, + 1405.0, + 1904.0, + 1405.0, + 1935.0, + 293.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1935.0, + 1177.0, + 1935.0, + 1177.0, + 1964.0, + 295.0, + 1964.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 409, + 453, + 1290, + 453, + 1290, + 1068, + 409, + 1068 + ], + "score": 0.98, + "html": "
TaskSoftD-NTM(1-step)DiscreteD-NTM(1-step)SoftD-NTM(3-steps)DiscreteD-NTM(3-steps)
10.000.000.000.00
23461.0459.3756.8755.62
55.6257.562.557.5
27.2924.8926.4527.08
513.5512.0815.8314.78
613.5414.3721.8713.33
78.546.258.7514.58
81.691.363.013.02
917.716.6637.7017.08
1026.0427.0826.8723.952.294.16
1120.413.952.5
120.410.830.20
133.121.044.795.83
1462.0858.3361.2560.620.05
1531.6626.250.62
1654.4748.5448.9548.9530.6236.04
1743.7531.8743.75
1833.7539.3736.66
192064.6369.2167.2365.46
1.250.001.450.00
Avg27.0224.9826.3624.05
" + }, + { + "category_id": 0, + "poly": [ + 296, + 225, + 986, + 225, + 986, + 263, + 296, + 263 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 299, + 292, + 1404, + 292, + 1404, + 442, + 299, + 442 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 864, + 2089, + 864, + 2112, + 836, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 6, + "poly": [ + 296, + 1090, + 1402, + 1090, + 1402, + 1152, + 296, + 1152 + ], + "score": 0.472 + }, + { + "category_id": 1, + "poly": [ + 296, + 1090, + 1402, + 1090, + 1402, + 1152, + 296, + 1152 + ], + "score": 0.463 + }, + { + "category_id": 6, + "poly": [ + 299, + 292, + 1404, + 292, + 1404, + 442, + 299, + 442 + ], + "score": 0.136 + }, + { + "category_id": 13, + "poly": [ + 569, + 1092, + 613, + 1092, + 613, + 1121, + 569, + 1121 + ], + "score": 0.8, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1025, + 1120, + 1090, + 1120, + 1090, + 1149, + 1025, + 1149 + ], + "score": 0.46, + "latex": "\\mathrm { B o W }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 991.0, + 222.0, + 991.0, + 268.0, + 291.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 568.0, + 1089.0, + 568.0, + 1125.0, + 295.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1089.0, + 1404.0, + 1089.0, + 1404.0, + 1125.0, + 614.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1115.0, + 1024.0, + 1115.0, + 1024.0, + 1158.0, + 293.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1115.0, + 1390.0, + 1115.0, + 1390.0, + 1158.0, + 1091.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 1409.0, + 293.0, + 1409.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1406.0, + 321.0, + 1406.0, + 359.0, + 294.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 409.0, + 737.0, + 409.0, + 737.0, + 448.0, + 293.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 1409.0, + 293.0, + 1409.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1406.0, + 321.0, + 1406.0, + 359.0, + 294.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 409.0, + 737.0, + 409.0, + 737.0, + 448.0, + 293.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 568.0, + 1089.0, + 568.0, + 1125.0, + 295.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1089.0, + 1404.0, + 1089.0, + 1404.0, + 1125.0, + 614.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1115.0, + 1024.0, + 1115.0, + 1024.0, + 1158.0, + 293.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1115.0, + 1390.0, + 1115.0, + 1390.0, + 1158.0, + 1091.0, + 1158.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/BkgYPREtPr/images/066175af26ade87ef420ad40c58c4452f8ece88a66ac76e90f88a068840ebd7c.jpg b/parse/train/BkgYPREtPr/images/066175af26ade87ef420ad40c58c4452f8ece88a66ac76e90f88a068840ebd7c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0acb6a8ec367cb0eaf20f3309abf128d9429d710 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/066175af26ade87ef420ad40c58c4452f8ece88a66ac76e90f88a068840ebd7c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:deaf4c8f7ea1f716f31f48097b45f85aa5f0b872c6a1c3b9172f05b7d0820b9e +size 9611 diff --git a/parse/train/BkgYPREtPr/images/12567491105ff78e397b8e0bafd6d07f4e0f00fdf6a056fc29a9afb37e1011a1.jpg b/parse/train/BkgYPREtPr/images/12567491105ff78e397b8e0bafd6d07f4e0f00fdf6a056fc29a9afb37e1011a1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1f9533e95a8d4751fbce2d5dbf3dfd5f52a64830 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/12567491105ff78e397b8e0bafd6d07f4e0f00fdf6a056fc29a9afb37e1011a1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0f379cc0964e50fc37e7878ee12029649cfdca9cb5bd69c6e811681be80c181 +size 3931 diff --git a/parse/train/BkgYPREtPr/images/1d5eaa40baa0a9adb3bdfa97f37c5c61623ab810e84ef379d09259709df9b0e4.jpg b/parse/train/BkgYPREtPr/images/1d5eaa40baa0a9adb3bdfa97f37c5c61623ab810e84ef379d09259709df9b0e4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a6d88967cf54eab5444fd658af18ba6458e12446 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/1d5eaa40baa0a9adb3bdfa97f37c5c61623ab810e84ef379d09259709df9b0e4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57122517f705adf0582127a4c2f48633e735cd17c5b7d0d5d6fe9ee7841ab62c +size 36490 diff --git a/parse/train/BkgYPREtPr/images/211aef153d54559dad958c17bfa2018004136b7685bc3cf6e29b51a98cb37382.jpg b/parse/train/BkgYPREtPr/images/211aef153d54559dad958c17bfa2018004136b7685bc3cf6e29b51a98cb37382.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f616a943d7d9ed6cb400fbd8f65a76e6b286c375 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/211aef153d54559dad958c17bfa2018004136b7685bc3cf6e29b51a98cb37382.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9aeda9217875a448d5a7daeb85cd1bfd8da21e6fd9ec3f11788ffad3b8261826 +size 4728 diff --git a/parse/train/BkgYPREtPr/images/244d9eb4162103278e38ebcf9f6c3fae6a695441df234562e62e061229592d61.jpg b/parse/train/BkgYPREtPr/images/244d9eb4162103278e38ebcf9f6c3fae6a695441df234562e62e061229592d61.jpg new file mode 100644 index 0000000000000000000000000000000000000000..58d1e7a3ffc6a9328874cc9e61926876e1d20c17 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/244d9eb4162103278e38ebcf9f6c3fae6a695441df234562e62e061229592d61.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08d1f902fbf8a19054f84348de02bac58e9c39177ffd5fe8fefaa43b67e632ef +size 37402 diff --git a/parse/train/BkgYPREtPr/images/2d14d0e2b29edf052b00c5624104bc69a4b813db77c87f132af3872ca1e0df83.jpg b/parse/train/BkgYPREtPr/images/2d14d0e2b29edf052b00c5624104bc69a4b813db77c87f132af3872ca1e0df83.jpg new file mode 100644 index 0000000000000000000000000000000000000000..65cf3522e2c42bb0697bfa5763fa47c09512f2e8 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/2d14d0e2b29edf052b00c5624104bc69a4b813db77c87f132af3872ca1e0df83.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1dbf741e3c92eb0bd460760d3ba31cc7d87a54b137dd5aa4ad433dde676e275 +size 9568 diff --git a/parse/train/BkgYPREtPr/images/3dbd1f5d1f519066236c9331cb51d2898b592d20f173cbb2c0c144ea285f871f.jpg b/parse/train/BkgYPREtPr/images/3dbd1f5d1f519066236c9331cb51d2898b592d20f173cbb2c0c144ea285f871f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ad5ef43fd308a2e7f60755f7d87f3a5267f40ac7 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/3dbd1f5d1f519066236c9331cb51d2898b592d20f173cbb2c0c144ea285f871f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c85a9a6523106894512a3ba245df0c1c5b2e8b267eab70480144ae2abb83e173 +size 80270 diff --git a/parse/train/BkgYPREtPr/images/46b7091903c117c2e472c306df82dac9b9b1ce8fccf50a4d154ed5cd567dd5b0.jpg b/parse/train/BkgYPREtPr/images/46b7091903c117c2e472c306df82dac9b9b1ce8fccf50a4d154ed5cd567dd5b0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a71de2d80b3423fa930e7c311d9cd0b1a2b8c2df --- /dev/null +++ b/parse/train/BkgYPREtPr/images/46b7091903c117c2e472c306df82dac9b9b1ce8fccf50a4d154ed5cd567dd5b0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03735df05f421223e9a61874a0fb51c06b149a0613db1ee8980850ca871f9d69 +size 7290 diff --git a/parse/train/BkgYPREtPr/images/47e66723c86bd642f34030c67e03d7d775c0518e6669ae9c9d21c25fdc0fa895.jpg b/parse/train/BkgYPREtPr/images/47e66723c86bd642f34030c67e03d7d775c0518e6669ae9c9d21c25fdc0fa895.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c04f54326df8b8a3aa1cca38cb478d59e12a64f9 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/47e66723c86bd642f34030c67e03d7d775c0518e6669ae9c9d21c25fdc0fa895.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:402aadb800ef5df93c55ea084afa2195519a5d1d864e548e4b054fa79e3be6ae +size 5261 diff --git a/parse/train/BkgYPREtPr/images/58eef280515505d2d2a197f0815c1b5e745e524840a019420424404c881865fc.jpg b/parse/train/BkgYPREtPr/images/58eef280515505d2d2a197f0815c1b5e745e524840a019420424404c881865fc.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ea6c208ad8c26fe033b468b2077a9296dc88f91e --- /dev/null +++ b/parse/train/BkgYPREtPr/images/58eef280515505d2d2a197f0815c1b5e745e524840a019420424404c881865fc.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a238a3004ec78c45454e06bc287781a5846f2fcddae1753f3533f80bff593aff +size 33121 diff --git a/parse/train/BkgYPREtPr/images/5c1db73dcb5b3fc5e3ef840365db4876bb70b01e700a39d154afa83332a0b7a8.jpg b/parse/train/BkgYPREtPr/images/5c1db73dcb5b3fc5e3ef840365db4876bb70b01e700a39d154afa83332a0b7a8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..540eb31a1936105fa782944197b19c34ddcc85b2 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/5c1db73dcb5b3fc5e3ef840365db4876bb70b01e700a39d154afa83332a0b7a8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87eebd251a584591bf0c6b112a874928aeed7bc94883464a5409b677685a662c +size 2984 diff --git a/parse/train/BkgYPREtPr/images/604b4047c708aa608825994b79054dc45866f894e762b2dea61dbb1e5d78f317.jpg b/parse/train/BkgYPREtPr/images/604b4047c708aa608825994b79054dc45866f894e762b2dea61dbb1e5d78f317.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a8a6d75456d7995e5931f6a82d04ad1c949bce8c --- /dev/null +++ b/parse/train/BkgYPREtPr/images/604b4047c708aa608825994b79054dc45866f894e762b2dea61dbb1e5d78f317.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d411adebd1402bce9b8002cc07f0280c80551e93dfa49b0e402d613c6863dd1 +size 6422 diff --git a/parse/train/BkgYPREtPr/images/69faf379f7b86e81a31ac7cde5d9a6d26b838717c5ed192578fa109d9039f679.jpg b/parse/train/BkgYPREtPr/images/69faf379f7b86e81a31ac7cde5d9a6d26b838717c5ed192578fa109d9039f679.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0cf0c429423f7c550675c48dcdb8a04f76420bae --- /dev/null +++ b/parse/train/BkgYPREtPr/images/69faf379f7b86e81a31ac7cde5d9a6d26b838717c5ed192578fa109d9039f679.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0478d8b0cad04d82780ab7ccd26be9a464f2c6e4e17647f771d1720cc775216 +size 53604 diff --git a/parse/train/BkgYPREtPr/images/7674f229951c2cac24c31140a8aaae74f4f711ada20d1e1d4b4ad44b0f8351dd.jpg b/parse/train/BkgYPREtPr/images/7674f229951c2cac24c31140a8aaae74f4f711ada20d1e1d4b4ad44b0f8351dd.jpg new file mode 100644 index 0000000000000000000000000000000000000000..338f76e77efce7bd644318830a64018db5f71204 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/7674f229951c2cac24c31140a8aaae74f4f711ada20d1e1d4b4ad44b0f8351dd.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cc218e8378bf6e8b271644bd98e84dc917487f7447e07105f9b02b9a5f7b1e4 +size 34559 diff --git a/parse/train/BkgYPREtPr/images/7d83acaa24389edd2280bbbaeec81b83c93918316c7c749434afef3ec1fa4a2e.jpg b/parse/train/BkgYPREtPr/images/7d83acaa24389edd2280bbbaeec81b83c93918316c7c749434afef3ec1fa4a2e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9d955ff364a709bd7b79dd6f068daf5d7c7989cc --- /dev/null +++ b/parse/train/BkgYPREtPr/images/7d83acaa24389edd2280bbbaeec81b83c93918316c7c749434afef3ec1fa4a2e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:752c2fe535d696e43477cf970b46418ba2ed6078ae9805a7e086bdea7f812fb4 +size 33468 diff --git a/parse/train/BkgYPREtPr/images/82006e85767afeac6b04f57786af22c9d5f9daee7cfe8516652bd6de02ac6884.jpg b/parse/train/BkgYPREtPr/images/82006e85767afeac6b04f57786af22c9d5f9daee7cfe8516652bd6de02ac6884.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8fae3cb96ad70e7b5207fd4cc817fb0aca2ea1cf --- /dev/null +++ b/parse/train/BkgYPREtPr/images/82006e85767afeac6b04f57786af22c9d5f9daee7cfe8516652bd6de02ac6884.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b72b6d182c7998f1b1d79f3dffdcb8d9fef6eda027944894100838f6adee6646 +size 3084 diff --git a/parse/train/BkgYPREtPr/images/83b5660a973c4dc436d8959be4cf1868b47997b26d7f9f5e681925629384c6b3.jpg b/parse/train/BkgYPREtPr/images/83b5660a973c4dc436d8959be4cf1868b47997b26d7f9f5e681925629384c6b3.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7c90035d8b5704d00eefdfbdda7a716c35eef2a4 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/83b5660a973c4dc436d8959be4cf1868b47997b26d7f9f5e681925629384c6b3.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3d529ed82abac2d8080f38c4235e7912e4973a4883fa44bc4ca49aa458cfd6c +size 52667 diff --git a/parse/train/BkgYPREtPr/images/8d1b84cee8f0bdd381c9e9b27d07f884439873a400ac83b5561d904c15297b6a.jpg b/parse/train/BkgYPREtPr/images/8d1b84cee8f0bdd381c9e9b27d07f884439873a400ac83b5561d904c15297b6a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d1f51438ba18781d21af709bf290d050550f69bf --- /dev/null +++ b/parse/train/BkgYPREtPr/images/8d1b84cee8f0bdd381c9e9b27d07f884439873a400ac83b5561d904c15297b6a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a506650c7aad27f5829b8445cbda014caa97a7f4be803db03d83a064418032d6 +size 6438 diff --git a/parse/train/BkgYPREtPr/images/8e9d95c519606a34e4b13b0a5a508d99139fce19b7ab0939cd5f87effadb965c.jpg b/parse/train/BkgYPREtPr/images/8e9d95c519606a34e4b13b0a5a508d99139fce19b7ab0939cd5f87effadb965c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..09e53a4a9c88d30981e55ee1aafa3de13d7b5816 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/8e9d95c519606a34e4b13b0a5a508d99139fce19b7ab0939cd5f87effadb965c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9513fb35ccee689642abb857bb377c377a98b41bf91bdbcf5b6ec9e1813457b +size 17306 diff --git a/parse/train/BkgYPREtPr/images/928b160a39136b7a95f56e26b6408e07ac879442583ef9c027670d3a898e4e6e.jpg b/parse/train/BkgYPREtPr/images/928b160a39136b7a95f56e26b6408e07ac879442583ef9c027670d3a898e4e6e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4e92bade0e2b457fba6ce717c94170e1102b3429 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/928b160a39136b7a95f56e26b6408e07ac879442583ef9c027670d3a898e4e6e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19a31fa90d202f0068b8e7803a9105ab6cff827a9d4882115370f514314fef17 +size 4490 diff --git a/parse/train/BkgYPREtPr/images/9a21ffa9e2d59c98ed0251c75ddf083c188ef56ae61e6a4c1d5a34cc6c2f612a.jpg b/parse/train/BkgYPREtPr/images/9a21ffa9e2d59c98ed0251c75ddf083c188ef56ae61e6a4c1d5a34cc6c2f612a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3b2cf24dc3228fe200c54ef13bc6bf0c6bec2702 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/9a21ffa9e2d59c98ed0251c75ddf083c188ef56ae61e6a4c1d5a34cc6c2f612a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6635d9f5e141d633a879db494b6c6955d7b783a1e9232c63d7a83ff797a5410 +size 34935 diff --git a/parse/train/BkgYPREtPr/images/a692b4e7c3b0e871013011cf7c192888522d08c6ef52db13d5aace12288d8481.jpg b/parse/train/BkgYPREtPr/images/a692b4e7c3b0e871013011cf7c192888522d08c6ef52db13d5aace12288d8481.jpg new file mode 100644 index 0000000000000000000000000000000000000000..13a742441e1c7d1115f290bb08f098689362ae67 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/a692b4e7c3b0e871013011cf7c192888522d08c6ef52db13d5aace12288d8481.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:990cdf11f790d5d4f706ec6f067e56228ab6a99355e5737d09b533fe47db5a0a +size 20135 diff --git a/parse/train/BkgYPREtPr/images/b984b2d7fc940f0c081de21685e85cab62c13266e79dfd784038cb20f8d48b4d.jpg b/parse/train/BkgYPREtPr/images/b984b2d7fc940f0c081de21685e85cab62c13266e79dfd784038cb20f8d48b4d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4d0757fd535304406b0b9c21b66e7e3c337aa323 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/b984b2d7fc940f0c081de21685e85cab62c13266e79dfd784038cb20f8d48b4d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3974978cc5b4c6cea30eb0272d94f48e4a8719ded8ca429a1e30bae825951335 +size 33913 diff --git a/parse/train/BkgYPREtPr/images/ba13579af38b65d9adb43b36d245df8956ec780539b558ba49a9813db37afa1a.jpg b/parse/train/BkgYPREtPr/images/ba13579af38b65d9adb43b36d245df8956ec780539b558ba49a9813db37afa1a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..92db69bda2d1a9d194d188f092e37ed0170353f8 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/ba13579af38b65d9adb43b36d245df8956ec780539b558ba49a9813db37afa1a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5c038833bd1352e1f08395a32fbf652690d0cb2d267ee48a83c8ec4c99f80e2 +size 35033 diff --git a/parse/train/BkgYPREtPr/images/bb84020d14394637248edf3dde128a9af425810b173e863d12de5f7badcf873f.jpg b/parse/train/BkgYPREtPr/images/bb84020d14394637248edf3dde128a9af425810b173e863d12de5f7badcf873f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..cfa05c05449b01983a3227736a56fb70f4ee9914 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/bb84020d14394637248edf3dde128a9af425810b173e863d12de5f7badcf873f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d79c95778a3695230caf4b393bed6cc1e0ddeee19ab9cf3c6af8cf82a7d84801 +size 35845 diff --git a/parse/train/BkgYPREtPr/images/c70f8eea76b7d41e606bd8ff096251202e2055959db8d56f7de1010669a612ae.jpg b/parse/train/BkgYPREtPr/images/c70f8eea76b7d41e606bd8ff096251202e2055959db8d56f7de1010669a612ae.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b9d42968b34b675ed161f57c7d3d4104d3a4c1a4 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/c70f8eea76b7d41e606bd8ff096251202e2055959db8d56f7de1010669a612ae.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d4b2438754a7db6554cb7a5717e63ad7a460ec01e7e28b453762efa3e73ce41 +size 3977 diff --git a/parse/train/BkgYPREtPr/images/d01575285c4fe0dbbb50cf80c42081c9b34cbce99ee011dfeae65ad2b7853cac.jpg b/parse/train/BkgYPREtPr/images/d01575285c4fe0dbbb50cf80c42081c9b34cbce99ee011dfeae65ad2b7853cac.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1954155519b2bca3d651bccc9ff432178667186a --- /dev/null +++ b/parse/train/BkgYPREtPr/images/d01575285c4fe0dbbb50cf80c42081c9b34cbce99ee011dfeae65ad2b7853cac.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a449b9967348d643aae150712162025664262e0bb0f4f9eea94610120281734 +size 35669 diff --git a/parse/train/BkgYPREtPr/images/d79ed8533c2414feaafacaffc61ee1cf39bf13c3bc22d69b657dfa347093f1b7.jpg b/parse/train/BkgYPREtPr/images/d79ed8533c2414feaafacaffc61ee1cf39bf13c3bc22d69b657dfa347093f1b7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..38fe8ec9a4e219b5edd276ace3fad9fea101b8e6 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/d79ed8533c2414feaafacaffc61ee1cf39bf13c3bc22d69b657dfa347093f1b7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02e037d995b12ccb2bad1676449f05ce2c1d4964f0f2ba92456ade82f5d7bb0f +size 36583 diff --git a/parse/train/BkgYPREtPr/images/d8edef04ebcaf8356a642dab32e6e9738edbcd86e2f6c04891aa9dd60b00e959.jpg b/parse/train/BkgYPREtPr/images/d8edef04ebcaf8356a642dab32e6e9738edbcd86e2f6c04891aa9dd60b00e959.jpg new file mode 100644 index 0000000000000000000000000000000000000000..da1dce2caf0d3227dcb33cfbbc99865f0ba33235 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/d8edef04ebcaf8356a642dab32e6e9738edbcd86e2f6c04891aa9dd60b00e959.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6435c69e336721fa772ddfcb0d561d558878952f08f71138a8b7b51426456f3 +size 53711 diff --git a/parse/train/BkgYPREtPr/images/d97e59307a592699aa87a2c136a20f8f58f31165ceb40016c1a9bfa840d010cb.jpg b/parse/train/BkgYPREtPr/images/d97e59307a592699aa87a2c136a20f8f58f31165ceb40016c1a9bfa840d010cb.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ef1a940c9c2830d03fc2aae57581d48d31e9d969 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/d97e59307a592699aa87a2c136a20f8f58f31165ceb40016c1a9bfa840d010cb.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:779b00ed4a5dd97ea376a03412919c286664691f40ed89ad77d4a3cd1eefd9a2 +size 63495 diff --git a/parse/train/BkgYPREtPr/images/da0653d940a2d7564549b2253a7aafae4fd8d9cbaa3f2890e6ccf4c2cc261742.jpg b/parse/train/BkgYPREtPr/images/da0653d940a2d7564549b2253a7aafae4fd8d9cbaa3f2890e6ccf4c2cc261742.jpg new file mode 100644 index 0000000000000000000000000000000000000000..276fc4a3d26b16383ccd954114411c0ff2ca29ad --- /dev/null +++ b/parse/train/BkgYPREtPr/images/da0653d940a2d7564549b2253a7aafae4fd8d9cbaa3f2890e6ccf4c2cc261742.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61c87a8431aa0b6f4de77e073d61083fede731e28962e96e95996df593f23088 +size 3564 diff --git a/parse/train/BkgYPREtPr/images/e8c24983c3fed7cf6a51524359b7491dbc3fe35a3d656cc0408f581fb320a6a0.jpg b/parse/train/BkgYPREtPr/images/e8c24983c3fed7cf6a51524359b7491dbc3fe35a3d656cc0408f581fb320a6a0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e7de76e1557baa2567daffbbd0d5ebae40c4cc9e --- /dev/null +++ b/parse/train/BkgYPREtPr/images/e8c24983c3fed7cf6a51524359b7491dbc3fe35a3d656cc0408f581fb320a6a0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3b92a5d3518539bfafe09344bc8ac574182e2e80bd17ae9f4018a1ef7aec7fb +size 37750 diff --git a/parse/train/BkgYPREtPr/images/eee0d8f6c04f4bfdbfb00b4d5eeb875bdcd1f04d436115517315983e6d9cad88.jpg b/parse/train/BkgYPREtPr/images/eee0d8f6c04f4bfdbfb00b4d5eeb875bdcd1f04d436115517315983e6d9cad88.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c534f735622e7ceee84f28b0e8e713a42dc57c0d --- /dev/null +++ b/parse/train/BkgYPREtPr/images/eee0d8f6c04f4bfdbfb00b4d5eeb875bdcd1f04d436115517315983e6d9cad88.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:954d96d88c211cef832c5472825b3af4bdce1a6de90c5d09a9f0dc02f5eb65aa +size 56219 diff --git a/parse/train/BkgYPREtPr/images/f39b053fb4377ba51e492d655adf677dbfc95921a2904d29437cf470b644fd76.jpg b/parse/train/BkgYPREtPr/images/f39b053fb4377ba51e492d655adf677dbfc95921a2904d29437cf470b644fd76.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a25b6eae2be9b400bebbd3da55dd90a6200be001 --- /dev/null +++ b/parse/train/BkgYPREtPr/images/f39b053fb4377ba51e492d655adf677dbfc95921a2904d29437cf470b644fd76.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d311fa55baad7acd2e1d3de7b04380cf42e2cbc9bbc700665c5a60001659543a +size 50180 diff --git a/parse/train/BkgYPREtPr/images/f4acf6f3d7bffc814663fb6df875bd940961e319078f2fc2fe7d07ac0fe6a8c2.jpg b/parse/train/BkgYPREtPr/images/f4acf6f3d7bffc814663fb6df875bd940961e319078f2fc2fe7d07ac0fe6a8c2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f385b035330a7ca34aade82f4a0dc4a3c1580c8c --- /dev/null +++ b/parse/train/BkgYPREtPr/images/f4acf6f3d7bffc814663fb6df875bd940961e319078f2fc2fe7d07ac0fe6a8c2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd9c0499f773a16c469402fdbe0985176e32556fd7dc18bcc6f18a01defdd62f +size 58077 diff --git a/parse/train/BkgYPREtPr/images/fc0f82d6f9d40437eaf549fc69e6d2a6e9f2529c8359d7eabe5e2d8357b51354.jpg b/parse/train/BkgYPREtPr/images/fc0f82d6f9d40437eaf549fc69e6d2a6e9f2529c8359d7eabe5e2d8357b51354.jpg new file mode 100644 index 0000000000000000000000000000000000000000..33790424e5990e060d0446842883a39199fc728e --- /dev/null +++ b/parse/train/BkgYPREtPr/images/fc0f82d6f9d40437eaf549fc69e6d2a6e9f2529c8359d7eabe5e2d8357b51354.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bccbbc946d4008b2a5350ae07a146a34443e7c88de8669beb91abc226f544c7d +size 33714 diff --git a/parse/train/ByldLrqlx/ByldLrqlx_content_list.json b/parse/train/ByldLrqlx/ByldLrqlx_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..851763785b05568bcddf1fc25a966ac5b0359407 --- /dev/null +++ b/parse/train/ByldLrqlx/ByldLrqlx_content_list.json @@ -0,0 +1,2470 @@ +[ + { + "type": "text", + "text": "DEEPCODER: LEARNING TO WRITE PROGRAMS ", + "text_level": 1, + "bbox": [ + 176, + 98, + 748, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matej Balog∗ Department of Engineering University of Cambridge ", + "bbox": [ + 184, + 145, + 364, + 188 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexander L. Gaunt, Marc Brockschmidt, Sebastian Nowozin, Daniel Tarlow Microsoft Research ", + "bbox": [ + 450, + 145, + 745, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 224, + 544, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We develop a first line of attack for solving programming competition-style problems from input-output examples using deep learning. The approach is to train a neural network to predict properties of the program that generated the outputs from the inputs. We use the neural network’s predictions to augment search techniques from the programming languages community, including enumerative search and an SMT-based solver. Empirically, we show that our approach leads to an order of magnitude speedup over the strong non-augmented baselines and a Recurrent Neural Network approach, and that we are able to solve problems of difficulty comparable to the simplest problems on programming competition websites. ", + "bbox": [ + 233, + 255, + 766, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 405, + 336, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A dream of artificial intelligence is to build systems that can write computer programs. Recently, there has been much interest in program-like neural network models (Graves et al., 2014; Weston et al., 2015; Kurach et al., 2015; Joulin & Mikolov, 2015; Grefenstette et al., 2015; Sukhbaatar et al., 2015; Neelakantan et al., 2016; Kaiser & Sutskever, 2016; Reed & de Freitas, 2016; Zaremba et al., 2016; Graves et al., 2016), but none of these can write programs; that is, they do not generate human-readable source code. Only very recently, Riedel et al. (2016); Bunel et al. (2016); Gaunt et al. (2016) explored the use of gradient descent to induce source code from input-output examples via differentiable interpreters, and Ling et al. (2016) explored the generation of source code from unstructured text descriptions. However, Gaunt et al. (2016) showed that differentiable interpreterbased program induction is inferior to discrete search-based techniques used by the programming languages community. We are then left with the question of how to make progress on program induction using machine learning techniques. ", + "bbox": [ + 174, + 436, + 825, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose two main ideas: (1) learn to induce programs; that is, use a corpus of program induction problems to learn strategies that generalize across problems, and (2) integrate neural network architectures with search-based techniques rather than replace them. ", + "bbox": [ + 174, + 611, + 825, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In more detail, we can contrast our approach to existing work on differentiable interpreters. In differentiable interpreters, the idea is to define a differentiable mapping from source code and inputs to outputs. After observing inputs and outputs, gradient descent can be used to search for a program that matches the input-output examples. This approach leverages gradient-based optimization, which has proven powerful for training neural networks, but each synthesis problem is still solved independently—solving many synthesis problems does not help to solve the next problem. ", + "bbox": [ + 174, + 660, + 825, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that machine learning can provide significant value towards solving Inductive Program Synthesis (IPS) by re-casting the problem as a big data problem. We show that training a neural network on a large number of generated IPS problems to predict cues from the problem description can help a search-based technique. In this work, we focus on predicting an order on the program space and show how to use it to guide search-based techniques that are common in the programming languages community. This approach has three desirable properties: first, we transform a difficult search problem into a supervised learning problem; second, we soften the effect of failures of the neural network by searching over program space rather than relying on a single prediction; and third, the neural network’s predictions are used to guide existing program synthesis systems, allowing us to use and improve on the best solvers from the programming languages community. Empirically, we show orders-of-magnitude improvements over optimized standard search techniques and a Recurrent Neural Network-based approach to the problem. ", + "bbox": [ + 174, + 750, + 825, + 890 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we define and instantiate a framework for using deep learning for program synthesis problems like ones appearing on programming competition websites. Our concrete contributions are: ", + "bbox": [ + 173, + 138, + 825, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. defining a programming language that is expressive enough to include real-world programming problems while being high-level enough to be predictable from input-output examples; 2. models for mapping sets of input-output examples to program properties; and 3. experiments that show an order of magnitude speedup over standard program synthesis techniques, which makes this approach feasible for solving problems of similar difficulty as the simplest problems that appear on programming competition websites. ", + "bbox": [ + 212, + 178, + 825, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND ON INDUCTIVE PROGRAM SYNTHESIS ", + "text_level": 1, + "bbox": [ + 174, + 291, + 647, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by providing background on Inductive Program Synthesis, including a brief overview of how it is typically formulated and solved in the programming languages community. ", + "bbox": [ + 178, + 321, + 823, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Inductive Program Synthesis (IPS) problem is the following: given input-output examples, produce a program that has behavior consistent with the examples. ", + "bbox": [ + 176, + 357, + 821, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building an IPS system requires solving two problems. First, the search problem: to find consistent programs we need to search over a suitable set of possible programs. We need to define the set (i.e., the program space) and search procedure. Second, the ranking problem: if there are multiple programs consistent with the input-output examples, which one do we return? Both of these problems are dependent on the specifics of the problem formulation. Thus, the first important decision in formulating an approach to program synthesis is the choice of a Domain Specific Language. ", + "bbox": [ + 174, + 392, + 825, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Specific Languages (DSLs). DSLs are programming languages that are suitable for a specialized domain but are more restrictive than full-featured programming languages. For example, one might disallow loops or other control flow, and only allow string data types and a small number of primitive operations like concatenation. Most of program synthesis research focuses on synthesizing programs in DSLs, because full-featured languages like $\\mathrm { C } { + + }$ enlarge the search space and complicate synthesis. Restricted DSLs can also enable more efficient special-purpose search algorithms. For example, if a DSL only allows concatenations of substrings of an input string, a dynamic programming algorithm can efficiently search over all possible programs (Polozov & Gulwani, 2015). The choice of DSL also affects the difficulty of the ranking problem. For example, in a DSL without if statements, the same algorithm is applied to all inputs, reducing the number of programs consistent with any set of input-output examples, and thus the ranking problem becomes easier. Of course, the restrictiveness of the chosen DSL also determines which problems the system can solve at all. ", + "bbox": [ + 174, + 491, + 826, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Search Techniques. There are many techniques for searching for programs consistent with inputoutput examples. Perhaps the simplest approach is to define a grammar and then enumerate all derivations of the grammar, checking each one for consistency with the examples. This approach can be combined with pruning based on types and other logical reasoning (Feser et al., 2015). While simple, these approaches can be implemented efficiently, and they can be surprisingly effective. ", + "bbox": [ + 174, + 672, + 825, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In restricted domains such as the concatenation example discussed above, special-purpose algorithms can be used. FlashMeta (Polozov & Gulwani, 2015) describes a framework for DSLs which allow decomposition of the search problem, e.g., where the production of an output string from an input string can be reduced to finding a program for producing the first part of the output and concatenating it with a program for producing the latter part of the output string. ", + "bbox": [ + 174, + 750, + 825, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another class of systems is based on Satisfiability Modulo Theories (SMT) solving. SMT combines SAT-style search with theories like arithmetic and inequalities, with the benefit that theory-dependent subproblems can be handled by special-purpose solvers. For example, a special-purpose solver can easily find integers $x , y$ such that $x < y$ and $y < - 1 0 0$ hold, whereas an enumeration strategy may need to consider many values before satisfying the constraints. Many program synthesis engines based on SMT solvers exist, e.g., Sketch (Solar-Lezama, 2008) and Brahma (Gulwani et al., 2011). They convert the semantics of a DSL into a set of constraints between variables representing the program and the input-output values, and then call an SMT solver to find a satisfying setting of the program variables. This approach shines when special-purpose reasoning can be leveraged, but complex DSLs can lead to very large constraint problems where constructing and manipulating the constraints can be a lot slower than an enumerative approach. ", + "bbox": [ + 174, + 825, + 825, + 922 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, stochastic local search can be employed to search over program space, and there is a long history of applying genetic algorithms to this problem. One of the most successful recent examples is the STOKE super-optimization system (Schkufza et al., 2016), which uses stochastic local search to find assembly programs that have the same semantics as an input program but execute faster. ", + "bbox": [ + 173, + 166, + 825, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ranking. While we focus on the search problem in this work, we briefly mention the ranking problem here. A popular choice for ranking is to choose the shortest program consistent with inputoutput examples (Gulwani, 2016). A more sophisticated approach is employed by FlashFill (Singh & Gulwani, 2015). It works in a manner similar to max-margin structured prediction, where known ground truth programs are given, and the learning task is to assign scores to programs such that the ground truth programs score higher than other programs that satisfy the input-output specification. ", + "bbox": [ + 174, + 238, + 825, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 LEARNING INDUCTIVE PROGRAM SYNTHESIS (LIPS) ", + "text_level": 1, + "bbox": [ + 176, + 343, + 650, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we outline the general approach that we follow in this work, which we call Learning Inductive Program Synthesis (LIPS). The details of our instantiation of LIPS appear in Sect. 4. The components of LIPS are (1) a DSL specification, (2) a data-generation procedure, (3) a machine learning model that maps from input-output examples to program attributes, and (4) a search procedure that searches program space in an order guided by the model from (3). The framework is related to the formulation of Menon et al. (2013); the relationship and key differences are discussed in Sect. 6. ", + "bbox": [ + 174, + 375, + 825, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) DSL and Attributes. The choice of DSL is important in LIPS, just as it is in any program synthesis system. It should be expressive enough to capture the problems that we wish to solve, but restricted as much as possible to limit the difficulty of the search. In LIPS we additionally specify an attribute function $\\mathcal { A }$ that maps programs $P$ of the DSL to finite attribute vectors $\\mathbf { \\boldsymbol { a } } \\overset { \\cdot } { = } \\mathbf { \\bar { A } } ( \\mathbf { \\boldsymbol { P } } )$ . (Attribute vectors of different programs need not have equal length.) Attributes serve as the link between the machine learning and the search component of LIPS: the machine learning model predicts a distribution $q ( \\pmb { a } \\mid \\mathcal { E } )$ , where $\\mathcal { E }$ is the set of input-output examples, and the search procedure aims to search over programs $P$ as ordered by $q ( { \\bar { \\mathcal { A } } } ( { \\bar { P } } ) \\mid \\varepsilon )$ . Thus an attribute is useful if it is both predictable from input-output examples, and if conditioning on its value significantly reduces the effective size of the search space. ", + "bbox": [ + 174, + 474, + 825, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Possible attributes are the (perhaps position-dependent) presence or absence of high-level functions (e.g., does the program contain or end in a call to SORT). Other possible attributes include control flow templates (e.g., the number of loops and conditionals). In the extreme case, one may set $\\mathcal { A }$ to the identity function, in which case the attribute is equivalent to the program; however, in our experiments we find that performance is improved by choosing a more abstract attribute function. ", + "bbox": [ + 174, + 621, + 825, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Data Generation. Step 2 is to generate a dataset $( ( P ^ { ( n ) } , \\pmb { a } ^ { ( n ) } , \\pmb { \\mathcal { E } } ^ { ( n ) } ) ) _ { n = 1 } ^ { N }$ of programs $P ^ { ( n ) }$ in the chosen DSL, their attributes $\\mathbf { \\pmb { a } } ^ { ( n ) }$ , and accompanying input-output examples ${ \\mathcal { E } } ^ { ( n ) }$ . Different approaches are possible, ranging from enumerating valid programs in the DSL and pruning, to training a more sophisticated generative model of programs in the DSL. The key in the LIPS formulation is to ensure that it is feasible to generate a large dataset (ideally millions of programs). ", + "bbox": [ + 174, + 707, + 825, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(3) Machine Learning Model. The machine learning problem is to learn a distribution of attributes given input-output examples, $q ( \\pmb { a } \\mid \\mathcal { E } )$ . There is freedom to explore a large space of models, so long as the input component can encode $\\mathcal { E }$ , and the output is a proper distribution over attributes (e.g., if attributes are a fixed-size binary vector, then a neural network with independent sigmoid outputs is appropriate; if attributes are variable size, then a recurrent neural network output could be used). Attributes are observed at training time, so training can use a maximum likelihood objective. ", + "bbox": [ + 174, + 795, + 825, + 880 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(4) Search. The aim of the search component is to interface with an existing solver, using the predicted $q ( \\pmb { a } \\mid \\mathcal { E } )$ to guide the search. We describe specific approaches in the next section. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 DEEPCODER ", + "text_level": 1, + "bbox": [ + 174, + 102, + 312, + 118 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we describe DeepCoder, our instantiation of LIPS including a choice of DSL, a data generation strategy, models for encoding input-output sets, and algorithms for searching over program space. ", + "bbox": [ + 176, + 133, + 826, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 DOMAIN SPECIFIC LANGUAGE AND ATTRIBUTES ", + "text_level": 1, + "bbox": [ + 173, + 180, + 553, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider binary attributes indicating the presence or absence of high-level functions in the target program. To make this effective, the chosen DSL needs to contain constructs that are not so low-level that they all appear in the vast majority of programs, but at the same time should be common enough so that predicting their occurrence from input-output examples can be learned successfully. ", + "bbox": [ + 174, + 207, + 825, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following this observation, our DSL is loosely inspired by query languages such as SQL or LINQ, where high-level functions are used in sequence to manipulate data. A program in our DSL is a sequence of function calls, where the result of each call initializes a fresh variable that is either a singleton integer or an integer array. Functions can be applied to any of the inputs or previously computed (intermediate) variables. The output of the program is the return value of the last function call, i.e., the last variable. See Fig. 1 for an example program of length $T = 4$ in our DSL. ", + "bbox": [ + 173, + 270, + 825, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/95255e32da9df33db9a517e94978033a462b93ebbbe24b571c4b80d5e2c7f8e6.jpg", + "image_caption": [ + "Figure 1: An example program in our DSL that takes a single integer array as its input. " + ], + "image_footnote": [], + "bbox": [ + 174, + 367, + 841, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, our DSL contains the first-order functions HEAD, LAST, TAKE, DROP, ACCESS, MINIMUM, MAXIMUM, REVERSE, SORT, SUM, and the higher-order functions MAP, FILTER, COUNT, ZIPWITH, SCANL1. Higher-order functions require suitable lambda functions for their behavior to be fully specified: for MAP our DSL provides lambdas $( + 1 )$ , $( - 1 )$ , $( \\star 2 )$ , $( / 2 )$ ), $( \\star \\ ( - 1 ) \\ )$ ), $( \\star \\star 2 )$ , $( \\star 3 )$ , $( / 3 )$ , $( \\star 4 )$ , $( / 4 )$ ; for FILTER and COUNT there are predicates $( > 0 )$ ), $( < 0 )$ , ( $\\scriptstyle { \\frac { 0 } { 0 } } 2 = = 0$ ), ( $\\scriptstyle { \\frac { 0 } { 0 } } 2 = = 1$ ) and for ZIPWITH and SCANL1 the DSL provides lambdas $( + ) , ( - ) , ( \\star ) ,$ MIN, MAX. A description of the semantics of all functions is provided in Appendix F. ", + "bbox": [ + 173, + 492, + 826, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that while the language only allows linear control flow, many of its functions do perform branching and looping internally (e.g., SORT, COUNT, ...). Examples of more sophisticated programs expressible in our DSL, which were inspired by the simplest problems appearing on programming competition websites, are shown in Appendix A. ", + "bbox": [ + 174, + 597, + 825, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 DATA GENERATION ", + "text_level": 1, + "bbox": [ + 176, + 671, + 349, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate a dataset, we enumerate programs in the DSL, heuristically pruning away those with easily detectable issues such as a redundant variable whose value does not affect the program output, or, more generally, existence of a shorter equivalent program (equivalence can be overapproximated by identical behavior on randomly or carefully chosen inputs). To generate valid inputs for a program, we enforce a constraint on the output value bounding integers to some predetermined range, and then propagate these constraints backward through the program to obtain a range of valid values for each input. If one of these ranges is empty, we discard the program. Otherwise, input-output pairs can be generated by picking inputs from the pre-computed valid ranges and executing the program to obtain the output values. The binary attribute vectors are easily computed from the program source codes. ", + "bbox": [ + 173, + 698, + 825, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 MACHINE LEARNING MODEL ", + "text_level": 1, + "bbox": [ + 176, + 840, + 419, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observe how the input-output data in Fig. 1 is informative of the functions appearing in the program: the values in the output are all negative, divisible by 4, they are sorted in decreasing order, and they happen to be multiples of numbers appearing in the input. Our aim is to learn to recognize such patterns in the input-output examples, and to leverage them to predict the presence or absence of individual functions. We employ neural networks to model and learn the mapping from input-output examples to attributes. We can think of these networks as consisting of two parts: ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. an encoder: a differentiable mapping from a set of $M$ input-output examples generated by a single program to a latent real-valued vector, and 2. a decoder: a differentiable mapping from the latent vector representing a set of $M$ inputoutput examples to predictions of the ground truth program’s attributes. ", + "bbox": [ + 212, + 143, + 825, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the encoder we use a simple feed-forward architecture. First, we represent the input and output types (singleton or array) by a one-hot-encoding, and we pad the inputs and outputs to a maximum length $L$ with a special NULL value. Second, each integer in the inputs and in the output is mapped to a learned embedding vector of size $E = 2 0$ . (The range of integers is restricted to a finite range and each embedding is parametrized individually.) Third, for each input-output example separately, we concatenate the embeddings of the input types, the inputs, the output type, and the output into a single (fixed-length) vector, and pass this vector through $H = 3$ hidden layers containing $K = 2 5 6$ sigmoid units each. The third hidden layer thus provides an encoding of each individual input-output example. Finally, for input-output examples in a set generated from the same program, we pool these representations together by simple arithmetic averaging. See Appendix C for more details. ", + "bbox": [ + 174, + 214, + 825, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The advantage of this encoder lies in its simplicity, and we found it reasonably easy to train. A disadvantage is that it requires an upper bound $L$ on the length of arrays appearing in the input and output. We confirmed that the chosen encoder architecture is sensible in that it performs empirically at least as well as an RNN encoder, a natural baseline, which may however be more difficult to train. ", + "bbox": [ + 174, + 361, + 825, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DeepCoder learns to predict presence or absence of individual functions of the DSL. We shall see this can already be exploited by various search techniques to large computational gains. We use a decoder that pre-multiplies the encoding of input-output examples by a learned $C \\times K$ matrix, where $C = 3 4$ is the number of functions in our DSL (higher-order functions and lambdas are predicted independently), and treats the resulting $C$ numbers as log-unnormalized probabilities (logits) of each function appearing in the source code. Fig. 2 shows the predictions a trained neural network made from 5 input-output examples for the program shown in Fig. 1. ", + "bbox": [ + 173, + 422, + 825, + 522 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3c080b2a8813d24f061fe69fe44372e58ba574c230d52185f026a848a7215ca.jpg", + "image_caption": [ + "Figure 2: Neural network predicts the probability of each function appearing in the source code. " + ], + "image_footnote": [], + "bbox": [ + 181, + 539, + 816, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4 SEARCH ", + "text_level": 1, + "bbox": [ + 174, + 642, + 272, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the central ideas of this work is to use a neural network to guide the search for a program consistent with a set of input-output examples instead of directly predicting the entire source code. This section briefly describes the search techniques and how they integrate the predicted attributes. ", + "bbox": [ + 174, + 669, + 825, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depth-first search (DFS). We use an optimized version of DFS to search over programs with a given maximum length $T$ (see Appendix D for details). When the search procedure extends a partial program by a new function, it has to try the functions in the DSL in some order. At this point DFS can opt to consider the functions as ordered by their predicted probabilities from the neural network. ", + "bbox": [ + 174, + 724, + 825, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "“Sort and add” enumeration. A stronger way of utilizing the predicted probabilities of functions in an enumerative search procedure is to use a Sort and add scheme, which maintains a set of active functions and performs DFS with the active function set only. Whenever the search fails, the next most probable function (or several) are added to the active set and the search restarts with this larger active set. Note that this scheme has the deficiency of potentially re-exploring some parts of the search space several times, which could be avoided by a more sophisticated search procedure. ", + "bbox": [ + 174, + 796, + 825, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sketch. Sketch (Solar-Lezama, 2008) is a successful SMT-based program synthesis tool from the programming languages research community. While its main use case is to synthesize programs by filling in “holes” in incomplete source code so as to match specified requirements, it is flexible enough for our use case as well. The function in each step and its arguments can be treated as the “holes”, and the requirement to be satisfied is consistency with the provided set of input-output examples. Sketch can utilize the neural network predictions in a Sort and add scheme as described above, as the possibilities for each function hole can be restricted to the current active set. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda ^ { 2 }$ . $\\lambda ^ { 2 }$ (Feser et al., 2015) is a program synthesis tool from the programming languages community that combines enumerative search with deduction to prune the search space. It is designed to infer small functional programs for data structure manipulation from input-output examples, by combining functions from a provided library. $\\lambda ^ { 2 }$ can be used in our framework using a Sort and add scheme as described above by choosing the library of functions according to the neural network predictions. ", + "bbox": [ + 174, + 188, + 825, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5 TRAINING LOSS FUNCTION ", + "text_level": 1, + "bbox": [ + 176, + 275, + 405, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the negative cross entropy loss to train the neural network described in Sect. 4.3, so that its predictions about each function can be interpreted as marginal probabilities. The LIPS framework dictates learning $q ( \\pmb { a } \\mid \\mathcal { E } )$ , the joint distribution of all attributes $\\textbf { \\em a }$ given the input-output examples, and it is not clear a priori how much DeepCoder loses by ignoring correlations between functions. However, under the simplifying assumption that the runtime of searching for a program of length $T$ with $C$ functions made available to a search routine is proportional to $\\check { C } ^ { T }$ , the following result for Sort and add procedures shows that their runtime can be optimized using marginal probabilities. ", + "bbox": [ + 173, + 301, + 825, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 1. For any fixed program length $T$ , the expected total runtime of a Sort and add search scheme can be upper bounded by a quantity that is minimized by adding the functions in the order of decreasing true marginal probabilities. ", + "bbox": [ + 174, + 402, + 821, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. Predicting source code functions from input-output examples can be seen as a multi-label classification problem, where each set of input-output examples is associated with a set of relevant labels (functions appearing in the ground truth source code). Dembczynski et al. (2010) showed that in multi-label classification under a so-called Rank loss, it is Bayes optimal to rank the labels according to their marginal probabilities. If the runtime of search with $C$ functions is proportional to $C ^ { T }$ , the total runtime of a Sort and add procedure can be monotonically transformed so that it is upper bounded by this Rank loss. See Appendix E for more details. □ ", + "bbox": [ + 174, + 460, + 825, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 578, + 326, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we report results from two categories of experiments. Our main experiments (Sect. 5.1) show that the LIPS framework can lead to significant performance gains in solving IPS by demonstrating such gains with DeepCoder. In Sect. 5.2 we illustrate the robustness of the method by demonstrating a strong kind of generalization ability across programs of different lengths. ", + "bbox": [ + 174, + 609, + 825, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 DEEPCODER COMPARED TO BASELINES", + "text_level": 1, + "bbox": [ + 176, + 681, + 493, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We trained a neural network as described in Sect. 4.3 to predict used functions from input-output examples and constructed a test set of $P = 5 0 0$ programs, guaranteed to be semantically disjoint from all programs on which the neural network was trained (similarly to the equivalence check described in Sect. 4.2, we have ensured that all test programs behave differently from all programs used during training on at least one input). For each test program we generated $M = 5$ input-output examples involving integers of magnitudes up to 256, passed the examples to the trained neural network, and fed the obtained predictions to the search procedures from Sect. 4.4. We also considered a RNN-based decoder generating programs using beam search (see Sect. 5.3 for details). To evaluate DeepCoder, we then recorded the time the search procedures needed to find a program consistent with the $M$ input-output examples. As a baseline, we also ran all search procedures using a simple prior as function probabilities, computed from their global incidence in the program corpus. ", + "bbox": [ + 174, + 708, + 825, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first, smaller-scale experiment (program search space size $\\sim 2 \\times 1 0 ^ { 6 }$ ) we trained the neural network on programs of length $T = 3$ , and the test programs were of the same length. Table 1 shows the per-task timeout required such that a solution could be found for given proportions of the test tasks (in time less than or equal to the timeout). For example, in a hypothetical test set with 4 tasks and runtimes of 3s, 2s, 1s, 4s, the timeout required to solve $50 \\%$ of tasks would be 2s. More detailed experimental results are discussed in Appendix B. ", + "bbox": [ + 176, + 867, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2b097877aeee4ab1671c81b735bc0ae0412b264777fb72d65ca97b67fdb09e11.jpg", + "table_caption": [ + "Table 1: Search speedups on programs of length $T = 3$ due to using neural network predictions. " + ], + "table_footnote": [], + "table_body": "
Timeout needed to solveDFSEnumeration12SketchBeam
20%40%60%20%40%60%20%40%60%20%40%20%
Baseline41ms126ms314ms80ms:335ms861ms18.9s49.6s 84.2s>103s>103s>103s
DeepCoder2.7ms33ms110ms1.3ms6.1ms27ms0.23s0.52s 13.5s2.13s455s292s
Speedup15.2×3.9×2.9×62.2×54.6×31.5×80.4×94.6×6.2×>467×>2.2×>3.4×
", + "bbox": [ + 173, + 127, + 821, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 226, + 823, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the main experiment, we tackled a large-scale problem of searching for programs consistent with input-output examples generated from programs of length $T = 5$ (search space size on the order of $1 0 ^ { \\overline { { 1 0 } } }$ ), supported by a neural network trained with programs of shorter length $T = 4$ . Here, we only consider $P = 1 0 0$ programs for reasons of computational efficiency, after having verified that this does not significantly affect the results in Table 1. The table in Fig. 3a shows significant speedups for DFS, Sort and add enumeration, and $\\lambda ^ { 2 }$ with Sort and add enumeration, the search techniques capable of solving the search problem in reasonable time frames. Note that Sort and add enumeration without the neural network (using prior probabilities of functions) exceeded the $1 0 ^ { 4 }$ second timeout in two cases, so the relative speedups shown are crude lower bounds. ", + "bbox": [ + 173, + 261, + 825, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7cd9438bdca3d6b3e822af385015998ac105636d940628b2f68f443f81b05e3.jpg", + "image_caption": [ + "Figure 3: Search speedups on programs of length $T = 5$ and influence of length of training programs. " + ], + "image_footnote": [], + "bbox": [ + 171, + 395, + 813, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We hypothesize that the substantially larger performance gains on Sort and add schemes as compared to gains on DFS can be explained by the fact that the choice of attribute function (predicting presence of functions anywhere in the program) and learning objective of the neural network are better matched to the Sort and add schemes. Indeed, a more appropriate attribute function for DFS would be one that is more informative of the functions appearing early in the program, since exploring an incorrect first function is costly with DFS. On the other hand, the discussion in Sect. 4.5 provides theoretical indication that ignoring the correlations between functions is not cataclysmic for Sort and add enumeration, since a Rank loss that upper bounds the Sort and add runtime can still be minimized. ", + "bbox": [ + 173, + 559, + 825, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Appendix G we analyse the performance of the neural networks used in these experiments, by investigating which attributes (program instructions) tend to be difficult to distinguish from each other. ", + "bbox": [ + 176, + 678, + 823, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 GENERALIZATION ACROSS PROGRAM LENGTHS ", + "text_level": 1, + "bbox": [ + 174, + 738, + 542, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the encoder’s generalization ability across programs of different lengths, we trained a network to predict used functions from input-output examples that were generated from programs of length $T _ { \\mathrm { t r a i n } } \\in \\{ 1 , . . . , 4 \\}$ . We then used each of these networks to predict functions on 5 test sets containing input-output examples generated from programs of lengths $T _ { \\mathrm { t e s t } } \\in \\{ 1 , . . . , 5 \\}$ , respectively. The test programs of a given length $T$ were semantically disjoint from all training programs of the same length $T$ and also from all training and test programs of shorter lengths $T ^ { \\prime } < T$ . ", + "bbox": [ + 173, + 762, + 825, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each of the combinations of $T _ { \\mathrm { t r a i n } }$ and $T _ { \\mathrm { t e s t } }$ , Sort and add enumerative search was run both with and without using the neural network’s predictions (in the latter case using prior probabilities) until it solved $2 0 \\%$ of the test set tasks. Fig. 3b shows the relative speedup of the solver having access to predictions from the trained neural networks. These results indicate that the neural networks are able to generalize beyond programs of the same length that they were trained on. This is partly due to the search procedure on top of their predictions, which has the opportunity to correct for the presence of functions that the neural network failed to predict. Note that a sequence-to-sequence model trained on programs of a fixed length could not be expected to exhibit this kind of generalization ability. ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 146 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 ALTERNATIVE MODELS ", + "text_level": 1, + "bbox": [ + 174, + 162, + 375, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Encoder We evaluated replacing the feed-forward architecture encoder (Sect. 4.3) with an RNN, a natural baseline. Using a GRU-based RNN we were able to achieve results almost as good as using the feed-forward architecture, but found the RNN encoder more difficult to train. ", + "bbox": [ + 174, + 188, + 825, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoder We also considered a purely neural network-based approach, where an RNN decoder is trained to predict the entire program token-by-token. We combined this with our feed-forward encoder by initializing the RNN using the pooled final layer of the encoder. We found it substantially more difficult to train an RNN decoder as compared to the independent binary classifiers employed above. Beam search was used to explore likely programs predicted by the RNN, but it only lead to a solution comparable with the other techniques when searching for programs of lengths $T \\le 2$ , where the search space size is very small (on the order of $1 0 ^ { 3 }$ ). Note that using an RNN for both the encoder and decoder corresponds to a standard sequence-to-sequence model. However, we do do not rule out that a more sophisticated RNN decoder or training procedure could be possibly more successful. ", + "bbox": [ + 174, + 246, + 825, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 390, + 344, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Machine Learning for Inductive Program Synthesis. There is relatively little work on using machine learning for programming by example. The most closely related work is that of Menon et al. (2013), in which a hand-coded set of features of input-output examples are used as “clues.” When a clue appears in the input-output examples (e.g., the output is a permutation of the input), it reweights the probabilities of productions in a probabilistic context free grammar by a learned amount. This work shares the idea of learning to guide the search over program space conditional on input-output examples. One difference is in the domains. Menon et al. (2013) operate on short string manipulation programs, where it is arguably easier to hand-code features to recognize patterns in the input-output examples (e.g., if the outputs are always permutations or substrings of the input). Our work shows that there are strong cues in patterns in input-output examples in the domain of numbers and lists. However, the main difference is the scale. Menon et al. (2013) learns from a small (280 examples), manually-constructed dataset, which limits the capacity of the machine learning model that can be trained. Thus, it forces the machine learning component to be relatively simple. Indeed, Menon et al. (2013) use a log-linear model and rely on hand-constructed features. LIPS automatically generates training data, which yields datasets with millions of programs and enables high-capacity deep learning models to be brought to bear on the problem. ", + "bbox": [ + 174, + 421, + 825, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning Representations of Program State. Piech et al. (2015) propose to learn joint embeddings of program states and programs to automatically extend teacher feedback to many similar programs in the MOOC setting. This work is similar in that it considers embedding program states, but the domain is different, and it otherwise specifically focuses on syntactic differences between semantically equivalent programs to provide stylistic feedback. Li et al. (2016) use graph neural networks (GNNs) to predict logical descriptions from program states, focusing on data structure shapes instead of numerical and list data. Such GNNs may be a suitable architecture to encode states appearing when extending our DSL to handle more complex data structures. ", + "bbox": [ + 174, + 659, + 825, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning to Infer. Very recently, Alemi et al. (2016) used neural sequence models in tandem with an automated theorem prover. Similar to our Sort and Add strategy, a neural network component is trained to select premises that the theorem prover can use to prove a theorem. A recent extension (Loos et al., 2017) is similar to our DFS enumeration strategy and uses a neural network to guide the proof search at intermediate steps. The main differences are in the domains, and that they train on an existing corpus of theorems. More broadly, if we view a DSL as defining a model and search as a form of inference algorithm, then there is a large body of work on using discriminatively-trained models to aid inference in generative models. Examples include Dayan et al. (1995); Kingma & Welling (2014); Shotton et al. (2013); Stuhlmuller et al. (2013); Heess et al. (2013); Jampani et al. ¨ (2015). ", + "bbox": [ + 174, + 785, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 DISCUSSION AND FUTURE WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 483, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented a framework for improving IPS systems by using neural networks to translate cues in input-output examples to guidance over where to search in program space. Our empirical results show that for many programs, this technique improves the runtime of a wide range of IPS baselines by 1-3 orders. We have found several problems in real online programming challenges that can be solved with a program in our language, which validates the relevance of the class of problems that we have studied in this work. In sum, this suggests that we have made significant progress towards being able to solve programming competition problems, and the machine learning component plays an important role in making it tractable. ", + "bbox": [ + 174, + 133, + 825, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "There remain some limitations, however. First, the programs we can synthesize are only the simplest problems on programming competition websites and are simpler than most competition problems. Many problems require more complex algorithmic solutions like dynamic programming and search, which are currently beyond our reach. Our chosen DSL currently cannot express solutions to many problems. To do so, it would need to be extended by adding more primitives and allow for more flexibility in program constructs (such as allowing loops). Second, we currently use five input-output examples with relatively large integer values (up to 256 in magnitude), which are probably more informative than typical (smaller) examples. While we remain optimistic about LIPS’s applicability as the DSL becomes more complex and the input-output examples become less informative, it remains to be seen what the magnitude of these effects are as we move towards solving large subsets of programming competition problems. ", + "bbox": [ + 174, + 251, + 825, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We foresee many extensions of DeepCoder. We are most interested in better data generation procedures by using generative models of source code, and to incorporate natural language problem descriptions to lessen the information burden required from input-output examples. In sum, DeepCoder represents a promising direction forward, and we are optimistic about the future prospects of using machine learning to synthesize programs. ", + "bbox": [ + 174, + 411, + 825, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 497, + 326, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to express their gratitude to Rishabh Singh and Jack Feser for their valuable guidance and help on using the Sketch and $\\lambda ^ { 2 }$ program synthesis systems. ", + "bbox": [ + 174, + 520, + 823, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 569, + 285, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex A. Alemi, Franc¸ois Chollet, Geoffrey Irving, Christian Szegedy, and Josef Urban. DeepMath - deep sequence models for premise selection. In Proocedings of the 29th Conference on Advances in Neural Information Processing Systems (NIPS), 2016. \nRudy R Bunel, Alban Desmaison, Pawan K Mudigonda, Pushmeet Kohli, and Philip Torr. Adaptive neural compilation. In Proceedings of the 29th Conference on Advances in Neural Information Processing Systems (NIPS), 2016. \nPeter Dayan, Geoffrey E Hinton, Radford M Neal, and Richard S Zemel. The Helmholtz machine. Neural computation, 7(5):889–904, 1995. \nKrzysztof Dembczynski, Willem Waegeman, Weiwei Cheng, and Eyke H ´ ullermeier. On label de- ¨ pendence and loss minimization in multi-label classification. Machine Learning, 88(1):5–45, 2012. \nKrzysztof J. Dembczynski, Weiwei Cheng, and Eyke Hllermeier. Bayes optimal multilabel classification via probabilistic classifier chains. In Proceedings of the 27th International Conference on Machine Learning (ICML), 2010. \nJohn K. Feser, Swarat Chaudhuri, and Isil Dillig. Synthesizing data structure transformations from input-output examples. In Proceedings of the 36th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), 2015. \nAlexander L. Gaunt, Marc Brockschmidt, Rishabh Singh, Nate Kushman, Pushmeet Kohli, Jonathan Taylor, and Daniel Tarlow. Terpret: A probabilistic programming language for program induction. CoRR, abs/1608.04428, 2016. URL http://arxiv.org/abs/1608.04428. ", + "bbox": [ + 171, + 590, + 828, + 925 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural Turing machines. CoRR, abs/1410.5401, 2014. URL http://arxiv.org/abs/1410.5401. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ et al. Hybrid computing using a neural network with dynamic external memory. Nature, 2016. ", + "bbox": [ + 176, + 140, + 825, + 183 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to transduce with unbounded memory. In Proceedings of the 28th Conference on Advances in Neural Information Processing Systems (NIPS), 2015. ", + "bbox": [ + 176, + 190, + 823, + 233 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sumit Gulwani. Programming by examples: Applications, algorithms, and ambiguity resolution. In Proceedings of the 8th International Joint Conference on Automated Reasoning (IJCAR), 2016. ", + "bbox": [ + 171, + 241, + 823, + 271 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sumit Gulwani, Susmit Jha, Ashish Tiwari, and Ramarathnam Venkatesan. Synthesis of loop-free programs. In Proceedings of the 32nd ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), 2011. ", + "bbox": [ + 176, + 277, + 825, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nicolas Heess, Daniel Tarlow, and John Winn. Learning to pass expectation propagation messages. In Proceedings of the 26th Conference on Advances in Neural Information Processing Systems (NIPS), 2013. ", + "bbox": [ + 174, + 328, + 825, + 371 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Varun Jampani, Sebastian Nowozin, Matthew Loper, and Peter V Gehler. The informed sampler: A discriminative approach to Bayesian inference in generative computer vision models. Computer Vision and Image Understanding, 136:32–44, 2015. ", + "bbox": [ + 174, + 378, + 823, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In Proceedings of the 28th Conference on Advances in Neural Information Processing Systems (NIPS), 2015. ", + "bbox": [ + 174, + 429, + 823, + 473 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In Proceedings of the 4th International Conference on Learning Representations, 2016. ", + "bbox": [ + 171, + 481, + 825, + 510 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Diederik P Kingma and Max Welling. Stochastic gradient VB and the variational auto-encoder. In Proceedings of the 2nd International Conference on Learning Representations (ICLR), 2014. ", + "bbox": [ + 173, + 517, + 825, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. In Proceedings of the 4th International Conference on Learning Representations 2016, 2015. ", + "bbox": [ + 173, + 554, + 823, + 584 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard S. Zemel. Gated graph sequence neural networks. In Proceedings of the 4th International Conference on Learning Representations (ICLR), 2016. ", + "bbox": [ + 176, + 590, + 825, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Wang Ling, Edward Grefenstette, Karl Moritz Hermann, Toma´s Ko ˇ cisk ˇ y, Andrew Senior, Fumin ´ Wang, and Phil Blunsom. Latent predictor networks for code generation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, 2016. ", + "bbox": [ + 176, + 641, + 823, + 685 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sarah M. Loos, Geoffrey Irving, Christian Szegedy, and Cezary Kaliszyk. Deep network guided proof search. CoRR, abs/1701.06972, 2017. URL http://arxiv.org/abs/1701.06972. ", + "bbox": [ + 171, + 691, + 825, + 722 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aditya Krishna Menon, Omer Tamuz, Sumit Gulwani, Butler W Lampson, and Adam Kalai. A machine learning framework for programming by example. In Proceedings of the International Conference on Machine Learning (ICML), 2013. ", + "bbox": [ + 176, + 728, + 826, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. Neural programmer: Inducing latent programs with gradient descent. In Proceedings of the 4th International Conference on Learning Representations (ICLR), 2016. ", + "bbox": [ + 174, + 780, + 823, + 823 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chris Piech, Jonathan Huang, Andy Nguyen, Mike Phulsuksombati, Mehran Sahami, and Leonidas J. Guibas. Learning program embeddings to propagate feedback on student code. In Proceedings of the 32nd International Conference on Machine Learning (ICML), 2015. ", + "bbox": [ + 176, + 830, + 825, + 873 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Oleksandr Polozov and Sumit Gulwani. FlashMeta: a framework for inductive program synthesis. In Proceedings of the International Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), 2015. ", + "bbox": [ + 174, + 881, + 826, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Scott E. Reed and Nando de Freitas. Neural programmer-interpreters. In Proceedings of the 4th International Conference on Learning Representations (ICLR), 2016. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sebastian Riedel, Matko Bosnjak, and Tim Rocktaschel. Programming with a differentiable forth ¨ interpreter. CoRR, abs/1605.06640, 2016. URL http://arxiv.org/abs/1605.06640. ", + "bbox": [ + 173, + 141, + 823, + 171 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Eric Schkufza, Rahul Sharma, and Alex Aiken. Stochastic program optimization. Commununications of the ACM, 59(2):114–122, 2016. ", + "bbox": [ + 171, + 178, + 823, + 208 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jamie Shotton, Toby Sharp, Alex Kipman, Andrew Fitzgibbon, Mark Finocchio, Andrew Blake, Mat Cook, and Richard Moore. Real-time human pose recognition in parts from single depth images. Communications of the ACM, 56(1):116–124, 2013. ", + "bbox": [ + 176, + 215, + 823, + 260 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rishabh Singh and Sumit Gulwani. Predicting a correct program in programming by example. In Proceedings of the 27th Conference on Computer Aided Verification (CAV), 2015. ", + "bbox": [ + 169, + 267, + 825, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Armando Solar-Lezama. Program Synthesis By Sketching. PhD thesis, EECS Dept., UC Berkeley, 2008. ", + "bbox": [ + 171, + 305, + 825, + 334 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Andreas Stuhlmuller, Jessica Taylor, and Noah D. Goodman. Learning stochastic inverses. In ¨ Proceedings of the 26th Conference on Advances in Neural Information Processing Systems (NIPS), 2013. ", + "bbox": [ + 174, + 343, + 825, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In Proceedings of the 28th Conference on Advances in Neural Information Processing Systems (NIPS), 2015. ", + "bbox": [ + 173, + 395, + 825, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015. ", + "bbox": [ + 171, + 446, + 823, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus. Learning simple algorithms from examples. In Proceedings of the 33nd International Conference on Machine Learning (ICML), 2016. ", + "bbox": [ + 174, + 484, + 825, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A EXAMPLE PROGRAMS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 393, + 118 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This section shows example programs in our Domain Specific Language (DSL), together with inputoutput examples and short descriptions. These programs have been inspired by simple tasks appearing on real programming competition websites, and are meant to illustrate the expressive power of our DSL. ", + "bbox": [ + 174, + 161, + 826, + 217 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Program 0: ", + "text_level": 1, + "bbox": [ + 183, + 236, + 266, + 247 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\mathrm { k } $ int b ← [int] c ← SORT b d ← TAKE k c e ← SUM d ", + "bbox": [ + 183, + 251, + 287, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Input-output example: \nInput: \n2, [3 5 4 7 5] \nOutput: \n[7] ", + "bbox": [ + 357, + 234, + 516, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Description: ", + "bbox": [ + 532, + 234, + 616, + 247 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A new shop near you is selling $n$ paintings. You have $k \\ < \\ n$ friends and you would like to buy each of your friends a painting from the shop. Return the minimal amount of money you will need to spend. ", + "bbox": [ + 532, + 247, + 823, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Program 1: ", + "text_level": 1, + "bbox": [ + 183, + 343, + 264, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\bar { w } $ [int] \nt ← [int] \nc ← MAP (\\*3) w \nd ← ZIPWITH (+) c t \ne ← MAXIMUM d \nInput-output example: \nInput: \n[6 2 4 7 9], \n[5 3 6 1 0] \nOutput: \n27 ", + "bbox": [ + 181, + 357, + 344, + 424 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 357, + 340, + 517, + 424 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Description: ", + "bbox": [ + 532, + 334, + 616, + 347 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In soccer leagues, match winners are awarded 3 points, losers 0 points, and both teams get 1 point in the case of a tie. Compute the number of points awarded to the winner of a league given two arrays $w , t$ of the same length, where $w [ i ]$ (resp. $t [ i ] ,$ ) is the number of times team $i$ won (resp. tied). ", + "bbox": [ + 532, + 347, + 823, + 444 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Program 2: ", + "text_level": 1, + "bbox": [ + 184, + 462, + 264, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "a ← [int] \nb ← [int] \nc ← ZIPWITH (-) b a $\\mathsf { d } \\gets \\mathbf { C o u v r } \\mathsf { T }$ $( > 0 )$ ) c ", + "bbox": [ + 181, + 477, + 346, + 531 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Input-output example: ", + "text_level": 1, + "bbox": [ + 357, + 462, + 517, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Input: \n[6 2 4 7 9], \n[5 3 2 1 0] \nOutput: \n4 ", + "bbox": [ + 357, + 477, + 472, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Description: ", + "bbox": [ + 532, + 462, + 614, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alice and Bob are comparing their results in a recent exam. Given their marks per question as two arrays $a$ and $b$ , count on how many questions Alice got more points than Bob. ", + "bbox": [ + 532, + 476, + 823, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Program 3: ", + "text_level": 1, + "bbox": [ + 184, + 560, + 264, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "h ← [int] \nb ← SCANL1 MIN h \nc ← ZIPWITH (-) h b $\\mathrm { ~ d ~ } $ FILTER $( > 0 )$ ) c \n$\\mathrm { ~ e ~ } { \\gets } \\mathbf { S } \\mathbf { U } \\mathbf { M } \\mathrm { ~ d ~ }$ \nInput-output example: \nInput: \n[8 5 7 2 5] \nOutput: \n5 ", + "bbox": [ + 183, + 574, + 344, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 357, + 560, + 519, + 632 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Description: ", + "bbox": [ + 531, + 559, + 614, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Perditia is very peculiar about her garden and wants that the trees standing in a row are all of non-increasing heights. Given the tree heights in centimeters in order of the row as an array h, compute how many centimeters she needs to trim the trees in total. ", + "bbox": [ + 532, + 574, + 821, + 655 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Program 4: ", + "text_level": 1, + "bbox": [ + 184, + 679, + 264, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "x ← [int] \ny ← [int] \n$\\mathbf { \\boldsymbol { \\mathrm { c } } } \\gets \\mathbf { \\boldsymbol { S } } \\mathbf { \\boldsymbol { O } } \\mathbf { R } \\mathbf { \\boldsymbol { T } } \\ \\mathbf { \\boldsymbol { x } }$ \nd ← SORT y \ne ← REVERSE d \nf ← ZIPWITH (\\*) d e \n$\\mathsf { g } \\gets \\mathsf { S U M } \\mathrm { ~ f ~ }$ \nInput-output example: \nInput: \n[7 3 8 2 5], \n[2 8 9 1 3] \nOutput: \n79 ", + "bbox": [ + 181, + 693, + 344, + 789 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 357, + 679, + 517, + 761 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Description: ", + "bbox": [ + 532, + 671, + 616, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Xavier and Yasmine are laying sticks to form non-overlapping rectangles on the ground. They both have fixed sets of pairs of sticks of certain lengths (represented as arrays x and y of numbers). Xavier only lays sticks parallel to the x axis, and Yasmine lays sticks only parallel to y axis. Compute the area their rectangles will cover at least. ", + "bbox": [ + 531, + 684, + 823, + 795 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Program 5: \n$\\ a \\gets$ [int] \n$\\ u \\text b \\gets$ REVERSE a \n$\\subset $ ZIPWITH MIN a b \nInput-output example: \nInput: \n[3 7 5 2 8] \nOutput: \n[3 2 5 2 3] ", + "bbox": [ + 181, + 834, + 346, + 890 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 357, + 833, + 517, + 904 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Description: ", + "text_level": 1, + "bbox": [ + 532, + 813, + 616, + 825 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A sequence called Billy is looking into the mirror, wondering how much weight it could lose by replacing any of its elements by their mirror images. Given a description of Billy as an array $b$ of length $n$ , return an array $c$ of minimal sum where each element $c [ i ]$ is either $b [ i ]$ or its mirror image $b [ n - i - 1 ]$ . ", + "bbox": [ + 532, + 825, + 821, + 924 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c0b9fa01cfad5a4a38406952d046a34d3bfd424df86c9c92982450ad23a4740b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Program 6: IO example: t←[int] Input: p←[int] [4 8 11 2], C←MAP(-1)t [2341] d←MAP(-1)p Output: e ←ZIPWITH(+)c d 1 f←MINIMUMeDescription: Umberto has a large collection of ties and match- ing pocket squares—too large, his wife says—and he needs to sell one pair. Given their values as arrays t and p,assuming that he sells the cheapest pair, and selling costs 2,how much will he lose from the sale?
Program 7: IO example: s←[int] Input: p←[int] [4723], C ← SCANL1(+)p [2131] d ←ZIPWITH(*)S C Output: e←SUMd 62Description: Zack always promised his n friends to buy them candy, but never did. Now he won the lottery and counts how often and how much candy he promised to his friends,obtaining arrays p (num- ber of promises) and s (number of promised sweets). He announces that to repay them,he will buy s[1]+s[2]+...+s[n] pieces of candy for the firstp[1]days,thens[2]+s[3]+...+s[n]for p[2] days,and so on,until he has fulfilled all promises. How much candy will he buy in total?
Program 8: IO example: s←[int] Input: b ←REVERSE S [12457] C ←ZIPWITH(-)b s Output: d ←FILTER(>O)C 9 e←SuMdDescription: Vivian loves rearranging things.Most of all,when she sees a row of heaps, she wants to make sure that each heap has more items than the one to its left. She is also obsessed with efficiency, so always moves the least possible number of items.Her dad really dislikes if she changes the size of heaps,so she only moves single items between them,making sure that the set of sizes of the heaps is the same as at the start; they are only in a different order. When you come in, you see heaps of sizes (of course,sizes strictly monotonically increasing)s[O],s[1],... s[n].What is the maximal number of items that Vivian could have moved?
", + "bbox": [ + 174, + 53, + 828, + 584 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6d29878119ef08e20bad2e228bb9a4b1140e7a88251e4c13cbe46fb4f2478484.jpg", + "image_caption": [ + "Fig. 4 shows the predictions made by a neural network trained on programs of length $T = 4$ that were ensured to be semantically disjoint from all 9 example programs shown in this section. For each task, the neural network was provided with 5 input-output examples. ", + "Figure 4: Predictions of a neural network on the 9 example programs described in this section. Numbers in squares would ideally be close to 1 (function is present in the ground truth source code), whereas all other numbers should ideally be close to 0 (function is not needed). " + ], + "image_footnote": [], + "bbox": [ + 178, + 656, + 818, + 782 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B EXPERIMENTAL RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 864, + 423, + 881 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results presented in Sect. 5.1 showcased the computational speedups obtained from the LIPS framework (using DeepCoder), as opposed to solving each program synthesis problem with only the information about global incidence of functions in source code available. For completeness, here we show plots of raw computation times of each search procedure to solve a given number of problems. ", + "bbox": [ + 174, + 895, + 825, + 924 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 5 shows the computation times of DFS, of Enumerative search with a Sort and add scheme, of the $\\lambda ^ { 2 }$ and Sketch solvers with a Sort and add scheme, and of Beam search, when searching for a program consistent with input-output examples generated from $P = 5 0 0$ different test programs of length $T = 3$ . As discussed in Sect. 5.1, these test programs were ensured to be semantically disjoint from all programs used to train the neural networks, as well as from all programs of shorter length (as discussed in Sect. 4.2). ", + "bbox": [ + 173, + 138, + 825, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/69617f5d9c5e8b7b5a7be3c5b88b0456a2865590e84e41413e9067abe2f41f17.jpg", + "image_caption": [ + "Figure 5: Number of test problems solved versus computation time. " + ], + "image_footnote": [], + "bbox": [ + 179, + 241, + 812, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The “steps” in the results for Beam search are due to our search strategy, which doubles the size of the considered beam until reaching the timeout (of 1000 seconds) and thus steps occur whenever the search for a beam of size $2 ^ { k }$ is finished. For $\\dot { \\lambda ^ { 2 } }$ , we observed that no solution for a given set of allowed functions was ever found after about 5 seconds (on the benchmark machines), but that $\\lambda ^ { 2 }$ continued to search. Hence, we introduced a hard timeout after 6 seconds for all but the last iterations of our Sort and add scheme. ", + "bbox": [ + 173, + 419, + 826, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 6 shows the computation times of DFS, Enumerative search with a Sort and add scheme, and $\\lambda ^ { 2 }$ with a Sort and add scheme when searching for programs consistent with input-output examples generated from $P = 1 0 0$ different test programs of length $T = 5$ . The neural network was trained on programs of length $T = 4$ . ", + "bbox": [ + 173, + 508, + 825, + 565 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8d83b6177f99db604465ccbe078168c96645752256eb9f14da14c4754aa158fd.jpg", + "image_caption": [ + "Figure 6: Number of test problems solved versus computation time. " + ], + "image_footnote": [], + "bbox": [ + 181, + 584, + 813, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C THE NEURAL NETWORK ", + "text_level": 1, + "bbox": [ + 176, + 753, + 415, + 770 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As briefly described in Sect. 4.3, we used the following simple feed-forward architecture encoder: ", + "bbox": [ + 178, + 787, + 812, + 803 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "• For each input-output example in the set generated from a single ground truth program: ", + "bbox": [ + 205, + 814, + 805, + 830 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "– Pad arrays appearing in the inputs and in the output to a maximum length $L = 2 0$ with a special NULL value. \n– Represent the type (singleton integer or integer array) of each input and of the output using a one-hot-encoding vector. Embed each integer in the valid integer range $( - 2 5 6$ to 255) using a learned embedding into $E = 2 0$ dimensional space. Also learn an embedding for the padding NULL value. ", + "bbox": [ + 248, + 837, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "– Concatenate the representations of the input types, the embeddings of integers in the inputs, the representation of the output type, and the embeddings of integers in the output into a single (fixed-length) vector. – Pass this vector through $H = 3$ hidden layers containing $K = 2 5 6$ sigmoid units each. ", + "bbox": [ + 248, + 103, + 825, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "• Pool the last hidden layer encodings of each input-output example together by simple arithmetic averaging. ", + "bbox": [ + 212, + 169, + 825, + 196 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Fig. 7 shows a schematic drawing of this encoder architecture, together with the decoder that performs independent binary classification for each function in the DSL, indicating whether or not it appears in the ground truth source code. ", + "bbox": [ + 173, + 209, + 826, + 252 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/1b4ae4043ae84d4ac119b02713a27cb85173d4a2d0a85116dc32199039ab2623.jpg", + "image_caption": [ + "Figure 7: Schematic representation of our feed-forward encoder, and the decoder. " + ], + "image_footnote": [], + "bbox": [ + 178, + 266, + 825, + 545 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "While DeepCoder learns to embed integers into a $E = 2 0$ dimensional space, we built the system up gradually, starting with a $E = 2$ dimensional space and only training on programs of length $T = 1$ . Such a small scale setting allowed easier investigation of the workings of the neural network, and indeed Fig. 8 below shows a learned embedding of integers in $\\mathbb { R } ^ { 2 }$ . The figure demonstrates that the network has learnt the concepts of number magnitude, sign (positive or negative) and evenness, presumably due to FILTER $( > 0 )$ , FILTER $( < 0 )$ , FILTER ( $\\scriptstyle { \\frac { 0 } { 0 } } 2 = = 0$ ) and FILTER ( $\\scriptstyle { \\frac { 0 } { 0 } } 2 = = 1$ ) all being among the programs on which the network was trained. ", + "bbox": [ + 173, + 590, + 826, + 688 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D DEPTH-FIRST SEARCH ", + "text_level": 1, + "bbox": [ + 176, + 709, + 400, + 726 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We use an optimized $\\mathrm { C } { + } { + }$ implementation of depth-first search (DFS) to search over programs with a given maximum length $T$ . In depth-first search, we start by choosing the first function (and its arguments) of a potential solution program, and then recursively consider all ways of filling in the rest of the program (up to length $T$ ), before moving on to a next choice of first instruction (if a solution has not yet been found). ", + "bbox": [ + 174, + 742, + 825, + 811 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A program is considered a solution if it is consistent with all $M = 5$ provided input-output examples. Note that this requires evaluating all candidate programs on the $M$ inputs and checking the results for equality with the provided $M$ respective outputs. Our implementation of DFS exploits the sequential structure of programs in our DSL by caching the results of evaluating all prefixes of the currently considered program on the example inputs, thus allowing efficient reuse of computation between candidate programs with common prefixes. ", + "bbox": [ + 174, + 818, + 825, + 902 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This allows us to explore the search space at roughly the speed of $\\sim 3 \\times 1 0 ^ { 6 }$ programs per second. ", + "bbox": [ + 171, + 909, + 818, + 924 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/94148f752106bd924a4f80c9c825128a3efe588caf08184e8765377bb5e169db.jpg", + "image_caption": [ + "Figure 8: A learned embedding of integers $\\{ - 2 5 6 , - 2 5 5 , \\ldots , - 1 , 0 , 1 , \\ldots , 2 5 5 \\}$ in $\\mathbb { R } ^ { 2 }$ . The color intensity corresponds to the magnitude of the embedded integer. " + ], + "image_footnote": [], + "bbox": [ + 178, + 106, + 818, + 344 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "When the search procedure extends a partial program by a new function, it has to try the functions in the DSL in some order. At this point DFS can opt to consider the functions as ordered by their predicted probabilities from the neural network. The probability of a function consisting of a higherorder function and a lambda is taken to be the minimum of the probabilities of the two constituent functions. ", + "bbox": [ + 173, + 416, + 826, + 487 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "E TRAINING LOSS FUNCTION ", + "text_level": 1, + "bbox": [ + 176, + 507, + 436, + 523 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In Sect. 4.5 we outlined a justification for using marginal probabilities of individual functions as a sensible intermediate representation to provide a solver employing a Sort and add scheme (we considered Enumerative search and the Sketch solver with this scheme). Here we provide a more detailed discussion. ", + "bbox": [ + 174, + 540, + 825, + 595 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Predicting program components from input-output examples can be cast as a multilabel classification problem, where each instance (set of input-output examples) is associated with a set of relevant labels (functions appearing in the code that generated the examples). We denote the number of labels (functions) by $C$ , and note that throughout this work $C = 3 4$ . ", + "bbox": [ + 174, + 603, + 825, + 660 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "When the task is to predict a subset of labels $\\mathbf { y } \\in \\{ 0 , 1 \\} ^ { C }$ , different loss functions can be employed to measure the prediction error of a classifier $\\mathbf { h } ( \\mathbf { x } )$ or ranking function $\\mathbf { f } \\left( \\mathbf { x } \\right)$ . Dembczynski et al. (2010) discuss the following three loss functions: ", + "bbox": [ + 176, + 665, + 825, + 708 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "• Hamming loss counts the number of labels that are predicted incorrectly by a classifier h: ", + "bbox": [ + 215, + 719, + 818, + 734 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/fa6196f9fbf8394f13f6328262bfa423b74b8377f605daf309401f0b4973e80a.jpg", + "text": "$$\nL _ { H } ( \\mathbf { y } , \\mathbf { h } ( \\mathbf { x } ) ) = \\sum _ { c = 1 } ^ { C } \\mathbb { 1 } _ { \\{ y _ { c } \\neq h _ { c } ( \\mathbf { x } ) \\} }\n$$", + "text_format": "latex", + "bbox": [ + 419, + 741, + 635, + 786 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "• Rank loss counts the number of label pairs violating the condition that relevant labels are ranked higher than irrelevant ones by a scoring function f: ", + "bbox": [ + 215, + 794, + 823, + 823 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/6c4a10361237269ae5e855a79a9b0252c35e0fc76a09bc0388230ba2537710ec.jpg", + "text": "$$\nL _ { r } ( \\mathbf { y } , \\mathbf { f } ( \\mathbf { x } ) ) = \\sum _ { ( i , j ) : y _ { i } = 1 , y _ { j } = 0 } ^ { C } \\mathbb { 1 } _ { \\{ f _ { i } < f _ { j } \\} }\n$$", + "text_format": "latex", + "bbox": [ + 400, + 829, + 655, + 876 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "• Subset Zero-One loss indicates whether all labels have been correctly predicted by $\\mathbf { h }$ : ", + "bbox": [ + 215, + 886, + 789, + 901 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/63f6d63361921ea6ea8466912257b8b5227d5af74f176ac36e92a36e47107a3e.jpg", + "text": "$$\nL _ { s } ( \\mathbf { y } , \\mathbf { h } ( \\mathbf { x } ) ) = \\mathbb { 1 } _ { \\{ \\mathbf { y } \\neq \\mathbf { h } ( \\mathbf { x } ) \\} }\n$$", + "text_format": "latex", + "bbox": [ + 441, + 907, + 614, + 926 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Dembczynski et al. (2010) proved that Bayes optimal decisions under the Hamming and Rank loss functions, i.e., decisions minimizing the expected loss under these loss functions, can be computed from marginal probabilities $p _ { c } ( y _ { c } | \\mathbf { x } )$ . This suggests that: ", + "bbox": [ + 174, + 103, + 825, + 147 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• Multilabel classification under these two loss functions may not benefit from considering dependencies between the labels. \n• ”Instead of minimizing the Rank loss directly, one can simply use any approach for single label prediction that properly estimates the marginal probabilities.” (Dembczynski et al., ´ 2012) ", + "bbox": [ + 215, + 156, + 825, + 231 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Training the neural network with the negative cross entropy loss function as the training objective is precisely a method for properly estimating the marginal probabilities of labels (functions appearing in source code). It is thus a sensible step in preparation for making predictions under a Rank loss. ", + "bbox": [ + 178, + 242, + 825, + 285 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "It remains to discuss the relationship between the Rank loss and the actual quantity we care about, which is the total runtime of a Sort and add search procedure. Recall the simplifying assumption that the runtime of searching for a program of length $T$ with $C$ functions made available to the search is proportional to $C ^ { T }$ , and consider a Sort and add search for a program of length $T$ , where the size of the active set is increased by 1 whenever the search fails. Starting with an active set of size 1, the total time until a solution is found can be upper bounded by ", + "bbox": [ + 173, + 291, + 825, + 375 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/385208351763d143575274950b7c6359d7104974069d4a34eedcf96bafee253d.jpg", + "text": "$$\n1 ^ { T } + 2 ^ { T } + \\cdot \\cdot \\cdot + C _ { A } ^ { T } \\leq C _ { A } ^ { T + 1 } \\leq C C _ { A } ^ { T }\n$$", + "text_format": "latex", + "bbox": [ + 367, + 380, + 629, + 398 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where $C _ { A }$ is the size of the active set when the search finally succeeds (i.e., when the active set finally contains all necessary functions for a solution to exist). Hence the total runtime of a Sort and add search can be upper bounded by a quantity that is proportional to $C _ { A } ^ { T }$ . ", + "bbox": [ + 173, + 402, + 823, + 446 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Now fix a valid program solution $P$ that requires $C _ { P }$ functions, and let $\\mathbf { y } _ { P } \\in \\{ 0 , 1 \\} ^ { C }$ be the indicator vector of functions used by $P$ . Let $D : = C _ { A } - C _ { P }$ be the number of redundant operations added into the active set until all operations from $P$ have been added. ", + "bbox": [ + 176, + 453, + 821, + 496 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Example 1. Suppose the labels, as sorted by decreasing predicted marginal probabilities $\\mathbf { f } \\left( \\mathbf { x } \\right)$ , are as follows: ", + "bbox": [ + 173, + 500, + 823, + 525 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/b6763a465e09eaf5ed0f8ceb822ac7f8dfd443e282b7c5de3bb72a65706d7f1f.jpg", + "text": "$$\n\\begin{array} { r l r l r } { 1 \\mathrm { ~ 1 ~ 1 ~ 1 ~ } } & { { } } & { 1 } & \\qquad 1 \\mathrm ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 ~ 0 0 ~ 0 ~ 0 ~ 0 ~ 0 0 \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 282, + 526, + 715, + 541 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Then the solution $P$ contains $C _ { P } = 6$ functions, but the active set needs to grow to size $C _ { A } = 1 1$ to include all of them, adding $D = 5$ redundant functions along the way. Note that the rank loss of the predictions $\\mathbf { f } \\left( \\mathbf { x } \\right)$ is $L _ { r } ( \\mathbf { y } _ { P } , \\mathbf { f } ( \\mathbf { x } ) ) = 2 + 5 = 7$ , as it double counts the two redundant functions which are scored higher than two relevant labels. ", + "bbox": [ + 173, + 545, + 826, + 601 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Noting that in general $L _ { r } ( \\mathbf { y } _ { P } , \\mathbf { f } ( \\mathbf { x } ) ) \\ge D$ , the previous upper bound on the runtime of Sort and add can be further upper bounded as follows: ", + "bbox": [ + 173, + 611, + 825, + 640 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/2ccda6d9ede732a7742ffa9a2afdab450031f1151b482106f9279b1f97de67ef.jpg", + "text": "$$\nC _ { A } ^ { T } = ( C _ { P } + D ) ^ { T } \\leq \\mathrm { c o n s t } + \\mathrm { c o n s t } \\times D ^ { T } \\leq \\mathrm { c o n s t } + \\mathrm { c o n s t } \\times L _ { r } ( \\mathbf { y } _ { P } , \\mathbf { f } ( \\mathbf { x } ) ) ^ { T }\n$$", + "text_format": "latex", + "bbox": [ + 243, + 643, + 754, + 664 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Hence we see that for a constant value of $T$ , this upper bound can be minimized by optimizing the Rank loss of the predictions $\\mathbf { f } \\left( \\mathbf { x } \\right)$ . Note also that $\\bar { L _ { r } } ( { \\bf y } _ { P } , { \\bf f } ( { \\bf x } ) ) = 0$ would imply $D = 0$ , in which case $C _ { A } = C _ { P }$ . ", + "bbox": [ + 173, + 667, + 825, + 710 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "F DOMAIN SPECIFIC LANGUAGE OF DEEPCODER", + "text_level": 1, + "bbox": [ + 174, + 729, + 599, + 746 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Here we provide a description of the semantics of our DSL from Sect. 4.1, both in English and as a Python implementation. Throughout, NULL is a special value that can be set e.g. to an integer outside the working integer range. ", + "bbox": [ + 174, + 760, + 825, + 804 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "First-order functions: ", + "bbox": [ + 174, + 810, + 313, + 824 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• HEAD :: [int] -> int lambda xs: xs[0] if len $( \\mathbf { \\nabla } _ { \\mathbf { X } } \\mathbf { S } \\mathbf { \\nabla } ) > 0$ else Null Given an array, returns its first element (or NULL if the array is empty). \n• LAST :: [int] $- >$ int lambda xs: xs[-1] if len $( \\mathbf { \\nabla } _ { \\mathbf { X } } \\mathbf { s } ) > 0$ else Null Given an array, returns its last element (or NULL if the array is empty). \n• TAKE :: int $- >$ [int] -> int lambda n, xs: xs[:n] Given an integer n and array xs, returns the array truncated after the $\\boldsymbol { \\mathrm { n } }$ -th element. (If the length of xs was no larger than n in the first place, it is returned without modification.) \n• DROP :: int $- >$ [int] $- >$ int lambda n, xs: xs[n:] Given an integer n and array $_ { \\textrm { x S } }$ , returns the array with the first n elements dropped. (If the length of xs was no larger than n in the first place, an empty array is returned.) \n• ACCESS :: int $- >$ [int] $- >$ int lambda n, xs: xs[n] if $\\mathrm { n } > = 0$ and len $( \\mathsf { x s } ) > \\mathsf { n }$ else Null Given an integer n and array $_ { \\textrm { x S } }$ , returns the $( \\mathrm { n } { + } 1 )$ -st element of $_ { \\textrm { x S } }$ . (If the length of xs was less than or equal to $\\boldsymbol { \\mathrm { n } }$ , the value NULL is returned instead.) \n• MINIMUM :: [int] $- >$ int lambda xs: min(xs) if len $( \\mathbf { \\nabla } _ { \\mathbf { X } } \\mathbf { S } ) > 0$ else Null Given an array, returns its minimum (or NULL if the array is empty). \n• MAXIMUM :: [int] $- >$ int lambda xs: max(xs) if len $( \\mathbf { \\nabla } _ { \\mathbf { X } } \\mathbf { S } ) > 0$ else Null Given an array, returns its maximum (or NULL if the array is empty). REVERSE :: [int] $- >$ [int] lambda xs: list(reversed(xs)) Given an array, returns its elements in reversed order. \n• SORT :: [int] -> [int] lambda xs: sorted(xs) Given an array, return its elements in non-decreasing order. SUM :: [int] $- >$ int lambda xs: sum(xs) Given an array, returns the sum of its elements. (The sum of an empty array is 0.) ", + "bbox": [ + 215, + 835, + 700, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "", + "bbox": [ + 214, + 103, + 825, + 517 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Higher-order functions: ", + "bbox": [ + 173, + 530, + 330, + 544 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "• MAP :: (int -> int) -> [int] -> [int] lambda f, xs: [f(x) for x in xs] Given a lambda function f mapping from integers to integers, and an array xs, returns the array resulting from applying f to each element of xs. \n• FILTER :: (int $- >$ bool) $- >$ [int] -> [int] lambda f, xs: [x for x in xs if f(x)] Given a predicate f mapping from integers to truth values, and an array xs, returns the elements of xs satisfying the predicate in their original order. \n• COUNT :: (int $- >$ bool) $- >$ [int] $- >$ int lambda f, xs: len([x for x in xs if f(x)]) Given a predicate f mapping from integers to truth values, and an array xs, returns the number of elements in xs satisfying the predicate. \n• ZIPWITH :: (int $- >$ int $- >$ int) $- >$ [int] -> [int] -> [int] lambda f, xs, ys: [f(x, y) for (x, y) in zip(xs, ys)] Given a lambda function f mapping integer pairs to integers, and two arrays xs and ys, returns the array resulting from applying f to corresponding elements of xs and ys. The length of the returned array is the minimum of the lengths of $_ { \\textrm { x S } }$ and ys. SCANL1 :: (int $- >$ int $- >$ int) $- >$ [int] -> [int] Given a lambda function f mapping integer pairs to integers, and an array $_ { \\textrm { x S } }$ , returns an array ys of the same length as xs and with its content defined by the recurrence $\\begin{array} { r l } { \\mathrm { y s } \\left[ 0 \\right] } & { { } = } \\end{array}$ $\\mathbf { x } \\mathbf { s } \\left[ 0 \\right] , \\mathbf { y } \\mathbf { s } \\left[ \\mathrm { n } \\right] \\mathbf { \\Psi } = \\mathbf { \\Psi } \\mathbf { f } \\left( \\mathbf { y } \\mathbf { s } \\left[ \\mathrm { n - 1 } \\right] \\right.$ , $\\mathtt { x s } [ \\mathtt { n } ]$ ) for $n \\geq 1$ . ", + "bbox": [ + 215, + 555, + 826, + 871 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "The INT INT lambdas $\\left( + 1 \\right) , \\left( - 1 \\right) , \\left( \\star 2 \\right) , \\left( / 2 \\right) , \\left( \\star \\left( - 1 \\right) \\right) , \\left( \\star \\star 2 \\right) , \\left( \\star 3 \\right) , \\left( / 3 \\right) , \\left( \\star 4 \\right) , \\left( / 4 \\right) , \\left( \\star 3 \\right) , \\left( \\star 4 \\right) .$ ) provided by our DSL map integers to integers in a self-explanatory manner. The $\\mathrm { I N T } { } \\mathrm { B O O L }$ lambdas $( > 0 )$ , $( < 0 )$ , ( $\\scriptstyle { \\frac { 0 } { 0 } } 2 = = 0$ ), ( $\\scriptstyle { \\frac { 0 } { 0 } } 2 = = 1$ ) respectively test positivity, negativity, evenness and oddness of the input integer value. Finally, the $\\mathrm { I N T } { } \\mathrm { I N T } { } \\mathrm { I N T }$ lambdas $( + ) , ( - ) , ( \\star )$ , MIN, MAX apply a function to a pair of integers and produce a single integer. ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "As an example, consider the function SCANL1 MAX, consisting of the higher-order function SCANL1 and the $\\mathrm { I N T } { } \\mathrm { I N T } { } \\mathrm { I N T }$ lambda MAX. Given an integer array a of length $L$ , this function computes the running maximum of the array a. Specifically, it returns an array $\\mathrm { b }$ of the same length $L$ whose $i$ -th element is the maximum of the first $i$ elements in $\\exists$ . ", + "bbox": [ + 174, + 138, + 825, + 194 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/da5620eb6e4f0a76e4439147498cbeb617797fc9a1d092bb07c51b51e15fdcb8.jpg", + "image_caption": [ + "Figure 9: Conditional confusion matrix for the neural network and test set of $P = 5 0 0$ programs of length $T = 3$ that were used to obtain the results presented in Table 1. Each cell contains the average false positive probability (in larger font) and the number of test programs from which this average was computed (smaller font, in brackets). The color intensity of each cell’s shading coresponds to the magnitude of the average false positive probability. " + ], + "image_footnote": [], + "bbox": [ + 176, + 208, + 818, + 707 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "G ANALYSIS OF TRAINED NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 174, + 823, + 576, + 838 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "We analyzed the performance of trained neural networks by investigating which program instructions tend to get confused by the networks. To this end, we looked at a generalization of confusion matrices to the multilabel classification setting: for each attribute in a ground truth program (rows) measure how likely each other attribute (columns) is predicted as a false positive. More formally, in this matrix the $( i , j )$ -entry is the average predicted probability of attribute $j$ among test programs that do possess attribute $i$ and do not possess attribute $j$ . Intuitively, the $i$ -th row of this matrix shows how the presence of attribute $i$ confuses the network into incorrectly predicting each other attribute $j$ . ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Figure 9 shows this conditional confusion matrix for the neural network and $P = 5 0 0$ program test set configuration used to obtain Table 1. We re-ordered the confusion matrix to try to expose block structure in the false positive probabilities, revealing groups of instructions that tend to be difficult to distinguish. Figure 10 show the conditional confusion matrix for the neural network used to obtain the table in Fig. 3a. While the results are somewhat noisy, we observe a few general tendencies: ", + "bbox": [ + 174, + 138, + 825, + 208 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "• There is increased confusion amongst instructions that select out a single element from an array: HEAD, LAST, ACCESS, MINIMUM, MAXIMUM. \n• Some common attributes get predicted more often regardless of the ground truth program: FILTER, $( > 0 )$ ), $( < 0 )$ , ( $\\scriptstyle { \\frac { \\circ } { \\circ } } 2 = = 1$ ), ( $\\scriptstyle { \\frac { 0 } { \\circ } } 2 = = 0$ ), MIN, MAX, (+), (-), ZIPWITH. \nThere are some groups of lambdas that are more difficult for the network to distinguish within: $( + )$ vs $( - ) ~ ; ~ ( + 1 )$ vs $( - 1 ) \\ ; \\ ( \\ / 2 )$ vs (/3) vs (/4). \n• When a program uses $( \\star \\star 2 )$ , the network often thinks it’s using $( \\star )$ , presumably because both can lead to large values in the output. ", + "bbox": [ + 215, + 219, + 825, + 348 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/a2f6e1ddd3c6cc3aed7aca84775a0568e1abe730347297f96f96da4cfd262c8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 368, + 818, + 864 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Figure 10: Conditional confusion matrix for the neural network and test set of $P = 5 0 0$ programs of length $T = 5$ . The presentation is the same as in Figure 9. ", + "bbox": [ + 173, + 877, + 823, + 906 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/parse/train/H1lj0nNFwB/H1lj0nNFwB.md b/parse/train/H1lj0nNFwB/H1lj0nNFwB.md new file mode 100644 index 0000000000000000000000000000000000000000..af267eedfbc9f88eca28ff5ba5ce7ba4eaed8fb6 --- /dev/null +++ b/parse/train/H1lj0nNFwB/H1lj0nNFwB.md @@ -0,0 +1,870 @@ +# THE IMPLICIT BIAS OF DEPTH: HOW INCREMENTAL LEARNING DRIVES GENERALIZATION + +Daniel Gissin, Shai Shalev-Shwartz, Amit Daniely +School of Computer Science +The Hebrew University +Jerusalem, Israel +{daniel.gissin,shais,amit.daniely}@mail.huji.ac.il + +# ABSTRACT + +A leading hypothesis for the surprising generalization of neural networks is that the dynamics of gradient descent bias the model towards simple solutions, by searching through the solution space in an incremental order of complexity. We formally define the notion of incremental learning dynamics and derive the conditions on depth and initialization for which this phenomenon arises in deep linear models. Our main theoretical contribution is a dynamical depth separation result, proving that while shallow models can exhibit incremental learning dynamics, they require the initialization to be exponentially small for these dynamics to present themselves. However, once the model becomes deeper, the dependence becomes polynomial and incremental learning can arise in more natural settings. We complement our theoretical findings by experimenting with deep matrix sensing, quadratic neural networks and with binary classification using diagonal and convolutional linear networks, showing all of these models exhibit incremental learning. + +# 1 INTRODUCTION + +Neural networks have led to a breakthrough in modern machine learning, allowing us to efficiently learn highly expressive models that still generalize to unseen data. The theoretical reasons for this success are still unclear, as the generalization capabilities of neural networks defy the classic statistical learning theory bounds. Since these bounds, which depend solely on the capacity of the learned model, are unable to account for the success of neural networks, we must examine additional properties of the learning process. One such property is the optimization algorithm - while neural networks can express a multitude of possible ERM solutions for a given training set, gradient-based methods with the right initialization may be implicitly biased towards certain solutions which generalize. + +A possible way such an implicit bias may present itself, is if gradient-based methods were to search the hypothesis space for possible solutions of gradually increasing complexity. This would suggest that while the hypothesis space itself is extremely complex, our search strategy favors the simplest solutions and thus generalizes. One of the leading results along these lines has been by Saxe et al. (2013), deriving an analytical solution for the gradient flow dynamics of deep linear networks and showing that for such models, the singular values converge at different rates, with larger values converging first. At the limit of infinitesimal initialization of the deep linear network, Gidel et al. (2019) show these dynamics exhibit a behavior of “incremental learning” - the singular values of the model are learned separately, one at a time. Our work generalizes these results to small but finite initialization scales. + +Incremental learning dynamics have also been explored in gradient descent applied to matrix completion and sensing with a factorized parameterization (Gunasekar et al. (2017), Arora et al. (2018), Woodworth et al. (2019)). When initialized with small Gaussian weights and trained with a small learning rate, such a model is able to successfully recover the low-rank matrix which labeled the data, even if the problem is highly over-determined and no additional regularization is applied. In their proof of low-rank recovery for such models, Li et al. (2017) show that the model remains lowrank throughout the optimization process, leading to the successful generalization. Additionally, + +![](images/14126dc03eaae57887ee79a34684cde80a2ac3af64c8443f5095c0fc7eaad71c.jpg) +Figure 1: Incremental learning dynamics in deep models. Each panel shows the evolution of the five largest values of $\sigma$ , the parameters of the induced model. All models were trained using gradient descent with a small initialization and learning rate, on a small training set such that there are multiple possible solutions. In all cases, the deep parameterization of the models lead to “incremental learning”, where the values are learned at different rates (larger values are learned first), leading to sparse solutions. (a) Depth 4 matrix sensing, $\sigma$ denotes singular values (see section 4.1). (b) Quadratic networks, $\sigma$ denotes singular values (see section 4.2). (c) Depth 3 diagonal networks, $\sigma$ denotes feature weights (see section 4.3). (d) Depth 3 circular-convolutional networks, $\sigma$ denotes amplitudes in the frequency domain of the feature weights (see appendix G). + +Arora et al. (2019) explore the dynamics of such models, showing the singular values are learned at different rates and that deeper models exhibit stronger incremental learning dynamics. Our work deals with a more simplified setting, allowing us to determine explicitly under which conditions depth leads to this dynamical phenomenon. + +Finally, the learning dynamics of nonlinear models have been studied as well. Combes et al. (2018) and Williams et al. (2019) study the gradient flow dynamics of shallow ReLU networks under restrictive distributional assumptions, Ronen et al. (2019) show that shallow networks learn functions of gradually increasing frequencies and Nakkiran et al. (2019) show how deep ReLU networks correlate with linear classifiers in the early stages of training. + +These findings, along with others, suggest that the generalization ability of deep networks is at least in part due to the incremental learning dynamics of gradient descent. Following this line of work, we begin by explicitly defining the notion of incremental learning for a toy model which exhibits this sort of behavior. Analyzing the dynamics of the model for gradient flow and gradient descent, we characterize the effect of the model’s depth and initialization scale on incremental learning, showing how deeper models allow for incremental learning in larger (realistic) initialization scales. Specifically, we show that a depth-2 model requires exponentially small initialization for incremental learning to occur, while deeper models only require the initialization to be polynomially small. + +Once incremental learning has been defined and characterized for the toy model, we generalize our results theoretically and empirically for larger linear and quadratic models. Examples of incremental learning in these models can be seen in figure 1, which we discuss further in section 4. + +# 2 DYNAMICAL ANALYSIS OF A TOY MODEL + +We begin by analyzing incremental learning for a simple model. This will allow us to gain a clear understanding of the phenomenon and the conditions for it, which we will later be able to apply to a variety of other models in which incremental learning is present. + +# 2.1 PRELIMINARIES + +Our simple linear model will be similar to the toy model analyzed by Woodworth et al. (2019). Our input space will be $\mathcal { X } = \mathbb { R } ^ { d }$ and the hypothesis space will be linear models with non-negative weights, such that: + +$$ +f _ { \sigma } ( { \boldsymbol { x } } ) = \langle \sigma , { \boldsymbol { x } } \rangle \qquad \sigma \in \mathbb { R } _ { \geq 0 } ^ { d } +$$ + +We will introduce depth into our model, by parameterizing $\sigma$ using $w \in \mathbb { R } _ { \geq 0 } ^ { d }$ in the following way: + +$$ +\forall i : \sigma _ { i } = w _ { i } ^ { N } +$$ + +Where $N$ represents the depth of the model. Since we restrict the model to having non-negative weights, this parameterization doesn’t change the expressiveness, but it does radically change it’s optimization dynamics. + +Assuming the data is labeled by some $\sigma ^ { * } \in \mathbb { R } _ { \geq 0 } ^ { d }$ , we will study the dynamics of this model for general under a depth-normalized1 squared loss over Gaussian inputs, which will allow us to derive our analytical solution: + +$$ +\ell _ { N } ( w ) = \frac { 1 } { 2 N ^ { 2 } } \mathbb { E } _ { x } [ ( \langle \sigma ^ { * } , x \rangle - \langle w ^ { N } , x \rangle ) ^ { 2 } ] = \frac { 1 } { 2 N ^ { 2 } } | | w ^ { N } - \sigma ^ { * } | | ^ { 2 } +$$ + +We will assume that our model is initialized uniformly with a tunable scaling factor, such that: + +$$ +\forall i : \ w _ { i } ( 0 ) = \ \sqrt [ N ] { \sigma _ { 0 } } +$$ + +# 2.2 GRADIENT FLOW ANALYTICAL SOLUTIONS + +Analyzing our toy model using gradient flow allows us to obtain an analytical solution for the dynamics of $\sigma ( t )$ along with the dynamics of the loss function for a general $N$ . For brevity, the following theorem refers only to $N = 1 , 2$ and $N \to \infty$ , however the solutions for $3 \leq N < \infty$ are similar in structure to $N \to \infty$ , but more complicated. We also assume $\sigma _ { i } ^ { * } > 0$ for brevity, however we can derive the solutions for $\sigma _ { i } ^ { * } = 0$ as well. Note that this result is a special case adaptation of the one presented in Saxe et al. (2013) for deep linear networks: + +Theorem 1. Minimizing the toy linear model described in (1) with gradient flow over the depth normalized squared loss (2), with Gaussian inputs and weights initialized as in (3) and assuming $\sigma _ { i } ^ { * } > 0$ leads to the following analytical solutions for different values of $N$ : + +$$ +\begin{array} { r l } & { \begin{array} { r l } { N = 1 : } & { \sigma _ { i } ( t ) = \sigma _ { i } ^ { * } + ( \sigma _ { 0 } - \sigma _ { i } ^ { * } ) e ^ { - t } } \\ { N = 2 : } & { \sigma _ { i } ( t ) = \frac { \sigma _ { 0 } \sigma _ { i } ^ { * } e ^ { \sigma _ { i } ^ { * } t } } { \sigma _ { 0 } \left( e ^ { \sigma _ { i } ^ { * } t } - 1 \right) + \sigma _ { i } ^ { * } } } \end{array} } \\ & { \begin{array} { r l } { N \to \infty : } & { t = \frac { 1 } { ( \sigma _ { i } ^ { * } ) ^ { 2 } } \log \big ( \frac { \sigma _ { i } ( t ) ( \sigma _ { 0 } - \sigma _ { i } ^ { * } ) } { \sigma _ { 0 } \left( \sigma _ { i } ( t ) - \sigma _ { i } ^ { * } \right) } \big ) - \frac { 1 } { \sigma _ { i } ^ { * } } \big ( \frac { 1 } { \sigma _ { i } ( t ) } - \frac { 1 } { \sigma _ { 0 } } \big ) } \end{array} } \end{array} +$$ + +Proof. The gradient flow equations for our model are the following: + +$$ +\dot { w } _ { i } = - \nabla _ { w _ { i } } \ell = \frac { 1 } { N } w _ { i } ^ { N - 1 } ( \sigma _ { i } ^ { * } - w _ { i } ^ { N } ) +$$ + +Given the dynamics of the $w$ parameters, we may use the chain rule to derive the dynamics of the induced model, $\sigma$ : + +$$ +\dot { \sigma } _ { i } = \frac { d \sigma _ { i } } { d w _ { i } } \dot { w } _ { i } = w ^ { 2 N - 2 } ( \sigma _ { i } ^ { * } - w _ { i } ^ { N } ) = \sigma _ { i } ^ { 2 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma _ { i } ) +$$ + +This differential equation is solvable for all $N$ , leading to the solutions in the theorem. Taking $N \to \infty$ in (4) leads to $\dot { \sigma } _ { i } = \sigma _ { i } ^ { 2 } ( \sigma _ { i } ^ { * } - \sigma _ { i } )$ , which is also solvable. + +![](images/c32e4b6822cc4e3b3801507a0f18a13c9706cba7f76cc5695c9ee740ff4a57cf.jpg) +Figure 2: Incremental learning dynamics in the toy model. Each panel shows the evolution of $\frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } }$ for $\sigma _ { i } ^ { * } \in \{ 1 2 , 6 , 4 , 3 \}$ according to the analytical solutions in theorem 1, under different depths and initializations. The first column has all values converging at the same rate. Notice how the deep parameterization with small initialization leads to distinct phases of learning, where values are learned incrementally (bottom-right). The shallow model’s much weaker incremental learning, even at small initialization scales (second column), is explained in theorem 2. + +Analyzing these solutions, we see how even in such a simple model depth causes different factors of the model to be learned at different rates. Specifically, values corresponding to larger optimal values converge faster, suggesting a form of incremental learning. This is most clear for $N = 2$ where the solution isn’t implicit, but is also the case for $N \geq 3$ , as we will see in the next subsection. + +These dynamics are depicted in figure 2, where we see the dynamics of the different values of $\sigma ( t )$ as learning progresses. When $N = 1$ , all values are learned at the same rate regardless of the initialization, while the deeper models are clearly biased towards learning the larger singular values first, especially at small initialization scales. + +Our model has only one optimal solution due to the population loss, but it is clear how this sort of dynamic can induce sparse solutions - if the model is able to fit the data after a small amount of learning phases, then it’s obtained result will be sparse. Alternatively, if $N = 1$ , we know that the dynamics will lead to the minimal $\ell _ { 2 }$ norm solution which is dense. We explore the sparsity inducing bias of our toy model by comparing it empirically2 to a greedy sparse approximation algorithm in appendix D, and give our theoretical results in the next section. + +# 3 INCREMENTAL LEARNING + +Equipped with analytical solutions for the dynamics of our model for every depth, we turn to study how the depth and initialization effect incremental learning. While Gidel et al. (2019) focuses on incremental learning in depth-2 models at the limit of $\sigma _ { 0 } 0$ , we will study the phenomenon for a general depth and for $\sigma _ { 0 } > 0$ . + +First, we will define the notion of incremental learning. Since all values of $\sigma$ are learned in parallel, we can’t expect one value to converge before the other moves at all (which happens for infinitesimal initialization as shown by Gidel et al. (2019)). We will need a more relaxed definition for incremental learning in finite initialization scales. + +Definition 1. Given two values $\sigma _ { i } , \sigma _ { j }$ such that $\sigma _ { i } ^ { * } > \sigma _ { j } ^ { * } > 0$ and both are initialized as $\sigma _ { i } ( 0 ) =$ $\sigma _ { j } ( 0 ) = \sigma _ { 0 } < \sigma _ { j } ^ { * }$ , and given two scalars $s \in ( 0 , \frac { 1 } { 4 } )$ and $f \in \left( \frac { 3 } { 4 } , 1 \right)$ , we call the learning of the values $( s , f )$ -incremental if there exists a t for which: + +$$ +\sigma _ { j } ( t ) \leq s \sigma _ { j } ^ { * } < f \sigma _ { i } ^ { * } \leq \sigma _ { i } ( t ) +$$ + +In words, two values have distinct learning phases if the first almost converges $( f \approx 1 )$ before the second changes by much $s \ll 1 \ r _ { \cdot }$ ). Note that for any $N , \sigma ( t )$ is monotonically increasing and so once $\sigma _ { j } ( t ) = s \sigma _ { j } ^ { * }$ , it will not decrease to allow further incremental learning. Given this definition of incremental learning, we turn to study the conditions that facilitate incremental learning in our toy model. + +Our main result is a dynamical depth separation result, showing that incremental learning is dependent on $\frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } }$ in different ways for different values of $N$ . The largest difference in dependence happens between $N = 2$ and $N = 3$ , where the dependence changes from exponential to polynomial: + +Theorem 2. Given two values $\sigma _ { i } , \sigma _ { j }$ of a toy linear model as in (1), where $\begin{array} { r } { \frac { { \sigma } _ { i } ^ { * } } { \sigma _ { j } ^ { * } } = r > 1 } \end{array}$ and the model is initialized as in (3), and given two scalars $s \in ( 0 , \frac { 1 } { 4 } )$ and $f \in \left( \frac { 3 } { 4 } , 1 \right)$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\sigma _ { 0 } ^ { t h }$ , is bounded in the following way: + +$$ +\begin{array} { r l r } { s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { s } { r f } \Big ) ^ { \frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \le \sigma _ { 0 } ^ { t h } \le s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { s } { r f } \Big ) ^ { \frac { 1 - s } { r - 1 } } } & { \qquad } & { N = 2 } \\ { s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \Big ) ^ { \frac { N } { N - 2 } } } & { \le \sigma _ { 0 } ^ { t h } \le s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { r - 1 } { r - s } \Big ) ^ { \frac { N } { N - 2 } } } & { \qquad } & { N \ge 3 } \end{array} +$$ + +Proof sketch (the full proof is given in appendix $A$ ). Rewriting the separable differential equation in (4) to calculate the time until $\sigma ( t ) = \alpha \sigma ^ { * }$ , we get the following: + +$$ +t _ { \alpha } ( \sigma ) = \int _ { \sigma _ { 0 } } ^ { \alpha \sigma ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma ^ { * } - \sigma ) } +$$ + +The condition for incremental learning is then the requirement that $t _ { f } ( \sigma _ { i } ) \leq t _ { s } ( \sigma _ { j } )$ , resulting in: + +$$ +\int _ { \sigma _ { 0 } } ^ { f \sigma _ { i } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma ) } \le \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma _ { j } ^ { * } - \sigma ) } +$$ + +We then relax/restrict the above condition to get a necessary/sufficient condition on $\sigma _ { 0 }$ , leading to a lower and upper bound on $\sigma _ { 0 } ^ { t h }$ . + +Note that the value determining the condition for incremental learning is $\frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } }$ same order of magnitude, then their ratio will be close to 1 and we will need a small initialization to obtain incremental learning. The dependence on the ratio changes with depth, and is exponential for $N = 2$ . This means that incremental learning, while possible for shallow models, is difficult to see in practice. This result explains why changing the initialization scale in figure 2 changes the dynamics of the $N \geq 3$ models, while not changing the dynamics for $N = 2$ noticeably. + +The next theorem extends part of our analysis to gradient descent, a more realistic setting than the infinitesimal learning rate of gradient flow: + +Theorem 3. Given two values $\sigma _ { i } , \sigma _ { j }$ of a depth-2 toy linear model as in (1), such that $\begin{array} { r } { \frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } } = r > 1 } \end{array}$ and the model is initialized as in (3), and given two scalars $s \in ( 0 , \frac { 1 } { 4 } )$ and $f \in ( \frac { 3 } { 4 } , 1 )$ , and assuming $\sigma _ { j } ^ { * } ~ \geq ~ 2 \sigma _ { 0 }$ , and assuming we optimize with gradient descent with a learning rate $\begin{array} { r } { \eta \ \leq \ \frac { c } { \sigma _ { 1 } ^ { * } } } \end{array}$ for $c < 2 ( { \sqrt { 2 } } - 1 )$ and $\sigma _ { 1 } ^ { * }$ the largest value of $\sigma ^ { * }$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\sigma _ { 0 } ^ { t h }$ , is lower and upper bounded in the following way: + +$$ +\frac { 1 } { 2 } \frac { s } { 1 - s } \sigma _ { j } ^ { * } \Bigl ( \frac { 1 - f } { 2 r f } \frac { s } { 1 - s } \Bigr ) ^ { \frac { 1 } { A - 1 } } \leq \sigma _ { 0 } ^ { t h } \leq \frac { s } { 1 - s } \sigma _ { j } ^ { * } \Bigl ( \frac { 1 - f } { f } \frac { s } { 1 - s } \Bigr ) ^ { \frac { 1 } { B - 1 } } +$$ + +Where $A$ and $B$ are defined as: + +$$ +A = \frac { \log \left( 1 - c \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } + c ^ { 2 } \big ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } { \log \left( 1 - c \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } - \frac { c ^ { 2 } } { 4 } \big ( \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } \qquad B = \frac { \log \left( 1 - c \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } - \frac { c ^ { 2 } } { 4 } \big ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } { \log \left( 1 - c \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } + c ^ { 2 } \big ( \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } +$$ + +We defer the proof to appendix B. + +Note that this result, while less elegant than the bounds of the gradient flow analysis, is similar in nature. Both $A$ and $B$ simplify to $r$ when we take their first order approximation around $c = 0$ , giving us similar bounds and showing that the condition on $\sigma _ { 0 }$ for $N = 2$ is exponential in gradient descent as well. + +While similar gradient descent results are harder to obtain for deeper models, we discuss the general effect of depth on the gradient decent dynamics in appendix C. + +# 4 INCREMENTAL LEARNING IN LARGER MODELS + +So far, we have only shown interesting properties of incremental learning caused by depth for a toy model. In this section, we will relate several deep models to our toy model and show how incremental learning presents itself in larger models as well. + +# 4.1 MATRIX SENSING + +The task of matrix sensing is a generalization of matrix completion, where our input space is $\mathcal { X } =$ $\mathbb { R } ^ { d \times d }$ and our model is a matrix $\mathbf { \bar { \boldsymbol { W } } } \in \mathbb { R } ^ { d \times d }$ , such that: + +$$ +f _ { W } ( A ) = \langle W , A \rangle = t r \big ( W ^ { T } A \big ) +$$ + +Following Arora et al. (2019), we introduce depth by parameterizing the model using a product of matrices and the following initialization scheme $( W _ { i } ^ { \setminus } \in \mathbb { R } ^ { d \times d } )$ ): + +$$ +\begin{array} { c } { { W = W _ { N } W _ { N - 1 } \cdot \cdot \cdot W _ { 1 } } } \\ { { \forall i \in [ N ] , ~ W _ { i } ( 0 ) = ~ \sqrt [ N ] { \sigma _ { 0 } } I } } \end{array} +$$ + +Note that when $d = 1$ , the deep matrix sensing model reduces to our toy model without weight sharing. We study the dynamics of the model under gradient flow over a depth-normalized squared loss, assuming the data is labeled by a matrix sensing model parameterized by a PSD $W ^ { \ast } \in \mathring { \mathbb { R } } ^ { d \times d }$ : + +$$ +\begin{array} { c } { \displaystyle \ell _ { N } ( W _ { N } W _ { N - 1 } \cdot \cdot \cdot W _ { 1 } ) = \displaystyle \frac { 1 } { 2 N } \mathbb { E } _ { A } [ ( \langle \left( W _ { N } W _ { N - 1 } \cdot \cdot \cdot W _ { 1 } \right) - W ^ { * } , A \rangle ^ { 2 } ] } \\ { = \displaystyle \frac { 1 } { 2 N } | | \big ( W _ { N } W _ { N - 1 } \cdot \cdot \cdot W _ { 1 } \big ) - W ^ { * } | | _ { F } ^ { 2 } } \end{array} +$$ + +The following theorem relates the deep matrix sensing model to our toy model, showing the two have the same dynamical equations: + +Theorem 4. Optimizing the deep matrix sensing model described in (5) with gradient flow over the depth normalized squared loss ((6)), with weights initialized as in (5) leads to the following dynamical equations for different values of $N$ : + +$$ +\dot { \sigma } _ { i } ( t ) = \sigma _ { i } ( t ) ^ { 2 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) +$$ + +Where $\sigma _ { i }$ and $\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vector. + +The proof follows that of Saxe et al. (2013) and Gidel et al. (2019) and is deferred to appendix E. + +Theorem 4 shows us that the bias towards sparse solutions introduced by depth in the toy model is equivalent to the bias for low-rank solutions in the matrix sensing task. This bias was studied in a more general setting in Arora et al. (2019), with empirical results supporting the effect of depth on the obtainment of low-rank solutions under a more natural loss and initialization scheme. We recreate and discuss these experiments and their connection to our analysis in appendix E, and an example of these dynamics in deep matrix sensing can also be seen in panel (a) of figure 1. + +# 4.2 QUADRATIC NEURAL NETWORKS + +By drawing connections between quadratic networks and matrix sensing (as in Soltanolkotabi et al. (2018)), we can extend our results to these nonlinear models. We will study a simplified quadratic network, where our input space is $\mathcal { X } = \mathbb { R } ^ { d }$ and the first layer is parameterized by a weight matrix $W \in \mathbb { R } ^ { d \times d }$ and followed by a quadratic activation function. The final layer will be a summation layer. We assume, like before, that the labeling function is a quadratic network parameterized by $\bar { W } ^ { * } \in \mathbb { R } ^ { d \times d }$ . Our model can be written in the following way, using the following orthogonal initialization scheme: + +$$ +\begin{array} { r l } & { f _ { W } ( x ) = \displaystyle \sum _ { i = 1 } ^ { d } ( w _ { i } ^ { T } x ) ^ { 2 } = x ^ { T } W ^ { T } W x = \langle W ^ { T } W , x x ^ { T } \rangle } \\ & { W _ { 0 } ^ { T } W _ { 0 } = \sigma _ { 0 } I } \end{array} +$$ + +Immediately, we see the similarity of the quadratic network to the deep matrix sensing model with $N = 2$ , where the input space is made up of rank-1 matrices. However, the change in input space forces us to optimize over a different loss function to reproduce the same dynamics: + +Definition 2. Given an input distribution over an input space $\mathcal { X }$ with a labeling function $y : \mathcal { X } \mathbb { R }$ and a hypothesis $h$ , the variance loss is defined in the following way: + +$$ +\ell _ { v a r } ( h ) = \frac { 1 } { 1 6 } \mathbb { E } _ { x } [ ( y ( x ) - h ( x ) ) ^ { 2 } ] - \frac { 1 } { 1 6 } \mathbb { E } _ { x } [ y ( x ) - h ( x ) ] ^ { 2 } +$$ + +Note that minimizing this loss function amounts to minimizing the variance of the error, while the squared loss minimizes the second moment of the error. We note that both loss functions have the same minimum for our problem, and the dynamics of the squared loss can be approximated in certain cases by the dynamics of the variance loss. For a complete discussion of the two losses, including the cases where the two losses have similar dynamics, we refer the reader to appendix F. + +Theorem 5. Minimizing the quadratic network described and initialized as in (7) with gradient flow over the variance loss defined in (2) leads to the following dynamical equations: + +$$ +\dot { \sigma } _ { i } ( t ) = \sigma _ { i } ( t ) ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) +$$ + +Where $\sigma _ { i }$ and $\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vector. + +We defer the proof to appendix F and note that these dynamics are the same as our depth-2 toy model, showing that shallow quadratic networks can exhibit incremental learning (albeit requiring a small initialization). + +# 4.3 DIAGONAL/CONVOLUTIONAL LINEAR NETWORKS + +While incremental learning has been described for deep linear networks in the past, it has been restricted to regression tasks. Here, we illustrate how incremental learning presents itself in binary classification, where implicit bias results have so far focused on convergence at $t \to \infty$ (Soudry et al. (2018), Nacson et al. (2018), Ji & Telgarsky (2019)). Deep linear networks with diagonal weight matrices have been shown to be biased towards sparse solutions when $N > 1$ in Gunasekar et al. (2018), and biased towards the max-margin solution for $N = 1$ . Instead of analyzing convergence at $t \infty$ , we intend to show that the model favors sparse solutions for the entire duration of optimization, and that this is due to the dynamics of incremental learning. + +Our theoretical illustration will use our toy model as in (1) (initialized as in (3)) as a special weightshared case of deep networks with diagonal weight matrices, and we will then show empirical results for the more general setting. We analyze the optimization dynamics of this model over a separable dataset $\{ x _ { i } , y _ { i } \} _ { i = 1 } ^ { m }$ where $y _ { i } \in \{ \pm 1 \}$ . We use the exponential loss $( \ell ( f ( x ) , y ) = e ^ { - y f ( x ) } )$ for the theoretical illustration and experiment on the exponential and logistic losses. + +Computing the gradient for the model over $w$ , the gradient flow dynamics for $\sigma$ become: + +$$ +\dot { \sigma } _ { i } = \frac { N ^ { 2 } } { m } \sigma _ { i } ^ { 2 - \frac { 2 } { N } } \sum _ { j = 1 } ^ { m } e ^ { - y _ { j } \langle \beta , x _ { j } \rangle } x _ { j } +$$ + +We see the same dynamical attenuation of small values of $\sigma$ that is seen in the regression model, caused by the multiplication by σ2− 2N From this, we can expect the same type of incremental learning to occur - weights of $\sigma$ will be learned incrementally until the dataset can be separated by the current support of $\sigma$ . Then, the dynamics strengthen the growth of the current support while relatively attenuating that of the other values. Since the data is separated, increasing the values of the current support reduces the loss and the magnitude of subsequent gradients, and so we should expect the support to remain the same and the model to converge to a sparse solution. + +Granted, the above description is just intuition, but panel (c) of figure 1 shows how it is born out in practice (similar results are obtained for the logistic loss). In appendix G we further explore this model, showing deeper networks have a stronger bias for sparsity. We also observe that the initialization scale plays a similar role as before - deep models are less biased towards sparsity when $\sigma _ { 0 }$ is large. + +In their work, Gunasekar et al. (2018) show an equivalence between the diagonal network and the circular-convolutional network in the frequency domain. According to their results, we should expect to see the same sparsity-bias of diagonal networks in convolutional networks, when looking at the Fourier coefficients of $\sigma$ . An example of this can be seen in panel (d) of figure 1, and we refer the reader to appendix $\mathbf { G }$ for a full discussion of their convolutional model and it’s incremental learning dynamics. + +# 5 CONCLUSION + +Gradient-based optimization for deep linear models has an implicit bias towards simple (sparse) solutions, caused by an incremental search strategy over the hypothesis space. Deeper models have a stronger tendency for incremental learning, exhibiting it in more realistic initialization scales. + +This dynamical phenomenon exists for the entire optimization process for regression as well as classification tasks, and for many types of models - diagonal networks, convolutional networks, matrix completion and even the nonlinear quadratic network. We believe this kind of dynamical analysis may be able to shed light on the generalization of deeper nonlinear neural networks as well, with shallow quadratic networks being only a first step towards that goal. + +# ACKNOWLEDGMENTS + +This research is supported by the European Research Council (TheoryDL project). + +# REFERENCES + +Sanjeev Arora, Nadav Cohen, and Elad Hazan. On the optimization of deep networks: Implicit acceleration by overparameterization. arXiv preprint arXiv:1802.06509, 2018. + +Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In Advances in Neural Information Processing Systems 32, pp. 7411–7422. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8960-implicit-regularization-in-deep-matrix-factorization.pdf. + +Remi Tachet des Combes, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848, 2018. + +Gauthier Gidel, Francis Bach, and Simon Lacoste-Julien. Implicit regularization of discrete gradient dynamics in linear neural networks. In Advances in Neural Information Processing Systems, pp. 3196–3206, 2019. + +Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit regularization in matrix factorization. In Advances in Neural Information Processing Systems, pp. 6151–6159, 2017. + +Suriya Gunasekar, Jason D Lee, Daniel Soudry, and Nati Srebro. Implicit bias of gradient descent on linear convolutional networks. In Advances in Neural Information Processing Systems, pp. 9461–9471, 2018. + +Ziwei Ji and Matus Telgarsky. The implicit bias of gradient descent on nonseparable data. In Conference on Learning Theory, pp. 1772–1798, 2019. + +Yuanzhi Li, Tengyu Ma, and Hongyang Zhang. Algorithmic regularization in over-parameterized matrix sensing and neural networks with quadratic activations. arXiv preprint arXiv:1712.09203, 2017. + +Mor Shpigel Nacson, Jason Lee, Suriya Gunasekar, Pedro HP Savarese, Nathan Srebro, and Daniel Soudry. Convergence of gradient descent on separable data. arXiv preprint arXiv:1803.01905, 2018. + +Preetum Nakkiran, Gal Kaplun, Dimitris Kalimeris, Tristan Yang, Benjamin L Edelman, Fred Zhang, and Boaz Barak. Sgd on neural networks learns functions of increasing complexity. arXiv preprint arXiv:1905.11604, 2019. + +Yagyensh Chandra Pati, Ramin Rezaiifar, and Perinkulam Sambamurthy Krishnaprasad. Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition. In Proceedings of 27th Asilomar conference on signals, systems and computers, pp. 40–44. IEEE, 1993. + +Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman. The convergence rate of neural networks for learned functions of different frequencies. In Advances in Neural Information Processing Systems 32, pp. 4763–4772. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8723-the-convergence-rate-of-neural-networks-for-learned-functions-of-different-fr pdf. + +Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013. + +Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. IEEE Transactions on Information Theory, 65(2):742–769, 2018. + +Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 19 (1):2822–2878, 2018. + +Francis Williams, Matthew Trager, Claudio Silva, Daniele Panozzo, Denis Zorin, and Joan Bruna. Gradient dynamics of shallow univariate relu networks. arXiv preprint arXiv:1906.07842, 2019. + +Blake Woodworth, Suriya Gunasekar, Jason Lee, Daniel Soudry, and Nathan Srebro. Kernel and deep regimes in overparametrized models. arXiv preprint arXiv:1906.05827, 2019. + +# A PROOF OF THEOREM 2 + +Theorem. Given two values $\sigma _ { i } , \sigma _ { j }$ of a toy linear model as in (1), such that $\begin{array} { r } { \frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } } = r > 1 } \end{array}$ and the model is initialized as in (3), and given two scalars $s \in ( 0 , \frac { 1 } { 4 } )$ and $f \in \left( \frac { 3 } { 4 } , 1 \right)$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\sigma _ { 0 } ^ { t h }$ , is lower and upper bounded in the following way: + +$$ +\begin{array} { r l r } { s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { s } { r f } \Big ) ^ { \frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \le \sigma _ { 0 } ^ { t h } \le s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { s } { r f } \Big ) ^ { \frac { 1 - s } { r - 1 } } } & { \qquad } & { N = 2 } \\ { s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \Big ) ^ { \frac { N } { N - 2 } } } & { \le \sigma _ { 0 } ^ { t h } \le s \sigma _ { j } ^ { * } \Big ( \displaystyle \frac { r - 1 } { r - s } \Big ) ^ { \frac { N } { N - 2 } } } & { \qquad } & { N \ge 3 } \end{array} +$$ + +Proof. Our strategy will be to define the time $t _ { \alpha }$ for which a value reaches a fraction $\alpha$ of it’s optimal value, and then require that $t _ { f } ( \sigma _ { i } ) \leq t _ { s } ( \sigma _ { j } )$ . We begin with recalling the differential equation which determines the dynamics of the model: + +$$ +\dot { \sigma } = \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma ^ { * } - \sigma ) +$$ + +Since the solution for $N \geq 3$ is implicit and difficult to manage in a general form, we will define $t _ { \alpha }$ using the integral of the differential equation. The equation is separable, and under initialization of $\sigma _ { 0 }$ we can describe $t _ { \alpha } ( \sigma )$ in the following way: + +$$ +t _ { \alpha } ( \sigma ) = \int _ { \sigma _ { 0 } } ^ { \alpha \sigma ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma ^ { * } - \sigma ) } +$$ + +Incremental learning takes place when $\sigma _ { i } ( t _ { f } ) = f \sigma _ { i } ^ { * }$ happens before $\sigma _ { j } ( t _ { s } ) = s \sigma _ { j } ^ { * }$ . We can write this condition in the following way: + +$$ +\int _ { \sigma _ { 0 } } ^ { f \sigma _ { i } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma ) } \le \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( \sigma _ { j } ^ { * } - \sigma ) } +$$ + +Plugging in $\sigma _ { i } = r \sigma _ { j }$ and rearranging, we get the following necessary and sufficient condition for incremental learning: + +$$ +\int _ { \sigma _ { 0 } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } ) } \leq r \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { \sigma _ { j } ^ { * } } ) } +$$ + +Our last step before relaxing and restricting our condition will be to split the integral on the left-hand side into two integrals: + +$$ +\int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } ) } + \int _ { s \sigma _ { j } ^ { * } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } ) } \le r \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { \sigma _ { j } ^ { * } } ) } +$$ + +At this point, we cannot solve this equation and isolate $\sigma _ { 0 }$ to obtain a clear threshold condition on it for incremental learning. Instead, we will relax/restrict the above condition to get a necessary/sufficient condition on $\sigma _ { 0 }$ , leading to a lower and upper bound on the threshold value of $\sigma _ { 0 }$ . + +# SUFFICIENT CONDITION + +To obtain a sufficient (but not necessary) condition on $\sigma _ { 0 }$ , we may make the condition stricter either by increasing the left-hand side or decreasing the right-hand side. We can increase the left-hand side by removing $r$ from the left-most integral’s denominator $( r > 1 )$ ) and then combine the left-most and right-most integrals: + +$$ +\int _ { s \sigma _ { j } ^ { * } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } ) } \le ( r - 1 ) \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { \sigma _ { j } ^ { * } } ) } +$$ + +Next, we note that the integration bounds give us a bound on $\sigma$ for either integral. This means we can replace $\begin{array} { r } { 1 - \frac { \sigma } { \sigma _ { j } ^ { * } } } \end{array}$ with 1 on the right-hand side, and replace $\begin{array} { r } { 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } } \end{array}$ with $1 - f$ on the left-hand side: + +$$ +\frac { 1 } { 1 - f } \int _ { s \sigma _ { j } ^ { * } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } } \le ( r - 1 ) \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } } +$$ + +We may now solve these integrals for every $N$ and isolate $\sigma _ { 0 }$ , obtaining the lower bound on $\sigma _ { 0 } ^ { t h }$ . We start with the case where $N = 2$ : + +$$ +\frac { 1 } { 1 - f } \big ( \log ( f r \sigma _ { j } ^ { * } ) - \log ( s \sigma _ { j } ^ { * } ) \big ) \leq ( r - 1 ) \big ( \log ( s \sigma _ { j } ^ { * } ) - \log ( \sigma _ { 0 } ) \big ) +$$ + +Rearranging to isolate $\sigma _ { 0 }$ , we obtain our result: + +$$ +\sigma _ { 0 } \leq s \sigma _ { j } ^ { * } \big ( { \frac { s } { r f } } \big ) ^ { \frac { 1 } { ( 1 - f ) ( r - 1 ) } } +$$ + +For the $N \geq 3$ case, we have the following after solving the integrals: + +$$ +\frac { 1 } { 1 - f } \Big ( \big ( \frac { 1 } { s \sigma _ { j } ^ { * } } \big ) ^ { 1 - \frac { 2 } { N } } - \big ( \frac { 1 } { r f \sigma _ { j } ^ { * } } \big ) ^ { 1 - \frac { 2 } { N } } \Big ) \leq ( r - 1 ) \Big ( \big ( \frac { 1 } { \sigma _ { 0 } } \big ) ^ { 1 - \frac { 2 } { N } } - \big ( \frac { 1 } { s \sigma _ { j } ^ { * } } \big ) ^ { 1 - \frac { 2 } { N } } \Big ) +$$ + +For simplicity we may further restrict the condition by removing the term $\Big ( \frac { 1 } { r f \sigma _ { j } ^ { * } } \Big ) ^ { 1 - \frac { 2 } { N } }$ 1− 2N . Solving for $\sigma _ { 0 }$ gives us the following: + +$$ +\sigma _ { 0 } \leq s \sigma _ { j } ^ { * } \Big ( \frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \Big ) ^ { \frac { N } { N - 2 } } +$$ + +# NECESSARY CONDITION + +To obtain a necessary (but not sufficient) condition on $\sigma _ { 0 }$ , we may relax the condition in (8) either by decreasing the left-hand side or increasing the right-hand side. We begin by rearranging the equation: + +$$ +\int _ { s \sigma _ { j } ^ { * } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } ) } \leq r \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { \sigma _ { j } ^ { * } } ) } - \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } ( 1 - \frac { \sigma } { r \sigma _ { j } ^ { * } } ) } +$$ + +Like before, we may use the integration bounds to bound $\sigma$ . Plugging in $\sigma = s \sigma _ { j } ^ { * }$ for all integrals decreases the left-hand side and increases the right-hand side, leading us to the following: + +$$ +\frac { r } { r - s } \int _ { s \sigma _ { j } ^ { * } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } } \le \big ( \frac { r } { 1 - s } - \frac { r } { r - s } \big ) \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } } +$$ + +Rearranging, we get the following inequality: + +$$ +\int _ { s \sigma _ { j } ^ { * } } ^ { f r \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } } \leq \frac { r - 1 } { 1 - s } \int _ { \sigma _ { 0 } } ^ { s \sigma _ { j } ^ { * } } \frac { d \sigma } { \sigma ^ { 2 - \frac { 2 } { N } } } +$$ + +We now solve the integrals for the different cases. For $N = 2$ , we have: + +$$ +\log ( f r \sigma _ { j } ^ { * } ) - \log ( s \sigma _ { j } ^ { * } ) \leq \frac { r - 1 } { 1 - s } \Bigl ( \log ( s \sigma _ { j } ^ { * } ) - \log ( \sigma _ { 0 } ) \Bigr ) +$$ + +Rearranging to isolate $\sigma _ { 0 }$ , we get our condition: + +$$ +\sigma _ { 0 } \leq s \sigma _ { j } ^ { * } \big ( \frac { s } { r f } \big ) ^ { \frac { 1 - s } { r - 1 } } +$$ + +Finally, for $N \geq 3$ , we solve the integrals to give us: + +$$ +\Big ( \big ( \frac { 1 } { s \sigma _ { j } ^ { * } } \big ) ^ { 1 - \frac { 2 } { N } } - \big ( \frac { 1 } { r f \sigma _ { j } ^ { * } } \big ) ^ { 1 - \frac { 2 } { N } } \Big ) \leq \frac { r - 1 } { 1 - s } \Big ( \big ( \frac { 1 } { \sigma _ { 0 } } \big ) ^ { 1 - \frac { 2 } { N } } - \big ( \frac { 1 } { s \sigma _ { j } ^ { * } } \big ) ^ { 1 - \frac { 2 } { N } } \Big ) +$$ + +Rearranging to isolate $\sigma _ { 0 }$ , we get our condition: + +$$ +\sigma _ { 0 } \leq s \sigma _ { j } ^ { * } \Big ( \frac { r - 1 } { r - s } \Big ) ^ { \frac { N } { N - 2 } } +$$ + +SUMMARY + +For a given $N$ , we derived a sufficient condition and a necessary condition on $\sigma _ { 0 }$ for $( s , f )$ - incremental learning. The necessary and sufficient condition on $\sigma _ { 0 }$ , which is the largest initialization value for which we see incremental learning (denoted $\sigma _ { 0 } ^ { t h \cdot }$ ), is between the two derived bounds. + +The precise bounds can possibly be improved a bit, but the asymptotic dependence on $r$ is the crux of the matter, showing the dependence on $r$ changes with depth with a substantial difference when we move from shallow models $N = 2$ ) to deeper ones $( N \geq 3$ ) + +# B PROOF OF THEOREM 3 + +Theorem. Given two values $\sigma _ { i } , \sigma _ { j }$ of a depth-2 toy linear model as in (1), such that $\begin{array} { r } { \frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } } = r > 1 } \end{array}$ and the model is initialized as in (3), and given two scalars $s \in ( 0 , \frac { 1 } { 4 } )$ and $f \in ( \frac { 3 } { 4 } , 1 )$ , and assuming $\sigma _ { j } ^ { * } ~ \geq ~ 2 \sigma _ { 0 }$ , and assuming we optimize with gradient descent with a learning rate $\begin{array} { r } { \eta \ \leq \ \frac { c } { \sigma _ { 1 } ^ { * } } } \end{array}$ for $c < 2 ( { \sqrt { 2 } } - 1 )$ and $\sigma _ { 1 } ^ { * }$ the largest value of $\sigma ^ { * }$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\sigma _ { 0 } ^ { t h }$ , is lower and upper bounded in the following way: + +$$ +\frac { 1 } { 2 } \frac { s } { 1 - s } \sigma _ { j } ^ { * } \Bigl ( \frac { 1 - f } { 2 r f } \frac { s } { 1 - s } \Bigr ) ^ { \frac { 1 } { A - 1 } } \leq \sigma _ { 0 } ^ { t h } \leq \frac { s } { 1 - s } \sigma _ { j } ^ { * } \Bigl ( \frac { 1 - f } { f } \frac { s } { 1 - s } \Bigr ) ^ { \frac { 1 } { B - 1 } } +$$ + +Where $A$ and $B$ are defined as: + +$$ +A = \frac { \log \left( 1 - c \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } + c ^ { 2 } \big ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } { \log \left( 1 - c \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } - \frac { c ^ { 2 } } { 4 } \big ( \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } \qquad B = \frac { \log \left( 1 - c \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } - \frac { c ^ { 2 } } { 4 } \big ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } { \log \left( 1 - c \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } + c ^ { 2 } \big ( \frac { \sigma _ { j } ^ { * } } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 } \right) } +$$ + +Proof. To show our result for gradient descent and $N = 2$ , we build on the proof techniques of theorem 3 of Gidel et al. (2019). We start by deriving the recurrence relation for the values $\sigma ( t )$ for general depth, when $t$ now stands for the iteration. Remembering that $w _ { i } ^ { n } = \sigma _ { i }$ , we write down the gradient update for $w _ { i } ( t )$ : + +$$ +w _ { i } ( t + 1 ) = w _ { i } ( t ) + \eta \frac { 1 } { N } w _ { i } ( t ) ^ { N - 1 } ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) = \sqrt [ N ] { \sigma _ { i } ( t ) } + \eta \frac { 1 } { N } \sigma _ { i } ( t ) ^ { 1 - \frac { 1 } { N } } ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) +$$ + +Raising $w _ { i } ( t )$ to the $N$ th power, we get the gradient update for the $\sigma$ values: + +$$ +r _ { i } ( t + 1 ) = \bigg ( \sqrt [ r ] { \sigma _ { i } ( t ) } + \eta \frac { 1 } { N } ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) \sigma _ { i } ( t ) ^ { 1 - \frac { 1 } { N } } \bigg ) ^ { N } = \sigma _ { i } ( t ) \Big ( 1 + \eta \frac { 1 } { N } \sigma _ { i } ( t ) ^ { 1 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) \Big ) ^ { N } +$$ + +Next, we will prove a simple lemma which gives us the maximal learning rate we will consider for the analysis, for which there is no overshooting (the values don’t grow larger than the optimal values). + +Lemma 1. For the gradient update in (9), assuming $\sigma _ { i } ( 0 ) < \sigma _ { i } ^ { * }$ , if $\begin{array} { r } { \eta \leq \left( \frac { 1 } { \sigma _ { 1 } ^ { * } } \right) ^ { 2 - \frac { 2 } { N } } } \end{array}$ and $N \geq 2$ , then: + +$$ +\forall t : \ \sigma _ { i } ( t ) \leq \sigma _ { i } ^ { * } +$$ + +Proof. Plugging in $\begin{array} { r } { \eta = c \bigg ( \frac { 1 } { \sigma _ { i } ^ { * } } \bigg ) ^ { 2 - \frac { 2 } { N } } } \end{array}$ for $c \leq 1$ , we have: + +$$ +\sigma _ { i } ( t + 1 ) = \sigma _ { i } ( t ) \Big ( 1 + \frac { c } { N } \Big ( \frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } } \Big ) ^ { 1 - \frac { 2 } { N } } \Big ( 1 - \big ( \frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } } \big ) \Big ) \Big ) ^ { N } \leq \sigma _ { i } ( t ) e ^ { \left( \frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } } \right) ^ { 1 - \frac { 2 } { N } } \Big ( 1 - \big ( \frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } } \big ) \Big ) } +$$ + +Defining ri = σiσ∗i a nd dividing both sides by $\sigma _ { i } ^ { * }$ , we have: + +$$ +r _ { i } ( t + 1 ) \leq r _ { i } ( t ) e ^ { r _ { i } ( t ) ^ { 1 - \frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) } +$$ + +It is enough to show that for any $0 \leq r \leq 1$ , we have that $r e ^ { r ^ { 1 - \frac { 2 } { N } } ( 1 - r ) } \leq 1$ , as over-shooting occurs when $r _ { i } ( \bar { t } ) > 1$ . Indeed, this function is monotonic increasing in $0 \leq r \leq 1$ (since the exponent is non negative), and equals 1 when $r = 1$ . Since $r = 1$ is a fixed point and no iteration that starts at $r < 1$ can cross 1, then $r _ { i } ( t ) \leq 1$ for any $t$ . This concludes our proof. □ + +Under this choice of learning rate, we can now obtain our incremental learning results for gradient descent when $N = 2$ . Our strategy will be bounding $\sigma _ { i } ( t )$ from below and above, which will give us a lower and upper bound for $t _ { \alpha } ( \sigma _ { i } )$ . Once we have these bounds, we will be able to describe either a necessary or a sufficient condition on $\sigma _ { 0 }$ for incremental learning, similar to theorem 2. + +The update rule for $N = 2$ is: + +$$ +\sigma _ { i } ( t + 1 ) = \sigma _ { i } ( t ) \Big ( 1 + \frac { 1 } { 2 } \eta ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) \Big ) ^ { 2 } +$$ + +Next, we plug in $\begin{array} { r } { \eta = \frac { c } { \sigma _ { 1 } ^ { * } } } \end{array}$ for $c < 2 ( \sqrt { 2 } - 1 ) < 1$ and denote $\begin{array} { r } { R _ { i } = \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } \leq 1 } \end{array}$ and $\begin{array} { r } { r _ { i } ( t ) = \frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } } \le 1 } \end{array}$ to get: + +$$ +\sigma _ { i } ( t + 1 ) = \sigma _ { i } ( t ) \Big ( 1 + \frac { c } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \Big ) ^ { 2 } +$$ + +Following theorem 3 of Gidel et al. (2019), we bound $\frac { 1 } { \sigma _ { i } ( t ) }$ : + +$$ +\begin{array} { r l } & { \frac { 1 } { \sigma _ { i } ( t + 1 ) } = \frac { 1 } { \sigma _ { i } ( t ) } \frac { 1 } { \left( 1 + \frac { \varepsilon } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \right) ^ { 2 } } } \\ & { \qquad = \frac { 1 } { \sigma _ { i } ( t ) } \frac { 1 } { 1 + \left( 1 - r _ { i } ( t ) \right) \left( \boldsymbol { { C R } } _ { i } + \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \left( 1 - r _ { i } ( t ) \right) \right) } } \\ & { \qquad \ge \frac { 1 } { \sigma _ { i } ( t ) } \frac { 1 } { 1 + \left( 1 - r _ { i } ( t ) \right) \left( \boldsymbol { { C R } } _ { i } + \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \right) } } \\ & { \qquad \ge \frac { 1 } { \sigma _ { i } ( t ) } \bigl ( 1 - ( 1 - r _ { i } ( t ) ) ( { c R } _ { i } + \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) \bigr ) } \\ & { \qquad = \frac { 1 } { \sigma _ { i } ( t ) } - ( \frac { 1 } { \sigma _ { i } ( t ) } - \frac { 1 } { \sigma _ { i } ^ { * } } ) ( { c R } _ { i } + \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) } \end{array} +$$ + +Where in the fourth line we use the inequality $\begin{array} { r } { \frac { 1 } { 1 + x } \geq 1 - x , \forall x \geq 0 } \end{array}$ . We may now subtract $\textstyle { \frac { 1 } { \sigma _ { i } ^ { * } } }$ from both sides to obtain: + +$$ +\frac { 1 } { \sigma _ { i } ( t ) } - \frac { 1 } { \sigma _ { i } ^ { * } } \geq \big ( \frac { 1 } { \sigma _ { i } ( t - 1 ) } - \frac { 1 } { \sigma _ { i } ^ { * } } \big ) \big ( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \big ) \geq \big ( \frac { 1 } { \sigma _ { 0 } } - \frac { 1 } { \sigma _ { i } ^ { * } } \big ) \big ( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \big ) ^ { t } +$$ + +We may now obtain a bound on $t _ { \alpha } ( \sigma _ { i } )$ by plugging in $\sigma _ { i } ( t ) = \alpha \sigma _ { i } ^ { * }$ and taking the log: + +$$ +\log \big ( \frac { 1 - \alpha } { \alpha \big ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 0 } } - 1 \big ) } \big ) \geq t _ { \alpha } \cdot \log \big ( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \big ) +$$ + +Rearranging (note that $\begin{array} { r } { \log \left( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \right) < 0 } \end{array}$ and that our choice of $c$ keeps the argument of the log positive), we get: + +$$ +t _ { \alpha } ( \sigma _ { i } ) \geq \frac { \log \big ( \frac { 1 - \alpha } { \alpha ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 0 } } - 1 ) } \big ) } { \log \big ( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \big ) } +$$ + +Next, we follow the same procedure for an upper bound. Starting with our update step: + +$$ +\begin{array} { r l r } { { \frac { 1 } { \sigma _ { i } ( t + 1 ) } = \frac { 1 } { \sigma _ { i } ( t ) } \frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) ( c R _ { i } + \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ( 1 - r _ { i } ( t ) ) ) } } } \\ & { } & { \leq \frac { 1 } { \sigma _ { i } ( t ) } \frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) c R _ { i } } } \\ & { } & { \leq \frac { 1 } { \sigma _ { i } ( t ) } ( 1 - ( 1 - r _ { i } ( t ) ) c R _ { i } + ( 1 - r _ { i } ( t ) ) ^ { 2 } c ^ { 2 } R _ { i } ^ { 2 } ) } \end{array} +$$ + +Where in the last line we use the inequality $\begin{array} { r } { \frac { 1 } { 1 + x } \leq 1 - x + x ^ { 2 } , \forall x \geq 0 } \end{array}$ . Subtracting $\textstyle { \frac { 1 } { \sigma _ { i } ^ { * } } }$ from both sides, we get: + +$$ +\frac { 1 } { \sigma _ { i } ( t ) } - \frac { 1 } { \sigma _ { i } ^ { * } } \leq \big ( \frac { 1 } { \sigma _ { i } ( t - 1 ) } - \frac { 1 } { \sigma _ { i } ^ { * } } \big ) \big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \big ) \leq \big ( \frac { 1 } { \sigma _ { 0 } } - \frac { 1 } { \sigma _ { i } ^ { * } } \big ) \big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \big ) ^ { t } +$$ + +Rearranging like before, we get the bound on the $\alpha$ -time: + +$$ +t _ { \alpha } ( \sigma _ { i } ) \leq \frac { \log \big ( \frac { 1 - \alpha } { \alpha ( \frac { \sigma _ { i } ^ { * } } { \sigma _ { 0 } } - 1 ) } \big ) } { \log \big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \big ) } +$$ + +Given these bounds, we would like to find the conditions on $\sigma _ { 0 }$ that allows for $( s , f )$ -incremental learning. We will find a sufficient condition and a necessary condition, like in the proof of theorem 2. + +# SUFFICIENT CONDITION + +A sufficient condition for incremental learning will be one which is possibly stricter than the exact condition. We can find such a condition by requiring the upper bound of $t _ { f } ( \sigma _ { i } )$ to be smaller than the lower bound on $t _ { s } ( \sigma _ { j } )$ . This becomes the following condition: + +$$ +\frac { \log \left( \frac { 1 - f } { f } \frac { \sigma _ { 0 } } { \sigma _ { i } ^ { * } - \sigma _ { 0 } } \right) } { \log \left( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \right) } \leq \frac { \log \left( \frac { 1 - s } { s } \frac { \sigma _ { 0 } } { \sigma _ { j } ^ { * } - \sigma _ { 0 } } \right) } { \log \left( 1 - c R _ { j } - \frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \right) } +$$ + +Defining $\begin{array} { r } { A = \frac { \log \big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \big ) } { \log \big ( 1 - c R _ { j } - \frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \big ) } } \end{array}$ and rearranging, we get the following: + +$$ +\log \big ( \frac { 1 - f } { f } \frac { \sigma _ { 0 } } { \sigma _ { i } ^ { * } - \sigma _ { 0 } } \big ) \geq A \log \big ( \frac { 1 - s } { s } \frac { \sigma _ { 0 } } { \sigma _ { j } ^ { * } - \sigma _ { 0 } } \big ) +$$ + +We may now take the exponent of both sides and rearrange again, remembering $\begin{array} { r } { \frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } } = r > 1 } \end{array}$ , to get the following condition: + +$$ +\frac { ( \sigma _ { j } ^ { * } - \sigma _ { 0 } ) ^ { A } } { \sigma _ { 0 } ^ { A - 1 } ( r \sigma _ { j } ^ { * } - \sigma _ { 0 } ) } \geq \frac { f } { 1 - f } ( \frac { 1 - s } { s } ) ^ { A } +$$ + +Now, we will add the very reasonable assumption that $\sigma _ { j } ^ { * } \geq 2 \sigma _ { 0 }$ , which allows us to replace $\frac { \sigma _ { j } ^ { * } - \sigma _ { 0 } } { r \sigma _ { j } ^ { * } - \sigma _ { 0 } }$ with $\textstyle { \frac { 1 } { 2 r } }$ and replace $( \sigma _ { j } ^ { * } - \sigma _ { 0 } ) ^ { A - 1 }$ with $\textstyle { \left( { \frac { 1 } { 2 } } \sigma _ { j } \right) } ^ { A - 1 }$ , only making the condition stricter. This simplifies the expression to the following: + +$$ +\big ( \frac { \sigma _ { j } ^ { * } } { \sigma _ { 0 } } \big ) ^ { A - 1 } \geq \frac { r f } { 1 - f } \big ( \frac { 2 - 2 s } { s } \big ) ^ { A } +$$ + +Now we can rearrange and isolate $\sigma _ { 0 }$ to get a sufficient condition for incremental learning: + +$$ +\sigma _ { 0 } \leq \frac { 1 } { 2 } \frac { s } { 1 - s } \sigma _ { j } ^ { * } \bigg ( \frac { 1 - f } { 2 r f } \frac { s } { 1 - s } \bigg ) ^ { \frac { 1 } { A - 1 } } +$$ + +# NECESSARY CONDITION + +A necessary condition for incremental learning will be one which is possibly more relaxed than the exact condition. We can find such a condition by requiring the lower bound of $t _ { f } ( \sigma _ { i } )$ to be smaller than the upper bound on $t _ { s } ( \sigma _ { j } )$ . This becomes the following condition: + +$$ +\frac { \log \left( \frac { 1 - f } { f } \frac { \sigma _ { 0 } } { \sigma _ { i } ^ { * } - \sigma _ { 0 } } \right) } { \log \left( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \right) } \leq \frac { \log \left( \frac { 1 - s } { s } \frac { \sigma _ { 0 } } { \sigma _ { j } ^ { * } - \sigma _ { 0 } } \right) } { \log \left( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \right) } +$$ + +Defining $\begin{array} { r } { B = \frac { \log \big ( 1 - c R _ { i } - \frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \big ) } { \log \big ( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \big ) } } \end{array}$ and rearranging, we get the following: + +$$ +\log \Big ( \frac { 1 - f } { f } \frac { \sigma _ { 0 } } { \sigma _ { i } ^ { * } - \sigma _ { 0 } } \Big ) \geq B \log \Big ( \frac { 1 - s } { s } \frac { \sigma _ { 0 } } { \sigma _ { j } ^ { * } - \sigma _ { 0 } } \Big ) +$$ + +We may now take the exponent of both sides and rearrange again, remembering $\begin{array} { r } { \frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } } = r > 1 } \end{array}$ , to get the following condition: + +$$ +\frac { ( \sigma _ { j } ^ { * } - \sigma _ { 0 } ) ^ { B } } { \sigma _ { 0 } ^ { B - 1 } ( r \sigma _ { j } ^ { * } - \sigma _ { 0 } ) } \geq \frac { f } { 1 - f } ( \frac { 1 - s } { s } ) ^ { B } +$$ + +We may now relax the condition further, by removing the $r$ from the denominator of the left-hand side and the $\sigma _ { 0 }$ from the numerator. This gives us the following: + +$$ +\big ( \frac { \sigma _ { j } ^ { * } } { \sigma _ { 0 } } \big ) ^ { B - 1 } \geq \frac { f } { 1 - f } \big ( \frac { 1 - s } { s } \big ) ^ { B } +$$ + +Finally, rearranging gives us the necessary condition: + +$$ +\sigma _ { 0 } \leq \frac { s } { 1 - s } \sigma _ { j } ^ { * } \Bigl ( \frac { 1 - f } { f } \frac { s } { 1 - s } \Bigr ) ^ { \frac { 1 } { B - 1 } } +$$ + +# C DISCUSSION OF GRADIENT DESCENT FOR GENERAL $N$ + +While we were able to generalize our result to gradient descent for $N = 2$ , our proof technique relies on the ability to get a non-implicit solution for $\sigma ( t )$ which we discretized and bounded. This is harder to generalize to larger values of $N$ , where the solution is implicit. Still, we can informally illustrate the effect of depth on the dynamics of gradient descent by approximating the update rule of the values. + +We start by reminding ourselves of the gradient descent update rule for $\sigma$ , for a learning rate $\eta =$ c 1σ ∗  2 − 2N : + +$$ +\sigma _ { i } ( t + 1 ) = \sigma _ { i } ( t ) \Big ( 1 + \frac { c } { N } \big ( \frac { 1 } { \sigma _ { 1 } ^ { * } } \big ) ^ { 2 - \frac { 2 } { N } } \sigma _ { i } ( t ) ^ { 1 - \frac { 2 } { N } } \big ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) \big ) \Big ) ^ { N } +$$ + +To compare two values in the same scales, we will divide both sides by the optimal value $\sigma _ { i } ^ { * }$ and look at the update step of the ratio $\begin{array} { r } { r _ { i } = \frac { \sigma _ { i } ( t ) } { \sigma _ { i } ^ { * } } } \end{array}$ , also denoting $\begin{array} { r } { R _ { i } = \frac { \sigma _ { i } ^ { * } } { \sigma _ { 1 } ^ { * } } } \end{array}$ : + +$$ +r _ { i } ( t + 1 ) = r _ { i } ( t ) \Big ( 1 + \frac { c } { N } R _ { i } ^ { 2 - \frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) \Big ) ^ { N } +$$ + +We will focus on the early stages of the optimization process, where $r \ll 1$ . This means we can neglect the $1 - r _ { i } ( t )$ term in the update step, giving us the approximate update step we will use to compare the general $i , j$ values: + +$$ +r _ { i } ( t + 1 ) \approx r _ { i } ( t ) \left( 1 + \frac { c } { N } R _ { i } ^ { 2 - \frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \frac { 2 } { N } } \right) ^ { N } +$$ + +We would like to compare the dynamics of $r _ { i }$ and $r _ { j }$ , which is difficult to do when the recurrence relation isn’t solvable. However, we can observe the first iteration of gradient descent and see how depth affects this iteration. Since we are dealing with variables which are ratios of different optimal values, the initial values of $r$ are different. Denoting $\begin{array} { r } { \mathbf { r } = \frac { \sigma _ { i } ^ { * } } { \sigma _ { j } ^ { * } } } \end{array}$ , we can describe the initialization of $r _ { j }$ using that of $r _ { i }$ : + +$$ +r _ { j } ( 0 ) = { \bf r } r _ { i } ( 0 ) +$$ + +Plugging in the initial conditions and noting that $R _ { i } = { \bf r } R _ { j }$ , we get: + +$$ +\begin{array} { r } { r _ { i } ( 1 ) \approx r _ { i } ( 0 ) \Big ( 1 + \frac { c R _ { i } ^ { 2 - \frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \frac { 2 } { N } } \Big ) ^ { N } } \\ { r _ { j } ( 1 ) \approx r _ { i } ( 0 ) \Big ( \sqrt [ N ] { \mathbf { r } } + \big ( \frac { 1 } { \mathbf { r } } \big ) ^ { \frac { N - 1 } { N } } \frac { c R _ { i } ^ { 2 - \frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \frac { 2 } { N } } \Big ) ^ { N } } \end{array} +$$ + +We see that the two ratios have a similar update, with the ratio of optimal values playing a role in how large the initial value is versus how large the added value is. When we use a small learning rate, we have a very small $c$ which means we can make a final approximation and neglect the higher order terms of $c$ : + +$$ +\begin{array} { r l } & { r _ { i } ( 1 ) \approx r _ { i } ( 0 ) + c R _ { i } ^ { 2 - \frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \frac { 2 } { N } } } \\ & { r _ { j } ( 1 ) \approx { \bf r } r _ { i } ( 0 ) + \big ( \frac { 1 } { { \bf r } } \big ) ^ { \frac { N - 1 } { N } } c R _ { i } ^ { 2 - \frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \frac { 2 } { N } } } \end{array} +$$ + +We can see that while the initial conditions favor $r _ { j }$ , the size of the update for $r _ { i }$ is larger by a factor of $r ^ { \frac { N - 1 } { N } }$ when the initialization and learning rates are small. This accumulates throughout the optimization, making $r _ { i }$ eventually converge faster than $r _ { j }$ . + +The effect of depth here is clear - the deeper the model, the larger the relative step size of $r _ { i }$ and the faster it converges relative to $r _ { j }$ . + +# D COMPARISON OF THE TOY MODEL AND OMP + +Learning our toy model, when it’s incremental learning is taken to the limit, can be described as an iterative procedure where at every step an additional feature is introduced such that it’s weight is non-zero and then the model is optimized over the current set of features. This description is also relevant for the sparse approximation algorithm orthogonal matching pursuit (Pati et al., 1993), where the next feature is greedily chosen to be the one which most improves the current model. + +While the toy model and OMP are very different algorithms for learning sparse linear models, we will show empirically that they behave similarly. This allows us to view incremental learning as a continuous-time extension of a greedy iterative algorithm. + +To allow for negative weights in our experiments, we augment our toy model as in the toy model of Woodworth et al. (2019). Our model will have the same induced form as before: + +$$ +f _ { \sigma } ( x ) = \langle \sigma , x \rangle +$$ + +However, we parameterize $\sigma$ using $w _ { + } , w _ { - } \in \mathbb { R } ^ { d }$ in the following way: + +$$ +\begin{array} { c } { \sigma _ { i } = w _ { + , i } ^ { N } - w _ { - , i } ^ { N } } \\ { \qquad \forall i , \ w _ { + , i } ( 0 ) = w _ { - , i } ( 0 ) = \sqrt [ N ] { \sigma _ { 0 } } } \end{array} +$$ + +We can now treat this algorithm as a sparse approximation pursuit algorithm - given a dictionary $D \in \mathbb { R } ^ { d \times n }$ and an example $x \in \mathbb { R } ^ { d }$ , we wish to find the sparsest $\alpha$ for which $D \alpha \approx x$ by minimizing the $\ell _ { 0 }$ norm of $\alpha$ subject to $\| D \alpha - x \| _ { 2 } ^ { 2 } = 0 ^ { 3 }$ . Under this setting, we can compare OMP to our toy model by comparing the sets of features that the two algorithms choose for a given example and dictionary. + +In figure 3 we run such a comparison. Using a dictionary of 1000 atoms and an example of dimensionality 80 sampled from a random hidden vector of a given sparsity $s$ , we run both algorithms and record the first $s$ features chosen4. + +![](images/850628cf8a6350a15f50e62eeedfcba28e4ac2fa035fa393674ec77b7c764190.jpg) +Figure 3: Empirical comparison of the dynamics of the toy model to OMP. The toy model has a depth of 5 and was initialized with a scale of 1e-4 and a learning rate of 3e-3. We compare the fraction of agreement between the sets of first $s$ features selected of the two algorithms for every given sparsity level $s$ , averaged over 100 experiments (the shaded regions are empirical standard deviations). For example, for sparsity level 3, we look at the sets of first 3 features selected by each algorithm and calculate the fraction of them that appear in both sets. + +For every sparsity $s$ , we plot the mean fraction of agreement between the sets of features chosen by OMP and the toy model over 100 experiments. We see that the two algorithms choose very similar features at the beginning, suggesting that the deep model approximates the discrete behavior of OMP. Only when the number of features increases do we see that the behavior of the two models begins to differ, caused by the fact that the toy model has a finite initialization scale and learning rate. + +These experiments demonstrate the similarity between the incremental learning of deep models and the discrete behavior of greedy approximation algorithms such as OMP. Adopting this view also allows us to put our finger on another strength of the dynamics of deep models - while greedy algorithms such as OMP require the analytical solution or approximation of every iterate, the dynamics of deep models are able to incrementally learn any differentiable function. For example, looking back at the matrix sensing task and the classification models in section 4, we see that while there isn’t an immediate and efficient extension of OMP for these settings, the dynamics of learning deep models extends naturally and exhibits the same incremental learning as OMP. + +# E INCREMENTAL LEARNING IN MATRIX SENSING + +# E.1 PROOF OF THEOREM 4 + +Theorem. Minimizing the deep matrix sensing model described in (5) with gradient flow over the depth normalized squared loss (6), with Gaussian inputs and weights initialized as in (5) leads to the following dynamical equations for different values of $N$ : + +$$ +\dot { \sigma } _ { i } ( t ) = \sigma _ { i } ( t ) ^ { 2 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) +$$ + +Where $\sigma _ { i }$ and $\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vectors. + +Proof. We will adapt the proof from Saxe et al. (2013) for multilayer linear networks. The gradient flow equations for $W _ { n } , n \in [ N ]$ are: + +$$ +\dot { W _ { n } } = { \frac { 1 } { N } } W _ { 1 : n - 1 } ^ { T } ( W ^ { * } - W ) W _ { n + 1 : N } ^ { T } +$$ + +Where we denote $\begin{array} { r } { W _ { j : k } = \prod _ { i = j } ^ { k } W _ { i } } \end{array}$ + +Since we assumed $W ^ { * }$ is (symmetric) PSD, there exists an orthogonal matrix $U$ for which $D ^ { * } =$ $U W ^ { * } U ^ { T }$ where $D ^ { * }$ is diagonal. Under the initialization in (3), $U$ diagonalizes all $W _ { n }$ matrices at initialization such that ${ \cal D } _ { n } \overline { { } } = U W _ { n } U ^ { T } = \sqrt [ N ] { \sigma _ { 0 } } I$ . Making this change of variables for all $W _ { n }$ , we get: + +$$ +\dot { W _ { n } } = \frac { 1 } { N } U ^ { T } D _ { 1 : n - 1 } U ( W ^ { * } - W ) U ^ { T } D _ { n + 1 : N } U +$$ + +Rearranging, we get a set of decoupled differential equations for the singular values of $W _ { n }$ + +$$ +\dot { D _ { n } } = { \frac { 1 } { N } } D _ { 1 : n - 1 } ( D ^ { * } - D ) D _ { n + 1 : N } +$$ + +Note that since these matrices are all diagonal at initialization, the above dynamics ensure that they remain diagonal throughout the optimization. Denoting $\sigma _ { n , i }$ as the $i ^ { \because }$ ’th singular value of $W _ { n }$ and $\sigma _ { i }$ as the $\overrightarrow { \imath } ^ { \prime }$ ’th singular value of $W$ , we get the following differential equation: + +$$ +\dot { \sigma } _ { n , i } = \frac { 1 } { N } ( \sigma _ { i } ^ { * } - \sigma _ { i } ) \prod _ { j \neq n } \sigma _ { j , i } +$$ + +Since we assume at initialization that $\forall n , m , i : \sigma _ { n , i } ( 0 ) = \sigma _ { m , i } ( 0 ) = \sqrt [ N ] { \sigma _ { 0 } }$ , the above dynamics are the same for all singular values and we get $\forall n , m , i : \sigma _ { n , i } ( t ) = \sigma _ { m , i } ( t ) = \sqrt [ N ] { \sigma _ { i } ( t ) }$ . We may now use this to calculate the dynamics of the singular value of $W$ , since they are the product the the singular values of all $W _ { n }$ matrices. Denoting $\begin{array} { r } { \sigma _ { - n , i } = \prod _ { k \neq n } \sigma _ { k , i } } \end{array}$ and using the chain rule: + +$$ +\dot { \sigma } _ { i } = \sum _ { n = 1 } ^ { N } \sigma _ { - n , i } \dot { \sigma } _ { n , i } = \sigma _ { i } ^ { 2 - \frac { 2 } { N } } ( \sigma _ { i } ^ { * } - \sigma _ { i } ) +$$ + +# E.2 EMPIRICAL EXAMINATION + +Our analytical results are only applicable for the population loss over Gaussian inputs. These conditions are far from the ones used in practice and studied in Arora et al. (2019), where the problem is over-determined and the weights are drawn from a Gaussian distribution with a small variance. To show our conclusions regarding incremental learning extend qualitatively to more natural settings, we empirically examine the deep matrix sensing model in this natural setting for different depths and initialization scales as seen in figure 4. + +Notice how incremental learning is exhibited even when the number of examples is much smaller than the number of parameters in the model. While we can’t rely on our theory for describing the exact dynamics of the optimization for these kinds of over-determined problems, the qualitative conclusions we get from it are still applicable. + +Another interesting phenomena we should note is that once the dataset becomes very small (the second row of the figure), we see all “currently active” singular values change at the beginning of every new phase (this is best seen in the bottom-right panel). This suggests that since there is more than one optimal solution, once we increase the current rank of our model it may find a solution that has a different set of singular values and vectors and thus all singular values change at the beginning of a new learning phase. This demonstrates the importance of incremental learning for obtaining sparse solutions - once the initialization conditions and depth are such that the learning phases are distinct, gradient descent finds the optimal rank- $i$ solution in every phase $i$ . For these dynamics to successfully recover the optimal solution at every phase, the phases need to be far enough apart from each other to allow for the singular values and vectors to change before the next phase begins. + +![](images/30d0053940d521183eae431a6073eb95e1a63934ce44ff4a2a09d223b98048e2.jpg) +Figure 4: Evolution of the top-5 singular values of the deep matrix sensing model, with Gaussian initialization with variance such that the initial singular values are in expectation 1e-4. The model’s size and data are in $\mathbb { R } ^ { 5 0 \times 5 0 }$ . The columns correspond to different parameterization depths, while the rows correspond to different dataset sizes. In both cases the problem is over-determined, since the number of examples is smaller than the number of parameters. Since the original matrix is rank-4, we can recognize an unsuccessful recovery when all five singular values are nonzero, as seen clearly for both depth-1 plots. + +F INCREMENTAL LEARNING IN QUADRATIC NETWORKS + +F.1 PROOF OF THEOREM 5 + +Theorem. Minimizing the quadratic network described and initialized as in (7) with gradient flow over the variance loss defined in (2) with Gaussian inputs leads to the following dynamical equations: + +$$ +\dot { \sigma } _ { i } ( t ) = \sigma _ { i } ( t ) ( \sigma _ { i } ^ { * } - \sigma _ { i } ( t ) ) +$$ + +Where $\sigma _ { i }$ and $\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vectors. + +Proof. Our proof will follow similar lines as the analysis of the deep matrix sensing model. Taking the expectation of the variance loss over Gaussian inputs for our model gives us: + +$$ +\begin{array} { l } { \displaystyle \ell _ { v a r } ( W ) = \frac { 1 } { 1 6 } \mathbb { E } _ { { x } } [ ( \langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \rangle ) ^ { 2 } ] - \frac { 1 } { 1 6 } \mathbb { E } _ { { x } } [ \langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \rangle ] ^ { 2 } } \\ { \displaystyle = \frac { 1 } { 8 } \| W _ { * } ^ { T } W _ { * } - W ^ { T } W \| _ { F } ^ { 2 } + \frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } - \frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\ { \displaystyle \quad = \frac { 1 } { 8 } \| W _ { * } ^ { T } W _ { * } - W ^ { T } W \| _ { F } ^ { 2 } } \end{array} +$$ + +Following the gradient flow dynamics over $W$ leads to the following differential equation: + +$$ +\dot { W } = \frac { 1 } { 2 } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) +$$ + +We can now calculate the gradient flow dynamics of $W ^ { T } W$ using the chain rule: + +$$ +W ^ { \dot { T } } W = W ^ { T } \dot { W } + \dot { W } ^ { T } W = \frac { 1 } { 2 } \Bigl ( W ^ { T } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) + ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) W ^ { T } W \Bigr ) +$$ + +Now, under our initialization $W _ { 0 } ^ { T } W _ { 0 } = \sigma _ { 0 } I \mathrm { . }$ , we get that $W ^ { T } W$ and $W _ { * } ^ { T } W _ { * }$ are simultaneously diagonalizable at initialization by some matrix $U$ , such that the following is true for diagonal $D$ and $D ^ { * }$ : + +$$ +D ( 0 ) = U ^ { T } W ( 0 ) ^ { T } W ( 0 ) U +$$ + +$$ +D ^ { * } = U ^ { T } W _ { * } ^ { T } W _ { * } U +$$ + +Multiplying equation (10) by $U$ and $U ^ { T }$ gives us the following dynamics for the singular values of $W ^ { T } W$ : + +$$ +\dot { D } = \frac { 1 } { 2 } \Big ( D ( D ^ { * } - D ) + ( D ^ { * } - D ) D \Big ) = D ( D ^ { * } - D ) +$$ + +These matrices are diagonal at initialization, and remain diagonal throughout the dynamics (the offdiagonal elements are static according to these equations). We may now look at the dynamics of a single diagonal element, noticing it is equivalent to the depth-2 toy model: + +$$ +\dot { \sigma } _ { i } = \sigma _ { i } ( \sigma _ { i } ^ { * } - \sigma _ { i } ) +$$ + +# F.2 DISCUSSION OF THE VARIANCE LOSS + +It may seem that the variance loss is an unnatural loss function to analyze, since it isn’t used in practice. While this is true, we will show how the dynamics of this loss function are an approximation of the square loss dynamics. + +We begin by describing the dynamics of both losses, showing how incremental learning can’t take place for quadratic networks as defined over the squared loss. Then, we show how adding a global bias to the quadratic network leads to similar dynamics for small initialization scales. + +# F.2.1 DYNAMICAL DERIVATION + +in the previous section, we derive the differential equations for the singular values of $W ^ { T } W$ under the variance loss: + +$$ +\dot { \sigma } _ { i } = \sigma _ { i } ( \sigma _ { i } ^ { * } - \sigma _ { i } ) +$$ + +We will now derive similar equations for the squared loss. The scaled squared loss in expectation over the Gaussian inputs is: + +$$ +\begin{array} { r l } & { \ell ( W ) = \displaystyle \frac { 1 } { 1 6 } \mathbb { E } _ { x } [ ( \langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \rangle ) ^ { 2 } ] } \\ & { \qquad = \displaystyle \frac { 1 } { 8 } | | W _ { * } ^ { T } W _ { * } - W ^ { T } W | | _ { F } ^ { 2 } + \frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \end{array} +$$ + +Defining $\Delta = W _ { * } ^ { T } W _ { * } - W ^ { T } W$ for brevity, the differential equations for $W$ become: + +$$ +\dot { W } = \frac { 1 } { 2 } W \Delta + \frac { 1 } { 4 } T r ( \Delta ) W +$$ + +Calculating the dynamics of $W ^ { T } W$ after noting that it is simultaneously diagonalizable with $W _ { * } ^ { T } W _ { * }$ (as in the derivation for the variance loss) leads to the following differential equations for the singular values of $W ^ { T } W$ : + +$$ +\dot { \sigma } _ { i } = \sigma _ { i } \Big ( \sigma _ { i } ^ { * } - \sigma _ { i } + \frac { 1 } { 2 } \sum _ { j } ( \sigma _ { j } ^ { * } - \sigma _ { j } ) \Big ) +$$ + +We see that the equations are now coupled and so we cannot solve them analytically. Another issue is that for our initialization, all singular values have very similar dynamics at initialization due to the coupling. For example, values corresponding to small optimal singular values grow much faster than in the variance loss dynamics, due to the effect the large optimal singular values have on them. + +We see from these equations that we shouldn’t expect quadratic networks optimized over the squared loss to exhibit incremental learning behavior. We next show how adding a global bias to our model can help. + +# F.2.2 VARIANCE LOSS APPROXIMATES THE SQUARED LOSS + +To see how the variance loss can have dynamics resembling those of the squared loss, we will add a global (trainable) bias to our model. This means our model is now parameterized by $W \in \mathbb { R } ^ { d \times d }$ and a scalar $b \in \mathbb { R }$ : + +$$ +f _ { W , b } ( x ) = \langle W ^ { T } W , x x ^ { T } \rangle + b +$$ + +We may now analyze gradient flow of the squared loss. Following the same methods as before, this leads us to the following differential equations: + +$$ +\begin{array} { l } { \displaystyle \dot { \sigma } _ { i } = \sigma _ { i } \Big ( \sigma _ { i } ^ { * } - \sigma _ { i } + \frac { 1 } { 2 } \sum _ { j } ( \sigma _ { j } ^ { * } - \sigma _ { j } ) + \frac { 1 } { 2 } ( b ^ { * } - b ) \Big ) } \\ { \displaystyle \dot { b } = \sum _ { j } ( \sigma _ { j } ^ { * } - \sigma _ { j } ) + b ^ { * } - b } \end{array} +$$ + +Notice how, if $b$ is at it’s optimum at a given time $\begin{array} { r } { ( b = \sum _ { j } ( \sigma _ { j } ^ { * } - \sigma _ { j } ) + b ^ { * } ) } \end{array}$ , the dynamics of $\sigma _ { i }$ align with those of the variance loss. Alternatively, when $b = b ^ { * }$ , we recover the dynamics of the squared loss without the global bias term. + +To convince ourselves that the dynamics of this model resemble those of the variance loss, we would need to explain why the global bias is at it’s optimum “most of the time”, such that the singular values don’t change much during the times when it is not at it’s optimum. + +Observing the differential equations for $b$ and for $\sigma _ { i }$ , we see they are similar (if we ignore the $\sigma _ { i } ^ { * } - \sigma _ { i }$ value which doesn’t change the order of magnitude of the entire expression when there haven’t been many learning phases yet). The only difference being a multiplication by $\sigma _ { i }$ . This means that we may informally write: + +$$ +\frac { \dot { \sigma } _ { i } ( t ) } { \dot { b } ( t ) } \approx \sigma _ { i } ( t ) +$$ + +Since at initialization and until the learning phase of $\sigma _ { i }$ takes place, we have $\sigma _ { i } ( t ) \ll 1$ , we see that the global bias optimizes much faster than the singular values for which the learning phase hasn’t begun yet. This means these singular values will remain small during the times in which the bias isn’t optimal, and so incremental learning can still take place (assuming a small enough initialization). + +![](images/0ec431bf9815c02f0689d4f67bcd2532ec6e6174e98a709848c535d46e23f897.jpg) +Figure 5: Quadratic model’s evolution of top-5 singular values for a rank-4 labeling function. The rows correspond to whether or not a global bias is introduced to the model. The first two columns are for a large dataset (one optimal solution) and the last two columns are for a small dataset (overdetermined problem). When a bias is introduced, it is initialized to it’s optimal value at initialization. Note how without the bias, the singular values are learned together and there is over-shooting of the optimal singular value caused by the coupling of the dynamics of the singular values. For the small datasets, we see that the model with no bias reaches a solution with a larger rank. Once a global bias is introduced, the dynamics become more incremental as in the analysis of the variance loss. Note that in this case the solution obtained for the small dataset is the optimal low-rank solution. + +Under these considerations, we say that the dynamics of the squared loss for a quadratic network with an added global bias resemble the idealized dynamics of the variance loss for a depth-2 linear model which we analyze formally in the paper. In figure 5 we experimentally show how adding a bias to a quadratic network does lead to incremental learning similar to the depth-2 toy model. + +# G INCREMENTAL LEARNING IN CLASSIFICATION + +# G.1 DIAGONAL NETWORKS + +In section 4.3 we viewed our toy model as a special case of the deep diagonal networks described in Gunasekar et al. (2018), expected to be biased towards sparse solutions. Figure 6 shows the dynamics of the largest values of $\sigma$ for different depths of the model. We see that the same type of incremental learning we saw in earlier models exists here as well - the features are learned one by one in deeper models, resulting in a sparse solution. The leftmost panel shows how the initialization scale plays a role here as well, with the solution being more sparse when the initialization is small. We should note that these results do not defy the results of Gunasekar et al. (2018) (from which we would expect the initialization not to matter), since their results deal with the solution at $t \to \infty$ . + +# G.2 CONVOLUTIONAL NETWORKS - PRELIMINARIES + +The linear circular-convolutional network of Gunasekar et al. (2018) deals with one-dimensional convolutions with the same number of outputs as inputs, such that the mapping from one hidden layer to the next is parameterized by $w _ { n }$ and defined to be: + +$$ +h _ { n } [ i ] = \sum _ { k = 0 } ^ { d - 1 } w _ { n } [ k ] h _ { n - 1 } [ ( i + k ) \mod d ] = ( h _ { n - 1 } \star w _ { n } ) [ i ] +$$ + +![](images/f24003e92c57d90fdbdbc73d05bd3dd33d93af75d8ddeff5e4ea5313c8d7c9a4.jpg) +Figure 6: Incremental learning in binary classification. A model as in section 4.3 is trained over 200 i.i.d. random Gaussian examples, where $d = 1 0 0$ . The data is labeled by a weight vector with 4 nonzero values, making the problem realizable with a sparse solution while the max-margin solution isn’t sparse. The left panel describes the obtained solution’s correlation with the sparse labeling vector for different depths and initializations. The results are averaged over 100 experiments, with shaded regions denoting empirical standard deviations. We see that depth-1 models reach results similar to the max-margin SVM solution as predicted by Gunasekar et al. (2018), while deeper models are highly correlated with the sparse solution, with this correlation increasing when the initialization scale is small. The other panels show the evolution of the absolute values of the top-5 weights of $\sigma$ for the smallest initialization scale. Note that as we increase the depth, incremental learning is clearly presented. + +The final layer is a fully connected layer parameterized by $w _ { N } \in \mathbb { R } ^ { d }$ , such that the final model can be written in the following way: + +$$ +f _ { \sigma } ( x ) = ( ( ( ( x \star w _ { 1 } ) \star w _ { 2 } ) \cdot \cdot \cdot ) \star w _ { N - 1 } ) ^ { T } w _ { N } +$$ + +Lemma 3 from Gunasekar et al. (2018) shows how we can relate the Fourier coefficients of the weight vectors to the Fourier coefficients of the linear model induced by the model: + +Lemma. For the circular-convolutional model as in (11): + +$$ +\hat { \sigma } = d i a g ( \hat { w } _ { 1 } ) \cdot \cdot \cdot d i a g ( \hat { w } _ { N - 1 } ) \hat { w } _ { N } , +$$ + +where for $n = 1 , 2 , . . . , N , \hat { w } _ { n } \in \mathbb { C } ^ { d }$ are the Fourier coefficients of the parameters $w _ { n } \in \mathbb { R } ^ { d }$ + +This lemma connects the convolutional network to the diagonal network, and thus we should expect to see the same incremental learning of the values of the diagonal network exhibited by the Fourier coefficients of the convolutional network. + +# G.3 CONVOLUTIONAL NETWORKS - EMPIRICAL EXAMINATION + +In figure 7 we see the same plots as in figure 6 but for the Fourier coefficients of the convolutional model. We see that even when the model is far from the toy parameterization (there is no weight sharing and the initialization is with random Gaussian weights), incremental learning is still clearly seen in the dynamics of the model. We see how the inherent reason for the sparsity towards sparse solution found in Gunasekar et al. (2018) is the result of the dynamics of the model - small amplitudes are attenuated while large ones are amplified. + +![](images/ce17411c14f982ed90cef0de02508e3a99c1bff2ecdf153320a808745f475e11.jpg) +Figure 7: Incremental learning in convolutional networks. A model as in appendix G is trained over 200 i.i.d. random Gaussian examples, where $d = 1 0 0$ . The weights are initialized randomly and the data is labeled by a weight vector with 4 nonzero frequencies, making the problem realizable with a sparse solution in the frequency domain. The left panel describes the obtained solution’s correlation in the frequency domain with the sparse labeling vector for different depths and initializations. The results are averaged over 9 experiments, with shaded regions denoting empirical standard deviations. We see that depth-1 models reach results similar to the max-margin SVM solution, while deeper models are highly correlated with the optimal sparse solution. The other panels show the evolution of the amplitudes of the top-5 frequencies of $\sigma$ for the smallest initialization scale. Note that as we increase the depth, incremental learning is clearly presented. \ No newline at end of file diff --git a/parse/train/H1lj0nNFwB/H1lj0nNFwB_content_list.json b/parse/train/H1lj0nNFwB/H1lj0nNFwB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6feb5d4433ca3b25d6ca8f271be9840f7e72c0a4 --- /dev/null +++ b/parse/train/H1lj0nNFwB/H1lj0nNFwB_content_list.json @@ -0,0 +1,3943 @@ +[ + { + "type": "text", + "text": "THE IMPLICIT BIAS OF DEPTH: HOW INCREMENTAL LEARNING DRIVES GENERALIZATION ", + "text_level": 1, + "bbox": [ + 174, + 98, + 820, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Gissin, Shai Shalev-Shwartz, Amit Daniely \nSchool of Computer Science \nThe Hebrew University \nJerusalem, Israel \n{daniel.gissin,shais,amit.daniely}@mail.huji.ac.il ", + "bbox": [ + 184, + 169, + 669, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 276, + 544, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A leading hypothesis for the surprising generalization of neural networks is that the dynamics of gradient descent bias the model towards simple solutions, by searching through the solution space in an incremental order of complexity. We formally define the notion of incremental learning dynamics and derive the conditions on depth and initialization for which this phenomenon arises in deep linear models. Our main theoretical contribution is a dynamical depth separation result, proving that while shallow models can exhibit incremental learning dynamics, they require the initialization to be exponentially small for these dynamics to present themselves. However, once the model becomes deeper, the dependence becomes polynomial and incremental learning can arise in more natural settings. We complement our theoretical findings by experimenting with deep matrix sensing, quadratic neural networks and with binary classification using diagonal and convolutional linear networks, showing all of these models exhibit incremental learning. ", + "bbox": [ + 233, + 309, + 764, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 530, + 336, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural networks have led to a breakthrough in modern machine learning, allowing us to efficiently learn highly expressive models that still generalize to unseen data. The theoretical reasons for this success are still unclear, as the generalization capabilities of neural networks defy the classic statistical learning theory bounds. Since these bounds, which depend solely on the capacity of the learned model, are unable to account for the success of neural networks, we must examine additional properties of the learning process. One such property is the optimization algorithm - while neural networks can express a multitude of possible ERM solutions for a given training set, gradient-based methods with the right initialization may be implicitly biased towards certain solutions which generalize. ", + "bbox": [ + 174, + 563, + 823, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A possible way such an implicit bias may present itself, is if gradient-based methods were to search the hypothesis space for possible solutions of gradually increasing complexity. This would suggest that while the hypothesis space itself is extremely complex, our search strategy favors the simplest solutions and thus generalizes. One of the leading results along these lines has been by Saxe et al. (2013), deriving an analytical solution for the gradient flow dynamics of deep linear networks and showing that for such models, the singular values converge at different rates, with larger values converging first. At the limit of infinitesimal initialization of the deep linear network, Gidel et al. (2019) show these dynamics exhibit a behavior of “incremental learning” - the singular values of the model are learned separately, one at a time. Our work generalizes these results to small but finite initialization scales. ", + "bbox": [ + 174, + 680, + 825, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incremental learning dynamics have also been explored in gradient descent applied to matrix completion and sensing with a factorized parameterization (Gunasekar et al. (2017), Arora et al. (2018), Woodworth et al. (2019)). When initialized with small Gaussian weights and trained with a small learning rate, such a model is able to successfully recover the low-rank matrix which labeled the data, even if the problem is highly over-determined and no additional regularization is applied. In their proof of low-rank recovery for such models, Li et al. (2017) show that the model remains lowrank throughout the optimization process, leading to the successful generalization. Additionally, ", + "bbox": [ + 174, + 827, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14126dc03eaae57887ee79a34684cde80a2ac3af64c8443f5095c0fc7eaad71c.jpg", + "image_caption": [ + "Figure 1: Incremental learning dynamics in deep models. Each panel shows the evolution of the five largest values of $\\sigma$ , the parameters of the induced model. All models were trained using gradient descent with a small initialization and learning rate, on a small training set such that there are multiple possible solutions. In all cases, the deep parameterization of the models lead to “incremental learning”, where the values are learned at different rates (larger values are learned first), leading to sparse solutions. (a) Depth 4 matrix sensing, $\\sigma$ denotes singular values (see section 4.1). (b) Quadratic networks, $\\sigma$ denotes singular values (see section 4.2). (c) Depth 3 diagonal networks, $\\sigma$ denotes feature weights (see section 4.3). (d) Depth 3 circular-convolutional networks, $\\sigma$ denotes amplitudes in the frequency domain of the feature weights (see appendix G). " + ], + "image_footnote": [], + "bbox": [ + 184, + 104, + 808, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Arora et al. (2019) explore the dynamics of such models, showing the singular values are learned at different rates and that deeper models exhibit stronger incremental learning dynamics. Our work deals with a more simplified setting, allowing us to determine explicitly under which conditions depth leads to this dynamical phenomenon. ", + "bbox": [ + 174, + 401, + 825, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, the learning dynamics of nonlinear models have been studied as well. Combes et al. (2018) and Williams et al. (2019) study the gradient flow dynamics of shallow ReLU networks under restrictive distributional assumptions, Ronen et al. (2019) show that shallow networks learn functions of gradually increasing frequencies and Nakkiran et al. (2019) show how deep ReLU networks correlate with linear classifiers in the early stages of training. ", + "bbox": [ + 174, + 465, + 823, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These findings, along with others, suggest that the generalization ability of deep networks is at least in part due to the incremental learning dynamics of gradient descent. Following this line of work, we begin by explicitly defining the notion of incremental learning for a toy model which exhibits this sort of behavior. Analyzing the dynamics of the model for gradient flow and gradient descent, we characterize the effect of the model’s depth and initialization scale on incremental learning, showing how deeper models allow for incremental learning in larger (realistic) initialization scales. Specifically, we show that a depth-2 model requires exponentially small initialization for incremental learning to occur, while deeper models only require the initialization to be polynomially small. ", + "bbox": [ + 174, + 542, + 825, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once incremental learning has been defined and characterized for the toy model, we generalize our results theoretically and empirically for larger linear and quadratic models. Examples of incremental learning in these models can be seen in figure 1, which we discuss further in section 4. ", + "bbox": [ + 176, + 660, + 825, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 DYNAMICAL ANALYSIS OF A TOY MODEL ", + "text_level": 1, + "bbox": [ + 174, + 723, + 555, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by analyzing incremental learning for a simple model. This will allow us to gain a clear understanding of the phenomenon and the conditions for it, which we will later be able to apply to a variety of other models in which incremental learning is present. ", + "bbox": [ + 174, + 756, + 825, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 174, + 818, + 326, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our simple linear model will be similar to the toy model analyzed by Woodworth et al. (2019). Our input space will be $\\mathcal { X } = \\mathbb { R } ^ { d }$ and the hypothesis space will be linear models with non-negative weights, such that: ", + "bbox": [ + 174, + 844, + 825, + 886 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/499720dd31f433eff0df02a5e5569bd3b2c566437ae3c79f1dbed5637444a0cd.jpg", + "text": "$$\nf _ { \\sigma } ( { \\boldsymbol { x } } ) = \\langle \\sigma , { \\boldsymbol { x } } \\rangle \\qquad \\sigma \\in \\mathbb { R } _ { \\geq 0 } ^ { d }\n$$", + "text_format": "latex", + "bbox": [ + 401, + 906, + 596, + 926 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We will introduce depth into our model, by parameterizing $\\sigma$ using $w \\in \\mathbb { R } _ { \\geq 0 } ^ { d }$ in the following way: ", + "bbox": [ + 171, + 102, + 821, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/21315d50d993256a7c23bc4317b615a3b2bef2d2c90d55cbde7842b28fdb5289.jpg", + "text": "$$\n\\forall i : \\sigma _ { i } = w _ { i } ^ { N }\n$$", + "text_format": "latex", + "bbox": [ + 449, + 136, + 547, + 155 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $N$ represents the depth of the model. Since we restrict the model to having non-negative weights, this parameterization doesn’t change the expressiveness, but it does radically change it’s optimization dynamics. ", + "bbox": [ + 173, + 172, + 825, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming the data is labeled by some $\\sigma ^ { * } \\in \\mathbb { R } _ { \\geq 0 } ^ { d }$ , we will study the dynamics of this model for general under a depth-normalized1 squared loss over Gaussian inputs, which will allow us to derive our analytical solution: ", + "bbox": [ + 173, + 222, + 825, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b4ccf63aab033eb8c116f99b44fa7690c8b7a66e9b5d6d770be36beeb019e707.jpg", + "text": "$$\n\\ell _ { N } ( w ) = \\frac { 1 } { 2 N ^ { 2 } } \\mathbb { E } _ { x } [ ( \\langle \\sigma ^ { * } , x \\rangle - \\langle w ^ { N } , x \\rangle ) ^ { 2 } ] = \\frac { 1 } { 2 N ^ { 2 } } | | w ^ { N } - \\sigma ^ { * } | | ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 290, + 285, + 705, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will assume that our model is initialized uniformly with a tunable scaling factor, such that: ", + "bbox": [ + 171, + 321, + 790, + 337 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/e952567285b6a26ea557f8a66e804174635f29b768fd1db13491785e3438b0dc.jpg", + "text": "$$\n\\forall i : \\ w _ { i } ( 0 ) = \\ \\sqrt [ N ] { \\sigma _ { 0 } }\n$$", + "text_format": "latex", + "bbox": [ + 431, + 356, + 566, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 GRADIENT FLOW ANALYTICAL SOLUTIONS ", + "text_level": 1, + "bbox": [ + 174, + 387, + 519, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analyzing our toy model using gradient flow allows us to obtain an analytical solution for the dynamics of $\\sigma ( t )$ along with the dynamics of the loss function for a general $N$ . For brevity, the following theorem refers only to $N = 1 , 2$ and $N \\to \\infty$ , however the solutions for $3 \\leq N < \\infty$ are similar in structure to $N \\to \\infty$ , but more complicated. We also assume $\\sigma _ { i } ^ { * } > 0$ for brevity, however we can derive the solutions for $\\sigma _ { i } ^ { * } = 0$ as well. Note that this result is a special case adaptation of the one presented in Saxe et al. (2013) for deep linear networks: ", + "bbox": [ + 173, + 412, + 825, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. Minimizing the toy linear model described in (1) with gradient flow over the depth normalized squared loss (2), with Gaussian inputs and weights initialized as in (3) and assuming $\\sigma _ { i } ^ { * } > 0$ leads to the following analytical solutions for different values of $N$ : ", + "bbox": [ + 173, + 501, + 825, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/afc342ab27650fa4f36a85ae581655158219108cacb9a75e1a24550ca023e60b.jpg", + "text": "$$\n\\begin{array} { r l } & { \\begin{array} { r l } { N = 1 : } & { \\sigma _ { i } ( t ) = \\sigma _ { i } ^ { * } + ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) e ^ { - t } } \\\\ { N = 2 : } & { \\sigma _ { i } ( t ) = \\frac { \\sigma _ { 0 } \\sigma _ { i } ^ { * } e ^ { \\sigma _ { i } ^ { * } t } } { \\sigma _ { 0 } \\left( e ^ { \\sigma _ { i } ^ { * } t } - 1 \\right) + \\sigma _ { i } ^ { * } } } \\end{array} } \\\\ & { \\begin{array} { r l } { N \\to \\infty : } & { t = \\frac { 1 } { ( \\sigma _ { i } ^ { * } ) ^ { 2 } } \\log \\big ( \\frac { \\sigma _ { i } ( t ) ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) } { \\sigma _ { 0 } \\left( \\sigma _ { i } ( t ) - \\sigma _ { i } ^ { * } \\right) } \\big ) - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { 0 } } \\big ) } \\end{array} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 297, + 563, + 699, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. The gradient flow equations for our model are the following: ", + "bbox": [ + 173, + 676, + 620, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/8b6f82c9fadbc78ec694e51acfb72295a15e0ada8ab4b7e03abf74b1bcfd0aa6.jpg", + "text": "$$\n\\dot { w } _ { i } = - \\nabla _ { w _ { i } } \\ell = \\frac { 1 } { N } w _ { i } ^ { N - 1 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } )\n$$", + "text_format": "latex", + "bbox": [ + 375, + 707, + 622, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the dynamics of the $w$ parameters, we may use the chain rule to derive the dynamics of the induced model, $\\sigma$ : ", + "bbox": [ + 174, + 744, + 823, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/e69396f523489b074fb323b0691c882f33574cad94f31b60a25ff774af8ce0fe.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } = \\frac { d \\sigma _ { i } } { d w _ { i } } \\dot { w } _ { i } = w ^ { 2 N - 2 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } ) = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )\n$$", + "text_format": "latex", + "bbox": [ + 320, + 789, + 678, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This differential equation is solvable for all $N$ , leading to the solutions in the theorem. Taking $N \\to \\infty$ in (4) leads to $\\dot { \\sigma } _ { i } = \\sigma _ { i } ^ { 2 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )$ , which is also solvable. ", + "bbox": [ + 173, + 829, + 823, + 858 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c32e4b6822cc4e3b3801507a0f18a13c9706cba7f76cc5695c9ee740ff4a57cf.jpg", + "image_caption": [ + "Figure 2: Incremental learning dynamics in the toy model. Each panel shows the evolution of $\\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } }$ for $\\sigma _ { i } ^ { * } \\in \\{ 1 2 , 6 , 4 , 3 \\}$ according to the analytical solutions in theorem 1, under different depths and initializations. The first column has all values converging at the same rate. Notice how the deep parameterization with small initialization leads to distinct phases of learning, where values are learned incrementally (bottom-right). The shallow model’s much weaker incremental learning, even at small initialization scales (second column), is explained in theorem 2. " + ], + "image_footnote": [], + "bbox": [ + 179, + 106, + 813, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analyzing these solutions, we see how even in such a simple model depth causes different factors of the model to be learned at different rates. Specifically, values corresponding to larger optimal values converge faster, suggesting a form of incremental learning. This is most clear for $N = 2$ where the solution isn’t implicit, but is also the case for $N \\geq 3$ , as we will see in the next subsection. ", + "bbox": [ + 174, + 492, + 825, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These dynamics are depicted in figure 2, where we see the dynamics of the different values of $\\sigma ( t )$ as learning progresses. When $N = 1$ , all values are learned at the same rate regardless of the initialization, while the deeper models are clearly biased towards learning the larger singular values first, especially at small initialization scales. ", + "bbox": [ + 174, + 555, + 823, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model has only one optimal solution due to the population loss, but it is clear how this sort of dynamic can induce sparse solutions - if the model is able to fit the data after a small amount of learning phases, then it’s obtained result will be sparse. Alternatively, if $N = 1$ , we know that the dynamics will lead to the minimal $\\ell _ { 2 }$ norm solution which is dense. We explore the sparsity inducing bias of our toy model by comparing it empirically2 to a greedy sparse approximation algorithm in appendix D, and give our theoretical results in the next section. ", + "bbox": [ + 174, + 617, + 825, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 INCREMENTAL LEARNING ", + "text_level": 1, + "bbox": [ + 178, + 726, + 423, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equipped with analytical solutions for the dynamics of our model for every depth, we turn to study how the depth and initialization effect incremental learning. While Gidel et al. (2019) focuses on incremental learning in depth-2 models at the limit of $\\sigma _ { 0 } 0$ , we will study the phenomenon for a general depth and for $\\sigma _ { 0 } > 0$ . ", + "bbox": [ + 174, + 760, + 825, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we will define the notion of incremental learning. Since all values of $\\sigma$ are learned in parallel, we can’t expect one value to converge before the other moves at all (which happens for infinitesimal initialization as shown by Gidel et al. (2019)). We will need a more relaxed definition for incremental learning in finite initialization scales. ", + "bbox": [ + 173, + 824, + 825, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1. Given two values $\\sigma _ { i } , \\sigma _ { j }$ such that $\\sigma _ { i } ^ { * } > \\sigma _ { j } ^ { * } > 0$ and both are initialized as $\\sigma _ { i } ( 0 ) =$ $\\sigma _ { j } ( 0 ) = \\sigma _ { 0 } < \\sigma _ { j } ^ { * }$ , and given two scalars $s \\in ( 0 , \\frac { 1 } { 4 } )$ and $f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)$ , we call the learning of the values $( s , f )$ -incremental if there exists a t for which: ", + "bbox": [ + 173, + 102, + 825, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/756427e823ee307e8aa880968cbf8646de56e5172a9941219f2ea7f263d8dcf7.jpg", + "text": "$$\n\\sigma _ { j } ( t ) \\leq s \\sigma _ { j } ^ { * } < f \\sigma _ { i } ^ { * } \\leq \\sigma _ { i } ( t )\n$$", + "text_format": "latex", + "bbox": [ + 403, + 170, + 593, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In words, two values have distinct learning phases if the first almost converges $( f \\approx 1 )$ before the second changes by much $s \\ll 1 \\ r _ { \\cdot }$ ). Note that for any $N , \\sigma ( t )$ is monotonically increasing and so once $\\sigma _ { j } ( t ) = s \\sigma _ { j } ^ { * }$ , it will not decrease to allow further incremental learning. Given this definition of incremental learning, we turn to study the conditions that facilitate incremental learning in our toy model. ", + "bbox": [ + 173, + 196, + 825, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our main result is a dynamical depth separation result, showing that incremental learning is dependent on $\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }$ in different ways for different values of $N$ . The largest difference in dependence happens between $N = 2$ and $N = 3$ , where the dependence changes from exponential to polynomial: ", + "bbox": [ + 173, + 273, + 825, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Given two values $\\sigma _ { i } , \\sigma _ { j }$ of a toy linear model as in (1), where $\\begin{array} { r } { \\frac { { \\sigma } _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}$ and the model is initialized as in (3), and given two scalars $s \\in ( 0 , \\frac { 1 } { 4 } )$ and $f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\\sigma _ { 0 } ^ { t h }$ , is bounded in the following way: ", + "bbox": [ + 173, + 329, + 826, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0bdfaf2010332f9ed8389cd05c9719d7b46af7a10d495ec58e608b955e39d0cf.jpg", + "text": "$$\n\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 246, + 420, + 751, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof sketch (the full proof is given in appendix $A$ ). Rewriting the separable differential equation in (4) to calculate the time until $\\sigma ( t ) = \\alpha \\sigma ^ { * }$ , we get the following: ", + "bbox": [ + 171, + 503, + 825, + 534 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/010182bab0f2f4e2fdb5ab3ced1eb9a629796460db21e636764dbcfa65bfd632.jpg", + "text": "$$\nt _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }\n$$", + "text_format": "latex", + "bbox": [ + 392, + 549, + 604, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The condition for incremental learning is then the requirement that $t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )$ , resulting in: ", + "bbox": [ + 171, + 597, + 805, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/45bbd2b211f6960978f4e641ef8b910b12b6c04083081c3da66ad1f666d598dd.jpg", + "text": "$$\n\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }\n$$", + "text_format": "latex", + "bbox": [ + 338, + 627, + 658, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then relax/restrict the above condition to get a necessary/sufficient condition on $\\sigma _ { 0 }$ , leading to a lower and upper bound on $\\sigma _ { 0 } ^ { t h }$ . ", + "bbox": [ + 173, + 678, + 825, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the value determining the condition for incremental learning is $\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }$ same order of magnitude, then their ratio will be close to 1 and we will need a small initialization to obtain incremental learning. The dependence on the ratio changes with depth, and is exponential for $N = 2$ . This means that incremental learning, while possible for shallow models, is difficult to see in practice. This result explains why changing the initialization scale in figure 2 changes the dynamics of the $N \\geq 3$ models, while not changing the dynamics for $N = 2$ noticeably. ", + "bbox": [ + 173, + 741, + 825, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The next theorem extends part of our analysis to gradient descent, a more realistic setting than the infinitesimal learning rate of gradient flow: ", + "bbox": [ + 171, + 838, + 823, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3. Given two values $\\sigma _ { i } , \\sigma _ { j }$ of a depth-2 toy linear model as in (1), such that $\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}$ and the model is initialized as in (3), and given two scalars $s \\in ( 0 , \\frac { 1 } { 4 } )$ and $f \\in ( \\frac { 3 } { 4 } , 1 )$ , and assuming $\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }$ , and assuming we optimize with gradient descent with a learning rate $\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}$ for $c < 2 ( { \\sqrt { 2 } } - 1 )$ and $\\sigma _ { 1 } ^ { * }$ the largest value of $\\sigma ^ { * }$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\\sigma _ { 0 } ^ { t h }$ , is lower and upper bounded in the following way: ", + "bbox": [ + 174, + 872, + 825, + 925 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 102, + 825, + 146 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/d67099a203e064b3678ebe86bcb9a99f2d540931c92f05481f5369b04fc29419.jpg", + "text": "$$\n\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }\n$$", + "text_format": "latex", + "bbox": [ + 271, + 164, + 723, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $A$ and $B$ are defined as: ", + "bbox": [ + 174, + 210, + 380, + 226 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/d5e177cebb8d5bd48f03d19faf96fcbba6231bbd13232b8eac6d06dc194bf4e1.jpg", + "text": "$$\nA = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }\n$$", + "text_format": "latex", + "bbox": [ + 272, + 244, + 723, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We defer the proof to appendix B. ", + "bbox": [ + 174, + 305, + 398, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that this result, while less elegant than the bounds of the gradient flow analysis, is similar in nature. Both $A$ and $B$ simplify to $r$ when we take their first order approximation around $c = 0$ , giving us similar bounds and showing that the condition on $\\sigma _ { 0 }$ for $N = 2$ is exponential in gradient descent as well. ", + "bbox": [ + 173, + 327, + 825, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While similar gradient descent results are harder to obtain for deeper models, we discuss the general effect of depth on the gradient decent dynamics in appendix C. ", + "bbox": [ + 171, + 390, + 823, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 INCREMENTAL LEARNING IN LARGER MODELS ", + "text_level": 1, + "bbox": [ + 173, + 440, + 602, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "So far, we have only shown interesting properties of incremental learning caused by depth for a toy model. In this section, we will relate several deep models to our toy model and show how incremental learning presents itself in larger models as well. ", + "bbox": [ + 173, + 472, + 826, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 MATRIX SENSING ", + "text_level": 1, + "bbox": [ + 174, + 534, + 339, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The task of matrix sensing is a generalization of matrix completion, where our input space is $\\mathcal { X } =$ $\\mathbb { R } ^ { d \\times d }$ and our model is a matrix $\\mathbf { \\bar { \\boldsymbol { W } } } \\in \\mathbb { R } ^ { d \\times d }$ , such that: ", + "bbox": [ + 173, + 560, + 823, + 589 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/dfbaa3e0800fc64e0f4d9dccac3558d2f6174b550e65d3f5fb315ee4e27ce3a0.jpg", + "text": "$$\nf _ { W } ( A ) = \\langle W , A \\rangle = t r \\big ( W ^ { T } A \\big )\n$$", + "text_format": "latex", + "bbox": [ + 392, + 609, + 604, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following Arora et al. (2019), we introduce depth by parameterizing the model using a product of matrices and the following initialization scheme $( W _ { i } ^ { \\setminus } \\in \\mathbb { R } ^ { d \\times d } )$ ): ", + "bbox": [ + 171, + 641, + 825, + 670 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/d4bcc48e64d0072b37aa4917ac74b963de477178d5b3f8c3099946938d4f034c.jpg", + "text": "$$\n\\begin{array} { c } { { W = W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } } } \\\\ { { \\forall i \\in [ N ] , ~ W _ { i } ( 0 ) = ~ \\sqrt [ N ] { \\sigma _ { 0 } } I } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 369, + 698, + 629, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that when $d = 1$ , the deep matrix sensing model reduces to our toy model without weight sharing. We study the dynamics of the model under gradient flow over a depth-normalized squared loss, assuming the data is labeled by a matrix sensing model parameterized by a PSD $W ^ { \\ast } \\in \\mathring { \\mathbb { R } } ^ { d \\times d }$ : ", + "bbox": [ + 174, + 751, + 825, + 795 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/c69af104dc366bd12c182e997c6bb6fcbb91a45e3e31d22b766831ccba4a2db6.jpg", + "text": "$$\n\\begin{array} { c } { \\displaystyle \\ell _ { N } ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } ) = \\displaystyle \\frac { 1 } { 2 N } \\mathbb { E } _ { A } [ ( \\langle \\left( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\right) - W ^ { * } , A \\rangle ^ { 2 } ] } \\\\ { = \\displaystyle \\frac { 1 } { 2 N } | | \\big ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\big ) - W ^ { * } | | _ { F } ^ { 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 269, + 820, + 727, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The following theorem relates the deep matrix sensing model to our toy model, showing the two have the same dynamical equations: ", + "bbox": [ + 173, + 895, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 4. Optimizing the deep matrix sensing model described in (5) with gradient flow over the depth normalized squared loss ((6)), with weights initialized as in (5) leads to the following dynamical equations for different values of $N$ : ", + "bbox": [ + 173, + 103, + 825, + 146 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/e869fb28a4b187879be9124717884221e3f7d404bdf21d63b6c6c622629dfca6.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )\n$$", + "text_format": "latex", + "bbox": [ + 395, + 162, + 601, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $\\sigma _ { i }$ and $\\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vector. ", + "bbox": [ + 174, + 190, + 823, + 219 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proof follows that of Saxe et al. (2013) and Gidel et al. (2019) and is deferred to appendix E. ", + "bbox": [ + 173, + 228, + 810, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 4 shows us that the bias towards sparse solutions introduced by depth in the toy model is equivalent to the bias for low-rank solutions in the matrix sensing task. This bias was studied in a more general setting in Arora et al. (2019), with empirical results supporting the effect of depth on the obtainment of low-rank solutions under a more natural loss and initialization scheme. We recreate and discuss these experiments and their connection to our analysis in appendix E, and an example of these dynamics in deep matrix sensing can also be seen in panel (a) of figure 1. ", + "bbox": [ + 173, + 251, + 825, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 QUADRATIC NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 351, + 447, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By drawing connections between quadratic networks and matrix sensing (as in Soltanolkotabi et al. (2018)), we can extend our results to these nonlinear models. We will study a simplified quadratic network, where our input space is $\\mathcal { X } = \\mathbb { R } ^ { d }$ and the first layer is parameterized by a weight matrix $W \\in \\mathbb { R } ^ { d \\times d }$ and followed by a quadratic activation function. The final layer will be a summation layer. We assume, like before, that the labeling function is a quadratic network parameterized by $\\bar { W } ^ { * } \\in \\mathbb { R } ^ { d \\times d }$ . Our model can be written in the following way, using the following orthogonal initialization scheme: ", + "bbox": [ + 173, + 376, + 825, + 474 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/77dfced7fa36f4674dc21569f5bb29aeb6c1e39e9cc33d763eab427083736522.jpg", + "text": "$$\n\\begin{array} { r l } & { f _ { W } ( x ) = \\displaystyle \\sum _ { i = 1 } ^ { d } ( w _ { i } ^ { T } x ) ^ { 2 } = x ^ { T } W ^ { T } W x = \\langle W ^ { T } W , x x ^ { T } \\rangle } \\\\ & { W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 320, + 496, + 679, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Immediately, we see the similarity of the quadratic network to the deep matrix sensing model with $N = 2$ , where the input space is made up of rank-1 matrices. However, the change in input space forces us to optimize over a different loss function to reproduce the same dynamics: ", + "bbox": [ + 174, + 569, + 823, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 2. Given an input distribution over an input space $\\mathcal { X }$ with a labeling function $y : \\mathcal { X } \\mathbb { R }$ and a hypothesis $h$ , the variance loss is defined in the following way: ", + "bbox": [ + 169, + 614, + 823, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/6b5b1b247a9b5d07cfca0e85c60c21dd0c8564cee84fb9a48e85832a27dea825.jpg", + "text": "$$\n\\ell _ { v a r } ( h ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( y ( x ) - h ( x ) ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ y ( x ) - h ( x ) ] ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 307, + 660, + 691, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that minimizing this loss function amounts to minimizing the variance of the error, while the squared loss minimizes the second moment of the error. We note that both loss functions have the same minimum for our problem, and the dynamics of the squared loss can be approximated in certain cases by the dynamics of the variance loss. For a complete discussion of the two losses, including the cases where the two losses have similar dynamics, we refer the reader to appendix F. ", + "bbox": [ + 174, + 696, + 825, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 5. Minimizing the quadratic network described and initialized as in (7) with gradient flow over the variance loss defined in (2) leads to the following dynamical equations: ", + "bbox": [ + 173, + 768, + 825, + 797 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/b7b36ff480a5c95d8b2916ff37e28783adb3d453b4e5b57103f77b1419768b14.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )\n$$", + "text_format": "latex", + "bbox": [ + 411, + 818, + 586, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $\\sigma _ { i }$ and $\\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vector. ", + "bbox": [ + 173, + 842, + 825, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We defer the proof to appendix F and note that these dynamics are the same as our depth-2 toy model, showing that shallow quadratic networks can exhibit incremental learning (albeit requiring a small initialization). ", + "bbox": [ + 174, + 881, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 DIAGONAL/CONVOLUTIONAL LINEAR NETWORKS ", + "text_level": 1, + "bbox": [ + 173, + 103, + 563, + 117 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While incremental learning has been described for deep linear networks in the past, it has been restricted to regression tasks. Here, we illustrate how incremental learning presents itself in binary classification, where implicit bias results have so far focused on convergence at $t \\to \\infty$ (Soudry et al. (2018), Nacson et al. (2018), Ji & Telgarsky (2019)). Deep linear networks with diagonal weight matrices have been shown to be biased towards sparse solutions when $N > 1$ in Gunasekar et al. (2018), and biased towards the max-margin solution for $N = 1$ . Instead of analyzing convergence at $t \\infty$ , we intend to show that the model favors sparse solutions for the entire duration of optimization, and that this is due to the dynamics of incremental learning. ", + "bbox": [ + 173, + 131, + 825, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our theoretical illustration will use our toy model as in (1) (initialized as in (3)) as a special weightshared case of deep networks with diagonal weight matrices, and we will then show empirical results for the more general setting. We analyze the optimization dynamics of this model over a separable dataset $\\{ x _ { i } , y _ { i } \\} _ { i = 1 } ^ { m }$ where $y _ { i } \\in \\{ \\pm 1 \\}$ . We use the exponential loss $( \\ell ( f ( x ) , y ) = e ^ { - y f ( x ) } )$ for the theoretical illustration and experiment on the exponential and logistic losses. ", + "bbox": [ + 174, + 248, + 825, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computing the gradient for the model over $w$ , the gradient flow dynamics for $\\sigma$ become: ", + "bbox": [ + 171, + 327, + 753, + 342 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/0bc426bdb9a32f2dac9a10eceb8c5499a734113b8de64f6cb7f5945ec02450ed.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } = \\frac { N ^ { 2 } } { m } \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } \\sum _ { j = 1 } ^ { m } e ^ { - y _ { j } \\langle \\beta , x _ { j } \\rangle } x _ { j }\n$$", + "text_format": "latex", + "bbox": [ + 388, + 362, + 607, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We see the same dynamical attenuation of small values of $\\sigma$ that is seen in the regression model, caused by the multiplication by σ2− 2N From this, we can expect the same type of incremental learning to occur - weights of $\\sigma$ will be learned incrementally until the dataset can be separated by the current support of $\\sigma$ . Then, the dynamics strengthen the growth of the current support while relatively attenuating that of the other values. Since the data is separated, increasing the values of the current support reduces the loss and the magnitude of subsequent gradients, and so we should expect the support to remain the same and the model to converge to a sparse solution. ", + "bbox": [ + 173, + 419, + 825, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Granted, the above description is just intuition, but panel (c) of figure 1 shows how it is born out in practice (similar results are obtained for the logistic loss). In appendix G we further explore this model, showing deeper networks have a stronger bias for sparsity. We also observe that the initialization scale plays a similar role as before - deep models are less biased towards sparsity when $\\sigma _ { 0 }$ is large. ", + "bbox": [ + 174, + 527, + 825, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In their work, Gunasekar et al. (2018) show an equivalence between the diagonal network and the circular-convolutional network in the frequency domain. According to their results, we should expect to see the same sparsity-bias of diagonal networks in convolutional networks, when looking at the Fourier coefficients of $\\sigma$ . An example of this can be seen in panel (d) of figure 1, and we refer the reader to appendix $\\mathbf { G }$ for a full discussion of their convolutional model and it’s incremental learning dynamics. ", + "bbox": [ + 174, + 604, + 825, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 712, + 318, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gradient-based optimization for deep linear models has an implicit bias towards simple (sparse) solutions, caused by an incremental search strategy over the hypothesis space. Deeper models have a stronger tendency for incremental learning, exhibiting it in more realistic initialization scales. ", + "bbox": [ + 174, + 746, + 823, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This dynamical phenomenon exists for the entire optimization process for regression as well as classification tasks, and for many types of models - diagonal networks, convolutional networks, matrix completion and even the nonlinear quadratic network. We believe this kind of dynamical analysis may be able to shed light on the generalization of deeper nonlinear neural networks as well, with shallow quadratic networks being only a first step towards that goal. ", + "bbox": [ + 174, + 795, + 823, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 885, + 326, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by the European Research Council (TheoryDL project). ", + "bbox": [ + 174, + 909, + 712, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 287, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sanjeev Arora, Nadav Cohen, and Elad Hazan. On the optimization of deep networks: Implicit acceleration by overparameterization. arXiv preprint arXiv:1802.06509, 2018. ", + "bbox": [ + 174, + 126, + 821, + 155 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In Advances in Neural Information Processing Systems 32, pp. 7411–7422. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8960-implicit-regularization-in-deep-matrix-factorization.pdf. ", + "bbox": [ + 173, + 164, + 825, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remi Tachet des Combes, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848, 2018. ", + "bbox": [ + 176, + 229, + 823, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gauthier Gidel, Francis Bach, and Simon Lacoste-Julien. Implicit regularization of discrete gradient dynamics in linear neural networks. In Advances in Neural Information Processing Systems, pp. 3196–3206, 2019. ", + "bbox": [ + 173, + 281, + 825, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit regularization in matrix factorization. In Advances in Neural Information Processing Systems, pp. 6151–6159, 2017. ", + "bbox": [ + 173, + 333, + 825, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Suriya Gunasekar, Jason D Lee, Daniel Soudry, and Nati Srebro. Implicit bias of gradient descent on linear convolutional networks. In Advances in Neural Information Processing Systems, pp. 9461–9471, 2018. ", + "bbox": [ + 173, + 386, + 825, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ziwei Ji and Matus Telgarsky. The implicit bias of gradient descent on nonseparable data. In Conference on Learning Theory, pp. 1772–1798, 2019. ", + "bbox": [ + 173, + 438, + 825, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yuanzhi Li, Tengyu Ma, and Hongyang Zhang. Algorithmic regularization in over-parameterized matrix sensing and neural networks with quadratic activations. arXiv preprint arXiv:1712.09203, 2017. ", + "bbox": [ + 173, + 476, + 823, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mor Shpigel Nacson, Jason Lee, Suriya Gunasekar, Pedro HP Savarese, Nathan Srebro, and Daniel Soudry. Convergence of gradient descent on separable data. arXiv preprint arXiv:1803.01905, 2018. ", + "bbox": [ + 174, + 527, + 823, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Preetum Nakkiran, Gal Kaplun, Dimitris Kalimeris, Tristan Yang, Benjamin L Edelman, Fred Zhang, and Boaz Barak. Sgd on neural networks learns functions of increasing complexity. arXiv preprint arXiv:1905.11604, 2019. ", + "bbox": [ + 178, + 579, + 825, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yagyensh Chandra Pati, Ramin Rezaiifar, and Perinkulam Sambamurthy Krishnaprasad. Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition. In Proceedings of 27th Asilomar conference on signals, systems and computers, pp. 40–44. IEEE, 1993. ", + "bbox": [ + 174, + 631, + 825, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman. The convergence rate of neural networks for learned functions of different frequencies. In Advances in Neural Information Processing Systems 32, pp. 4763–4772. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8723-the-convergence-rate-of-neural-networks-for-learned-functions-of-different-fr pdf. ", + "bbox": [ + 171, + 698, + 996, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013. ", + "bbox": [ + 171, + 791, + 823, + 820 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. IEEE Transactions on Information Theory, 65(2):742–769, 2018. ", + "bbox": [ + 176, + 829, + 825, + 872 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 19 (1):2822–2878, 2018. ", + "bbox": [ + 174, + 881, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Francis Williams, Matthew Trager, Claudio Silva, Daniele Panozzo, Denis Zorin, and Joan Bruna. Gradient dynamics of shallow univariate relu networks. arXiv preprint arXiv:1906.07842, 2019. ", + "bbox": [ + 173, + 102, + 823, + 132 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Blake Woodworth, Suriya Gunasekar, Jason Lee, Daniel Soudry, and Nathan Srebro. Kernel and deep regimes in overparametrized models. arXiv preprint arXiv:1906.05827, 2019. ", + "bbox": [ + 174, + 138, + 825, + 170 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A PROOF OF THEOREM 2 ", + "text_level": 1, + "bbox": [ + 176, + 193, + 400, + 210 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Theorem. Given two values $\\sigma _ { i } , \\sigma _ { j }$ of a toy linear model as in (1), such that $\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}$ and the model is initialized as in (3), and given two scalars $s \\in ( 0 , \\frac { 1 } { 4 } )$ and $f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\\sigma _ { 0 } ^ { t h }$ , is lower and upper bounded in the following way: ", + "bbox": [ + 173, + 222, + 826, + 291 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/8a0b254cfd784cab90491532e9bfa6e3bbe6c232bb42638b93c4f5af2af1436f.jpg", + "text": "$$\n\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 246, + 314, + 751, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Proof. Our strategy will be to define the time $t _ { \\alpha }$ for which a value reaches a fraction $\\alpha$ of it’s optimal value, and then require that $t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )$ . We begin with recalling the differential equation which determines the dynamics of the model: ", + "bbox": [ + 174, + 397, + 825, + 440 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/247f451707f15659e57f94c99d582f7d5c0df14f15bb12b92e1b3e3e297da7ca.jpg", + "text": "$$\n\\dot { \\sigma } = \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma )\n$$", + "text_format": "latex", + "bbox": [ + 429, + 455, + 566, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since the solution for $N \\geq 3$ is implicit and difficult to manage in a general form, we will define $t _ { \\alpha }$ using the integral of the differential equation. The equation is separable, and under initialization of $\\sigma _ { 0 }$ we can describe $t _ { \\alpha } ( \\sigma )$ in the following way: ", + "bbox": [ + 173, + 486, + 825, + 529 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/4ff2011e8893ed9bb5cee419ee302a60aff79ef738e7c946be049ea2688e2405.jpg", + "text": "$$\nt _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }\n$$", + "text_format": "latex", + "bbox": [ + 392, + 544, + 604, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Incremental learning takes place when $\\sigma _ { i } ( t _ { f } ) = f \\sigma _ { i } ^ { * }$ happens before $\\sigma _ { j } ( t _ { s } ) = s \\sigma _ { j } ^ { * }$ . We can write this condition in the following way: ", + "bbox": [ + 174, + 592, + 823, + 622 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/31e330b6bdd3b57f2f78fe7c9959612954102e9f4dd0c887df02e2c4286879b5.jpg", + "text": "$$\n\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }\n$$", + "text_format": "latex", + "bbox": [ + 338, + 636, + 658, + 678 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Plugging in $\\sigma _ { i } = r \\sigma _ { j }$ and rearranging, we get the following necessary and sufficient condition for incremental learning: ", + "bbox": [ + 174, + 686, + 825, + 715 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/0665e01f255ebd93529f2cfb9ec005b44f08bd939e9bc0acbcb0b6855cb8397c.jpg", + "text": "$$\n\\int _ { \\sigma _ { 0 } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }\n$$", + "text_format": "latex", + "bbox": [ + 326, + 731, + 669, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our last step before relaxing and restricting our condition will be to split the integral on the left-hand side into two integrals: ", + "bbox": [ + 173, + 784, + 826, + 813 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/2d0910f469d5ab8099d3e7561097b2dc6dbda9317da20e4fa4db434c2157b4e5.jpg", + "text": "$$\n\\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } + \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }\n$$", + "text_format": "latex", + "bbox": [ + 241, + 828, + 754, + 873 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "At this point, we cannot solve this equation and isolate $\\sigma _ { 0 }$ to obtain a clear threshold condition on it for incremental learning. Instead, we will relax/restrict the above condition to get a necessary/sufficient condition on $\\sigma _ { 0 }$ , leading to a lower and upper bound on the threshold value of $\\sigma _ { 0 }$ . ", + "bbox": [ + 174, + 881, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SUFFICIENT CONDITION ", + "text_level": 1, + "bbox": [ + 174, + 103, + 344, + 118 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To obtain a sufficient (but not necessary) condition on $\\sigma _ { 0 }$ , we may make the condition stricter either by increasing the left-hand side or decreasing the right-hand side. We can increase the left-hand side by removing $r$ from the left-most integral’s denominator $( r > 1 )$ ) and then combine the left-most and right-most integrals: ", + "bbox": [ + 173, + 128, + 826, + 185 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/e838aa694500dfac4c54e68c27c20a5779a2e0f05b3746a5c346dbc69cceb6ca.jpg", + "text": "$$\n\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }\n$$", + "text_format": "latex", + "bbox": [ + 307, + 202, + 691, + 247 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Next, we note that the integration bounds give us a bound on $\\sigma$ for either integral. This means we can replace $\\begin{array} { r } { 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } } \\end{array}$ with 1 on the right-hand side, and replace $\\begin{array} { r } { 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } } \\end{array}$ with $1 - f$ on the left-hand side: ", + "bbox": [ + 173, + 255, + 825, + 301 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/f005d94ad21671d3896b0f5f73445479860092c93e94c6d30bd718edf61a3cef.jpg", + "text": "$$\n\\frac { 1 } { 1 - f } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }\n$$", + "text_format": "latex", + "bbox": [ + 348, + 318, + 650, + 358 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We may now solve these integrals for every $N$ and isolate $\\sigma _ { 0 }$ , obtaining the lower bound on $\\sigma _ { 0 } ^ { t h }$ . We start with the case where $N = 2$ : ", + "bbox": [ + 171, + 369, + 825, + 397 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/ed37d967a247a7869497d836ecb1529455eb4ab2b29346cca432e01ee7cc8b5f.jpg", + "text": "$$\n\\frac { 1 } { 1 - f } \\big ( \\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\big ) \\leq ( r - 1 ) \\big ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\big )\n$$", + "text_format": "latex", + "bbox": [ + 284, + 414, + 712, + 446 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rearranging to isolate $\\sigma _ { 0 }$ , we obtain our result: ", + "bbox": [ + 173, + 455, + 483, + 469 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/c3c5fdd56841a4a51a8156094280ed16da7dfce8d3de6ac369a84cd1f1c553b9.jpg", + "text": "$$\n\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( { \\frac { s } { r f } } \\big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } }\n$$", + "text_format": "latex", + "bbox": [ + 415, + 486, + 581, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the $N \\geq 3$ case, we have the following after solving the integrals: ", + "bbox": [ + 171, + 525, + 630, + 541 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/549d15c01baa23c359a2f826a5e804905209affd1a36eef36205bd5a2547ba92.jpg", + "text": "$$\n\\frac { 1 } { 1 - f } \\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq ( r - 1 ) \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )\n$$", + "text_format": "latex", + "bbox": [ + 250, + 555, + 745, + 592 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For simplicity we may further restrict the condition by removing the term $\\Big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } }$ \u00011− 2N . Solving for $\\sigma _ { 0 }$ gives us the following: ", + "bbox": [ + 174, + 602, + 825, + 640 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/69b4202b14fed56c585a7999a284af0a6a5a79b02826cf7d9933ddbba295cad6.jpg", + "text": "$$\n\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 375, + 655, + 620, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "NECESSARY CONDITION ", + "text_level": 1, + "bbox": [ + 174, + 705, + 348, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To obtain a necessary (but not sufficient) condition on $\\sigma _ { 0 }$ , we may relax the condition in (8) either by decreasing the left-hand side or increasing the right-hand side. We begin by rearranging the equation: ", + "bbox": [ + 173, + 731, + 825, + 773 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/813b870fe15e38f3325a5a02b85ccc357d691c80b99b6fa649998cfc632f0aeb.jpg", + "text": "$$\n\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) } - \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) }\n$$", + "text_format": "latex", + "bbox": [ + 240, + 789, + 754, + 834 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Like before, we may use the integration bounds to bound $\\sigma$ . Plugging in $\\sigma = s \\sigma _ { j } ^ { * }$ for all integrals decreases the left-hand side and increases the right-hand side, leading us to the following: ", + "bbox": [ + 171, + 842, + 823, + 872 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/a3eb3dee43b36a19272b64864fcdf199a5779479c58a832565c195196b45b2dd.jpg", + "text": "$$\n\\frac { r } { r - s } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le \\big ( \\frac { r } { 1 - s } - \\frac { r } { r - s } \\big ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }\n$$", + "text_format": "latex", + "bbox": [ + 316, + 887, + 679, + 928 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rearranging, we get the following inequality: ", + "bbox": [ + 174, + 103, + 472, + 118 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/1d24111e000c0abc2af85b0afad77e68d6ed714d550b4c053e14679c34b80c2c.jpg", + "text": "$$\n\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\leq \\frac { r - 1 } { 1 - s } \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }\n$$", + "text_format": "latex", + "bbox": [ + 374, + 135, + 624, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We now solve the integrals for the different cases. For $N = 2$ , we have: ", + "bbox": [ + 173, + 188, + 643, + 203 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/29ea47288f5b6c5a3f43f605c76489c0d20bc983b0b3f4d1823726fb8f42f9d1.jpg", + "text": "$$\n\\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\leq \\frac { r - 1 } { 1 - s } \\Bigl ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\Bigr )\n$$", + "text_format": "latex", + "bbox": [ + 316, + 219, + 679, + 251 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Rearranging to isolate $\\sigma _ { 0 }$ , we get our condition: ", + "bbox": [ + 173, + 261, + 488, + 276 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/f78019dc3ba334eb806b7a9a5fb6400862bf625da840cbab9b9df16d4244f0df.jpg", + "text": "$$\n\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( \\frac { s } { r f } \\big ) ^ { \\frac { 1 - s } { r - 1 } }\n$$", + "text_format": "latex", + "bbox": [ + 434, + 294, + 562, + 325 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Finally, for $N \\geq 3$ , we solve the integrals to give us: ", + "bbox": [ + 173, + 335, + 517, + 352 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/1df08ba2c837b07f2d05916c13fad65fc1ca43d84edb1061197795569d1659ac.jpg", + "text": "$$\n\\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq \\frac { r - 1 } { 1 - s } \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )\n$$", + "text_format": "latex", + "bbox": [ + 276, + 367, + 720, + 404 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Rearranging to isolate $\\sigma _ { 0 }$ , we get our condition: ", + "bbox": [ + 174, + 414, + 486, + 429 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/c400becfc7d03fa51e6ed901fe1b391c865f5df9d121b21a818720e762fcbfb5.jpg", + "text": "$$\n\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 421, + 445, + 576, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "SUMMARY ", + "bbox": [ + 174, + 496, + 250, + 510 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For a given $N$ , we derived a sufficient condition and a necessary condition on $\\sigma _ { 0 }$ for $( s , f )$ - incremental learning. The necessary and sufficient condition on $\\sigma _ { 0 }$ , which is the largest initialization value for which we see incremental learning (denoted $\\sigma _ { 0 } ^ { t h \\cdot }$ ), is between the two derived bounds. ", + "bbox": [ + 174, + 521, + 825, + 564 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The precise bounds can possibly be improved a bit, but the asymptotic dependence on $r$ is the crux of the matter, showing the dependence on $r$ changes with depth with a substantial difference when we move from shallow models $N = 2$ ) to deeper ones $( N \\geq 3$ ) ", + "bbox": [ + 174, + 570, + 825, + 613 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B PROOF OF THEOREM 3 ", + "text_level": 1, + "bbox": [ + 176, + 655, + 398, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Theorem. Given two values $\\sigma _ { i } , \\sigma _ { j }$ of a depth-2 toy linear model as in (1), such that $\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}$ and the model is initialized as in (3), and given two scalars $s \\in ( 0 , \\frac { 1 } { 4 } )$ and $f \\in ( \\frac { 3 } { 4 } , 1 )$ , and assuming $\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }$ , and assuming we optimize with gradient descent with a learning rate $\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}$ for $c < 2 ( { \\sqrt { 2 } } - 1 )$ and $\\sigma _ { 1 } ^ { * }$ the largest value of $\\sigma ^ { * }$ , then the largest initialization value for which the learning phases of the values are $( s , f )$ -incremental, denoted $\\sigma _ { 0 } ^ { t h }$ , is lower and upper bounded in the following way: ", + "bbox": [ + 173, + 684, + 826, + 785 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/f989af24e020caa89a073422b6ed1af5eea8a8a77a43f865802bb8988011016b.jpg", + "text": "$$\n\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }\n$$", + "text_format": "latex", + "bbox": [ + 271, + 801, + 723, + 835 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Where $A$ and $B$ are defined as: ", + "bbox": [ + 176, + 847, + 380, + 861 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/bf85fc793530d30b5a4cd608c3212d1a16361d8dd660a5771a094371a743ba85.jpg", + "text": "$$\nA = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }\n$$", + "text_format": "latex", + "bbox": [ + 272, + 880, + 723, + 928 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proof. To show our result for gradient descent and $N = 2$ , we build on the proof techniques of theorem 3 of Gidel et al. (2019). We start by deriving the recurrence relation for the values $\\sigma ( t )$ for general depth, when $t$ now stands for the iteration. Remembering that $w _ { i } ^ { n } = \\sigma _ { i }$ , we write down the gradient update for $w _ { i } ( t )$ : ", + "bbox": [ + 173, + 102, + 825, + 161 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/9772893184c9e2aa969e20cefc687450254670e52098c17ec1a940af2aebee97.jpg", + "text": "$$\nw _ { i } ( t + 1 ) = w _ { i } ( t ) + \\eta \\frac { 1 } { N } w _ { i } ( t ) ^ { N - 1 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )\n$$", + "text_format": "latex", + "bbox": [ + 202, + 184, + 794, + 215 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Raising $w _ { i } ( t )$ to the $N$ th power, we get the gradient update for the $\\sigma$ values: ", + "bbox": [ + 173, + 229, + 673, + 244 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/17860e6e9702f9121662d68c4defc1a0556ff5952ae0f2e2b7f801965256cf26.jpg", + "text": "$$\nr _ { i } ( t + 1 ) = \\bigg ( \\sqrt [ r ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } \\bigg ) ^ { N } = \\sigma _ { i } ( t ) \\Big ( 1 + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { N }\n$$", + "text_format": "latex", + "bbox": [ + 181, + 268, + 825, + 300 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Next, we will prove a simple lemma which gives us the maximal learning rate we will consider for the analysis, for which there is no overshooting (the values don’t grow larger than the optimal values). ", + "bbox": [ + 173, + 318, + 825, + 361 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lemma 1. For the gradient update in (9), assuming $\\sigma _ { i } ( 0 ) < \\sigma _ { i } ^ { * }$ , if $\\begin{array} { r } { \\eta \\leq \\left( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\right) ^ { 2 - \\frac { 2 } { N } } } \\end{array}$ and $N \\geq 2$ , then: ", + "bbox": [ + 173, + 369, + 825, + 410 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/a5f408efe5ec5d8ec43a138f796555e1c81732745262dfee382c4d5e4c585170.jpg", + "text": "$$\n\\forall t : \\ \\sigma _ { i } ( t ) \\leq \\sigma _ { i } ^ { * }\n$$", + "text_format": "latex", + "bbox": [ + 442, + 429, + 553, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof. Plugging in $\\begin{array} { r } { \\eta = c \\bigg ( \\frac { 1 } { \\sigma _ { i } ^ { * } } \\bigg ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}$ for $c \\leq 1$ , we have: ", + "bbox": [ + 173, + 460, + 542, + 491 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/c42971020f57e61baf48bc9dc18de1d65dd7ce01fb564abf820ad4d07f8eeb05.jpg", + "text": "$$\n\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\Big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) \\Big ) ^ { N } \\leq \\sigma _ { i } ( t ) e ^ { \\left( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\right) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) }\n$$", + "text_format": "latex", + "bbox": [ + 194, + 510, + 802, + 554 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Defining ri = σiσ∗i a nd dividing both sides by $\\sigma _ { i } ^ { * }$ , we have: ", + "bbox": [ + 171, + 568, + 553, + 585 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/7d5aa12762aef96b649c1121280d46213f820f4d300cf9dc461f87a38a34882e.jpg", + "text": "$$\nr _ { i } ( t + 1 ) \\leq r _ { i } ( t ) e ^ { r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) }\n$$", + "text_format": "latex", + "bbox": [ + 382, + 599, + 614, + 623 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is enough to show that for any $0 \\leq r \\leq 1$ , we have that $r e ^ { r ^ { 1 - \\frac { 2 } { N } } ( 1 - r ) } \\leq 1$ , as over-shooting occurs when $r _ { i } ( \\bar { t } ) > 1$ . Indeed, this function is monotonic increasing in $0 \\leq r \\leq 1$ (since the exponent is non negative), and equals 1 when $r = 1$ . Since $r = 1$ is a fixed point and no iteration that starts at $r < 1$ can cross 1, then $r _ { i } ( t ) \\leq 1$ for any $t$ . This concludes our proof. □ ", + "bbox": [ + 173, + 635, + 825, + 696 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Under this choice of learning rate, we can now obtain our incremental learning results for gradient descent when $N = 2$ . Our strategy will be bounding $\\sigma _ { i } ( t )$ from below and above, which will give us a lower and upper bound for $t _ { \\alpha } ( \\sigma _ { i } )$ . Once we have these bounds, we will be able to describe either a necessary or a sufficient condition on $\\sigma _ { 0 }$ for incremental learning, similar to theorem 2. ", + "bbox": [ + 173, + 712, + 825, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The update rule for $N = 2$ is: ", + "bbox": [ + 174, + 775, + 370, + 790 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/cd05808cb73e514a483f4c120952578fd482e56fe54ba33c984b736407f3fce0.jpg", + "text": "$$\n\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { 1 } { 2 } \\eta ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 359, + 806, + 637, + 837 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Next, we plug in $\\begin{array} { r } { \\eta = \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}$ for $c < 2 ( \\sqrt { 2 } - 1 ) < 1$ and denote $\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\leq 1 } \\end{array}$ and $\\begin{array} { r } { r _ { i } ( t ) = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\le 1 } \\end{array}$ to get: ", + "bbox": [ + 171, + 848, + 826, + 883 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/1cf77218caa89b2240727d215e87107e30d27bd14515fd06a4c85302fd063f79.jpg", + "text": "$$\n\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 361, + 898, + 635, + 929 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Following theorem 3 of Gidel et al. (2019), we bound $\\frac { 1 } { \\sigma _ { i } ( t ) }$ : ", + "bbox": [ + 173, + 102, + 566, + 122 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/869f7f8a106e05ed77f77284a7b85cc8974315bd7f3d07ed533e53a5664702a0.jpg", + "text": "$$\n\\begin{array} { r l } & { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { \\left( 1 + \\frac { \\varepsilon } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\right) ^ { 2 } } } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\left( 1 - r _ { i } ( t ) \\right) \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\bigl ( 1 - ( 1 - r _ { i } ( t ) ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) \\bigr ) } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } - ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 302, + 138, + 696, + 324 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Where in the fourth line we use the inequality $\\begin{array} { r } { \\frac { 1 } { 1 + x } \\geq 1 - x , \\forall x \\geq 0 } \\end{array}$ . We may now subtract $\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }$ from both sides to obtain: ", + "bbox": [ + 174, + 332, + 821, + 366 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/398c960671dfcf23edd99904d0af487602e5cc56a57b02a3af1c0c9387dc9e55.jpg", + "text": "$$\n\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\geq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) \\geq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) ^ { t }\n$$", + "text_format": "latex", + "bbox": [ + 209, + 381, + 789, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We may now obtain a bound on $t _ { \\alpha } ( \\sigma _ { i } )$ by plugging in $\\sigma _ { i } ( t ) = \\alpha \\sigma _ { i } ^ { * }$ and taking the log: ", + "bbox": [ + 171, + 425, + 741, + 443 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/61f62d820fb5bc742ea13b173f735ebc48de539c2b889c93337334c3f931ed0f.jpg", + "text": "$$\n\\log \\big ( \\frac { 1 - \\alpha } { \\alpha \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 \\big ) } \\big ) \\geq t _ { \\alpha } \\cdot \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big )\n$$", + "text_format": "latex", + "bbox": [ + 341, + 457, + 655, + 497 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Rearranging (note that $\\begin{array} { r } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) < 0 } \\end{array}$ and that our choice of $c$ keeps the argument of the log positive), we get: ", + "bbox": [ + 176, + 506, + 823, + 537 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/3031ae89a58278b4232cdfab63ef9cc0ada25c71dd93dd2708ce51130ce4fd48.jpg", + "text": "$$\nt _ { \\alpha } ( \\sigma _ { i } ) \\geq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) }\n$$", + "text_format": "latex", + "bbox": [ + 388, + 553, + 609, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Next, we follow the same procedure for an upper bound. Starting with our update step: ", + "bbox": [ + 174, + 609, + 740, + 626 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/503a6deb3e4a5432aff9b363a5e322c1b39f60bdf4f59fb5e2375cbda9131a22.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) ( c R _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ( 1 - r _ { i } ( t ) ) ) } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) c R _ { i } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } ( 1 - ( 1 - r _ { i } ( t ) ) c R _ { i } + ( 1 - r _ { i } ( t ) ) ^ { 2 } c ^ { 2 } R _ { i } ^ { 2 } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 300, + 645, + 696, + 751 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Where in the last line we use the inequality $\\begin{array} { r } { \\frac { 1 } { 1 + x } \\leq 1 - x + x ^ { 2 } , \\forall x \\geq 0 } \\end{array}$ . Subtracting $\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }$ from both sides, we get: ", + "bbox": [ + 173, + 758, + 823, + 794 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/1fbf41e7558ebc51e0a148769c816886b852351549695e6fb49eeedbdc559603.jpg", + "text": "$$\n\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\leq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) \\leq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) ^ { t }\n$$", + "text_format": "latex", + "bbox": [ + 212, + 808, + 784, + 843 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Rearranging like before, we get the bound on the $\\alpha$ -time: ", + "bbox": [ + 173, + 849, + 549, + 866 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/32e30bd54300ce13fc724e28bbf6bc24de046cdc14df93642dc289b37d7c0551.jpg", + "text": "$$\nt _ { \\alpha } ( \\sigma _ { i } ) \\leq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) }\n$$", + "text_format": "latex", + "bbox": [ + 390, + 881, + 607, + 929 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Given these bounds, we would like to find the conditions on $\\sigma _ { 0 }$ that allows for $( s , f )$ -incremental learning. We will find a sufficient condition and a necessary condition, like in the proof of theorem 2. ", + "bbox": [ + 173, + 102, + 825, + 146 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "SUFFICIENT CONDITION ", + "text_level": 1, + "bbox": [ + 176, + 164, + 344, + 179 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A sufficient condition for incremental learning will be one which is possibly stricter than the exact condition. We can find such a condition by requiring the upper bound of $t _ { f } ( \\sigma _ { i } )$ to be smaller than the lower bound on $t _ { s } ( \\sigma _ { j } )$ . This becomes the following condition: ", + "bbox": [ + 174, + 190, + 825, + 234 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/fcf156d6d7052f7e5308846663a328233ce16c1034b484297216fe9ee4cd5444.jpg", + "text": "$$\n\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\right) }\n$$", + "text_format": "latex", + "bbox": [ + 333, + 251, + 663, + 296 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Defining $\\begin{array} { r } { A = \\frac { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\big ) } } \\end{array}$ and rearranging, we get the following: ", + "bbox": [ + 171, + 308, + 645, + 342 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/7a792ed0cd3ecf2fbb22f0057d9e3acfdfc0c9fecf1c1ba4103d93667df9c07a.jpg", + "text": "$$\n\\log \\big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\big ) \\geq A \\log \\big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\big )\n$$", + "text_format": "latex", + "bbox": [ + 339, + 349, + 656, + 386 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We may now take the exponent of both sides and rearrange again, remembering $\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}$ , to get the following condition: ", + "bbox": [ + 173, + 400, + 825, + 435 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/3f6348d01377d5ba7a1a4664d06de2954d942de52d7e8b71a9fb2b9419c0ce51.jpg", + "text": "$$\n\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A } } { \\sigma _ { 0 } ^ { A - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { A }\n$$", + "text_format": "latex", + "bbox": [ + 375, + 453, + 622, + 493 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Now, we will add the very reasonable assumption that $\\sigma _ { j } ^ { * } \\geq 2 \\sigma _ { 0 }$ , which allows us to replace $\\frac { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } { r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } }$ with $\\textstyle { \\frac { 1 } { 2 r } }$ and replace $( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A - 1 }$ with $\\textstyle { \\left( { \\frac { 1 } { 2 } } \\sigma _ { j } \\right) } ^ { A - 1 }$ , only making the condition stricter. This simplifies the expression to the following: ", + "bbox": [ + 173, + 508, + 825, + 560 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/db3295364432591d6e09a05005d50a52ee38d2c1f5988c2b10f614df7a6bcc2b.jpg", + "text": "$$\n\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { A - 1 } \\geq \\frac { r f } { 1 - f } \\big ( \\frac { 2 - 2 s } { s } \\big ) ^ { A }\n$$", + "text_format": "latex", + "bbox": [ + 397, + 578, + 601, + 613 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Now we can rearrange and isolate $\\sigma _ { 0 }$ to get a sufficient condition for incremental learning: ", + "bbox": [ + 169, + 623, + 766, + 640 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/44f3a5c73f179cd18e12f16bd10686e7817ca87b0ee31991b60e390d70c493d6.jpg", + "text": "$$\n\\sigma _ { 0 } \\leq \\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\bigg ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\bigg ) ^ { \\frac { 1 } { A - 1 } }\n$$", + "text_format": "latex", + "bbox": [ + 379, + 656, + 619, + 691 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "NECESSARY CONDITION ", + "text_level": 1, + "bbox": [ + 174, + 707, + 346, + 722 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A necessary condition for incremental learning will be one which is possibly more relaxed than the exact condition. We can find such a condition by requiring the lower bound of $t _ { f } ( \\sigma _ { i } )$ to be smaller than the upper bound on $t _ { s } ( \\sigma _ { j } )$ . This becomes the following condition: ", + "bbox": [ + 174, + 733, + 825, + 777 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/ce276208cd2811997f831b9309860b2e7c24a4126dae02a8fc16713e5572078f.jpg", + "text": "$$\n\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\right) }\n$$", + "text_format": "latex", + "bbox": [ + 333, + 795, + 663, + 838 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Defining $\\begin{array} { r } { B = \\frac { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\big ) } } \\end{array}$ and rearranging, we get the following: ", + "bbox": [ + 171, + 851, + 643, + 885 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/24ed8ee8ddb98f1166441d9d420ef87a20a63f89ff8f5cb56035bba81892b10f.jpg", + "text": "$$\n\\log \\Big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\Big ) \\geq B \\log \\Big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\Big )\n$$", + "text_format": "latex", + "bbox": [ + 339, + 892, + 656, + 928 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We may now take the exponent of both sides and rearrange again, remembering $\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}$ , to get the following condition: ", + "bbox": [ + 173, + 101, + 825, + 137 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/2613fd0283081b65387794b66fbb9672bab6c8f7554511298d90a8b93755f048.jpg", + "text": "$$\n\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { B } } { \\sigma _ { 0 } ^ { B - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { B }\n$$", + "text_format": "latex", + "bbox": [ + 375, + 154, + 622, + 194 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We may now relax the condition further, by removing the $r$ from the denominator of the left-hand side and the $\\sigma _ { 0 }$ from the numerator. This gives us the following: ", + "bbox": [ + 169, + 202, + 825, + 232 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/30f112edadd193c88e289a66765d3240020ce807caafce3dbf5e0f30105d1c89.jpg", + "text": "$$\n\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { B - 1 } \\geq \\frac { f } { 1 - f } \\big ( \\frac { 1 - s } { s } \\big ) ^ { B }\n$$", + "text_format": "latex", + "bbox": [ + 400, + 247, + 598, + 281 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Finally, rearranging gives us the necessary condition: ", + "bbox": [ + 173, + 290, + 522, + 305 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/b6b0bb317c61a5ecad46f8d8fb71fefbd5011b7089bb664b3d6f457e8bea89fa.jpg", + "text": "$$\n\\sigma _ { 0 } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }\n$$", + "text_format": "latex", + "bbox": [ + 383, + 320, + 611, + 356 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C DISCUSSION OF GRADIENT DESCENT FOR GENERAL $N$ ", + "text_level": 1, + "bbox": [ + 174, + 397, + 668, + 415 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "While we were able to generalize our result to gradient descent for $N = 2$ , our proof technique relies on the ability to get a non-implicit solution for $\\sigma ( t )$ which we discretized and bounded. This is harder to generalize to larger values of $N$ , where the solution is implicit. Still, we can informally illustrate the effect of depth on the dynamics of gradient descent by approximating the update rule of the values. ", + "bbox": [ + 173, + 429, + 825, + 501 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We start by reminding ourselves of the gradient descent update rule for $\\sigma$ , for a learning rate $\\eta =$ c \u0000 1σ ∗ \u0001 2 − 2N : ", + "bbox": [ + 173, + 506, + 821, + 542 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/0cd0c1812f44087bd78975b3a0595876da98bd39fcbd6616177755d12e70f2c2.jpg", + "text": "$$\n\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\big ( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 - \\frac { 2 } { N } } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) \\big ) \\Big ) ^ { N }\n$$", + "text_format": "latex", + "bbox": [ + 292, + 558, + 704, + 590 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "To compare two values in the same scales, we will divide both sides by the optimal value $\\sigma _ { i } ^ { * }$ and look at the update step of the ratio $\\begin{array} { r } { r _ { i } = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } } \\end{array}$ , also denoting $\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } } \\end{array}$ : ", + "bbox": [ + 173, + 598, + 826, + 635 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/1fe0c7b717c4daecb4d27f5eddfff0ec23efbe5f581056d1196de31176422610.jpg", + "text": "$$\nr _ { i } ( t + 1 ) = r _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { N }\n$$", + "text_format": "latex", + "bbox": [ + 312, + 648, + 684, + 679 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We will focus on the early stages of the optimization process, where $r \\ll 1$ . This means we can neglect the $1 - r _ { i } ( t )$ term in the update step, giving us the approximate update step we will use to compare the general $i , j$ values: ", + "bbox": [ + 173, + 686, + 825, + 731 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/70832c4c34788b1125286d215c6c50fe4da301fcd150899ab014edb54372ac84.jpg", + "text": "$$\nr _ { i } ( t + 1 ) \\approx r _ { i } ( t ) \\left( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\right) ^ { N }\n$$", + "text_format": "latex", + "bbox": [ + 348, + 746, + 648, + 776 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We would like to compare the dynamics of $r _ { i }$ and $r _ { j }$ , which is difficult to do when the recurrence relation isn’t solvable. However, we can observe the first iteration of gradient descent and see how depth affects this iteration. Since we are dealing with variables which are ratios of different optimal values, the initial values of $r$ are different. Denoting $\\begin{array} { r } { \\mathbf { r } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } } \\end{array}$ , we can describe the initialization of $r _ { j }$ using that of $r _ { i }$ : ", + "bbox": [ + 173, + 785, + 825, + 864 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/732ca989aad5d07fc705b83e0d0badcf61f9e527c5faefc1129c5848d57079b3.jpg", + "text": "$$\nr _ { j } ( 0 ) = { \\bf r } r _ { i } ( 0 )\n$$", + "text_format": "latex", + "bbox": [ + 449, + 882, + 549, + 900 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Plugging in the initial conditions and noting that $R _ { i } = { \\bf r } R _ { j }$ , we get: ", + "bbox": [ + 173, + 909, + 617, + 925 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/36f6da11812ef6f6803b6b456473767caaae9b4e7c828ae1174c883796119f3b.jpg", + "text": "$$\n\\begin{array} { r } { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( 1 + \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\\\ { r _ { j } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( \\sqrt [ N ] { \\mathbf { r } } + \\big ( \\frac { 1 } { \\mathbf { r } } \\big ) ^ { \\frac { N - 1 } { N } } \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 325, + 116, + 673, + 193 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We see that the two ratios have a similar update, with the ratio of optimal values playing a role in how large the initial value is versus how large the added value is. When we use a small learning rate, we have a very small $c$ which means we can make a final approximation and neglect the higher order terms of $c$ : ", + "bbox": [ + 173, + 196, + 826, + 253 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/0599ec72d1268d76f947ebd27669a0dcf1aff387d0b69787069a2b1ceaeb5261.jpg", + "text": "$$\n\\begin{array} { r l } & { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) + c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\\\ & { r _ { j } ( 1 ) \\approx { \\bf r } r _ { i } ( 0 ) + \\big ( \\frac { 1 } { { \\bf r } } \\big ) ^ { \\frac { N - 1 } { N } } c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 356, + 272, + 642, + 329 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We can see that while the initial conditions favor $r _ { j }$ , the size of the update for $r _ { i }$ is larger by a factor of $r ^ { \\frac { N - 1 } { N } }$ when the initialization and learning rates are small. This accumulates throughout the optimization, making $r _ { i }$ eventually converge faster than $r _ { j }$ . ", + "bbox": [ + 174, + 334, + 826, + 382 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The effect of depth here is clear - the deeper the model, the larger the relative step size of $r _ { i }$ and the faster it converges relative to $r _ { j }$ . ", + "bbox": [ + 173, + 387, + 823, + 417 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D COMPARISON OF THE TOY MODEL AND OMP ", + "text_level": 1, + "bbox": [ + 174, + 434, + 593, + 452 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Learning our toy model, when it’s incremental learning is taken to the limit, can be described as an iterative procedure where at every step an additional feature is introduced such that it’s weight is non-zero and then the model is optimized over the current set of features. This description is also relevant for the sparse approximation algorithm orthogonal matching pursuit (Pati et al., 1993), where the next feature is greedily chosen to be the one which most improves the current model. ", + "bbox": [ + 173, + 465, + 825, + 536 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "While the toy model and OMP are very different algorithms for learning sparse linear models, we will show empirically that they behave similarly. This allows us to view incremental learning as a continuous-time extension of a greedy iterative algorithm. ", + "bbox": [ + 174, + 542, + 825, + 585 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "To allow for negative weights in our experiments, we augment our toy model as in the toy model of Woodworth et al. (2019). Our model will have the same induced form as before: ", + "bbox": [ + 171, + 592, + 823, + 621 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/8e6cbc889da30afebafa57fb8cd73c4527400fe432a1c124d0ac410d7a8b5389.jpg", + "text": "$$\nf _ { \\sigma } ( x ) = \\langle \\sigma , x \\rangle\n$$", + "text_format": "latex", + "bbox": [ + 447, + 640, + 550, + 656 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "However, we parameterize $\\sigma$ using $w _ { + } , w _ { - } \\in \\mathbb { R } ^ { d }$ in the following way: ", + "bbox": [ + 173, + 662, + 642, + 679 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/b7b5e6a4c4974a05a300ad6b46052fb88d306a51df08491dd7e32e93e7e9ef14.jpg", + "text": "$$\n\\begin{array} { c } { \\sigma _ { i } = w _ { + , i } ^ { N } - w _ { - , i } ^ { N } } \\\\ { \\qquad \\forall i , \\ w _ { + , i } ( 0 ) = w _ { - , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 390, + 699, + 607, + 739 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We can now treat this algorithm as a sparse approximation pursuit algorithm - given a dictionary $D \\in \\mathbb { R } ^ { d \\times n }$ and an example $x \\in \\mathbb { R } ^ { d }$ , we wish to find the sparsest $\\alpha$ for which $D \\alpha \\approx x$ by minimizing the $\\ell _ { 0 }$ norm of $\\alpha$ subject to $\\| D \\alpha - x \\| _ { 2 } ^ { 2 } = 0 ^ { 3 }$ . Under this setting, we can compare OMP to our toy model by comparing the sets of features that the two algorithms choose for a given example and dictionary. ", + "bbox": [ + 173, + 746, + 825, + 816 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "In figure 3 we run such a comparison. Using a dictionary of 1000 atoms and an example of dimensionality 80 sampled from a random hidden vector of a given sparsity $s$ , we run both algorithms and record the first $s$ features chosen4. ", + "bbox": [ + 173, + 823, + 825, + 866 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/850628cf8a6350a15f50e62eeedfcba28e4ac2fa035fa393674ec77b7c764190.jpg", + "image_caption": [ + "Figure 3: Empirical comparison of the dynamics of the toy model to OMP. The toy model has a depth of 5 and was initialized with a scale of 1e-4 and a learning rate of 3e-3. We compare the fraction of agreement between the sets of first $s$ features selected of the two algorithms for every given sparsity level $s$ , averaged over 100 experiments (the shaded regions are empirical standard deviations). For example, for sparsity level 3, we look at the sets of first 3 features selected by each algorithm and calculate the fraction of them that appear in both sets. " + ], + "image_footnote": [], + "bbox": [ + 316, + 135, + 648, + 383 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "For every sparsity $s$ , we plot the mean fraction of agreement between the sets of features chosen by OMP and the toy model over 100 experiments. We see that the two algorithms choose very similar features at the beginning, suggesting that the deep model approximates the discrete behavior of OMP. Only when the number of features increases do we see that the behavior of the two models begins to differ, caused by the fact that the toy model has a finite initialization scale and learning rate. ", + "bbox": [ + 173, + 518, + 825, + 602 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "These experiments demonstrate the similarity between the incremental learning of deep models and the discrete behavior of greedy approximation algorithms such as OMP. Adopting this view also allows us to put our finger on another strength of the dynamics of deep models - while greedy algorithms such as OMP require the analytical solution or approximation of every iterate, the dynamics of deep models are able to incrementally learn any differentiable function. For example, looking back at the matrix sensing task and the classification models in section 4, we see that while there isn’t an immediate and efficient extension of OMP for these settings, the dynamics of learning deep models extends naturally and exhibits the same incremental learning as OMP. ", + "bbox": [ + 173, + 608, + 825, + 720 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "E INCREMENTAL LEARNING IN MATRIX SENSING ", + "text_level": 1, + "bbox": [ + 174, + 742, + 602, + 758 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "E.1 PROOF OF THEOREM 4 ", + "text_level": 1, + "bbox": [ + 174, + 775, + 374, + 789 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Theorem. Minimizing the deep matrix sensing model described in (5) with gradient flow over the depth normalized squared loss (6), with Gaussian inputs and weights initialized as in (5) leads to the following dynamical equations for different values of $N$ : ", + "bbox": [ + 174, + 801, + 823, + 844 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/26ecb4ca6a2dc65c84ab0680a444b7f51428246ecc707196e8cd0b3d744a6a73.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )\n$$", + "text_format": "latex", + "bbox": [ + 395, + 863, + 601, + 883 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Where $\\sigma _ { i }$ and $\\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vectors. ", + "bbox": [ + 174, + 895, + 825, + 922 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Proof. We will adapt the proof from Saxe et al. (2013) for multilayer linear networks. The gradient flow equations for $W _ { n } , n \\in [ N ]$ are: ", + "bbox": [ + 171, + 103, + 825, + 133 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/e42e2d276924324141d11217a52cfca45b365a8a314885ae029a3964c269613c.jpg", + "text": "$$\n\\dot { W _ { n } } = { \\frac { 1 } { N } } W _ { 1 : n - 1 } ^ { T } ( W ^ { * } - W ) W _ { n + 1 : N } ^ { T }\n$$", + "text_format": "latex", + "bbox": [ + 372, + 148, + 624, + 179 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Where we denote $\\begin{array} { r } { W _ { j : k } = \\prod _ { i = j } ^ { k } W _ { i } } \\end{array}$ ", + "bbox": [ + 174, + 189, + 413, + 208 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Since we assumed $W ^ { * }$ is (symmetric) PSD, there exists an orthogonal matrix $U$ for which $D ^ { * } =$ $U W ^ { * } U ^ { T }$ where $D ^ { * }$ is diagonal. Under the initialization in (3), $U$ diagonalizes all $W _ { n }$ matrices at initialization such that ${ \\cal D } _ { n } \\overline { { } } = U W _ { n } U ^ { T } = \\sqrt [ N ] { \\sigma _ { 0 } } I$ . Making this change of variables for all $W _ { n }$ , we get: ", + "bbox": [ + 173, + 213, + 825, + 270 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/77582bad9b550554ae4889585a494609a229ea62d2a7364ae430273bc229eb84.jpg", + "text": "$$\n\\dot { W _ { n } } = \\frac { 1 } { N } U ^ { T } D _ { 1 : n - 1 } U ( W ^ { * } - W ) U ^ { T } D _ { n + 1 : N } U\n$$", + "text_format": "latex", + "bbox": [ + 338, + 286, + 658, + 316 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Rearranging, we get a set of decoupled differential equations for the singular values of $W _ { n }$ ", + "bbox": [ + 169, + 324, + 771, + 339 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/1fce7a2eb8034aece84f701ddd076d1426b3bea6ce8aece2702e5e3d034d46c8.jpg", + "text": "$$\n\\dot { D _ { n } } = { \\frac { 1 } { N } } D _ { 1 : n - 1 } ( D ^ { * } - D ) D _ { n + 1 : N }\n$$", + "text_format": "latex", + "bbox": [ + 379, + 354, + 617, + 386 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Note that since these matrices are all diagonal at initialization, the above dynamics ensure that they remain diagonal throughout the optimization. Denoting $\\sigma _ { n , i }$ as the $i ^ { \\because }$ ’th singular value of $W _ { n }$ and $\\sigma _ { i }$ as the $\\overrightarrow { \\imath } ^ { \\prime }$ ’th singular value of $W$ , we get the following differential equation: ", + "bbox": [ + 173, + 393, + 825, + 436 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/7ae62e0de9170e192f9c400a9e2263d45b0ae45e6eadd34184f1b74df67142cb.jpg", + "text": "$$\n\\dot { \\sigma } _ { n , i } = \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ) \\prod _ { j \\neq n } \\sigma _ { j , i }\n$$", + "text_format": "latex", + "bbox": [ + 403, + 453, + 593, + 492 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Since we assume at initialization that $\\forall n , m , i : \\sigma _ { n , i } ( 0 ) = \\sigma _ { m , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } }$ , the above dynamics are the same for all singular values and we get $\\forall n , m , i : \\sigma _ { n , i } ( t ) = \\sigma _ { m , i } ( t ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) }$ . We may now use this to calculate the dynamics of the singular value of $W$ , since they are the product the the singular values of all $W _ { n }$ matrices. Denoting $\\begin{array} { r } { \\sigma _ { - n , i } = \\prod _ { k \\neq n } \\sigma _ { k , i } } \\end{array}$ and using the chain rule: ", + "bbox": [ + 173, + 501, + 825, + 563 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/cb13b4e605be597a0d56eaacac3386f61d1aa8e54c7d344f34dbc1ca5c323726.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } = \\sum _ { n = 1 } ^ { N } \\sigma _ { - n , i } \\dot { \\sigma } _ { n , i } = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )\n$$", + "text_format": "latex", + "bbox": [ + 369, + 577, + 629, + 621 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E.2 EMPIRICAL EXAMINATION ", + "text_level": 1, + "bbox": [ + 174, + 660, + 401, + 675 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Our analytical results are only applicable for the population loss over Gaussian inputs. These conditions are far from the ones used in practice and studied in Arora et al. (2019), where the problem is over-determined and the weights are drawn from a Gaussian distribution with a small variance. To show our conclusions regarding incremental learning extend qualitatively to more natural settings, we empirically examine the deep matrix sensing model in this natural setting for different depths and initialization scales as seen in figure 4. ", + "bbox": [ + 173, + 685, + 825, + 771 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Notice how incremental learning is exhibited even when the number of examples is much smaller than the number of parameters in the model. While we can’t rely on our theory for describing the exact dynamics of the optimization for these kinds of over-determined problems, the qualitative conclusions we get from it are still applicable. ", + "bbox": [ + 174, + 776, + 825, + 833 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Another interesting phenomena we should note is that once the dataset becomes very small (the second row of the figure), we see all “currently active” singular values change at the beginning of every new phase (this is best seen in the bottom-right panel). This suggests that since there is more than one optimal solution, once we increase the current rank of our model it may find a solution that has a different set of singular values and vectors and thus all singular values change at the beginning of a new learning phase. This demonstrates the importance of incremental learning for obtaining sparse solutions - once the initialization conditions and depth are such that the learning phases are distinct, gradient descent finds the optimal rank- $i$ solution in every phase $i$ . For these dynamics to successfully recover the optimal solution at every phase, the phases need to be far enough apart from each other to allow for the singular values and vectors to change before the next phase begins. ", + "bbox": [ + 173, + 839, + 825, + 924 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/30d0053940d521183eae431a6073eb95e1a63934ce44ff4a2a09d223b98048e2.jpg", + "image_caption": [ + "Figure 4: Evolution of the top-5 singular values of the deep matrix sensing model, with Gaussian initialization with variance such that the initial singular values are in expectation 1e-4. The model’s size and data are in $\\mathbb { R } ^ { 5 0 \\times 5 0 }$ . The columns correspond to different parameterization depths, while the rows correspond to different dataset sizes. In both cases the problem is over-determined, since the number of examples is smaller than the number of parameters. Since the original matrix is rank-4, we can recognize an unsuccessful recovery when all five singular values are nonzero, as seen clearly for both depth-1 plots. " + ], + "image_footnote": [], + "bbox": [ + 181, + 102, + 818, + 349 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 492, + 825, + 549 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "F INCREMENTAL LEARNING IN QUADRATIC NETWORKS ", + "bbox": [ + 174, + 568, + 656, + 585 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "F.1 PROOF OF THEOREM 5 ", + "bbox": [ + 174, + 598, + 372, + 613 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Theorem. Minimizing the quadratic network described and initialized as in (7) with gradient flow over the variance loss defined in (2) with Gaussian inputs leads to the following dynamical equations: ", + "bbox": [ + 174, + 625, + 825, + 667 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/2c3723674660a152f88205f30c94626451ce52242b7952c621e4f5f1203b0993.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )\n$$", + "text_format": "latex", + "bbox": [ + 411, + 686, + 586, + 704 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Where $\\sigma _ { i }$ and $\\sigma _ { i } ^ { * }$ are the ith singular values of $W$ and $W ^ { * }$ , respectively, corresponding to the same singular vectors. ", + "bbox": [ + 173, + 712, + 823, + 741 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Proof. Our proof will follow similar lines as the analysis of the deep matrix sensing model. Taking the expectation of the variance loss over Gaussian inputs for our model gives us: ", + "bbox": [ + 173, + 756, + 821, + 786 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/33ab1a67e9b6848f6dc0f8ea52ab043b39cdff03886a7d0ecfe294b486f6701f.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\ell _ { v a r } ( W ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ] ^ { 2 } } \\\\ { \\displaystyle = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } - \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\\\ { \\displaystyle \\quad = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 184, + 808, + 813, + 900 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Following the gradient flow dynamics over $W$ leads to the following differential equation: ", + "bbox": [ + 168, + 909, + 764, + 924 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/ca35af0726e0ad8d69915e77afade54f7497d063e6eac2091df983edb20d80c1.jpg", + "text": "$$\n\\dot { W } = \\frac { 1 } { 2 } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W )\n$$", + "text_format": "latex", + "bbox": [ + 397, + 113, + 599, + 143 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We can now calculate the gradient flow dynamics of $W ^ { T } W$ using the chain rule: ", + "bbox": [ + 173, + 154, + 702, + 170 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/b1d6d2741c130e57df027aa47d1150d94b31ad4d63b01c7f415ce7b07f6dd70f.jpg", + "text": "$$\nW ^ { \\dot { T } } W = W ^ { T } \\dot { W } + \\dot { W } ^ { T } W = \\frac { 1 } { 2 } \\Bigl ( W ^ { T } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) + ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) W ^ { T } W \\Bigr )\n$$", + "text_format": "latex", + "bbox": [ + 181, + 193, + 787, + 224 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Now, under our initialization $W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I \\mathrm { . }$ , we get that $W ^ { T } W$ and $W _ { * } ^ { T } W _ { * }$ are simultaneously diagonalizable at initialization by some matrix $U$ , such that the following is true for diagonal $D$ and $D ^ { * }$ : ", + "bbox": [ + 174, + 237, + 825, + 280 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/a9c4edfa8ae905caf3cd41b198a1de4a32d1b65aedff0615d3fa20f5435c5a9d.jpg", + "text": "$$\nD ( 0 ) = U ^ { T } W ( 0 ) ^ { T } W ( 0 ) U\n$$", + "text_format": "latex", + "bbox": [ + 406, + 297, + 591, + 318 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/e9b51eae83acb8130f277fe19d230dedd5977780133543aba25c77ac48bd25e4.jpg", + "text": "$$\nD ^ { * } = U ^ { T } W _ { * } ^ { T } W _ { * } U\n$$", + "text_format": "latex", + "bbox": [ + 431, + 338, + 566, + 357 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Multiplying equation (10) by $U$ and $U ^ { T }$ gives us the following dynamics for the singular values of $W ^ { T } W$ : ", + "bbox": [ + 171, + 366, + 825, + 396 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/f1d394e05a751c376879c6cd9bd491a4c5719defb884874224541df06e35b646.jpg", + "text": "$$\n\\dot { D } = \\frac { 1 } { 2 } \\Big ( D ( D ^ { * } - D ) + ( D ^ { * } - D ) D \\Big ) = D ( D ^ { * } - D )\n$$", + "text_format": "latex", + "bbox": [ + 320, + 411, + 678, + 443 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "These matrices are diagonal at initialization, and remain diagonal throughout the dynamics (the offdiagonal elements are static according to these equations). We may now look at the dynamics of a single diagonal element, noticing it is equivalent to the depth-2 toy model: ", + "bbox": [ + 174, + 455, + 825, + 500 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/c12c96e75c3ba1ff7288ec15a7ba1234aa02106bcc232f412bf93e76e878e246.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )\n$$", + "text_format": "latex", + "bbox": [ + 439, + 517, + 558, + 535 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "F.2 DISCUSSION OF THE VARIANCE LOSS ", + "text_level": 1, + "bbox": [ + 176, + 577, + 475, + 592 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "It may seem that the variance loss is an unnatural loss function to analyze, since it isn’t used in practice. While this is true, we will show how the dynamics of this loss function are an approximation of the square loss dynamics. ", + "bbox": [ + 174, + 602, + 823, + 645 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We begin by describing the dynamics of both losses, showing how incremental learning can’t take place for quadratic networks as defined over the squared loss. Then, we show how adding a global bias to the quadratic network leads to similar dynamics for small initialization scales. ", + "bbox": [ + 174, + 651, + 823, + 694 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "F.2.1 DYNAMICAL DERIVATION ", + "text_level": 1, + "bbox": [ + 174, + 708, + 408, + 723 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "in the previous section, we derive the differential equations for the singular values of $W ^ { T } W$ under the variance loss: ", + "bbox": [ + 173, + 733, + 823, + 762 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/67100e570fb3f2842452968d2b157ad3a4def67113be508f28b081040a0c2ff8.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )\n$$", + "text_format": "latex", + "bbox": [ + 439, + 781, + 558, + 799 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We will now derive similar equations for the squared loss. The scaled squared loss in expectation over the Gaussian inputs is: ", + "bbox": [ + 171, + 808, + 825, + 838 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/e082b4c28eb6e542bc04000b3c7f672951bbfd35a72df9d25e6cef7a6eef3549.jpg", + "text": "$$\n\\begin{array} { r l } & { \\ell ( W ) = \\displaystyle \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] } \\\\ & { \\qquad = \\displaystyle \\frac { 1 } { 8 } | | W _ { * } ^ { T } W _ { * } - W ^ { T } W | | _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 292, + 862, + 705, + 925 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Defining $\\Delta = W _ { * } ^ { T } W _ { * } - W ^ { T } W$ for brevity, the differential equations for $W$ become: ", + "bbox": [ + 173, + 102, + 735, + 119 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/890dcfaeec2b118f56d3a15c72b6c45a986472d6e54582ca2f59ae9d6c614c05.jpg", + "text": "$$\n\\dot { W } = \\frac { 1 } { 2 } W \\Delta + \\frac { 1 } { 4 } T r ( \\Delta ) W\n$$", + "text_format": "latex", + "bbox": [ + 406, + 133, + 589, + 165 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Calculating the dynamics of $W ^ { T } W$ after noting that it is simultaneously diagonalizable with $W _ { * } ^ { T } W _ { * }$ (as in the derivation for the variance loss) leads to the following differential equations for the singular values of $W ^ { T } W$ : ", + "bbox": [ + 173, + 175, + 825, + 218 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/3621aa3a2f9b0623dfe6b0accc893aadfeee631cbc164503d5c89a3b5e4f1d24.jpg", + "text": "$$\n\\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) \\Big )\n$$", + "text_format": "latex", + "bbox": [ + 372, + 234, + 624, + 273 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We see that the equations are now coupled and so we cannot solve them analytically. Another issue is that for our initialization, all singular values have very similar dynamics at initialization due to the coupling. For example, values corresponding to small optimal singular values grow much faster than in the variance loss dynamics, due to the effect the large optimal singular values have on them. ", + "bbox": [ + 173, + 284, + 825, + 340 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We see from these equations that we shouldn’t expect quadratic networks optimized over the squared loss to exhibit incremental learning behavior. We next show how adding a global bias to our model can help. ", + "bbox": [ + 173, + 347, + 825, + 390 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "F.2.2 VARIANCE LOSS APPROXIMATES THE SQUARED LOSS ", + "text_level": 1, + "bbox": [ + 173, + 404, + 604, + 420 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "To see how the variance loss can have dynamics resembling those of the squared loss, we will add a global (trainable) bias to our model. This means our model is now parameterized by $W \\in \\mathbb { R } ^ { d \\times d }$ and a scalar $b \\in \\mathbb { R }$ : ", + "bbox": [ + 174, + 429, + 825, + 472 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/d5cad89cfe3e2d31a2c1c6cb4521849140245aafe938867c89b4a7515e0ad247.jpg", + "text": "$$\nf _ { W , b } ( x ) = \\langle W ^ { T } W , x x ^ { T } \\rangle + b\n$$", + "text_format": "latex", + "bbox": [ + 400, + 489, + 598, + 508 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We may now analyze gradient flow of the squared loss. Following the same methods as before, this leads us to the following differential equations: ", + "bbox": [ + 176, + 518, + 825, + 547 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/19e998aeb0cd8260a3e51c339ee0ab9bd808b022a4849845dbd64c732e80dfb1.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + \\frac { 1 } { 2 } ( b ^ { * } - b ) \\Big ) } \\\\ { \\displaystyle \\dot { b } = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } - b } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 331, + 571, + 665, + 648 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Notice how, if $b$ is at it’s optimum at a given time $\\begin{array} { r } { ( b = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } ) } \\end{array}$ , the dynamics of $\\sigma _ { i }$ align with those of the variance loss. Alternatively, when $b = b ^ { * }$ , we recover the dynamics of the squared loss without the global bias term. ", + "bbox": [ + 176, + 661, + 823, + 707 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "To convince ourselves that the dynamics of this model resemble those of the variance loss, we would need to explain why the global bias is at it’s optimum “most of the time”, such that the singular values don’t change much during the times when it is not at it’s optimum. ", + "bbox": [ + 173, + 713, + 825, + 756 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Observing the differential equations for $b$ and for $\\sigma _ { i }$ , we see they are similar (if we ignore the $\\sigma _ { i } ^ { * } - \\sigma _ { i }$ value which doesn’t change the order of magnitude of the entire expression when there haven’t been many learning phases yet). The only difference being a multiplication by $\\sigma _ { i }$ . This means that we may informally write: ", + "bbox": [ + 173, + 761, + 825, + 818 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/a4a92114e23afb1dfe38d059b447a751c3212fc653e0e70d2347ecca2789b023.jpg", + "text": "$$\n\\frac { \\dot { \\sigma } _ { i } ( t ) } { \\dot { b } ( t ) } \\approx \\sigma _ { i } ( t )\n$$", + "text_format": "latex", + "bbox": [ + 452, + 834, + 547, + 871 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Since at initialization and until the learning phase of $\\sigma _ { i }$ takes place, we have $\\sigma _ { i } ( t ) \\ll 1$ , we see that the global bias optimizes much faster than the singular values for which the learning phase hasn’t begun yet. This means these singular values will remain small during the times in which the bias isn’t optimal, and so incremental learning can still take place (assuming a small enough initialization). ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/0ec431bf9815c02f0689d4f67bcd2532ec6e6174e98a709848c535d46e23f897.jpg", + "image_caption": [ + "Figure 5: Quadratic model’s evolution of top-5 singular values for a rank-4 labeling function. The rows correspond to whether or not a global bias is introduced to the model. The first two columns are for a large dataset (one optimal solution) and the last two columns are for a small dataset (overdetermined problem). When a bias is introduced, it is initialized to it’s optimal value at initialization. Note how without the bias, the singular values are learned together and there is over-shooting of the optimal singular value caused by the coupling of the dynamics of the singular values. For the small datasets, we see that the model with no bias reaches a solution with a larger rank. Once a global bias is introduced, the dynamics become more incremental as in the analysis of the variance loss. Note that in this case the solution obtained for the small dataset is the optimal low-rank solution. " + ], + "image_footnote": [], + "bbox": [ + 222, + 99, + 774, + 315 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 489, + 823, + 518 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Under these considerations, we say that the dynamics of the squared loss for a quadratic network with an added global bias resemble the idealized dynamics of the variance loss for a depth-2 linear model which we analyze formally in the paper. In figure 5 we experimentally show how adding a bias to a quadratic network does lead to incremental learning similar to the depth-2 toy model. ", + "bbox": [ + 174, + 525, + 825, + 580 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "G INCREMENTAL LEARNING IN CLASSIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 603, + 596, + 619 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "G.1 DIAGONAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 636, + 380, + 651 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "In section 4.3 we viewed our toy model as a special case of the deep diagonal networks described in Gunasekar et al. (2018), expected to be biased towards sparse solutions. Figure 6 shows the dynamics of the largest values of $\\sigma$ for different depths of the model. We see that the same type of incremental learning we saw in earlier models exists here as well - the features are learned one by one in deeper models, resulting in a sparse solution. The leftmost panel shows how the initialization scale plays a role here as well, with the solution being more sparse when the initialization is small. We should note that these results do not defy the results of Gunasekar et al. (2018) (from which we would expect the initialization not to matter), since their results deal with the solution at $t \\to \\infty$ . ", + "bbox": [ + 174, + 664, + 825, + 775 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "G.2 CONVOLUTIONAL NETWORKS - PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 174, + 795, + 553, + 809 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "The linear circular-convolutional network of Gunasekar et al. (2018) deals with one-dimensional convolutions with the same number of outputs as inputs, such that the mapping from one hidden layer to the next is parameterized by $w _ { n }$ and defined to be: ", + "bbox": [ + 174, + 821, + 825, + 864 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/b332b917ac73178d8b5090101a64bebb24e89ab47a251c24e0dea18ce83bf948.jpg", + "text": "$$\nh _ { n } [ i ] = \\sum _ { k = 0 } ^ { d - 1 } w _ { n } [ k ] h _ { n - 1 } [ ( i + k ) \\mod d ] = ( h _ { n - 1 } \\star w _ { n } ) [ i ]\n$$", + "text_format": "latex", + "bbox": [ + 300, + 885, + 697, + 928 + ], + "page_idx": 22 + }, + { + "type": "image", + "img_path": "images/f24003e92c57d90fdbdbc73d05bd3dd33d93af75d8ddeff5e4ea5313c8d7c9a4.jpg", + "image_caption": [ + "Figure 6: Incremental learning in binary classification. A model as in section 4.3 is trained over 200 i.i.d. random Gaussian examples, where $d = 1 0 0$ . The data is labeled by a weight vector with 4 nonzero values, making the problem realizable with a sparse solution while the max-margin solution isn’t sparse. The left panel describes the obtained solution’s correlation with the sparse labeling vector for different depths and initializations. The results are averaged over 100 experiments, with shaded regions denoting empirical standard deviations. We see that depth-1 models reach results similar to the max-margin SVM solution as predicted by Gunasekar et al. (2018), while deeper models are highly correlated with the sparse solution, with this correlation increasing when the initialization scale is small. The other panels show the evolution of the absolute values of the top-5 weights of $\\sigma$ for the smallest initialization scale. Note that as we increase the depth, incremental learning is clearly presented. " + ], + "image_footnote": [], + "bbox": [ + 178, + 103, + 818, + 224 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "The final layer is a fully connected layer parameterized by $w _ { N } \\in \\mathbb { R } ^ { d }$ , such that the final model can be written in the following way: ", + "bbox": [ + 174, + 424, + 823, + 454 + ], + "page_idx": 23 + }, + { + "type": "equation", + "img_path": "images/bc8a2f81d3ff4ed13d4edfc3f0d635663a8474a933f6acac03db3a299b44ba82.jpg", + "text": "$$\nf _ { \\sigma } ( x ) = ( ( ( ( x \\star w _ { 1 } ) \\star w _ { 2 } ) \\cdot \\cdot \\cdot ) \\star w _ { N - 1 } ) ^ { T } w _ { N }\n$$", + "text_format": "latex", + "bbox": [ + 344, + 470, + 651, + 491 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "Lemma 3 from Gunasekar et al. (2018) shows how we can relate the Fourier coefficients of the weight vectors to the Fourier coefficients of the linear model induced by the model: ", + "bbox": [ + 173, + 500, + 826, + 529 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "Lemma. For the circular-convolutional model as in (11): ", + "bbox": [ + 173, + 534, + 553, + 547 + ], + "page_idx": 23 + }, + { + "type": "equation", + "img_path": "images/db754dd72fc0c7a055bdeb7bd57f99f82e160c8768d31cc0fe73539600e6d0ca.jpg", + "text": "$$\n\\hat { \\sigma } = d i a g ( \\hat { w } _ { 1 } ) \\cdot \\cdot \\cdot d i a g ( \\hat { w } _ { N - 1 } ) \\hat { w } _ { N } ,\n$$", + "text_format": "latex", + "bbox": [ + 380, + 566, + 614, + 584 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "where for $n = 1 , 2 , . . . , N , \\hat { w } _ { n } \\in \\mathbb { C } ^ { d }$ are the Fourier coefficients of the parameters $w _ { n } \\in \\mathbb { R } ^ { d }$ ", + "bbox": [ + 171, + 593, + 769, + 609 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "This lemma connects the convolutional network to the diagonal network, and thus we should expect to see the same incremental learning of the values of the diagonal network exhibited by the Fourier coefficients of the convolutional network. ", + "bbox": [ + 173, + 621, + 825, + 662 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "G.3 CONVOLUTIONAL NETWORKS - EMPIRICAL EXAMINATION ", + "text_level": 1, + "bbox": [ + 174, + 680, + 627, + 695 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "In figure 7 we see the same plots as in figure 6 but for the Fourier coefficients of the convolutional model. We see that even when the model is far from the toy parameterization (there is no weight sharing and the initialization is with random Gaussian weights), incremental learning is still clearly seen in the dynamics of the model. We see how the inherent reason for the sparsity towards sparse solution found in Gunasekar et al. (2018) is the result of the dynamics of the model - small amplitudes are attenuated while large ones are amplified. ", + "bbox": [ + 173, + 705, + 825, + 790 + ], + "page_idx": 23 + }, + { + "type": "image", + "img_path": "images/ce17411c14f982ed90cef0de02508e3a99c1bff2ecdf153320a808745f475e11.jpg", + "image_caption": [ + "Figure 7: Incremental learning in convolutional networks. A model as in appendix G is trained over 200 i.i.d. random Gaussian examples, where $d = 1 0 0$ . The weights are initialized randomly and the data is labeled by a weight vector with 4 nonzero frequencies, making the problem realizable with a sparse solution in the frequency domain. The left panel describes the obtained solution’s correlation in the frequency domain with the sparse labeling vector for different depths and initializations. The results are averaged over 9 experiments, with shaded regions denoting empirical standard deviations. We see that depth-1 models reach results similar to the max-margin SVM solution, while deeper models are highly correlated with the optimal sparse solution. The other panels show the evolution of the amplitudes of the top-5 frequencies of $\\sigma$ for the smallest initialization scale. Note that as we increase the depth, incremental learning is clearly presented. " + ], + "image_footnote": [], + "bbox": [ + 230, + 376, + 761, + 496 + ], + "page_idx": 24 + } +] \ No newline at end of file diff --git a/parse/train/H1lj0nNFwB/H1lj0nNFwB_middle.json b/parse/train/H1lj0nNFwB/H1lj0nNFwB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..2829bc40730c7e6a54f2c0fc653466bb3bac5e45 --- /dev/null +++ b/parse/train/H1lj0nNFwB/H1lj0nNFwB_middle.json @@ -0,0 +1,66534 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 502, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "score": 1.0, + "content": "THE IMPLICIT BIAS OF DEPTH: HOW INCREMENTAL", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 388, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 388, + 118 + ], + "score": 1.0, + "content": "LEARNING DRIVES GENERALIZATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 410, + 190 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 326, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 326, + 147 + ], + "score": 1.0, + "content": "Daniel Gissin, Shai Shalev-Shwartz, Amit Daniely", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 145, + 230, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 230, + 158 + ], + "score": 1.0, + "content": "School of Computer Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 155, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 208, + 169 + ], + "score": 1.0, + "content": "The Hebrew University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 167, + 182, + 180 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 182, + 180 + ], + "score": 1.0, + "content": "Jerusalem, Israel", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 179, + 411, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 179, + 411, + 191 + ], + "score": 1.0, + "content": "{daniel.gissin,shais,amit.daniely}@mail.huji.ac.il", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 218, + 335, + 233 + ], + "spans": [ + { + "bbox": [ + 276, + 218, + 335, + 233 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 245, + 468, + 398 + ], + "lines": [ + { + "bbox": [ + 142, + 244, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 244, + 470, + 257 + ], + "score": 1.0, + "content": "A leading hypothesis for the surprising generalization of neural networks is that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 254, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 269 + ], + "score": 1.0, + "content": "the dynamics of gradient descent bias the model towards simple solutions, by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "searching through the solution space in an incremental order of complexity. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 277, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 469, + 290 + ], + "score": 1.0, + "content": "formally define the notion of incremental learning dynamics and derive the condi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "tions on depth and initialization for which this phenomenon arises in deep linear", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 311 + ], + "score": 1.0, + "content": "models. Our main theoretical contribution is a dynamical depth separation re-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "sult, proving that while shallow models can exhibit incremental learning dynam-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "ics, they require the initialization to be exponentially small for these dynamics to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "present themselves. However, once the model becomes deeper, the dependence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "score": 1.0, + "content": "becomes polynomial and incremental learning can arise in more natural settings.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "score": 1.0, + "content": "We complement our theoretical findings by experimenting with deep matrix sens-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "score": 1.0, + "content": "ing, quadratic neural networks and with binary classification using diagonal and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 376, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 469, + 388 + ], + "score": 1.0, + "content": "convolutional linear networks, showing all of these models exhibit incremental", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 180, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 180, + 401 + ], + "score": 1.0, + "content": "learning.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 420, + 206, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 208, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 208, + 436 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 504, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "Neural networks have led to a breakthrough in modern machine learning, allowing us to efficiently", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "learn highly expressive models that still generalize to unseen data. The theoretical reasons for this", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "success are still unclear, as the generalization capabilities of neural networks defy the classic statis-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 479, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 490 + ], + "score": 1.0, + "content": "tical learning theory bounds. Since these bounds, which depend solely on the capacity of the learned", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "model, are unable to account for the success of neural networks, we must examine additional proper-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "ties of the learning process. One such property is the optimization algorithm - while neural networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "score": 1.0, + "content": "can express a multitude of possible ERM solutions for a given training set, gradient-based methods", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 523, + 490, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 490, + 535 + ], + "score": 1.0, + "content": "with the right initialization may be implicitly biased towards certain solutions which generalize.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "A possible way such an implicit bias may present itself, is if gradient-based methods were to search", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "the hypothesis space for possible solutions of gradually increasing complexity. This would suggest", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "that while the hypothesis space itself is extremely complex, our search strategy favors the simplest", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "score": 1.0, + "content": "solutions and thus generalizes. One of the leading results along these lines has been by Saxe et al.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "(2013), deriving an analytical solution for the gradient flow dynamics of deep linear networks and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "showing that for such models, the singular values converge at different rates, with larger values", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "converging first. At the limit of infinitesimal initialization of the deep linear network, Gidel et al.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "(2019) show these dynamics exhibit a behavior of “incremental learning” - the singular values of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "model are learned separately, one at a time. Our work generalizes these results to small but finite", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 637, + 187, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 187, + 650 + ], + "score": 1.0, + "content": "initialization scales.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 504, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 504, + 667 + ], + "score": 1.0, + "content": "Incremental learning dynamics have also been explored in gradient descent applied to matrix com-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 504, + 678 + ], + "score": 1.0, + "content": "pletion and sensing with a factorized parameterization (Gunasekar et al. (2017), Arora et al. (2018),", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Woodworth et al. (2019)). When initialized with small Gaussian weights and trained with a small", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "learning rate, such a model is able to successfully recover the low-rank matrix which labeled the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "data, even if the problem is highly over-determined and no additional regularization is applied. In", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "their proof of low-rank recovery for such models, Li et al. (2017) show that the model remains low-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "rank throughout the optimization process, leading to the successful generalization. Additionally,", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 502, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "score": 1.0, + "content": "THE IMPLICIT BIAS OF DEPTH: HOW INCREMENTAL", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 388, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 388, + 118 + ], + "score": 1.0, + "content": "LEARNING DRIVES GENERALIZATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 113, + 134, + 410, + 190 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 326, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 326, + 147 + ], + "score": 1.0, + "content": "Daniel Gissin, Shai Shalev-Shwartz, Amit Daniely", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 145, + 230, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 230, + 158 + ], + "score": 1.0, + "content": "School of Computer Science", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 155, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 208, + 169 + ], + "score": 1.0, + "content": "The Hebrew University", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 167, + 182, + 180 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 182, + 180 + ], + "score": 1.0, + "content": "Jerusalem, Israel", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 179, + 411, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 179, + 411, + 191 + ], + "score": 1.0, + "content": "{daniel.gissin,shais,amit.daniely}@mail.huji.ac.il", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 111, + 134, + 411, + 191 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 218, + 335, + 233 + ], + "spans": [ + { + "bbox": [ + 276, + 218, + 335, + 233 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 245, + 468, + 398 + ], + "lines": [ + { + "bbox": [ + 142, + 244, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 244, + 470, + 257 + ], + "score": 1.0, + "content": "A leading hypothesis for the surprising generalization of neural networks is that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 254, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 469, + 269 + ], + "score": 1.0, + "content": "the dynamics of gradient descent bias the model towards simple solutions, by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "searching through the solution space in an incremental order of complexity. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 277, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 469, + 290 + ], + "score": 1.0, + "content": "formally define the notion of incremental learning dynamics and derive the condi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "tions on depth and initialization for which this phenomenon arises in deep linear", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 469, + 311 + ], + "score": 1.0, + "content": "models. Our main theoretical contribution is a dynamical depth separation re-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "sult, proving that while shallow models can exhibit incremental learning dynam-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "ics, they require the initialization to be exponentially small for these dynamics to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 345 + ], + "score": 1.0, + "content": "present themselves. However, once the model becomes deeper, the dependence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 469, + 356 + ], + "score": 1.0, + "content": "becomes polynomial and incremental learning can arise in more natural settings.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "score": 1.0, + "content": "We complement our theoretical findings by experimenting with deep matrix sens-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 470, + 378 + ], + "score": 1.0, + "content": "ing, quadratic neural networks and with binary classification using diagonal and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 376, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 469, + 388 + ], + "score": 1.0, + "content": "convolutional linear networks, showing all of these models exhibit incremental", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 180, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 180, + 401 + ], + "score": 1.0, + "content": "learning.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 14.5, + "bbox_fs": [ + 141, + 244, + 470, + 401 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 420, + 206, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 208, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 208, + 436 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 504, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "Neural networks have led to a breakthrough in modern machine learning, allowing us to efficiently", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "learn highly expressive models that still generalize to unseen data. The theoretical reasons for this", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 468, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 479 + ], + "score": 1.0, + "content": "success are still unclear, as the generalization capabilities of neural networks defy the classic statis-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 479, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 490 + ], + "score": 1.0, + "content": "tical learning theory bounds. Since these bounds, which depend solely on the capacity of the learned", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "model, are unable to account for the success of neural networks, we must examine additional proper-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "ties of the learning process. One such property is the optimization algorithm - while neural networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "score": 1.0, + "content": "can express a multitude of possible ERM solutions for a given training set, gradient-based methods", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 523, + 490, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 490, + 535 + ], + "score": 1.0, + "content": "with the right initialization may be implicitly biased towards certain solutions which generalize.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 445, + 506, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 539, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "A possible way such an implicit bias may present itself, is if gradient-based methods were to search", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "the hypothesis space for possible solutions of gradually increasing complexity. This would suggest", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "that while the hypothesis space itself is extremely complex, our search strategy favors the simplest", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 584 + ], + "score": 1.0, + "content": "solutions and thus generalizes. One of the leading results along these lines has been by Saxe et al.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "(2013), deriving an analytical solution for the gradient flow dynamics of deep linear networks and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "showing that for such models, the singular values converge at different rates, with larger values", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "converging first. At the limit of infinitesimal initialization of the deep linear network, Gidel et al.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "(2019) show these dynamics exhibit a behavior of “incremental learning” - the singular values of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "model are learned separately, one at a time. Our work generalizes these results to small but finite", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 637, + 187, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 187, + 650 + ], + "score": 1.0, + "content": "initialization scales.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 540, + 505, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 504, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 504, + 667 + ], + "score": 1.0, + "content": "Incremental learning dynamics have also been explored in gradient descent applied to matrix com-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 504, + 678 + ], + "score": 1.0, + "content": "pletion and sensing with a factorized parameterization (Gunasekar et al. (2017), Arora et al. (2018),", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Woodworth et al. (2019)). When initialized with small Gaussian weights and trained with a small", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "learning rate, such a model is able to successfully recover the low-rank matrix which labeled the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "data, even if the problem is highly over-determined and no additional regularization is applied. In", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "their proof of low-rank recovery for such models, Li et al. (2017) show that the model remains low-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "rank throughout the optimization process, leading to the successful generalization. Additionally,", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 655, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 83, + 495, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 83, + 495, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 83, + 495, + 189 + ], + "spans": [ + { + "bbox": [ + 113, + 83, + 495, + 189 + ], + "score": 0.972, + "type": "image", + "image_path": "14126dc03eaae57887ee79a34684cde80a2ac3af64c8443f5095c0fc7eaad71c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 83, + 495, + 118.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 118.33333333333334, + 495, + 153.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 153.66666666666669, + 495, + 189.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 505, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "Figure 1: Incremental learning dynamics in deep models. Each panel shows the evolution of the five", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 176, + 221 + ], + "score": 1.0, + "content": "largest values of", + "type": "text" + }, + { + "bbox": [ + 176, + 210, + 183, + 218 + ], + "score": 0.74, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 208, + 505, + 221 + ], + "score": 1.0, + "content": ", the parameters of the induced model. All models were trained using gradient", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "descent with a small initialization and learning rate, on a small training set such that there are mul-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "tiple possible solutions. In all cases, the deep parameterization of the models lead to “incremental", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "learning”, where the values are learned at different rates (larger values are learned first), leading", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 307, + 265 + ], + "score": 1.0, + "content": "to sparse solutions. (a) Depth 4 matrix sensing,", + "type": "text" + }, + { + "bbox": [ + 307, + 254, + 315, + 262 + ], + "score": 0.7, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "denotes singular values (see section 4.1). (b)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 190, + 276 + ], + "score": 1.0, + "content": "Quadratic networks,", + "type": "text" + }, + { + "bbox": [ + 190, + 264, + 198, + 273 + ], + "score": 0.74, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 263, + 496, + 276 + ], + "score": 1.0, + "content": "denotes singular values (see section 4.2). (c) Depth 3 diagonal networks,", + "type": "text" + }, + { + "bbox": [ + 497, + 265, + 504, + 273 + ], + "score": 0.71, + "content": "\\sigma", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 462, + 287 + ], + "score": 1.0, + "content": "denotes feature weights (see section 4.3). (d) Depth 3 circular-convolutional networks,", + "type": "text" + }, + { + "bbox": [ + 463, + 276, + 470, + 284 + ], + "score": 0.75, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 285, + 414, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 414, + 298 + ], + "score": 1.0, + "content": "amplitudes in the frequency domain of the feature weights (see appendix G).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 107, + 318, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "Arora et al. (2019) explore the dynamics of such models, showing the singular values are learned", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "at different rates and that deeper models exhibit stronger incremental learning dynamics. Our work", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "deals with a more simplified setting, allowing us to determine explicitly under which conditions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 352, + 281, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 281, + 365 + ], + "score": 1.0, + "content": "depth leads to this dynamical phenomenon.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 504, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Finally, the learning dynamics of nonlinear models have been studied as well. Combes et al. (2018)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "and Williams et al. (2019) study the gradient flow dynamics of shallow ReLU networks under re-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "strictive distributional assumptions, Ronen et al. (2019) show that shallow networks learn functions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "of gradually increasing frequencies and Nakkiran et al. (2019) show how deep ReLU networks cor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 412, + 339, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 339, + 426 + ], + "score": 1.0, + "content": "relate with linear classifiers in the early stages of training.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "These findings, along with others, suggest that the generalization ability of deep networks is at least", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "score": 1.0, + "content": "in part due to the incremental learning dynamics of gradient descent. Following this line of work,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "we begin by explicitly defining the notion of incremental learning for a toy model which exhibits", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "this sort of behavior. Analyzing the dynamics of the model for gradient flow and gradient descent,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 472, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 487 + ], + "score": 1.0, + "content": "we characterize the effect of the model’s depth and initialization scale on incremental learning,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "showing how deeper models allow for incremental learning in larger (realistic) initialization scales.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "Specifically, we show that a depth-2 model requires exponentially small initialization for incremental", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 507, + 486, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 486, + 519 + ], + "score": 1.0, + "content": "learning to occur, while deeper models only require the initialization to be polynomially small.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 108, + 523, + 505, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "Once incremental learning has been defined and characterized for the toy model, we generalize our", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "results theoretically and empirically for larger linear and quadratic models. Examples of incremental", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 545, + 454, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 454, + 558 + ], + "score": 1.0, + "content": "learning in these models can be seen in figure 1, which we discuss further in section 4.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 573, + 340, + 586 + ], + "lines": [ + { + "bbox": [ + 104, + 573, + 342, + 588 + ], + "spans": [ + { + "bbox": [ + 104, + 573, + 342, + 588 + ], + "score": 1.0, + "content": "2 DYNAMICAL ANALYSIS OF A TOY MODEL", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "We begin by analyzing incremental learning for a simple model. This will allow us to gain a clear", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "understanding of the phenomenon and the conditions for it, which we will later be able to apply to a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 622, + 366, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 366, + 634 + ], + "score": 1.0, + "content": "variety of other models in which incremental learning is present.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 107, + 648, + 200, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 647, + 201, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 201, + 660 + ], + "score": 1.0, + "content": "2.1 PRELIMINARIES", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 669, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Our simple linear model will be similar to the toy model analyzed by Woodworth et al. (2019).", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 203, + 693 + ], + "score": 1.0, + "content": "Our input space will be", + "type": "text" + }, + { + "bbox": [ + 204, + 680, + 239, + 690 + ], + "score": 0.91, + "content": "\\mathcal { X } = \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "and the hypothesis space will be linear models with non-negative", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 691, + 184, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 184, + 703 + ], + "score": 1.0, + "content": "weights, such that:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 718, + 365, + 734 + ], + "lines": [ + { + "bbox": [ + 246, + 718, + 365, + 734 + ], + "spans": [ + { + "bbox": [ + 246, + 718, + 365, + 734 + ], + "score": 0.92, + "content": "f _ { \\sigma } ( { \\boldsymbol { x } } ) = \\langle \\sigma , { \\boldsymbol { x } } \\rangle \\qquad \\sigma \\in \\mathbb { R } _ { \\geq 0 } ^ { d }", + "type": "interline_equation", + "image_path": "499720dd31f433eff0df02a5e5569bd3b2c566437ae3c79f1dbed5637444a0cd.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 246, + 718, + 365, + 734 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 83, + 495, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 83, + 495, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 83, + 495, + 189 + ], + "spans": [ + { + "bbox": [ + 113, + 83, + 495, + 189 + ], + "score": 0.972, + "type": "image", + "image_path": "14126dc03eaae57887ee79a34684cde80a2ac3af64c8443f5095c0fc7eaad71c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 83, + 495, + 118.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 118.33333333333334, + 495, + 153.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 153.66666666666669, + 495, + 189.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 505, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "Figure 1: Incremental learning dynamics in deep models. Each panel shows the evolution of the five", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 176, + 221 + ], + "score": 1.0, + "content": "largest values of", + "type": "text" + }, + { + "bbox": [ + 176, + 210, + 183, + 218 + ], + "score": 0.74, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 208, + 505, + 221 + ], + "score": 1.0, + "content": ", the parameters of the induced model. All models were trained using gradient", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "descent with a small initialization and learning rate, on a small training set such that there are mul-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "tiple possible solutions. In all cases, the deep parameterization of the models lead to “incremental", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "learning”, where the values are learned at different rates (larger values are learned first), leading", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 307, + 265 + ], + "score": 1.0, + "content": "to sparse solutions. (a) Depth 4 matrix sensing,", + "type": "text" + }, + { + "bbox": [ + 307, + 254, + 315, + 262 + ], + "score": 0.7, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "denotes singular values (see section 4.1). (b)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 190, + 276 + ], + "score": 1.0, + "content": "Quadratic networks,", + "type": "text" + }, + { + "bbox": [ + 190, + 264, + 198, + 273 + ], + "score": 0.74, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 263, + 496, + 276 + ], + "score": 1.0, + "content": "denotes singular values (see section 4.2). (c) Depth 3 diagonal networks,", + "type": "text" + }, + { + "bbox": [ + 497, + 265, + 504, + 273 + ], + "score": 0.71, + "content": "\\sigma", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 462, + 287 + ], + "score": 1.0, + "content": "denotes feature weights (see section 4.3). (d) Depth 3 circular-convolutional networks,", + "type": "text" + }, + { + "bbox": [ + 463, + 276, + 470, + 284 + ], + "score": 0.75, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 285, + 414, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 414, + 298 + ], + "score": 1.0, + "content": "amplitudes in the frequency domain of the feature weights (see appendix G).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 107, + 318, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "Arora et al. (2019) explore the dynamics of such models, showing the singular values are learned", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "at different rates and that deeper models exhibit stronger incremental learning dynamics. Our work", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "deals with a more simplified setting, allowing us to determine explicitly under which conditions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 352, + 281, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 281, + 365 + ], + "score": 1.0, + "content": "depth leads to this dynamical phenomenon.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 319, + 506, + 365 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 504, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Finally, the learning dynamics of nonlinear models have been studied as well. Combes et al. (2018)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "and Williams et al. (2019) study the gradient flow dynamics of shallow ReLU networks under re-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "strictive distributional assumptions, Ronen et al. (2019) show that shallow networks learn functions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "of gradually increasing frequencies and Nakkiran et al. (2019) show how deep ReLU networks cor-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 412, + 339, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 339, + 426 + ], + "score": 1.0, + "content": "relate with linear classifiers in the early stages of training.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 369, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "These findings, along with others, suggest that the generalization ability of deep networks is at least", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "score": 1.0, + "content": "in part due to the incremental learning dynamics of gradient descent. Following this line of work,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "we begin by explicitly defining the notion of incremental learning for a toy model which exhibits", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "this sort of behavior. Analyzing the dynamics of the model for gradient flow and gradient descent,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 472, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 487 + ], + "score": 1.0, + "content": "we characterize the effect of the model’s depth and initialization scale on incremental learning,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "showing how deeper models allow for incremental learning in larger (realistic) initialization scales.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "Specifically, we show that a depth-2 model requires exponentially small initialization for incremental", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 507, + 486, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 486, + 519 + ], + "score": 1.0, + "content": "learning to occur, while deeper models only require the initialization to be polynomially small.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 429, + 506, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 523, + 505, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "Once incremental learning has been defined and characterized for the toy model, we generalize our", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "results theoretically and empirically for larger linear and quadratic models. Examples of incremental", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 545, + 454, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 454, + 558 + ], + "score": 1.0, + "content": "learning in these models can be seen in figure 1, which we discuss further in section 4.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 522, + 505, + 558 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 573, + 340, + 586 + ], + "lines": [ + { + "bbox": [ + 104, + 573, + 342, + 588 + ], + "spans": [ + { + "bbox": [ + 104, + 573, + 342, + 588 + ], + "score": 1.0, + "content": "2 DYNAMICAL ANALYSIS OF A TOY MODEL", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "We begin by analyzing incremental learning for a simple model. This will allow us to gain a clear", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "understanding of the phenomenon and the conditions for it, which we will later be able to apply to a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 622, + 366, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 366, + 634 + ], + "score": 1.0, + "content": "variety of other models in which incremental learning is present.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 599, + 506, + 634 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 648, + 200, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 647, + 201, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 201, + 660 + ], + "score": 1.0, + "content": "2.1 PRELIMINARIES", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 669, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Our simple linear model will be similar to the toy model analyzed by Woodworth et al. (2019).", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 203, + 693 + ], + "score": 1.0, + "content": "Our input space will be", + "type": "text" + }, + { + "bbox": [ + 204, + 680, + 239, + 690 + ], + "score": 0.91, + "content": "\\mathcal { X } = \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "and the hypothesis space will be linear models with non-negative", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 691, + 184, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 184, + 703 + ], + "score": 1.0, + "content": "weights, such that:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 668, + 505, + 703 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 718, + 365, + 734 + ], + "lines": [ + { + "bbox": [ + 246, + 718, + 365, + 734 + ], + "spans": [ + { + "bbox": [ + 246, + 718, + 365, + 734 + ], + "score": 0.92, + "content": "f _ { \\sigma } ( { \\boldsymbol { x } } ) = \\langle \\sigma , { \\boldsymbol { x } } \\rangle \\qquad \\sigma \\in \\mathbb { R } _ { \\geq 0 } ^ { d }", + "type": "interline_equation", + "image_path": "499720dd31f433eff0df02a5e5569bd3b2c566437ae3c79f1dbed5637444a0cd.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 246, + 718, + 365, + 734 + ], + "spans": [], + "index": 40 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 503, + 96 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 343, + 98 + ], + "score": 1.0, + "content": "We will introduce depth into our model, by parameterizing", + "type": "text" + }, + { + "bbox": [ + 343, + 84, + 351, + 93 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 78, + 375, + 98 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 376, + 82, + 414, + 96 + ], + "score": 0.96, + "content": "w \\in \\mathbb { R } _ { \\geq 0 } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "in the following way:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 108, + 335, + 123 + ], + "lines": [ + { + "bbox": [ + 275, + 108, + 335, + 123 + ], + "spans": [ + { + "bbox": [ + 275, + 108, + 335, + 123 + ], + "score": 0.91, + "content": "\\forall i : \\sigma _ { i } = w _ { i } ^ { N }", + "type": "interline_equation", + "image_path": "21315d50d993256a7c23bc4317b615a3b2bef2d2c90d55cbde7842b28fdb5289.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 275, + 108, + 335, + 123 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 137, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 137, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 136, + 151 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 137, + 139, + 147, + 148 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 137, + 505, + 151 + ], + "score": 1.0, + "content": "represents the depth of the model. Since we restrict the model to having non-negative", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "weights, this parameterization doesn’t change the expressiveness, but it does radically change it’s", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 161, + 202, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 202, + 172 + ], + "score": 1.0, + "content": "optimization dynamics.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 101, + 170, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 101, + 170, + 266, + 198 + ], + "score": 1.0, + "content": "Assuming the data is labeled by some", + "type": "text" + }, + { + "bbox": [ + 266, + 176, + 311, + 191 + ], + "score": 0.94, + "content": "\\sigma ^ { * } \\in \\mathbb { R } _ { \\geq 0 } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 170, + 505, + 198 + ], + "score": 1.0, + "content": ", we will study the dynamics of this model for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 139, + 203 + ], + "score": 1.0, + "content": "general", + "type": "text" + }, + { + "bbox": [ + 150, + 190, + 506, + 203 + ], + "score": 1.0, + "content": "under a depth-normalized1 squared loss over Gaussian inputs, which will allow us to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 201, + 228, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 228, + 214 + ], + "score": 1.0, + "content": "derive our analytical solution:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 178, + 226, + 432, + 250 + ], + "lines": [ + { + "bbox": [ + 178, + 226, + 432, + 250 + ], + "spans": [ + { + "bbox": [ + 178, + 226, + 432, + 250 + ], + "score": 0.92, + "content": "\\ell _ { N } ( w ) = \\frac { 1 } { 2 N ^ { 2 } } \\mathbb { E } _ { x } [ ( \\langle \\sigma ^ { * } , x \\rangle - \\langle w ^ { N } , x \\rangle ) ^ { 2 } ] = \\frac { 1 } { 2 N ^ { 2 } } | | w ^ { N } - \\sigma ^ { * } | | ^ { 2 }", + "type": "interline_equation", + "image_path": "b4ccf63aab033eb8c116f99b44fa7690c8b7a66e9b5d6d770be36beeb019e707.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 178, + 226, + 432, + 250 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 255, + 484, + 267 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 484, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 484, + 269 + ], + "score": 1.0, + "content": "We will assume that our model is initialized uniformly with a tunable scaling factor, such that:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 282, + 347, + 296 + ], + "lines": [ + { + "bbox": [ + 264, + 282, + 347, + 296 + ], + "spans": [ + { + "bbox": [ + 264, + 282, + 347, + 296 + ], + "score": 0.9, + "content": "\\forall i : \\ w _ { i } ( 0 ) = \\ \\sqrt [ N ] { \\sigma _ { 0 } }", + "type": "interline_equation", + "image_path": "e952567285b6a26ea557f8a66e804174635f29b768fd1db13491785e3438b0dc.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 264, + 282, + 347, + 296 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 307, + 318, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 318, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 318, + 321 + ], + "score": 1.0, + "content": "2.2 GRADIENT FLOW ANALYTICAL SOLUTIONS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "Analyzing our toy model using gradient flow allows us to obtain an analytical solution for the dy-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 151, + 352 + ], + "score": 1.0, + "content": "namics of", + "type": "text" + }, + { + "bbox": [ + 151, + 339, + 169, + 352 + ], + "score": 0.92, + "content": "\\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 339, + 420, + 352 + ], + "score": 1.0, + "content": "along with the dynamics of the loss function for a general", + "type": "text" + }, + { + "bbox": [ + 420, + 340, + 430, + 349 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 339, + 505, + 352 + ], + "score": 1.0, + "content": ". For brevity, the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 238, + 363 + ], + "score": 1.0, + "content": "following theorem refers only to", + "type": "text" + }, + { + "bbox": [ + 239, + 351, + 276, + 362 + ], + "score": 0.91, + "content": "N = 1 , 2", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 349, + 294, + 363 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 295, + 351, + 330, + 361 + ], + "score": 0.9, + "content": "N \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 349, + 437, + 363 + ], + "score": 1.0, + "content": ", however the solutions for", + "type": "text" + }, + { + "bbox": [ + 438, + 351, + 489, + 361 + ], + "score": 0.91, + "content": "3 \\leq N < \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 194, + 374 + ], + "score": 1.0, + "content": "similar in structure to", + "type": "text" + }, + { + "bbox": [ + 194, + 362, + 230, + 371 + ], + "score": 0.9, + "content": "N \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 361, + 391, + 374 + ], + "score": 1.0, + "content": ", but more complicated. We also assume", + "type": "text" + }, + { + "bbox": [ + 392, + 362, + 421, + 374 + ], + "score": 0.92, + "content": "\\sigma _ { i } ^ { * } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "for brevity, however", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 373, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 233, + 384 + ], + "score": 1.0, + "content": "we can derive the solutions for", + "type": "text" + }, + { + "bbox": [ + 233, + 373, + 264, + 384 + ], + "score": 0.92, + "content": "\\sigma _ { i } ^ { * } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 373, + 506, + 384 + ], + "score": 1.0, + "content": "as well. Note that this result is a special case adaptation of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 384, + 363, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 363, + 395 + ], + "score": 1.0, + "content": "the one presented in Saxe et al. (2013) for deep linear networks:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Theorem 1. Minimizing the toy linear model described in (1) with gradient flow over the depth", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "normalized squared loss (2), with Gaussian inputs and weights initialized as in (3) and assuming", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 419, + 410, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 137, + 431 + ], + "score": 0.91, + "content": "\\sigma _ { i } ^ { * } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 419, + 395, + 432 + ], + "score": 1.0, + "content": "leads to the following analytical solutions for different values of", + "type": "text" + }, + { + "bbox": [ + 396, + 420, + 406, + 429 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 419, + 410, + 432 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 446, + 428, + 517 + ], + "lines": [ + { + "bbox": [ + 182, + 446, + 428, + 517 + ], + "spans": [ + { + "bbox": [ + 182, + 446, + 428, + 517 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\begin{array} { r l } { N = 1 : } & { \\sigma _ { i } ( t ) = \\sigma _ { i } ^ { * } + ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) e ^ { - t } } \\\\ { N = 2 : } & { \\sigma _ { i } ( t ) = \\frac { \\sigma _ { 0 } \\sigma _ { i } ^ { * } e ^ { \\sigma _ { i } ^ { * } t } } { \\sigma _ { 0 } \\left( e ^ { \\sigma _ { i } ^ { * } t } - 1 \\right) + \\sigma _ { i } ^ { * } } } \\end{array} } \\\\ & { \\begin{array} { r l } { N \\to \\infty : } & { t = \\frac { 1 } { ( \\sigma _ { i } ^ { * } ) ^ { 2 } } \\log \\big ( \\frac { \\sigma _ { i } ( t ) ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) } { \\sigma _ { 0 } \\left( \\sigma _ { i } ( t ) - \\sigma _ { i } ^ { * } \\right) } \\big ) - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { 0 } } \\big ) } \\end{array} } \\end{array}", + "type": "interline_equation", + "image_path": "afc342ab27650fa4f36a85ae581655158219108cacb9a75e1a24550ca023e60b.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 182, + 446, + 428, + 469.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 182, + 469.6666666666667, + 428, + 493.33333333333337 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 182, + 493.33333333333337, + 428, + 517.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 380, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 381, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 381, + 551 + ], + "score": 1.0, + "content": "Proof. The gradient flow equations for our model are the following:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 560, + 381, + 584 + ], + "lines": [ + { + "bbox": [ + 230, + 560, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 230, + 560, + 381, + 584 + ], + "score": 0.94, + "content": "\\dot { w } _ { i } = - \\nabla _ { w _ { i } } \\ell = \\frac { 1 } { N } w _ { i } ^ { N - 1 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } )", + "type": "interline_equation", + "image_path": "8b6f82c9fadbc78ec694e51acfb72295a15e0ada8ab4b7e03abf74b1bcfd0aa6.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 230, + 560, + 381, + 584 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 504, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 216, + 603 + ], + "score": 1.0, + "content": "Given the dynamics of the", + "type": "text" + }, + { + "bbox": [ + 216, + 592, + 225, + 600 + ], + "score": 0.78, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 589, + 506, + 603 + ], + "score": 1.0, + "content": "parameters, we may use the chain rule to derive the dynamics of the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 601, + 182, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 170, + 613 + ], + "score": 1.0, + "content": "induced model,", + "type": "text" + }, + { + "bbox": [ + 171, + 603, + 178, + 611 + ], + "score": 0.73, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 601, + 182, + 613 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 625, + 415, + 651 + ], + "lines": [ + { + "bbox": [ + 196, + 625, + 415, + 651 + ], + "spans": [ + { + "bbox": [ + 196, + 625, + 415, + 651 + ], + "score": 0.93, + "content": "\\dot { \\sigma } _ { i } = \\frac { d \\sigma _ { i } } { d w _ { i } } \\dot { w } _ { i } = w ^ { 2 N - 2 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } ) = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "e69396f523489b074fb323b0691c882f33574cad94f31b60a25ff774af8ce0fe.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 196, + 625, + 415, + 651 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 657, + 504, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 290, + 671 + ], + "score": 1.0, + "content": "This differential equation is solvable for all", + "type": "text" + }, + { + "bbox": [ + 290, + 658, + 300, + 667 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 655, + 506, + 671 + ], + "score": 1.0, + "content": ", leading to the solutions in the theorem. Taking", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 667, + 371, + 680 + ], + "spans": [ + { + "bbox": [ + 107, + 669, + 142, + 678 + ], + "score": 0.85, + "content": "N \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 667, + 201, + 680 + ], + "score": 1.0, + "content": "in (4) leads to", + "type": "text" + }, + { + "bbox": [ + 201, + 667, + 275, + 680 + ], + "score": 0.93, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ^ { 2 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 667, + 371, + 680 + ], + "score": 1.0, + "content": ", which is also solvable.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 701, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "1This normalization is used for mathematical convenience to have solutions of different depths exhibit", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "similar time scales in their dynamics. Equivalently, we can derive the solutions for the regular square loss and", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 315, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 315, + 732 + ], + "score": 1.0, + "content": "then use different time scalings in the dynamical analysis.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 495, + 685, + 504, + 695 + ], + "lines": [ + { + "bbox": [ + 496, + 686, + 504, + 695 + ], + "spans": [ + { + "bbox": [ + 496, + 686, + 504, + 695 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 503, + 96 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 343, + 98 + ], + "score": 1.0, + "content": "We will introduce depth into our model, by parameterizing", + "type": "text" + }, + { + "bbox": [ + 343, + 84, + 351, + 93 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 78, + 375, + 98 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 376, + 82, + 414, + 96 + ], + "score": 0.96, + "content": "w \\in \\mathbb { R } _ { \\geq 0 } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "in the following way:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 104, + 78, + 505, + 98 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 108, + 335, + 123 + ], + "lines": [ + { + "bbox": [ + 275, + 108, + 335, + 123 + ], + "spans": [ + { + "bbox": [ + 275, + 108, + 335, + 123 + ], + "score": 0.91, + "content": "\\forall i : \\sigma _ { i } = w _ { i } ^ { N }", + "type": "interline_equation", + "image_path": "21315d50d993256a7c23bc4317b615a3b2bef2d2c90d55cbde7842b28fdb5289.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 275, + 108, + 335, + 123 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 137, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 106, + 137, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 136, + 151 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 137, + 139, + 147, + 148 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 137, + 505, + 151 + ], + "score": 1.0, + "content": "represents the depth of the model. Since we restrict the model to having non-negative", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "weights, this parameterization doesn’t change the expressiveness, but it does radically change it’s", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 161, + 202, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 202, + 172 + ], + "score": 1.0, + "content": "optimization dynamics.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 137, + 506, + 172 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 101, + 170, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 101, + 170, + 266, + 198 + ], + "score": 1.0, + "content": "Assuming the data is labeled by some", + "type": "text" + }, + { + "bbox": [ + 266, + 176, + 311, + 191 + ], + "score": 0.94, + "content": "\\sigma ^ { * } \\in \\mathbb { R } _ { \\geq 0 } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 170, + 505, + 198 + ], + "score": 1.0, + "content": ", we will study the dynamics of this model for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 139, + 203 + ], + "score": 1.0, + "content": "general", + "type": "text" + }, + { + "bbox": [ + 150, + 190, + 506, + 203 + ], + "score": 1.0, + "content": "under a depth-normalized1 squared loss over Gaussian inputs, which will allow us to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 201, + 228, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 228, + 214 + ], + "score": 1.0, + "content": "derive our analytical solution:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 101, + 170, + 506, + 214 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 178, + 226, + 432, + 250 + ], + "lines": [ + { + "bbox": [ + 178, + 226, + 432, + 250 + ], + "spans": [ + { + "bbox": [ + 178, + 226, + 432, + 250 + ], + "score": 0.92, + "content": "\\ell _ { N } ( w ) = \\frac { 1 } { 2 N ^ { 2 } } \\mathbb { E } _ { x } [ ( \\langle \\sigma ^ { * } , x \\rangle - \\langle w ^ { N } , x \\rangle ) ^ { 2 } ] = \\frac { 1 } { 2 N ^ { 2 } } | | w ^ { N } - \\sigma ^ { * } | | ^ { 2 }", + "type": "interline_equation", + "image_path": "b4ccf63aab033eb8c116f99b44fa7690c8b7a66e9b5d6d770be36beeb019e707.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 178, + 226, + 432, + 250 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 255, + 484, + 267 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 484, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 484, + 269 + ], + "score": 1.0, + "content": "We will assume that our model is initialized uniformly with a tunable scaling factor, such that:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 255, + 484, + 269 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 282, + 347, + 296 + ], + "lines": [ + { + "bbox": [ + 264, + 282, + 347, + 296 + ], + "spans": [ + { + "bbox": [ + 264, + 282, + 347, + 296 + ], + "score": 0.9, + "content": "\\forall i : \\ w _ { i } ( 0 ) = \\ \\sqrt [ N ] { \\sigma _ { 0 } }", + "type": "interline_equation", + "image_path": "e952567285b6a26ea557f8a66e804174635f29b768fd1db13491785e3438b0dc.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 264, + 282, + 347, + 296 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 307, + 318, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 318, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 318, + 321 + ], + "score": 1.0, + "content": "2.2 GRADIENT FLOW ANALYTICAL SOLUTIONS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "Analyzing our toy model using gradient flow allows us to obtain an analytical solution for the dy-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 151, + 352 + ], + "score": 1.0, + "content": "namics of", + "type": "text" + }, + { + "bbox": [ + 151, + 339, + 169, + 352 + ], + "score": 0.92, + "content": "\\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 339, + 420, + 352 + ], + "score": 1.0, + "content": "along with the dynamics of the loss function for a general", + "type": "text" + }, + { + "bbox": [ + 420, + 340, + 430, + 349 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 339, + 505, + 352 + ], + "score": 1.0, + "content": ". For brevity, the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 238, + 363 + ], + "score": 1.0, + "content": "following theorem refers only to", + "type": "text" + }, + { + "bbox": [ + 239, + 351, + 276, + 362 + ], + "score": 0.91, + "content": "N = 1 , 2", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 349, + 294, + 363 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 295, + 351, + 330, + 361 + ], + "score": 0.9, + "content": "N \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 349, + 437, + 363 + ], + "score": 1.0, + "content": ", however the solutions for", + "type": "text" + }, + { + "bbox": [ + 438, + 351, + 489, + 361 + ], + "score": 0.91, + "content": "3 \\leq N < \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 194, + 374 + ], + "score": 1.0, + "content": "similar in structure to", + "type": "text" + }, + { + "bbox": [ + 194, + 362, + 230, + 371 + ], + "score": 0.9, + "content": "N \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 361, + 391, + 374 + ], + "score": 1.0, + "content": ", but more complicated. We also assume", + "type": "text" + }, + { + "bbox": [ + 392, + 362, + 421, + 374 + ], + "score": 0.92, + "content": "\\sigma _ { i } ^ { * } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "for brevity, however", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 373, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 233, + 384 + ], + "score": 1.0, + "content": "we can derive the solutions for", + "type": "text" + }, + { + "bbox": [ + 233, + 373, + 264, + 384 + ], + "score": 0.92, + "content": "\\sigma _ { i } ^ { * } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 373, + 506, + 384 + ], + "score": 1.0, + "content": "as well. Note that this result is a special case adaptation of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 384, + 363, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 363, + 395 + ], + "score": 1.0, + "content": "the one presented in Saxe et al. (2013) for deep linear networks:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 328, + 506, + 395 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Theorem 1. Minimizing the toy linear model described in (1) with gradient flow over the depth", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 421 + ], + "score": 1.0, + "content": "normalized squared loss (2), with Gaussian inputs and weights initialized as in (3) and assuming", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 419, + 410, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 137, + 431 + ], + "score": 0.91, + "content": "\\sigma _ { i } ^ { * } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 419, + 395, + 432 + ], + "score": 1.0, + "content": "leads to the following analytical solutions for different values of", + "type": "text" + }, + { + "bbox": [ + 396, + 420, + 406, + 429 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 419, + 410, + 432 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 396, + 505, + 432 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 446, + 428, + 517 + ], + "lines": [ + { + "bbox": [ + 182, + 446, + 428, + 517 + ], + "spans": [ + { + "bbox": [ + 182, + 446, + 428, + 517 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\begin{array} { r l } { N = 1 : } & { \\sigma _ { i } ( t ) = \\sigma _ { i } ^ { * } + ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) e ^ { - t } } \\\\ { N = 2 : } & { \\sigma _ { i } ( t ) = \\frac { \\sigma _ { 0 } \\sigma _ { i } ^ { * } e ^ { \\sigma _ { i } ^ { * } t } } { \\sigma _ { 0 } \\left( e ^ { \\sigma _ { i } ^ { * } t } - 1 \\right) + \\sigma _ { i } ^ { * } } } \\end{array} } \\\\ & { \\begin{array} { r l } { N \\to \\infty : } & { t = \\frac { 1 } { ( \\sigma _ { i } ^ { * } ) ^ { 2 } } \\log \\big ( \\frac { \\sigma _ { i } ( t ) ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) } { \\sigma _ { 0 } \\left( \\sigma _ { i } ( t ) - \\sigma _ { i } ^ { * } \\right) } \\big ) - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { 0 } } \\big ) } \\end{array} } \\end{array}", + "type": "interline_equation", + "image_path": "afc342ab27650fa4f36a85ae581655158219108cacb9a75e1a24550ca023e60b.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 182, + 446, + 428, + 469.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 182, + 469.6666666666667, + 428, + 493.33333333333337 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 182, + 493.33333333333337, + 428, + 517.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 380, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 381, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 381, + 551 + ], + "score": 1.0, + "content": "Proof. The gradient flow equations for our model are the following:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 534, + 381, + 551 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 560, + 381, + 584 + ], + "lines": [ + { + "bbox": [ + 230, + 560, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 230, + 560, + 381, + 584 + ], + "score": 0.94, + "content": "\\dot { w } _ { i } = - \\nabla _ { w _ { i } } \\ell = \\frac { 1 } { N } w _ { i } ^ { N - 1 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } )", + "type": "interline_equation", + "image_path": "8b6f82c9fadbc78ec694e51acfb72295a15e0ada8ab4b7e03abf74b1bcfd0aa6.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 230, + 560, + 381, + 584 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 504, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 216, + 603 + ], + "score": 1.0, + "content": "Given the dynamics of the", + "type": "text" + }, + { + "bbox": [ + 216, + 592, + 225, + 600 + ], + "score": 0.78, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 589, + 506, + 603 + ], + "score": 1.0, + "content": "parameters, we may use the chain rule to derive the dynamics of the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 601, + 182, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 170, + 613 + ], + "score": 1.0, + "content": "induced model,", + "type": "text" + }, + { + "bbox": [ + 171, + 603, + 178, + 611 + ], + "score": 0.73, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 601, + 182, + 613 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 589, + 506, + 613 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 625, + 415, + 651 + ], + "lines": [ + { + "bbox": [ + 196, + 625, + 415, + 651 + ], + "spans": [ + { + "bbox": [ + 196, + 625, + 415, + 651 + ], + "score": 0.93, + "content": "\\dot { \\sigma } _ { i } = \\frac { d \\sigma _ { i } } { d w _ { i } } \\dot { w } _ { i } = w ^ { 2 N - 2 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } ) = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "e69396f523489b074fb323b0691c882f33574cad94f31b60a25ff774af8ce0fe.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 196, + 625, + 415, + 651 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 657, + 504, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 290, + 671 + ], + "score": 1.0, + "content": "This differential equation is solvable for all", + "type": "text" + }, + { + "bbox": [ + 290, + 658, + 300, + 667 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 655, + 506, + 671 + ], + "score": 1.0, + "content": ", leading to the solutions in the theorem. Taking", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 667, + 371, + 680 + ], + "spans": [ + { + "bbox": [ + 107, + 669, + 142, + 678 + ], + "score": 0.85, + "content": "N \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 667, + 201, + 680 + ], + "score": 1.0, + "content": "in (4) leads to", + "type": "text" + }, + { + "bbox": [ + 201, + 667, + 275, + 680 + ], + "score": 0.93, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ^ { 2 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 667, + 371, + 680 + ], + "score": 1.0, + "content": ", which is also solvable.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 655, + 506, + 680 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 84, + 498, + 275 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 84, + 498, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 84, + 498, + 275 + ], + "spans": [ + { + "bbox": [ + 110, + 84, + 498, + 275 + ], + "score": 0.972, + "type": "image", + "image_path": "c32e4b6822cc4e3b3801507a0f18a13c9706cba7f76cc5695c9ee740ff4a57cf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 84, + 498, + 147.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 147.66666666666666, + 498, + 211.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 211.33333333333331, + 498, + 275.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 295, + 505, + 365 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 103, + 293, + 504, + 311 + ], + "spans": [ + { + "bbox": [ + 103, + 293, + 483, + 311 + ], + "score": 1.0, + "content": "Figure 2: Incremental learning dynamics in the toy model. Each panel shows the evolution of", + "type": "text" + }, + { + "bbox": [ + 484, + 293, + 504, + 311 + ], + "score": 0.92, + "content": "\\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 122, + 322 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 309, + 199, + 322 + ], + "score": 0.92, + "content": "\\sigma _ { i } ^ { * } \\in \\{ 1 2 , 6 , 4 , 3 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "according to the analytical solutions in theorem 1, under different depths", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "and initializations. The first column has all values converging at the same rate. Notice how the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "deep parameterization with small initialization leads to distinct phases of learning, where values are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "learned incrementally (bottom-right). The shallow model’s much weaker incremental learning, even", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 354, + 397, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 397, + 366 + ], + "score": 1.0, + "content": "at small initialization scales (second column), is explained in theorem 2.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "Analyzing these solutions, we see how even in such a simple model depth causes different factors of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "the model to be learned at different rates. Specifically, values corresponding to larger optimal values", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 433, + 424 + ], + "score": 1.0, + "content": "converge faster, suggesting a form of incremental learning. This is most clear for", + "type": "text" + }, + { + "bbox": [ + 433, + 412, + 462, + 422 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "where the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 423, + 470, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 288, + 435 + ], + "score": 1.0, + "content": "solution isn’t implicit, but is also the case for", + "type": "text" + }, + { + "bbox": [ + 289, + 423, + 317, + 434 + ], + "score": 0.9, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 423, + 470, + 435 + ], + "score": 1.0, + "content": ", as we will see in the next subsection.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 485, + 452 + ], + "score": 1.0, + "content": "These dynamics are depicted in figure 2, where we see the dynamics of the different values of", + "type": "text" + }, + { + "bbox": [ + 486, + 440, + 504, + 452 + ], + "score": 0.9, + "content": "\\sigma ( t )", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 451, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 234, + 463 + ], + "score": 1.0, + "content": "as learning progresses. When", + "type": "text" + }, + { + "bbox": [ + 234, + 451, + 267, + 461 + ], + "score": 0.89, + "content": "N = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 451, + 506, + 463 + ], + "score": 1.0, + "content": ", all values are learned at the same rate regardless of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "initialization, while the deeper models are clearly biased towards learning the larger singular values", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 473, + 285, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 285, + 485 + ], + "score": 1.0, + "content": "first, especially at small initialization scales.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "Our model has only one optimal solution due to the population loss, but it is clear how this sort of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "dynamic can induce sparse solutions - if the model is able to fit the data after a small amount of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 512, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 399, + 523 + ], + "score": 1.0, + "content": "learning phases, then it’s obtained result will be sparse. Alternatively, if", + "type": "text" + }, + { + "bbox": [ + 399, + 512, + 429, + 522 + ], + "score": 0.9, + "content": "N = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 512, + 505, + 523 + ], + "score": 1.0, + "content": ", we know that the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 520, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 242, + 537 + ], + "score": 1.0, + "content": "dynamics will lead to the minimal", + "type": "text" + }, + { + "bbox": [ + 242, + 523, + 252, + 533 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 520, + 506, + 537 + ], + "score": 1.0, + "content": "norm solution which is dense. We explore the sparsity inducing", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "bias of our toy model by comparing it empirically2 to a greedy sparse approximation algorithm in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 545, + 361, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 361, + 556 + ], + "score": 1.0, + "content": "appendix D, and give our theoretical results in the next section.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 109, + 575, + 259, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 261, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 261, + 590 + ], + "score": 1.0, + "content": "3 INCREMENTAL LEARNING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 602, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "Equipped with analytical solutions for the dynamics of our model for every depth, we turn to study", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "score": 1.0, + "content": "how the depth and initialization effect incremental learning. While Gidel et al. (2019) focuses on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 322, + 637 + ], + "score": 1.0, + "content": "incremental learning in depth-2 models at the limit of", + "type": "text" + }, + { + "bbox": [ + 323, + 626, + 355, + 636 + ], + "score": 0.91, + "content": "\\sigma _ { 0 } 0", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 624, + 506, + 637 + ], + "score": 1.0, + "content": ", we will study the phenomenon for a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 636, + 228, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 194, + 648 + ], + "score": 1.0, + "content": "general depth and for", + "type": "text" + }, + { + "bbox": [ + 194, + 636, + 223, + 647 + ], + "score": 0.91, + "content": "\\sigma _ { 0 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 636, + 228, + 648 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 653, + 505, + 697 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 405, + 665 + ], + "score": 1.0, + "content": "First, we will define the notion of incremental learning. Since all values of", + "type": "text" + }, + { + "bbox": [ + 405, + 655, + 412, + 663 + ], + "score": 0.74, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 652, + 505, + 665 + ], + "score": 1.0, + "content": "are learned in parallel,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "score": 1.0, + "content": "we can’t expect one value to converge before the other moves at all (which happens for infinitesimal", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "initialization as shown by Gidel et al. (2019)). We will need a more relaxed definition for incremental", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 686, + 256, + 698 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 256, + 698 + ], + "score": 1.0, + "content": "learning in finite initialization scales.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 108, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "2The code for reproducing all of our experiments can be found in https://github.com/dsgissin/", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "score": 1.0, + "content": "Incremental-Learning", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 84, + 498, + 275 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 84, + 498, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 84, + 498, + 275 + ], + "spans": [ + { + "bbox": [ + 110, + 84, + 498, + 275 + ], + "score": 0.972, + "type": "image", + "image_path": "c32e4b6822cc4e3b3801507a0f18a13c9706cba7f76cc5695c9ee740ff4a57cf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 84, + 498, + 147.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 147.66666666666666, + 498, + 211.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 211.33333333333331, + 498, + 275.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 295, + 505, + 365 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 103, + 293, + 504, + 311 + ], + "spans": [ + { + "bbox": [ + 103, + 293, + 483, + 311 + ], + "score": 1.0, + "content": "Figure 2: Incremental learning dynamics in the toy model. Each panel shows the evolution of", + "type": "text" + }, + { + "bbox": [ + 484, + 293, + 504, + 311 + ], + "score": 0.92, + "content": "\\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 122, + 322 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 309, + 199, + 322 + ], + "score": 0.92, + "content": "\\sigma _ { i } ^ { * } \\in \\{ 1 2 , 6 , 4 , 3 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "according to the analytical solutions in theorem 1, under different depths", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "and initializations. The first column has all values converging at the same rate. Notice how the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "deep parameterization with small initialization leads to distinct phases of learning, where values are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "learned incrementally (bottom-right). The shallow model’s much weaker incremental learning, even", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 354, + 397, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 397, + 366 + ], + "score": 1.0, + "content": "at small initialization scales (second column), is explained in theorem 2.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "Analyzing these solutions, we see how even in such a simple model depth causes different factors of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "the model to be learned at different rates. Specifically, values corresponding to larger optimal values", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 433, + 424 + ], + "score": 1.0, + "content": "converge faster, suggesting a form of incremental learning. This is most clear for", + "type": "text" + }, + { + "bbox": [ + 433, + 412, + 462, + 422 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "where the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 423, + 470, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 288, + 435 + ], + "score": 1.0, + "content": "solution isn’t implicit, but is also the case for", + "type": "text" + }, + { + "bbox": [ + 289, + 423, + 317, + 434 + ], + "score": 0.9, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 423, + 470, + 435 + ], + "score": 1.0, + "content": ", as we will see in the next subsection.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 390, + 506, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 485, + 452 + ], + "score": 1.0, + "content": "These dynamics are depicted in figure 2, where we see the dynamics of the different values of", + "type": "text" + }, + { + "bbox": [ + 486, + 440, + 504, + 452 + ], + "score": 0.9, + "content": "\\sigma ( t )", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 451, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 234, + 463 + ], + "score": 1.0, + "content": "as learning progresses. When", + "type": "text" + }, + { + "bbox": [ + 234, + 451, + 267, + 461 + ], + "score": 0.89, + "content": "N = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 451, + 506, + 463 + ], + "score": 1.0, + "content": ", all values are learned at the same rate regardless of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "initialization, while the deeper models are clearly biased towards learning the larger singular values", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 473, + 285, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 285, + 485 + ], + "score": 1.0, + "content": "first, especially at small initialization scales.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 439, + 506, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "Our model has only one optimal solution due to the population loss, but it is clear how this sort of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "dynamic can induce sparse solutions - if the model is able to fit the data after a small amount of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 512, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 399, + 523 + ], + "score": 1.0, + "content": "learning phases, then it’s obtained result will be sparse. Alternatively, if", + "type": "text" + }, + { + "bbox": [ + 399, + 512, + 429, + 522 + ], + "score": 0.9, + "content": "N = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 512, + 505, + 523 + ], + "score": 1.0, + "content": ", we know that the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 520, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 242, + 537 + ], + "score": 1.0, + "content": "dynamics will lead to the minimal", + "type": "text" + }, + { + "bbox": [ + 242, + 523, + 252, + 533 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 520, + 506, + 537 + ], + "score": 1.0, + "content": "norm solution which is dense. We explore the sparsity inducing", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "bias of our toy model by comparing it empirically2 to a greedy sparse approximation algorithm in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 545, + 361, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 361, + 556 + ], + "score": 1.0, + "content": "appendix D, and give our theoretical results in the next section.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 489, + 506, + 556 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 575, + 259, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 261, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 261, + 590 + ], + "score": 1.0, + "content": "3 INCREMENTAL LEARNING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 602, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 615 + ], + "score": 1.0, + "content": "Equipped with analytical solutions for the dynamics of our model for every depth, we turn to study", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "score": 1.0, + "content": "how the depth and initialization effect incremental learning. While Gidel et al. (2019) focuses on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 322, + 637 + ], + "score": 1.0, + "content": "incremental learning in depth-2 models at the limit of", + "type": "text" + }, + { + "bbox": [ + 323, + 626, + 355, + 636 + ], + "score": 0.91, + "content": "\\sigma _ { 0 } 0", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 624, + 506, + 637 + ], + "score": 1.0, + "content": ", we will study the phenomenon for a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 636, + 228, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 194, + 648 + ], + "score": 1.0, + "content": "general depth and for", + "type": "text" + }, + { + "bbox": [ + 194, + 636, + 223, + 647 + ], + "score": 0.91, + "content": "\\sigma _ { 0 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 636, + 228, + 648 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 602, + 506, + 648 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 653, + 505, + 697 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 405, + 665 + ], + "score": 1.0, + "content": "First, we will define the notion of incremental learning. Since all values of", + "type": "text" + }, + { + "bbox": [ + 405, + 655, + 412, + 663 + ], + "score": 0.74, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 652, + 505, + 665 + ], + "score": 1.0, + "content": "are learned in parallel,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "score": 1.0, + "content": "we can’t expect one value to converge before the other moves at all (which happens for infinitesimal", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "initialization as shown by Gidel et al. (2019)). We will need a more relaxed definition for incremental", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 686, + 256, + 698 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 256, + 698 + ], + "score": 1.0, + "content": "learning in finite initialization scales.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 652, + 505, + 698 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 120 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 236, + 96 + ], + "score": 1.0, + "content": "Definition 1. Given two values", + "type": "text" + }, + { + "bbox": [ + 236, + 84, + 261, + 95 + ], + "score": 0.89, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 81, + 302, + 96 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 302, + 83, + 360, + 96 + ], + "score": 0.93, + "content": "\\sigma _ { i } ^ { * } > \\sigma _ { j } ^ { * } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 81, + 469, + 96 + ], + "score": 1.0, + "content": "and both are initialized as", + "type": "text" + }, + { + "bbox": [ + 470, + 82, + 505, + 95 + ], + "score": 0.91, + "content": "\\sigma _ { i } ( 0 ) =", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 95, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 107, + 95, + 183, + 109 + ], + "score": 0.91, + "content": "\\sigma _ { j } ( 0 ) = \\sigma _ { 0 } < \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 95, + 280, + 109 + ], + "score": 1.0, + "content": ", and given two scalars", + "type": "text" + }, + { + "bbox": [ + 280, + 95, + 324, + 109 + ], + "score": 0.92, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 95, + 344, + 109 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 345, + 95, + 389, + 109 + ], + "score": 0.93, + "content": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 95, + 505, + 109 + ], + "score": 1.0, + "content": ", we call the learning of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 108, + 323, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 135, + 121 + ], + "score": 1.0, + "content": "values", + "type": "text" + }, + { + "bbox": [ + 135, + 109, + 158, + 120 + ], + "score": 0.9, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 108, + 323, + 121 + ], + "score": 1.0, + "content": "-incremental if there exists a t for which:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 135, + 363, + 150 + ], + "lines": [ + { + "bbox": [ + 247, + 135, + 363, + 150 + ], + "spans": [ + { + "bbox": [ + 247, + 135, + 363, + 150 + ], + "score": 0.91, + "content": "\\sigma _ { j } ( t ) \\leq s \\sigma _ { j } ^ { * } < f \\sigma _ { i } ^ { * } \\leq \\sigma _ { i } ( t )", + "type": "interline_equation", + "image_path": "756427e823ee307e8aa880968cbf8646de56e5172a9941219f2ea7f263d8dcf7.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 247, + 135, + 363, + 150 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 156, + 505, + 212 + ], + "lines": [ + { + "bbox": [ + 105, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 428, + 169 + ], + "score": 1.0, + "content": "In words, two values have distinct learning phases if the first almost converges", + "type": "text" + }, + { + "bbox": [ + 428, + 157, + 459, + 168 + ], + "score": 0.86, + "content": "( f \\approx 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "before the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 167, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 213, + 180 + ], + "score": 1.0, + "content": "second changes by much", + "type": "text" + }, + { + "bbox": [ + 214, + 168, + 244, + 178 + ], + "score": 0.82, + "content": "s \\ll 1 \\ r _ { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 167, + 325, + 180 + ], + "score": 1.0, + "content": "). Note that for any", + "type": "text" + }, + { + "bbox": [ + 326, + 168, + 359, + 180 + ], + "score": 0.58, + "content": "N , \\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 167, + 506, + 180 + ], + "score": 1.0, + "content": "is monotonically increasing and so", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 178, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 178, + 127, + 191 + ], + "score": 1.0, + "content": "once", + "type": "text" + }, + { + "bbox": [ + 128, + 178, + 178, + 192 + ], + "score": 0.92, + "content": "\\sigma _ { j } ( t ) = s \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 178, + 506, + 191 + ], + "score": 1.0, + "content": ", it will not decrease to allow further incremental learning. Given this definition of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 188, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 203 + ], + "score": 1.0, + "content": "incremental learning, we turn to study the conditions that facilitate incremental learning in our toy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 200, + 137, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 137, + 213 + ], + "score": 1.0, + "content": "model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "Our main result is a dynamical depth separation result, showing that incremental learning is depen-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 228, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 138, + 249 + ], + "score": 1.0, + "content": "dent on", + "type": "text" + }, + { + "bbox": [ + 139, + 228, + 150, + 247 + ], + "score": 0.87, + "content": "\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 229, + 308, + 249 + ], + "score": 1.0, + "content": "in different ways for different values of", + "type": "text" + }, + { + "bbox": [ + 308, + 231, + 318, + 241 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 229, + 505, + 249 + ], + "score": 1.0, + "content": ". The largest difference in dependence happens", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 245, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 142, + 258 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 142, + 246, + 171, + 255 + ], + "score": 0.86, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 245, + 189, + 258 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 246, + 217, + 256 + ], + "score": 0.9, + "content": "N = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 245, + 478, + 258 + ], + "score": 1.0, + "content": ", where the dependence changes from exponential to polynomial:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 506, + 315 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 234, + 281 + ], + "score": 1.0, + "content": "Theorem 2. Given two values", + "type": "text" + }, + { + "bbox": [ + 234, + 264, + 258, + 275 + ], + "score": 0.89, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 258, + 417, + 281 + ], + "score": 1.0, + "content": "of a toy linear model as in (1), where", + "type": "text" + }, + { + "bbox": [ + 417, + 259, + 470, + 279 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\frac { { \\sigma } _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 258, + 505, + 281 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 277, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 322, + 292 + ], + "score": 1.0, + "content": "model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 322, + 277, + 367, + 292 + ], + "score": 0.93, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 278, + 388, + 292 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 388, + 278, + 434, + 291 + ], + "score": 0.92, + "content": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 278, + 505, + 292 + ], + "score": 1.0, + "content": ", then the largest", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 102, + 286, + 508, + 308 + ], + "spans": [ + { + "bbox": [ + 102, + 286, + 374, + 308 + ], + "score": 1.0, + "content": "initialization value for which the learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 374, + 292, + 397, + 304 + ], + "score": 0.91, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 286, + 486, + 308 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 486, + 291, + 501, + 304 + ], + "score": 0.89, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 286, + 508, + 308 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 302, + 239, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 239, + 315 + ], + "score": 1.0, + "content": "is bounded in the following way:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 151, + 333, + 460, + 390 + ], + "lines": [ + { + "bbox": [ + 151, + 333, + 460, + 390 + ], + "spans": [ + { + "bbox": [ + 151, + 333, + 460, + 390 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}", + "type": "interline_equation", + "image_path": "0bdfaf2010332f9ed8389cd05c9719d7b46af7a10d495ec58e608b955e39d0cf.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 151, + 333, + 460, + 352.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 151, + 352.0, + 460, + 371.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 151, + 371.0, + 460, + 390.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 399, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 297, + 413 + ], + "score": 1.0, + "content": "Proof sketch (the full proof is given in appendix", + "type": "text" + }, + { + "bbox": [ + 298, + 401, + 305, + 410 + ], + "score": 0.47, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "). Rewriting the separable differential equation in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 410, + 365, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 225, + 425 + ], + "score": 1.0, + "content": "(4) to calculate the time until", + "type": "text" + }, + { + "bbox": [ + 225, + 411, + 273, + 423 + ], + "score": 0.93, + "content": "\\sigma ( t ) = \\alpha \\sigma ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 410, + 365, + 425 + ], + "score": 1.0, + "content": ", we get the following:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 435, + 370, + 466 + ], + "lines": [ + { + "bbox": [ + 240, + 435, + 370, + 466 + ], + "spans": [ + { + "bbox": [ + 240, + 435, + 370, + 466 + ], + "score": 0.94, + "content": "t _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "010182bab0f2f4e2fdb5ab3ced1eb9a629796460db21e636764dbcfa65bfd632.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 435, + 370, + 450.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 240, + 450.5, + 370, + 466.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 473, + 493, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 495, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 375, + 488 + ], + "score": 1.0, + "content": "The condition for incremental learning is then the requirement that", + "type": "text" + }, + { + "bbox": [ + 375, + 474, + 440, + 486 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 471, + 495, + 488 + ], + "score": 1.0, + "content": ", resulting in:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 497, + 403, + 530 + ], + "lines": [ + { + "bbox": [ + 207, + 497, + 403, + 530 + ], + "spans": [ + { + "bbox": [ + 207, + 497, + 403, + 530 + ], + "score": 0.93, + "content": "\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "45bbd2b211f6960978f4e641ef8b910b12b6c04083081c3da66ad1f666d598dd.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 497, + 403, + 513.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 207, + 513.5, + 403, + 530.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 441, + 551 + ], + "score": 1.0, + "content": "We then relax/restrict the above condition to get a necessary/sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 441, + 540, + 452, + 549 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 537, + 506, + 551 + ], + "score": 1.0, + "content": ", leading to a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 103, + 544, + 235, + 565 + ], + "spans": [ + { + "bbox": [ + 103, + 544, + 213, + 565 + ], + "score": 1.0, + "content": "lower and upper bound on", + "type": "text" + }, + { + "bbox": [ + 214, + 548, + 228, + 561 + ], + "score": 0.91, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 544, + 235, + 565 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 587, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 407, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 394, + 602 + ], + "score": 1.0, + "content": "Note that the value determining the condition for incremental learning is", + "type": "text" + }, + { + "bbox": [ + 395, + 587, + 407, + 606 + ], + "score": 0.78, + "content": "\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "same order of magnitude, then their ratio will be close to 1 and we will need a small initialization", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "to obtain incremental learning. The dependence on the ratio changes with depth, and is exponential", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 624, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 121, + 639 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 626, + 153, + 636 + ], + "score": 0.89, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 624, + 506, + 639 + ], + "score": 1.0, + "content": ". This means that incremental learning, while possible for shallow models, is difficult", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "to see in practice. This result explains why changing the initialization scale in figure 2 changes the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 648, + 459, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 172, + 660 + ], + "score": 1.0, + "content": "dynamics of the", + "type": "text" + }, + { + "bbox": [ + 173, + 648, + 201, + 659 + ], + "score": 0.9, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 648, + 385, + 660 + ], + "score": 1.0, + "content": "models, while not changing the dynamics for", + "type": "text" + }, + { + "bbox": [ + 385, + 648, + 413, + 658 + ], + "score": 0.89, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 648, + 459, + 660 + ], + "score": 1.0, + "content": "noticeably.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 105, + 664, + 504, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "The next theorem extends part of our analysis to gradient descent, a more realistic setting than the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 676, + 280, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 280, + 688 + ], + "score": 1.0, + "content": "infinitesimal learning rate of gradient flow:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 231, + 712 + ], + "score": 1.0, + "content": "Theorem 3. Given two values", + "type": "text" + }, + { + "bbox": [ + 232, + 695, + 256, + 705 + ], + "score": 0.87, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 689, + 454, + 712 + ], + "score": 1.0, + "content": "of a depth-2 toy linear model as in (1), such that", + "type": "text" + }, + { + "bbox": [ + 455, + 690, + 504, + 709 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 706, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 342, + 723 + ], + "score": 1.0, + "content": "and the model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 342, + 708, + 383, + 722 + ], + "score": 0.92, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 706, + 401, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 402, + 708, + 443, + 722 + ], + "score": 0.93, + "content": "f \\in ( \\frac { 3 } { 4 } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 706, + 506, + 723 + ], + "score": 1.0, + "content": ", and assuming", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 719, + 507, + 735 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 152, + 734 + ], + "score": 0.92, + "content": "\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 719, + 451, + 735 + ], + "score": 1.0, + "content": ", and assuming we optimize with gradient descent with a learning rate", + "type": "text" + }, + { + "bbox": [ + 452, + 720, + 488, + 735 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 719, + 507, + 735 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 495, + 566, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 496, + 568, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 496, + 568, + 504, + 576 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 120 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 236, + 96 + ], + "score": 1.0, + "content": "Definition 1. Given two values", + "type": "text" + }, + { + "bbox": [ + 236, + 84, + 261, + 95 + ], + "score": 0.89, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 81, + 302, + 96 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 302, + 83, + 360, + 96 + ], + "score": 0.93, + "content": "\\sigma _ { i } ^ { * } > \\sigma _ { j } ^ { * } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 81, + 469, + 96 + ], + "score": 1.0, + "content": "and both are initialized as", + "type": "text" + }, + { + "bbox": [ + 470, + 82, + 505, + 95 + ], + "score": 0.91, + "content": "\\sigma _ { i } ( 0 ) =", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 95, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 107, + 95, + 183, + 109 + ], + "score": 0.91, + "content": "\\sigma _ { j } ( 0 ) = \\sigma _ { 0 } < \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 95, + 280, + 109 + ], + "score": 1.0, + "content": ", and given two scalars", + "type": "text" + }, + { + "bbox": [ + 280, + 95, + 324, + 109 + ], + "score": 0.92, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 95, + 344, + 109 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 345, + 95, + 389, + 109 + ], + "score": 0.93, + "content": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 95, + 505, + 109 + ], + "score": 1.0, + "content": ", we call the learning of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 108, + 323, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 135, + 121 + ], + "score": 1.0, + "content": "values", + "type": "text" + }, + { + "bbox": [ + 135, + 109, + 158, + 120 + ], + "score": 0.9, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 108, + 323, + 121 + ], + "score": 1.0, + "content": "-incremental if there exists a t for which:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 81, + 505, + 121 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 135, + 363, + 150 + ], + "lines": [ + { + "bbox": [ + 247, + 135, + 363, + 150 + ], + "spans": [ + { + "bbox": [ + 247, + 135, + 363, + 150 + ], + "score": 0.91, + "content": "\\sigma _ { j } ( t ) \\leq s \\sigma _ { j } ^ { * } < f \\sigma _ { i } ^ { * } \\leq \\sigma _ { i } ( t )", + "type": "interline_equation", + "image_path": "756427e823ee307e8aa880968cbf8646de56e5172a9941219f2ea7f263d8dcf7.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 247, + 135, + 363, + 150 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 156, + 505, + 212 + ], + "lines": [ + { + "bbox": [ + 105, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 428, + 169 + ], + "score": 1.0, + "content": "In words, two values have distinct learning phases if the first almost converges", + "type": "text" + }, + { + "bbox": [ + 428, + 157, + 459, + 168 + ], + "score": 0.86, + "content": "( f \\approx 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "before the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 167, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 213, + 180 + ], + "score": 1.0, + "content": "second changes by much", + "type": "text" + }, + { + "bbox": [ + 214, + 168, + 244, + 178 + ], + "score": 0.82, + "content": "s \\ll 1 \\ r _ { \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 167, + 325, + 180 + ], + "score": 1.0, + "content": "). Note that for any", + "type": "text" + }, + { + "bbox": [ + 326, + 168, + 359, + 180 + ], + "score": 0.58, + "content": "N , \\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 167, + 506, + 180 + ], + "score": 1.0, + "content": "is monotonically increasing and so", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 178, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 178, + 127, + 191 + ], + "score": 1.0, + "content": "once", + "type": "text" + }, + { + "bbox": [ + 128, + 178, + 178, + 192 + ], + "score": 0.92, + "content": "\\sigma _ { j } ( t ) = s \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 178, + 506, + 191 + ], + "score": 1.0, + "content": ", it will not decrease to allow further incremental learning. Given this definition of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 188, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 203 + ], + "score": 1.0, + "content": "incremental learning, we turn to study the conditions that facilitate incremental learning in our toy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 200, + 137, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 137, + 213 + ], + "score": 1.0, + "content": "model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 104, + 156, + 506, + 213 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "Our main result is a dynamical depth separation result, showing that incremental learning is depen-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 228, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 138, + 249 + ], + "score": 1.0, + "content": "dent on", + "type": "text" + }, + { + "bbox": [ + 139, + 228, + 150, + 247 + ], + "score": 0.87, + "content": "\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 229, + 308, + 249 + ], + "score": 1.0, + "content": "in different ways for different values of", + "type": "text" + }, + { + "bbox": [ + 308, + 231, + 318, + 241 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 229, + 505, + 249 + ], + "score": 1.0, + "content": ". The largest difference in dependence happens", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 245, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 142, + 258 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 142, + 246, + 171, + 255 + ], + "score": 0.86, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 245, + 189, + 258 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 246, + 217, + 256 + ], + "score": 0.9, + "content": "N = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 245, + 478, + 258 + ], + "score": 1.0, + "content": ", where the dependence changes from exponential to polynomial:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 216, + 505, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 506, + 315 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 234, + 281 + ], + "score": 1.0, + "content": "Theorem 2. Given two values", + "type": "text" + }, + { + "bbox": [ + 234, + 264, + 258, + 275 + ], + "score": 0.89, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 258, + 417, + 281 + ], + "score": 1.0, + "content": "of a toy linear model as in (1), where", + "type": "text" + }, + { + "bbox": [ + 417, + 259, + 470, + 279 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\frac { { \\sigma } _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 258, + 505, + 281 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 277, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 322, + 292 + ], + "score": 1.0, + "content": "model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 322, + 277, + 367, + 292 + ], + "score": 0.93, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 278, + 388, + 292 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 388, + 278, + 434, + 291 + ], + "score": 0.92, + "content": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 278, + 505, + 292 + ], + "score": 1.0, + "content": ", then the largest", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 102, + 286, + 508, + 308 + ], + "spans": [ + { + "bbox": [ + 102, + 286, + 374, + 308 + ], + "score": 1.0, + "content": "initialization value for which the learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 374, + 292, + 397, + 304 + ], + "score": 0.91, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 286, + 486, + 308 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 486, + 291, + 501, + 304 + ], + "score": 0.89, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 286, + 508, + 308 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 302, + 239, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 239, + 315 + ], + "score": 1.0, + "content": "is bounded in the following way:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 102, + 258, + 508, + 315 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 151, + 333, + 460, + 390 + ], + "lines": [ + { + "bbox": [ + 151, + 333, + 460, + 390 + ], + "spans": [ + { + "bbox": [ + 151, + 333, + 460, + 390 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}", + "type": "interline_equation", + "image_path": "0bdfaf2010332f9ed8389cd05c9719d7b46af7a10d495ec58e608b955e39d0cf.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 151, + 333, + 460, + 352.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 151, + 352.0, + 460, + 371.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 151, + 371.0, + 460, + 390.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 399, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 297, + 413 + ], + "score": 1.0, + "content": "Proof sketch (the full proof is given in appendix", + "type": "text" + }, + { + "bbox": [ + 298, + 401, + 305, + 410 + ], + "score": 0.47, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "). Rewriting the separable differential equation in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 410, + 365, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 225, + 425 + ], + "score": 1.0, + "content": "(4) to calculate the time until", + "type": "text" + }, + { + "bbox": [ + 225, + 411, + 273, + 423 + ], + "score": 0.93, + "content": "\\sigma ( t ) = \\alpha \\sigma ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 410, + 365, + 425 + ], + "score": 1.0, + "content": ", we get the following:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 399, + 506, + 425 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 435, + 370, + 466 + ], + "lines": [ + { + "bbox": [ + 240, + 435, + 370, + 466 + ], + "spans": [ + { + "bbox": [ + 240, + 435, + 370, + 466 + ], + "score": 0.94, + "content": "t _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "010182bab0f2f4e2fdb5ab3ced1eb9a629796460db21e636764dbcfa65bfd632.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 435, + 370, + 450.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 240, + 450.5, + 370, + 466.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 473, + 493, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 495, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 375, + 488 + ], + "score": 1.0, + "content": "The condition for incremental learning is then the requirement that", + "type": "text" + }, + { + "bbox": [ + 375, + 474, + 440, + 486 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 471, + 495, + 488 + ], + "score": 1.0, + "content": ", resulting in:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 471, + 495, + 488 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 497, + 403, + 530 + ], + "lines": [ + { + "bbox": [ + 207, + 497, + 403, + 530 + ], + "spans": [ + { + "bbox": [ + 207, + 497, + 403, + 530 + ], + "score": 0.93, + "content": "\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "45bbd2b211f6960978f4e641ef8b910b12b6c04083081c3da66ad1f666d598dd.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 497, + 403, + 513.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 207, + 513.5, + 403, + 530.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 441, + 551 + ], + "score": 1.0, + "content": "We then relax/restrict the above condition to get a necessary/sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 441, + 540, + 452, + 549 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 537, + 506, + 551 + ], + "score": 1.0, + "content": ", leading to a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 103, + 544, + 235, + 565 + ], + "spans": [ + { + "bbox": [ + 103, + 544, + 213, + 565 + ], + "score": 1.0, + "content": "lower and upper bound on", + "type": "text" + }, + { + "bbox": [ + 214, + 548, + 228, + 561 + ], + "score": 0.91, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 544, + 235, + 565 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 103, + 537, + 506, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 587, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 407, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 394, + 602 + ], + "score": 1.0, + "content": "Note that the value determining the condition for incremental learning is", + "type": "text" + }, + { + "bbox": [ + 395, + 587, + 407, + 606 + ], + "score": 0.78, + "content": "\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "same order of magnitude, then their ratio will be close to 1 and we will need a small initialization", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "to obtain incremental learning. The dependence on the ratio changes with depth, and is exponential", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 624, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 121, + 639 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 626, + 153, + 636 + ], + "score": 0.89, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 624, + 506, + 639 + ], + "score": 1.0, + "content": ". This means that incremental learning, while possible for shallow models, is difficult", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "to see in practice. This result explains why changing the initialization scale in figure 2 changes the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 648, + 459, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 172, + 660 + ], + "score": 1.0, + "content": "dynamics of the", + "type": "text" + }, + { + "bbox": [ + 173, + 648, + 201, + 659 + ], + "score": 0.9, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 648, + 385, + 660 + ], + "score": 1.0, + "content": "models, while not changing the dynamics for", + "type": "text" + }, + { + "bbox": [ + 385, + 648, + 413, + 658 + ], + "score": 0.89, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 648, + 459, + 660 + ], + "score": 1.0, + "content": "noticeably.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 587, + 506, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 664, + 504, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "The next theorem extends part of our analysis to gradient descent, a more realistic setting than the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 676, + 280, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 280, + 688 + ], + "score": 1.0, + "content": "infinitesimal learning rate of gradient flow:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 664, + 505, + 688 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 231, + 712 + ], + "score": 1.0, + "content": "Theorem 3. Given two values", + "type": "text" + }, + { + "bbox": [ + 232, + 695, + 256, + 705 + ], + "score": 0.87, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 689, + 454, + 712 + ], + "score": 1.0, + "content": "of a depth-2 toy linear model as in (1), such that", + "type": "text" + }, + { + "bbox": [ + 455, + 690, + 504, + 709 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 706, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 342, + 723 + ], + "score": 1.0, + "content": "and the model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 342, + 708, + 383, + 722 + ], + "score": 0.92, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 706, + 401, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 402, + 708, + 443, + 722 + ], + "score": 0.93, + "content": "f \\in ( \\frac { 3 } { 4 } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 706, + 506, + 723 + ], + "score": 1.0, + "content": ", and assuming", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 719, + 507, + 735 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 152, + 734 + ], + "score": 0.92, + "content": "\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 719, + 451, + 735 + ], + "score": 1.0, + "content": ", and assuming we optimize with gradient descent with a learning rate", + "type": "text" + }, + { + "bbox": [ + 452, + 720, + 488, + 735 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 719, + 507, + 735 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 172, + 95 + ], + "score": 0.91, + "content": "c < 2 ( { \\sqrt { 2 } } - 1 )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 172, + 81, + 192, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 192, + 83, + 204, + 95 + ], + "score": 0.88, + "content": "\\sigma _ { 1 } ^ { * }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 204, + 81, + 287, + 95 + ], + "score": 1.0, + "content": "the largest value of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 287, + 83, + 299, + 93 + ], + "score": 0.85, + "content": "\\sigma ^ { * }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 299, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ", then the largest initialization value for which the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 91, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 244, + 108 + ], + "score": 1.0, + "content": "learning phases of the values are", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 245, + 94, + 268, + 106 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 268, + 91, + 357, + 108 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 358, + 93, + 372, + 105 + ], + "score": 0.9, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 373, + 91, + 506, + 108 + ], + "score": 1.0, + "content": ", is lower and upper bounded in", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 183, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 183, + 118 + ], + "score": 1.0, + "content": "the following way:", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 689, + 507, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 107, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 172, + 95 + ], + "score": 0.91, + "content": "c < 2 ( { \\sqrt { 2 } } - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 81, + 192, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 192, + 83, + 204, + 95 + ], + "score": 0.88, + "content": "\\sigma _ { 1 } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 81, + 287, + 95 + ], + "score": 1.0, + "content": "the largest value of", + "type": "text" + }, + { + "bbox": [ + 287, + 83, + 299, + 93 + ], + "score": 0.85, + "content": "\\sigma ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ", then the largest initialization value for which the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 91, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 244, + 108 + ], + "score": 1.0, + "content": "learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 245, + 94, + 268, + 106 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 91, + 357, + 108 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 358, + 93, + 372, + 105 + ], + "score": 0.9, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 91, + 506, + 108 + ], + "score": 1.0, + "content": ", is lower and upper bounded in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 183, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 183, + 118 + ], + "score": 1.0, + "content": "the following way:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 130, + 443, + 158 + ], + "lines": [ + { + "bbox": [ + 166, + 130, + 443, + 158 + ], + "spans": [ + { + "bbox": [ + 166, + 130, + 443, + 158 + ], + "score": 0.92, + "content": "\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }", + "type": "interline_equation", + "image_path": "d67099a203e064b3678ebe86bcb9a99f2d540931c92f05481f5369b04fc29419.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 166, + 130, + 443, + 158 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 233, + 179 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 234, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 135, + 181 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 168, + 143, + 177 + ], + "score": 0.73, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 165, + 162, + 181 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 168, + 171, + 177 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 165, + 234, + 181 + ], + "score": 1.0, + "content": "are defined as:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 194, + 443, + 233 + ], + "lines": [ + { + "bbox": [ + 167, + 194, + 443, + 233 + ], + "spans": [ + { + "bbox": [ + 167, + 194, + 443, + 233 + ], + "score": 0.94, + "content": "A = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "d5e177cebb8d5bd48f03d19faf96fcbba6231bbd13232b8eac6d06dc194bf4e1.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 167, + 194, + 443, + 207.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 167, + 207.0, + 443, + 220.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 167, + 220.0, + 443, + 233.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 242, + 244, + 254 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 245, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 245, + 256 + ], + "score": 1.0, + "content": "We defer the proof to appendix B.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "Note that this result, while less elegant than the bounds of the gradient flow analysis, is similar in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 271, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 162, + 282 + ], + "score": 1.0, + "content": "nature. Both", + "type": "text" + }, + { + "bbox": [ + 162, + 271, + 171, + 280 + ], + "score": 0.74, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 271, + 190, + 282 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 191, + 271, + 200, + 280 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 271, + 249, + 282 + ], + "score": 1.0, + "content": "simplify to", + "type": "text" + }, + { + "bbox": [ + 249, + 273, + 256, + 280 + ], + "score": 0.71, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 271, + 474, + 282 + ], + "score": 1.0, + "content": "when we take their first order approximation around", + "type": "text" + }, + { + "bbox": [ + 474, + 271, + 501, + 281 + ], + "score": 0.89, + "content": "c = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 271, + 505, + 282 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 282, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 344, + 294 + ], + "score": 1.0, + "content": "giving us similar bounds and showing that the condition on", + "type": "text" + }, + { + "bbox": [ + 345, + 283, + 356, + 293 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 282, + 371, + 294 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 372, + 282, + 401, + 291 + ], + "score": 0.92, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 282, + 506, + 294 + ], + "score": 1.0, + "content": "is exponential in gradient", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 171, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 171, + 303 + ], + "score": 1.0, + "content": "descent as well.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 105, + 309, + 504, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "While similar gradient descent results are harder to obtain for deeper models, we discuss the general", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 320, + 359, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 359, + 333 + ], + "score": 1.0, + "content": "effect of depth on the gradient decent dynamics in appendix C.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 106, + 349, + 369, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 369, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 369, + 363 + ], + "score": 1.0, + "content": "4 INCREMENTAL LEARNING IN LARGER MODELS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 374, + 506, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "score": 1.0, + "content": "So far, we have only shown interesting properties of incremental learning caused by depth for a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "toy model. In this section, we will relate several deep models to our toy model and show how", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 397, + 349, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 349, + 410 + ], + "score": 1.0, + "content": "incremental learning presents itself in larger models as well.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 107, + 423, + 208, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 209, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 209, + 436 + ], + "score": 1.0, + "content": "4.1 MATRIX SENSING", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 504, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 483, + 457 + ], + "score": 1.0, + "content": "The task of matrix sensing is a generalization of matrix completion, where our input space is", + "type": "text" + }, + { + "bbox": [ + 483, + 445, + 506, + 456 + ], + "score": 0.83, + "content": "\\mathcal { X } =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 453, + 327, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 130, + 465 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 453, + 236, + 468 + ], + "score": 1.0, + "content": "and our model is a matrix", + "type": "text" + }, + { + "bbox": [ + 236, + 455, + 282, + 466 + ], + "score": 0.92, + "content": "\\mathbf { \\bar { \\boldsymbol { W } } } \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 453, + 327, + 468 + ], + "score": 1.0, + "content": ", such that:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 483, + 370, + 499 + ], + "lines": [ + { + "bbox": [ + 240, + 483, + 370, + 499 + ], + "spans": [ + { + "bbox": [ + 240, + 483, + 370, + 499 + ], + "score": 0.91, + "content": "f _ { W } ( A ) = \\langle W , A \\rangle = t r \\big ( W ^ { T } A \\big )", + "type": "interline_equation", + "image_path": "dfbaa3e0800fc64e0f4d9dccac3558d2f6174b550e65d3f5fb315ee4e27ce3a0.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 240, + 483, + 370, + 499 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 508, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "Following Arora et al. (2019), we introduce depth by parameterizing the model using a product of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 518, + 360, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 302, + 531 + ], + "score": 1.0, + "content": "matrices and the following initialization scheme", + "type": "text" + }, + { + "bbox": [ + 302, + 518, + 354, + 531 + ], + "score": 0.91, + "content": "( W _ { i } ^ { \\setminus } \\in \\mathbb { R } ^ { d \\times d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 518, + 360, + 531 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 553, + 385, + 585 + ], + "lines": [ + { + "bbox": [ + 226, + 553, + 385, + 585 + ], + "spans": [ + { + "bbox": [ + 226, + 553, + 385, + 585 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { W = W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } } } \\\\ { { \\forall i \\in [ N ] , ~ W _ { i } ( 0 ) = ~ \\sqrt [ N ] { \\sigma _ { 0 } } I } } \\end{array}", + "type": "interline_equation", + "image_path": "d4bcc48e64d0072b37aa4917ac74b963de477178d5b3f8c3099946938d4f034c.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 553, + 385, + 569.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 226, + 569.0, + 385, + 585.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 595, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 173, + 608 + ], + "score": 1.0, + "content": "Note that when", + "type": "text" + }, + { + "bbox": [ + 173, + 596, + 201, + 606 + ], + "score": 0.9, + "content": "d = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 594, + 505, + 608 + ], + "score": 1.0, + "content": ", the deep matrix sensing model reduces to our toy model without weight", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "sharing. We study the dynamics of the model under gradient flow over a depth-normalized squared", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 616, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 448, + 630 + ], + "score": 1.0, + "content": "loss, assuming the data is labeled by a matrix sensing model parameterized by a PSD", + "type": "text" + }, + { + "bbox": [ + 448, + 617, + 499, + 628 + ], + "score": 0.92, + "content": "W ^ { \\ast } \\in \\mathring { \\mathbb { R } } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 616, + 505, + 630 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 650, + 445, + 700 + ], + "lines": [ + { + "bbox": [ + 165, + 650, + 445, + 700 + ], + "spans": [ + { + "bbox": [ + 165, + 650, + 445, + 700 + ], + "score": 0.94, + "content": "\\begin{array} { c } { \\displaystyle \\ell _ { N } ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } ) = \\displaystyle \\frac { 1 } { 2 N } \\mathbb { E } _ { A } [ ( \\langle \\left( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\right) - W ^ { * } , A \\rangle ^ { 2 } ] } \\\\ { = \\displaystyle \\frac { 1 } { 2 N } | | \\big ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\big ) - W ^ { * } | | _ { F } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "c69af104dc366bd12c182e997c6bb6fcbb91a45e3e31d22b766831ccba4a2db6.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 165, + 650, + 445, + 666.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 165, + 666.6666666666666, + 445, + 683.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 165, + 683.3333333333333, + 445, + 699.9999999999999 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "The following theorem relates the deep matrix sensing model to our toy model, showing the two", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 253, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 253, + 733 + ], + "score": 1.0, + "content": "have the same dynamical equations:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 104, + 81, + 506, + 118 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 130, + 443, + 158 + ], + "lines": [ + { + "bbox": [ + 166, + 130, + 443, + 158 + ], + "spans": [ + { + "bbox": [ + 166, + 130, + 443, + 158 + ], + "score": 0.92, + "content": "\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }", + "type": "interline_equation", + "image_path": "d67099a203e064b3678ebe86bcb9a99f2d540931c92f05481f5369b04fc29419.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 166, + 130, + 443, + 158 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 233, + 179 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 234, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 135, + 181 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 168, + 143, + 177 + ], + "score": 0.73, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 165, + 162, + 181 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 168, + 171, + 177 + ], + "score": 0.79, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 165, + 234, + 181 + ], + "score": 1.0, + "content": "are defined as:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 165, + 234, + 181 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 194, + 443, + 233 + ], + "lines": [ + { + "bbox": [ + 167, + 194, + 443, + 233 + ], + "spans": [ + { + "bbox": [ + 167, + 194, + 443, + 233 + ], + "score": 0.94, + "content": "A = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "d5e177cebb8d5bd48f03d19faf96fcbba6231bbd13232b8eac6d06dc194bf4e1.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 167, + 194, + 443, + 207.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 167, + 207.0, + 443, + 220.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 167, + 220.0, + 443, + 233.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 242, + 244, + 254 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 245, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 245, + 256 + ], + "score": 1.0, + "content": "We defer the proof to appendix B.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 241, + 245, + 256 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "Note that this result, while less elegant than the bounds of the gradient flow analysis, is similar in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 271, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 162, + 282 + ], + "score": 1.0, + "content": "nature. Both", + "type": "text" + }, + { + "bbox": [ + 162, + 271, + 171, + 280 + ], + "score": 0.74, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 271, + 190, + 282 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 191, + 271, + 200, + 280 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 271, + 249, + 282 + ], + "score": 1.0, + "content": "simplify to", + "type": "text" + }, + { + "bbox": [ + 249, + 273, + 256, + 280 + ], + "score": 0.71, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 271, + 474, + 282 + ], + "score": 1.0, + "content": "when we take their first order approximation around", + "type": "text" + }, + { + "bbox": [ + 474, + 271, + 501, + 281 + ], + "score": 0.89, + "content": "c = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 271, + 505, + 282 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 282, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 344, + 294 + ], + "score": 1.0, + "content": "giving us similar bounds and showing that the condition on", + "type": "text" + }, + { + "bbox": [ + 345, + 283, + 356, + 293 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 282, + 371, + 294 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 372, + 282, + 401, + 291 + ], + "score": 0.92, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 282, + 506, + 294 + ], + "score": 1.0, + "content": "is exponential in gradient", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 171, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 171, + 303 + ], + "score": 1.0, + "content": "descent as well.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 259, + 506, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 309, + 504, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "While similar gradient descent results are harder to obtain for deeper models, we discuss the general", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 320, + 359, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 359, + 333 + ], + "score": 1.0, + "content": "effect of depth on the gradient decent dynamics in appendix C.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 308, + 505, + 333 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 349, + 369, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 369, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 369, + 363 + ], + "score": 1.0, + "content": "4 INCREMENTAL LEARNING IN LARGER MODELS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 374, + 506, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 388 + ], + "score": 1.0, + "content": "So far, we have only shown interesting properties of incremental learning caused by depth for a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "toy model. In this section, we will relate several deep models to our toy model and show how", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 397, + 349, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 349, + 410 + ], + "score": 1.0, + "content": "incremental learning presents itself in larger models as well.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 374, + 505, + 410 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 423, + 208, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 209, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 209, + 436 + ], + "score": 1.0, + "content": "4.1 MATRIX SENSING", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 504, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 483, + 457 + ], + "score": 1.0, + "content": "The task of matrix sensing is a generalization of matrix completion, where our input space is", + "type": "text" + }, + { + "bbox": [ + 483, + 445, + 506, + 456 + ], + "score": 0.83, + "content": "\\mathcal { X } =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 453, + 327, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 130, + 465 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 453, + 236, + 468 + ], + "score": 1.0, + "content": "and our model is a matrix", + "type": "text" + }, + { + "bbox": [ + 236, + 455, + 282, + 466 + ], + "score": 0.92, + "content": "\\mathbf { \\bar { \\boldsymbol { W } } } \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 453, + 327, + 468 + ], + "score": 1.0, + "content": ", such that:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 443, + 506, + 468 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 483, + 370, + 499 + ], + "lines": [ + { + "bbox": [ + 240, + 483, + 370, + 499 + ], + "spans": [ + { + "bbox": [ + 240, + 483, + 370, + 499 + ], + "score": 0.91, + "content": "f _ { W } ( A ) = \\langle W , A \\rangle = t r \\big ( W ^ { T } A \\big )", + "type": "interline_equation", + "image_path": "dfbaa3e0800fc64e0f4d9dccac3558d2f6174b550e65d3f5fb315ee4e27ce3a0.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 240, + 483, + 370, + 499 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 508, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "Following Arora et al. (2019), we introduce depth by parameterizing the model using a product of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 518, + 360, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 302, + 531 + ], + "score": 1.0, + "content": "matrices and the following initialization scheme", + "type": "text" + }, + { + "bbox": [ + 302, + 518, + 354, + 531 + ], + "score": 0.91, + "content": "( W _ { i } ^ { \\setminus } \\in \\mathbb { R } ^ { d \\times d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 518, + 360, + 531 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 508, + 506, + 531 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 553, + 385, + 585 + ], + "lines": [ + { + "bbox": [ + 226, + 553, + 385, + 585 + ], + "spans": [ + { + "bbox": [ + 226, + 553, + 385, + 585 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { W = W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } } } \\\\ { { \\forall i \\in [ N ] , ~ W _ { i } ( 0 ) = ~ \\sqrt [ N ] { \\sigma _ { 0 } } I } } \\end{array}", + "type": "interline_equation", + "image_path": "d4bcc48e64d0072b37aa4917ac74b963de477178d5b3f8c3099946938d4f034c.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 553, + 385, + 569.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 226, + 569.0, + 385, + 585.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 595, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 173, + 608 + ], + "score": 1.0, + "content": "Note that when", + "type": "text" + }, + { + "bbox": [ + 173, + 596, + 201, + 606 + ], + "score": 0.9, + "content": "d = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 594, + 505, + 608 + ], + "score": 1.0, + "content": ", the deep matrix sensing model reduces to our toy model without weight", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "sharing. We study the dynamics of the model under gradient flow over a depth-normalized squared", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 616, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 104, + 616, + 448, + 630 + ], + "score": 1.0, + "content": "loss, assuming the data is labeled by a matrix sensing model parameterized by a PSD", + "type": "text" + }, + { + "bbox": [ + 448, + 617, + 499, + 628 + ], + "score": 0.92, + "content": "W ^ { \\ast } \\in \\mathring { \\mathbb { R } } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 616, + 505, + 630 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 594, + 505, + 630 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 650, + 445, + 700 + ], + "lines": [ + { + "bbox": [ + 165, + 650, + 445, + 700 + ], + "spans": [ + { + "bbox": [ + 165, + 650, + 445, + 700 + ], + "score": 0.94, + "content": "\\begin{array} { c } { \\displaystyle \\ell _ { N } ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } ) = \\displaystyle \\frac { 1 } { 2 N } \\mathbb { E } _ { A } [ ( \\langle \\left( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\right) - W ^ { * } , A \\rangle ^ { 2 } ] } \\\\ { = \\displaystyle \\frac { 1 } { 2 N } | | \\big ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\big ) - W ^ { * } | | _ { F } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "c69af104dc366bd12c182e997c6bb6fcbb91a45e3e31d22b766831ccba4a2db6.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 165, + 650, + 445, + 666.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 165, + 666.6666666666666, + 445, + 683.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 165, + 683.3333333333333, + 445, + 699.9999999999999 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "The following theorem relates the deep matrix sensing model to our toy model, showing the two", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 253, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 253, + 733 + ], + "score": 1.0, + "content": "have the same dynamical equations:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 708, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Theorem 4. Optimizing the deep matrix sensing model described in (5) with gradient flow over", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "score": 1.0, + "content": "the depth normalized squared loss ((6)), with weights initialized as in (5) leads to the following", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 280, + 117 + ], + "score": 1.0, + "content": "dynamical equations for different values of", + "type": "text" + }, + { + "bbox": [ + 280, + 105, + 290, + 114 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 104, + 294, + 117 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 129, + 368, + 145 + ], + "lines": [ + { + "bbox": [ + 242, + 129, + 368, + 145 + ], + "spans": [ + { + "bbox": [ + 242, + 129, + 368, + 145 + ], + "score": 0.92, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "e869fb28a4b187879be9124717884221e3f7d404bdf21d63b6c6c622629dfca6.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 242, + 129, + 368, + 145 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 151, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 135, + 165 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 153, + 145, + 163 + ], + "score": 0.83, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 150, + 163, + 165 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 153, + 175, + 164 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 150, + 294, + 165 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 152, + 306, + 162 + ], + "score": 0.69, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 150, + 325, + 165 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 152, + 341, + 162 + ], + "score": 0.86, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 150, + 506, + 165 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 163, + 171, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 171, + 174 + ], + "score": 1.0, + "content": "singular vector.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 496, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 495, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 495, + 195 + ], + "score": 1.0, + "content": "The proof follows that of Saxe et al. (2013) and Gidel et al. (2019) and is deferred to appendix E.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 199, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "score": 1.0, + "content": "Theorem 4 shows us that the bias towards sparse solutions introduced by depth in the toy model is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "equivalent to the bias for low-rank solutions in the matrix sensing task. This bias was studied in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 504, + 233 + ], + "score": 1.0, + "content": "a more general setting in Arora et al. (2019), with empirical results supporting the effect of depth", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "on the obtainment of low-rank solutions under a more natural loss and initialization scheme. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "recreate and discuss these experiments and their connection to our analysis in appendix E, and an", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 254, + 471, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 471, + 266 + ], + "score": 1.0, + "content": "example of these dynamics in deep matrix sensing can also be seen in panel (a) of figure 1.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 278, + 274, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 276, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 276, + 291 + ], + "score": 1.0, + "content": "4.2 QUADRATIC NEURAL NETWORKS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "By drawing connections between quadratic networks and matrix sensing (as in Soltanolkotabi et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "(2018)), we can extend our results to these nonlinear models. We will study a simplified quadratic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 319, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 319, + 245, + 334 + ], + "score": 1.0, + "content": "network, where our input space is", + "type": "text" + }, + { + "bbox": [ + 246, + 320, + 282, + 331 + ], + "score": 0.92, + "content": "\\mathcal { X } = \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 319, + 506, + 334 + ], + "score": 1.0, + "content": "and the first layer is parameterized by a weight matrix", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 330, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 155, + 342 + ], + "score": 0.92, + "content": "W \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 330, + 505, + 345 + ], + "score": 1.0, + "content": "and followed by a quadratic activation function. The final layer will be a summation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "layer. We assume, like before, that the labeling function is a quadratic network parameterized by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 351, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 107, + 353, + 162, + 364 + ], + "score": 0.92, + "content": "\\bar { W } ^ { * } \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 351, + 506, + 368 + ], + "score": 1.0, + "content": ". Our model can be written in the following way, using the following orthogonal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 363, + 195, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 195, + 376 + ], + "score": 1.0, + "content": "initialization scheme:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 393, + 416, + 445 + ], + "lines": [ + { + "bbox": [ + 196, + 393, + 416, + 445 + ], + "spans": [ + { + "bbox": [ + 196, + 393, + 416, + 445 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { f _ { W } ( x ) = \\displaystyle \\sum _ { i = 1 } ^ { d } ( w _ { i } ^ { T } x ) ^ { 2 } = x ^ { T } W ^ { T } W x = \\langle W ^ { T } W , x x ^ { T } \\rangle } \\\\ & { W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I } \\end{array}", + "type": "interline_equation", + "image_path": "77dfced7fa36f4674dc21569f5bb29aeb6c1e39e9cc33d763eab427083736522.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 393, + 416, + 406.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 196, + 406.0, + 416, + 419.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 196, + 419.0, + 416, + 432.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 196, + 432.0, + 416, + 445.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 504, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "Immediately, we see the similarity of the quadratic network to the deep matrix sensing model with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 461, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 107, + 463, + 137, + 473 + ], + "score": 0.86, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 461, + 506, + 476 + ], + "score": 1.0, + "content": ", where the input space is made up of rank-1 matrices. However, the change in input space", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 473, + 443, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 443, + 487 + ], + "score": 1.0, + "content": "forces us to optimize over a different loss function to reproduce the same dynamics:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 104, + 487, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 351, + 500 + ], + "score": 1.0, + "content": "Definition 2. Given an input distribution over an input space", + "type": "text" + }, + { + "bbox": [ + 351, + 488, + 361, + 497 + ], + "score": 0.78, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 487, + 458, + 500 + ], + "score": 1.0, + "content": "with a labeling function", + "type": "text" + }, + { + "bbox": [ + 458, + 488, + 504, + 499 + ], + "score": 0.9, + "content": "y : \\mathcal { X } \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 497, + 384, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 176, + 512 + ], + "score": 1.0, + "content": "and a hypothesis", + "type": "text" + }, + { + "bbox": [ + 176, + 499, + 182, + 508 + ], + "score": 0.74, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 497, + 384, + 512 + ], + "score": 1.0, + "content": ", the variance loss is defined in the following way:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 523, + 423, + 546 + ], + "lines": [ + { + "bbox": [ + 188, + 523, + 423, + 546 + ], + "spans": [ + { + "bbox": [ + 188, + 523, + 423, + 546 + ], + "score": 0.93, + "content": "\\ell _ { v a r } ( h ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( y ( x ) - h ( x ) ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ y ( x ) - h ( x ) ] ^ { 2 }", + "type": "interline_equation", + "image_path": "6b5b1b247a9b5d07cfca0e85c60c21dd0c8564cee84fb9a48e85832a27dea825.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 188, + 523, + 423, + 546 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "Note that minimizing this loss function amounts to minimizing the variance of the error, while the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "squared loss minimizes the second moment of the error. We note that both loss functions have the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "same minimum for our problem, and the dynamics of the squared loss can be approximated in certain", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 583, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 599 + ], + "score": 1.0, + "content": "cases by the dynamics of the variance loss. For a complete discussion of the two losses, including", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 596, + 462, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 462, + 609 + ], + "score": 1.0, + "content": "the cases where the two losses have similar dynamics, we refer the reader to appendix F.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 623 + ], + "score": 1.0, + "content": "Theorem 5. Minimizing the quadratic network described and initialized as in (7) with gradient flow", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 621, + 429, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 429, + 634 + ], + "score": 1.0, + "content": "over the variance loss defined in (2) leads to the following dynamical equations:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 648, + 359, + 662 + ], + "lines": [ + { + "bbox": [ + 252, + 648, + 359, + 662 + ], + "spans": [ + { + "bbox": [ + 252, + 648, + 359, + 662 + ], + "score": 0.92, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "b7b36ff480a5c95d8b2916ff37e28783adb3d453b4e5b57103f77b1419768b14.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 252, + 648, + 359, + 662 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 135, + 682 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 669, + 145, + 679 + ], + "score": 0.83, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 666, + 163, + 682 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 669, + 175, + 680 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 666, + 294, + 682 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 668, + 306, + 678 + ], + "score": 0.68, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 666, + 325, + 682 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 668, + 341, + 678 + ], + "score": 0.86, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 666, + 505, + 682 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 679, + 171, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 171, + 691 + ], + "score": 1.0, + "content": "singular vector.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "We defer the proof to appendix F and note that these dynamics are the same as our depth-2 toy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "model, showing that shallow quadratic networks can exhibit incremental learning (albeit requiring a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 721, + 189, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 189, + 732 + ], + "score": 1.0, + "content": "small initialization).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Theorem 4. Optimizing the deep matrix sensing model described in (5) with gradient flow over", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 107 + ], + "score": 1.0, + "content": "the depth normalized squared loss ((6)), with weights initialized as in (5) leads to the following", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 280, + 117 + ], + "score": 1.0, + "content": "dynamical equations for different values of", + "type": "text" + }, + { + "bbox": [ + 280, + 105, + 290, + 114 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 104, + 294, + 117 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 506, + 117 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 129, + 368, + 145 + ], + "lines": [ + { + "bbox": [ + 242, + 129, + 368, + 145 + ], + "spans": [ + { + "bbox": [ + 242, + 129, + 368, + 145 + ], + "score": 0.92, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "e869fb28a4b187879be9124717884221e3f7d404bdf21d63b6c6c622629dfca6.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 242, + 129, + 368, + 145 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 151, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 135, + 165 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 153, + 145, + 163 + ], + "score": 0.83, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 150, + 163, + 165 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 153, + 175, + 164 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 150, + 294, + 165 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 152, + 306, + 162 + ], + "score": 0.69, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 150, + 325, + 165 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 152, + 341, + 162 + ], + "score": 0.86, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 150, + 506, + 165 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 163, + 171, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 171, + 174 + ], + "score": 1.0, + "content": "singular vector.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 150, + 506, + 174 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 496, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 495, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 495, + 195 + ], + "score": 1.0, + "content": "The proof follows that of Saxe et al. (2013) and Gidel et al. (2019) and is deferred to appendix E.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 181, + 495, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 199, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "score": 1.0, + "content": "Theorem 4 shows us that the bias towards sparse solutions introduced by depth in the toy model is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "equivalent to the bias for low-rank solutions in the matrix sensing task. This bias was studied in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 504, + 233 + ], + "score": 1.0, + "content": "a more general setting in Arora et al. (2019), with empirical results supporting the effect of depth", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "on the obtainment of low-rank solutions under a more natural loss and initialization scheme. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "recreate and discuss these experiments and their connection to our analysis in appendix E, and an", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 254, + 471, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 471, + 266 + ], + "score": 1.0, + "content": "example of these dynamics in deep matrix sensing can also be seen in panel (a) of figure 1.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 198, + 506, + 266 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 278, + 274, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 276, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 276, + 291 + ], + "score": 1.0, + "content": "4.2 QUADRATIC NEURAL NETWORKS", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "By drawing connections between quadratic networks and matrix sensing (as in Soltanolkotabi et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "(2018)), we can extend our results to these nonlinear models. We will study a simplified quadratic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 319, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 319, + 245, + 334 + ], + "score": 1.0, + "content": "network, where our input space is", + "type": "text" + }, + { + "bbox": [ + 246, + 320, + 282, + 331 + ], + "score": 0.92, + "content": "\\mathcal { X } = \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 319, + 506, + 334 + ], + "score": 1.0, + "content": "and the first layer is parameterized by a weight matrix", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 330, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 155, + 342 + ], + "score": 0.92, + "content": "W \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 330, + 505, + 345 + ], + "score": 1.0, + "content": "and followed by a quadratic activation function. The final layer will be a summation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "layer. We assume, like before, that the labeling function is a quadratic network parameterized by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 351, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 107, + 353, + 162, + 364 + ], + "score": 0.92, + "content": "\\bar { W } ^ { * } \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 351, + 506, + 368 + ], + "score": 1.0, + "content": ". Our model can be written in the following way, using the following orthogonal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 363, + 195, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 195, + 376 + ], + "score": 1.0, + "content": "initialization scheme:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 299, + 506, + 376 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 393, + 416, + 445 + ], + "lines": [ + { + "bbox": [ + 196, + 393, + 416, + 445 + ], + "spans": [ + { + "bbox": [ + 196, + 393, + 416, + 445 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { f _ { W } ( x ) = \\displaystyle \\sum _ { i = 1 } ^ { d } ( w _ { i } ^ { T } x ) ^ { 2 } = x ^ { T } W ^ { T } W x = \\langle W ^ { T } W , x x ^ { T } \\rangle } \\\\ & { W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I } \\end{array}", + "type": "interline_equation", + "image_path": "77dfced7fa36f4674dc21569f5bb29aeb6c1e39e9cc33d763eab427083736522.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 393, + 416, + 406.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 196, + 406.0, + 416, + 419.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 196, + 419.0, + 416, + 432.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 196, + 432.0, + 416, + 445.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 504, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "Immediately, we see the similarity of the quadratic network to the deep matrix sensing model with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 461, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 107, + 463, + 137, + 473 + ], + "score": 0.86, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 461, + 506, + 476 + ], + "score": 1.0, + "content": ", where the input space is made up of rank-1 matrices. However, the change in input space", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 473, + 443, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 443, + 487 + ], + "score": 1.0, + "content": "forces us to optimize over a different loss function to reproduce the same dynamics:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 451, + 506, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 487, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 351, + 500 + ], + "score": 1.0, + "content": "Definition 2. Given an input distribution over an input space", + "type": "text" + }, + { + "bbox": [ + 351, + 488, + 361, + 497 + ], + "score": 0.78, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 487, + 458, + 500 + ], + "score": 1.0, + "content": "with a labeling function", + "type": "text" + }, + { + "bbox": [ + 458, + 488, + 504, + 499 + ], + "score": 0.9, + "content": "y : \\mathcal { X } \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 497, + 384, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 176, + 512 + ], + "score": 1.0, + "content": "and a hypothesis", + "type": "text" + }, + { + "bbox": [ + 176, + 499, + 182, + 508 + ], + "score": 0.74, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 497, + 384, + 512 + ], + "score": 1.0, + "content": ", the variance loss is defined in the following way:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 487, + 504, + 512 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 523, + 423, + 546 + ], + "lines": [ + { + "bbox": [ + 188, + 523, + 423, + 546 + ], + "spans": [ + { + "bbox": [ + 188, + 523, + 423, + 546 + ], + "score": 0.93, + "content": "\\ell _ { v a r } ( h ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( y ( x ) - h ( x ) ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ y ( x ) - h ( x ) ] ^ { 2 }", + "type": "interline_equation", + "image_path": "6b5b1b247a9b5d07cfca0e85c60c21dd0c8564cee84fb9a48e85832a27dea825.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 188, + 523, + 423, + 546 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "Note that minimizing this loss function amounts to minimizing the variance of the error, while the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "squared loss minimizes the second moment of the error. We note that both loss functions have the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "same minimum for our problem, and the dynamics of the squared loss can be approximated in certain", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 583, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 599 + ], + "score": 1.0, + "content": "cases by the dynamics of the variance loss. For a complete discussion of the two losses, including", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 596, + 462, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 462, + 609 + ], + "score": 1.0, + "content": "the cases where the two losses have similar dynamics, we refer the reader to appendix F.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 552, + 506, + 609 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 623 + ], + "score": 1.0, + "content": "Theorem 5. Minimizing the quadratic network described and initialized as in (7) with gradient flow", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 621, + 429, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 429, + 634 + ], + "score": 1.0, + "content": "over the variance loss defined in (2) leads to the following dynamical equations:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 609, + 505, + 634 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 648, + 359, + 662 + ], + "lines": [ + { + "bbox": [ + 252, + 648, + 359, + 662 + ], + "spans": [ + { + "bbox": [ + 252, + 648, + 359, + 662 + ], + "score": 0.92, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "b7b36ff480a5c95d8b2916ff37e28783adb3d453b4e5b57103f77b1419768b14.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 252, + 648, + 359, + 662 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 135, + 682 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 669, + 145, + 679 + ], + "score": 0.83, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 666, + 163, + 682 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 669, + 175, + 680 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 666, + 294, + 682 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 668, + 306, + 678 + ], + "score": 0.68, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 666, + 325, + 682 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 668, + 341, + 678 + ], + "score": 0.86, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 666, + 505, + 682 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 679, + 171, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 171, + 691 + ], + "score": 1.0, + "content": "singular vector.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 666, + 505, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "We defer the proof to appendix F and note that these dynamics are the same as our depth-2 toy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "model, showing that shallow quadratic networks can exhibit incremental learning (albeit requiring a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 721, + 189, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 189, + 732 + ], + "score": 1.0, + "content": "small initialization).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 697, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 345, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 347, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 347, + 95 + ], + "score": 1.0, + "content": "4.3 DIAGONAL/CONVOLUTIONAL LINEAR NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 103, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 117 + ], + "score": 1.0, + "content": "While incremental learning has been described for deep linear networks in the past, it has been", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "restricted to regression tasks. Here, we illustrate how incremental learning presents itself in binary", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 418, + 138 + ], + "score": 1.0, + "content": "classification, where implicit bias results have so far focused on convergence at", + "type": "text" + }, + { + "bbox": [ + 419, + 127, + 449, + 136 + ], + "score": 0.9, + "content": "t \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "(Soudry et al.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "(2018), Nacson et al. (2018), Ji & Telgarsky (2019)). Deep linear networks with diagonal weight", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 148, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 393, + 160 + ], + "score": 1.0, + "content": "matrices have been shown to be biased towards sparse solutions when", + "type": "text" + }, + { + "bbox": [ + 393, + 148, + 424, + 158 + ], + "score": 0.9, + "content": "N > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "in Gunasekar et al.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 333, + 172 + ], + "score": 1.0, + "content": "(2018), and biased towards the max-margin solution for", + "type": "text" + }, + { + "bbox": [ + 334, + 159, + 363, + 169 + ], + "score": 0.9, + "content": "N = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 158, + 505, + 172 + ], + "score": 1.0, + "content": ". Instead of analyzing convergence", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 170, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 117, + 182 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 118, + 171, + 153, + 180 + ], + "score": 0.91, + "content": "t \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 170, + 506, + 182 + ], + "score": 1.0, + "content": ", we intend to show that the model favors sparse solutions for the entire duration of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 403, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 403, + 194 + ], + "score": 1.0, + "content": "optimization, and that this is due to the dynamics of incremental learning.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 504, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 504, + 211 + ], + "score": 1.0, + "content": "Our theoretical illustration will use our toy model as in (1) (initialized as in (3)) as a special weight-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "score": 1.0, + "content": "shared case of deep networks with diagonal weight matrices, and we will then show empirical results", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 234 + ], + "score": 1.0, + "content": "for the more general setting. We analyze the optimization dynamics of this model over a separable", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 103, + 227, + 507, + 248 + ], + "spans": [ + { + "bbox": [ + 103, + 227, + 137, + 248 + ], + "score": 1.0, + "content": "dataset", + "type": "text" + }, + { + "bbox": [ + 138, + 231, + 183, + 244 + ], + "score": 0.94, + "content": "\\{ x _ { i } , y _ { i } \\} _ { i = 1 } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 227, + 212, + 248 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 212, + 231, + 258, + 244 + ], + "score": 0.93, + "content": "y _ { i } \\in \\{ \\pm 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 227, + 380, + 248 + ], + "score": 1.0, + "content": ". We use the exponential loss", + "type": "text" + }, + { + "bbox": [ + 380, + 230, + 473, + 244 + ], + "score": 0.91, + "content": "( \\ell ( f ( x ) , y ) = e ^ { - y f ( x ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 227, + 507, + 248 + ], + "score": 1.0, + "content": "for the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 414, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 414, + 255 + ], + "score": 1.0, + "content": "theoretical illustration and experiment on the exponential and logistic losses.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 105, + 259, + 461, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 462, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 280, + 273 + ], + "score": 1.0, + "content": "Computing the gradient for the model over", + "type": "text" + }, + { + "bbox": [ + 280, + 262, + 288, + 270 + ], + "score": 0.79, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 259, + 416, + 273 + ], + "score": 1.0, + "content": ", the gradient flow dynamics for", + "type": "text" + }, + { + "bbox": [ + 417, + 262, + 424, + 270 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 259, + 462, + 273 + ], + "score": 1.0, + "content": "become:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 287, + 372, + 321 + ], + "lines": [ + { + "bbox": [ + 238, + 287, + 372, + 321 + ], + "spans": [ + { + "bbox": [ + 238, + 287, + 372, + 321 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { i } = \\frac { N ^ { 2 } } { m } \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } \\sum _ { j = 1 } ^ { m } e ^ { - y _ { j } \\langle \\beta , x _ { j } \\rangle } x _ { j }", + "type": "interline_equation", + "image_path": "0bc426bdb9a32f2dac9a10eceb8c5499a734113b8de64f6cb7f5945ec02450ed.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 238, + 287, + 372, + 304.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 238, + 304.0, + 372, + 321.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 332, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 346, + 345 + ], + "score": 1.0, + "content": "We see the same dynamical attenuation of small values of", + "type": "text" + }, + { + "bbox": [ + 347, + 335, + 354, + 343 + ], + "score": 0.79, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "that is seen in the regression model,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 340, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 240, + 364 + ], + "score": 1.0, + "content": "caused by the multiplication by", + "type": "text" + }, + { + "bbox": [ + 244, + 340, + 275, + 357 + ], + "score": 1.0, + "content": "σ2− 2N", + "type": "text" + }, + { + "bbox": [ + 271, + 345, + 506, + 361 + ], + "score": 1.0, + "content": "From this, we can expect the same type of incremental", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 358, + 504, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 228, + 371 + ], + "score": 1.0, + "content": "learning to occur - weights of", + "type": "text" + }, + { + "bbox": [ + 228, + 361, + 236, + 369 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 358, + 504, + 371 + ], + "score": 1.0, + "content": "will be learned incrementally until the dataset can be separated by", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 198, + 381 + ], + "score": 1.0, + "content": "the current support of", + "type": "text" + }, + { + "bbox": [ + 199, + 371, + 206, + 379 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 370, + 505, + 381 + ], + "score": 1.0, + "content": ". Then, the dynamics strengthen the growth of the current support while", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "score": 1.0, + "content": "relatively attenuating that of the other values. Since the data is separated, increasing the values of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "the current support reduces the loss and the magnitude of subsequent gradients, and so we should", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 402, + 449, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 449, + 415 + ], + "score": 1.0, + "content": "expect the support to remain the same and the model to converge to a sparse solution.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "Granted, the above description is just intuition, but panel (c) of figure 1 shows how it is born out", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "in practice (similar results are obtained for the logistic loss). In appendix G we further explore", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "score": 1.0, + "content": "this model, showing deeper networks have a stronger bias for sparsity. We also observe that the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "initialization scale plays a similar role as before - deep models are less biased towards sparsity when", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 463, + 153, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 118, + 474 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 463, + 153, + 477 + ], + "score": 1.0, + "content": "is large.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "In their work, Gunasekar et al. (2018) show an equivalence between the diagonal network and the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 492, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 503 + ], + "score": 1.0, + "content": "circular-convolutional network in the frequency domain. According to their results, we should ex-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "pect to see the same sparsity-bias of diagonal networks in convolutional networks, when looking at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 210, + 525 + ], + "score": 1.0, + "content": "the Fourier coefficients of", + "type": "text" + }, + { + "bbox": [ + 210, + 514, + 218, + 523 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 513, + 505, + 525 + ], + "score": 1.0, + "content": ". An example of this can be seen in panel (d) of figure 1, and we refer the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 522, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 104, + 522, + 182, + 538 + ], + "score": 1.0, + "content": "reader to appendix", + "type": "text" + }, + { + "bbox": [ + 183, + 524, + 192, + 534 + ], + "score": 0.27, + "content": "\\mathbf { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 522, + 506, + 538 + ], + "score": 1.0, + "content": "for a full discussion of their convolutional model and it’s incremental learning", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 534, + 150, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 150, + 549 + ], + "score": 1.0, + "content": "dynamics.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 564, + 195, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 197, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 197, + 581 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 504, + 625 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "Gradient-based optimization for deep linear models has an implicit bias towards simple (sparse)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "score": 1.0, + "content": "solutions, caused by an incremental search strategy over the hypothesis space. Deeper models have", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 613, + 488, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 488, + 626 + ], + "score": 1.0, + "content": "a stronger tendency for incremental learning, exhibiting it in more realistic initialization scales.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 630, + 504, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 629, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 505, + 643 + ], + "score": 1.0, + "content": "This dynamical phenomenon exists for the entire optimization process for regression as well as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "classification tasks, and for many types of models - diagonal networks, convolutional networks,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 653, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 664 + ], + "score": 1.0, + "content": "matrix completion and even the nonlinear quadratic network. We believe this kind of dynamical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 663, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 506, + 675 + ], + "score": 1.0, + "content": "analysis may be able to shed light on the generalization of deeper nonlinear neural networks as well,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 675, + 399, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 399, + 686 + ], + "score": 1.0, + "content": "with shallow quadratic networks being only a first step towards that goal.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "title", + "bbox": [ + 108, + 701, + 200, + 711 + ], + "lines": [ + { + "bbox": [ + 107, + 701, + 200, + 712 + ], + "spans": [ + { + "bbox": [ + 107, + 701, + 200, + 712 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 436, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 438, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 438, + 735 + ], + "score": 1.0, + "content": "This research is supported by the European Research Council (TheoryDL project).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 345, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 347, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 347, + 95 + ], + "score": 1.0, + "content": "4.3 DIAGONAL/CONVOLUTIONAL LINEAR NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 104, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 103, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 117 + ], + "score": 1.0, + "content": "While incremental learning has been described for deep linear networks in the past, it has been", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "restricted to regression tasks. Here, we illustrate how incremental learning presents itself in binary", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 418, + 138 + ], + "score": 1.0, + "content": "classification, where implicit bias results have so far focused on convergence at", + "type": "text" + }, + { + "bbox": [ + 419, + 127, + 449, + 136 + ], + "score": 0.9, + "content": "t \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "(Soudry et al.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 149 + ], + "score": 1.0, + "content": "(2018), Nacson et al. (2018), Ji & Telgarsky (2019)). Deep linear networks with diagonal weight", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 148, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 393, + 160 + ], + "score": 1.0, + "content": "matrices have been shown to be biased towards sparse solutions when", + "type": "text" + }, + { + "bbox": [ + 393, + 148, + 424, + 158 + ], + "score": 0.9, + "content": "N > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "in Gunasekar et al.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 333, + 172 + ], + "score": 1.0, + "content": "(2018), and biased towards the max-margin solution for", + "type": "text" + }, + { + "bbox": [ + 334, + 159, + 363, + 169 + ], + "score": 0.9, + "content": "N = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 158, + 505, + 172 + ], + "score": 1.0, + "content": ". Instead of analyzing convergence", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 170, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 117, + 182 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 118, + 171, + 153, + 180 + ], + "score": 0.91, + "content": "t \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 170, + 506, + 182 + ], + "score": 1.0, + "content": ", we intend to show that the model favors sparse solutions for the entire duration of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 403, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 403, + 194 + ], + "score": 1.0, + "content": "optimization, and that this is due to the dynamics of incremental learning.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5, + "bbox_fs": [ + 104, + 103, + 506, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 504, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 504, + 211 + ], + "score": 1.0, + "content": "Our theoretical illustration will use our toy model as in (1) (initialized as in (3)) as a special weight-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 504, + 221 + ], + "score": 1.0, + "content": "shared case of deep networks with diagonal weight matrices, and we will then show empirical results", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 234 + ], + "score": 1.0, + "content": "for the more general setting. We analyze the optimization dynamics of this model over a separable", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 103, + 227, + 507, + 248 + ], + "spans": [ + { + "bbox": [ + 103, + 227, + 137, + 248 + ], + "score": 1.0, + "content": "dataset", + "type": "text" + }, + { + "bbox": [ + 138, + 231, + 183, + 244 + ], + "score": 0.94, + "content": "\\{ x _ { i } , y _ { i } \\} _ { i = 1 } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 227, + 212, + 248 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 212, + 231, + 258, + 244 + ], + "score": 0.93, + "content": "y _ { i } \\in \\{ \\pm 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 227, + 380, + 248 + ], + "score": 1.0, + "content": ". We use the exponential loss", + "type": "text" + }, + { + "bbox": [ + 380, + 230, + 473, + 244 + ], + "score": 0.91, + "content": "( \\ell ( f ( x ) , y ) = e ^ { - y f ( x ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 227, + 507, + 248 + ], + "score": 1.0, + "content": "for the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 414, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 414, + 255 + ], + "score": 1.0, + "content": "theoretical illustration and experiment on the exponential and logistic losses.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 103, + 197, + 507, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 259, + 461, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 462, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 280, + 273 + ], + "score": 1.0, + "content": "Computing the gradient for the model over", + "type": "text" + }, + { + "bbox": [ + 280, + 262, + 288, + 270 + ], + "score": 0.79, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 259, + 416, + 273 + ], + "score": 1.0, + "content": ", the gradient flow dynamics for", + "type": "text" + }, + { + "bbox": [ + 417, + 262, + 424, + 270 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 259, + 462, + 273 + ], + "score": 1.0, + "content": "become:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 259, + 462, + 273 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 287, + 372, + 321 + ], + "lines": [ + { + "bbox": [ + 238, + 287, + 372, + 321 + ], + "spans": [ + { + "bbox": [ + 238, + 287, + 372, + 321 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { i } = \\frac { N ^ { 2 } } { m } \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } \\sum _ { j = 1 } ^ { m } e ^ { - y _ { j } \\langle \\beta , x _ { j } \\rangle } x _ { j }", + "type": "interline_equation", + "image_path": "0bc426bdb9a32f2dac9a10eceb8c5499a734113b8de64f6cb7f5945ec02450ed.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 238, + 287, + 372, + 304.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 238, + 304.0, + 372, + 321.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 332, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 346, + 345 + ], + "score": 1.0, + "content": "We see the same dynamical attenuation of small values of", + "type": "text" + }, + { + "bbox": [ + 347, + 335, + 354, + 343 + ], + "score": 0.79, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "that is seen in the regression model,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 340, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 240, + 364 + ], + "score": 1.0, + "content": "caused by the multiplication by", + "type": "text" + }, + { + "bbox": [ + 244, + 340, + 275, + 357 + ], + "score": 1.0, + "content": "σ2− 2N", + "type": "text" + }, + { + "bbox": [ + 271, + 345, + 506, + 361 + ], + "score": 1.0, + "content": "From this, we can expect the same type of incremental", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 358, + 504, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 228, + 371 + ], + "score": 1.0, + "content": "learning to occur - weights of", + "type": "text" + }, + { + "bbox": [ + 228, + 361, + 236, + 369 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 358, + 504, + 371 + ], + "score": 1.0, + "content": "will be learned incrementally until the dataset can be separated by", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 198, + 381 + ], + "score": 1.0, + "content": "the current support of", + "type": "text" + }, + { + "bbox": [ + 199, + 371, + 206, + 379 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 370, + 505, + 381 + ], + "score": 1.0, + "content": ". Then, the dynamics strengthen the growth of the current support while", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "score": 1.0, + "content": "relatively attenuating that of the other values. Since the data is separated, increasing the values of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "the current support reduces the loss and the magnitude of subsequent gradients, and so we should", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 402, + 449, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 449, + 415 + ], + "score": 1.0, + "content": "expect the support to remain the same and the model to converge to a sparse solution.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 332, + 506, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "Granted, the above description is just intuition, but panel (c) of figure 1 shows how it is born out", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "in practice (similar results are obtained for the logistic loss). In appendix G we further explore", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 454 + ], + "score": 1.0, + "content": "this model, showing deeper networks have a stronger bias for sparsity. We also observe that the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "initialization scale plays a similar role as before - deep models are less biased towards sparsity when", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 463, + 153, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 118, + 474 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 463, + 153, + 477 + ], + "score": 1.0, + "content": "is large.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 419, + 506, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "In their work, Gunasekar et al. (2018) show an equivalence between the diagonal network and the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 492, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 503 + ], + "score": 1.0, + "content": "circular-convolutional network in the frequency domain. According to their results, we should ex-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "pect to see the same sparsity-bias of diagonal networks in convolutional networks, when looking at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 210, + 525 + ], + "score": 1.0, + "content": "the Fourier coefficients of", + "type": "text" + }, + { + "bbox": [ + 210, + 514, + 218, + 523 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 513, + 505, + 525 + ], + "score": 1.0, + "content": ". An example of this can be seen in panel (d) of figure 1, and we refer the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 522, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 104, + 522, + 182, + 538 + ], + "score": 1.0, + "content": "reader to appendix", + "type": "text" + }, + { + "bbox": [ + 183, + 524, + 192, + 534 + ], + "score": 0.27, + "content": "\\mathbf { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 522, + 506, + 538 + ], + "score": 1.0, + "content": "for a full discussion of their convolutional model and it’s incremental learning", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 534, + 150, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 150, + 549 + ], + "score": 1.0, + "content": "dynamics.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 480, + 506, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 564, + 195, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 197, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 197, + 581 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 504, + 625 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "Gradient-based optimization for deep linear models has an implicit bias towards simple (sparse)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 615 + ], + "score": 1.0, + "content": "solutions, caused by an incremental search strategy over the hypothesis space. Deeper models have", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 613, + 488, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 488, + 626 + ], + "score": 1.0, + "content": "a stronger tendency for incremental learning, exhibiting it in more realistic initialization scales.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 591, + 506, + 626 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 630, + 504, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 629, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 505, + 643 + ], + "score": 1.0, + "content": "This dynamical phenomenon exists for the entire optimization process for regression as well as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "classification tasks, and for many types of models - diagonal networks, convolutional networks,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 653, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 664 + ], + "score": 1.0, + "content": "matrix completion and even the nonlinear quadratic network. We believe this kind of dynamical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 663, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 506, + 675 + ], + "score": 1.0, + "content": "analysis may be able to shed light on the generalization of deeper nonlinear neural networks as well,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 675, + 399, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 399, + 686 + ], + "score": 1.0, + "content": "with shallow quadratic networks being only a first step towards that goal.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 629, + 506, + 686 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 701, + 200, + 711 + ], + "lines": [ + { + "bbox": [ + 107, + 701, + 200, + 712 + ], + "spans": [ + { + "bbox": [ + 107, + 701, + 200, + 712 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 436, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 438, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 438, + 735 + ], + "score": 1.0, + "content": "This research is supported by the European Research Council (TheoryDL project).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 718, + 438, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 503, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Sanjeev Arora, Nadav Cohen, and Elad Hazan. On the optimization of deep networks: Implicit", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 433, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 433, + 123 + ], + "score": 1.0, + "content": "acceleration by overparameterization. arXiv preprint arXiv:1802.06509, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 130, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 143 + ], + "score": 1.0, + "content": "Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "matrix factorization. In Advances in Neural Information Processing Systems 32, pp.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 152, + 504, + 164 + ], + "spans": [ + { + "bbox": [ + 115, + 152, + 504, + 164 + ], + "score": 1.0, + "content": "7411–7422. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 490, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 490, + 176 + ], + "score": 1.0, + "content": "8960-implicit-regularization-in-deep-matrix-factorization.pdf.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "Remi Tachet des Combes, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 202, + 143, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 143, + 217 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 505, + 236 + ], + "score": 1.0, + "content": "Gauthier Gidel, Francis Bach, and Simon Lacoste-Julien. Implicit regularization of discrete gradient", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "dynamics in linear neural networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 245, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 192, + 257 + ], + "score": 1.0, + "content": "3196–3206, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 504, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 504, + 278 + ], + "score": 1.0, + "content": "Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 275, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 505, + 290 + ], + "score": 1.0, + "content": "Implicit regularization in matrix factorization. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 287, + 243, + 299 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 243, + 299 + ], + "score": 1.0, + "content": "Systems, pp. 6151–6159, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Suriya Gunasekar, Jason D Lee, Daniel Soudry, and Nati Srebro. Implicit bias of gradient descent", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 316, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 505, + 331 + ], + "score": 1.0, + "content": "on linear convolutional networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 327, + 191, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 327, + 191, + 339 + ], + "score": 1.0, + "content": "9461–9471, 2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "Ziwei Ji and Matus Telgarsky. The implicit bias of gradient descent on nonseparable data. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 358, + 339, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 358, + 339, + 370 + ], + "score": 1.0, + "content": "Conference on Learning Theory, pp. 1772–1798, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "Yuanzhi Li, Tengyu Ma, and Hongyang Zhang. Algorithmic regularization in over-parameterized", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "matrix sensing and neural networks with quadratic activations. arXiv preprint arXiv:1712.09203,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 399, + 144, + 412 + ], + "spans": [ + { + "bbox": [ + 115, + 399, + 144, + 412 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "Mor Shpigel Nacson, Jason Lee, Suriya Gunasekar, Pedro HP Savarese, Nathan Srebro, and Daniel", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "Soudry. Convergence of gradient descent on separable data. arXiv preprint arXiv:1803.01905,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 439, + 144, + 453 + ], + "spans": [ + { + "bbox": [ + 114, + 439, + 144, + 453 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 109, + 459, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "Preetum Nakkiran, Gal Kaplun, Dimitris Kalimeris, Tristan Yang, Benjamin L Edelman, Fred", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "Zhang, and Boaz Barak. Sgd on neural networks learns functions of increasing complexity. arXiv", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 482, + 254, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 482, + 254, + 493 + ], + "score": 1.0, + "content": "preprint arXiv:1905.11604, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Yagyensh Chandra Pati, Ramin Rezaiifar, and Perinkulam Sambamurthy Krishnaprasad. Orthogonal", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "matching pursuit: Recursive function approximation with applications to wavelet decomposition.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "In Proceedings of 27th Asilomar conference on signals, systems and computers, pp. 40–44. IEEE,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 532, + 143, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 143, + 546 + ], + "score": 1.0, + "content": "1993.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 105, + 553, + 610, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 425, + 565 + ], + "score": 1.0, + "content": "Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman.", + "type": "text" + }, + { + "bbox": [ + 446, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "The conver-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 114, + 563, + 477, + 576 + ], + "score": 1.0, + "content": "gence rate of neural networks for learned functions of different frequencies.", + "type": "text" + }, + { + "bbox": [ + 492, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 114, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems 32, pp. 4763–4772. Cur-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 584, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 114, + 585, + 265, + 599 + ], + "score": 1.0, + "content": "ran Associates, Inc., 2019.", + "type": "text" + }, + { + "bbox": [ + 299, + 584, + 505, + 599 + ], + "score": 1.0, + "content": "URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 597, + 613, + 609 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 613, + 609 + ], + "score": 1.0, + "content": "8723-the-convergence-rate-of-neural-networks-for-learned-functions-of-different-fr", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 113, + 609, + 140, + 619 + ], + "spans": [ + { + "bbox": [ + 113, + 609, + 140, + 619 + ], + "score": 1.0, + "content": "pdf.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 627, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynam-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 638, + 460, + 651 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 460, + 651 + ], + "score": 1.0, + "content": "ics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "score": 1.0, + "content": "Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "landscape of over-parameterized shallow neural networks. IEEE Transactions on Information", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 678, + 239, + 692 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 239, + 692 + ], + "score": 1.0, + "content": "Theory, 65(2):742–769, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The im-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "plicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 19", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 206, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 206, + 732 + ], + "score": 1.0, + "content": "(1):2822–2878, 2018.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 503, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Sanjeev Arora, Nadav Cohen, and Elad Hazan. On the optimization of deep networks: Implicit", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 433, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 433, + 123 + ], + "score": 1.0, + "content": "acceleration by overparameterization. arXiv preprint arXiv:1802.06509, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 100, + 505, + 123 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 130, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 143 + ], + "score": 1.0, + "content": "Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "matrix factorization. In Advances in Neural Information Processing Systems 32, pp.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 152, + 504, + 164 + ], + "spans": [ + { + "bbox": [ + 115, + 152, + 504, + 164 + ], + "score": 1.0, + "content": "7411–7422. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 490, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 490, + 176 + ], + "score": 1.0, + "content": "8960-implicit-regularization-in-deep-matrix-factorization.pdf.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 129, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "Remi Tachet des Combes, Mohammad Pezeshki, Samira Shabanian, Aaron Courville, and Yoshua", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Bengio. On the learning dynamics of deep neural networks. arXiv preprint arXiv:1809.06848,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 202, + 143, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 143, + 217 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 181, + 505, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 224, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 505, + 236 + ], + "score": 1.0, + "content": "Gauthier Gidel, Francis Bach, and Simon Lacoste-Julien. Implicit regularization of discrete gradient", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "dynamics in linear neural networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 245, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 192, + 257 + ], + "score": 1.0, + "content": "3196–3206, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 224, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 504, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 504, + 278 + ], + "score": 1.0, + "content": "Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 275, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 505, + 290 + ], + "score": 1.0, + "content": "Implicit regularization in matrix factorization. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 287, + 243, + 299 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 243, + 299 + ], + "score": 1.0, + "content": "Systems, pp. 6151–6159, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 264, + 505, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Suriya Gunasekar, Jason D Lee, Daniel Soudry, and Nati Srebro. Implicit bias of gradient descent", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 316, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 505, + 331 + ], + "score": 1.0, + "content": "on linear convolutional networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 327, + 191, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 327, + 191, + 339 + ], + "score": 1.0, + "content": "9461–9471, 2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 306, + 505, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "Ziwei Ji and Matus Telgarsky. The implicit bias of gradient descent on nonseparable data. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 358, + 339, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 358, + 339, + 370 + ], + "score": 1.0, + "content": "Conference on Learning Theory, pp. 1772–1798, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 347, + 505, + 370 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "Yuanzhi Li, Tengyu Ma, and Hongyang Zhang. Algorithmic regularization in over-parameterized", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "matrix sensing and neural networks with quadratic activations. arXiv preprint arXiv:1712.09203,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 399, + 144, + 412 + ], + "spans": [ + { + "bbox": [ + 115, + 399, + 144, + 412 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 376, + 506, + 412 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "Mor Shpigel Nacson, Jason Lee, Suriya Gunasekar, Pedro HP Savarese, Nathan Srebro, and Daniel", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "Soudry. Convergence of gradient descent on separable data. arXiv preprint arXiv:1803.01905,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 439, + 144, + 453 + ], + "spans": [ + { + "bbox": [ + 114, + 439, + 144, + 453 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 418, + 506, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 459, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "Preetum Nakkiran, Gal Kaplun, Dimitris Kalimeris, Tristan Yang, Benjamin L Edelman, Fred", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 115, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "Zhang, and Boaz Barak. Sgd on neural networks learns functions of increasing complexity. arXiv", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 482, + 254, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 482, + 254, + 493 + ], + "score": 1.0, + "content": "preprint arXiv:1905.11604, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 458, + 506, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Yagyensh Chandra Pati, Ramin Rezaiifar, and Perinkulam Sambamurthy Krishnaprasad. Orthogonal", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "matching pursuit: Recursive function approximation with applications to wavelet decomposition.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "In Proceedings of 27th Asilomar conference on signals, systems and computers, pp. 40–44. IEEE,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 532, + 143, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 143, + 546 + ], + "score": 1.0, + "content": "1993.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 501, + 506, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 553, + 610, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 425, + 565 + ], + "score": 1.0, + "content": "Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman.", + "type": "text" + }, + { + "bbox": [ + 446, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "The conver-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 114, + 563, + 477, + 576 + ], + "score": 1.0, + "content": "gence rate of neural networks for learned functions of different frequencies.", + "type": "text" + }, + { + "bbox": [ + 492, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 114, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems 32, pp. 4763–4772. Cur-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 584, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 114, + 585, + 265, + 599 + ], + "score": 1.0, + "content": "ran Associates, Inc., 2019.", + "type": "text" + }, + { + "bbox": [ + 299, + 584, + 505, + 599 + ], + "score": 1.0, + "content": "URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 597, + 613, + 609 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 613, + 609 + ], + "score": 1.0, + "content": "8723-the-convergence-rate-of-neural-networks-for-learned-functions-of-different-fr", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 113, + 609, + 140, + 619 + ], + "spans": [ + { + "bbox": [ + 113, + 609, + 140, + 619 + ], + "score": 1.0, + "content": "pdf.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 552, + 613, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 627, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynam-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 638, + 460, + 651 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 460, + 651 + ], + "score": 1.0, + "content": "ics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 627, + 505, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "score": 1.0, + "content": "Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "landscape of over-parameterized shallow neural networks. IEEE Transactions on Information", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 678, + 239, + 692 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 239, + 692 + ], + "score": 1.0, + "content": "Theory, 65(2):742–769, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 656, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gunasekar, and Nathan Srebro. The im-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "plicit bias of gradient descent on separable data. The Journal of Machine Learning Research, 19", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 206, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 206, + 732 + ], + "score": 1.0, + "content": "(1):2822–2878, 2018.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 699, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Francis Williams, Matthew Trager, Claudio Silva, Daniele Panozzo, Denis Zorin, and Joan Bruna.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 93, + 502, + 105 + ], + "spans": [ + { + "bbox": [ + 117, + 93, + 502, + 105 + ], + "score": 1.0, + "content": "Gradient dynamics of shallow univariate relu networks. arXiv preprint arXiv:1906.07842, 2019.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Blake Woodworth, Suriya Gunasekar, Jason Lee, Daniel Soudry, and Nathan Srebro. Kernel and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 451, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 451, + 135 + ], + "score": 1.0, + "content": "deep regimes in overparametrized models. arXiv preprint arXiv:1906.05827, 2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 153, + 245, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 246, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 246, + 169 + ], + "score": 1.0, + "content": "A PROOF OF THEOREM 2", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 506, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 225, + 198 + ], + "score": 1.0, + "content": "Theorem. Given two values", + "type": "text" + }, + { + "bbox": [ + 225, + 181, + 250, + 191 + ], + "score": 0.9, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 174, + 419, + 198 + ], + "score": 1.0, + "content": "of a toy linear model as in (1), such that", + "type": "text" + }, + { + "bbox": [ + 419, + 177, + 471, + 195 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 174, + 505, + 198 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 194, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 322, + 210 + ], + "score": 1.0, + "content": "model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 322, + 194, + 367, + 208 + ], + "score": 0.93, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 194, + 388, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 388, + 195, + 434, + 208 + ], + "score": 0.92, + "content": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 194, + 506, + 210 + ], + "score": 1.0, + "content": ", then the largest", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 204, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 104, + 204, + 374, + 223 + ], + "score": 1.0, + "content": "initialization value for which the learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 374, + 208, + 397, + 220 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 204, + 486, + 223 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 486, + 208, + 501, + 220 + ], + "score": 0.89, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 204, + 506, + 223 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 218, + 308, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 308, + 232 + ], + "score": 1.0, + "content": "is lower and upper bounded in the following way:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 151, + 249, + 460, + 306 + ], + "lines": [ + { + "bbox": [ + 151, + 249, + 460, + 306 + ], + "spans": [ + { + "bbox": [ + 151, + 249, + 460, + 306 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}", + "type": "interline_equation", + "image_path": "8a0b254cfd784cab90491532e9bfa6e3bbe6c232bb42638b93c4f5af2af1436f.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 151, + 249, + 460, + 268.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 151, + 268.0, + 460, + 287.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 151, + 287.0, + 460, + 306.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 286, + 328 + ], + "score": 1.0, + "content": "Proof. Our strategy will be to define the time", + "type": "text" + }, + { + "bbox": [ + 286, + 317, + 297, + 326 + ], + "score": 0.88, + "content": "t _ { \\alpha }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 315, + 439, + 328 + ], + "score": 1.0, + "content": "for which a value reaches a fraction", + "type": "text" + }, + { + "bbox": [ + 439, + 317, + 447, + 325 + ], + "score": 0.78, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "of it’s optimal", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 216, + 340 + ], + "score": 1.0, + "content": "value, and then require that", + "type": "text" + }, + { + "bbox": [ + 216, + 326, + 281, + 339 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 326, + 505, + 340 + ], + "score": 1.0, + "content": ". We begin with recalling the differential equation which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 337, + 264, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 264, + 349 + ], + "score": 1.0, + "content": "determines the dynamics of the model:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 263, + 361, + 347, + 378 + ], + "lines": [ + { + "bbox": [ + 263, + 361, + 347, + 378 + ], + "spans": [ + { + "bbox": [ + 263, + 361, + 347, + 378 + ], + "score": 0.9, + "content": "\\dot { \\sigma } = \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma )", + "type": "interline_equation", + "image_path": "247f451707f15659e57f94c99d582f7d5c0df14f15bb12b92e1b3e3e297da7ca.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 263, + 361, + 347, + 378 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 194, + 398 + ], + "score": 1.0, + "content": "Since the solution for", + "type": "text" + }, + { + "bbox": [ + 195, + 385, + 223, + 396 + ], + "score": 0.91, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 384, + 493, + 398 + ], + "score": 1.0, + "content": "is implicit and difficult to manage in a general form, we will define", + "type": "text" + }, + { + "bbox": [ + 493, + 386, + 504, + 396 + ], + "score": 0.86, + "content": "t _ { \\alpha }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "using the integral of the differential equation. The equation is separable, and under initialization of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 407, + 299, + 421 + ], + "spans": [ + { + "bbox": [ + 107, + 409, + 118, + 418 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 407, + 185, + 421 + ], + "score": 1.0, + "content": "we can describe", + "type": "text" + }, + { + "bbox": [ + 185, + 407, + 209, + 419 + ], + "score": 0.93, + "content": "t _ { \\alpha } ( \\sigma )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 407, + 299, + 421 + ], + "score": 1.0, + "content": "in the following way:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 431, + 370, + 462 + ], + "lines": [ + { + "bbox": [ + 240, + 431, + 370, + 462 + ], + "spans": [ + { + "bbox": [ + 240, + 431, + 370, + 462 + ], + "score": 0.93, + "content": "t _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "4ff2011e8893ed9bb5cee419ee302a60aff79ef738e7c946be049ea2688e2405.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 431, + 370, + 446.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 240, + 446.5, + 370, + 462.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 264, + 483 + ], + "score": 1.0, + "content": "Incremental learning takes place when", + "type": "text" + }, + { + "bbox": [ + 264, + 469, + 322, + 482 + ], + "score": 0.93, + "content": "\\sigma _ { i } ( t _ { f } ) = f \\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 468, + 387, + 483 + ], + "score": 1.0, + "content": "happens before", + "type": "text" + }, + { + "bbox": [ + 388, + 470, + 444, + 483 + ], + "score": 0.93, + "content": "\\sigma _ { j } ( t _ { s } ) = s \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 468, + 505, + 483 + ], + "score": 1.0, + "content": ". We can write", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 479, + 251, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 251, + 495 + ], + "score": 1.0, + "content": "this condition in the following way:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 504, + 403, + 537 + ], + "lines": [ + { + "bbox": [ + 207, + 504, + 403, + 537 + ], + "spans": [ + { + "bbox": [ + 207, + 504, + 403, + 537 + ], + "score": 0.92, + "content": "\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "31e330b6bdd3b57f2f78fe7c9959612954102e9f4dd0c887df02e2c4286879b5.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 504, + 403, + 520.5 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 207, + 520.5, + 403, + 537.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 156, + 558 + ], + "score": 1.0, + "content": "Plugging in", + "type": "text" + }, + { + "bbox": [ + 156, + 546, + 195, + 557 + ], + "score": 0.89, + "content": "\\sigma _ { i } = r \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "and rearranging, we get the following necessary and sufficient condition for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 554, + 194, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 194, + 569 + ], + "score": 1.0, + "content": "incremental learning:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 579, + 410, + 615 + ], + "lines": [ + { + "bbox": [ + 200, + 579, + 410, + 615 + ], + "spans": [ + { + "bbox": [ + 200, + 579, + 410, + 615 + ], + "score": 0.93, + "content": "\\int _ { \\sigma _ { 0 } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "0665e01f255ebd93529f2cfb9ec005b44f08bd939e9bc0acbcb0b6855cb8397c.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 200, + 579, + 410, + 597.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 200, + 597.0, + 410, + 615.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 506, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Our last step before relaxing and restricting our condition will be to split the integral on the left-hand", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 632, + 200, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 200, + 645 + ], + "score": 1.0, + "content": "side into two integrals:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 656, + 462, + 692 + ], + "lines": [ + { + "bbox": [ + 148, + 656, + 462, + 692 + ], + "spans": [ + { + "bbox": [ + 148, + 656, + 462, + 692 + ], + "score": 0.94, + "content": "\\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } + \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "2d0910f469d5ab8099d3e7561097b2dc6dbda9317da20e4fa4db434c2157b4e5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 148, + 656, + 462, + 668.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 148, + 668.0, + 462, + 680.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 148, + 680.0, + 462, + 692.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 339, + 711 + ], + "score": 1.0, + "content": "At this point, we cannot solve this equation and isolate", + "type": "text" + }, + { + "bbox": [ + 339, + 701, + 350, + 710 + ], + "score": 0.83, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "to obtain a clear threshold condition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "on it for incremental learning. Instead, we will relax/restrict the above condition to get a neces-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 494, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 218, + 734 + ], + "score": 1.0, + "content": "sary/sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 218, + 722, + 229, + 732 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 720, + 478, + 734 + ], + "score": 1.0, + "content": ", leading to a lower and upper bound on the threshold value of", + "type": "text" + }, + { + "bbox": [ + 478, + 722, + 489, + 732 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 720, + 494, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Francis Williams, Matthew Trager, Claudio Silva, Daniele Panozzo, Denis Zorin, and Joan Bruna.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 93, + 502, + 105 + ], + "spans": [ + { + "bbox": [ + 117, + 93, + 502, + 105 + ], + "score": 1.0, + "content": "Gradient dynamics of shallow univariate relu networks. arXiv preprint arXiv:1906.07842, 2019.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Blake Woodworth, Suriya Gunasekar, Jason Lee, Daniel Soudry, and Nathan Srebro. Kernel and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 451, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 451, + 135 + ], + "score": 1.0, + "content": "deep regimes in overparametrized models. arXiv preprint arXiv:1906.05827, 2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 112, + 505, + 135 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 153, + 245, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 246, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 246, + 169 + ], + "score": 1.0, + "content": "A PROOF OF THEOREM 2", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 506, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 225, + 198 + ], + "score": 1.0, + "content": "Theorem. Given two values", + "type": "text" + }, + { + "bbox": [ + 225, + 181, + 250, + 191 + ], + "score": 0.9, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 174, + 419, + 198 + ], + "score": 1.0, + "content": "of a toy linear model as in (1), such that", + "type": "text" + }, + { + "bbox": [ + 419, + 177, + 471, + 195 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 174, + 505, + 198 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 194, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 322, + 210 + ], + "score": 1.0, + "content": "model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 322, + 194, + 367, + 208 + ], + "score": 0.93, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 194, + 388, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 388, + 195, + 434, + 208 + ], + "score": 0.92, + "content": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 194, + 506, + 210 + ], + "score": 1.0, + "content": ", then the largest", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 204, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 104, + 204, + 374, + 223 + ], + "score": 1.0, + "content": "initialization value for which the learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 374, + 208, + 397, + 220 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 204, + 486, + 223 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 486, + 208, + 501, + 220 + ], + "score": 0.89, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 204, + 506, + 223 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 218, + 308, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 308, + 232 + ], + "score": 1.0, + "content": "is lower and upper bounded in the following way:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 174, + 506, + 232 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 151, + 249, + 460, + 306 + ], + "lines": [ + { + "bbox": [ + 151, + 249, + 460, + 306 + ], + "spans": [ + { + "bbox": [ + 151, + 249, + 460, + 306 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}", + "type": "interline_equation", + "image_path": "8a0b254cfd784cab90491532e9bfa6e3bbe6c232bb42638b93c4f5af2af1436f.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 151, + 249, + 460, + 268.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 151, + 268.0, + 460, + 287.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 151, + 287.0, + 460, + 306.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 286, + 328 + ], + "score": 1.0, + "content": "Proof. Our strategy will be to define the time", + "type": "text" + }, + { + "bbox": [ + 286, + 317, + 297, + 326 + ], + "score": 0.88, + "content": "t _ { \\alpha }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 315, + 439, + 328 + ], + "score": 1.0, + "content": "for which a value reaches a fraction", + "type": "text" + }, + { + "bbox": [ + 439, + 317, + 447, + 325 + ], + "score": 0.78, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "of it’s optimal", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 216, + 340 + ], + "score": 1.0, + "content": "value, and then require that", + "type": "text" + }, + { + "bbox": [ + 216, + 326, + 281, + 339 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 326, + 505, + 340 + ], + "score": 1.0, + "content": ". We begin with recalling the differential equation which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 337, + 264, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 264, + 349 + ], + "score": 1.0, + "content": "determines the dynamics of the model:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 315, + 505, + 349 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 263, + 361, + 347, + 378 + ], + "lines": [ + { + "bbox": [ + 263, + 361, + 347, + 378 + ], + "spans": [ + { + "bbox": [ + 263, + 361, + 347, + 378 + ], + "score": 0.9, + "content": "\\dot { \\sigma } = \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma )", + "type": "interline_equation", + "image_path": "247f451707f15659e57f94c99d582f7d5c0df14f15bb12b92e1b3e3e297da7ca.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 263, + 361, + 347, + 378 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 194, + 398 + ], + "score": 1.0, + "content": "Since the solution for", + "type": "text" + }, + { + "bbox": [ + 195, + 385, + 223, + 396 + ], + "score": 0.91, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 384, + 493, + 398 + ], + "score": 1.0, + "content": "is implicit and difficult to manage in a general form, we will define", + "type": "text" + }, + { + "bbox": [ + 493, + 386, + 504, + 396 + ], + "score": 0.86, + "content": "t _ { \\alpha }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "using the integral of the differential equation. The equation is separable, and under initialization of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 407, + 299, + 421 + ], + "spans": [ + { + "bbox": [ + 107, + 409, + 118, + 418 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 407, + 185, + 421 + ], + "score": 1.0, + "content": "we can describe", + "type": "text" + }, + { + "bbox": [ + 185, + 407, + 209, + 419 + ], + "score": 0.93, + "content": "t _ { \\alpha } ( \\sigma )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 407, + 299, + 421 + ], + "score": 1.0, + "content": "in the following way:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 384, + 505, + 421 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 431, + 370, + 462 + ], + "lines": [ + { + "bbox": [ + 240, + 431, + 370, + 462 + ], + "spans": [ + { + "bbox": [ + 240, + 431, + 370, + 462 + ], + "score": 0.93, + "content": "t _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "4ff2011e8893ed9bb5cee419ee302a60aff79ef738e7c946be049ea2688e2405.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 431, + 370, + 446.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 240, + 446.5, + 370, + 462.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 264, + 483 + ], + "score": 1.0, + "content": "Incremental learning takes place when", + "type": "text" + }, + { + "bbox": [ + 264, + 469, + 322, + 482 + ], + "score": 0.93, + "content": "\\sigma _ { i } ( t _ { f } ) = f \\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 468, + 387, + 483 + ], + "score": 1.0, + "content": "happens before", + "type": "text" + }, + { + "bbox": [ + 388, + 470, + 444, + 483 + ], + "score": 0.93, + "content": "\\sigma _ { j } ( t _ { s } ) = s \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 468, + 505, + 483 + ], + "score": 1.0, + "content": ". We can write", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 479, + 251, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 251, + 495 + ], + "score": 1.0, + "content": "this condition in the following way:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 468, + 505, + 495 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 504, + 403, + 537 + ], + "lines": [ + { + "bbox": [ + 207, + 504, + 403, + 537 + ], + "spans": [ + { + "bbox": [ + 207, + 504, + 403, + 537 + ], + "score": 0.92, + "content": "\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }", + "type": "interline_equation", + "image_path": "31e330b6bdd3b57f2f78fe7c9959612954102e9f4dd0c887df02e2c4286879b5.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 504, + 403, + 520.5 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 207, + 520.5, + 403, + 537.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 156, + 558 + ], + "score": 1.0, + "content": "Plugging in", + "type": "text" + }, + { + "bbox": [ + 156, + 546, + 195, + 557 + ], + "score": 0.89, + "content": "\\sigma _ { i } = r \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "and rearranging, we get the following necessary and sufficient condition for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 554, + 194, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 194, + 569 + ], + "score": 1.0, + "content": "incremental learning:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 543, + 505, + 569 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 579, + 410, + 615 + ], + "lines": [ + { + "bbox": [ + 200, + 579, + 410, + 615 + ], + "spans": [ + { + "bbox": [ + 200, + 579, + 410, + 615 + ], + "score": 0.93, + "content": "\\int _ { \\sigma _ { 0 } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "0665e01f255ebd93529f2cfb9ec005b44f08bd939e9bc0acbcb0b6855cb8397c.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 200, + 579, + 410, + 597.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 200, + 597.0, + 410, + 615.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 506, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Our last step before relaxing and restricting our condition will be to split the integral on the left-hand", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 632, + 200, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 200, + 645 + ], + "score": 1.0, + "content": "side into two integrals:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 621, + 505, + 645 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 656, + 462, + 692 + ], + "lines": [ + { + "bbox": [ + 148, + 656, + 462, + 692 + ], + "spans": [ + { + "bbox": [ + 148, + 656, + 462, + 692 + ], + "score": 0.94, + "content": "\\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } + \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "2d0910f469d5ab8099d3e7561097b2dc6dbda9317da20e4fa4db434c2157b4e5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 148, + 656, + 462, + 668.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 148, + 668.0, + 462, + 680.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 148, + 680.0, + 462, + 692.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 339, + 711 + ], + "score": 1.0, + "content": "At this point, we cannot solve this equation and isolate", + "type": "text" + }, + { + "bbox": [ + 339, + 701, + 350, + 710 + ], + "score": 0.83, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "to obtain a clear threshold condition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "on it for incremental learning. Instead, we will relax/restrict the above condition to get a neces-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 494, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 218, + 734 + ], + "score": 1.0, + "content": "sary/sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 218, + 722, + 229, + 732 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 720, + 478, + 734 + ], + "score": 1.0, + "content": ", leading to a lower and upper bound on the threshold value of", + "type": "text" + }, + { + "bbox": [ + 478, + 722, + 489, + 732 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 720, + 494, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 699, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 211, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 212, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 212, + 95 + ], + "score": 1.0, + "content": "SUFFICIENT CONDITION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 506, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 322, + 116 + ], + "score": 1.0, + "content": "To obtain a sufficient (but not necessary) condition on", + "type": "text" + }, + { + "bbox": [ + 323, + 105, + 334, + 114 + ], + "score": 0.86, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 102, + 506, + 116 + ], + "score": 1.0, + "content": ", we may make the condition stricter either", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "by increasing the left-hand side or decreasing the right-hand side. We can increase the left-hand side", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 160, + 138 + ], + "score": 1.0, + "content": "by removing", + "type": "text" + }, + { + "bbox": [ + 161, + 127, + 167, + 135 + ], + "score": 0.71, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 125, + 342, + 138 + ], + "score": 1.0, + "content": "from the left-most integral’s denominator", + "type": "text" + }, + { + "bbox": [ + 342, + 126, + 371, + 136 + ], + "score": 0.85, + "content": "( r > 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 125, + 506, + 138 + ], + "score": 1.0, + "content": ") and then combine the left-most", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 136, + 207, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 207, + 148 + ], + "score": 1.0, + "content": "and right-most integrals:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 160, + 423, + 196 + ], + "lines": [ + { + "bbox": [ + 188, + 160, + 423, + 196 + ], + "spans": [ + { + "bbox": [ + 188, + 160, + 423, + 196 + ], + "score": 0.94, + "content": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "e838aa694500dfac4c54e68c27c20a5779a2e0f05b3746a5c346dbc69cceb6ca.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 160, + 423, + 178.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 178.0, + 423, + 196.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 202, + 505, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 356, + 214 + ], + "score": 1.0, + "content": "Next, we note that the integration bounds give us a bound on", + "type": "text" + }, + { + "bbox": [ + 356, + 205, + 363, + 212 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "for either integral. This means we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 211, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 154, + 225 + ], + "score": 1.0, + "content": "can replace", + "type": "text" + }, + { + "bbox": [ + 155, + 213, + 185, + 230 + ], + "score": 0.92, + "content": "\\begin{array} { r } { 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 211, + 356, + 233 + ], + "score": 1.0, + "content": "with 1 on the right-hand side, and replace", + "type": "text" + }, + { + "bbox": [ + 357, + 213, + 390, + 230 + ], + "score": 0.92, + "content": "\\begin{array} { r } { 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 213, + 412, + 232 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 413, + 213, + 438, + 225 + ], + "score": 0.9, + "content": "1 - f", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 213, + 505, + 232 + ], + "score": 1.0, + "content": "on the left-hand", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 226, + 130, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 130, + 240 + ], + "score": 1.0, + "content": "side:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 252, + 398, + 284 + ], + "lines": [ + { + "bbox": [ + 213, + 252, + 398, + 284 + ], + "spans": [ + { + "bbox": [ + 213, + 252, + 398, + 284 + ], + "score": 0.93, + "content": "\\frac { 1 } { 1 - f } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }", + "type": "interline_equation", + "image_path": "f005d94ad21671d3896b0f5f73445479860092c93e94c6d30bd718edf61a3cef.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 213, + 252, + 398, + 268.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 213, + 268.0, + 398, + 284.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 293, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 507, + 309 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 280, + 309 + ], + "score": 1.0, + "content": "We may now solve these integrals for every", + "type": "text" + }, + { + "bbox": [ + 281, + 294, + 291, + 303 + ], + "score": 0.86, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 289, + 336, + 309 + ], + "score": 1.0, + "content": "and isolate", + "type": "text" + }, + { + "bbox": [ + 336, + 295, + 348, + 304 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 289, + 470, + 309 + ], + "score": 1.0, + "content": ", obtaining the lower bound on", + "type": "text" + }, + { + "bbox": [ + 470, + 292, + 485, + 305 + ], + "score": 0.92, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 289, + 507, + 309 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 241, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 207, + 316 + ], + "score": 1.0, + "content": "start with the case where", + "type": "text" + }, + { + "bbox": [ + 208, + 304, + 236, + 314 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 303, + 241, + 316 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 174, + 328, + 436, + 354 + ], + "lines": [ + { + "bbox": [ + 174, + 328, + 436, + 354 + ], + "spans": [ + { + "bbox": [ + 174, + 328, + 436, + 354 + ], + "score": 0.92, + "content": "\\frac { 1 } { 1 - f } \\big ( \\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\big ) \\leq ( r - 1 ) \\big ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\big )", + "type": "interline_equation", + "image_path": "ed37d967a247a7869497d836ecb1529455eb4ab2b29346cca432e01ee7cc8b5f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 174, + 328, + 436, + 354 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 296, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 297, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 197, + 374 + ], + "score": 1.0, + "content": "Rearranging to isolate", + "type": "text" + }, + { + "bbox": [ + 197, + 363, + 208, + 372 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 360, + 297, + 374 + ], + "score": 1.0, + "content": ", we obtain our result:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 254, + 385, + 356, + 410 + ], + "lines": [ + { + "bbox": [ + 254, + 385, + 356, + 410 + ], + "spans": [ + { + "bbox": [ + 254, + 385, + 356, + 410 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( { \\frac { s } { r f } } \\big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } }", + "type": "interline_equation", + "image_path": "c3c5fdd56841a4a51a8156094280ed16da7dfce8d3de6ac369a84cd1f1c553b9.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 254, + 385, + 356, + 410 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 416, + 386, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 415, + 387, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 137, + 430 + ], + "score": 1.0, + "content": "For the", + "type": "text" + }, + { + "bbox": [ + 137, + 417, + 166, + 428 + ], + "score": 0.9, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 415, + 387, + 430 + ], + "score": 1.0, + "content": "case, we have the following after solving the integrals:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 440, + 456, + 469 + ], + "lines": [ + { + "bbox": [ + 153, + 440, + 456, + 469 + ], + "spans": [ + { + "bbox": [ + 153, + 440, + 456, + 469 + ], + "score": 0.93, + "content": "\\frac { 1 } { 1 - f } \\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq ( r - 1 ) \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )", + "type": "interline_equation", + "image_path": "549d15c01baa23c359a2f826a5e804905209affd1a36eef36205bd5a2547ba92.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 153, + 440, + 456, + 449.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 153, + 449.6666666666667, + 456, + 459.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 153, + 459.33333333333337, + 456, + 469.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 404, + 494 + ], + "score": 1.0, + "content": "For simplicity we may further restrict the condition by removing the term", + "type": "text" + }, + { + "bbox": [ + 404, + 476, + 452, + 497 + ], + "score": 0.93, + "content": "\\Big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 474, + 506, + 495 + ], + "score": 1.0, + "content": "\u00011− 2N . Solving for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 495, + 212, + 508 + ], + "spans": [ + { + "bbox": [ + 107, + 497, + 118, + 506 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 495, + 212, + 508 + ], + "score": 1.0, + "content": "gives us the following:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 519, + 380, + 548 + ], + "lines": [ + { + "bbox": [ + 230, + 519, + 380, + 548 + ], + "spans": [ + { + "bbox": [ + 230, + 519, + 380, + 548 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } }", + "type": "interline_equation", + "image_path": "69b4202b14fed56c585a7999a284af0a6a5a79b02826cf7d9933ddbba295cad6.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 519, + 380, + 533.5 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 230, + 533.5, + 380, + 548.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 559, + 213, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 213, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 213, + 571 + ], + "score": 1.0, + "content": "NECESSARY CONDITION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 326, + 592 + ], + "score": 1.0, + "content": "To obtain a necessary (but not sufficient) condition on", + "type": "text" + }, + { + "bbox": [ + 327, + 582, + 338, + 591 + ], + "score": 0.86, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 579, + 505, + 592 + ], + "score": 1.0, + "content": ", we may relax the condition in (8) either", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "by decreasing the left-hand side or increasing the right-hand side. We begin by rearranging the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 601, + 147, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 601, + 147, + 615 + ], + "score": 1.0, + "content": "equation:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 625, + 462, + 661 + ], + "lines": [ + { + "bbox": [ + 147, + 625, + 462, + 661 + ], + "spans": [ + { + "bbox": [ + 147, + 625, + 462, + 661 + ], + "score": 0.94, + "content": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) } - \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "813b870fe15e38f3325a5a02b85ccc357d691c80b99b6fa649998cfc632f0aeb.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 147, + 625, + 462, + 637.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 147, + 637.0, + 462, + 649.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 147, + 649.0, + 462, + 661.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 667, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 340, + 680 + ], + "score": 1.0, + "content": "Like before, we may use the integration bounds to bound", + "type": "text" + }, + { + "bbox": [ + 340, + 670, + 347, + 678 + ], + "score": 0.75, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 667, + 402, + 680 + ], + "score": 1.0, + "content": ". Plugging in", + "type": "text" + }, + { + "bbox": [ + 402, + 668, + 439, + 681 + ], + "score": 0.92, + "content": "\\sigma = s \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "for all integrals", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 466, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 466, + 692 + ], + "score": 1.0, + "content": "decreases the left-hand side and increases the right-hand side, leading us to the following:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 703, + 416, + 735 + ], + "lines": [ + { + "bbox": [ + 194, + 703, + 416, + 735 + ], + "spans": [ + { + "bbox": [ + 194, + 703, + 416, + 735 + ], + "score": 0.93, + "content": "\\frac { r } { r - s } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le \\big ( \\frac { r } { 1 - s } - \\frac { r } { r - s } \\big ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }", + "type": "interline_equation", + "image_path": "a3eb3dee43b36a19272b64864fcdf199a5779479c58a832565c195196b45b2dd.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 194, + 703, + 416, + 719.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 194, + 719.0, + 416, + 735.0 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 211, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 212, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 212, + 95 + ], + "score": 1.0, + "content": "SUFFICIENT CONDITION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 506, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 322, + 116 + ], + "score": 1.0, + "content": "To obtain a sufficient (but not necessary) condition on", + "type": "text" + }, + { + "bbox": [ + 323, + 105, + 334, + 114 + ], + "score": 0.86, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 102, + 506, + 116 + ], + "score": 1.0, + "content": ", we may make the condition stricter either", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "by increasing the left-hand side or decreasing the right-hand side. We can increase the left-hand side", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 160, + 138 + ], + "score": 1.0, + "content": "by removing", + "type": "text" + }, + { + "bbox": [ + 161, + 127, + 167, + 135 + ], + "score": 0.71, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 125, + 342, + 138 + ], + "score": 1.0, + "content": "from the left-most integral’s denominator", + "type": "text" + }, + { + "bbox": [ + 342, + 126, + 371, + 136 + ], + "score": 0.85, + "content": "( r > 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 125, + 506, + 138 + ], + "score": 1.0, + "content": ") and then combine the left-most", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 136, + 207, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 207, + 148 + ], + "score": 1.0, + "content": "and right-most integrals:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 102, + 506, + 148 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 160, + 423, + 196 + ], + "lines": [ + { + "bbox": [ + 188, + 160, + 423, + 196 + ], + "spans": [ + { + "bbox": [ + 188, + 160, + 423, + 196 + ], + "score": 0.94, + "content": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "e838aa694500dfac4c54e68c27c20a5779a2e0f05b3746a5c346dbc69cceb6ca.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 160, + 423, + 178.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 178.0, + 423, + 196.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 202, + 505, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 356, + 214 + ], + "score": 1.0, + "content": "Next, we note that the integration bounds give us a bound on", + "type": "text" + }, + { + "bbox": [ + 356, + 205, + 363, + 212 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "for either integral. This means we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 211, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 154, + 225 + ], + "score": 1.0, + "content": "can replace", + "type": "text" + }, + { + "bbox": [ + 155, + 213, + 185, + 230 + ], + "score": 0.92, + "content": "\\begin{array} { r } { 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 211, + 356, + 233 + ], + "score": 1.0, + "content": "with 1 on the right-hand side, and replace", + "type": "text" + }, + { + "bbox": [ + 357, + 213, + 390, + 230 + ], + "score": 0.92, + "content": "\\begin{array} { r } { 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 213, + 412, + 232 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 413, + 213, + 438, + 225 + ], + "score": 0.9, + "content": "1 - f", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 213, + 505, + 232 + ], + "score": 1.0, + "content": "on the left-hand", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 226, + 130, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 130, + 240 + ], + "score": 1.0, + "content": "side:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 201, + 506, + 240 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 252, + 398, + 284 + ], + "lines": [ + { + "bbox": [ + 213, + 252, + 398, + 284 + ], + "spans": [ + { + "bbox": [ + 213, + 252, + 398, + 284 + ], + "score": 0.93, + "content": "\\frac { 1 } { 1 - f } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }", + "type": "interline_equation", + "image_path": "f005d94ad21671d3896b0f5f73445479860092c93e94c6d30bd718edf61a3cef.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 213, + 252, + 398, + 268.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 213, + 268.0, + 398, + 284.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 293, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 507, + 309 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 280, + 309 + ], + "score": 1.0, + "content": "We may now solve these integrals for every", + "type": "text" + }, + { + "bbox": [ + 281, + 294, + 291, + 303 + ], + "score": 0.86, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 289, + 336, + 309 + ], + "score": 1.0, + "content": "and isolate", + "type": "text" + }, + { + "bbox": [ + 336, + 295, + 348, + 304 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 289, + 470, + 309 + ], + "score": 1.0, + "content": ", obtaining the lower bound on", + "type": "text" + }, + { + "bbox": [ + 470, + 292, + 485, + 305 + ], + "score": 0.92, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 289, + 507, + 309 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 241, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 207, + 316 + ], + "score": 1.0, + "content": "start with the case where", + "type": "text" + }, + { + "bbox": [ + 208, + 304, + 236, + 314 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 303, + 241, + 316 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 289, + 507, + 316 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 174, + 328, + 436, + 354 + ], + "lines": [ + { + "bbox": [ + 174, + 328, + 436, + 354 + ], + "spans": [ + { + "bbox": [ + 174, + 328, + 436, + 354 + ], + "score": 0.92, + "content": "\\frac { 1 } { 1 - f } \\big ( \\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\big ) \\leq ( r - 1 ) \\big ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\big )", + "type": "interline_equation", + "image_path": "ed37d967a247a7869497d836ecb1529455eb4ab2b29346cca432e01ee7cc8b5f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 174, + 328, + 436, + 354 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 296, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 297, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 197, + 374 + ], + "score": 1.0, + "content": "Rearranging to isolate", + "type": "text" + }, + { + "bbox": [ + 197, + 363, + 208, + 372 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 360, + 297, + 374 + ], + "score": 1.0, + "content": ", we obtain our result:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 360, + 297, + 374 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 254, + 385, + 356, + 410 + ], + "lines": [ + { + "bbox": [ + 254, + 385, + 356, + 410 + ], + "spans": [ + { + "bbox": [ + 254, + 385, + 356, + 410 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( { \\frac { s } { r f } } \\big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } }", + "type": "interline_equation", + "image_path": "c3c5fdd56841a4a51a8156094280ed16da7dfce8d3de6ac369a84cd1f1c553b9.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 254, + 385, + 356, + 410 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 416, + 386, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 415, + 387, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 137, + 430 + ], + "score": 1.0, + "content": "For the", + "type": "text" + }, + { + "bbox": [ + 137, + 417, + 166, + 428 + ], + "score": 0.9, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 415, + 387, + 430 + ], + "score": 1.0, + "content": "case, we have the following after solving the integrals:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 415, + 387, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 440, + 456, + 469 + ], + "lines": [ + { + "bbox": [ + 153, + 440, + 456, + 469 + ], + "spans": [ + { + "bbox": [ + 153, + 440, + 456, + 469 + ], + "score": 0.93, + "content": "\\frac { 1 } { 1 - f } \\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq ( r - 1 ) \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )", + "type": "interline_equation", + "image_path": "549d15c01baa23c359a2f826a5e804905209affd1a36eef36205bd5a2547ba92.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 153, + 440, + 456, + 449.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 153, + 449.6666666666667, + 456, + 459.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 153, + 459.33333333333337, + 456, + 469.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 404, + 494 + ], + "score": 1.0, + "content": "For simplicity we may further restrict the condition by removing the term", + "type": "text" + }, + { + "bbox": [ + 404, + 476, + 452, + 497 + ], + "score": 0.93, + "content": "\\Big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 474, + 506, + 495 + ], + "score": 1.0, + "content": "\u00011− 2N . Solving for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 495, + 212, + 508 + ], + "spans": [ + { + "bbox": [ + 107, + 497, + 118, + 506 + ], + "score": 0.81, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 495, + 212, + 508 + ], + "score": 1.0, + "content": "gives us the following:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 474, + 506, + 508 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 519, + 380, + 548 + ], + "lines": [ + { + "bbox": [ + 230, + 519, + 380, + 548 + ], + "spans": [ + { + "bbox": [ + 230, + 519, + 380, + 548 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } }", + "type": "interline_equation", + "image_path": "69b4202b14fed56c585a7999a284af0a6a5a79b02826cf7d9933ddbba295cad6.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 519, + 380, + 533.5 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 230, + 533.5, + 380, + 548.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 559, + 213, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 213, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 213, + 571 + ], + "score": 1.0, + "content": "NECESSARY CONDITION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 326, + 592 + ], + "score": 1.0, + "content": "To obtain a necessary (but not sufficient) condition on", + "type": "text" + }, + { + "bbox": [ + 327, + 582, + 338, + 591 + ], + "score": 0.86, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 579, + 505, + 592 + ], + "score": 1.0, + "content": ", we may relax the condition in (8) either", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "by decreasing the left-hand side or increasing the right-hand side. We begin by rearranging the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 601, + 147, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 601, + 147, + 615 + ], + "score": 1.0, + "content": "equation:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 104, + 579, + 505, + 615 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 625, + 462, + 661 + ], + "lines": [ + { + "bbox": [ + 147, + 625, + 462, + 661 + ], + "spans": [ + { + "bbox": [ + 147, + 625, + 462, + 661 + ], + "score": 0.94, + "content": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) } - \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) }", + "type": "interline_equation", + "image_path": "813b870fe15e38f3325a5a02b85ccc357d691c80b99b6fa649998cfc632f0aeb.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 147, + 625, + 462, + 637.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 147, + 637.0, + 462, + 649.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 147, + 649.0, + 462, + 661.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 667, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 340, + 680 + ], + "score": 1.0, + "content": "Like before, we may use the integration bounds to bound", + "type": "text" + }, + { + "bbox": [ + 340, + 670, + 347, + 678 + ], + "score": 0.75, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 667, + 402, + 680 + ], + "score": 1.0, + "content": ". Plugging in", + "type": "text" + }, + { + "bbox": [ + 402, + 668, + 439, + 681 + ], + "score": 0.92, + "content": "\\sigma = s \\sigma _ { j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "for all integrals", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 466, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 466, + 692 + ], + "score": 1.0, + "content": "decreases the left-hand side and increases the right-hand side, leading us to the following:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 667, + 505, + 692 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 703, + 416, + 735 + ], + "lines": [ + { + "bbox": [ + 194, + 703, + 416, + 735 + ], + "spans": [ + { + "bbox": [ + 194, + 703, + 416, + 735 + ], + "score": 0.93, + "content": "\\frac { r } { r - s } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le \\big ( \\frac { r } { 1 - s } - \\frac { r } { r - s } \\big ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }", + "type": "interline_equation", + "image_path": "a3eb3dee43b36a19272b64864fcdf199a5779479c58a832565c195196b45b2dd.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 194, + 703, + 416, + 719.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 194, + 719.0, + 416, + 735.0 + ], + "spans": [], + "index": 35 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 289, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 290, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 290, + 97 + ], + "score": 1.0, + "content": "Rearranging, we get the following inequality:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 107, + 382, + 140 + ], + "lines": [ + { + "bbox": [ + 229, + 107, + 382, + 140 + ], + "spans": [ + { + "bbox": [ + 229, + 107, + 382, + 140 + ], + "score": 0.93, + "content": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\leq \\frac { r - 1 } { 1 - s } \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }", + "type": "interline_equation", + "image_path": "1d24111e000c0abc2af85b0afad77e68d6ed714d550b4c053e14679c34b80c2c.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 107, + 382, + 123.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 229, + 123.5, + 382, + 140.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 394, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 148, + 395, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 324, + 162 + ], + "score": 1.0, + "content": "We now solve the integrals for the different cases. For", + "type": "text" + }, + { + "bbox": [ + 325, + 150, + 353, + 159 + ], + "score": 0.92, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 148, + 395, + 162 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 174, + 416, + 199 + ], + "lines": [ + { + "bbox": [ + 194, + 174, + 416, + 199 + ], + "spans": [ + { + "bbox": [ + 194, + 174, + 416, + 199 + ], + "score": 0.9, + "content": "\\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\leq \\frac { r - 1 } { 1 - s } \\Bigl ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\Bigr )", + "type": "interline_equation", + "image_path": "29ea47288f5b6c5a3f43f605c76489c0d20bc983b0b3f4d1823726fb8f42f9d1.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 194, + 174, + 416, + 199 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 207, + 299, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 299, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 197, + 221 + ], + "score": 1.0, + "content": "Rearranging to isolate", + "type": "text" + }, + { + "bbox": [ + 197, + 210, + 208, + 219 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 207, + 299, + 221 + ], + "score": 1.0, + "content": ", we get our condition:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 233, + 344, + 258 + ], + "lines": [ + { + "bbox": [ + 266, + 233, + 344, + 258 + ], + "spans": [ + { + "bbox": [ + 266, + 233, + 344, + 258 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( \\frac { s } { r f } \\big ) ^ { \\frac { 1 - s } { r - 1 } }", + "type": "interline_equation", + "image_path": "f78019dc3ba334eb806b7a9a5fb6400862bf625da840cbab9b9df16d4244f0df.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 266, + 233, + 344, + 258 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 317, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 318, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 153, + 280 + ], + "score": 1.0, + "content": "Finally, for", + "type": "text" + }, + { + "bbox": [ + 153, + 267, + 182, + 277 + ], + "score": 0.91, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 264, + 318, + 280 + ], + "score": 1.0, + "content": ", we solve the integrals to give us:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 291, + 441, + 320 + ], + "lines": [ + { + "bbox": [ + 169, + 291, + 441, + 320 + ], + "spans": [ + { + "bbox": [ + 169, + 291, + 441, + 320 + ], + "score": 0.91, + "content": "\\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq \\frac { r - 1 } { 1 - s } \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )", + "type": "interline_equation", + "image_path": "1df08ba2c837b07f2d05916c13fad65fc1ca43d84edb1061197795569d1659ac.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 169, + 291, + 441, + 320 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 298, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 299, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 197, + 342 + ], + "score": 1.0, + "content": "Rearranging to isolate", + "type": "text" + }, + { + "bbox": [ + 197, + 331, + 208, + 339 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 327, + 299, + 342 + ], + "score": 1.0, + "content": ", we get our condition:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 353, + 353, + 380 + ], + "lines": [ + { + "bbox": [ + 258, + 353, + 353, + 380 + ], + "spans": [ + { + "bbox": [ + 258, + 353, + 353, + 380 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } }", + "type": "interline_equation", + "image_path": "c400becfc7d03fa51e6ed901fe1b391c865f5df9d121b21a818720e762fcbfb5.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 258, + 353, + 353, + 380 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 153, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 154, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 154, + 405 + ], + "score": 1.0, + "content": "SUMMARY", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 413, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 159, + 426 + ], + "score": 1.0, + "content": "For a given", + "type": "text" + }, + { + "bbox": [ + 160, + 414, + 170, + 424 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 413, + 446, + 426 + ], + "score": 1.0, + "content": ", we derived a sufficient condition and a necessary condition on", + "type": "text" + }, + { + "bbox": [ + 447, + 416, + 458, + 425 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 413, + 477, + 426 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 478, + 414, + 501, + 425 + ], + "score": 0.93, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 359, + 437 + ], + "score": 1.0, + "content": "incremental learning. The necessary and sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 360, + 426, + 370, + 435 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 425, + 505, + 437 + ], + "score": 1.0, + "content": ", which is the largest initialization", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 434, + 488, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 322, + 449 + ], + "score": 1.0, + "content": "value for which we see incremental learning (denoted", + "type": "text" + }, + { + "bbox": [ + 323, + 435, + 339, + 447 + ], + "score": 0.87, + "content": "\\sigma _ { 0 } ^ { t h \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 434, + 488, + 449 + ], + "score": 1.0, + "content": "), is between the two derived bounds.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 453, + 465 + ], + "score": 1.0, + "content": "The precise bounds can possibly be improved a bit, but the asymptotic dependence on", + "type": "text" + }, + { + "bbox": [ + 453, + 455, + 460, + 462 + ], + "score": 0.71, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "is the crux", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 276, + 476 + ], + "score": 1.0, + "content": "of the matter, showing the dependence on", + "type": "text" + }, + { + "bbox": [ + 277, + 466, + 283, + 473 + ], + "score": 0.64, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "changes with depth with a substantial difference when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 474, + 365, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 234, + 487 + ], + "score": 1.0, + "content": "we move from shallow models", + "type": "text" + }, + { + "bbox": [ + 235, + 474, + 264, + 485 + ], + "score": 0.85, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 474, + 330, + 487 + ], + "score": 1.0, + "content": ") to deeper ones", + "type": "text" + }, + { + "bbox": [ + 330, + 474, + 361, + 486 + ], + "score": 0.86, + "content": "( N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 474, + 365, + 487 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 519, + 244, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 245, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 245, + 534 + ], + "score": 1.0, + "content": "B PROOF OF THEOREM 3", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 506, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 504, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 225, + 558 + ], + "score": 1.0, + "content": "Theorem. Given two values", + "type": "text" + }, + { + "bbox": [ + 225, + 547, + 250, + 558 + ], + "score": 0.9, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 543, + 452, + 558 + ], + "score": 1.0, + "content": "of a depth-2 toy linear model as in (1), such that", + "type": "text" + }, + { + "bbox": [ + 452, + 542, + 504, + 561 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 342, + 575 + ], + "score": 1.0, + "content": "and the model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 342, + 560, + 383, + 574 + ], + "score": 0.92, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 560, + 401, + 575 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 402, + 560, + 443, + 574 + ], + "score": 0.92, + "content": "f \\in ( \\frac { 3 } { 4 } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 560, + 506, + 575 + ], + "score": 1.0, + "content": ", and assuming", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 571, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 107, + 572, + 152, + 586 + ], + "score": 0.93, + "content": "\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 571, + 451, + 586 + ], + "score": 1.0, + "content": ", and assuming we optimize with gradient descent with a learning rate", + "type": "text" + }, + { + "bbox": [ + 452, + 572, + 488, + 587 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 571, + 505, + 586 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 107, + 586, + 172, + 600 + ], + "score": 0.91, + "content": "c < 2 ( { \\sqrt { 2 } } - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 586, + 192, + 600 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 192, + 588, + 204, + 600 + ], + "score": 0.89, + "content": "\\sigma _ { 1 } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 586, + 287, + 600 + ], + "score": 1.0, + "content": "the largest value of", + "type": "text" + }, + { + "bbox": [ + 288, + 588, + 299, + 598 + ], + "score": 0.85, + "content": "\\sigma ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 586, + 505, + 600 + ], + "score": 1.0, + "content": ", then the largest initialization value for which the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 596, + 507, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 244, + 613 + ], + "score": 1.0, + "content": "learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 245, + 599, + 268, + 611 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 596, + 358, + 613 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 358, + 598, + 372, + 611 + ], + "score": 0.9, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 596, + 507, + 613 + ], + "score": 1.0, + "content": ", is lower and upper bounded in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 609, + 183, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 183, + 623 + ], + "score": 1.0, + "content": "the following way:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 635, + 443, + 662 + ], + "lines": [ + { + "bbox": [ + 166, + 635, + 443, + 662 + ], + "spans": [ + { + "bbox": [ + 166, + 635, + 443, + 662 + ], + "score": 0.93, + "content": "\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }", + "type": "interline_equation", + "image_path": "f989af24e020caa89a073422b6ed1af5eea8a8a77a43f865802bb8988011016b.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 166, + 635, + 443, + 662 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 671, + 233, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 234, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 135, + 684 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 672, + 144, + 681 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 669, + 162, + 684 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 672, + 171, + 681 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 669, + 234, + 684 + ], + "score": 1.0, + "content": "are defined as:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 697, + 443, + 735 + ], + "lines": [ + { + "bbox": [ + 167, + 697, + 443, + 735 + ], + "spans": [ + { + "bbox": [ + 167, + 697, + 443, + 735 + ], + "score": 0.93, + "content": "A = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "bf85fc793530d30b5a4cd608c3212d1a16361d8dd660a5771a094371a743ba85.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 167, + 697, + 443, + 709.6666666666666 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 167, + 709.6666666666666, + 443, + 722.3333333333333 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 167, + 722.3333333333333, + 443, + 734.9999999999999 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 491, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 496, + 493, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 496, + 493, + 505, + 502 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 289, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 290, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 290, + 97 + ], + "score": 1.0, + "content": "Rearranging, we get the following inequality:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 80, + 290, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 107, + 382, + 140 + ], + "lines": [ + { + "bbox": [ + 229, + 107, + 382, + 140 + ], + "spans": [ + { + "bbox": [ + 229, + 107, + 382, + 140 + ], + "score": 0.93, + "content": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\leq \\frac { r - 1 } { 1 - s } \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }", + "type": "interline_equation", + "image_path": "1d24111e000c0abc2af85b0afad77e68d6ed714d550b4c053e14679c34b80c2c.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 107, + 382, + 123.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 229, + 123.5, + 382, + 140.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 394, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 148, + 395, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 324, + 162 + ], + "score": 1.0, + "content": "We now solve the integrals for the different cases. For", + "type": "text" + }, + { + "bbox": [ + 325, + 150, + 353, + 159 + ], + "score": 0.92, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 148, + 395, + 162 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 148, + 395, + 162 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 174, + 416, + 199 + ], + "lines": [ + { + "bbox": [ + 194, + 174, + 416, + 199 + ], + "spans": [ + { + "bbox": [ + 194, + 174, + 416, + 199 + ], + "score": 0.9, + "content": "\\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\leq \\frac { r - 1 } { 1 - s } \\Bigl ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\Bigr )", + "type": "interline_equation", + "image_path": "29ea47288f5b6c5a3f43f605c76489c0d20bc983b0b3f4d1823726fb8f42f9d1.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 194, + 174, + 416, + 199 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 207, + 299, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 299, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 197, + 221 + ], + "score": 1.0, + "content": "Rearranging to isolate", + "type": "text" + }, + { + "bbox": [ + 197, + 210, + 208, + 219 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 207, + 299, + 221 + ], + "score": 1.0, + "content": ", we get our condition:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 207, + 299, + 221 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 233, + 344, + 258 + ], + "lines": [ + { + "bbox": [ + 266, + 233, + 344, + 258 + ], + "spans": [ + { + "bbox": [ + 266, + 233, + 344, + 258 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( \\frac { s } { r f } \\big ) ^ { \\frac { 1 - s } { r - 1 } }", + "type": "interline_equation", + "image_path": "f78019dc3ba334eb806b7a9a5fb6400862bf625da840cbab9b9df16d4244f0df.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 266, + 233, + 344, + 258 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 317, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 318, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 153, + 280 + ], + "score": 1.0, + "content": "Finally, for", + "type": "text" + }, + { + "bbox": [ + 153, + 267, + 182, + 277 + ], + "score": 0.91, + "content": "N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 264, + 318, + 280 + ], + "score": 1.0, + "content": ", we solve the integrals to give us:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 264, + 318, + 280 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 291, + 441, + 320 + ], + "lines": [ + { + "bbox": [ + 169, + 291, + 441, + 320 + ], + "spans": [ + { + "bbox": [ + 169, + 291, + 441, + 320 + ], + "score": 0.91, + "content": "\\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq \\frac { r - 1 } { 1 - s } \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )", + "type": "interline_equation", + "image_path": "1df08ba2c837b07f2d05916c13fad65fc1ca43d84edb1061197795569d1659ac.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 169, + 291, + 441, + 320 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 298, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 299, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 197, + 342 + ], + "score": 1.0, + "content": "Rearranging to isolate", + "type": "text" + }, + { + "bbox": [ + 197, + 331, + 208, + 339 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 327, + 299, + 342 + ], + "score": 1.0, + "content": ", we get our condition:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 327, + 299, + 342 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 353, + 353, + 380 + ], + "lines": [ + { + "bbox": [ + 258, + 353, + 353, + 380 + ], + "spans": [ + { + "bbox": [ + 258, + 353, + 353, + 380 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } }", + "type": "interline_equation", + "image_path": "c400becfc7d03fa51e6ed901fe1b391c865f5df9d121b21a818720e762fcbfb5.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 258, + 353, + 353, + 380 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 393, + 153, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 154, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 154, + 405 + ], + "score": 1.0, + "content": "SUMMARY", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 393, + 154, + 405 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 413, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 159, + 426 + ], + "score": 1.0, + "content": "For a given", + "type": "text" + }, + { + "bbox": [ + 160, + 414, + 170, + 424 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 413, + 446, + 426 + ], + "score": 1.0, + "content": ", we derived a sufficient condition and a necessary condition on", + "type": "text" + }, + { + "bbox": [ + 447, + 416, + 458, + 425 + ], + "score": 0.84, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 413, + 477, + 426 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 478, + 414, + 501, + 425 + ], + "score": 0.93, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 359, + 437 + ], + "score": 1.0, + "content": "incremental learning. The necessary and sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 360, + 426, + 370, + 435 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 425, + 505, + 437 + ], + "score": 1.0, + "content": ", which is the largest initialization", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 434, + 488, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 322, + 449 + ], + "score": 1.0, + "content": "value for which we see incremental learning (denoted", + "type": "text" + }, + { + "bbox": [ + 323, + 435, + 339, + 447 + ], + "score": 0.87, + "content": "\\sigma _ { 0 } ^ { t h \\cdot }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 434, + 488, + 449 + ], + "score": 1.0, + "content": "), is between the two derived bounds.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 413, + 505, + 449 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 453, + 465 + ], + "score": 1.0, + "content": "The precise bounds can possibly be improved a bit, but the asymptotic dependence on", + "type": "text" + }, + { + "bbox": [ + 453, + 455, + 460, + 462 + ], + "score": 0.71, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "is the crux", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 276, + 476 + ], + "score": 1.0, + "content": "of the matter, showing the dependence on", + "type": "text" + }, + { + "bbox": [ + 277, + 466, + 283, + 473 + ], + "score": 0.64, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "changes with depth with a substantial difference when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 474, + 365, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 234, + 487 + ], + "score": 1.0, + "content": "we move from shallow models", + "type": "text" + }, + { + "bbox": [ + 235, + 474, + 264, + 485 + ], + "score": 0.85, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 474, + 330, + 487 + ], + "score": 1.0, + "content": ") to deeper ones", + "type": "text" + }, + { + "bbox": [ + 330, + 474, + 361, + 486 + ], + "score": 0.86, + "content": "( N \\geq 3", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 474, + 365, + 487 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 451, + 505, + 487 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 519, + 244, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 245, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 245, + 534 + ], + "score": 1.0, + "content": "B PROOF OF THEOREM 3", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 506, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 504, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 225, + 558 + ], + "score": 1.0, + "content": "Theorem. Given two values", + "type": "text" + }, + { + "bbox": [ + 225, + 547, + 250, + 558 + ], + "score": 0.9, + "content": "\\sigma _ { i } , \\sigma _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 543, + 452, + 558 + ], + "score": 1.0, + "content": "of a depth-2 toy linear model as in (1), such that", + "type": "text" + }, + { + "bbox": [ + 452, + 542, + 504, + 561 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 342, + 575 + ], + "score": 1.0, + "content": "and the model is initialized as in (3), and given two scalars", + "type": "text" + }, + { + "bbox": [ + 342, + 560, + 383, + 574 + ], + "score": 0.92, + "content": "s \\in ( 0 , \\frac { 1 } { 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 560, + 401, + 575 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 402, + 560, + 443, + 574 + ], + "score": 0.92, + "content": "f \\in ( \\frac { 3 } { 4 } , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 560, + 506, + 575 + ], + "score": 1.0, + "content": ", and assuming", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 571, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 107, + 572, + 152, + 586 + ], + "score": 0.93, + "content": "\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 571, + 451, + 586 + ], + "score": 1.0, + "content": ", and assuming we optimize with gradient descent with a learning rate", + "type": "text" + }, + { + "bbox": [ + 452, + 572, + 488, + 587 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 571, + 505, + 586 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 107, + 586, + 172, + 600 + ], + "score": 0.91, + "content": "c < 2 ( { \\sqrt { 2 } } - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 586, + 192, + 600 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 192, + 588, + 204, + 600 + ], + "score": 0.89, + "content": "\\sigma _ { 1 } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 586, + 287, + 600 + ], + "score": 1.0, + "content": "the largest value of", + "type": "text" + }, + { + "bbox": [ + 288, + 588, + 299, + 598 + ], + "score": 0.85, + "content": "\\sigma ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 586, + 505, + 600 + ], + "score": 1.0, + "content": ", then the largest initialization value for which the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 596, + 507, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 244, + 613 + ], + "score": 1.0, + "content": "learning phases of the values are", + "type": "text" + }, + { + "bbox": [ + 245, + 599, + 268, + 611 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 596, + 358, + 613 + ], + "score": 1.0, + "content": "-incremental, denoted", + "type": "text" + }, + { + "bbox": [ + 358, + 598, + 372, + 611 + ], + "score": 0.9, + "content": "\\sigma _ { 0 } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 596, + 507, + 613 + ], + "score": 1.0, + "content": ", is lower and upper bounded in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 609, + 183, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 183, + 623 + ], + "score": 1.0, + "content": "the following way:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 542, + 507, + 623 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 635, + 443, + 662 + ], + "lines": [ + { + "bbox": [ + 166, + 635, + 443, + 662 + ], + "spans": [ + { + "bbox": [ + 166, + 635, + 443, + 662 + ], + "score": 0.93, + "content": "\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }", + "type": "interline_equation", + "image_path": "f989af24e020caa89a073422b6ed1af5eea8a8a77a43f865802bb8988011016b.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 166, + 635, + 443, + 662 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 671, + 233, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 234, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 135, + 684 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 672, + 144, + 681 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 669, + 162, + 684 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 672, + 171, + 681 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 669, + 234, + 684 + ], + "score": 1.0, + "content": "are defined as:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 669, + 234, + 684 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 697, + 443, + 735 + ], + "lines": [ + { + "bbox": [ + 167, + 697, + 443, + 735 + ], + "spans": [ + { + "bbox": [ + 167, + 697, + 443, + 735 + ], + "score": 0.93, + "content": "A = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "bf85fc793530d30b5a4cd608c3212d1a16361d8dd660a5771a094371a743ba85.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 167, + 697, + 443, + 709.6666666666666 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 167, + 709.6666666666666, + 443, + 722.3333333333333 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 167, + 722.3333333333333, + 443, + 734.9999999999999 + ], + "spans": [], + "index": 29 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 318, + 94 + ], + "score": 1.0, + "content": "Proof. To show our result for gradient descent and", + "type": "text" + }, + { + "bbox": [ + 318, + 83, + 351, + 93 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 82, + 505, + 94 + ], + "score": 1.0, + "content": ", we build on the proof techniques of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 471, + 106 + ], + "score": 1.0, + "content": "theorem 3 of Gidel et al. (2019). We start by deriving the recurrence relation for the values", + "type": "text" + }, + { + "bbox": [ + 471, + 93, + 489, + 106 + ], + "score": 0.91, + "content": "\\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 189, + 117 + ], + "score": 1.0, + "content": "general depth, when", + "type": "text" + }, + { + "bbox": [ + 190, + 106, + 195, + 114 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 105, + 388, + 117 + ], + "score": 1.0, + "content": "now stands for the iteration. Remembering that", + "type": "text" + }, + { + "bbox": [ + 388, + 105, + 424, + 117 + ], + "score": 0.92, + "content": "w _ { i } ^ { n } = \\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 105, + 505, + 117 + ], + "score": 1.0, + "content": ", we write down the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 212, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 185, + 128 + ], + "score": 1.0, + "content": "gradient update for", + "type": "text" + }, + { + "bbox": [ + 185, + 115, + 208, + 127 + ], + "score": 0.92, + "content": "w _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 115, + 212, + 128 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "interline_equation", + "bbox": [ + 124, + 146, + 486, + 171 + ], + "lines": [ + { + "bbox": [ + 124, + 146, + 486, + 171 + ], + "spans": [ + { + "bbox": [ + 124, + 146, + 486, + 171 + ], + "score": 0.92, + "content": "w _ { i } ( t + 1 ) = w _ { i } ( t ) + \\eta \\frac { 1 } { N } w _ { i } ( t ) ^ { N - 1 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "9772893184c9e2aa969e20cefc687450254670e52098c17ec1a940af2aebee97.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 124, + 146, + 486, + 171 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 182, + 412, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 412, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 139, + 196 + ], + "score": 1.0, + "content": "Raising", + "type": "text" + }, + { + "bbox": [ + 140, + 182, + 162, + 194 + ], + "score": 0.93, + "content": "w _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 181, + 189, + 196 + ], + "score": 1.0, + "content": "to the", + "type": "text" + }, + { + "bbox": [ + 189, + 183, + 198, + 192 + ], + "score": 0.71, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 181, + 373, + 196 + ], + "score": 1.0, + "content": "th power, we get the gradient update for the", + "type": "text" + }, + { + "bbox": [ + 374, + 185, + 381, + 192 + ], + "score": 0.81, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 181, + 412, + 196 + ], + "score": 1.0, + "content": "values:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 213, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 111, + 213, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 111, + 213, + 505, + 238 + ], + "score": 0.89, + "content": "r _ { i } ( t + 1 ) = \\bigg ( \\sqrt [ r ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } \\bigg ) ^ { N } = \\sigma _ { i } ( t ) \\Big ( 1 + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { N }", + "type": "interline_equation", + "image_path": "17860e6e9702f9121662d68c4defc1a0556ff5952ae0f2e2b7f801965256cf26.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 213, + 505, + 238 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "Next, we will prove a simple lemma which gives us the maximal learning rate we will consider", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "for the analysis, for which there is no overshooting (the values don’t grow larger than the optimal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 272, + 141, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 141, + 288 + ], + "score": 1.0, + "content": "values).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 323, + 314 + ], + "score": 1.0, + "content": "Lemma 1. For the gradient update in (9), assuming", + "type": "text" + }, + { + "bbox": [ + 324, + 298, + 372, + 311 + ], + "score": 0.89, + "content": "\\sigma _ { i } ( 0 ) < \\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 289, + 385, + 314 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 386, + 291, + 450, + 315 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\eta \\leq \\left( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\right) ^ { 2 - \\frac { 2 } { N } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 289, + 470, + 314 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 470, + 298, + 501, + 310 + ], + "score": 0.76, + "content": "N \\geq 2", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 289, + 506, + 314 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 311, + 131, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 311, + 131, + 326 + ], + "score": 1.0, + "content": "then:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 340, + 339, + 354 + ], + "lines": [ + { + "bbox": [ + 271, + 340, + 339, + 354 + ], + "spans": [ + { + "bbox": [ + 271, + 340, + 339, + 354 + ], + "score": 0.82, + "content": "\\forall t : \\ \\sigma _ { i } ( t ) \\leq \\sigma _ { i } ^ { * }", + "type": "interline_equation", + "image_path": "a5f408efe5ec5d8ec43a138f796555e1c81732745262dfee382c4d5e4c585170.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 271, + 340, + 339, + 354 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 332, + 389 + ], + "lines": [ + { + "bbox": [ + 103, + 365, + 333, + 390 + ], + "spans": [ + { + "bbox": [ + 103, + 366, + 185, + 390 + ], + "score": 1.0, + "content": "Proof. Plugging in", + "type": "text" + }, + { + "bbox": [ + 185, + 365, + 250, + 389 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\eta = c \\bigg ( \\frac { 1 } { \\sigma _ { i } ^ { * } } \\bigg ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 371, + 267, + 385 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 267, + 374, + 290, + 384 + ], + "score": 0.87, + "content": "c \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 371, + 333, + 385 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 404, + 491, + 439 + ], + "lines": [ + { + "bbox": [ + 119, + 404, + 491, + 439 + ], + "spans": [ + { + "bbox": [ + 119, + 404, + 491, + 439 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\Big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) \\Big ) ^ { N } \\leq \\sigma _ { i } ( t ) e ^ { \\left( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\right) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) }", + "type": "interline_equation", + "image_path": "c42971020f57e61baf48bc9dc18de1d65dd7ce01fb564abf820ad4d07f8eeb05.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 119, + 404, + 491, + 415.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 119, + 415.6666666666667, + 491, + 427.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 119, + 427.33333333333337, + 491, + 439.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 450, + 339, + 464 + ], + "lines": [ + { + "bbox": [ + 104, + 443, + 339, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 186, + 469 + ], + "score": 1.0, + "content": "Defining ri = σiσ∗i a", + "type": "text" + }, + { + "bbox": [ + 178, + 450, + 286, + 463 + ], + "score": 1.0, + "content": "nd dividing both sides by", + "type": "text" + }, + { + "bbox": [ + 286, + 451, + 298, + 462 + ], + "score": 0.9, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 450, + 339, + 463 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 475, + 376, + 494 + ], + "lines": [ + { + "bbox": [ + 234, + 475, + 376, + 494 + ], + "spans": [ + { + "bbox": [ + 234, + 475, + 376, + 494 + ], + "score": 0.91, + "content": "r _ { i } ( t + 1 ) \\leq r _ { i } ( t ) e ^ { r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) }", + "type": "interline_equation", + "image_path": "7d5aa12762aef96b649c1121280d46213f820f4d300cf9dc461f87a38a34882e.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 234, + 475, + 376, + 494 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 101, + 498, + 509, + 524 + ], + "spans": [ + { + "bbox": [ + 101, + 498, + 236, + 524 + ], + "score": 1.0, + "content": "It is enough to show that for any", + "type": "text" + }, + { + "bbox": [ + 236, + 507, + 279, + 518 + ], + "score": 0.91, + "content": "0 \\leq r \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 498, + 334, + 524 + ], + "score": 1.0, + "content": ", we have that", + "type": "text" + }, + { + "bbox": [ + 334, + 502, + 405, + 518 + ], + "score": 0.95, + "content": "r e ^ { r ^ { 1 - \\frac { 2 } { N } } ( 1 - r ) } \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 498, + 509, + 524 + ], + "score": 1.0, + "content": ", as over-shooting occurs", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 131, + 531 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 518, + 170, + 530 + ], + "score": 0.93, + "content": "r _ { i } ( \\bar { t } ) > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 516, + 369, + 531 + ], + "score": 1.0, + "content": ". Indeed, this function is monotonic increasing in", + "type": "text" + }, + { + "bbox": [ + 370, + 519, + 414, + 529 + ], + "score": 0.89, + "content": "0 \\leq r \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "(since the exponent is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 529, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 243, + 541 + ], + "score": 1.0, + "content": "non negative), and equals 1 when", + "type": "text" + }, + { + "bbox": [ + 244, + 530, + 268, + 539 + ], + "score": 0.9, + "content": "r = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 529, + 298, + 541 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 299, + 530, + 324, + 540 + ], + "score": 0.9, + "content": "r = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 529, + 506, + 541 + ], + "score": 1.0, + "content": "is a fixed point and no iteration that starts at", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 131, + 550 + ], + "score": 0.88, + "content": "r < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 540, + 200, + 552 + ], + "score": 1.0, + "content": "can cross 1, then", + "type": "text" + }, + { + "bbox": [ + 201, + 540, + 240, + 552 + ], + "score": 0.93, + "content": "r _ { i } ( t ) \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 540, + 271, + 552 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 272, + 541, + 276, + 550 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 540, + 385, + 552 + ], + "score": 1.0, + "content": ". This concludes our proof.", + "type": "text" + }, + { + "bbox": [ + 495, + 541, + 505, + 550 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Under this choice of learning rate, we can now obtain our incremental learning results for gradient", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 163, + 588 + ], + "score": 1.0, + "content": "descent when", + "type": "text" + }, + { + "bbox": [ + 164, + 576, + 193, + 586 + ], + "score": 0.89, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 575, + 322, + 588 + ], + "score": 1.0, + "content": ". Our strategy will be bounding", + "type": "text" + }, + { + "bbox": [ + 322, + 576, + 343, + 588 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "from below and above, which will give", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 238, + 599 + ], + "score": 1.0, + "content": "us a lower and upper bound for", + "type": "text" + }, + { + "bbox": [ + 239, + 587, + 266, + 598 + ], + "score": 0.92, + "content": "t _ { \\alpha } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 587, + 505, + 599 + ], + "score": 1.0, + "content": ". Once we have these bounds, we will be able to describe", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 597, + 489, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 289, + 610 + ], + "score": 1.0, + "content": "either a necessary or a sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 290, + 599, + 301, + 609 + ], + "score": 0.86, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 597, + 489, + 610 + ], + "score": 1.0, + "content": "for incremental learning, similar to theorem 2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 227, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 227, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 185, + 626 + ], + "score": 1.0, + "content": "The update rule for", + "type": "text" + }, + { + "bbox": [ + 186, + 614, + 214, + 624 + ], + "score": 0.91, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 613, + 227, + 626 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 639, + 390, + 663 + ], + "lines": [ + { + "bbox": [ + 220, + 639, + 390, + 663 + ], + "spans": [ + { + "bbox": [ + 220, + 639, + 390, + 663 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { 1 } { 2 } \\eta ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { 2 }", + "type": "interline_equation", + "image_path": "cd05808cb73e514a483f4c120952578fd482e56fe54ba33c984b736407f3fce0.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 220, + 639, + 390, + 663 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 672, + 506, + 700 + ], + "lines": [ + { + "bbox": [ + 102, + 668, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 102, + 668, + 175, + 693 + ], + "score": 1.0, + "content": "Next, we plug in", + "type": "text" + }, + { + "bbox": [ + 175, + 675, + 206, + 690 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\eta = \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 668, + 222, + 693 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 222, + 673, + 302, + 687 + ], + "score": 0.93, + "content": "c < 2 ( \\sqrt { 2 } - 1 ) < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 668, + 349, + 693 + ], + "score": 1.0, + "content": "and denote", + "type": "text" + }, + { + "bbox": [ + 349, + 672, + 404, + 690 + ], + "score": 0.94, + "content": "\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\leq 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 668, + 422, + 693 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 422, + 672, + 494, + 690 + ], + "score": 0.94, + "content": "\\begin{array} { r } { r _ { i } ( t ) = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\le 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 668, + 507, + 693 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 687, + 124, + 699 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 124, + 699 + ], + "score": 1.0, + "content": "get:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 712, + 389, + 736 + ], + "lines": [ + { + "bbox": [ + 221, + 712, + 389, + 736 + ], + "spans": [ + { + "bbox": [ + 221, + 712, + 389, + 736 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { 2 }", + "type": "interline_equation", + "image_path": "1cf77218caa89b2240727d215e87107e30d27bd14515fd06a4c85302fd063f79.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 221, + 712, + 389, + 736 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 318, + 94 + ], + "score": 1.0, + "content": "Proof. To show our result for gradient descent and", + "type": "text" + }, + { + "bbox": [ + 318, + 83, + 351, + 93 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 82, + 505, + 94 + ], + "score": 1.0, + "content": ", we build on the proof techniques of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 471, + 106 + ], + "score": 1.0, + "content": "theorem 3 of Gidel et al. (2019). We start by deriving the recurrence relation for the values", + "type": "text" + }, + { + "bbox": [ + 471, + 93, + 489, + 106 + ], + "score": 0.91, + "content": "\\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 189, + 117 + ], + "score": 1.0, + "content": "general depth, when", + "type": "text" + }, + { + "bbox": [ + 190, + 106, + 195, + 114 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 105, + 388, + 117 + ], + "score": 1.0, + "content": "now stands for the iteration. Remembering that", + "type": "text" + }, + { + "bbox": [ + 388, + 105, + 424, + 117 + ], + "score": 0.92, + "content": "w _ { i } ^ { n } = \\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 105, + 505, + 117 + ], + "score": 1.0, + "content": ", we write down the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 212, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 185, + 128 + ], + "score": 1.0, + "content": "gradient update for", + "type": "text" + }, + { + "bbox": [ + 185, + 115, + 208, + 127 + ], + "score": 0.92, + "content": "w _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 115, + 212, + 128 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 505, + 128 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 124, + 146, + 486, + 171 + ], + "lines": [ + { + "bbox": [ + 124, + 146, + 486, + 171 + ], + "spans": [ + { + "bbox": [ + 124, + 146, + 486, + 171 + ], + "score": 0.92, + "content": "w _ { i } ( t + 1 ) = w _ { i } ( t ) + \\eta \\frac { 1 } { N } w _ { i } ( t ) ^ { N - 1 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "9772893184c9e2aa969e20cefc687450254670e52098c17ec1a940af2aebee97.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 124, + 146, + 486, + 171 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 182, + 412, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 412, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 139, + 196 + ], + "score": 1.0, + "content": "Raising", + "type": "text" + }, + { + "bbox": [ + 140, + 182, + 162, + 194 + ], + "score": 0.93, + "content": "w _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 181, + 189, + 196 + ], + "score": 1.0, + "content": "to the", + "type": "text" + }, + { + "bbox": [ + 189, + 183, + 198, + 192 + ], + "score": 0.71, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 181, + 373, + 196 + ], + "score": 1.0, + "content": "th power, we get the gradient update for the", + "type": "text" + }, + { + "bbox": [ + 374, + 185, + 381, + 192 + ], + "score": 0.81, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 181, + 412, + 196 + ], + "score": 1.0, + "content": "values:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 181, + 412, + 196 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 213, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 111, + 213, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 111, + 213, + 505, + 238 + ], + "score": 0.89, + "content": "r _ { i } ( t + 1 ) = \\bigg ( \\sqrt [ r ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } \\bigg ) ^ { N } = \\sigma _ { i } ( t ) \\Big ( 1 + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { N }", + "type": "interline_equation", + "image_path": "17860e6e9702f9121662d68c4defc1a0556ff5952ae0f2e2b7f801965256cf26.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 213, + 505, + 238 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "Next, we will prove a simple lemma which gives us the maximal learning rate we will consider", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "for the analysis, for which there is no overshooting (the values don’t grow larger than the optimal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 272, + 141, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 141, + 288 + ], + "score": 1.0, + "content": "values).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 252, + 505, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 323, + 314 + ], + "score": 1.0, + "content": "Lemma 1. For the gradient update in (9), assuming", + "type": "text" + }, + { + "bbox": [ + 324, + 298, + 372, + 311 + ], + "score": 0.89, + "content": "\\sigma _ { i } ( 0 ) < \\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 289, + 385, + 314 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 386, + 291, + 450, + 315 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\eta \\leq \\left( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\right) ^ { 2 - \\frac { 2 } { N } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 289, + 470, + 314 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 470, + 298, + 501, + 310 + ], + "score": 0.76, + "content": "N \\geq 2", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 289, + 506, + 314 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 311, + 131, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 311, + 131, + 326 + ], + "score": 1.0, + "content": "then:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 289, + 506, + 326 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 340, + 339, + 354 + ], + "lines": [ + { + "bbox": [ + 271, + 340, + 339, + 354 + ], + "spans": [ + { + "bbox": [ + 271, + 340, + 339, + 354 + ], + "score": 0.82, + "content": "\\forall t : \\ \\sigma _ { i } ( t ) \\leq \\sigma _ { i } ^ { * }", + "type": "interline_equation", + "image_path": "a5f408efe5ec5d8ec43a138f796555e1c81732745262dfee382c4d5e4c585170.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 271, + 340, + 339, + 354 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 332, + 389 + ], + "lines": [ + { + "bbox": [ + 103, + 365, + 333, + 390 + ], + "spans": [ + { + "bbox": [ + 103, + 366, + 185, + 390 + ], + "score": 1.0, + "content": "Proof. Plugging in", + "type": "text" + }, + { + "bbox": [ + 185, + 365, + 250, + 389 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\eta = c \\bigg ( \\frac { 1 } { \\sigma _ { i } ^ { * } } \\bigg ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 371, + 267, + 385 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 267, + 374, + 290, + 384 + ], + "score": 0.87, + "content": "c \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 371, + 333, + 385 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 103, + 365, + 333, + 390 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 404, + 491, + 439 + ], + "lines": [ + { + "bbox": [ + 119, + 404, + 491, + 439 + ], + "spans": [ + { + "bbox": [ + 119, + 404, + 491, + 439 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\Big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) \\Big ) ^ { N } \\leq \\sigma _ { i } ( t ) e ^ { \\left( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\right) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) }", + "type": "interline_equation", + "image_path": "c42971020f57e61baf48bc9dc18de1d65dd7ce01fb564abf820ad4d07f8eeb05.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 119, + 404, + 491, + 415.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 119, + 415.6666666666667, + 491, + 427.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 119, + 427.33333333333337, + 491, + 439.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 450, + 339, + 464 + ], + "lines": [ + { + "bbox": [ + 104, + 443, + 339, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 186, + 469 + ], + "score": 1.0, + "content": "Defining ri = σiσ∗i a", + "type": "text" + }, + { + "bbox": [ + 178, + 450, + 286, + 463 + ], + "score": 1.0, + "content": "nd dividing both sides by", + "type": "text" + }, + { + "bbox": [ + 286, + 451, + 298, + 462 + ], + "score": 0.9, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 450, + 339, + 463 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 443, + 339, + 469 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 475, + 376, + 494 + ], + "lines": [ + { + "bbox": [ + 234, + 475, + 376, + 494 + ], + "spans": [ + { + "bbox": [ + 234, + 475, + 376, + 494 + ], + "score": 0.91, + "content": "r _ { i } ( t + 1 ) \\leq r _ { i } ( t ) e ^ { r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) }", + "type": "interline_equation", + "image_path": "7d5aa12762aef96b649c1121280d46213f820f4d300cf9dc461f87a38a34882e.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 234, + 475, + 376, + 494 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 101, + 498, + 509, + 524 + ], + "spans": [ + { + "bbox": [ + 101, + 498, + 236, + 524 + ], + "score": 1.0, + "content": "It is enough to show that for any", + "type": "text" + }, + { + "bbox": [ + 236, + 507, + 279, + 518 + ], + "score": 0.91, + "content": "0 \\leq r \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 498, + 334, + 524 + ], + "score": 1.0, + "content": ", we have that", + "type": "text" + }, + { + "bbox": [ + 334, + 502, + 405, + 518 + ], + "score": 0.95, + "content": "r e ^ { r ^ { 1 - \\frac { 2 } { N } } ( 1 - r ) } \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 498, + 509, + 524 + ], + "score": 1.0, + "content": ", as over-shooting occurs", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 131, + 531 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 518, + 170, + 530 + ], + "score": 0.93, + "content": "r _ { i } ( \\bar { t } ) > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 516, + 369, + 531 + ], + "score": 1.0, + "content": ". Indeed, this function is monotonic increasing in", + "type": "text" + }, + { + "bbox": [ + 370, + 519, + 414, + 529 + ], + "score": 0.89, + "content": "0 \\leq r \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "(since the exponent is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 529, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 243, + 541 + ], + "score": 1.0, + "content": "non negative), and equals 1 when", + "type": "text" + }, + { + "bbox": [ + 244, + 530, + 268, + 539 + ], + "score": 0.9, + "content": "r = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 529, + 298, + 541 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 299, + 530, + 324, + 540 + ], + "score": 0.9, + "content": "r = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 529, + 506, + 541 + ], + "score": 1.0, + "content": "is a fixed point and no iteration that starts at", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 131, + 550 + ], + "score": 0.88, + "content": "r < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 540, + 200, + 552 + ], + "score": 1.0, + "content": "can cross 1, then", + "type": "text" + }, + { + "bbox": [ + 201, + 540, + 240, + 552 + ], + "score": 0.93, + "content": "r _ { i } ( t ) \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 540, + 271, + 552 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 272, + 541, + 276, + 550 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 540, + 385, + 552 + ], + "score": 1.0, + "content": ". This concludes our proof.", + "type": "text" + }, + { + "bbox": [ + 495, + 541, + 505, + 550 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 101, + 498, + 509, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Under this choice of learning rate, we can now obtain our incremental learning results for gradient", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 163, + 588 + ], + "score": 1.0, + "content": "descent when", + "type": "text" + }, + { + "bbox": [ + 164, + 576, + 193, + 586 + ], + "score": 0.89, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 575, + 322, + 588 + ], + "score": 1.0, + "content": ". Our strategy will be bounding", + "type": "text" + }, + { + "bbox": [ + 322, + 576, + 343, + 588 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "from below and above, which will give", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 238, + 599 + ], + "score": 1.0, + "content": "us a lower and upper bound for", + "type": "text" + }, + { + "bbox": [ + 239, + 587, + 266, + 598 + ], + "score": 0.92, + "content": "t _ { \\alpha } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 587, + 505, + 599 + ], + "score": 1.0, + "content": ". Once we have these bounds, we will be able to describe", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 597, + 489, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 289, + 610 + ], + "score": 1.0, + "content": "either a necessary or a sufficient condition on", + "type": "text" + }, + { + "bbox": [ + 290, + 599, + 301, + 609 + ], + "score": 0.86, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 597, + 489, + 610 + ], + "score": 1.0, + "content": "for incremental learning, similar to theorem 2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 565, + 505, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 227, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 227, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 185, + 626 + ], + "score": 1.0, + "content": "The update rule for", + "type": "text" + }, + { + "bbox": [ + 186, + 614, + 214, + 624 + ], + "score": 0.91, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 613, + 227, + 626 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 613, + 227, + 626 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 639, + 390, + 663 + ], + "lines": [ + { + "bbox": [ + 220, + 639, + 390, + 663 + ], + "spans": [ + { + "bbox": [ + 220, + 639, + 390, + 663 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { 1 } { 2 } \\eta ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { 2 }", + "type": "interline_equation", + "image_path": "cd05808cb73e514a483f4c120952578fd482e56fe54ba33c984b736407f3fce0.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 220, + 639, + 390, + 663 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 672, + 506, + 700 + ], + "lines": [ + { + "bbox": [ + 102, + 668, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 102, + 668, + 175, + 693 + ], + "score": 1.0, + "content": "Next, we plug in", + "type": "text" + }, + { + "bbox": [ + 175, + 675, + 206, + 690 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\eta = \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 668, + 222, + 693 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 222, + 673, + 302, + 687 + ], + "score": 0.93, + "content": "c < 2 ( \\sqrt { 2 } - 1 ) < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 668, + 349, + 693 + ], + "score": 1.0, + "content": "and denote", + "type": "text" + }, + { + "bbox": [ + 349, + 672, + 404, + 690 + ], + "score": 0.94, + "content": "\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\leq 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 668, + 422, + 693 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 422, + 672, + 494, + 690 + ], + "score": 0.94, + "content": "\\begin{array} { r } { r _ { i } ( t ) = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\le 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 668, + 507, + 693 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 687, + 124, + 699 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 124, + 699 + ], + "score": 1.0, + "content": "get:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 102, + 668, + 507, + 699 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 712, + 389, + 736 + ], + "lines": [ + { + "bbox": [ + 221, + 712, + 389, + 736 + ], + "spans": [ + { + "bbox": [ + 221, + 712, + 389, + 736 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { 2 }", + "type": "interline_equation", + "image_path": "1cf77218caa89b2240727d215e87107e30d27bd14515fd06a4c85302fd063f79.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 221, + 712, + 389, + 736 + ], + "spans": [], + "index": 31 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 347, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 346, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "score": 1.0, + "content": "Following theorem 3 of Gidel et al. (2019), we bound", + "type": "text" + }, + { + "bbox": [ + 323, + 81, + 343, + 98 + ], + "score": 0.93, + "content": "\\frac { 1 } { \\sigma _ { i } ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 86, + 346, + 99 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 110, + 426, + 257 + ], + "lines": [ + { + "bbox": [ + 185, + 110, + 426, + 257 + ], + "spans": [ + { + "bbox": [ + 185, + 110, + 426, + 257 + ], + "score": 0.96, + "content": "\\begin{array} { r l } & { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { \\left( 1 + \\frac { \\varepsilon } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\right) ^ { 2 } } } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\left( 1 - r _ { i } ( t ) \\right) \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\bigl ( 1 - ( 1 - r _ { i } ( t ) ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) \\bigr ) } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } - ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "869f7f8a106e05ed77f77284a7b85cc8974315bd7f3d07ed533e53a5664702a0.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 110, + 426, + 128.375 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 185, + 128.375, + 426, + 146.75 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 185, + 146.75, + 426, + 165.125 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 185, + 165.125, + 426, + 183.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 185, + 183.5, + 426, + 201.875 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 185, + 201.875, + 426, + 220.25 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 185, + 220.25, + 426, + 238.625 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 185, + 238.625, + 426, + 257.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 503, + 290 + ], + "lines": [ + { + "bbox": [ + 104, + 262, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 297, + 282 + ], + "score": 1.0, + "content": "Where in the fourth line we use the inequality", + "type": "text" + }, + { + "bbox": [ + 297, + 263, + 394, + 279 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\frac { 1 } { 1 + x } \\geq 1 - x , \\forall x \\geq 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 262, + 491, + 282 + ], + "score": 1.0, + "content": ". We may now subtract", + "type": "text" + }, + { + "bbox": [ + 491, + 263, + 504, + 280 + ], + "score": 0.89, + "content": "\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 277, + 211, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 211, + 290 + ], + "score": 1.0, + "content": "from both sides to obtain:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 302, + 483, + 331 + ], + "lines": [ + { + "bbox": [ + 128, + 302, + 483, + 331 + ], + "spans": [ + { + "bbox": [ + 128, + 302, + 483, + 331 + ], + "score": 0.93, + "content": "\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\geq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) \\geq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) ^ { t }", + "type": "interline_equation", + "image_path": "398c960671dfcf23edd99904d0af487602e5cc56a57b02a3af1c0c9387dc9e55.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 128, + 302, + 483, + 331 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 337, + 454, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 455, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 235, + 352 + ], + "score": 1.0, + "content": "We may now obtain a bound on", + "type": "text" + }, + { + "bbox": [ + 235, + 338, + 262, + 350 + ], + "score": 0.93, + "content": "t _ { \\alpha } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 335, + 324, + 352 + ], + "score": 1.0, + "content": "by plugging in", + "type": "text" + }, + { + "bbox": [ + 325, + 338, + 376, + 350 + ], + "score": 0.93, + "content": "\\sigma _ { i } ( t ) = \\alpha \\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 335, + 455, + 352 + ], + "score": 1.0, + "content": "and taking the log:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 362, + 401, + 394 + ], + "lines": [ + { + "bbox": [ + 209, + 362, + 401, + 394 + ], + "spans": [ + { + "bbox": [ + 209, + 362, + 401, + 394 + ], + "score": 0.92, + "content": "\\log \\big ( \\frac { 1 - \\alpha } { \\alpha \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 \\big ) } \\big ) \\geq t _ { \\alpha } \\cdot \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big )", + "type": "interline_equation", + "image_path": "61f62d820fb5bc742ea13b173f735ebc48de539c2b889c93337334c3f931ed0f.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 209, + 362, + 401, + 394 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 504, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 200, + 417 + ], + "score": 1.0, + "content": "Rearranging (note that", + "type": "text" + }, + { + "bbox": [ + 200, + 401, + 313, + 416 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) < 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 401, + 406, + 417 + ], + "score": 1.0, + "content": "and that our choice of", + "type": "text" + }, + { + "bbox": [ + 406, + 406, + 412, + 413 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 401, + 506, + 417 + ], + "score": 1.0, + "content": "keeps the argument of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 413, + 207, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 207, + 428 + ], + "score": 1.0, + "content": "the log positive), we get:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 438, + 373, + 477 + ], + "lines": [ + { + "bbox": [ + 238, + 438, + 373, + 477 + ], + "spans": [ + { + "bbox": [ + 238, + 438, + 373, + 477 + ], + "score": 0.94, + "content": "t _ { \\alpha } ( \\sigma _ { i } ) \\geq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) }", + "type": "interline_equation", + "image_path": "3031ae89a58278b4232cdfab63ef9cc0ada25c71dd93dd2708ce51130ce4fd48.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 238, + 438, + 373, + 457.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 238, + 457.5, + 373, + 477.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 453, + 496 + ], + "lines": [ + { + "bbox": [ + 104, + 481, + 456, + 498 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 456, + 498 + ], + "score": 1.0, + "content": "Next, we follow the same procedure for an upper bound. Starting with our update step:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 511, + 426, + 595 + ], + "lines": [ + { + "bbox": [ + 184, + 511, + 426, + 595 + ], + "spans": [ + { + "bbox": [ + 184, + 511, + 426, + 595 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) ( c R _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ( 1 - r _ { i } ( t ) ) ) } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) c R _ { i } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } ( 1 - ( 1 - r _ { i } ( t ) ) c R _ { i } + ( 1 - r _ { i } ( t ) ) ^ { 2 } c ^ { 2 } R _ { i } ^ { 2 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "503a6deb3e4a5432aff9b363a5e322c1b39f60bdf4f59fb5e2375cbda9131a22.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 184, + 511, + 426, + 527.8 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 184, + 527.8, + 426, + 544.5999999999999 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 184, + 544.5999999999999, + 426, + 561.3999999999999 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 184, + 561.3999999999999, + 426, + 578.1999999999998 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 184, + 578.1999999999998, + 426, + 594.9999999999998 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 104, + 599, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 104, + 599, + 282, + 621 + ], + "score": 1.0, + "content": "Where in the last line we use the inequality", + "type": "text" + }, + { + "bbox": [ + 282, + 602, + 396, + 617 + ], + "score": 0.77, + "content": "\\begin{array} { r } { \\frac { 1 } { 1 + x } \\leq 1 - x + x ^ { 2 } , \\forall x \\geq 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 599, + 450, + 621 + ], + "score": 1.0, + "content": ". Subtracting", + "type": "text" + }, + { + "bbox": [ + 450, + 602, + 461, + 618 + ], + "score": 0.91, + "content": "\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 599, + 507, + 621 + ], + "score": 1.0, + "content": "from both", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 615, + 164, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 164, + 630 + ], + "score": 1.0, + "content": "sides, we get:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 640, + 480, + 668 + ], + "lines": [ + { + "bbox": [ + 130, + 640, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 480, + 668 + ], + "score": 0.92, + "content": "\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\leq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) \\leq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) ^ { t }", + "type": "interline_equation", + "image_path": "1fbf41e7558ebc51e0a148769c816886b852351549695e6fb49eeedbdc559603.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 130, + 640, + 480, + 668 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 336, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 673, + 337, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 304, + 687 + ], + "score": 1.0, + "content": "Rearranging like before, we get the bound on the", + "type": "text" + }, + { + "bbox": [ + 304, + 676, + 312, + 684 + ], + "score": 0.78, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 673, + 337, + 687 + ], + "score": 1.0, + "content": "-time:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 698, + 372, + 736 + ], + "lines": [ + { + "bbox": [ + 239, + 698, + 372, + 736 + ], + "spans": [ + { + "bbox": [ + 239, + 698, + 372, + 736 + ], + "score": 0.94, + "content": "t _ { \\alpha } ( \\sigma _ { i } ) \\leq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) }", + "type": "interline_equation", + "image_path": "32e30bd54300ce13fc724e28bbf6bc24de046cdc14df93642dc289b37d7c0551.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 698, + 372, + 717.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 239, + 717.0, + 372, + 736.0 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 347, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 346, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "score": 1.0, + "content": "Following theorem 3 of Gidel et al. (2019), we bound", + "type": "text" + }, + { + "bbox": [ + 323, + 81, + 343, + 98 + ], + "score": 0.93, + "content": "\\frac { 1 } { \\sigma _ { i } ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 86, + 346, + 99 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 81, + 346, + 99 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 110, + 426, + 257 + ], + "lines": [ + { + "bbox": [ + 185, + 110, + 426, + 257 + ], + "spans": [ + { + "bbox": [ + 185, + 110, + 426, + 257 + ], + "score": 0.96, + "content": "\\begin{array} { r l } & { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { \\left( 1 + \\frac { \\varepsilon } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\right) ^ { 2 } } } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\left( 1 - r _ { i } ( t ) \\right) \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\bigl ( 1 - ( 1 - r _ { i } ( t ) ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) \\bigr ) } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } - ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "869f7f8a106e05ed77f77284a7b85cc8974315bd7f3d07ed533e53a5664702a0.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 110, + 426, + 128.375 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 185, + 128.375, + 426, + 146.75 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 185, + 146.75, + 426, + 165.125 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 185, + 165.125, + 426, + 183.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 185, + 183.5, + 426, + 201.875 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 185, + 201.875, + 426, + 220.25 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 185, + 220.25, + 426, + 238.625 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 185, + 238.625, + 426, + 257.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 503, + 290 + ], + "lines": [ + { + "bbox": [ + 104, + 262, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 297, + 282 + ], + "score": 1.0, + "content": "Where in the fourth line we use the inequality", + "type": "text" + }, + { + "bbox": [ + 297, + 263, + 394, + 279 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\frac { 1 } { 1 + x } \\geq 1 - x , \\forall x \\geq 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 262, + 491, + 282 + ], + "score": 1.0, + "content": ". We may now subtract", + "type": "text" + }, + { + "bbox": [ + 491, + 263, + 504, + 280 + ], + "score": 0.89, + "content": "\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 277, + 211, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 211, + 290 + ], + "score": 1.0, + "content": "from both sides to obtain:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 262, + 504, + 290 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 302, + 483, + 331 + ], + "lines": [ + { + "bbox": [ + 128, + 302, + 483, + 331 + ], + "spans": [ + { + "bbox": [ + 128, + 302, + 483, + 331 + ], + "score": 0.93, + "content": "\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\geq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) \\geq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) ^ { t }", + "type": "interline_equation", + "image_path": "398c960671dfcf23edd99904d0af487602e5cc56a57b02a3af1c0c9387dc9e55.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 128, + 302, + 483, + 331 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 337, + 454, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 455, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 235, + 352 + ], + "score": 1.0, + "content": "We may now obtain a bound on", + "type": "text" + }, + { + "bbox": [ + 235, + 338, + 262, + 350 + ], + "score": 0.93, + "content": "t _ { \\alpha } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 335, + 324, + 352 + ], + "score": 1.0, + "content": "by plugging in", + "type": "text" + }, + { + "bbox": [ + 325, + 338, + 376, + 350 + ], + "score": 0.93, + "content": "\\sigma _ { i } ( t ) = \\alpha \\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 335, + 455, + 352 + ], + "score": 1.0, + "content": "and taking the log:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 335, + 455, + 352 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 362, + 401, + 394 + ], + "lines": [ + { + "bbox": [ + 209, + 362, + 401, + 394 + ], + "spans": [ + { + "bbox": [ + 209, + 362, + 401, + 394 + ], + "score": 0.92, + "content": "\\log \\big ( \\frac { 1 - \\alpha } { \\alpha \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 \\big ) } \\big ) \\geq t _ { \\alpha } \\cdot \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big )", + "type": "interline_equation", + "image_path": "61f62d820fb5bc742ea13b173f735ebc48de539c2b889c93337334c3f931ed0f.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 209, + 362, + 401, + 394 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 504, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 200, + 417 + ], + "score": 1.0, + "content": "Rearranging (note that", + "type": "text" + }, + { + "bbox": [ + 200, + 401, + 313, + 416 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) < 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 401, + 406, + 417 + ], + "score": 1.0, + "content": "and that our choice of", + "type": "text" + }, + { + "bbox": [ + 406, + 406, + 412, + 413 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 401, + 506, + 417 + ], + "score": 1.0, + "content": "keeps the argument of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 413, + 207, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 207, + 428 + ], + "score": 1.0, + "content": "the log positive), we get:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 401, + 506, + 428 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 238, + 438, + 373, + 477 + ], + "lines": [ + { + "bbox": [ + 238, + 438, + 373, + 477 + ], + "spans": [ + { + "bbox": [ + 238, + 438, + 373, + 477 + ], + "score": 0.94, + "content": "t _ { \\alpha } ( \\sigma _ { i } ) \\geq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) }", + "type": "interline_equation", + "image_path": "3031ae89a58278b4232cdfab63ef9cc0ada25c71dd93dd2708ce51130ce4fd48.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 238, + 438, + 373, + 457.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 238, + 457.5, + 373, + 477.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 453, + 496 + ], + "lines": [ + { + "bbox": [ + 104, + 481, + 456, + 498 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 456, + 498 + ], + "score": 1.0, + "content": "Next, we follow the same procedure for an upper bound. Starting with our update step:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 481, + 456, + 498 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 511, + 426, + 595 + ], + "lines": [ + { + "bbox": [ + 184, + 511, + 426, + 595 + ], + "spans": [ + { + "bbox": [ + 184, + 511, + 426, + 595 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) ( c R _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ( 1 - r _ { i } ( t ) ) ) } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) c R _ { i } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } ( 1 - ( 1 - r _ { i } ( t ) ) c R _ { i } + ( 1 - r _ { i } ( t ) ) ^ { 2 } c ^ { 2 } R _ { i } ^ { 2 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "503a6deb3e4a5432aff9b363a5e322c1b39f60bdf4f59fb5e2375cbda9131a22.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 184, + 511, + 426, + 527.8 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 184, + 527.8, + 426, + 544.5999999999999 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 184, + 544.5999999999999, + 426, + 561.3999999999999 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 184, + 561.3999999999999, + 426, + 578.1999999999998 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 184, + 578.1999999999998, + 426, + 594.9999999999998 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 104, + 599, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 104, + 599, + 282, + 621 + ], + "score": 1.0, + "content": "Where in the last line we use the inequality", + "type": "text" + }, + { + "bbox": [ + 282, + 602, + 396, + 617 + ], + "score": 0.77, + "content": "\\begin{array} { r } { \\frac { 1 } { 1 + x } \\leq 1 - x + x ^ { 2 } , \\forall x \\geq 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 599, + 450, + 621 + ], + "score": 1.0, + "content": ". Subtracting", + "type": "text" + }, + { + "bbox": [ + 450, + 602, + 461, + 618 + ], + "score": 0.91, + "content": "\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 599, + 507, + 621 + ], + "score": 1.0, + "content": "from both", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 615, + 164, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 164, + 630 + ], + "score": 1.0, + "content": "sides, we get:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 599, + 507, + 630 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 640, + 480, + 668 + ], + "lines": [ + { + "bbox": [ + 130, + 640, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 480, + 668 + ], + "score": 0.92, + "content": "\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\leq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) \\leq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) ^ { t }", + "type": "interline_equation", + "image_path": "1fbf41e7558ebc51e0a148769c816886b852351549695e6fb49eeedbdc559603.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 130, + 640, + 480, + 668 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 336, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 673, + 337, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 304, + 687 + ], + "score": 1.0, + "content": "Rearranging like before, we get the bound on the", + "type": "text" + }, + { + "bbox": [ + 304, + 676, + 312, + 684 + ], + "score": 0.78, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 673, + 337, + 687 + ], + "score": 1.0, + "content": "-time:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 673, + 337, + 687 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 698, + 372, + 736 + ], + "lines": [ + { + "bbox": [ + 239, + 698, + 372, + 736 + ], + "spans": [ + { + "bbox": [ + 239, + 698, + 372, + 736 + ], + "score": 0.94, + "content": "t _ { \\alpha } ( \\sigma _ { i } ) \\leq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) }", + "type": "interline_equation", + "image_path": "32e30bd54300ce13fc724e28bbf6bc24de046cdc14df93642dc289b37d7c0551.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 698, + 372, + 717.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 239, + 717.0, + 372, + 736.0 + ], + "spans": [], + "index": 29 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "Given these bounds, we would like to find the conditions on", + "type": "text" + }, + { + "bbox": [ + 354, + 84, + 366, + 94 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 82, + 430, + 95 + ], + "score": 1.0, + "content": "that allows for", + "type": "text" + }, + { + "bbox": [ + 430, + 83, + 453, + 95 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "-incremental", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "learning. We will find a sufficient condition and a necessary condition, like in the proof of theorem", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 118, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 118, + 117 + ], + "score": 1.0, + "content": "2.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 130, + 211, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 212, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 212, + 143 + ], + "score": 1.0, + "content": "SUFFICIENT CONDITION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 151, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 165 + ], + "score": 1.0, + "content": "A sufficient condition for incremental learning will be one which is possibly stricter than the exact", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 162, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 402, + 175 + ], + "score": 1.0, + "content": "condition. We can find such a condition by requiring the upper bound of", + "type": "text" + }, + { + "bbox": [ + 402, + 163, + 429, + 175 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 162, + 504, + 175 + ], + "score": 1.0, + "content": "to be smaller than", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 173, + 372, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 186, + 186 + ], + "score": 1.0, + "content": "the lower bound on", + "type": "text" + }, + { + "bbox": [ + 186, + 174, + 212, + 186 + ], + "score": 0.92, + "content": "t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 173, + 372, + 186 + ], + "score": 1.0, + "content": ". This becomes the following condition:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 199, + 406, + 235 + ], + "lines": [ + { + "bbox": [ + 204, + 199, + 406, + 235 + ], + "spans": [ + { + "bbox": [ + 204, + 199, + 406, + 235 + ], + "score": 0.92, + "content": "\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "fcf156d6d7052f7e5308846663a328233ce16c1034b484297216fe9ee4cd5444.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 199, + 406, + 217.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 204, + 217.0, + 406, + 235.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 244, + 395, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 395, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 144, + 264 + ], + "score": 1.0, + "content": "Defining", + "type": "text" + }, + { + "bbox": [ + 144, + 244, + 238, + 271 + ], + "score": 0.95, + "content": "\\begin{array} { r } { A = \\frac { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\big ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 249, + 395, + 266 + ], + "score": 1.0, + "content": "and rearranging, we get the following:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 277, + 402, + 306 + ], + "lines": [ + { + "bbox": [ + 208, + 277, + 402, + 306 + ], + "spans": [ + { + "bbox": [ + 208, + 277, + 402, + 306 + ], + "score": 0.93, + "content": "\\log \\big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\big ) \\geq A \\log \\big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\big )", + "type": "interline_equation", + "image_path": "7a792ed0cd3ecf2fbb22f0057d9e3acfdfc0c9fecf1c1ba4103d93667df9c07a.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 277, + 402, + 291.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 208, + 291.5, + 402, + 306.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 427, + 337 + ], + "score": 1.0, + "content": "We may now take the exponent of both sides and rearrange again, remembering", + "type": "text" + }, + { + "bbox": [ + 428, + 316, + 476, + 335 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 313, + 505, + 337 + ], + "score": 1.0, + "content": ", to get", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 333, + 205, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 205, + 345 + ], + "score": 1.0, + "content": "the following condition:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 359, + 381, + 391 + ], + "lines": [ + { + "bbox": [ + 230, + 359, + 381, + 391 + ], + "spans": [ + { + "bbox": [ + 230, + 359, + 381, + 391 + ], + "score": 0.94, + "content": "\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A } } { \\sigma _ { 0 } ^ { A - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { A }", + "type": "interline_equation", + "image_path": "3f6348d01377d5ba7a1a4664d06de2954d942de52d7e8b71a9fb2b9419c0ce51.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 359, + 381, + 375.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 230, + 375.0, + 381, + 391.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 503, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 322, + 425 + ], + "score": 1.0, + "content": "Now, we will add the very reasonable assumption that", + "type": "text" + }, + { + "bbox": [ + 322, + 405, + 362, + 419 + ], + "score": 0.93, + "content": "\\sigma _ { j } ^ { * } \\geq 2 \\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 400, + 473, + 425 + ], + "score": 1.0, + "content": ", which allows us to replace", + "type": "text" + }, + { + "bbox": [ + 473, + 401, + 503, + 421 + ], + "score": 0.93, + "content": "\\frac { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } { r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 101, + 416, + 509, + 439 + ], + "spans": [ + { + "bbox": [ + 101, + 416, + 127, + 439 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 420, + 137, + 434 + ], + "score": 0.9, + "content": "\\textstyle { \\frac { 1 } { 2 r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 416, + 186, + 439 + ], + "score": 1.0, + "content": "and replace", + "type": "text" + }, + { + "bbox": [ + 186, + 420, + 240, + 434 + ], + "score": 0.91, + "content": "( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 416, + 261, + 439 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 262, + 420, + 302, + 434 + ], + "score": 0.92, + "content": "\\textstyle { \\left( { \\frac { 1 } { 2 } } \\sigma _ { j } \\right) } ^ { A - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 416, + 509, + 439 + ], + "score": 1.0, + "content": ", only making the condition stricter. This simplifies", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 432, + 235, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 235, + 446 + ], + "score": 1.0, + "content": "the expression to the following:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 458, + 368, + 486 + ], + "lines": [ + { + "bbox": [ + 243, + 458, + 368, + 486 + ], + "spans": [ + { + "bbox": [ + 243, + 458, + 368, + 486 + ], + "score": 0.94, + "content": "\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { A - 1 } \\geq \\frac { r f } { 1 - f } \\big ( \\frac { 2 - 2 s } { s } \\big ) ^ { A }", + "type": "interline_equation", + "image_path": "db3295364432591d6e09a05005d50a52ee38d2c1f5988c2b10f614df7a6bcc2b.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 458, + 368, + 472.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 243, + 472.0, + 368, + 486.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 494, + 469, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 470, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 244, + 508 + ], + "score": 1.0, + "content": "Now we can rearrange and isolate", + "type": "text" + }, + { + "bbox": [ + 244, + 497, + 255, + 506 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 493, + 470, + 508 + ], + "score": 1.0, + "content": "to get a sufficient condition for incremental learning:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 520, + 379, + 548 + ], + "lines": [ + { + "bbox": [ + 232, + 520, + 379, + 548 + ], + "spans": [ + { + "bbox": [ + 232, + 520, + 379, + 548 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq \\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\bigg ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\bigg ) ^ { \\frac { 1 } { A - 1 } }", + "type": "interline_equation", + "image_path": "44f3a5c73f179cd18e12f16bd10686e7817ca87b0ee31991b60e390d70c493d6.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 232, + 520, + 379, + 534.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 232, + 534.0, + 379, + 548.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 560, + 212, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 213, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 213, + 573 + ], + "score": 1.0, + "content": "NECESSARY CONDITION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 595 + ], + "score": 1.0, + "content": "A necessary condition for incremental learning will be one which is possibly more relaxed than the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 423, + 605 + ], + "score": 1.0, + "content": "exact condition. We can find such a condition by requiring the lower bound of", + "type": "text" + }, + { + "bbox": [ + 423, + 593, + 449, + 605 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "to be smaller", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 604, + 393, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 205, + 616 + ], + "score": 1.0, + "content": "than the upper bound on", + "type": "text" + }, + { + "bbox": [ + 206, + 604, + 232, + 616 + ], + "score": 0.92, + "content": "t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 604, + 393, + 616 + ], + "score": 1.0, + "content": ". This becomes the following condition:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 630, + 406, + 664 + ], + "lines": [ + { + "bbox": [ + 204, + 630, + 406, + 664 + ], + "spans": [ + { + "bbox": [ + 204, + 630, + 406, + 664 + ], + "score": 0.92, + "content": "\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "ce276208cd2811997f831b9309860b2e7c24a4126dae02a8fc16713e5572078f.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 630, + 406, + 647.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 204, + 647.0, + 406, + 664.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 674, + 394, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 672, + 395, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 144, + 695 + ], + "score": 1.0, + "content": "Defining", + "type": "text" + }, + { + "bbox": [ + 144, + 674, + 238, + 701 + ], + "score": 0.95, + "content": "\\begin{array} { r } { B = \\frac { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\big ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 672, + 395, + 695 + ], + "score": 1.0, + "content": "and rearranging, we get the following:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 707, + 402, + 735 + ], + "lines": [ + { + "bbox": [ + 208, + 707, + 402, + 735 + ], + "spans": [ + { + "bbox": [ + 208, + 707, + 402, + 735 + ], + "score": 0.92, + "content": "\\log \\Big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\Big ) \\geq B \\log \\Big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\Big )", + "type": "interline_equation", + "image_path": "24ed8ee8ddb98f1166441d9d420ef87a20a63f89ff8f5cb56035bba81892b10f.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 707, + 402, + 721.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 208, + 721.0, + 402, + 735.0 + ], + "spans": [], + "index": 32 + } + ] + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "Given these bounds, we would like to find the conditions on", + "type": "text" + }, + { + "bbox": [ + 354, + 84, + 366, + 94 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 82, + 430, + 95 + ], + "score": 1.0, + "content": "that allows for", + "type": "text" + }, + { + "bbox": [ + 430, + 83, + 453, + 95 + ], + "score": 0.92, + "content": "( s , f )", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "-incremental", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "learning. We will find a sufficient condition and a necessary condition, like in the proof of theorem", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 118, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 118, + 117 + ], + "score": 1.0, + "content": "2.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 82, + 505, + 117 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 130, + 211, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 212, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 212, + 143 + ], + "score": 1.0, + "content": "SUFFICIENT CONDITION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 151, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 165 + ], + "score": 1.0, + "content": "A sufficient condition for incremental learning will be one which is possibly stricter than the exact", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 162, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 402, + 175 + ], + "score": 1.0, + "content": "condition. We can find such a condition by requiring the upper bound of", + "type": "text" + }, + { + "bbox": [ + 402, + 163, + 429, + 175 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 162, + 504, + 175 + ], + "score": 1.0, + "content": "to be smaller than", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 173, + 372, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 186, + 186 + ], + "score": 1.0, + "content": "the lower bound on", + "type": "text" + }, + { + "bbox": [ + 186, + 174, + 212, + 186 + ], + "score": 0.92, + "content": "t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 173, + 372, + 186 + ], + "score": 1.0, + "content": ". This becomes the following condition:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 150, + 506, + 186 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 199, + 406, + 235 + ], + "lines": [ + { + "bbox": [ + 204, + 199, + 406, + 235 + ], + "spans": [ + { + "bbox": [ + 204, + 199, + 406, + 235 + ], + "score": 0.92, + "content": "\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "fcf156d6d7052f7e5308846663a328233ce16c1034b484297216fe9ee4cd5444.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 199, + 406, + 217.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 204, + 217.0, + 406, + 235.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 244, + 395, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 395, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 144, + 264 + ], + "score": 1.0, + "content": "Defining", + "type": "text" + }, + { + "bbox": [ + 144, + 244, + 238, + 271 + ], + "score": 0.95, + "content": "\\begin{array} { r } { A = \\frac { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\big ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 249, + 395, + 266 + ], + "score": 1.0, + "content": "and rearranging, we get the following:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 244, + 395, + 271 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 277, + 402, + 306 + ], + "lines": [ + { + "bbox": [ + 208, + 277, + 402, + 306 + ], + "spans": [ + { + "bbox": [ + 208, + 277, + 402, + 306 + ], + "score": 0.93, + "content": "\\log \\big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\big ) \\geq A \\log \\big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\big )", + "type": "interline_equation", + "image_path": "7a792ed0cd3ecf2fbb22f0057d9e3acfdfc0c9fecf1c1ba4103d93667df9c07a.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 277, + 402, + 291.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 208, + 291.5, + 402, + 306.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 427, + 337 + ], + "score": 1.0, + "content": "We may now take the exponent of both sides and rearrange again, remembering", + "type": "text" + }, + { + "bbox": [ + 428, + 316, + 476, + 335 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 313, + 505, + 337 + ], + "score": 1.0, + "content": ", to get", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 333, + 205, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 205, + 345 + ], + "score": 1.0, + "content": "the following condition:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 313, + 505, + 345 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 359, + 381, + 391 + ], + "lines": [ + { + "bbox": [ + 230, + 359, + 381, + 391 + ], + "spans": [ + { + "bbox": [ + 230, + 359, + 381, + 391 + ], + "score": 0.94, + "content": "\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A } } { \\sigma _ { 0 } ^ { A - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { A }", + "type": "interline_equation", + "image_path": "3f6348d01377d5ba7a1a4664d06de2954d942de52d7e8b71a9fb2b9419c0ce51.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 359, + 381, + 375.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 230, + 375.0, + 381, + 391.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 503, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 322, + 425 + ], + "score": 1.0, + "content": "Now, we will add the very reasonable assumption that", + "type": "text" + }, + { + "bbox": [ + 322, + 405, + 362, + 419 + ], + "score": 0.93, + "content": "\\sigma _ { j } ^ { * } \\geq 2 \\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 400, + 473, + 425 + ], + "score": 1.0, + "content": ", which allows us to replace", + "type": "text" + }, + { + "bbox": [ + 473, + 401, + 503, + 421 + ], + "score": 0.93, + "content": "\\frac { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } { r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 101, + 416, + 509, + 439 + ], + "spans": [ + { + "bbox": [ + 101, + 416, + 127, + 439 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 420, + 137, + 434 + ], + "score": 0.9, + "content": "\\textstyle { \\frac { 1 } { 2 r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 416, + 186, + 439 + ], + "score": 1.0, + "content": "and replace", + "type": "text" + }, + { + "bbox": [ + 186, + 420, + 240, + 434 + ], + "score": 0.91, + "content": "( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 416, + 261, + 439 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 262, + 420, + 302, + 434 + ], + "score": 0.92, + "content": "\\textstyle { \\left( { \\frac { 1 } { 2 } } \\sigma _ { j } \\right) } ^ { A - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 416, + 509, + 439 + ], + "score": 1.0, + "content": ", only making the condition stricter. This simplifies", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 432, + 235, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 235, + 446 + ], + "score": 1.0, + "content": "the expression to the following:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 101, + 400, + 509, + 446 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 458, + 368, + 486 + ], + "lines": [ + { + "bbox": [ + 243, + 458, + 368, + 486 + ], + "spans": [ + { + "bbox": [ + 243, + 458, + 368, + 486 + ], + "score": 0.94, + "content": "\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { A - 1 } \\geq \\frac { r f } { 1 - f } \\big ( \\frac { 2 - 2 s } { s } \\big ) ^ { A }", + "type": "interline_equation", + "image_path": "db3295364432591d6e09a05005d50a52ee38d2c1f5988c2b10f614df7a6bcc2b.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 458, + 368, + 472.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 243, + 472.0, + 368, + 486.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 494, + 469, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 470, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 244, + 508 + ], + "score": 1.0, + "content": "Now we can rearrange and isolate", + "type": "text" + }, + { + "bbox": [ + 244, + 497, + 255, + 506 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 493, + 470, + 508 + ], + "score": 1.0, + "content": "to get a sufficient condition for incremental learning:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 493, + 470, + 508 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 520, + 379, + 548 + ], + "lines": [ + { + "bbox": [ + 232, + 520, + 379, + 548 + ], + "spans": [ + { + "bbox": [ + 232, + 520, + 379, + 548 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq \\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\bigg ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\bigg ) ^ { \\frac { 1 } { A - 1 } }", + "type": "interline_equation", + "image_path": "44f3a5c73f179cd18e12f16bd10686e7817ca87b0ee31991b60e390d70c493d6.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 232, + 520, + 379, + 534.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 232, + 534.0, + 379, + 548.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 560, + 212, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 213, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 213, + 573 + ], + "score": 1.0, + "content": "NECESSARY CONDITION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 595 + ], + "score": 1.0, + "content": "A necessary condition for incremental learning will be one which is possibly more relaxed than the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 423, + 605 + ], + "score": 1.0, + "content": "exact condition. We can find such a condition by requiring the lower bound of", + "type": "text" + }, + { + "bbox": [ + 423, + 593, + 449, + 605 + ], + "score": 0.93, + "content": "t _ { f } ( \\sigma _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "to be smaller", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 604, + 393, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 205, + 616 + ], + "score": 1.0, + "content": "than the upper bound on", + "type": "text" + }, + { + "bbox": [ + 206, + 604, + 232, + 616 + ], + "score": 0.92, + "content": "t _ { s } ( \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 604, + 393, + 616 + ], + "score": 1.0, + "content": ". This becomes the following condition:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 580, + 505, + 616 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 630, + 406, + 664 + ], + "lines": [ + { + "bbox": [ + 204, + 630, + 406, + 664 + ], + "spans": [ + { + "bbox": [ + 204, + 630, + 406, + 664 + ], + "score": 0.92, + "content": "\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\right) }", + "type": "interline_equation", + "image_path": "ce276208cd2811997f831b9309860b2e7c24a4126dae02a8fc16713e5572078f.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 630, + 406, + 647.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 204, + 647.0, + 406, + 664.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 674, + 394, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 672, + 395, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 144, + 695 + ], + "score": 1.0, + "content": "Defining", + "type": "text" + }, + { + "bbox": [ + 144, + 674, + 238, + 701 + ], + "score": 0.95, + "content": "\\begin{array} { r } { B = \\frac { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\big ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 672, + 395, + 695 + ], + "score": 1.0, + "content": "and rearranging, we get the following:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 672, + 395, + 701 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 707, + 402, + 735 + ], + "lines": [ + { + "bbox": [ + 208, + 707, + 402, + 735 + ], + "spans": [ + { + "bbox": [ + 208, + 707, + 402, + 735 + ], + "score": 0.92, + "content": "\\log \\Big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\Big ) \\geq B \\log \\Big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\Big )", + "type": "interline_equation", + "image_path": "24ed8ee8ddb98f1166441d9d420ef87a20a63f89ff8f5cb56035bba81892b10f.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 707, + 402, + 721.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 208, + 721.0, + 402, + 735.0 + ], + "spans": [], + "index": 32 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 80, + 505, + 109 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 427, + 102 + ], + "score": 1.0, + "content": "We may now take the exponent of both sides and rearrange again, remembering", + "type": "text" + }, + { + "bbox": [ + 428, + 80, + 476, + 99 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 77, + 506, + 102 + ], + "score": 1.0, + "content": ", to get", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 205, + 109 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 205, + 109 + ], + "score": 1.0, + "content": "the following condition:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 122, + 381, + 154 + ], + "lines": [ + { + "bbox": [ + 230, + 122, + 381, + 154 + ], + "spans": [ + { + "bbox": [ + 230, + 122, + 381, + 154 + ], + "score": 0.93, + "content": "\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { B } } { \\sigma _ { 0 } ^ { B - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { B }", + "type": "interline_equation", + "image_path": "2613fd0283081b65387794b66fbb9672bab6c8f7554511298d90a8b93755f048.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 122, + 381, + 138.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 230, + 138.0, + 381, + 154.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 160, + 505, + 184 + ], + "lines": [ + { + "bbox": [ + 107, + 161, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 161, + 342, + 172 + ], + "score": 1.0, + "content": "We may now relax the condition further, by removing the", + "type": "text" + }, + { + "bbox": [ + 342, + 163, + 348, + 171 + ], + "score": 0.77, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 161, + 505, + 172 + ], + "score": 1.0, + "content": "from the denominator of the left-hand", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 171, + 365, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 156, + 185 + ], + "score": 1.0, + "content": "side and the", + "type": "text" + }, + { + "bbox": [ + 156, + 173, + 168, + 183 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 171, + 365, + 185 + ], + "score": 1.0, + "content": "from the numerator. This gives us the following:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 196, + 366, + 223 + ], + "lines": [ + { + "bbox": [ + 245, + 196, + 366, + 223 + ], + "spans": [ + { + "bbox": [ + 245, + 196, + 366, + 223 + ], + "score": 0.94, + "content": "\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { B - 1 } \\geq \\frac { f } { 1 - f } \\big ( \\frac { 1 - s } { s } \\big ) ^ { B }", + "type": "interline_equation", + "image_path": "30f112edadd193c88e289a66765d3240020ce807caafce3dbf5e0f30105d1c89.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 245, + 196, + 366, + 223 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 320, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 321, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 321, + 244 + ], + "score": 1.0, + "content": "Finally, rearranging gives us the necessary condition:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 254, + 374, + 282 + ], + "lines": [ + { + "bbox": [ + 235, + 254, + 374, + 282 + ], + "spans": [ + { + "bbox": [ + 235, + 254, + 374, + 282 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }", + "type": "interline_equation", + "image_path": "b6b0bb317c61a5ecad46f8d8fb71fefbd5011b7089bb664b3d6f457e8bea89fa.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 235, + 254, + 374, + 282 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 315, + 409, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 409, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 397, + 331 + ], + "score": 1.0, + "content": "C DISCUSSION OF GRADIENT DESCENT FOR GENERAL", + "type": "text" + }, + { + "bbox": [ + 398, + 317, + 409, + 327 + ], + "score": 0.74, + "content": "N", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 385, + 353 + ], + "score": 1.0, + "content": "While we were able to generalize our result to gradient descent for", + "type": "text" + }, + { + "bbox": [ + 385, + 342, + 417, + 351 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 341, + 505, + 353 + ], + "score": 1.0, + "content": ", our proof technique", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 320, + 364 + ], + "score": 1.0, + "content": "relies on the ability to get a non-implicit solution for", + "type": "text" + }, + { + "bbox": [ + 320, + 352, + 338, + 364 + ], + "score": 0.92, + "content": "\\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "which we discretized and bounded. This", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 363, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 272, + 375 + ], + "score": 1.0, + "content": "is harder to generalize to larger values of", + "type": "text" + }, + { + "bbox": [ + 272, + 363, + 282, + 373 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 363, + 504, + 375 + ], + "score": 1.0, + "content": ", where the solution is implicit. Still, we can informally", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "illustrate the effect of depth on the dynamics of gradient descent by approximating the update rule", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 383, + 162, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 162, + 397 + ], + "score": 1.0, + "content": "of the values.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 503, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 398, + 415 + ], + "score": 1.0, + "content": "We start by reminding ourselves of the gradient descent update rule for", + "type": "text" + }, + { + "bbox": [ + 398, + 405, + 405, + 412 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 399, + 486, + 415 + ], + "score": 1.0, + "content": ", for a learning rate", + "type": "text" + }, + { + "bbox": [ + 486, + 403, + 506, + 414 + ], + "score": 0.82, + "content": "\\eta =", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 101, + 410, + 155, + 434 + ], + "spans": [ + { + "bbox": [ + 101, + 410, + 150, + 434 + ], + "score": 1.0, + "content": "c \u0000 1σ ∗ \u0001 2 − 2N", + "type": "text" + }, + { + "bbox": [ + 151, + 411, + 155, + 427 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 442, + 431, + 468 + ], + "lines": [ + { + "bbox": [ + 179, + 442, + 431, + 468 + ], + "spans": [ + { + "bbox": [ + 179, + 442, + 431, + 468 + ], + "score": 0.91, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\big ( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 - \\frac { 2 } { N } } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) \\big ) \\Big ) ^ { N }", + "type": "interline_equation", + "image_path": "0cd0c1812f44087bd78975b3a0595876da98bd39fcbd6616177755d12e70f2c2.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 179, + 442, + 431, + 468 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 506, + 503 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 474, + 488 + ], + "score": 1.0, + "content": "To compare two values in the same scales, we will divide both sides by the optimal value", + "type": "text" + }, + { + "bbox": [ + 475, + 476, + 486, + 487 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 474, + 505, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 486, + 387, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 245, + 502 + ], + "score": 1.0, + "content": "look at the update step of the ratio", + "type": "text" + }, + { + "bbox": [ + 245, + 487, + 286, + 505 + ], + "score": 0.94, + "content": "\\begin{array} { r } { r _ { i } = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 486, + 346, + 505 + ], + "score": 1.0, + "content": ", also denoting", + "type": "text" + }, + { + "bbox": [ + 347, + 486, + 383, + 504 + ], + "score": 0.93, + "content": "\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 486, + 387, + 505 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 514, + 419, + 538 + ], + "lines": [ + { + "bbox": [ + 191, + 514, + 419, + 538 + ], + "spans": [ + { + "bbox": [ + 191, + 514, + 419, + 538 + ], + "score": 0.91, + "content": "r _ { i } ( t + 1 ) = r _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { N }", + "type": "interline_equation", + "image_path": "1fe0c7b717c4daecb4d27f5eddfff0ec23efbe5f581056d1196de31176422610.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 191, + 514, + 419, + 538 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 544, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 389, + 558 + ], + "score": 1.0, + "content": "We will focus on the early stages of the optimization process, where", + "type": "text" + }, + { + "bbox": [ + 389, + 545, + 418, + 555 + ], + "score": 0.89, + "content": "r \\ll 1", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 543, + 506, + 558 + ], + "score": 1.0, + "content": ". This means we can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 153, + 569 + ], + "score": 1.0, + "content": "neglect the", + "type": "text" + }, + { + "bbox": [ + 153, + 556, + 191, + 568 + ], + "score": 0.92, + "content": "1 - r _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "term in the update step, giving us the approximate update step we will use to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 567, + 235, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 190, + 579 + ], + "score": 1.0, + "content": "compare the general", + "type": "text" + }, + { + "bbox": [ + 190, + 568, + 203, + 578 + ], + "score": 0.89, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 567, + 235, + 579 + ], + "score": 1.0, + "content": "values:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 591, + 397, + 615 + ], + "lines": [ + { + "bbox": [ + 213, + 591, + 397, + 615 + ], + "spans": [ + { + "bbox": [ + 213, + 591, + 397, + 615 + ], + "score": 0.92, + "content": "r _ { i } ( t + 1 ) \\approx r _ { i } ( t ) \\left( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\right) ^ { N }", + "type": "interline_equation", + "image_path": "70832c4c34788b1125286d215c6c50fe4da301fcd150899ab014edb54372ac84.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 213, + 591, + 397, + 615 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 283, + 635 + ], + "score": 1.0, + "content": "We would like to compare the dynamics of", + "type": "text" + }, + { + "bbox": [ + 284, + 624, + 293, + 633 + ], + "score": 0.85, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 621, + 311, + 635 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 312, + 624, + 321, + 635 + ], + "score": 0.85, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 621, + 506, + 635 + ], + "score": 1.0, + "content": ", which is difficult to do when the recurrence", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "relation isn’t solvable. However, we can observe the first iteration of gradient descent and see how", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 645, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 107, + 645, + 505, + 656 + ], + "score": 1.0, + "content": "depth affects this iteration. Since we are dealing with variables which are ratios of different optimal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 651, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 215, + 677 + ], + "score": 1.0, + "content": "values, the initial values of", + "type": "text" + }, + { + "bbox": [ + 216, + 660, + 221, + 668 + ], + "score": 0.75, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 651, + 316, + 677 + ], + "score": 1.0, + "content": "are different. Denoting", + "type": "text" + }, + { + "bbox": [ + 316, + 655, + 346, + 675 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { r } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 651, + 494, + 677 + ], + "score": 1.0, + "content": ", we can describe the initialization of", + "type": "text" + }, + { + "bbox": [ + 494, + 660, + 504, + 671 + ], + "score": 0.76, + "content": "r _ { j }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 672, + 172, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 159, + 685 + ], + "score": 1.0, + "content": "using that of", + "type": "text" + }, + { + "bbox": [ + 159, + 674, + 168, + 684 + ], + "score": 0.81, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 672, + 172, + 685 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 699, + 336, + 713 + ], + "lines": [ + { + "bbox": [ + 275, + 699, + 336, + 713 + ], + "spans": [ + { + "bbox": [ + 275, + 699, + 336, + 713 + ], + "score": 0.89, + "content": "r _ { j } ( 0 ) = { \\bf r } r _ { i } ( 0 )", + "type": "interline_equation", + "image_path": "732ca989aad5d07fc705b83e0d0badcf61f9e527c5faefc1129c5848d57079b3.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 275, + 699, + 336, + 713 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 378, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 379, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 302, + 735 + ], + "score": 1.0, + "content": "Plugging in the initial conditions and noting that", + "type": "text" + }, + { + "bbox": [ + 302, + 721, + 343, + 733 + ], + "score": 0.92, + "content": "R _ { i } = { \\bf r } R _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 718, + 379, + 735 + ], + "score": 1.0, + "content": ", we get:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 289, + 505, + 300 + ], + "lines": [ + { + "bbox": [ + 496, + 291, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 496, + 291, + 504, + 299 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 80, + 505, + 109 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 427, + 102 + ], + "score": 1.0, + "content": "We may now take the exponent of both sides and rearrange again, remembering", + "type": "text" + }, + { + "bbox": [ + 428, + 80, + 476, + 99 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 77, + 506, + 102 + ], + "score": 1.0, + "content": ", to get", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 205, + 109 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 205, + 109 + ], + "score": 1.0, + "content": "the following condition:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 77, + 506, + 109 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 122, + 381, + 154 + ], + "lines": [ + { + "bbox": [ + 230, + 122, + 381, + 154 + ], + "spans": [ + { + "bbox": [ + 230, + 122, + 381, + 154 + ], + "score": 0.93, + "content": "\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { B } } { \\sigma _ { 0 } ^ { B - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { B }", + "type": "interline_equation", + "image_path": "2613fd0283081b65387794b66fbb9672bab6c8f7554511298d90a8b93755f048.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 122, + 381, + 138.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 230, + 138.0, + 381, + 154.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 160, + 505, + 184 + ], + "lines": [ + { + "bbox": [ + 107, + 161, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 161, + 342, + 172 + ], + "score": 1.0, + "content": "We may now relax the condition further, by removing the", + "type": "text" + }, + { + "bbox": [ + 342, + 163, + 348, + 171 + ], + "score": 0.77, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 161, + 505, + 172 + ], + "score": 1.0, + "content": "from the denominator of the left-hand", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 171, + 365, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 156, + 185 + ], + "score": 1.0, + "content": "side and the", + "type": "text" + }, + { + "bbox": [ + 156, + 173, + 168, + 183 + ], + "score": 0.85, + "content": "\\sigma _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 171, + 365, + 185 + ], + "score": 1.0, + "content": "from the numerator. This gives us the following:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 161, + 505, + 185 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 196, + 366, + 223 + ], + "lines": [ + { + "bbox": [ + 245, + 196, + 366, + 223 + ], + "spans": [ + { + "bbox": [ + 245, + 196, + 366, + 223 + ], + "score": 0.94, + "content": "\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { B - 1 } \\geq \\frac { f } { 1 - f } \\big ( \\frac { 1 - s } { s } \\big ) ^ { B }", + "type": "interline_equation", + "image_path": "30f112edadd193c88e289a66765d3240020ce807caafce3dbf5e0f30105d1c89.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 245, + 196, + 366, + 223 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 320, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 321, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 321, + 244 + ], + "score": 1.0, + "content": "Finally, rearranging gives us the necessary condition:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 229, + 321, + 244 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 254, + 374, + 282 + ], + "lines": [ + { + "bbox": [ + 235, + 254, + 374, + 282 + ], + "spans": [ + { + "bbox": [ + 235, + 254, + 374, + 282 + ], + "score": 0.93, + "content": "\\sigma _ { 0 } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }", + "type": "interline_equation", + "image_path": "b6b0bb317c61a5ecad46f8d8fb71fefbd5011b7089bb664b3d6f457e8bea89fa.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 235, + 254, + 374, + 282 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 315, + 409, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 409, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 397, + 331 + ], + "score": 1.0, + "content": "C DISCUSSION OF GRADIENT DESCENT FOR GENERAL", + "type": "text" + }, + { + "bbox": [ + 398, + 317, + 409, + 327 + ], + "score": 0.74, + "content": "N", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 385, + 353 + ], + "score": 1.0, + "content": "While we were able to generalize our result to gradient descent for", + "type": "text" + }, + { + "bbox": [ + 385, + 342, + 417, + 351 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 341, + 505, + 353 + ], + "score": 1.0, + "content": ", our proof technique", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 320, + 364 + ], + "score": 1.0, + "content": "relies on the ability to get a non-implicit solution for", + "type": "text" + }, + { + "bbox": [ + 320, + 352, + 338, + 364 + ], + "score": 0.92, + "content": "\\sigma ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "which we discretized and bounded. This", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 363, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 272, + 375 + ], + "score": 1.0, + "content": "is harder to generalize to larger values of", + "type": "text" + }, + { + "bbox": [ + 272, + 363, + 282, + 373 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 363, + 504, + 375 + ], + "score": 1.0, + "content": ", where the solution is implicit. Still, we can informally", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 387 + ], + "score": 1.0, + "content": "illustrate the effect of depth on the dynamics of gradient descent by approximating the update rule", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 383, + 162, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 162, + 397 + ], + "score": 1.0, + "content": "of the values.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 341, + 505, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 503, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 398, + 415 + ], + "score": 1.0, + "content": "We start by reminding ourselves of the gradient descent update rule for", + "type": "text" + }, + { + "bbox": [ + 398, + 405, + 405, + 412 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 399, + 486, + 415 + ], + "score": 1.0, + "content": ", for a learning rate", + "type": "text" + }, + { + "bbox": [ + 486, + 403, + 506, + 414 + ], + "score": 0.82, + "content": "\\eta =", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 101, + 410, + 155, + 434 + ], + "spans": [ + { + "bbox": [ + 101, + 410, + 150, + 434 + ], + "score": 1.0, + "content": "c \u0000 1σ ∗ \u0001 2 − 2N", + "type": "text" + }, + { + "bbox": [ + 151, + 411, + 155, + 427 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 101, + 399, + 506, + 434 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 442, + 431, + 468 + ], + "lines": [ + { + "bbox": [ + 179, + 442, + 431, + 468 + ], + "spans": [ + { + "bbox": [ + 179, + 442, + 431, + 468 + ], + "score": 0.91, + "content": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\big ( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 - \\frac { 2 } { N } } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) \\big ) \\Big ) ^ { N }", + "type": "interline_equation", + "image_path": "0cd0c1812f44087bd78975b3a0595876da98bd39fcbd6616177755d12e70f2c2.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 179, + 442, + 431, + 468 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 506, + 503 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 474, + 488 + ], + "score": 1.0, + "content": "To compare two values in the same scales, we will divide both sides by the optimal value", + "type": "text" + }, + { + "bbox": [ + 475, + 476, + 486, + 487 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 474, + 505, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 486, + 387, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 245, + 502 + ], + "score": 1.0, + "content": "look at the update step of the ratio", + "type": "text" + }, + { + "bbox": [ + 245, + 487, + 286, + 505 + ], + "score": 0.94, + "content": "\\begin{array} { r } { r _ { i } = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 486, + 346, + 505 + ], + "score": 1.0, + "content": ", also denoting", + "type": "text" + }, + { + "bbox": [ + 347, + 486, + 383, + 504 + ], + "score": 0.93, + "content": "\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 486, + 387, + 505 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 474, + 505, + 505 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 514, + 419, + 538 + ], + "lines": [ + { + "bbox": [ + 191, + 514, + 419, + 538 + ], + "spans": [ + { + "bbox": [ + 191, + 514, + 419, + 538 + ], + "score": 0.91, + "content": "r _ { i } ( t + 1 ) = r _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { N }", + "type": "interline_equation", + "image_path": "1fe0c7b717c4daecb4d27f5eddfff0ec23efbe5f581056d1196de31176422610.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 191, + 514, + 419, + 538 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 544, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 389, + 558 + ], + "score": 1.0, + "content": "We will focus on the early stages of the optimization process, where", + "type": "text" + }, + { + "bbox": [ + 389, + 545, + 418, + 555 + ], + "score": 0.89, + "content": "r \\ll 1", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 543, + 506, + 558 + ], + "score": 1.0, + "content": ". This means we can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 153, + 569 + ], + "score": 1.0, + "content": "neglect the", + "type": "text" + }, + { + "bbox": [ + 153, + 556, + 191, + 568 + ], + "score": 0.92, + "content": "1 - r _ { i } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "term in the update step, giving us the approximate update step we will use to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 567, + 235, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 190, + 579 + ], + "score": 1.0, + "content": "compare the general", + "type": "text" + }, + { + "bbox": [ + 190, + 568, + 203, + 578 + ], + "score": 0.89, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 567, + 235, + 579 + ], + "score": 1.0, + "content": "values:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 543, + 506, + 579 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 591, + 397, + 615 + ], + "lines": [ + { + "bbox": [ + 213, + 591, + 397, + 615 + ], + "spans": [ + { + "bbox": [ + 213, + 591, + 397, + 615 + ], + "score": 0.92, + "content": "r _ { i } ( t + 1 ) \\approx r _ { i } ( t ) \\left( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\right) ^ { N }", + "type": "interline_equation", + "image_path": "70832c4c34788b1125286d215c6c50fe4da301fcd150899ab014edb54372ac84.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 213, + 591, + 397, + 615 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 283, + 635 + ], + "score": 1.0, + "content": "We would like to compare the dynamics of", + "type": "text" + }, + { + "bbox": [ + 284, + 624, + 293, + 633 + ], + "score": 0.85, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 621, + 311, + 635 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 312, + 624, + 321, + 635 + ], + "score": 0.85, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 621, + 506, + 635 + ], + "score": 1.0, + "content": ", which is difficult to do when the recurrence", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "relation isn’t solvable. However, we can observe the first iteration of gradient descent and see how", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 645, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 107, + 645, + 505, + 656 + ], + "score": 1.0, + "content": "depth affects this iteration. Since we are dealing with variables which are ratios of different optimal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 651, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 215, + 677 + ], + "score": 1.0, + "content": "values, the initial values of", + "type": "text" + }, + { + "bbox": [ + 216, + 660, + 221, + 668 + ], + "score": 0.75, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 651, + 316, + 677 + ], + "score": 1.0, + "content": "are different. Denoting", + "type": "text" + }, + { + "bbox": [ + 316, + 655, + 346, + 675 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { r } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 651, + 494, + 677 + ], + "score": 1.0, + "content": ", we can describe the initialization of", + "type": "text" + }, + { + "bbox": [ + 494, + 660, + 504, + 671 + ], + "score": 0.76, + "content": "r _ { j }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 672, + 172, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 159, + 685 + ], + "score": 1.0, + "content": "using that of", + "type": "text" + }, + { + "bbox": [ + 159, + 674, + 168, + 684 + ], + "score": 0.81, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 672, + 172, + 685 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 621, + 506, + 685 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 699, + 336, + 713 + ], + "lines": [ + { + "bbox": [ + 275, + 699, + 336, + 713 + ], + "spans": [ + { + "bbox": [ + 275, + 699, + 336, + 713 + ], + "score": 0.89, + "content": "r _ { j } ( 0 ) = { \\bf r } r _ { i } ( 0 )", + "type": "interline_equation", + "image_path": "732ca989aad5d07fc705b83e0d0badcf61f9e527c5faefc1129c5848d57079b3.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 275, + 699, + 336, + 713 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 378, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 718, + 379, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 302, + 735 + ], + "score": 1.0, + "content": "Plugging in the initial conditions and noting that", + "type": "text" + }, + { + "bbox": [ + 302, + 721, + 343, + 733 + ], + "score": 0.92, + "content": "R _ { i } = { \\bf r } R _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 718, + 379, + 735 + ], + "score": 1.0, + "content": ", we get:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 718, + 379, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 199, + 92, + 412, + 153 + ], + "lines": [ + { + "bbox": [ + 199, + 92, + 412, + 153 + ], + "spans": [ + { + "bbox": [ + 199, + 92, + 412, + 153 + ], + "score": 0.93, + "content": "\\begin{array} { r } { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( 1 + \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\\\ { r _ { j } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( \\sqrt [ N ] { \\mathbf { r } } + \\big ( \\frac { 1 } { \\mathbf { r } } \\big ) ^ { \\frac { N - 1 } { N } } \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\end{array}", + "type": "interline_equation", + "image_path": "36f6da11812ef6f6803b6b456473767caaae9b4e7c828ae1174c883796119f3b.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 199, + 92, + 412, + 107.25 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 199, + 107.25, + 412, + 122.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 199, + 122.5, + 412, + 137.75 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 199, + 137.75, + 412, + 153.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 156, + 506, + 201 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "We see that the two ratios have a similar update, with the ratio of optimal values playing a role in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 166, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 104, + 166, + 506, + 181 + ], + "score": 1.0, + "content": "how large the initial value is versus how large the added value is. When we use a small learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 212, + 191 + ], + "score": 1.0, + "content": "rate, we have a very small", + "type": "text" + }, + { + "bbox": [ + 212, + 182, + 218, + 189 + ], + "score": 0.74, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "which means we can make a final approximation and neglect the higher", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 190, + 176, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 165, + 201 + ], + "score": 1.0, + "content": "order terms of", + "type": "text" + }, + { + "bbox": [ + 166, + 193, + 171, + 200 + ], + "score": 0.72, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 190, + 176, + 201 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 216, + 393, + 261 + ], + "lines": [ + { + "bbox": [ + 218, + 216, + 393, + 261 + ], + "spans": [ + { + "bbox": [ + 218, + 216, + 393, + 261 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) + c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\\\ & { r _ { j } ( 1 ) \\approx { \\bf r } r _ { i } ( 0 ) + \\big ( \\frac { 1 } { { \\bf r } } \\big ) ^ { \\frac { N - 1 } { N } } c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}", + "type": "interline_equation", + "image_path": "0599ec72d1268d76f947ebd27669a0dcf1aff387d0b69787069a2b1ceaeb5261.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 218, + 216, + 393, + 231.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 218, + 231.0, + 393, + 246.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 218, + 246.0, + 393, + 261.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 506, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 313, + 279 + ], + "score": 1.0, + "content": "We can see that while the initial conditions favor", + "type": "text" + }, + { + "bbox": [ + 313, + 267, + 323, + 278 + ], + "score": 0.84, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 264, + 436, + 279 + ], + "score": 1.0, + "content": ", the size of the update for", + "type": "text" + }, + { + "bbox": [ + 436, + 268, + 445, + 277 + ], + "score": 0.84, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 264, + 505, + 279 + ], + "score": 1.0, + "content": "is larger by a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 277, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 142, + 293 + ], + "score": 1.0, + "content": "factor of", + "type": "text" + }, + { + "bbox": [ + 143, + 277, + 165, + 290 + ], + "score": 0.93, + "content": "r ^ { \\frac { N - 1 } { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 277, + 506, + 293 + ], + "score": 1.0, + "content": "when the initialization and learning rates are small. This accumulates throughout the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 289, + 345, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 194, + 305 + ], + "score": 1.0, + "content": "optimization, making", + "type": "text" + }, + { + "bbox": [ + 194, + 292, + 203, + 302 + ], + "score": 0.83, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 289, + 330, + 305 + ], + "score": 1.0, + "content": "eventually converge faster than", + "type": "text" + }, + { + "bbox": [ + 331, + 293, + 340, + 303 + ], + "score": 0.84, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 289, + 345, + 305 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 504, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 463, + 320 + ], + "score": 1.0, + "content": "The effect of depth here is clear - the deeper the model, the larger the relative step size of", + "type": "text" + }, + { + "bbox": [ + 463, + 309, + 472, + 319 + ], + "score": 0.84, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 237, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 223, + 333 + ], + "score": 1.0, + "content": "faster it converges relative to", + "type": "text" + }, + { + "bbox": [ + 223, + 321, + 233, + 331 + ], + "score": 0.84, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 317, + 237, + 333 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 344, + 363, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 364, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 364, + 360 + ], + "score": 1.0, + "content": "D COMPARISON OF THE TOY MODEL AND OMP", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Learning our toy model, when it’s incremental learning is taken to the limit, can be described as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "an iterative procedure where at every step an additional feature is introduced such that it’s weight", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 392, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 403 + ], + "score": 1.0, + "content": "is non-zero and then the model is optimized over the current set of features. This description is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "also relevant for the sparse approximation algorithm orthogonal matching pursuit (Pati et al., 1993),", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 413, + 489, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 489, + 426 + ], + "score": 1.0, + "content": "where the next feature is greedily chosen to be the one which most improves the current model.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "While the toy model and OMP are very different algorithms for learning sparse linear models, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 442, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 506, + 454 + ], + "score": 1.0, + "content": "will show empirically that they behave similarly. This allows us to view incremental learning as a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 339, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 339, + 465 + ], + "score": 1.0, + "content": "continuous-time extension of a greedy iterative algorithm.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 105, + 469, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "To allow for negative weights in our experiments, we augment our toy model as in the toy model of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 480, + 429, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 429, + 492 + ], + "score": 1.0, + "content": "Woodworth et al. (2019). Our model will have the same induced form as before:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 274, + 507, + 337, + 520 + ], + "lines": [ + { + "bbox": [ + 274, + 507, + 337, + 520 + ], + "spans": [ + { + "bbox": [ + 274, + 507, + 337, + 520 + ], + "score": 0.91, + "content": "f _ { \\sigma } ( x ) = \\langle \\sigma , x \\rangle", + "type": "interline_equation", + "image_path": "8e6cbc889da30afebafa57fb8cd73c4527400fe432a1c124d0ac410d7a8b5389.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 274, + 507, + 337, + 520 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 393, + 538 + ], + "lines": [ + { + "bbox": [ + 104, + 523, + 394, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 215, + 541 + ], + "score": 1.0, + "content": "However, we parameterize", + "type": "text" + }, + { + "bbox": [ + 215, + 529, + 222, + 536 + ], + "score": 0.79, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 523, + 248, + 541 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 248, + 525, + 305, + 538 + ], + "score": 0.92, + "content": "w _ { + } , w _ { - } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 523, + 394, + 541 + ], + "score": 1.0, + "content": "in the following way:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 554, + 372, + 586 + ], + "lines": [ + { + "bbox": [ + 239, + 554, + 372, + 586 + ], + "spans": [ + { + "bbox": [ + 239, + 554, + 372, + 586 + ], + "score": 0.91, + "content": "\\begin{array} { c } { \\sigma _ { i } = w _ { + , i } ^ { N } - w _ { - , i } ^ { N } } \\\\ { \\qquad \\forall i , \\ w _ { + , i } ( 0 ) = w _ { - , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } } } \\end{array}", + "type": "interline_equation", + "image_path": "b7b5e6a4c4974a05a300ad6b46052fb88d306a51df08491dd7e32e93e7e9ef14.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 554, + 372, + 570.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 239, + 570.0, + 372, + 586.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 606 + ], + "score": 1.0, + "content": "We can now treat this algorithm as a sparse approximation pursuit algorithm - given a dictionary", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 600, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 107, + 602, + 151, + 613 + ], + "score": 0.92, + "content": "D \\in \\mathbb { R } ^ { d \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 600, + 216, + 617 + ], + "score": 1.0, + "content": "and an example", + "type": "text" + }, + { + "bbox": [ + 216, + 602, + 246, + 613 + ], + "score": 0.91, + "content": "x \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 600, + 360, + 617 + ], + "score": 1.0, + "content": ", we wish to find the sparsest", + "type": "text" + }, + { + "bbox": [ + 360, + 605, + 368, + 613 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 600, + 408, + 617 + ], + "score": 1.0, + "content": "for which", + "type": "text" + }, + { + "bbox": [ + 409, + 603, + 444, + 613 + ], + "score": 0.9, + "content": "D \\alpha \\approx x", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 600, + 505, + 617 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 612, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 121, + 628 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 614, + 131, + 625 + ], + "score": 0.87, + "content": "\\ell _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 612, + 167, + 628 + ], + "score": 1.0, + "content": "norm of", + "type": "text" + }, + { + "bbox": [ + 167, + 615, + 175, + 624 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 612, + 218, + 628 + ], + "score": 1.0, + "content": "subject to", + "type": "text" + }, + { + "bbox": [ + 218, + 613, + 290, + 626 + ], + "score": 0.93, + "content": "\\| D \\alpha - x \\| _ { 2 } ^ { 2 } = 0 ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 612, + 505, + 628 + ], + "score": 1.0, + "content": ". Under this setting, we can compare OMP to our toy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "model by comparing the sets of features that the two algorithms choose for a given example and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 634, + 151, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 151, + 650 + ], + "score": 1.0, + "content": "dictionary.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 653, + 504, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 504, + 665 + ], + "score": 1.0, + "content": "In figure 3 we run such a comparison. Using a dictionary of 1000 atoms and an example of dimen-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 384, + 677 + ], + "score": 1.0, + "content": "sionality 80 sampled from a random hidden vector of a given sparsity", + "type": "text" + }, + { + "bbox": [ + 384, + 666, + 390, + 673 + ], + "score": 0.69, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 663, + 506, + 677 + ], + "score": 1.0, + "content": ", we run both algorithms and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 675, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 167, + 685 + ], + "score": 1.0, + "content": "record the first", + "type": "text" + }, + { + "bbox": [ + 167, + 677, + 173, + 685 + ], + "score": 0.73, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 675, + 244, + 685 + ], + "score": 1.0, + "content": "features chosen4.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 690, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 688, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 118, + 688, + 505, + 703 + ], + "score": 1.0, + "content": "3Note that this problem is equivalent to learning the toy model over the squared loss, where the examples in", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 700, + 324, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 324, + 713 + ], + "score": 1.0, + "content": "the original learning problem play the role of the dictionary.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 356, + 723 + ], + "score": 1.0, + "content": "4For the toy model, we define a feature to be “chosen” once it’s", + "type": "text" + }, + { + "bbox": [ + 357, + 713, + 366, + 721 + ], + "score": 0.81, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "value passes some small threshold in", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 720, + 163, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 163, + 732 + ], + "score": 1.0, + "content": "absolute value.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 199, + 92, + 412, + 153 + ], + "lines": [ + { + "bbox": [ + 199, + 92, + 412, + 153 + ], + "spans": [ + { + "bbox": [ + 199, + 92, + 412, + 153 + ], + "score": 0.93, + "content": "\\begin{array} { r } { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( 1 + \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\\\ { r _ { j } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( \\sqrt [ N ] { \\mathbf { r } } + \\big ( \\frac { 1 } { \\mathbf { r } } \\big ) ^ { \\frac { N - 1 } { N } } \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\end{array}", + "type": "interline_equation", + "image_path": "36f6da11812ef6f6803b6b456473767caaae9b4e7c828ae1174c883796119f3b.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 199, + 92, + 412, + 107.25 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 199, + 107.25, + 412, + 122.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 199, + 122.5, + 412, + 137.75 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 199, + 137.75, + 412, + 153.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 156, + 506, + 201 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 506, + 169 + ], + "score": 1.0, + "content": "We see that the two ratios have a similar update, with the ratio of optimal values playing a role in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 166, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 104, + 166, + 506, + 181 + ], + "score": 1.0, + "content": "how large the initial value is versus how large the added value is. When we use a small learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 212, + 191 + ], + "score": 1.0, + "content": "rate, we have a very small", + "type": "text" + }, + { + "bbox": [ + 212, + 182, + 218, + 189 + ], + "score": 0.74, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "which means we can make a final approximation and neglect the higher", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 190, + 176, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 165, + 201 + ], + "score": 1.0, + "content": "order terms of", + "type": "text" + }, + { + "bbox": [ + 166, + 193, + 171, + 200 + ], + "score": 0.72, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 190, + 176, + 201 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 156, + 506, + 201 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 216, + 393, + 261 + ], + "lines": [ + { + "bbox": [ + 218, + 216, + 393, + 261 + ], + "spans": [ + { + "bbox": [ + 218, + 216, + 393, + 261 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) + c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\\\ & { r _ { j } ( 1 ) \\approx { \\bf r } r _ { i } ( 0 ) + \\big ( \\frac { 1 } { { \\bf r } } \\big ) ^ { \\frac { N - 1 } { N } } c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}", + "type": "interline_equation", + "image_path": "0599ec72d1268d76f947ebd27669a0dcf1aff387d0b69787069a2b1ceaeb5261.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 218, + 216, + 393, + 231.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 218, + 231.0, + 393, + 246.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 218, + 246.0, + 393, + 261.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 506, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 313, + 279 + ], + "score": 1.0, + "content": "We can see that while the initial conditions favor", + "type": "text" + }, + { + "bbox": [ + 313, + 267, + 323, + 278 + ], + "score": 0.84, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 264, + 436, + 279 + ], + "score": 1.0, + "content": ", the size of the update for", + "type": "text" + }, + { + "bbox": [ + 436, + 268, + 445, + 277 + ], + "score": 0.84, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 264, + 505, + 279 + ], + "score": 1.0, + "content": "is larger by a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 277, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 142, + 293 + ], + "score": 1.0, + "content": "factor of", + "type": "text" + }, + { + "bbox": [ + 143, + 277, + 165, + 290 + ], + "score": 0.93, + "content": "r ^ { \\frac { N - 1 } { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 277, + 506, + 293 + ], + "score": 1.0, + "content": "when the initialization and learning rates are small. This accumulates throughout the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 289, + 345, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 194, + 305 + ], + "score": 1.0, + "content": "optimization, making", + "type": "text" + }, + { + "bbox": [ + 194, + 292, + 203, + 302 + ], + "score": 0.83, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 289, + 330, + 305 + ], + "score": 1.0, + "content": "eventually converge faster than", + "type": "text" + }, + { + "bbox": [ + 331, + 293, + 340, + 303 + ], + "score": 0.84, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 289, + 345, + 305 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 264, + 506, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 504, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 463, + 320 + ], + "score": 1.0, + "content": "The effect of depth here is clear - the deeper the model, the larger the relative step size of", + "type": "text" + }, + { + "bbox": [ + 463, + 309, + 472, + 319 + ], + "score": 0.84, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 237, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 223, + 333 + ], + "score": 1.0, + "content": "faster it converges relative to", + "type": "text" + }, + { + "bbox": [ + 223, + 321, + 233, + 331 + ], + "score": 0.84, + "content": "r _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 317, + 237, + 333 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 306, + 506, + 333 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 344, + 363, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 364, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 364, + 360 + ], + "score": 1.0, + "content": "D COMPARISON OF THE TOY MODEL AND OMP", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Learning our toy model, when it’s incremental learning is taken to the limit, can be described as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "an iterative procedure where at every step an additional feature is introduced such that it’s weight", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 392, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 403 + ], + "score": 1.0, + "content": "is non-zero and then the model is optimized over the current set of features. This description is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "also relevant for the sparse approximation algorithm orthogonal matching pursuit (Pati et al., 1993),", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 413, + 489, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 489, + 426 + ], + "score": 1.0, + "content": "where the next feature is greedily chosen to be the one which most improves the current model.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 369, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "While the toy model and OMP are very different algorithms for learning sparse linear models, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 442, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 506, + 454 + ], + "score": 1.0, + "content": "will show empirically that they behave similarly. This allows us to view incremental learning as a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 339, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 339, + 465 + ], + "score": 1.0, + "content": "continuous-time extension of a greedy iterative algorithm.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 430, + 506, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 469, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "To allow for negative weights in our experiments, we augment our toy model as in the toy model of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 480, + 429, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 429, + 492 + ], + "score": 1.0, + "content": "Woodworth et al. (2019). Our model will have the same induced form as before:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 468, + 505, + 492 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 274, + 507, + 337, + 520 + ], + "lines": [ + { + "bbox": [ + 274, + 507, + 337, + 520 + ], + "spans": [ + { + "bbox": [ + 274, + 507, + 337, + 520 + ], + "score": 0.91, + "content": "f _ { \\sigma } ( x ) = \\langle \\sigma , x \\rangle", + "type": "interline_equation", + "image_path": "8e6cbc889da30afebafa57fb8cd73c4527400fe432a1c124d0ac410d7a8b5389.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 274, + 507, + 337, + 520 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 393, + 538 + ], + "lines": [ + { + "bbox": [ + 104, + 523, + 394, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 215, + 541 + ], + "score": 1.0, + "content": "However, we parameterize", + "type": "text" + }, + { + "bbox": [ + 215, + 529, + 222, + 536 + ], + "score": 0.79, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 523, + 248, + 541 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 248, + 525, + 305, + 538 + ], + "score": 0.92, + "content": "w _ { + } , w _ { - } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 523, + 394, + 541 + ], + "score": 1.0, + "content": "in the following way:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 523, + 394, + 541 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 554, + 372, + 586 + ], + "lines": [ + { + "bbox": [ + 239, + 554, + 372, + 586 + ], + "spans": [ + { + "bbox": [ + 239, + 554, + 372, + 586 + ], + "score": 0.91, + "content": "\\begin{array} { c } { \\sigma _ { i } = w _ { + , i } ^ { N } - w _ { - , i } ^ { N } } \\\\ { \\qquad \\forall i , \\ w _ { + , i } ( 0 ) = w _ { - , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } } } \\end{array}", + "type": "interline_equation", + "image_path": "b7b5e6a4c4974a05a300ad6b46052fb88d306a51df08491dd7e32e93e7e9ef14.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 554, + 372, + 570.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 239, + 570.0, + 372, + 586.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 505, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 606 + ], + "score": 1.0, + "content": "We can now treat this algorithm as a sparse approximation pursuit algorithm - given a dictionary", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 600, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 107, + 602, + 151, + 613 + ], + "score": 0.92, + "content": "D \\in \\mathbb { R } ^ { d \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 600, + 216, + 617 + ], + "score": 1.0, + "content": "and an example", + "type": "text" + }, + { + "bbox": [ + 216, + 602, + 246, + 613 + ], + "score": 0.91, + "content": "x \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 600, + 360, + 617 + ], + "score": 1.0, + "content": ", we wish to find the sparsest", + "type": "text" + }, + { + "bbox": [ + 360, + 605, + 368, + 613 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 600, + 408, + 617 + ], + "score": 1.0, + "content": "for which", + "type": "text" + }, + { + "bbox": [ + 409, + 603, + 444, + 613 + ], + "score": 0.9, + "content": "D \\alpha \\approx x", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 600, + 505, + 617 + ], + "score": 1.0, + "content": "by minimizing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 612, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 121, + 628 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 614, + 131, + 625 + ], + "score": 0.87, + "content": "\\ell _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 612, + 167, + 628 + ], + "score": 1.0, + "content": "norm of", + "type": "text" + }, + { + "bbox": [ + 167, + 615, + 175, + 624 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 612, + 218, + 628 + ], + "score": 1.0, + "content": "subject to", + "type": "text" + }, + { + "bbox": [ + 218, + 613, + 290, + 626 + ], + "score": 0.93, + "content": "\\| D \\alpha - x \\| _ { 2 } ^ { 2 } = 0 ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 612, + 505, + 628 + ], + "score": 1.0, + "content": ". Under this setting, we can compare OMP to our toy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "model by comparing the sets of features that the two algorithms choose for a given example and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 634, + 151, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 151, + 650 + ], + "score": 1.0, + "content": "dictionary.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 590, + 505, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 653, + 504, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 504, + 665 + ], + "score": 1.0, + "content": "In figure 3 we run such a comparison. Using a dictionary of 1000 atoms and an example of dimen-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 384, + 677 + ], + "score": 1.0, + "content": "sionality 80 sampled from a random hidden vector of a given sparsity", + "type": "text" + }, + { + "bbox": [ + 384, + 666, + 390, + 673 + ], + "score": 0.69, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 663, + 506, + 677 + ], + "score": 1.0, + "content": ", we run both algorithms and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 675, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 167, + 685 + ], + "score": 1.0, + "content": "record the first", + "type": "text" + }, + { + "bbox": [ + 167, + 677, + 173, + 685 + ], + "score": 0.73, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 675, + 244, + 685 + ], + "score": 1.0, + "content": "features chosen4.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 653, + 506, + 685 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 107, + 397, + 304 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 194, + 107, + 397, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 194, + 107, + 397, + 304 + ], + "spans": [ + { + "bbox": [ + 194, + 107, + 397, + 304 + ], + "score": 0.97, + "type": "image", + "image_path": "850628cf8a6350a15f50e62eeedfcba28e4ac2fa035fa393674ec77b7c764190.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 194, + 107, + 397, + 120.13333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 194, + 120.13333333333333, + 397, + 133.26666666666665 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 194, + 133.26666666666665, + 397, + 146.39999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 194, + 146.39999999999998, + 397, + 159.5333333333333 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 194, + 159.5333333333333, + 397, + 172.66666666666663 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 194, + 172.66666666666663, + 397, + 185.79999999999995 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 194, + 185.79999999999995, + 397, + 198.93333333333328 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 194, + 198.93333333333328, + 397, + 212.0666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 194, + 212.0666666666666, + 397, + 225.19999999999993 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 194, + 225.19999999999993, + 397, + 238.33333333333326 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 194, + 238.33333333333326, + 397, + 251.46666666666658 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 194, + 251.46666666666658, + 397, + 264.5999999999999 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 194, + 264.5999999999999, + 397, + 277.73333333333323 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 194, + 277.73333333333323, + 397, + 290.86666666666656 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 194, + 290.86666666666656, + 397, + 303.9999999999999 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 321, + 505, + 388 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "Figure 3: Empirical comparison of the dynamics of the toy model to OMP. The toy model has a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "depth of 5 and was initialized with a scale of 1e-4 and a learning rate of 3e-3. We compare the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 342, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 342, + 297, + 358 + ], + "score": 1.0, + "content": "fraction of agreement between the sets of first", + "type": "text" + }, + { + "bbox": [ + 297, + 346, + 303, + 354 + ], + "score": 0.62, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 342, + 505, + 358 + ], + "score": 1.0, + "content": "features selected of the two algorithms for every", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 188, + 367 + ], + "score": 1.0, + "content": "given sparsity level", + "type": "text" + }, + { + "bbox": [ + 189, + 357, + 195, + 365 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 355, + 505, + 367 + ], + "score": 1.0, + "content": ", averaged over 100 experiments (the shaded regions are empirical standard", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "deviations). For example, for sparsity level 3, we look at the sets of first 3 features selected by each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 376, + 381, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 381, + 389 + ], + "score": 1.0, + "content": "algorithm and calculate the fraction of them that appear in both sets.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + } + ], + "index": 12.25 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 182, + 423 + ], + "score": 1.0, + "content": "For every sparsity", + "type": "text" + }, + { + "bbox": [ + 183, + 414, + 189, + 421 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 412, + 505, + 423 + ], + "score": 1.0, + "content": ", we plot the mean fraction of agreement between the sets of features chosen", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "by OMP and the toy model over 100 experiments. We see that the two algorithms choose very", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "similar features at the beginning, suggesting that the deep model approximates the discrete behavior", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 444, + 504, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 504, + 456 + ], + "score": 1.0, + "content": "of OMP. Only when the number of features increases do we see that the behavior of the two models", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 454, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 506, + 469 + ], + "score": 1.0, + "content": "begins to differ, caused by the fact that the toy model has a finite initialization scale and learning", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 466, + 127, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 127, + 478 + ], + "score": 1.0, + "content": "rate.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "These experiments demonstrate the similarity between the incremental learning of deep models and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "the discrete behavior of greedy approximation algorithms such as OMP. Adopting this view also", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 505, + 504, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 517 + ], + "score": 1.0, + "content": "allows us to put our finger on another strength of the dynamics of deep models - while greedy algo-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "rithms such as OMP require the analytical solution or approximation of every iterate, the dynamics", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "of deep models are able to incrementally learn any differentiable function. For example, looking", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "back at the matrix sensing task and the classification models in section 4, we see that while there", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 547, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 104, + 547, + 506, + 562 + ], + "score": 1.0, + "content": "isn’t an immediate and efficient extension of OMP for these settings, the dynamics of learning deep", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 559, + 417, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 417, + 572 + ], + "score": 1.0, + "content": "models extends naturally and exhibits the same incremental learning as OMP.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 588, + 369, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 370, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 370, + 603 + ], + "score": 1.0, + "content": "E INCREMENTAL LEARNING IN MATRIX SENSING", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 107, + 614, + 229, + 625 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 230, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 230, + 626 + ], + "score": 1.0, + "content": "E.1 PROOF OF THEOREM 4", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 635, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "Theorem. Minimizing the deep matrix sensing model described in (5) with gradient flow over the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 646, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 506, + 658 + ], + "score": 1.0, + "content": "depth normalized squared loss (6), with Gaussian inputs and weights initialized as in (5) leads to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 657, + 350, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 335, + 670 + ], + "score": 1.0, + "content": "the following dynamical equations for different values of", + "type": "text" + }, + { + "bbox": [ + 335, + 658, + 345, + 667 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 657, + 350, + 670 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 684, + 368, + 700 + ], + "lines": [ + { + "bbox": [ + 242, + 684, + 368, + 700 + ], + "spans": [ + { + "bbox": [ + 242, + 684, + 368, + 700 + ], + "score": 0.91, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "26ecb4ca6a2dc65c84ab0680a444b7f51428246ecc707196e8cd0b3d744a6a73.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 242, + 684, + 368, + 700 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 135, + 723 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 711, + 145, + 721 + ], + "score": 0.82, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 708, + 164, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 711, + 175, + 722 + ], + "score": 0.88, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 708, + 294, + 723 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 711, + 306, + 720 + ], + "score": 0.77, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 708, + 325, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 711, + 341, + 720 + ], + "score": 0.85, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 708, + 505, + 723 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "score": 1.0, + "content": "singular vectors.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 107, + 397, + 304 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 194, + 107, + 397, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 194, + 107, + 397, + 304 + ], + "spans": [ + { + "bbox": [ + 194, + 107, + 397, + 304 + ], + "score": 0.97, + "type": "image", + "image_path": "850628cf8a6350a15f50e62eeedfcba28e4ac2fa035fa393674ec77b7c764190.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 194, + 107, + 397, + 120.13333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 194, + 120.13333333333333, + 397, + 133.26666666666665 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 194, + 133.26666666666665, + 397, + 146.39999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 194, + 146.39999999999998, + 397, + 159.5333333333333 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 194, + 159.5333333333333, + 397, + 172.66666666666663 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 194, + 172.66666666666663, + 397, + 185.79999999999995 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 194, + 185.79999999999995, + 397, + 198.93333333333328 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 194, + 198.93333333333328, + 397, + 212.0666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 194, + 212.0666666666666, + 397, + 225.19999999999993 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 194, + 225.19999999999993, + 397, + 238.33333333333326 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 194, + 238.33333333333326, + 397, + 251.46666666666658 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 194, + 251.46666666666658, + 397, + 264.5999999999999 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 194, + 264.5999999999999, + 397, + 277.73333333333323 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 194, + 277.73333333333323, + 397, + 290.86666666666656 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 194, + 290.86666666666656, + 397, + 303.9999999999999 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 321, + 505, + 388 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "Figure 3: Empirical comparison of the dynamics of the toy model to OMP. The toy model has a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "depth of 5 and was initialized with a scale of 1e-4 and a learning rate of 3e-3. We compare the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 342, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 342, + 297, + 358 + ], + "score": 1.0, + "content": "fraction of agreement between the sets of first", + "type": "text" + }, + { + "bbox": [ + 297, + 346, + 303, + 354 + ], + "score": 0.62, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 342, + 505, + 358 + ], + "score": 1.0, + "content": "features selected of the two algorithms for every", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 188, + 367 + ], + "score": 1.0, + "content": "given sparsity level", + "type": "text" + }, + { + "bbox": [ + 189, + 357, + 195, + 365 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 355, + 505, + 367 + ], + "score": 1.0, + "content": ", averaged over 100 experiments (the shaded regions are empirical standard", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "deviations). For example, for sparsity level 3, we look at the sets of first 3 features selected by each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 376, + 381, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 381, + 389 + ], + "score": 1.0, + "content": "algorithm and calculate the fraction of them that appear in both sets.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + } + ], + "index": 12.25 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 182, + 423 + ], + "score": 1.0, + "content": "For every sparsity", + "type": "text" + }, + { + "bbox": [ + 183, + 414, + 189, + 421 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 412, + 505, + 423 + ], + "score": 1.0, + "content": ", we plot the mean fraction of agreement between the sets of features chosen", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "by OMP and the toy model over 100 experiments. We see that the two algorithms choose very", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "similar features at the beginning, suggesting that the deep model approximates the discrete behavior", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 444, + 504, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 504, + 456 + ], + "score": 1.0, + "content": "of OMP. Only when the number of features increases do we see that the behavior of the two models", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 454, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 506, + 469 + ], + "score": 1.0, + "content": "begins to differ, caused by the fact that the toy model has a finite initialization scale and learning", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 466, + 127, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 127, + 478 + ], + "score": 1.0, + "content": "rate.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 412, + 506, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "These experiments demonstrate the similarity between the incremental learning of deep models and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "the discrete behavior of greedy approximation algorithms such as OMP. Adopting this view also", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 505, + 504, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 504, + 517 + ], + "score": 1.0, + "content": "allows us to put our finger on another strength of the dynamics of deep models - while greedy algo-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "rithms such as OMP require the analytical solution or approximation of every iterate, the dynamics", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "of deep models are able to incrementally learn any differentiable function. For example, looking", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "back at the matrix sensing task and the classification models in section 4, we see that while there", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 547, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 104, + 547, + 506, + 562 + ], + "score": 1.0, + "content": "isn’t an immediate and efficient extension of OMP for these settings, the dynamics of learning deep", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 559, + 417, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 417, + 572 + ], + "score": 1.0, + "content": "models extends naturally and exhibits the same incremental learning as OMP.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 483, + 506, + 572 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 588, + 369, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 370, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 370, + 603 + ], + "score": 1.0, + "content": "E INCREMENTAL LEARNING IN MATRIX SENSING", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 107, + 614, + 229, + 625 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 230, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 230, + 626 + ], + "score": 1.0, + "content": "E.1 PROOF OF THEOREM 4", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 635, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "Theorem. Minimizing the deep matrix sensing model described in (5) with gradient flow over the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 646, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 506, + 658 + ], + "score": 1.0, + "content": "depth normalized squared loss (6), with Gaussian inputs and weights initialized as in (5) leads to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 657, + 350, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 335, + 670 + ], + "score": 1.0, + "content": "the following dynamical equations for different values of", + "type": "text" + }, + { + "bbox": [ + 335, + 658, + 345, + 667 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 657, + 350, + 670 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 635, + 506, + 670 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 684, + 368, + 700 + ], + "lines": [ + { + "bbox": [ + 242, + 684, + 368, + 700 + ], + "spans": [ + { + "bbox": [ + 242, + 684, + 368, + 700 + ], + "score": 0.91, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "26ecb4ca6a2dc65c84ab0680a444b7f51428246ecc707196e8cd0b3d744a6a73.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 242, + 684, + 368, + 700 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 135, + 723 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 711, + 145, + 721 + ], + "score": 0.82, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 708, + 164, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 711, + 175, + 722 + ], + "score": 0.88, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 708, + 294, + 723 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 711, + 306, + 720 + ], + "score": 0.77, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 708, + 325, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 711, + 341, + 720 + ], + "score": 0.85, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 708, + 505, + 723 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "score": 1.0, + "content": "singular vectors.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Proof. We will adapt the proof from Saxe et al. (2013) for multilayer linear networks. The gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 257, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 181, + 107 + ], + "score": 1.0, + "content": "flow equations for", + "type": "text" + }, + { + "bbox": [ + 182, + 94, + 237, + 106 + ], + "score": 0.9, + "content": "W _ { n } , n \\in [ N ]", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 92, + 257, + 107 + ], + "score": 1.0, + "content": "are:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 118, + 382, + 142 + ], + "lines": [ + { + "bbox": [ + 228, + 118, + 382, + 142 + ], + "spans": [ + { + "bbox": [ + 228, + 118, + 382, + 142 + ], + "score": 0.94, + "content": "\\dot { W _ { n } } = { \\frac { 1 } { N } } W _ { 1 : n - 1 } ^ { T } ( W ^ { * } - W ) W _ { n + 1 : N } ^ { T }", + "type": "interline_equation", + "image_path": "e42e2d276924324141d11217a52cfca45b365a8a314885ae029a3964c269613c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 228, + 118, + 382, + 142 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 150, + 253, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 250, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 179, + 167 + ], + "score": 1.0, + "content": "Where we denote", + "type": "text" + }, + { + "bbox": [ + 179, + 149, + 250, + 165 + ], + "score": 0.95, + "content": "\\begin{array} { r } { W _ { j : k } = \\prod _ { i = j } ^ { k } W _ { i } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 169, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 184, + 182 + ], + "score": 1.0, + "content": "Since we assumed", + "type": "text" + }, + { + "bbox": [ + 184, + 170, + 200, + 180 + ], + "score": 0.89, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 169, + 424, + 182 + ], + "score": 1.0, + "content": "is (symmetric) PSD, there exists an orthogonal matrix", + "type": "text" + }, + { + "bbox": [ + 425, + 170, + 434, + 180 + ], + "score": 0.81, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 169, + 478, + 182 + ], + "score": 1.0, + "content": "for which", + "type": "text" + }, + { + "bbox": [ + 478, + 170, + 505, + 181 + ], + "score": 0.9, + "content": "D ^ { * } =", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 179, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 107, + 180, + 145, + 191 + ], + "score": 0.9, + "content": "U W ^ { * } U ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 179, + 174, + 194 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 174, + 181, + 188, + 191 + ], + "score": 0.88, + "content": "D ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 179, + 364, + 194 + ], + "score": 1.0, + "content": "is diagonal. Under the initialization in (3),", + "type": "text" + }, + { + "bbox": [ + 365, + 181, + 374, + 191 + ], + "score": 0.78, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 179, + 441, + 194 + ], + "score": 1.0, + "content": "diagonalizes all", + "type": "text" + }, + { + "bbox": [ + 441, + 181, + 457, + 192 + ], + "score": 0.9, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 179, + 506, + 194 + ], + "score": 1.0, + "content": "matrices at", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 190, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 199, + 205 + ], + "score": 1.0, + "content": "initialization such that", + "type": "text" + }, + { + "bbox": [ + 199, + 191, + 306, + 204 + ], + "score": 0.93, + "content": "{ \\cal D } _ { n } \\overline { { } } = U W _ { n } U ^ { T } = \\sqrt [ N ] { \\sigma _ { 0 } } I", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 190, + 470, + 205 + ], + "score": 1.0, + "content": ". Making this change of variables for all", + "type": "text" + }, + { + "bbox": [ + 471, + 192, + 486, + 203 + ], + "score": 0.9, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 190, + 506, + 205 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 203, + 125, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 203, + 125, + 217 + ], + "score": 1.0, + "content": "get:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 227, + 403, + 251 + ], + "lines": [ + { + "bbox": [ + 207, + 227, + 403, + 251 + ], + "spans": [ + { + "bbox": [ + 207, + 227, + 403, + 251 + ], + "score": 0.93, + "content": "\\dot { W _ { n } } = \\frac { 1 } { N } U ^ { T } D _ { 1 : n - 1 } U ( W ^ { * } - W ) U ^ { T } D _ { n + 1 : N } U", + "type": "interline_equation", + "image_path": "77582bad9b550554ae4889585a494609a229ea62d2a7364ae430273bc229eb84.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 207, + 227, + 403, + 251 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 257, + 472, + 269 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 469, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 453, + 272 + ], + "score": 1.0, + "content": "Rearranging, we get a set of decoupled differential equations for the singular values of", + "type": "text" + }, + { + "bbox": [ + 453, + 258, + 469, + 269 + ], + "score": 0.89, + "content": "W _ { n }", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 281, + 378, + 306 + ], + "lines": [ + { + "bbox": [ + 232, + 281, + 378, + 306 + ], + "spans": [ + { + "bbox": [ + 232, + 281, + 378, + 306 + ], + "score": 0.94, + "content": "\\dot { D _ { n } } = { \\frac { 1 } { N } } D _ { 1 : n - 1 } ( D ^ { * } - D ) D _ { n + 1 : N }", + "type": "interline_equation", + "image_path": "1fce7a2eb8034aece84f701ddd076d1426b3bea6ce8aece2702e5e3d034d46c8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 232, + 281, + 378, + 306 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 326 + ], + "score": 1.0, + "content": "Note that since these matrices are all diagonal at initialization, the above dynamics ensure that they", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 324, + 504, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 329, + 337 + ], + "score": 1.0, + "content": "remain diagonal throughout the optimization. Denoting", + "type": "text" + }, + { + "bbox": [ + 330, + 325, + 347, + 336 + ], + "score": 0.89, + "content": "\\sigma _ { n , i }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 324, + 373, + 337 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 374, + 325, + 379, + 334 + ], + "score": 0.68, + "content": "i ^ { \\because }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 324, + 459, + 337 + ], + "score": 1.0, + "content": "’th singular value of", + "type": "text" + }, + { + "bbox": [ + 460, + 324, + 475, + 335 + ], + "score": 0.9, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 324, + 493, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 494, + 325, + 504, + 335 + ], + "score": 0.83, + "content": "\\sigma _ { i }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 335, + 404, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 132, + 347 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 132, + 335, + 138, + 344 + ], + "score": 0.59, + "content": "\\overrightarrow { \\imath } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 335, + 218, + 347 + ], + "score": 1.0, + "content": "’th singular value of", + "type": "text" + }, + { + "bbox": [ + 218, + 335, + 230, + 344 + ], + "score": 0.8, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 335, + 404, + 347 + ], + "score": 1.0, + "content": ", we get the following differential equation:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 359, + 363, + 390 + ], + "lines": [ + { + "bbox": [ + 247, + 359, + 363, + 390 + ], + "spans": [ + { + "bbox": [ + 247, + 359, + 363, + 390 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { n , i } = \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ) \\prod _ { j \\neq n } \\sigma _ { j , i }", + "type": "interline_equation", + "image_path": "7ae62e0de9170e192f9c400a9e2263d45b0ae45e6eadd34184f1b74df67142cb.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 359, + 363, + 374.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 247, + 374.5, + 363, + 390.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 259, + 411 + ], + "score": 1.0, + "content": "Since we assume at initialization that", + "type": "text" + }, + { + "bbox": [ + 259, + 397, + 418, + 410 + ], + "score": 0.87, + "content": "\\forall n , m , i : \\sigma _ { n , i } ( 0 ) = \\sigma _ { m , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 397, + 505, + 411 + ], + "score": 1.0, + "content": ", the above dynamics", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 410, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 297, + 425 + ], + "score": 1.0, + "content": "are the same for all singular values and we get", + "type": "text" + }, + { + "bbox": [ + 297, + 410, + 464, + 423 + ], + "score": 0.88, + "content": "\\forall n , m , i : \\sigma _ { n , i } ( t ) = \\sigma _ { m , i } ( t ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 410, + 505, + 425 + ], + "score": 1.0, + "content": ". We may", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 356, + 434 + ], + "score": 1.0, + "content": "now use this to calculate the dynamics of the singular value of", + "type": "text" + }, + { + "bbox": [ + 356, + 423, + 367, + 432 + ], + "score": 0.78, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 422, + 505, + 434 + ], + "score": 1.0, + "content": ", since they are the product the the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 433, + 471, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 192, + 448 + ], + "score": 1.0, + "content": "singular values of all", + "type": "text" + }, + { + "bbox": [ + 192, + 433, + 208, + 444 + ], + "score": 0.89, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 433, + 288, + 448 + ], + "score": 1.0, + "content": "matrices. Denoting", + "type": "text" + }, + { + "bbox": [ + 289, + 433, + 368, + 447 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sigma _ { - n , i } = \\prod _ { k \\neq n } \\sigma _ { k , i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 433, + 471, + 448 + ], + "score": 1.0, + "content": "and using the chain rule:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 457, + 385, + 492 + ], + "lines": [ + { + "bbox": [ + 226, + 457, + 385, + 492 + ], + "spans": [ + { + "bbox": [ + 226, + 457, + 385, + 492 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { i } = \\sum _ { n = 1 } ^ { N } \\sigma _ { - n , i } \\dot { \\sigma } _ { n , i } = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "cb13b4e605be597a0d56eaacac3386f61d1aa8e54c7d344f34dbc1ca5c323726.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 457, + 385, + 474.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 226, + 474.5, + 385, + 492.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 523, + 246, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 247, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 247, + 536 + ], + "score": 1.0, + "content": "E.2 EMPIRICAL EXAMINATION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 543, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 504, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 504, + 557 + ], + "score": 1.0, + "content": "Our analytical results are only applicable for the population loss over Gaussian inputs. These condi-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "tions are far from the ones used in practice and studied in Arora et al. (2019), where the problem is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "over-determined and the weights are drawn from a Gaussian distribution with a small variance. To", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "show our conclusions regarding incremental learning extend qualitatively to more natural settings,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "we empirically examine the deep matrix sensing model in this natural setting for different depths", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 599, + 280, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 280, + 612 + ], + "score": 1.0, + "content": "and initialization scales as seen in figure 4.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Notice how incremental learning is exhibited even when the number of examples is much smaller", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "than the number of parameters in the model. While we can’t rely on our theory for describing the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "exact dynamics of the optimization for these kinds of over-determined problems, the qualitative", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 649, + 293, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 293, + 661 + ], + "score": 1.0, + "content": "conclusions we get from it are still applicable.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Another interesting phenomena we should note is that once the dataset becomes very small (the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "second row of the figure), we see all “currently active” singular values change at the beginning of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "every new phase (this is best seen in the bottom-right panel). This suggests that since there is more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "than one optimal solution, once we increase the current rank of our model it may find a solution that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "has a different set of singular values and vectors and thus all singular values change at the beginning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "of a new learning phase. This demonstrates the importance of incremental learning for obtaining", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 499, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 495, + 500, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 495, + 500, + 505, + 510 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Proof. We will adapt the proof from Saxe et al. (2013) for multilayer linear networks. The gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 257, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 181, + 107 + ], + "score": 1.0, + "content": "flow equations for", + "type": "text" + }, + { + "bbox": [ + 182, + 94, + 237, + 106 + ], + "score": 0.9, + "content": "W _ { n } , n \\in [ N ]", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 92, + 257, + 107 + ], + "score": 1.0, + "content": "are:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 107 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 118, + 382, + 142 + ], + "lines": [ + { + "bbox": [ + 228, + 118, + 382, + 142 + ], + "spans": [ + { + "bbox": [ + 228, + 118, + 382, + 142 + ], + "score": 0.94, + "content": "\\dot { W _ { n } } = { \\frac { 1 } { N } } W _ { 1 : n - 1 } ^ { T } ( W ^ { * } - W ) W _ { n + 1 : N } ^ { T }", + "type": "interline_equation", + "image_path": "e42e2d276924324141d11217a52cfca45b365a8a314885ae029a3964c269613c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 228, + 118, + 382, + 142 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 150, + 253, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 250, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 179, + 167 + ], + "score": 1.0, + "content": "Where we denote", + "type": "text" + }, + { + "bbox": [ + 179, + 149, + 250, + 165 + ], + "score": 0.95, + "content": "\\begin{array} { r } { W _ { j : k } = \\prod _ { i = j } ^ { k } W _ { i } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 148, + 250, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 169, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 184, + 182 + ], + "score": 1.0, + "content": "Since we assumed", + "type": "text" + }, + { + "bbox": [ + 184, + 170, + 200, + 180 + ], + "score": 0.89, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 169, + 424, + 182 + ], + "score": 1.0, + "content": "is (symmetric) PSD, there exists an orthogonal matrix", + "type": "text" + }, + { + "bbox": [ + 425, + 170, + 434, + 180 + ], + "score": 0.81, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 169, + 478, + 182 + ], + "score": 1.0, + "content": "for which", + "type": "text" + }, + { + "bbox": [ + 478, + 170, + 505, + 181 + ], + "score": 0.9, + "content": "D ^ { * } =", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 179, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 107, + 180, + 145, + 191 + ], + "score": 0.9, + "content": "U W ^ { * } U ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 179, + 174, + 194 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 174, + 181, + 188, + 191 + ], + "score": 0.88, + "content": "D ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 179, + 364, + 194 + ], + "score": 1.0, + "content": "is diagonal. Under the initialization in (3),", + "type": "text" + }, + { + "bbox": [ + 365, + 181, + 374, + 191 + ], + "score": 0.78, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 179, + 441, + 194 + ], + "score": 1.0, + "content": "diagonalizes all", + "type": "text" + }, + { + "bbox": [ + 441, + 181, + 457, + 192 + ], + "score": 0.9, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 179, + 506, + 194 + ], + "score": 1.0, + "content": "matrices at", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 190, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 199, + 205 + ], + "score": 1.0, + "content": "initialization such that", + "type": "text" + }, + { + "bbox": [ + 199, + 191, + 306, + 204 + ], + "score": 0.93, + "content": "{ \\cal D } _ { n } \\overline { { } } = U W _ { n } U ^ { T } = \\sqrt [ N ] { \\sigma _ { 0 } } I", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 190, + 470, + 205 + ], + "score": 1.0, + "content": ". Making this change of variables for all", + "type": "text" + }, + { + "bbox": [ + 471, + 192, + 486, + 203 + ], + "score": 0.9, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 190, + 506, + 205 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 203, + 125, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 203, + 125, + 217 + ], + "score": 1.0, + "content": "get:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 169, + 506, + 217 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 227, + 403, + 251 + ], + "lines": [ + { + "bbox": [ + 207, + 227, + 403, + 251 + ], + "spans": [ + { + "bbox": [ + 207, + 227, + 403, + 251 + ], + "score": 0.93, + "content": "\\dot { W _ { n } } = \\frac { 1 } { N } U ^ { T } D _ { 1 : n - 1 } U ( W ^ { * } - W ) U ^ { T } D _ { n + 1 : N } U", + "type": "interline_equation", + "image_path": "77582bad9b550554ae4889585a494609a229ea62d2a7364ae430273bc229eb84.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 207, + 227, + 403, + 251 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 257, + 472, + 269 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 469, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 453, + 272 + ], + "score": 1.0, + "content": "Rearranging, we get a set of decoupled differential equations for the singular values of", + "type": "text" + }, + { + "bbox": [ + 453, + 258, + 469, + 269 + ], + "score": 0.89, + "content": "W _ { n }", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 255, + 469, + 272 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 281, + 378, + 306 + ], + "lines": [ + { + "bbox": [ + 232, + 281, + 378, + 306 + ], + "spans": [ + { + "bbox": [ + 232, + 281, + 378, + 306 + ], + "score": 0.94, + "content": "\\dot { D _ { n } } = { \\frac { 1 } { N } } D _ { 1 : n - 1 } ( D ^ { * } - D ) D _ { n + 1 : N }", + "type": "interline_equation", + "image_path": "1fce7a2eb8034aece84f701ddd076d1426b3bea6ce8aece2702e5e3d034d46c8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 232, + 281, + 378, + 306 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 326 + ], + "score": 1.0, + "content": "Note that since these matrices are all diagonal at initialization, the above dynamics ensure that they", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 324, + 504, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 329, + 337 + ], + "score": 1.0, + "content": "remain diagonal throughout the optimization. Denoting", + "type": "text" + }, + { + "bbox": [ + 330, + 325, + 347, + 336 + ], + "score": 0.89, + "content": "\\sigma _ { n , i }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 324, + 373, + 337 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 374, + 325, + 379, + 334 + ], + "score": 0.68, + "content": "i ^ { \\because }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 324, + 459, + 337 + ], + "score": 1.0, + "content": "’th singular value of", + "type": "text" + }, + { + "bbox": [ + 460, + 324, + 475, + 335 + ], + "score": 0.9, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 324, + 493, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 494, + 325, + 504, + 335 + ], + "score": 0.83, + "content": "\\sigma _ { i }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 335, + 404, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 132, + 347 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 132, + 335, + 138, + 344 + ], + "score": 0.59, + "content": "\\overrightarrow { \\imath } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 335, + 218, + 347 + ], + "score": 1.0, + "content": "’th singular value of", + "type": "text" + }, + { + "bbox": [ + 218, + 335, + 230, + 344 + ], + "score": 0.8, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 335, + 404, + 347 + ], + "score": 1.0, + "content": ", we get the following differential equation:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 311, + 505, + 347 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 359, + 363, + 390 + ], + "lines": [ + { + "bbox": [ + 247, + 359, + 363, + 390 + ], + "spans": [ + { + "bbox": [ + 247, + 359, + 363, + 390 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { n , i } = \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ) \\prod _ { j \\neq n } \\sigma _ { j , i }", + "type": "interline_equation", + "image_path": "7ae62e0de9170e192f9c400a9e2263d45b0ae45e6eadd34184f1b74df67142cb.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 359, + 363, + 374.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 247, + 374.5, + 363, + 390.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 259, + 411 + ], + "score": 1.0, + "content": "Since we assume at initialization that", + "type": "text" + }, + { + "bbox": [ + 259, + 397, + 418, + 410 + ], + "score": 0.87, + "content": "\\forall n , m , i : \\sigma _ { n , i } ( 0 ) = \\sigma _ { m , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 397, + 505, + 411 + ], + "score": 1.0, + "content": ", the above dynamics", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 410, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 297, + 425 + ], + "score": 1.0, + "content": "are the same for all singular values and we get", + "type": "text" + }, + { + "bbox": [ + 297, + 410, + 464, + 423 + ], + "score": 0.88, + "content": "\\forall n , m , i : \\sigma _ { n , i } ( t ) = \\sigma _ { m , i } ( t ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 410, + 505, + 425 + ], + "score": 1.0, + "content": ". We may", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 356, + 434 + ], + "score": 1.0, + "content": "now use this to calculate the dynamics of the singular value of", + "type": "text" + }, + { + "bbox": [ + 356, + 423, + 367, + 432 + ], + "score": 0.78, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 422, + 505, + 434 + ], + "score": 1.0, + "content": ", since they are the product the the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 433, + 471, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 192, + 448 + ], + "score": 1.0, + "content": "singular values of all", + "type": "text" + }, + { + "bbox": [ + 192, + 433, + 208, + 444 + ], + "score": 0.89, + "content": "W _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 433, + 288, + 448 + ], + "score": 1.0, + "content": "matrices. Denoting", + "type": "text" + }, + { + "bbox": [ + 289, + 433, + 368, + 447 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sigma _ { - n , i } = \\prod _ { k \\neq n } \\sigma _ { k , i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 433, + 471, + 448 + ], + "score": 1.0, + "content": "and using the chain rule:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 397, + 505, + 448 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 457, + 385, + 492 + ], + "lines": [ + { + "bbox": [ + 226, + 457, + 385, + 492 + ], + "spans": [ + { + "bbox": [ + 226, + 457, + 385, + 492 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { i } = \\sum _ { n = 1 } ^ { N } \\sigma _ { - n , i } \\dot { \\sigma } _ { n , i } = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "cb13b4e605be597a0d56eaacac3386f61d1aa8e54c7d344f34dbc1ca5c323726.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 457, + 385, + 474.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 226, + 474.5, + 385, + 492.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 523, + 246, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 247, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 247, + 536 + ], + "score": 1.0, + "content": "E.2 EMPIRICAL EXAMINATION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 543, + 505, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 504, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 504, + 557 + ], + "score": 1.0, + "content": "Our analytical results are only applicable for the population loss over Gaussian inputs. These condi-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "tions are far from the ones used in practice and studied in Arora et al. (2019), where the problem is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "over-determined and the weights are drawn from a Gaussian distribution with a small variance. To", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "show our conclusions regarding incremental learning extend qualitatively to more natural settings,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "we empirically examine the deep matrix sensing model in this natural setting for different depths", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 599, + 280, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 280, + 612 + ], + "score": 1.0, + "content": "and initialization scales as seen in figure 4.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 544, + 505, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Notice how incremental learning is exhibited even when the number of examples is much smaller", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "than the number of parameters in the model. While we can’t rely on our theory for describing the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "exact dynamics of the optimization for these kinds of over-determined problems, the qualitative", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 649, + 293, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 293, + 661 + ], + "score": 1.0, + "content": "conclusions we get from it are still applicable.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 616, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Another interesting phenomena we should note is that once the dataset becomes very small (the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "second row of the figure), we see all “currently active” singular values change at the beginning of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "every new phase (this is best seen in the bottom-right panel). This suggests that since there is more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "than one optimal solution, once we increase the current rank of our model it may find a solution that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "has a different set of singular values and vectors and thus all singular values change at the beginning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "of a new learning phase. This demonstrates the importance of incremental learning for obtaining", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "sparse solutions - once the initialization conditions and depth are such that the learning phases are", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 301, + 414 + ], + "score": 1.0, + "content": "distinct, gradient descent finds the optimal rank-", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 302, + 402, + 306, + 411 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 307, + 401, + 403, + 414 + ], + "score": 1.0, + "content": "solution in every phase", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 403, + 402, + 408, + 411 + ], + "score": 0.3, + "content": "i", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 408, + 401, + 505, + 414 + ], + "score": 1.0, + "content": ". For these dynamics to", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "successfully recover the optimal solution at every phase, the phases need to be far enough apart from", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 423, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 482, + 436 + ], + "score": 1.0, + "content": "each other to allow for the singular values and vectors to change before the next phase begins.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 81, + 501, + 277 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 81, + 501, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 81, + 501, + 277 + ], + "spans": [ + { + "bbox": [ + 111, + 81, + 501, + 277 + ], + "score": 0.976, + "type": "image", + "image_path": "30d0053940d521183eae431a6073eb95e1a63934ce44ff4a2a09d223b98048e2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 81, + 501, + 146.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 146.33333333333331, + 501, + 211.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 211.66666666666663, + 501, + 276.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 293, + 505, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Figure 4: Evolution of the top-5 singular values of the deep matrix sensing model, with Gaussian", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "score": 1.0, + "content": "initialization with variance such that the initial singular values are in expectation 1e-4. The model’s", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 312, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 104, + 312, + 184, + 329 + ], + "score": 1.0, + "content": "size and data are in", + "type": "text" + }, + { + "bbox": [ + 184, + 314, + 215, + 325 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { 5 0 \\times 5 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 312, + 506, + 329 + ], + "score": 1.0, + "content": ". The columns correspond to different parameterization depths, while the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 326, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "rows correspond to different dataset sizes. In both cases the problem is over-determined, since the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 337, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 505, + 348 + ], + "score": 1.0, + "content": "number of examples is smaller than the number of parameters. Since the original matrix is rank-4,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "we can recognize an unsuccessful recovery when all five singular values are nonzero, as seen clearly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 358, + 198, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 198, + 372 + ], + "score": 1.0, + "content": "for both depth-1 plots.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "sparse solutions - once the initialization conditions and depth are such that the learning phases are", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 301, + 414 + ], + "score": 1.0, + "content": "distinct, gradient descent finds the optimal rank-", + "type": "text" + }, + { + "bbox": [ + 302, + 402, + 306, + 411 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 401, + 403, + 414 + ], + "score": 1.0, + "content": "solution in every phase", + "type": "text" + }, + { + "bbox": [ + 403, + 402, + 408, + 411 + ], + "score": 0.3, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 401, + 505, + 414 + ], + "score": 1.0, + "content": ". For these dynamics to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "successfully recover the optimal solution at every phase, the phases need to be far enough apart from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 423, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 482, + 436 + ], + "score": 1.0, + "content": "each other to allow for the singular values and vectors to change before the next phase begins.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 402, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 403, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 403, + 466 + ], + "score": 1.0, + "content": "F INCREMENTAL LEARNING IN QUADRATIC NETWORKS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 228, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 228, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 228, + 487 + ], + "score": 1.0, + "content": "F.1 PROOF OF THEOREM 5", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "Theorem. Minimizing the quadratic network described and initialized as in (7) with gradient flow", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "over the variance loss defined in (2) with Gaussian inputs leads to the following dynamical equa-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 517, + 132, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 132, + 531 + ], + "score": 1.0, + "content": "tions:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 544, + 359, + 558 + ], + "lines": [ + { + "bbox": [ + 252, + 544, + 359, + 558 + ], + "spans": [ + { + "bbox": [ + 252, + 544, + 359, + 558 + ], + "score": 0.9, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "2c3723674660a152f88205f30c94626451ce52242b7952c621e4f5f1203b0993.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 252, + 544, + 359, + 558 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 504, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 135, + 579 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 567, + 145, + 576 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 563, + 163, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 566, + 175, + 577 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 563, + 294, + 579 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 565, + 306, + 576 + ], + "score": 0.75, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 563, + 325, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 565, + 341, + 576 + ], + "score": 0.87, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 563, + 506, + 579 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 576, + 176, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 176, + 589 + ], + "score": 1.0, + "content": "singular vectors.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 599, + 503, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 614 + ], + "score": 1.0, + "content": "Proof. Our proof will follow similar lines as the analysis of the deep matrix sensing model. Taking", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 610, + 429, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 429, + 624 + ], + "score": 1.0, + "content": "the expectation of the variance loss over Gaussian inputs for our model gives us:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 640, + 498, + 713 + ], + "lines": [ + { + "bbox": [ + 113, + 640, + 498, + 713 + ], + "spans": [ + { + "bbox": [ + 113, + 640, + 498, + 713 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\ell _ { v a r } ( W ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ] ^ { 2 } } \\\\ { \\displaystyle = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } - \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\\\ { \\displaystyle \\quad = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "33ab1a67e9b6848f6dc0f8ea52ab043b39cdff03886a7d0ecfe294b486f6701f.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 113, + 640, + 498, + 664.3333333333334 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 113, + 664.3333333333334, + 498, + 688.6666666666667 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 113, + 688.6666666666667, + 498, + 713.0000000000001 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 103, + 720, + 468, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 468, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 280, + 734 + ], + "score": 1.0, + "content": "Following the gradient flow dynamics over", + "type": "text" + }, + { + "bbox": [ + 280, + 722, + 291, + 730 + ], + "score": 0.77, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 719, + 468, + 734 + ], + "score": 1.0, + "content": "leads to the following differential equation:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 81, + 501, + 277 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 81, + 501, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 81, + 501, + 277 + ], + "spans": [ + { + "bbox": [ + 111, + 81, + 501, + 277 + ], + "score": 0.976, + "type": "image", + "image_path": "30d0053940d521183eae431a6073eb95e1a63934ce44ff4a2a09d223b98048e2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 81, + 501, + 146.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 146.33333333333331, + 501, + 211.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 211.66666666666663, + 501, + 276.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 293, + 505, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Figure 4: Evolution of the top-5 singular values of the deep matrix sensing model, with Gaussian", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "score": 1.0, + "content": "initialization with variance such that the initial singular values are in expectation 1e-4. The model’s", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 312, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 104, + 312, + 184, + 329 + ], + "score": 1.0, + "content": "size and data are in", + "type": "text" + }, + { + "bbox": [ + 184, + 314, + 215, + 325 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { 5 0 \\times 5 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 312, + 506, + 329 + ], + "score": 1.0, + "content": ". The columns correspond to different parameterization depths, while the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 326, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "rows correspond to different dataset sizes. In both cases the problem is over-determined, since the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 337, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 505, + 348 + ], + "score": 1.0, + "content": "number of examples is smaller than the number of parameters. Since the original matrix is rank-4,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "we can recognize an unsuccessful recovery when all five singular values are nonzero, as seen clearly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 358, + 198, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 198, + 372 + ], + "score": 1.0, + "content": "for both depth-1 plots.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 435 + ], + "lines": [], + "index": 11.5, + "bbox_fs": [ + 105, + 390, + 506, + 436 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 402, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 403, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 403, + 466 + ], + "score": 1.0, + "content": "F INCREMENTAL LEARNING IN QUADRATIC NETWORKS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 449, + 403, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 228, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 228, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 228, + 487 + ], + "score": 1.0, + "content": "F.1 PROOF OF THEOREM 5", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 474, + 228, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "Theorem. Minimizing the quadratic network described and initialized as in (7) with gradient flow", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "over the variance loss defined in (2) with Gaussian inputs leads to the following dynamical equa-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 517, + 132, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 132, + 531 + ], + "score": 1.0, + "content": "tions:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 495, + 505, + 531 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 544, + 359, + 558 + ], + "lines": [ + { + "bbox": [ + 252, + 544, + 359, + 558 + ], + "spans": [ + { + "bbox": [ + 252, + 544, + 359, + 558 + ], + "score": 0.9, + "content": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )", + "type": "interline_equation", + "image_path": "2c3723674660a152f88205f30c94626451ce52242b7952c621e4f5f1203b0993.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 252, + 544, + 359, + 558 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 504, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 135, + 579 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 135, + 567, + 145, + 576 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 563, + 163, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 566, + 175, + 577 + ], + "score": 0.89, + "content": "\\sigma _ { i } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 563, + 294, + 579 + ], + "score": 1.0, + "content": "are the ith singular values of", + "type": "text" + }, + { + "bbox": [ + 295, + 565, + 306, + 576 + ], + "score": 0.75, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 563, + 325, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 325, + 565, + 341, + 576 + ], + "score": 0.87, + "content": "W ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 563, + 506, + 579 + ], + "score": 1.0, + "content": ", respectively, corresponding to the same", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 576, + 176, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 176, + 589 + ], + "score": 1.0, + "content": "singular vectors.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 563, + 506, + 589 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 599, + 503, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 614 + ], + "score": 1.0, + "content": "Proof. Our proof will follow similar lines as the analysis of the deep matrix sensing model. Taking", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 610, + 429, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 429, + 624 + ], + "score": 1.0, + "content": "the expectation of the variance loss over Gaussian inputs for our model gives us:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 598, + 505, + 624 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 640, + 498, + 713 + ], + "lines": [ + { + "bbox": [ + 113, + 640, + 498, + 713 + ], + "spans": [ + { + "bbox": [ + 113, + 640, + 498, + 713 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\ell _ { v a r } ( W ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ] ^ { 2 } } \\\\ { \\displaystyle = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } - \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\\\ { \\displaystyle \\quad = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "33ab1a67e9b6848f6dc0f8ea52ab043b39cdff03886a7d0ecfe294b486f6701f.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 113, + 640, + 498, + 664.3333333333334 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 113, + 664.3333333333334, + 498, + 688.6666666666667 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 113, + 688.6666666666667, + 498, + 713.0000000000001 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 103, + 720, + 468, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 468, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 280, + 734 + ], + "score": 1.0, + "content": "Following the gradient flow dynamics over", + "type": "text" + }, + { + "bbox": [ + 280, + 722, + 291, + 730 + ], + "score": 0.77, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 719, + 468, + 734 + ], + "score": 1.0, + "content": "leads to the following differential equation:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 719, + 468, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 243, + 90, + 367, + 114 + ], + "lines": [ + { + "bbox": [ + 243, + 90, + 367, + 114 + ], + "spans": [ + { + "bbox": [ + 243, + 90, + 367, + 114 + ], + "score": 0.93, + "content": "\\dot { W } = \\frac { 1 } { 2 } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W )", + "type": "interline_equation", + "image_path": "ca35af0726e0ad8d69915e77afade54f7497d063e6eac2091df983edb20d80c1.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 243, + 90, + 367, + 114 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 430, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 431, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 317, + 137 + ], + "score": 1.0, + "content": "We can now calculate the gradient flow dynamics of", + "type": "text" + }, + { + "bbox": [ + 317, + 122, + 346, + 133 + ], + "score": 0.91, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 120, + 431, + 137 + ], + "score": 1.0, + "content": "using the chain rule:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 153, + 482, + 178 + ], + "lines": [ + { + "bbox": [ + 111, + 153, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 111, + 153, + 482, + 178 + ], + "score": 0.94, + "content": "W ^ { \\dot { T } } W = W ^ { T } \\dot { W } + \\dot { W } ^ { T } W = \\frac { 1 } { 2 } \\Bigl ( W ^ { T } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) + ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) W ^ { T } W \\Bigr )", + "type": "interline_equation", + "image_path": "b1d6d2741c130e57df027aa47d1150d94b31ad4d63b01c7f415ce7b07f6dd70f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 111, + 153, + 482, + 178 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 188, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 227, + 202 + ], + "score": 1.0, + "content": "Now, under our initialization", + "type": "text" + }, + { + "bbox": [ + 227, + 189, + 291, + 201 + ], + "score": 0.91, + "content": "W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I \\mathrm { . }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 187, + 344, + 202 + ], + "score": 1.0, + "content": ", we get that", + "type": "text" + }, + { + "bbox": [ + 344, + 189, + 373, + 200 + ], + "score": 0.9, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 187, + 392, + 202 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 393, + 189, + 425, + 201 + ], + "score": 0.92, + "content": "W _ { * } ^ { T } W _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 187, + 505, + 202 + ], + "score": 1.0, + "content": "are simultaneously", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 200, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 293, + 212 + ], + "score": 1.0, + "content": "diagonalizable at initialization by some matrix", + "type": "text" + }, + { + "bbox": [ + 294, + 201, + 302, + 210 + ], + "score": 0.83, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 200, + 477, + 212 + ], + "score": 1.0, + "content": ", such that the following is true for diagonal", + "type": "text" + }, + { + "bbox": [ + 478, + 201, + 487, + 210 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 200, + 505, + 212 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 210, + 127, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 212, + 120, + 221 + ], + "score": 0.84, + "content": "D ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 210, + 127, + 223 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 236, + 362, + 252 + ], + "lines": [ + { + "bbox": [ + 249, + 236, + 362, + 252 + ], + "spans": [ + { + "bbox": [ + 249, + 236, + 362, + 252 + ], + "score": 0.84, + "content": "D ( 0 ) = U ^ { T } W ( 0 ) ^ { T } W ( 0 ) U", + "type": "interline_equation", + "image_path": "a9c4edfa8ae905caf3cd41b198a1de4a32d1b65aedff0615d3fa20f5435c5a9d.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 249, + 236, + 362, + 252 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 268, + 347, + 283 + ], + "lines": [ + { + "bbox": [ + 264, + 268, + 347, + 283 + ], + "spans": [ + { + "bbox": [ + 264, + 268, + 347, + 283 + ], + "score": 0.89, + "content": "D ^ { * } = U ^ { T } W _ { * } ^ { T } W _ { * } U", + "type": "interline_equation", + "image_path": "e9b51eae83acb8130f277fe19d230dedd5977780133543aba25c77ac48bd25e4.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 264, + 268, + 347, + 283 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 290, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 226, + 305 + ], + "score": 1.0, + "content": "Multiplying equation (10) by", + "type": "text" + }, + { + "bbox": [ + 226, + 292, + 235, + 302 + ], + "score": 0.84, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 289, + 253, + 305 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 254, + 291, + 269, + 302 + ], + "score": 0.89, + "content": "U ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 289, + 506, + 305 + ], + "score": 1.0, + "content": "gives us the following dynamics for the singular values of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 301, + 141, + 315 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 135, + 313 + ], + "score": 0.86, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 301, + 141, + 315 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 326, + 415, + 351 + ], + "lines": [ + { + "bbox": [ + 196, + 326, + 415, + 351 + ], + "spans": [ + { + "bbox": [ + 196, + 326, + 415, + 351 + ], + "score": 0.92, + "content": "\\dot { D } = \\frac { 1 } { 2 } \\Big ( D ( D ^ { * } - D ) + ( D ^ { * } - D ) D \\Big ) = D ( D ^ { * } - D )", + "type": "interline_equation", + "image_path": "f1d394e05a751c376879c6cd9bd491a4c5719defb884874224541df06e35b646.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 196, + 326, + 415, + 351 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 504, + 374 + ], + "score": 1.0, + "content": "These matrices are diagonal at initialization, and remain diagonal throughout the dynamics (the off-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "diagonal elements are static according to these equations). We may now look at the dynamics of a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 384, + 405, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 405, + 396 + ], + "score": 1.0, + "content": "single diagonal element, noticing it is equivalent to the depth-2 toy model:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 269, + 410, + 342, + 424 + ], + "lines": [ + { + "bbox": [ + 269, + 410, + 342, + 424 + ], + "spans": [ + { + "bbox": [ + 269, + 410, + 342, + 424 + ], + "score": 0.91, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "c12c96e75c3ba1ff7288ec15a7ba1234aa02106bcc232f412bf93e76e878e246.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 269, + 410, + 342, + 424 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 457, + 291, + 469 + ], + "lines": [ + { + "bbox": [ + 104, + 456, + 293, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 456, + 293, + 470 + ], + "score": 1.0, + "content": "F.2 DISCUSSION OF THE VARIANCE LOSS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "It may seem that the variance loss is an unnatural loss function to analyze, since it isn’t used in prac-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "tice. While this is true, we will show how the dynamics of this loss function are an approximation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 499, + 221, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 221, + 513 + ], + "score": 1.0, + "content": "of the square loss dynamics.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "We begin by describing the dynamics of both losses, showing how incremental learning can’t take", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "place for quadratic networks as defined over the squared loss. Then, we show how adding a global", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 538, + 448, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 448, + 551 + ], + "score": 1.0, + "content": "bias to the quadratic network leads to similar dynamics for small initialization scales.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 561, + 250, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 251, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 251, + 574 + ], + "score": 1.0, + "content": "F.2.1 DYNAMICAL DERIVATION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 449, + 594 + ], + "score": 1.0, + "content": "in the previous section, we derive the differential equations for the singular values of", + "type": "text" + }, + { + "bbox": [ + 450, + 581, + 478, + 591 + ], + "score": 0.9, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "under", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 592, + 178, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 178, + 604 + ], + "score": 1.0, + "content": "the variance loss:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 269, + 619, + 342, + 633 + ], + "lines": [ + { + "bbox": [ + 269, + 619, + 342, + 633 + ], + "spans": [ + { + "bbox": [ + 269, + 619, + 342, + 633 + ], + "score": 0.88, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "67100e570fb3f2842452968d2b157ad3a4def67113be508f28b081040a0c2ff8.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 269, + 619, + 342, + 633 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 640, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "We will now derive similar equations for the squared loss. The scaled squared loss in expectation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 653, + 219, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 219, + 664 + ], + "score": 1.0, + "content": "over the Gaussian inputs is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 683, + 432, + 733 + ], + "lines": [ + { + "bbox": [ + 179, + 683, + 432, + 733 + ], + "spans": [ + { + "bbox": [ + 179, + 683, + 432, + 733 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\ell ( W ) = \\displaystyle \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] } \\\\ & { \\qquad = \\displaystyle \\frac { 1 } { 8 } | | W _ { * } ^ { T } W _ { * } - W ^ { T } W | | _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "e082b4c28eb6e542bc04000b3c7f672951bbfd35a72df9d25e6cef7a6eef3549.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 179, + 683, + 432, + 699.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 179, + 699.6666666666666, + 432, + 716.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 179, + 716.3333333333333, + 432, + 732.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 432, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 495, + 434, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 495, + 434, + 505, + 444 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 243, + 90, + 367, + 114 + ], + "lines": [ + { + "bbox": [ + 243, + 90, + 367, + 114 + ], + "spans": [ + { + "bbox": [ + 243, + 90, + 367, + 114 + ], + "score": 0.93, + "content": "\\dot { W } = \\frac { 1 } { 2 } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W )", + "type": "interline_equation", + "image_path": "ca35af0726e0ad8d69915e77afade54f7497d063e6eac2091df983edb20d80c1.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 243, + 90, + 367, + 114 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 430, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 431, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 317, + 137 + ], + "score": 1.0, + "content": "We can now calculate the gradient flow dynamics of", + "type": "text" + }, + { + "bbox": [ + 317, + 122, + 346, + 133 + ], + "score": 0.91, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 120, + 431, + 137 + ], + "score": 1.0, + "content": "using the chain rule:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 120, + 431, + 137 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 153, + 482, + 178 + ], + "lines": [ + { + "bbox": [ + 111, + 153, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 111, + 153, + 482, + 178 + ], + "score": 0.94, + "content": "W ^ { \\dot { T } } W = W ^ { T } \\dot { W } + \\dot { W } ^ { T } W = \\frac { 1 } { 2 } \\Bigl ( W ^ { T } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) + ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) W ^ { T } W \\Bigr )", + "type": "interline_equation", + "image_path": "b1d6d2741c130e57df027aa47d1150d94b31ad4d63b01c7f415ce7b07f6dd70f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 111, + 153, + 482, + 178 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 188, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 227, + 202 + ], + "score": 1.0, + "content": "Now, under our initialization", + "type": "text" + }, + { + "bbox": [ + 227, + 189, + 291, + 201 + ], + "score": 0.91, + "content": "W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I \\mathrm { . }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 187, + 344, + 202 + ], + "score": 1.0, + "content": ", we get that", + "type": "text" + }, + { + "bbox": [ + 344, + 189, + 373, + 200 + ], + "score": 0.9, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 187, + 392, + 202 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 393, + 189, + 425, + 201 + ], + "score": 0.92, + "content": "W _ { * } ^ { T } W _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 187, + 505, + 202 + ], + "score": 1.0, + "content": "are simultaneously", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 200, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 293, + 212 + ], + "score": 1.0, + "content": "diagonalizable at initialization by some matrix", + "type": "text" + }, + { + "bbox": [ + 294, + 201, + 302, + 210 + ], + "score": 0.83, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 200, + 477, + 212 + ], + "score": 1.0, + "content": ", such that the following is true for diagonal", + "type": "text" + }, + { + "bbox": [ + 478, + 201, + 487, + 210 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 200, + 505, + 212 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 210, + 127, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 212, + 120, + 221 + ], + "score": 0.84, + "content": "D ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 210, + 127, + 223 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 187, + 505, + 223 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 236, + 362, + 252 + ], + "lines": [ + { + "bbox": [ + 249, + 236, + 362, + 252 + ], + "spans": [ + { + "bbox": [ + 249, + 236, + 362, + 252 + ], + "score": 0.84, + "content": "D ( 0 ) = U ^ { T } W ( 0 ) ^ { T } W ( 0 ) U", + "type": "interline_equation", + "image_path": "a9c4edfa8ae905caf3cd41b198a1de4a32d1b65aedff0615d3fa20f5435c5a9d.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 249, + 236, + 362, + 252 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 268, + 347, + 283 + ], + "lines": [ + { + "bbox": [ + 264, + 268, + 347, + 283 + ], + "spans": [ + { + "bbox": [ + 264, + 268, + 347, + 283 + ], + "score": 0.89, + "content": "D ^ { * } = U ^ { T } W _ { * } ^ { T } W _ { * } U", + "type": "interline_equation", + "image_path": "e9b51eae83acb8130f277fe19d230dedd5977780133543aba25c77ac48bd25e4.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 264, + 268, + 347, + 283 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 290, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 104, + 289, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 226, + 305 + ], + "score": 1.0, + "content": "Multiplying equation (10) by", + "type": "text" + }, + { + "bbox": [ + 226, + 292, + 235, + 302 + ], + "score": 0.84, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 289, + 253, + 305 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 254, + 291, + 269, + 302 + ], + "score": 0.89, + "content": "U ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 289, + 506, + 305 + ], + "score": 1.0, + "content": "gives us the following dynamics for the singular values of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 301, + 141, + 315 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 135, + 313 + ], + "score": 0.86, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 301, + 141, + 315 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 289, + 506, + 315 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 196, + 326, + 415, + 351 + ], + "lines": [ + { + "bbox": [ + 196, + 326, + 415, + 351 + ], + "spans": [ + { + "bbox": [ + 196, + 326, + 415, + 351 + ], + "score": 0.92, + "content": "\\dot { D } = \\frac { 1 } { 2 } \\Big ( D ( D ^ { * } - D ) + ( D ^ { * } - D ) D \\Big ) = D ( D ^ { * } - D )", + "type": "interline_equation", + "image_path": "f1d394e05a751c376879c6cd9bd491a4c5719defb884874224541df06e35b646.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 196, + 326, + 415, + 351 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 504, + 374 + ], + "score": 1.0, + "content": "These matrices are diagonal at initialization, and remain diagonal throughout the dynamics (the off-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "diagonal elements are static according to these equations). We may now look at the dynamics of a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 384, + 405, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 405, + 396 + ], + "score": 1.0, + "content": "single diagonal element, noticing it is equivalent to the depth-2 toy model:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 362, + 506, + 396 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 269, + 410, + 342, + 424 + ], + "lines": [ + { + "bbox": [ + 269, + 410, + 342, + 424 + ], + "spans": [ + { + "bbox": [ + 269, + 410, + 342, + 424 + ], + "score": 0.91, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "c12c96e75c3ba1ff7288ec15a7ba1234aa02106bcc232f412bf93e76e878e246.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 269, + 410, + 342, + 424 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 457, + 291, + 469 + ], + "lines": [ + { + "bbox": [ + 104, + 456, + 293, + 470 + ], + "spans": [ + { + "bbox": [ + 104, + 456, + 293, + 470 + ], + "score": 1.0, + "content": "F.2 DISCUSSION OF THE VARIANCE LOSS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "It may seem that the variance loss is an unnatural loss function to analyze, since it isn’t used in prac-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "tice. While this is true, we will show how the dynamics of this loss function are an approximation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 499, + 221, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 221, + 513 + ], + "score": 1.0, + "content": "of the square loss dynamics.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 477, + 505, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "We begin by describing the dynamics of both losses, showing how incremental learning can’t take", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "place for quadratic networks as defined over the squared loss. Then, we show how adding a global", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 538, + 448, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 448, + 551 + ], + "score": 1.0, + "content": "bias to the quadratic network leads to similar dynamics for small initialization scales.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 516, + 505, + 551 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 561, + 250, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 251, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 251, + 574 + ], + "score": 1.0, + "content": "F.2.1 DYNAMICAL DERIVATION", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 449, + 594 + ], + "score": 1.0, + "content": "in the previous section, we derive the differential equations for the singular values of", + "type": "text" + }, + { + "bbox": [ + 450, + 581, + 478, + 591 + ], + "score": 0.9, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "under", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 592, + 178, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 178, + 604 + ], + "score": 1.0, + "content": "the variance loss:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 581, + 505, + 604 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 269, + 619, + 342, + 633 + ], + "lines": [ + { + "bbox": [ + 269, + 619, + 342, + 633 + ], + "spans": [ + { + "bbox": [ + 269, + 619, + 342, + 633 + ], + "score": 0.88, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )", + "type": "interline_equation", + "image_path": "67100e570fb3f2842452968d2b157ad3a4def67113be508f28b081040a0c2ff8.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 269, + 619, + 342, + 633 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 640, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "We will now derive similar equations for the squared loss. The scaled squared loss in expectation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 653, + 219, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 219, + 664 + ], + "score": 1.0, + "content": "over the Gaussian inputs is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 640, + 505, + 664 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 683, + 432, + 733 + ], + "lines": [ + { + "bbox": [ + 179, + 683, + 432, + 733 + ], + "spans": [ + { + "bbox": [ + 179, + 683, + 432, + 733 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\ell ( W ) = \\displaystyle \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] } \\\\ & { \\qquad = \\displaystyle \\frac { 1 } { 8 } | | W _ { * } ^ { T } W _ { * } - W ^ { T } W | | _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "e082b4c28eb6e542bc04000b3c7f672951bbfd35a72df9d25e6cef7a6eef3549.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 179, + 683, + 432, + 699.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 179, + 699.6666666666666, + 432, + 716.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 179, + 716.3333333333333, + 432, + 732.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 450, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 452, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 144, + 97 + ], + "score": 1.0, + "content": "Defining", + "type": "text" + }, + { + "bbox": [ + 144, + 82, + 238, + 94 + ], + "score": 0.93, + "content": "\\Delta = W _ { * } ^ { T } W _ { * } - W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 79, + 401, + 97 + ], + "score": 1.0, + "content": "for brevity, the differential equations for", + "type": "text" + }, + { + "bbox": [ + 402, + 83, + 413, + 93 + ], + "score": 0.69, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 79, + 452, + 97 + ], + "score": 1.0, + "content": "become:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 106, + 361, + 131 + ], + "lines": [ + { + "bbox": [ + 249, + 106, + 361, + 131 + ], + "spans": [ + { + "bbox": [ + 249, + 106, + 361, + 131 + ], + "score": 0.94, + "content": "\\dot { W } = \\frac { 1 } { 2 } W \\Delta + \\frac { 1 } { 4 } T r ( \\Delta ) W", + "type": "interline_equation", + "image_path": "890dcfaeec2b118f56d3a15c72b6c45a986472d6e54582ca2f59ae9d6c614c05.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 249, + 106, + 361, + 131 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 139, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 139, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 219, + 153 + ], + "score": 1.0, + "content": "Calculating the dynamics of", + "type": "text" + }, + { + "bbox": [ + 219, + 139, + 248, + 150 + ], + "score": 0.91, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 139, + 471, + 153 + ], + "score": 1.0, + "content": "after noting that it is simultaneously diagonalizable with", + "type": "text" + }, + { + "bbox": [ + 472, + 139, + 504, + 152 + ], + "score": 0.91, + "content": "W _ { * } ^ { T } W _ { * }", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 1.0, + "content": "(as in the derivation for the variance loss) leads to the following differential equations for the singular", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 162, + 178, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 145, + 173 + ], + "score": 1.0, + "content": "values of", + "type": "text" + }, + { + "bbox": [ + 145, + 162, + 174, + 172 + ], + "score": 0.91, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 162, + 178, + 173 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 186, + 382, + 217 + ], + "lines": [ + { + "bbox": [ + 228, + 186, + 382, + 217 + ], + "spans": [ + { + "bbox": [ + 228, + 186, + 382, + 217 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) \\Big )", + "type": "interline_equation", + "image_path": "3621aa3a2f9b0623dfe6b0accc893aadfeee631cbc164503d5c89a3b5e4f1d24.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 186, + 382, + 201.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 228, + 201.5, + 382, + 217.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 237 + ], + "score": 1.0, + "content": "We see that the equations are now coupled and so we cannot solve them analytically. Another issue", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "is that for our initialization, all singular values have very similar dynamics at initialization due to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "the coupling. For example, values corresponding to small optimal singular values grow much faster", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 258, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 504, + 271 + ], + "score": 1.0, + "content": "than in the variance loss dynamics, due to the effect the large optimal singular values have on them.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 504, + 288 + ], + "score": 1.0, + "content": "We see from these equations that we shouldn’t expect quadratic networks optimized over the squared", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "loss to exhibit incremental learning behavior. We next show how adding a global bias to our model", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 297, + 145, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 145, + 311 + ], + "score": 1.0, + "content": "can help.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 106, + 320, + 370, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 370, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 370, + 335 + ], + "score": 1.0, + "content": "F.2.2 VARIANCE LOSS APPROXIMATES THE SQUARED LOSS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "To see how the variance loss can have dynamics resembling those of the squared loss, we will add", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 103, + 350, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 103, + 350, + 456, + 364 + ], + "score": 1.0, + "content": "a global (trainable) bias to our model. This means our model is now parameterized by", + "type": "text" + }, + { + "bbox": [ + 456, + 351, + 504, + 362 + ], + "score": 0.92, + "content": "W \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 362, + 185, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 156, + 373 + ], + "score": 1.0, + "content": "and a scalar", + "type": "text" + }, + { + "bbox": [ + 156, + 362, + 181, + 372 + ], + "score": 0.9, + "content": "b \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 362, + 185, + 373 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 388, + 366, + 403 + ], + "lines": [ + { + "bbox": [ + 245, + 388, + 366, + 403 + ], + "spans": [ + { + "bbox": [ + 245, + 388, + 366, + 403 + ], + "score": 0.9, + "content": "f _ { W , b } ( x ) = \\langle W ^ { T } W , x x ^ { T } \\rangle + b", + "type": "interline_equation", + "image_path": "d5cad89cfe3e2d31a2c1c6cb4521849140245aafe938867c89b4a7515e0ad247.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 245, + 388, + 366, + 403 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 411, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "We may now analyze gradient flow of the squared loss. Following the same methods as before, this", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 422, + 297, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 297, + 435 + ], + "score": 1.0, + "content": "leads us to the following differential equations:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 453, + 407, + 514 + ], + "lines": [ + { + "bbox": [ + 203, + 453, + 407, + 514 + ], + "spans": [ + { + "bbox": [ + 203, + 453, + 407, + 514 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + \\frac { 1 } { 2 } ( b ^ { * } - b ) \\Big ) } \\\\ { \\displaystyle \\dot { b } = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } - b } \\end{array}", + "type": "interline_equation", + "image_path": "19e998aeb0cd8260a3e51c339ee0ab9bd808b022a4849845dbd64c732e80dfb1.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 453, + 407, + 468.25 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 203, + 468.25, + 407, + 483.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 203, + 483.5, + 407, + 498.75 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 203, + 498.75, + 407, + 514.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 524, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 104, + 523, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 165, + 540 + ], + "score": 1.0, + "content": "Notice how, if", + "type": "text" + }, + { + "bbox": [ + 165, + 526, + 171, + 535 + ], + "score": 0.79, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 523, + 304, + 540 + ], + "score": 1.0, + "content": "is at it’s optimum at a given time", + "type": "text" + }, + { + "bbox": [ + 304, + 525, + 402, + 539 + ], + "score": 0.9, + "content": "\\begin{array} { r } { ( b = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 523, + 471, + 540 + ], + "score": 1.0, + "content": ", the dynamics of", + "type": "text" + }, + { + "bbox": [ + 471, + 527, + 482, + 536 + ], + "score": 0.85, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 523, + 506, + 540 + ], + "score": 1.0, + "content": "align", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 313, + 550 + ], + "score": 1.0, + "content": "with those of the variance loss. Alternatively, when", + "type": "text" + }, + { + "bbox": [ + 313, + 538, + 340, + 547 + ], + "score": 0.87, + "content": "b = b ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 537, + 505, + 550 + ], + "score": 1.0, + "content": ", we recover the dynamics of the squared", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 548, + 241, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 241, + 561 + ], + "score": 1.0, + "content": "loss without the global bias term.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "To convince ourselves that the dynamics of this model resemble those of the variance loss, we would", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "need to explain why the global bias is at it’s optimum “most of the time”, such that the singular values", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 587, + 374, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 374, + 600 + ], + "score": 1.0, + "content": "don’t change much during the times when it is not at it’s optimum.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 264, + 617 + ], + "score": 1.0, + "content": "Observing the differential equations for", + "type": "text" + }, + { + "bbox": [ + 265, + 605, + 270, + 614 + ], + "score": 0.74, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 603, + 301, + 617 + ], + "score": 1.0, + "content": "and for", + "type": "text" + }, + { + "bbox": [ + 301, + 606, + 311, + 615 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 603, + 473, + 617 + ], + "score": 1.0, + "content": ", we see they are similar (if we ignore the", + "type": "text" + }, + { + "bbox": [ + 474, + 605, + 504, + 616 + ], + "score": 0.91, + "content": "\\sigma _ { i } ^ { * } - \\sigma _ { i }", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "value which doesn’t change the order of magnitude of the entire expression when there haven’t been", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 407, + 639 + ], + "score": 1.0, + "content": "many learning phases yet). The only difference being a multiplication by", + "type": "text" + }, + { + "bbox": [ + 407, + 628, + 417, + 637 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 627, + 505, + 639 + ], + "score": 1.0, + "content": ". This means that we", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 637, + 196, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 196, + 649 + ], + "score": 1.0, + "content": "may informally write:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "interline_equation", + "bbox": [ + 277, + 661, + 335, + 690 + ], + "lines": [ + { + "bbox": [ + 277, + 661, + 335, + 690 + ], + "spans": [ + { + "bbox": [ + 277, + 661, + 335, + 690 + ], + "score": 0.93, + "content": "\\frac { \\dot { \\sigma } _ { i } ( t ) } { \\dot { b } ( t ) } \\approx \\sigma _ { i } ( t )", + "type": "interline_equation", + "image_path": "a4a92114e23afb1dfe38d059b447a751c3212fc653e0e70d2347ecca2789b023.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 277, + 661, + 335, + 675.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 277, + 675.5, + 335, + 690.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 325, + 712 + ], + "score": 1.0, + "content": "Since at initialization and until the learning phase of", + "type": "text" + }, + { + "bbox": [ + 325, + 701, + 335, + 710 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 698, + 425, + 712 + ], + "score": 1.0, + "content": "takes place, we have", + "type": "text" + }, + { + "bbox": [ + 425, + 699, + 470, + 711 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t ) \\ll 1", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 698, + 505, + 712 + ], + "score": 1.0, + "content": ", we see", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "that the global bias optimizes much faster than the singular values for which the learning phase", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "hasn’t begun yet. This means these singular values will remain small during the times in which", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 450, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 452, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 144, + 97 + ], + "score": 1.0, + "content": "Defining", + "type": "text" + }, + { + "bbox": [ + 144, + 82, + 238, + 94 + ], + "score": 0.93, + "content": "\\Delta = W _ { * } ^ { T } W _ { * } - W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 79, + 401, + 97 + ], + "score": 1.0, + "content": "for brevity, the differential equations for", + "type": "text" + }, + { + "bbox": [ + 402, + 83, + 413, + 93 + ], + "score": 0.69, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 79, + 452, + 97 + ], + "score": 1.0, + "content": "become:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 104, + 79, + 452, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 106, + 361, + 131 + ], + "lines": [ + { + "bbox": [ + 249, + 106, + 361, + 131 + ], + "spans": [ + { + "bbox": [ + 249, + 106, + 361, + 131 + ], + "score": 0.94, + "content": "\\dot { W } = \\frac { 1 } { 2 } W \\Delta + \\frac { 1 } { 4 } T r ( \\Delta ) W", + "type": "interline_equation", + "image_path": "890dcfaeec2b118f56d3a15c72b6c45a986472d6e54582ca2f59ae9d6c614c05.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 249, + 106, + 361, + 131 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 139, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 139, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 219, + 153 + ], + "score": 1.0, + "content": "Calculating the dynamics of", + "type": "text" + }, + { + "bbox": [ + 219, + 139, + 248, + 150 + ], + "score": 0.91, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 139, + 471, + 153 + ], + "score": 1.0, + "content": "after noting that it is simultaneously diagonalizable with", + "type": "text" + }, + { + "bbox": [ + 472, + 139, + 504, + 152 + ], + "score": 0.91, + "content": "W _ { * } ^ { T } W _ { * }", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 1.0, + "content": "(as in the derivation for the variance loss) leads to the following differential equations for the singular", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 162, + 178, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 145, + 173 + ], + "score": 1.0, + "content": "values of", + "type": "text" + }, + { + "bbox": [ + 145, + 162, + 174, + 172 + ], + "score": 0.91, + "content": "W ^ { T } W", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 162, + 178, + 173 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 139, + 505, + 173 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 186, + 382, + 217 + ], + "lines": [ + { + "bbox": [ + 228, + 186, + 382, + 217 + ], + "spans": [ + { + "bbox": [ + 228, + 186, + 382, + 217 + ], + "score": 0.94, + "content": "\\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) \\Big )", + "type": "interline_equation", + "image_path": "3621aa3a2f9b0623dfe6b0accc893aadfeee631cbc164503d5c89a3b5e4f1d24.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 186, + 382, + 201.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 228, + 201.5, + 382, + 217.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 237 + ], + "score": 1.0, + "content": "We see that the equations are now coupled and so we cannot solve them analytically. Another issue", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "is that for our initialization, all singular values have very similar dynamics at initialization due to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "the coupling. For example, values corresponding to small optimal singular values grow much faster", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 258, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 504, + 271 + ], + "score": 1.0, + "content": "than in the variance loss dynamics, due to the effect the large optimal singular values have on them.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 226, + 506, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 504, + 288 + ], + "score": 1.0, + "content": "We see from these equations that we shouldn’t expect quadratic networks optimized over the squared", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "loss to exhibit incremental learning behavior. We next show how adding a global bias to our model", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 297, + 145, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 145, + 311 + ], + "score": 1.0, + "content": "can help.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 275, + 505, + 311 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 320, + 370, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 370, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 370, + 335 + ], + "score": 1.0, + "content": "F.2.2 VARIANCE LOSS APPROXIMATES THE SQUARED LOSS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "To see how the variance loss can have dynamics resembling those of the squared loss, we will add", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 103, + 350, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 103, + 350, + 456, + 364 + ], + "score": 1.0, + "content": "a global (trainable) bias to our model. This means our model is now parameterized by", + "type": "text" + }, + { + "bbox": [ + 456, + 351, + 504, + 362 + ], + "score": 0.92, + "content": "W \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 362, + 185, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 156, + 373 + ], + "score": 1.0, + "content": "and a scalar", + "type": "text" + }, + { + "bbox": [ + 156, + 362, + 181, + 372 + ], + "score": 0.9, + "content": "b \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 362, + 185, + 373 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 103, + 340, + 505, + 373 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 388, + 366, + 403 + ], + "lines": [ + { + "bbox": [ + 245, + 388, + 366, + 403 + ], + "spans": [ + { + "bbox": [ + 245, + 388, + 366, + 403 + ], + "score": 0.9, + "content": "f _ { W , b } ( x ) = \\langle W ^ { T } W , x x ^ { T } \\rangle + b", + "type": "interline_equation", + "image_path": "d5cad89cfe3e2d31a2c1c6cb4521849140245aafe938867c89b4a7515e0ad247.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 245, + 388, + 366, + 403 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 411, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "We may now analyze gradient flow of the squared loss. Following the same methods as before, this", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 422, + 297, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 297, + 435 + ], + "score": 1.0, + "content": "leads us to the following differential equations:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 411, + 505, + 435 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 453, + 407, + 514 + ], + "lines": [ + { + "bbox": [ + 203, + 453, + 407, + 514 + ], + "spans": [ + { + "bbox": [ + 203, + 453, + 407, + 514 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + \\frac { 1 } { 2 } ( b ^ { * } - b ) \\Big ) } \\\\ { \\displaystyle \\dot { b } = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } - b } \\end{array}", + "type": "interline_equation", + "image_path": "19e998aeb0cd8260a3e51c339ee0ab9bd808b022a4849845dbd64c732e80dfb1.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 453, + 407, + 468.25 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 203, + 468.25, + 407, + 483.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 203, + 483.5, + 407, + 498.75 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 203, + 498.75, + 407, + 514.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 524, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 104, + 523, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 165, + 540 + ], + "score": 1.0, + "content": "Notice how, if", + "type": "text" + }, + { + "bbox": [ + 165, + 526, + 171, + 535 + ], + "score": 0.79, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 523, + 304, + 540 + ], + "score": 1.0, + "content": "is at it’s optimum at a given time", + "type": "text" + }, + { + "bbox": [ + 304, + 525, + 402, + 539 + ], + "score": 0.9, + "content": "\\begin{array} { r } { ( b = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 523, + 471, + 540 + ], + "score": 1.0, + "content": ", the dynamics of", + "type": "text" + }, + { + "bbox": [ + 471, + 527, + 482, + 536 + ], + "score": 0.85, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 523, + 506, + 540 + ], + "score": 1.0, + "content": "align", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 313, + 550 + ], + "score": 1.0, + "content": "with those of the variance loss. Alternatively, when", + "type": "text" + }, + { + "bbox": [ + 313, + 538, + 340, + 547 + ], + "score": 0.87, + "content": "b = b ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 537, + 505, + 550 + ], + "score": 1.0, + "content": ", we recover the dynamics of the squared", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 548, + 241, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 241, + 561 + ], + "score": 1.0, + "content": "loss without the global bias term.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 523, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "To convince ourselves that the dynamics of this model resemble those of the variance loss, we would", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "need to explain why the global bias is at it’s optimum “most of the time”, such that the singular values", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 587, + 374, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 374, + 600 + ], + "score": 1.0, + "content": "don’t change much during the times when it is not at it’s optimum.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 565, + 505, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 264, + 617 + ], + "score": 1.0, + "content": "Observing the differential equations for", + "type": "text" + }, + { + "bbox": [ + 265, + 605, + 270, + 614 + ], + "score": 0.74, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 603, + 301, + 617 + ], + "score": 1.0, + "content": "and for", + "type": "text" + }, + { + "bbox": [ + 301, + 606, + 311, + 615 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 603, + 473, + 617 + ], + "score": 1.0, + "content": ", we see they are similar (if we ignore the", + "type": "text" + }, + { + "bbox": [ + 474, + 605, + 504, + 616 + ], + "score": 0.91, + "content": "\\sigma _ { i } ^ { * } - \\sigma _ { i }", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "value which doesn’t change the order of magnitude of the entire expression when there haven’t been", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 407, + 639 + ], + "score": 1.0, + "content": "many learning phases yet). The only difference being a multiplication by", + "type": "text" + }, + { + "bbox": [ + 407, + 628, + 417, + 637 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 627, + 505, + 639 + ], + "score": 1.0, + "content": ". This means that we", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 637, + 196, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 196, + 649 + ], + "score": 1.0, + "content": "may informally write:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 603, + 505, + 649 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 277, + 661, + 335, + 690 + ], + "lines": [ + { + "bbox": [ + 277, + 661, + 335, + 690 + ], + "spans": [ + { + "bbox": [ + 277, + 661, + 335, + 690 + ], + "score": 0.93, + "content": "\\frac { \\dot { \\sigma } _ { i } ( t ) } { \\dot { b } ( t ) } \\approx \\sigma _ { i } ( t )", + "type": "interline_equation", + "image_path": "a4a92114e23afb1dfe38d059b447a751c3212fc653e0e70d2347ecca2789b023.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 277, + 661, + 335, + 675.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 277, + 675.5, + 335, + 690.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 325, + 712 + ], + "score": 1.0, + "content": "Since at initialization and until the learning phase of", + "type": "text" + }, + { + "bbox": [ + 325, + 701, + 335, + 710 + ], + "score": 0.84, + "content": "\\sigma _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 698, + 425, + 712 + ], + "score": 1.0, + "content": "takes place, we have", + "type": "text" + }, + { + "bbox": [ + 425, + 699, + 470, + 711 + ], + "score": 0.92, + "content": "\\sigma _ { i } ( t ) \\ll 1", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 698, + 505, + 712 + ], + "score": 1.0, + "content": ", we see", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "that the global bias optimizes much faster than the singular values for which the learning phase", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "hasn’t begun yet. This means these singular values will remain small during the times in which", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "score": 1.0, + "content": "the bias isn’t optimal, and so incremental learning can still take place (assuming a small enough", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 397, + 165, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 165, + 411 + ], + "score": 1.0, + "content": "initialization).", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 79, + 474, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 79, + 474, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 79, + 474, + 250 + ], + "spans": [ + { + "bbox": [ + 136, + 79, + 474, + 250 + ], + "score": 0.977, + "type": "image", + "image_path": "0ec431bf9815c02f0689d4f67bcd2532ec6e6174e98a709848c535d46e23f897.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 79, + 474, + 136.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 136.0, + 474, + 193.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 193.0, + 474, + 250.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 265, + 505, + 364 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "score": 1.0, + "content": "Figure 5: Quadratic model’s evolution of top-5 singular values for a rank-4 labeling function. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 276, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 104, + 276, + 506, + 289 + ], + "score": 1.0, + "content": "rows correspond to whether or not a global bias is introduced to the model. The first two columns", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 287, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 504, + 299 + ], + "score": 1.0, + "content": "are for a large dataset (one optimal solution) and the last two columns are for a small dataset (over-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 298, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 504, + 310 + ], + "score": 1.0, + "content": "determined problem). When a bias is introduced, it is initialized to it’s optimal value at initialization.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "Note how without the bias, the singular values are learned together and there is over-shooting of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "optimal singular value caused by the coupling of the dynamics of the singular values. For the small", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 331, + 504, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 504, + 343 + ], + "score": 1.0, + "content": "datasets, we see that the model with no bias reaches a solution with a larger rank. Once a global bias", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "is introduced, the dynamics become more incremental as in the analysis of the variance loss. Note", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 353, + 471, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 471, + 365 + ], + "score": 1.0, + "content": "that in this case the solution obtained for the small dataset is the optimal low-rank solution.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "score": 1.0, + "content": "the bias isn’t optimal, and so incremental learning can still take place (assuming a small enough", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 397, + 165, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 165, + 411 + ], + "score": 1.0, + "content": "initialization).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "Under these considerations, we say that the dynamics of the squared loss for a quadratic network", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "with an added global bias resemble the idealized dynamics of the variance loss for a depth-2 linear", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "model which we analyze formally in the paper. In figure 5 we experimentally show how adding a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 449, + 484, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 484, + 461 + ], + "score": 1.0, + "content": "bias to a quadratic network does lead to incremental learning similar to the depth-2 toy model.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 478, + 365, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 366, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 366, + 493 + ], + "score": 1.0, + "content": "G INCREMENTAL LEARNING IN CLASSIFICATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 504, + 233, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 234, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 234, + 517 + ], + "score": 1.0, + "content": "G.1 DIAGONAL NETWORKS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "In section 4.3 we viewed our toy model as a special case of the deep diagonal networks described", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "in Gunasekar et al. (2018), expected to be biased towards sparse solutions. Figure 6 shows the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 241, + 560 + ], + "score": 1.0, + "content": "dynamics of the largest values of", + "type": "text" + }, + { + "bbox": [ + 241, + 550, + 249, + 558 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 548, + 506, + 560 + ], + "score": 1.0, + "content": "for different depths of the model. We see that the same type of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "incremental learning we saw in earlier models exists here as well - the features are learned one by", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "one in deeper models, resulting in a sparse solution. The leftmost panel shows how the initialization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 581, + 504, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 504, + 593 + ], + "score": 1.0, + "content": "scale plays a role here as well, with the solution being more sparse when the initialization is small.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "We should note that these results do not defy the results of Gunasekar et al. (2018) (from which we", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 603, + 494, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 459, + 615 + ], + "score": 1.0, + "content": "would expect the initialization not to matter), since their results deal with the solution at", + "type": "text" + }, + { + "bbox": [ + 460, + 604, + 489, + 613 + ], + "score": 0.89, + "content": "t \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 603, + 494, + 615 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 107, + 630, + 339, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 629, + 340, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 340, + 643 + ], + "score": 1.0, + "content": "G.2 CONVOLUTIONAL NETWORKS - PRELIMINARIES", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "The linear circular-convolutional network of Gunasekar et al. (2018) deals with one-dimensional", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 663, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 504, + 673 + ], + "score": 1.0, + "content": "convolutions with the same number of outputs as inputs, such that the mapping from one hidden", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 673, + 342, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 254, + 686 + ], + "score": 1.0, + "content": "layer to the next is parameterized by", + "type": "text" + }, + { + "bbox": [ + 254, + 675, + 267, + 685 + ], + "score": 0.85, + "content": "w _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 673, + 342, + 686 + ], + "score": 1.0, + "content": "and defined to be:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 701, + 427, + 735 + ], + "lines": [ + { + "bbox": [ + 184, + 701, + 427, + 735 + ], + "spans": [ + { + "bbox": [ + 184, + 701, + 427, + 735 + ], + "score": 0.94, + "content": "h _ { n } [ i ] = \\sum _ { k = 0 } ^ { d - 1 } w _ { n } [ k ] h _ { n - 1 } [ ( i + k ) \\mod d ] = ( h _ { n - 1 } \\star w _ { n } ) [ i ]", + "type": "interline_equation", + "image_path": "b332b917ac73178d8b5090101a64bebb24e89ab47a251c24e0dea18ce83bf948.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 701, + 427, + 718.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 184, + 718.0, + 427, + 735.0 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 79, + 474, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 79, + 474, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 79, + 474, + 250 + ], + "spans": [ + { + "bbox": [ + 136, + 79, + 474, + 250 + ], + "score": 0.977, + "type": "image", + "image_path": "0ec431bf9815c02f0689d4f67bcd2532ec6e6174e98a709848c535d46e23f897.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 79, + 474, + 136.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 136.0, + 474, + 193.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 193.0, + 474, + 250.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 265, + 505, + 364 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "score": 1.0, + "content": "Figure 5: Quadratic model’s evolution of top-5 singular values for a rank-4 labeling function. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 276, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 104, + 276, + 506, + 289 + ], + "score": 1.0, + "content": "rows correspond to whether or not a global bias is introduced to the model. The first two columns", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 287, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 504, + 299 + ], + "score": 1.0, + "content": "are for a large dataset (one optimal solution) and the last two columns are for a small dataset (over-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 298, + 504, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 504, + 310 + ], + "score": 1.0, + "content": "determined problem). When a bias is introduced, it is initialized to it’s optimal value at initialization.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 322 + ], + "score": 1.0, + "content": "Note how without the bias, the singular values are learned together and there is over-shooting of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "optimal singular value caused by the coupling of the dynamics of the singular values. For the small", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 331, + 504, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 504, + 343 + ], + "score": 1.0, + "content": "datasets, we see that the model with no bias reaches a solution with a larger rank. Once a global bias", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "is introduced, the dynamics become more incremental as in the analysis of the variance loss. Note", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 353, + 471, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 471, + 365 + ], + "score": 1.0, + "content": "that in this case the solution obtained for the small dataset is the optimal low-rank solution.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 106, + 388, + 504, + 411 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 105, + 386, + 505, + 411 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "Under these considerations, we say that the dynamics of the squared loss for a quadratic network", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "with an added global bias resemble the idealized dynamics of the variance loss for a depth-2 linear", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "model which we analyze formally in the paper. In figure 5 we experimentally show how adding a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 449, + 484, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 484, + 461 + ], + "score": 1.0, + "content": "bias to a quadratic network does lead to incremental learning similar to the depth-2 toy model.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 415, + 506, + 461 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 478, + 365, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 366, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 366, + 493 + ], + "score": 1.0, + "content": "G INCREMENTAL LEARNING IN CLASSIFICATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 504, + 233, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 234, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 234, + 517 + ], + "score": 1.0, + "content": "G.1 DIAGONAL NETWORKS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "In section 4.3 we viewed our toy model as a special case of the deep diagonal networks described", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "in Gunasekar et al. (2018), expected to be biased towards sparse solutions. Figure 6 shows the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 241, + 560 + ], + "score": 1.0, + "content": "dynamics of the largest values of", + "type": "text" + }, + { + "bbox": [ + 241, + 550, + 249, + 558 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 548, + 506, + 560 + ], + "score": 1.0, + "content": "for different depths of the model. We see that the same type of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "incremental learning we saw in earlier models exists here as well - the features are learned one by", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "one in deeper models, resulting in a sparse solution. The leftmost panel shows how the initialization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 581, + 504, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 504, + 593 + ], + "score": 1.0, + "content": "scale plays a role here as well, with the solution being more sparse when the initialization is small.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "We should note that these results do not defy the results of Gunasekar et al. (2018) (from which we", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 603, + 494, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 459, + 615 + ], + "score": 1.0, + "content": "would expect the initialization not to matter), since their results deal with the solution at", + "type": "text" + }, + { + "bbox": [ + 460, + 604, + 489, + 613 + ], + "score": 0.89, + "content": "t \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 603, + 494, + 615 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 527, + 506, + 615 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 630, + 339, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 629, + 340, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 340, + 643 + ], + "score": 1.0, + "content": "G.2 CONVOLUTIONAL NETWORKS - PRELIMINARIES", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 651, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "The linear circular-convolutional network of Gunasekar et al. (2018) deals with one-dimensional", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 663, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 504, + 673 + ], + "score": 1.0, + "content": "convolutions with the same number of outputs as inputs, such that the mapping from one hidden", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 673, + 342, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 254, + 686 + ], + "score": 1.0, + "content": "layer to the next is parameterized by", + "type": "text" + }, + { + "bbox": [ + 254, + 675, + 267, + 685 + ], + "score": 0.85, + "content": "w _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 673, + 342, + 686 + ], + "score": 1.0, + "content": "and defined to be:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 651, + 505, + 686 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 701, + 427, + 735 + ], + "lines": [ + { + "bbox": [ + 184, + 701, + 427, + 735 + ], + "spans": [ + { + "bbox": [ + 184, + 701, + 427, + 735 + ], + "score": 0.94, + "content": "h _ { n } [ i ] = \\sum _ { k = 0 } ^ { d - 1 } w _ { n } [ k ] h _ { n - 1 } [ ( i + k ) \\mod d ] = ( h _ { n - 1 } \\star w _ { n } ) [ i ]", + "type": "interline_equation", + "image_path": "b332b917ac73178d8b5090101a64bebb24e89ab47a251c24e0dea18ce83bf948.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 701, + 427, + 718.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 184, + 718.0, + 427, + 735.0 + ], + "spans": [], + "index": 33 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 501, + 178 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 82, + 501, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 82, + 501, + 178 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 501, + 178 + ], + "score": 0.964, + "type": "image", + "image_path": "f24003e92c57d90fdbdbc73d05bd3dd33d93af75d8ddeff5e4ea5313c8d7c9a4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 82, + 501, + 114.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 114.0, + 501, + 146.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 146.0, + 501, + 178.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 194, + 506, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Figure 6: Incremental learning in binary classification. A model as in section 4.3 is trained over 200", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 205, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 273, + 218 + ], + "score": 1.0, + "content": "i.i.d. random Gaussian examples, where", + "type": "text" + }, + { + "bbox": [ + 274, + 205, + 311, + 216 + ], + "score": 0.9, + "content": "d = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 205, + 506, + 218 + ], + "score": 1.0, + "content": ". The data is labeled by a weight vector with 4", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 215, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 215, + 506, + 230 + ], + "score": 1.0, + "content": "nonzero values, making the problem realizable with a sparse solution while the max-margin solution", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 225, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 241 + ], + "score": 1.0, + "content": "isn’t sparse. The left panel describes the obtained solution’s correlation with the sparse labeling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "vector for different depths and initializations. The results are averaged over 100 experiments, with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 249, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 261 + ], + "score": 1.0, + "content": "shaded regions denoting empirical standard deviations. We see that depth-1 models reach results", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "similar to the max-margin SVM solution as predicted by Gunasekar et al. (2018), while deeper", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "models are highly correlated with the sparse solution, with this correlation increasing when the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 281, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 295 + ], + "score": 1.0, + "content": "initialization scale is small. The other panels show the evolution of the absolute values of the top-5", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 152, + 305 + ], + "score": 1.0, + "content": "weights of", + "type": "text" + }, + { + "bbox": [ + 152, + 295, + 160, + 303 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 293, + 506, + 305 + ], + "score": 1.0, + "content": "for the smallest initialization scale. Note that as we increase the depth, incremental", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 223, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 223, + 316 + ], + "score": 1.0, + "content": "learning is clearly presented.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 343, + 350 + ], + "score": 1.0, + "content": "The final layer is a fully connected layer parameterized by", + "type": "text" + }, + { + "bbox": [ + 343, + 336, + 383, + 348 + ], + "score": 0.92, + "content": "w _ { N } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 335, + 505, + 350 + ], + "score": 1.0, + "content": ", such that the final model can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 346, + 237, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 237, + 361 + ], + "score": 1.0, + "content": "be written in the following way:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 373, + 399, + 389 + ], + "lines": [ + { + "bbox": [ + 211, + 373, + 399, + 389 + ], + "spans": [ + { + "bbox": [ + 211, + 373, + 399, + 389 + ], + "score": 0.9, + "content": "f _ { \\sigma } ( x ) = ( ( ( ( x \\star w _ { 1 } ) \\star w _ { 2 } ) \\cdot \\cdot \\cdot ) \\star w _ { N - 1 } ) ^ { T } w _ { N }", + "type": "interline_equation", + "image_path": "bc8a2f81d3ff4ed13d4edfc3f0d635663a8474a933f6acac03db3a299b44ba82.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 211, + 373, + 399, + 389 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 506, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "Lemma 3 from Gunasekar et al. (2018) shows how we can relate the Fourier coefficients of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 407, + 442, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 442, + 420 + ], + "score": 1.0, + "content": "weight vectors to the Fourier coefficients of the linear model induced by the model:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 339, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 340, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 340, + 436 + ], + "score": 1.0, + "content": "Lemma. For the circular-convolutional model as in (11):", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 449, + 376, + 463 + ], + "lines": [ + { + "bbox": [ + 233, + 449, + 376, + 463 + ], + "spans": [ + { + "bbox": [ + 233, + 449, + 376, + 463 + ], + "score": 0.91, + "content": "\\hat { \\sigma } = d i a g ( \\hat { w } _ { 1 } ) \\cdot \\cdot \\cdot d i a g ( \\hat { w } _ { N - 1 } ) \\hat { w } _ { N } ,", + "type": "interline_equation", + "image_path": "db754dd72fc0c7a055bdeb7bd57f99f82e160c8768d31cc0fe73539600e6d0ca.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 233, + 449, + 376, + 463 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 470, + 471, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 469, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 146, + 485 + ], + "score": 1.0, + "content": "where for", + "type": "text" + }, + { + "bbox": [ + 147, + 470, + 249, + 483 + ], + "score": 0.91, + "content": "n = 1 , 2 , . . . , N , \\hat { w } _ { n } \\in \\mathbb { C } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 469, + 432, + 485 + ], + "score": 1.0, + "content": "are the Fourier coefficients of the parameters", + "type": "text" + }, + { + "bbox": [ + 432, + 470, + 469, + 483 + ], + "score": 0.91, + "content": "w _ { n } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 492, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "This lemma connects the convolutional network to the diagonal network, and thus we should expect", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "to see the same incremental learning of the values of the diagonal network exhibited by the Fourier", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 514, + 274, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 274, + 525 + ], + "score": 1.0, + "content": "coefficients of the convolutional network.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 539, + 384, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 385, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 385, + 552 + ], + "score": 1.0, + "content": "G.3 CONVOLUTIONAL NETWORKS - EMPIRICAL EXAMINATION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "In figure 7 we see the same plots as in figure 6 but for the Fourier coefficients of the convolutional", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "model. We see that even when the model is far from the toy parameterization (there is no weight", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "sharing and the initialization is with random Gaussian weights), incremental learning is still clearly", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "seen in the dynamics of the model. We see how the inherent reason for the sparsity towards sparse", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 604, + 504, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 504, + 615 + ], + "score": 1.0, + "content": "solution found in Gunasekar et al. (2018) is the result of the dynamics of the model - small ampli-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 614, + 313, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 313, + 627 + ], + "score": 1.0, + "content": "tudes are attenuated while large ones are amplified.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 501, + 178 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 82, + 501, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 82, + 501, + 178 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 501, + 178 + ], + "score": 0.964, + "type": "image", + "image_path": "f24003e92c57d90fdbdbc73d05bd3dd33d93af75d8ddeff5e4ea5313c8d7c9a4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 82, + 501, + 114.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 114.0, + 501, + 146.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 146.0, + 501, + 178.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 194, + 506, + 316 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Figure 6: Incremental learning in binary classification. A model as in section 4.3 is trained over 200", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 205, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 273, + 218 + ], + "score": 1.0, + "content": "i.i.d. random Gaussian examples, where", + "type": "text" + }, + { + "bbox": [ + 274, + 205, + 311, + 216 + ], + "score": 0.9, + "content": "d = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 205, + 506, + 218 + ], + "score": 1.0, + "content": ". The data is labeled by a weight vector with 4", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 215, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 104, + 215, + 506, + 230 + ], + "score": 1.0, + "content": "nonzero values, making the problem realizable with a sparse solution while the max-margin solution", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 225, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 241 + ], + "score": 1.0, + "content": "isn’t sparse. The left panel describes the obtained solution’s correlation with the sparse labeling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "vector for different depths and initializations. The results are averaged over 100 experiments, with", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 249, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 261 + ], + "score": 1.0, + "content": "shaded regions denoting empirical standard deviations. We see that depth-1 models reach results", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "similar to the max-margin SVM solution as predicted by Gunasekar et al. (2018), while deeper", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "models are highly correlated with the sparse solution, with this correlation increasing when the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 281, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 295 + ], + "score": 1.0, + "content": "initialization scale is small. The other panels show the evolution of the absolute values of the top-5", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 152, + 305 + ], + "score": 1.0, + "content": "weights of", + "type": "text" + }, + { + "bbox": [ + 152, + 295, + 160, + 303 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 293, + 506, + 305 + ], + "score": 1.0, + "content": "for the smallest initialization scale. Note that as we increase the depth, incremental", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 223, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 223, + 316 + ], + "score": 1.0, + "content": "learning is clearly presented.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 343, + 350 + ], + "score": 1.0, + "content": "The final layer is a fully connected layer parameterized by", + "type": "text" + }, + { + "bbox": [ + 343, + 336, + 383, + 348 + ], + "score": 0.92, + "content": "w _ { N } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 335, + 505, + 350 + ], + "score": 1.0, + "content": ", such that the final model can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 346, + 237, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 237, + 361 + ], + "score": 1.0, + "content": "be written in the following way:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 335, + 505, + 361 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 373, + 399, + 389 + ], + "lines": [ + { + "bbox": [ + 211, + 373, + 399, + 389 + ], + "spans": [ + { + "bbox": [ + 211, + 373, + 399, + 389 + ], + "score": 0.9, + "content": "f _ { \\sigma } ( x ) = ( ( ( ( x \\star w _ { 1 } ) \\star w _ { 2 } ) \\cdot \\cdot \\cdot ) \\star w _ { N - 1 } ) ^ { T } w _ { N }", + "type": "interline_equation", + "image_path": "bc8a2f81d3ff4ed13d4edfc3f0d635663a8474a933f6acac03db3a299b44ba82.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 211, + 373, + 399, + 389 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 506, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "Lemma 3 from Gunasekar et al. (2018) shows how we can relate the Fourier coefficients of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 407, + 442, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 442, + 420 + ], + "score": 1.0, + "content": "weight vectors to the Fourier coefficients of the linear model induced by the model:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 396, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 339, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 340, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 340, + 436 + ], + "score": 1.0, + "content": "Lemma. For the circular-convolutional model as in (11):", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 421, + 340, + 436 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 449, + 376, + 463 + ], + "lines": [ + { + "bbox": [ + 233, + 449, + 376, + 463 + ], + "spans": [ + { + "bbox": [ + 233, + 449, + 376, + 463 + ], + "score": 0.91, + "content": "\\hat { \\sigma } = d i a g ( \\hat { w } _ { 1 } ) \\cdot \\cdot \\cdot d i a g ( \\hat { w } _ { N - 1 } ) \\hat { w } _ { N } ,", + "type": "interline_equation", + "image_path": "db754dd72fc0c7a055bdeb7bd57f99f82e160c8768d31cc0fe73539600e6d0ca.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 233, + 449, + 376, + 463 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 470, + 471, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 469, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 146, + 485 + ], + "score": 1.0, + "content": "where for", + "type": "text" + }, + { + "bbox": [ + 147, + 470, + 249, + 483 + ], + "score": 0.91, + "content": "n = 1 , 2 , . . . , N , \\hat { w } _ { n } \\in \\mathbb { C } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 469, + 432, + 485 + ], + "score": 1.0, + "content": "are the Fourier coefficients of the parameters", + "type": "text" + }, + { + "bbox": [ + 432, + 470, + 469, + 483 + ], + "score": 0.91, + "content": "w _ { n } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 469, + 469, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 492, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "This lemma connects the convolutional network to the diagonal network, and thus we should expect", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "to see the same incremental learning of the values of the diagonal network exhibited by the Fourier", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 514, + 274, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 274, + 525 + ], + "score": 1.0, + "content": "coefficients of the convolutional network.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 491, + 505, + 525 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 539, + 384, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 385, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 385, + 552 + ], + "score": 1.0, + "content": "G.3 CONVOLUTIONAL NETWORKS - EMPIRICAL EXAMINATION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "In figure 7 we see the same plots as in figure 6 but for the Fourier coefficients of the convolutional", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "model. We see that even when the model is far from the toy parameterization (there is no weight", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "sharing and the initialization is with random Gaussian weights), incremental learning is still clearly", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "seen in the dynamics of the model. We see how the inherent reason for the sparsity towards sparse", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 604, + 504, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 504, + 615 + ], + "score": 1.0, + "content": "solution found in Gunasekar et al. (2018) is the result of the dynamics of the model - small ampli-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 614, + 313, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 313, + 627 + ], + "score": 1.0, + "content": "tudes are attenuated while large ones are amplified.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 560, + 505, + 627 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 298, + 466, + 393 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 141, + 298, + 466, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 298, + 466, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 466, + 393 + ], + "score": 0.966, + "type": "image", + "image_path": "ce17411c14f982ed90cef0de02508e3a99c1bff2ecdf153320a808745f475e11.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 141, + 298, + 466, + 329.6666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 141, + 329.6666666666667, + 466, + 361.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 141, + 361.33333333333337, + 466, + 393.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 406, + 505, + 517 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "Figure 7: Incremental learning in convolutional networks. A model as in appendix G is trained over", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 285, + 429 + ], + "score": 1.0, + "content": "200 i.i.d. random Gaussian examples, where", + "type": "text" + }, + { + "bbox": [ + 286, + 417, + 320, + 428 + ], + "score": 0.9, + "content": "d = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 417, + 505, + 429 + ], + "score": 1.0, + "content": ". The weights are initialized randomly and the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "data is labeled by a weight vector with 4 nonzero frequencies, making the problem realizable with a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "sparse solution in the frequency domain. The left panel describes the obtained solution’s correlation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 451, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 462 + ], + "score": 1.0, + "content": "in the frequency domain with the sparse labeling vector for different depths and initializations. The", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "results are averaged over 9 experiments, with shaded regions denoting empirical standard deviations.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "score": 1.0, + "content": "We see that depth-1 models reach results similar to the max-margin SVM solution, while deeper", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 484, + 504, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 504, + 495 + ], + "score": 1.0, + "content": "models are highly correlated with the optimal sparse solution. The other panels show the evolution", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 288, + 506 + ], + "score": 1.0, + "content": "of the amplitudes of the top-5 frequencies of", + "type": "text" + }, + { + "bbox": [ + 288, + 496, + 295, + 504 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "for the smallest initialization scale. Note that as we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 505, + 350, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 350, + 518 + ], + "score": 1.0, + "content": "increase the depth, incremental learning is clearly presented.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + } + ], + "index": 4.25 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 298, + 466, + 393 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 141, + 298, + 466, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 298, + 466, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 466, + 393 + ], + "score": 0.966, + "type": "image", + "image_path": "ce17411c14f982ed90cef0de02508e3a99c1bff2ecdf153320a808745f475e11.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 141, + 298, + 466, + 329.6666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 141, + 329.6666666666667, + 466, + 361.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 141, + 361.33333333333337, + 466, + 393.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 406, + 505, + 517 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "Figure 7: Incremental learning in convolutional networks. A model as in appendix G is trained over", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 285, + 429 + ], + "score": 1.0, + "content": "200 i.i.d. random Gaussian examples, where", + "type": "text" + }, + { + "bbox": [ + 286, + 417, + 320, + 428 + ], + "score": 0.9, + "content": "d = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 417, + 505, + 429 + ], + "score": 1.0, + "content": ". The weights are initialized randomly and the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "data is labeled by a weight vector with 4 nonzero frequencies, making the problem realizable with a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "sparse solution in the frequency domain. The left panel describes the obtained solution’s correlation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 451, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 462 + ], + "score": 1.0, + "content": "in the frequency domain with the sparse labeling vector for different depths and initializations. The", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "results are averaged over 9 experiments, with shaded regions denoting empirical standard deviations.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "score": 1.0, + "content": "We see that depth-1 models reach results similar to the max-margin SVM solution, while deeper", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 484, + 504, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 504, + 495 + ], + "score": 1.0, + "content": "models are highly correlated with the optimal sparse solution. The other panels show the evolution", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 288, + 506 + ], + "score": 1.0, + "content": "of the amplitudes of the top-5 frequencies of", + "type": "text" + }, + { + "bbox": [ + 288, + 496, + 295, + 504 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "for the smallest initialization scale. Note that as we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 505, + 350, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 350, + 518 + ], + "score": 1.0, + "content": "increase the depth, incremental learning is clearly presented.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + } + ], + "index": 4.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/H1lj0nNFwB/H1lj0nNFwB_model.json b/parse/train/H1lj0nNFwB/H1lj0nNFwB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c94ed300e07490ddbf78e64cda44dbfabe68af37 --- /dev/null +++ b/parse/train/H1lj0nNFwB/H1lj0nNFwB_model.json @@ -0,0 +1,36236 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 681, + 1302, + 681, + 1302, + 1107, + 398, + 1107 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1499, + 1403, + 1499, + 1403, + 1803, + 298, + 1803 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1239, + 1402, + 1239, + 1402, + 1484, + 298, + 1484 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1820, + 1402, + 1820, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 314, + 374, + 1141, + 374, + 1141, + 530, + 314, + 530 + ], + "score": 0.965 + }, + { + "category_id": 0, + "poly": [ + 299, + 217, + 1397, + 217, + 1397, + 324, + 299, + 324 + ], + "score": 0.963 + }, + { + "category_id": 0, + "poly": [ + 302, + 1168, + 573, + 1168, + 573, + 1203, + 302, + 1203 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 773, + 610, + 926, + 610, + 926, + 644, + 773, + 644 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 104, + 299, + 104 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.618 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 216.0, + 1405.0, + 216.0, + 1405.0, + 272.0, + 294.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 272.0, + 1079.0, + 272.0, + 1079.0, + 329.0, + 293.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1165.0, + 579.0, + 1165.0, + 579.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 607.0, + 933.0, + 607.0, + 933.0, + 648.0, + 768.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 680.0, + 1307.0, + 680.0, + 1307.0, + 715.0, + 395.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 708.0, + 1305.0, + 708.0, + 1305.0, + 749.0, + 392.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 743.0, + 1305.0, + 743.0, + 1305.0, + 775.0, + 395.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 772.0, + 1305.0, + 772.0, + 1305.0, + 806.0, + 394.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 803.0, + 1305.0, + 803.0, + 1305.0, + 836.0, + 393.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 832.0, + 1305.0, + 832.0, + 1305.0, + 866.0, + 393.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 865.0, + 1305.0, + 865.0, + 1305.0, + 897.0, + 395.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 894.0, + 1307.0, + 894.0, + 1307.0, + 929.0, + 393.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 925.0, + 1306.0, + 925.0, + 1306.0, + 960.0, + 393.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 955.0, + 1305.0, + 955.0, + 1305.0, + 990.0, + 393.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 985.0, + 1305.0, + 985.0, + 1305.0, + 1019.0, + 394.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1014.0, + 1307.0, + 1014.0, + 1307.0, + 1052.0, + 393.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1047.0, + 1305.0, + 1047.0, + 1305.0, + 1079.0, + 394.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1075.0, + 501.0, + 1075.0, + 501.0, + 1114.0, + 393.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1532.0, + 296.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 1405.0, + 1530.0, + 1405.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1592.0, + 1403.0, + 1592.0, + 1403.0, + 1623.0, + 296.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1621.0, + 1403.0, + 1621.0, + 1403.0, + 1652.0, + 296.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1654.0, + 1404.0, + 1654.0, + 1404.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1715.0, + 292.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1714.0, + 1404.0, + 1714.0, + 1404.0, + 1746.0, + 296.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1777.0, + 292.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1771.0, + 522.0, + 1771.0, + 522.0, + 1806.0, + 295.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1237.0, + 1404.0, + 1237.0, + 1404.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 1406.0, + 1270.0, + 1406.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1300.0, + 1404.0, + 1300.0, + 1404.0, + 1333.0, + 292.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1333.0, + 1403.0, + 1333.0, + 1403.0, + 1363.0, + 296.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 1404.0, + 1361.0, + 1404.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 1406.0, + 1391.0, + 1406.0, + 1425.0, + 293.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1423.0, + 1406.0, + 1423.0, + 1406.0, + 1456.0, + 293.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1454.0, + 1362.0, + 1454.0, + 1362.0, + 1487.0, + 296.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 1402.0, + 1820.0, + 1402.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1402.0, + 1851.0, + 1402.0, + 1884.0, + 293.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1883.0, + 1403.0, + 1883.0, + 1403.0, + 1914.0, + 297.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1914.0, + 1404.0, + 1914.0, + 1404.0, + 1944.0, + 296.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1975.0, + 296.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 373.0, + 906.0, + 373.0, + 906.0, + 411.0, + 311.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 405.0, + 640.0, + 405.0, + 640.0, + 439.0, + 312.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 433.0, + 580.0, + 433.0, + 580.0, + 472.0, + 310.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 466.0, + 508.0, + 466.0, + 508.0, + 500.0, + 310.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 498.0, + 1143.0, + 498.0, + 1143.0, + 532.0, + 313.0, + 532.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1195, + 1405, + 1195, + 1405, + 1441, + 298, + 1441 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1025, + 1402, + 1025, + 1402, + 1180, + 298, + 1180 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1666, + 1404, + 1666, + 1404, + 1760, + 299, + 1760 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 886, + 1403, + 886, + 1403, + 1010, + 298, + 1010 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1859, + 1403, + 1859, + 1403, + 1952, + 299, + 1952 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 314, + 232, + 1375, + 232, + 1375, + 525, + 314, + 525 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1455, + 1404, + 1455, + 1404, + 1548, + 300, + 1548 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 296, + 548, + 1405, + 548, + 1405, + 825, + 296, + 825 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 686, + 1997, + 1014, + 1997, + 1014, + 2043, + 686, + 2043 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 299, + 1594, + 947, + 1594, + 947, + 1630, + 299, + 1630 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 1801, + 557, + 1801, + 557, + 1832, + 299, + 1832 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 2004, + 1400, + 2004, + 1400, + 2034, + 1366, + 2034 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.733 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.27 + }, + { + "category_id": 14, + "poly": [ + 685, + 1997, + 1015, + 1997, + 1015, + 2041, + 685, + 2041 + ], + "score": 0.92, + "latex": "f _ { \\sigma } ( { \\boldsymbol { x } } ) = \\langle \\sigma , { \\boldsymbol { x } } \\rangle \\qquad \\sigma \\in \\mathbb { R } _ { \\geq 0 } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 567, + 1890, + 665, + 1890, + 665, + 1918, + 567, + 1918 + ], + "score": 0.91, + "latex": "\\mathcal { X } = \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1287, + 768, + 1308, + 768, + 1308, + 790, + 1287, + 790 + ], + "score": 0.75, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 530, + 736, + 551, + 736, + 551, + 759, + 530, + 759 + ], + "score": 0.74, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 490, + 585, + 510, + 585, + 510, + 607, + 490, + 607 + ], + "score": 0.74, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1381, + 737, + 1401, + 737, + 1401, + 759, + 1381, + 759 + ], + "score": 0.71, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 855, + 707, + 876, + 707, + 876, + 730, + 855, + 730 + ], + "score": 0.7, + "latex": "\\sigma" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 247.0, + 366.0, + 247.0, + 366.0, + 272.0, + 339.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 243.0, + 646.0, + 243.0, + 646.0, + 270.0, + 611.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 234.0, + 918.0, + 234.0, + 918.0, + 257.0, + 887.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 231.0, + 1199.0, + 231.0, + 1199.0, + 258.0, + 1159.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 272.0, + 367.0, + 272.0, + 367.0, + 365.0, + 310.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 283.0, + 656.0, + 283.0, + 656.0, + 371.0, + 576.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 274.0, + 915.0, + 274.0, + 915.0, + 363.0, + 859.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 261.0, + 1199.0, + 261.0, + 1199.0, + 362.0, + 1129.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 359.0, + 363.0, + 359.0, + 363.0, + 370.0, + 352.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 363.0, + 646.0, + 363.0, + 646.0, + 391.0, + 611.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 364.0, + 915.0, + 364.0, + 915.0, + 383.0, + 895.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 353.0, + 1198.0, + 353.0, + 1198.0, + 376.0, + 1170.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 382.0, + 364.0, + 382.0, + 364.0, + 399.0, + 350.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 381.0, + 1199.0, + 381.0, + 1199.0, + 409.0, + 1167.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 406.0, + 366.0, + 406.0, + 366.0, + 428.0, + 347.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 403.0, + 647.0, + 403.0, + 647.0, + 430.0, + 611.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 407.0, + 916.0, + 407.0, + 916.0, + 426.0, + 897.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 413.0, + 1196.0, + 413.0, + 1196.0, + 435.0, + 1176.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 424.0, + 384.0, + 424.0, + 384.0, + 445.0, + 366.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 423.0, + 472.0, + 423.0, + 472.0, + 447.0, + 420.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 423.0, + 542.0, + 423.0, + 542.0, + 447.0, + 491.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 424.0, + 662.0, + 424.0, + 662.0, + 446.0, + 643.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 424.0, + 745.0, + 424.0, + 745.0, + 446.0, + 703.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 423.0, + 822.0, + 423.0, + 822.0, + 447.0, + 770.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 425.0, + 933.0, + 425.0, + 933.0, + 446.0, + 913.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 423.0, + 1020.0, + 423.0, + 1020.0, + 447.0, + 973.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 423.0, + 1095.0, + 423.0, + 1095.0, + 447.0, + 1043.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 425.0, + 1213.0, + 425.0, + 1213.0, + 446.0, + 1194.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 423.0, + 1290.0, + 423.0, + 1290.0, + 447.0, + 1242.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 423.0, + 1356.0, + 423.0, + 1356.0, + 447.0, + 1304.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 436.0, + 506.0, + 436.0, + 506.0, + 472.0, + 404.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 436.0, + 782.0, + 436.0, + 782.0, + 472.0, + 679.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 437.0, + 1055.0, + 437.0, + 1055.0, + 472.0, + 952.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 437.0, + 1333.0, + 437.0, + 1333.0, + 472.0, + 1228.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 488.0, + 533.0, + 488.0, + 533.0, + 527.0, + 333.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 490.0, + 806.0, + 490.0, + 806.0, + 526.0, + 607.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 489.0, + 1077.0, + 489.0, + 1077.0, + 528.0, + 884.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 490.0, + 1374.0, + 490.0, + 1374.0, + 526.0, + 1135.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 550.0, + 1405.0, + 550.0, + 1405.0, + 583.0, + 295.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 578.0, + 489.0, + 578.0, + 489.0, + 615.0, + 294.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 578.0, + 1405.0, + 578.0, + 1405.0, + 615.0, + 511.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 608.0, + 1405.0, + 608.0, + 1405.0, + 646.0, + 293.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 639.0, + 1405.0, + 639.0, + 1405.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 707.0, + 294.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 854.0, + 701.0, + 854.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 701.0, + 1405.0, + 701.0, + 1405.0, + 738.0, + 877.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 529.0, + 731.0, + 529.0, + 768.0, + 295.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 731.0, + 1380.0, + 731.0, + 1380.0, + 768.0, + 552.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 768.0, + 1402.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 761.0, + 1286.0, + 761.0, + 1286.0, + 798.0, + 294.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 761.0, + 1405.0, + 761.0, + 1405.0, + 798.0, + 1309.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 793.0, + 1151.0, + 793.0, + 1151.0, + 829.0, + 296.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1592.0, + 951.0, + 1592.0, + 951.0, + 1636.0, + 291.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1798.0, + 559.0, + 1798.0, + 559.0, + 1836.0, + 294.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1194.0, + 1406.0, + 1194.0, + 1406.0, + 1231.0, + 295.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1227.0, + 1402.0, + 1227.0, + 1402.0, + 1260.0, + 295.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1256.0, + 1403.0, + 1256.0, + 1403.0, + 1289.0, + 293.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1313.0, + 1405.0, + 1313.0, + 1405.0, + 1354.0, + 293.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1348.0, + 1403.0, + 1348.0, + 1403.0, + 1382.0, + 295.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1379.0, + 1405.0, + 1379.0, + 1405.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1409.0, + 1351.0, + 1409.0, + 1351.0, + 1443.0, + 295.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1027.0, + 1404.0, + 1027.0, + 1404.0, + 1060.0, + 296.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1057.0, + 1404.0, + 1057.0, + 1404.0, + 1091.0, + 296.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1088.0, + 1404.0, + 1088.0, + 1404.0, + 1121.0, + 296.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1118.0, + 1403.0, + 1118.0, + 1403.0, + 1152.0, + 296.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1145.0, + 943.0, + 1145.0, + 943.0, + 1186.0, + 292.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1666.0, + 1405.0, + 1666.0, + 1405.0, + 1703.0, + 295.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1696.0, + 1408.0, + 1696.0, + 1408.0, + 1735.0, + 293.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1729.0, + 1018.0, + 1729.0, + 1018.0, + 1763.0, + 297.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 887.0, + 1404.0, + 887.0, + 1404.0, + 920.0, + 296.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 918.0, + 1407.0, + 918.0, + 1407.0, + 954.0, + 293.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 948.0, + 1403.0, + 948.0, + 1403.0, + 984.0, + 293.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 979.0, + 783.0, + 979.0, + 783.0, + 1015.0, + 295.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1857.0, + 1403.0, + 1857.0, + 1403.0, + 1895.0, + 295.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1889.0, + 566.0, + 1889.0, + 566.0, + 1925.0, + 294.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1889.0, + 1405.0, + 1889.0, + 1405.0, + 1925.0, + 666.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1920.0, + 512.0, + 1920.0, + 512.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 1405.0, + 1452.0, + 1405.0, + 1491.0, + 294.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 1404.0, + 1486.0, + 1404.0, + 1520.0, + 295.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1516.0, + 1262.0, + 1516.0, + 1262.0, + 1550.0, + 295.0, + 1550.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 911, + 1405, + 911, + 1405, + 1098, + 296, + 1098 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 490, + 1405, + 490, + 1405, + 594, + 297, + 594 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 383, + 1405, + 383, + 1405, + 478, + 297, + 478 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 506, + 1244, + 1193, + 1244, + 1193, + 1437, + 506, + 1437 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 639, + 1558, + 1059, + 1558, + 1059, + 1624, + 639, + 1624 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 295, + 1103, + 1403, + 1103, + 1403, + 1199, + 295, + 1199 + ], + "score": 0.955 + }, + { + "category_id": 2, + "poly": [ + 298, + 1948, + 1404, + 1948, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 495, + 629, + 1204, + 629, + 1204, + 695, + 495, + 695 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 1640, + 1402, + 1640, + 1402, + 1703, + 298, + 1703 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 1825, + 1401, + 1825, + 1401, + 1890, + 297, + 1890 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 544, + 1741, + 1155, + 1741, + 1155, + 1808, + 544, + 1808 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 293, + 710, + 1345, + 710, + 1345, + 744, + 293, + 744 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 295, + 1489, + 1057, + 1489, + 1057, + 1523, + 295, + 1523 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 736, + 784, + 963, + 784, + 963, + 824, + 736, + 824 + ], + "score": 0.928 + }, + { + "category_id": 8, + "poly": [ + 765, + 301, + 930, + 301, + 930, + 343, + 765, + 343 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 293, + 226, + 1398, + 226, + 1398, + 267, + 293, + 267 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 298, + 855, + 884, + 855, + 884, + 888, + 298, + 888 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.917 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1760, + 1400, + 1760, + 1400, + 1789, + 1366, + 1789 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1366, + 789, + 1400, + 789, + 1400, + 819, + 1366, + 819 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 648, + 1400, + 648, + 1400, + 678, + 1366, + 678 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 1375, + 1904, + 1402, + 1904, + 1402, + 1932, + 1375, + 1932 + ], + "score": 0.794 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.675 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.494 + }, + { + "category_id": 13, + "poly": [ + 1045, + 228, + 1152, + 228, + 1152, + 269, + 1045, + 269 + ], + "score": 0.96, + "latex": "w \\in \\mathbb { R } _ { \\geq 0 } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 640, + 1557, + 1060, + 1557, + 1060, + 1624, + 640, + 1624 + ], + "score": 0.94, + "latex": "\\dot { w } _ { i } = - \\nabla _ { w _ { i } } \\ell = \\frac { 1 } { N } w _ { i } ^ { N - 1 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } )" + }, + { + "category_id": 13, + "poly": [ + 740, + 491, + 866, + 491, + 866, + 532, + 740, + 532 + ], + "score": 0.94, + "latex": "\\sigma ^ { * } \\in \\mathbb { R } _ { \\geq 0 } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 506, + 1241, + 1190, + 1241, + 1190, + 1438, + 506, + 1438 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\begin{array} { r l } { N = 1 : } & { \\sigma _ { i } ( t ) = \\sigma _ { i } ^ { * } + ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) e ^ { - t } } \\\\ { N = 2 : } & { \\sigma _ { i } ( t ) = \\frac { \\sigma _ { 0 } \\sigma _ { i } ^ { * } e ^ { \\sigma _ { i } ^ { * } t } } { \\sigma _ { 0 } \\left( e ^ { \\sigma _ { i } ^ { * } t } - 1 \\right) + \\sigma _ { i } ^ { * } } } \\end{array} } \\\\ & { \\begin{array} { r l } { N \\to \\infty : } & { t = \\frac { 1 } { ( \\sigma _ { i } ^ { * } ) ^ { 2 } } \\log \\big ( \\frac { \\sigma _ { i } ( t ) ( \\sigma _ { 0 } - \\sigma _ { i } ^ { * } ) } { \\sigma _ { 0 } \\left( \\sigma _ { i } ( t ) - \\sigma _ { i } ^ { * } \\right) } \\big ) - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { 0 } } \\big ) } \\end{array} } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 546, + 1738, + 1155, + 1738, + 1155, + 1809, + 546, + 1809 + ], + "score": 0.93, + "latex": "\\dot { \\sigma } _ { i } = \\frac { d \\sigma _ { i } } { d w _ { i } } \\dot { w } _ { i } = w ^ { 2 N - 2 } ( \\sigma _ { i } ^ { * } - w _ { i } ^ { N } ) = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 561, + 1855, + 764, + 1855, + 764, + 1890, + 561, + 1890 + ], + "score": 0.93, + "latex": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ^ { 2 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1089, + 1007, + 1172, + 1007, + 1172, + 1039, + 1089, + 1039 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ^ { * } > 0" + }, + { + "category_id": 14, + "poly": [ + 496, + 628, + 1202, + 628, + 1202, + 695, + 496, + 695 + ], + "score": 0.92, + "latex": "\\ell _ { N } ( w ) = \\frac { 1 } { 2 N ^ { 2 } } \\mathbb { E } _ { x } [ ( \\langle \\sigma ^ { * } , x \\rangle - \\langle w ^ { N } , x \\rangle ) ^ { 2 } ] = \\frac { 1 } { 2 N ^ { 2 } } | | w ^ { N } - \\sigma ^ { * } | | ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 649, + 1037, + 736, + 1037, + 736, + 1068, + 649, + 1068 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ^ { * } = 0" + }, + { + "category_id": 13, + "poly": [ + 421, + 944, + 472, + 944, + 472, + 978, + 421, + 978 + ], + "score": 0.92, + "latex": "\\sigma ( t )" + }, + { + "category_id": 14, + "poly": [ + 766, + 300, + 933, + 300, + 933, + 342, + 766, + 342 + ], + "score": 0.91, + "latex": "\\forall i : \\sigma _ { i } = w _ { i } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 975, + 1360, + 975, + 1360, + 1005, + 1217, + 1005 + ], + "score": 0.91, + "latex": "3 \\leq N < \\infty" + }, + { + "category_id": 13, + "poly": [ + 297, + 1166, + 381, + 1166, + 381, + 1199, + 297, + 1199 + ], + "score": 0.91, + "latex": "\\sigma _ { i } ^ { * } > 0" + }, + { + "category_id": 13, + "poly": [ + 664, + 975, + 769, + 975, + 769, + 1006, + 664, + 1006 + ], + "score": 0.91, + "latex": "N = 1 , 2" + }, + { + "category_id": 14, + "poly": [ + 735, + 785, + 965, + 785, + 965, + 824, + 735, + 824 + ], + "score": 0.9, + "latex": "\\forall i : \\ w _ { i } ( 0 ) = \\ \\sqrt [ N ] { \\sigma _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 541, + 1006, + 640, + 1006, + 640, + 1033, + 541, + 1033 + ], + "score": 0.9, + "latex": "N \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 820, + 976, + 918, + 976, + 918, + 1003, + 820, + 1003 + ], + "score": 0.9, + "latex": "N \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 298, + 1859, + 396, + 1859, + 396, + 1884, + 298, + 1884 + ], + "score": 0.85, + "latex": "N \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 1168, + 946, + 1196, + 946, + 1196, + 972, + 1168, + 972 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 807, + 1828, + 835, + 1828, + 835, + 1854, + 807, + 1854 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 381, + 387, + 409, + 387, + 409, + 413, + 381, + 413 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 389, + 531, + 418, + 531, + 418, + 558, + 389, + 558 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1167, + 1128, + 1167, + 1128, + 1193, + 1100, + 1193 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 602, + 1647, + 625, + 1647, + 625, + 1668, + 602, + 1668 + ], + "score": 0.78, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 954, + 236, + 975, + 236, + 975, + 259, + 954, + 259 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 475, + 1676, + 495, + 1676, + 495, + 1698, + 475, + 1698 + ], + "score": 0.73, + "latex": "\\sigma" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1983.0, + 328.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1977.0, + 1405.0, + 1977.0, + 1405.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 877.0, + 2004.0, + 877.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 853.0, + 886.0, + 853.0, + 886.0, + 892.0, + 293.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 1908.0, + 1402.0, + 1908.0, + 1402.0, + 1932.0, + 1378.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 912.0, + 1403.0, + 912.0, + 1403.0, + 948.0, + 295.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 944.0, + 420.0, + 944.0, + 420.0, + 979.0, + 293.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 944.0, + 1167.0, + 944.0, + 1167.0, + 979.0, + 473.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 944.0, + 1405.0, + 944.0, + 1405.0, + 979.0, + 1197.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 972.0, + 663.0, + 972.0, + 663.0, + 1010.0, + 293.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 972.0, + 819.0, + 972.0, + 819.0, + 1010.0, + 770.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 972.0, + 1216.0, + 972.0, + 1216.0, + 1010.0, + 919.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 972.0, + 1406.0, + 972.0, + 1406.0, + 1010.0, + 1361.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1003.0, + 540.0, + 1003.0, + 540.0, + 1040.0, + 293.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1003.0, + 1088.0, + 1003.0, + 1088.0, + 1040.0, + 641.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1003.0, + 1406.0, + 1003.0, + 1406.0, + 1040.0, + 1173.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 648.0, + 1037.0, + 648.0, + 1069.0, + 295.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1069.0, + 737.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1067.0, + 1010.0, + 1067.0, + 1010.0, + 1099.0, + 296.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 474.0, + 739.0, + 474.0, + 739.0, + 551.0, + 283.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 474.0, + 1404.0, + 474.0, + 1404.0, + 551.0, + 867.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 528.0, + 388.0, + 528.0, + 388.0, + 566.0, + 292.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 528.0, + 1406.0, + 528.0, + 1406.0, + 566.0, + 419.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 561.0, + 635.0, + 561.0, + 635.0, + 597.0, + 295.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 383.0, + 380.0, + 383.0, + 380.0, + 421.0, + 295.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 383.0, + 1405.0, + 383.0, + 1405.0, + 421.0, + 410.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 415.0, + 1406.0, + 415.0, + 1406.0, + 452.0, + 294.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 448.0, + 563.0, + 448.0, + 563.0, + 480.0, + 293.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1101.0, + 1405.0, + 1101.0, + 1405.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1133.0, + 1405.0, + 1133.0, + 1405.0, + 1172.0, + 292.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1166.0, + 296.0, + 1166.0, + 296.0, + 1200.0, + 293.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1166.0, + 1099.0, + 1166.0, + 1099.0, + 1200.0, + 382.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1166.0, + 1141.0, + 1166.0, + 1141.0, + 1200.0, + 1129.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 601.0, + 1637.0, + 601.0, + 1677.0, + 294.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1637.0, + 1406.0, + 1637.0, + 1406.0, + 1677.0, + 626.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1670.0, + 474.0, + 1670.0, + 474.0, + 1703.0, + 296.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1670.0, + 508.0, + 1670.0, + 508.0, + 1703.0, + 496.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 806.0, + 1820.0, + 806.0, + 1864.0, + 293.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1864.0, + 836.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1855.0, + 560.0, + 1855.0, + 560.0, + 1891.0, + 397.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1855.0, + 1032.0, + 1855.0, + 1032.0, + 1891.0, + 765.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 709.0, + 1346.0, + 709.0, + 1346.0, + 748.0, + 297.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1484.0, + 1060.0, + 1484.0, + 1060.0, + 1531.0, + 293.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 219.0, + 953.0, + 219.0, + 953.0, + 274.0, + 291.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 219.0, + 1044.0, + 219.0, + 1044.0, + 274.0, + 976.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 219.0, + 1404.0, + 219.0, + 1404.0, + 274.0, + 1153.0, + 274.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1361, + 1404, + 1361, + 1404, + 1544, + 298, + 1544 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1084, + 1404, + 1084, + 1404, + 1206, + 299, + 1206 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1223, + 1402, + 1223, + 1402, + 1345, + 298, + 1345 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1674, + 1403, + 1674, + 1403, + 1798, + 299, + 1798 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1814, + 1403, + 1814, + 1403, + 1937, + 297, + 1937 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 308, + 234, + 1386, + 234, + 1386, + 766, + 308, + 766 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 295, + 820, + 1405, + 820, + 1405, + 1015, + 295, + 1015 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 303, + 1599, + 721, + 1599, + 721, + 1634, + 303, + 1634 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.774 + }, + { + "category_id": 2, + "poly": [ + 301, + 1978, + 1403, + 1978, + 1403, + 2034, + 301, + 2034 + ], + "score": 0.667 + }, + { + "category_id": 2, + "poly": [ + 303, + 1977, + 1402, + 1977, + 1402, + 2034, + 303, + 2034 + ], + "score": 0.601 + }, + { + "category_id": 13, + "poly": [ + 340, + 860, + 553, + 860, + 553, + 895, + 340, + 895 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ^ { * } \\in \\{ 1 2 , 6 , 4 , 3 \\}" + }, + { + "category_id": 13, + "poly": [ + 1345, + 815, + 1400, + 815, + 1400, + 865, + 1345, + 865 + ], + "score": 0.92, + "latex": "\\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 898, + 1739, + 987, + 1739, + 987, + 1767, + 898, + 1767 + ], + "score": 0.91, + "latex": "\\sigma _ { 0 } 0" + }, + { + "category_id": 13, + "poly": [ + 540, + 1769, + 622, + 1769, + 622, + 1798, + 540, + 1798 + ], + "score": 0.91, + "latex": "\\sigma _ { 0 } > 0" + }, + { + "category_id": 13, + "poly": [ + 1350, + 1223, + 1400, + 1223, + 1400, + 1257, + 1350, + 1257 + ], + "score": 0.9, + "latex": "\\sigma ( t )" + }, + { + "category_id": 13, + "poly": [ + 803, + 1176, + 881, + 1176, + 881, + 1206, + 803, + 1206 + ], + "score": 0.9, + "latex": "N \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 1205, + 1146, + 1285, + 1146, + 1285, + 1173, + 1205, + 1173 + ], + "score": 0.9, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1423, + 1193, + 1423, + 1193, + 1451, + 1111, + 1451 + ], + "score": 0.9, + "latex": "N = 1" + }, + { + "category_id": 13, + "poly": [ + 652, + 1254, + 743, + 1254, + 743, + 1282, + 652, + 1282 + ], + "score": 0.89, + "latex": "N = 1" + }, + { + "category_id": 13, + "poly": [ + 674, + 1453, + 702, + 1453, + 702, + 1483, + 674, + 1483 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1821, + 1147, + 1821, + 1147, + 1842, + 1127, + 1842 + ], + "score": 0.74, + "latex": "\\sigma" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 235.0, + 592.0, + 235.0, + 592.0, + 270.0, + 436.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 232.0, + 851.0, + 232.0, + 851.0, + 272.0, + 692.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 232.0, + 1110.0, + 232.0, + 1110.0, + 272.0, + 950.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 232.0, + 1365.0, + 232.0, + 1365.0, + 272.0, + 1209.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 253.0, + 438.0, + 253.0, + 438.0, + 296.0, + 367.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 256.0, + 694.0, + 256.0, + 694.0, + 294.0, + 625.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 256.0, + 949.0, + 256.0, + 949.0, + 295.0, + 884.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 256.0, + 1210.0, + 256.0, + 1210.0, + 294.0, + 1142.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 300.0, + 429.0, + 300.0, + 429.0, + 340.0, + 367.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 300.0, + 686.0, + 300.0, + 686.0, + 340.0, + 625.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 300.0, + 943.0, + 300.0, + 943.0, + 340.0, + 884.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 301.0, + 1200.0, + 301.0, + 1200.0, + 337.0, + 1144.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 327.0, + 436.0, + 327.0, + 436.0, + 401.0, + 301.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 342.0, + 598.0, + 342.0, + 598.0, + 377.0, + 557.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 346.0, + 686.0, + 346.0, + 686.0, + 382.0, + 626.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 346.0, + 943.0, + 346.0, + 943.0, + 382.0, + 884.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 346.0, + 1200.0, + 346.0, + 1200.0, + 382.0, + 1144.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 367.0, + 587.0, + 367.0, + 587.0, + 401.0, + 559.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 388.0, + 429.0, + 388.0, + 429.0, + 427.0, + 367.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 391.0, + 586.0, + 391.0, + 586.0, + 426.0, + 557.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 390.0, + 684.0, + 390.0, + 684.0, + 426.0, + 626.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 390.0, + 943.0, + 390.0, + 943.0, + 426.0, + 884.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 390.0, + 1200.0, + 390.0, + 1200.0, + 426.0, + 1144.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 416.0, + 589.0, + 416.0, + 589.0, + 452.0, + 555.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 432.0, + 436.0, + 432.0, + 436.0, + 470.0, + 367.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 432.0, + 697.0, + 432.0, + 697.0, + 470.0, + 625.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 435.0, + 951.0, + 435.0, + 951.0, + 469.0, + 885.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 435.0, + 1213.0, + 435.0, + 1213.0, + 469.0, + 1143.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 459.0, + 445.0, + 459.0, + 445.0, + 486.0, + 419.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 455.0, + 594.0, + 455.0, + 594.0, + 489.0, + 464.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 456.0, + 705.0, + 456.0, + 705.0, + 488.0, + 676.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 454.0, + 852.0, + 454.0, + 852.0, + 489.0, + 724.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 457.0, + 962.0, + 457.0, + 962.0, + 489.0, + 934.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 452.0, + 1113.0, + 452.0, + 1113.0, + 491.0, + 979.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 459.0, + 1219.0, + 459.0, + 1219.0, + 486.0, + 1194.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 456.0, + 1369.0, + 456.0, + 1369.0, + 490.0, + 1242.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 489.0, + 605.0, + 489.0, + 605.0, + 528.0, + 420.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 489.0, + 864.0, + 489.0, + 864.0, + 528.0, + 679.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 489.0, + 1123.0, + 489.0, + 1123.0, + 528.0, + 937.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 490.0, + 1380.0, + 490.0, + 1380.0, + 526.0, + 1197.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 515.0, + 433.0, + 515.0, + 433.0, + 549.0, + 370.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 515.0, + 689.0, + 515.0, + 689.0, + 549.0, + 626.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 515.0, + 944.0, + 515.0, + 944.0, + 549.0, + 885.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 515.0, + 1203.0, + 515.0, + 1203.0, + 551.0, + 1143.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 554.0, + 429.0, + 554.0, + 429.0, + 598.0, + 368.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 558.0, + 685.0, + 558.0, + 685.0, + 594.0, + 626.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 557.0, + 943.0, + 557.0, + 943.0, + 596.0, + 884.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 557.0, + 1202.0, + 557.0, + 1202.0, + 596.0, + 1143.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 584.0, + 437.0, + 584.0, + 437.0, + 657.0, + 301.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 601.0, + 685.0, + 601.0, + 685.0, + 637.0, + 626.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 601.0, + 943.0, + 601.0, + 943.0, + 637.0, + 885.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 601.0, + 1200.0, + 601.0, + 1200.0, + 637.0, + 1144.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 643.0, + 427.0, + 643.0, + 427.0, + 684.0, + 367.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 643.0, + 685.0, + 643.0, + 685.0, + 683.0, + 625.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 644.0, + 940.0, + 644.0, + 940.0, + 681.0, + 885.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 643.0, + 1202.0, + 643.0, + 1202.0, + 683.0, + 1143.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 687.0, + 437.0, + 687.0, + 437.0, + 723.0, + 367.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 689.0, + 685.0, + 689.0, + 685.0, + 723.0, + 626.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 689.0, + 943.0, + 689.0, + 943.0, + 723.0, + 885.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 689.0, + 1202.0, + 689.0, + 1202.0, + 723.0, + 1144.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 705.0, + 1203.0, + 705.0, + 1203.0, + 715.0, + 1193.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 715.0, + 445.0, + 715.0, + 445.0, + 742.0, + 419.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 709.0, + 594.0, + 709.0, + 594.0, + 746.0, + 465.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 714.0, + 701.0, + 714.0, + 701.0, + 742.0, + 678.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 710.0, + 801.0, + 710.0, + 801.0, + 743.0, + 740.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 711.0, + 885.0, + 711.0, + 885.0, + 746.0, + 822.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 715.0, + 960.0, + 715.0, + 960.0, + 742.0, + 935.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 710.0, + 1122.0, + 710.0, + 1122.0, + 746.0, + 986.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 715.0, + 1219.0, + 715.0, + 1219.0, + 742.0, + 1194.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 710.0, + 1375.0, + 710.0, + 1375.0, + 746.0, + 1237.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 730.0, + 546.0, + 730.0, + 546.0, + 768.0, + 481.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 730.0, + 803.0, + 730.0, + 803.0, + 768.0, + 742.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 732.0, + 1062.0, + 732.0, + 1062.0, + 767.0, + 999.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 730.0, + 1320.0, + 730.0, + 1320.0, + 768.0, + 1257.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 815.0, + 1344.0, + 815.0, + 1344.0, + 866.0, + 288.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 815.0, + 1408.0, + 815.0, + 1408.0, + 866.0, + 1401.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 339.0, + 858.0, + 339.0, + 895.0, + 294.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 858.0, + 1405.0, + 858.0, + 1405.0, + 895.0, + 554.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 892.0, + 1405.0, + 892.0, + 1405.0, + 925.0, + 295.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 921.0, + 1406.0, + 921.0, + 1406.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 950.0, + 1406.0, + 950.0, + 1406.0, + 990.0, + 292.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 985.0, + 1103.0, + 985.0, + 1103.0, + 1018.0, + 294.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1595.0, + 726.0, + 1595.0, + 726.0, + 1640.0, + 292.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2012.0, + 329.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 604.0, + 2004.0, + 604.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2013.0, + 328.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 604.0, + 2004.0, + 604.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1360.0, + 1408.0, + 1360.0, + 1408.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1391.0, + 1408.0, + 1391.0, + 1408.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1110.0, + 1423.0, + 1110.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1455.0, + 1194.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1445.0, + 673.0, + 1445.0, + 673.0, + 1493.0, + 291.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1445.0, + 1406.0, + 1445.0, + 1406.0, + 1493.0, + 703.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1514.0, + 1003.0, + 1514.0, + 1003.0, + 1547.0, + 293.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1085.0, + 1406.0, + 1085.0, + 1406.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1116.0, + 1404.0, + 1116.0, + 1404.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1204.0, + 1146.0, + 1204.0, + 1179.0, + 294.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1179.0, + 1286.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 802.0, + 1176.0, + 802.0, + 1209.0, + 295.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1176.0, + 1306.0, + 1176.0, + 1306.0, + 1209.0, + 882.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1349.0, + 1222.0, + 1349.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1222.0, + 1404.0, + 1222.0, + 1404.0, + 1257.0, + 1401.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1253.0, + 651.0, + 1253.0, + 651.0, + 1287.0, + 292.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1253.0, + 1407.0, + 1253.0, + 1407.0, + 1287.0, + 744.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 1404.0, + 1283.0, + 1404.0, + 1318.0, + 293.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1315.0, + 792.0, + 1315.0, + 792.0, + 1348.0, + 294.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 1404.0, + 1674.0, + 1404.0, + 1710.0, + 294.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1703.0, + 1407.0, + 1703.0, + 1407.0, + 1742.0, + 292.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1735.0, + 897.0, + 1735.0, + 897.0, + 1772.0, + 293.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1735.0, + 1408.0, + 1735.0, + 1408.0, + 1772.0, + 988.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 539.0, + 1768.0, + 539.0, + 1801.0, + 294.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1768.0, + 635.0, + 1768.0, + 635.0, + 1801.0, + 623.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1813.0, + 1126.0, + 1813.0, + 1126.0, + 1849.0, + 293.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1813.0, + 1404.0, + 1813.0, + 1404.0, + 1849.0, + 1148.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1845.0, + 1404.0, + 1845.0, + 1404.0, + 1878.0, + 295.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1877.0, + 1404.0, + 1877.0, + 1404.0, + 1909.0, + 295.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1907.0, + 713.0, + 1907.0, + 713.0, + 1939.0, + 295.0, + 1939.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1632, + 1405, + 1632, + 1405, + 1834, + 296, + 1834 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 434, + 1405, + 434, + 1405, + 590, + 297, + 590 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 727, + 1407, + 727, + 1407, + 875, + 296, + 875 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 603, + 1405, + 603, + 1405, + 716, + 296, + 716 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 1405, + 227, + 1405, + 336, + 296, + 336 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1922, + 1404, + 1922, + 1404, + 2037, + 298, + 2037 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 579, + 1385, + 1119, + 1385, + 1119, + 1475, + 579, + 1475 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 295, + 1494, + 1403, + 1494, + 1403, + 1560, + 295, + 1560 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 292, + 1109, + 1403, + 1109, + 1403, + 1176, + 292, + 1176 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 668, + 1212, + 1029, + 1212, + 1029, + 1297, + 668, + 1297 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 689, + 376, + 1010, + 376, + 1010, + 419, + 689, + 419 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 293, + 1846, + 1402, + 1846, + 1402, + 1911, + 293, + 1911 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 292, + 1316, + 1371, + 1316, + 1371, + 1352, + 292, + 1352 + ], + "score": 0.933 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 8, + "poly": [ + 421, + 926, + 1173, + 926, + 1173, + 1083, + 421, + 1083 + ], + "score": 0.833 + }, + { + "category_id": 2, + "poly": [ + 1375, + 1573, + 1402, + 1573, + 1402, + 1601, + 1375, + 1601 + ], + "score": 0.796 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2112, + 841, + 2112 + ], + "score": 0.758 + }, + { + "category_id": 8, + "poly": [ + 427, + 926, + 1086, + 926, + 1086, + 1083, + 427, + 1083 + ], + "score": 0.519 + }, + { + "category_id": 14, + "poly": [ + 669, + 1209, + 1029, + 1209, + 1029, + 1296, + 669, + 1296 + ], + "score": 0.94, + "latex": "t _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }" + }, + { + "category_id": 13, + "poly": [ + 896, + 772, + 1021, + 772, + 1021, + 812, + 896, + 812 + ], + "score": 0.93, + "latex": "s \\in ( 0 , \\frac { 1 } { 4 } )" + }, + { + "category_id": 14, + "poly": [ + 576, + 1383, + 1120, + 1383, + 1120, + 1473, + 576, + 1473 + ], + "score": 0.93, + "latex": "\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }" + }, + { + "category_id": 14, + "poly": [ + 422, + 927, + 1280, + 927, + 1280, + 1085, + 422, + 1085 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 626, + 1143, + 759, + 1143, + 759, + 1176, + 626, + 1176 + ], + "score": 0.93, + "latex": "\\sigma ( t ) = \\alpha \\sigma ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 841, + 232, + 1000, + 232, + 1000, + 267, + 841, + 267 + ], + "score": 0.93, + "latex": "\\sigma _ { i } ^ { * } > \\sigma _ { j } ^ { * } > 0" + }, + { + "category_id": 13, + "poly": [ + 959, + 265, + 1083, + 265, + 1083, + 304, + 959, + 304 + ], + "score": 0.93, + "latex": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)" + }, + { + "category_id": 13, + "poly": [ + 1044, + 1317, + 1224, + 1317, + 1224, + 1352, + 1044, + 1352 + ], + "score": 0.93, + "latex": "t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 1264, + 1917, + 1402, + 1917, + 1402, + 1971, + 1264, + 1971 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1967, + 1233, + 1967, + 1233, + 2006, + 1117, + 2006 + ], + "score": 0.93, + "latex": "f \\in ( \\frac { 3 } { 4 } , 1 )" + }, + { + "category_id": 13, + "poly": [ + 952, + 1968, + 1065, + 1968, + 1065, + 2007, + 952, + 2007 + ], + "score": 0.92, + "latex": "s \\in ( 0 , \\frac { 1 } { 4 } )" + }, + { + "category_id": 13, + "poly": [ + 1256, + 2002, + 1357, + 2002, + 1357, + 2043, + 1256, + 2043 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 2002, + 423, + 2002, + 423, + 2039, + 298, + 2039 + ], + "score": 0.92, + "latex": "\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 780, + 266, + 902, + 266, + 902, + 304, + 780, + 304 + ], + "score": 0.92, + "latex": "s \\in ( 0 , \\frac { 1 } { 4 } )" + }, + { + "category_id": 13, + "poly": [ + 1079, + 774, + 1206, + 774, + 1206, + 811, + 1079, + 811 + ], + "score": 0.92, + "latex": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)" + }, + { + "category_id": 13, + "poly": [ + 356, + 497, + 497, + 497, + 497, + 535, + 356, + 535 + ], + "score": 0.92, + "latex": "\\sigma _ { j } ( t ) = s \\sigma _ { j } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 298, + 266, + 509, + 266, + 509, + 303, + 298, + 303 + ], + "score": 0.91, + "latex": "\\sigma _ { j } ( 0 ) = \\sigma _ { 0 } < \\sigma _ { j } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1306, + 230, + 1405, + 230, + 1405, + 265, + 1306, + 265 + ], + "score": 0.91, + "latex": "\\sigma _ { i } ( 0 ) =" + }, + { + "category_id": 14, + "poly": [ + 688, + 376, + 1011, + 376, + 1011, + 418, + 688, + 418 + ], + "score": 0.91, + "latex": "\\sigma _ { j } ( t ) \\leq s \\sigma _ { j } ^ { * } < f \\sigma _ { i } ^ { * } \\leq \\sigma _ { i } ( t )" + }, + { + "category_id": 13, + "poly": [ + 1161, + 722, + 1308, + 722, + 1308, + 776, + 1161, + 776 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\frac { { \\sigma } _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1040, + 812, + 1104, + 812, + 1104, + 845, + 1040, + 845 + ], + "score": 0.91, + "latex": "( s , f )" + }, + { + "category_id": 13, + "poly": [ + 595, + 1524, + 636, + 1524, + 636, + 1559, + 595, + 1559 + ], + "score": 0.91, + "latex": "\\sigma _ { 0 } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 376, + 304, + 440, + 304, + 440, + 336, + 376, + 336 + ], + "score": 0.9, + "latex": "( s , f )" + }, + { + "category_id": 13, + "poly": [ + 481, + 1801, + 561, + 1801, + 561, + 1831, + 481, + 1831 + ], + "score": 0.9, + "latex": "N \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 526, + 684, + 605, + 684, + 605, + 712, + 526, + 712 + ], + "score": 0.9, + "latex": "N = 3" + }, + { + "category_id": 13, + "poly": [ + 1352, + 809, + 1392, + 809, + 1392, + 845, + 1352, + 845 + ], + "score": 0.89, + "latex": "\\sigma _ { 0 } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 339, + 1740, + 426, + 1740, + 426, + 1768, + 339, + 1768 + ], + "score": 0.89, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 658, + 236, + 726, + 236, + 726, + 265, + 658, + 265 + ], + "score": 0.89, + "latex": "\\sigma _ { i } , \\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1071, + 1801, + 1149, + 1801, + 1149, + 1828, + 1071, + 1828 + ], + "score": 0.89, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 652, + 735, + 719, + 735, + 719, + 765, + 652, + 765 + ], + "score": 0.89, + "latex": "\\sigma _ { i } , \\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 387, + 635, + 419, + 635, + 419, + 687, + 387, + 687 + ], + "score": 0.87, + "latex": "\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 645, + 1931, + 713, + 1931, + 713, + 1961, + 645, + 1961 + ], + "score": 0.87, + "latex": "\\sigma _ { i } , \\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 397, + 686, + 476, + 686, + 476, + 711, + 397, + 711 + ], + "score": 0.86, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 1190, + 437, + 1277, + 437, + 1277, + 469, + 1190, + 469 + ], + "score": 0.86, + "latex": "( f \\approx 1 )" + }, + { + "category_id": 13, + "poly": [ + 1226, + 1501, + 1257, + 1501, + 1257, + 1526, + 1226, + 1526 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 595, + 468, + 679, + 468, + 679, + 496, + 595, + 496 + ], + "score": 0.82, + "latex": "s \\ll 1 \\ r _ { \\cdot }" + }, + { + "category_id": 13, + "poly": [ + 857, + 643, + 885, + 643, + 885, + 670, + 857, + 670 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1098, + 1631, + 1133, + 1631, + 1133, + 1684, + 1098, + 1684 + ], + "score": 0.78, + "latex": "\\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 906, + 467, + 999, + 467, + 999, + 500, + 906, + 500 + ], + "score": 0.58, + "latex": "N , \\sigma ( t )" + }, + { + "category_id": 13, + "poly": [ + 828, + 1115, + 849, + 1115, + 849, + 1139, + 828, + 1139 + ], + "score": 0.47, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 907, + 467, + 935, + 467, + 935, + 495, + 907, + 495 + ], + "score": 0.29, + "latex": "N" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1380.0, + 1578.0, + 1401.0, + 1578.0, + 1401.0, + 1600.0, + 1380.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1636.0, + 1097.0, + 1636.0, + 1097.0, + 1674.0, + 294.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1403.0, + 1679.0, + 1403.0, + 1713.0, + 294.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1710.0, + 1403.0, + 1710.0, + 1403.0, + 1744.0, + 295.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1736.0, + 338.0, + 1736.0, + 338.0, + 1775.0, + 293.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1736.0, + 1406.0, + 1736.0, + 1406.0, + 1775.0, + 427.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1804.0, + 295.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1802.0, + 480.0, + 1802.0, + 480.0, + 1835.0, + 296.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1802.0, + 1070.0, + 1802.0, + 1070.0, + 1835.0, + 562.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1802.0, + 1277.0, + 1802.0, + 1277.0, + 1835.0, + 1150.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 435.0, + 1189.0, + 435.0, + 1189.0, + 472.0, + 294.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 435.0, + 1406.0, + 435.0, + 1406.0, + 472.0, + 1278.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 465.0, + 594.0, + 465.0, + 594.0, + 501.0, + 294.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 465.0, + 905.0, + 465.0, + 905.0, + 501.0, + 680.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 465.0, + 1406.0, + 465.0, + 1406.0, + 501.0, + 1000.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 495.0, + 355.0, + 495.0, + 355.0, + 532.0, + 291.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 495.0, + 1408.0, + 495.0, + 1408.0, + 532.0, + 498.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 524.0, + 1403.0, + 524.0, + 1403.0, + 565.0, + 292.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 557.0, + 382.0, + 557.0, + 382.0, + 592.0, + 294.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 719.0, + 651.0, + 719.0, + 651.0, + 783.0, + 296.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 719.0, + 1160.0, + 719.0, + 1160.0, + 783.0, + 720.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 719.0, + 1404.0, + 719.0, + 1404.0, + 783.0, + 1309.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 773.0, + 895.0, + 773.0, + 895.0, + 813.0, + 293.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 773.0, + 1078.0, + 773.0, + 1078.0, + 813.0, + 1022.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 773.0, + 1405.0, + 773.0, + 1405.0, + 813.0, + 1207.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 797.0, + 1039.0, + 797.0, + 1039.0, + 858.0, + 286.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 797.0, + 1351.0, + 797.0, + 1351.0, + 858.0, + 1105.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 797.0, + 1412.0, + 797.0, + 1412.0, + 858.0, + 1393.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 841.0, + 664.0, + 841.0, + 664.0, + 877.0, + 294.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 602.0, + 1403.0, + 602.0, + 1403.0, + 641.0, + 294.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 638.0, + 386.0, + 638.0, + 386.0, + 693.0, + 294.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 638.0, + 856.0, + 638.0, + 856.0, + 693.0, + 420.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 638.0, + 1405.0, + 638.0, + 1405.0, + 693.0, + 886.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 683.0, + 396.0, + 683.0, + 396.0, + 717.0, + 295.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 683.0, + 525.0, + 683.0, + 525.0, + 717.0, + 477.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 683.0, + 1328.0, + 683.0, + 1328.0, + 717.0, + 606.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 225.0, + 657.0, + 225.0, + 657.0, + 267.0, + 291.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 225.0, + 840.0, + 225.0, + 840.0, + 267.0, + 727.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 225.0, + 1305.0, + 225.0, + 1305.0, + 267.0, + 1001.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 265.0, + 297.0, + 265.0, + 297.0, + 305.0, + 294.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 265.0, + 779.0, + 265.0, + 779.0, + 305.0, + 510.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 265.0, + 958.0, + 265.0, + 958.0, + 305.0, + 903.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 265.0, + 1405.0, + 265.0, + 1405.0, + 305.0, + 1084.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 301.0, + 375.0, + 301.0, + 375.0, + 338.0, + 296.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 301.0, + 898.0, + 301.0, + 898.0, + 338.0, + 441.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1914.0, + 644.0, + 1914.0, + 644.0, + 1980.0, + 293.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1914.0, + 1263.0, + 1914.0, + 1263.0, + 1980.0, + 714.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1963.0, + 951.0, + 1963.0, + 951.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1963.0, + 1116.0, + 1963.0, + 1116.0, + 2009.0, + 1066.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1963.0, + 1406.0, + 1963.0, + 1406.0, + 2009.0, + 1234.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1998.0, + 297.0, + 1998.0, + 297.0, + 2044.0, + 292.0, + 2044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1998.0, + 1255.0, + 1998.0, + 1255.0, + 2044.0, + 424.0, + 2044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 1998.0, + 1409.0, + 1998.0, + 1409.0, + 2044.0, + 1358.0, + 2044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1492.0, + 1225.0, + 1492.0, + 1225.0, + 1531.0, + 294.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1492.0, + 1406.0, + 1492.0, + 1406.0, + 1531.0, + 1258.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1513.0, + 594.0, + 1513.0, + 594.0, + 1571.0, + 288.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1513.0, + 653.0, + 1513.0, + 653.0, + 1571.0, + 637.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1109.0, + 827.0, + 1109.0, + 827.0, + 1149.0, + 294.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1109.0, + 1406.0, + 1109.0, + 1406.0, + 1149.0, + 850.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1139.0, + 625.0, + 1139.0, + 625.0, + 1182.0, + 293.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1139.0, + 1014.0, + 1139.0, + 1014.0, + 1182.0, + 760.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1846.0, + 1404.0, + 1846.0, + 1404.0, + 1882.0, + 296.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1880.0, + 778.0, + 1880.0, + 778.0, + 1912.0, + 296.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1310.0, + 1043.0, + 1310.0, + 1043.0, + 1357.0, + 294.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1310.0, + 1376.0, + 1310.0, + 1376.0, + 1357.0, + 1225.0, + 1357.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1041, + 1406, + 1041, + 1406, + 1137, + 297, + 1137 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 721, + 1405, + 721, + 1405, + 845, + 296, + 845 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1654, + 1404, + 1654, + 1404, + 1750, + 298, + 1750 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 226, + 1405, + 226, + 1405, + 324, + 298, + 324 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 466, + 538, + 1232, + 538, + 1232, + 650, + 466, + 650 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 628, + 1538, + 1065, + 1538, + 1065, + 1623, + 628, + 1623 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 293, + 1412, + 1403, + 1412, + 1403, + 1476, + 293, + 1476 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 296, + 1235, + 1402, + 1235, + 1402, + 1298, + 296, + 1298 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1403, + 1971, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 294, + 859, + 1402, + 859, + 1402, + 925, + 294, + 925 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 463, + 364, + 1231, + 364, + 1231, + 441, + 463, + 441 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 463, + 1806, + 1237, + 1806, + 1237, + 1942, + 463, + 1942 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 672, + 1343, + 1029, + 1343, + 1029, + 1388, + 672, + 1388 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 297, + 970, + 1025, + 970, + 1025, + 1009, + 297, + 1009 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 674, + 678, + 674, + 678, + 708, + 298, + 708 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 299, + 465, + 648, + 465, + 648, + 498, + 299, + 498 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 299, + 1176, + 579, + 1176, + 579, + 1208, + 299, + 1208 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.918 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1544, + 1400, + 1544, + 1400, + 1575, + 1365, + 1575 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1859, + 1400, + 1859, + 1400, + 1888, + 1366, + 1888 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2113, + 840, + 2113 + ], + "score": 0.822 + }, + { + "category_id": 8, + "poly": [ + 462, + 1806, + 1239, + 1806, + 1239, + 1942, + 462, + 1942 + ], + "score": 0.098 + }, + { + "category_id": 14, + "poly": [ + 464, + 539, + 1232, + 539, + 1232, + 649, + 464, + 649 + ], + "score": 0.94, + "latex": "A = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }" + }, + { + "category_id": 14, + "poly": [ + 461, + 1806, + 1238, + 1806, + 1238, + 1947, + 461, + 1947 + ], + "score": 0.94, + "latex": "\\begin{array} { c } { \\displaystyle \\ell _ { N } ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } ) = \\displaystyle \\frac { 1 } { 2 N } \\mathbb { E } _ { A } [ ( \\langle \\left( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\right) - W ^ { * } , A \\rangle ^ { 2 } ] } \\\\ { = \\displaystyle \\frac { 1 } { 2 N } | | \\big ( W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } \\big ) - W ^ { * } | | _ { F } ^ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 657, + 1264, + 785, + 1264, + 785, + 1295, + 657, + 1295 + ], + "score": 0.92, + "latex": "\\mathbf { \\bar { \\boldsymbol { W } } } \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1714, + 1387, + 1714, + 1387, + 1745, + 1246, + 1745 + ], + "score": 0.92, + "latex": "W ^ { \\ast } \\in \\mathring { \\mathbb { R } } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1034, + 784, + 1114, + 784, + 1114, + 811, + 1034, + 811 + ], + "score": 0.92, + "latex": "N = 2" + }, + { + "category_id": 14, + "poly": [ + 463, + 363, + 1232, + 363, + 1232, + 440, + 463, + 440 + ], + "score": 0.92, + "latex": "\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 681, + 262, + 745, + 262, + 745, + 295, + 681, + 295 + ], + "score": 0.92, + "latex": "( s , f )" + }, + { + "category_id": 13, + "poly": [ + 841, + 1441, + 985, + 1441, + 985, + 1475, + 841, + 1475 + ], + "score": 0.91, + "latex": "( W _ { i } ^ { \\setminus } \\in \\mathbb { R } ^ { d \\times d } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 226, + 479, + 226, + 479, + 264, + 298, + 264 + ], + "score": 0.91, + "latex": "c < 2 ( { \\sqrt { 2 } } - 1 )" + }, + { + "category_id": 14, + "poly": [ + 669, + 1342, + 1029, + 1342, + 1029, + 1388, + 669, + 1388 + ], + "score": 0.91, + "latex": "f _ { W } ( A ) = \\langle W , A \\rangle = t r \\big ( W ^ { T } A \\big )" + }, + { + "category_id": 13, + "poly": [ + 482, + 1657, + 559, + 1657, + 559, + 1684, + 482, + 1684 + ], + "score": 0.9, + "latex": "d = 1" + }, + { + "category_id": 13, + "poly": [ + 995, + 259, + 1036, + 259, + 1036, + 294, + 995, + 294 + ], + "score": 0.9, + "latex": "\\sigma _ { 0 } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1264, + 362, + 1264, + 362, + 1294, + 297, + 1294 + ], + "score": 0.9, + "latex": "\\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 14, + "poly": [ + 628, + 1538, + 1070, + 1538, + 1070, + 1625, + 628, + 1625 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { W = W _ { N } W _ { N - 1 } \\cdot \\cdot \\cdot W _ { 1 } } } \\\\ { { \\forall i \\in [ N ] , ~ W _ { i } ( 0 ) = ~ \\sqrt [ N ] { \\sigma _ { 0 } } I } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1319, + 754, + 1393, + 754, + 1393, + 782, + 1319, + 782 + ], + "score": 0.89, + "latex": "c = 0" + }, + { + "category_id": 13, + "poly": [ + 535, + 232, + 568, + 232, + 568, + 264, + 535, + 264 + ], + "score": 0.88, + "latex": "\\sigma _ { 1 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 799, + 233, + 831, + 233, + 831, + 260, + 799, + 260 + ], + "score": 0.85, + "latex": "\\sigma ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 959, + 787, + 991, + 787, + 991, + 814, + 959, + 814 + ], + "score": 0.84, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1343, + 1237, + 1406, + 1237, + 1406, + 1268, + 1343, + 1268 + ], + "score": 0.83, + "latex": "\\mathcal { X } =" + }, + { + "category_id": 13, + "poly": [ + 531, + 754, + 557, + 754, + 557, + 780, + 531, + 780 + ], + "score": 0.8, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 452, + 468, + 477, + 468, + 477, + 494, + 452, + 494 + ], + "score": 0.79, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 452, + 754, + 477, + 754, + 477, + 780, + 452, + 780 + ], + "score": 0.74, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 376, + 468, + 399, + 468, + 399, + 494, + 376, + 494 + ], + "score": 0.73, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 694, + 759, + 712, + 759, + 712, + 780, + 694, + 780 + ], + "score": 0.71, + "latex": "r" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 969.0, + 1025.0, + 969.0, + 1025.0, + 1011.0, + 293.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 581.0, + 1176.0, + 581.0, + 1212.0, + 295.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1041.0, + 1404.0, + 1041.0, + 1404.0, + 1078.0, + 294.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1074.0, + 1404.0, + 1074.0, + 1404.0, + 1109.0, + 295.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1105.0, + 970.0, + 1105.0, + 970.0, + 1139.0, + 295.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 722.0, + 1405.0, + 722.0, + 1405.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 753.0, + 451.0, + 753.0, + 451.0, + 785.0, + 294.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 753.0, + 530.0, + 753.0, + 530.0, + 785.0, + 478.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 753.0, + 693.0, + 753.0, + 693.0, + 785.0, + 558.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 753.0, + 1318.0, + 753.0, + 1318.0, + 785.0, + 713.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 753.0, + 1405.0, + 753.0, + 1405.0, + 785.0, + 1394.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 784.0, + 958.0, + 784.0, + 958.0, + 818.0, + 293.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 784.0, + 1033.0, + 784.0, + 1033.0, + 818.0, + 992.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 784.0, + 1406.0, + 784.0, + 1406.0, + 818.0, + 1115.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 815.0, + 476.0, + 815.0, + 476.0, + 844.0, + 295.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1652.0, + 481.0, + 1652.0, + 481.0, + 1691.0, + 292.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1691.0, + 560.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1686.0, + 1404.0, + 1686.0, + 1404.0, + 1720.0, + 294.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1713.0, + 1245.0, + 1713.0, + 1245.0, + 1752.0, + 289.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1388.0, + 1713.0, + 1404.0, + 1713.0, + 1404.0, + 1752.0, + 1388.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 297.0, + 227.0, + 297.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 227.0, + 534.0, + 227.0, + 534.0, + 265.0, + 480.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 227.0, + 798.0, + 227.0, + 798.0, + 265.0, + 569.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 265.0, + 832.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 254.0, + 680.0, + 254.0, + 680.0, + 301.0, + 291.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 254.0, + 994.0, + 254.0, + 994.0, + 301.0, + 746.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 254.0, + 1408.0, + 254.0, + 1408.0, + 301.0, + 1037.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 510.0, + 290.0, + 510.0, + 330.0, + 294.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1412.0, + 1406.0, + 1412.0, + 1406.0, + 1448.0, + 295.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1440.0, + 840.0, + 1440.0, + 840.0, + 1477.0, + 292.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1440.0, + 1002.0, + 1440.0, + 1002.0, + 1477.0, + 986.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 1342.0, + 1233.0, + 1342.0, + 1270.0, + 294.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1259.0, + 296.0, + 1259.0, + 296.0, + 1301.0, + 291.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1259.0, + 656.0, + 1259.0, + 656.0, + 1301.0, + 363.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1259.0, + 911.0, + 1259.0, + 911.0, + 1301.0, + 786.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 703.0, + 2000.0, + 703.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 857.0, + 1404.0, + 857.0, + 1404.0, + 896.0, + 294.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 891.0, + 998.0, + 891.0, + 998.0, + 927.0, + 296.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 671.0, + 681.0, + 671.0, + 681.0, + 713.0, + 296.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 461.0, + 375.0, + 461.0, + 375.0, + 503.0, + 297.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 461.0, + 451.0, + 461.0, + 451.0, + 503.0, + 400.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 461.0, + 651.0, + 461.0, + 651.0, + 503.0, + 478.0, + 503.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 829, + 1404, + 829, + 1404, + 1047, + 296, + 1047 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 553, + 1404, + 553, + 1404, + 740, + 297, + 740 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1534, + 1405, + 1534, + 1405, + 1689, + 298, + 1689 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1255, + 1402, + 1255, + 1402, + 1348, + 299, + 1348 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1402, + 1941, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 548, + 1097, + 1157, + 1097, + 1157, + 1233, + 548, + 1233 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1404, + 228, + 1404, + 324, + 297, + 324 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 522, + 1454, + 1175, + 1454, + 1175, + 1520, + 522, + 1520 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 674, + 361, + 1026, + 361, + 1026, + 406, + 674, + 406 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 1854, + 1404, + 1854, + 1404, + 1920, + 296, + 1920 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 289, + 1355, + 1401, + 1355, + 1401, + 1419, + 289, + 1419 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 299, + 421, + 1402, + 421, + 1402, + 485, + 299, + 485 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 295, + 1694, + 1403, + 1694, + 1403, + 1758, + 295, + 1758 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 700, + 1799, + 999, + 1799, + 999, + 1840, + 700, + 1840 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 297, + 505, + 1379, + 505, + 1379, + 540, + 297, + 540 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 105, + 299, + 105 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 301, + 774, + 763, + 774, + 763, + 805, + 301, + 805 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1127, + 1400, + 1127, + 1400, + 1158, + 1366, + 1158 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.711 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.184 + }, + { + "category_id": 14, + "poly": [ + 545, + 1092, + 1157, + 1092, + 1157, + 1238, + 545, + 1238 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { f _ { W } ( x ) = \\displaystyle \\sum _ { i = 1 } ^ { d } ( w _ { i } ^ { T } x ) ^ { 2 } = x ^ { T } W ^ { T } W x = \\langle W ^ { T } W , x x ^ { T } \\rangle } \\\\ & { W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 523, + 1453, + 1176, + 1453, + 1176, + 1519, + 523, + 1519 + ], + "score": 0.93, + "latex": "\\ell _ { v a r } ( h ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( y ( x ) - h ( x ) ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ y ( x ) - h ( x ) ] ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 921, + 433, + 921, + 433, + 952, + 297, + 952 + ], + "score": 0.92, + "latex": "W \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 298, + 981, + 452, + 981, + 452, + 1013, + 298, + 1013 + ], + "score": 0.92, + "latex": "\\bar { W } ^ { * } \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 684, + 891, + 785, + 891, + 785, + 920, + 684, + 920 + ], + "score": 0.92, + "latex": "\\mathcal { X } = \\mathbb { R } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 674, + 360, + 1024, + 360, + 1024, + 403, + 674, + 403 + ], + "score": 0.92, + "latex": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )" + }, + { + "category_id": 14, + "poly": [ + 701, + 1800, + 998, + 1800, + 998, + 1839, + 701, + 1839 + ], + "score": 0.92, + "latex": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )" + }, + { + "category_id": 13, + "poly": [ + 1274, + 1356, + 1401, + 1356, + 1401, + 1388, + 1274, + 1388 + ], + "score": 0.9, + "latex": "y : \\mathcal { X } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 456, + 425, + 487, + 425, + 487, + 456, + 456, + 456 + ], + "score": 0.89, + "latex": "\\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 456, + 1859, + 487, + 1859, + 487, + 1890, + 456, + 1890 + ], + "score": 0.89, + "latex": "\\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 905, + 1858, + 948, + 1858, + 948, + 1885, + 905, + 1885 + ], + "score": 0.86, + "latex": "W ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 904, + 423, + 948, + 423, + 948, + 451, + 904, + 451 + ], + "score": 0.86, + "latex": "W ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1287, + 381, + 1287, + 381, + 1314, + 298, + 1314 + ], + "score": 0.86, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 376, + 427, + 403, + 427, + 403, + 453, + 376, + 453 + ], + "score": 0.83, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 376, + 1861, + 404, + 1861, + 404, + 1887, + 376, + 1887 + ], + "score": 0.83, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 779, + 293, + 807, + 293, + 807, + 319, + 779, + 319 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 977, + 1357, + 1004, + 1357, + 1004, + 1383, + 977, + 1383 + ], + "score": 0.78, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 491, + 1388, + 508, + 1388, + 508, + 1413, + 491, + 1413 + ], + "score": 0.74, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 820, + 423, + 852, + 423, + 852, + 451, + 820, + 451 + ], + "score": 0.69, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 820, + 1857, + 852, + 1857, + 852, + 1885, + 820, + 1885 + ], + "score": 0.68, + "latex": "W" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 770.0, + 769.0, + 770.0, + 769.0, + 810.0, + 292.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 831.0, + 1404.0, + 831.0, + 1404.0, + 866.0, + 294.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 862.0, + 1404.0, + 862.0, + 1404.0, + 897.0, + 295.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 888.0, + 683.0, + 888.0, + 683.0, + 928.0, + 290.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 888.0, + 1406.0, + 888.0, + 1406.0, + 928.0, + 786.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 919.0, + 1405.0, + 919.0, + 1405.0, + 959.0, + 434.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 950.0, + 1404.0, + 950.0, + 1404.0, + 989.0, + 293.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 975.0, + 297.0, + 975.0, + 297.0, + 1024.0, + 291.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 975.0, + 1408.0, + 975.0, + 1408.0, + 1024.0, + 453.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 542.0, + 1011.0, + 542.0, + 1047.0, + 294.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1406.0, + 552.0, + 1406.0, + 589.0, + 294.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 585.0, + 1406.0, + 585.0, + 1406.0, + 620.0, + 294.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 617.0, + 1402.0, + 617.0, + 1402.0, + 649.0, + 292.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 647.0, + 1405.0, + 647.0, + 1405.0, + 679.0, + 295.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 676.0, + 1405.0, + 676.0, + 1405.0, + 711.0, + 292.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 707.0, + 1311.0, + 707.0, + 1311.0, + 741.0, + 294.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1536.0, + 1405.0, + 1536.0, + 1405.0, + 1569.0, + 296.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1565.0, + 1405.0, + 1565.0, + 1405.0, + 1600.0, + 293.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1597.0, + 1405.0, + 1597.0, + 1405.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1622.0, + 1406.0, + 1622.0, + 1406.0, + 1666.0, + 292.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1657.0, + 1284.0, + 1657.0, + 1284.0, + 1694.0, + 295.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1254.0, + 1404.0, + 1254.0, + 1404.0, + 1292.0, + 294.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 297.0, + 1283.0, + 297.0, + 1324.0, + 293.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1283.0, + 1407.0, + 1283.0, + 1407.0, + 1324.0, + 382.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1316.0, + 1233.0, + 1316.0, + 1233.0, + 1353.0, + 294.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1938.0, + 1403.0, + 1938.0, + 1403.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 525.0, + 2003.0, + 525.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 258.0, + 1405.0, + 258.0, + 1405.0, + 299.0, + 292.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 778.0, + 291.0, + 778.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 291.0, + 819.0, + 291.0, + 819.0, + 325.0, + 808.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 375.0, + 1851.0, + 375.0, + 1895.0, + 294.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1851.0, + 455.0, + 1851.0, + 455.0, + 1895.0, + 405.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1851.0, + 819.0, + 1851.0, + 819.0, + 1895.0, + 488.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1851.0, + 904.0, + 1851.0, + 904.0, + 1895.0, + 853.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1895.0, + 949.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1888.0, + 476.0, + 1888.0, + 476.0, + 1922.0, + 295.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1354.0, + 976.0, + 1354.0, + 976.0, + 1390.0, + 296.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1354.0, + 1273.0, + 1354.0, + 1273.0, + 1390.0, + 1005.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1354.0, + 1406.0, + 1354.0, + 1406.0, + 1390.0, + 1402.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1383.0, + 490.0, + 1383.0, + 490.0, + 1423.0, + 293.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1383.0, + 1067.0, + 1383.0, + 1067.0, + 1423.0, + 509.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 417.0, + 375.0, + 417.0, + 375.0, + 461.0, + 294.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 417.0, + 455.0, + 417.0, + 455.0, + 461.0, + 404.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 417.0, + 819.0, + 417.0, + 819.0, + 461.0, + 488.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 417.0, + 903.0, + 417.0, + 903.0, + 461.0, + 853.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 417.0, + 1406.0, + 417.0, + 1406.0, + 461.0, + 949.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 453.0, + 476.0, + 453.0, + 476.0, + 486.0, + 295.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1404.0, + 1693.0, + 1404.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 1194.0, + 1726.0, + 1194.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 504.0, + 1377.0, + 504.0, + 1377.0, + 544.0, + 296.0, + 544.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 289, + 1403, + 289, + 1403, + 536, + 297, + 536 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 923, + 1404, + 923, + 1404, + 1151, + 297, + 1151 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1333, + 1403, + 1333, + 1403, + 1518, + 298, + 1518 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 549, + 1404, + 549, + 1404, + 708, + 298, + 708 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1163, + 1403, + 1163, + 1403, + 1320, + 298, + 1320 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1752, + 1402, + 1752, + 1402, + 1905, + 299, + 1905 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1643, + 1402, + 1643, + 1402, + 1737, + 299, + 1737 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 661, + 800, + 1035, + 800, + 1035, + 891, + 661, + 891 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 294, + 721, + 1281, + 721, + 1281, + 755, + 294, + 755 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 299, + 2002, + 1212, + 2002, + 1212, + 2034, + 299, + 2034 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 300, + 1569, + 543, + 1569, + 543, + 1606, + 300, + 1606 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 301, + 1948, + 557, + 1948, + 557, + 1977, + 301, + 1977 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.781 + }, + { + "category_id": 0, + "poly": [ + 297, + 229, + 961, + 229, + 961, + 261, + 297, + 261 + ], + "score": 0.472 + }, + { + "category_id": 14, + "poly": [ + 663, + 799, + 1036, + 799, + 1036, + 893, + 663, + 893 + ], + "score": 0.94, + "latex": "\\dot { \\sigma } _ { i } = \\frac { N ^ { 2 } } { m } \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } \\sum _ { j = 1 } ^ { m } e ^ { - y _ { j } \\langle \\beta , x _ { j } \\rangle } x _ { j }" + }, + { + "category_id": 13, + "poly": [ + 384, + 644, + 510, + 644, + 510, + 678, + 384, + 678 + ], + "score": 0.94, + "latex": "\\{ x _ { i } , y _ { i } \\} _ { i = 1 } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 668, + 955, + 739, + 955, + 739, + 1001, + 668, + 1001 + ], + "score": 0.93, + "latex": "\\sigma _ { i } ^ { 2 - \\frac { 2 } { N } }" + }, + { + "category_id": 13, + "poly": [ + 591, + 644, + 718, + 644, + 718, + 678, + 591, + 678 + ], + "score": 0.93, + "latex": "y _ { i } \\in \\{ \\pm 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 328, + 476, + 426, + 476, + 426, + 502, + 328, + 502 + ], + "score": 0.91, + "latex": "t \\infty" + }, + { + "category_id": 13, + "poly": [ + 1058, + 641, + 1316, + 641, + 1316, + 678, + 1058, + 678 + ], + "score": 0.91, + "latex": "( \\ell ( f ( x ) , y ) = e ^ { - y f ( x ) } )" + }, + { + "category_id": 13, + "poly": [ + 1094, + 413, + 1180, + 413, + 1180, + 441, + 1094, + 441 + ], + "score": 0.9, + "latex": "N > 1" + }, + { + "category_id": 13, + "poly": [ + 1164, + 354, + 1249, + 354, + 1249, + 380, + 1164, + 380 + ], + "score": 0.9, + "latex": "t \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 928, + 443, + 1010, + 443, + 1010, + 471, + 928, + 471 + ], + "score": 0.9, + "latex": "N = 1" + }, + { + "category_id": 13, + "poly": [ + 297, + 1292, + 328, + 1292, + 328, + 1318, + 297, + 1318 + ], + "score": 0.81, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 779, + 729, + 802, + 729, + 802, + 750, + 779, + 750 + ], + "score": 0.79, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 964, + 931, + 984, + 931, + 984, + 953, + 964, + 953 + ], + "score": 0.79, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1159, + 729, + 1179, + 729, + 1179, + 750, + 1159, + 750 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 636, + 1003, + 656, + 1003, + 656, + 1025, + 636, + 1025 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 553, + 1033, + 573, + 1033, + 573, + 1055, + 553, + 1055 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 586, + 1430, + 606, + 1430, + 606, + 1453, + 586, + 1453 + ], + "score": 0.76, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 509, + 1457, + 534, + 1457, + 534, + 1485, + 509, + 1485 + ], + "score": 0.27, + "latex": "\\mathbf { G }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1566.0, + 549.0, + 1566.0, + 549.0, + 1614.0, + 292.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1949.0, + 558.0, + 1949.0, + 558.0, + 1979.0, + 299.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 964.0, + 229.0, + 964.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 288.0, + 1405.0, + 288.0, + 1405.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 319.0, + 1404.0, + 319.0, + 1404.0, + 357.0, + 290.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 352.0, + 1163.0, + 352.0, + 1163.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 386.0, + 1250.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 416.0, + 295.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 415.0, + 1093.0, + 415.0, + 1093.0, + 445.0, + 295.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 415.0, + 1404.0, + 415.0, + 1404.0, + 445.0, + 1181.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 440.0, + 927.0, + 440.0, + 927.0, + 478.0, + 294.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 440.0, + 1405.0, + 440.0, + 1405.0, + 478.0, + 1011.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 474.0, + 327.0, + 474.0, + 327.0, + 508.0, + 293.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 474.0, + 1408.0, + 474.0, + 1408.0, + 508.0, + 427.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 503.0, + 1120.0, + 503.0, + 1120.0, + 540.0, + 293.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 924.0, + 963.0, + 924.0, + 963.0, + 960.0, + 296.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 924.0, + 1404.0, + 924.0, + 1404.0, + 960.0, + 985.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 955.0, + 667.0, + 955.0, + 667.0, + 1012.0, + 290.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 1005.0, + 754.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 997.0, + 635.0, + 997.0, + 635.0, + 1033.0, + 294.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 997.0, + 1402.0, + 997.0, + 1402.0, + 1033.0, + 657.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1029.0, + 552.0, + 1029.0, + 552.0, + 1061.0, + 296.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1061.0, + 574.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1056.0, + 1406.0, + 1056.0, + 1406.0, + 1091.0, + 292.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1406.0, + 1087.0, + 1406.0, + 1123.0, + 294.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1119.0, + 1249.0, + 1119.0, + 1249.0, + 1154.0, + 294.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 945.5, + 766.0, + 945.5, + 766.0, + 994.0, + 678.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1335.0, + 1404.0, + 1335.0, + 1404.0, + 1367.0, + 294.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1367.0, + 1404.0, + 1367.0, + 1404.0, + 1399.0, + 296.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1395.0, + 1405.0, + 1395.0, + 1405.0, + 1431.0, + 293.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1427.0, + 585.0, + 1427.0, + 585.0, + 1459.0, + 296.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1459.0, + 607.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1452.0, + 508.0, + 1452.0, + 508.0, + 1496.0, + 291.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1452.0, + 1407.0, + 1452.0, + 1407.0, + 1496.0, + 535.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1484.0, + 419.0, + 1484.0, + 419.0, + 1525.0, + 294.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 549.0, + 1402.0, + 549.0, + 1402.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 581.0, + 1402.0, + 581.0, + 1402.0, + 615.0, + 296.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 608.0, + 1405.0, + 608.0, + 1405.0, + 650.0, + 293.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 631.0, + 383.0, + 631.0, + 383.0, + 691.0, + 288.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 631.0, + 590.0, + 631.0, + 590.0, + 691.0, + 511.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 631.0, + 1057.0, + 631.0, + 1057.0, + 691.0, + 719.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 631.0, + 1411.0, + 631.0, + 1411.0, + 691.0, + 1317.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1151.0, + 673.0, + 1151.0, + 711.0, + 294.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1164.0, + 1407.0, + 1164.0, + 1407.0, + 1201.0, + 294.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1196.0, + 1403.0, + 1196.0, + 1403.0, + 1229.0, + 296.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1225.0, + 1405.0, + 1225.0, + 1405.0, + 1263.0, + 294.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1294.0, + 292.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1287.0, + 296.0, + 1287.0, + 296.0, + 1326.0, + 293.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1287.0, + 427.0, + 1287.0, + 427.0, + 1326.0, + 329.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 1404.0, + 1749.0, + 1404.0, + 1788.0, + 294.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1784.0, + 1403.0, + 1784.0, + 1403.0, + 1817.0, + 297.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1814.0, + 1403.0, + 1814.0, + 1403.0, + 1847.0, + 294.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1843.0, + 1406.0, + 1843.0, + 1406.0, + 1877.0, + 294.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1875.0, + 1110.0, + 1875.0, + 1110.0, + 1908.0, + 297.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1643.0, + 1406.0, + 1643.0, + 1406.0, + 1680.0, + 294.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1675.0, + 1406.0, + 1675.0, + 1406.0, + 1709.0, + 295.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1705.0, + 1357.0, + 1705.0, + 1357.0, + 1741.0, + 292.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 720.0, + 778.0, + 720.0, + 778.0, + 759.0, + 295.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 720.0, + 1158.0, + 720.0, + 1158.0, + 759.0, + 803.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 720.0, + 1285.0, + 720.0, + 1285.0, + 759.0, + 1180.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1996.0, + 1219.0, + 1996.0, + 1219.0, + 2044.0, + 293.0, + 2044.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1391, + 1403, + 1391, + 1403, + 1515, + 299, + 1515 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1402, + 1941, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 297, + 362, + 1404, + 362, + 1404, + 487, + 297, + 487 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 300, + 507, + 1401, + 507, + 1401, + 601, + 300, + 601 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 303, + 1277, + 1403, + 1277, + 1403, + 1371, + 303, + 1371 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 298, + 1162, + 1402, + 1162, + 1402, + 1257, + 298, + 1257 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 297, + 622, + 1405, + 622, + 1405, + 715, + 297, + 715 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 297, + 736, + 1405, + 736, + 1405, + 831, + 297, + 831 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 297, + 850, + 1405, + 850, + 1405, + 944, + 297, + 944 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 294, + 1742, + 1402, + 1742, + 1402, + 1807, + 294, + 1807 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 294, + 1537, + 1696, + 1537, + 1696, + 1723, + 294, + 1723 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 301, + 1826, + 1403, + 1826, + 1403, + 1920, + 301, + 1920 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 297, + 1049, + 1401, + 1049, + 1401, + 1142, + 297, + 1142 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 295, + 965, + 1403, + 965, + 1403, + 1030, + 295, + 1030 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 299, + 279, + 1399, + 279, + 1399, + 342, + 299, + 342 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 261, + 300, + 261 + ], + "score": 0.821 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.769 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 490.0, + 229.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1392.0, + 1405.0, + 1392.0, + 1405.0, + 1429.0, + 297.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1425.0, + 1403.0, + 1425.0, + 1403.0, + 1458.0, + 322.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1454.0, + 1407.0, + 1454.0, + 1407.0, + 1490.0, + 320.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1479.0, + 398.0, + 1479.0, + 398.0, + 1519.0, + 320.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1402.0, + 1942.0, + 1402.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 320.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 573.0, + 2002.0, + 573.0, + 2034.0, + 322.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 360.0, + 1405.0, + 360.0, + 1405.0, + 399.0, + 294.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 394.0, + 1405.0, + 394.0, + 1405.0, + 430.0, + 321.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 423.0, + 1402.0, + 423.0, + 1402.0, + 458.0, + 322.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 455.0, + 1362.0, + 455.0, + 1362.0, + 491.0, + 322.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1405.0, + 505.0, + 1405.0, + 542.0, + 294.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 539.0, + 1405.0, + 539.0, + 1405.0, + 573.0, + 323.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 563.0, + 399.0, + 563.0, + 399.0, + 604.0, + 320.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1406.0, + 1274.0, + 1406.0, + 1313.0, + 294.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1308.0, + 1404.0, + 1308.0, + 1404.0, + 1343.0, + 322.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1341.0, + 706.0, + 1341.0, + 706.0, + 1371.0, + 320.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1163.0, + 1404.0, + 1163.0, + 1404.0, + 1197.0, + 296.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1191.0, + 1407.0, + 1191.0, + 1407.0, + 1230.0, + 320.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1220.0, + 402.0, + 1220.0, + 402.0, + 1261.0, + 319.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 623.0, + 1403.0, + 623.0, + 1403.0, + 656.0, + 296.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 691.0, + 321.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 683.0, + 534.0, + 683.0, + 534.0, + 714.0, + 324.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 735.0, + 1402.0, + 735.0, + 1402.0, + 773.0, + 295.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 764.0, + 1405.0, + 764.0, + 1405.0, + 807.0, + 320.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 798.0, + 677.0, + 798.0, + 677.0, + 832.0, + 322.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 851.0, + 1405.0, + 851.0, + 1405.0, + 885.0, + 296.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 878.0, + 1405.0, + 878.0, + 1405.0, + 921.0, + 320.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 910.0, + 533.0, + 910.0, + 533.0, + 944.0, + 323.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1742.0, + 1403.0, + 1742.0, + 1403.0, + 1778.0, + 296.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1773.0, + 1280.0, + 1773.0, + 1280.0, + 1809.0, + 321.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1535.0, + 1181.0, + 1535.0, + 1181.0, + 1572.0, + 293.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1534.0, + 1405.0, + 1534.0, + 1405.0, + 1574.0, + 1240.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1566.0, + 1325.0, + 1566.0, + 1325.0, + 1602.0, + 319.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1567.0, + 1408.0, + 1567.0, + 1408.0, + 1602.0, + 1367.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1596.0, + 1408.0, + 1596.0, + 1408.0, + 1633.0, + 319.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1627.0, + 737.0, + 1627.0, + 737.0, + 1664.0, + 318.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1624.0, + 1403.0, + 1624.0, + 1403.0, + 1666.0, + 831.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1660.0, + 1704.0, + 1660.0, + 1704.0, + 1693.0, + 322.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1692.0, + 390.0, + 1692.0, + 390.0, + 1721.0, + 316.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1824.0, + 1403.0, + 1824.0, + 1403.0, + 1862.0, + 294.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1858.0, + 1403.0, + 1858.0, + 1403.0, + 1892.0, + 324.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1886.0, + 666.0, + 1886.0, + 666.0, + 1924.0, + 323.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1047.0, + 1406.0, + 1047.0, + 1406.0, + 1086.0, + 295.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1081.0, + 1405.0, + 1081.0, + 1405.0, + 1115.0, + 321.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1109.0, + 400.0, + 1109.0, + 400.0, + 1145.0, + 320.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 965.0, + 1405.0, + 965.0, + 1405.0, + 1001.0, + 295.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 995.0, + 942.0, + 995.0, + 942.0, + 1030.0, + 324.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 279.0, + 1404.0, + 279.0, + 1404.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 311.0, + 1203.0, + 311.0, + 1203.0, + 344.0, + 321.0, + 344.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 490, + 1407, + 490, + 1407, + 643, + 297, + 643 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 877, + 1404, + 877, + 1404, + 970, + 299, + 970 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1070, + 1405, + 1070, + 1405, + 1166, + 297, + 1166 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1399, + 1941, + 1399, + 2036, + 299, + 2036 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 557, + 1612, + 1138, + 1612, + 1138, + 1709, + 557, + 1709 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 413, + 1825, + 1284, + 1825, + 1284, + 1923, + 413, + 1923 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 579, + 1405, + 1119, + 1405, + 1119, + 1495, + 579, + 1495 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 299, + 1304, + 1402, + 1304, + 1402, + 1371, + 299, + 1371 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 1512, + 1403, + 1512, + 1403, + 1577, + 298, + 1577 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 668, + 1202, + 1030, + 1202, + 1030, + 1287, + 668, + 1287 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 297, + 1726, + 1406, + 1726, + 1406, + 1791, + 297, + 1791 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 732, + 1008, + 966, + 1008, + 966, + 1052, + 732, + 1052 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 1401, + 227, + 1401, + 294, + 296, + 294 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 376, + 298, + 376 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 106, + 299, + 106 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 301, + 427, + 683, + 427, + 683, + 466, + 301, + 466 + ], + "score": 0.908 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1854, + 1400, + 1854, + 1400, + 1885, + 1366, + 1885 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 866, + 2087, + 866, + 2113, + 836, + 2113 + ], + "score": 0.877 + }, + { + "category_id": 8, + "poly": [ + 426, + 693, + 1082, + 693, + 1082, + 850, + 426, + 850 + ], + "score": 0.746 + }, + { + "category_id": 8, + "poly": [ + 422, + 692, + 1151, + 692, + 1151, + 849, + 422, + 849 + ], + "score": 0.687 + }, + { + "category_id": 9, + "poly": [ + 1194, + 716, + 1277, + 716, + 1277, + 748, + 1194, + 748 + ], + "score": 0.155 + }, + { + "category_id": 13, + "poly": [ + 1166, + 492, + 1309, + 492, + 1309, + 544, + 1166, + 544 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 412, + 1823, + 1286, + 1823, + 1286, + 1923, + 412, + 1923 + ], + "score": 0.94, + "latex": "\\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } + \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }" + }, + { + "category_id": 14, + "poly": [ + 422, + 692, + 1280, + 692, + 1280, + 851, + 422, + 851 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { s } { r f } \\Big ) ^ { \\frac { 1 - s } { r - 1 } } } & { \\qquad } & { N = 2 } \\\\ { s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\le \\sigma _ { 0 } ^ { t h } \\le s \\sigma _ { j } ^ { * } \\Big ( \\displaystyle \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } } } & { \\qquad } & { N \\ge 3 } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 669, + 1199, + 1030, + 1199, + 1030, + 1285, + 669, + 1285 + ], + "score": 0.93, + "latex": "t _ { \\alpha } ( \\sigma ) = \\int _ { \\sigma _ { 0 } } ^ { \\alpha \\sigma ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma ) }" + }, + { + "category_id": 13, + "poly": [ + 735, + 1305, + 895, + 1305, + 895, + 1341, + 735, + 1341 + ], + "score": 0.93, + "latex": "\\sigma _ { i } ( t _ { f } ) = f \\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 896, + 540, + 1021, + 540, + 1021, + 579, + 896, + 579 + ], + "score": 0.93, + "latex": "s \\in ( 0 , \\frac { 1 } { 4 } )" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1306, + 1234, + 1306, + 1234, + 1344, + 1078, + 1344 + ], + "score": 0.93, + "latex": "\\sigma _ { j } ( t _ { s } ) = s \\sigma _ { j } ^ { * }" + }, + { + "category_id": 14, + "poly": [ + 557, + 1610, + 1139, + 1610, + 1139, + 1709, + 557, + 1709 + ], + "score": 0.93, + "latex": "\\int _ { \\sigma _ { 0 } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }" + }, + { + "category_id": 13, + "poly": [ + 601, + 908, + 781, + 908, + 781, + 943, + 601, + 943 + ], + "score": 0.93, + "latex": "t _ { f } ( \\sigma _ { i } ) \\leq t _ { s } ( \\sigma _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 516, + 1133, + 583, + 1133, + 583, + 1166, + 516, + 1166 + ], + "score": 0.93, + "latex": "t _ { \\alpha } ( \\sigma )" + }, + { + "category_id": 14, + "poly": [ + 576, + 1402, + 1120, + 1402, + 1120, + 1494, + 576, + 1494 + ], + "score": 0.92, + "latex": "\\int _ { \\sigma _ { 0 } } ^ { f \\sigma _ { i } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma ) } \\le \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { j } ^ { * } - \\sigma ) }" + }, + { + "category_id": 13, + "poly": [ + 1079, + 542, + 1206, + 542, + 1206, + 579, + 1079, + 579 + ], + "score": 0.92, + "latex": "f \\in \\left( \\frac { 3 } { 4 } , 1 \\right)" + }, + { + "category_id": 13, + "poly": [ + 1040, + 580, + 1104, + 580, + 1104, + 613, + 1040, + 613 + ], + "score": 0.92, + "latex": "( s , f )" + }, + { + "category_id": 13, + "poly": [ + 542, + 1072, + 621, + 1072, + 621, + 1102, + 542, + 1102 + ], + "score": 0.91, + "latex": "N \\geq 3" + }, + { + "category_id": 14, + "poly": [ + 733, + 1005, + 965, + 1005, + 965, + 1051, + 733, + 1051 + ], + "score": 0.9, + "latex": "\\dot { \\sigma } = \\sigma ^ { 2 - \\frac { 2 } { N } } ( \\sigma ^ { * } - \\sigma )" + }, + { + "category_id": 13, + "poly": [ + 627, + 504, + 695, + 504, + 695, + 533, + 627, + 533 + ], + "score": 0.9, + "latex": "\\sigma _ { i } , \\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1352, + 578, + 1392, + 578, + 1392, + 612, + 1352, + 612 + ], + "score": 0.89, + "latex": "\\sigma _ { 0 } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 435, + 1518, + 544, + 1518, + 544, + 1548, + 435, + 1548 + ], + "score": 0.89, + "latex": "\\sigma _ { i } = r \\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 797, + 881, + 825, + 881, + 825, + 908, + 797, + 908 + ], + "score": 0.88, + "latex": "t _ { \\alpha }" + }, + { + "category_id": 13, + "poly": [ + 1371, + 1074, + 1400, + 1074, + 1400, + 1102, + 1371, + 1102 + ], + "score": 0.86, + "latex": "t _ { \\alpha }" + }, + { + "category_id": 13, + "poly": [ + 1329, + 2007, + 1359, + 2007, + 1359, + 2035, + 1329, + 2035 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 944, + 1948, + 974, + 1948, + 974, + 1973, + 944, + 1973 + ], + "score": 0.83, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1137, + 328, + 1137, + 328, + 1163, + 298, + 1163 + ], + "score": 0.81, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 607, + 2007, + 637, + 2007, + 637, + 2034, + 607, + 2034 + ], + "score": 0.81, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1221, + 883, + 1243, + 883, + 1243, + 905, + 1221, + 905 + ], + "score": 0.78, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 426.0, + 684.0, + 426.0, + 684.0, + 470.0, + 294.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 485.0, + 626.0, + 485.0, + 626.0, + 552.0, + 296.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 485.0, + 1165.0, + 485.0, + 1165.0, + 552.0, + 696.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 485.0, + 1404.0, + 485.0, + 1404.0, + 552.0, + 1310.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 539.0, + 895.0, + 539.0, + 895.0, + 584.0, + 292.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 539.0, + 1078.0, + 539.0, + 1078.0, + 584.0, + 1022.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 539.0, + 1408.0, + 539.0, + 1408.0, + 584.0, + 1207.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 568.0, + 1039.0, + 568.0, + 1039.0, + 622.0, + 290.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 568.0, + 1351.0, + 568.0, + 1351.0, + 622.0, + 1105.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 568.0, + 1408.0, + 568.0, + 1408.0, + 622.0, + 1393.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 607.0, + 856.0, + 607.0, + 856.0, + 647.0, + 292.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 875.0, + 796.0, + 875.0, + 796.0, + 913.0, + 294.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 875.0, + 1220.0, + 875.0, + 1220.0, + 913.0, + 826.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 913.0, + 1244.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 908.0, + 600.0, + 908.0, + 600.0, + 945.0, + 294.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 908.0, + 1405.0, + 908.0, + 1405.0, + 945.0, + 782.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 938.0, + 735.0, + 938.0, + 735.0, + 971.0, + 294.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1068.0, + 541.0, + 1068.0, + 541.0, + 1107.0, + 294.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1068.0, + 1370.0, + 1068.0, + 1370.0, + 1107.0, + 622.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1068.0, + 1405.0, + 1068.0, + 1405.0, + 1107.0, + 1401.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1103.0, + 1403.0, + 1103.0, + 1403.0, + 1134.0, + 297.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1133.0, + 297.0, + 1133.0, + 297.0, + 1171.0, + 294.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1133.0, + 515.0, + 1133.0, + 515.0, + 1171.0, + 329.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1133.0, + 831.0, + 1133.0, + 831.0, + 1171.0, + 584.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 943.0, + 1942.0, + 943.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 975.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 606.0, + 2002.0, + 606.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 2002.0, + 1328.0, + 2002.0, + 1328.0, + 2040.0, + 638.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 2002.0, + 1373.0, + 2002.0, + 1373.0, + 2040.0, + 1360.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1302.0, + 734.0, + 1302.0, + 734.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1302.0, + 1077.0, + 1302.0, + 1077.0, + 1343.0, + 896.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1302.0, + 1404.0, + 1302.0, + 1404.0, + 1343.0, + 1235.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1332.0, + 699.0, + 1332.0, + 699.0, + 1376.0, + 294.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1510.0, + 434.0, + 1510.0, + 434.0, + 1550.0, + 293.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1510.0, + 1405.0, + 1510.0, + 1405.0, + 1550.0, + 545.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1539.0, + 541.0, + 1539.0, + 541.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1726.0, + 1404.0, + 1726.0, + 1404.0, + 1762.0, + 296.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1757.0, + 557.0, + 1757.0, + 557.0, + 1793.0, + 294.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1403.0, + 228.0, + 1403.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 261.0, + 1397.0, + 261.0, + 1397.0, + 293.0, + 325.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 312.0, + 1404.0, + 312.0, + 1404.0, + 345.0, + 296.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 342.0, + 1255.0, + 342.0, + 1255.0, + 376.0, + 323.0, + 376.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 286, + 1406, + 286, + 1406, + 411, + 297, + 411 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1610, + 1405, + 1610, + 1405, + 1704, + 297, + 1704 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 562, + 1405, + 562, + 1405, + 666, + 297, + 666 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1325, + 1404, + 1325, + 1404, + 1409, + 298, + 1409 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 524, + 447, + 1173, + 447, + 1173, + 544, + 524, + 544 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 412, + 1739, + 1283, + 1739, + 1283, + 1837, + 412, + 1837 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 639, + 1444, + 1056, + 1444, + 1056, + 1524, + 639, + 1524 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 595, + 701, + 1102, + 701, + 1102, + 790, + 595, + 790 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 292, + 1855, + 1402, + 1855, + 1402, + 1921, + 292, + 1921 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 294, + 814, + 1403, + 814, + 1403, + 876, + 294, + 876 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 540, + 1954, + 1156, + 1954, + 1156, + 2044, + 540, + 2044 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 708, + 1071, + 987, + 1071, + 987, + 1138, + 708, + 1138 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 427, + 1222, + 1270, + 1222, + 1270, + 1304, + 427, + 1304 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 294, + 1157, + 1073, + 1157, + 1073, + 1192, + 294, + 1192 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 484, + 913, + 1213, + 913, + 1213, + 985, + 484, + 985 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 297, + 1003, + 823, + 1003, + 823, + 1036, + 297, + 1036 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 588, + 229, + 588, + 262, + 299, + 262 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 299, + 1553, + 592, + 1553, + 592, + 1585, + 299, + 1585 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 863, + 2087, + 863, + 2113, + 835, + 2113 + ], + "score": 0.862 + }, + { + "category_id": 14, + "poly": [ + 523, + 445, + 1175, + 445, + 1175, + 545, + 523, + 545 + ], + "score": 0.94, + "latex": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) }" + }, + { + "category_id": 14, + "poly": [ + 410, + 1737, + 1286, + 1737, + 1286, + 1838, + 410, + 1838 + ], + "score": 0.94, + "latex": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) } \\leq r \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } ) } - \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } ( 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } ) }" + }, + { + "category_id": 14, + "poly": [ + 541, + 1954, + 1158, + 1954, + 1158, + 2044, + 541, + 2044 + ], + "score": 0.93, + "latex": "\\frac { r } { r - s } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le \\big ( \\frac { r } { 1 - s } - \\frac { r } { r - s } \\big ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }" + }, + { + "category_id": 14, + "poly": [ + 707, + 1071, + 990, + 1071, + 990, + 1139, + 707, + 1139 + ], + "score": 0.93, + "latex": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( { \\frac { s } { r f } } \\big ) ^ { \\frac { 1 } { ( 1 - f ) ( r - 1 ) } }" + }, + { + "category_id": 14, + "poly": [ + 427, + 1224, + 1268, + 1224, + 1268, + 1303, + 427, + 1303 + ], + "score": 0.93, + "latex": "\\frac { 1 } { 1 - f } \\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq ( r - 1 ) \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )" + }, + { + "category_id": 14, + "poly": [ + 593, + 701, + 1107, + 701, + 1107, + 790, + 593, + 790 + ], + "score": 0.93, + "latex": "\\frac { 1 } { 1 - f } \\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\le ( r - 1 ) \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }" + }, + { + "category_id": 13, + "poly": [ + 1124, + 1323, + 1257, + 1323, + 1257, + 1382, + 1124, + 1382 + ], + "score": 0.93, + "latex": "\\Big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } }" + }, + { + "category_id": 14, + "poly": [ + 640, + 1443, + 1058, + 1443, + 1058, + 1523, + 640, + 1523 + ], + "score": 0.93, + "latex": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { ( 1 - f ) ( r - 1 ) } { 1 + ( 1 - f ) ( r - 1 ) } \\Big ) ^ { \\frac { N } { N - 2 } }" + }, + { + "category_id": 14, + "poly": [ + 486, + 913, + 1212, + 913, + 1212, + 985, + 486, + 985 + ], + "score": 0.92, + "latex": "\\frac { 1 } { 1 - f } \\big ( \\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\big ) \\leq ( r - 1 ) \\big ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 1307, + 813, + 1348, + 813, + 1348, + 849, + 1307, + 849 + ], + "score": 0.92, + "latex": "\\sigma _ { 0 } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 992, + 594, + 1086, + 594, + 1086, + 639, + 992, + 639 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { 1 - \\frac { \\sigma } { r \\sigma _ { j } ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1119, + 1858, + 1221, + 1858, + 1221, + 1893, + 1119, + 1893 + ], + "score": 0.92, + "latex": "\\sigma = s \\sigma _ { j } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 431, + 594, + 514, + 594, + 514, + 639, + 431, + 639 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { 1 - \\frac { \\sigma } { \\sigma _ { j } ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 383, + 1160, + 463, + 1160, + 463, + 1189, + 383, + 1189 + ], + "score": 0.9, + "latex": "N \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 578, + 847, + 657, + 847, + 657, + 873, + 578, + 873 + ], + "score": 0.9, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 1148, + 594, + 1217, + 594, + 1217, + 626, + 1148, + 626 + ], + "score": 0.9, + "latex": "1 - f" + }, + { + "category_id": 13, + "poly": [ + 909, + 1618, + 939, + 1618, + 939, + 1642, + 909, + 1642 + ], + "score": 0.86, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 898, + 294, + 928, + 294, + 928, + 318, + 898, + 318 + ], + "score": 0.86, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 781, + 817, + 809, + 817, + 809, + 843, + 781, + 843 + ], + "score": 0.86, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 936, + 821, + 967, + 821, + 967, + 846, + 936, + 846 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 952, + 350, + 1032, + 350, + 1032, + 379, + 952, + 379 + ], + "score": 0.85, + "latex": "( r > 1 )" + }, + { + "category_id": 13, + "poly": [ + 549, + 1010, + 579, + 1010, + 579, + 1034, + 549, + 1034 + ], + "score": 0.84, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1382, + 328, + 1382, + 328, + 1408, + 298, + 1408 + ], + "score": 0.81, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 990, + 570, + 1010, + 570, + 1010, + 591, + 990, + 591 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 946, + 1863, + 965, + 1863, + 965, + 1884, + 946, + 1884 + ], + "score": 0.75, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 448, + 355, + 465, + 355, + 465, + 376, + 448, + 376 + ], + "score": 0.71, + "latex": "r" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 589.0, + 227.0, + 589.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1552.0, + 593.0, + 1552.0, + 593.0, + 1588.0, + 295.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 284.0, + 897.0, + 284.0, + 897.0, + 323.0, + 292.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 284.0, + 1406.0, + 284.0, + 1406.0, + 323.0, + 929.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 318.0, + 1404.0, + 318.0, + 1404.0, + 350.0, + 294.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 348.0, + 447.0, + 348.0, + 447.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 348.0, + 951.0, + 348.0, + 951.0, + 385.0, + 466.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 348.0, + 1407.0, + 348.0, + 1407.0, + 385.0, + 1033.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 380.0, + 577.0, + 380.0, + 577.0, + 413.0, + 295.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1611.0, + 908.0, + 1611.0, + 908.0, + 1645.0, + 295.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1611.0, + 1405.0, + 1611.0, + 1405.0, + 1645.0, + 940.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 1403.0, + 1642.0, + 1403.0, + 1676.0, + 295.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1671.0, + 409.0, + 1671.0, + 409.0, + 1709.0, + 291.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 560.0, + 989.0, + 560.0, + 989.0, + 597.0, + 292.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 560.0, + 1406.0, + 560.0, + 1406.0, + 597.0, + 1011.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 595.0, + 430.0, + 595.0, + 430.0, + 625.0, + 293.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 587.0, + 991.0, + 587.0, + 991.0, + 648.0, + 515.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 592.0, + 1147.0, + 592.0, + 1147.0, + 647.0, + 1087.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 592.0, + 1403.0, + 592.0, + 1403.0, + 647.0, + 1218.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 629.0, + 362.0, + 629.0, + 362.0, + 669.0, + 291.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 1123.0, + 1334.0, + 1123.0, + 1373.0, + 293.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1376.0, + 297.0, + 1376.0, + 297.0, + 1413.0, + 294.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1376.0, + 591.0, + 1376.0, + 591.0, + 1413.0, + 329.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1319.0, + 1407.0, + 1319.0, + 1407.0, + 1376.0, + 1185.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1339.0, + 1195.0, + 1339.0, + 1195.0, + 1383.5, + 1125.0, + 1383.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1854.0, + 945.0, + 1854.0, + 945.0, + 1891.0, + 294.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1854.0, + 1118.0, + 1854.0, + 1118.0, + 1891.0, + 966.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1854.0, + 1404.0, + 1854.0, + 1404.0, + 1891.0, + 1222.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1883.0, + 1296.0, + 1883.0, + 1296.0, + 1924.0, + 294.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 804.0, + 780.0, + 804.0, + 780.0, + 860.0, + 289.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 804.0, + 935.0, + 804.0, + 935.0, + 860.0, + 810.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 804.0, + 1306.0, + 804.0, + 1306.0, + 860.0, + 968.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 804.0, + 1411.0, + 804.0, + 1411.0, + 860.0, + 1349.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 577.0, + 844.0, + 577.0, + 879.0, + 294.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 844.0, + 670.0, + 844.0, + 670.0, + 879.0, + 658.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1153.0, + 382.0, + 1153.0, + 382.0, + 1197.0, + 293.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1153.0, + 1075.0, + 1153.0, + 1075.0, + 1197.0, + 464.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1001.0, + 548.0, + 1001.0, + 548.0, + 1039.0, + 295.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1001.0, + 825.0, + 1001.0, + 825.0, + 1039.0, + 580.0, + 1039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1508, + 1407, + 1508, + 1407, + 1729, + 295, + 1729 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1256, + 1404, + 1256, + 1404, + 1351, + 298, + 1351 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1148, + 1404, + 1148, + 1404, + 1243, + 298, + 1243 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 465, + 1934, + 1232, + 1934, + 1232, + 2046, + 465, + 2046 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 471, + 809, + 1226, + 809, + 1226, + 890, + 471, + 890 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 639, + 301, + 1059, + 301, + 1059, + 390, + 639, + 390 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 537, + 486, + 1157, + 486, + 1157, + 553, + 537, + 553 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 740, + 649, + 955, + 649, + 955, + 717, + 740, + 717 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 718, + 985, + 979, + 985, + 979, + 1059, + 718, + 1059 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 463, + 1765, + 1231, + 1765, + 1231, + 1842, + 463, + 1842 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 296, + 414, + 1095, + 414, + 1095, + 448, + 296, + 448 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 297, + 577, + 831, + 577, + 831, + 610, + 297, + 610 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 296, + 740, + 882, + 740, + 882, + 775, + 296, + 775 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 298, + 913, + 830, + 913, + 830, + 946, + 298, + 946 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 300, + 1864, + 648, + 1864, + 648, + 1897, + 300, + 1897 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 805, + 229, + 805, + 263, + 298, + 263 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 300, + 1443, + 680, + 1443, + 680, + 1480, + 300, + 1480 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 817, + 73, + 817, + 106, + 297, + 106 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1366, + 1403, + 1366, + 1403, + 1395, + 1374, + 1395 + ], + "score": 0.807 + }, + { + "category_id": 1, + "poly": [ + 298, + 1092, + 426, + 1092, + 426, + 1123, + 298, + 1123 + ], + "score": 0.602 + }, + { + "category_id": 0, + "poly": [ + 298, + 1092, + 426, + 1092, + 426, + 1123, + 298, + 1123 + ], + "score": 0.363 + }, + { + "category_id": 13, + "poly": [ + 1256, + 1591, + 1356, + 1591, + 1356, + 1632, + 1256, + 1632 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\eta \\ \\leq \\ \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 465, + 1938, + 1231, + 1938, + 1231, + 2044, + 465, + 2044 + ], + "score": 0.93, + "latex": "A = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } \\qquad B = \\frac { \\log \\left( 1 - c \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } - \\frac { c ^ { 2 } } { 4 } \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) } { \\log \\left( 1 - c \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } + c ^ { 2 } \\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 } \\right) }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 1152, + 1392, + 1152, + 1392, + 1183, + 1328, + 1183 + ], + "score": 0.93, + "latex": "( s , f )" + }, + { + "category_id": 14, + "poly": [ + 463, + 1765, + 1232, + 1765, + 1232, + 1841, + 463, + 1841 + ], + "score": 0.93, + "latex": "\\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { A - 1 } } \\leq \\sigma _ { 0 } ^ { t h } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1591, + 423, + 1591, + 423, + 1628, + 298, + 1628 + ], + "score": 0.93, + "latex": "\\sigma _ { j } ^ { * } ~ \\geq ~ 2 \\sigma _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 741, + 648, + 957, + 648, + 957, + 718, + 741, + 718 + ], + "score": 0.93, + "latex": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\big ( \\frac { s } { r f } \\big ) ^ { \\frac { 1 - s } { r - 1 } }" + }, + { + "category_id": 14, + "poly": [ + 637, + 299, + 1062, + 299, + 1062, + 390, + 637, + 390 + ], + "score": 0.93, + "latex": "\\int _ { s \\sigma _ { j } ^ { * } } ^ { f r \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } } \\leq \\frac { r - 1 } { 1 - s } \\int _ { \\sigma _ { 0 } } ^ { s \\sigma _ { j } ^ { * } } \\frac { d \\sigma } { \\sigma ^ { 2 - \\frac { 2 } { N } } }" + }, + { + "category_id": 14, + "poly": [ + 717, + 982, + 981, + 982, + 981, + 1058, + 717, + 1058 + ], + "score": 0.93, + "latex": "\\sigma _ { 0 } \\leq s \\sigma _ { j } ^ { * } \\Big ( \\frac { r - 1 } { r - s } \\Big ) ^ { \\frac { N } { N - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 1258, + 1507, + 1401, + 1507, + 1401, + 1561, + 1258, + 1561 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1557, + 1233, + 1557, + 1233, + 1595, + 1117, + 1595 + ], + "score": 0.92, + "latex": "f \\in ( \\frac { 3 } { 4 } , 1 )" + }, + { + "category_id": 13, + "poly": [ + 903, + 417, + 981, + 417, + 981, + 444, + 903, + 444 + ], + "score": 0.92, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 951, + 1557, + 1065, + 1557, + 1065, + 1595, + 951, + 1595 + ], + "score": 0.92, + "latex": "s \\in ( 0 , \\frac { 1 } { 4 } )" + }, + { + "category_id": 13, + "poly": [ + 681, + 1665, + 745, + 1665, + 745, + 1698, + 681, + 1698 + ], + "score": 0.92, + "latex": "( s , f )" + }, + { + "category_id": 14, + "poly": [ + 470, + 811, + 1226, + 811, + 1226, + 890, + 470, + 890 + ], + "score": 0.91, + "latex": "\\Big ( \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { r f \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big ) \\leq \\frac { r - 1 } { 1 - s } \\Big ( \\big ( \\frac { 1 } { \\sigma _ { 0 } } \\big ) ^ { 1 - \\frac { 2 } { N } } - \\big ( \\frac { 1 } { s \\sigma _ { j } ^ { * } } \\big ) ^ { 1 - \\frac { 2 } { N } } \\Big )" + }, + { + "category_id": 13, + "poly": [ + 427, + 743, + 506, + 743, + 506, + 772, + 427, + 772 + ], + "score": 0.91, + "latex": "N \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 298, + 1630, + 478, + 1630, + 478, + 1667, + 298, + 1667 + ], + "score": 0.91, + "latex": "c < 2 ( { \\sqrt { 2 } } - 1 )" + }, + { + "category_id": 13, + "poly": [ + 996, + 1663, + 1036, + 1663, + 1036, + 1698, + 996, + 1698 + ], + "score": 0.9, + "latex": "\\sigma _ { 0 } ^ { t h }" + }, + { + "category_id": 14, + "poly": [ + 539, + 484, + 1157, + 484, + 1157, + 554, + 539, + 554 + ], + "score": 0.9, + "latex": "\\log ( f r \\sigma _ { j } ^ { * } ) - \\log ( s \\sigma _ { j } ^ { * } ) \\leq \\frac { r - 1 } { 1 - s } \\Bigl ( \\log ( s \\sigma _ { j } ^ { * } ) - \\log ( \\sigma _ { 0 } ) \\Bigr )" + }, + { + "category_id": 13, + "poly": [ + 627, + 1521, + 695, + 1521, + 695, + 1550, + 627, + 1550 + ], + "score": 0.9, + "latex": "\\sigma _ { i } , \\sigma _ { j }" + }, + { + "category_id": 13, + "poly": [ + 535, + 1635, + 568, + 1635, + 568, + 1667, + 535, + 1667 + ], + "score": 0.89, + "latex": "\\sigma _ { 1 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 898, + 1209, + 943, + 1209, + 943, + 1244, + 898, + 1244 + ], + "score": 0.87, + "latex": "\\sigma _ { 0 } ^ { t h \\cdot }" + }, + { + "category_id": 13, + "poly": [ + 918, + 1319, + 1003, + 1319, + 1003, + 1350, + 918, + 1350 + ], + "score": 0.86, + "latex": "( N \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 800, + 1636, + 831, + 1636, + 831, + 1662, + 800, + 1662 + ], + "score": 0.85, + "latex": "\\sigma ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 549, + 920, + 579, + 920, + 579, + 944, + 549, + 944 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 1185, + 1030, + 1185, + 1030, + 1211, + 1000, + 1211 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 653, + 1319, + 734, + 1319, + 734, + 1348, + 653, + 1348 + ], + "score": 0.85, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1156, + 1274, + 1156, + 1274, + 1181, + 1242, + 1181 + ], + "score": 0.84, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 549, + 584, + 579, + 584, + 579, + 609, + 549, + 609 + ], + "score": 0.84, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 452, + 1867, + 477, + 1867, + 477, + 1893, + 452, + 1893 + ], + "score": 0.8, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 445, + 1152, + 474, + 1152, + 474, + 1179, + 445, + 1179 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1261, + 1265, + 1278, + 1265, + 1278, + 1285, + 1261, + 1285 + ], + "score": 0.71, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 376, + 1867, + 400, + 1867, + 400, + 1893, + 376, + 1893 + ], + "score": 0.71, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 770, + 1295, + 788, + 1295, + 788, + 1316, + 770, + 1316 + ], + "score": 0.64, + "latex": "r" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1441.0, + 681.0, + 1441.0, + 681.0, + 1485.0, + 293.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 1370.0, + 1403.0, + 1370.0, + 1403.0, + 1396.0, + 1378.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1092.0, + 429.0, + 1092.0, + 429.0, + 1125.0, + 297.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1511.0, + 626.0, + 1511.0, + 626.0, + 1552.0, + 296.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1511.0, + 1257.0, + 1511.0, + 1257.0, + 1552.0, + 696.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1556.0, + 950.0, + 1556.0, + 950.0, + 1598.0, + 292.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1556.0, + 1116.0, + 1556.0, + 1116.0, + 1598.0, + 1066.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1556.0, + 1406.0, + 1556.0, + 1406.0, + 1598.0, + 1234.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1588.0, + 297.0, + 1588.0, + 297.0, + 1630.0, + 292.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1588.0, + 1255.0, + 1588.0, + 1255.0, + 1630.0, + 424.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 1588.0, + 1405.0, + 1588.0, + 1405.0, + 1630.0, + 1357.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1630.0, + 297.0, + 1630.0, + 297.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1630.0, + 534.0, + 1630.0, + 534.0, + 1667.0, + 479.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1630.0, + 799.0, + 1630.0, + 799.0, + 1667.0, + 569.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1667.0, + 832.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1657.0, + 680.0, + 1657.0, + 680.0, + 1705.0, + 290.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1657.0, + 995.0, + 1657.0, + 995.0, + 1705.0, + 746.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1657.0, + 1409.0, + 1657.0, + 1409.0, + 1705.0, + 1037.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1692.0, + 509.0, + 1692.0, + 509.0, + 1733.0, + 293.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 1260.0, + 1255.0, + 1260.0, + 1294.0, + 293.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1294.0, + 1279.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1289.0, + 769.0, + 1289.0, + 769.0, + 1323.0, + 296.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1289.0, + 1404.0, + 1289.0, + 1404.0, + 1323.0, + 789.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1319.0, + 652.0, + 1319.0, + 652.0, + 1353.0, + 296.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1319.0, + 917.0, + 1319.0, + 917.0, + 1353.0, + 735.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1319.0, + 1014.0, + 1319.0, + 1014.0, + 1353.0, + 1004.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1148.0, + 444.0, + 1148.0, + 444.0, + 1185.0, + 293.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1148.0, + 1241.0, + 1148.0, + 1241.0, + 1185.0, + 475.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 1148.0, + 1327.0, + 1148.0, + 1327.0, + 1185.0, + 1275.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1148.0, + 1403.0, + 1148.0, + 1403.0, + 1185.0, + 1393.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1181.0, + 999.0, + 1181.0, + 999.0, + 1215.0, + 294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1181.0, + 1405.0, + 1181.0, + 1405.0, + 1215.0, + 1031.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 897.0, + 1206.0, + 897.0, + 1248.0, + 294.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1206.0, + 1357.0, + 1206.0, + 1357.0, + 1248.0, + 944.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 412.0, + 902.0, + 412.0, + 902.0, + 451.0, + 296.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 412.0, + 1098.0, + 412.0, + 1098.0, + 451.0, + 982.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 575.0, + 548.0, + 575.0, + 548.0, + 614.0, + 296.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 575.0, + 833.0, + 575.0, + 833.0, + 614.0, + 580.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 735.0, + 426.0, + 735.0, + 426.0, + 780.0, + 294.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 735.0, + 884.0, + 735.0, + 884.0, + 780.0, + 507.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 911.0, + 548.0, + 911.0, + 548.0, + 950.0, + 295.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 911.0, + 833.0, + 911.0, + 833.0, + 950.0, + 580.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1861.0, + 375.0, + 1861.0, + 375.0, + 1901.0, + 297.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1861.0, + 451.0, + 1861.0, + 451.0, + 1901.0, + 401.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1861.0, + 650.0, + 1861.0, + 650.0, + 1901.0, + 478.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 808.0, + 224.0, + 808.0, + 270.0, + 293.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1092.0, + 429.0, + 1092.0, + 429.0, + 1125.0, + 297.0, + 1125.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 227, + 1404, + 227, + 1404, + 356, + 297, + 356 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1399, + 1405, + 1399, + 1405, + 1535, + 297, + 1535 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1568, + 1404, + 1568, + 1404, + 1694, + 297, + 1694 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 701, + 1405, + 701, + 1405, + 795, + 297, + 795 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 814, + 1404, + 814, + 1404, + 903, + 296, + 903 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 611, + 1778, + 1086, + 1778, + 1086, + 1844, + 611, + 1844 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 331, + 1127, + 1364, + 1127, + 1364, + 1223, + 331, + 1223 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 612, + 1980, + 1085, + 1980, + 1085, + 2045, + 612, + 2045 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 1016, + 923, + 1016, + 923, + 1082, + 295, + 1082 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 292, + 1869, + 1407, + 1869, + 1407, + 1945, + 292, + 1945 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 651, + 1323, + 1046, + 1323, + 1046, + 1375, + 651, + 1375 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 345, + 409, + 1352, + 409, + 1352, + 477, + 345, + 477 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 295, + 506, + 1146, + 506, + 1146, + 541, + 295, + 541 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 754, + 945, + 940, + 945, + 940, + 986, + 754, + 986 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 294, + 1251, + 944, + 1251, + 944, + 1290, + 294, + 1290 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 299, + 1706, + 632, + 1706, + 632, + 1739, + 299, + 1739 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 106, + 298, + 106 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 296, + 593, + 1392, + 593, + 1392, + 663, + 296, + 663 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2114, + 836, + 2114 + ], + "score": 0.868 + }, + { + "category_id": 9, + "poly": [ + 1365, + 658, + 1400, + 658, + 1400, + 687, + 1365, + 687 + ], + "score": 0.808 + }, + { + "category_id": 13, + "poly": [ + 930, + 1397, + 1127, + 1397, + 1127, + 1440, + 930, + 1440 + ], + "score": 0.95, + "latex": "r e ^ { r ^ { 1 - \\frac { 2 } { N } } ( 1 - r ) } \\leq 1" + }, + { + "category_id": 13, + "poly": [ + 972, + 1867, + 1124, + 1867, + 1124, + 1917, + 972, + 1917 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } \\leq 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1174, + 1867, + 1373, + 1867, + 1373, + 1917, + 1174, + 1917 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { r _ { i } ( t ) = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\le 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 618, + 1871, + 840, + 1871, + 840, + 1910, + 618, + 1910 + ], + "score": 0.93, + "latex": "c < 2 ( \\sqrt { 2 } - 1 ) < 1" + }, + { + "category_id": 13, + "poly": [ + 559, + 1501, + 667, + 1501, + 667, + 1535, + 559, + 1535 + ], + "score": 0.93, + "latex": "r _ { i } ( t ) \\leq 1" + }, + { + "category_id": 13, + "poly": [ + 365, + 1440, + 474, + 1440, + 474, + 1474, + 365, + 1474 + ], + "score": 0.93, + "latex": "r _ { i } ( \\bar { t } ) > 1" + }, + { + "category_id": 13, + "poly": [ + 488, + 1876, + 574, + 1876, + 574, + 1917, + 488, + 1917 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\eta = \\frac { c } { \\sigma _ { 1 } ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 389, + 507, + 452, + 507, + 452, + 541, + 389, + 541 + ], + "score": 0.93, + "latex": "w _ { i } ( t )" + }, + { + "category_id": 13, + "poly": [ + 515, + 1016, + 697, + 1016, + 697, + 1083, + 515, + 1083 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\eta = c \\bigg ( \\frac { 1 } { \\sigma _ { i } ^ { * } } \\bigg ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 401, + 1254, + 494, + 1254, + 494, + 1295, + 401, + 1295 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { r _ { i } = \\frac { \\sigma _ { i } } { \\sigma _ { i } ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 664, + 1631, + 739, + 1631, + 739, + 1663, + 664, + 1663 + ], + "score": 0.92, + "latex": "t _ { \\alpha } ( \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1079, + 293, + 1180, + 293, + 1180, + 325, + 1079, + 325 + ], + "score": 0.92, + "latex": "w _ { i } ^ { n } = \\sigma _ { i }" + }, + { + "category_id": 14, + "poly": [ + 331, + 1124, + 1365, + 1124, + 1365, + 1222, + 331, + 1222 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\Big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\Big ) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) \\Big ) ^ { N } \\leq \\sigma _ { i } ( t ) e ^ { \\left( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\right) ^ { 1 - \\frac { 2 } { N } } \\Big ( 1 - \\big ( \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } \\big ) \\Big ) }" + }, + { + "category_id": 13, + "poly": [ + 515, + 322, + 578, + 322, + 578, + 355, + 515, + 355 + ], + "score": 0.92, + "latex": "w _ { i } ( t )" + }, + { + "category_id": 13, + "poly": [ + 896, + 1600, + 955, + 1600, + 955, + 1634, + 896, + 1634 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ( t )" + }, + { + "category_id": 14, + "poly": [ + 615, + 1978, + 1083, + 1978, + 1083, + 2045, + 615, + 2045 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { 2 }" + }, + { + "category_id": 14, + "poly": [ + 612, + 1776, + 1085, + 1776, + 1085, + 1843, + 612, + 1843 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { 1 } { 2 } \\eta ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { 2 }" + }, + { + "category_id": 14, + "poly": [ + 345, + 408, + 1352, + 408, + 1352, + 476, + 345, + 476 + ], + "score": 0.92, + "latex": "w _ { i } ( t + 1 ) = w _ { i } ( t ) + \\eta \\frac { 1 } { N } w _ { i } ( t ) ^ { N - 1 } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )" + }, + { + "category_id": 13, + "poly": [ + 1310, + 261, + 1361, + 261, + 1361, + 295, + 1310, + 295 + ], + "score": 0.91, + "latex": "\\sigma ( t )" + }, + { + "category_id": 13, + "poly": [ + 1073, + 809, + 1250, + 809, + 1250, + 875, + 1073, + 875 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\eta \\leq \\left( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\right) ^ { 2 - \\frac { 2 } { N } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 517, + 1708, + 596, + 1708, + 596, + 1735, + 517, + 1735 + ], + "score": 0.91, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 658, + 1411, + 775, + 1411, + 775, + 1441, + 658, + 1441 + ], + "score": 0.91, + "latex": "0 \\leq r \\leq 1" + }, + { + "category_id": 14, + "poly": [ + 651, + 1321, + 1046, + 1321, + 1046, + 1374, + 651, + 1374 + ], + "score": 0.91, + "latex": "r _ { i } ( t + 1 ) \\leq r _ { i } ( t ) e ^ { r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) }" + }, + { + "category_id": 13, + "poly": [ + 797, + 1255, + 828, + 1255, + 828, + 1286, + 797, + 1286 + ], + "score": 0.9, + "latex": "\\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 886, + 231, + 975, + 231, + 975, + 259, + 886, + 259 + ], + "score": 0.9, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 831, + 1473, + 902, + 1473, + 902, + 1500, + 831, + 1500 + ], + "score": 0.9, + "latex": "r = 1" + }, + { + "category_id": 13, + "poly": [ + 678, + 1473, + 747, + 1473, + 747, + 1499, + 678, + 1499 + ], + "score": 0.9, + "latex": "r = 1" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1442, + 1151, + 1442, + 1151, + 1472, + 1028, + 1472 + ], + "score": 0.89, + "latex": "0 \\leq r \\leq 1" + }, + { + "category_id": 13, + "poly": [ + 456, + 1601, + 538, + 1601, + 538, + 1629, + 456, + 1629 + ], + "score": 0.89, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 900, + 829, + 1035, + 829, + 1035, + 865, + 900, + 865 + ], + "score": 0.89, + "latex": "\\sigma _ { i } ( 0 ) < \\sigma _ { i } ^ { * }" + }, + { + "category_id": 14, + "poly": [ + 310, + 594, + 1404, + 594, + 1404, + 662, + 310, + 662 + ], + "score": 0.89, + "latex": "r _ { i } ( t + 1 ) = \\bigg ( \\sqrt [ r ] { \\sigma _ { i } ( t ) } + \\eta \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\sigma _ { i } ( t ) ^ { 1 - \\frac { 1 } { N } } \\bigg ) ^ { N } = \\sigma _ { i } ( t ) \\Big ( 1 + \\eta \\frac { 1 } { N } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) ) \\Big ) ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1503, + 365, + 1503, + 365, + 1530, + 297, + 1530 + ], + "score": 0.88, + "latex": "r < 1" + }, + { + "category_id": 13, + "poly": [ + 743, + 1039, + 808, + 1039, + 808, + 1068, + 743, + 1068 + ], + "score": 0.87, + "latex": "c \\leq 1" + }, + { + "category_id": 13, + "poly": [ + 806, + 1665, + 837, + 1665, + 837, + 1692, + 806, + 1692 + ], + "score": 0.86, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 753, + 945, + 942, + 945, + 942, + 984, + 753, + 984 + ], + "score": 0.82, + "latex": "\\forall t : \\ \\sigma _ { i } ( t ) \\leq \\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1039, + 514, + 1059, + 514, + 1059, + 535, + 1039, + 535 + ], + "score": 0.81, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1308, + 829, + 1393, + 829, + 1393, + 863, + 1308, + 863 + ], + "score": 0.76, + "latex": "N \\geq 2" + }, + { + "category_id": 13, + "poly": [ + 756, + 1505, + 769, + 1505, + 769, + 1529, + 756, + 1529 + ], + "score": 0.72, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 526, + 509, + 552, + 509, + 552, + 535, + 526, + 535 + ], + "score": 0.71, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 528, + 295, + 543, + 295, + 543, + 319, + 528, + 319 + ], + "score": 0.71, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 885.0, + 229.0, + 885.0, + 263.0, + 295.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 263.0, + 976.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1309.0, + 262.0, + 1309.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 295.0, + 1362.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 292.0, + 527.0, + 292.0, + 527.0, + 326.0, + 292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 292.0, + 1078.0, + 292.0, + 1078.0, + 326.0, + 544.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 292.0, + 1405.0, + 292.0, + 1405.0, + 326.0, + 1181.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 514.0, + 320.0, + 514.0, + 358.0, + 293.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 320.0, + 591.0, + 320.0, + 591.0, + 358.0, + 579.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 1386.0, + 657.0, + 1386.0, + 657.0, + 1456.0, + 282.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1386.0, + 929.0, + 1386.0, + 929.0, + 1456.0, + 776.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1386.0, + 1414.0, + 1386.0, + 1414.0, + 1456.0, + 1128.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1435.0, + 364.0, + 1435.0, + 364.0, + 1477.0, + 292.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1435.0, + 1027.0, + 1435.0, + 1027.0, + 1477.0, + 475.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1435.0, + 1406.0, + 1435.0, + 1406.0, + 1477.0, + 1152.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1471.0, + 677.0, + 1471.0, + 677.0, + 1505.0, + 294.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1471.0, + 830.0, + 1471.0, + 830.0, + 1505.0, + 748.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1471.0, + 1407.0, + 1471.0, + 1407.0, + 1505.0, + 903.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1500.0, + 296.0, + 1500.0, + 296.0, + 1535.0, + 292.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1500.0, + 558.0, + 1500.0, + 558.0, + 1535.0, + 366.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1500.0, + 755.0, + 1500.0, + 755.0, + 1535.0, + 668.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1500.0, + 1072.0, + 1500.0, + 1072.0, + 1535.0, + 770.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1504.0, + 1403.0, + 1504.0, + 1403.0, + 1530.0, + 1376.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1570.0, + 1405.0, + 1570.0, + 1405.0, + 1603.0, + 296.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1598.0, + 455.0, + 1598.0, + 455.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1598.0, + 895.0, + 1598.0, + 895.0, + 1636.0, + 539.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1598.0, + 1405.0, + 1598.0, + 1405.0, + 1636.0, + 956.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1631.0, + 663.0, + 1631.0, + 663.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1631.0, + 1404.0, + 1631.0, + 1404.0, + 1664.0, + 740.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1660.0, + 805.0, + 1660.0, + 805.0, + 1697.0, + 294.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1660.0, + 1360.0, + 1660.0, + 1360.0, + 1697.0, + 838.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 702.0, + 1405.0, + 702.0, + 1405.0, + 736.0, + 295.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 731.0, + 1403.0, + 731.0, + 1403.0, + 767.0, + 294.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 757.0, + 393.0, + 757.0, + 393.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 804.0, + 899.0, + 804.0, + 899.0, + 874.0, + 290.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 804.0, + 1072.0, + 804.0, + 1072.0, + 874.0, + 1036.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 804.0, + 1307.0, + 804.0, + 1307.0, + 874.0, + 1251.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 804.0, + 1407.0, + 804.0, + 1407.0, + 874.0, + 1394.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 864.0, + 366.0, + 864.0, + 366.0, + 908.0, + 290.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1019.0, + 514.0, + 1019.0, + 514.0, + 1085.0, + 288.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1032.0, + 742.0, + 1032.0, + 742.0, + 1072.0, + 699.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1032.0, + 926.0, + 1032.0, + 926.0, + 1072.0, + 809.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1857.0, + 487.0, + 1857.0, + 487.0, + 1926.0, + 285.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1857.0, + 617.0, + 1857.0, + 617.0, + 1926.0, + 575.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1857.0, + 971.0, + 1857.0, + 971.0, + 1926.0, + 841.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1857.0, + 1173.0, + 1857.0, + 1173.0, + 1926.0, + 1125.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 1857.0, + 1410.0, + 1857.0, + 1410.0, + 1926.0, + 1374.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1911.0, + 347.0, + 1911.0, + 347.0, + 1944.0, + 289.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 505.0, + 388.0, + 505.0, + 388.0, + 545.0, + 296.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 505.0, + 525.0, + 505.0, + 525.0, + 545.0, + 453.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 505.0, + 1038.0, + 505.0, + 1038.0, + 545.0, + 553.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 505.0, + 1147.0, + 505.0, + 1147.0, + 545.0, + 1060.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1251.0, + 796.0, + 1251.0, + 796.0, + 1288.0, + 495.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1251.0, + 944.0, + 1251.0, + 944.0, + 1288.0, + 829.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1231.5, + 518.0, + 1231.5, + 518.0, + 1303.0, + 289.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1704.0, + 516.0, + 1704.0, + 516.0, + 1741.0, + 294.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1704.0, + 633.0, + 1704.0, + 633.0, + 1741.0, + 597.0, + 1741.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 512, + 1420, + 1187, + 1420, + 1187, + 1651, + 512, + 1651 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 514, + 308, + 1185, + 308, + 1185, + 711, + 514, + 711 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 662, + 1940, + 1034, + 1940, + 1034, + 2046, + 662, + 2046 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 661, + 1220, + 1035, + 1220, + 1035, + 1329, + 661, + 1329 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 581, + 1008, + 1116, + 1008, + 1116, + 1095, + 581, + 1095 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 298, + 733, + 1398, + 733, + 1398, + 807, + 298, + 807 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 301, + 1115, + 1401, + 1115, + 1401, + 1186, + 301, + 1186 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1672, + 1402, + 1672, + 1402, + 1748, + 296, + 1748 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 357, + 842, + 1343, + 842, + 1343, + 922, + 357, + 922 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 362, + 1780, + 1335, + 1780, + 1335, + 1856, + 362, + 1856 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 297, + 1871, + 934, + 1871, + 934, + 1906, + 297, + 1906 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 1342, + 1259, + 1342, + 1259, + 1378, + 298, + 1378 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 296, + 226, + 965, + 226, + 965, + 270, + 296, + 270 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 294, + 937, + 1262, + 937, + 1262, + 975, + 294, + 975 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 817, + 73, + 817, + 106, + 297, + 106 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.87 + }, + { + "category_id": 14, + "poly": [ + 515, + 308, + 1184, + 308, + 1184, + 714, + 515, + 714 + ], + "score": 0.96, + "latex": "\\begin{array} { r l } & { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { \\left( 1 + \\frac { \\varepsilon } { 2 } R _ { i } ( 1 - r _ { i } ( t ) ) \\right) ^ { 2 } } } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\left( 1 - r _ { i } ( t ) \\right) \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + \\left( 1 - r _ { i } ( t ) \\right) \\left( \\boldsymbol { { C R } } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } } \\\\ & { \\qquad \\ge \\frac { 1 } { \\sigma _ { i } ( t ) } \\bigl ( 1 - ( 1 - r _ { i } ( t ) ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) \\bigr ) } \\\\ & { \\qquad = \\frac { 1 } { \\sigma _ { i } ( t ) } - ( \\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } ) ( { c R } _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 513, + 1421, + 1186, + 1421, + 1186, + 1654, + 513, + 1654 + ], + "score": 0.94, + "latex": "\\begin{array} { r l r } { { \\frac { 1 } { \\sigma _ { i } ( t + 1 ) } = \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) ( c R _ { i } + \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } ( 1 - r _ { i } ( t ) ) ) } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } \\frac { 1 } { 1 + ( 1 - r _ { i } ( t ) ) c R _ { i } } } \\\\ & { } & { \\leq \\frac { 1 } { \\sigma _ { i } ( t ) } ( 1 - ( 1 - r _ { i } ( t ) ) c R _ { i } + ( 1 - r _ { i } ( t ) ) ^ { 2 } c ^ { 2 } R _ { i } ^ { 2 } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 665, + 1940, + 1034, + 1940, + 1034, + 2045, + 665, + 2045 + ], + "score": 0.94, + "latex": "t _ { \\alpha } ( \\sigma _ { i } ) \\leq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) }" + }, + { + "category_id": 14, + "poly": [ + 663, + 1219, + 1037, + 1219, + 1037, + 1327, + 663, + 1327 + ], + "score": 0.94, + "latex": "t _ { \\alpha } ( \\sigma _ { i } ) \\geq \\frac { \\log \\big ( \\frac { 1 - \\alpha } { \\alpha ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 ) } \\big ) } { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) }" + }, + { + "category_id": 13, + "poly": [ + 903, + 939, + 1045, + 939, + 1045, + 973, + 903, + 973 + ], + "score": 0.93, + "latex": "\\sigma _ { i } ( t ) = \\alpha \\sigma _ { i } ^ { * }" + }, + { + "category_id": 14, + "poly": [ + 356, + 841, + 1343, + 841, + 1343, + 921, + 356, + 921 + ], + "score": 0.93, + "latex": "\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\geq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) \\geq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 557, + 1114, + 870, + 1114, + 870, + 1158, + 557, + 1158 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) < 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 655, + 939, + 730, + 939, + 730, + 973, + 655, + 973 + ], + "score": 0.93, + "latex": "t _ { \\alpha } ( \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 898, + 227, + 953, + 227, + 953, + 273, + 898, + 273 + ], + "score": 0.93, + "latex": "\\frac { 1 } { \\sigma _ { i } ( t ) }" + }, + { + "category_id": 14, + "poly": [ + 362, + 1780, + 1336, + 1780, + 1336, + 1856, + 362, + 1856 + ], + "score": 0.92, + "latex": "\\frac { 1 } { \\sigma _ { i } ( t ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\leq \\big ( \\frac { 1 } { \\sigma _ { i } ( t - 1 ) } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) \\leq \\big ( \\frac { 1 } { \\sigma _ { 0 } } - \\frac { 1 } { \\sigma _ { i } ^ { * } } \\big ) \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) ^ { t }" + }, + { + "category_id": 14, + "poly": [ + 581, + 1007, + 1116, + 1007, + 1116, + 1095, + 581, + 1095 + ], + "score": 0.92, + "latex": "\\log \\big ( \\frac { 1 - \\alpha } { \\alpha \\big ( \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 0 } } - 1 \\big ) } \\big ) \\geq t _ { \\alpha } \\cdot \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big )" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1674, + 1283, + 1674, + 1283, + 1719, + 1251, + 1719 + ], + "score": 0.91, + "latex": "\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 733, + 1401, + 733, + 1401, + 779, + 1366, + 779 + ], + "score": 0.89, + "latex": "\\textstyle { \\frac { 1 } { \\sigma _ { i } ^ { * } } }" + }, + { + "category_id": 13, + "poly": [ + 827, + 732, + 1096, + 732, + 1096, + 776, + 827, + 776 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\frac { 1 } { 1 + x } \\geq 1 - x , \\forall x \\geq 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 847, + 1880, + 867, + 1880, + 867, + 1900, + 847, + 1900 + ], + "score": 0.78, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1130, + 1128, + 1145, + 1128, + 1145, + 1149, + 1130, + 1149 + ], + "score": 0.77, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 786, + 1673, + 1102, + 1673, + 1102, + 1716, + 786, + 1716 + ], + "score": 0.77, + "latex": "\\begin{array} { r } { \\frac { 1 } { 1 + x } \\leq 1 - x + x ^ { 2 } , \\forall x \\geq 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1675, + 1101, + 1675, + 1101, + 1708, + 1013, + 1708 + ], + "score": 0.35, + "latex": "\\forall x \\geq 0" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 728.0, + 826.0, + 728.0, + 826.0, + 784.0, + 290.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 728.0, + 1365.0, + 728.0, + 1365.0, + 784.0, + 1097.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 770.0, + 588.0, + 770.0, + 588.0, + 808.0, + 294.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1114.0, + 556.0, + 1114.0, + 556.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1114.0, + 1129.0, + 1114.0, + 1129.0, + 1160.0, + 871.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1114.0, + 1407.0, + 1114.0, + 1407.0, + 1160.0, + 1146.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1149.0, + 577.0, + 1149.0, + 577.0, + 1191.0, + 295.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1665.0, + 785.0, + 1665.0, + 785.0, + 1725.0, + 289.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1665.0, + 1250.0, + 1665.0, + 1250.0, + 1725.0, + 1103.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1665.0, + 1411.0, + 1665.0, + 1411.0, + 1725.0, + 1284.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 456.0, + 1710.0, + 456.0, + 1751.0, + 293.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1870.0, + 846.0, + 1870.0, + 846.0, + 1909.0, + 294.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1870.0, + 937.0, + 1870.0, + 937.0, + 1909.0, + 868.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1337.0, + 1267.0, + 1337.0, + 1267.0, + 1385.0, + 291.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 897.0, + 228.0, + 897.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 239.0, + 962.0, + 239.0, + 962.0, + 277.0, + 954.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 932.0, + 654.0, + 932.0, + 654.0, + 980.0, + 293.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 932.0, + 902.0, + 932.0, + 902.0, + 980.0, + 731.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 932.0, + 1264.0, + 932.0, + 1264.0, + 980.0, + 1046.0, + 980.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1120, + 1404, + 1120, + 1404, + 1236, + 297, + 1236 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1615, + 1405, + 1615, + 1405, + 1712, + 298, + 1712 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 882, + 1405, + 882, + 1405, + 960, + 297, + 960 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 678, + 1275, + 1021, + 1275, + 1021, + 1351, + 678, + 1351 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 646, + 1448, + 1050, + 1448, + 1050, + 1524, + 646, + 1524 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 571, + 554, + 1126, + 554, + 1126, + 653, + 571, + 653 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 570, + 1750, + 1127, + 1750, + 1127, + 1848, + 570, + 1848 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 639, + 1000, + 1060, + 1000, + 1060, + 1089, + 639, + 1089 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 580, + 1967, + 1116, + 1967, + 1116, + 2043, + 580, + 2043 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 293, + 1873, + 1096, + 1873, + 1096, + 1948, + 293, + 1948 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 581, + 773, + 1115, + 773, + 1115, + 850, + 581, + 850 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 294, + 680, + 1098, + 680, + 1098, + 755, + 294, + 755 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 420, + 1405, + 420, + 1405, + 517, + 298, + 517 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 1405, + 227, + 1405, + 324, + 296, + 324 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 290, + 1374, + 1304, + 1374, + 1304, + 1409, + 290, + 1409 + ], + "score": 0.934 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 106, + 298, + 106 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 300, + 363, + 587, + 363, + 587, + 396, + 300, + 396 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 299, + 1558, + 591, + 1558, + 591, + 1590, + 299, + 1590 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.873 + }, + { + "category_id": 13, + "poly": [ + 402, + 1873, + 662, + 1873, + 662, + 1948, + 402, + 1948 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { B = \\frac { \\log \\big ( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\big ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 402, + 680, + 662, + 680, + 662, + 755, + 402, + 755 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { A = \\frac { \\log \\big ( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\big ) } { \\log \\big ( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\big ) } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 640, + 999, + 1060, + 999, + 1060, + 1088, + 640, + 1088 + ], + "score": 0.94, + "latex": "\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A } } { \\sigma _ { 0 } ^ { A - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { A }" + }, + { + "category_id": 14, + "poly": [ + 676, + 1274, + 1023, + 1274, + 1023, + 1350, + 676, + 1350 + ], + "score": 0.94, + "latex": "\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { A - 1 } \\geq \\frac { r f } { 1 - f } \\big ( \\frac { 2 - 2 s } { s } \\big ) ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 896, + 1127, + 1007, + 1127, + 1007, + 1165, + 896, + 1165 + ], + "score": 0.93, + "latex": "\\sigma _ { j } ^ { * } \\geq 2 \\sigma _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 646, + 1447, + 1053, + 1447, + 1053, + 1523, + 646, + 1523 + ], + "score": 0.93, + "latex": "\\sigma _ { 0 } \\leq \\frac { 1 } { 2 } \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\bigg ( \\frac { 1 - f } { 2 r f } \\frac { s } { 1 - s } \\bigg ) ^ { \\frac { 1 } { A - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 1118, + 454, + 1192, + 454, + 1192, + 487, + 1118, + 487 + ], + "score": 0.93, + "latex": "t _ { f } ( \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1648, + 1249, + 1648, + 1249, + 1683, + 1176, + 1683 + ], + "score": 0.93, + "latex": "t _ { f } ( \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1316, + 1116, + 1399, + 1116, + 1399, + 1172, + 1316, + 1172 + ], + "score": 0.93, + "latex": "\\frac { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } { r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } }" + }, + { + "category_id": 14, + "poly": [ + 580, + 772, + 1118, + 772, + 1118, + 850, + 580, + 850 + ], + "score": 0.93, + "latex": "\\log \\big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\big ) \\geq A \\log \\big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\big )" + }, + { + "category_id": 13, + "poly": [ + 518, + 484, + 591, + 484, + 591, + 518, + 518, + 518 + ], + "score": 0.92, + "latex": "t _ { s } ( \\sigma _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 573, + 1678, + 646, + 1678, + 646, + 1713, + 573, + 1713 + ], + "score": 0.92, + "latex": "t _ { s } ( \\sigma _ { j } )" + }, + { + "category_id": 14, + "poly": [ + 580, + 1966, + 1118, + 1966, + 1118, + 2043, + 580, + 2043 + ], + "score": 0.92, + "latex": "\\log \\Big ( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\Big ) \\geq B \\log \\Big ( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\Big )" + }, + { + "category_id": 13, + "poly": [ + 1189, + 879, + 1324, + 879, + 1324, + 933, + 1189, + 933 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 569, + 1750, + 1130, + 1750, + 1130, + 1846, + 569, + 1846 + ], + "score": 0.92, + "latex": "\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } - \\frac { c ^ { 2 } } { 4 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } + c ^ { 2 } R _ { j } ^ { 2 } \\right) }" + }, + { + "category_id": 13, + "poly": [ + 728, + 1168, + 841, + 1168, + 841, + 1208, + 728, + 1208 + ], + "score": 0.92, + "latex": "\\textstyle { \\left( { \\frac { 1 } { 2 } } \\sigma _ { j } \\right) } ^ { A - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 231, + 1260, + 231, + 1260, + 264, + 1196, + 264 + ], + "score": 0.92, + "latex": "( s , f )" + }, + { + "category_id": 14, + "poly": [ + 569, + 554, + 1130, + 554, + 1130, + 653, + 569, + 653 + ], + "score": 0.92, + "latex": "\\frac { \\log \\left( \\frac { 1 - f } { f } \\frac { \\sigma _ { 0 } } { \\sigma _ { i } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { i } + c ^ { 2 } R _ { i } ^ { 2 } \\right) } \\leq \\frac { \\log \\left( \\frac { 1 - s } { s } \\frac { \\sigma _ { 0 } } { \\sigma _ { j } ^ { * } - \\sigma _ { 0 } } \\right) } { \\log \\left( 1 - c R _ { j } - \\frac { c ^ { 2 } } { 4 } R _ { j } ^ { 2 } \\right) }" + }, + { + "category_id": 13, + "poly": [ + 518, + 1168, + 668, + 1168, + 668, + 1208, + 518, + 1208 + ], + "score": 0.91, + "latex": "( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { A - 1 }" + }, + { + "category_id": 13, + "poly": [ + 354, + 1169, + 383, + 1169, + 383, + 1208, + 354, + 1208 + ], + "score": 0.9, + "latex": "\\textstyle { \\frac { 1 } { 2 r } }" + }, + { + "category_id": 13, + "poly": [ + 679, + 1382, + 710, + 1382, + 710, + 1406, + 679, + 1406 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 986, + 235, + 1018, + 235, + 1018, + 262, + 986, + 262 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 363.0, + 590.0, + 363.0, + 590.0, + 399.0, + 296.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1557.0, + 593.0, + 1557.0, + 593.0, + 1593.0, + 295.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1113.0, + 895.0, + 1113.0, + 895.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1113.0, + 1315.0, + 1113.0, + 1315.0, + 1181.0, + 1008.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 1113.0, + 1404.0, + 1113.0, + 1404.0, + 1181.0, + 1400.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1157.0, + 353.0, + 1157.0, + 353.0, + 1222.0, + 283.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1157.0, + 517.0, + 1157.0, + 517.0, + 1222.0, + 384.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1157.0, + 727.0, + 1157.0, + 727.0, + 1222.0, + 669.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1157.0, + 1416.0, + 1157.0, + 1416.0, + 1222.0, + 842.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1200.0, + 654.0, + 1200.0, + 654.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1613.0, + 1405.0, + 1613.0, + 1405.0, + 1654.0, + 292.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1647.0, + 1175.0, + 1647.0, + 1175.0, + 1682.0, + 295.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1682.0, + 1250.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1678.0, + 572.0, + 1678.0, + 572.0, + 1713.0, + 296.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1678.0, + 1092.0, + 1678.0, + 1092.0, + 1713.0, + 647.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1188.0, + 872.0, + 1188.0, + 938.0, + 294.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 872.0, + 1405.0, + 872.0, + 1405.0, + 938.0, + 1325.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 927.0, + 570.0, + 927.0, + 570.0, + 961.0, + 296.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1868.0, + 401.0, + 1868.0, + 401.0, + 1932.0, + 293.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1868.0, + 1098.0, + 1868.0, + 1098.0, + 1932.0, + 663.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 695.0, + 401.0, + 695.0, + 401.0, + 735.0, + 293.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 694.0, + 1098.0, + 694.0, + 1098.0, + 739.0, + 663.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 418.0, + 1406.0, + 418.0, + 1406.0, + 459.0, + 292.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 452.0, + 1117.0, + 452.0, + 1117.0, + 487.0, + 295.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 452.0, + 1402.0, + 452.0, + 1402.0, + 487.0, + 1193.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 483.0, + 517.0, + 483.0, + 517.0, + 518.0, + 296.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 483.0, + 1036.0, + 483.0, + 1036.0, + 518.0, + 592.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 985.0, + 230.0, + 985.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 230.0, + 1195.0, + 230.0, + 1195.0, + 264.0, + 1019.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 230.0, + 1403.0, + 230.0, + 1403.0, + 264.0, + 1261.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 289.0, + 329.0, + 289.0, + 329.0, + 326.0, + 291.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1371.0, + 678.0, + 1371.0, + 678.0, + 1413.0, + 292.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1371.0, + 1306.0, + 1371.0, + 1306.0, + 1413.0, + 711.0, + 1413.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1728, + 1405, + 1728, + 1405, + 1903, + 296, + 1903 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 946, + 1405, + 946, + 1405, + 1103, + 296, + 1103 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1513, + 1405, + 1513, + 1405, + 1609, + 297, + 1609 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1319, + 1406, + 1319, + 1406, + 1399, + 297, + 1399 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 224, + 1405, + 224, + 1405, + 303, + 297, + 303 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 639, + 340, + 1060, + 340, + 1060, + 428, + 639, + 428 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 683, + 547, + 1014, + 547, + 1014, + 622, + 683, + 622 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 495, + 1228, + 1197, + 1228, + 1197, + 1300, + 495, + 1300 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 657, + 709, + 1039, + 709, + 1039, + 786, + 657, + 786 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 592, + 1644, + 1101, + 1644, + 1101, + 1711, + 592, + 1711 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 291, + 446, + 1403, + 446, + 1403, + 512, + 291, + 512 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 295, + 1116, + 1399, + 1116, + 1399, + 1196, + 295, + 1196 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 530, + 1429, + 1162, + 1429, + 1162, + 1496, + 530, + 1496 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 2001, + 1051, + 2001, + 1051, + 2037, + 295, + 2037 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 295, + 639, + 891, + 639, + 891, + 674, + 295, + 674 + ], + "score": 0.928 + }, + { + "category_id": 8, + "poly": [ + 763, + 1944, + 936, + 1944, + 936, + 1983, + 763, + 1983 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 298, + 877, + 1138, + 877, + 1138, + 916, + 298, + 916 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 1374, + 803, + 1404, + 803, + 1404, + 834, + 1374, + 834 + ], + "score": 0.849 + }, + { + "category_id": 13, + "poly": [ + 682, + 1353, + 796, + 1353, + 796, + 1403, + 682, + 1403 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { r _ { i } = \\frac { \\sigma _ { i } ( t ) } { \\sigma _ { i } ^ { * } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 683, + 546, + 1018, + 546, + 1018, + 621, + 683, + 621 + ], + "score": 0.94, + "latex": "\\big ( \\frac { \\sigma _ { j } ^ { * } } { \\sigma _ { 0 } } \\big ) ^ { B - 1 } \\geq \\frac { f } { 1 - f } \\big ( \\frac { 1 - s } { s } \\big ) ^ { B }" + }, + { + "category_id": 13, + "poly": [ + 964, + 1352, + 1064, + 1352, + 1064, + 1402, + 964, + 1402 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { R _ { i } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { 1 } ^ { * } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 639, + 339, + 1061, + 339, + 1061, + 428, + 639, + 428 + ], + "score": 0.93, + "latex": "\\frac { ( \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) ^ { B } } { \\sigma _ { 0 } ^ { B - 1 } ( r \\sigma _ { j } ^ { * } - \\sigma _ { 0 } ) } \\geq \\frac { f } { 1 - f } ( \\frac { 1 - s } { s } ) ^ { B }" + }, + { + "category_id": 14, + "poly": [ + 655, + 707, + 1041, + 707, + 1041, + 785, + 655, + 785 + ], + "score": 0.93, + "latex": "\\sigma _ { 0 } \\leq \\frac { s } { 1 - s } \\sigma _ { j } ^ { * } \\Bigl ( \\frac { 1 - f } { f } \\frac { s } { 1 - s } \\Bigr ) ^ { \\frac { 1 } { B - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1147, + 421, + 1147, + 421, + 1201, + 298, + 1201 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } { c \\Big ( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\Big ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 879, + 1822, + 962, + 1822, + 962, + 1875, + 879, + 1875 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathbf { r } = \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1189, + 223, + 1324, + 223, + 1324, + 277, + 1189, + 277 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\frac { \\sigma _ { i } ^ { * } } { \\sigma _ { j } ^ { * } } = r > 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 841, + 2004, + 954, + 2004, + 954, + 2038, + 841, + 2038 + ], + "score": 0.92, + "latex": "R _ { i } = { \\bf r } R _ { j }" + }, + { + "category_id": 13, + "poly": [ + 427, + 1545, + 531, + 1545, + 531, + 1578, + 427, + 1578 + ], + "score": 0.92, + "latex": "1 - r _ { i } ( t )" + }, + { + "category_id": 13, + "poly": [ + 890, + 979, + 941, + 979, + 941, + 1013, + 890, + 1013 + ], + "score": 0.92, + "latex": "\\sigma ( t )" + }, + { + "category_id": 14, + "poly": [ + 593, + 1643, + 1103, + 1643, + 1103, + 1710, + 593, + 1710 + ], + "score": 0.92, + "latex": "r _ { i } ( t + 1 ) \\approx r _ { i } ( t ) \\left( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\right) ^ { N }" + }, + { + "category_id": 14, + "poly": [ + 498, + 1228, + 1198, + 1228, + 1198, + 1301, + 498, + 1301 + ], + "score": 0.91, + "latex": "\\sigma _ { i } ( t + 1 ) = \\sigma _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } \\big ( \\frac { 1 } { \\sigma _ { 1 } ^ { * } } \\big ) ^ { 2 - \\frac { 2 } { N } } \\sigma _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } \\big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) \\big ) \\Big ) ^ { N }" + }, + { + "category_id": 14, + "poly": [ + 533, + 1428, + 1164, + 1428, + 1164, + 1495, + 533, + 1495 + ], + "score": 0.91, + "latex": "r _ { i } ( t + 1 ) = r _ { i } ( t ) \\Big ( 1 + \\frac { c } { N } R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( t ) ^ { 1 - \\frac { 2 } { N } } ( 1 - r _ { i } ( t ) ) \\Big ) ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1072, + 950, + 1160, + 950, + 1160, + 977, + 1072, + 977 + ], + "score": 0.9, + "latex": "N = 2" + }, + { + "category_id": 14, + "poly": [ + 764, + 1943, + 936, + 1943, + 936, + 1982, + 764, + 1982 + ], + "score": 0.89, + "latex": "r _ { j } ( 0 ) = { \\bf r } r _ { i } ( 0 )" + }, + { + "category_id": 13, + "poly": [ + 529, + 1578, + 566, + 1578, + 566, + 1607, + 529, + 1607 + ], + "score": 0.89, + "latex": "i , j" + }, + { + "category_id": 13, + "poly": [ + 1320, + 1324, + 1351, + 1324, + 1351, + 1355, + 1320, + 1355 + ], + "score": 0.89, + "latex": "\\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1082, + 1516, + 1162, + 1516, + 1162, + 1544, + 1082, + 1544 + ], + "score": 0.89, + "latex": "r \\ll 1" + }, + { + "category_id": 13, + "poly": [ + 789, + 1735, + 815, + 1735, + 815, + 1761, + 789, + 1761 + ], + "score": 0.85, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 867, + 1736, + 894, + 1736, + 894, + 1764, + 867, + 1764 + ], + "score": 0.85, + "latex": "r _ { j }" + }, + { + "category_id": 13, + "poly": [ + 436, + 482, + 467, + 482, + 467, + 509, + 436, + 509 + ], + "score": 0.85, + "latex": "\\sigma _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1121, + 1406, + 1121, + 1406, + 1150, + 1351, + 1150 + ], + "score": 0.82, + "latex": "\\eta =" + }, + { + "category_id": 13, + "poly": [ + 757, + 1011, + 785, + 1011, + 785, + 1037, + 757, + 1037 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 443, + 1874, + 467, + 1874, + 467, + 1900, + 443, + 1900 + ], + "score": 0.81, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 951, + 454, + 968, + 454, + 968, + 475, + 951, + 475 + ], + "score": 0.77, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1374, + 1836, + 1400, + 1836, + 1400, + 1864, + 1374, + 1864 + ], + "score": 0.76, + "latex": "r _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1108, + 1125, + 1127, + 1125, + 1127, + 1145, + 1108, + 1145 + ], + "score": 0.76, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 600, + 1835, + 615, + 1835, + 615, + 1856, + 600, + 1856 + ], + "score": 0.75, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1106, + 882, + 1138, + 882, + 1138, + 911, + 1106, + 911 + ], + "score": 0.74, + "latex": "N" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 876.0, + 1105.0, + 876.0, + 1105.0, + 920.0, + 296.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1380.0, + 809.0, + 1402.0, + 809.0, + 1402.0, + 832.0, + 1380.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1727.0, + 788.0, + 1727.0, + 788.0, + 1765.0, + 295.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1727.0, + 866.0, + 1727.0, + 866.0, + 1765.0, + 816.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1765.0, + 895.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1792.0, + 295.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1793.0, + 1403.0, + 1793.0, + 1403.0, + 1824.0, + 298.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1811.0, + 599.0, + 1811.0, + 599.0, + 1882.0, + 295.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1811.0, + 878.0, + 1811.0, + 878.0, + 1882.0, + 616.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1811.0, + 1373.0, + 1811.0, + 1373.0, + 1882.0, + 963.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1868.0, + 442.0, + 1868.0, + 442.0, + 1904.0, + 294.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1868.0, + 478.0, + 1868.0, + 478.0, + 1904.0, + 468.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 948.0, + 1071.0, + 948.0, + 1071.0, + 982.0, + 296.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 948.0, + 1405.0, + 948.0, + 1405.0, + 982.0, + 1161.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 979.0, + 889.0, + 979.0, + 889.0, + 1012.0, + 295.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 979.0, + 1403.0, + 979.0, + 1403.0, + 1012.0, + 942.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1010.0, + 756.0, + 1010.0, + 756.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1010.0, + 1402.0, + 1010.0, + 1402.0, + 1043.0, + 786.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1037.0, + 1405.0, + 1037.0, + 1405.0, + 1077.0, + 293.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1066.0, + 452.0, + 1066.0, + 452.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1511.0, + 1081.0, + 1511.0, + 1081.0, + 1550.0, + 294.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1511.0, + 1406.0, + 1511.0, + 1406.0, + 1550.0, + 1163.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1544.0, + 426.0, + 1544.0, + 426.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1544.0, + 1406.0, + 1544.0, + 1406.0, + 1582.0, + 532.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1577.0, + 528.0, + 1577.0, + 528.0, + 1611.0, + 295.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1577.0, + 655.0, + 1577.0, + 655.0, + 1611.0, + 567.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1317.0, + 1319.0, + 1317.0, + 1319.0, + 1358.0, + 294.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1317.0, + 1404.0, + 1317.0, + 1404.0, + 1358.0, + 1352.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1360.0, + 681.0, + 1360.0, + 681.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1351.0, + 963.0, + 1351.0, + 963.0, + 1405.0, + 797.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1351.0, + 1076.0, + 1351.0, + 1076.0, + 1405.0, + 1065.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 216.0, + 1188.0, + 216.0, + 1188.0, + 286.0, + 294.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 216.0, + 1406.0, + 216.0, + 1406.0, + 286.0, + 1325.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 270.0, + 570.0, + 270.0, + 570.0, + 305.0, + 295.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 448.0, + 950.0, + 448.0, + 950.0, + 480.0, + 298.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 448.0, + 1404.0, + 448.0, + 1404.0, + 480.0, + 969.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 476.0, + 435.0, + 476.0, + 435.0, + 515.0, + 293.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 476.0, + 1014.0, + 476.0, + 1014.0, + 515.0, + 468.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1111.0, + 1107.0, + 1111.0, + 1107.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1111.0, + 1350.0, + 1111.0, + 1350.0, + 1155.0, + 1128.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1142.0, + 433.0, + 1142.0, + 433.0, + 1187.0, + 422.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1140.0, + 418.0, + 1140.0, + 418.0, + 1207.5, + 283.0, + 1207.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1996.0, + 840.0, + 1996.0, + 840.0, + 2042.0, + 293.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1996.0, + 1053.0, + 1996.0, + 1053.0, + 2042.0, + 955.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 638.0, + 892.0, + 638.0, + 892.0, + 678.0, + 295.0, + 678.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1025, + 1405, + 1025, + 1405, + 1182, + 297, + 1182 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1195, + 1404, + 1195, + 1404, + 1290, + 298, + 1290 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1644, + 1404, + 1644, + 1404, + 1799, + 297, + 1799 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 738, + 1406, + 738, + 1406, + 842, + 298, + 842 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 436, + 1408, + 436, + 1408, + 561, + 297, + 561 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 552, + 259, + 1144, + 259, + 1144, + 422, + 552, + 422 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 603, + 605, + 1090, + 605, + 1090, + 722, + 603, + 722 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 297, + 1813, + 1405, + 1813, + 1405, + 1907, + 297, + 1907 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 666, + 1541, + 1035, + 1541, + 1035, + 1629, + 666, + 1629 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 294, + 1304, + 1402, + 1304, + 1402, + 1368, + 294, + 1368 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 297, + 854, + 1402, + 854, + 1402, + 920, + 297, + 920 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 764, + 1409, + 936, + 1409, + 936, + 1448, + 764, + 1448 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 296, + 1461, + 1092, + 1461, + 1092, + 1497, + 296, + 1497 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 298, + 958, + 1011, + 958, + 1011, + 997, + 298, + 997 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 105, + 298, + 105 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 296, + 1918, + 1404, + 1918, + 1404, + 2034, + 296, + 2034 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 864, + 2087, + 864, + 2112, + 836, + 2112 + ], + "score": 0.863 + }, + { + "category_id": 14, + "poly": [ + 554, + 258, + 1145, + 258, + 1145, + 427, + 554, + 427 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( 1 + \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\\\ { r _ { j } ( 1 ) \\approx r _ { i } ( 0 ) \\Big ( \\sqrt [ N ] { \\mathbf { r } } + \\big ( \\frac { 1 } { \\mathbf { r } } \\big ) ^ { \\frac { N - 1 } { N } } \\frac { c R _ { i } ^ { 2 - \\frac { 2 } { N } } } { N } r _ { i } ( 0 ) ^ { 1 - \\frac { 2 } { N } } \\Big ) ^ { N } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 606, + 602, + 1093, + 602, + 1093, + 725, + 606, + 725 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { r _ { i } ( 1 ) \\approx r _ { i } ( 0 ) + c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\\\ & { r _ { j } ( 1 ) \\approx { \\bf r } r _ { i } ( 0 ) + \\big ( \\frac { 1 } { { \\bf r } } \\big ) ^ { \\frac { N - 1 } { N } } c R _ { i } ^ { 2 - \\frac { 2 } { N } } r _ { i } ( 0 ) ^ { 2 - \\frac { 2 } { N } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 398, + 771, + 461, + 771, + 461, + 806, + 398, + 806 + ], + "score": 0.93, + "latex": "r ^ { \\frac { N - 1 } { N } }" + }, + { + "category_id": 13, + "poly": [ + 608, + 1705, + 807, + 1705, + 807, + 1740, + 608, + 1740 + ], + "score": 0.93, + "latex": "\\| D \\alpha - x \\| _ { 2 } ^ { 2 } = 0 ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1674, + 421, + 1674, + 421, + 1704, + 298, + 1704 + ], + "score": 0.92, + "latex": "D \\in \\mathbb { R } ^ { d \\times n }" + }, + { + "category_id": 13, + "poly": [ + 690, + 1461, + 848, + 1461, + 848, + 1496, + 690, + 1496 + ], + "score": 0.92, + "latex": "w _ { + } , w _ { - } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1674, + 686, + 1674, + 686, + 1704, + 601, + 1704 + ], + "score": 0.91, + "latex": "x \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 762, + 1409, + 938, + 1409, + 938, + 1447, + 762, + 1447 + ], + "score": 0.91, + "latex": "f _ { \\sigma } ( x ) = \\langle \\sigma , x \\rangle" + }, + { + "category_id": 14, + "poly": [ + 666, + 1539, + 1034, + 1539, + 1034, + 1630, + 666, + 1630 + ], + "score": 0.91, + "latex": "\\begin{array} { c } { \\sigma _ { i } = w _ { + , i } ^ { N } - w _ { - , i } ^ { N } } \\\\ { \\qquad \\forall i , \\ w _ { + , i } ( 0 ) = w _ { - , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1677, + 1234, + 1677, + 1234, + 1704, + 1137, + 1704 + ], + "score": 0.9, + "latex": "D \\alpha \\approx x" + }, + { + "category_id": 13, + "poly": [ + 338, + 1708, + 366, + 1708, + 366, + 1737, + 338, + 1737 + ], + "score": 0.87, + "latex": "\\ell _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 871, + 744, + 898, + 744, + 898, + 774, + 871, + 774 + ], + "score": 0.84, + "latex": "r _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 861, + 1313, + 861, + 1313, + 887, + 1288, + 887 + ], + "score": 0.84, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 622, + 893, + 648, + 893, + 648, + 921, + 622, + 921 + ], + "score": 0.84, + "latex": "r _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 745, + 1238, + 745, + 1238, + 770, + 1213, + 770 + ], + "score": 0.84, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 920, + 814, + 946, + 814, + 946, + 843, + 920, + 843 + ], + "score": 0.84, + "latex": "r _ { j }" + }, + { + "category_id": 13, + "poly": [ + 541, + 813, + 565, + 813, + 565, + 839, + 541, + 839 + ], + "score": 0.83, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 992, + 1983, + 1017, + 1983, + 1017, + 2004, + 992, + 2004 + ], + "score": 0.81, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1001, + 1681, + 1023, + 1681, + 1023, + 1703, + 1001, + 1703 + ], + "score": 0.8, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 466, + 1711, + 488, + 1711, + 488, + 1735, + 466, + 1735 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 599, + 1470, + 619, + 1470, + 619, + 1491, + 599, + 1491 + ], + "score": 0.79, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 591, + 506, + 606, + 506, + 606, + 526, + 591, + 526 + ], + "score": 0.74, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 466, + 1881, + 482, + 1881, + 482, + 1903, + 466, + 1903 + ], + "score": 0.73, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 462, + 537, + 477, + 537, + 477, + 556, + 462, + 556 + ], + "score": 0.72, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1852, + 1084, + 1852, + 1084, + 1872, + 1069, + 1872 + ], + "score": 0.69, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 1013.0, + 957.0, + 1013.0, + 1001.0, + 295.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1954.0, + 329.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1945.0, + 902.0, + 1945.0, + 902.0, + 1981.0, + 295.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1971.0, + 991.0, + 1971.0, + 991.0, + 2011.0, + 328.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2011.0, + 1018.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 453.0, + 2001.0, + 453.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1027.0, + 1405.0, + 1027.0, + 1405.0, + 1061.0, + 296.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1058.0, + 1402.0, + 1058.0, + 1402.0, + 1091.0, + 295.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1089.0, + 1405.0, + 1089.0, + 1405.0, + 1122.0, + 294.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1118.0, + 1405.0, + 1118.0, + 1405.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1149.0, + 1359.0, + 1149.0, + 1359.0, + 1186.0, + 295.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1195.0, + 1405.0, + 1195.0, + 1405.0, + 1233.0, + 294.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1228.0, + 1408.0, + 1228.0, + 1408.0, + 1262.0, + 296.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1258.0, + 944.0, + 1258.0, + 944.0, + 1292.0, + 294.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1640.0, + 1404.0, + 1640.0, + 1404.0, + 1684.0, + 295.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 297.0, + 1668.0, + 297.0, + 1714.0, + 294.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1668.0, + 600.0, + 1668.0, + 600.0, + 1714.0, + 422.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1668.0, + 1000.0, + 1668.0, + 1000.0, + 1714.0, + 687.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1668.0, + 1136.0, + 1668.0, + 1136.0, + 1714.0, + 1024.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1668.0, + 1405.0, + 1668.0, + 1405.0, + 1714.0, + 1235.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1700.0, + 337.0, + 1700.0, + 337.0, + 1745.0, + 292.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1700.0, + 465.0, + 1700.0, + 465.0, + 1745.0, + 367.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1700.0, + 607.0, + 1700.0, + 607.0, + 1745.0, + 489.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1745.0, + 808.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1735.0, + 1405.0, + 1735.0, + 1405.0, + 1772.0, + 294.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1762.0, + 422.0, + 1762.0, + 422.0, + 1807.0, + 293.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 736.0, + 870.0, + 736.0, + 870.0, + 775.0, + 295.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 736.0, + 1212.0, + 736.0, + 1212.0, + 775.0, + 899.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 736.0, + 1404.0, + 736.0, + 1404.0, + 775.0, + 1239.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 772.0, + 397.0, + 772.0, + 397.0, + 816.0, + 292.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 772.0, + 1406.0, + 772.0, + 1406.0, + 816.0, + 462.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 805.0, + 540.0, + 805.0, + 540.0, + 849.0, + 293.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 805.0, + 919.0, + 805.0, + 919.0, + 849.0, + 566.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 805.0, + 959.0, + 805.0, + 959.0, + 849.0, + 947.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 435.0, + 1406.0, + 435.0, + 1406.0, + 470.0, + 295.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 463.0, + 1406.0, + 463.0, + 1406.0, + 504.0, + 291.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 500.0, + 590.0, + 500.0, + 590.0, + 532.0, + 294.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 500.0, + 1405.0, + 500.0, + 1405.0, + 532.0, + 607.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 529.0, + 461.0, + 529.0, + 461.0, + 559.0, + 296.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 529.0, + 489.0, + 529.0, + 489.0, + 559.0, + 478.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1814.0, + 1401.0, + 1814.0, + 1401.0, + 1848.0, + 295.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1843.0, + 1068.0, + 1843.0, + 1068.0, + 1881.0, + 294.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1843.0, + 1406.0, + 1843.0, + 1406.0, + 1881.0, + 1085.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1876.0, + 465.0, + 1876.0, + 465.0, + 1905.0, + 293.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1876.0, + 678.0, + 1876.0, + 678.0, + 1905.0, + 483.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1301.0, + 1405.0, + 1301.0, + 1405.0, + 1339.0, + 294.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1336.0, + 1193.0, + 1336.0, + 1193.0, + 1368.0, + 297.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 852.0, + 1287.0, + 852.0, + 1287.0, + 890.0, + 295.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 852.0, + 1406.0, + 852.0, + 1406.0, + 890.0, + 1314.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 621.0, + 882.0, + 621.0, + 925.0, + 294.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 882.0, + 659.0, + 882.0, + 659.0, + 925.0, + 649.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1453.0, + 598.0, + 1453.0, + 598.0, + 1505.0, + 291.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1453.0, + 689.0, + 1453.0, + 689.0, + 1505.0, + 620.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1453.0, + 1097.0, + 1453.0, + 1097.0, + 1505.0, + 849.0, + 1505.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1341, + 1405, + 1341, + 1405, + 1587, + 297, + 1587 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1143, + 1404, + 1143, + 1404, + 1325, + 297, + 1325 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 541, + 298, + 1104, + 298, + 1104, + 846, + 541, + 846 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 296, + 893, + 1404, + 893, + 1404, + 1079, + 296, + 1079 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 675, + 1901, + 1026, + 1901, + 1026, + 1946, + 675, + 1946 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1403, + 1971, + 1403, + 2033, + 299, + 2033 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 298, + 1765, + 1402, + 1765, + 1402, + 1859, + 298, + 1859 + ], + "score": 0.939 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 299, + 1706, + 637, + 1706, + 637, + 1738, + 299, + 1738 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.862 + }, + { + "category_id": 0, + "poly": [ + 298, + 1634, + 1026, + 1634, + 1026, + 1671, + 298, + 1671 + ], + "score": 0.842 + }, + { + "category_id": 14, + "poly": [ + 674, + 1900, + 1024, + 1900, + 1024, + 1945, + 674, + 1945 + ], + "score": 0.91, + "latex": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )" + }, + { + "category_id": 13, + "poly": [ + 457, + 1976, + 487, + 1976, + 487, + 2006, + 457, + 2006 + ], + "score": 0.88, + "latex": "\\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 905, + 1975, + 948, + 1975, + 948, + 2000, + 905, + 2000 + ], + "score": 0.85, + "latex": "W ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 376, + 1977, + 403, + 1977, + 403, + 2003, + 376, + 2003 + ], + "score": 0.82, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 932, + 1828, + 959, + 1828, + 959, + 1854, + 932, + 1854 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 820, + 1975, + 851, + 1975, + 851, + 2000, + 820, + 2000 + ], + "score": 0.77, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 826, + 962, + 844, + 962, + 844, + 984, + 826, + 984 + ], + "score": 0.62, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 509, + 1151, + 525, + 1151, + 525, + 1171, + 509, + 1171 + ], + "score": 0.6, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 525, + 993, + 542, + 993, + 542, + 1015, + 525, + 1015 + ], + "score": 0.6, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 303.0, + 624.0, + 303.0, + 624.0, + 344.0, + 570.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 325.0, + 589.0, + 325.0, + 589.0, + 761.0, + 529.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 365.0, + 626.0, + 365.0, + 626.0, + 411.0, + 568.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 432.0, + 623.0, + 432.0, + 623.0, + 470.0, + 569.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 476.0, + 825.0, + 476.0, + 825.0, + 504.0, + 795.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 496.0, + 624.0, + 496.0, + 624.0, + 538.0, + 569.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 559.0, + 624.0, + 559.0, + 624.0, + 600.0, + 567.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 622.0, + 624.0, + 622.0, + 624.0, + 664.0, + 569.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 685.0, + 623.0, + 685.0, + 623.0, + 724.0, + 571.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 747.0, + 624.0, + 747.0, + 624.0, + 790.0, + 570.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 785.0, + 631.0, + 785.0, + 631.0, + 817.0, + 609.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 791.0, + 713.0, + 791.0, + 713.0, + 818.0, + 690.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 782.0, + 804.0, + 782.0, + 804.0, + 821.0, + 765.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 779.0, + 887.0, + 779.0, + 887.0, + 822.0, + 846.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 781.0, + 970.0, + 781.0, + 970.0, + 822.0, + 928.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 783.0, + 1049.0, + 783.0, + 1049.0, + 822.0, + 1011.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 803.0, + 1002.0, + 803.0, + 1002.0, + 853.0, + 717.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 638.0, + 992.0, + 638.0, + 992.0, + 660.5, + 907.0, + 660.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 893.0, + 1407.0, + 893.0, + 1407.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 925.0, + 1406.0, + 925.0, + 1406.0, + 962.0, + 295.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 951.0, + 825.0, + 951.0, + 825.0, + 995.0, + 291.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 951.0, + 1405.0, + 951.0, + 1405.0, + 995.0, + 845.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 987.0, + 524.0, + 987.0, + 524.0, + 1020.0, + 293.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1020.0, + 543.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1019.0, + 1404.0, + 1019.0, + 1404.0, + 1051.0, + 295.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1047.0, + 1059.0, + 1047.0, + 1059.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1707.0, + 640.0, + 1707.0, + 640.0, + 1740.0, + 296.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1633.0, + 1029.0, + 1633.0, + 1029.0, + 1675.0, + 293.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1343.0, + 1405.0, + 1343.0, + 1405.0, + 1376.0, + 295.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1373.0, + 1405.0, + 1373.0, + 1405.0, + 1406.0, + 295.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1404.0, + 1402.0, + 1404.0, + 1402.0, + 1438.0, + 295.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1464.0, + 1405.0, + 1464.0, + 1405.0, + 1500.0, + 292.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1492.0, + 1405.0, + 1492.0, + 1405.0, + 1530.0, + 291.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1522.0, + 1406.0, + 1522.0, + 1406.0, + 1563.0, + 291.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1555.0, + 1161.0, + 1555.0, + 1161.0, + 1589.0, + 294.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1145.0, + 508.0, + 1145.0, + 508.0, + 1176.0, + 296.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1145.0, + 1404.0, + 1145.0, + 1404.0, + 1176.0, + 526.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1171.0, + 1405.0, + 1171.0, + 1405.0, + 1211.0, + 292.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1205.0, + 1405.0, + 1205.0, + 1405.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1235.0, + 1401.0, + 1235.0, + 1401.0, + 1267.0, + 295.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1263.0, + 1406.0, + 1263.0, + 1406.0, + 1303.0, + 291.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1295.0, + 353.0, + 1295.0, + 353.0, + 1330.0, + 292.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 375.0, + 1968.0, + 375.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1968.0, + 456.0, + 1968.0, + 456.0, + 2010.0, + 404.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1968.0, + 819.0, + 1968.0, + 819.0, + 2010.0, + 488.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1968.0, + 904.0, + 1968.0, + 904.0, + 2010.0, + 852.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2010.0, + 949.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 489.0, + 2003.0, + 489.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1765.0, + 1406.0, + 1765.0, + 1406.0, + 1802.0, + 294.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1796.0, + 1406.0, + 1796.0, + 1406.0, + 1830.0, + 296.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1825.0, + 931.0, + 1825.0, + 931.0, + 1863.0, + 293.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1825.0, + 974.0, + 1825.0, + 974.0, + 1863.0, + 960.0, + 1863.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1511, + 1404, + 1511, + 1404, + 1698, + 297, + 1698 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1849, + 1405, + 1849, + 1405, + 2035, + 297, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 1104, + 1405, + 1104, + 1405, + 1240, + 296, + 1240 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 471, + 1405, + 471, + 1405, + 597, + 297, + 597 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1711, + 1403, + 1711, + 1403, + 1836, + 298, + 1836 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 868, + 1403, + 868, + 1403, + 963, + 297, + 963 + ], + "score": 0.977 + }, + { + "category_id": 8, + "poly": [ + 688, + 999, + 1009, + 999, + 1009, + 1083, + 688, + 1083 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 629, + 1276, + 1070, + 1276, + 1070, + 1366, + 629, + 1366 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 647, + 787, + 1049, + 787, + 1049, + 850, + 647, + 850 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 294, + 228, + 1403, + 228, + 1403, + 296, + 294, + 296 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 577, + 633, + 1123, + 633, + 1123, + 698, + 577, + 698 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 636, + 331, + 1061, + 331, + 1061, + 395, + 636, + 395 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 290, + 715, + 1313, + 715, + 1313, + 749, + 290, + 749 + ], + "score": 0.937 + }, + { + "category_id": 0, + "poly": [ + 298, + 1455, + 685, + 1455, + 685, + 1488, + 298, + 1488 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 299, + 417, + 704, + 417, + 704, + 459, + 299, + 459 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1387, + 1403, + 1387, + 1403, + 1417, + 1374, + 1417 + ], + "score": 0.825 + }, + { + "category_id": 13, + "poly": [ + 499, + 416, + 696, + 416, + 696, + 461, + 499, + 461 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { W _ { j : k } = \\prod _ { i = j } ^ { k } W _ { i } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 688, + 998, + 1011, + 998, + 1011, + 1085, + 688, + 1085 + ], + "score": 0.94, + "latex": "\\dot { \\sigma } _ { n , i } = \\frac { 1 } { N } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ) \\prod _ { j \\neq n } \\sigma _ { j , i }" + }, + { + "category_id": 14, + "poly": [ + 630, + 1272, + 1070, + 1272, + 1070, + 1368, + 630, + 1368 + ], + "score": 0.94, + "latex": "\\dot { \\sigma } _ { i } = \\sum _ { n = 1 } ^ { N } \\sigma _ { - n , i } \\dot { \\sigma } _ { n , i } = \\sigma _ { i } ^ { 2 - \\frac { 2 } { N } } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )" + }, + { + "category_id": 14, + "poly": [ + 635, + 328, + 1062, + 328, + 1062, + 395, + 635, + 395 + ], + "score": 0.94, + "latex": "\\dot { W _ { n } } = { \\frac { 1 } { N } } W _ { 1 : n - 1 } ^ { T } ( W ^ { * } - W ) W _ { n + 1 : N } ^ { T }" + }, + { + "category_id": 14, + "poly": [ + 647, + 783, + 1051, + 783, + 1051, + 851, + 647, + 851 + ], + "score": 0.94, + "latex": "\\dot { D _ { n } } = { \\frac { 1 } { N } } D _ { 1 : n - 1 } ( D ^ { * } - D ) D _ { n + 1 : N }" + }, + { + "category_id": 13, + "poly": [ + 803, + 1203, + 1023, + 1203, + 1023, + 1243, + 803, + 1243 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\sigma _ { - n , i } = \\prod _ { k \\neq n } \\sigma _ { k , i } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 554, + 531, + 852, + 531, + 852, + 569, + 554, + 569 + ], + "score": 0.93, + "latex": "{ \\cal D } _ { n } \\overline { { } } = U W _ { n } U ^ { T } = \\sqrt [ N ] { \\sigma _ { 0 } } I" + }, + { + "category_id": 14, + "poly": [ + 577, + 631, + 1120, + 631, + 1120, + 698, + 577, + 698 + ], + "score": 0.93, + "latex": "\\dot { W _ { n } } = \\frac { 1 } { N } U ^ { T } D _ { 1 : n - 1 } U ( W ^ { * } - W ) U ^ { T } D _ { n + 1 : N } U" + }, + { + "category_id": 13, + "poly": [ + 298, + 502, + 403, + 502, + 403, + 532, + 298, + 532 + ], + "score": 0.9, + "latex": "U W ^ { * } U ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 1278, + 901, + 1322, + 901, + 1322, + 931, + 1278, + 931 + ], + "score": 0.9, + "latex": "W _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1329, + 473, + 1404, + 473, + 1404, + 503, + 1329, + 503 + ], + "score": 0.9, + "latex": "D ^ { * } =" + }, + { + "category_id": 13, + "poly": [ + 1309, + 535, + 1352, + 535, + 1352, + 564, + 1309, + 564 + ], + "score": 0.9, + "latex": "W _ { n }" + }, + { + "category_id": 13, + "poly": [ + 506, + 262, + 660, + 262, + 660, + 295, + 506, + 295 + ], + "score": 0.9, + "latex": "W _ { n } , n \\in [ N ]" + }, + { + "category_id": 13, + "poly": [ + 1226, + 504, + 1271, + 504, + 1271, + 534, + 1226, + 534 + ], + "score": 0.9, + "latex": "W _ { n }" + }, + { + "category_id": 13, + "poly": [ + 535, + 1205, + 579, + 1205, + 579, + 1235, + 535, + 1235 + ], + "score": 0.89, + "latex": "W _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1261, + 718, + 1304, + 718, + 1304, + 748, + 1261, + 748 + ], + "score": 0.89, + "latex": "W _ { n }" + }, + { + "category_id": 13, + "poly": [ + 513, + 473, + 558, + 473, + 558, + 501, + 513, + 501 + ], + "score": 0.89, + "latex": "W ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 917, + 904, + 965, + 904, + 965, + 934, + 917, + 934 + ], + "score": 0.89, + "latex": "\\sigma _ { n , i }" + }, + { + "category_id": 13, + "poly": [ + 485, + 505, + 524, + 505, + 524, + 531, + 485, + 531 + ], + "score": 0.88, + "latex": "D ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 826, + 1141, + 1290, + 1141, + 1290, + 1177, + 826, + 1177 + ], + "score": 0.88, + "latex": "\\forall n , m , i : \\sigma _ { n , i } ( t ) = \\sigma _ { m , i } ( t ) = \\sqrt [ N ] { \\sigma _ { i } ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 721, + 1105, + 1163, + 1105, + 1163, + 1141, + 721, + 1141 + ], + "score": 0.87, + "latex": "\\forall n , m , i : \\sigma _ { n , i } ( 0 ) = \\sigma _ { m , i } ( 0 ) = \\sqrt [ N ] { \\sigma _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1373, + 904, + 1400, + 904, + 1400, + 931, + 1373, + 931 + ], + "score": 0.83, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1181, + 474, + 1206, + 474, + 1206, + 500, + 1181, + 500 + ], + "score": 0.81, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 608, + 932, + 639, + 932, + 639, + 958, + 608, + 958 + ], + "score": 0.8, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 1014, + 505, + 1039, + 505, + 1039, + 531, + 1014, + 531 + ], + "score": 0.78, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 991, + 1176, + 1022, + 1176, + 1022, + 1201, + 991, + 1201 + ], + "score": 0.78, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 1039, + 903, + 1055, + 903, + 1055, + 928, + 1039, + 928 + ], + "score": 0.68, + "latex": "i ^ { \\because }" + }, + { + "category_id": 13, + "poly": [ + 368, + 933, + 385, + 933, + 385, + 958, + 368, + 958 + ], + "score": 0.59, + "latex": "\\overrightarrow { \\imath } ^ { \\prime }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1456.0, + 687.0, + 1456.0, + 687.0, + 1489.0, + 296.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1391.0, + 1403.0, + 1391.0, + 1403.0, + 1419.0, + 1377.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1512.0, + 1402.0, + 1512.0, + 1402.0, + 1548.0, + 295.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1544.0, + 1404.0, + 1544.0, + 1404.0, + 1576.0, + 296.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1606.0, + 296.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1640.0, + 295.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1634.0, + 1405.0, + 1634.0, + 1405.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1665.0, + 779.0, + 1665.0, + 779.0, + 1701.0, + 295.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1851.0, + 1403.0, + 1851.0, + 1403.0, + 1883.0, + 297.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1406.0, + 1883.0, + 1406.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1104.0, + 720.0, + 1104.0, + 720.0, + 1142.0, + 295.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1104.0, + 1405.0, + 1104.0, + 1405.0, + 1142.0, + 1164.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1140.0, + 825.0, + 1140.0, + 825.0, + 1182.0, + 291.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1140.0, + 1405.0, + 1140.0, + 1405.0, + 1182.0, + 1291.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1173.0, + 990.0, + 1173.0, + 990.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1173.0, + 1403.0, + 1173.0, + 1403.0, + 1208.0, + 1023.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1203.0, + 534.0, + 1203.0, + 534.0, + 1245.0, + 294.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1203.0, + 802.0, + 1203.0, + 802.0, + 1245.0, + 580.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1203.0, + 1310.0, + 1203.0, + 1310.0, + 1245.0, + 1024.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 470.0, + 512.0, + 470.0, + 512.0, + 507.0, + 294.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 470.0, + 1180.0, + 470.0, + 1180.0, + 507.0, + 559.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 470.0, + 1328.0, + 470.0, + 1328.0, + 507.0, + 1207.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 499.0, + 297.0, + 499.0, + 297.0, + 539.0, + 292.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 499.0, + 484.0, + 499.0, + 484.0, + 539.0, + 404.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 499.0, + 1013.0, + 499.0, + 1013.0, + 539.0, + 525.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 499.0, + 1225.0, + 499.0, + 1225.0, + 539.0, + 1040.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 499.0, + 1407.0, + 499.0, + 1407.0, + 539.0, + 1272.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 529.0, + 553.0, + 529.0, + 553.0, + 571.0, + 292.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 529.0, + 1308.0, + 529.0, + 1308.0, + 571.0, + 853.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 529.0, + 1406.0, + 529.0, + 1406.0, + 571.0, + 1353.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 565.0, + 349.0, + 565.0, + 349.0, + 603.0, + 291.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 1404.0, + 1744.0, + 1404.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1403.0, + 1773.0, + 1403.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1805.0, + 815.0, + 1805.0, + 815.0, + 1838.0, + 296.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 864.0, + 1404.0, + 864.0, + 1404.0, + 907.0, + 292.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 900.0, + 916.0, + 900.0, + 916.0, + 937.0, + 292.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 900.0, + 1038.0, + 900.0, + 1038.0, + 937.0, + 966.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 900.0, + 1277.0, + 900.0, + 1277.0, + 937.0, + 1056.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 900.0, + 1372.0, + 900.0, + 1372.0, + 937.0, + 1323.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 900.0, + 1405.0, + 900.0, + 1405.0, + 937.0, + 1401.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 931.0, + 367.0, + 931.0, + 367.0, + 965.0, + 295.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 931.0, + 607.0, + 931.0, + 607.0, + 965.0, + 386.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 931.0, + 1123.0, + 931.0, + 1123.0, + 965.0, + 640.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 505.0, + 258.0, + 505.0, + 299.0, + 294.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 258.0, + 714.0, + 258.0, + 714.0, + 299.0, + 661.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 710.0, + 1260.0, + 710.0, + 1260.0, + 756.0, + 292.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 710.0, + 1319.0, + 710.0, + 1319.0, + 756.0, + 1305.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 412.0, + 498.0, + 412.0, + 498.0, + 466.0, + 292.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 412.0, + 710.0, + 412.0, + 710.0, + 466.0, + 697.0, + 466.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 310, + 226, + 1392, + 226, + 1392, + 771, + 310, + 771 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1085, + 1405, + 1085, + 1405, + 1209, + 298, + 1209 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 815, + 1405, + 815, + 1405, + 1030, + 296, + 1030 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 314, + 1778, + 1385, + 1778, + 1385, + 1978, + 314, + 1978 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1377, + 1405, + 1377, + 1405, + 1470, + 298, + 1470 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 296, + 1666, + 1399, + 1666, + 1399, + 1731, + 296, + 1731 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 1569, + 1402, + 1569, + 1402, + 1633, + 297, + 1633 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 700, + 1512, + 998, + 1512, + 998, + 1554, + 700, + 1554 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 287, + 2000, + 1300, + 2000, + 1300, + 2036, + 287, + 2036 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 866, + 2087, + 866, + 2113, + 835, + 2113 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 299, + 1251, + 1118, + 1251, + 1118, + 1289, + 299, + 1289 + ], + "score": 0.76 + }, + { + "category_id": 1, + "poly": [ + 299, + 1319, + 635, + 1319, + 635, + 1352, + 299, + 1352 + ], + "score": 0.738 + }, + { + "category_id": 0, + "poly": [ + 299, + 1251, + 1118, + 1251, + 1118, + 1289, + 299, + 1289 + ], + "score": 0.205 + }, + { + "category_id": 0, + "poly": [ + 299, + 1319, + 635, + 1319, + 635, + 1352, + 299, + 1352 + ], + "score": 0.183 + }, + { + "category_id": 14, + "poly": [ + 314, + 1778, + 1386, + 1778, + 1386, + 1982, + 314, + 1982 + ], + "score": 0.94, + "latex": "\\begin{array} { l } { \\displaystyle \\ell _ { v a r } ( W ) = \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] - \\frac { 1 } { 1 6 } \\mathbb { E } _ { { x } } [ \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ] ^ { 2 } } \\\\ { \\displaystyle = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } - \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\\\ { \\displaystyle \\quad = \\frac { 1 } { 8 } \\| W _ { * } ^ { T } W _ { * } - W ^ { T } W \\| _ { F } ^ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 513, + 874, + 598, + 874, + 598, + 905, + 513, + 905 + ], + "score": 0.9, + "latex": "\\mathbb { R } ^ { 5 0 \\times 5 0 }" + }, + { + "category_id": 14, + "poly": [ + 700, + 1513, + 998, + 1513, + 998, + 1552, + 700, + 1552 + ], + "score": 0.9, + "latex": "\\dot { \\sigma } _ { i } ( t ) = \\sigma _ { i } ( t ) ( \\sigma _ { i } ^ { * } - \\sigma _ { i } ( t ) )" + }, + { + "category_id": 13, + "poly": [ + 456, + 1574, + 487, + 1574, + 487, + 1605, + 456, + 1605 + ], + "score": 0.89, + "latex": "\\sigma _ { i } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 905, + 1572, + 948, + 1572, + 948, + 1600, + 905, + 1600 + ], + "score": 0.87, + "latex": "W ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 376, + 1576, + 403, + 1576, + 403, + 1602, + 376, + 1602 + ], + "score": 0.84, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 780, + 2008, + 811, + 2008, + 811, + 2030, + 780, + 2030 + ], + "score": 0.77, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 820, + 1572, + 852, + 1572, + 852, + 1600, + 820, + 1600 + ], + "score": 0.75, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 839, + 1119, + 852, + 1119, + 852, + 1144, + 839, + 1144 + ], + "score": 0.68, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1122, + 1119, + 1135, + 1119, + 1135, + 1144, + 1122, + 1144 + ], + "score": 0.3, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 229.0, + 536.0, + 229.0, + 536.0, + 254.0, + 383.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 229.0, + 807.0, + 229.0, + 807.0, + 254.0, + 655.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 229.0, + 1078.0, + 229.0, + 1078.0, + 254.0, + 925.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 229.0, + 1349.0, + 229.0, + 1349.0, + 254.0, + 1196.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 259.0, + 341.0, + 259.0, + 341.0, + 270.0, + 331.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 253.0, + 616.0, + 253.0, + 616.0, + 272.0, + 592.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 259.0, + 885.0, + 259.0, + 885.0, + 278.0, + 863.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 259.0, + 1156.0, + 259.0, + 1156.0, + 278.0, + 1134.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 299.0, + 341.0, + 299.0, + 341.0, + 311.0, + 331.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 287.0, + 616.0, + 287.0, + 616.0, + 306.0, + 592.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 292.0, + 885.0, + 292.0, + 885.0, + 311.0, + 863.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 292.0, + 1158.0, + 292.0, + 1158.0, + 311.0, + 1134.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 322.0, + 616.0, + 322.0, + 616.0, + 339.0, + 599.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 326.0, + 885.0, + 326.0, + 885.0, + 341.0, + 869.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 326.0, + 1156.0, + 326.0, + 1156.0, + 342.0, + 1140.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 333.0, + 335.0, + 333.0, + 335.0, + 398.0, + 315.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 337.0, + 342.0, + 337.0, + 342.0, + 354.0, + 329.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 355.0, + 615.0, + 355.0, + 615.0, + 373.0, + 599.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 360.0, + 883.0, + 360.0, + 883.0, + 373.0, + 872.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 359.0, + 1155.0, + 359.0, + 1155.0, + 374.0, + 1140.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 378.0, + 341.0, + 378.0, + 341.0, + 390.0, + 330.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 389.0, + 616.0, + 389.0, + 616.0, + 405.0, + 600.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 394.0, + 883.0, + 394.0, + 883.0, + 404.0, + 872.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 394.0, + 1154.0, + 394.0, + 1154.0, + 404.0, + 1144.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 418.0, + 342.0, + 418.0, + 342.0, + 430.0, + 331.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 422.0, + 615.0, + 422.0, + 615.0, + 439.0, + 599.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 425.0, + 883.0, + 425.0, + 883.0, + 438.0, + 872.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 424.0, + 1155.0, + 424.0, + 1155.0, + 439.0, + 1140.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 456.0, + 345.0, + 456.0, + 345.0, + 472.0, + 330.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 456.0, + 616.0, + 456.0, + 616.0, + 472.0, + 599.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 456.0, + 887.0, + 456.0, + 887.0, + 472.0, + 871.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 456.0, + 1155.0, + 456.0, + 1155.0, + 472.0, + 1142.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 476.0, + 365.0, + 476.0, + 365.0, + 492.0, + 350.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 476.0, + 456.0, + 476.0, + 456.0, + 493.0, + 381.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 476.0, + 498.0, + 476.0, + 498.0, + 493.0, + 461.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 476.0, + 537.0, + 476.0, + 537.0, + 493.0, + 501.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 476.0, + 580.0, + 476.0, + 580.0, + 493.0, + 538.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 477.0, + 636.0, + 477.0, + 636.0, + 492.0, + 621.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 475.0, + 847.0, + 475.0, + 847.0, + 495.0, + 640.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 475.0, + 1126.0, + 475.0, + 1126.0, + 495.0, + 890.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 477.0, + 1176.0, + 477.0, + 1176.0, + 492.0, + 1161.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 475.0, + 1374.0, + 475.0, + 1374.0, + 495.0, + 1187.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 496.0, + 532.0, + 496.0, + 532.0, + 516.0, + 389.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 496.0, + 803.0, + 496.0, + 803.0, + 516.0, + 660.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 496.0, + 1073.0, + 496.0, + 1073.0, + 516.0, + 931.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 496.0, + 1344.0, + 496.0, + 1344.0, + 516.0, + 1201.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 526.0, + 885.0, + 526.0, + 885.0, + 545.0, + 863.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 524.0, + 919.0, + 524.0, + 919.0, + 534.0, + 908.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 526.0, + 1156.0, + 526.0, + 1156.0, + 546.0, + 1134.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 526.0, + 1223.0, + 526.0, + 1223.0, + 536.0, + 1208.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 531.0, + 345.0, + 531.0, + 345.0, + 550.0, + 319.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 539.0, + 616.0, + 539.0, + 616.0, + 559.0, + 592.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 559.0, + 885.0, + 559.0, + 885.0, + 578.0, + 863.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 559.0, + 1156.0, + 559.0, + 1156.0, + 578.0, + 1134.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 568.0, + 346.0, + 568.0, + 346.0, + 587.0, + 319.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 577.0, + 615.0, + 577.0, + 615.0, + 593.0, + 599.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 596.0, + 342.0, + 596.0, + 342.0, + 661.0, + 304.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 611.0, + 617.0, + 611.0, + 617.0, + 631.0, + 597.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 592.0, + 885.0, + 592.0, + 885.0, + 608.0, + 869.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 592.0, + 1156.0, + 592.0, + 1156.0, + 608.0, + 1140.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 623.0, + 887.0, + 623.0, + 887.0, + 640.0, + 869.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 623.0, + 1158.0, + 623.0, + 1158.0, + 640.0, + 1140.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 642.0, + 344.0, + 642.0, + 344.0, + 662.0, + 319.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 648.0, + 616.0, + 648.0, + 616.0, + 665.0, + 600.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 656.0, + 885.0, + 656.0, + 885.0, + 671.0, + 869.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 656.0, + 1156.0, + 656.0, + 1156.0, + 671.0, + 1142.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 681.0, + 345.0, + 681.0, + 345.0, + 700.0, + 319.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 684.0, + 615.0, + 684.0, + 615.0, + 701.0, + 599.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 687.0, + 884.0, + 687.0, + 884.0, + 704.0, + 869.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 687.0, + 1155.0, + 687.0, + 1155.0, + 704.0, + 1140.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 718.0, + 345.0, + 718.0, + 345.0, + 738.0, + 319.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 720.0, + 616.0, + 720.0, + 616.0, + 737.0, + 600.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 720.0, + 888.0, + 720.0, + 888.0, + 737.0, + 871.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 720.0, + 1158.0, + 720.0, + 1158.0, + 737.0, + 1140.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 739.0, + 365.0, + 739.0, + 365.0, + 755.0, + 350.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 737.0, + 429.0, + 737.0, + 429.0, + 759.0, + 392.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 737.0, + 482.0, + 737.0, + 482.0, + 758.0, + 445.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 737.0, + 536.0, + 737.0, + 536.0, + 759.0, + 500.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 739.0, + 588.0, + 739.0, + 588.0, + 757.0, + 554.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 739.0, + 636.0, + 739.0, + 636.0, + 755.0, + 620.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 738.0, + 860.0, + 738.0, + 860.0, + 758.0, + 644.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 742.0, + 905.0, + 742.0, + 905.0, + 753.0, + 893.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 739.0, + 1126.0, + 739.0, + 1126.0, + 755.0, + 913.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 739.0, + 1176.0, + 739.0, + 1176.0, + 755.0, + 1160.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 739.0, + 1234.0, + 739.0, + 1234.0, + 757.0, + 1195.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 739.0, + 1281.0, + 739.0, + 1281.0, + 757.0, + 1240.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 739.0, + 1325.0, + 739.0, + 1325.0, + 757.0, + 1284.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 739.0, + 1371.0, + 739.0, + 1371.0, + 757.0, + 1331.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 752.0, + 487.0, + 752.0, + 487.0, + 772.0, + 436.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 752.0, + 757.0, + 752.0, + 757.0, + 772.0, + 706.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 752.0, + 1028.0, + 752.0, + 1028.0, + 772.0, + 977.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 752.0, + 1298.0, + 752.0, + 1298.0, + 772.0, + 1247.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 815.0, + 1403.0, + 815.0, + 1403.0, + 850.0, + 295.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 847.0, + 1403.0, + 847.0, + 1403.0, + 878.0, + 295.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 868.0, + 512.0, + 868.0, + 512.0, + 914.0, + 290.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 868.0, + 1408.0, + 868.0, + 1408.0, + 914.0, + 599.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 908.0, + 1405.0, + 908.0, + 1405.0, + 941.0, + 293.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 938.0, + 1403.0, + 938.0, + 1403.0, + 969.0, + 296.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 968.0, + 1403.0, + 968.0, + 1403.0, + 1002.0, + 294.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 996.0, + 552.0, + 996.0, + 552.0, + 1035.0, + 293.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1248.0, + 1120.0, + 1248.0, + 1120.0, + 1295.0, + 292.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1319.0, + 635.0, + 1319.0, + 635.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1085.0, + 1406.0, + 1085.0, + 1406.0, + 1121.0, + 292.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1115.0, + 838.0, + 1115.0, + 838.0, + 1151.0, + 293.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1115.0, + 1121.0, + 1115.0, + 1121.0, + 1151.0, + 853.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1115.0, + 1405.0, + 1115.0, + 1405.0, + 1151.0, + 1136.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1148.0, + 1403.0, + 1148.0, + 1403.0, + 1181.0, + 296.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1176.0, + 1341.0, + 1176.0, + 1341.0, + 1212.0, + 293.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1375.0, + 1403.0, + 1375.0, + 1403.0, + 1414.0, + 295.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1407.0, + 1403.0, + 1407.0, + 1403.0, + 1444.0, + 293.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 369.0, + 1438.0, + 369.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1697.0, + 1192.0, + 1697.0, + 1192.0, + 1735.0, + 294.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1566.0, + 375.0, + 1566.0, + 375.0, + 1609.0, + 294.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1566.0, + 455.0, + 1566.0, + 455.0, + 1609.0, + 404.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1566.0, + 819.0, + 1566.0, + 819.0, + 1609.0, + 488.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1566.0, + 904.0, + 1566.0, + 904.0, + 1609.0, + 853.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1566.0, + 1406.0, + 1566.0, + 1406.0, + 1609.0, + 949.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 489.0, + 1601.0, + 489.0, + 1638.0, + 293.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1998.0, + 779.0, + 1998.0, + 779.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1998.0, + 1301.0, + 1998.0, + 1301.0, + 2039.0, + 812.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1248.0, + 1120.0, + 1248.0, + 1120.0, + 1295.0, + 292.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1319.0, + 635.0, + 1319.0, + 635.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1005, + 1405, + 1005, + 1405, + 1100, + 298, + 1100 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 523, + 1403, + 523, + 1403, + 618, + 298, + 618 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 499, + 1898, + 1201, + 1898, + 1201, + 2032, + 499, + 2032 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1435, + 1400, + 1435, + 1400, + 1530, + 298, + 1530 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 1327, + 1401, + 1327, + 1401, + 1421, + 299, + 1421 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 294, + 1780, + 1403, + 1780, + 1403, + 1845, + 294, + 1845 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 297, + 1614, + 1400, + 1614, + 1400, + 1678, + 297, + 1678 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 679, + 254, + 1022, + 254, + 1022, + 316, + 679, + 316 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 545, + 913, + 1155, + 913, + 1155, + 975, + 545, + 975 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 292, + 808, + 1405, + 808, + 1405, + 874, + 292, + 874 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 746, + 1721, + 951, + 1721, + 951, + 1761, + 746, + 1761 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 295, + 339, + 1196, + 339, + 1196, + 376, + 295, + 376 + ], + "score": 0.934 + }, + { + "category_id": 8, + "poly": [ + 313, + 428, + 1333, + 428, + 1333, + 495, + 313, + 495 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 746, + 1142, + 952, + 1142, + 952, + 1181, + 746, + 1181 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 299, + 1561, + 696, + 1561, + 696, + 1594, + 299, + 1594 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 736, + 746, + 963, + 746, + 963, + 789, + 736, + 789 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 695, + 658, + 1005, + 658, + 1005, + 701, + 695, + 701 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 1270, + 810, + 1270, + 810, + 1303, + 300, + 1303 + ], + "score": 0.906 + }, + { + "category_id": 9, + "poly": [ + 1354, + 445, + 1400, + 445, + 1400, + 480, + 1354, + 480 + ], + "score": 0.846 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1202, + 1403, + 1202, + 1403, + 1232, + 1374, + 1232 + ], + "score": 0.843 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.605 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 863, + 2087, + 863, + 2113, + 834, + 2113 + ], + "score": 0.254 + }, + { + "category_id": 14, + "poly": [ + 499, + 1899, + 1200, + 1899, + 1200, + 2037, + 499, + 2037 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\ell ( W ) = \\displaystyle \\frac { 1 } { 1 6 } \\mathbb { E } _ { x } [ ( \\langle W _ { * } ^ { T } W _ { * } - W ^ { T } W , x x ^ { T } \\rangle ) ^ { 2 } ] } \\\\ & { \\qquad = \\displaystyle \\frac { 1 } { 8 } | | W _ { * } ^ { T } W _ { * } - W ^ { T } W | | _ { F } ^ { 2 } + \\frac { 1 } { 1 6 } T r ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) ^ { 2 } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 311, + 427, + 1341, + 427, + 1341, + 495, + 311, + 495 + ], + "score": 0.94, + "latex": "W ^ { \\dot { T } } W = W ^ { T } \\dot { W } + \\dot { W } ^ { T } W = \\frac { 1 } { 2 } \\Bigl ( W ^ { T } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) + ( W _ { * } ^ { T } W _ { * } - W ^ { T } W ) W ^ { T } W \\Bigr )" + }, + { + "category_id": 14, + "poly": [ + 677, + 250, + 1022, + 250, + 1022, + 317, + 677, + 317 + ], + "score": 0.93, + "latex": "\\dot { W } = \\frac { 1 } { 2 } W ( W _ { * } ^ { T } W _ { * } - W ^ { T } W )" + }, + { + "category_id": 14, + "poly": [ + 545, + 908, + 1154, + 908, + 1154, + 975, + 545, + 975 + ], + "score": 0.92, + "latex": "\\dot { D } = \\frac { 1 } { 2 } \\Big ( D ( D ^ { * } - D ) + ( D ^ { * } - D ) D \\Big ) = D ( D ^ { * } - D )" + }, + { + "category_id": 13, + "poly": [ + 1092, + 525, + 1181, + 525, + 1181, + 560, + 1092, + 560 + ], + "score": 0.92, + "latex": "W _ { * } ^ { T } W _ { * }" + }, + { + "category_id": 13, + "poly": [ + 633, + 525, + 810, + 525, + 810, + 561, + 633, + 561 + ], + "score": 0.91, + "latex": "W _ { 0 } ^ { T } W _ { 0 } = \\sigma _ { 0 } I \\mathrm { . }" + }, + { + "category_id": 14, + "poly": [ + 748, + 1141, + 951, + 1141, + 951, + 1179, + 748, + 1179 + ], + "score": 0.91, + "latex": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 882, + 340, + 963, + 340, + 963, + 371, + 882, + 371 + ], + "score": 0.91, + "latex": "W ^ { T } W" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1614, + 1330, + 1614, + 1330, + 1644, + 1250, + 1644 + ], + "score": 0.9, + "latex": "W ^ { T } W" + }, + { + "category_id": 13, + "poly": [ + 957, + 525, + 1037, + 525, + 1037, + 556, + 957, + 556 + ], + "score": 0.9, + "latex": "W ^ { T } W" + }, + { + "category_id": 13, + "poly": [ + 706, + 810, + 748, + 810, + 748, + 839, + 706, + 839 + ], + "score": 0.89, + "latex": "U ^ { T }" + }, + { + "category_id": 14, + "poly": [ + 736, + 746, + 964, + 746, + 964, + 788, + 736, + 788 + ], + "score": 0.89, + "latex": "D ^ { * } = U ^ { T } W _ { * } ^ { T } W _ { * } U" + }, + { + "category_id": 14, + "poly": [ + 748, + 1721, + 951, + 1721, + 951, + 1759, + 748, + 1759 + ], + "score": 0.88, + "latex": "\\dot { \\sigma } _ { i } = \\sigma _ { i } ( \\sigma _ { i } ^ { * } - \\sigma _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 842, + 377, + 842, + 377, + 870, + 298, + 870 + ], + "score": 0.86, + "latex": "W ^ { T } W" + }, + { + "category_id": 14, + "poly": [ + 694, + 658, + 1007, + 658, + 1007, + 701, + 694, + 701 + ], + "score": 0.84, + "latex": "D ( 0 ) = U ^ { T } W ( 0 ) ^ { T } W ( 0 ) U" + }, + { + "category_id": 13, + "poly": [ + 630, + 813, + 654, + 813, + 654, + 839, + 630, + 839 + ], + "score": 0.84, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 298, + 590, + 336, + 590, + 336, + 616, + 298, + 616 + ], + "score": 0.84, + "latex": "D ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 817, + 559, + 841, + 559, + 841, + 585, + 817, + 585 + ], + "score": 0.83, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 1328, + 559, + 1354, + 559, + 1354, + 585, + 1328, + 585 + ], + "score": 0.83, + "latex": "D" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 699.0, + 1559.0, + 699.0, + 1597.0, + 294.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1267.0, + 814.0, + 1267.0, + 814.0, + 1307.0, + 291.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1206.0, + 1403.0, + 1206.0, + 1403.0, + 1234.0, + 1377.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2125.0, + 830.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1006.0, + 1402.0, + 1006.0, + 1402.0, + 1040.0, + 296.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1071.0, + 296.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1067.0, + 1127.0, + 1067.0, + 1127.0, + 1101.0, + 295.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 522.0, + 632.0, + 522.0, + 632.0, + 563.0, + 292.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 522.0, + 956.0, + 522.0, + 956.0, + 563.0, + 811.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 522.0, + 1091.0, + 522.0, + 1091.0, + 563.0, + 1038.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 522.0, + 1404.0, + 522.0, + 1404.0, + 563.0, + 1182.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 557.0, + 816.0, + 557.0, + 816.0, + 591.0, + 294.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 557.0, + 1327.0, + 557.0, + 1327.0, + 591.0, + 842.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 557.0, + 1404.0, + 557.0, + 1404.0, + 591.0, + 1355.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 586.0, + 297.0, + 586.0, + 297.0, + 622.0, + 293.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 586.0, + 353.0, + 586.0, + 353.0, + 622.0, + 337.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1436.0, + 1404.0, + 1436.0, + 1404.0, + 1470.0, + 296.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 1404.0, + 1467.0, + 1404.0, + 1501.0, + 293.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1497.0, + 1247.0, + 1497.0, + 1247.0, + 1531.0, + 293.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1325.0, + 1403.0, + 1325.0, + 1403.0, + 1363.0, + 293.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1357.0, + 1405.0, + 1357.0, + 1405.0, + 1393.0, + 294.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1387.0, + 616.0, + 1387.0, + 616.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1780.0, + 1404.0, + 1780.0, + 1404.0, + 1816.0, + 296.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1814.0, + 609.0, + 1814.0, + 609.0, + 1846.0, + 296.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1614.0, + 1249.0, + 1614.0, + 1249.0, + 1650.0, + 293.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1614.0, + 1404.0, + 1614.0, + 1404.0, + 1650.0, + 1331.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1646.0, + 495.0, + 1646.0, + 495.0, + 1679.0, + 295.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 805.0, + 629.0, + 805.0, + 629.0, + 849.0, + 291.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 805.0, + 705.0, + 805.0, + 705.0, + 849.0, + 655.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 805.0, + 1408.0, + 805.0, + 1408.0, + 849.0, + 749.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 837.0, + 297.0, + 837.0, + 297.0, + 876.0, + 294.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 837.0, + 394.0, + 837.0, + 394.0, + 876.0, + 378.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 335.0, + 881.0, + 335.0, + 881.0, + 381.0, + 293.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 335.0, + 1199.0, + 335.0, + 1199.0, + 381.0, + 964.0, + 381.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1677, + 1404, + 1677, + 1404, + 1802, + 297, + 1802 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 626, + 1404, + 626, + 1404, + 752, + 297, + 752 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 388, + 1405, + 388, + 1405, + 482, + 297, + 482 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1941, + 1403, + 1941, + 1403, + 2036, + 298, + 2036 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 765, + 1404, + 765, + 1404, + 860, + 297, + 860 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1570, + 1404, + 1570, + 1404, + 1665, + 297, + 1665 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 565, + 1260, + 1129, + 1260, + 1129, + 1426, + 565, + 1426 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 300, + 1458, + 1400, + 1458, + 1400, + 1556, + 300, + 1556 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 299, + 945, + 1404, + 945, + 1404, + 1039, + 299, + 1039 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 1143, + 1403, + 1143, + 1403, + 1208, + 300, + 1208 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 635, + 520, + 1062, + 520, + 1062, + 602, + 635, + 602 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 696, + 298, + 1002, + 298, + 1002, + 366, + 696, + 366 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 770, + 1839, + 931, + 1839, + 931, + 1918, + 770, + 1918 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 685, + 1079, + 1018, + 1079, + 1018, + 1122, + 685, + 1122 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 295, + 225, + 1252, + 225, + 1252, + 264, + 295, + 264 + ], + "score": 0.928 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 817, + 73, + 817, + 106, + 297, + 106 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 297, + 891, + 1028, + 891, + 1028, + 926, + 297, + 926 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2113, + 835, + 2113 + ], + "score": 0.875 + }, + { + "category_id": 14, + "poly": [ + 636, + 519, + 1062, + 519, + 1062, + 605, + 636, + 605 + ], + "score": 0.94, + "latex": "\\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) \\Big )" + }, + { + "category_id": 14, + "poly": [ + 694, + 297, + 1005, + 297, + 1005, + 365, + 694, + 365 + ], + "score": 0.94, + "latex": "\\dot { W } = \\frac { 1 } { 2 } W \\Delta + \\frac { 1 } { 4 } T r ( \\Delta ) W" + }, + { + "category_id": 14, + "poly": [ + 565, + 1261, + 1133, + 1261, + 1133, + 1430, + 565, + 1430 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { \\displaystyle \\dot { \\sigma } _ { i } = \\sigma _ { i } \\Big ( \\sigma _ { i } ^ { * } - \\sigma _ { i } + \\frac { 1 } { 2 } \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + \\frac { 1 } { 2 } ( b ^ { * } - b ) \\Big ) } \\\\ { \\displaystyle \\dot { b } = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } - b } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 402, + 228, + 662, + 228, + 662, + 263, + 402, + 263 + ], + "score": 0.93, + "latex": "\\Delta = W _ { * } ^ { T } W _ { * } - W ^ { T } W" + }, + { + "category_id": 14, + "poly": [ + 770, + 1838, + 932, + 1838, + 932, + 1918, + 770, + 1918 + ], + "score": 0.93, + "latex": "\\frac { \\dot { \\sigma } _ { i } ( t ) } { \\dot { b } ( t ) } \\approx \\sigma _ { i } ( t )" + }, + { + "category_id": 13, + "poly": [ + 1182, + 1942, + 1307, + 1942, + 1307, + 1975, + 1182, + 1975 + ], + "score": 0.92, + "latex": "\\sigma _ { i } ( t ) \\ll 1" + }, + { + "category_id": 13, + "poly": [ + 1269, + 975, + 1400, + 975, + 1400, + 1006, + 1269, + 1006 + ], + "score": 0.92, + "latex": "W \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 388, + 1400, + 388, + 1400, + 423, + 1312, + 423 + ], + "score": 0.91, + "latex": "W _ { * } ^ { T } W _ { * }" + }, + { + "category_id": 13, + "poly": [ + 610, + 388, + 690, + 388, + 690, + 419, + 610, + 419 + ], + "score": 0.91, + "latex": "W ^ { T } W" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1681, + 1400, + 1681, + 1400, + 1713, + 1317, + 1713 + ], + "score": 0.91, + "latex": "\\sigma _ { i } ^ { * } - \\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 405, + 450, + 484, + 450, + 484, + 480, + 405, + 480 + ], + "score": 0.91, + "latex": "W ^ { T } W" + }, + { + "category_id": 14, + "poly": [ + 683, + 1078, + 1017, + 1078, + 1017, + 1122, + 683, + 1122 + ], + "score": 0.9, + "latex": "f _ { W , b } ( x ) = \\langle W ^ { T } W , x x ^ { T } \\rangle + b" + }, + { + "category_id": 13, + "poly": [ + 847, + 1459, + 1118, + 1459, + 1118, + 1498, + 847, + 1498 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { ( b = \\sum _ { j } ( \\sigma _ { j } ^ { * } - \\sigma _ { j } ) + b ^ { * } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 435, + 1008, + 503, + 1008, + 503, + 1036, + 435, + 1036 + ], + "score": 0.9, + "latex": "b \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 872, + 1495, + 947, + 1495, + 947, + 1522, + 872, + 1522 + ], + "score": 0.87, + "latex": "b = b ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1311, + 1464, + 1339, + 1464, + 1339, + 1490, + 1311, + 1490 + ], + "score": 0.85, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 838, + 1686, + 866, + 1686, + 866, + 1710, + 838, + 1710 + ], + "score": 0.84, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1745, + 1160, + 1745, + 1160, + 1771, + 1132, + 1771 + ], + "score": 0.84, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 904, + 1948, + 932, + 1948, + 932, + 1973, + 904, + 1973 + ], + "score": 0.84, + "latex": "\\sigma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 460, + 1462, + 475, + 1462, + 475, + 1487, + 460, + 1487 + ], + "score": 0.79, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 737, + 1681, + 752, + 1681, + 752, + 1707, + 737, + 1707 + ], + "score": 0.74, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 1117, + 232, + 1149, + 232, + 1149, + 259, + 1117, + 259 + ], + "score": 0.69, + "latex": "W" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 888.0, + 1029.0, + 888.0, + 1029.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1675.0, + 736.0, + 1675.0, + 736.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1675.0, + 837.0, + 1675.0, + 837.0, + 1716.0, + 753.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1675.0, + 1316.0, + 1675.0, + 1316.0, + 1716.0, + 867.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1675.0, + 1405.0, + 1675.0, + 1405.0, + 1716.0, + 1401.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1710.0, + 1404.0, + 1710.0, + 1404.0, + 1743.0, + 296.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 1131.0, + 1742.0, + 1131.0, + 1775.0, + 295.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1775.0, + 1161.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 547.0, + 1772.0, + 547.0, + 1805.0, + 293.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 628.0, + 1405.0, + 628.0, + 1405.0, + 661.0, + 296.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 657.0, + 1406.0, + 657.0, + 1406.0, + 693.0, + 292.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 723.0, + 294.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 718.0, + 1402.0, + 718.0, + 1402.0, + 754.0, + 294.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 388.0, + 609.0, + 388.0, + 609.0, + 425.0, + 295.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 388.0, + 1311.0, + 388.0, + 1311.0, + 425.0, + 691.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 420.0, + 1405.0, + 420.0, + 1405.0, + 454.0, + 296.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 450.0, + 404.0, + 450.0, + 404.0, + 481.0, + 293.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 450.0, + 497.0, + 450.0, + 497.0, + 481.0, + 485.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 903.0, + 1939.0, + 903.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1939.0, + 1181.0, + 1939.0, + 1181.0, + 1979.0, + 933.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1979.0, + 1308.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 766.0, + 1402.0, + 766.0, + 1402.0, + 800.0, + 296.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 794.0, + 1404.0, + 794.0, + 1404.0, + 831.0, + 294.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 825.0, + 405.0, + 825.0, + 405.0, + 865.0, + 292.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1571.0, + 1404.0, + 1571.0, + 1404.0, + 1605.0, + 295.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1633.0, + 1039.0, + 1633.0, + 1039.0, + 1667.0, + 295.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1455.0, + 459.0, + 1455.0, + 459.0, + 1501.0, + 291.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1455.0, + 846.0, + 1455.0, + 846.0, + 1501.0, + 476.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1455.0, + 1310.0, + 1455.0, + 1310.0, + 1501.0, + 1119.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 1455.0, + 1408.0, + 1455.0, + 1408.0, + 1501.0, + 1340.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1493.0, + 871.0, + 1493.0, + 871.0, + 1528.0, + 296.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1493.0, + 1404.0, + 1493.0, + 1404.0, + 1528.0, + 948.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1523.0, + 670.0, + 1523.0, + 670.0, + 1560.0, + 294.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 946.0, + 1404.0, + 946.0, + 1404.0, + 980.0, + 294.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 973.0, + 1268.0, + 973.0, + 1268.0, + 1013.0, + 288.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 973.0, + 1407.0, + 973.0, + 1407.0, + 1013.0, + 1401.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1008.0, + 434.0, + 1008.0, + 434.0, + 1038.0, + 295.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1008.0, + 515.0, + 1008.0, + 515.0, + 1038.0, + 504.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1143.0, + 1404.0, + 1143.0, + 1404.0, + 1179.0, + 296.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1174.0, + 825.0, + 1174.0, + 825.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 401.0, + 222.0, + 401.0, + 271.0, + 291.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 222.0, + 1116.0, + 222.0, + 1116.0, + 271.0, + 663.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 222.0, + 1257.0, + 222.0, + 1257.0, + 271.0, + 1150.0, + 271.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1462, + 1405, + 1462, + 1405, + 1708, + 298, + 1708 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 379, + 222, + 1318, + 222, + 1318, + 696, + 379, + 696 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1156, + 1404, + 1156, + 1404, + 1280, + 298, + 1280 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1810, + 1405, + 1810, + 1405, + 1903, + 298, + 1903 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 737, + 1405, + 737, + 1405, + 1013, + 296, + 1013 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 511, + 1952, + 1188, + 1952, + 1188, + 2043, + 511, + 2043 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 297, + 1078, + 1400, + 1078, + 1400, + 1142, + 297, + 1142 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 300, + 1402, + 648, + 1402, + 648, + 1435, + 300, + 1435 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 300, + 1330, + 1015, + 1330, + 1015, + 1366, + 300, + 1366 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.856 + }, + { + "category_id": 0, + "poly": [ + 298, + 1751, + 942, + 1751, + 942, + 1782, + 298, + 1782 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 298, + 1751, + 942, + 1751, + 942, + 1782, + 298, + 1782 + ], + "score": 0.154 + }, + { + "category_id": 14, + "poly": [ + 513, + 1948, + 1188, + 1948, + 1188, + 2044, + 513, + 2044 + ], + "score": 0.94, + "latex": "h _ { n } [ i ] = \\sum _ { k = 0 } ^ { d - 1 } w _ { n } [ k ] h _ { n - 1 } [ ( i + k ) \\mod d ] = ( h _ { n - 1 } \\star w _ { n } ) [ i ]" + }, + { + "category_id": 13, + "poly": [ + 1278, + 1679, + 1361, + 1679, + 1361, + 1704, + 1278, + 1704 + ], + "score": 0.89, + "latex": "t \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 707, + 1875, + 744, + 1875, + 744, + 1903, + 707, + 1903 + ], + "score": 0.85, + "latex": "w _ { n }" + }, + { + "category_id": 13, + "poly": [ + 672, + 1528, + 693, + 1528, + 693, + 1551, + 672, + 1551 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 231.0, + 593.0, + 231.0, + 593.0, + 248.0, + 440.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 231.0, + 824.0, + 231.0, + 824.0, + 248.0, + 672.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 229.0, + 1054.0, + 229.0, + 1054.0, + 249.0, + 906.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 231.0, + 1284.0, + 231.0, + 1284.0, + 248.0, + 1136.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 241.0, + 1115.0, + 241.0, + 1115.0, + 260.0, + 1086.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 268.0, + 417.0, + 268.0, + 417.0, + 285.0, + 394.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 262.0, + 646.0, + 262.0, + 646.0, + 279.0, + 624.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 262.0, + 879.0, + 262.0, + 879.0, + 278.0, + 856.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 262.0, + 1113.0, + 262.0, + 1113.0, + 282.0, + 1087.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 288.0, + 647.0, + 288.0, + 647.0, + 304.0, + 624.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 284.0, + 1114.0, + 284.0, + 1114.0, + 303.0, + 1087.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 297.0, + 418.0, + 297.0, + 418.0, + 316.0, + 392.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 312.0, + 649.0, + 312.0, + 649.0, + 331.0, + 623.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 299.0, + 881.0, + 299.0, + 881.0, + 318.0, + 855.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 305.0, + 1113.0, + 305.0, + 1113.0, + 325.0, + 1087.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 328.0, + 416.0, + 328.0, + 416.0, + 344.0, + 393.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 327.0, + 1113.0, + 327.0, + 1113.0, + 347.0, + 1087.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 339.0, + 653.0, + 339.0, + 653.0, + 358.0, + 623.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 338.0, + 881.0, + 338.0, + 881.0, + 357.0, + 854.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 349.0, + 1115.0, + 349.0, + 1115.0, + 368.0, + 1087.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 364.0, + 652.0, + 364.0, + 652.0, + 383.0, + 623.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 377.0, + 881.0, + 377.0, + 881.0, + 396.0, + 855.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 371.0, + 1113.0, + 371.0, + 1113.0, + 390.0, + 1087.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 385.0, + 419.0, + 385.0, + 419.0, + 404.0, + 393.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 391.0, + 648.0, + 391.0, + 648.0, + 407.0, + 624.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 394.0, + 1113.0, + 394.0, + 1113.0, + 413.0, + 1087.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 416.0, + 420.0, + 416.0, + 420.0, + 435.0, + 393.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 417.0, + 648.0, + 417.0, + 648.0, + 434.0, + 626.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 418.0, + 879.0, + 418.0, + 879.0, + 434.0, + 856.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 416.0, + 1113.0, + 416.0, + 1113.0, + 435.0, + 1087.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 436.0, + 433.0, + 436.0, + 433.0, + 450.0, + 419.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 434.0, + 497.0, + 434.0, + 497.0, + 452.0, + 440.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 434.0, + 615.0, + 434.0, + 615.0, + 451.0, + 500.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 436.0, + 666.0, + 436.0, + 666.0, + 450.0, + 650.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 434.0, + 853.0, + 434.0, + 853.0, + 451.0, + 668.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 436.0, + 895.0, + 436.0, + 895.0, + 450.0, + 882.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 433.0, + 952.0, + 433.0, + 952.0, + 452.0, + 921.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 434.0, + 1002.0, + 434.0, + 1002.0, + 452.0, + 970.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 434.0, + 1050.0, + 434.0, + 1050.0, + 452.0, + 1018.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 436.0, + 1127.0, + 436.0, + 1127.0, + 450.0, + 1114.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 434.0, + 1184.0, + 434.0, + 1184.0, + 452.0, + 1153.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 434.0, + 1232.0, + 434.0, + 1232.0, + 452.0, + 1200.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 434.0, + 1280.0, + 434.0, + 1280.0, + 452.0, + 1248.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 444.0, + 1002.0, + 444.0, + 1002.0, + 462.0, + 957.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 444.0, + 1234.0, + 444.0, + 1234.0, + 462.0, + 1189.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 462.0, + 598.0, + 462.0, + 598.0, + 482.0, + 435.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 462.0, + 830.0, + 462.0, + 830.0, + 482.0, + 667.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 462.0, + 1058.0, + 462.0, + 1058.0, + 482.0, + 902.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 462.0, + 1290.0, + 462.0, + 1290.0, + 482.0, + 1132.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 475.0, + 423.0, + 475.0, + 423.0, + 495.0, + 392.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 476.0, + 655.0, + 476.0, + 655.0, + 495.0, + 622.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 486.0, + 1114.0, + 486.0, + 1114.0, + 505.0, + 1087.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 510.0, + 419.0, + 510.0, + 419.0, + 529.0, + 392.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 510.0, + 653.0, + 510.0, + 653.0, + 529.0, + 623.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 504.0, + 880.0, + 504.0, + 880.0, + 520.0, + 856.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 529.0, + 1113.0, + 529.0, + 1113.0, + 544.0, + 1088.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 535.0, + 429.0, + 535.0, + 429.0, + 612.0, + 371.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 544.0, + 653.0, + 544.0, + 653.0, + 563.0, + 623.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 539.0, + 881.0, + 539.0, + 881.0, + 558.0, + 854.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 579.0, + 654.0, + 579.0, + 654.0, + 598.0, + 623.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 575.0, + 881.0, + 575.0, + 881.0, + 594.0, + 854.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 569.0, + 1113.0, + 569.0, + 1113.0, + 585.0, + 1088.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 613.0, + 419.0, + 613.0, + 419.0, + 632.0, + 393.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 613.0, + 652.0, + 613.0, + 652.0, + 632.0, + 623.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 611.0, + 881.0, + 611.0, + 881.0, + 631.0, + 855.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 607.0, + 1115.0, + 607.0, + 1115.0, + 627.0, + 1087.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 648.0, + 419.0, + 648.0, + 419.0, + 667.0, + 393.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 649.0, + 650.0, + 649.0, + 650.0, + 665.0, + 626.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 649.0, + 880.0, + 649.0, + 880.0, + 665.0, + 856.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 649.0, + 1111.0, + 649.0, + 1111.0, + 665.0, + 1088.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 667.0, + 433.0, + 667.0, + 433.0, + 682.0, + 419.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 665.0, + 543.0, + 665.0, + 543.0, + 684.0, + 445.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 667.0, + 613.0, + 667.0, + 613.0, + 682.0, + 549.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 668.0, + 667.0, + 668.0, + 667.0, + 681.0, + 650.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 666.0, + 839.0, + 666.0, + 839.0, + 683.0, + 669.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 667.0, + 895.0, + 667.0, + 895.0, + 682.0, + 882.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 666.0, + 943.0, + 666.0, + 943.0, + 684.0, + 910.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 667.0, + 982.0, + 667.0, + 982.0, + 682.0, + 952.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 667.0, + 1020.0, + 667.0, + 1020.0, + 682.0, + 988.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 665.0, + 1060.0, + 665.0, + 1060.0, + 683.0, + 1027.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 667.0, + 1128.0, + 667.0, + 1128.0, + 682.0, + 1114.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 666.0, + 1202.0, + 666.0, + 1202.0, + 684.0, + 1137.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 667.0, + 1303.0, + 667.0, + 1303.0, + 682.0, + 1205.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 675.0, + 540.0, + 675.0, + 540.0, + 696.0, + 494.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 675.0, + 772.0, + 675.0, + 772.0, + 696.0, + 726.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 677.0, + 1003.0, + 677.0, + 1003.0, + 694.0, + 957.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 677.0, + 1234.0, + 677.0, + 1234.0, + 695.0, + 1189.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 739.0, + 1405.0, + 739.0, + 1405.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 767.0, + 1406.0, + 767.0, + 1406.0, + 803.0, + 291.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 799.0, + 1401.0, + 799.0, + 1401.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 829.0, + 1402.0, + 829.0, + 1402.0, + 863.0, + 294.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 857.0, + 1405.0, + 857.0, + 1405.0, + 895.0, + 293.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 891.0, + 1405.0, + 891.0, + 1405.0, + 925.0, + 293.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 921.0, + 1402.0, + 921.0, + 1402.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 1403.0, + 952.0, + 1403.0, + 985.0, + 294.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 981.0, + 1310.0, + 981.0, + 1310.0, + 1016.0, + 294.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1400.0, + 651.0, + 1400.0, + 651.0, + 1438.0, + 295.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1327.0, + 1019.0, + 1327.0, + 1019.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2124.0, + 828.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1748.0, + 946.0, + 1748.0, + 946.0, + 1787.0, + 294.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1464.0, + 1405.0, + 1464.0, + 1405.0, + 1497.0, + 295.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1494.0, + 1405.0, + 1494.0, + 1405.0, + 1527.0, + 293.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1524.0, + 671.0, + 1524.0, + 671.0, + 1557.0, + 295.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1524.0, + 1406.0, + 1524.0, + 1406.0, + 1557.0, + 694.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1555.0, + 1405.0, + 1555.0, + 1405.0, + 1589.0, + 295.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1619.0, + 295.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1615.0, + 1402.0, + 1615.0, + 1402.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1644.0, + 1405.0, + 1644.0, + 1405.0, + 1679.0, + 293.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1676.0, + 1277.0, + 1676.0, + 1277.0, + 1710.0, + 295.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1676.0, + 1373.0, + 1676.0, + 1373.0, + 1710.0, + 1362.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1155.0, + 1404.0, + 1155.0, + 1404.0, + 1190.0, + 295.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1188.0, + 1405.0, + 1188.0, + 1405.0, + 1220.0, + 296.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1216.0, + 1408.0, + 1216.0, + 1408.0, + 1254.0, + 292.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1248.0, + 1345.0, + 1248.0, + 1345.0, + 1281.0, + 296.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1809.0, + 1405.0, + 1809.0, + 1405.0, + 1844.0, + 295.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1842.0, + 1402.0, + 1842.0, + 1402.0, + 1872.0, + 296.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1872.0, + 706.0, + 1872.0, + 706.0, + 1906.0, + 296.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1872.0, + 952.0, + 1872.0, + 952.0, + 1906.0, + 745.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1074.0, + 1404.0, + 1074.0, + 1404.0, + 1116.0, + 292.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1105.0, + 460.0, + 1105.0, + 460.0, + 1144.0, + 294.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1748.0, + 946.0, + 1748.0, + 946.0, + 1787.0, + 294.0, + 1787.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1554, + 1404, + 1554, + 1404, + 1741, + 297, + 1741 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1367, + 1403, + 1367, + 1403, + 1461, + 297, + 1461 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 305, + 228, + 1394, + 228, + 1394, + 495, + 305, + 495 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 298, + 935, + 1400, + 935, + 1400, + 1000, + 298, + 1000 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 1102, + 1406, + 1102, + 1406, + 1165, + 295, + 1165 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 590, + 1040, + 1109, + 1040, + 1109, + 1082, + 590, + 1082 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 648, + 1249, + 1049, + 1249, + 1049, + 1288, + 648, + 1288 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 292, + 1308, + 1311, + 1308, + 1311, + 1344, + 292, + 1344 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 297, + 1175, + 944, + 1175, + 944, + 1207, + 297, + 1207 + ], + "score": 0.918 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1045, + 1399, + 1045, + 1399, + 1076, + 1352, + 1076 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 299, + 1498, + 1069, + 1498, + 1069, + 1532, + 299, + 1532 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2112, + 834, + 2112 + ], + "score": 0.854 + }, + { + "category_id": 4, + "poly": [ + 295, + 539, + 1406, + 539, + 1406, + 878, + 295, + 878 + ], + "score": 0.783 + }, + { + "category_id": 4, + "poly": [ + 296, + 539, + 1405, + 539, + 1405, + 878, + 296, + 878 + ], + "score": 0.72 + }, + { + "category_id": 13, + "poly": [ + 954, + 935, + 1066, + 935, + 1066, + 968, + 954, + 968 + ], + "score": 0.92, + "latex": "w _ { N } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 648, + 1249, + 1047, + 1249, + 1047, + 1288, + 648, + 1288 + ], + "score": 0.91, + "latex": "\\hat { \\sigma } = d i a g ( \\hat { w } _ { 1 } ) \\cdot \\cdot \\cdot d i a g ( \\hat { w } _ { N - 1 } ) \\hat { w } _ { N } ," + }, + { + "category_id": 13, + "poly": [ + 409, + 1308, + 692, + 1308, + 692, + 1343, + 409, + 1343 + ], + "score": 0.91, + "latex": "n = 1 , 2 , . . . , N , \\hat { w } _ { n } \\in \\mathbb { C } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1308, + 1305, + 1308, + 1305, + 1342, + 1202, + 1342 + ], + "score": 0.91, + "latex": "w _ { n } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 762, + 572, + 865, + 572, + 865, + 600, + 762, + 600 + ], + "score": 0.9, + "latex": "d = 1 0 0" + }, + { + "category_id": 14, + "poly": [ + 588, + 1038, + 1111, + 1038, + 1111, + 1081, + 588, + 1081 + ], + "score": 0.9, + "latex": "f _ { \\sigma } ( x ) = ( ( ( ( x \\star w _ { 1 } ) \\star w _ { 2 } ) \\cdot \\cdot \\cdot ) \\star w _ { N - 1 } ) ^ { T } w _ { N }" + }, + { + "category_id": 13, + "poly": [ + 424, + 820, + 445, + 820, + 445, + 842, + 424, + 842 + ], + "score": 0.76, + "latex": "\\sigma" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 228.0, + 563.0, + 228.0, + 563.0, + 256.0, + 375.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 226.0, + 783.0, + 226.0, + 783.0, + 260.0, + 698.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 226.0, + 1053.0, + 226.0, + 1053.0, + 260.0, + 968.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 226.0, + 1326.0, + 226.0, + 1326.0, + 260.0, + 1240.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 250.0, + 368.0, + 250.0, + 368.0, + 273.0, + 321.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 257.0, + 433.0, + 257.0, + 433.0, + 280.0, + 391.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 252.0, + 636.0, + 252.0, + 636.0, + 272.0, + 617.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 257.0, + 906.0, + 257.0, + 906.0, + 279.0, + 883.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 247.0, + 1180.0, + 247.0, + 1180.0, + 270.0, + 1143.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 277.0, + 365.0, + 277.0, + 365.0, + 412.0, + 304.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 271.0, + 428.0, + 271.0, + 428.0, + 322.0, + 392.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 282.0, + 633.0, + 282.0, + 633.0, + 428.0, + 592.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 285.0, + 905.0, + 285.0, + 905.0, + 425.0, + 859.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 270.0, + 1178.0, + 270.0, + 1178.0, + 431.0, + 1119.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 322.0, + 902.0, + 322.0, + 902.0, + 333.0, + 892.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 317.0, + 1176.0, + 317.0, + 1176.0, + 341.0, + 1143.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 333.0, + 363.0, + 333.0, + 363.0, + 351.0, + 327.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 328.0, + 633.0, + 328.0, + 633.0, + 345.0, + 618.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 353.0, + 902.0, + 353.0, + 902.0, + 363.0, + 892.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 341.0, + 1178.0, + 341.0, + 1178.0, + 364.0, + 1144.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 358.0, + 364.0, + 358.0, + 364.0, + 379.0, + 326.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 364.0, + 633.0, + 364.0, + 633.0, + 381.0, + 618.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 367.0, + 1175.0, + 367.0, + 1175.0, + 386.0, + 1149.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 384.0, + 364.0, + 384.0, + 364.0, + 406.0, + 326.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 381.0, + 904.0, + 381.0, + 904.0, + 393.0, + 893.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 390.0, + 1175.0, + 390.0, + 1175.0, + 410.0, + 1150.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 405.0, + 633.0, + 405.0, + 633.0, + 415.0, + 622.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 411.0, + 365.0, + 411.0, + 365.0, + 433.0, + 326.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 411.0, + 902.0, + 411.0, + 902.0, + 424.0, + 892.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 412.0, + 1176.0, + 412.0, + 1176.0, + 433.0, + 1150.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 437.0, + 365.0, + 437.0, + 365.0, + 460.0, + 324.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 438.0, + 635.0, + 438.0, + 635.0, + 455.0, + 619.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 438.0, + 908.0, + 438.0, + 908.0, + 455.0, + 890.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 434.0, + 1179.0, + 434.0, + 1179.0, + 458.0, + 1149.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 455.0, + 392.0, + 455.0, + 392.0, + 477.0, + 358.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 456.0, + 422.0, + 456.0, + 422.0, + 476.0, + 402.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 455.0, + 461.0, + 455.0, + 461.0, + 475.0, + 441.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 455.0, + 498.0, + 455.0, + 498.0, + 475.0, + 478.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 457.0, + 536.0, + 457.0, + 536.0, + 474.0, + 519.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 457.0, + 575.0, + 457.0, + 575.0, + 474.0, + 557.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 457.0, + 654.0, + 457.0, + 654.0, + 474.0, + 638.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 455.0, + 838.0, + 455.0, + 838.0, + 476.0, + 660.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 457.0, + 925.0, + 457.0, + 925.0, + 474.0, + 909.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 455.0, + 1108.0, + 455.0, + 1108.0, + 476.0, + 931.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 457.0, + 1196.0, + 457.0, + 1196.0, + 474.0, + 1180.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 453.0, + 1382.0, + 453.0, + 1382.0, + 477.0, + 1201.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 467.0, + 503.0, + 467.0, + 503.0, + 498.0, + 437.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 470.0, + 783.0, + 470.0, + 783.0, + 495.0, + 700.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 470.0, + 1053.0, + 470.0, + 1053.0, + 495.0, + 970.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 470.0, + 1324.0, + 470.0, + 1324.0, + 495.0, + 1241.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1497.0, + 1070.0, + 1497.0, + 1070.0, + 1534.0, + 296.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 872.0, + 2084.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1404.0, + 539.0, + 1404.0, + 574.0, + 294.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 571.0, + 761.0, + 571.0, + 761.0, + 606.0, + 293.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 571.0, + 1406.0, + 571.0, + 1406.0, + 606.0, + 866.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 599.0, + 1406.0, + 599.0, + 1406.0, + 639.0, + 290.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 627.0, + 1406.0, + 627.0, + 1406.0, + 671.0, + 292.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 1404.0, + 662.0, + 1404.0, + 696.0, + 293.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 692.0, + 1406.0, + 692.0, + 1406.0, + 727.0, + 294.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 720.0, + 1407.0, + 720.0, + 1407.0, + 761.0, + 292.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 752.0, + 1406.0, + 752.0, + 1406.0, + 788.0, + 292.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1406.0, + 782.0, + 1406.0, + 820.0, + 293.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 815.0, + 423.0, + 815.0, + 423.0, + 849.0, + 295.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 815.0, + 1406.0, + 815.0, + 1406.0, + 849.0, + 446.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 846.0, + 622.0, + 846.0, + 622.0, + 881.0, + 294.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1405.0, + 539.0, + 1405.0, + 574.0, + 294.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 571.0, + 761.0, + 571.0, + 761.0, + 606.0, + 294.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 571.0, + 1406.0, + 571.0, + 1406.0, + 606.0, + 866.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 599.0, + 1406.0, + 599.0, + 1406.0, + 639.0, + 290.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 627.0, + 1406.0, + 627.0, + 1406.0, + 671.0, + 291.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 1405.0, + 662.0, + 1405.0, + 696.0, + 293.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 692.0, + 1405.0, + 692.0, + 1405.0, + 727.0, + 294.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 720.0, + 1406.0, + 720.0, + 1406.0, + 761.0, + 291.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 752.0, + 1403.0, + 752.0, + 1403.0, + 788.0, + 294.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1405.0, + 782.0, + 1405.0, + 820.0, + 293.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 815.0, + 423.0, + 815.0, + 423.0, + 849.0, + 295.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 815.0, + 1405.0, + 815.0, + 1405.0, + 849.0, + 446.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 620.0, + 844.0, + 620.0, + 879.0, + 294.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1557.0, + 1404.0, + 1557.0, + 1404.0, + 1589.0, + 295.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1586.0, + 1405.0, + 1586.0, + 1405.0, + 1622.0, + 294.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1615.0, + 1405.0, + 1615.0, + 1405.0, + 1653.0, + 292.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1649.0, + 1405.0, + 1649.0, + 1405.0, + 1681.0, + 295.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1678.0, + 1401.0, + 1678.0, + 1401.0, + 1710.0, + 295.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1708.0, + 871.0, + 1708.0, + 871.0, + 1744.0, + 294.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1365.0, + 1405.0, + 1365.0, + 1405.0, + 1403.0, + 294.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1433.0, + 293.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1430.0, + 763.0, + 1430.0, + 763.0, + 1460.0, + 295.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 933.0, + 953.0, + 933.0, + 953.0, + 973.0, + 293.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 933.0, + 1405.0, + 933.0, + 1405.0, + 973.0, + 1067.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 962.0, + 659.0, + 962.0, + 659.0, + 1005.0, + 293.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1137.0, + 293.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1132.0, + 1228.0, + 1132.0, + 1228.0, + 1167.0, + 294.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 408.0, + 1305.0, + 408.0, + 1348.0, + 294.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1305.0, + 1201.0, + 1305.0, + 1201.0, + 1348.0, + 693.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1305.0, + 1316.0, + 1305.0, + 1316.0, + 1348.0, + 1306.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1171.0, + 945.0, + 1171.0, + 945.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 394, + 828, + 1297, + 828, + 1297, + 1093, + 394, + 1093 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 296, + 1130, + 1405, + 1130, + 1405, + 1437, + 296, + 1437 + ], + "score": 0.947 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 13, + "poly": [ + 795, + 1161, + 891, + 1161, + 891, + 1189, + 795, + 1189 + ], + "score": 0.9, + "latex": "d = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 801, + 1378, + 822, + 1378, + 822, + 1401, + 801, + 1401 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 828.0, + 605.0, + 828.0, + 605.0, + 851.0, + 455.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 827.0, + 783.0, + 827.0, + 783.0, + 852.0, + 722.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 827.0, + 1006.0, + 827.0, + 1006.0, + 852.0, + 945.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 827.0, + 1229.0, + 827.0, + 1229.0, + 852.0, + 1167.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 863.0, + 439.0, + 863.0, + 439.0, + 882.0, + 410.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 844.0, + 488.0, + 844.0, + 488.0, + 895.0, + 458.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 851.0, + 663.0, + 851.0, + 663.0, + 869.0, + 632.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 859.0, + 882.0, + 859.0, + 882.0, + 875.0, + 862.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 879.0, + 664.0, + 879.0, + 664.0, + 897.0, + 632.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 871.0, + 1107.0, + 871.0, + 1107.0, + 890.0, + 1081.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 887.0, + 439.0, + 887.0, + 439.0, + 1001.0, + 392.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 892.0, + 660.0, + 892.0, + 660.0, + 1013.0, + 616.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 890.0, + 881.0, + 890.0, + 881.0, + 1018.0, + 846.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 893.0, + 1106.0, + 893.0, + 1106.0, + 1019.0, + 1063.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 936.0, + 437.0, + 936.0, + 437.0, + 954.0, + 410.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 933.0, + 662.0, + 933.0, + 662.0, + 951.0, + 633.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 941.0, + 1105.0, + 941.0, + 1105.0, + 957.0, + 1086.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 948.0, + 881.0, + 948.0, + 881.0, + 969.0, + 864.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 959.0, + 438.0, + 959.0, + 438.0, + 978.0, + 410.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 962.0, + 660.0, + 962.0, + 660.0, + 978.0, + 639.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 985.0, + 437.0, + 985.0, + 437.0, + 1001.0, + 412.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 988.0, + 661.0, + 988.0, + 661.0, + 1005.0, + 639.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 981.0, + 881.0, + 981.0, + 881.0, + 998.0, + 863.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 975.0, + 1106.0, + 975.0, + 1106.0, + 992.0, + 1087.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1008.0, + 438.0, + 1008.0, + 438.0, + 1028.0, + 410.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1016.0, + 661.0, + 1016.0, + 661.0, + 1033.0, + 639.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1012.0, + 881.0, + 1012.0, + 881.0, + 1029.0, + 864.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1008.0, + 1106.0, + 1008.0, + 1106.0, + 1026.0, + 1086.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1032.0, + 438.0, + 1032.0, + 438.0, + 1052.0, + 410.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1043.0, + 661.0, + 1043.0, + 661.0, + 1060.0, + 639.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1044.0, + 882.0, + 1044.0, + 882.0, + 1059.0, + 868.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1044.0, + 1105.0, + 1044.0, + 1105.0, + 1058.0, + 1091.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1060.0, + 460.0, + 1060.0, + 460.0, + 1079.0, + 433.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1062.0, + 564.0, + 1062.0, + 564.0, + 1097.0, + 499.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1063.0, + 621.0, + 1063.0, + 621.0, + 1077.0, + 608.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1063.0, + 675.0, + 1063.0, + 675.0, + 1077.0, + 663.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1060.0, + 809.0, + 1060.0, + 809.0, + 1093.0, + 694.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1061.0, + 850.0, + 1061.0, + 850.0, + 1079.0, + 814.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1062.0, + 899.0, + 1062.0, + 899.0, + 1077.0, + 885.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1060.0, + 941.0, + 1060.0, + 941.0, + 1079.0, + 909.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1060.0, + 1078.0, + 1060.0, + 1078.0, + 1093.0, + 943.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1061.0, + 1295.0, + 1061.0, + 1295.0, + 1093.0, + 1107.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 857.5, + 1063.0, + 857.5, + 1063.0, + 864.5, + 1046.0, + 864.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1128.0, + 1407.0, + 1128.0, + 1407.0, + 1165.0, + 293.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1161.0, + 794.0, + 1161.0, + 794.0, + 1193.0, + 295.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1193.0, + 892.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 1407.0, + 1191.0, + 1407.0, + 1226.0, + 294.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1220.0, + 1405.0, + 1220.0, + 1405.0, + 1257.0, + 293.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1254.0, + 1403.0, + 1254.0, + 1403.0, + 1286.0, + 295.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1310.0, + 1405.0, + 1310.0, + 1405.0, + 1348.0, + 294.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1345.0, + 1402.0, + 1345.0, + 1402.0, + 1376.0, + 295.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1371.0, + 800.0, + 1371.0, + 800.0, + 1408.0, + 293.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1371.0, + 1406.0, + 1371.0, + 1406.0, + 1408.0, + 823.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 974.0, + 1404.0, + 974.0, + 1440.0, + 294.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/HyMS8iRcK7/HyMS8iRcK7.md b/parse/train/HyMS8iRcK7/HyMS8iRcK7.md new file mode 100644 index 0000000000000000000000000000000000000000..3f578aef1b5d554991e4c54913e96a7290ee9000 --- /dev/null +++ b/parse/train/HyMS8iRcK7/HyMS8iRcK7.md @@ -0,0 +1,415 @@ +# SEQUENCE MODELLING WITH AUTO-ADDRESSINGAND RECURRENT MEMORY INTEGRATING NETWORKS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Processing sequential data with long term dependencies and learn complex transitions are two major challenges in many deep learning applications. In this paper, we introduce a novel architecture, the Auto-addressing and Recurrent Memory Integrating Network (ARMIN) to address these issues. The ARMIN explicitly stores previous hidden states and recurrently integrate useful past states into current time-step by an efficient memory addressing mechanism. Compared to existing memory networks, the ARMIN is more light-weight and inference-time efficient. Our network can be trained on small slices of long sequential data, and thus, can boost its training speed. Experiments on various tasks demonstrate the efficiency of the ARMIN architecture. Codes and models will be available. + +# 1 INTRODUCTION + +Recurrent neural networks, such as the Long Short-Term Memory (LSTM) (Hochreiter & Schmidhuber, 1997) and Gated Recurrent Unit (GRU) (Cho et al., 2014) have shown promising performance for processing sequential data. However, it’s known that RNNs suffer from gradient vanishing problem. Moreover, as pointed out by Rae et al. (2016), the number of parameters grows proportionally to the square of the size of the hidden units, which carry the historical information. Recent memorybased approaches exhibit potential to address these issues, by decoupling memory capacity from model parameters, and backpropagating the gradients through the memory. + +Neural Turing Machine (NTM) (Graves et al., 2014) first emerged as a recurrent model that incorporates external memory abilities. NTM maintains a memory matrix, and at every time-step, the network reads and writes (with erasing) to the memory matrix using certain soft-attentional mechanism, controlled by an LSTM that produces read and write vectors. NTM and its successor, the Differentiable Neural Computer(Graves et al., 2016), have shown success on some algorithmic tasks such as copying, priority sorting and some real-world tasks such as question answering. But one limitation of the NTM is that due to its smooth read and write mechanism, NTM has to do propagations on the entire memory, usually causing huge amount of memory consumption. To this end, Rae et al. (2016) proposes the Sparse Access Memory(SAM) network, by thresholding memory modifications to a sparse subset, i.e. all read and write operations are limited to several memory words. This allows memory and time efficient propagations while maintaining NTM’s performance. However, these external memory models have relatively complicated memory addressing mechanisms, making them suitable for only large-scale memory. Moreover, the RNN in these models plays a simple role of being a controller, but the gradient vanishing problem of RNN itself is not given attention. In contrast to these networks, our network uses a light-weight read mechanism to address a small external memory, and relieves the gradient vanishing problem by directly concatenating memory word to output. + +Inspired by prior memory models, efforts have been made to build a bridge between simple RNNs and complicated memory models. Kurach et al. (2015) propose the Neural Random-access Machines (NRAM) that can manipulate and dereference pointers to an external variable-size random-access memory. Danihelka et al. (2016) improve LSTM with ideas from Holographic Reduced Representations (Plate, 2003) that enables key-value storage of data. Grave et al. (2016) propose a method of augmenting LSTM by storing previous (hidden state, input word) pairs in memory and using the current hidden state as a query vector to recover historical input words. This method requires no backpropagation through memory and is well-suited to word-level language tasks. Grefenstette et al. (2015); Dyer et al. (2015); Joulin & Mikolov (2015) augment RNNs with a stack structure that works as a natural parsing tool, and use them to process algorithmic and nature language processing (NLP) tasks; nonetheless, the speed of stack-augmented RNNs is rather slow due to multiple push-pop operations at every time-step. Ke et al. (2018) propose the Sparse Attentive Backtracking (SAB) architecture, which recalls a few past hidden sates at every time-step and do “mental” backpropogations to the nearby hidden states with respect to the past hidden states. Gulcehre et al. (2017) propose the TARDIS network, which is an LSTM-resembled RNN that directly stores a fixed number of previous hidden states with learned key parameters for memory addressing. At every time-step, the network reads out one historical state $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ from the memory, and uses it, along with the input $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and last hidden state $h _ { t - 1 }$ , to produce new hidden state $h _ { t }$ , then overwrites $\boldsymbol { h } _ { t }$ to the location of $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ . TARDIS optionally uses the gumbel-softmax estimator(Maddison et al., 2016; Jang et al., 2016) to sample the location of $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ (we will explain this in section 2.1), so the whole network is differentiable. However, the TARDIS still involves some hand-crafted memory addressing method and its addressing mechanism requires a considerable amount of memory and time consumption. + +Inspired by the TARDIS, we introduce the ARMIN architecture, a more light-weight and memorybased architecture with a very simple memory addressing mechanism and in-depth modification of the existing LSTM structure. Concretely, our contributions are as follows: + +• We distill previous memory addressing methods and propose a simple yet effective memory addressing mechanism for the external memory, namely Auto-addressing, by encoding the information for memory addressing directly via the inputs $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and the hidden states $\boldsymbol { h } _ { t - 1 }$ . • We propose a novel recurrent cell that combines the gating advantages of LSTM and allows direct gradient backpropagation from output to memory. With only a single cell, it learns better representations than many hierarchical RNN structures. • We show in our ablation study that the ARMIN is robust to small iteration lengths when training long sequential data, which enables training with large batch sizes and boost its training speed. We further demonstrate in char-level language modelling task that the ARMIN obtain $4 0 \%$ training speed gain than LSTM while keeping similar performance and memory consumption. + +• We demonstrate competitive results on various tasks while keeping efficient time and memory consumption during training and inference time. + +# 2 BACKGROUND + +# 2.1 GUMBEL-SOFTMAX ESTIMATOR + +Categorical distribution is a natural choice for representing discrete structure in the world; however, it’s rarely used in neural networks due to its inability to backpropagate through samples(Jang et al., 2016). To this end, Maddison et al. (2016) and Jang et al. (2016) propose a continuous relaxation of categorical distribution and the corresponding gumbel-softmax gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample. Specifically, given a probability distribution $\pmb { p } = ( \pi _ { 1 } , \pi _ { 2 } , . . . , \pi _ { k } )$ over $k$ categories, the gumbel-softmax estimator produces an one-hot sample vector $\textbf { { y } }$ with its $i$ -th element calculated as follows: + +$$ +y _ { i } = { \frac { \exp ( ( \log ( \pi _ { i } ) + g _ { i } ) / \tau ) } { \sum _ { j = 1 } ^ { k } \exp ( ( \log ( \pi _ { j } ) + g _ { j } ) / \tau ) } } \quad { \mathrm { f o r ~ } } i = 1 , 2 , . . . , k , +$$ + +where $g _ { 1 } , . . . , g _ { k }$ are i.i.d samples drawn from Gumbel distribution(Gumbel, 1954): + +$$ +g _ { i } = - \log ( - \log ( u _ { i } ) ) \ , u _ { i } \sim \mathrm { U n i f o r m } ( 0 , 1 ) , +$$ + +and $\tau$ is the temperature parameter. In practice, we usually start at a high temperature and anneal to a small but non-zero temperature (Jang et al., 2016). + +# 3 AUTO-ADDRESSING AND RECURRENT MEMORY INTEGRATING NETWORK + +In this section, we describe the structure of the ARMIN network as shown in Figure 1. It consists of a recurrent cell and a small external memory that stores historical hidden states. While processing + +![](images/eab2a73d816507916b458d579eb538fc7d49d69aba42b4a31414c981c7c3217e.jpg) +Figure 1: The ARMIN structure. At each time-step, the ARMIN performs read operation, cell processing and write operation in chronological order: (a) It reads out a historical hidden state $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ from memory with an one-hot read vector produced via passing $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $\boldsymbol { h } _ { t - 1 }$ to a fully connected layer followed by a gumbel-softmax function. (b) The ARMIN cell receives $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ , $h _ { t - 1 }$ and $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ as inputs and outputs $\mathbf { } _ { o _ { t } }$ and $\boldsymbol { h } _ { t }$ . $\mathbf { } _ { o _ { t } }$ is passed to output layers, and $h _ { t }$ is passed to next time-step. (c) $\boldsymbol { h } _ { t }$ is written to the previous location of $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ . + +sequential data, the ARMIN performs reading from memory, cell processing and writing to memory operations in chronological order during each time-step. In the following subsections, we first explain the structure of the recurrent cell, and then discuss the read and write operations. + +# 3.1 THE RECURRENT CELL OF ARMIN + +Inspired by classical LSTM structure (please refer to APPENDIX A for details in RNN and LSTM structures), we propose a novel memory-augmented recurrent cell structure, namely the ARMIN cell. At every time-step, it takes in an input $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ , the last hidden state $\boldsymbol { h } _ { t - 1 }$ and a recovered historical hidden state $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ chosen by a read operation, and produces an output vector $\mathbf { } _ { o _ { t } }$ and the new hidden state $\boldsymbol { h } _ { t }$ . The computation process is as follows: + +$$ +\left\{ { g } _ { t } ^ { h } \atop { g } _ { t } ^ { r } \right\} = \left\{ { \sigma } \atop { \sigma } \right\} W _ { i g } [ x _ { t } , h _ { t - 1 } , r _ { t } ] + b _ { i g } , +$$ + +$$ +h _ { t - 1 } ^ { g } = { \pmb g } _ { t } ^ { h } \circ { \pmb h } _ { t - 1 } , +$$ + +$$ +\begin{array} { r } { \pmb { r } _ { t } ^ { g } = \pmb { g } _ { t } ^ { r } \circ \pmb { r } _ { t } , } \end{array} +$$ + +$$ +\left\{ \begin{array} { c } { i _ { t } } \\ { f _ { t } } \\ { g _ { t } } \\ { o _ { t } ^ { h } } \\ { o _ { t } ^ { r } } \end{array} \right\} = \left\{ \begin{array} { c } { \sigma } \\ { \sigma } \\ { \operatorname { t a n h } } \\ { \sigma } \\ { \sigma } \\ { \sigma } \end{array} \right\} W _ { g o } [ \boldsymbol { x } _ { t } , h _ { t - 1 } ^ { g } , \boldsymbol { r } _ { t } ^ { g } ] + b _ { g o } , +$$ + +$$ +\pmb { h } _ { t } = \pmb { f } _ { t } \circ \pmb { h } _ { t - 1 } + i _ { t } \circ \pmb { g } _ { t } , +$$ + +$$ +\begin{array} { r } { \pmb { o } _ { t } = \left[ \pmb { o } _ { t } ^ { h } \circ \operatorname { t a n h } ( \pmb { h } _ { t } ) , \pmb { o } _ { t } ^ { r } \circ \operatorname { t a n h } ( \pmb { r } _ { t } ) \right] . } \end{array} +$$ + +where $h _ { t - 1 } , h _ { t } \in \mathbb { R } ^ { d _ { h } } , { \boldsymbol { r } } _ { t } \in \mathbb { R } ^ { d _ { r } } , { \boldsymbol { x } } _ { t } \in \mathbb { R } ^ { d _ { i } }$ , and $\begin{array} { r c l } { W _ { i g } } & { \in } & { \mathbb { R } ^ { 2 d _ { h } \times ( d _ { i } + d _ { h } + d _ { r } ) } , W _ { g o } } & { \in } \end{array}$ $\begin{array} { r } { \mathbb { R } ^ { ( 4 d _ { h } + d _ { r } ) \times ( d _ { i } + d _ { h } + d _ { r } ) } , \pmb { b } _ { i g } \ \in \ \mathbb { R } ^ { d _ { h } + d _ { r } } , \pmb { b } _ { g o } \ \in \ \mathbb { R } ^ { 4 d _ { h } + d _ { r } } , \pmb { o } _ { t } \ \in \ \mathbb { R } ^ { d _ { h } + d _ { r } } } \end{array}$ . We refer to $d _ { h }$ as the hidden size of the recurrent cell of ARMIN. Usually we have $d _ { h } = d _ { r }$ and allocate equal number of weight parameters for $\boldsymbol { h } _ { t }$ and $\mathbf { \nabla } _ { \mathbf { \boldsymbol { r } } _ { t } }$ . + +In equation $3 \sim 5$ , two gates are calculated to control the information flow for $h _ { t - 1 }$ and $r _ { t }$ respectively, generating gated hidden state $h _ { t - 1 } ^ { g }$ and historical state $\boldsymbol { r } _ { t } ^ { g }$ ; using this method, we can filter out the irrelevant information for the current time-step. Then as shown in equation 6, we compute the input gate $\mathbf { \delta } _ { i _ { t } }$ , forget gate $\pmb { f } _ { t }$ , cell state $\mathbf { \pmb { g } } _ { t }$ and output gate $\mathbf { \sigma } _ { o _ { t } ^ { h } } ^ { h }$ for the new hidden state just like in classical LSTM structure. Additionally, an output gate ${ \mathbf { } } _ { o _ { t } ^ { r } }$ for historical state $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ is computed. Next in equation 7, we compute new hidden state $h _ { t }$ that is the sum of $h _ { t - 1 }$ and cell state $\mathbf { \sigma } _ { \mathbf { \sigma } _ { \mathbf { \sigma } _ { \mathbf { \lambda } } } } \mathbf { \sigma } _ { \mathbf { \sigma } _ { \mathbf { \lambda } } } \mathbf { \sigma } _ { \mathbf { \lambda } _ { \mathbf { \lambda } } } \mathbf { \sigma } _ { \mathbf { \lambda } _ { \mathbf { \lambda } } } \mathbf { \sigma } _ { \mathbf { \lambda } _ { \mathbf { \lambda } } } \mathbf { \sigma } _ { \mathbf { \lambda } _ { \mathbf { \lambda } } } \mathbf { \sigma } _ { \mathbf { \lambda } _ { \mathbf { \lambda } } } \mathbf { \sigma } _ { \mathbf { \lambda } _ { \lambda } }$ , leveraged by forget gate $\pmb { f } _ { t }$ and input gate $\dot { \mathbf { \delta } } _ { \mathbf { \delta } }$ . Finally in equation 8, we calculate the output of this timestep, which is the concatenation of the gated contents from $h _ { t }$ and $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ . Using this method, we can selectively backpropagate gradients from the output to the historical state $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ . We involve $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ (or $\boldsymbol { r } _ { t } ^ { g }$ ) in all gates’ computations. + +Intuitively, the $\boldsymbol { h } _ { t - 1 }$ acts as the old working memory, and $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ is treated as the long-term memory. The cell processes them with the input $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ to generate the new working memory $\boldsymbol { h } _ { t }$ and the output $\mathbf { } _ { o _ { t } }$ . More specifically, each $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ is a summary of historical hidden states selected by auto-addressing mechanism. The ARMIN cell learns to recurrently integrate the summary of long-term information from $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ into the working memory $\boldsymbol { h } _ { t }$ . We will demonstrate the efficiency of recurrent memory integration in language modelling tasks. + +Unlike the LSTM that passes a tuple of $\left( \boldsymbol { h } _ { t } , \boldsymbol { c } _ { t } \right)$ to the next recurrence and outputs $\boldsymbol { h } _ { t }$ at the same time, the ARMIN only passes one hidden state $h _ { t }$ and separately uses $\mathbf { } _ { o _ { t } }$ as output, which has 2 benefits: one is that it takes lower memory consumption to store one hidden state rather than two, and the other is we can decouple the information needed for the output $\mathbf { } _ \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf \mathbf { } \mathbf \mathbf { } \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf { } \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf \mathbf $ , from the information (i.e. $\boldsymbol { h } _ { t }$ ) that is needed for memorizing and being used at later time-steps. Furthermore, we list 3 important differences between the ARMIN cell and TARDIS cell, please refer to APPENDIX B.1. + +# 3.2 READ OPERATION WITH AUTO-ADDRESSING + +The ARMIN maintains a memory matrix $M \in \mathbb { R } ^ { n _ { m e m } \times d _ { h } }$ , where the constant $n _ { m e m }$ denotes the number of memory slots. At each recurrence, the ARMIN chooses a historical state $r _ { t }$ from memory according to the information in $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $\boldsymbol { h } _ { t - 1 }$ ,which is formulated as follows: + +$$ +\begin{array} { c } { { \displaystyle s _ { t } = \mathrm { g u m b e l - s o f t m a x } \big ( W _ { s } [ \ : x _ { t } \ : , h _ { t - 1 } \ : ] + b _ { s } \big ) \ : , } } \\ { { { } } } \\ { { \displaystyle r _ { t } = \sum _ { i = 0 } ^ { n - 1 } s _ { t } ( i ) M ( i , : ) \ : . } } \end{array} +$$ + +where $W _ { s } \ \in \ \mathbb { R } ^ { n _ { m e m } \times ( d _ { i } + d _ { h } ) }$ , $b _ { s } \in \mathbb { R } ^ { n _ { m e m } }$ , $\mathbf { \Delta } _ { \mathbf { \mathcal { S } } _ { t } }$ is a one-hot vector sampled by gumbel-softmax function, $s _ { t } ( i )$ denotes the $i$ -th element of $\mathbf { \boldsymbol { s } } _ { t }$ , $M ( i , : )$ denotes the $i$ -th row of $M$ . + +As opposed to previous memory networks such as Santoro et al. (2016) and Gulcehre et al. (2017), we don’t use any extra usage vectors and learnable key parameters that manually encode memory addressing information to assist memory addressing. As our ablation study (Appendix E) shows, the hidden state $\boldsymbol { h } _ { t - 1 }$ is sufficient to encode the historical memory accessing information. Furthermore, the TARDIS addressing mechanism requires concatenating $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $\boldsymbol { h } _ { t }$ to every memory cell, which usually causes large extra memory and time consumption in practical scenes. Please refer to APPENDIX B.2 for more precise comparison regarding addressing mechanism of the ARMIN and TARDIS networks. + +# 3.3 WRITE OPERATION + +After the reading and cell processing stages, the ARMIN writes the new hidden state $\boldsymbol { h } _ { t }$ to the memory $M$ . Following the TARDIS network (Gulcehre et al., 2017), we simply overwrite $h _ { t }$ to the memory slot where we just read out the $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ (for conditions where $d _ { h }$ is not equal to $d _ { r }$ , we first use a linear layer to transform $\boldsymbol { h } _ { t }$ from $d _ { h }$ dimension to $d _ { r }$ dimension). But at the initial time-steps, we write the hidden states to the empty memory slots, until all empty slots are filled with historical states. In this way, we can maximally preserve useful historical information, because the ARMIN cell can learn to copy the useful information from $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ to $\boldsymbol { h } _ { t }$ , and then write $h _ { t }$ to the previous location of $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ . The read/write mechanism can be viewed as a form of a skip-connection of hidden states. By training the whole network, the ARMIN can easily learn long-term dependencies via direct access to historical hidden states. From this point of view, our network is similar to the Skip RNN proposed by Campos Camunez et al. (2018), which differs from our network in that they use a binary state update gate to select whether the state of the RNN will be updated or copied from the previous time-step. + +# 4 EXPERIMENTS + +We evaluate our model on algorithmic tasks, character-level language modelling task and temporal action detection and proposal task. We compare our network with NTM (with an LSTM controller), TARDIS and vanilla LSTM networks. We also directly compare the auto-addressing against the TARDIS addressing mechanism, by replacing auto-addressing with the TARDIS addressing method in ARMIN. We refer to this network as ARMIN $^ +$ TARDIS-addr in our experiments. Please refer to APPENDIX C for some universal setups used in all our experiments below. + +# 4.1 ALGORITHMIC TASKS + +Along with the NTM, Graves et al. (2014) introduced a set of synthetic algorithm tasks which consists of copy, repeat copy, associative recall, N-gram and priority sort tasks. Here we use four out of five of these tasks (excluding the N-gram task) to examine if our network can choose correct time-steps from the past and effectively make use of them. Please refer to APPENDIX C.1 for more explanations and hyperparameter setups with respect to these tasks. Note that we have a strong LSTM baseline with about 4 times larger parameter count than memory networks, and the training of our copy and priority sort tasks are more challenging than the original tasks in Graves et al. (2014), for having more than 2 times longer input and target sequences. The models are optimized to minimize the averaged binary cross-entropy loss on the target multi-bit binary sequences. Following Gulcehre et al. (2017); Campos Camunez et al. (2018), in all tasks, we consider a task solved if the validation loss is at least two orders of magnitude below the initial cost which is around 0.70, i.e. the validation loss (the validation set is generated randomly) converges to less than 0.01 (with less than $3 0 \%$ in 10 consecutive validation of sharp losses that are higher than 0.01 ). In our evaluation, we are interested in if a model can successfully solve the task in 100k iterations and the elapsed training time and iterations till the model succeeds on the task. For all networks we ensure the similar total parameter counts. All important hyperparameters are shown in APPENDIX C.1. We run all experiments with batch size of 1 under 2.8 GHz Intel Core i7 CPU and report the wall clock time and the number of iterations for solved tasks of each model. For models that fail to converge to less than 0.01 loss in $1 0 0 \mathrm { k }$ iterations, we report the average loss of the final 10 validations (denoted in underlines) and the elapsed training time for 100k iterations. The results are shown in Table 1. The training curves for each task are shown in APPENDIX D. + +The results show that the ARMIN can solve 3 out of 4 tasks in a fast converge speed. By comparing the ARMIN with NTM, we observe that the NTM is able to almost solve all of the tasks, but with a much slower training speed compared to the ARMIN, for example, the training time of NTM is 5 times larger than ARMIN to solve the copy task. In fact, the training speed of ARMIN for each iteration is about $3 \sim 4$ times faster than the NTM’s speed. We observe that the TARDIS and ARMIN $\dotplus$ TARDIS-addr fail to solve 3 out 4 tasks in the given $1 0 0 \mathrm { k }$ iterations. By comparing them with the ARMIN, we confirm the efficiency of the auto-addressing of ARMIN, for bringing a fast converge speed and short training time. We also find that the LSTM, when given enough parameter counts (4 times larger than memory networks), is able to solve the priority sort task quickly and converge to a better performance than TARDIS and ARMIN $^ +$ TARDIS in other tasks. We believe there are 2 possible reasons for the poor performance of the TARDIS addressing mechanism: a) The tanh activation (see equation 21 in APPENDIX B.2) might cause the gradient vanishing problem when it cooperates with the gumbel-softmax function. b) The multiple inputs of $\pmb { x } _ { t } , \pmb { h } _ { t } , \pmb { M } _ { t } [ i ] , \pmb { u } _ { t }$ in equation 21 cause the network hard to initialize to keep equal input and output variances for the linear layer and might also cause gradient dispersion. The auto-addressing has neither of these two issues and can better integrately backpropogate gradients, therefore it can perform well and quickly on these tasks. However, we would like to point out 2 limitations of the auto-addressing: a) The associative recall task shows the auto-addressing might have problems in establishing indirect references among memory data. b) The auto-addressing is relatively hard to generalize to unseen longer sequence in tasks such as copy and repeat copy due to its simplicity, whereas the NTM is able to do that. We leave these 2 issues for future work. + +Table 1: The average elapsed time(min) and iterations ${ \mathrm { ( k ) } }$ of different networks till task solved in given $1 0 0 \mathrm { k }$ iterations. The wall clock training time for 100k iterations and the average loss of final 10 validations (denoted in underlines) are shown for the unsolved tasks. + +
CopyRepeat CopyAssociative RecallPriority Sort
timeiter.timeiter.timeiter.timeiter.
LSTM660.359450.0160.3253237.0
NTM3212.41710.01419.63600.012
TARDIS1570.451970.1663422660.33013362.6
ARMIN+TARDIS-addr1520.410950.297640.33615171.4
ARMIN67.63667.8330.0523733.4
+ +Table 2: Bits-per-character on Penn Treebank and enwik8 test set. + +
ModelPTBenwik8
BPCParamsBPCParams
LSTM1.36一1.45
LSTM+Zoneout (Krueger et al., 2016)1.2711
LSTM+Layer Norm(1000 units)11.2674.26M1一
LSTM+Layer Norm+Zoneout(1024 units)²1.244.79M一
HM-LSTM+Layer Norm (Chung et al.,2016)1.2411.3235M
HyperLSTM+Layer Norm(Ha et al., 2016)1.21914.41M1.34026.5M
NASCell (Zoph & Le,2016)1.21416.28M
IndRNN (21 layers) (Li et al.,2018)1.2111
Recurrent Highway Hypernetwork (Suarez, 2017)1.19一 15.5M11
Fast-Slow LSTM (Mujika et al., 2017)1.1907.2M1.277一 27M
NTM(800 units, nhead = 1, Pdropout=0.6)
TARDIS(paper)1.5358.28M1
1.25一
TARDIS(1000 units,Pdropout=0.6)1.2689.2M一
ARMIN+TARDIS-addr(800 units,Pdropout=0.6)1.22310.2M
ARMIN (500 units, Pdropout=0.4) (Ours)1.2364.03M
ARMIN (800 units, Pdropout=0.6) (Ours)1.2029.8M
ARMIN(800 units,2 layer) (Ours)111.33121.6M
+ +# 4.2 CHARACTER-LEVEL LANGUAGE MODELLING + +The character-level language modelling task consists of predicting the probability distribution of the next character given all the previous ones, and we use the widely used Penn Treebank and Hutter Prize Wikipedia (also known as enwik8) datasets for this task. We compare our network with vanilla LSTMs and other memory networks on Penn Treebank, and compare with some popular RNN variants on both datasets. We ensure similar parameter counts and hyperparameters for all memory networks. Following prior works, we apply truncated backpropagation through time(TBPTT) (Rumelhart et al., 1986; Elman, 1990) to approximate the gradients: at each iteration, the network predict the next 150 characters, and the hidden state $\boldsymbol { h } _ { t }$ and memory state $M$ are passed to the next iteration. The gradients are truncated between different iterations. Please refer to APPENDIX C.2 for more experimental details. + +The results are shown in table 2. On Penn Treebank dataset, our best performing network achieves competitive 1.202 BPC on Penn Treebank dataset, which is the best single cell performance that we are aware of, with relatively small parameter count. By comparing TARDIS and ARMIN $^ +$ TARDISaddr, we observe about 4 points of improvement, which shows the efficiency of the ARMIN cell. By comparing ARMIN $+$ TARDIS-addr, we observe a further improvement of around 2 points, which shows the efficiency of the auto-addressing mechanism. We also find that the same architecture of NTM that performs well in algorithmic tasks fails to converge to a lower BPC than vanilla LSTM, even with various regularizations we add to NTM. + +The results on Penn Treebank show our single ARMIN cell learns better representations than many hierarchical RNN structures, such as the HM-LSTM, 2-Layer HyperLSTM and 21 layer IndRNN. Our network is outperformed by Recurrent Highway Hypernetwork and Fast-slow LSTM which are both multi-scale and deep transition RNNs and are state-of-the-art results without any auxiliary technics on this dataset. By saying “better representations”, we refer to the concatenation of the gated contents from $\boldsymbol { h } _ { t }$ and $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ as in equation 8. For example, if we remove the gated contents of $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ from $\mathbf { } _ { o _ { t } }$ , the ARMIN undergoes a BPC performance drop from 1.202 to 1.220, which is still better than the best performing BPC of 1.24 of the LSTM. We believe the rest of the performance gain comes from the recurrent memory integration of the ARMIN cell, which also favors a paradigm of deep transition, as is shown by the success of the deep transition RNNs on this task. For more ablation study regarding the auto-addressing mechanism and the ARMIN cell please refer to APPENDIX E (where we show the robustness of ARMIN to small TBPTT length). + +The result on enwik8 demonstrates a simple 2-layer ARMIN can achieve competitive BPC performance of 1.33, with less parameter count compared to the HyperLSTM and HM-LSTM. By constructing deeper ARMIN network or even combining with other multi-scale and hierarchical RNN architectures, we believe the performance can be further improved. + +# 4.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14 + +In this subsection, we evaluate the ARMIN on a more complicated real-world task concerning video analysis, i.e. the temporal action detection and proposals, which consists of taking an input video and producing a set of temporal intervals that are likely to contain human actions (Buch et al., 2017). We use the Single-Stream Temporal Action Proposal(SST) (Buch et al., 2017) as our framework, and evaluate the model on the THUMOS’ 14 dataset (Jiang et al., 2014), which contains $^ { 2 0 + }$ hours of video with 200 train/validation and 213 test untrimmed video sequences. The SST consists of a GRU and a fully connected output layer. In our experiment, we replace the GRU encoder in SST with a single ARMIN cell, and we call the modified network MA-SST. For fairness, we also implement a modified SST with LSTM encoder and compare them with the original SST in Buch et al. (2017). Please refer to APPENDIX C.3 for more experimental details. + +Table 3: Comparison of proposal generation performance in terms of recall at 1000 proposals. We choose $\mathrm { \ t I o U { = } 0 } . 6$ and 0.8 to compare for consistency with Buch et al. (2017). Our MA-SST achieves highest recall at $\mathrm { \ t I o U { = } } 0 . 8$ . +Table 4: Comparison for performance and cost of different setups on Penn Treebank dataset. + +
ModelARMINLSTM
Setup1212
Hidden sizenparam(M)nmemTtruncbatch sizeMemory(GB)Speed(chars/s)BPC5005501k1k
4.0254.814.794.791
101
5050100150
384300128128
3.493.562.363.27
98k75k71k70k
1.2381.2261.271.24
+ +
ModeltIoU=0.6tIoU=0.8
SST (Paper)0.9200.672
SST (We implement)0.8970.732
MA-SST(Ours)0.9110.759
+ +The average recall under different proposal numbers and tIoUs3 are depicted in figure 2. MA-SST outperforms the original SST and our SST simultaneously. To the best of our knowledge, MA-SST achieves highest performance of recall at around $\mathrm { \ t I o U { = } } 0 . 8$ (see table 3), even if other competitive non-RNN networks (Escorcia et al., 2016; Caba Heilbron et al., 2016; Gao et al., 2017; Guo et al., 2018) are taken into account. The MA-SST tends to precisely generate proposals that has a high overlapping area with the ground truth action intervals, leading to high recall performance at high tIoUs. We believe the MA-SST has better performance than the SST mainly because of the direct access to historical states from which it can better locate the starting frame of an action. + +![](images/80e05beebf3cb64569df7a0d731d988352907332c1e2de7be2c1637e50fd4897.jpg) +Figure 2: Comparison of proposal generation performance. For data generation, we use the code offered in Buch et al. (2017) for consistency. (Left) The average number of proposals v.s. average recall for $\mathrm { t I o U } \geq 0 . 5$ . The MA-SST outperforms our SST when average number of proposals is less than 500, and outperforms the original SST when average number of proposals is more than 500. (Right) Different tIoU v.s. recall of 1000 average proposals. The MA-SST strongly outperforms the original SST and outperform our SST by a maximal margin of $2 . 7 \%$ . + +# 5 TOWARDS MORE LIGHT-WEIGHT RECURRENT MEMORY NETWORKS + +It’s known that the speed and computer memory consumption have important influences on the practicability of an RNN, especially when external memory is involved. We have shown in some of the algorithmic tasks that the ARMIN can be trained 3 4 times faster than the NTM in terms of wall clock time. Next we conduct a more thorough comparison experiment among the vanilla LSTM and memory networks under different hidden sizes. The experiment is based on the character-level language modelling task on Penn Treebank dataset, without the loss of generality. For a fair comparison, we implement all networks with near-optimal implementations using Pytorch, an efficient deep-learning framework. We use the aforementioned experiment setup, and keep all memory matrices the same size, i.e. $2 0 \times d _ { h }$ . We run the experiment using single-precision floating point calculations under a Titan XP GPU that has 12GB memory space. The results are depicted in figure 3. From the results we observe 2 important phenomena: a) The ARMIN consistently outperform other memory networks shown in the graph in terms of running speed both at training and inference stages, and the main contribution to this comes from the simple auto-addressing mechanism of the ARMIN. Moreover, at inference stage, we can replace the memory matrix with a list of discrete memory slots, and update memory by simply replacing the old hidden states with the new ones, furthermore, we can replace the gumbel-softmax function with argmax. Using these methods, ARMIN’s inference speed obtains significant improvement than in training stage. b) The ARMIN outperforms TARDIS and ARMIN $\dotplus$ TARDIS-addr in terms of training memory consumption, mainly because the TARDIS addressing mechanism requires concatenating $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $\boldsymbol { h } _ { t - 1 }$ to every memory cell for parallelization. At inference stage, we observe TARDIS have smaller memory consumption than ARMIN and ARMIN $+ { \bf \Phi }$ TARDIS-addr, mainly because TARDIS has smaller parameter counts under the same hidden size. However, we would like to point out that under the same parameter counts the ARMIN usually outperforms TARDIS as we show in algorithm and language modelling tasks. We also observe the NTM has the largest memory consumption at inference time, mainly because its complex addressing mechanism. + +We have shown in the ablation study (see APPENDIX E) that our network is robust to small iteration length when trained with TBPTT. We realize that we can use this feature to enable large batch size, and boost ARMIN’s training speed while keeping better performance than LSTM. To validate this benefit, we conduct an experiment to compare the performance and cost of ARMIN against our best performing LSTM. We also provide the test result of the same LSTM under TBPTT length $T _ { t r u n c } = 1 0 0$ for better comparing. The hyperparameter setups and results are shown in Table 4. + +By comparing ARMIN setup 1 with LSTM setup 2, we observe that with only $6 . 7 \%$ more memory consumption, we obtain $4 0 \%$ training speed gain while keeping a slightly better BPC performance and $1 6 \%$ less parameter count; by comparing ARMIN setup 2 with LSTM setup 2, we observe that with only $8 . 8 \%$ more memory consumption, we obtain about 1.5 points of BPC performance gain and $7 . 1 \%$ training speed gain under similar parameter count. + +![](images/fb603389200962c80fad3cc21f0db454db3b03532b7db183bfb4f4b91ccf1b51.jpg) +Figure 3: The running speed and memory consumption at the training and inference stages $( n _ { m e m } = 2 0 )$ ). The solid blocks shows the memory consumption and the curves denote the running speed in characters/s. (a) shows the training stage, and (b) shows the inference stage (note that ARMIN $+$ TARDIS-addr has basically the same memory consumption with ARMIN, so it’s not shown in the graph). + +# 6 CONCLUSION + +In this paper, we have introduced the ARMIN, a light-weight and memory-augmented RNN architecture with a novel ARMIN cell. The ARMIN incorporates an efficient external memory with the light-weight auto-addressing mechanism. We demonstrate competitive performance of ARMIN in various tasks, and shows the generality of our model. Our ablation study suggests the efficacy of our external memory and addressing mechanism, and notably, our network is robust to shortlength TBPTT which enables using large batch size to speed up the training and further increase performance on sequence modelling tasks. Further research may lead to the efficient hierarchical and multi-scale structures of the ARMIN and its successful applications in Seq2Seq models and complicated reasoning tasks. + +# REFERENCES + +Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. + +Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst: Single-stream temporal action proposals. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6373–6382. IEEE, 2017. + +Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Fast temporal activity proposals for efficient detection of human actions in untrimmed videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1914–1923, 2016. + +Victor Campos Camunez, Brendan Jou, Xavier Giro Nieto, Jordi Torres Vi ´ nals, and Shih-Fu Chang. ˜ Skip rnn: learning to skip state updates in recurrent neural networks. In Sixth International Conference on Learning Representations: Monday April 30-Thursday May 03, 2018, Vancouver Convention Center, Vancouver:[proceedings], pp. 1–17, 2018. + +Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨ ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. + +Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Gated feedback recurrent neural networks. In International Conference on Machine Learning, pp. 2067–2075, 2015. + +Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. Hierarchical multiscale recurrent neural networks. arXiv preprint arXiv:1609.01704, 2016. + +Ivo Danihelka, Greg Wayne, Benigno Uria, Nal Kalchbrenner, and Alex Graves. Associative long short-term memory. In Proceedings of the 33rd International Conference on International Conference on Machine Learning-Volume 48, pp. 1986–1994. JMLR. org, 2016. + +Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A Smith. Transition-based dependency parsing with stack long short-term memory. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), volume 1, pp. 334–343, 2015. + +Jeffrey L Elman. Finding structure in time. Cognitive science, 14(2):179–211, 1990. + +Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep action proposals for action understanding. In European Conference on Computer Vision, pp. 768–784. Springer, 2016. + +Jiyang Gao, Zhenheng Yang, Chen Sun, Kan Chen, and Ram Nevatia. Turn tap: Temporal unit regression network for temporal action proposals. In Computer Vision (ICCV), 2017 IEEE International Conference on, pp. 3648–3656. IEEE, 2017. + +Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a continuous cache. arXiv preprint arXiv:1612.04426, 2016. + +Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. + +Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538 (7626):471, 2016. + +Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp. 1828–1836, 2015. + +Caglar Gulcehre, Sarath Chandar, and Yoshua Bengio. Memory augmented neural networks with wormhole connections. arXiv preprint arXiv:1701.08718, 2017. + +Emil Julius Gumbel. Statistical theory of extreme values and some practical applications: a series of lectures. Number 33. US Govt. Print. Office, 1954. + +Dashan Guo, Wei Li, and Xiangzhong Fang. Fully convolutional network for multiscale temporal action proposals. IEEE Transactions on Multimedia, 2018. + +David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. + +Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. + +YG Jiang, Jingen Liu, A Roshan Zamir, G Toderici, I Laptev, Mubarak Shah, and Rahul Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. + +Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In Advances in neural information processing systems, pp. 190–198, 2015. + +Nan Rosemary Ke, Anirudh Goyal, Olexa Bilaniuk, Jonathan Binas, Michael C Mozer, Chris Pal, and Yoshua Bengio. Sparse attentive backtracking: Temporal creditassignment through reminding. arXiv preprint arXiv:1809.03702, 2018. + +Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +David Krueger, Tegan Maharaj, Janos Kramar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary Ke, Anirudh Goyal, Yoshua Bengio, Aaron Courville, and Christopher Pal. Zoneout: Regularizing rnns by randomly preserving hidden activations. 2016. + +Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. arXiv preprint arXiv:1511.06392, 2015. + +Shuai Li, Wanqing Li, Chris Cook, Ce Zhu, and Yanbo Gao. Independently recurrent neural network (indrnn): Building a longer and deeper rnn. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5457–5466, 2018. + +Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. + +Toma´s Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and Jan Cer- ˇ nocky. Subword language modeling with neural networks. preprint (http://www. fit. vutbr. cz/imikolov/rnnlm/char. pdf), 8, 2012. + +Asier Mujika, Florian Meier, and Angelika Steger. Fast-slow recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 5915–5924, 2017. + +Tony A Plate. Holographic reduced representation: Distributed representation for cognitive structures. 2003. + +Jack Rae, Jonathan J Hunt, Ivo Danihelka, Timothy Harley, Andrew W Senior, Gregory Wayne, Alex Graves, and Tim Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes. In Advances in Neural Information Processing Systems, pp. 3621–3629, 2016. + +David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by backpropagating errors. nature, 323(6088):533, 1986. + +Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Oneshot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016. + +Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958, 2014. + +Joseph Suarez. Language modeling with recurrent highway hypernetworks. In Advances in Neural Information Processing Systems, pp. 3267–3276, 2017. + +Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26– 31, 2012. + +Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pp. 4489–4497, 2015. + +Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. + +# APPENDIX A RNN AND LONG SHORT-TERM MEMORY + +A recurrent neural network (RNN) is a class of neural network that recurrently processes a sequence of inputs $\{ \pmb { x } _ { 1 } , \pmb { x } _ { 2 } , . . . , \pmb { x } _ { T } \}$ , and returns a sequence of outputs $\left\{ y _ { 1 } , y _ { 2 } , . . . , y _ { T } \right\}$ where $\pmb { x } _ { t } \in \mathbb { R } ^ { d _ { i } }$ , $\pmb { y } _ { t } \in \mathbb { R } ^ { d _ { o } }$ , and $d _ { i } , d _ { o }$ is the input size and output size, respectively. In a vanilla RNN, $\mathbf { \nabla } _ { \mathbf { \boldsymbol { y } } _ { i } }$ is given by following equations: + +$$ +\begin{array} { r } { \pmb { h } _ { t } = \sigma ( \pmb { U } \pmb { x } _ { t } + \pmb { W } \pmb { h } _ { t - 1 } + \pmb { b } ) , } \end{array} +$$ + +$$ +\pmb { y } _ { t } = \pmb { V } \pmb { h } _ { t } + \pmb { c } , +$$ + +where the hidden state $\pmb { h } _ { t } ~ \in ~ \mathbb { R } ^ { d _ { h } }$ is passed to the next recurrence. $d _ { h }$ is the hidden size, and $U , W , V , b , c$ are learnable parameters. + +The LSTM (Hochreiter & Schmidhuber, 1997) is designed to overcome the gradient vanishing and exploding problem in vanilla RNN and learn long-term dependencies, by passing a tuple of $\left( { { h _ { t } } , { c _ { t } } } \right)$ to the next recurrence. Its computation process is defined as follows: + +$$ +\left\{ \begin{array} { c } { i _ { t } } \\ { f _ { t } } \\ { g _ { t } } \\ { o _ { t } } \end{array} \right\} = \left\{ \begin{array} { c } { \sigma } \\ { \sigma } \\ { \operatorname { t a n h } } \\ { \sigma } \end{array} \right\} W _ { 0 } [ \ : x _ { t } , h _ { t - 1 } \ : ] + b _ { 0 } , +$$ + +$$ +\begin{array} { r } { \pmb { c } _ { t } = \pmb { f } _ { t } \circ \pmb { c } _ { t - 1 } + \pmb { i } _ { t } \circ \pmb { g } _ { t } , } \\ { \pmb { h } _ { t } = \pmb { o } _ { t } \circ \operatorname { t a n h } ( \pmb { c } _ { t } ) , } \end{array} +$$ + +where $\circ$ is the element-wise product, $W _ { 0 } \in \mathbb { R } ^ { 4 d _ { h } \times ( d _ { i } + d _ { h } ) }$ , $b _ { 0 } \in \mathbb { R } ^ { 4 d _ { h } }$ , $i _ { t } , f _ { t } , g _ { t } o _ { t }$ are the input gate, forget gate, cell state, output gate at time $t$ , respectively, and they control the information flow in the LSTM. + +# APPENDIX B DIFFERENCES BETWEEN TARDIS AND ARMIN + +The major differences between ARMIN and TARDIS networks are the recurrent cell computation and read operation. We show these differences respectively in subsections below: + +# B.1 CELL COMPUTATIONS + +The cell computation process of TARDIS is as follow: + +$$ +\left\{ \begin{array} { l } { i _ { t } } \\ { f _ { t } } \\ { o _ { t } } \end{array} \right\} = \left\{ \begin{array} { l } { \sigma } \\ { \sigma } \\ { \sigma } \end{array} \right\} W [ \ : x _ { t } , h _ { t - 1 } , r _ { t } \ : ] + b \ : , +$$ + +$$ +\left\{ \begin{array} { l } \alpha _ { t } \right\} = \left\{ \begin{array} { l l } { \mathrm { g u m b e l - s i g m o i d } } \\ { \mathrm { g u m b e l - s i g m o i d } } \end{array} \right\} \left\{ \begin{array} { l } { \pmb { w } ^ { \alpha \top } } \\ { \pmb { w } ^ { \beta \top } } \end{array} \right\} \left[ \begin{array} { l } { \pmb { x } _ { t } , \pmb { h } _ { t - 1 } , \pmb { r } _ { t } } \end{array} \right] , \end{array} +$$ + +$$ +g _ { t } = \operatorname { t a n h } ( W ^ { g } [ x _ { t } , \alpha _ { t } h _ { t - 1 } , \beta _ { t } { r } _ { t } ] + b ^ { g } ) , +$$ + +$$ +\pmb { c } _ { t } = \pmb { f } _ { t } \circ \pmb { c } _ { t - 1 } + i _ { t } \circ \pmb { g } _ { t } , +$$ + +$$ +h _ { t } = o _ { t } \circ \operatorname { t a n h } ( c _ { t } ) , +$$ + +where $h _ { t - 1 } , h _ { t } , c _ { t } , i _ { t } , f _ { t } , o _ { t } , g _ { t } \in \mathbb { R } ^ { d _ { h } } , { r } _ { t } \in \mathbb { R } ^ { d _ { r } } , { \pmb x } _ { t } \in \mathbb { R } ^ { d _ { i } } , { \pmb w } ^ { \alpha } , { \pmb w } ^ { \beta } \in \mathbb { R } ^ { d _ { i } + d _ { h } + d _ { r } }$ , and $W \in$ R3dh×(di+dh+dr) $, \pmb { W } ^ { g } \in \mathbb { R } ^ { d _ { h } \times ( d _ { i } + d _ { h } + d _ { r } ) } , \pmb { b } \in \mathbb { R } ^ { 3 d _ { h } } , \pmb { b } ^ { g } \in \mathbb { R } ^ { d _ { h } }$ . The output of TARDIS is the concatenation of $\boldsymbol { h } _ { t }$ and $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ . + +There are 3 important points that distinguish the ARMIN cell from TARDIS cell, which are as follows: + +• The TARDIS uses two binary scalar gates $\alpha _ { t } , \beta _ { t }$ to control the information flow in equation 17. The two scalar gates are activated using gumbel-sigmoid (which is similar to gumbel-softmax and the output is close to binary), whereas the ARMIN uses two soft vector gates to control the information flow, which is more flexible and enable larger model capacity. For example, in the charlevel language modelling experiments, if we replace the sigmoid activations in equation 3 with the element-wise gumbel-sigmoid, the BPC performance immediately drops from 1.202 to 1.235. + +• The TADIS apply information control only when calculating the cell state $\mathbf { \nabla } _ { \mathbf { \boldsymbol { g } } _ { t } }$ , whereas the ARMIN apply information control immediately after the read operation. Our main motivation to this change is that if we don’t filter out irrelevant information in $\mathbf { \nabla } _ { \mathbf { r } _ { t } }$ and $h _ { t }$ at first, these noisy information will cause the weight parameters hard to learn. Furthermore, if the read operation chooses the wrong past state, we can block the information flow for $\mathbf { } _ { \mathbf { } } ^ { \mathbf { } } \mathbf { \Delta } \mathbf { r } _ { t }$ at the first place. + +• The ARMIN has an extra gate $\mathbf { \delta } _ { o _ { t } ^ { r } }$ to control the output from $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ as in equation 8, which the TARDIS doesn’t have. If we remove this gate and directly output $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ like in TARDIS, we get a BPC performance drop from 1.202 to 1.219 in char-level language modelling task. + +Finally, despite that the ARMIN has 3 more vector gates than the TARDIS, we have shown in the $p \mathrm { M N I S T }$ and Penn Treebank experiments that the ARMIN can still outperform the TARDIS under similar parameter counts. + +# B.2 READ OPERATION + +The memory matrix $M _ { t }$ of TARDIS has disjoint address section $\mathbf { \boldsymbol { A } } _ { t } ~ \in ~ \mathbb { R } ^ { n _ { m e m } \times d _ { k } }$ and content section $C _ { t } \ \stackrel { \cdot } { \in } \ \mathbb { R } ^ { n _ { m e m } \times d _ { r } }$ , $M _ { t } = [ A _ { t } ; C _ { t } ]$ . The controller reads both the address and the content parts of the memory, but it will only write into the content section of the memory. + +The TARDIS paper explored two ways to sample the read locations——using reinforce and using gumbel-softmax function, respectively. The gumbel-softmax way shows a better performance and its read operation can be formulated as follow: + +$$ +\begin{array} { c } { { \pi _ { t } [ i ] = a ^ { \top } \operatorname { t a n h } ( W ^ { \gamma } [ h _ { t } , x _ { t } , M _ { t } [ i ] , { \boldsymbol u } _ { t } ] + b ^ { \gamma } ) , } } \\ { { s _ { t } = \mathrm { g u m b e l - s o f t m a x } ( \pi _ { t } ) . } } \end{array} +$$ + +where $\{ a , W ^ { \gamma } , b ^ { \gamma } \}$ are learnable parameters (note that in our experiments, we always set the dimension of $\textbf { \em a }$ to about 1/4 of the hidden size) and $\mathbf { \Delta } \mathbf { u } _ { t }$ is a hand-crafted usage vector which denotes the frequency of accesses to each cell in the memory. + +There are two important differences with respect to TARDIS addressing mechanism and our autoaddressing mechanism: + +• In practice, equation 21 requires concatenating $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $h _ { t }$ to every memory cells $M _ { t } [ i ]$ for parallelization, which often causes significant extra memory and time consumption compared to the auto-addressing of ARMIN, as we show in the algorithmic tasks and char-level language modelling task. + +• The auto-addressing omits the learnable address section $\pmb { A } _ { t }$ , the hand-crafted $\mathbf { \pmb { u } } _ { t }$ vector, the tanh activation and an auxiliary vector $^ { a }$ compared to TARDIS addressing mechanism. The TARDIS also requires manually subtracting 100 in unnormalized probability at last read location to avoid repeated read, whereas the auto-addressing doesn’t need to. With these simplifications, we found the auto-addressing mechanism easier to train and it leads to a much faster converge speed, as we show in the algorithmic tasks. We also observe a BPC performance growth from 1.223 to 1.202 in char-level language modelling task when we switch from ARMIN $^ +$ TARDIS-addr to ARMIN with auto-addressing. + +# APPENDIX C EXPERIMENT SETUPS + +We first describe some universal setups in all our experiments (unless we mention them again with different setup). Our model can be trained end-to-end, due to the utilization of gumbel-softmax function. As pointed out in Maddison et al. (2016), when the temperature $\tau \leq ( n - 1 ) ^ { - 1 }$ , the density function of gumbel-softmax estimator is log-convex, which gives us a good guarantee for optimization. However, smaller $\tau$ will also bring higher variance of the gradients (Jang et al., 2016). So in all experiments, we initialize the reciprocal of $\tau$ as 1, and increase it by 1 after each epoch, until the reciprocal equals to $n - 1$ . This scheme has shown better performance than other schemes. + +Table 5: Hyperparameters of the models for algorithmic tasks. + +
ModelHidden Sizememory SizeParam Count
LSTM300一376k
NTM120128×2088k
TARDIS12050×3290k
ARMIN+TARDIS10050×3291k
ARMIN10050×3290k
+ +Table 6: Hyperparameters of the algorithmic tasks. + +
TaskSetup
Copycopy length =1~50
Repeat Copycopyl length=1~10 number of repeats = 1~10
Associative Recallitem size=3×6 item number = 2~6
Priority SortSort length = 40
+ +We apply Layer Normalization(Ba et al., 2016) to all networks. In the ARMIN case, layer norm is applied to equation 3, 6 and 7, which we find important to regulate the hidden state when scale grows. To avoid overfitting, we apply dropout (Srivastava et al., 2014) to the input and output layers, and Zoneout(Krueger et al., 2016) is applied for recurrent connections. The networks are trained with Adam optimizer (Kingma & Ba, 2014). All weight matrices in RNNs are initialized to orthogonal matrices, and the bias of the forget gate $\pmb { f } _ { t }$ is initialized to 1. For all models using TARDIS addressing mechanism, we set the memory key size to about 1/5 of the memory content size. + +# C.1 ALGORITHMIC TASKS + +In all algorithmic experiments below, following Graves et al. (2014), we apply RMSProp optimizer (Tieleman & Hinton, 2012) with a momentum of 0.9. The learning rate is 0.0001 and gradient norms are clipped to 10.0. No regularization is used as the size of the networks is small. Following Graves et al. (2014), we have learned bias of initial hidden states and memory matrices for all memory networks. For TARDIS, ARMIN $+ { \bf \Phi }$ TARDIS-addr and ARMIN we increase the reciprocal of temperature $\tau$ by 1 for every 200 iterations. The hyperparamters for models and tasks are shown in Table 5 and Table 6, respectively. We explain these tasks in detail in subsections below: + +# C.1.1 COPY + +The copy task tests whether a recurrent model can correctly store and recall a long sequence of arbitrary information. In our experiment, the networks are given randomly generated eight-bit binary sequences with its length ranging from 1 to 50, followed by a finish indicator. Then the networks are asked to output the same sequence with the input. + +# C.1.2 REPEAT COPY + +The repeat copy task is a repeat version of the copy task. The main motivation was to see if a recurrent model could memorize input information and repeatedly reuse them at later time-steps. In + +our experiment, the networks are asked to repeatedly copy the randomly generated input sequence with length ranging from 1 to 10 to output for 1 to 10 times. + +# C.1.3 ASSOCIATIVE RECALL + +This task test whether a memory network has the ability to retrieve memory data indirectly. The networks are given a sequence of items and then a query with one of the items, and are asked to output the subsequent item. In our experiments, each item consists of three six-bit binary vectors, and the number of items are ranging from 2 to 6. + +# C.1.4 PRIORITY SORT + +This task test whether a recurrent model can do priority sort——a classic elementary algorithm. This task is more challenging than previous ones as it not only require memorizing the input sequences but also making efficient use of the mutual relations of the items in the input sequences. In our experiments, the networks are given 40 eight-bit random binary keys with scalar priority values in a sequence, and are asked to output the top 30 keys in descending order of priority. + +C.2 CHARACTER-LEVEL LANGUAGE MODELLING ON PENN TREENBANK DATASET + +# C.2.1 PENN TREEBANK DATASET + +We use the train/validation/test split outlined in Mikolov et al. (2012). The train/validation/test batch size are 128/64/1 respectively. Following prior works on this task (Ha et al., 2016; Suarez, 2017; Mujika et al., 2017), the model is optimized over the cross-entropy loss between the predictions and the training labels, and evaluated using a bits-per-character measure. We use a single ARMIN cell of different hidden sizes, with a fully connected layer to output the predictive probabilities. The input embedding size $d _ { i }$ is set to 128. The number of memory slots $n _ { m e m }$ is 20. The zoneout probability is 0.3. We train the networks for 200 epochs with a learning rate of 0.002, and decay with a factor of 10 at the last 20 epochs. We clip the gradients to a maximum norm of 1.0. + +# C.2.2 HUTTER PRIZE WIKIPEDIA DATASET + +We use the train/validation/test split outlined in Chung et al. (2015). The train/validation/test batch size are 128/64/1 respectively, and the vocabulary size is 205. The zoneout and dropout probabilities are 0.3 and 0.2, respectively. The input embedding size is 256. Our ARMIN network has 2 layer and 800 units at each layer, connected by a linear layer to convert 1600-dimensional features from layer 1 to 256 dimension as input to layer 2. $n _ { m e m }$ for each layer is set to 10. We train the networks for 50 epochs with a learning rate of 0.001, and decay with a factor of 10 at the last 10 peochs. We clip the gradients to a maximum norm of 1.0. + +# C.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14 + +The SST takes in sequential inputs of a video and outputs the existing probability predictions of action proposals with different time lengths that end at each time-step, and uses a multi-label loss to optimize its parameters. The sequential inputs are 4096-dimensional features extracted from every 16 consecutive video frames using C3D network (Tran et al., 2015), followed by PCA to downsize the input feature dimensionality. + +Our experiment setup is basically the same as the SST paper, except that we use a simple fully connected layer to downsize the input features instead of using PCA. The input size of the two networks are both set to 256. The MA-SST have 256 hidden units, while the SST with LSTM have 512 hidden units. This ensures the 2 networks have roughly the same parameter count. The dropout and zoneout probabilities are both set to 0.2 and 0.3. $n _ { m e m }$ of the MA-SST is set to 35. We train the networks for 300 epochs with initial learning rate of 0.002, and decay with a factor of 10 at the last 60 epochs. + +![](images/50001ffd2185dd7661fc9aba2d9d75316cd989dcf8de377293cc7dbfab09dd11.jpg) +Figure 4: Copy task. + +![](images/4550719af5c0f00a37cd0f179fc1ec6a044179f516a3a2932ce1e6936572202f.jpg) +Figure 5: Repeat copy task. + +![](images/040994a58270f2a89222cd44b6a7cbcbc8407ddd4dfecf1d485eb625bbffc5c4.jpg) +Figure 6: Associative recall task. + +![](images/4cb5157c10e72c964f7ac67806728f1c6305e97e3dd30e339a3d94d5e8a27dc8.jpg) +Figure 7: Priority sort task. + +Table 7: Ablation on Penn Treebank $\scriptstyle ( d _ { h } = 8 0 0 )$ . + +
Abaltion ModelBPC
No Control Gates1.354
LSTM (TBPTT length=50) ARMIN(TBPTT length=50)1.390 1.220
Random Read1.224
Queue-style Write1.352
Independent Read/write1.245
SmoothRead/write1.212
+ +APPENDIX D TRAINING CURVES ON ALGORITHMIC TASKS + +# APPENDIX E ABLATION STUDY ON MEMORY AND ADDRESSING MECHANISM + +In this subsection, we show the efficacy of our external memory and addressing mechanism by ablation experiments. We choose the char-level language modelling task to do the ablation research as it is a relatively standard benchmark task and has appropriate medium scale. We use the aforementioned experiment setup. To validate the efficacy of our external memory, we conduct the experiments as follows and the results are depicted in table 7: + +• Short TBPTT Length We shorten the truncated length in TBPTT from 150 to 50. The gradients are truncated so we can’t do backpropagation on hidden states or memory from last iteration, but the ARMIN still has direct access to historical hidden states. By comparing the performance of the LSTM and ARMIN under such circumstances, we can know the efficacy of the external memory. The results show that the LSTM has a dramatic performance drop of 1.390 BPC while the ARMIN only has a minor performance drop of 1.220 BPC. The results confirm that the direct access to memory is crucial when the networks can’t access historical information by backpropagating gradients. And notably, our experiment shows that the ARMIN is very robust to short-length TBPTT, i.e. small iteration length, which means lower memory consumption. It enables using very large batch size to speed up the training and further increasing performance on sequence modelling tasks. + +• No Control Gates We remove the control gate $\pmb { g } _ { t } ^ { h }$ and $\pmb { g } _ { t } ^ { r }$ in ARMIN cell as depicted in equation 3, and directly use $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $\boldsymbol { h } _ { t - 1 }$ to compute the gates in equation 6. The result shows a BPC of 1.354 that has a dramatic performance drop compared the 1.202 BPC of normal ARMIN. From this, we deduce that the control gates successfully filter out irrelevant information. Without the control gates, the historical state $\mathbf { } _ { \mathbf { } \mathbf { } \mathbf { } \mathbf { } r _ { t } }$ would have bring more noise than useful information. + +To validate the efficacy of the addressing mechanism of the ARMIN, we do the following modifications to the read/write operations of the ARMIN, and the results are also depicted in table 7: + +• Random Read The ARMIN randomly chooses a historical state while reading. The result shows a BPC of 1.224, compared to the 1.202 BPC of normal ARMIN. It proves that the read regime we proposed in section 3.2 is effective and can choose more useful historical state than random read. + +• Queue-style Write While writing, the ARMIN follows a fashion of first-in-first-out, i.e. always chooses the oldest memory slot to write. It makes the memory $M$ stubbornly store the hidden states of the last 20 time-steps. The result shows a BPC of 1.352, which proves our normal read/write mechanism has much more flexibility and can learn long-term patterns. + +• Independent Read/write While writing, the ARMIN doesn’t write to the last read memory slot, but independently choose a memory slot to overwrite according to $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { t } }$ and $h _ { t }$ , in a similar fashion with the read operation we presented in section 3.2. The result shows a BPC of 1.245, proving that the normal writing regime in the ARMIN can better learn long-term patterns and maximally preserve historical information, as we explain in section 3.3. + +• Smooth Read/write We replace the gumbel-softmax function with the vanilla softmax function, so the ARMIN reads and writes in a smooth way. The results show a minor performance drop of 1.212 BPC. The result shows that our discrete read/write regime is still better. Moreover, the smooth read/write regime would consume much more memory at both training and inference stages. \ No newline at end of file diff --git a/parse/train/HyMS8iRcK7/HyMS8iRcK7_content_list.json b/parse/train/HyMS8iRcK7/HyMS8iRcK7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1b4197b1623e037a473ea2fba4ab21ae374529a4 --- /dev/null +++ b/parse/train/HyMS8iRcK7/HyMS8iRcK7_content_list.json @@ -0,0 +1,2167 @@ +[ + { + "type": "text", + "text": "SEQUENCE MODELLING WITH AUTO-ADDRESSINGAND RECURRENT MEMORY INTEGRATING NETWORKS", + "text_level": 1, + "bbox": [ + 176, + 99, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Processing sequential data with long term dependencies and learn complex transitions are two major challenges in many deep learning applications. In this paper, we introduce a novel architecture, the Auto-addressing and Recurrent Memory Integrating Network (ARMIN) to address these issues. The ARMIN explicitly stores previous hidden states and recurrently integrate useful past states into current time-step by an efficient memory addressing mechanism. Compared to existing memory networks, the ARMIN is more light-weight and inference-time efficient. Our network can be trained on small slices of long sequential data, and thus, can boost its training speed. Experiments on various tasks demonstrate the efficiency of the ARMIN architecture. Codes and models will be available. ", + "bbox": [ + 233, + 266, + 764, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 433, + 336, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recurrent neural networks, such as the Long Short-Term Memory (LSTM) (Hochreiter & Schmidhuber, 1997) and Gated Recurrent Unit (GRU) (Cho et al., 2014) have shown promising performance for processing sequential data. However, it’s known that RNNs suffer from gradient vanishing problem. Moreover, as pointed out by Rae et al. (2016), the number of parameters grows proportionally to the square of the size of the hidden units, which carry the historical information. Recent memorybased approaches exhibit potential to address these issues, by decoupling memory capacity from model parameters, and backpropagating the gradients through the memory. ", + "bbox": [ + 174, + 464, + 823, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Turing Machine (NTM) (Graves et al., 2014) first emerged as a recurrent model that incorporates external memory abilities. NTM maintains a memory matrix, and at every time-step, the network reads and writes (with erasing) to the memory matrix using certain soft-attentional mechanism, controlled by an LSTM that produces read and write vectors. NTM and its successor, the Differentiable Neural Computer(Graves et al., 2016), have shown success on some algorithmic tasks such as copying, priority sorting and some real-world tasks such as question answering. But one limitation of the NTM is that due to its smooth read and write mechanism, NTM has to do propagations on the entire memory, usually causing huge amount of memory consumption. To this end, Rae et al. (2016) proposes the Sparse Access Memory(SAM) network, by thresholding memory modifications to a sparse subset, i.e. all read and write operations are limited to several memory words. This allows memory and time efficient propagations while maintaining NTM’s performance. However, these external memory models have relatively complicated memory addressing mechanisms, making them suitable for only large-scale memory. Moreover, the RNN in these models plays a simple role of being a controller, but the gradient vanishing problem of RNN itself is not given attention. In contrast to these networks, our network uses a light-weight read mechanism to address a small external memory, and relieves the gradient vanishing problem by directly concatenating memory word to output. ", + "bbox": [ + 174, + 569, + 825, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by prior memory models, efforts have been made to build a bridge between simple RNNs and complicated memory models. Kurach et al. (2015) propose the Neural Random-access Machines (NRAM) that can manipulate and dereference pointers to an external variable-size random-access memory. Danihelka et al. (2016) improve LSTM with ideas from Holographic Reduced Representations (Plate, 2003) that enables key-value storage of data. Grave et al. (2016) propose a method of augmenting LSTM by storing previous (hidden state, input word) pairs in memory and using the current hidden state as a query vector to recover historical input words. This method requires no backpropagation through memory and is well-suited to word-level language tasks. Grefenstette et al. (2015); Dyer et al. (2015); Joulin & Mikolov (2015) augment RNNs with a stack structure that works as a natural parsing tool, and use them to process algorithmic and nature language processing (NLP) tasks; nonetheless, the speed of stack-augmented RNNs is rather slow due to multiple push-pop operations at every time-step. Ke et al. (2018) propose the Sparse Attentive Backtracking (SAB) architecture, which recalls a few past hidden sates at every time-step and do “mental” backpropogations to the nearby hidden states with respect to the past hidden states. Gulcehre et al. (2017) propose the TARDIS network, which is an LSTM-resembled RNN that directly stores a fixed number of previous hidden states with learned key parameters for memory addressing. At every time-step, the network reads out one historical state $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ from the memory, and uses it, along with the input $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and last hidden state $h _ { t - 1 }$ , to produce new hidden state $h _ { t }$ , then overwrites $\\boldsymbol { h } _ { t }$ to the location of $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ . TARDIS optionally uses the gumbel-softmax estimator(Maddison et al., 2016; Jang et al., 2016) to sample the location of $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ (we will explain this in section 2.1), so the whole network is differentiable. However, the TARDIS still involves some hand-crafted memory addressing method and its addressing mechanism requires a considerable amount of memory and time consumption. ", + "bbox": [ + 174, + 811, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the TARDIS, we introduce the ARMIN architecture, a more light-weight and memorybased architecture with a very simple memory addressing mechanism and in-depth modification of the existing LSTM structure. Concretely, our contributions are as follows: ", + "bbox": [ + 174, + 304, + 821, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We distill previous memory addressing methods and propose a simple yet effective memory addressing mechanism for the external memory, namely Auto-addressing, by encoding the information for memory addressing directly via the inputs $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and the hidden states $\\boldsymbol { h } _ { t - 1 }$ . • We propose a novel recurrent cell that combines the gating advantages of LSTM and allows direct gradient backpropagation from output to memory. With only a single cell, it learns better representations than many hierarchical RNN structures. • We show in our ablation study that the ARMIN is robust to small iteration lengths when training long sequential data, which enables training with large batch sizes and boost its training speed. We further demonstrate in char-level language modelling task that the ARMIN obtain $4 0 \\%$ training speed gain than LSTM while keeping similar performance and memory consumption. ", + "bbox": [ + 173, + 361, + 826, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We demonstrate competitive results on various tasks while keeping efficient time and memory consumption during training and inference time. ", + "bbox": [ + 171, + 513, + 823, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 174, + 563, + 326, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 GUMBEL-SOFTMAX ESTIMATOR ", + "text_level": 1, + "bbox": [ + 174, + 594, + 434, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Categorical distribution is a natural choice for representing discrete structure in the world; however, it’s rarely used in neural networks due to its inability to backpropagate through samples(Jang et al., 2016). To this end, Maddison et al. (2016) and Jang et al. (2016) propose a continuous relaxation of categorical distribution and the corresponding gumbel-softmax gradient estimator that replaces the non-differentiable sample from a categorical distribution with a differentiable sample. Specifically, given a probability distribution $\\pmb { p } = ( \\pi _ { 1 } , \\pi _ { 2 } , . . . , \\pi _ { k } )$ over $k$ categories, the gumbel-softmax estimator produces an one-hot sample vector $\\textbf { { y } }$ with its $i$ -th element calculated as follows: ", + "bbox": [ + 173, + 619, + 825, + 718 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/3d7fbb0ffd4812f959da86e645aa96d65bbe99701c988143802f6d35011fd4d2.jpg", + "text": "$$\ny _ { i } = { \\frac { \\exp ( ( \\log ( \\pi _ { i } ) + g _ { i } ) / \\tau ) } { \\sum _ { j = 1 } ^ { k } \\exp ( ( \\log ( \\pi _ { j } ) + g _ { j } ) / \\tau ) } } \\quad { \\mathrm { f o r ~ } } i = 1 , 2 , . . . , k ,\n$$", + "text_format": "latex", + "bbox": [ + 315, + 724, + 679, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $g _ { 1 } , . . . , g _ { k }$ are i.i.d samples drawn from Gumbel distribution(Gumbel, 1954): ", + "bbox": [ + 173, + 768, + 714, + 785 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/f9b200eff054dd7a1f090b16bac57a8cbc7a58c52f126fa82b11ce37ab6acf42.jpg", + "text": "$$\ng _ { i } = - \\log ( - \\log ( u _ { i } ) ) \\ , u _ { i } \\sim \\mathrm { U n i f o r m } ( 0 , 1 ) ,\n$$", + "text_format": "latex", + "bbox": [ + 349, + 791, + 647, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and $\\tau$ is the temperature parameter. In practice, we usually start at a high temperature and anneal to a small but non-zero temperature (Jang et al., 2016). ", + "bbox": [ + 169, + 815, + 823, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 AUTO-ADDRESSING AND RECURRENT MEMORY INTEGRATING NETWORK", + "text_level": 1, + "bbox": [ + 171, + 863, + 818, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we describe the structure of the ARMIN network as shown in Figure 1. It consists of a recurrent cell and a small external memory that stores historical hidden states. While processing ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eab2a73d816507916b458d579eb538fc7d49d69aba42b4a31414c981c7c3217e.jpg", + "image_caption": [ + "Figure 1: The ARMIN structure. At each time-step, the ARMIN performs read operation, cell processing and write operation in chronological order: (a) It reads out a historical hidden state $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ from memory with an one-hot read vector produced via passing $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $\\boldsymbol { h } _ { t - 1 }$ to a fully connected layer followed by a gumbel-softmax function. (b) The ARMIN cell receives $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ , $h _ { t - 1 }$ and $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ as inputs and outputs $\\mathbf { } _ { o _ { t } }$ and $\\boldsymbol { h } _ { t }$ . $\\mathbf { } _ { o _ { t } }$ is passed to output layers, and $h _ { t }$ is passed to next time-step. (c) $\\boldsymbol { h } _ { t }$ is written to the previous location of $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ . " + ], + "image_footnote": [], + "bbox": [ + 228, + 114, + 750, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sequential data, the ARMIN performs reading from memory, cell processing and writing to memory operations in chronological order during each time-step. In the following subsections, we first explain the structure of the recurrent cell, and then discuss the read and write operations. ", + "bbox": [ + 174, + 467, + 825, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 THE RECURRENT CELL OF ARMIN", + "text_level": 1, + "bbox": [ + 176, + 526, + 459, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by classical LSTM structure (please refer to APPENDIX A for details in RNN and LSTM structures), we propose a novel memory-augmented recurrent cell structure, namely the ARMIN cell. At every time-step, it takes in an input $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ , the last hidden state $\\boldsymbol { h } _ { t - 1 }$ and a recovered historical hidden state $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ chosen by a read operation, and produces an output vector $\\mathbf { } _ { o _ { t } }$ and the new hidden state $\\boldsymbol { h } _ { t }$ . The computation process is as follows: ", + "bbox": [ + 173, + 551, + 825, + 622 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/1a2381c9e716db06877af5cbb16644d9f37ab3bbcdff250345c9426035fa0588.jpg", + "text": "$$\n\\left\\{ { g } _ { t } ^ { h } \\atop { g } _ { t } ^ { r } \\right\\} = \\left\\{ { \\sigma } \\atop { \\sigma } \\right\\} W _ { i g } [ x _ { t } , h _ { t - 1 } , r _ { t } ] + b _ { i g } ,\n$$", + "text_format": "latex", + "bbox": [ + 359, + 638, + 640, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/217d67193e54a23659454a2544a34c38d6fc260d6bad8abdfb195eba32d5d806.jpg", + "text": "$$\nh _ { t - 1 } ^ { g } = { \\pmb g } _ { t } ^ { h } \\circ { \\pmb h } _ { t - 1 } ,\n$$", + "text_format": "latex", + "bbox": [ + 431, + 680, + 563, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/2ff3978396c1892799181c33d24e40b07974e748cecce11677c3c5fc54205888.jpg", + "text": "$$\n\\begin{array} { r } { \\pmb { r } _ { t } ^ { g } = \\pmb { g } _ { t } ^ { r } \\circ \\pmb { r } _ { t } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 450, + 702, + 547, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/159cf38197a23b05f8f396953f74402cbfc9572e8bbc0deba36b532733921968.jpg", + "text": "$$\n\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } ^ { h } } \\\\ { o _ { t } ^ { r } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W _ { g o } [ \\boldsymbol { x } _ { t } , h _ { t - 1 } ^ { g } , \\boldsymbol { r } _ { t } ^ { g } ] + b _ { g o } ,\n$$", + "text_format": "latex", + "bbox": [ + 339, + 723, + 658, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/34a091eaff95e86551a79e6eb6b3b2b7571b87ba5d72f231a10afbfeff0581e3.jpg", + "text": "$$\n\\pmb { h } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { h } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ,\n$$", + "text_format": "latex", + "bbox": [ + 405, + 815, + 588, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/58570aefc636c1ba6609da63ca2076995d0b6a9d09702509a5cdbb78c6ea9728.jpg", + "text": "$$\n\\begin{array} { r } { \\pmb { o } _ { t } = \\left[ \\pmb { o } _ { t } ^ { h } \\circ \\operatorname { t a n h } ( \\pmb { h } _ { t } ) , \\pmb { o } _ { t } ^ { r } \\circ \\operatorname { t a n h } ( \\pmb { r } _ { t } ) \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 366, + 835, + 632, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h _ { t - 1 } , h _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { \\boldsymbol { r } } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\boldsymbol { x } } _ { t } \\in \\mathbb { R } ^ { d _ { i } }$ , and $\\begin{array} { r c l } { W _ { i g } } & { \\in } & { \\mathbb { R } ^ { 2 d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , W _ { g o } } & { \\in } \\end{array}$ $\\begin{array} { r } { \\mathbb { R } ^ { ( 4 d _ { h } + d _ { r } ) \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } _ { i g } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } , \\pmb { b } _ { g o } \\ \\in \\ \\mathbb { R } ^ { 4 d _ { h } + d _ { r } } , \\pmb { o } _ { t } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } } \\end{array}$ . We refer to $d _ { h }$ as the hidden size of the recurrent cell of ARMIN. Usually we have $d _ { h } = d _ { r }$ and allocate equal number of weight parameters for $\\boldsymbol { h } _ { t }$ and $\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { r } } _ { t } }$ . ", + "bbox": [ + 173, + 863, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In equation $3 \\sim 5$ , two gates are calculated to control the information flow for $h _ { t - 1 }$ and $r _ { t }$ respectively, generating gated hidden state $h _ { t - 1 } ^ { g }$ and historical state $\\boldsymbol { r } _ { t } ^ { g }$ ; using this method, we can filter out the irrelevant information for the current time-step. Then as shown in equation 6, we compute the input gate $\\mathbf { \\delta } _ { i _ { t } }$ , forget gate $\\pmb { f } _ { t }$ , cell state $\\mathbf { \\pmb { g } } _ { t }$ and output gate $\\mathbf { \\sigma } _ { o _ { t } ^ { h } } ^ { h }$ for the new hidden state just like in classical LSTM structure. Additionally, an output gate ${ \\mathbf { } } _ { o _ { t } ^ { r } }$ for historical state $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ is computed. Next in equation 7, we compute new hidden state $h _ { t }$ that is the sum of $h _ { t - 1 }$ and cell state $\\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } } \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\lambda } }$ , leveraged by forget gate $\\pmb { f } _ { t }$ and input gate $\\dot { \\mathbf { \\delta } } _ { \\mathbf { \\delta } }$ . Finally in equation 8, we calculate the output of this timestep, which is the concatenation of the gated contents from $h _ { t }$ and $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ . Using this method, we can selectively backpropagate gradients from the output to the historical state $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ . We involve $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ (or $\\boldsymbol { r } _ { t } ^ { g }$ ) in all gates’ computations. ", + "bbox": [ + 173, + 103, + 825, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, the $\\boldsymbol { h } _ { t - 1 }$ acts as the old working memory, and $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ is treated as the long-term memory. The cell processes them with the input $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ to generate the new working memory $\\boldsymbol { h } _ { t }$ and the output $\\mathbf { } _ { o _ { t } }$ . More specifically, each $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ is a summary of historical hidden states selected by auto-addressing mechanism. The ARMIN cell learns to recurrently integrate the summary of long-term information from $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ into the working memory $\\boldsymbol { h } _ { t }$ . We will demonstrate the efficiency of recurrent memory integration in language modelling tasks. ", + "bbox": [ + 173, + 250, + 825, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the LSTM that passes a tuple of $\\left( \\boldsymbol { h } _ { t } , \\boldsymbol { c } _ { t } \\right)$ to the next recurrence and outputs $\\boldsymbol { h } _ { t }$ at the same time, the ARMIN only passes one hidden state $h _ { t }$ and separately uses $\\mathbf { } _ { o _ { t } }$ as output, which has 2 benefits: one is that it takes lower memory consumption to store one hidden state rather than two, and the other is we can decouple the information needed for the output $\\mathbf { } _ \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf $ , from the information (i.e. $\\boldsymbol { h } _ { t }$ ) that is needed for memorizing and being used at later time-steps. Furthermore, we list 3 important differences between the ARMIN cell and TARDIS cell, please refer to APPENDIX B.1. ", + "bbox": [ + 174, + 339, + 825, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 READ OPERATION WITH AUTO-ADDRESSING", + "text_level": 1, + "bbox": [ + 178, + 440, + 522, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ARMIN maintains a memory matrix $M \\in \\mathbb { R } ^ { n _ { m e m } \\times d _ { h } }$ , where the constant $n _ { m e m }$ denotes the number of memory slots. At each recurrence, the ARMIN chooses a historical state $r _ { t }$ from memory according to the information in $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $\\boldsymbol { h } _ { t - 1 }$ ,which is formulated as follows: ", + "bbox": [ + 174, + 465, + 823, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/d20fa90dbab7b45d436ec9d1fe2698fcfc15d2cb48c0e332be86d0aefcd202b3.jpg", + "text": "$$\n\\begin{array} { c } { { \\displaystyle s _ { t } = \\mathrm { g u m b e l - s o f t m a x } \\big ( W _ { s } [ \\ : x _ { t } \\ : , h _ { t - 1 } \\ : ] + b _ { s } \\big ) \\ : , } } \\\\ { { { } } } \\\\ { { \\displaystyle r _ { t } = \\sum _ { i = 0 } ^ { n - 1 } s _ { t } ( i ) M ( i , : ) \\ : . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 341, + 526, + 655, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W _ { s } \\ \\in \\ \\mathbb { R } ^ { n _ { m e m } \\times ( d _ { i } + d _ { h } ) }$ , $b _ { s } \\in \\mathbb { R } ^ { n _ { m e m } }$ , $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { S } } _ { t } }$ is a one-hot vector sampled by gumbel-softmax function, $s _ { t } ( i )$ denotes the $i$ -th element of $\\mathbf { \\boldsymbol { s } } _ { t }$ , $M ( i , : )$ denotes the $i$ -th row of $M$ . ", + "bbox": [ + 176, + 594, + 821, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As opposed to previous memory networks such as Santoro et al. (2016) and Gulcehre et al. (2017), we don’t use any extra usage vectors and learnable key parameters that manually encode memory addressing information to assist memory addressing. As our ablation study (Appendix E) shows, the hidden state $\\boldsymbol { h } _ { t - 1 }$ is sufficient to encode the historical memory accessing information. Furthermore, the TARDIS addressing mechanism requires concatenating $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $\\boldsymbol { h } _ { t }$ to every memory cell, which usually causes large extra memory and time consumption in practical scenes. Please refer to APPENDIX B.2 for more precise comparison regarding addressing mechanism of the ARMIN and TARDIS networks. ", + "bbox": [ + 173, + 630, + 825, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 WRITE OPERATION", + "text_level": 1, + "bbox": [ + 176, + 758, + 346, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the reading and cell processing stages, the ARMIN writes the new hidden state $\\boldsymbol { h } _ { t }$ to the memory $M$ . Following the TARDIS network (Gulcehre et al., 2017), we simply overwrite $h _ { t }$ to the memory slot where we just read out the $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ (for conditions where $d _ { h }$ is not equal to $d _ { r }$ , we first use a linear layer to transform $\\boldsymbol { h } _ { t }$ from $d _ { h }$ dimension to $d _ { r }$ dimension). But at the initial time-steps, we write the hidden states to the empty memory slots, until all empty slots are filled with historical states. In this way, we can maximally preserve useful historical information, because the ARMIN cell can learn to copy the useful information from $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ to $\\boldsymbol { h } _ { t }$ , and then write $h _ { t }$ to the previous location of $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ . The read/write mechanism can be viewed as a form of a skip-connection of hidden states. By training the whole network, the ARMIN can easily learn long-term dependencies via direct access to historical hidden states. From this point of view, our network is similar to the Skip RNN proposed by Campos Camunez et al. (2018), which differs from our network in that they use a binary state update gate to select whether the state of the RNN will be updated or copied from the previous time-step. ", + "bbox": [ + 174, + 784, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 165, + 326, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model on algorithmic tasks, character-level language modelling task and temporal action detection and proposal task. We compare our network with NTM (with an LSTM controller), TARDIS and vanilla LSTM networks. We also directly compare the auto-addressing against the TARDIS addressing mechanism, by replacing auto-addressing with the TARDIS addressing method in ARMIN. We refer to this network as ARMIN $^ +$ TARDIS-addr in our experiments. Please refer to APPENDIX C for some universal setups used in all our experiments below. ", + "bbox": [ + 174, + 196, + 825, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 ALGORITHMIC TASKS ", + "text_level": 1, + "bbox": [ + 176, + 297, + 364, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Along with the NTM, Graves et al. (2014) introduced a set of synthetic algorithm tasks which consists of copy, repeat copy, associative recall, N-gram and priority sort tasks. Here we use four out of five of these tasks (excluding the N-gram task) to examine if our network can choose correct time-steps from the past and effectively make use of them. Please refer to APPENDIX C.1 for more explanations and hyperparameter setups with respect to these tasks. Note that we have a strong LSTM baseline with about 4 times larger parameter count than memory networks, and the training of our copy and priority sort tasks are more challenging than the original tasks in Graves et al. (2014), for having more than 2 times longer input and target sequences. The models are optimized to minimize the averaged binary cross-entropy loss on the target multi-bit binary sequences. Following Gulcehre et al. (2017); Campos Camunez et al. (2018), in all tasks, we consider a task solved if the validation loss is at least two orders of magnitude below the initial cost which is around 0.70, i.e. the validation loss (the validation set is generated randomly) converges to less than 0.01 (with less than $3 0 \\%$ in 10 consecutive validation of sharp losses that are higher than 0.01 ). In our evaluation, we are interested in if a model can successfully solve the task in 100k iterations and the elapsed training time and iterations till the model succeeds on the task. For all networks we ensure the similar total parameter counts. All important hyperparameters are shown in APPENDIX C.1. We run all experiments with batch size of 1 under 2.8 GHz Intel Core i7 CPU and report the wall clock time and the number of iterations for solved tasks of each model. For models that fail to converge to less than 0.01 loss in $1 0 0 \\mathrm { k }$ iterations, we report the average loss of the final 10 validations (denoted in underlines) and the elapsed training time for 100k iterations. The results are shown in Table 1. The training curves for each task are shown in APPENDIX D. ", + "bbox": [ + 174, + 323, + 825, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results show that the ARMIN can solve 3 out of 4 tasks in a fast converge speed. By comparing the ARMIN with NTM, we observe that the NTM is able to almost solve all of the tasks, but with a much slower training speed compared to the ARMIN, for example, the training time of NTM is 5 times larger than ARMIN to solve the copy task. In fact, the training speed of ARMIN for each iteration is about $3 \\sim 4$ times faster than the NTM’s speed. We observe that the TARDIS and ARMIN $\\dotplus$ TARDIS-addr fail to solve 3 out 4 tasks in the given $1 0 0 \\mathrm { k }$ iterations. By comparing them with the ARMIN, we confirm the efficiency of the auto-addressing of ARMIN, for bringing a fast converge speed and short training time. We also find that the LSTM, when given enough parameter counts (4 times larger than memory networks), is able to solve the priority sort task quickly and converge to a better performance than TARDIS and ARMIN $^ +$ TARDIS in other tasks. We believe there are 2 possible reasons for the poor performance of the TARDIS addressing mechanism: a) The tanh activation (see equation 21 in APPENDIX B.2) might cause the gradient vanishing problem when it cooperates with the gumbel-softmax function. b) The multiple inputs of $\\pmb { x } _ { t } , \\pmb { h } _ { t } , \\pmb { M } _ { t } [ i ] , \\pmb { u } _ { t }$ in equation 21 cause the network hard to initialize to keep equal input and output variances for the linear layer and might also cause gradient dispersion. The auto-addressing has neither of these two issues and can better integrately backpropogate gradients, therefore it can perform well and quickly on these tasks. However, we would like to point out 2 limitations of the auto-addressing: a) The associative recall task shows the auto-addressing might have problems in establishing indirect references among memory data. b) The auto-addressing is relatively hard to generalize to unseen longer sequence in tasks such as copy and repeat copy due to its simplicity, whereas the NTM is able to do that. We leave these 2 issues for future work. ", + "bbox": [ + 174, + 621, + 825, + 718 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3e9670eb5cd4fc0e6980a395033241124564e6ed25b13ade887f0170637515a7.jpg", + "table_caption": [ + "Table 1: The average elapsed time(min) and iterations ${ \\mathrm { ( k ) } }$ of different networks till task solved in given $1 0 0 \\mathrm { k }$ iterations. The wall clock training time for 100k iterations and the average loss of final 10 validations (denoted in underlines) are shown for the unsolved tasks. " + ], + "table_footnote": [], + "table_body": "
CopyRepeat CopyAssociative RecallPriority Sort
timeiter.timeiter.timeiter.timeiter.
LSTM660.359450.0160.3253237.0
NTM3212.41710.01419.63600.012
TARDIS1570.451970.1663422660.33013362.6
ARMIN+TARDIS-addr1520.410950.297640.33615171.4
ARMIN67.63667.8330.0523733.4
", + "bbox": [ + 197, + 800, + 800, + 915 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4160d0d481224f188526c1216f05763bf331261cc0ba4a079eac59cba78e5098.jpg", + "table_caption": [ + "Table 2: Bits-per-character on Penn Treebank and enwik8 test set. " + ], + "table_footnote": [], + "table_body": "
ModelPTBenwik8
BPCParamsBPCParams
LSTM1.36一1.45
LSTM+Zoneout (Krueger et al., 2016)1.2711
LSTM+Layer Norm(1000 units)11.2674.26M1一
LSTM+Layer Norm+Zoneout(1024 units)²1.244.79M一
HM-LSTM+Layer Norm (Chung et al.,2016)1.2411.3235M
HyperLSTM+Layer Norm(Ha et al., 2016)1.21914.41M1.34026.5M
NASCell (Zoph & Le,2016)1.21416.28M
IndRNN (21 layers) (Li et al.,2018)1.2111
Recurrent Highway Hypernetwork (Suarez, 2017)1.19一 15.5M11
Fast-Slow LSTM (Mujika et al., 2017)1.1907.2M1.277一 27M
NTM(800 units, nhead = 1, Pdropout=0.6)
TARDIS(paper)1.5358.28M1
1.25一
TARDIS(1000 units,Pdropout=0.6)1.2689.2M一
ARMIN+TARDIS-addr(800 units,Pdropout=0.6)1.22310.2M
ARMIN (500 units, Pdropout=0.4) (Ours)1.2364.03M
ARMIN (800 units, Pdropout=0.6) (Ours)1.2029.8M
ARMIN(800 units,2 layer) (Ours)111.33121.6M
", + "bbox": [ + 199, + 132, + 797, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 460, + 825, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 CHARACTER-LEVEL LANGUAGE MODELLING ", + "text_level": 1, + "bbox": [ + 174, + 672, + 526, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The character-level language modelling task consists of predicting the probability distribution of the next character given all the previous ones, and we use the widely used Penn Treebank and Hutter Prize Wikipedia (also known as enwik8) datasets for this task. We compare our network with vanilla LSTMs and other memory networks on Penn Treebank, and compare with some popular RNN variants on both datasets. We ensure similar parameter counts and hyperparameters for all memory networks. Following prior works, we apply truncated backpropagation through time(TBPTT) (Rumelhart et al., 1986; Elman, 1990) to approximate the gradients: at each iteration, the network predict the next 150 characters, and the hidden state $\\boldsymbol { h } _ { t }$ and memory state $M$ are passed to the next iteration. The gradients are truncated between different iterations. Please refer to APPENDIX C.2 for more experimental details. ", + "bbox": [ + 174, + 698, + 825, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in table 2. On Penn Treebank dataset, our best performing network achieves competitive 1.202 BPC on Penn Treebank dataset, which is the best single cell performance that we are aware of, with relatively small parameter count. By comparing TARDIS and ARMIN $^ +$ TARDISaddr, we observe about 4 points of improvement, which shows the efficiency of the ARMIN cell. By comparing ARMIN $+$ TARDIS-addr, we observe a further improvement of around 2 points, which shows the efficiency of the auto-addressing mechanism. We also find that the same architecture of NTM that performs well in algorithmic tasks fails to converge to a lower BPC than vanilla LSTM, even with various regularizations we add to NTM. ", + "bbox": [ + 176, + 844, + 825, + 886 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on Penn Treebank show our single ARMIN cell learns better representations than many hierarchical RNN structures, such as the HM-LSTM, 2-Layer HyperLSTM and 21 layer IndRNN. Our network is outperformed by Recurrent Highway Hypernetwork and Fast-slow LSTM which are both multi-scale and deep transition RNNs and are state-of-the-art results without any auxiliary technics on this dataset. By saying “better representations”, we refer to the concatenation of the gated contents from $\\boldsymbol { h } _ { t }$ and $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ as in equation 8. For example, if we remove the gated contents of $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ from $\\mathbf { } _ { o _ { t } }$ , the ARMIN undergoes a BPC performance drop from 1.202 to 1.220, which is still better than the best performing BPC of 1.24 of the LSTM. We believe the rest of the performance gain comes from the recurrent memory integration of the ARMIN cell, which also favors a paradigm of deep transition, as is shown by the success of the deep transition RNNs on this task. For more ablation study regarding the auto-addressing mechanism and the ARMIN cell please refer to APPENDIX E (where we show the robustness of ARMIN to small TBPTT length). ", + "bbox": [ + 174, + 180, + 825, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The result on enwik8 demonstrates a simple 2-layer ARMIN can achieve competitive BPC performance of 1.33, with less parameter count compared to the HyperLSTM and HM-LSTM. By constructing deeper ARMIN network or even combining with other multi-scale and hierarchical RNN architectures, we believe the performance can be further improved. ", + "bbox": [ + 176, + 354, + 823, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14 ", + "text_level": 1, + "bbox": [ + 176, + 426, + 684, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we evaluate the ARMIN on a more complicated real-world task concerning video analysis, i.e. the temporal action detection and proposals, which consists of taking an input video and producing a set of temporal intervals that are likely to contain human actions (Buch et al., 2017). We use the Single-Stream Temporal Action Proposal(SST) (Buch et al., 2017) as our framework, and evaluate the model on the THUMOS’ 14 dataset (Jiang et al., 2014), which contains $^ { 2 0 + }$ hours of video with 200 train/validation and 213 test untrimmed video sequences. The SST consists of a GRU and a fully connected output layer. In our experiment, we replace the GRU encoder in SST with a single ARMIN cell, and we call the modified network MA-SST. For fairness, we also implement a modified SST with LSTM encoder and compare them with the original SST in Buch et al. (2017). Please refer to APPENDIX C.3 for more experimental details. ", + "bbox": [ + 174, + 452, + 825, + 590 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9dee7e00530b4ee45916348ee8b30316fa8b7e2b2b83fe8150f9f7711f240679.jpg", + "table_caption": [ + "Table 3: Comparison of proposal generation performance in terms of recall at 1000 proposals. We choose $\\mathrm { \\ t I o U { = } 0 } . 6$ and 0.8 to compare for consistency with Buch et al. (2017). Our MA-SST achieves highest recall at $\\mathrm { \\ t I o U { = } } 0 . 8$ . ", + "Table 4: Comparison for performance and cost of different setups on Penn Treebank dataset. " + ], + "table_footnote": [], + "table_body": "
ModelARMINLSTM
Setup1212
Hidden sizenparam(M)nmemTtruncbatch sizeMemory(GB)Speed(chars/s)BPC5005501k1k
4.0254.814.794.791
101
5050100150
384300128128
3.493.562.363.27
98k75k71k70k
1.2381.2261.271.24
", + "bbox": [ + 504, + 631, + 826, + 792 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/850273973a2f428712c3ae412e72577c47e4474b112a60383bcfb3009629e5e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModeltIoU=0.6tIoU=0.8
SST (Paper)0.9200.672
SST (We implement)0.8970.732
MA-SST(Ours)0.9110.759
", + "bbox": [ + 176, + 702, + 496, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The average recall under different proposal numbers and tIoUs3 are depicted in figure 2. MA-SST outperforms the original SST and our SST simultaneously. To the best of our knowledge, MA-SST achieves highest performance of recall at around $\\mathrm { \\ t I o U { = } } 0 . 8$ (see table 3), even if other competitive non-RNN networks (Escorcia et al., 2016; Caba Heilbron et al., 2016; Gao et al., 2017; Guo et al., 2018) are taken into account. The MA-SST tends to precisely generate proposals that has a high overlapping area with the ground truth action intervals, leading to high recall performance at high tIoUs. We believe the MA-SST has better performance than the SST mainly because of the direct access to historical states from which it can better locate the starting frame of an action. ", + "bbox": [ + 173, + 804, + 825, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80e05beebf3cb64569df7a0d731d988352907332c1e2de7be2c1637e50fd4897.jpg", + "image_caption": [ + "Figure 2: Comparison of proposal generation performance. For data generation, we use the code offered in Buch et al. (2017) for consistency. (Left) The average number of proposals v.s. average recall for $\\mathrm { t I o U } \\geq 0 . 5$ . The MA-SST outperforms our SST when average number of proposals is less than 500, and outperforms the original SST when average number of proposals is more than 500. (Right) Different tIoU v.s. recall of 1000 average proposals. The MA-SST strongly outperforms the original SST and outperform our SST by a maximal margin of $2 . 7 \\%$ . " + ], + "image_footnote": [], + "bbox": [ + 183, + 98, + 810, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 402, + 823, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 TOWARDS MORE LIGHT-WEIGHT RECURRENT MEMORY NETWORKS", + "text_level": 1, + "bbox": [ + 173, + 455, + 758, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It’s known that the speed and computer memory consumption have important influences on the practicability of an RNN, especially when external memory is involved. We have shown in some of the algorithmic tasks that the ARMIN can be trained 3 4 times faster than the NTM in terms of wall clock time. Next we conduct a more thorough comparison experiment among the vanilla LSTM and memory networks under different hidden sizes. The experiment is based on the character-level language modelling task on Penn Treebank dataset, without the loss of generality. For a fair comparison, we implement all networks with near-optimal implementations using Pytorch, an efficient deep-learning framework. We use the aforementioned experiment setup, and keep all memory matrices the same size, i.e. $2 0 \\times d _ { h }$ . We run the experiment using single-precision floating point calculations under a Titan XP GPU that has 12GB memory space. The results are depicted in figure 3. From the results we observe 2 important phenomena: a) The ARMIN consistently outperform other memory networks shown in the graph in terms of running speed both at training and inference stages, and the main contribution to this comes from the simple auto-addressing mechanism of the ARMIN. Moreover, at inference stage, we can replace the memory matrix with a list of discrete memory slots, and update memory by simply replacing the old hidden states with the new ones, furthermore, we can replace the gumbel-softmax function with argmax. Using these methods, ARMIN’s inference speed obtains significant improvement than in training stage. b) The ARMIN outperforms TARDIS and ARMIN $\\dotplus$ TARDIS-addr in terms of training memory consumption, mainly because the TARDIS addressing mechanism requires concatenating $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $\\boldsymbol { h } _ { t - 1 }$ to every memory cell for parallelization. At inference stage, we observe TARDIS have smaller memory consumption than ARMIN and ARMIN $+ { \\bf \\Phi }$ TARDIS-addr, mainly because TARDIS has smaller parameter counts under the same hidden size. However, we would like to point out that under the same parameter counts the ARMIN usually outperforms TARDIS as we show in algorithm and language modelling tasks. We also observe the NTM has the largest memory consumption at inference time, mainly because its complex addressing mechanism. ", + "bbox": [ + 174, + 488, + 825, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have shown in the ablation study (see APPENDIX E) that our network is robust to small iteration length when trained with TBPTT. We realize that we can use this feature to enable large batch size, and boost ARMIN’s training speed while keeping better performance than LSTM. To validate this benefit, we conduct an experiment to compare the performance and cost of ARMIN against our best performing LSTM. We also provide the test result of the same LSTM under TBPTT length $T _ { t r u n c } = 1 0 0$ for better comparing. The hyperparameter setups and results are shown in Table 4. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By comparing ARMIN setup 1 with LSTM setup 2, we observe that with only $6 . 7 \\%$ more memory consumption, we obtain $4 0 \\%$ training speed gain while keeping a slightly better BPC performance and $1 6 \\%$ less parameter count; by comparing ARMIN setup 2 with LSTM setup 2, we observe that with only $8 . 8 \\%$ more memory consumption, we obtain about 1.5 points of BPC performance gain and $7 . 1 \\%$ training speed gain under similar parameter count. ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fb603389200962c80fad3cc21f0db454db3b03532b7db183bfb4f4b91ccf1b51.jpg", + "image_caption": [ + "Figure 3: The running speed and memory consumption at the training and inference stages $( n _ { m e m } = 2 0 )$ ). The solid blocks shows the memory consumption and the curves denote the running speed in characters/s. (a) shows the training stage, and (b) shows the inference stage (note that ARMIN $+$ TARDIS-addr has basically the same memory consumption with ARMIN, so it’s not shown in the graph). " + ], + "image_footnote": [], + "bbox": [ + 184, + 196, + 813, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 469, + 318, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have introduced the ARMIN, a light-weight and memory-augmented RNN architecture with a novel ARMIN cell. The ARMIN incorporates an efficient external memory with the light-weight auto-addressing mechanism. We demonstrate competitive performance of ARMIN in various tasks, and shows the generality of our model. Our ablation study suggests the efficacy of our external memory and addressing mechanism, and notably, our network is robust to shortlength TBPTT which enables using large batch size to speed up the training and further increase performance on sequence modelling tasks. Further research may lead to the efficient hierarchical and multi-scale structures of the ARMIN and its successful applications in Seq2Seq models and complicated reasoning tasks. ", + "bbox": [ + 173, + 502, + 825, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 648, + 285, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. ", + "bbox": [ + 173, + 671, + 823, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst: Single-stream temporal action proposals. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6373–6382. IEEE, 2017. ", + "bbox": [ + 176, + 710, + 823, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Fast temporal activity proposals for efficient detection of human actions in untrimmed videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1914–1923, 2016. ", + "bbox": [ + 174, + 762, + 823, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Victor Campos Camunez, Brendan Jou, Xavier Giro Nieto, Jordi Torres Vi ´ nals, and Shih-Fu Chang. ˜ Skip rnn: learning to skip state updates in recurrent neural networks. In Sixth International Conference on Learning Representations: Monday April 30-Thursday May 03, 2018, Vancouver Convention Center, Vancouver:[proceedings], pp. 1–17, 2018. ", + "bbox": [ + 181, + 814, + 821, + 872 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨ ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Gated feedback recurrent neural networks. In International Conference on Machine Learning, pp. 2067–2075, 2015. ", + "bbox": [ + 169, + 103, + 823, + 133 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. Hierarchical multiscale recurrent neural networks. arXiv preprint arXiv:1609.01704, 2016. ", + "bbox": [ + 171, + 142, + 823, + 171 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ivo Danihelka, Greg Wayne, Benigno Uria, Nal Kalchbrenner, and Alex Graves. Associative long short-term memory. In Proceedings of the 33rd International Conference on International Conference on Machine Learning-Volume 48, pp. 1986–1994. JMLR. org, 2016. ", + "bbox": [ + 173, + 181, + 825, + 226 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A Smith. Transition-based dependency parsing with stack long short-term memory. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), volume 1, pp. 334–343, 2015. ", + "bbox": [ + 174, + 234, + 825, + 292 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jeffrey L Elman. Finding structure in time. Cognitive science, 14(2):179–211, 1990. ", + "bbox": [ + 173, + 301, + 728, + 318 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep action proposals for action understanding. In European Conference on Computer Vision, pp. 768–784. Springer, 2016. ", + "bbox": [ + 174, + 327, + 823, + 371 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiyang Gao, Zhenheng Yang, Chen Sun, Kan Chen, and Ram Nevatia. Turn tap: Temporal unit regression network for temporal action proposals. In Computer Vision (ICCV), 2017 IEEE International Conference on, pp. 3648–3656. IEEE, 2017. ", + "bbox": [ + 173, + 381, + 825, + 424 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a continuous cache. arXiv preprint arXiv:1612.04426, 2016. ", + "bbox": [ + 171, + 434, + 823, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014. ", + "bbox": [ + 173, + 473, + 825, + 502 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´ et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538 (7626):471, 2016. ", + "bbox": [ + 174, + 512, + 825, + 569 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp. 1828–1836, 2015. ", + "bbox": [ + 173, + 580, + 823, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Caglar Gulcehre, Sarath Chandar, and Yoshua Bengio. Memory augmented neural networks with wormhole connections. arXiv preprint arXiv:1701.08718, 2017. ", + "bbox": [ + 173, + 633, + 821, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Emil Julius Gumbel. Statistical theory of extreme values and some practical applications: a series of lectures. Number 33. US Govt. Print. Office, 1954. ", + "bbox": [ + 174, + 672, + 823, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dashan Guo, Wei Li, and Xiangzhong Fang. Fully convolutional network for multiscale temporal action proposals. IEEE Transactions on Multimedia, 2018. ", + "bbox": [ + 174, + 712, + 823, + 741 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. ", + "bbox": [ + 173, + 751, + 815, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 171, + 776, + 823, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. ", + "bbox": [ + 173, + 816, + 823, + 845 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "YG Jiang, Jingen Liu, A Roshan Zamir, G Toderici, I Laptev, Mubarak Shah, and Rahul Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. ", + "bbox": [ + 171, + 856, + 823, + 885 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In Advances in neural information processing systems, pp. 190–198, 2015. ", + "bbox": [ + 174, + 895, + 820, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nan Rosemary Ke, Anirudh Goyal, Olexa Bilaniuk, Jonathan Binas, Michael C Mozer, Chris Pal, and Yoshua Bengio. Sparse attentive backtracking: Temporal creditassignment through reminding. arXiv preprint arXiv:1809.03702, 2018. ", + "bbox": [ + 174, + 103, + 821, + 146 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 174, + 159, + 823, + 188 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Krueger, Tegan Maharaj, Janos Kramar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary Ke, Anirudh Goyal, Yoshua Bengio, Aaron Courville, and Christopher Pal. Zoneout: Regularizing rnns by randomly preserving hidden activations. 2016. ", + "bbox": [ + 174, + 202, + 821, + 244 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. arXiv preprint arXiv:1511.06392, 2015. ", + "bbox": [ + 173, + 257, + 823, + 286 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Shuai Li, Wanqing Li, Chris Cook, Ce Zhu, and Yanbo Gao. Independently recurrent neural network (indrnn): Building a longer and deeper rnn. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5457–5466, 2018. ", + "bbox": [ + 173, + 299, + 826, + 342 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. ", + "bbox": [ + 171, + 354, + 823, + 385 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Toma´s Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and Jan Cer- ˇ nocky. Subword language modeling with neural networks. preprint (http://www. fit. vutbr. cz/imikolov/rnnlm/char. pdf), 8, 2012. ", + "bbox": [ + 174, + 397, + 825, + 440 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Asier Mujika, Florian Meier, and Angelika Steger. Fast-slow recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 5915–5924, 2017. ", + "bbox": [ + 169, + 453, + 823, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tony A Plate. Holographic reduced representation: Distributed representation for cognitive structures. 2003. ", + "bbox": [ + 173, + 494, + 821, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jack Rae, Jonathan J Hunt, Ivo Danihelka, Timothy Harley, Andrew W Senior, Gregory Wayne, Alex Graves, and Tim Lillicrap. Scaling memory-augmented neural networks with sparse reads and writes. In Advances in Neural Information Processing Systems, pp. 3621–3629, 2016. ", + "bbox": [ + 173, + 536, + 825, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by backpropagating errors. nature, 323(6088):533, 1986. ", + "bbox": [ + 169, + 593, + 823, + 622 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Oneshot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016. ", + "bbox": [ + 173, + 635, + 825, + 665 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958, 2014. ", + "bbox": [ + 176, + 678, + 823, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Joseph Suarez. Language modeling with recurrent highway hypernetworks. In Advances in Neural Information Processing Systems, pp. 3267–3276, 2017. ", + "bbox": [ + 173, + 733, + 823, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26– 31, 2012. ", + "bbox": [ + 173, + 775, + 823, + 818 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pp. 4489–4497, 2015. ", + "bbox": [ + 173, + 832, + 821, + 875 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. ", + "bbox": [ + 173, + 887, + 823, + 916 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "APPENDIX A RNN AND LONG SHORT-TERM MEMORY", + "text_level": 1, + "bbox": [ + 174, + 101, + 645, + 119 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A recurrent neural network (RNN) is a class of neural network that recurrently processes a sequence of inputs $\\{ \\pmb { x } _ { 1 } , \\pmb { x } _ { 2 } , . . . , \\pmb { x } _ { T } \\}$ , and returns a sequence of outputs $\\left\\{ y _ { 1 } , y _ { 2 } , . . . , y _ { T } \\right\\}$ where $\\pmb { x } _ { t } \\in \\mathbb { R } ^ { d _ { i } }$ , $\\pmb { y } _ { t } \\in \\mathbb { R } ^ { d _ { o } }$ , and $d _ { i } , d _ { o }$ is the input size and output size, respectively. In a vanilla RNN, $\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { y } } _ { i } }$ is given by following equations: ", + "bbox": [ + 173, + 133, + 825, + 190 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/804120d85e7408afb1231cfe5c6429e74bfe78e522a34d4170c11f8c2955407d.jpg", + "text": "$$\n\\begin{array} { r } { \\pmb { h } _ { t } = \\sigma ( \\pmb { U } \\pmb { x } _ { t } + \\pmb { W } \\pmb { h } _ { t - 1 } + \\pmb { b } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 393, + 191, + 601, + 209 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/5a5db4a4e05fdc2a17d697e8d5c04f697686435e05ad87f9bd6a5cdda4e3d158.jpg", + "text": "$$\n\\pmb { y } _ { t } = \\pmb { V } \\pmb { h } _ { t } + \\pmb { c } ,\n$$", + "text_format": "latex", + "bbox": [ + 442, + 215, + 553, + 232 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where the hidden state $\\pmb { h } _ { t } ~ \\in ~ \\mathbb { R } ^ { d _ { h } }$ is passed to the next recurrence. $d _ { h }$ is the hidden size, and $U , W , V , b , c$ are learnable parameters. ", + "bbox": [ + 173, + 238, + 826, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The LSTM (Hochreiter & Schmidhuber, 1997) is designed to overcome the gradient vanishing and exploding problem in vanilla RNN and learn long-term dependencies, by passing a tuple of $\\left( { { h _ { t } } , { c _ { t } } } \\right)$ to the next recurrence. Its computation process is defined as follows: ", + "bbox": [ + 173, + 272, + 825, + 315 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/ad5becc79b9d06847187e14ed27650233d88b8c613e95d1246b529200f1a4b81.jpg", + "text": "$$\n\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\end{array} \\right\\} W _ { 0 } [ \\ : x _ { t } , h _ { t - 1 } \\ : ] + b _ { 0 } ,\n$$", + "text_format": "latex", + "bbox": [ + 357, + 324, + 640, + 393 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/7e19929b1d6fbfaad66140021ec43d3b6a8e073f83af83a84b1d6717e68449a7.jpg", + "text": "$$\n\\begin{array} { r } { \\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + \\pmb { i } _ { t } \\circ \\pmb { g } _ { t } , } \\\\ { \\pmb { h } _ { t } = \\pmb { o } _ { t } \\circ \\operatorname { t a n h } ( \\pmb { c } _ { t } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 410, + 401, + 586, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $\\circ$ is the element-wise product, $W _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } \\times ( d _ { i } + d _ { h } ) }$ , $b _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } }$ , $i _ { t } , f _ { t } , g _ { t } o _ { t }$ are the input gate, forget gate, cell state, output gate at time $t$ , respectively, and they control the information flow in the LSTM. ", + "bbox": [ + 174, + 450, + 825, + 494 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "APPENDIX B DIFFERENCES BETWEEN TARDIS AND ARMIN ", + "text_level": 1, + "bbox": [ + 176, + 516, + 700, + 534 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The major differences between ARMIN and TARDIS networks are the recurrent cell computation and read operation. We show these differences respectively in subsections below: ", + "bbox": [ + 173, + 549, + 825, + 578 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B.1 CELL COMPUTATIONS ", + "text_level": 1, + "bbox": [ + 174, + 597, + 369, + 611 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The cell computation process of TARDIS is as follow: ", + "bbox": [ + 174, + 622, + 531, + 638 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/cfc57efefd21d85bfcfac35b79da40758e958d2a8066e4b1422a2ef153775068.jpg", + "text": "$$\n\\left\\{ \\begin{array} { l } { i _ { t } } \\\\ { f _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { l } { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W [ \\ : x _ { t } , h _ { t - 1 } , r _ { t } \\ : ] + b \\ : ,\n$$", + "text_format": "latex", + "bbox": [ + 364, + 657, + 633, + 709 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/71101ff4ff32c3594e94661b93bf7d8f5fafd4bb91c5b49209a91828cd7b2ce5.jpg", + "text": "$$\n\\left\\{ \\begin{array} { l } \\alpha _ { t } \\right\\} = \\left\\{ \\begin{array} { l l } { \\mathrm { g u m b e l - s i g m o i d } } \\\\ { \\mathrm { g u m b e l - s i g m o i d } } \\end{array} \\right\\} \\left\\{ \\begin{array} { l } { \\pmb { w } ^ { \\alpha \\top } } \\\\ { \\pmb { w } ^ { \\beta \\top } } \\end{array} \\right\\} \\left[ \\begin{array} { l } { \\pmb { x } _ { t } , \\pmb { h } _ { t - 1 } , \\pmb { r } _ { t } } \\end{array} \\right] , \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 315, + 717, + 681, + 760 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/b95d0d07a5d5e4b4e3a719b99db47ce964570506dd1749bd579e978cf18903e1.jpg", + "text": "$$\ng _ { t } = \\operatorname { t a n h } ( W ^ { g } [ x _ { t } , \\alpha _ { t } h _ { t - 1 } , \\beta _ { t } { r } _ { t } ] + b ^ { g } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 361, + 766, + 633, + 785 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/d3e8560b4874992f9a038ec70f36d44c13ab400568e9a9915d7768bafc583dd1.jpg", + "text": "$$\n\\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ,\n$$", + "text_format": "latex", + "bbox": [ + 406, + 791, + 588, + 809 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/21307b6c091a805149c17bb66defa7ee661ed633dcc7e674f35b9b58ce3a98bc.jpg", + "text": "$$\nh _ { t } = o _ { t } \\circ \\operatorname { t a n h } ( c _ { t } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 428, + 814, + 568, + 832 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $h _ { t - 1 } , h _ { t } , c _ { t } , i _ { t } , f _ { t } , o _ { t } , g _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { r } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\pmb x } _ { t } \\in \\mathbb { R } ^ { d _ { i } } , { \\pmb w } ^ { \\alpha } , { \\pmb w } ^ { \\beta } \\in \\mathbb { R } ^ { d _ { i } + d _ { h } + d _ { r } }$ , and $W \\in$ R3dh×(di+dh+dr) $, \\pmb { W } ^ { g } \\in \\mathbb { R } ^ { d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } \\in \\mathbb { R } ^ { 3 d _ { h } } , \\pmb { b } ^ { g } \\in \\mathbb { R } ^ { d _ { h } }$ . The output of TARDIS is the concatenation of $\\boldsymbol { h } _ { t }$ and $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ . ", + "bbox": [ + 173, + 843, + 823, + 888 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "There are 3 important points that distinguish the ARMIN cell from TARDIS cell, which are as follows: ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "• The TARDIS uses two binary scalar gates $\\alpha _ { t } , \\beta _ { t }$ to control the information flow in equation 17. The two scalar gates are activated using gumbel-sigmoid (which is similar to gumbel-softmax and the output is close to binary), whereas the ARMIN uses two soft vector gates to control the information flow, which is more flexible and enable larger model capacity. For example, in the charlevel language modelling experiments, if we replace the sigmoid activations in equation 3 with the element-wise gumbel-sigmoid, the BPC performance immediately drops from 1.202 to 1.235. ", + "bbox": [ + 174, + 103, + 823, + 188 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "• The TADIS apply information control only when calculating the cell state $\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { g } } _ { t } }$ , whereas the ARMIN apply information control immediately after the read operation. Our main motivation to this change is that if we don’t filter out irrelevant information in $\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }$ and $h _ { t }$ at first, these noisy information will cause the weight parameters hard to learn. Furthermore, if the read operation chooses the wrong past state, we can block the information flow for $\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }$ at the first place. ", + "bbox": [ + 174, + 191, + 825, + 262 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "• The ARMIN has an extra gate $\\mathbf { \\delta } _ { o _ { t } ^ { r } }$ to control the output from $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ as in equation 8, which the TARDIS doesn’t have. If we remove this gate and directly output $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ like in TARDIS, we get a BPC performance drop from 1.202 to 1.219 in char-level language modelling task. ", + "bbox": [ + 173, + 266, + 821, + 309 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Finally, despite that the ARMIN has 3 more vector gates than the TARDIS, we have shown in the $p \\mathrm { M N I S T }$ and Penn Treebank experiments that the ARMIN can still outperform the TARDIS under similar parameter counts. ", + "bbox": [ + 174, + 321, + 825, + 363 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.2 READ OPERATION", + "text_level": 1, + "bbox": [ + 176, + 381, + 344, + 395 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The memory matrix $M _ { t }$ of TARDIS has disjoint address section $\\mathbf { \\boldsymbol { A } } _ { t } ~ \\in ~ \\mathbb { R } ^ { n _ { m e m } \\times d _ { k } }$ and content section $C _ { t } \\ \\stackrel { \\cdot } { \\in } \\ \\mathbb { R } ^ { n _ { m e m } \\times d _ { r } }$ , $M _ { t } = [ A _ { t } ; C _ { t } ]$ . The controller reads both the address and the content parts of the memory, but it will only write into the content section of the memory. ", + "bbox": [ + 174, + 407, + 825, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The TARDIS paper explored two ways to sample the read locations——using reinforce and using gumbel-softmax function, respectively. The gumbel-softmax way shows a better performance and its read operation can be formulated as follow: ", + "bbox": [ + 174, + 457, + 823, + 500 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/93c16322c16a353d9081e01081b4078d9a8374ee48f55d406ae85acf775e9d61.jpg", + "text": "$$\n\\begin{array} { c } { { \\pi _ { t } [ i ] = a ^ { \\top } \\operatorname { t a n h } ( W ^ { \\gamma } [ h _ { t } , x _ { t } , M _ { t } [ i ] , { \\boldsymbol u } _ { t } ] + b ^ { \\gamma } ) , } } \\\\ { { s _ { t } = \\mathrm { g u m b e l - s o f t m a x } ( \\pi _ { t } ) . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 341, + 516, + 655, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where $\\{ a , W ^ { \\gamma } , b ^ { \\gamma } \\}$ are learnable parameters (note that in our experiments, we always set the dimension of $\\textbf { \\em a }$ to about 1/4 of the hidden size) and $\\mathbf { \\Delta } \\mathbf { u } _ { t }$ is a hand-crafted usage vector which denotes the frequency of accesses to each cell in the memory. ", + "bbox": [ + 176, + 565, + 823, + 608 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "There are two important differences with respect to TARDIS addressing mechanism and our autoaddressing mechanism: ", + "bbox": [ + 173, + 614, + 821, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "• In practice, equation 21 requires concatenating $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $h _ { t }$ to every memory cells $M _ { t } [ i ]$ for parallelization, which often causes significant extra memory and time consumption compared to the auto-addressing of ARMIN, as we show in the algorithmic tasks and char-level language modelling task. ", + "bbox": [ + 174, + 656, + 825, + 712 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "• The auto-addressing omits the learnable address section $\\pmb { A } _ { t }$ , the hand-crafted $\\mathbf { \\pmb { u } } _ { t }$ vector, the tanh activation and an auxiliary vector $^ { a }$ compared to TARDIS addressing mechanism. The TARDIS also requires manually subtracting 100 in unnormalized probability at last read location to avoid repeated read, whereas the auto-addressing doesn’t need to. With these simplifications, we found the auto-addressing mechanism easier to train and it leads to a much faster converge speed, as we show in the algorithmic tasks. We also observe a BPC performance growth from 1.223 to 1.202 in char-level language modelling task when we switch from ARMIN $^ +$ TARDIS-addr to ARMIN with auto-addressing. ", + "bbox": [ + 174, + 717, + 825, + 829 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "APPENDIX C EXPERIMENT SETUPS ", + "text_level": 1, + "bbox": [ + 178, + 849, + 482, + 866 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We first describe some universal setups in all our experiments (unless we mention them again with different setup). Our model can be trained end-to-end, due to the utilization of gumbel-softmax function. As pointed out in Maddison et al. (2016), when the temperature $\\tau \\leq ( n - 1 ) ^ { - 1 }$ , the density function of gumbel-softmax estimator is log-convex, which gives us a good guarantee for optimization. However, smaller $\\tau$ will also bring higher variance of the gradients (Jang et al., 2016). So in all experiments, we initialize the reciprocal of $\\tau$ as 1, and increase it by 1 after each epoch, until the reciprocal equals to $n - 1$ . This scheme has shown better performance than other schemes. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/78261b201b724626971a34ad21df1c98e3584c9aec6b3dee323832c606b2286f.jpg", + "table_caption": [ + "Table 5: Hyperparameters of the models for algorithmic tasks. " + ], + "table_footnote": [], + "table_body": "
ModelHidden Sizememory SizeParam Count
LSTM300一376k
NTM120128×2088k
TARDIS12050×3290k
ARMIN+TARDIS10050×3291k
ARMIN10050×3290k
", + "bbox": [ + 320, + 132, + 678, + 239 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/3cd63d277de72abed816edd48b7040e7ae1701681a9dcf4a3cd4fb53edb5f175.jpg", + "table_caption": [ + "Table 6: Hyperparameters of the algorithmic tasks. " + ], + "table_footnote": [], + "table_body": "
TaskSetup
Copycopy length =1~50
Repeat Copycopyl length=1~10 number of repeats = 1~10
Associative Recallitem size=3×6 item number = 2~6
Priority SortSort length = 40
", + "bbox": [ + 330, + 291, + 666, + 401 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 435, + 825, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We apply Layer Normalization(Ba et al., 2016) to all networks. In the ARMIN case, layer norm is applied to equation 3, 6 and 7, which we find important to regulate the hidden state when scale grows. To avoid overfitting, we apply dropout (Srivastava et al., 2014) to the input and output layers, and Zoneout(Krueger et al., 2016) is applied for recurrent connections. The networks are trained with Adam optimizer (Kingma & Ba, 2014). All weight matrices in RNNs are initialized to orthogonal matrices, and the bias of the forget gate $\\pmb { f } _ { t }$ is initialized to 1. For all models using TARDIS addressing mechanism, we set the memory key size to about 1/5 of the memory content size. ", + "bbox": [ + 174, + 497, + 825, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C.1 ALGORITHMIC TASKS ", + "text_level": 1, + "bbox": [ + 176, + 628, + 367, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In all algorithmic experiments below, following Graves et al. (2014), we apply RMSProp optimizer (Tieleman & Hinton, 2012) with a momentum of 0.9. The learning rate is 0.0001 and gradient norms are clipped to 10.0. No regularization is used as the size of the networks is small. Following Graves et al. (2014), we have learned bias of initial hidden states and memory matrices for all memory networks. For TARDIS, ARMIN $+ { \\bf \\Phi }$ TARDIS-addr and ARMIN we increase the reciprocal of temperature $\\tau$ by 1 for every 200 iterations. The hyperparamters for models and tasks are shown in Table 5 and Table 6, respectively. We explain these tasks in detail in subsections below: ", + "bbox": [ + 174, + 655, + 825, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C.1.1 COPY ", + "text_level": 1, + "bbox": [ + 174, + 771, + 271, + 785 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The copy task tests whether a recurrent model can correctly store and recall a long sequence of arbitrary information. In our experiment, the networks are given randomly generated eight-bit binary sequences with its length ranging from 1 to 50, followed by a finish indicator. Then the networks are asked to output the same sequence with the input. ", + "bbox": [ + 174, + 796, + 825, + 853 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C.1.2 REPEAT COPY ", + "text_level": 1, + "bbox": [ + 174, + 869, + 330, + 885 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The repeat copy task is a repeat version of the copy task. The main motivation was to see if a recurrent model could memorize input information and repeatedly reuse them at later time-steps. In ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "our experiment, the networks are asked to repeatedly copy the randomly generated input sequence with length ranging from 1 to 10 to output for 1 to 10 times. ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C.1.3 ASSOCIATIVE RECALL ", + "text_level": 1, + "bbox": [ + 174, + 154, + 387, + 169 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This task test whether a memory network has the ability to retrieve memory data indirectly. The networks are given a sequence of items and then a query with one of the items, and are asked to output the subsequent item. In our experiments, each item consists of three six-bit binary vectors, and the number of items are ranging from 2 to 6. ", + "bbox": [ + 174, + 181, + 825, + 237 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C.1.4 PRIORITY SORT ", + "text_level": 1, + "bbox": [ + 174, + 260, + 341, + 275 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This task test whether a recurrent model can do priority sort——a classic elementary algorithm. This task is more challenging than previous ones as it not only require memorizing the input sequences but also making efficient use of the mutual relations of the items in the input sequences. In our experiments, the networks are given 40 eight-bit random binary keys with scalar priority values in a sequence, and are asked to output the top 30 keys in descending order of priority. ", + "bbox": [ + 174, + 287, + 825, + 357 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C.2 CHARACTER-LEVEL LANGUAGE MODELLING ON PENN TREENBANK DATASET ", + "bbox": [ + 176, + 382, + 751, + 396 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C.2.1 PENN TREEBANK DATASET ", + "text_level": 1, + "bbox": [ + 176, + 410, + 416, + 424 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We use the train/validation/test split outlined in Mikolov et al. (2012). The train/validation/test batch size are 128/64/1 respectively. Following prior works on this task (Ha et al., 2016; Suarez, 2017; Mujika et al., 2017), the model is optimized over the cross-entropy loss between the predictions and the training labels, and evaluated using a bits-per-character measure. We use a single ARMIN cell of different hidden sizes, with a fully connected layer to output the predictive probabilities. The input embedding size $d _ { i }$ is set to 128. The number of memory slots $n _ { m e m }$ is 20. The zoneout probability is 0.3. We train the networks for 200 epochs with a learning rate of 0.002, and decay with a factor of 10 at the last 20 epochs. We clip the gradients to a maximum norm of 1.0. ", + "bbox": [ + 174, + 438, + 825, + 549 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C.2.2 HUTTER PRIZE WIKIPEDIA DATASET ", + "text_level": 1, + "bbox": [ + 176, + 571, + 478, + 585 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We use the train/validation/test split outlined in Chung et al. (2015). The train/validation/test batch size are 128/64/1 respectively, and the vocabulary size is 205. The zoneout and dropout probabilities are 0.3 and 0.2, respectively. The input embedding size is 256. Our ARMIN network has 2 layer and 800 units at each layer, connected by a linear layer to convert 1600-dimensional features from layer 1 to 256 dimension as input to layer 2. $n _ { m e m }$ for each layer is set to 10. We train the networks for 50 epochs with a learning rate of 0.001, and decay with a factor of 10 at the last 10 peochs. We clip the gradients to a maximum norm of 1.0. ", + "bbox": [ + 174, + 598, + 825, + 695 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14 ", + "text_level": 1, + "bbox": [ + 176, + 720, + 686, + 734 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The SST takes in sequential inputs of a video and outputs the existing probability predictions of action proposals with different time lengths that end at each time-step, and uses a multi-label loss to optimize its parameters. The sequential inputs are 4096-dimensional features extracted from every 16 consecutive video frames using C3D network (Tran et al., 2015), followed by PCA to downsize the input feature dimensionality. ", + "bbox": [ + 174, + 750, + 825, + 819 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Our experiment setup is basically the same as the SST paper, except that we use a simple fully connected layer to downsize the input features instead of using PCA. The input size of the two networks are both set to 256. The MA-SST have 256 hidden units, while the SST with LSTM have 512 hidden units. This ensures the 2 networks have roughly the same parameter count. The dropout and zoneout probabilities are both set to 0.2 and 0.3. $n _ { m e m }$ of the MA-SST is set to 35. We train the networks for 300 epochs with initial learning rate of 0.002, and decay with a factor of 10 at the last 60 epochs. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/50001ffd2185dd7661fc9aba2d9d75316cd989dcf8de377293cc7dbfab09dd11.jpg", + "image_caption": [ + "Figure 4: Copy task. " + ], + "image_footnote": [], + "bbox": [ + 181, + 109, + 816, + 320 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/4550719af5c0f00a37cd0f179fc1ec6a044179f516a3a2932ce1e6936572202f.jpg", + "image_caption": [ + "Figure 5: Repeat copy task. " + ], + "image_footnote": [], + "bbox": [ + 183, + 376, + 810, + 587 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/040994a58270f2a89222cd44b6a7cbcbc8407ddd4dfecf1d485eb625bbffc5c4.jpg", + "image_caption": [ + "Figure 6: Associative recall task. " + ], + "image_footnote": [], + "bbox": [ + 181, + 695, + 805, + 905 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/4cb5157c10e72c964f7ac67806728f1c6305e97e3dd30e339a3d94d5e8a27dc8.jpg", + "image_caption": [ + "Figure 7: Priority sort task. " + ], + "image_footnote": [], + "bbox": [ + 183, + 111, + 816, + 320 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/2c719e270172f3bfec5784ed1d4baff76a9ce737a4b574ccd9f5a06794ee2ff7.jpg", + "table_caption": [ + "Table 7: Ablation on Penn Treebank $\\scriptstyle ( d _ { h } = 8 0 0 )$ . " + ], + "table_footnote": [], + "table_body": "
Abaltion ModelBPC
No Control Gates1.354
LSTM (TBPTT length=50) ARMIN(TBPTT length=50)1.390 1.220
Random Read1.224
Queue-style Write1.352
Independent Read/write1.245
SmoothRead/write1.212
", + "bbox": [ + 364, + 375, + 632, + 496 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "APPENDIX D TRAINING CURVES ON ALGORITHMIC TASKS ", + "bbox": [ + 176, + 521, + 671, + 536 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "APPENDIX E ABLATION STUDY ON MEMORY AND ADDRESSING MECHANISM ", + "text_level": 1, + "bbox": [ + 176, + 555, + 821, + 570 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "In this subsection, we show the efficacy of our external memory and addressing mechanism by ablation experiments. We choose the char-level language modelling task to do the ablation research as it is a relatively standard benchmark task and has appropriate medium scale. We use the aforementioned experiment setup. To validate the efficacy of our external memory, we conduct the experiments as follows and the results are depicted in table 7: ", + "bbox": [ + 174, + 585, + 825, + 655 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• Short TBPTT Length We shorten the truncated length in TBPTT from 150 to 50. The gradients are truncated so we can’t do backpropagation on hidden states or memory from last iteration, but the ARMIN still has direct access to historical hidden states. By comparing the performance of the LSTM and ARMIN under such circumstances, we can know the efficacy of the external memory. The results show that the LSTM has a dramatic performance drop of 1.390 BPC while the ARMIN only has a minor performance drop of 1.220 BPC. The results confirm that the direct access to memory is crucial when the networks can’t access historical information by backpropagating gradients. And notably, our experiment shows that the ARMIN is very robust to short-length TBPTT, i.e. small iteration length, which means lower memory consumption. It enables using very large batch size to speed up the training and further increasing performance on sequence modelling tasks. ", + "bbox": [ + 173, + 667, + 825, + 808 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• No Control Gates We remove the control gate $\\pmb { g } _ { t } ^ { h }$ and $\\pmb { g } _ { t } ^ { r }$ in ARMIN cell as depicted in equation 3, and directly use $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $\\boldsymbol { h } _ { t - 1 }$ to compute the gates in equation 6. The result shows a BPC of 1.354 that has a dramatic performance drop compared the 1.202 BPC of normal ARMIN. From this, we deduce that the control gates successfully filter out irrelevant information. Without the control gates, the historical state $\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }$ would have bring more noise than useful information. ", + "bbox": [ + 174, + 811, + 823, + 882 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "To validate the efficacy of the addressing mechanism of the ARMIN, we do the following modifications to the read/write operations of the ARMIN, and the results are also depicted in table 7: ", + "bbox": [ + 174, + 895, + 820, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• Random Read The ARMIN randomly chooses a historical state while reading. The result shows a BPC of 1.224, compared to the 1.202 BPC of normal ARMIN. It proves that the read regime we proposed in section 3.2 is effective and can choose more useful historical state than random read. ", + "bbox": [ + 173, + 103, + 823, + 145 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "• Queue-style Write While writing, the ARMIN follows a fashion of first-in-first-out, i.e. always chooses the oldest memory slot to write. It makes the memory $M$ stubbornly store the hidden states of the last 20 time-steps. The result shows a BPC of 1.352, which proves our normal read/write mechanism has much more flexibility and can learn long-term patterns. ", + "bbox": [ + 174, + 150, + 825, + 205 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "• Independent Read/write While writing, the ARMIN doesn’t write to the last read memory slot, but independently choose a memory slot to overwrite according to $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }$ and $h _ { t }$ , in a similar fashion with the read operation we presented in section 3.2. The result shows a BPC of 1.245, proving that the normal writing regime in the ARMIN can better learn long-term patterns and maximally preserve historical information, as we explain in section 3.3. ", + "bbox": [ + 174, + 210, + 825, + 280 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "• Smooth Read/write We replace the gumbel-softmax function with the vanilla softmax function, so the ARMIN reads and writes in a smooth way. The results show a minor performance drop of 1.212 BPC. The result shows that our discrete read/write regime is still better. Moreover, the smooth read/write regime would consume much more memory at both training and inference stages. ", + "bbox": [ + 174, + 285, + 825, + 340 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/parse/train/HyMS8iRcK7/HyMS8iRcK7_middle.json b/parse/train/HyMS8iRcK7/HyMS8iRcK7_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..ed8123a480a92f73a47f8683a25e0e69165f34e2 --- /dev/null +++ b/parse/train/HyMS8iRcK7/HyMS8iRcK7_middle.json @@ -0,0 +1,46965 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 505, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 505, + 99 + ], + "score": 1.0, + "content": "SEQUENCE MODELLING WITH AUTO-ADDRESSING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 117 + ], + "score": 1.0, + "content": "AND RECURRENT MEMORY INTEGRATING NETWORKS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 211, + 468, + 321 + ], + "lines": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "score": 1.0, + "content": "Processing sequential data with long term dependencies and learn complex transi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "tions are two major challenges in many deep learning applications. In this paper,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "score": 1.0, + "content": "we introduce a novel architecture, the Auto-addressing and Recurrent Memory", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "Integrating Network (ARMIN) to address these issues. The ARMIN explicitly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "score": 1.0, + "content": "stores previous hidden states and recurrently integrate useful past states into cur-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "rent time-step by an efficient memory addressing mechanism. Compared to ex-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "score": 1.0, + "content": "isting memory networks, the ARMIN is more light-weight and inference-time ef-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "ficient. Our network can be trained on small slices of long sequential data, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 300, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 142, + 300, + 469, + 312 + ], + "score": 1.0, + "content": "thus, can boost its training speed. Experiments on various tasks demonstrate the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 443, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 443, + 322 + ], + "score": 1.0, + "content": "efficiency of the ARMIN architecture. Codes and models will be available.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 343, + 206, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 208, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 208, + 359 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 368, + 504, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Recurrent neural networks, such as the Long Short-Term Memory (LSTM) (Hochreiter & Schmid-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "huber, 1997) and Gated Recurrent Unit (GRU) (Cho et al., 2014) have shown promising performance", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "for processing sequential data. However, it’s known that RNNs suffer from gradient vanishing prob-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 401, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 504, + 415 + ], + "score": 1.0, + "content": "lem. Moreover, as pointed out by Rae et al. (2016), the number of parameters grows proportionally", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "score": 1.0, + "content": "to the square of the size of the hidden units, which carry the historical information. Recent memory-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "based approaches exhibit potential to address these issues, by decoupling memory capacity from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 408, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 408, + 448 + ], + "score": 1.0, + "content": "model parameters, and backpropagating the gradients through the memory.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "Neural Turing Machine (NTM) (Graves et al., 2014) first emerged as a recurrent model that incor-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "porates external memory abilities. NTM maintains a memory matrix, and at every time-step, the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 473, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 504, + 486 + ], + "score": 1.0, + "content": "network reads and writes (with erasing) to the memory matrix using certain soft-attentional mech-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "anism, controlled by an LSTM that produces read and write vectors. NTM and its successor, the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "Differentiable Neural Computer(Graves et al., 2016), have shown success on some algorithmic tasks", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "such as copying, priority sorting and some real-world tasks such as question answering. But one lim-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "itation of the NTM is that due to its smooth read and write mechanism, NTM has to do propagations", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "on the entire memory, usually causing huge amount of memory consumption. To this end, Rae et al.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "(2016) proposes the Sparse Access Memory(SAM) network, by thresholding memory modifications", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "to a sparse subset, i.e. all read and write operations are limited to several memory words. This", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "allows memory and time efficient propagations while maintaining NTM’s performance. However,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "these external memory models have relatively complicated memory addressing mechanisms, mak-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "ing them suitable for only large-scale memory. Moreover, the RNN in these models plays a simple", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "role of being a controller, but the gradient vanishing problem of RNN itself is not given attention.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "In contrast to these networks, our network uses a light-weight read mechanism to address a small", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "external memory, and relieves the gradient vanishing problem by directly concatenating memory", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 169, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 169, + 640 + ], + "score": 1.0, + "content": "word to output.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "Inspired by prior memory models, efforts have been made to build a bridge between simple RNNs", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "and complicated memory models. Kurach et al. (2015) propose the Neural Random-access Machines", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "(NRAM) that can manipulate and dereference pointers to an external variable-size random-access", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "memory. Danihelka et al. (2016) improve LSTM with ideas from Holographic Reduced Represen-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "tations (Plate, 2003) that enables key-value storage of data. Grave et al. (2016) propose a method", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "of augmenting LSTM by storing previous (hidden state, input word) pairs in memory and using", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "the current hidden state as a query vector to recover historical input words. This method requires", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "no backpropagation through memory and is well-suited to word-level language tasks. Grefenstette", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 505, + 99 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 505, + 99 + ], + "score": 1.0, + "content": "SEQUENCE MODELLING WITH AUTO-ADDRESSING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 117 + ], + "score": 1.0, + "content": "AND RECURRENT MEMORY INTEGRATING NETWORKS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 211, + 468, + 321 + ], + "lines": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "score": 1.0, + "content": "Processing sequential data with long term dependencies and learn complex transi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "tions are two major challenges in many deep learning applications. In this paper,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "score": 1.0, + "content": "we introduce a novel architecture, the Auto-addressing and Recurrent Memory", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "Integrating Network (ARMIN) to address these issues. The ARMIN explicitly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "score": 1.0, + "content": "stores previous hidden states and recurrently integrate useful past states into cur-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "rent time-step by an efficient memory addressing mechanism. Compared to ex-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "score": 1.0, + "content": "isting memory networks, the ARMIN is more light-weight and inference-time ef-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "ficient. Our network can be trained on small slices of long sequential data, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 300, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 142, + 300, + 469, + 312 + ], + "score": 1.0, + "content": "thus, can boost its training speed. Experiments on various tasks demonstrate the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 443, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 443, + 322 + ], + "score": 1.0, + "content": "efficiency of the ARMIN architecture. Codes and models will be available.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9.5, + "bbox_fs": [ + 141, + 212, + 469, + 322 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 343, + 206, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 208, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 208, + 359 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 368, + 504, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Recurrent neural networks, such as the Long Short-Term Memory (LSTM) (Hochreiter & Schmid-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "huber, 1997) and Gated Recurrent Unit (GRU) (Cho et al., 2014) have shown promising performance", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "for processing sequential data. However, it’s known that RNNs suffer from gradient vanishing prob-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 401, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 504, + 415 + ], + "score": 1.0, + "content": "lem. Moreover, as pointed out by Rae et al. (2016), the number of parameters grows proportionally", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 426 + ], + "score": 1.0, + "content": "to the square of the size of the hidden units, which carry the historical information. Recent memory-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "based approaches exhibit potential to address these issues, by decoupling memory capacity from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 408, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 408, + 448 + ], + "score": 1.0, + "content": "model parameters, and backpropagating the gradients through the memory.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 369, + 506, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "Neural Turing Machine (NTM) (Graves et al., 2014) first emerged as a recurrent model that incor-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "porates external memory abilities. NTM maintains a memory matrix, and at every time-step, the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 473, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 504, + 486 + ], + "score": 1.0, + "content": "network reads and writes (with erasing) to the memory matrix using certain soft-attentional mech-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "anism, controlled by an LSTM that produces read and write vectors. NTM and its successor, the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "Differentiable Neural Computer(Graves et al., 2016), have shown success on some algorithmic tasks", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "such as copying, priority sorting and some real-world tasks such as question answering. But one lim-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "itation of the NTM is that due to its smooth read and write mechanism, NTM has to do propagations", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "on the entire memory, usually causing huge amount of memory consumption. To this end, Rae et al.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "(2016) proposes the Sparse Access Memory(SAM) network, by thresholding memory modifications", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "to a sparse subset, i.e. all read and write operations are limited to several memory words. This", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "allows memory and time efficient propagations while maintaining NTM’s performance. However,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "these external memory models have relatively complicated memory addressing mechanisms, mak-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "ing them suitable for only large-scale memory. Moreover, the RNN in these models plays a simple", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "role of being a controller, but the gradient vanishing problem of RNN itself is not given attention.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "In contrast to these networks, our network uses a light-weight read mechanism to address a small", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "external memory, and relieves the gradient vanishing problem by directly concatenating memory", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 169, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 169, + 640 + ], + "score": 1.0, + "content": "word to output.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 452, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "Inspired by prior memory models, efforts have been made to build a bridge between simple RNNs", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "and complicated memory models. Kurach et al. (2015) propose the Neural Random-access Machines", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "(NRAM) that can manipulate and dereference pointers to an external variable-size random-access", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "memory. Danihelka et al. (2016) improve LSTM with ideas from Holographic Reduced Represen-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "tations (Plate, 2003) that enables key-value storage of data. Grave et al. (2016) propose a method", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "of augmenting LSTM by storing previous (hidden state, input word) pairs in memory and using", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "the current hidden state as a query vector to recover historical input words. This method requires", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "no backpropagation through memory and is well-suited to word-level language tasks. Grefenstette", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "et al. (2015); Dyer et al. (2015); Joulin & Mikolov (2015) augment RNNs with a stack structure that", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "works as a natural parsing tool, and use them to process algorithmic and nature language process-", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "ing (NLP) tasks; nonetheless, the speed of stack-augmented RNNs is rather slow due to multiple", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "push-pop operations at every time-step. Ke et al. (2018) propose the Sparse Attentive Backtrack-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "ing (SAB) architecture, which recalls a few past hidden sates at every time-step and do “mental”", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "backpropogations to the nearby hidden states with respect to the past hidden states. Gulcehre et al.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 161 + ], + "score": 1.0, + "content": "(2017) propose the TARDIS network, which is an LSTM-resembled RNN that directly stores a", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 506, + 172 + ], + "score": 1.0, + "content": "fixed number of previous hidden states with learned key parameters for memory addressing. At ev-", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 328, + 183 + ], + "score": 1.0, + "content": "ery time-step, the network reads out one historical state", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 329, + 172, + 339, + 181 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 339, + 169, + 505, + 183 + ], + "score": 1.0, + "content": "from the memory, and uses it, along with", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 182, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 145, + 193 + ], + "score": 1.0, + "content": "the input", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 145, + 183, + 156, + 192 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 157, + 182, + 243, + 193 + ], + "score": 1.0, + "content": "and last hidden state", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 243, + 182, + 265, + 192 + ], + "score": 0.91, + "content": "h _ { t - 1 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 265, + 182, + 385, + 193 + ], + "score": 1.0, + "content": ", to produce new hidden state", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 385, + 182, + 397, + 192 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 397, + 182, + 466, + 193 + ], + "score": 1.0, + "content": ", then overwrites", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 466, + 182, + 478, + 192 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 478, + 182, + 505, + 193 + ], + "score": 1.0, + "content": "to the", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 152, + 206 + ], + "score": 1.0, + "content": "location of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 152, + 194, + 162, + 203 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 163, + 191, + 506, + 206 + ], + "score": 1.0, + "content": ". TARDIS optionally uses the gumbel-softmax estimator(Maddison et al., 2016; Jang", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 255, + 216 + ], + "score": 1.0, + "content": "et al., 2016) to sample the location of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 255, + 204, + 265, + 214 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 265, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "(we will explain this in section 2.1), so the whole network is", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "differentiable. However, the TARDIS still involves some hand-crafted memory addressing method", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 495, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 495, + 238 + ], + "score": 1.0, + "content": "and its addressing mechanism requires a considerable amount of memory and time consumption.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 644, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "et al. (2015); Dyer et al. (2015); Joulin & Mikolov (2015) augment RNNs with a stack structure that", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "works as a natural parsing tool, and use them to process algorithmic and nature language process-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "ing (NLP) tasks; nonetheless, the speed of stack-augmented RNNs is rather slow due to multiple", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "push-pop operations at every time-step. Ke et al. (2018) propose the Sparse Attentive Backtrack-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "ing (SAB) architecture, which recalls a few past hidden sates at every time-step and do “mental”", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "backpropogations to the nearby hidden states with respect to the past hidden states. Gulcehre et al.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 161 + ], + "score": 1.0, + "content": "(2017) propose the TARDIS network, which is an LSTM-resembled RNN that directly stores a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 506, + 172 + ], + "score": 1.0, + "content": "fixed number of previous hidden states with learned key parameters for memory addressing. At ev-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 328, + 183 + ], + "score": 1.0, + "content": "ery time-step, the network reads out one historical state", + "type": "text" + }, + { + "bbox": [ + 329, + 172, + 339, + 181 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 169, + 505, + 183 + ], + "score": 1.0, + "content": "from the memory, and uses it, along with", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 182, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 145, + 193 + ], + "score": 1.0, + "content": "the input", + "type": "text" + }, + { + "bbox": [ + 145, + 183, + 156, + 192 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 182, + 243, + 193 + ], + "score": 1.0, + "content": "and last hidden state", + "type": "text" + }, + { + "bbox": [ + 243, + 182, + 265, + 192 + ], + "score": 0.91, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 182, + 385, + 193 + ], + "score": 1.0, + "content": ", to produce new hidden state", + "type": "text" + }, + { + "bbox": [ + 385, + 182, + 397, + 192 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 182, + 466, + 193 + ], + "score": 1.0, + "content": ", then overwrites", + "type": "text" + }, + { + "bbox": [ + 466, + 182, + 478, + 192 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 182, + 505, + 193 + ], + "score": 1.0, + "content": "to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 152, + 206 + ], + "score": 1.0, + "content": "location of", + "type": "text" + }, + { + "bbox": [ + 152, + 194, + 162, + 203 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 191, + 506, + 206 + ], + "score": 1.0, + "content": ". TARDIS optionally uses the gumbel-softmax estimator(Maddison et al., 2016; Jang", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 255, + 216 + ], + "score": 1.0, + "content": "et al., 2016) to sample the location of", + "type": "text" + }, + { + "bbox": [ + 255, + 204, + 265, + 214 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "(we will explain this in section 2.1), so the whole network is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "differentiable. However, the TARDIS still involves some hand-crafted memory addressing method", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 495, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 495, + 238 + ], + "score": 1.0, + "content": "and its addressing mechanism requires a considerable amount of memory and time consumption.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 241, + 503, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 255 + ], + "score": 1.0, + "content": "Inspired by the TARDIS, we introduce the ARMIN architecture, a more light-weight and memory-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "based architecture with a very simple memory addressing mechanism and in-depth modification of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 404, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 404, + 276 + ], + "score": 1.0, + "content": "the existing LSTM structure. Concretely, our contributions are as follows:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 104, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "• We distill previous memory addressing methods and propose a simple yet effective memory ad-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "dressing mechanism for the external memory, namely Auto-addressing, by encoding the information", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 414, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 290, + 321 + ], + "score": 1.0, + "content": "for memory addressing directly via the inputs", + "type": "text" + }, + { + "bbox": [ + 290, + 309, + 302, + 319 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 308, + 388, + 321 + ], + "score": 1.0, + "content": "and the hidden states", + "type": "text" + }, + { + "bbox": [ + 389, + 308, + 410, + 320 + ], + "score": 0.9, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 308, + 414, + 321 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "• We propose a novel recurrent cell that combines the gating advantages of LSTM and allows", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "direct gradient backpropagation from output to memory. With only a single cell, it learns better", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 331, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 331, + 357 + ], + "score": 1.0, + "content": "representations than many hierarchical RNN structures.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 358, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 104, + 358, + 506, + 373 + ], + "score": 1.0, + "content": "• We show in our ablation study that the ARMIN is robust to small iteration lengths when training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "long sequential data, which enables training with large batch sizes and boost its training speed.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 450, + 395 + ], + "score": 1.0, + "content": "We further demonstrate in char-level language modelling task that the ARMIN obtain", + "type": "text" + }, + { + "bbox": [ + 451, + 381, + 470, + 392 + ], + "score": 0.87, + "content": "4 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "training", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 393, + 451, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 451, + 406 + ], + "score": 1.0, + "content": "speed gain than LSTM while keeping similar performance and memory consumption.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 105, + 407, + 504, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 421 + ], + "score": 1.0, + "content": "• We demonstrate competitive results on various tasks while keeping efficient time and memory", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 419, + 302, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 302, + 431 + ], + "score": 1.0, + "content": "consumption during training and inference time.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 107, + 446, + 200, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 444, + 201, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 201, + 462 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 107, + 471, + 266, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 267, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 267, + 484 + ], + "score": 1.0, + "content": "2.1 GUMBEL-SOFTMAX ESTIMATOR", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "Categorical distribution is a natural choice for representing discrete structure in the world; however,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "it’s rarely used in neural networks due to its inability to backpropagate through samples(Jang et al.,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "2016). To this end, Maddison et al. (2016) and Jang et al. (2016) propose a continuous relaxation of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "categorical distribution and the corresponding gumbel-softmax gradient estimator that replaces the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "non-differentiable sample from a categorical distribution with a differentiable sample. Specifically,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 230, + 560 + ], + "score": 1.0, + "content": "given a probability distribution", + "type": "text" + }, + { + "bbox": [ + 231, + 547, + 311, + 558 + ], + "score": 0.91, + "content": "\\pmb { p } = ( \\pi _ { 1 } , \\pi _ { 2 } , . . . , \\pi _ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 546, + 331, + 560 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 331, + 547, + 338, + 556 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "categories, the gumbel-softmax estimator", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 558, + 429, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 248, + 570 + ], + "score": 1.0, + "content": "produces an one-hot sample vector", + "type": "text" + }, + { + "bbox": [ + 248, + 560, + 255, + 569 + ], + "score": 0.82, + "content": "\\textbf { { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 558, + 289, + 570 + ], + "score": 1.0, + "content": "with its", + "type": "text" + }, + { + "bbox": [ + 289, + 559, + 293, + 568 + ], + "score": 0.8, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 558, + 429, + 570 + ], + "score": 1.0, + "content": "-th element calculated as follows:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 574, + 416, + 605 + ], + "lines": [ + { + "bbox": [ + 193, + 574, + 416, + 605 + ], + "spans": [ + { + "bbox": [ + 193, + 574, + 416, + 605 + ], + "score": 0.93, + "content": "y _ { i } = { \\frac { \\exp ( ( \\log ( \\pi _ { i } ) + g _ { i } ) / \\tau ) } { \\sum _ { j = 1 } ^ { k } \\exp ( ( \\log ( \\pi _ { j } ) + g _ { j } ) / \\tau ) } } \\quad { \\mathrm { f o r ~ } } i = 1 , 2 , . . . , k ,", + "type": "interline_equation", + "image_path": "3d7fbb0ffd4812f959da86e645aa96d65bbe99701c988143802f6d35011fd4d2.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 574, + 416, + 589.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 193, + 589.5, + 416, + 605.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 437, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 438, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 133, + 623 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 612, + 170, + 622 + ], + "score": 0.9, + "content": "g _ { 1 } , . . . , g _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 608, + 438, + 623 + ], + "score": 1.0, + "content": "are i.i.d samples drawn from Gumbel distribution(Gumbel, 1954):", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 627, + 396, + 640 + ], + "lines": [ + { + "bbox": [ + 214, + 627, + 396, + 640 + ], + "spans": [ + { + "bbox": [ + 214, + 627, + 396, + 640 + ], + "score": 0.93, + "content": "g _ { i } = - \\log ( - \\log ( u _ { i } ) ) \\ , u _ { i } \\sim \\mathrm { U n i f o r m } ( 0 , 1 ) ,", + "type": "interline_equation", + "image_path": "f9b200eff054dd7a1f090b16bac57a8cbc7a58c52f126fa82b11ce37ab6acf42.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 214, + 627, + 396, + 640 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 646, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 123, + 659 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 648, + 130, + 656 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 644, + 505, + 659 + ], + "score": 1.0, + "content": "is the temperature parameter. In practice, we usually start at a high temperature and anneal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 657, + 327, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 327, + 669 + ], + "score": 1.0, + "content": "to a small but non-zero temperature (Jang et al., 2016).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "title", + "bbox": [ + 105, + 684, + 501, + 698 + ], + "lines": [ + { + "bbox": [ + 105, + 684, + 502, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 502, + 698 + ], + "score": 1.0, + "content": "3 AUTO-ADDRESSING AND RECURRENT MEMORY INTEGRATING NETWORK", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "In this section, we describe the structure of the ARMIN network as shown in Figure 1. It consists of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "a recurrent cell and a small external memory that stores historical hidden states. While processing", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 237 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 104, + 82, + 506, + 238 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 241, + 503, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 255 + ], + "score": 1.0, + "content": "Inspired by the TARDIS, we introduce the ARMIN architecture, a more light-weight and memory-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "based architecture with a very simple memory addressing mechanism and in-depth modification of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 404, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 404, + 276 + ], + "score": 1.0, + "content": "the existing LSTM structure. Concretely, our contributions are as follows:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 240, + 505, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 286, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 104, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "• We distill previous memory addressing methods and propose a simple yet effective memory ad-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "dressing mechanism for the external memory, namely Auto-addressing, by encoding the information", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 414, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 290, + 321 + ], + "score": 1.0, + "content": "for memory addressing directly via the inputs", + "type": "text" + }, + { + "bbox": [ + 290, + 309, + 302, + 319 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 308, + 388, + 321 + ], + "score": 1.0, + "content": "and the hidden states", + "type": "text" + }, + { + "bbox": [ + 389, + 308, + 410, + 320 + ], + "score": 0.9, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 308, + 414, + 321 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "• We propose a novel recurrent cell that combines the gating advantages of LSTM and allows", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "direct gradient backpropagation from output to memory. With only a single cell, it learns better", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 331, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 331, + 357 + ], + "score": 1.0, + "content": "representations than many hierarchical RNN structures.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 358, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 104, + 358, + 506, + 373 + ], + "score": 1.0, + "content": "• We show in our ablation study that the ARMIN is robust to small iteration lengths when training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "long sequential data, which enables training with large batch sizes and boost its training speed.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 450, + 395 + ], + "score": 1.0, + "content": "We further demonstrate in char-level language modelling task that the ARMIN obtain", + "type": "text" + }, + { + "bbox": [ + 451, + 381, + 470, + 392 + ], + "score": 0.87, + "content": "4 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "training", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 393, + 451, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 451, + 406 + ], + "score": 1.0, + "content": "speed gain than LSTM while keeping similar performance and memory consumption.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 285, + 506, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 407, + 504, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 421 + ], + "score": 1.0, + "content": "• We demonstrate competitive results on various tasks while keeping efficient time and memory", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 419, + 302, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 302, + 431 + ], + "score": 1.0, + "content": "consumption during training and inference time.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 405, + 505, + 431 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 446, + 200, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 444, + 201, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 201, + 462 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 107, + 471, + 266, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 267, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 267, + 484 + ], + "score": 1.0, + "content": "2.1 GUMBEL-SOFTMAX ESTIMATOR", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "Categorical distribution is a natural choice for representing discrete structure in the world; however,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "it’s rarely used in neural networks due to its inability to backpropagate through samples(Jang et al.,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "2016). To this end, Maddison et al. (2016) and Jang et al. (2016) propose a continuous relaxation of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "categorical distribution and the corresponding gumbel-softmax gradient estimator that replaces the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "non-differentiable sample from a categorical distribution with a differentiable sample. Specifically,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 230, + 560 + ], + "score": 1.0, + "content": "given a probability distribution", + "type": "text" + }, + { + "bbox": [ + 231, + 547, + 311, + 558 + ], + "score": 0.91, + "content": "\\pmb { p } = ( \\pi _ { 1 } , \\pi _ { 2 } , . . . , \\pi _ { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 546, + 331, + 560 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 331, + 547, + 338, + 556 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "categories, the gumbel-softmax estimator", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 558, + 429, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 248, + 570 + ], + "score": 1.0, + "content": "produces an one-hot sample vector", + "type": "text" + }, + { + "bbox": [ + 248, + 560, + 255, + 569 + ], + "score": 0.82, + "content": "\\textbf { { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 558, + 289, + 570 + ], + "score": 1.0, + "content": "with its", + "type": "text" + }, + { + "bbox": [ + 289, + 559, + 293, + 568 + ], + "score": 0.8, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 558, + 429, + 570 + ], + "score": 1.0, + "content": "-th element calculated as follows:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 104, + 491, + 506, + 570 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 574, + 416, + 605 + ], + "lines": [ + { + "bbox": [ + 193, + 574, + 416, + 605 + ], + "spans": [ + { + "bbox": [ + 193, + 574, + 416, + 605 + ], + "score": 0.93, + "content": "y _ { i } = { \\frac { \\exp ( ( \\log ( \\pi _ { i } ) + g _ { i } ) / \\tau ) } { \\sum _ { j = 1 } ^ { k } \\exp ( ( \\log ( \\pi _ { j } ) + g _ { j } ) / \\tau ) } } \\quad { \\mathrm { f o r ~ } } i = 1 , 2 , . . . , k ,", + "type": "interline_equation", + "image_path": "3d7fbb0ffd4812f959da86e645aa96d65bbe99701c988143802f6d35011fd4d2.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 574, + 416, + 589.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 193, + 589.5, + 416, + 605.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 437, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 438, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 133, + 623 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 612, + 170, + 622 + ], + "score": 0.9, + "content": "g _ { 1 } , . . . , g _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 608, + 438, + 623 + ], + "score": 1.0, + "content": "are i.i.d samples drawn from Gumbel distribution(Gumbel, 1954):", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 608, + 438, + 623 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 627, + 396, + 640 + ], + "lines": [ + { + "bbox": [ + 214, + 627, + 396, + 640 + ], + "spans": [ + { + "bbox": [ + 214, + 627, + 396, + 640 + ], + "score": 0.93, + "content": "g _ { i } = - \\log ( - \\log ( u _ { i } ) ) \\ , u _ { i } \\sim \\mathrm { U n i f o r m } ( 0 , 1 ) ,", + "type": "interline_equation", + "image_path": "f9b200eff054dd7a1f090b16bac57a8cbc7a58c52f126fa82b11ce37ab6acf42.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 214, + 627, + 396, + 640 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 646, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 123, + 659 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 648, + 130, + 656 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 644, + 505, + 659 + ], + "score": 1.0, + "content": "is the temperature parameter. In practice, we usually start at a high temperature and anneal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 657, + 327, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 327, + 669 + ], + "score": 1.0, + "content": "to a small but non-zero temperature (Jang et al., 2016).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 644, + 505, + 669 + ] + }, + { + "type": "title", + "bbox": [ + 105, + 684, + 501, + 698 + ], + "lines": [ + { + "bbox": [ + 105, + 684, + 502, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 502, + 698 + ], + "score": 1.0, + "content": "3 AUTO-ADDRESSING AND RECURRENT MEMORY INTEGRATING NETWORK", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "In this section, we describe the structure of the ARMIN network as shown in Figure 1. It consists of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "a recurrent cell and a small external memory that stores historical hidden states. While processing", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 104, + 710, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 91, + 459, + 277 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 140, + 91, + 459, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 459, + 277 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 459, + 277 + ], + "score": 0.968, + "type": "image", + "image_path": "eab2a73d816507916b458d579eb538fc7d49d69aba42b4a31414c981c7c3217e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 140, + 91, + 459, + 153.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 140, + 153.0, + 459, + 215.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 140, + 215.0, + 459, + 277.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 281, + 505, + 349 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Figure 1: The ARMIN structure. At each time-step, the ARMIN performs read operation, cell", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 293, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 493, + 307 + ], + "score": 1.0, + "content": "processing and write operation in chronological order: (a) It reads out a historical hidden state", + "type": "text" + }, + { + "bbox": [ + 494, + 295, + 504, + 304 + ], + "score": 0.82, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 367, + 317 + ], + "score": 1.0, + "content": "from memory with an one-hot read vector produced via passing", + "type": "text" + }, + { + "bbox": [ + 368, + 306, + 379, + 315 + ], + "score": 0.86, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 304, + 398, + 317 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 304, + 420, + 316 + ], + "score": 0.91, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "to a fully connected", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 424, + 328 + ], + "score": 1.0, + "content": "layer followed by a gumbel-softmax function. (b) The ARMIN cell receives", + "type": "text" + }, + { + "bbox": [ + 424, + 317, + 435, + 327 + ], + "score": 0.66, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 315, + 440, + 328 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 440, + 316, + 462, + 327 + ], + "score": 0.8, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 315, + 482, + 328 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 482, + 317, + 492, + 327 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 183, + 339 + ], + "score": 1.0, + "content": "inputs and outputs", + "type": "text" + }, + { + "bbox": [ + 183, + 328, + 194, + 337 + ], + "score": 0.85, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 326, + 212, + 339 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 212, + 327, + 223, + 337 + ], + "score": 0.85, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 326, + 228, + 339 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 229, + 327, + 240, + 337 + ], + "score": 0.82, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 326, + 365, + 339 + ], + "score": 1.0, + "content": "is passed to output layers, and", + "type": "text" + }, + { + "bbox": [ + 366, + 327, + 377, + 337 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "is passed to next time-step. (c)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 337, + 280, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 118, + 348 + ], + "score": 0.86, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 337, + 265, + 351 + ], + "score": 1.0, + "content": "is written to the previous location of", + "type": "text" + }, + { + "bbox": [ + 266, + 339, + 276, + 348 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 337, + 280, + 351 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "sequential data, the ARMIN performs reading from memory, cell processing and writing to mem-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "ory operations in chronological order during each time-step. In the following subsections, we first", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 392, + 462, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 462, + 406 + ], + "score": 1.0, + "content": "explain the structure of the recurrent cell, and then discuss the read and write operations.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 108, + 417, + 281, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 281, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 281, + 430 + ], + "score": 1.0, + "content": "3.1 THE RECURRENT CELL OF ARMIN", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 437, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "Inspired by classical LSTM structure (please refer to APPENDIX A for details in RNN and LSTM", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "structures), we propose a novel memory-augmented recurrent cell structure, namely the ARMIN", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 281, + 473 + ], + "score": 1.0, + "content": "cell. At every time-step, it takes in an input", + "type": "text" + }, + { + "bbox": [ + 281, + 462, + 292, + 471 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 461, + 377, + 473 + ], + "score": 1.0, + "content": ", the last hidden state", + "type": "text" + }, + { + "bbox": [ + 378, + 461, + 399, + 472 + ], + "score": 0.92, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "and a recovered historical", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 158, + 484 + ], + "score": 1.0, + "content": "hidden state", + "type": "text" + }, + { + "bbox": [ + 158, + 472, + 168, + 482 + ], + "score": 0.84, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 471, + 411, + 484 + ], + "score": 1.0, + "content": "chosen by a read operation, and produces an output vector", + "type": "text" + }, + { + "bbox": [ + 411, + 473, + 421, + 482 + ], + "score": 0.86, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "and the new hidden", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 482, + 300, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 127, + 495 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 127, + 482, + 138, + 493 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 483, + 300, + 495 + ], + "score": 1.0, + "content": ". The computation process is as follows:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 506, + 392, + 536 + ], + "lines": [ + { + "bbox": [ + 220, + 506, + 392, + 536 + ], + "spans": [ + { + "bbox": [ + 220, + 506, + 392, + 536 + ], + "score": 0.89, + "content": "\\left\\{ { g } _ { t } ^ { h } \\atop { g } _ { t } ^ { r } \\right\\} = \\left\\{ { \\sigma } \\atop { \\sigma } \\right\\} W _ { i g } [ x _ { t } , h _ { t - 1 } , r _ { t } ] + b _ { i g } ,", + "type": "interline_equation", + "image_path": "1a2381c9e716db06877af5cbb16644d9f37ab3bbcdff250345c9426035fa0588.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 506, + 392, + 521.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 220, + 521.0, + 392, + 536.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 539, + 345, + 554 + ], + "lines": [ + { + "bbox": [ + 264, + 539, + 345, + 554 + ], + "spans": [ + { + "bbox": [ + 264, + 539, + 345, + 554 + ], + "score": 0.71, + "content": "h _ { t - 1 } ^ { g } = { \\pmb g } _ { t } ^ { h } \\circ { \\pmb h } _ { t - 1 } ,", + "type": "interline_equation", + "image_path": "217d67193e54a23659454a2544a34c38d6fc260d6bad8abdfb195eba32d5d806.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 264, + 539, + 345, + 554 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 276, + 556, + 335, + 570 + ], + "lines": [ + { + "bbox": [ + 276, + 556, + 335, + 570 + ], + "spans": [ + { + "bbox": [ + 276, + 556, + 335, + 570 + ], + "score": 0.66, + "content": "\\begin{array} { r } { \\pmb { r } _ { t } ^ { g } = \\pmb { g } _ { t } ^ { r } \\circ \\pmb { r } _ { t } , } \\end{array}", + "type": "interline_equation", + "image_path": "2ff3978396c1892799181c33d24e40b07974e748cecce11677c3c5fc54205888.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 276, + 556, + 335, + 570 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 573, + 403, + 644 + ], + "lines": [ + { + "bbox": [ + 208, + 573, + 403, + 644 + ], + "spans": [ + { + "bbox": [ + 208, + 573, + 403, + 644 + ], + "score": 0.89, + "content": "\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } ^ { h } } \\\\ { o _ { t } ^ { r } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W _ { g o } [ \\boldsymbol { x } _ { t } , h _ { t - 1 } ^ { g } , \\boldsymbol { r } _ { t } ^ { g } ] + b _ { g o } ,", + "type": "interline_equation", + "image_path": "159cf38197a23b05f8f396953f74402cbfc9572e8bbc0deba36b532733921968.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 208, + 573, + 403, + 587.2 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 208, + 587.2, + 403, + 601.4000000000001 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 208, + 601.4000000000001, + 403, + 615.6000000000001 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 208, + 615.6000000000001, + 403, + 629.8000000000002 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 208, + 629.8000000000002, + 403, + 644.0000000000002 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 646, + 360, + 659 + ], + "lines": [ + { + "bbox": [ + 248, + 646, + 360, + 659 + ], + "spans": [ + { + "bbox": [ + 248, + 646, + 360, + 659 + ], + "score": 0.65, + "content": "\\pmb { h } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { h } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ,", + "type": "interline_equation", + "image_path": "34a091eaff95e86551a79e6eb6b3b2b7571b87ba5d72f231a10afbfeff0581e3.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 248, + 646, + 360, + 659 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 662, + 387, + 677 + ], + "lines": [ + { + "bbox": [ + 224, + 662, + 387, + 677 + ], + "spans": [ + { + "bbox": [ + 224, + 662, + 387, + 677 + ], + "score": 0.49, + "content": "\\begin{array} { r } { \\pmb { o } _ { t } = \\left[ \\pmb { o } _ { t } ^ { h } \\circ \\operatorname { t a n h } ( \\pmb { h } _ { t } ) , \\pmb { o } _ { t } ^ { r } \\circ \\operatorname { t a n h } ( \\pmb { r } _ { t } ) \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "58570aefc636c1ba6609da63ca2076995d0b6a9d09702509a5cdbb78c6ea9728.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 224, + 662, + 387, + 677 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 684, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 682, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 136, + 700 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 685, + 322, + 697 + ], + "score": 0.89, + "content": "h _ { t - 1 } , h _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { \\boldsymbol { r } } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\boldsymbol { x } } _ { t } \\in \\mathbb { R } ^ { d _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 682, + 350, + 700 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 351, + 684, + 505, + 698 + ], + "score": 0.89, + "content": "\\begin{array} { r c l } { W _ { i g } } & { \\in } & { \\mathbb { R } ^ { 2 d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , W _ { g o } } & { \\in } \\end{array}", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 694, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 107, + 698, + 399, + 711 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\mathbb { R } ^ { ( 4 d _ { h } + d _ { r } ) \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } _ { i g } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } , \\pmb { b } _ { g o } \\ \\in \\ \\mathbb { R } ^ { 4 d _ { h } + d _ { r } } , \\pmb { o } _ { t } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 694, + 462, + 713 + ], + "score": 1.0, + "content": ". We refer to", + "type": "text" + }, + { + "bbox": [ + 463, + 699, + 474, + 710 + ], + "score": 0.87, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 694, + 507, + 713 + ], + "score": 1.0, + "content": "as the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 360, + 723 + ], + "score": 1.0, + "content": "hidden size of the recurrent cell of ARMIN. Usually we have", + "type": "text" + }, + { + "bbox": [ + 360, + 710, + 395, + 721 + ], + "score": 0.9, + "content": "d _ { h } = d _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "and allocate equal number", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 719, + 250, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 207, + 733 + ], + "score": 1.0, + "content": "of weight parameters for", + "type": "text" + }, + { + "bbox": [ + 207, + 721, + 218, + 732 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 719, + 236, + 733 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 722, + 246, + 732 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { r } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 719, + 250, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 91, + 459, + 277 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 140, + 91, + 459, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 459, + 277 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 459, + 277 + ], + "score": 0.968, + "type": "image", + "image_path": "eab2a73d816507916b458d579eb538fc7d49d69aba42b4a31414c981c7c3217e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 140, + 91, + 459, + 153.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 140, + 153.0, + 459, + 215.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 140, + 215.0, + 459, + 277.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 281, + 505, + 349 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Figure 1: The ARMIN structure. At each time-step, the ARMIN performs read operation, cell", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 293, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 493, + 307 + ], + "score": 1.0, + "content": "processing and write operation in chronological order: (a) It reads out a historical hidden state", + "type": "text" + }, + { + "bbox": [ + 494, + 295, + 504, + 304 + ], + "score": 0.82, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 367, + 317 + ], + "score": 1.0, + "content": "from memory with an one-hot read vector produced via passing", + "type": "text" + }, + { + "bbox": [ + 368, + 306, + 379, + 315 + ], + "score": 0.86, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 304, + 398, + 317 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 304, + 420, + 316 + ], + "score": 0.91, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "to a fully connected", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 424, + 328 + ], + "score": 1.0, + "content": "layer followed by a gumbel-softmax function. (b) The ARMIN cell receives", + "type": "text" + }, + { + "bbox": [ + 424, + 317, + 435, + 327 + ], + "score": 0.66, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 315, + 440, + 328 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 440, + 316, + 462, + 327 + ], + "score": 0.8, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 315, + 482, + 328 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 482, + 317, + 492, + 327 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 183, + 339 + ], + "score": 1.0, + "content": "inputs and outputs", + "type": "text" + }, + { + "bbox": [ + 183, + 328, + 194, + 337 + ], + "score": 0.85, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 326, + 212, + 339 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 212, + 327, + 223, + 337 + ], + "score": 0.85, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 326, + 228, + 339 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 229, + 327, + 240, + 337 + ], + "score": 0.82, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 326, + 365, + 339 + ], + "score": 1.0, + "content": "is passed to output layers, and", + "type": "text" + }, + { + "bbox": [ + 366, + 327, + 377, + 337 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "is passed to next time-step. (c)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 337, + 280, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 118, + 348 + ], + "score": 0.86, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 337, + 265, + 351 + ], + "score": 1.0, + "content": "is written to the previous location of", + "type": "text" + }, + { + "bbox": [ + 266, + 339, + 276, + 348 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 337, + 280, + 351 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "sequential data, the ARMIN performs reading from memory, cell processing and writing to mem-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "ory operations in chronological order during each time-step. In the following subsections, we first", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 392, + 462, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 462, + 406 + ], + "score": 1.0, + "content": "explain the structure of the recurrent cell, and then discuss the read and write operations.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 370, + 505, + 406 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 417, + 281, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 281, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 281, + 430 + ], + "score": 1.0, + "content": "3.1 THE RECURRENT CELL OF ARMIN", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 437, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "Inspired by classical LSTM structure (please refer to APPENDIX A for details in RNN and LSTM", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "structures), we propose a novel memory-augmented recurrent cell structure, namely the ARMIN", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 281, + 473 + ], + "score": 1.0, + "content": "cell. At every time-step, it takes in an input", + "type": "text" + }, + { + "bbox": [ + 281, + 462, + 292, + 471 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 461, + 377, + 473 + ], + "score": 1.0, + "content": ", the last hidden state", + "type": "text" + }, + { + "bbox": [ + 378, + 461, + 399, + 472 + ], + "score": 0.92, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "and a recovered historical", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 158, + 484 + ], + "score": 1.0, + "content": "hidden state", + "type": "text" + }, + { + "bbox": [ + 158, + 472, + 168, + 482 + ], + "score": 0.84, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 471, + 411, + 484 + ], + "score": 1.0, + "content": "chosen by a read operation, and produces an output vector", + "type": "text" + }, + { + "bbox": [ + 411, + 473, + 421, + 482 + ], + "score": 0.86, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "and the new hidden", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 482, + 300, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 127, + 495 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 127, + 482, + 138, + 493 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 483, + 300, + 495 + ], + "score": 1.0, + "content": ". The computation process is as follows:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 438, + 505, + 495 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 506, + 392, + 536 + ], + "lines": [ + { + "bbox": [ + 220, + 506, + 392, + 536 + ], + "spans": [ + { + "bbox": [ + 220, + 506, + 392, + 536 + ], + "score": 0.89, + "content": "\\left\\{ { g } _ { t } ^ { h } \\atop { g } _ { t } ^ { r } \\right\\} = \\left\\{ { \\sigma } \\atop { \\sigma } \\right\\} W _ { i g } [ x _ { t } , h _ { t - 1 } , r _ { t } ] + b _ { i g } ,", + "type": "interline_equation", + "image_path": "1a2381c9e716db06877af5cbb16644d9f37ab3bbcdff250345c9426035fa0588.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 506, + 392, + 521.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 220, + 521.0, + 392, + 536.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 264, + 539, + 345, + 554 + ], + "lines": [ + { + "bbox": [ + 264, + 539, + 345, + 554 + ], + "spans": [ + { + "bbox": [ + 264, + 539, + 345, + 554 + ], + "score": 0.71, + "content": "h _ { t - 1 } ^ { g } = { \\pmb g } _ { t } ^ { h } \\circ { \\pmb h } _ { t - 1 } ,", + "type": "interline_equation", + "image_path": "217d67193e54a23659454a2544a34c38d6fc260d6bad8abdfb195eba32d5d806.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 264, + 539, + 345, + 554 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 276, + 556, + 335, + 570 + ], + "lines": [ + { + "bbox": [ + 276, + 556, + 335, + 570 + ], + "spans": [ + { + "bbox": [ + 276, + 556, + 335, + 570 + ], + "score": 0.66, + "content": "\\begin{array} { r } { \\pmb { r } _ { t } ^ { g } = \\pmb { g } _ { t } ^ { r } \\circ \\pmb { r } _ { t } , } \\end{array}", + "type": "interline_equation", + "image_path": "2ff3978396c1892799181c33d24e40b07974e748cecce11677c3c5fc54205888.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 276, + 556, + 335, + 570 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 573, + 403, + 644 + ], + "lines": [ + { + "bbox": [ + 208, + 573, + 403, + 644 + ], + "spans": [ + { + "bbox": [ + 208, + 573, + 403, + 644 + ], + "score": 0.89, + "content": "\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } ^ { h } } \\\\ { o _ { t } ^ { r } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W _ { g o } [ \\boldsymbol { x } _ { t } , h _ { t - 1 } ^ { g } , \\boldsymbol { r } _ { t } ^ { g } ] + b _ { g o } ,", + "type": "interline_equation", + "image_path": "159cf38197a23b05f8f396953f74402cbfc9572e8bbc0deba36b532733921968.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 208, + 573, + 403, + 587.2 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 208, + 587.2, + 403, + 601.4000000000001 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 208, + 601.4000000000001, + 403, + 615.6000000000001 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 208, + 615.6000000000001, + 403, + 629.8000000000002 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 208, + 629.8000000000002, + 403, + 644.0000000000002 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 646, + 360, + 659 + ], + "lines": [ + { + "bbox": [ + 248, + 646, + 360, + 659 + ], + "spans": [ + { + "bbox": [ + 248, + 646, + 360, + 659 + ], + "score": 0.65, + "content": "\\pmb { h } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { h } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ,", + "type": "interline_equation", + "image_path": "34a091eaff95e86551a79e6eb6b3b2b7571b87ba5d72f231a10afbfeff0581e3.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 248, + 646, + 360, + 659 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 662, + 387, + 677 + ], + "lines": [ + { + "bbox": [ + 224, + 662, + 387, + 677 + ], + "spans": [ + { + "bbox": [ + 224, + 662, + 387, + 677 + ], + "score": 0.49, + "content": "\\begin{array} { r } { \\pmb { o } _ { t } = \\left[ \\pmb { o } _ { t } ^ { h } \\circ \\operatorname { t a n h } ( \\pmb { h } _ { t } ) , \\pmb { o } _ { t } ^ { r } \\circ \\operatorname { t a n h } ( \\pmb { r } _ { t } ) \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "58570aefc636c1ba6609da63ca2076995d0b6a9d09702509a5cdbb78c6ea9728.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 224, + 662, + 387, + 677 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 684, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 682, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 136, + 700 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 685, + 322, + 697 + ], + "score": 0.89, + "content": "h _ { t - 1 } , h _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { \\boldsymbol { r } } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\boldsymbol { x } } _ { t } \\in \\mathbb { R } ^ { d _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 682, + 350, + 700 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 351, + 684, + 505, + 698 + ], + "score": 0.89, + "content": "\\begin{array} { r c l } { W _ { i g } } & { \\in } & { \\mathbb { R } ^ { 2 d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , W _ { g o } } & { \\in } \\end{array}", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 694, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 107, + 698, + 399, + 711 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\mathbb { R } ^ { ( 4 d _ { h } + d _ { r } ) \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } _ { i g } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } , \\pmb { b } _ { g o } \\ \\in \\ \\mathbb { R } ^ { 4 d _ { h } + d _ { r } } , \\pmb { o } _ { t } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 694, + 462, + 713 + ], + "score": 1.0, + "content": ". We refer to", + "type": "text" + }, + { + "bbox": [ + 463, + 699, + 474, + 710 + ], + "score": 0.87, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 694, + 507, + 713 + ], + "score": 1.0, + "content": "as the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 360, + 723 + ], + "score": 1.0, + "content": "hidden size of the recurrent cell of ARMIN. Usually we have", + "type": "text" + }, + { + "bbox": [ + 360, + 710, + 395, + 721 + ], + "score": 0.9, + "content": "d _ { h } = d _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "and allocate equal number", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 719, + 250, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 207, + 733 + ], + "score": 1.0, + "content": "of weight parameters for", + "type": "text" + }, + { + "bbox": [ + 207, + 721, + 218, + 732 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 719, + 236, + 733 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 722, + 246, + 732 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { r } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 719, + 250, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 682, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 154, + 95 + ], + "score": 1.0, + "content": "In equation", + "type": "text" + }, + { + "bbox": [ + 155, + 83, + 179, + 93 + ], + "score": 0.87, + "content": "3 \\sim 5", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 82, + 423, + 95 + ], + "score": 1.0, + "content": ", two gates are calculated to control the information flow for", + "type": "text" + }, + { + "bbox": [ + 424, + 83, + 444, + 94 + ], + "score": 0.92, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 82, + 462, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 463, + 84, + 472, + 93 + ], + "score": 0.83, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "respec-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 91, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 255, + 108 + ], + "score": 1.0, + "content": "tively, generating gated hidden state", + "type": "text" + }, + { + "bbox": [ + 255, + 93, + 277, + 106 + ], + "score": 0.92, + "content": "h _ { t - 1 } ^ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 91, + 357, + 108 + ], + "score": 1.0, + "content": "and historical state", + "type": "text" + }, + { + "bbox": [ + 358, + 93, + 369, + 105 + ], + "score": 0.89, + "content": "\\boldsymbol { r } _ { t } ^ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 91, + 506, + 108 + ], + "score": 1.0, + "content": "; using this method, we can filter", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "out the irrelevant information for the current time-step. Then as shown in equation 6, we compute", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 163, + 128 + ], + "score": 1.0, + "content": "the input gate", + "type": "text" + }, + { + "bbox": [ + 163, + 116, + 172, + 127 + ], + "score": 0.85, + "content": "\\mathbf { \\delta } _ { i _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 115, + 221, + 128 + ], + "score": 1.0, + "content": ", forget gate", + "type": "text" + }, + { + "bbox": [ + 221, + 116, + 231, + 127 + ], + "score": 0.87, + "content": "\\pmb { f } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 115, + 272, + 128 + ], + "score": 1.0, + "content": ", cell state", + "type": "text" + }, + { + "bbox": [ + 273, + 117, + 283, + 127 + ], + "score": 0.85, + "content": "\\mathbf { \\pmb { g } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 115, + 348, + 128 + ], + "score": 1.0, + "content": "and output gate", + "type": "text" + }, + { + "bbox": [ + 349, + 115, + 361, + 127 + ], + "score": 0.9, + "content": "\\mathbf { \\sigma } _ { o _ { t } ^ { h } } ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "for the new hidden state just like in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 328, + 138 + ], + "score": 1.0, + "content": "classical LSTM structure. Additionally, an output gate", + "type": "text" + }, + { + "bbox": [ + 329, + 127, + 340, + 138 + ], + "score": 0.89, + "content": "{ \\mathbf { } } _ { o _ { t } ^ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 126, + 416, + 138 + ], + "score": 1.0, + "content": "for historical state", + "type": "text" + }, + { + "bbox": [ + 416, + 128, + 427, + 137 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "is computed. Next", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 285, + 150 + ], + "score": 1.0, + "content": "in equation 7, we compute new hidden state", + "type": "text" + }, + { + "bbox": [ + 286, + 138, + 297, + 148 + ], + "score": 0.89, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 137, + 371, + 150 + ], + "score": 1.0, + "content": "that is the sum of", + "type": "text" + }, + { + "bbox": [ + 371, + 137, + 393, + 149 + ], + "score": 0.91, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 137, + 449, + 150 + ], + "score": 1.0, + "content": "and cell state", + "type": "text" + }, + { + "bbox": [ + 450, + 139, + 460, + 149 + ], + "score": 0.84, + "content": "\\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } } \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\lambda } }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 137, + 505, + 150 + ], + "score": 1.0, + "content": ", leveraged", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 167, + 161 + ], + "score": 1.0, + "content": "by forget gate", + "type": "text" + }, + { + "bbox": [ + 168, + 149, + 178, + 160 + ], + "score": 0.86, + "content": "\\pmb { f } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 148, + 242, + 161 + ], + "score": 1.0, + "content": "and input gate", + "type": "text" + }, + { + "bbox": [ + 243, + 149, + 251, + 159 + ], + "score": 0.86, + "content": "\\dot { \\mathbf { \\delta } } _ { \\mathbf { \\delta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 148, + 505, + 161 + ], + "score": 1.0, + "content": ". Finally in equation 8, we calculate the output of this time-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 348, + 172 + ], + "score": 1.0, + "content": "step, which is the concatenation of the gated contents from", + "type": "text" + }, + { + "bbox": [ + 348, + 160, + 360, + 170 + ], + "score": 0.87, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 159, + 378, + 172 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 160, + 388, + 170 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 159, + 505, + 172 + ], + "score": 1.0, + "content": ". Using this method, we can", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 417, + 183 + ], + "score": 1.0, + "content": "selectively backpropagate gradients from the output to the historical state", + "type": "text" + }, + { + "bbox": [ + 417, + 172, + 427, + 181 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 169, + 479, + 183 + ], + "score": 1.0, + "content": ". We involve", + "type": "text" + }, + { + "bbox": [ + 479, + 172, + 489, + 181 + ], + "score": 0.83, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 169, + 506, + 183 + ], + "score": 1.0, + "content": "(or", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 230, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 118, + 194 + ], + "score": 0.82, + "content": "\\boldsymbol { r } _ { t } ^ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 181, + 230, + 194 + ], + "score": 1.0, + "content": ") in all gates’ computations.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 168, + 211 + ], + "score": 1.0, + "content": "Intuitively, the", + "type": "text" + }, + { + "bbox": [ + 168, + 198, + 190, + 210 + ], + "score": 0.91, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 197, + 345, + 211 + ], + "score": 1.0, + "content": "acts as the old working memory, and", + "type": "text" + }, + { + "bbox": [ + 346, + 200, + 356, + 209 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 197, + 505, + 211 + ], + "score": 1.0, + "content": "is treated as the long-term memory.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 265, + 222 + ], + "score": 1.0, + "content": "The cell processes them with the input", + "type": "text" + }, + { + "bbox": [ + 265, + 211, + 276, + 220 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 208, + 431, + 222 + ], + "score": 1.0, + "content": "to generate the new working memory", + "type": "text" + }, + { + "bbox": [ + 432, + 210, + 443, + 220 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "and the output", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 117, + 231 + ], + "score": 0.82, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 219, + 217, + 234 + ], + "score": 1.0, + "content": ". More specifically, each", + "type": "text" + }, + { + "bbox": [ + 218, + 221, + 228, + 231 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 219, + 506, + 234 + ], + "score": 1.0, + "content": "is a summary of historical hidden states selected by auto-addressing", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "mechanism. The ARMIN cell learns to recurrently integrate the summary of long-term information", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 129, + 255 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 244, + 140, + 253 + ], + "score": 0.83, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 241, + 250, + 255 + ], + "score": 1.0, + "content": "into the working memory", + "type": "text" + }, + { + "bbox": [ + 251, + 242, + 262, + 253 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 241, + 505, + 255 + ], + "score": 1.0, + "content": ". We will demonstrate the efficiency of recurrent memory", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 268, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 268, + 266 + ], + "score": 1.0, + "content": "integration in language modelling tasks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 269, + 282 + ], + "score": 1.0, + "content": "Unlike the LSTM that passes a tuple of", + "type": "text" + }, + { + "bbox": [ + 270, + 270, + 301, + 282 + ], + "score": 0.93, + "content": "\\left( \\boldsymbol { h } _ { t } , \\boldsymbol { c } _ { t } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 269, + 443, + 282 + ], + "score": 1.0, + "content": "to the next recurrence and outputs", + "type": "text" + }, + { + "bbox": [ + 443, + 270, + 455, + 281 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "at the same", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 302, + 293 + ], + "score": 1.0, + "content": "time, the ARMIN only passes one hidden state", + "type": "text" + }, + { + "bbox": [ + 302, + 281, + 313, + 292 + ], + "score": 0.87, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 281, + 397, + 293 + ], + "score": 1.0, + "content": "and separately uses", + "type": "text" + }, + { + "bbox": [ + 397, + 283, + 408, + 292 + ], + "score": 0.85, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "as output, which has 2", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "benefits: one is that it takes lower memory consumption to store one hidden state rather than two,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 402, + 315 + ], + "score": 1.0, + "content": "and the other is we can decouple the information needed for the output", + "type": "text" + }, + { + "bbox": [ + 402, + 304, + 412, + 314 + ], + "score": 0.85, + "content": "\\mathbf { } _ \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 303, + 505, + 315 + ], + "score": 1.0, + "content": ", from the information", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 126, + 326 + ], + "score": 1.0, + "content": "(i.e.", + "type": "text" + }, + { + "bbox": [ + 126, + 314, + 138, + 325 + ], + "score": 0.7, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 313, + 505, + 326 + ], + "score": 1.0, + "content": ") that is needed for memorizing and being used at later time-steps. Furthermore, we list 3", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 502, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 502, + 336 + ], + "score": 1.0, + "content": "important differences between the ARMIN cell and TARDIS cell, please refer to APPENDIX B.1.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 109, + 349, + 320, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 321, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 321, + 362 + ], + "score": 1.0, + "content": "3.2 READ OPERATION WITH AUTO-ADDRESSING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 275, + 383 + ], + "score": 1.0, + "content": "The ARMIN maintains a memory matrix", + "type": "text" + }, + { + "bbox": [ + 276, + 369, + 346, + 380 + ], + "score": 0.91, + "content": "M \\in \\mathbb { R } ^ { n _ { m e m } \\times d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 367, + 429, + 383 + ], + "score": 1.0, + "content": ", where the constant", + "type": "text" + }, + { + "bbox": [ + 430, + 372, + 455, + 381 + ], + "score": 0.89, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 367, + 506, + 383 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 437, + 395 + ], + "score": 1.0, + "content": "number of memory slots. At each recurrence, the ARMIN chooses a historical state", + "type": "text" + }, + { + "bbox": [ + 438, + 383, + 447, + 392 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 379, + 506, + 395 + ], + "score": 1.0, + "content": "from memory", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 392, + 417, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 232, + 404 + ], + "score": 1.0, + "content": "according to the information in", + "type": "text" + }, + { + "bbox": [ + 233, + 393, + 244, + 403 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 392, + 262, + 404 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 262, + 392, + 284, + 404 + ], + "score": 0.9, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 392, + 417, + 404 + ], + "score": 1.0, + "content": ",which is formulated as follows:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 417, + 401, + 468 + ], + "lines": [ + { + "bbox": [ + 209, + 417, + 401, + 468 + ], + "spans": [ + { + "bbox": [ + 209, + 417, + 401, + 468 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\displaystyle s _ { t } = \\mathrm { g u m b e l - s o f t m a x } \\big ( W _ { s } [ \\ : x _ { t } \\ : , h _ { t - 1 } \\ : ] + b _ { s } \\big ) \\ : , } } \\\\ { { { } } } \\\\ { { \\displaystyle r _ { t } = \\sum _ { i = 0 } ^ { n - 1 } s _ { t } ( i ) M ( i , : ) \\ : . } } \\end{array}", + "type": "interline_equation", + "image_path": "d20fa90dbab7b45d436ec9d1fe2698fcfc15d2cb48c0e332be86d0aefcd202b3.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 209, + 417, + 401, + 434.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 209, + 434.0, + 401, + 451.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 209, + 451.0, + 401, + 468.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 471, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 134, + 486 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 470, + 229, + 483 + ], + "score": 0.87, + "content": "W _ { s } \\ \\in \\ \\mathbb { R } ^ { n _ { m e m } \\times ( d _ { i } + d _ { h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 468, + 234, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 235, + 471, + 289, + 483 + ], + "score": 0.84, + "content": "b _ { s } \\in \\mathbb { R } ^ { n _ { m e m } }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 468, + 294, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 295, + 473, + 305, + 483 + ], + "score": 0.73, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { S } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 468, + 505, + 486 + ], + "score": 1.0, + "content": "is a one-hot vector sampled by gumbel-softmax", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 482, + 435, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 145, + 495 + ], + "score": 1.0, + "content": "function,", + "type": "text" + }, + { + "bbox": [ + 145, + 483, + 165, + 495 + ], + "score": 0.91, + "content": "s _ { t } ( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 482, + 214, + 495 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 214, + 484, + 219, + 493 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 482, + 276, + 495 + ], + "score": 1.0, + "content": "-th element of", + "type": "text" + }, + { + "bbox": [ + 276, + 484, + 286, + 494 + ], + "score": 0.75, + "content": "\\mathbf { \\boldsymbol { s } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 482, + 289, + 495 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 290, + 483, + 322, + 495 + ], + "score": 0.91, + "content": "M ( i , : )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 482, + 370, + 495 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 371, + 484, + 375, + 493 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 482, + 416, + 495 + ], + "score": 1.0, + "content": "-th row of", + "type": "text" + }, + { + "bbox": [ + 416, + 483, + 430, + 493 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 482, + 435, + 495 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "As opposed to previous memory networks such as Santoro et al. (2016) and Gulcehre et al. (2017),", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "we don’t use any extra usage vectors and learnable key parameters that manually encode memory", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "addressing information to assist memory addressing. As our ablation study (Appendix E) shows,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 172, + 545 + ], + "score": 1.0, + "content": "the hidden state", + "type": "text" + }, + { + "bbox": [ + 172, + 533, + 194, + 544 + ], + "score": 0.92, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "is sufficient to encode the historical memory accessing information. Further-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 371, + 557 + ], + "score": 1.0, + "content": "more, the TARDIS addressing mechanism requires concatenating", + "type": "text" + }, + { + "bbox": [ + 372, + 545, + 383, + 555 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 543, + 401, + 557 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 544, + 413, + 555 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "to every memory cell,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "which usually causes large extra memory and time consumption in practical scenes. Please refer to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "APPENDIX B.2 for more precise comparison regarding addressing mechanism of the ARMIN and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 576, + 185, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 185, + 588 + ], + "score": 1.0, + "content": "TARDIS networks.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 601, + 212, + 612 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 214, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 214, + 613 + ], + "score": 1.0, + "content": "3.3 WRITE OPERATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 464, + 634 + ], + "score": 1.0, + "content": "After the reading and cell processing stages, the ARMIN writes the new hidden state", + "type": "text" + }, + { + "bbox": [ + 465, + 622, + 476, + 633 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "to the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 143, + 646 + ], + "score": 1.0, + "content": "memory", + "type": "text" + }, + { + "bbox": [ + 143, + 633, + 156, + 643 + ], + "score": 0.59, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 633, + 481, + 646 + ], + "score": 1.0, + "content": ". Following the TARDIS network (Gulcehre et al., 2017), we simply overwrite", + "type": "text" + }, + { + "bbox": [ + 481, + 633, + 493, + 644 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 283, + 657 + ], + "score": 1.0, + "content": "the memory slot where we just read out the", + "type": "text" + }, + { + "bbox": [ + 284, + 645, + 294, + 655 + ], + "score": 0.84, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 643, + 384, + 657 + ], + "score": 1.0, + "content": "(for conditions where", + "type": "text" + }, + { + "bbox": [ + 385, + 644, + 396, + 655 + ], + "score": 0.88, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 643, + 457, + 657 + ], + "score": 1.0, + "content": "is not equal to", + "type": "text" + }, + { + "bbox": [ + 458, + 644, + 468, + 655 + ], + "score": 0.88, + "content": "d _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 643, + 506, + 657 + ], + "score": 1.0, + "content": ", we first", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 228, + 669 + ], + "score": 1.0, + "content": "use a linear layer to transform", + "type": "text" + }, + { + "bbox": [ + 228, + 655, + 239, + 666 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 654, + 262, + 669 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 263, + 655, + 274, + 666 + ], + "score": 0.88, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 654, + 329, + 669 + ], + "score": 1.0, + "content": "dimension to", + "type": "text" + }, + { + "bbox": [ + 330, + 655, + 340, + 666 + ], + "score": 0.88, + "content": "d _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "dimension). But at the initial time-steps,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "we write the hidden states to the empty memory slots, until all empty slots are filled with historical", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "states. In this way, we can maximally preserve useful historical information, because the ARMIN", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 304, + 700 + ], + "score": 1.0, + "content": "cell can learn to copy the useful information from", + "type": "text" + }, + { + "bbox": [ + 304, + 689, + 313, + 699 + ], + "score": 0.83, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 687, + 324, + 700 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 325, + 688, + 336, + 699 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 687, + 397, + 700 + ], + "score": 1.0, + "content": ", and then write", + "type": "text" + }, + { + "bbox": [ + 398, + 688, + 409, + 699 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "to the previous location", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 117, + 712 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 700, + 127, + 710 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". The read/write mechanism can be viewed as a form of a skip-connection of hidden states. By", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "training the whole network, the ARMIN can easily learn long-term dependencies via direct access", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "to historical hidden states. From this point of view, our network is similar to the Skip RNN proposed", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 44.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 154, + 95 + ], + "score": 1.0, + "content": "In equation", + "type": "text" + }, + { + "bbox": [ + 155, + 83, + 179, + 93 + ], + "score": 0.87, + "content": "3 \\sim 5", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 82, + 423, + 95 + ], + "score": 1.0, + "content": ", two gates are calculated to control the information flow for", + "type": "text" + }, + { + "bbox": [ + 424, + 83, + 444, + 94 + ], + "score": 0.92, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 82, + 462, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 463, + 84, + 472, + 93 + ], + "score": 0.83, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "respec-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 91, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 255, + 108 + ], + "score": 1.0, + "content": "tively, generating gated hidden state", + "type": "text" + }, + { + "bbox": [ + 255, + 93, + 277, + 106 + ], + "score": 0.92, + "content": "h _ { t - 1 } ^ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 91, + 357, + 108 + ], + "score": 1.0, + "content": "and historical state", + "type": "text" + }, + { + "bbox": [ + 358, + 93, + 369, + 105 + ], + "score": 0.89, + "content": "\\boldsymbol { r } _ { t } ^ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 91, + 506, + 108 + ], + "score": 1.0, + "content": "; using this method, we can filter", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "out the irrelevant information for the current time-step. Then as shown in equation 6, we compute", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 163, + 128 + ], + "score": 1.0, + "content": "the input gate", + "type": "text" + }, + { + "bbox": [ + 163, + 116, + 172, + 127 + ], + "score": 0.85, + "content": "\\mathbf { \\delta } _ { i _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 115, + 221, + 128 + ], + "score": 1.0, + "content": ", forget gate", + "type": "text" + }, + { + "bbox": [ + 221, + 116, + 231, + 127 + ], + "score": 0.87, + "content": "\\pmb { f } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 115, + 272, + 128 + ], + "score": 1.0, + "content": ", cell state", + "type": "text" + }, + { + "bbox": [ + 273, + 117, + 283, + 127 + ], + "score": 0.85, + "content": "\\mathbf { \\pmb { g } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 115, + 348, + 128 + ], + "score": 1.0, + "content": "and output gate", + "type": "text" + }, + { + "bbox": [ + 349, + 115, + 361, + 127 + ], + "score": 0.9, + "content": "\\mathbf { \\sigma } _ { o _ { t } ^ { h } } ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "for the new hidden state just like in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 328, + 138 + ], + "score": 1.0, + "content": "classical LSTM structure. Additionally, an output gate", + "type": "text" + }, + { + "bbox": [ + 329, + 127, + 340, + 138 + ], + "score": 0.89, + "content": "{ \\mathbf { } } _ { o _ { t } ^ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 126, + 416, + 138 + ], + "score": 1.0, + "content": "for historical state", + "type": "text" + }, + { + "bbox": [ + 416, + 128, + 427, + 137 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "is computed. Next", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 285, + 150 + ], + "score": 1.0, + "content": "in equation 7, we compute new hidden state", + "type": "text" + }, + { + "bbox": [ + 286, + 138, + 297, + 148 + ], + "score": 0.89, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 137, + 371, + 150 + ], + "score": 1.0, + "content": "that is the sum of", + "type": "text" + }, + { + "bbox": [ + 371, + 137, + 393, + 149 + ], + "score": 0.91, + "content": "h _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 137, + 449, + 150 + ], + "score": 1.0, + "content": "and cell state", + "type": "text" + }, + { + "bbox": [ + 450, + 139, + 460, + 149 + ], + "score": 0.84, + "content": "\\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } } \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\lambda } }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 137, + 505, + 150 + ], + "score": 1.0, + "content": ", leveraged", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 167, + 161 + ], + "score": 1.0, + "content": "by forget gate", + "type": "text" + }, + { + "bbox": [ + 168, + 149, + 178, + 160 + ], + "score": 0.86, + "content": "\\pmb { f } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 148, + 242, + 161 + ], + "score": 1.0, + "content": "and input gate", + "type": "text" + }, + { + "bbox": [ + 243, + 149, + 251, + 159 + ], + "score": 0.86, + "content": "\\dot { \\mathbf { \\delta } } _ { \\mathbf { \\delta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 148, + 505, + 161 + ], + "score": 1.0, + "content": ". Finally in equation 8, we calculate the output of this time-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 348, + 172 + ], + "score": 1.0, + "content": "step, which is the concatenation of the gated contents from", + "type": "text" + }, + { + "bbox": [ + 348, + 160, + 360, + 170 + ], + "score": 0.87, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 159, + 378, + 172 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 160, + 388, + 170 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 159, + 505, + 172 + ], + "score": 1.0, + "content": ". Using this method, we can", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 417, + 183 + ], + "score": 1.0, + "content": "selectively backpropagate gradients from the output to the historical state", + "type": "text" + }, + { + "bbox": [ + 417, + 172, + 427, + 181 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 169, + 479, + 183 + ], + "score": 1.0, + "content": ". We involve", + "type": "text" + }, + { + "bbox": [ + 479, + 172, + 489, + 181 + ], + "score": 0.83, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 169, + 506, + 183 + ], + "score": 1.0, + "content": "(or", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 230, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 118, + 194 + ], + "score": 0.82, + "content": "\\boldsymbol { r } _ { t } ^ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 181, + 230, + 194 + ], + "score": 1.0, + "content": ") in all gates’ computations.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 104, + 82, + 506, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 168, + 211 + ], + "score": 1.0, + "content": "Intuitively, the", + "type": "text" + }, + { + "bbox": [ + 168, + 198, + 190, + 210 + ], + "score": 0.91, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 197, + 345, + 211 + ], + "score": 1.0, + "content": "acts as the old working memory, and", + "type": "text" + }, + { + "bbox": [ + 346, + 200, + 356, + 209 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 197, + 505, + 211 + ], + "score": 1.0, + "content": "is treated as the long-term memory.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 265, + 222 + ], + "score": 1.0, + "content": "The cell processes them with the input", + "type": "text" + }, + { + "bbox": [ + 265, + 211, + 276, + 220 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 208, + 431, + 222 + ], + "score": 1.0, + "content": "to generate the new working memory", + "type": "text" + }, + { + "bbox": [ + 432, + 210, + 443, + 220 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "and the output", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 117, + 231 + ], + "score": 0.82, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 219, + 217, + 234 + ], + "score": 1.0, + "content": ". More specifically, each", + "type": "text" + }, + { + "bbox": [ + 218, + 221, + 228, + 231 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 219, + 506, + 234 + ], + "score": 1.0, + "content": "is a summary of historical hidden states selected by auto-addressing", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "mechanism. The ARMIN cell learns to recurrently integrate the summary of long-term information", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 129, + 255 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 244, + 140, + 253 + ], + "score": 0.83, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 241, + 250, + 255 + ], + "score": 1.0, + "content": "into the working memory", + "type": "text" + }, + { + "bbox": [ + 251, + 242, + 262, + 253 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 241, + 505, + 255 + ], + "score": 1.0, + "content": ". We will demonstrate the efficiency of recurrent memory", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 268, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 268, + 266 + ], + "score": 1.0, + "content": "integration in language modelling tasks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 197, + 506, + 266 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 269, + 282 + ], + "score": 1.0, + "content": "Unlike the LSTM that passes a tuple of", + "type": "text" + }, + { + "bbox": [ + 270, + 270, + 301, + 282 + ], + "score": 0.93, + "content": "\\left( \\boldsymbol { h } _ { t } , \\boldsymbol { c } _ { t } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 269, + 443, + 282 + ], + "score": 1.0, + "content": "to the next recurrence and outputs", + "type": "text" + }, + { + "bbox": [ + 443, + 270, + 455, + 281 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "at the same", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 302, + 293 + ], + "score": 1.0, + "content": "time, the ARMIN only passes one hidden state", + "type": "text" + }, + { + "bbox": [ + 302, + 281, + 313, + 292 + ], + "score": 0.87, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 281, + 397, + 293 + ], + "score": 1.0, + "content": "and separately uses", + "type": "text" + }, + { + "bbox": [ + 397, + 283, + 408, + 292 + ], + "score": 0.85, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "as output, which has 2", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "benefits: one is that it takes lower memory consumption to store one hidden state rather than two,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 402, + 315 + ], + "score": 1.0, + "content": "and the other is we can decouple the information needed for the output", + "type": "text" + }, + { + "bbox": [ + 402, + 304, + 412, + 314 + ], + "score": 0.85, + "content": "\\mathbf { } _ \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 303, + 505, + 315 + ], + "score": 1.0, + "content": ", from the information", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 126, + 326 + ], + "score": 1.0, + "content": "(i.e.", + "type": "text" + }, + { + "bbox": [ + 126, + 314, + 138, + 325 + ], + "score": 0.7, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 313, + 505, + 326 + ], + "score": 1.0, + "content": ") that is needed for memorizing and being used at later time-steps. Furthermore, we list 3", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 502, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 502, + 336 + ], + "score": 1.0, + "content": "important differences between the ARMIN cell and TARDIS cell, please refer to APPENDIX B.1.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 269, + 506, + 336 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 349, + 320, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 321, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 321, + 362 + ], + "score": 1.0, + "content": "3.2 READ OPERATION WITH AUTO-ADDRESSING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 275, + 383 + ], + "score": 1.0, + "content": "The ARMIN maintains a memory matrix", + "type": "text" + }, + { + "bbox": [ + 276, + 369, + 346, + 380 + ], + "score": 0.91, + "content": "M \\in \\mathbb { R } ^ { n _ { m e m } \\times d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 367, + 429, + 383 + ], + "score": 1.0, + "content": ", where the constant", + "type": "text" + }, + { + "bbox": [ + 430, + 372, + 455, + 381 + ], + "score": 0.89, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 367, + 506, + 383 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 437, + 395 + ], + "score": 1.0, + "content": "number of memory slots. At each recurrence, the ARMIN chooses a historical state", + "type": "text" + }, + { + "bbox": [ + 438, + 383, + 447, + 392 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 379, + 506, + 395 + ], + "score": 1.0, + "content": "from memory", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 392, + 417, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 232, + 404 + ], + "score": 1.0, + "content": "according to the information in", + "type": "text" + }, + { + "bbox": [ + 233, + 393, + 244, + 403 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 392, + 262, + 404 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 262, + 392, + 284, + 404 + ], + "score": 0.9, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 392, + 417, + 404 + ], + "score": 1.0, + "content": ",which is formulated as follows:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 367, + 506, + 404 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 417, + 401, + 468 + ], + "lines": [ + { + "bbox": [ + 209, + 417, + 401, + 468 + ], + "spans": [ + { + "bbox": [ + 209, + 417, + 401, + 468 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\displaystyle s _ { t } = \\mathrm { g u m b e l - s o f t m a x } \\big ( W _ { s } [ \\ : x _ { t } \\ : , h _ { t - 1 } \\ : ] + b _ { s } \\big ) \\ : , } } \\\\ { { { } } } \\\\ { { \\displaystyle r _ { t } = \\sum _ { i = 0 } ^ { n - 1 } s _ { t } ( i ) M ( i , : ) \\ : . } } \\end{array}", + "type": "interline_equation", + "image_path": "d20fa90dbab7b45d436ec9d1fe2698fcfc15d2cb48c0e332be86d0aefcd202b3.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 209, + 417, + 401, + 434.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 209, + 434.0, + 401, + 451.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 209, + 451.0, + 401, + 468.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 471, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 134, + 486 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 470, + 229, + 483 + ], + "score": 0.87, + "content": "W _ { s } \\ \\in \\ \\mathbb { R } ^ { n _ { m e m } \\times ( d _ { i } + d _ { h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 468, + 234, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 235, + 471, + 289, + 483 + ], + "score": 0.84, + "content": "b _ { s } \\in \\mathbb { R } ^ { n _ { m e m } }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 468, + 294, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 295, + 473, + 305, + 483 + ], + "score": 0.73, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { S } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 468, + 505, + 486 + ], + "score": 1.0, + "content": "is a one-hot vector sampled by gumbel-softmax", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 482, + 435, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 145, + 495 + ], + "score": 1.0, + "content": "function,", + "type": "text" + }, + { + "bbox": [ + 145, + 483, + 165, + 495 + ], + "score": 0.91, + "content": "s _ { t } ( i )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 482, + 214, + 495 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 214, + 484, + 219, + 493 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 482, + 276, + 495 + ], + "score": 1.0, + "content": "-th element of", + "type": "text" + }, + { + "bbox": [ + 276, + 484, + 286, + 494 + ], + "score": 0.75, + "content": "\\mathbf { \\boldsymbol { s } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 482, + 289, + 495 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 290, + 483, + 322, + 495 + ], + "score": 0.91, + "content": "M ( i , : )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 482, + 370, + 495 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 371, + 484, + 375, + 493 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 482, + 416, + 495 + ], + "score": 1.0, + "content": "-th row of", + "type": "text" + }, + { + "bbox": [ + 416, + 483, + 430, + 493 + ], + "score": 0.81, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 482, + 435, + 495 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 468, + 505, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "As opposed to previous memory networks such as Santoro et al. (2016) and Gulcehre et al. (2017),", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "we don’t use any extra usage vectors and learnable key parameters that manually encode memory", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "addressing information to assist memory addressing. As our ablation study (Appendix E) shows,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 172, + 545 + ], + "score": 1.0, + "content": "the hidden state", + "type": "text" + }, + { + "bbox": [ + 172, + 533, + 194, + 544 + ], + "score": 0.92, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "is sufficient to encode the historical memory accessing information. Further-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 371, + 557 + ], + "score": 1.0, + "content": "more, the TARDIS addressing mechanism requires concatenating", + "type": "text" + }, + { + "bbox": [ + 372, + 545, + 383, + 555 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 543, + 401, + 557 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 544, + 413, + 555 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "to every memory cell,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "which usually causes large extra memory and time consumption in practical scenes. Please refer to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "APPENDIX B.2 for more precise comparison regarding addressing mechanism of the ARMIN and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 576, + 185, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 185, + 588 + ], + "score": 1.0, + "content": "TARDIS networks.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 500, + 506, + 588 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 601, + 212, + 612 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 214, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 214, + 613 + ], + "score": 1.0, + "content": "3.3 WRITE OPERATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 464, + 634 + ], + "score": 1.0, + "content": "After the reading and cell processing stages, the ARMIN writes the new hidden state", + "type": "text" + }, + { + "bbox": [ + 465, + 622, + 476, + 633 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "to the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 143, + 646 + ], + "score": 1.0, + "content": "memory", + "type": "text" + }, + { + "bbox": [ + 143, + 633, + 156, + 643 + ], + "score": 0.59, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 633, + 481, + 646 + ], + "score": 1.0, + "content": ". Following the TARDIS network (Gulcehre et al., 2017), we simply overwrite", + "type": "text" + }, + { + "bbox": [ + 481, + 633, + 493, + 644 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 283, + 657 + ], + "score": 1.0, + "content": "the memory slot where we just read out the", + "type": "text" + }, + { + "bbox": [ + 284, + 645, + 294, + 655 + ], + "score": 0.84, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 643, + 384, + 657 + ], + "score": 1.0, + "content": "(for conditions where", + "type": "text" + }, + { + "bbox": [ + 385, + 644, + 396, + 655 + ], + "score": 0.88, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 643, + 457, + 657 + ], + "score": 1.0, + "content": "is not equal to", + "type": "text" + }, + { + "bbox": [ + 458, + 644, + 468, + 655 + ], + "score": 0.88, + "content": "d _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 643, + 506, + 657 + ], + "score": 1.0, + "content": ", we first", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 228, + 669 + ], + "score": 1.0, + "content": "use a linear layer to transform", + "type": "text" + }, + { + "bbox": [ + 228, + 655, + 239, + 666 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 654, + 262, + 669 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 263, + 655, + 274, + 666 + ], + "score": 0.88, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 654, + 329, + 669 + ], + "score": 1.0, + "content": "dimension to", + "type": "text" + }, + { + "bbox": [ + 330, + 655, + 340, + 666 + ], + "score": 0.88, + "content": "d _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "dimension). But at the initial time-steps,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "we write the hidden states to the empty memory slots, until all empty slots are filled with historical", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "states. In this way, we can maximally preserve useful historical information, because the ARMIN", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 304, + 700 + ], + "score": 1.0, + "content": "cell can learn to copy the useful information from", + "type": "text" + }, + { + "bbox": [ + 304, + 689, + 313, + 699 + ], + "score": 0.83, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 687, + 324, + 700 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 325, + 688, + 336, + 699 + ], + "score": 0.87, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 687, + 397, + 700 + ], + "score": 1.0, + "content": ", and then write", + "type": "text" + }, + { + "bbox": [ + 398, + 688, + 409, + 699 + ], + "score": 0.88, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "to the previous location", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 117, + 712 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 700, + 127, + 710 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 699, + 505, + 712 + ], + "score": 1.0, + "content": ". The read/write mechanism can be viewed as a form of a skip-connection of hidden states. By", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "training the whole network, the ARMIN can easily learn long-term dependencies via direct access", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "to historical hidden states. From this point of view, our network is similar to the Skip RNN proposed", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "by Campos Camunez et al. (2018), which differs from our network in that they use a binary state", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "update gate to select whether the state of the RNN will be updated or copied from the previous", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 149, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 149, + 118 + ], + "score": 1.0, + "content": "time-step.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 622, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "by Campos Camunez et al. (2018), which differs from our network in that they use a binary state", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "update gate to select whether the state of the RNN will be updated or copied from the previous", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 149, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 149, + 118 + ], + "score": 1.0, + "content": "time-step.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 131, + 200, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 201, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 201, + 146 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 156, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "We evaluate our model on algorithmic tasks, character-level language modelling task and temporal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "action detection and proposal task. We compare our network with NTM (with an LSTM controller),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "TARDIS and vanilla LSTM networks. We also directly compare the auto-addressing against the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "TARDIS addressing mechanism, by replacing auto-addressing with the TARDIS addressing method", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 300, + 212 + ], + "score": 1.0, + "content": "in ARMIN. We refer to this network as ARMIN", + "type": "text" + }, + { + "bbox": [ + 300, + 202, + 306, + 210 + ], + "score": 0.44, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "TARDIS-addr in our experiments. Please refer to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 211, + 410, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 410, + 223 + ], + "score": 1.0, + "content": "APPENDIX C for some universal setups used in all our experiments below.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 236, + 223, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 225, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 225, + 248 + ], + "score": 1.0, + "content": "4.1 ALGORITHMIC TASKS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 256, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 504, + 268 + ], + "score": 1.0, + "content": "Along with the NTM, Graves et al. (2014) introduced a set of synthetic algorithm tasks which con-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "sists of copy, repeat copy, associative recall, N-gram and priority sort tasks. Here we use four out", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "score": 1.0, + "content": "of five of these tasks (excluding the N-gram task) to examine if our network can choose correct", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 290, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 301 + ], + "score": 1.0, + "content": "time-steps from the past and effectively make use of them. Please refer to APPENDIX C.1 for more", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 299, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 314 + ], + "score": 1.0, + "content": "explanations and hyperparameter setups with respect to these tasks. Note that we have a strong", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "LSTM baseline with about 4 times larger parameter count than memory networks, and the train-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "ing of our copy and priority sort tasks are more challenging than the original tasks in Graves et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "(2014), for having more than 2 times longer input and target sequences. The models are optimized to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "minimize the averaged binary cross-entropy loss on the target multi-bit binary sequences. Following", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "Gulcehre et al. (2017); Campos Camunez et al. (2018), in all tasks, we consider a task solved if the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 365, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 379 + ], + "score": 1.0, + "content": "validation loss is at least two orders of magnitude below the initial cost which is around 0.70, i.e.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "the validation loss (the validation set is generated randomly) converges to less than 0.01 (with less", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 126, + 401 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 126, + 388, + 146, + 399 + ], + "score": 0.88, + "content": "3 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "in 10 consecutive validation of sharp losses that are higher than 0.01 ). In our evaluation,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "we are interested in if a model can successfully solve the task in 100k iterations and the elapsed", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "score": 1.0, + "content": "training time and iterations till the model succeeds on the task. For all networks we ensure the sim-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "ilar total parameter counts. All important hyperparameters are shown in APPENDIX C.1. We run", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 431, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 445 + ], + "score": 1.0, + "content": "all experiments with batch size of 1 under 2.8 GHz Intel Core i7 CPU and report the wall clock time", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "score": 1.0, + "content": "and the number of iterations for solved tasks of each model. For models that fail to converge to less", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 176, + 466 + ], + "score": 1.0, + "content": "than 0.01 loss in", + "type": "text" + }, + { + "bbox": [ + 177, + 454, + 198, + 464 + ], + "score": 0.32, + "content": "1 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "iterations, we report the average loss of the final 10 validations (denoted in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "underlines) and the elapsed training time for 100k iterations. The results are shown in Table 1. The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 477, + 338, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 338, + 487 + ], + "score": 1.0, + "content": "training curves for each task are shown in APPENDIX D.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "The results show that the ARMIN can solve 3 out of 4 tasks in a fast converge speed. By comparing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "the ARMIN with NTM, we observe that the NTM is able to almost solve all of the tasks, but with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "a much slower training speed compared to the ARMIN, for example, the training time of NTM is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "5 times larger than ARMIN to solve the copy task. In fact, the training speed of ARMIN for each", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 179, + 549 + ], + "score": 1.0, + "content": "iteration is about", + "type": "text" + }, + { + "bbox": [ + 180, + 537, + 209, + 547 + ], + "score": 0.88, + "content": "3 \\sim 4", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "times faster than the NTM’s speed. We observe that the TARDIS and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 140, + 560 + ], + "score": 1.0, + "content": "ARMIN", + "type": "text" + }, + { + "bbox": [ + 140, + 549, + 147, + 557 + ], + "score": 0.54, + "content": "\\dotplus", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 547, + 357, + 560 + ], + "score": 1.0, + "content": "TARDIS-addr fail to solve 3 out 4 tasks in the given", + "type": "text" + }, + { + "bbox": [ + 357, + 547, + 379, + 558 + ], + "score": 0.62, + "content": "1 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "iterations. By comparing them", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "with the ARMIN, we confirm the efficiency of the auto-addressing of ARMIN, for bringing a fast", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + }, + { + "type": "table", + "bbox": [ + 121, + 634, + 490, + 725 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 586, + 504, + 620 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 329, + 599 + ], + "score": 1.0, + "content": "Table 1: The average elapsed time(min) and iterations", + "type": "text" + }, + { + "bbox": [ + 329, + 587, + 341, + 597 + ], + "score": 0.49, + "content": "{ \\mathrm { ( k ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "of different networks till task solved in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 131, + 610 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 131, + 598, + 153, + 608 + ], + "score": 0.35, + "content": "1 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "iterations. The wall clock training time for 100k iterations and the average loss of final", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 609, + 394, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 394, + 620 + ], + "score": 1.0, + "content": "10 validations (denoted in underlines) are shown for the unsolved tasks.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "table_body", + "bbox": [ + 121, + 634, + 490, + 725 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 634, + 490, + 725 + ], + "spans": [ + { + "bbox": [ + 121, + 634, + 490, + 725 + ], + "score": 0.985, + "html": "
CopyRepeat CopyAssociative RecallPriority Sort
timeiter.timeiter.timeiter.timeiter.
LSTM660.359450.0160.3253237.0
NTM3212.41710.01419.63600.012
TARDIS1570.451970.1663422660.33013362.6
ARMIN+TARDIS-addr1520.410950.297640.33615171.4
ARMIN67.63667.8330.0523733.4
", + "type": "table", + "image_path": "3e9670eb5cd4fc0e6980a395033241124564e6ed25b13ade887f0170637515a7.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 121, + 634, + 490, + 664.3333333333334 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 121, + 664.3333333333334, + 490, + 694.6666666666667 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 121, + 694.6666666666667, + 490, + 725.0000000000001 + ], + "spans": [], + "index": 44 + } + ] + } + ], + "index": 41.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 118 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 131, + 200, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 201, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 201, + 146 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 156, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "We evaluate our model on algorithmic tasks, character-level language modelling task and temporal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "action detection and proposal task. We compare our network with NTM (with an LSTM controller),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "TARDIS and vanilla LSTM networks. We also directly compare the auto-addressing against the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "TARDIS addressing mechanism, by replacing auto-addressing with the TARDIS addressing method", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 300, + 212 + ], + "score": 1.0, + "content": "in ARMIN. We refer to this network as ARMIN", + "type": "text" + }, + { + "bbox": [ + 300, + 202, + 306, + 210 + ], + "score": 0.44, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "TARDIS-addr in our experiments. Please refer to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 211, + 410, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 410, + 223 + ], + "score": 1.0, + "content": "APPENDIX C for some universal setups used in all our experiments below.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 156, + 505, + 223 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 236, + 223, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 225, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 225, + 248 + ], + "score": 1.0, + "content": "4.1 ALGORITHMIC TASKS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 256, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 504, + 268 + ], + "score": 1.0, + "content": "Along with the NTM, Graves et al. (2014) introduced a set of synthetic algorithm tasks which con-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "sists of copy, repeat copy, associative recall, N-gram and priority sort tasks. Here we use four out", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "score": 1.0, + "content": "of five of these tasks (excluding the N-gram task) to examine if our network can choose correct", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 290, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 301 + ], + "score": 1.0, + "content": "time-steps from the past and effectively make use of them. Please refer to APPENDIX C.1 for more", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 299, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 314 + ], + "score": 1.0, + "content": "explanations and hyperparameter setups with respect to these tasks. Note that we have a strong", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "LSTM baseline with about 4 times larger parameter count than memory networks, and the train-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "ing of our copy and priority sort tasks are more challenging than the original tasks in Graves et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "(2014), for having more than 2 times longer input and target sequences. The models are optimized to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "minimize the averaged binary cross-entropy loss on the target multi-bit binary sequences. Following", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "Gulcehre et al. (2017); Campos Camunez et al. (2018), in all tasks, we consider a task solved if the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 365, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 379 + ], + "score": 1.0, + "content": "validation loss is at least two orders of magnitude below the initial cost which is around 0.70, i.e.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "the validation loss (the validation set is generated randomly) converges to less than 0.01 (with less", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 126, + 401 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 126, + 388, + 146, + 399 + ], + "score": 0.88, + "content": "3 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "in 10 consecutive validation of sharp losses that are higher than 0.01 ). In our evaluation,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "we are interested in if a model can successfully solve the task in 100k iterations and the elapsed", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "score": 1.0, + "content": "training time and iterations till the model succeeds on the task. For all networks we ensure the sim-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 434 + ], + "score": 1.0, + "content": "ilar total parameter counts. All important hyperparameters are shown in APPENDIX C.1. We run", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 431, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 445 + ], + "score": 1.0, + "content": "all experiments with batch size of 1 under 2.8 GHz Intel Core i7 CPU and report the wall clock time", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "score": 1.0, + "content": "and the number of iterations for solved tasks of each model. For models that fail to converge to less", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 176, + 466 + ], + "score": 1.0, + "content": "than 0.01 loss in", + "type": "text" + }, + { + "bbox": [ + 177, + 454, + 198, + 464 + ], + "score": 0.32, + "content": "1 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "iterations, we report the average loss of the final 10 validations (denoted in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "underlines) and the elapsed training time for 100k iterations. The results are shown in Table 1. The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 477, + 338, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 338, + 487 + ], + "score": 1.0, + "content": "training curves for each task are shown in APPENDIX D.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 257, + 506, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "The results show that the ARMIN can solve 3 out of 4 tasks in a fast converge speed. By comparing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "the ARMIN with NTM, we observe that the NTM is able to almost solve all of the tasks, but with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "a much slower training speed compared to the ARMIN, for example, the training time of NTM is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "5 times larger than ARMIN to solve the copy task. In fact, the training speed of ARMIN for each", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 179, + 549 + ], + "score": 1.0, + "content": "iteration is about", + "type": "text" + }, + { + "bbox": [ + 180, + 537, + 209, + 547 + ], + "score": 0.88, + "content": "3 \\sim 4", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "times faster than the NTM’s speed. We observe that the TARDIS and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 140, + 560 + ], + "score": 1.0, + "content": "ARMIN", + "type": "text" + }, + { + "bbox": [ + 140, + 549, + 147, + 557 + ], + "score": 0.54, + "content": "\\dotplus", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 547, + 357, + 560 + ], + "score": 1.0, + "content": "TARDIS-addr fail to solve 3 out 4 tasks in the given", + "type": "text" + }, + { + "bbox": [ + 357, + 547, + 379, + 558 + ], + "score": 0.62, + "content": "1 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "iterations. By comparing them", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "with the ARMIN, we confirm the efficiency of the auto-addressing of ARMIN, for bringing a fast", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "converge speed and short training time. We also find that the LSTM, when given enough parameter", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "counts (4 times larger than memory networks), is able to solve the priority sort task quickly and", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 352, + 400 + ], + "score": 1.0, + "content": "converge to a better performance than TARDIS and ARMIN", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 353, + 389, + 359, + 397 + ], + "score": 0.39, + "content": "^ +", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 360, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "TARDIS in other tasks. We believe", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "there are 2 possible reasons for the poor performance of the TARDIS addressing mechanism: a) The", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 410, + 504, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 504, + 421 + ], + "score": 1.0, + "content": "tanh activation (see equation 21 in APPENDIX B.2) might cause the gradient vanishing problem", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 420, + 503, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 435, + 434 + ], + "score": 1.0, + "content": "when it cooperates with the gumbel-softmax function. b) The multiple inputs of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 435, + 420, + 503, + 433 + ], + "score": 0.88, + "content": "\\pmb { x } _ { t } , \\pmb { h } _ { t } , \\pmb { M } _ { t } [ i ] , \\pmb { u } _ { t }", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "in equation 21 cause the network hard to initialize to keep equal input and output variances for", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "the linear layer and might also cause gradient dispersion. The auto-addressing has neither of these", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 453, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 466 + ], + "score": 1.0, + "content": "two issues and can better integrately backpropogate gradients, therefore it can perform well and", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "quickly on these tasks. However, we would like to point out 2 limitations of the auto-addressing: a)", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "The associative recall task shows the auto-addressing might have problems in establishing indirect", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "score": 1.0, + "content": "references among memory data. b) The auto-addressing is relatively hard to generalize to unseen", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "longer sequence in tasks such as copy and repeat copy due to its simplicity, whereas the NTM is", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 509, + 329, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 329, + 520 + ], + "score": 1.0, + "content": "able to do that. We leave these 2 issues for future work.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 491, + 505, + 570 + ] + }, + { + "type": "table", + "bbox": [ + 121, + 634, + 490, + 725 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 586, + 504, + 620 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 329, + 599 + ], + "score": 1.0, + "content": "Table 1: The average elapsed time(min) and iterations", + "type": "text" + }, + { + "bbox": [ + 329, + 587, + 341, + 597 + ], + "score": 0.49, + "content": "{ \\mathrm { ( k ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "of different networks till task solved in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 131, + 610 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 131, + 598, + 153, + 608 + ], + "score": 0.35, + "content": "1 0 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "iterations. The wall clock training time for 100k iterations and the average loss of final", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 609, + 394, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 394, + 620 + ], + "score": 1.0, + "content": "10 validations (denoted in underlines) are shown for the unsolved tasks.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "table_body", + "bbox": [ + 121, + 634, + 490, + 725 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 634, + 490, + 725 + ], + "spans": [ + { + "bbox": [ + 121, + 634, + 490, + 725 + ], + "score": 0.985, + "html": "
CopyRepeat CopyAssociative RecallPriority Sort
timeiter.timeiter.timeiter.timeiter.
LSTM660.359450.0160.3253237.0
NTM3212.41710.01419.63600.012
TARDIS1570.451970.1663422660.33013362.6
ARMIN+TARDIS-addr1520.410950.297640.33615171.4
ARMIN67.63667.8330.0523733.4
", + "type": "table", + "image_path": "3e9670eb5cd4fc0e6980a395033241124564e6ed25b13ade887f0170637515a7.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 121, + 634, + 490, + 664.3333333333334 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 121, + 664.3333333333334, + 490, + 694.6666666666667 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 121, + 694.6666666666667, + 490, + 725.0000000000001 + ], + "spans": [], + "index": 44 + } + ] + } + ], + "index": 41.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 105, + 488, + 342 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 172, + 81, + 437, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 80, + 438, + 93 + ], + "spans": [ + { + "bbox": [ + 173, + 80, + 438, + 93 + ], + "score": 1.0, + "content": "Table 2: Bits-per-character on Penn Treebank and enwik8 test set.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 122, + 105, + 488, + 342 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 105, + 488, + 342 + ], + "spans": [ + { + "bbox": [ + 122, + 105, + 488, + 342 + ], + "score": 0.983, + "html": "
ModelPTBenwik8
BPCParamsBPCParams
LSTM1.36一1.45
LSTM+Zoneout (Krueger et al., 2016)1.2711
LSTM+Layer Norm(1000 units)11.2674.26M1一
LSTM+Layer Norm+Zoneout(1024 units)²1.244.79M一
HM-LSTM+Layer Norm (Chung et al.,2016)1.2411.3235M
HyperLSTM+Layer Norm(Ha et al., 2016)1.21914.41M1.34026.5M
NASCell (Zoph & Le,2016)1.21416.28M
IndRNN (21 layers) (Li et al.,2018)1.2111
Recurrent Highway Hypernetwork (Suarez, 2017)1.19一 15.5M11
Fast-Slow LSTM (Mujika et al., 2017)1.1907.2M1.277一 27M
NTM(800 units, nhead = 1, Pdropout=0.6)
TARDIS(paper)1.5358.28M1
1.25一
TARDIS(1000 units,Pdropout=0.6)1.2689.2M一
ARMIN+TARDIS-addr(800 units,Pdropout=0.6)1.22310.2M
ARMIN (500 units, Pdropout=0.4) (Ours)1.2364.03M
ARMIN (800 units, Pdropout=0.6) (Ours)1.2029.8M
ARMIN(800 units,2 layer) (Ours)111.33121.6M
", + "type": "table", + "image_path": "4160d0d481224f188526c1216f05763bf331261cc0ba4a079eac59cba78e5098.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 122, + 105, + 488, + 184.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 184.0, + 488, + 263.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 263.0, + 488, + 342.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "converge speed and short training time. We also find that the LSTM, when given enough parameter", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "counts (4 times larger than memory networks), is able to solve the priority sort task quickly and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 352, + 400 + ], + "score": 1.0, + "content": "converge to a better performance than TARDIS and ARMIN", + "type": "text" + }, + { + "bbox": [ + 353, + 389, + 359, + 397 + ], + "score": 0.39, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "TARDIS in other tasks. We believe", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "there are 2 possible reasons for the poor performance of the TARDIS addressing mechanism: a) The", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 410, + 504, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 504, + 421 + ], + "score": 1.0, + "content": "tanh activation (see equation 21 in APPENDIX B.2) might cause the gradient vanishing problem", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 420, + 503, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 435, + 434 + ], + "score": 1.0, + "content": "when it cooperates with the gumbel-softmax function. b) The multiple inputs of", + "type": "text" + }, + { + "bbox": [ + 435, + 420, + 503, + 433 + ], + "score": 0.88, + "content": "\\pmb { x } _ { t } , \\pmb { h } _ { t } , \\pmb { M } _ { t } [ i ] , \\pmb { u } _ { t }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "in equation 21 cause the network hard to initialize to keep equal input and output variances for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "the linear layer and might also cause gradient dispersion. The auto-addressing has neither of these", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 453, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 466 + ], + "score": 1.0, + "content": "two issues and can better integrately backpropogate gradients, therefore it can perform well and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "quickly on these tasks. However, we would like to point out 2 limitations of the auto-addressing: a)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "The associative recall task shows the auto-addressing might have problems in establishing indirect", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 499 + ], + "score": 1.0, + "content": "references among memory data. b) The auto-addressing is relatively hard to generalize to unseen", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "longer sequence in tasks such as copy and repeat copy due to its simplicity, whereas the NTM is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 509, + 329, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 329, + 520 + ], + "score": 1.0, + "content": "able to do that. We leave these 2 issues for future work.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 533, + 322, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 323, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 323, + 545 + ], + "score": 1.0, + "content": "4.2 CHARACTER-LEVEL LANGUAGE MODELLING", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "The character-level language modelling task consists of predicting the probability distribution of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "next character given all the previous ones, and we use the widely used Penn Treebank and Hutter", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "Prize Wikipedia (also known as enwik8) datasets for this task. We compare our network with vanilla", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "LSTMs and other memory networks on Penn Treebank, and compare with some popular RNN", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 598, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 504, + 609 + ], + "score": 1.0, + "content": "variants on both datasets. We ensure similar parameter counts and hyperparameters for all mem-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "ory networks. Following prior works, we apply truncated backpropagation through time(TBPTT)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "(Rumelhart et al., 1986; Elman, 1990) to approximate the gradients: at each iteration, the network", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 630, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 315, + 642 + ], + "score": 1.0, + "content": "predict the next 150 characters, and the hidden state", + "type": "text" + }, + { + "bbox": [ + 316, + 631, + 327, + 641 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 630, + 402, + 642 + ], + "score": 1.0, + "content": "and memory state", + "type": "text" + }, + { + "bbox": [ + 402, + 630, + 416, + 640 + ], + "score": 0.53, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 630, + 506, + 642 + ], + "score": 1.0, + "content": "are passed to the next", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "iteration. The gradients are truncated between different iterations. Please refer to APPENDIX C.2", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 653, + 228, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 228, + 664 + ], + "score": 1.0, + "content": "for more experimental details.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "The results are shown in table 2. On Penn Treebank dataset, our best performing network achieves", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "competitive 1.202 BPC on Penn Treebank dataset, which is the best single cell performance that we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 459, + 703 + ], + "score": 1.0, + "content": "are aware of, with relatively small parameter count. By comparing TARDIS and ARMIN", + "type": "text" + }, + { + "bbox": [ + 459, + 692, + 465, + 701 + ], + "score": 0.42, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "TARDIS-", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 117, + 710, + 475, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 708, + 254, + 722 + ], + "spans": [ + { + "bbox": [ + 119, + 708, + 254, + 722 + ], + "score": 1.0, + "content": "1As implemented in Ha et al. (2016).", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 475, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 475, + 734 + ], + "score": 1.0, + "content": "2Our implementation. We don’t see performance growth when we further increase the hidden size.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 105, + 488, + 342 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 172, + 81, + 437, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 80, + 438, + 93 + ], + "spans": [ + { + "bbox": [ + 173, + 80, + 438, + 93 + ], + "score": 1.0, + "content": "Table 2: Bits-per-character on Penn Treebank and enwik8 test set.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 122, + 105, + 488, + 342 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 105, + 488, + 342 + ], + "spans": [ + { + "bbox": [ + 122, + 105, + 488, + 342 + ], + "score": 0.983, + "html": "
ModelPTBenwik8
BPCParamsBPCParams
LSTM1.36一1.45
LSTM+Zoneout (Krueger et al., 2016)1.2711
LSTM+Layer Norm(1000 units)11.2674.26M1一
LSTM+Layer Norm+Zoneout(1024 units)²1.244.79M一
HM-LSTM+Layer Norm (Chung et al.,2016)1.2411.3235M
HyperLSTM+Layer Norm(Ha et al., 2016)1.21914.41M1.34026.5M
NASCell (Zoph & Le,2016)1.21416.28M
IndRNN (21 layers) (Li et al.,2018)1.2111
Recurrent Highway Hypernetwork (Suarez, 2017)1.19一 15.5M11
Fast-Slow LSTM (Mujika et al., 2017)1.1907.2M1.277一 27M
NTM(800 units, nhead = 1, Pdropout=0.6)
TARDIS(paper)1.5358.28M1
1.25一
TARDIS(1000 units,Pdropout=0.6)1.2689.2M一
ARMIN+TARDIS-addr(800 units,Pdropout=0.6)1.22310.2M
ARMIN (500 units, Pdropout=0.4) (Ours)1.2364.03M
ARMIN (800 units, Pdropout=0.6) (Ours)1.2029.8M
ARMIN(800 units,2 layer) (Ours)111.33121.6M
", + "type": "table", + "image_path": "4160d0d481224f188526c1216f05763bf331261cc0ba4a079eac59cba78e5098.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 122, + 105, + 488, + 184.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 184.0, + 488, + 263.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 263.0, + 488, + 342.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 505, + 518 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 365, + 506, + 520 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 533, + 322, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 323, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 323, + 545 + ], + "score": 1.0, + "content": "4.2 CHARACTER-LEVEL LANGUAGE MODELLING", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "The character-level language modelling task consists of predicting the probability distribution of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "next character given all the previous ones, and we use the widely used Penn Treebank and Hutter", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "Prize Wikipedia (also known as enwik8) datasets for this task. We compare our network with vanilla", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "LSTMs and other memory networks on Penn Treebank, and compare with some popular RNN", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 598, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 504, + 609 + ], + "score": 1.0, + "content": "variants on both datasets. We ensure similar parameter counts and hyperparameters for all mem-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "ory networks. Following prior works, we apply truncated backpropagation through time(TBPTT)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "(Rumelhart et al., 1986; Elman, 1990) to approximate the gradients: at each iteration, the network", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 630, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 315, + 642 + ], + "score": 1.0, + "content": "predict the next 150 characters, and the hidden state", + "type": "text" + }, + { + "bbox": [ + 316, + 631, + 327, + 641 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 630, + 402, + 642 + ], + "score": 1.0, + "content": "and memory state", + "type": "text" + }, + { + "bbox": [ + 402, + 630, + 416, + 640 + ], + "score": 0.53, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 630, + 506, + 642 + ], + "score": 1.0, + "content": "are passed to the next", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "iteration. The gradients are truncated between different iterations. Please refer to APPENDIX C.2", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 653, + 228, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 228, + 664 + ], + "score": 1.0, + "content": "for more experimental details.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 553, + 506, + 664 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "The results are shown in table 2. On Penn Treebank dataset, our best performing network achieves", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "competitive 1.202 BPC on Penn Treebank dataset, which is the best single cell performance that we", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 459, + 703 + ], + "score": 1.0, + "content": "are aware of, with relatively small parameter count. By comparing TARDIS and ARMIN", + "type": "text" + }, + { + "bbox": [ + 459, + 692, + 465, + 701 + ], + "score": 0.42, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "TARDIS-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "addr, we observe about 4 points of improvement, which shows the efficiency of the ARMIN cell. By", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 186, + 106 + ], + "score": 1.0, + "content": "comparing ARMIN", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 186, + 95, + 192, + 103 + ], + "score": 0.36, + "content": "+", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 193, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "TARDIS-addr, we observe a further improvement of around 2 points, which", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "shows the efficiency of the auto-addressing mechanism. We also find that the same architecture of", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "NTM that performs well in algorithmic tasks fails to converge to a lower BPC than vanilla LSTM,", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 309, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 309, + 139 + ], + "score": 1.0, + "content": "even with various regularizations we add to NTM.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 668, + 505, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "addr, we observe about 4 points of improvement, which shows the efficiency of the ARMIN cell. By", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 186, + 106 + ], + "score": 1.0, + "content": "comparing ARMIN", + "type": "text" + }, + { + "bbox": [ + 186, + 95, + 192, + 103 + ], + "score": 0.36, + "content": "+", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "TARDIS-addr, we observe a further improvement of around 2 points, which", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "shows the efficiency of the auto-addressing mechanism. We also find that the same architecture of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "NTM that performs well in algorithmic tasks fails to converge to a lower BPC than vanilla LSTM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 309, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 309, + 139 + ], + "score": 1.0, + "content": "even with various regularizations we add to NTM.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "The results on Penn Treebank show our single ARMIN cell learns better representations than many", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 504, + 166 + ], + "score": 1.0, + "content": "hierarchical RNN structures, such as the HM-LSTM, 2-Layer HyperLSTM and 21 layer IndRNN.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Our network is outperformed by Recurrent Highway Hypernetwork and Fast-slow LSTM which are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "both multi-scale and deep transition RNNs and are state-of-the-art results without any auxiliary tech-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "nics on this dataset. By saying “better representations”, we refer to the concatenation of the gated", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 164, + 210 + ], + "score": 1.0, + "content": "contents from", + "type": "text" + }, + { + "bbox": [ + 164, + 198, + 176, + 209 + ], + "score": 0.89, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 198, + 194, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 194, + 199, + 204, + 209 + ], + "score": 0.86, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 198, + 471, + 210 + ], + "score": 1.0, + "content": "as in equation 8. For example, if we remove the gated contents of", + "type": "text" + }, + { + "bbox": [ + 472, + 199, + 482, + 209 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 211, + 117, + 220 + ], + "score": 0.82, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 209, + 505, + 222 + ], + "score": 1.0, + "content": ", the ARMIN undergoes a BPC performance drop from 1.202 to 1.220, which is still better than", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "the best performing BPC of 1.24 of the LSTM. We believe the rest of the performance gain comes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 229, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 104, + 229, + 506, + 245 + ], + "score": 1.0, + "content": "from the recurrent memory integration of the ARMIN cell, which also favors a paradigm of deep", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "transition, as is shown by the success of the deep transition RNNs on this task. For more ablation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 265 + ], + "score": 1.0, + "content": "study regarding the auto-addressing mechanism and the ARMIN cell please refer to APPENDIX E", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 263, + 380, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 380, + 277 + ], + "score": 1.0, + "content": "(where we show the robustness of ARMIN to small TBPTT length).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 108, + 281, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 504, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 504, + 294 + ], + "score": 1.0, + "content": "The result on enwik8 demonstrates a simple 2-layer ARMIN can achieve competitive BPC perfor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "mance of 1.33, with less parameter count compared to the HyperLSTM and HM-LSTM. By con-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "structing deeper ARMIN network or even combining with other multi-scale and hierarchical RNN", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 314, + 375, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 375, + 326 + ], + "score": 1.0, + "content": "architectures, we believe the performance can be further improved.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 338, + 419, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 421, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 421, + 351 + ], + "score": 1.0, + "content": "4.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 358, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "In this subsection, we evaluate the ARMIN on a more complicated real-world task concerning video", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "score": 1.0, + "content": "analysis, i.e. the temporal action detection and proposals, which consists of taking an input video", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "and producing a set of temporal intervals that are likely to contain human actions (Buch et al., 2017).", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "We use the Single-Stream Temporal Action Proposal(SST) (Buch et al., 2017) as our framework, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 450, + 415 + ], + "score": 1.0, + "content": "evaluate the model on the THUMOS’ 14 dataset (Jiang et al., 2014), which contains", + "type": "text" + }, + { + "bbox": [ + 450, + 403, + 468, + 413 + ], + "score": 0.82, + "content": "^ { 2 0 + }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "hours of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "video with 200 train/validation and 213 test untrimmed video sequences. The SST consists of a GRU", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "and a fully connected output layer. In our experiment, we replace the GRU encoder in SST with a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "single ARMIN cell, and we call the modified network MA-SST. For fairness, we also implement a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 447, + 504, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 504, + 459 + ], + "score": 1.0, + "content": "modified SST with LSTM encoder and compare them with the original SST in Buch et al. (2017).", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 356, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 356, + 469 + ], + "score": 1.0, + "content": "Please refer to APPENDIX C.3 for more experimental details.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26.5 + }, + { + "type": "table", + "bbox": [ + 309, + 500, + 506, + 628 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 495, + 306, + 550 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 494, + 307, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 307, + 507 + ], + "score": 1.0, + "content": "Table 3: Comparison of proposal generation per-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 506, + 307, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 307, + 518 + ], + "score": 1.0, + "content": "formance in terms of recall at 1000 proposals.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 517, + 307, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 153, + 528 + ], + "score": 1.0, + "content": "We choose", + "type": "text" + }, + { + "bbox": [ + 153, + 517, + 191, + 528 + ], + "score": 0.35, + "content": "\\mathrm { \\ t I o U { = } 0 } . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 517, + 307, + 528 + ], + "score": 1.0, + "content": "and 0.8 to compare for con-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 308, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 308, + 540 + ], + "score": 1.0, + "content": "sistency with Buch et al. (2017). Our MA-SST", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 539, + 250, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 208, + 550 + ], + "score": 1.0, + "content": "achieves highest recall at", + "type": "text" + }, + { + "bbox": [ + 209, + 540, + 246, + 550 + ], + "score": 0.52, + "content": "\\mathrm { \\ t I o U { = } } 0 . 8", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 539, + 250, + 550 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "table_caption", + "bbox": [ + 306, + 476, + 506, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 507, + 488 + ], + "score": 1.0, + "content": "Table 4: Comparison for performance and cost of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 487, + 477, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 477, + 498 + ], + "score": 1.0, + "content": "different setups on Penn Treebank dataset.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "table_body", + "bbox": [ + 309, + 500, + 506, + 628 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 309, + 500, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 309, + 500, + 506, + 628 + ], + "score": 0.978, + "html": "
ModelARMINLSTM
Setup1212
Hidden sizenparam(M)nmemTtruncbatch sizeMemory(GB)Speed(chars/s)BPC5005501k1k
4.0254.814.794.791
101
5050100150
384300128128
3.493.562.363.27
98k75k71k70k
1.2381.2261.271.24
", + "type": "table", + "image_path": "9dee7e00530b4ee45916348ee8b30316fa8b7e2b2b83fe8150f9f7711f240679.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 309, + 500, + 506, + 514.2222222222222 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 309, + 514.2222222222222, + 506, + 528.4444444444443 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 309, + 528.4444444444443, + 506, + 542.6666666666665 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 309, + 542.6666666666665, + 506, + 556.8888888888887 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 309, + 556.8888888888887, + 506, + 571.1111111111109 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 309, + 571.1111111111109, + 506, + 585.333333333333 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 309, + 585.333333333333, + 506, + 599.5555555555552 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 309, + 599.5555555555552, + 506, + 613.7777777777774 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 309, + 613.7777777777774, + 506, + 627.9999999999995 + ], + "spans": [], + "index": 51 + } + ] + } + ], + "index": 37.5 + }, + { + "type": "table", + "bbox": [ + 108, + 556, + 304, + 608 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 556, + 304, + 608 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 556, + 304, + 608 + ], + "spans": [ + { + "bbox": [ + 108, + 556, + 304, + 608 + ], + "score": 0.95, + "html": "
ModeltIoU=0.6tIoU=0.8
SST (Paper)0.9200.672
SST (We implement)0.8970.732
MA-SST(Ours)0.9110.759
", + "type": "table", + "image_path": "850273973a2f428712c3ae412e72577c47e4474b112a60383bcfb3009629e5e7.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 108, + 556, + 304, + 569.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 108, + 569.0, + 304, + 582.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 108, + 582.0, + 304, + 595.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 108, + 595.0, + 304, + 608.0 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 649 + ], + "score": 1.0, + "content": "The average recall under different proposal numbers and tIoUs3 are depicted in figure 2. MA-SST", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "outperforms the original SST and our SST simultaneously. To the best of our knowledge, MA-SST", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 306, + 672 + ], + "score": 1.0, + "content": "achieves highest performance of recall at around", + "type": "text" + }, + { + "bbox": [ + 307, + 660, + 345, + 671 + ], + "score": 0.47, + "content": "\\mathrm { \\ t I o U { = } } 0 . 8", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "(see table 3), even if other competitive", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "non-RNN networks (Escorcia et al., 2016; Caba Heilbron et al., 2016; Gao et al., 2017; Guo et al.,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "2018) are taken into account. The MA-SST tends to precisely generate proposals that has a high", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "overlapping area with the ground truth action intervals, leading to high recall performance at high", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 54.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 712, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "3tIoU denotes temporal Intersection over Union between a proposal and its maximally overlapped ground", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "truth action. Applying bigger tIoU threshold make the generated proposals higher in quality but less in quantity.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 137 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "The results on Penn Treebank show our single ARMIN cell learns better representations than many", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 504, + 166 + ], + "score": 1.0, + "content": "hierarchical RNN structures, such as the HM-LSTM, 2-Layer HyperLSTM and 21 layer IndRNN.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Our network is outperformed by Recurrent Highway Hypernetwork and Fast-slow LSTM which are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "both multi-scale and deep transition RNNs and are state-of-the-art results without any auxiliary tech-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "nics on this dataset. By saying “better representations”, we refer to the concatenation of the gated", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 164, + 210 + ], + "score": 1.0, + "content": "contents from", + "type": "text" + }, + { + "bbox": [ + 164, + 198, + 176, + 209 + ], + "score": 0.89, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 198, + 194, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 194, + 199, + 204, + 209 + ], + "score": 0.86, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 198, + 471, + 210 + ], + "score": 1.0, + "content": "as in equation 8. For example, if we remove the gated contents of", + "type": "text" + }, + { + "bbox": [ + 472, + 199, + 482, + 209 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 211, + 117, + 220 + ], + "score": 0.82, + "content": "\\mathbf { } _ { o _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 209, + 505, + 222 + ], + "score": 1.0, + "content": ", the ARMIN undergoes a BPC performance drop from 1.202 to 1.220, which is still better than", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "the best performing BPC of 1.24 of the LSTM. We believe the rest of the performance gain comes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 229, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 104, + 229, + 506, + 245 + ], + "score": 1.0, + "content": "from the recurrent memory integration of the ARMIN cell, which also favors a paradigm of deep", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "transition, as is shown by the success of the deep transition RNNs on this task. For more ablation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 265 + ], + "score": 1.0, + "content": "study regarding the auto-addressing mechanism and the ARMIN cell please refer to APPENDIX E", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 263, + 380, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 380, + 277 + ], + "score": 1.0, + "content": "(where we show the robustness of ARMIN to small TBPTT length).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 143, + 506, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 281, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 504, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 504, + 294 + ], + "score": 1.0, + "content": "The result on enwik8 demonstrates a simple 2-layer ARMIN can achieve competitive BPC perfor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "mance of 1.33, with less parameter count compared to the HyperLSTM and HM-LSTM. By con-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "structing deeper ARMIN network or even combining with other multi-scale and hierarchical RNN", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 314, + 375, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 375, + 326 + ], + "score": 1.0, + "content": "architectures, we believe the performance can be further improved.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 280, + 506, + 326 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 338, + 419, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 421, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 421, + 351 + ], + "score": 1.0, + "content": "4.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 358, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "In this subsection, we evaluate the ARMIN on a more complicated real-world task concerning video", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "score": 1.0, + "content": "analysis, i.e. the temporal action detection and proposals, which consists of taking an input video", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "and producing a set of temporal intervals that are likely to contain human actions (Buch et al., 2017).", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "We use the Single-Stream Temporal Action Proposal(SST) (Buch et al., 2017) as our framework, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 450, + 415 + ], + "score": 1.0, + "content": "evaluate the model on the THUMOS’ 14 dataset (Jiang et al., 2014), which contains", + "type": "text" + }, + { + "bbox": [ + 450, + 403, + 468, + 413 + ], + "score": 0.82, + "content": "^ { 2 0 + }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "hours of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "video with 200 train/validation and 213 test untrimmed video sequences. The SST consists of a GRU", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "and a fully connected output layer. In our experiment, we replace the GRU encoder in SST with a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "single ARMIN cell, and we call the modified network MA-SST. For fairness, we also implement a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 447, + 504, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 504, + 459 + ], + "score": 1.0, + "content": "modified SST with LSTM encoder and compare them with the original SST in Buch et al. (2017).", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 356, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 356, + 469 + ], + "score": 1.0, + "content": "Please refer to APPENDIX C.3 for more experimental details.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 358, + 506, + 469 + ] + }, + { + "type": "table", + "bbox": [ + 309, + 500, + 506, + 628 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 495, + 306, + 550 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 494, + 307, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 307, + 507 + ], + "score": 1.0, + "content": "Table 3: Comparison of proposal generation per-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 506, + 307, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 307, + 518 + ], + "score": 1.0, + "content": "formance in terms of recall at 1000 proposals.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 517, + 307, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 153, + 528 + ], + "score": 1.0, + "content": "We choose", + "type": "text" + }, + { + "bbox": [ + 153, + 517, + 191, + 528 + ], + "score": 0.35, + "content": "\\mathrm { \\ t I o U { = } 0 } . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 517, + 307, + 528 + ], + "score": 1.0, + "content": "and 0.8 to compare for con-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 308, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 308, + 540 + ], + "score": 1.0, + "content": "sistency with Buch et al. (2017). Our MA-SST", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 539, + 250, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 208, + 550 + ], + "score": 1.0, + "content": "achieves highest recall at", + "type": "text" + }, + { + "bbox": [ + 209, + 540, + 246, + 550 + ], + "score": 0.52, + "content": "\\mathrm { \\ t I o U { = } } 0 . 8", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 539, + 250, + 550 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "table_caption", + "bbox": [ + 306, + 476, + 506, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 507, + 488 + ], + "score": 1.0, + "content": "Table 4: Comparison for performance and cost of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 487, + 477, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 477, + 498 + ], + "score": 1.0, + "content": "different setups on Penn Treebank dataset.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "table_body", + "bbox": [ + 309, + 500, + 506, + 628 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 309, + 500, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 309, + 500, + 506, + 628 + ], + "score": 0.978, + "html": "
ModelARMINLSTM
Setup1212
Hidden sizenparam(M)nmemTtruncbatch sizeMemory(GB)Speed(chars/s)BPC5005501k1k
4.0254.814.794.791
101
5050100150
384300128128
3.493.562.363.27
98k75k71k70k
1.2381.2261.271.24
", + "type": "table", + "image_path": "9dee7e00530b4ee45916348ee8b30316fa8b7e2b2b83fe8150f9f7711f240679.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 309, + 500, + 506, + 514.2222222222222 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 309, + 514.2222222222222, + 506, + 528.4444444444443 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 309, + 528.4444444444443, + 506, + 542.6666666666665 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 309, + 542.6666666666665, + 506, + 556.8888888888887 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 309, + 556.8888888888887, + 506, + 571.1111111111109 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 309, + 571.1111111111109, + 506, + 585.333333333333 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 309, + 585.333333333333, + 506, + 599.5555555555552 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 309, + 599.5555555555552, + 506, + 613.7777777777774 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 309, + 613.7777777777774, + 506, + 627.9999999999995 + ], + "spans": [], + "index": 51 + } + ] + } + ], + "index": 37.5 + }, + { + "type": "table", + "bbox": [ + 108, + 556, + 304, + 608 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 556, + 304, + 608 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 556, + 304, + 608 + ], + "spans": [ + { + "bbox": [ + 108, + 556, + 304, + 608 + ], + "score": 0.95, + "html": "
ModeltIoU=0.6tIoU=0.8
SST (Paper)0.9200.672
SST (We implement)0.8970.732
MA-SST(Ours)0.9110.759
", + "type": "table", + "image_path": "850273973a2f428712c3ae412e72577c47e4474b112a60383bcfb3009629e5e7.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 108, + 556, + 304, + 569.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 108, + 569.0, + 304, + 582.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 108, + 582.0, + 304, + 595.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 108, + 595.0, + 304, + 608.0 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 649 + ], + "score": 1.0, + "content": "The average recall under different proposal numbers and tIoUs3 are depicted in figure 2. MA-SST", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "outperforms the original SST and our SST simultaneously. To the best of our knowledge, MA-SST", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 306, + 672 + ], + "score": 1.0, + "content": "achieves highest performance of recall at around", + "type": "text" + }, + { + "bbox": [ + 307, + 660, + 345, + 671 + ], + "score": 0.47, + "content": "\\mathrm { \\ t I o U { = } } 0 . 8", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "(see table 3), even if other competitive", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "non-RNN networks (Escorcia et al., 2016; Caba Heilbron et al., 2016; Gao et al., 2017; Guo et al.,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "2018) are taken into account. The MA-SST tends to precisely generate proposals that has a high", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "overlapping area with the ground truth action intervals, leading to high recall performance at high", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "tIoUs. We believe the MA-SST has better performance than the SST mainly because of the direct", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 330, + 458, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 458, + 344 + ], + "score": 1.0, + "content": "access to historical states from which it can better locate the starting frame of an action.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 54.5, + "bbox_fs": [ + 105, + 637, + 506, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 78, + 496, + 222 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 78, + 496, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 78, + 496, + 222 + ], + "spans": [ + { + "bbox": [ + 112, + 78, + 496, + 222 + ], + "score": 0.968, + "type": "image", + "image_path": "80e05beebf3cb64569df7a0d731d988352907332c1e2de7be2c1637e50fd4897.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 78, + 496, + 126.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 126.0, + 496, + 174.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 174.0, + 496, + 222.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 230, + 505, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "Figure 2: Comparison of proposal generation performance. For data generation, we use the code", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 240, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 104, + 240, + 506, + 256 + ], + "score": 1.0, + "content": "offered in Buch et al. (2017) for consistency. (Left) The average number of proposals v.s. average", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 145, + 265 + ], + "score": 1.0, + "content": "recall for", + "type": "text" + }, + { + "bbox": [ + 145, + 253, + 190, + 263 + ], + "score": 0.87, + "content": "\\mathrm { t I o U } \\geq 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 252, + 505, + 265 + ], + "score": 1.0, + "content": ". The MA-SST outperforms our SST when average number of proposals is less", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "than 500, and outperforms the original SST when average number of proposals is more than 500.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "(Right) Different tIoU v.s. recall of 1000 average proposals. The MA-SST strongly outperforms the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 285, + 383, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 357, + 297 + ], + "score": 1.0, + "content": "original SST and outperform our SST by a maximal margin of", + "type": "text" + }, + { + "bbox": [ + 357, + 285, + 379, + 296 + ], + "score": 0.87, + "content": "2 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 285, + 383, + 297 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 105, + 319, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "tIoUs. We believe the MA-SST has better performance than the SST mainly because of the direct", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 330, + 458, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 458, + 344 + ], + "score": 1.0, + "content": "access to historical states from which it can better locate the starting frame of an action.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 106, + 361, + 464, + 372 + ], + "lines": [ + { + "bbox": [ + 104, + 359, + 467, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 359, + 467, + 375 + ], + "score": 1.0, + "content": "5 TOWARDS MORE LIGHT-WEIGHT RECURRENT MEMORY NETWORKS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "It’s known that the speed and computer memory consumption have important influences on the prac-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "ticability of an RNN, especially when external memory is involved. We have shown in some of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "score": 1.0, + "content": "algorithmic tasks that the ARMIN can be trained 3 4 times faster than the NTM in terms of wall clock", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 418, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 432 + ], + "score": 1.0, + "content": "time. Next we conduct a more thorough comparison experiment among the vanilla LSTM and mem-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "ory networks under different hidden sizes. The experiment is based on the character-level language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "modelling task on Penn Treebank dataset, without the loss of generality. For a fair comparison, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 450, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 466 + ], + "score": 1.0, + "content": "implement all networks with near-optimal implementations using Pytorch, an efficient deep-learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "framework. We use the aforementioned experiment setup, and keep all memory matrices the same", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 473, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 144, + 487 + ], + "score": 1.0, + "content": "size, i.e.", + "type": "text" + }, + { + "bbox": [ + 144, + 474, + 178, + 484 + ], + "score": 0.91, + "content": "2 0 \\times d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 473, + 506, + 487 + ], + "score": 1.0, + "content": ". We run the experiment using single-precision floating point calculations under", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "a Titan XP GPU that has 12GB memory space. The results are depicted in figure 3. From the re-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "sults we observe 2 important phenomena: a) The ARMIN consistently outperform other memory", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "networks shown in the graph in terms of running speed both at training and inference stages, and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 517, + 504, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 504, + 530 + ], + "score": 1.0, + "content": "the main contribution to this comes from the simple auto-addressing mechanism of the ARMIN.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "Moreover, at inference stage, we can replace the memory matrix with a list of discrete memory", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "slots, and update memory by simply replacing the old hidden states with the new ones, furthermore,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "we can replace the gumbel-softmax function with argmax. Using these methods, ARMIN’s infer-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "score": 1.0, + "content": "ence speed obtains significant improvement than in training stage. b) The ARMIN outperforms", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 570, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 196, + 586 + ], + "score": 1.0, + "content": "TARDIS and ARMIN", + "type": "text" + }, + { + "bbox": [ + 196, + 573, + 203, + 582 + ], + "score": 0.42, + "content": "\\dotplus", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 570, + 505, + 586 + ], + "score": 1.0, + "content": "TARDIS-addr in terms of training memory consumption, mainly because", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 347, + 597 + ], + "score": 1.0, + "content": "the TARDIS addressing mechanism requires concatenating", + "type": "text" + }, + { + "bbox": [ + 347, + 585, + 359, + 594 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 583, + 378, + 597 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 583, + 400, + 595 + ], + "score": 0.92, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "to every memory cell for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "parallelization. At inference stage, we observe TARDIS have smaller memory consumption than", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 193, + 618 + ], + "score": 1.0, + "content": "ARMIN and ARMIN", + "type": "text" + }, + { + "bbox": [ + 193, + 606, + 200, + 614 + ], + "score": 0.45, + "content": "+ { \\bf \\Phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "TARDIS-addr, mainly because TARDIS has smaller parameter counts under", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "the same hidden size. However, we would like to point out that under the same parameter counts", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "the ARMIN usually outperforms TARDIS as we show in algorithm and language modelling tasks.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "We also observe the NTM has the largest memory consumption at inference time, mainly because", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 649, + 250, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 250, + 661 + ], + "score": 1.0, + "content": "its complex addressing mechanism.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "We have shown in the ablation study (see APPENDIX E) that our network is robust to small iteration", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "length when trained with TBPTT. We realize that we can use this feature to enable large batch size,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "and boost ARMIN’s training speed while keeping better performance than LSTM. To validate this", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "benefit, we conduct an experiment to compare the performance and cost of ARMIN against our", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "best performing LSTM. We also provide the test result of the same LSTM under TBPTT length", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 165, + 732 + ], + "score": 0.91, + "content": "T _ { t r u n c } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "for better comparing. The hyperparameter setups and results are shown in Table 4.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 78, + 496, + 222 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 78, + 496, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 78, + 496, + 222 + ], + "spans": [ + { + "bbox": [ + 112, + 78, + 496, + 222 + ], + "score": 0.968, + "type": "image", + "image_path": "80e05beebf3cb64569df7a0d731d988352907332c1e2de7be2c1637e50fd4897.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 78, + 496, + 126.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 126.0, + 496, + 174.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 174.0, + 496, + 222.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 230, + 505, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "Figure 2: Comparison of proposal generation performance. For data generation, we use the code", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 240, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 104, + 240, + 506, + 256 + ], + "score": 1.0, + "content": "offered in Buch et al. (2017) for consistency. (Left) The average number of proposals v.s. average", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 145, + 265 + ], + "score": 1.0, + "content": "recall for", + "type": "text" + }, + { + "bbox": [ + 145, + 253, + 190, + 263 + ], + "score": 0.87, + "content": "\\mathrm { t I o U } \\geq 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 252, + 505, + 265 + ], + "score": 1.0, + "content": ". The MA-SST outperforms our SST when average number of proposals is less", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "than 500, and outperforms the original SST when average number of proposals is more than 500.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "(Right) Different tIoU v.s. recall of 1000 average proposals. The MA-SST strongly outperforms the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 285, + 383, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 357, + 297 + ], + "score": 1.0, + "content": "original SST and outperform our SST by a maximal margin of", + "type": "text" + }, + { + "bbox": [ + 357, + 285, + 379, + 296 + ], + "score": 0.87, + "content": "2 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 285, + 383, + 297 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 105, + 319, + 504, + 342 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 105, + 319, + 505, + 344 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 106, + 361, + 464, + 372 + ], + "lines": [ + { + "bbox": [ + 104, + 359, + 467, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 359, + 467, + 375 + ], + "score": 1.0, + "content": "5 TOWARDS MORE LIGHT-WEIGHT RECURRENT MEMORY NETWORKS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "It’s known that the speed and computer memory consumption have important influences on the prac-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "ticability of an RNN, especially when external memory is involved. We have shown in some of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "score": 1.0, + "content": "algorithmic tasks that the ARMIN can be trained 3 4 times faster than the NTM in terms of wall clock", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 418, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 432 + ], + "score": 1.0, + "content": "time. Next we conduct a more thorough comparison experiment among the vanilla LSTM and mem-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "ory networks under different hidden sizes. The experiment is based on the character-level language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "modelling task on Penn Treebank dataset, without the loss of generality. For a fair comparison, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 450, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 466 + ], + "score": 1.0, + "content": "implement all networks with near-optimal implementations using Pytorch, an efficient deep-learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "framework. We use the aforementioned experiment setup, and keep all memory matrices the same", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 473, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 144, + 487 + ], + "score": 1.0, + "content": "size, i.e.", + "type": "text" + }, + { + "bbox": [ + 144, + 474, + 178, + 484 + ], + "score": 0.91, + "content": "2 0 \\times d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 473, + 506, + 487 + ], + "score": 1.0, + "content": ". We run the experiment using single-precision floating point calculations under", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "a Titan XP GPU that has 12GB memory space. The results are depicted in figure 3. From the re-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "sults we observe 2 important phenomena: a) The ARMIN consistently outperform other memory", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "networks shown in the graph in terms of running speed both at training and inference stages, and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 517, + 504, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 504, + 530 + ], + "score": 1.0, + "content": "the main contribution to this comes from the simple auto-addressing mechanism of the ARMIN.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "Moreover, at inference stage, we can replace the memory matrix with a list of discrete memory", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "slots, and update memory by simply replacing the old hidden states with the new ones, furthermore,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "we can replace the gumbel-softmax function with argmax. Using these methods, ARMIN’s infer-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "score": 1.0, + "content": "ence speed obtains significant improvement than in training stage. b) The ARMIN outperforms", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 570, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 196, + 586 + ], + "score": 1.0, + "content": "TARDIS and ARMIN", + "type": "text" + }, + { + "bbox": [ + 196, + 573, + 203, + 582 + ], + "score": 0.42, + "content": "\\dotplus", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 570, + 505, + 586 + ], + "score": 1.0, + "content": "TARDIS-addr in terms of training memory consumption, mainly because", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 347, + 597 + ], + "score": 1.0, + "content": "the TARDIS addressing mechanism requires concatenating", + "type": "text" + }, + { + "bbox": [ + 347, + 585, + 359, + 594 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 583, + 378, + 597 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 583, + 400, + 595 + ], + "score": 0.92, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "to every memory cell for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "parallelization. At inference stage, we observe TARDIS have smaller memory consumption than", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 193, + 618 + ], + "score": 1.0, + "content": "ARMIN and ARMIN", + "type": "text" + }, + { + "bbox": [ + 193, + 606, + 200, + 614 + ], + "score": 0.45, + "content": "+ { \\bf \\Phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "TARDIS-addr, mainly because TARDIS has smaller parameter counts under", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "the same hidden size. However, we would like to point out that under the same parameter counts", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "the ARMIN usually outperforms TARDIS as we show in algorithm and language modelling tasks.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "We also observe the NTM has the largest memory consumption at inference time, mainly because", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 649, + 250, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 250, + 661 + ], + "score": 1.0, + "content": "its complex addressing mechanism.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 386, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "We have shown in the ablation study (see APPENDIX E) that our network is robust to small iteration", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "length when trained with TBPTT. We realize that we can use this feature to enable large batch size,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "and boost ARMIN’s training speed while keeping better performance than LSTM. To validate this", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "benefit, we conduct an experiment to compare the performance and cost of ARMIN against our", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "best performing LSTM. We also provide the test result of the same LSTM under TBPTT length", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 165, + 732 + ], + "score": 0.91, + "content": "T _ { t r u n c } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "for better comparing. The hyperparameter setups and results are shown in Table 4.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 664, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 422, + 96 + ], + "score": 1.0, + "content": "By comparing ARMIN setup 1 with LSTM setup 2, we observe that with only", + "type": "text" + }, + { + "bbox": [ + 423, + 82, + 445, + 93 + ], + "score": 0.86, + "content": "6 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "more memory", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 205, + 107 + ], + "score": 1.0, + "content": "consumption, we obtain", + "type": "text" + }, + { + "bbox": [ + 205, + 93, + 226, + 104 + ], + "score": 0.89, + "content": "4 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "training speed gain while keeping a slightly better BPC performance", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 123, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 104, + 143, + 115 + ], + "score": 0.89, + "content": "1 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "less parameter count; by comparing ARMIN setup 2 with LSTM setup 2, we observe that", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 147, + 129 + ], + "score": 1.0, + "content": "with only", + "type": "text" + }, + { + "bbox": [ + 148, + 115, + 170, + 126 + ], + "score": 0.87, + "content": "8 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 113, + 506, + 129 + ], + "score": 1.0, + "content": "more memory consumption, we obtain about 1.5 points of BPC performance gain", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 350, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 123, + 140 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 126, + 146, + 137 + ], + "score": 0.87, + "content": "7 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 126, + 350, + 140 + ], + "score": 1.0, + "content": "training speed gain under similar parameter count.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 113, + 156, + 498, + 293 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 156, + 498, + 293 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 156, + 498, + 293 + ], + "spans": [ + { + "bbox": [ + 113, + 156, + 498, + 293 + ], + "score": 0.964, + "type": "image", + "image_path": "fb603389200962c80fad3cc21f0db454db3b03532b7db183bfb4f4b91ccf1b51.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 113, + 156, + 498, + 201.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 201.66666666666666, + 498, + 247.33333333333331 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 113, + 247.33333333333331, + 498, + 293.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 297, + 505, + 352 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "Figure 3: The running speed and memory consumption at the training and inference stages", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 108, + 308, + 163, + 319 + ], + "score": 0.87, + "content": "( n _ { m e m } = 2 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "). The solid blocks shows the memory consumption and the curves denote the run-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "ning speed in characters/s. (a) shows the training stage, and (b) shows the inference stage (note", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 328, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 158, + 343 + ], + "score": 1.0, + "content": "that ARMIN", + "type": "text" + }, + { + "bbox": [ + 158, + 331, + 164, + 339 + ], + "score": 0.38, + "content": "+", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 328, + 506, + 343 + ], + "score": 1.0, + "content": "TARDIS-addr has basically the same memory consumption with ARMIN, so it’s not", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 341, + 191, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 191, + 354 + ], + "score": 1.0, + "content": "shown in the graph).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + } + ], + "index": 8.0 + }, + { + "type": "title", + "bbox": [ + 107, + 372, + 195, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 197, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 197, + 388 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 398, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 504, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 504, + 410 + ], + "score": 1.0, + "content": "In this paper, we have introduced the ARMIN, a light-weight and memory-augmented RNN ar-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 408, + 504, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 504, + 421 + ], + "score": 1.0, + "content": "chitecture with a novel ARMIN cell. The ARMIN incorporates an efficient external memory with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 419, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 434 + ], + "score": 1.0, + "content": "the light-weight auto-addressing mechanism. We demonstrate competitive performance of ARMIN", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 429, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 104, + 429, + 505, + 444 + ], + "score": 1.0, + "content": "in various tasks, and shows the generality of our model. Our ablation study suggests the efficacy", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "of our external memory and addressing mechanism, and notably, our network is robust to short-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "length TBPTT which enables using large batch size to speed up the training and further increase", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "score": 1.0, + "content": "performance on sequence modelling tasks. Further research may lead to the efficient hierarchical", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "and multi-scale structures of the ARMIN and its successful applications in Seq2Seq models and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 223, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 223, + 498 + ], + "score": 1.0, + "content": "complicated reasoning tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 514, + 175, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 176, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 176, + 527 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 504, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 546 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 544, + 220, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 220, + 556 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 108, + 563, + 504, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Single-stream temporal action proposals. In 2017 IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 585, + 351, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 351, + 597 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), pp. 6373–6382. IEEE, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Fast temporal activity propos-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "als for efficient detection of human actions in untrimmed videos. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 626, + 430, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 430, + 639 + ], + "score": 1.0, + "content": "conference on computer vision and pattern recognition, pp. 1914–1923, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 111, + 645, + 503, + 691 + ], + "lines": [ + { + "bbox": [ + 108, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 108, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "Victor Campos Camunez, Brendan Jou, Xavier Giro Nieto, Jordi Torres Vi ´ nals, and Shih-Fu Chang. ˜", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 658, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 116, + 658, + 506, + 669 + ], + "score": 1.0, + "content": "Skip rnn: learning to skip state updates in recurrent neural networks. In Sixth International", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "score": 1.0, + "content": "Conference on Learning Representations: Monday April 30-Thursday May 03, 2018, Vancouver", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 680, + 366, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 366, + 691 + ], + "score": 1.0, + "content": "Convention Center, Vancouver:[proceedings], pp. 1–17, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 720, + 412, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 412, + 732 + ], + "score": 1.0, + "content": "for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 422, + 96 + ], + "score": 1.0, + "content": "By comparing ARMIN setup 1 with LSTM setup 2, we observe that with only", + "type": "text" + }, + { + "bbox": [ + 423, + 82, + 445, + 93 + ], + "score": 0.86, + "content": "6 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "more memory", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 205, + 107 + ], + "score": 1.0, + "content": "consumption, we obtain", + "type": "text" + }, + { + "bbox": [ + 205, + 93, + 226, + 104 + ], + "score": 0.89, + "content": "4 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "training speed gain while keeping a slightly better BPC performance", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 123, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 104, + 143, + 115 + ], + "score": 0.89, + "content": "1 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "less parameter count; by comparing ARMIN setup 2 with LSTM setup 2, we observe that", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 113, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 147, + 129 + ], + "score": 1.0, + "content": "with only", + "type": "text" + }, + { + "bbox": [ + 148, + 115, + 170, + 126 + ], + "score": 0.87, + "content": "8 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 113, + 506, + 129 + ], + "score": 1.0, + "content": "more memory consumption, we obtain about 1.5 points of BPC performance gain", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 350, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 123, + 140 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 126, + 146, + 137 + ], + "score": 0.87, + "content": "7 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 126, + 350, + 140 + ], + "score": 1.0, + "content": "training speed gain under similar parameter count.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 140 + ] + }, + { + "type": "image", + "bbox": [ + 113, + 156, + 498, + 293 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 156, + 498, + 293 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 156, + 498, + 293 + ], + "spans": [ + { + "bbox": [ + 113, + 156, + 498, + 293 + ], + "score": 0.964, + "type": "image", + "image_path": "fb603389200962c80fad3cc21f0db454db3b03532b7db183bfb4f4b91ccf1b51.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 113, + 156, + 498, + 201.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 201.66666666666666, + 498, + 247.33333333333331 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 113, + 247.33333333333331, + 498, + 293.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 297, + 505, + 352 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "Figure 3: The running speed and memory consumption at the training and inference stages", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 108, + 308, + 163, + 319 + ], + "score": 0.87, + "content": "( n _ { m e m } = 2 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "). The solid blocks shows the memory consumption and the curves denote the run-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "ning speed in characters/s. (a) shows the training stage, and (b) shows the inference stage (note", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 328, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 158, + 343 + ], + "score": 1.0, + "content": "that ARMIN", + "type": "text" + }, + { + "bbox": [ + 158, + 331, + 164, + 339 + ], + "score": 0.38, + "content": "+", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 328, + 506, + 343 + ], + "score": 1.0, + "content": "TARDIS-addr has basically the same memory consumption with ARMIN, so it’s not", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 341, + 191, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 191, + 354 + ], + "score": 1.0, + "content": "shown in the graph).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + } + ], + "index": 8.0 + }, + { + "type": "title", + "bbox": [ + 107, + 372, + 195, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 197, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 197, + 388 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 398, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 504, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 504, + 410 + ], + "score": 1.0, + "content": "In this paper, we have introduced the ARMIN, a light-weight and memory-augmented RNN ar-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 408, + 504, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 504, + 421 + ], + "score": 1.0, + "content": "chitecture with a novel ARMIN cell. The ARMIN incorporates an efficient external memory with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 419, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 506, + 434 + ], + "score": 1.0, + "content": "the light-weight auto-addressing mechanism. We demonstrate competitive performance of ARMIN", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 429, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 104, + 429, + 505, + 444 + ], + "score": 1.0, + "content": "in various tasks, and shows the generality of our model. Our ablation study suggests the efficacy", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "of our external memory and addressing mechanism, and notably, our network is robust to short-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "length TBPTT which enables using large batch size to speed up the training and further increase", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "score": 1.0, + "content": "performance on sequence modelling tasks. Further research may lead to the efficient hierarchical", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "and multi-scale structures of the ARMIN and its successful applications in Seq2Seq models and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 223, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 223, + 498 + ], + "score": 1.0, + "content": "complicated reasoning tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 398, + 506, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 514, + 175, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 176, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 176, + 527 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 504, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 546 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 544, + 220, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 220, + 556 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 531, + 505, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 563, + 504, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Single-stream temporal action proposals. In 2017 IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 585, + 351, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 351, + 597 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), pp. 6373–6382. IEEE, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 563, + 505, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Fast temporal activity propos-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "als for efficient detection of human actions in untrimmed videos. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 626, + 430, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 430, + 639 + ], + "score": 1.0, + "content": "conference on computer vision and pattern recognition, pp. 1914–1923, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 604, + 506, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 645, + 503, + 691 + ], + "lines": [ + { + "bbox": [ + 108, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 108, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "Victor Campos Camunez, Brendan Jou, Xavier Giro Nieto, Jordi Torres Vi ´ nals, and Shih-Fu Chang. ˜", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 658, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 116, + 658, + 506, + 669 + ], + "score": 1.0, + "content": "Skip rnn: learning to skip state updates in recurrent neural networks. In Sixth International", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 681 + ], + "score": 1.0, + "content": "Conference on Learning Representations: Monday April 30-Thursday May 03, 2018, Vancouver", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 680, + 366, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 366, + 691 + ], + "score": 1.0, + "content": "Convention Center, Vancouver:[proceedings], pp. 1–17, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 108, + 645, + 506, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 720, + 412, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 412, + 732 + ], + "score": 1.0, + "content": "for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Gated feedback recurrent", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 481, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 481, + 106 + ], + "score": 1.0, + "content": "neural networks. In International Conference on Machine Learning, pp. 2067–2075, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 113, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. Hierarchical multiscale recurrent neural net-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 125, + 308, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 125, + 308, + 136 + ], + "score": 1.0, + "content": "works. arXiv preprint arXiv:1609.01704, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 144, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 158 + ], + "score": 1.0, + "content": "Ivo Danihelka, Greg Wayne, Benigno Uria, Nal Kalchbrenner, and Alex Graves. Associative long", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "short-term memory. In Proceedings of the 33rd International Conference on International Con-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 167, + 423, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 423, + 179 + ], + "score": 1.0, + "content": "ference on Machine Learning-Volume 48, pp. 1986–1994. JMLR. org, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A Smith. Transition-based", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "dependency parsing with stack long short-term memory. In Proceedings of the 53rd Annual Meet-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "ing of the Association for Computational Linguistics and the 7th International Joint Conference", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 220, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 220, + 478, + 232 + ], + "score": 1.0, + "content": "on Natural Language Processing (Volume 1: Long Papers), volume 1, pp. 334–343, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 446, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 446, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 446, + 254 + ], + "score": 1.0, + "content": "Jeffrey L Elman. Finding structure in time. Cognitive science, 14(2):179–211, 1990.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 274 + ], + "score": 1.0, + "content": "Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 270, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 114, + 270, + 505, + 286 + ], + "score": 1.0, + "content": "action proposals for action understanding. In European Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 282, + 221, + 294 + ], + "spans": [ + { + "bbox": [ + 116, + 282, + 221, + 294 + ], + "score": 1.0, + "content": "768–784. Springer, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 316 + ], + "score": 1.0, + "content": "Jiyang Gao, Zhenheng Yang, Chen Sun, Kan Chen, and Ram Nevatia. Turn tap: Temporal unit", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 115, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "regression network for temporal action proposals. In Computer Vision (ICCV), 2017 IEEE Inter-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 325, + 331, + 337 + ], + "spans": [ + { + "bbox": [ + 116, + 325, + 331, + 337 + ], + "score": 1.0, + "content": "national Conference on, pp. 3648–3656. IEEE, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 344, + 504, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 356, + 353, + 368 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 353, + 368 + ], + "score": 1.0, + "content": "continuous cache. arXiv preprint arXiv:1612.04426, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 389 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 386, + 214, + 397 + ], + "spans": [ + { + "bbox": [ + 116, + 386, + 214, + 397 + ], + "score": 1.0, + "content": "arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 115, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 439, + 191, + 452 + ], + "spans": [ + { + "bbox": [ + 115, + 439, + 191, + 452 + ], + "score": 1.0, + "content": "(7626):471, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "score": 1.0, + "content": "Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 469, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 506, + 485 + ], + "score": 1.0, + "content": "transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 482, + 191, + 493 + ], + "spans": [ + { + "bbox": [ + 117, + 482, + 191, + 493 + ], + "score": 1.0, + "content": "1828–1836, 2015.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Sarath Chandar, and Yoshua Bengio. Memory augmented neural networks with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 513, + 376, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 376, + 524 + ], + "score": 1.0, + "content": "wormhole connections. arXiv preprint arXiv:1701.08718, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Emil Julius Gumbel. Statistical theory of extreme values and some practical applications: a series", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 544, + 333, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 544, + 333, + 556 + ], + "score": 1.0, + "content": "of lectures. Number 33. US Govt. Print. Office, 1954.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 504, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "Dashan Guo, Wei Li, and Xiangzhong Fang. Fully convolutional network for multiscale temporal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 576, + 353, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 576, + 353, + 587 + ], + "score": 1.0, + "content": "action proposals. IEEE Transactions on Multimedia, 2018.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 499, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 501, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 501, + 610 + ], + "score": 1.0, + "content": "David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 615, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 627, + 191, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 191, + 639 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 647, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 658, + 254, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 254, + 670 + ], + "score": 1.0, + "content": "preprint arXiv:1611.01144, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 504, + 690 + ], + "score": 1.0, + "content": "YG Jiang, Jingen Liu, A Roshan Zamir, G Toderici, I Laptev, Mubarak Shah, and Rahul Sukthankar.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 689, + 424, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 689, + 424, + 702 + ], + "score": 1.0, + "content": "Thumos challenge: Action recognition with a large number of classes, 2014.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 439, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 439, + 732 + ], + "score": 1.0, + "content": "nets. In Advances in neural information processing systems, pp. 190–198, 2015.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Junyoung Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Gated feedback recurrent", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 481, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 481, + 106 + ], + "score": 1.0, + "content": "neural networks. In International Conference on Machine Learning, pp. 2067–2075, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 113, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. Hierarchical multiscale recurrent neural net-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 125, + 308, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 125, + 308, + 136 + ], + "score": 1.0, + "content": "works. arXiv preprint arXiv:1609.01704, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 114, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 144, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 158 + ], + "score": 1.0, + "content": "Ivo Danihelka, Greg Wayne, Benigno Uria, Nal Kalchbrenner, and Alex Graves. Associative long", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "short-term memory. In Proceedings of the 33rd International Conference on International Con-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 167, + 423, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 423, + 179 + ], + "score": 1.0, + "content": "ference on Machine Learning-Volume 48, pp. 1986–1994. JMLR. org, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 143, + 506, + 179 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A Smith. Transition-based", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 199, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 199, + 505, + 210 + ], + "score": 1.0, + "content": "dependency parsing with stack long short-term memory. In Proceedings of the 53rd Annual Meet-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "ing of the Association for Computational Linguistics and the 7th International Joint Conference", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 220, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 220, + 478, + 232 + ], + "score": 1.0, + "content": "on Natural Language Processing (Volume 1: Long Papers), volume 1, pp. 334–343, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 186, + 506, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 446, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 446, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 446, + 254 + ], + "score": 1.0, + "content": "Jeffrey L Elman. Finding structure in time. Cognitive science, 14(2):179–211, 1990.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 239, + 446, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 274 + ], + "score": 1.0, + "content": "Victor Escorcia, Fabian Caba Heilbron, Juan Carlos Niebles, and Bernard Ghanem. Daps: Deep", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 270, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 114, + 270, + 505, + 286 + ], + "score": 1.0, + "content": "action proposals for action understanding. In European Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 282, + 221, + 294 + ], + "spans": [ + { + "bbox": [ + 116, + 282, + 221, + 294 + ], + "score": 1.0, + "content": "768–784. Springer, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 258, + 506, + 294 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 316 + ], + "score": 1.0, + "content": "Jiyang Gao, Zhenheng Yang, Chen Sun, Kan Chen, and Ram Nevatia. Turn tap: Temporal unit", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 115, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "regression network for temporal action proposals. In Computer Vision (ICCV), 2017 IEEE Inter-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 325, + 331, + 337 + ], + "spans": [ + { + "bbox": [ + 116, + 325, + 331, + 337 + ], + "score": 1.0, + "content": "national Conference on, pp. 3648–3656. IEEE, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 301, + 506, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 344, + 504, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "Edouard Grave, Armand Joulin, and Nicolas Usunier. Improving neural language models with a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 356, + 353, + 368 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 353, + 368 + ], + "score": 1.0, + "content": "continuous cache. arXiv preprint arXiv:1612.04426, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 343, + 506, + 368 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 389 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 386, + 214, + 397 + ], + "spans": [ + { + "bbox": [ + 116, + 386, + 214, + 397 + ], + "score": 1.0, + "content": "arXiv:1410.5401, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 374, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 115, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "Barwinska, Sergio G ´ omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, ´", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 439, + 191, + 452 + ], + "spans": [ + { + "bbox": [ + 115, + 439, + 191, + 452 + ], + "score": 1.0, + "content": "(7626):471, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 406, + 506, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "score": 1.0, + "content": "Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 469, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 506, + 485 + ], + "score": 1.0, + "content": "transduce with unbounded memory. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 482, + 191, + 493 + ], + "spans": [ + { + "bbox": [ + 117, + 482, + 191, + 493 + ], + "score": 1.0, + "content": "1828–1836, 2015.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 459, + 506, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Sarath Chandar, and Yoshua Bengio. Memory augmented neural networks with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 513, + 376, + 524 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 376, + 524 + ], + "score": 1.0, + "content": "wormhole connections. arXiv preprint arXiv:1701.08718, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 501, + 505, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Emil Julius Gumbel. Statistical theory of extreme values and some practical applications: a series", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 544, + 333, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 544, + 333, + 556 + ], + "score": 1.0, + "content": "of lectures. Number 33. US Govt. Print. Office, 1954.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 533, + 505, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 504, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "Dashan Guo, Wei Li, and Xiangzhong Fang. Fully convolutional network for multiscale temporal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 576, + 353, + 587 + ], + "spans": [ + { + "bbox": [ + 116, + 576, + 353, + 587 + ], + "score": 1.0, + "content": "action proposals. IEEE Transactions on Multimedia, 2018.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 564, + 505, + 587 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 499, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 501, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 501, + 610 + ], + "score": 1.0, + "content": "David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 594, + 501, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 615, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 627, + 191, + 639 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 191, + 639 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 615, + 505, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 647, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 658, + 254, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 254, + 670 + ], + "score": 1.0, + "content": "preprint arXiv:1611.01144, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 646, + 505, + 670 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 504, + 690 + ], + "score": 1.0, + "content": "YG Jiang, Jingen Liu, A Roshan Zamir, G Toderici, I Laptev, Mubarak Shah, and Rahul Sukthankar.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 689, + 424, + 702 + ], + "spans": [ + { + "bbox": [ + 116, + 689, + 424, + 702 + ], + "score": 1.0, + "content": "Thumos challenge: Action recognition with a large number of classes, 2014.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 678, + 504, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 439, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 439, + 732 + ], + "score": 1.0, + "content": "nets. In Advances in neural information processing systems, pp. 190–198, 2015.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Nan Rosemary Ke, Anirudh Goyal, Olexa Bilaniuk, Jonathan Binas, Michael C Mozer, Chris Pal,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 504, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 504, + 107 + ], + "score": 1.0, + "content": "and Yoshua Bengio. Sparse attentive backtracking: Temporal creditassignment through remind-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 297, + 117 + ], + "score": 1.0, + "content": "ing. arXiv preprint arXiv:1809.03702, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 504, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 125, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 140 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 138, + 214, + 150 + ], + "spans": [ + { + "bbox": [ + 116, + 138, + 214, + 150 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 160, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "score": 1.0, + "content": "David Krueger, Tegan Maharaj, Janos Kramar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 170, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 170, + 505, + 185 + ], + "score": 1.0, + "content": "Ke, Anirudh Goyal, Yoshua Bengio, Aaron Courville, and Christopher Pal. Zoneout: Regularizing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 337, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 337, + 195 + ], + "score": 1.0, + "content": "rnns by randomly preserving hidden activations. 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 504, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 216 + ], + "score": 1.0, + "content": "Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. arXiv", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 215, + 254, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 215, + 254, + 227 + ], + "score": 1.0, + "content": "preprint arXiv:1511.06392, 2015.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 506, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "Shuai Li, Wanqing Li, Chris Cook, Ce Zhu, and Yanbo Gao. Independently recurrent neural network", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 248, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 506, + 262 + ], + "score": 1.0, + "content": "(indrnn): Building a longer and deeper rnn. In Proceedings of the IEEE Conference on Computer", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 260, + 335, + 272 + ], + "spans": [ + { + "bbox": [ + 117, + 260, + 335, + 272 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, pp. 5457–5466, 2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 504, + 305 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 293, + 440, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 440, + 305 + ], + "score": 1.0, + "content": "relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Toma´s Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and Jan Cer- ˇ", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 326, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 117, + 326, + 504, + 339 + ], + "score": 1.0, + "content": "nocky. Subword language modeling with neural networks. preprint (http://www. fit. vutbr.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 117, + 337, + 269, + 349 + ], + "spans": [ + { + "bbox": [ + 117, + 337, + 269, + 349 + ], + "score": 1.0, + "content": "cz/imikolov/rnnlm/char. pdf), 8, 2012.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 104, + 359, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "Asier Mujika, Florian Meier, and Angelika Steger. Fast-slow recurrent neural networks. In Advances", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 370, + 380, + 383 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 380, + 383 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 5915–5924, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 392, + 503, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 406 + ], + "score": 1.0, + "content": "Tony A Plate. Holographic reduced representation: Distributed representation for cognitive struc-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 403, + 167, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 167, + 415 + ], + "score": 1.0, + "content": "tures. 2003.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 104, + 425, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 505, + 440 + ], + "score": 1.0, + "content": "Jack Rae, Jonathan J Hunt, Ivo Danihelka, Timothy Harley, Andrew W Senior, Gregory Wayne,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 436, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 505, + 450 + ], + "score": 1.0, + "content": "Alex Graves, and Tim Lillicrap. Scaling memory-augmented neural networks with sparse reads", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 448, + 478, + 460 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 478, + 460 + ], + "score": 1.0, + "content": "and writes. In Advances in Neural Information Processing Systems, pp. 3621–3629, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 104, + 470, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 504, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 504, + 483 + ], + "score": 1.0, + "content": "David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 482, + 316, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 482, + 316, + 493 + ], + "score": 1.0, + "content": "propagating errors. nature, 323(6088):533, 1986.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "shot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 537, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 559, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 297, + 571 + ], + "score": 1.0, + "content": "Learning Research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "Joseph Suarez. Language modeling with recurrent highway hypernetworks. In Advances in Neural", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 592, + 340, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 592, + 340, + 605 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 3267–3276, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 504, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 628 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 626, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 505, + 638 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26–", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 636, + 158, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 158, + 649 + ], + "score": 1.0, + "content": "31, 2012.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 659, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "score": 1.0, + "content": "Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spa-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "tiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 681, + 333, + 693 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 333, + 693 + ], + "score": 1.0, + "content": "conference on computer vision, pp. 4489–4497, 2015.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 703, + 504, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 702, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 505, + 717 + ], + "score": 1.0, + "content": "Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 714, + 219, + 726 + ], + "spans": [ + { + "bbox": [ + 116, + 714, + 219, + 726 + ], + "score": 1.0, + "content": "arXiv:1611.01578, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Nan Rosemary Ke, Anirudh Goyal, Olexa Bilaniuk, Jonathan Binas, Michael C Mozer, Chris Pal,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 504, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 504, + 107 + ], + "score": 1.0, + "content": "and Yoshua Bengio. Sparse attentive backtracking: Temporal creditassignment through remind-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 297, + 117 + ], + "score": 1.0, + "content": "ing. arXiv preprint arXiv:1809.03702, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 504, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 125, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 140 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 138, + 214, + 150 + ], + "spans": [ + { + "bbox": [ + 116, + 138, + 214, + 150 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 125, + 505, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 160, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 174 + ], + "score": 1.0, + "content": "David Krueger, Tegan Maharaj, Janos Kramar, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 170, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 170, + 505, + 185 + ], + "score": 1.0, + "content": "Ke, Anirudh Goyal, Yoshua Bengio, Aaron Courville, and Christopher Pal. Zoneout: Regularizing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 337, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 337, + 195 + ], + "score": 1.0, + "content": "rnns by randomly preserving hidden activations. 2016.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 159, + 505, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 504, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 216 + ], + "score": 1.0, + "content": "Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. arXiv", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 215, + 254, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 215, + 254, + 227 + ], + "score": 1.0, + "content": "preprint arXiv:1511.06392, 2015.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 205, + 505, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 506, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "Shuai Li, Wanqing Li, Chris Cook, Ce Zhu, and Yanbo Gao. Independently recurrent neural network", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 248, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 506, + 262 + ], + "score": 1.0, + "content": "(indrnn): Building a longer and deeper rnn. In Proceedings of the IEEE Conference on Computer", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 260, + 335, + 272 + ], + "spans": [ + { + "bbox": [ + 117, + 260, + 335, + 272 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition, pp. 5457–5466, 2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 237, + 506, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 504, + 305 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 293, + 440, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 440, + 305 + ], + "score": 1.0, + "content": "relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 281, + 505, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Toma´s Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and Jan Cer- ˇ", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 326, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 117, + 326, + 504, + 339 + ], + "score": 1.0, + "content": "nocky. Subword language modeling with neural networks. preprint (http://www. fit. vutbr.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 117, + 337, + 269, + 349 + ], + "spans": [ + { + "bbox": [ + 117, + 337, + 269, + 349 + ], + "score": 1.0, + "content": "cz/imikolov/rnnlm/char. pdf), 8, 2012.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 315, + 505, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 359, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "Asier Mujika, Florian Meier, and Angelika Steger. Fast-slow recurrent neural networks. In Advances", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 370, + 380, + 383 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 380, + 383 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 5915–5924, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 359, + 505, + 383 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 392, + 503, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 406 + ], + "score": 1.0, + "content": "Tony A Plate. Holographic reduced representation: Distributed representation for cognitive struc-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 403, + 167, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 167, + 415 + ], + "score": 1.0, + "content": "tures. 2003.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 391, + 505, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 104, + 425, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 505, + 440 + ], + "score": 1.0, + "content": "Jack Rae, Jonathan J Hunt, Ivo Danihelka, Timothy Harley, Andrew W Senior, Gregory Wayne,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 436, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 505, + 450 + ], + "score": 1.0, + "content": "Alex Graves, and Tim Lillicrap. Scaling memory-augmented neural networks with sparse reads", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 448, + 478, + 460 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 478, + 460 + ], + "score": 1.0, + "content": "and writes. In Advances in Neural Information Processing Systems, pp. 3621–3629, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 425, + 505, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 470, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 504, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 504, + 483 + ], + "score": 1.0, + "content": "David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 482, + 316, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 482, + 316, + 493 + ], + "score": 1.0, + "content": "propagating errors. nature, 323(6088):533, 1986.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 470, + 504, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. One-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "shot learning with memory-augmented neural networks. arXiv preprint arXiv:1605.06065, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 503, + 505, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 537, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 559, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 297, + 571 + ], + "score": 1.0, + "content": "Learning Research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 537, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "Joseph Suarez. Language modeling with recurrent highway hypernetworks. In Advances in Neural", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 592, + 340, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 592, + 340, + 605 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 3267–3276, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 581, + 506, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 504, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 628 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 626, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 505, + 638 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2):26–", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 636, + 158, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 158, + 649 + ], + "score": 1.0, + "content": "31, 2012.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 613, + 505, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 659, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "score": 1.0, + "content": "Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spa-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 116, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "tiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 681, + 333, + 693 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 333, + 693 + ], + "score": 1.0, + "content": "conference on computer vision, pp. 4489–4497, 2015.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 658, + 505, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 703, + 504, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 702, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 505, + 717 + ], + "score": 1.0, + "content": "Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 714, + 219, + 726 + ], + "spans": [ + { + "bbox": [ + 116, + 714, + 219, + 726 + ], + "score": 1.0, + "content": "arXiv:1611.01578, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 702, + 505, + 726 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 395, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 396, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 396, + 96 + ], + "score": 1.0, + "content": "APPENDIX A RNN AND LONG SHORT-TERM MEMORY", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 505, + 151 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "A recurrent neural network (RNN) is a class of neural network that recurrently processes a sequence", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 116, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 103, + 116, + 147, + 132 + ], + "score": 1.0, + "content": "of inputs", + "type": "text" + }, + { + "bbox": [ + 147, + 118, + 214, + 130 + ], + "score": 0.92, + "content": "\\{ \\pmb { x } _ { 1 } , \\pmb { x } _ { 2 } , . . . , \\pmb { x } _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 116, + 362, + 132 + ], + "score": 1.0, + "content": ", and returns a sequence of outputs", + "type": "text" + }, + { + "bbox": [ + 362, + 118, + 428, + 130 + ], + "score": 0.92, + "content": "\\left\\{ y _ { 1 } , y _ { 2 } , . . . , y _ { T } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 116, + 458, + 132 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 458, + 117, + 501, + 129 + ], + "score": 0.92, + "content": "\\pmb { x } _ { t } \\in \\mathbb { R } ^ { d _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 116, + 506, + 132 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 107, + 128, + 145, + 140 + ], + "score": 0.91, + "content": "\\pmb { y } _ { t } \\in \\mathbb { R } ^ { d _ { o } }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 128, + 166, + 142 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 166, + 130, + 190, + 140 + ], + "score": 0.84, + "content": "d _ { i } , d _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 128, + 447, + 142 + ], + "score": 1.0, + "content": "is the input size and output size, respectively. In a vanilla RNN,", + "type": "text" + }, + { + "bbox": [ + 447, + 131, + 457, + 140 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { y } } _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 128, + 505, + 142 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 139, + 191, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 191, + 153 + ], + "score": 1.0, + "content": "following equations:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 152, + 368, + 166 + ], + "lines": [ + { + "bbox": [ + 241, + 152, + 368, + 166 + ], + "spans": [ + { + "bbox": [ + 241, + 152, + 368, + 166 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\pmb { h } _ { t } = \\sigma ( \\pmb { U } \\pmb { x } _ { t } + \\pmb { W } \\pmb { h } _ { t - 1 } + \\pmb { b } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "804120d85e7408afb1231cfe5c6429e74bfe78e522a34d4170c11f8c2955407d.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 241, + 152, + 368, + 166 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 171, + 339, + 184 + ], + "lines": [ + { + "bbox": [ + 271, + 171, + 339, + 184 + ], + "spans": [ + { + "bbox": [ + 271, + 171, + 339, + 184 + ], + "score": 0.81, + "content": "\\pmb { y } _ { t } = \\pmb { V } \\pmb { h } _ { t } + \\pmb { c } ,", + "type": "interline_equation", + "image_path": "5a5db4a4e05fdc2a17d697e8d5c04f697686435e05ad87f9bd6a5cdda4e3d158.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 271, + 171, + 339, + 184 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 189, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 204, + 201 + ], + "score": 1.0, + "content": "where the hidden state", + "type": "text" + }, + { + "bbox": [ + 204, + 188, + 249, + 200 + ], + "score": 0.93, + "content": "\\pmb { h } _ { t } ~ \\in ~ \\mathbb { R } ^ { d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 187, + 393, + 201 + ], + "score": 1.0, + "content": "is passed to the next recurrence.", + "type": "text" + }, + { + "bbox": [ + 394, + 190, + 405, + 200 + ], + "score": 0.89, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "is the hidden size, and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 199, + 271, + 213 + ], + "spans": [ + { + "bbox": [ + 107, + 200, + 167, + 212 + ], + "score": 0.91, + "content": "U , W , V , b , c", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 199, + 271, + 213 + ], + "score": 1.0, + "content": "are learnable parameters.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "The LSTM (Hochreiter & Schmidhuber, 1997) is designed to overcome the gradient vanishing and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 227, + 504, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 472, + 241 + ], + "score": 1.0, + "content": "exploding problem in vanilla RNN and learn long-term dependencies, by passing a tuple of", + "type": "text" + }, + { + "bbox": [ + 472, + 228, + 504, + 240 + ], + "score": 0.93, + "content": "\\left( { { h _ { t } } , { c _ { t } } } \\right)", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 383, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 383, + 251 + ], + "score": 1.0, + "content": "to the next recurrence. Its computation process is defined as follows:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 257, + 392, + 312 + ], + "lines": [ + { + "bbox": [ + 219, + 257, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 219, + 257, + 392, + 312 + ], + "score": 0.94, + "content": "\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\end{array} \\right\\} W _ { 0 } [ \\ : x _ { t } , h _ { t - 1 } \\ : ] + b _ { 0 } ,", + "type": "interline_equation", + "image_path": "ad5becc79b9d06847187e14ed27650233d88b8c613e95d1246b529200f1a4b81.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 219, + 257, + 392, + 275.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 219, + 275.3333333333333, + 392, + 293.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 219, + 293.66666666666663, + 392, + 311.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 251, + 318, + 359, + 354 + ], + "lines": [ + { + "bbox": [ + 252, + 318, + 359, + 354 + ], + "spans": [ + { + "bbox": [ + 252, + 318, + 359, + 354 + ], + "score": 0.29, + "content": "\\begin{array} { r } { \\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + \\pmb { i } _ { t } \\circ \\pmb { g } _ { t } , } \\\\ { \\pmb { h } _ { t } = \\pmb { o } _ { t } \\circ \\operatorname { t a n h } ( \\pmb { c } _ { t } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "7e19929b1d6fbfaad66140021ec43d3b6a8e073f83af83a84b1d6717e68449a7.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 251, + 318, + 359, + 336.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 251, + 336.0, + 359, + 354.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 104, + 354, + 507, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 354, + 133, + 372 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 361, + 141, + 369 + ], + "score": 0.49, + "content": "\\circ", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 354, + 260, + 372 + ], + "score": 1.0, + "content": "is the element-wise product,", + "type": "text" + }, + { + "bbox": [ + 261, + 357, + 346, + 370 + ], + "score": 0.91, + "content": "W _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } \\times ( d _ { i } + d _ { h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 354, + 351, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 352, + 358, + 397, + 370 + ], + "score": 0.81, + "content": "b _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 354, + 403, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 403, + 358, + 449, + 370 + ], + "score": 0.69, + "content": "i _ { t } , f _ { t } , g _ { t } o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 354, + 507, + 372 + ], + "score": 1.0, + "content": "are the input", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 293, + 381 + ], + "score": 1.0, + "content": "gate, forget gate, cell state, output gate at time", + "type": "text" + }, + { + "bbox": [ + 293, + 371, + 298, + 380 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 370, + 505, + 381 + ], + "score": 1.0, + "content": ", respectively, and they control the information flow", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 381, + 163, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 163, + 391 + ], + "score": 1.0, + "content": "in the LSTM.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 409, + 429, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 430, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 430, + 425 + ], + "score": 1.0, + "content": "APPENDIX B DIFFERENCES BETWEEN TARDIS AND ARMIN", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "The major differences between ARMIN and TARDIS networks are the recurrent cell computation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 446, + 433, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 433, + 460 + ], + "score": 1.0, + "content": "and read operation. We show these differences respectively in subsections below:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 107, + 473, + 226, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 227, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 227, + 486 + ], + "score": 1.0, + "content": "B.1 CELL COMPUTATIONS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 325, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 326, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 326, + 507 + ], + "score": 1.0, + "content": "The cell computation process of TARDIS is as follow:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 521, + 388, + 562 + ], + "lines": [ + { + "bbox": [ + 223, + 521, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 223, + 521, + 388, + 562 + ], + "score": 0.93, + "content": "\\left\\{ \\begin{array} { l } { i _ { t } } \\\\ { f _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { l } { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W [ \\ : x _ { t } , h _ { t - 1 } , r _ { t } \\ : ] + b \\ : ,", + "type": "interline_equation", + "image_path": "cfc57efefd21d85bfcfac35b79da40758e958d2a8066e4b1422a2ef153775068.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 521, + 388, + 541.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 223, + 541.5, + 388, + 562.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 568, + 417, + 602 + ], + "lines": [ + { + "bbox": [ + 193, + 568, + 417, + 602 + ], + "spans": [ + { + "bbox": [ + 193, + 568, + 417, + 602 + ], + "score": 0.91, + "content": "\\left\\{ \\begin{array} { l } \\alpha _ { t } \\right\\} = \\left\\{ \\begin{array} { l l } { \\mathrm { g u m b e l - s i g m o i d } } \\\\ { \\mathrm { g u m b e l - s i g m o i d } } \\end{array} \\right\\} \\left\\{ \\begin{array} { l } { \\pmb { w } ^ { \\alpha \\top } } \\\\ { \\pmb { w } ^ { \\beta \\top } } \\end{array} \\right\\} \\left[ \\begin{array} { l } { \\pmb { x } _ { t } , \\pmb { h } _ { t - 1 } , \\pmb { r } _ { t } } \\end{array} \\right] , \\end{array}", + "type": "interline_equation", + "image_path": "71101ff4ff32c3594e94661b93bf7d8f5fafd4bb91c5b49209a91828cd7b2ce5.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 568, + 417, + 585.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 193, + 585.0, + 417, + 602.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 607, + 388, + 622 + ], + "lines": [ + { + "bbox": [ + 221, + 607, + 388, + 622 + ], + "spans": [ + { + "bbox": [ + 221, + 607, + 388, + 622 + ], + "score": 0.64, + "content": "g _ { t } = \\operatorname { t a n h } ( W ^ { g } [ x _ { t } , \\alpha _ { t } h _ { t - 1 } , \\beta _ { t } { r } _ { t } ] + b ^ { g } ) ,", + "type": "interline_equation", + "image_path": "b95d0d07a5d5e4b4e3a719b99db47ce964570506dd1749bd579e978cf18903e1.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 221, + 607, + 388, + 622 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 627, + 360, + 641 + ], + "lines": [ + { + "bbox": [ + 249, + 627, + 360, + 641 + ], + "spans": [ + { + "bbox": [ + 249, + 627, + 360, + 641 + ], + "score": 0.56, + "content": "\\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ,", + "type": "interline_equation", + "image_path": "d3e8560b4874992f9a038ec70f36d44c13ab400568e9a9915d7768bafc583dd1.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 249, + 627, + 360, + 641 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 645, + 348, + 659 + ], + "lines": [ + { + "bbox": [ + 262, + 645, + 348, + 659 + ], + "spans": [ + { + "bbox": [ + 262, + 645, + 348, + 659 + ], + "score": 0.33, + "content": "h _ { t } = o _ { t } \\circ \\operatorname { t a n h } ( c _ { t } ) ,", + "type": "interline_equation", + "image_path": "21307b6c091a805149c17bb66defa7ee661ed633dcc7e674f35b9b58ce3a98bc.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 262, + 645, + 348, + 659 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 104, + 666, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 133, + 683 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 669, + 455, + 681 + ], + "score": 0.81, + "content": "h _ { t - 1 } , h _ { t } , c _ { t } , i _ { t } , f _ { t } , o _ { t } , g _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { r } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\pmb x } _ { t } \\in \\mathbb { R } ^ { d _ { i } } , { \\pmb w } ^ { \\alpha } , { \\pmb w } ^ { \\beta } \\in \\mathbb { R } ^ { d _ { i } + d _ { h } + d _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 666, + 479, + 683 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 479, + 669, + 505, + 681 + ], + "score": 0.86, + "content": "W \\in", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 678, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 177, + 694 + ], + "score": 1.0, + "content": "R3dh×(di+dh+dr)", + "type": "text" + }, + { + "bbox": [ + 177, + 681, + 374, + 693 + ], + "score": 0.88, + "content": ", \\pmb { W } ^ { g } \\in \\mathbb { R } ^ { d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } \\in \\mathbb { R } ^ { 3 d _ { h } } , \\pmb { b } ^ { g } \\in \\mathbb { R } ^ { d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 678, + 506, + 694 + ], + "score": 1.0, + "content": ". The output of TARDIS is the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 692, + 219, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 175, + 705 + ], + "score": 1.0, + "content": "concatenation of", + "type": "text" + }, + { + "bbox": [ + 175, + 693, + 186, + 704 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 692, + 204, + 705 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 205, + 694, + 214, + 704 + ], + "score": 0.83, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 692, + 219, + 705 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "There are 3 important points that distinguish the ARMIN cell from TARDIS cell, which are as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 719, + 142, + 733 + ], + "spans": [ + { + "bbox": [ + 104, + 719, + 142, + 733 + ], + "score": 1.0, + "content": "follows:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 395, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 396, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 396, + 96 + ], + "score": 1.0, + "content": "APPENDIX A RNN AND LONG SHORT-TERM MEMORY", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 505, + 151 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "A recurrent neural network (RNN) is a class of neural network that recurrently processes a sequence", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 103, + 116, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 103, + 116, + 147, + 132 + ], + "score": 1.0, + "content": "of inputs", + "type": "text" + }, + { + "bbox": [ + 147, + 118, + 214, + 130 + ], + "score": 0.92, + "content": "\\{ \\pmb { x } _ { 1 } , \\pmb { x } _ { 2 } , . . . , \\pmb { x } _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 116, + 362, + 132 + ], + "score": 1.0, + "content": ", and returns a sequence of outputs", + "type": "text" + }, + { + "bbox": [ + 362, + 118, + 428, + 130 + ], + "score": 0.92, + "content": "\\left\\{ y _ { 1 } , y _ { 2 } , . . . , y _ { T } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 116, + 458, + 132 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 458, + 117, + 501, + 129 + ], + "score": 0.92, + "content": "\\pmb { x } _ { t } \\in \\mathbb { R } ^ { d _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 116, + 506, + 132 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 107, + 128, + 145, + 140 + ], + "score": 0.91, + "content": "\\pmb { y } _ { t } \\in \\mathbb { R } ^ { d _ { o } }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 128, + 166, + 142 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 166, + 130, + 190, + 140 + ], + "score": 0.84, + "content": "d _ { i } , d _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 128, + 447, + 142 + ], + "score": 1.0, + "content": "is the input size and output size, respectively. In a vanilla RNN,", + "type": "text" + }, + { + "bbox": [ + 447, + 131, + 457, + 140 + ], + "score": 0.86, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { y } } _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 128, + 505, + 142 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 139, + 191, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 191, + 153 + ], + "score": 1.0, + "content": "following equations:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 103, + 106, + 506, + 153 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 152, + 368, + 166 + ], + "lines": [ + { + "bbox": [ + 241, + 152, + 368, + 166 + ], + "spans": [ + { + "bbox": [ + 241, + 152, + 368, + 166 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\pmb { h } _ { t } = \\sigma ( \\pmb { U } \\pmb { x } _ { t } + \\pmb { W } \\pmb { h } _ { t - 1 } + \\pmb { b } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "804120d85e7408afb1231cfe5c6429e74bfe78e522a34d4170c11f8c2955407d.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 241, + 152, + 368, + 166 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 171, + 339, + 184 + ], + "lines": [ + { + "bbox": [ + 271, + 171, + 339, + 184 + ], + "spans": [ + { + "bbox": [ + 271, + 171, + 339, + 184 + ], + "score": 0.81, + "content": "\\pmb { y } _ { t } = \\pmb { V } \\pmb { h } _ { t } + \\pmb { c } ,", + "type": "interline_equation", + "image_path": "5a5db4a4e05fdc2a17d697e8d5c04f697686435e05ad87f9bd6a5cdda4e3d158.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 271, + 171, + 339, + 184 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 189, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 204, + 201 + ], + "score": 1.0, + "content": "where the hidden state", + "type": "text" + }, + { + "bbox": [ + 204, + 188, + 249, + 200 + ], + "score": 0.93, + "content": "\\pmb { h } _ { t } ~ \\in ~ \\mathbb { R } ^ { d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 187, + 393, + 201 + ], + "score": 1.0, + "content": "is passed to the next recurrence.", + "type": "text" + }, + { + "bbox": [ + 394, + 190, + 405, + 200 + ], + "score": 0.89, + "content": "d _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "is the hidden size, and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 199, + 271, + 213 + ], + "spans": [ + { + "bbox": [ + 107, + 200, + 167, + 212 + ], + "score": 0.91, + "content": "U , W , V , b , c", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 199, + 271, + 213 + ], + "score": 1.0, + "content": "are learnable parameters.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 187, + 505, + 213 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "The LSTM (Hochreiter & Schmidhuber, 1997) is designed to overcome the gradient vanishing and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 227, + 504, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 472, + 241 + ], + "score": 1.0, + "content": "exploding problem in vanilla RNN and learn long-term dependencies, by passing a tuple of", + "type": "text" + }, + { + "bbox": [ + 472, + 228, + 504, + 240 + ], + "score": 0.93, + "content": "\\left( { { h _ { t } } , { c _ { t } } } \\right)", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 383, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 383, + 251 + ], + "score": 1.0, + "content": "to the next recurrence. Its computation process is defined as follows:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 217, + 505, + 251 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 257, + 392, + 312 + ], + "lines": [ + { + "bbox": [ + 219, + 257, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 219, + 257, + 392, + 312 + ], + "score": 0.94, + "content": "\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\end{array} \\right\\} W _ { 0 } [ \\ : x _ { t } , h _ { t - 1 } \\ : ] + b _ { 0 } ,", + "type": "interline_equation", + "image_path": "ad5becc79b9d06847187e14ed27650233d88b8c613e95d1246b529200f1a4b81.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 219, + 257, + 392, + 275.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 219, + 275.3333333333333, + 392, + 293.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 219, + 293.66666666666663, + 392, + 311.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 251, + 318, + 359, + 354 + ], + "lines": [ + { + "bbox": [ + 252, + 318, + 359, + 354 + ], + "spans": [ + { + "bbox": [ + 252, + 318, + 359, + 354 + ], + "score": 0.29, + "content": "\\begin{array} { r } { \\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + \\pmb { i } _ { t } \\circ \\pmb { g } _ { t } , } \\\\ { \\pmb { h } _ { t } = \\pmb { o } _ { t } \\circ \\operatorname { t a n h } ( \\pmb { c } _ { t } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "7e19929b1d6fbfaad66140021ec43d3b6a8e073f83af83a84b1d6717e68449a7.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 251, + 318, + 359, + 336.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 251, + 336.0, + 359, + 354.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 104, + 354, + 507, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 354, + 133, + 372 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 361, + 141, + 369 + ], + "score": 0.49, + "content": "\\circ", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 354, + 260, + 372 + ], + "score": 1.0, + "content": "is the element-wise product,", + "type": "text" + }, + { + "bbox": [ + 261, + 357, + 346, + 370 + ], + "score": 0.91, + "content": "W _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } \\times ( d _ { i } + d _ { h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 354, + 351, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 352, + 358, + 397, + 370 + ], + "score": 0.81, + "content": "b _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 354, + 403, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 403, + 358, + 449, + 370 + ], + "score": 0.69, + "content": "i _ { t } , f _ { t } , g _ { t } o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 354, + 507, + 372 + ], + "score": 1.0, + "content": "are the input", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 293, + 381 + ], + "score": 1.0, + "content": "gate, forget gate, cell state, output gate at time", + "type": "text" + }, + { + "bbox": [ + 293, + 371, + 298, + 380 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 370, + 505, + 381 + ], + "score": 1.0, + "content": ", respectively, and they control the information flow", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 381, + 163, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 163, + 391 + ], + "score": 1.0, + "content": "in the LSTM.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 354, + 507, + 391 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 409, + 429, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 430, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 430, + 425 + ], + "score": 1.0, + "content": "APPENDIX B DIFFERENCES BETWEEN TARDIS AND ARMIN", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "The major differences between ARMIN and TARDIS networks are the recurrent cell computation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 446, + 433, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 433, + 460 + ], + "score": 1.0, + "content": "and read operation. We show these differences respectively in subsections below:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 435, + 505, + 460 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 473, + 226, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 227, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 227, + 486 + ], + "score": 1.0, + "content": "B.1 CELL COMPUTATIONS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 325, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 326, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 326, + 507 + ], + "score": 1.0, + "content": "The cell computation process of TARDIS is as follow:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 493, + 326, + 507 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 521, + 388, + 562 + ], + "lines": [ + { + "bbox": [ + 223, + 521, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 223, + 521, + 388, + 562 + ], + "score": 0.93, + "content": "\\left\\{ \\begin{array} { l } { i _ { t } } \\\\ { f _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { l } { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W [ \\ : x _ { t } , h _ { t - 1 } , r _ { t } \\ : ] + b \\ : ,", + "type": "interline_equation", + "image_path": "cfc57efefd21d85bfcfac35b79da40758e958d2a8066e4b1422a2ef153775068.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 521, + 388, + 541.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 223, + 541.5, + 388, + 562.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 568, + 417, + 602 + ], + "lines": [ + { + "bbox": [ + 193, + 568, + 417, + 602 + ], + "spans": [ + { + "bbox": [ + 193, + 568, + 417, + 602 + ], + "score": 0.91, + "content": "\\left\\{ \\begin{array} { l } \\alpha _ { t } \\right\\} = \\left\\{ \\begin{array} { l l } { \\mathrm { g u m b e l - s i g m o i d } } \\\\ { \\mathrm { g u m b e l - s i g m o i d } } \\end{array} \\right\\} \\left\\{ \\begin{array} { l } { \\pmb { w } ^ { \\alpha \\top } } \\\\ { \\pmb { w } ^ { \\beta \\top } } \\end{array} \\right\\} \\left[ \\begin{array} { l } { \\pmb { x } _ { t } , \\pmb { h } _ { t - 1 } , \\pmb { r } _ { t } } \\end{array} \\right] , \\end{array}", + "type": "interline_equation", + "image_path": "71101ff4ff32c3594e94661b93bf7d8f5fafd4bb91c5b49209a91828cd7b2ce5.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 568, + 417, + 585.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 193, + 585.0, + 417, + 602.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 607, + 388, + 622 + ], + "lines": [ + { + "bbox": [ + 221, + 607, + 388, + 622 + ], + "spans": [ + { + "bbox": [ + 221, + 607, + 388, + 622 + ], + "score": 0.64, + "content": "g _ { t } = \\operatorname { t a n h } ( W ^ { g } [ x _ { t } , \\alpha _ { t } h _ { t - 1 } , \\beta _ { t } { r } _ { t } ] + b ^ { g } ) ,", + "type": "interline_equation", + "image_path": "b95d0d07a5d5e4b4e3a719b99db47ce964570506dd1749bd579e978cf18903e1.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 221, + 607, + 388, + 622 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 627, + 360, + 641 + ], + "lines": [ + { + "bbox": [ + 249, + 627, + 360, + 641 + ], + "spans": [ + { + "bbox": [ + 249, + 627, + 360, + 641 + ], + "score": 0.56, + "content": "\\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ,", + "type": "interline_equation", + "image_path": "d3e8560b4874992f9a038ec70f36d44c13ab400568e9a9915d7768bafc583dd1.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 249, + 627, + 360, + 641 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 262, + 645, + 348, + 659 + ], + "lines": [ + { + "bbox": [ + 262, + 645, + 348, + 659 + ], + "spans": [ + { + "bbox": [ + 262, + 645, + 348, + 659 + ], + "score": 0.33, + "content": "h _ { t } = o _ { t } \\circ \\operatorname { t a n h } ( c _ { t } ) ,", + "type": "interline_equation", + "image_path": "21307b6c091a805149c17bb66defa7ee661ed633dcc7e674f35b9b58ce3a98bc.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 262, + 645, + 348, + 659 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 104, + 666, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 133, + 683 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 669, + 455, + 681 + ], + "score": 0.81, + "content": "h _ { t - 1 } , h _ { t } , c _ { t } , i _ { t } , f _ { t } , o _ { t } , g _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { r } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\pmb x } _ { t } \\in \\mathbb { R } ^ { d _ { i } } , { \\pmb w } ^ { \\alpha } , { \\pmb w } ^ { \\beta } \\in \\mathbb { R } ^ { d _ { i } + d _ { h } + d _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 666, + 479, + 683 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 479, + 669, + 505, + 681 + ], + "score": 0.86, + "content": "W \\in", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 678, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 177, + 694 + ], + "score": 1.0, + "content": "R3dh×(di+dh+dr)", + "type": "text" + }, + { + "bbox": [ + 177, + 681, + 374, + 693 + ], + "score": 0.88, + "content": ", \\pmb { W } ^ { g } \\in \\mathbb { R } ^ { d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } \\in \\mathbb { R } ^ { 3 d _ { h } } , \\pmb { b } ^ { g } \\in \\mathbb { R } ^ { d _ { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 678, + 506, + 694 + ], + "score": 1.0, + "content": ". The output of TARDIS is the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 692, + 219, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 175, + 705 + ], + "score": 1.0, + "content": "concatenation of", + "type": "text" + }, + { + "bbox": [ + 175, + 693, + 186, + 704 + ], + "score": 0.88, + "content": "\\boldsymbol { h } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 692, + 204, + 705 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 205, + 694, + 214, + 704 + ], + "score": 0.83, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 692, + 219, + 705 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 666, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "There are 3 important points that distinguish the ARMIN cell from TARDIS cell, which are as", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 719, + 142, + 733 + ], + "spans": [ + { + "bbox": [ + 104, + 719, + 142, + 733 + ], + "score": 1.0, + "content": "follows:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 708, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 288, + 95 + ], + "score": 1.0, + "content": "• The TARDIS uses two binary scalar gates", + "type": "text" + }, + { + "bbox": [ + 289, + 83, + 313, + 94 + ], + "score": 0.91, + "content": "\\alpha _ { t } , \\beta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "to control the information flow in equation 17.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "The two scalar gates are activated using gumbel-sigmoid (which is similar to gumbel-softmax and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "the output is close to binary), whereas the ARMIN uses two soft vector gates to control the infor-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "mation flow, which is more flexible and enable larger model capacity. For example, in the char-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "level language modelling experiments, if we replace the sigmoid activations in equation 3 with the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 483, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 483, + 150 + ], + "score": 1.0, + "content": "element-wise gumbel-sigmoid, the BPC performance immediately drops from 1.202 to 1.235.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 406, + 165 + ], + "score": 1.0, + "content": "• The TADIS apply information control only when calculating the cell state", + "type": "text" + }, + { + "bbox": [ + 407, + 154, + 417, + 164 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { g } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 152, + 505, + 165 + ], + "score": 1.0, + "content": ", whereas the ARMIN", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "apply information control immediately after the read operation. Our main motivation to this change", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 317, + 186 + ], + "score": 1.0, + "content": "is that if we don’t filter out irrelevant information in", + "type": "text" + }, + { + "bbox": [ + 318, + 176, + 328, + 185 + ], + "score": 0.84, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 174, + 346, + 186 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 346, + 175, + 358, + 185 + ], + "score": 0.89, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "at first, these noisy information will", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 184, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 506, + 200 + ], + "score": 1.0, + "content": "cause the weight parameters hard to learn. Furthermore, if the read operation chooses the wrong", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 196, + 381, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 301, + 209 + ], + "score": 1.0, + "content": "past state, we can block the information flow for", + "type": "text" + }, + { + "bbox": [ + 302, + 198, + 312, + 207 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 196, + 381, + 209 + ], + "score": 1.0, + "content": "at the first place.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 503, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 248, + 224 + ], + "score": 1.0, + "content": "• The ARMIN has an extra gate", + "type": "text" + }, + { + "bbox": [ + 248, + 212, + 259, + 223 + ], + "score": 0.88, + "content": "\\mathbf { \\delta } _ { o _ { t } ^ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 210, + 375, + 224 + ], + "score": 1.0, + "content": "to control the output from", + "type": "text" + }, + { + "bbox": [ + 375, + 214, + 385, + 222 + ], + "score": 0.86, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 210, + 505, + 224 + ], + "score": 1.0, + "content": "as in equation 8, which the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 369, + 235 + ], + "score": 1.0, + "content": "TARDIS doesn’t have. If we remove this gate and directly output", + "type": "text" + }, + { + "bbox": [ + 369, + 224, + 379, + 234 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "like in TARDIS, we get a BPC", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 233, + 416, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 416, + 246 + ], + "score": 1.0, + "content": "performance drop from 1.202 to 1.219 in char-level language modelling task.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "Finally, despite that the ARMIN has 3 more vector gates than the TARDIS, we have shown in the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 144, + 278 + ], + "score": 0.41, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "and Penn Treebank experiments that the ARMIN can still outperform the TARDIS under", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 277, + 210, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 210, + 290 + ], + "score": 1.0, + "content": "similar parameter counts.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 211, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 213, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 213, + 315 + ], + "score": 1.0, + "content": "B.2 READ OPERATION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 192, + 335 + ], + "score": 1.0, + "content": "The memory matrix", + "type": "text" + }, + { + "bbox": [ + 193, + 324, + 209, + 334 + ], + "score": 0.86, + "content": "M _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 321, + 379, + 335 + ], + "score": 1.0, + "content": "of TARDIS has disjoint address section", + "type": "text" + }, + { + "bbox": [ + 379, + 322, + 452, + 334 + ], + "score": 0.91, + "content": "\\mathbf { \\boldsymbol { A } } _ { t } ~ \\in ~ \\mathbb { R } ^ { n _ { m e m } \\times d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "and content", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 138, + 347 + ], + "score": 1.0, + "content": "section", + "type": "text" + }, + { + "bbox": [ + 138, + 334, + 208, + 345 + ], + "score": 0.86, + "content": "C _ { t } \\ \\stackrel { \\cdot } { \\in } \\ \\mathbb { R } ^ { n _ { m e m } \\times d _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 332, + 213, + 347 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 214, + 334, + 279, + 346 + ], + "score": 0.91, + "content": "M _ { t } = [ A _ { t } ; C _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 332, + 506, + 347 + ], + "score": 1.0, + "content": ". The controller reads both the address and the content", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 434, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 434, + 359 + ], + "score": 1.0, + "content": "parts of the memory, but it will only write into the content section of the memory.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 504, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 376 + ], + "score": 1.0, + "content": "The TARDIS paper explored two ways to sample the read locations——using reinforce and using", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "gumbel-softmax function, respectively. The gumbel-softmax way shows a better performance and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 384, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 294, + 397 + ], + "score": 1.0, + "content": "its read operation can be formulated as follow:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 409, + 401, + 441 + ], + "lines": [ + { + "bbox": [ + 209, + 409, + 401, + 441 + ], + "spans": [ + { + "bbox": [ + 209, + 409, + 401, + 441 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\pi _ { t } [ i ] = a ^ { \\top } \\operatorname { t a n h } ( W ^ { \\gamma } [ h _ { t } , x _ { t } , M _ { t } [ i ] , { \\boldsymbol u } _ { t } ] + b ^ { \\gamma } ) , } } \\\\ { { s _ { t } = \\mathrm { g u m b e l - s o f t m a x } ( \\pi _ { t } ) . } } \\end{array}", + "type": "interline_equation", + "image_path": "93c16322c16a353d9081e01081b4078d9a8374ee48f55d406ae85acf775e9d61.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 409, + 401, + 425.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 209, + 425.0, + 401, + 441.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 448, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 449, + 504, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 134, + 461 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 449, + 188, + 461 + ], + "score": 0.93, + "content": "\\{ a , W ^ { \\gamma } , b ^ { \\gamma } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 449, + 504, + 461 + ], + "score": 1.0, + "content": "are learnable parameters (note that in our experiments, we always set the di-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 460, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 154, + 472 + ], + "score": 1.0, + "content": "mension of", + "type": "text" + }, + { + "bbox": [ + 154, + 462, + 162, + 470 + ], + "score": 0.77, + "content": "\\textbf { \\em a }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 460, + 309, + 472 + ], + "score": 1.0, + "content": "to about 1/4 of the hidden size) and", + "type": "text" + }, + { + "bbox": [ + 309, + 461, + 320, + 471 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } \\mathbf { u } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 460, + 506, + 472 + ], + "score": 1.0, + "content": "is a hand-crafted usage vector which denotes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 470, + 321, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 321, + 484 + ], + "score": 1.0, + "content": "the frequency of accesses to each cell in the memory.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 503, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 504, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 504, + 501 + ], + "score": 1.0, + "content": "There are two important differences with respect to TARDIS addressing mechanism and our auto-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 499, + 203, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 203, + 511 + ], + "score": 1.0, + "content": "addressing mechanism:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 520, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 308, + 534 + ], + "score": 1.0, + "content": "• In practice, equation 21 requires concatenating", + "type": "text" + }, + { + "bbox": [ + 309, + 523, + 320, + 532 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 521, + 339, + 534 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 339, + 522, + 350, + 532 + ], + "score": 0.89, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 521, + 445, + 534 + ], + "score": 1.0, + "content": "to every memory cells", + "type": "text" + }, + { + "bbox": [ + 446, + 521, + 470, + 533 + ], + "score": 0.92, + "content": "M _ { t } [ i ]", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "for par-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "allelization, which often causes significant extra memory and time consumption compared to the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "auto-addressing of ARMIN, as we show in the algorithmic tasks and char-level language modelling", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 128, + 565 + ], + "score": 1.0, + "content": "task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 341, + 581 + ], + "score": 1.0, + "content": "• The auto-addressing omits the learnable address section", + "type": "text" + }, + { + "bbox": [ + 342, + 569, + 354, + 580 + ], + "score": 0.89, + "content": "\\pmb { A } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 568, + 425, + 581 + ], + "score": 1.0, + "content": ", the hand-crafted", + "type": "text" + }, + { + "bbox": [ + 426, + 570, + 437, + 580 + ], + "score": 0.85, + "content": "\\mathbf { \\pmb { u } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "vector, the tanh", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 580, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 246, + 591 + ], + "score": 1.0, + "content": "activation and an auxiliary vector", + "type": "text" + }, + { + "bbox": [ + 246, + 582, + 254, + 590 + ], + "score": 0.65, + "content": "^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 580, + 505, + 591 + ], + "score": 1.0, + "content": "compared to TARDIS addressing mechanism. The TARDIS", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "also requires manually subtracting 100 in unnormalized probability at last read location to avoid", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "repeated read, whereas the auto-addressing doesn’t need to. With these simplifications, we found", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "the auto-addressing mechanism easier to train and it leads to a much faster converge speed, as we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "show in the algorithmic tasks. We also observe a BPC performance growth from 1.223 to 1.202 in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 373, + 646 + ], + "score": 1.0, + "content": "char-level language modelling task when we switch from ARMIN", + "type": "text" + }, + { + "bbox": [ + 374, + 636, + 380, + 645 + ], + "score": 0.48, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "TARDIS-addr to ARMIN with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 645, + 175, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 175, + 659 + ], + "score": 1.0, + "content": "auto-addressing.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5 + }, + { + "type": "title", + "bbox": [ + 109, + 673, + 295, + 686 + ], + "lines": [ + { + "bbox": [ + 107, + 672, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 107, + 672, + 297, + 688 + ], + "score": 1.0, + "content": "APPENDIX C EXPERIMENT SETUPS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "We first describe some universal setups in all our experiments (unless we mention them again with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "different setup). Our model can be trained end-to-end, due to the utilization of gumbel-softmax", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 417, + 733 + ], + "score": 1.0, + "content": "function. As pointed out in Maddison et al. (2016), when the temperature", + "type": "text" + }, + { + "bbox": [ + 418, + 720, + 484, + 732 + ], + "score": 0.93, + "content": "\\tau \\leq ( n - 1 ) ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 720, + 506, + 733 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 288, + 95 + ], + "score": 1.0, + "content": "• The TARDIS uses two binary scalar gates", + "type": "text" + }, + { + "bbox": [ + 289, + 83, + 313, + 94 + ], + "score": 0.91, + "content": "\\alpha _ { t } , \\beta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "to control the information flow in equation 17.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "The two scalar gates are activated using gumbel-sigmoid (which is similar to gumbel-softmax and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "the output is close to binary), whereas the ARMIN uses two soft vector gates to control the infor-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "mation flow, which is more flexible and enable larger model capacity. For example, in the char-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "level language modelling experiments, if we replace the sigmoid activations in equation 3 with the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 483, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 483, + 150 + ], + "score": 1.0, + "content": "element-wise gumbel-sigmoid, the BPC performance immediately drops from 1.202 to 1.235.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 82, + 506, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 505, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 406, + 165 + ], + "score": 1.0, + "content": "• The TADIS apply information control only when calculating the cell state", + "type": "text" + }, + { + "bbox": [ + 407, + 154, + 417, + 164 + ], + "score": 0.85, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { g } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 152, + 505, + 165 + ], + "score": 1.0, + "content": ", whereas the ARMIN", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "apply information control immediately after the read operation. Our main motivation to this change", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 317, + 186 + ], + "score": 1.0, + "content": "is that if we don’t filter out irrelevant information in", + "type": "text" + }, + { + "bbox": [ + 318, + 176, + 328, + 185 + ], + "score": 0.84, + "content": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 174, + 346, + 186 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 346, + 175, + 358, + 185 + ], + "score": 0.89, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "at first, these noisy information will", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 184, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 506, + 200 + ], + "score": 1.0, + "content": "cause the weight parameters hard to learn. Furthermore, if the read operation chooses the wrong", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 196, + 381, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 301, + 209 + ], + "score": 1.0, + "content": "past state, we can block the information flow for", + "type": "text" + }, + { + "bbox": [ + 302, + 198, + 312, + 207 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 196, + 381, + 209 + ], + "score": 1.0, + "content": "at the first place.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 152, + 506, + 209 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 503, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 248, + 224 + ], + "score": 1.0, + "content": "• The ARMIN has an extra gate", + "type": "text" + }, + { + "bbox": [ + 248, + 212, + 259, + 223 + ], + "score": 0.88, + "content": "\\mathbf { \\delta } _ { o _ { t } ^ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 210, + 375, + 224 + ], + "score": 1.0, + "content": "to control the output from", + "type": "text" + }, + { + "bbox": [ + 375, + 214, + 385, + 222 + ], + "score": 0.86, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 210, + 505, + 224 + ], + "score": 1.0, + "content": "as in equation 8, which the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 369, + 235 + ], + "score": 1.0, + "content": "TARDIS doesn’t have. If we remove this gate and directly output", + "type": "text" + }, + { + "bbox": [ + 369, + 224, + 379, + 234 + ], + "score": 0.85, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "like in TARDIS, we get a BPC", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 233, + 416, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 416, + 246 + ], + "score": 1.0, + "content": "performance drop from 1.202 to 1.219 in char-level language modelling task.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 210, + 505, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "Finally, despite that the ARMIN has 3 more vector gates than the TARDIS, we have shown in the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 144, + 278 + ], + "score": 0.41, + "content": "p \\mathrm { M N I S T }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "and Penn Treebank experiments that the ARMIN can still outperform the TARDIS under", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 277, + 210, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 210, + 290 + ], + "score": 1.0, + "content": "similar parameter counts.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 255, + 505, + 290 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 211, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 213, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 213, + 315 + ], + "score": 1.0, + "content": "B.2 READ OPERATION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 192, + 335 + ], + "score": 1.0, + "content": "The memory matrix", + "type": "text" + }, + { + "bbox": [ + 193, + 324, + 209, + 334 + ], + "score": 0.86, + "content": "M _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 321, + 379, + 335 + ], + "score": 1.0, + "content": "of TARDIS has disjoint address section", + "type": "text" + }, + { + "bbox": [ + 379, + 322, + 452, + 334 + ], + "score": 0.91, + "content": "\\mathbf { \\boldsymbol { A } } _ { t } ~ \\in ~ \\mathbb { R } ^ { n _ { m e m } \\times d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "and content", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 138, + 347 + ], + "score": 1.0, + "content": "section", + "type": "text" + }, + { + "bbox": [ + 138, + 334, + 208, + 345 + ], + "score": 0.86, + "content": "C _ { t } \\ \\stackrel { \\cdot } { \\in } \\ \\mathbb { R } ^ { n _ { m e m } \\times d _ { r } }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 332, + 213, + 347 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 214, + 334, + 279, + 346 + ], + "score": 0.91, + "content": "M _ { t } = [ A _ { t } ; C _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 332, + 506, + 347 + ], + "score": 1.0, + "content": ". The controller reads both the address and the content", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 344, + 434, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 434, + 359 + ], + "score": 1.0, + "content": "parts of the memory, but it will only write into the content section of the memory.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 321, + 506, + 359 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 504, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 376 + ], + "score": 1.0, + "content": "The TARDIS paper explored two ways to sample the read locations——using reinforce and using", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "gumbel-softmax function, respectively. The gumbel-softmax way shows a better performance and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 384, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 294, + 397 + ], + "score": 1.0, + "content": "its read operation can be formulated as follow:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 361, + 506, + 397 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 409, + 401, + 441 + ], + "lines": [ + { + "bbox": [ + 209, + 409, + 401, + 441 + ], + "spans": [ + { + "bbox": [ + 209, + 409, + 401, + 441 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\pi _ { t } [ i ] = a ^ { \\top } \\operatorname { t a n h } ( W ^ { \\gamma } [ h _ { t } , x _ { t } , M _ { t } [ i ] , { \\boldsymbol u } _ { t } ] + b ^ { \\gamma } ) , } } \\\\ { { s _ { t } = \\mathrm { g u m b e l - s o f t m a x } ( \\pi _ { t } ) . } } \\end{array}", + "type": "interline_equation", + "image_path": "93c16322c16a353d9081e01081b4078d9a8374ee48f55d406ae85acf775e9d61.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 409, + 401, + 425.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 209, + 425.0, + 401, + 441.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 448, + 504, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 449, + 504, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 134, + 461 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 449, + 188, + 461 + ], + "score": 0.93, + "content": "\\{ a , W ^ { \\gamma } , b ^ { \\gamma } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 449, + 504, + 461 + ], + "score": 1.0, + "content": "are learnable parameters (note that in our experiments, we always set the di-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 460, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 154, + 472 + ], + "score": 1.0, + "content": "mension of", + "type": "text" + }, + { + "bbox": [ + 154, + 462, + 162, + 470 + ], + "score": 0.77, + "content": "\\textbf { \\em a }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 460, + 309, + 472 + ], + "score": 1.0, + "content": "to about 1/4 of the hidden size) and", + "type": "text" + }, + { + "bbox": [ + 309, + 461, + 320, + 471 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } \\mathbf { u } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 460, + 506, + 472 + ], + "score": 1.0, + "content": "is a hand-crafted usage vector which denotes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 470, + 321, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 321, + 484 + ], + "score": 1.0, + "content": "the frequency of accesses to each cell in the memory.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 449, + 506, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 503, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 504, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 504, + 501 + ], + "score": 1.0, + "content": "There are two important differences with respect to TARDIS addressing mechanism and our auto-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 499, + 203, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 203, + 511 + ], + "score": 1.0, + "content": "addressing mechanism:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 487, + 504, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 520, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 308, + 534 + ], + "score": 1.0, + "content": "• In practice, equation 21 requires concatenating", + "type": "text" + }, + { + "bbox": [ + 309, + 523, + 320, + 532 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 521, + 339, + 534 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 339, + 522, + 350, + 532 + ], + "score": 0.89, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 521, + 445, + 534 + ], + "score": 1.0, + "content": "to every memory cells", + "type": "text" + }, + { + "bbox": [ + 446, + 521, + 470, + 533 + ], + "score": 0.92, + "content": "M _ { t } [ i ]", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "for par-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "allelization, which often causes significant extra memory and time consumption compared to the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "auto-addressing of ARMIN, as we show in the algorithmic tasks and char-level language modelling", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 128, + 565 + ], + "score": 1.0, + "content": "task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 521, + 505, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 341, + 581 + ], + "score": 1.0, + "content": "• The auto-addressing omits the learnable address section", + "type": "text" + }, + { + "bbox": [ + 342, + 569, + 354, + 580 + ], + "score": 0.89, + "content": "\\pmb { A } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 568, + 425, + 581 + ], + "score": 1.0, + "content": ", the hand-crafted", + "type": "text" + }, + { + "bbox": [ + 426, + 570, + 437, + 580 + ], + "score": 0.85, + "content": "\\mathbf { \\pmb { u } } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "vector, the tanh", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 580, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 246, + 591 + ], + "score": 1.0, + "content": "activation and an auxiliary vector", + "type": "text" + }, + { + "bbox": [ + 246, + 582, + 254, + 590 + ], + "score": 0.65, + "content": "^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 580, + 505, + 591 + ], + "score": 1.0, + "content": "compared to TARDIS addressing mechanism. The TARDIS", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "also requires manually subtracting 100 in unnormalized probability at last read location to avoid", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "repeated read, whereas the auto-addressing doesn’t need to. With these simplifications, we found", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "the auto-addressing mechanism easier to train and it leads to a much faster converge speed, as we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "show in the algorithmic tasks. We also observe a BPC performance growth from 1.223 to 1.202 in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 373, + 646 + ], + "score": 1.0, + "content": "char-level language modelling task when we switch from ARMIN", + "type": "text" + }, + { + "bbox": [ + 374, + 636, + 380, + 645 + ], + "score": 0.48, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "TARDIS-addr to ARMIN with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 645, + 175, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 175, + 659 + ], + "score": 1.0, + "content": "auto-addressing.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 568, + 506, + 659 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 673, + 295, + 686 + ], + "lines": [ + { + "bbox": [ + 107, + 672, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 107, + 672, + 297, + 688 + ], + "score": 1.0, + "content": "APPENDIX C EXPERIMENT SETUPS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "We first describe some universal setups in all our experiments (unless we mention them again with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "different setup). Our model can be trained end-to-end, due to the utilization of gumbel-softmax", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 417, + 733 + ], + "score": 1.0, + "content": "function. As pointed out in Maddison et al. (2016), when the temperature", + "type": "text" + }, + { + "bbox": [ + 418, + 720, + 484, + 732 + ], + "score": 0.93, + "content": "\\tau \\leq ( n - 1 ) ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 720, + 506, + 733 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "density function of gumbel-softmax estimator is log-convex, which gives us a good guarantee for", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 234, + 368 + ], + "score": 1.0, + "content": "optimization. However, smaller", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 234, + 358, + 241, + 366 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 241, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "will also bring higher variance of the gradients (Jang et al., 2016).", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 320, + 379 + ], + "score": 1.0, + "content": "So in all experiments, we initialize the reciprocal of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 320, + 369, + 327, + 377 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 328, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "as 1, and increase it by 1 after each epoch,", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 222, + 390 + ], + "score": 1.0, + "content": "until the reciprocal equals to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 222, + 379, + 245, + 388 + ], + "score": 0.85, + "content": "n - 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 246, + 378, + 505, + 390 + ], + "score": 1.0, + "content": ". This scheme has shown better performance than other schemes.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 699, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 196, + 105, + 415, + 190 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 181, + 80, + 430, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 180, + 79, + 430, + 94 + ], + "spans": [ + { + "bbox": [ + 180, + 79, + 430, + 94 + ], + "score": 1.0, + "content": "Table 5: Hyperparameters of the models for algorithmic tasks.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 196, + 105, + 415, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 105, + 415, + 190 + ], + "spans": [ + { + "bbox": [ + 196, + 105, + 415, + 190 + ], + "score": 0.972, + "html": "
ModelHidden Sizememory SizeParam Count
LSTM300一376k
NTM120128×2088k
TARDIS12050×3290k
ARMIN+TARDIS10050×3291k
ARMIN10050×3290k
", + "type": "table", + "image_path": "78261b201b724626971a34ad21df1c98e3584c9aec6b3dee323832c606b2286f.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 105, + 415, + 119.16666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 196, + 119.16666666666667, + 415, + 133.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 196, + 133.33333333333334, + 415, + 147.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 196, + 147.5, + 415, + 161.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 196, + 161.66666666666666, + 415, + 175.83333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 196, + 175.83333333333331, + 415, + 189.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 202, + 231, + 408, + 318 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 203, + 207, + 408, + 219 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 202, + 204, + 408, + 221 + ], + "spans": [ + { + "bbox": [ + 202, + 204, + 408, + 221 + ], + "score": 1.0, + "content": "Table 6: Hyperparameters of the algorithmic tasks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 202, + 231, + 408, + 318 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 202, + 231, + 408, + 318 + ], + "spans": [ + { + "bbox": [ + 202, + 231, + 408, + 318 + ], + "score": 0.974, + "html": "
TaskSetup
Copycopy length =1~50
Repeat Copycopyl length=1~10 number of repeats = 1~10
Associative Recallitem size=3×6 item number = 2~6
Priority SortSort length = 40
", + "type": "table", + "image_path": "3cd63d277de72abed816edd48b7040e7ae1701681a9dcf4a3cd4fb53edb5f175.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 231, + 408, + 245.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 202, + 245.5, + 408, + 260.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 202, + 260.0, + 408, + 274.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 202, + 274.5, + 408, + 289.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 202, + 289.0, + 408, + 303.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 202, + 303.5, + 408, + 318.0 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 8.75 + }, + { + "type": "text", + "bbox": [ + 108, + 345, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "density function of gumbel-softmax estimator is log-convex, which gives us a good guarantee for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 234, + 368 + ], + "score": 1.0, + "content": "optimization. However, smaller", + "type": "text" + }, + { + "bbox": [ + 234, + 358, + 241, + 366 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "will also bring higher variance of the gradients (Jang et al., 2016).", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 320, + 379 + ], + "score": 1.0, + "content": "So in all experiments, we initialize the reciprocal of", + "type": "text" + }, + { + "bbox": [ + 320, + 369, + 327, + 377 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "as 1, and increase it by 1 after each epoch,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 222, + 390 + ], + "score": 1.0, + "content": "until the reciprocal equals to", + "type": "text" + }, + { + "bbox": [ + 222, + 379, + 245, + 388 + ], + "score": 0.85, + "content": "n - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 378, + 505, + 390 + ], + "score": 1.0, + "content": ". This scheme has shown better performance than other schemes.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "We apply Layer Normalization(Ba et al., 2016) to all networks. In the ARMIN case, layer norm", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "is applied to equation 3, 6 and 7, which we find important to regulate the hidden state when scale", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "grows. To avoid overfitting, we apply dropout (Srivastava et al., 2014) to the input and output", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "layers, and Zoneout(Krueger et al., 2016) is applied for recurrent connections. The networks are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 439, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 451 + ], + "score": 1.0, + "content": "trained with Adam optimizer (Kingma & Ba, 2014). All weight matrices in RNNs are initialized", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 448, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 329, + 463 + ], + "score": 1.0, + "content": "to orthogonal matrices, and the bias of the forget gate", + "type": "text" + }, + { + "bbox": [ + 330, + 450, + 340, + 461 + ], + "score": 0.87, + "content": "\\pmb { f } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 448, + 505, + 463 + ], + "score": 1.0, + "content": "is initialized to 1. For all models using", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "TARDIS addressing mechanism, we set the memory key size to about 1/5 of the memory content", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 471, + 128, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 128, + 484 + ], + "score": 1.0, + "content": "size.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 225, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 227, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 227, + 511 + ], + "score": 1.0, + "content": "C.1 ALGORITHMIC TASKS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "In all algorithmic experiments below, following Graves et al. (2014), we apply RMSProp optimizer", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "(Tieleman & Hinton, 2012) with a momentum of 0.9. The learning rate is 0.0001 and gradient", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 541, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 505, + 556 + ], + "score": 1.0, + "content": "norms are clipped to 10.0. No regularization is used as the size of the networks is small. Following", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "Graves et al. (2014), we have learned bias of initial hidden states and memory matrices for all", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 273, + 576 + ], + "score": 1.0, + "content": "memory networks. For TARDIS, ARMIN", + "type": "text" + }, + { + "bbox": [ + 273, + 565, + 280, + 573 + ], + "score": 0.46, + "content": "+ { \\bf \\Phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "TARDIS-addr and ARMIN we increase the reciprocal of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 575, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 156, + 586 + ], + "score": 1.0, + "content": "temperature", + "type": "text" + }, + { + "bbox": [ + 156, + 576, + 164, + 585 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 575, + 506, + 586 + ], + "score": 1.0, + "content": "by 1 for every 200 iterations. The hyperparamters for models and tasks are shown in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 586, + 458, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 458, + 598 + ], + "score": 1.0, + "content": "Table 5 and Table 6, respectively. We explain these tasks in detail in subsections below:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 611, + 166, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 610, + 168, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 168, + 624 + ], + "score": 1.0, + "content": "C.1.1 COPY", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "score": 1.0, + "content": "The copy task tests whether a recurrent model can correctly store and recall a long sequence of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 641, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 504, + 655 + ], + "score": 1.0, + "content": "arbitrary information. In our experiment, the networks are given randomly generated eight-bit binary", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "sequences with its length ranging from 1 to 50, followed by a finish indicator. Then the networks", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 666, + 321, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 321, + 676 + ], + "score": 1.0, + "content": "are asked to output the same sequence with the input.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 202, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 204, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 204, + 703 + ], + "score": 1.0, + "content": "C.1.2 REPEAT COPY", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "The repeat copy task is a repeat version of the copy task. The main motivation was to see if a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "recurrent model could memorize input information and repeatedly reuse them at later time-steps. In", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 196, + 105, + 415, + 190 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 181, + 80, + 430, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 180, + 79, + 430, + 94 + ], + "spans": [ + { + "bbox": [ + 180, + 79, + 430, + 94 + ], + "score": 1.0, + "content": "Table 5: Hyperparameters of the models for algorithmic tasks.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 196, + 105, + 415, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 105, + 415, + 190 + ], + "spans": [ + { + "bbox": [ + 196, + 105, + 415, + 190 + ], + "score": 0.972, + "html": "
ModelHidden Sizememory SizeParam Count
LSTM300一376k
NTM120128×2088k
TARDIS12050×3290k
ARMIN+TARDIS10050×3291k
ARMIN10050×3290k
", + "type": "table", + "image_path": "78261b201b724626971a34ad21df1c98e3584c9aec6b3dee323832c606b2286f.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 105, + 415, + 119.16666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 196, + 119.16666666666667, + 415, + 133.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 196, + 133.33333333333334, + 415, + 147.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 196, + 147.5, + 415, + 161.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 196, + 161.66666666666666, + 415, + 175.83333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 196, + 175.83333333333331, + 415, + 189.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 202, + 231, + 408, + 318 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 203, + 207, + 408, + 219 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 202, + 204, + 408, + 221 + ], + "spans": [ + { + "bbox": [ + 202, + 204, + 408, + 221 + ], + "score": 1.0, + "content": "Table 6: Hyperparameters of the algorithmic tasks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 202, + 231, + 408, + 318 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 202, + 231, + 408, + 318 + ], + "spans": [ + { + "bbox": [ + 202, + 231, + 408, + 318 + ], + "score": 0.974, + "html": "
TaskSetup
Copycopy length =1~50
Repeat Copycopyl length=1~10 number of repeats = 1~10
Associative Recallitem size=3×6 item number = 2~6
Priority SortSort length = 40
", + "type": "table", + "image_path": "3cd63d277de72abed816edd48b7040e7ae1701681a9dcf4a3cd4fb53edb5f175.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 231, + 408, + 245.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 202, + 245.5, + 408, + 260.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 202, + 260.0, + 408, + 274.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 202, + 274.5, + 408, + 289.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 202, + 289.0, + 408, + 303.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 202, + 303.5, + 408, + 318.0 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 8.75 + }, + { + "type": "text", + "bbox": [ + 108, + 345, + 505, + 389 + ], + "lines": [], + "index": 15.5, + "bbox_fs": [ + 106, + 345, + 505, + 390 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "We apply Layer Normalization(Ba et al., 2016) to all networks. In the ARMIN case, layer norm", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "is applied to equation 3, 6 and 7, which we find important to regulate the hidden state when scale", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "grows. To avoid overfitting, we apply dropout (Srivastava et al., 2014) to the input and output", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "layers, and Zoneout(Krueger et al., 2016) is applied for recurrent connections. The networks are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 439, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 451 + ], + "score": 1.0, + "content": "trained with Adam optimizer (Kingma & Ba, 2014). All weight matrices in RNNs are initialized", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 448, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 329, + 463 + ], + "score": 1.0, + "content": "to orthogonal matrices, and the bias of the forget gate", + "type": "text" + }, + { + "bbox": [ + 330, + 450, + 340, + 461 + ], + "score": 0.87, + "content": "\\pmb { f } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 448, + 505, + 463 + ], + "score": 1.0, + "content": "is initialized to 1. For all models using", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "TARDIS addressing mechanism, we set the memory key size to about 1/5 of the memory content", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 471, + 128, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 128, + 484 + ], + "score": 1.0, + "content": "size.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 394, + 506, + 484 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 225, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 227, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 227, + 511 + ], + "score": 1.0, + "content": "C.1 ALGORITHMIC TASKS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "In all algorithmic experiments below, following Graves et al. (2014), we apply RMSProp optimizer", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "(Tieleman & Hinton, 2012) with a momentum of 0.9. The learning rate is 0.0001 and gradient", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 541, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 505, + 556 + ], + "score": 1.0, + "content": "norms are clipped to 10.0. No regularization is used as the size of the networks is small. Following", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "Graves et al. (2014), we have learned bias of initial hidden states and memory matrices for all", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 273, + 576 + ], + "score": 1.0, + "content": "memory networks. For TARDIS, ARMIN", + "type": "text" + }, + { + "bbox": [ + 273, + 565, + 280, + 573 + ], + "score": 0.46, + "content": "+ { \\bf \\Phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "TARDIS-addr and ARMIN we increase the reciprocal of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 575, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 156, + 586 + ], + "score": 1.0, + "content": "temperature", + "type": "text" + }, + { + "bbox": [ + 156, + 576, + 164, + 585 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 575, + 506, + 586 + ], + "score": 1.0, + "content": "by 1 for every 200 iterations. The hyperparamters for models and tasks are shown in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 586, + 458, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 458, + 598 + ], + "score": 1.0, + "content": "Table 5 and Table 6, respectively. We explain these tasks in detail in subsections below:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 519, + 506, + 598 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 611, + 166, + 622 + ], + "lines": [ + { + "bbox": [ + 105, + 610, + 168, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 168, + 624 + ], + "score": 1.0, + "content": "C.1.1 COPY", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "score": 1.0, + "content": "The copy task tests whether a recurrent model can correctly store and recall a long sequence of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 641, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 504, + 655 + ], + "score": 1.0, + "content": "arbitrary information. In our experiment, the networks are given randomly generated eight-bit binary", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "sequences with its length ranging from 1 to 50, followed by a finish indicator. Then the networks", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 666, + 321, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 321, + 676 + ], + "score": 1.0, + "content": "are asked to output the same sequence with the input.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 630, + 506, + 676 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 202, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 204, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 204, + 703 + ], + "score": 1.0, + "content": "C.1.2 REPEAT COPY", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "The repeat copy task is a repeat version of the copy task. The main motivation was to see if a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "recurrent model could memorize input information and repeatedly reuse them at later time-steps. In", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 709, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 96 + ], + "score": 1.0, + "content": "our experiment, the networks are asked to repeatedly copy the randomly generated input sequence", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 348, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 348, + 106 + ], + "score": 1.0, + "content": "with length ranging from 1 to 10 to output for 1 to 10 times.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 122, + 237, + 134 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 239, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 239, + 135 + ], + "score": 1.0, + "content": "C.1.3 ASSOCIATIVE RECALL", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "This task test whether a memory network has the ability to retrieve memory data indirectly. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "networks are given a sequence of items and then a query with one of the items, and are asked to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "score": 1.0, + "content": "output the subsequent item. In our experiments, each item consists of three six-bit binary vectors,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 177, + 303, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 303, + 189 + ], + "score": 1.0, + "content": "and the number of items are ranging from 2 to 6.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 107, + 206, + 209, + 218 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 211, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 211, + 219 + ], + "score": 1.0, + "content": "C.1.4 PRIORITY SORT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "This task test whether a recurrent model can do priority sort——a classic elementary algorithm. This", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "task is more challenging than previous ones as it not only require memorizing the input sequences", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "but also making efficient use of the mutual relations of the items in the input sequences. In our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "experiments, the networks are given 40 eight-bit random binary keys with scalar priority values in a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 271, + 432, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 432, + 286 + ], + "score": 1.0, + "content": "sequence, and are asked to output the top 30 keys in descending order of priority.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 303, + 460, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 462, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 462, + 315 + ], + "score": 1.0, + "content": "C.2 CHARACTER-LEVEL LANGUAGE MODELLING ON PENN TREENBANK DATASET", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 325, + 255, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 257, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 257, + 338 + ], + "score": 1.0, + "content": "C.2.1 PENN TREEBANK DATASET", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 347, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "We use the train/validation/test split outlined in Mikolov et al. (2012). The train/validation/test batch", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "score": 1.0, + "content": "size are 128/64/1 respectively. Following prior works on this task (Ha et al., 2016; Suarez, 2017;", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Mujika et al., 2017), the model is optimized over the cross-entropy loss between the predictions and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "score": 1.0, + "content": "the training labels, and evaluated using a bits-per-character measure. We use a single ARMIN cell of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "different hidden sizes, with a fully connected layer to output the predictive probabilities. The input", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 171, + 415 + ], + "score": 1.0, + "content": "embedding size", + "type": "text" + }, + { + "bbox": [ + 171, + 402, + 181, + 413 + ], + "score": 0.87, + "content": "d _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 402, + 354, + 415 + ], + "score": 1.0, + "content": "is set to 128. The number of memory slots", + "type": "text" + }, + { + "bbox": [ + 355, + 403, + 380, + 413 + ], + "score": 0.89, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "is 20. The zoneout probability", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "is 0.3. We train the networks for 200 epochs with a learning rate of 0.002, and decay with a factor", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 424, + 414, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 414, + 435 + ], + "score": 1.0, + "content": "of 10 at the last 20 epochs. We clip the gradients to a maximum norm of 1.0.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 453, + 293, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 295, + 465 + ], + "score": 1.0, + "content": "C.2.2 HUTTER PRIZE WIKIPEDIA DATASET", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "We use the train/validation/test split outlined in Chung et al. (2015). The train/validation/test batch", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "size are 128/64/1 respectively, and the vocabulary size is 205. The zoneout and dropout probabilities", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "score": 1.0, + "content": "are 0.3 and 0.2, respectively. The input embedding size is 256. Our ARMIN network has 2 layer", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "score": 1.0, + "content": "and 800 units at each layer, connected by a linear layer to convert 1600-dimensional features from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 284, + 532 + ], + "score": 1.0, + "content": "layer 1 to 256 dimension as input to layer 2.", + "type": "text" + }, + { + "bbox": [ + 284, + 520, + 309, + 530 + ], + "score": 0.86, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "for each layer is set to 10. We train the networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "for 50 epochs with a learning rate of 0.001, and decay with a factor of 10 at the last 10 peochs. We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 540, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 289, + 552 + ], + "score": 1.0, + "content": "clip the gradients to a maximum norm of 1.0.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 420, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 423, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 423, + 584 + ], + "score": 1.0, + "content": "C.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "The SST takes in sequential inputs of a video and outputs the existing probability predictions of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "action proposals with different time lengths that end at each time-step, and uses a multi-label loss to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "score": 1.0, + "content": "optimize its parameters. The sequential inputs are 4096-dimensional features extracted from every", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "16 consecutive video frames using C3D network (Tran et al., 2015), followed by PCA to downsize", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 638, + 237, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 237, + 651 + ], + "score": 1.0, + "content": "the input feature dimensionality.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Our experiment setup is basically the same as the SST paper, except that we use a simple fully", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "connected layer to downsize the input features instead of using PCA. The input size of the two", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "networks are both set to 256. The MA-SST have 256 hidden units, while the SST with LSTM have", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "512 hidden units. This ensures the 2 networks have roughly the same parameter count. The dropout", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 323, + 712 + ], + "score": 1.0, + "content": "and zoneout probabilities are both set to 0.2 and 0.3.", + "type": "text" + }, + { + "bbox": [ + 324, + 700, + 349, + 710 + ], + "score": 0.85, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "of the MA-SST is set to 35. We train", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the networks for 300 epochs with initial learning rate of 0.002, and decay with a factor of 10 at the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 168, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 168, + 733 + ], + "score": 1.0, + "content": "last 60 epochs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 96 + ], + "score": 1.0, + "content": "our experiment, the networks are asked to repeatedly copy the randomly generated input sequence", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 348, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 348, + 106 + ], + "score": 1.0, + "content": "with length ranging from 1 to 10 to output for 1 to 10 times.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 504, + 106 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 122, + 237, + 134 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 239, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 239, + 135 + ], + "score": 1.0, + "content": "C.1.3 ASSOCIATIVE RECALL", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "This task test whether a memory network has the ability to retrieve memory data indirectly. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "networks are given a sequence of items and then a query with one of the items, and are asked to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 179 + ], + "score": 1.0, + "content": "output the subsequent item. In our experiments, each item consists of three six-bit binary vectors,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 177, + 303, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 303, + 189 + ], + "score": 1.0, + "content": "and the number of items are ranging from 2 to 6.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 145, + 505, + 189 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 206, + 209, + 218 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 211, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 211, + 219 + ], + "score": 1.0, + "content": "C.1.4 PRIORITY SORT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "This task test whether a recurrent model can do priority sort——a classic elementary algorithm. This", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "task is more challenging than previous ones as it not only require memorizing the input sequences", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "but also making efficient use of the mutual relations of the items in the input sequences. In our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 506, + 274 + ], + "score": 1.0, + "content": "experiments, the networks are given 40 eight-bit random binary keys with scalar priority values in a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 271, + 432, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 432, + 286 + ], + "score": 1.0, + "content": "sequence, and are asked to output the top 30 keys in descending order of priority.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 228, + 506, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 303, + 460, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 462, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 462, + 315 + ], + "score": 1.0, + "content": "C.2 CHARACTER-LEVEL LANGUAGE MODELLING ON PENN TREENBANK DATASET", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 302, + 462, + 315 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 325, + 255, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 257, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 257, + 338 + ], + "score": 1.0, + "content": "C.2.1 PENN TREEBANK DATASET", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 347, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "We use the train/validation/test split outlined in Mikolov et al. (2012). The train/validation/test batch", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "score": 1.0, + "content": "size are 128/64/1 respectively. Following prior works on this task (Ha et al., 2016; Suarez, 2017;", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "Mujika et al., 2017), the model is optimized over the cross-entropy loss between the predictions and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "score": 1.0, + "content": "the training labels, and evaluated using a bits-per-character measure. We use a single ARMIN cell of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "different hidden sizes, with a fully connected layer to output the predictive probabilities. The input", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 171, + 415 + ], + "score": 1.0, + "content": "embedding size", + "type": "text" + }, + { + "bbox": [ + 171, + 402, + 181, + 413 + ], + "score": 0.87, + "content": "d _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 402, + 354, + 415 + ], + "score": 1.0, + "content": "is set to 128. The number of memory slots", + "type": "text" + }, + { + "bbox": [ + 355, + 403, + 380, + 413 + ], + "score": 0.89, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "is 20. The zoneout probability", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "is 0.3. We train the networks for 200 epochs with a learning rate of 0.002, and decay with a factor", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 424, + 414, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 414, + 435 + ], + "score": 1.0, + "content": "of 10 at the last 20 epochs. We clip the gradients to a maximum norm of 1.0.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 347, + 506, + 435 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 453, + 293, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 295, + 465 + ], + "score": 1.0, + "content": "C.2.2 HUTTER PRIZE WIKIPEDIA DATASET", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "We use the train/validation/test split outlined in Chung et al. (2015). The train/validation/test batch", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "size are 128/64/1 respectively, and the vocabulary size is 205. The zoneout and dropout probabilities", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "score": 1.0, + "content": "are 0.3 and 0.2, respectively. The input embedding size is 256. Our ARMIN network has 2 layer", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "score": 1.0, + "content": "and 800 units at each layer, connected by a linear layer to convert 1600-dimensional features from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 284, + 532 + ], + "score": 1.0, + "content": "layer 1 to 256 dimension as input to layer 2.", + "type": "text" + }, + { + "bbox": [ + 284, + 520, + 309, + 530 + ], + "score": 0.86, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "for each layer is set to 10. We train the networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "for 50 epochs with a learning rate of 0.001, and decay with a factor of 10 at the last 10 peochs. We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 540, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 289, + 552 + ], + "score": 1.0, + "content": "clip the gradients to a maximum norm of 1.0.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 474, + 506, + 552 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 420, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 423, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 423, + 584 + ], + "score": 1.0, + "content": "C.3 TEMPORAL ACTION DETECTION AND PROPOSALS ON THUMOS’ 14", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "The SST takes in sequential inputs of a video and outputs the existing probability predictions of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "action proposals with different time lengths that end at each time-step, and uses a multi-label loss to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "score": 1.0, + "content": "optimize its parameters. The sequential inputs are 4096-dimensional features extracted from every", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "16 consecutive video frames using C3D network (Tran et al., 2015), followed by PCA to downsize", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 638, + 237, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 237, + 651 + ], + "score": 1.0, + "content": "the input feature dimensionality.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 593, + 505, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Our experiment setup is basically the same as the SST paper, except that we use a simple fully", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "connected layer to downsize the input features instead of using PCA. The input size of the two", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "networks are both set to 256. The MA-SST have 256 hidden units, while the SST with LSTM have", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "512 hidden units. This ensures the 2 networks have roughly the same parameter count. The dropout", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 323, + 712 + ], + "score": 1.0, + "content": "and zoneout probabilities are both set to 0.2 and 0.3.", + "type": "text" + }, + { + "bbox": [ + 324, + 700, + 349, + 710 + ], + "score": 0.85, + "content": "n _ { m e m }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "of the MA-SST is set to 35. We train", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the networks for 300 epochs with initial learning rate of 0.002, and decay with a factor of 10 at the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 168, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 168, + 733 + ], + "score": 1.0, + "content": "last 60 epochs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 87, + 500, + 254 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 87, + 500, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 87, + 500, + 254 + ], + "spans": [ + { + "bbox": [ + 111, + 87, + 500, + 254 + ], + "score": 0.966, + "type": "image", + "image_path": "50001ffd2185dd7661fc9aba2d9d75316cd989dcf8de377293cc7dbfab09dd11.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 87, + 500, + 142.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 142.66666666666666, + 500, + 198.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 198.33333333333331, + 500, + 253.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 263, + 257, + 347, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 262, + 256, + 348, + 271 + ], + "spans": [ + { + "bbox": [ + 262, + 256, + 348, + 271 + ], + "score": 1.0, + "content": "Figure 4: Copy task.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 112, + 298, + 496, + 465 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 298, + 496, + 465 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 298, + 496, + 465 + ], + "spans": [ + { + "bbox": [ + 112, + 298, + 496, + 465 + ], + "score": 0.971, + "type": "image", + "image_path": "4550719af5c0f00a37cd0f179fc1ec6a044179f516a3a2932ce1e6936572202f.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 112, + 298, + 496, + 353.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 112, + 353.6666666666667, + 496, + 409.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 409.33333333333337, + 496, + 465.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 250, + 468, + 361, + 480 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 248, + 466, + 361, + 483 + ], + "spans": [ + { + "bbox": [ + 248, + 466, + 361, + 483 + ], + "score": 1.0, + "content": "Figure 5: Repeat copy task.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "image", + "bbox": [ + 111, + 551, + 493, + 717 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 551, + 493, + 717 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 111, + 551, + 493, + 717 + ], + "spans": [ + { + "bbox": [ + 111, + 551, + 493, + 717 + ], + "score": 0.968, + "type": "image", + "image_path": "040994a58270f2a89222cd44b6a7cbcbc8407ddd4dfecf1d485eb625bbffc5c4.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 111, + 551, + 493, + 606.3333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 606.3333333333334, + 493, + 661.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 111, + 661.6666666666667, + 493, + 717.0000000000001 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 239, + 720, + 372, + 732 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 238, + 719, + 372, + 732 + ], + "spans": [ + { + "bbox": [ + 238, + 719, + 372, + 732 + ], + "score": 1.0, + "content": "Figure 6: Associative recall task.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 10.0 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 87, + 500, + 254 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 87, + 500, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 87, + 500, + 254 + ], + "spans": [ + { + "bbox": [ + 111, + 87, + 500, + 254 + ], + "score": 0.966, + "type": "image", + "image_path": "50001ffd2185dd7661fc9aba2d9d75316cd989dcf8de377293cc7dbfab09dd11.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 87, + 500, + 142.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 142.66666666666666, + 500, + 198.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 198.33333333333331, + 500, + 253.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 263, + 257, + 347, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 262, + 256, + 348, + 271 + ], + "spans": [ + { + "bbox": [ + 262, + 256, + 348, + 271 + ], + "score": 1.0, + "content": "Figure 4: Copy task.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 112, + 298, + 496, + 465 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 298, + 496, + 465 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 298, + 496, + 465 + ], + "spans": [ + { + "bbox": [ + 112, + 298, + 496, + 465 + ], + "score": 0.971, + "type": "image", + "image_path": "4550719af5c0f00a37cd0f179fc1ec6a044179f516a3a2932ce1e6936572202f.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 112, + 298, + 496, + 353.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 112, + 353.6666666666667, + 496, + 409.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 409.33333333333337, + 496, + 465.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 250, + 468, + 361, + 480 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 248, + 466, + 361, + 483 + ], + "spans": [ + { + "bbox": [ + 248, + 466, + 361, + 483 + ], + "score": 1.0, + "content": "Figure 5: Repeat copy task.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "image", + "bbox": [ + 111, + 551, + 493, + 717 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 551, + 493, + 717 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 111, + 551, + 493, + 717 + ], + "spans": [ + { + "bbox": [ + 111, + 551, + 493, + 717 + ], + "score": 0.968, + "type": "image", + "image_path": "040994a58270f2a89222cd44b6a7cbcbc8407ddd4dfecf1d485eb625bbffc5c4.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 111, + 551, + 493, + 606.3333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 606.3333333333334, + 493, + 661.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 111, + 661.6666666666667, + 493, + 717.0000000000001 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 239, + 720, + 372, + 732 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 238, + 719, + 372, + 732 + ], + "spans": [ + { + "bbox": [ + 238, + 719, + 372, + 732 + ], + "score": 1.0, + "content": "Figure 6: Associative recall task.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + } + ], + "index": 10.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 88, + 500, + 254 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 88, + 500, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 88, + 500, + 254 + ], + "spans": [ + { + "bbox": [ + 112, + 88, + 500, + 254 + ], + "score": 0.967, + "type": "image", + "image_path": "4cb5157c10e72c964f7ac67806728f1c6305e97e3dd30e339a3d94d5e8a27dc8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 88, + 500, + 143.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 143.33333333333334, + 500, + 198.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 198.66666666666669, + 500, + 254.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 250, + 257, + 360, + 269 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 250, + 257, + 361, + 271 + ], + "spans": [ + { + "bbox": [ + 250, + 257, + 361, + 271 + ], + "score": 1.0, + "content": "Figure 7: Priority sort task.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 223, + 297, + 387, + 393 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 210, + 277, + 400, + 289 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 210, + 275, + 400, + 291 + ], + "spans": [ + { + "bbox": [ + 210, + 275, + 361, + 291 + ], + "score": 1.0, + "content": "Table 7: Ablation on Penn Treebank", + "type": "text" + }, + { + "bbox": [ + 361, + 278, + 395, + 289 + ], + "score": 0.81, + "content": "\\scriptstyle ( d _ { h } = 8 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 275, + 400, + 291 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 223, + 297, + 387, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 223, + 297, + 387, + 393 + ], + "spans": [ + { + "bbox": [ + 223, + 297, + 387, + 393 + ], + "score": 0.94, + "html": "
Abaltion ModelBPC
No Control Gates1.354
LSTM (TBPTT length=50) ARMIN(TBPTT length=50)1.390 1.220
Random Read1.224
Queue-style Write1.352
Independent Read/write1.245
SmoothRead/write1.212
", + "type": "table", + "image_path": "2c719e270172f3bfec5784ed1d4baff76a9ce737a4b574ccd9f5a06794ee2ff7.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 223, + 297, + 387, + 310.7142857142857 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 223, + 310.7142857142857, + 387, + 324.42857142857144 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 223, + 324.42857142857144, + 387, + 338.14285714285717 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 223, + 338.14285714285717, + 387, + 351.8571428571429 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 223, + 351.8571428571429, + 387, + 365.5714285714286 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 223, + 365.5714285714286, + 387, + 379.28571428571433 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 223, + 379.28571428571433, + 387, + 393.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 108, + 413, + 411, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 413, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 413, + 427 + ], + "score": 1.0, + "content": "APPENDIX D TRAINING CURVES ON ALGORITHMIC TASKS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 440, + 503, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 453 + ], + "score": 1.0, + "content": "APPENDIX E ABLATION STUDY ON MEMORY AND ADDRESSING MECHANISM", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "In this subsection, we show the efficacy of our external memory and addressing mechanism by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "ablation experiments. We choose the char-level language modelling task to do the ablation re-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "search as it is a relatively standard benchmark task and has appropriate medium scale. We use the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "aforementioned experiment setup. To validate the efficacy of our external memory, we conduct the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 509, + 353, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 353, + 519 + ], + "score": 1.0, + "content": "experiments as follows and the results are depicted in table 7:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "• Short TBPTT Length We shorten the truncated length in TBPTT from 150 to 50. The gradients", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 542, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 553 + ], + "score": 1.0, + "content": "are truncated so we can’t do backpropagation on hidden states or memory from last iteration, but", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "the ARMIN still has direct access to historical hidden states. By comparing the performance of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 562, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 577 + ], + "score": 1.0, + "content": "LSTM and ARMIN under such circumstances, we can know the efficacy of the external memory.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "The results show that the LSTM has a dramatic performance drop of 1.390 BPC while the ARMIN", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "only has a minor performance drop of 1.220 BPC. The results confirm that the direct access to mem-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "ory is crucial when the networks can’t access historical information by backpropagating gradients.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "And notably, our experiment shows that the ARMIN is very robust to short-length TBPTT, i.e. small", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "score": 1.0, + "content": "iteration length, which means lower memory consumption. It enables using very large batch size to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 629, + 454, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 454, + 641 + ], + "score": 1.0, + "content": "speed up the training and further increasing performance on sequence modelling tasks.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 305, + 657 + ], + "score": 1.0, + "content": "• No Control Gates We remove the control gate", + "type": "text" + }, + { + "bbox": [ + 305, + 643, + 317, + 656 + ], + "score": 0.9, + "content": "\\pmb { g } _ { t } ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 642, + 335, + 657 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 335, + 644, + 347, + 656 + ], + "score": 0.89, + "content": "\\pmb { g } _ { t } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 642, + 506, + 657 + ], + "score": 1.0, + "content": "in ARMIN cell as depicted in equation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 181, + 668 + ], + "score": 1.0, + "content": "3, and directly use", + "type": "text" + }, + { + "bbox": [ + 181, + 656, + 192, + 666 + ], + "score": 0.86, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 654, + 209, + 668 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 210, + 655, + 231, + 666 + ], + "score": 0.91, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "to compute the gates in equation 6. The result shows a BPC of 1.354", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "that has a dramatic performance drop compared the 1.202 BPC of normal ARMIN. From this, we", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "score": 1.0, + "content": "deduce that the control gates successfully filter out irrelevant information. Without the control gates,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 688, + 409, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 181, + 700 + ], + "score": 1.0, + "content": "the historical state", + "type": "text" + }, + { + "bbox": [ + 181, + 689, + 191, + 699 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 688, + 409, + 700 + ], + "score": 1.0, + "content": "would have bring more noise than useful information.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "To validate the efficacy of the addressing mechanism of the ARMIN, we do the following modifica-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 720, + 475, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 475, + 732 + ], + "score": 1.0, + "content": "tions to the read/write operations of the ARMIN, and the results are also depicted in table 7:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 88, + 500, + 254 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 88, + 500, + 254 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 88, + 500, + 254 + ], + "spans": [ + { + "bbox": [ + 112, + 88, + 500, + 254 + ], + "score": 0.967, + "type": "image", + "image_path": "4cb5157c10e72c964f7ac67806728f1c6305e97e3dd30e339a3d94d5e8a27dc8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 88, + 500, + 143.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 143.33333333333334, + 500, + 198.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 198.66666666666669, + 500, + 254.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 250, + 257, + 360, + 269 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 250, + 257, + 361, + 271 + ], + "spans": [ + { + "bbox": [ + 250, + 257, + 361, + 271 + ], + "score": 1.0, + "content": "Figure 7: Priority sort task.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 223, + 297, + 387, + 393 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 210, + 277, + 400, + 289 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 210, + 275, + 400, + 291 + ], + "spans": [ + { + "bbox": [ + 210, + 275, + 361, + 291 + ], + "score": 1.0, + "content": "Table 7: Ablation on Penn Treebank", + "type": "text" + }, + { + "bbox": [ + 361, + 278, + 395, + 289 + ], + "score": 0.81, + "content": "\\scriptstyle ( d _ { h } = 8 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 275, + 400, + 291 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 223, + 297, + 387, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 223, + 297, + 387, + 393 + ], + "spans": [ + { + "bbox": [ + 223, + 297, + 387, + 393 + ], + "score": 0.94, + "html": "
Abaltion ModelBPC
No Control Gates1.354
LSTM (TBPTT length=50) ARMIN(TBPTT length=50)1.390 1.220
Random Read1.224
Queue-style Write1.352
Independent Read/write1.245
SmoothRead/write1.212
", + "type": "table", + "image_path": "2c719e270172f3bfec5784ed1d4baff76a9ce737a4b574ccd9f5a06794ee2ff7.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 223, + 297, + 387, + 310.7142857142857 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 223, + 310.7142857142857, + 387, + 324.42857142857144 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 223, + 324.42857142857144, + 387, + 338.14285714285717 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 223, + 338.14285714285717, + 387, + 351.8571428571429 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 223, + 351.8571428571429, + 387, + 365.5714285714286 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 223, + 365.5714285714286, + 387, + 379.28571428571433 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 223, + 379.28571428571433, + 387, + 393.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 108, + 413, + 411, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 413, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 413, + 427 + ], + "score": 1.0, + "content": "APPENDIX D TRAINING CURVES ON ALGORITHMIC TASKS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 412, + 413, + 427 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 440, + 503, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 453 + ], + "score": 1.0, + "content": "APPENDIX E ABLATION STUDY ON MEMORY AND ADDRESSING MECHANISM", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "In this subsection, we show the efficacy of our external memory and addressing mechanism by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "ablation experiments. We choose the char-level language modelling task to do the ablation re-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "search as it is a relatively standard benchmark task and has appropriate medium scale. We use the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "aforementioned experiment setup. To validate the efficacy of our external memory, we conduct the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 509, + 353, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 353, + 519 + ], + "score": 1.0, + "content": "experiments as follows and the results are depicted in table 7:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 464, + 505, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "• Short TBPTT Length We shorten the truncated length in TBPTT from 150 to 50. The gradients", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 542, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 553 + ], + "score": 1.0, + "content": "are truncated so we can’t do backpropagation on hidden states or memory from last iteration, but", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "the ARMIN still has direct access to historical hidden states. By comparing the performance of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 562, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 577 + ], + "score": 1.0, + "content": "LSTM and ARMIN under such circumstances, we can know the efficacy of the external memory.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "The results show that the LSTM has a dramatic performance drop of 1.390 BPC while the ARMIN", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "only has a minor performance drop of 1.220 BPC. The results confirm that the direct access to mem-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "ory is crucial when the networks can’t access historical information by backpropagating gradients.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "And notably, our experiment shows that the ARMIN is very robust to short-length TBPTT, i.e. small", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "score": 1.0, + "content": "iteration length, which means lower memory consumption. It enables using very large batch size to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 629, + 454, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 454, + 641 + ], + "score": 1.0, + "content": "speed up the training and further increasing performance on sequence modelling tasks.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 530, + 505, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 305, + 657 + ], + "score": 1.0, + "content": "• No Control Gates We remove the control gate", + "type": "text" + }, + { + "bbox": [ + 305, + 643, + 317, + 656 + ], + "score": 0.9, + "content": "\\pmb { g } _ { t } ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 642, + 335, + 657 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 335, + 644, + 347, + 656 + ], + "score": 0.89, + "content": "\\pmb { g } _ { t } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 642, + 506, + 657 + ], + "score": 1.0, + "content": "in ARMIN cell as depicted in equation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 181, + 668 + ], + "score": 1.0, + "content": "3, and directly use", + "type": "text" + }, + { + "bbox": [ + 181, + 656, + 192, + 666 + ], + "score": 0.86, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 654, + 209, + 668 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 210, + 655, + 231, + 666 + ], + "score": 0.91, + "content": "\\boldsymbol { h } _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "to compute the gates in equation 6. The result shows a BPC of 1.354", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "that has a dramatic performance drop compared the 1.202 BPC of normal ARMIN. From this, we", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "score": 1.0, + "content": "deduce that the control gates successfully filter out irrelevant information. Without the control gates,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 688, + 409, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 181, + 700 + ], + "score": 1.0, + "content": "the historical state", + "type": "text" + }, + { + "bbox": [ + 181, + 689, + 191, + 699 + ], + "score": 0.84, + "content": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 688, + 409, + 700 + ], + "score": 1.0, + "content": "would have bring more noise than useful information.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 642, + 506, + 700 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "To validate the efficacy of the addressing mechanism of the ARMIN, we do the following modifica-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 720, + 475, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 475, + 732 + ], + "score": 1.0, + "content": "tions to the read/write operations of the ARMIN, and the results are also depicted in table 7:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 709, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "• Random Read The ARMIN randomly chooses a historical state while reading. The result shows", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "a BPC of 1.224, compared to the 1.202 BPC of normal ARMIN. It proves that the read regime we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 496, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 496, + 117 + ], + "score": 1.0, + "content": "proposed in section 3.2 is effective and can choose more useful historical state than random read.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 119, + 505, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 118, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 505, + 132 + ], + "score": 1.0, + "content": "• Queue-style Write While writing, the ARMIN follows a fashion of first-in-first-out, i.e. always", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 355, + 142 + ], + "score": 1.0, + "content": "chooses the oldest memory slot to write. It makes the memory", + "type": "text" + }, + { + "bbox": [ + 356, + 131, + 369, + 141 + ], + "score": 0.65, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 131, + 505, + 142 + ], + "score": 1.0, + "content": "stubbornly store the hidden states", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 141, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 504, + 153 + ], + "score": 1.0, + "content": "of the last 20 time-steps. The result shows a BPC of 1.352, which proves our normal read/write", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 392, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 392, + 164 + ], + "score": 1.0, + "content": "mechanism has much more flexibility and can learn long-term patterns.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "score": 1.0, + "content": "• Independent Read/write While writing, the ARMIN doesn’t write to the last read memory slot,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 378, + 190 + ], + "score": 1.0, + "content": "but independently choose a memory slot to overwrite according to", + "type": "text" + }, + { + "bbox": [ + 378, + 180, + 389, + 189 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 178, + 408, + 190 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 408, + 178, + 420, + 189 + ], + "score": 0.87, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 178, + 505, + 190 + ], + "score": 1.0, + "content": ", in a similar fashion", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "with the read operation we presented in section 3.2. The result shows a BPC of 1.245, proving that", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "score": 1.0, + "content": "the normal writing regime in the ARMIN can better learn long-term patterns and maximally preserve", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 313, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 313, + 223 + ], + "score": 1.0, + "content": "historical information, as we explain in section 3.3.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "• Smooth Read/write We replace the gumbel-softmax function with the vanilla softmax function,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "so the ARMIN reads and writes in a smooth way. The results show a minor performance drop of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "1.212 BPC. The result shows that our discrete read/write regime is still better. Moreover, the smooth", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 477, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 477, + 272 + ], + "score": 1.0, + "content": "read/write regime would consume much more memory at both training and inference stages.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "• Random Read The ARMIN randomly chooses a historical state while reading. The result shows", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "a BPC of 1.224, compared to the 1.202 BPC of normal ARMIN. It proves that the read regime we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 496, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 496, + 117 + ], + "score": 1.0, + "content": "proposed in section 3.2 is effective and can choose more useful historical state than random read.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 119, + 505, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 118, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 505, + 132 + ], + "score": 1.0, + "content": "• Queue-style Write While writing, the ARMIN follows a fashion of first-in-first-out, i.e. always", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 355, + 142 + ], + "score": 1.0, + "content": "chooses the oldest memory slot to write. It makes the memory", + "type": "text" + }, + { + "bbox": [ + 356, + 131, + 369, + 141 + ], + "score": 0.65, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 131, + 505, + 142 + ], + "score": 1.0, + "content": "stubbornly store the hidden states", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 141, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 504, + 153 + ], + "score": 1.0, + "content": "of the last 20 time-steps. The result shows a BPC of 1.352, which proves our normal read/write", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 392, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 392, + 164 + ], + "score": 1.0, + "content": "mechanism has much more flexibility and can learn long-term patterns.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 118, + 505, + 164 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "score": 1.0, + "content": "• Independent Read/write While writing, the ARMIN doesn’t write to the last read memory slot,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 378, + 190 + ], + "score": 1.0, + "content": "but independently choose a memory slot to overwrite according to", + "type": "text" + }, + { + "bbox": [ + 378, + 180, + 389, + 189 + ], + "score": 0.85, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 178, + 408, + 190 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 408, + 178, + 420, + 189 + ], + "score": 0.87, + "content": "h _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 178, + 505, + 190 + ], + "score": 1.0, + "content": ", in a similar fashion", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "with the read operation we presented in section 3.2. The result shows a BPC of 1.245, proving that", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "score": 1.0, + "content": "the normal writing regime in the ARMIN can better learn long-term patterns and maximally preserve", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 313, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 313, + 223 + ], + "score": 1.0, + "content": "historical information, as we explain in section 3.3.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 167, + 505, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "• Smooth Read/write We replace the gumbel-softmax function with the vanilla softmax function,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "so the ARMIN reads and writes in a smooth way. The results show a minor performance drop of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "1.212 BPC. The result shows that our discrete read/write regime is still better. Moreover, the smooth", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 477, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 477, + 272 + ], + "score": 1.0, + "content": "read/write regime would consume much more memory at both training and inference stages.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 226, + 505, + 272 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/HyMS8iRcK7/HyMS8iRcK7_model.json b/parse/train/HyMS8iRcK7/HyMS8iRcK7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f991eff01179c0b36e5db50296abb58413a4537d --- /dev/null +++ b/parse/train/HyMS8iRcK7/HyMS8iRcK7_model.json @@ -0,0 +1,24914 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 588, + 1301, + 588, + 1301, + 894, + 398, + 894 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1024, + 1402, + 1024, + 1402, + 1240, + 299, + 1240 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1254, + 1404, + 1254, + 1404, + 1773, + 298, + 1773 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1402, + 1788, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 301, + 220, + 1400, + 220, + 1400, + 324, + 301, + 324 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.936 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 302, + 955, + 573, + 955, + 573, + 990, + 302, + 990 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 773, + 519, + 926, + 519, + 926, + 553, + 773, + 553 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.644 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 218.0, + 1405.0, + 218.0, + 1405.0, + 275.0, + 296.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 276.0, + 1404.0, + 276.0, + 1404.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 579.0, + 952.0, + 579.0, + 999.0, + 294.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 518.0, + 932.0, + 518.0, + 932.0, + 556.0, + 769.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 591.0, + 1303.0, + 591.0, + 1303.0, + 624.0, + 395.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 618.0, + 1305.0, + 618.0, + 1305.0, + 658.0, + 392.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 649.0, + 1305.0, + 649.0, + 1305.0, + 687.0, + 393.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 680.0, + 1303.0, + 680.0, + 1303.0, + 718.0, + 392.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 714.0, + 1304.0, + 714.0, + 1304.0, + 746.0, + 394.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 744.0, + 1305.0, + 744.0, + 1305.0, + 776.0, + 392.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 773.0, + 1305.0, + 773.0, + 1305.0, + 805.0, + 393.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 803.0, + 1305.0, + 803.0, + 1305.0, + 836.0, + 394.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 835.0, + 1305.0, + 835.0, + 1305.0, + 868.0, + 395.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 864.0, + 1232.0, + 864.0, + 1232.0, + 895.0, + 394.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1026.0, + 1403.0, + 1026.0, + 1403.0, + 1060.0, + 294.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1054.0, + 1406.0, + 1054.0, + 1406.0, + 1093.0, + 292.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1084.0, + 1406.0, + 1084.0, + 1406.0, + 1123.0, + 292.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 1402.0, + 1116.0, + 1402.0, + 1153.0, + 293.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1143.0, + 1406.0, + 1143.0, + 1406.0, + 1186.0, + 292.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1176.0, + 1406.0, + 1176.0, + 1406.0, + 1217.0, + 292.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1206.0, + 1135.0, + 1206.0, + 1135.0, + 1247.0, + 293.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1292.0, + 294.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1324.0, + 291.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1316.0, + 1402.0, + 1316.0, + 1402.0, + 1352.0, + 293.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1348.0, + 1404.0, + 1348.0, + 1404.0, + 1381.0, + 296.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1378.0, + 1405.0, + 1378.0, + 1405.0, + 1414.0, + 294.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 1404.0, + 1408.0, + 1404.0, + 1443.0, + 293.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1436.0, + 1404.0, + 1436.0, + 1404.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1504.0, + 294.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1499.0, + 1406.0, + 1499.0, + 1406.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1532.0, + 1404.0, + 1532.0, + 1404.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1559.0, + 1406.0, + 1559.0, + 1406.0, + 1597.0, + 293.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1404.0, + 1590.0, + 1404.0, + 1626.0, + 294.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 1404.0, + 1621.0, + 1404.0, + 1657.0, + 294.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 1404.0, + 1650.0, + 1404.0, + 1686.0, + 293.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1404.0, + 1681.0, + 1404.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1709.0, + 1404.0, + 1709.0, + 1404.0, + 1752.0, + 293.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1744.0, + 470.0, + 1744.0, + 470.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1790.0, + 1404.0, + 1790.0, + 1404.0, + 1823.0, + 296.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1406.0, + 1821.0, + 1406.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1404.0, + 1911.0, + 1404.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1981.0, + 292.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1406.0, + 2004.0, + 1406.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 659, + 298, + 659 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1366, + 1405, + 1366, + 1405, + 1583, + 297, + 1583 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 672, + 1398, + 672, + 1398, + 766, + 299, + 766 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 536, + 1597, + 1162, + 1597, + 1162, + 1682, + 536, + 1682 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 290, + 1795, + 1400, + 1795, + 1400, + 1860, + 290, + 1860 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1400, + 1971, + 1400, + 2034, + 296, + 2034 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 598, + 1743, + 1099, + 1743, + 1099, + 1782, + 598, + 1782 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 295, + 1694, + 1216, + 1694, + 1216, + 1728, + 295, + 1728 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 293, + 1133, + 1400, + 1133, + 1400, + 1195, + 293, + 1195 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 299, + 1240, + 557, + 1240, + 557, + 1276, + 299, + 1276 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 293, + 1901, + 1394, + 1901, + 1394, + 1939, + 293, + 1939 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1618, + 1400, + 1618, + 1400, + 1649, + 1366, + 1649 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1747, + 1399, + 1747, + 1399, + 1776, + 1366, + 1776 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 299, + 1310, + 740, + 1310, + 740, + 1341, + 299, + 1341 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.724 + }, + { + "category_id": 1, + "poly": [ + 296, + 795, + 1406, + 795, + 1406, + 1123, + 296, + 1123 + ], + "score": 0.681 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.13 + }, + { + "category_id": 14, + "poly": [ + 597, + 1742, + 1101, + 1742, + 1101, + 1780, + 597, + 1780 + ], + "score": 0.93, + "latex": "g _ { i } = - \\log ( - \\log ( u _ { i } ) ) \\ , u _ { i } \\sim \\mathrm { U n i f o r m } ( 0 , 1 ) ," + }, + { + "category_id": 14, + "poly": [ + 538, + 1595, + 1158, + 1595, + 1158, + 1683, + 538, + 1683 + ], + "score": 0.93, + "latex": "y _ { i } = { \\frac { \\exp ( ( \\log ( \\pi _ { i } ) + g _ { i } ) / \\tau ) } { \\sum _ { j = 1 } ^ { k } \\exp ( ( \\log ( \\pi _ { j } ) + g _ { j } ) / \\tau ) } } \\quad { \\mathrm { f o r ~ } } i = 1 , 2 , . . . , k ," + }, + { + "category_id": 13, + "poly": [ + 677, + 506, + 737, + 506, + 737, + 536, + 677, + 536 + ], + "score": 0.91, + "latex": "h _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 642, + 1520, + 865, + 1520, + 865, + 1552, + 642, + 1552 + ], + "score": 0.91, + "latex": "\\pmb { p } = ( \\pi _ { 1 } , \\pi _ { 2 } , . . . , \\pi _ { k } )" + }, + { + "category_id": 13, + "poly": [ + 1081, + 858, + 1141, + 858, + 1141, + 889, + 1081, + 889 + ], + "score": 0.9, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1702, + 474, + 1702, + 474, + 1728, + 372, + 1728 + ], + "score": 0.9, + "latex": "g _ { 1 } , . . . , g _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1072, + 506, + 1104, + 506, + 1104, + 535, + 1072, + 535 + ], + "score": 0.88, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1296, + 506, + 1329, + 506, + 1329, + 535, + 1296, + 535 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 404, + 509, + 436, + 509, + 436, + 535, + 404, + 535 + ], + "score": 0.87, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1253, + 1061, + 1308, + 1061, + 1308, + 1091, + 1253, + 1091 + ], + "score": 0.87, + "latex": "4 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 710, + 569, + 737, + 569, + 737, + 596, + 710, + 596 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 914, + 479, + 942, + 479, + 942, + 505, + 914, + 505 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 808, + 861, + 840, + 861, + 840, + 887, + 808, + 887 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 424, + 539, + 452, + 539, + 452, + 566, + 424, + 566 + ], + "score": 0.85, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 690, + 1557, + 710, + 1557, + 710, + 1582, + 690, + 1582 + ], + "score": 0.82, + "latex": "\\textbf { { y } }" + }, + { + "category_id": 13, + "poly": [ + 804, + 1555, + 816, + 1555, + 816, + 1578, + 804, + 1578 + ], + "score": 0.8, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 922, + 1521, + 941, + 1521, + 941, + 1547, + 922, + 1547 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 344, + 1802, + 363, + 1802, + 363, + 1823, + 344, + 1823 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1235.0, + 561.0, + 1235.0, + 561.0, + 1284.0, + 291.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1900.0, + 1396.0, + 1900.0, + 1396.0, + 1940.0, + 293.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1309.0, + 744.0, + 1309.0, + 744.0, + 1345.0, + 295.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 230.0, + 1407.0, + 230.0, + 1407.0, + 264.0, + 292.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 325.0, + 293.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 323.0, + 1405.0, + 323.0, + 1405.0, + 355.0, + 293.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 388.0, + 292.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 1407.0, + 383.0, + 1407.0, + 419.0, + 294.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 412.0, + 1408.0, + 412.0, + 1408.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 441.0, + 1407.0, + 441.0, + 1407.0, + 478.0, + 293.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 472.0, + 913.0, + 472.0, + 913.0, + 510.0, + 292.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 472.0, + 1405.0, + 472.0, + 1405.0, + 510.0, + 943.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 506.0, + 403.0, + 506.0, + 403.0, + 538.0, + 294.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 506.0, + 676.0, + 506.0, + 676.0, + 538.0, + 437.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 506.0, + 1071.0, + 506.0, + 1071.0, + 538.0, + 738.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 506.0, + 1295.0, + 506.0, + 1295.0, + 538.0, + 1105.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 506.0, + 1405.0, + 506.0, + 1405.0, + 538.0, + 1330.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 531.0, + 423.0, + 531.0, + 423.0, + 573.0, + 291.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 531.0, + 1407.0, + 531.0, + 1407.0, + 573.0, + 453.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 565.0, + 709.0, + 565.0, + 709.0, + 601.0, + 293.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 565.0, + 1405.0, + 565.0, + 1405.0, + 601.0, + 738.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 593.0, + 1405.0, + 593.0, + 1405.0, + 630.0, + 293.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 625.0, + 1376.0, + 625.0, + 1376.0, + 662.0, + 293.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1364.0, + 1405.0, + 1364.0, + 1405.0, + 1405.0, + 292.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1396.0, + 1403.0, + 1396.0, + 1403.0, + 1434.0, + 292.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1425.0, + 1407.0, + 1425.0, + 1407.0, + 1465.0, + 292.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1494.0, + 295.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1487.0, + 1405.0, + 1487.0, + 1405.0, + 1527.0, + 291.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1519.0, + 641.0, + 1519.0, + 641.0, + 1557.0, + 294.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1519.0, + 921.0, + 1519.0, + 921.0, + 1557.0, + 866.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1519.0, + 1407.0, + 1519.0, + 1407.0, + 1557.0, + 942.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 689.0, + 1550.0, + 689.0, + 1585.0, + 294.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1550.0, + 803.0, + 1550.0, + 803.0, + 1585.0, + 711.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1550.0, + 1192.0, + 1550.0, + 1192.0, + 1585.0, + 817.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 669.0, + 1403.0, + 669.0, + 1403.0, + 711.0, + 293.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1404.0, + 702.0, + 1404.0, + 737.0, + 294.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 735.0, + 1123.0, + 735.0, + 1123.0, + 769.0, + 295.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1791.0, + 343.0, + 1791.0, + 343.0, + 1833.0, + 292.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1791.0, + 1405.0, + 1791.0, + 1405.0, + 1833.0, + 364.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1827.0, + 909.0, + 1827.0, + 909.0, + 1860.0, + 295.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1999.0, + 1405.0, + 1999.0, + 1405.0, + 2041.0, + 291.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1691.0, + 371.0, + 1691.0, + 371.0, + 1733.0, + 293.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1691.0, + 1217.0, + 1691.0, + 1217.0, + 1733.0, + 475.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1127.0, + 1404.0, + 1127.0, + 1404.0, + 1172.0, + 292.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1164.0, + 839.0, + 1164.0, + 839.0, + 1199.0, + 295.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 792.0, + 1404.0, + 792.0, + 1404.0, + 833.0, + 291.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 826.0, + 1404.0, + 826.0, + 1404.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 856.0, + 807.0, + 856.0, + 807.0, + 892.0, + 293.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 856.0, + 1080.0, + 856.0, + 1080.0, + 892.0, + 841.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 856.0, + 1152.0, + 856.0, + 1152.0, + 892.0, + 1142.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 896.0, + 1404.0, + 896.0, + 1404.0, + 932.0, + 294.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 927.0, + 1404.0, + 927.0, + 1404.0, + 964.0, + 294.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 920.0, + 959.0, + 920.0, + 993.0, + 294.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 995.0, + 1406.0, + 995.0, + 1406.0, + 1038.0, + 291.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1031.0, + 1404.0, + 1031.0, + 1404.0, + 1065.0, + 294.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1059.0, + 1252.0, + 1059.0, + 1252.0, + 1099.0, + 293.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1059.0, + 1406.0, + 1059.0, + 1406.0, + 1099.0, + 1309.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1092.0, + 1253.0, + 1092.0, + 1253.0, + 1129.0, + 294.0, + 1129.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1216, + 1403, + 1216, + 1403, + 1372, + 297, + 1372 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1029, + 1404, + 1029, + 1404, + 1124, + 298, + 1124 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1901, + 1404, + 1901, + 1404, + 2036, + 297, + 2036 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 295, + 783, + 1405, + 783, + 1405, + 970, + 295, + 970 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 391, + 254, + 1276, + 254, + 1276, + 771, + 391, + 771 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 586, + 1594, + 1122, + 1594, + 1122, + 1789, + 586, + 1789 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 616, + 1411, + 1090, + 1411, + 1090, + 1490, + 616, + 1490 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 1160, + 781, + 1160, + 781, + 1192, + 300, + 1192 + ], + "score": 0.903 + }, + { + "category_id": 8, + "poly": [ + 625, + 1841, + 1069, + 1841, + 1069, + 1881, + 625, + 1881 + ], + "score": 0.899 + }, + { + "category_id": 8, + "poly": [ + 692, + 1797, + 1005, + 1797, + 1005, + 1833, + 692, + 1833 + ], + "score": 0.895 + }, + { + "category_id": 8, + "poly": [ + 765, + 1549, + 931, + 1549, + 931, + 1585, + 765, + 1585 + ], + "score": 0.887 + }, + { + "category_id": 8, + "poly": [ + 734, + 1501, + 960, + 1501, + 960, + 1539, + 734, + 1539 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1676, + 1401, + 1676, + 1401, + 1706, + 1366, + 1706 + ], + "score": 0.874 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1551, + 1401, + 1551, + 1401, + 1581, + 1366, + 1581 + ], + "score": 0.873 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1435, + 1401, + 1435, + 1401, + 1466, + 1365, + 1466 + ], + "score": 0.872 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1798, + 1400, + 1798, + 1400, + 1830, + 1366, + 1830 + ], + "score": 0.869 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1846, + 1400, + 1846, + 1400, + 1875, + 1366, + 1875 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1505, + 1401, + 1505, + 1401, + 1534, + 1366, + 1534 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2112, + 841, + 2112 + ], + "score": 0.678 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2112, + 841, + 2112 + ], + "score": 0.438 + }, + { + "category_id": 13, + "poly": [ + 1050, + 1281, + 1109, + 1281, + 1109, + 1312, + 1050, + 1312 + ], + "score": 0.92, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1108, + 847, + 1169, + 847, + 1169, + 879, + 1108, + 879 + ], + "score": 0.91, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1002, + 1974, + 1099, + 1974, + 1099, + 2003, + 1002, + 2003 + ], + "score": 0.9, + "latex": "d _ { h } = d _ { r }" + }, + { + "category_id": 13, + "poly": [ + 380, + 1903, + 897, + 1903, + 897, + 1938, + 380, + 1938 + ], + "score": 0.89, + "latex": "h _ { t - 1 } , h _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { \\boldsymbol { r } } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\boldsymbol { x } } _ { t } \\in \\mathbb { R } ^ { d _ { i } }" + }, + { + "category_id": 14, + "poly": [ + 580, + 1593, + 1121, + 1593, + 1121, + 1789, + 580, + 1789 + ], + "score": 0.89, + "latex": "\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } ^ { h } } \\\\ { o _ { t } ^ { r } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W _ { g o } [ \\boldsymbol { x } _ { t } , h _ { t - 1 } ^ { g } , \\boldsymbol { r } _ { t } ^ { g } ] + b _ { g o } ," + }, + { + "category_id": 14, + "poly": [ + 612, + 1408, + 1089, + 1408, + 1089, + 1489, + 612, + 1489 + ], + "score": 0.89, + "latex": "\\left\\{ { g } _ { t } ^ { h } \\atop { g } _ { t } ^ { r } \\right\\} = \\left\\{ { \\sigma } \\atop { \\sigma } \\right\\} W _ { i g } [ x _ { t } , h _ { t - 1 } , r _ { t } ] + b _ { i g } ," + }, + { + "category_id": 13, + "poly": [ + 975, + 1902, + 1403, + 1902, + 1403, + 1940, + 975, + 1940 + ], + "score": 0.89, + "latex": "\\begin{array} { r c l } { W _ { i g } } & { \\in } & { \\mathbb { R } ^ { 2 d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , W _ { g o } } & { \\in } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 299, + 1939, + 1111, + 1939, + 1111, + 1977, + 299, + 1977 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\mathbb { R } ^ { ( 4 d _ { h } + d _ { r } ) \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } _ { i g } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } , \\pmb { b } _ { g o } \\ \\in \\ \\mathbb { R } ^ { 4 d _ { h } + d _ { r } } , \\pmb { o } _ { t } \\ \\in \\ \\mathbb { R } ^ { d _ { h } + d _ { r } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 577, + 2004, + 608, + 2004, + 608, + 2034, + 577, + 2034 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1017, + 909, + 1048, + 909, + 1048, + 938, + 1017, + 938 + ], + "score": 0.88, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1944, + 1319, + 1944, + 1319, + 1973, + 1287, + 1973 + ], + "score": 0.87, + "latex": "d _ { h }" + }, + { + "category_id": 13, + "poly": [ + 355, + 1341, + 386, + 1341, + 386, + 1371, + 355, + 1371 + ], + "score": 0.87, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 782, + 1284, + 813, + 1284, + 813, + 1310, + 782, + 1310 + ], + "score": 0.87, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1314, + 1171, + 1314, + 1171, + 1341, + 1143, + 1341 + ], + "score": 0.86, + "latex": "\\mathbf { } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 296, + 939, + 328, + 939, + 328, + 969, + 296, + 969 + ], + "score": 0.86, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1023, + 851, + 1055, + 851, + 1055, + 877, + 1023, + 877 + ], + "score": 0.86, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 739, + 942, + 767, + 942, + 767, + 969, + 739, + 969 + ], + "score": 0.86, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 859, + 584, + 897, + 584, + 897, + 612, + 859, + 612 + ], + "score": 0.85, + "latex": "h _ { t _ { 1 } }" + }, + { + "category_id": 13, + "poly": [ + 659, + 2007, + 686, + 2007, + 686, + 2034, + 659, + 2034 + ], + "score": 0.85, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { r } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 510, + 912, + 539, + 912, + 539, + 938, + 510, + 938 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 591, + 909, + 622, + 909, + 622, + 938, + 591, + 938 + ], + "score": 0.85, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 440, + 1313, + 469, + 1313, + 469, + 1341, + 440, + 1341 + ], + "score": 0.84, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 882, + 1369, + 882, + 1369, + 909, + 1340, + 909 + ], + "score": 0.84, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 637, + 911, + 667, + 911, + 667, + 938, + 637, + 938 + ], + "score": 0.82, + "latex": "\\mathbf { } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1373, + 821, + 1400, + 821, + 1400, + 847, + 1373, + 847 + ], + "score": 0.82, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1224, + 878, + 1285, + 878, + 1285, + 910, + 1224, + 910 + ], + "score": 0.8, + "latex": "h _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 860, + 621, + 897, + 621, + 897, + 648, + 860, + 648 + ], + "score": 0.79, + "latex": "h _ { t _ { 2 } }" + }, + { + "category_id": 14, + "poly": [ + 734, + 1498, + 961, + 1498, + 961, + 1541, + 734, + 1541 + ], + "score": 0.71, + "latex": "h _ { t - 1 } ^ { g } = { \\pmb g } _ { t } ^ { h } \\circ { \\pmb h } _ { t - 1 } ," + }, + { + "category_id": 14, + "poly": [ + 767, + 1547, + 931, + 1547, + 931, + 1585, + 767, + 1585 + ], + "score": 0.66, + "latex": "\\begin{array} { r } { \\pmb { r } _ { t } ^ { g } = \\pmb { g } _ { t } ^ { r } \\circ \\pmb { r } _ { t } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1179, + 882, + 1211, + 882, + 1211, + 909, + 1179, + 909 + ], + "score": 0.66, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 14, + "poly": [ + 691, + 1796, + 1002, + 1796, + 1002, + 1831, + 691, + 1831 + ], + "score": 0.65, + "latex": "\\pmb { h } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { h } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ," + }, + { + "category_id": 13, + "poly": [ + 859, + 692, + 898, + 692, + 898, + 723, + 859, + 723 + ], + "score": 0.56, + "latex": "h _ { t _ { n } }" + }, + { + "category_id": 14, + "poly": [ + 624, + 1839, + 1076, + 1839, + 1076, + 1881, + 624, + 1881 + ], + "score": 0.49, + "latex": "\\begin{array} { r } { \\pmb { o } _ { t } = \\left[ \\pmb { o } _ { t } ^ { h } \\circ \\operatorname { t a n h } ( \\pmb { h } _ { t } ) , \\pmb { o } _ { t } ^ { r } \\circ \\operatorname { t a n h } ( \\pmb { r } _ { t } ) \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 865, + 261, + 895, + 261, + 895, + 289, + 865, + 289 + ], + "score": 0.37, + "latex": "\\mathbf { \\delta } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 863, + 501, + 893, + 501, + 893, + 526, + 863, + 526 + ], + "score": 0.32, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1299, + 1904, + 1403, + 1904, + 1403, + 1941, + 1299, + 1941 + ], + "score": 0.3, + "latex": "W _ { g o } \\in" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 1405.0, + 784.0, + 1405.0, + 820.0, + 294.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 815.0, + 1372.0, + 815.0, + 1372.0, + 853.0, + 292.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 815.0, + 1406.0, + 815.0, + 1406.0, + 853.0, + 1401.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 846.0, + 1022.0, + 846.0, + 1022.0, + 882.0, + 293.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 846.0, + 1107.0, + 846.0, + 1107.0, + 882.0, + 1056.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 846.0, + 1406.0, + 846.0, + 1406.0, + 882.0, + 1170.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 1178.0, + 877.0, + 1178.0, + 912.0, + 294.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 877.0, + 1223.0, + 877.0, + 1223.0, + 912.0, + 1212.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 877.0, + 1339.0, + 877.0, + 1339.0, + 912.0, + 1286.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 877.0, + 1405.0, + 877.0, + 1405.0, + 912.0, + 1370.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 907.0, + 509.0, + 907.0, + 509.0, + 943.0, + 293.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 907.0, + 590.0, + 907.0, + 590.0, + 943.0, + 540.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 907.0, + 636.0, + 907.0, + 636.0, + 943.0, + 623.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 907.0, + 1016.0, + 907.0, + 1016.0, + 943.0, + 668.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 907.0, + 1406.0, + 907.0, + 1406.0, + 943.0, + 1049.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 937.0, + 738.0, + 937.0, + 738.0, + 975.0, + 329.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 937.0, + 778.0, + 937.0, + 778.0, + 975.0, + 768.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 255.0, + 864.0, + 255.0, + 864.0, + 299.0, + 859.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 255.0, + 902.0, + 255.0, + 902.0, + 299.0, + 896.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 289.0, + 892.0, + 289.0, + 892.0, + 320.0, + 867.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 336.0, + 473.0, + 336.0, + 473.0, + 376.0, + 412.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 367.0, + 504.0, + 367.0, + 504.0, + 408.0, + 388.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 385.0, + 704.0, + 385.0, + 704.0, + 436.0, + 650.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 358.0, + 927.0, + 358.0, + 927.0, + 424.0, + 826.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 367.0, + 1166.0, + 367.0, + 1166.0, + 418.0, + 1033.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 358.0, + 1252.0, + 358.0, + 1252.0, + 396.0, + 1187.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 421.0, + 690.0, + 421.0, + 690.0, + 456.0, + 657.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 466.0, + 704.0, + 466.0, + 704.0, + 493.0, + 575.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 445.0, + 895.0, + 445.0, + 895.0, + 493.0, + 861.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 484.0, + 676.0, + 484.0, + 676.0, + 515.0, + 608.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 490.0, + 862.0, + 490.0, + 862.0, + 532.0, + 856.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 490.0, + 901.0, + 490.0, + 901.0, + 532.0, + 894.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 494.0, + 1158.0, + 494.0, + 1158.0, + 535.0, + 1087.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 524.0, + 897.0, + 524.0, + 897.0, + 574.0, + 862.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 568.0, + 544.0, + 568.0, + 544.0, + 606.0, + 455.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 573.0, + 858.0, + 573.0, + 858.0, + 621.0, + 853.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 573.0, + 906.0, + 573.0, + 906.0, + 621.0, + 898.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 597.0, + 644.0, + 597.0, + 644.0, + 700.0, + 606.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 605.0, + 676.0, + 605.0, + 676.0, + 693.0, + 638.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 610.0, + 859.0, + 610.0, + 859.0, + 657.0, + 852.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 610.0, + 907.0, + 610.0, + 907.0, + 657.0, + 898.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 630.0, + 718.0, + 630.0, + 718.0, + 671.0, + 679.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 659.0, + 900.0, + 659.0, + 900.0, + 680.0, + 851.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 739.0, + 929.0, + 739.0, + 929.0, + 781.0, + 828.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 351.5, + 643.0, + 351.5, + 643.0, + 391.5, + 515.0, + 391.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 388.5, + 1276.0, + 388.5, + 1276.0, + 425.5, + 1165.0, + 425.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.75, + 688.5, + 900.75, + 688.5, + 900.75, + 727.0, + 855.75, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1160.0, + 783.0, + 1160.0, + 783.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1217.0, + 1405.0, + 1217.0, + 1405.0, + 1254.0, + 295.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1249.0, + 1405.0, + 1249.0, + 1405.0, + 1283.0, + 292.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1281.0, + 781.0, + 1281.0, + 781.0, + 1314.0, + 295.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1281.0, + 1049.0, + 1281.0, + 1049.0, + 1314.0, + 814.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1281.0, + 1404.0, + 1281.0, + 1404.0, + 1314.0, + 1110.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1309.0, + 439.0, + 1309.0, + 439.0, + 1346.0, + 292.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1309.0, + 1142.0, + 1309.0, + 1142.0, + 1346.0, + 470.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1309.0, + 1405.0, + 1309.0, + 1405.0, + 1346.0, + 1172.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1342.0, + 354.0, + 1342.0, + 354.0, + 1375.0, + 295.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1342.0, + 835.0, + 1342.0, + 835.0, + 1375.0, + 387.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 1404.0, + 1029.0, + 1404.0, + 1067.0, + 293.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1061.0, + 1405.0, + 1061.0, + 1405.0, + 1097.0, + 292.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1091.0, + 1284.0, + 1091.0, + 1284.0, + 1128.0, + 292.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1895.0, + 379.0, + 1895.0, + 379.0, + 1946.0, + 290.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1895.0, + 974.0, + 1895.0, + 974.0, + 1946.0, + 898.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1895.0, + 1411.0, + 1895.0, + 1411.0, + 1946.0, + 1404.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1929.0, + 298.0, + 1929.0, + 298.0, + 1982.0, + 289.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1929.0, + 1286.0, + 1929.0, + 1286.0, + 1982.0, + 1112.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1929.0, + 1409.0, + 1929.0, + 1409.0, + 1982.0, + 1320.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1967.0, + 1001.0, + 1967.0, + 1001.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1967.0, + 1406.0, + 1967.0, + 1406.0, + 2009.0, + 1100.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 576.0, + 1999.0, + 576.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1999.0, + 658.0, + 1999.0, + 658.0, + 2038.0, + 609.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1999.0, + 697.0, + 1999.0, + 697.0, + 2038.0, + 687.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1388, + 1404, + 1388, + 1404, + 1633, + 297, + 1633 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 536, + 297, + 536 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1727, + 1404, + 1727, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 749, + 1403, + 749, + 1403, + 935, + 298, + 935 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 551, + 1403, + 551, + 1403, + 736, + 297, + 736 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1027, + 1402, + 1027, + 1402, + 1122, + 299, + 1122 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1309, + 1399, + 1309, + 1399, + 1375, + 300, + 1375 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 715, + 1212, + 982, + 1212, + 982, + 1299, + 715, + 1299 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 301, + 1671, + 590, + 1671, + 590, + 1702, + 301, + 1702 + ], + "score": 0.911 + }, + { + "category_id": 8, + "poly": [ + 587, + 1165, + 1115, + 1165, + 1115, + 1202, + 587, + 1202 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1239, + 1399, + 1239, + 1399, + 1269, + 1353, + 1269 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 303, + 972, + 889, + 972, + 889, + 1003, + 303, + 1003 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1168, + 1399, + 1168, + 1399, + 1198, + 1366, + 1198 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.762 + }, + { + "category_id": 13, + "poly": [ + 750, + 750, + 838, + 750, + 838, + 784, + 750, + 784 + ], + "score": 0.93, + "latex": "\\left( \\boldsymbol { h } _ { t } , \\boldsymbol { c } _ { t } \\right)" + }, + { + "category_id": 13, + "poly": [ + 710, + 260, + 770, + 260, + 770, + 296, + 710, + 296 + ], + "score": 0.92, + "latex": "h _ { t - 1 } ^ { g }" + }, + { + "category_id": 13, + "poly": [ + 1178, + 232, + 1235, + 232, + 1235, + 263, + 1178, + 263 + ], + "score": 0.92, + "latex": "h _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 479, + 1481, + 539, + 1481, + 539, + 1512, + 479, + 1512 + ], + "score": 0.92, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 767, + 1027, + 962, + 1027, + 962, + 1058, + 767, + 1058 + ], + "score": 0.91, + "latex": "M \\in \\mathbb { R } ^ { n _ { m e m } \\times d _ { h } }" + }, + { + "category_id": 13, + "poly": [ + 468, + 552, + 529, + 552, + 529, + 584, + 468, + 584 + ], + "score": 0.91, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1033, + 383, + 1093, + 383, + 1093, + 415, + 1033, + 415 + ], + "score": 0.91, + "latex": "h _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 806, + 1343, + 895, + 1343, + 895, + 1376, + 806, + 1376 + ], + "score": 0.91, + "latex": "M ( i , : )" + }, + { + "category_id": 13, + "poly": [ + 404, + 1343, + 460, + 1343, + 460, + 1376, + 404, + 1376 + ], + "score": 0.91, + "latex": "s _ { t } ( i )" + }, + { + "category_id": 13, + "poly": [ + 970, + 320, + 1003, + 320, + 1003, + 354, + 970, + 354 + ], + "score": 0.9, + "latex": "\\mathbf { \\sigma } _ { o _ { t } ^ { h } } ^ { h }" + }, + { + "category_id": 13, + "poly": [ + 730, + 1091, + 790, + 1091, + 790, + 1123, + 730, + 1123 + ], + "score": 0.9, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 14, + "poly": [ + 582, + 1160, + 1116, + 1160, + 1116, + 1302, + 582, + 1302 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { \\displaystyle s _ { t } = \\mathrm { g u m b e l - s o f t m a x } \\big ( W _ { s } [ \\ : x _ { t } \\ : , h _ { t - 1 } \\ : ] + b _ { s } \\big ) \\ : , } } \\\\ { { { } } } \\\\ { { \\displaystyle r _ { t } = \\sum _ { i = 0 } ^ { n - 1 } s _ { t } ( i ) M ( i , : ) \\ : . } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 795, + 384, + 827, + 384, + 827, + 413, + 795, + 413 + ], + "score": 0.89, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 914, + 355, + 945, + 355, + 945, + 386, + 914, + 386 + ], + "score": 0.89, + "latex": "{ \\mathbf { } } _ { o _ { t } ^ { r } }" + }, + { + "category_id": 13, + "poly": [ + 995, + 260, + 1026, + 260, + 1026, + 294, + 995, + 294 + ], + "score": 0.89, + "latex": "\\boldsymbol { r } _ { t } ^ { g }" + }, + { + "category_id": 13, + "poly": [ + 1195, + 1034, + 1265, + 1034, + 1265, + 1060, + 1195, + 1060 + ], + "score": 0.89, + "latex": "n _ { m e m }" + }, + { + "category_id": 13, + "poly": [ + 731, + 1822, + 763, + 1822, + 763, + 1851, + 731, + 1851 + ], + "score": 0.88, + "latex": "d _ { h }" + }, + { + "category_id": 13, + "poly": [ + 1233, + 752, + 1265, + 752, + 1265, + 781, + 1233, + 781 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 917, + 1821, + 947, + 1821, + 947, + 1851, + 917, + 1851 + ], + "score": 0.88, + "latex": "d _ { r }" + }, + { + "category_id": 13, + "poly": [ + 1338, + 1761, + 1370, + 1761, + 1370, + 1790, + 1338, + 1790 + ], + "score": 0.88, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1730, + 1324, + 1730, + 1324, + 1760, + 1292, + 1760 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 1791, + 1102, + 1791, + 1102, + 1820, + 1070, + 1820 + ], + "score": 0.88, + "latex": "d _ { h }" + }, + { + "category_id": 13, + "poly": [ + 1106, + 1912, + 1137, + 1912, + 1137, + 1942, + 1106, + 1942 + ], + "score": 0.88, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 1513, + 1149, + 1513, + 1149, + 1542, + 1116, + 1542 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 635, + 1821, + 666, + 1821, + 666, + 1851, + 635, + 1851 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1273, + 1791, + 1301, + 1791, + 1301, + 1820, + 1273, + 1820 + ], + "score": 0.88, + "latex": "d _ { r }" + }, + { + "category_id": 13, + "poly": [ + 375, + 1308, + 637, + 1308, + 637, + 1343, + 375, + 1343 + ], + "score": 0.87, + "latex": "W _ { s } \\ \\in \\ \\mathbb { R } ^ { n _ { m e m } \\times ( d _ { i } + d _ { h } ) }" + }, + { + "category_id": 13, + "poly": [ + 616, + 323, + 644, + 323, + 644, + 353, + 616, + 353 + ], + "score": 0.87, + "latex": "\\pmb { f } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 841, + 783, + 871, + 783, + 871, + 812, + 841, + 812 + ], + "score": 0.87, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1200, + 584, + 1232, + 584, + 1232, + 613, + 1200, + 613 + ], + "score": 0.87, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 903, + 1913, + 934, + 1913, + 934, + 1942, + 903, + 1942 + ], + "score": 0.87, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 968, + 445, + 1000, + 445, + 1000, + 474, + 968, + 474 + ], + "score": 0.87, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 698, + 674, + 730, + 674, + 730, + 704, + 698, + 704 + ], + "score": 0.87, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 431, + 232, + 499, + 232, + 499, + 260, + 431, + 260 + ], + "score": 0.87, + "latex": "3 \\sim 5" + }, + { + "category_id": 13, + "poly": [ + 467, + 415, + 496, + 415, + 496, + 445, + 467, + 445 + ], + "score": 0.86, + "latex": "\\pmb { f } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 675, + 415, + 699, + 415, + 699, + 444, + 675, + 444 + ], + "score": 0.86, + "latex": "\\dot { \\mathbf { \\delta } } _ { \\mathbf { \\delta } }" + }, + { + "category_id": 13, + "poly": [ + 738, + 587, + 769, + 587, + 769, + 613, + 738, + 613 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 648, + 1094, + 680, + 1094, + 680, + 1121, + 648, + 1121 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1104, + 787, + 1134, + 787, + 1134, + 812, + 1104, + 812 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 759, + 326, + 788, + 326, + 788, + 354, + 759, + 354 + ], + "score": 0.85, + "latex": "\\mathbf { \\pmb { g } } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1034, + 1516, + 1065, + 1516, + 1065, + 1542, + 1034, + 1542 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1118, + 847, + 1146, + 847, + 1146, + 873, + 1118, + 873 + ], + "score": 0.85, + "latex": "\\mathbf { } _ \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf " + }, + { + "category_id": 13, + "poly": [ + 454, + 323, + 479, + 323, + 479, + 353, + 454, + 353 + ], + "score": 0.85, + "latex": "\\mathbf { \\delta } _ { i _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1158, + 356, + 1187, + 356, + 1187, + 383, + 1158, + 383 + ], + "score": 0.85, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1064, + 1243, + 1064, + 1243, + 1091, + 1217, + 1091 + ], + "score": 0.85, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 962, + 557, + 990, + 557, + 990, + 582, + 962, + 582 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 327, + 1946, + 354, + 1946, + 354, + 1973, + 327, + 1973 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1052, + 447, + 1080, + 447, + 1080, + 474, + 1052, + 474 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 606, + 616, + 634, + 616, + 634, + 643, + 606, + 643 + ], + "score": 0.84, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 653, + 1310, + 805, + 1310, + 805, + 1343, + 653, + 1343 + ], + "score": 0.84, + "latex": "b _ { s } \\in \\mathbb { R } ^ { n _ { m e m } }" + }, + { + "category_id": 13, + "poly": [ + 1250, + 387, + 1279, + 387, + 1279, + 415, + 1250, + 415 + ], + "score": 0.84, + "latex": "\\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } } \\mathbf { \\sigma } _ { \\mathbf { \\sigma } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\mathbf { \\lambda } } } \\mathbf { \\sigma } _ { \\mathbf { \\lambda } _ { \\lambda } }" + }, + { + "category_id": 13, + "poly": [ + 789, + 1793, + 818, + 1793, + 818, + 1820, + 789, + 1820 + ], + "score": 0.84, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1160, + 479, + 1187, + 479, + 1187, + 505, + 1160, + 505 + ], + "score": 0.84, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 479, + 1361, + 479, + 1361, + 505, + 1333, + 505 + ], + "score": 0.83, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1287, + 236, + 1313, + 236, + 1313, + 261, + 1287, + 261 + ], + "score": 0.83, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 360, + 678, + 390, + 678, + 390, + 704, + 360, + 704 + ], + "score": 0.83, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 846, + 1916, + 872, + 1916, + 872, + 1942, + 846, + 1942 + ], + "score": 0.83, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 617, + 326, + 617, + 326, + 643, + 297, + 643 + ], + "score": 0.82, + "latex": "\\mathbf { } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 296, + 504, + 330, + 504, + 330, + 539, + 296, + 539 + ], + "score": 0.82, + "latex": "\\boldsymbol { r } _ { t } ^ { g }" + }, + { + "category_id": 13, + "poly": [ + 1158, + 1344, + 1195, + 1344, + 1195, + 1370, + 1158, + 1370 + ], + "score": 0.81, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 596, + 1346, + 609, + 1346, + 609, + 1370, + 596, + 1370 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 768, + 1347, + 795, + 1347, + 795, + 1373, + 768, + 1373 + ], + "score": 0.75, + "latex": "\\mathbf { \\boldsymbol { s } } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1031, + 1345, + 1044, + 1345, + 1044, + 1370, + 1031, + 1370 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 820, + 1315, + 848, + 1315, + 848, + 1343, + 820, + 1343 + ], + "score": 0.73, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { S } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 351, + 873, + 386, + 873, + 386, + 904, + 351, + 904 + ], + "score": 0.7, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 399, + 1760, + 436, + 1760, + 436, + 1787, + 399, + 1787 + ], + "score": 0.59, + "latex": "M" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1669.0, + 596.0, + 1669.0, + 596.0, + 1705.0, + 295.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 971.0, + 893.0, + 971.0, + 893.0, + 1006.0, + 296.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 1404.0, + 1390.0, + 1404.0, + 1423.0, + 294.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 1405.0, + 1418.0, + 1405.0, + 1457.0, + 292.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1449.0, + 1405.0, + 1449.0, + 1405.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1481.0, + 478.0, + 1481.0, + 478.0, + 1514.0, + 295.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1481.0, + 1404.0, + 1481.0, + 1404.0, + 1514.0, + 540.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1509.0, + 1033.0, + 1509.0, + 1033.0, + 1548.0, + 294.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1509.0, + 1115.0, + 1509.0, + 1115.0, + 1548.0, + 1066.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1509.0, + 1404.0, + 1509.0, + 1404.0, + 1548.0, + 1150.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1539.0, + 1406.0, + 1539.0, + 1406.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 1407.0, + 1572.0, + 1407.0, + 1609.0, + 294.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1601.0, + 516.0, + 1601.0, + 516.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 430.0, + 230.0, + 430.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 230.0, + 1177.0, + 230.0, + 1177.0, + 265.0, + 500.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 230.0, + 1286.0, + 230.0, + 1286.0, + 265.0, + 1236.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 265.0, + 1314.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 255.0, + 709.0, + 255.0, + 709.0, + 301.0, + 291.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 255.0, + 994.0, + 255.0, + 994.0, + 301.0, + 771.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 255.0, + 1408.0, + 255.0, + 1408.0, + 301.0, + 1027.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 327.0, + 295.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 453.0, + 320.0, + 453.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 320.0, + 615.0, + 320.0, + 615.0, + 356.0, + 480.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 320.0, + 758.0, + 320.0, + 758.0, + 356.0, + 645.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 320.0, + 969.0, + 320.0, + 969.0, + 356.0, + 789.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 320.0, + 1406.0, + 320.0, + 1406.0, + 356.0, + 1004.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 351.0, + 913.0, + 351.0, + 913.0, + 386.0, + 295.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 351.0, + 1157.0, + 351.0, + 1157.0, + 386.0, + 946.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 386.0, + 1188.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 382.0, + 794.0, + 382.0, + 794.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 382.0, + 1032.0, + 382.0, + 1032.0, + 418.0, + 828.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 382.0, + 1249.0, + 382.0, + 1249.0, + 418.0, + 1094.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 418.0, + 1280.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 466.0, + 413.0, + 466.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 413.0, + 674.0, + 413.0, + 674.0, + 449.0, + 497.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 449.0, + 700.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 967.0, + 444.0, + 967.0, + 478.0, + 292.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 444.0, + 1051.0, + 444.0, + 1051.0, + 478.0, + 1001.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 444.0, + 1405.0, + 444.0, + 1405.0, + 478.0, + 1081.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 470.0, + 1159.0, + 470.0, + 1159.0, + 511.0, + 292.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 470.0, + 1332.0, + 470.0, + 1332.0, + 511.0, + 1188.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 470.0, + 1407.0, + 470.0, + 1407.0, + 511.0, + 1362.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 503.0, + 641.0, + 503.0, + 641.0, + 540.0, + 331.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 1291.0, + 1728.0, + 1291.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1763.0, + 1325.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1760.0, + 398.0, + 1760.0, + 398.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1760.0, + 1337.0, + 1760.0, + 1337.0, + 1795.0, + 437.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1795.0, + 1371.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 788.0, + 1788.0, + 788.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1788.0, + 1069.0, + 1788.0, + 1069.0, + 1825.0, + 819.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1788.0, + 1272.0, + 1788.0, + 1272.0, + 1825.0, + 1103.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1788.0, + 1406.0, + 1788.0, + 1406.0, + 1825.0, + 1302.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 634.0, + 1817.0, + 634.0, + 1859.0, + 292.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1817.0, + 730.0, + 1817.0, + 730.0, + 1859.0, + 667.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1817.0, + 916.0, + 1817.0, + 916.0, + 1859.0, + 764.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1817.0, + 1408.0, + 1817.0, + 1408.0, + 1859.0, + 948.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1406.0, + 1881.0, + 1406.0, + 1916.0, + 292.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 845.0, + 1910.0, + 845.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1910.0, + 902.0, + 1910.0, + 902.0, + 1947.0, + 873.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1910.0, + 1105.0, + 1910.0, + 1105.0, + 1947.0, + 935.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1947.0, + 1138.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 326.0, + 1942.0, + 326.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1978.0, + 355.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 748.0, + 749.0, + 748.0, + 749.0, + 786.0, + 293.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 748.0, + 1232.0, + 748.0, + 1232.0, + 786.0, + 839.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 748.0, + 1405.0, + 748.0, + 1405.0, + 786.0, + 1266.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 840.0, + 781.0, + 840.0, + 815.0, + 294.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 781.0, + 1103.0, + 781.0, + 1103.0, + 815.0, + 872.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 781.0, + 1405.0, + 781.0, + 1405.0, + 815.0, + 1135.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 809.0, + 1407.0, + 809.0, + 1407.0, + 847.0, + 293.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 843.0, + 1117.0, + 843.0, + 1117.0, + 875.0, + 296.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 843.0, + 1404.0, + 843.0, + 1404.0, + 875.0, + 1147.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 871.0, + 350.0, + 871.0, + 350.0, + 906.0, + 293.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 906.0, + 387.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 902.0, + 1395.0, + 902.0, + 1395.0, + 936.0, + 293.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 549.0, + 467.0, + 549.0, + 467.0, + 588.0, + 292.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 549.0, + 961.0, + 549.0, + 961.0, + 588.0, + 530.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 549.0, + 1405.0, + 549.0, + 1405.0, + 588.0, + 991.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 737.0, + 580.0, + 737.0, + 618.0, + 294.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 580.0, + 1199.0, + 580.0, + 1199.0, + 618.0, + 770.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 580.0, + 1405.0, + 580.0, + 1405.0, + 618.0, + 1233.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 610.0, + 296.0, + 610.0, + 296.0, + 651.0, + 292.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 610.0, + 605.0, + 610.0, + 605.0, + 651.0, + 327.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 610.0, + 1407.0, + 610.0, + 1407.0, + 651.0, + 635.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 640.0, + 1405.0, + 640.0, + 1405.0, + 678.0, + 292.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 670.0, + 359.0, + 670.0, + 359.0, + 711.0, + 294.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 670.0, + 697.0, + 670.0, + 697.0, + 711.0, + 391.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 670.0, + 1404.0, + 670.0, + 1404.0, + 711.0, + 731.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 703.0, + 747.0, + 703.0, + 747.0, + 739.0, + 295.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 766.0, + 1022.0, + 766.0, + 1066.0, + 292.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1022.0, + 1194.0, + 1022.0, + 1194.0, + 1066.0, + 963.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1022.0, + 1407.0, + 1022.0, + 1407.0, + 1066.0, + 1266.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1055.0, + 1216.0, + 1055.0, + 1216.0, + 1098.0, + 293.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1055.0, + 1406.0, + 1055.0, + 1406.0, + 1098.0, + 1244.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1090.0, + 647.0, + 1090.0, + 647.0, + 1124.0, + 295.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1090.0, + 729.0, + 1090.0, + 729.0, + 1124.0, + 681.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1090.0, + 1159.0, + 1090.0, + 1159.0, + 1124.0, + 791.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1302.0, + 374.0, + 1302.0, + 374.0, + 1350.0, + 293.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1302.0, + 652.0, + 1302.0, + 652.0, + 1350.0, + 638.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1302.0, + 819.0, + 1302.0, + 819.0, + 1350.0, + 806.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1302.0, + 1405.0, + 1302.0, + 1405.0, + 1350.0, + 849.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1341.0, + 403.0, + 1341.0, + 403.0, + 1377.0, + 296.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1341.0, + 595.0, + 1341.0, + 595.0, + 1377.0, + 461.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1341.0, + 767.0, + 1341.0, + 767.0, + 1377.0, + 610.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1341.0, + 805.0, + 1341.0, + 805.0, + 1377.0, + 796.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1341.0, + 1030.0, + 1341.0, + 1030.0, + 1377.0, + 896.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1341.0, + 1157.0, + 1341.0, + 1157.0, + 1377.0, + 1045.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1341.0, + 1209.0, + 1341.0, + 1209.0, + 1377.0, + 1196.0, + 1377.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 338, + 1763, + 1363, + 1763, + 1363, + 2016, + 338, + 2016 + ], + "score": 0.985, + "html": "
CopyRepeat CopyAssociative RecallPriority Sort
timeiter.timeiter.timeiter.timeiter.
LSTM660.359450.0160.3253237.0
NTM3212.41710.01419.63600.012
TARDIS1570.451970.1663422660.33013362.6
ARMIN+TARDIS-addr1520.410950.297640.33615171.4
ARMIN67.63667.8330.0523733.4
" + }, + { + "category_id": 1, + "poly": [ + 298, + 712, + 1404, + 712, + 1404, + 1353, + 298, + 1353 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 434, + 1404, + 434, + 1404, + 618, + 298, + 618 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1369, + 1403, + 1369, + 1403, + 1583, + 298, + 1583 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 323, + 299, + 323 + ], + "score": 0.957 + }, + { + "category_id": 6, + "poly": [ + 300, + 1630, + 1400, + 1630, + 1400, + 1724, + 300, + 1724 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 301, + 366, + 557, + 366, + 557, + 402, + 301, + 402 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 300, + 657, + 622, + 657, + 622, + 686, + 300, + 686 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2113, + 840, + 2113 + ], + "score": 0.747 + }, + { + "category_id": 13, + "poly": [ + 352, + 1078, + 407, + 1078, + 407, + 1109, + 352, + 1109 + ], + "score": 0.88, + "latex": "3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 500, + 1492, + 582, + 1492, + 582, + 1520, + 500, + 1520 + ], + "score": 0.88, + "latex": "3 \\sim 4" + }, + { + "category_id": 13, + "poly": [ + 994, + 1522, + 1053, + 1522, + 1053, + 1551, + 994, + 1551 + ], + "score": 0.62, + "latex": "1 0 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 390, + 1525, + 409, + 1525, + 409, + 1549, + 390, + 1549 + ], + "score": 0.54, + "latex": "\\dotplus" + }, + { + "category_id": 13, + "poly": [ + 916, + 1632, + 949, + 1632, + 949, + 1661, + 916, + 1661 + ], + "score": 0.49, + "latex": "{ \\mathrm { ( k ) } }" + }, + { + "category_id": 13, + "poly": [ + 835, + 563, + 852, + 563, + 852, + 585, + 835, + 585 + ], + "score": 0.44, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 366, + 1662, + 426, + 1662, + 426, + 1691, + 366, + 1691 + ], + "score": 0.35, + "latex": "1 0 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 492, + 1262, + 551, + 1262, + 551, + 1291, + 492, + 1291 + ], + "score": 0.32, + "latex": "1 0 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1630.0, + 915.0, + 1630.0, + 915.0, + 1664.0, + 295.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1664.0, + 950.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1660.0, + 365.0, + 1660.0, + 365.0, + 1696.0, + 294.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1660.0, + 1405.0, + 1660.0, + 1405.0, + 1696.0, + 427.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1693.0, + 1096.0, + 1693.0, + 1096.0, + 1723.0, + 297.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 364.0, + 561.0, + 364.0, + 561.0, + 406.0, + 292.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 655.0, + 626.0, + 655.0, + 626.0, + 690.0, + 295.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 715.0, + 1402.0, + 715.0, + 1402.0, + 747.0, + 297.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 746.0, + 1405.0, + 746.0, + 1405.0, + 778.0, + 296.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 773.0, + 1406.0, + 773.0, + 1406.0, + 811.0, + 293.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 806.0, + 1405.0, + 806.0, + 1405.0, + 838.0, + 296.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 832.0, + 1406.0, + 832.0, + 1406.0, + 874.0, + 292.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 865.0, + 1404.0, + 865.0, + 1404.0, + 901.0, + 294.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 897.0, + 1404.0, + 897.0, + 1404.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 926.0, + 1406.0, + 926.0, + 1406.0, + 962.0, + 294.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 953.0, + 1408.0, + 953.0, + 1408.0, + 995.0, + 291.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 985.0, + 1406.0, + 985.0, + 1406.0, + 1022.0, + 293.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 1405.0, + 1016.0, + 1405.0, + 1053.0, + 293.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 1405.0, + 1048.0, + 1405.0, + 1084.0, + 294.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 351.0, + 1076.0, + 351.0, + 1114.0, + 293.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1114.0, + 408.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1108.0, + 1404.0, + 1108.0, + 1404.0, + 1143.0, + 294.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1141.0, + 1404.0, + 1141.0, + 1404.0, + 1173.0, + 296.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1170.0, + 1405.0, + 1170.0, + 1405.0, + 1206.0, + 294.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 1405.0, + 1198.0, + 1405.0, + 1237.0, + 293.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1229.0, + 1402.0, + 1229.0, + 1402.0, + 1265.0, + 294.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1259.0, + 491.0, + 1259.0, + 491.0, + 1296.0, + 293.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1259.0, + 1405.0, + 1259.0, + 1405.0, + 1296.0, + 552.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1292.0, + 1405.0, + 1292.0, + 1405.0, + 1325.0, + 297.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1325.0, + 941.0, + 1325.0, + 941.0, + 1354.0, + 294.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 435.0, + 1405.0, + 435.0, + 1405.0, + 470.0, + 296.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 466.0, + 1404.0, + 466.0, + 1404.0, + 502.0, + 294.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 496.0, + 1405.0, + 496.0, + 1405.0, + 530.0, + 296.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 526.0, + 1405.0, + 526.0, + 1405.0, + 562.0, + 294.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 559.0, + 834.0, + 559.0, + 834.0, + 591.0, + 296.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 559.0, + 1405.0, + 559.0, + 1405.0, + 591.0, + 853.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 587.0, + 1139.0, + 587.0, + 1139.0, + 621.0, + 295.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1366.0, + 1403.0, + 1366.0, + 1403.0, + 1405.0, + 295.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1401.0, + 1404.0, + 1401.0, + 1404.0, + 1432.0, + 296.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1431.0, + 1405.0, + 1431.0, + 1405.0, + 1465.0, + 293.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 1405.0, + 1461.0, + 1405.0, + 1496.0, + 294.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1489.0, + 499.0, + 1489.0, + 499.0, + 1525.0, + 293.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1489.0, + 1405.0, + 1489.0, + 1405.0, + 1525.0, + 583.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1520.0, + 389.0, + 1520.0, + 389.0, + 1557.0, + 293.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1520.0, + 993.0, + 1520.0, + 993.0, + 1557.0, + 410.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1520.0, + 1405.0, + 1520.0, + 1405.0, + 1557.0, + 1054.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1552.0, + 1405.0, + 1552.0, + 1405.0, + 1586.0, + 296.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 288.0, + 416.0, + 288.0, + 416.0, + 330.0, + 294.0, + 330.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 340, + 292, + 1358, + 292, + 1358, + 950, + 340, + 950 + ], + "score": 0.983, + "html": "
ModelPTBenwik8
BPCParamsBPCParams
LSTM1.36一1.45
LSTM+Zoneout (Krueger et al., 2016)1.2711
LSTM+Layer Norm(1000 units)11.2674.26M1一
LSTM+Layer Norm+Zoneout(1024 units)²1.244.79M一
HM-LSTM+Layer Norm (Chung et al.,2016)1.2411.3235M
HyperLSTM+Layer Norm(Ha et al., 2016)1.21914.41M1.34026.5M
NASCell (Zoph & Le,2016)1.21416.28M
IndRNN (21 layers) (Li et al.,2018)1.2111
Recurrent Highway Hypernetwork (Suarez, 2017)1.19一 15.5M11
Fast-Slow LSTM (Mujika et al., 2017)1.1907.2M1.277一 27M
NTM(800 units, nhead = 1, Pdropout=0.6)
TARDIS(paper)1.5358.28M1
1.25一
TARDIS(1000 units,Pdropout=0.6)1.2689.2M一
ARMIN+TARDIS-addr(800 units,Pdropout=0.6)1.22310.2M
ARMIN (500 units, Pdropout=0.4) (Ours)1.2364.03M
ARMIN (800 units, Pdropout=0.6) (Ours)1.2029.8M
ARMIN(800 units,2 layer) (Ours)111.33121.6M
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1016, + 1404, + 1016, + 1404, + 1441, + 297, + 1441 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1537, + 1404, + 1537, + 1404, + 1843, + 298, + 1843 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 301, + 1859, + 1404, + 1859, + 1404, + 1952, + 301, + 1952 + ], + "score": 0.962 + }, + { + "category_id": 2, + "poly": [ + 326, + 1974, + 1321, + 1974, + 1321, + 2034, + 326, + 2034 + ], + "score": 0.935 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.906 + }, + { + "category_id": 6, + "poly": [ + 480, + 225, + 1216, + 225, + 1216, + 257, + 480, + 257 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.797 + }, + { + "category_id": 0, + "poly": [ + 299, + 1483, + 895, + 1483, + 895, + 1511, + 299, + 1511 + ], + "score": 0.708 + }, + { + "category_id": 0, + "poly": [ + 299, + 1483, + 895, + 1483, + 895, + 1511, + 299, + 1511 + ], + "score": 0.143 + }, + { + "category_id": 13, + "poly": [ + 878, + 1753, + 910, + 1753, + 910, + 1783, + 878, + 1783 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1210, + 1169, + 1399, + 1169, + 1399, + 1203, + 1210, + 1203 + ], + "score": 0.88, + "latex": "\\pmb { x } _ { t } , \\pmb { h } _ { t } , \\pmb { M } _ { t } [ i ] , \\pmb { u } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 588, + 793, + 739, + 793, + 739, + 823, + 588, + 823 + ], + "score": 0.73, + "latex": "p _ { d r o p o u t } { = } 0 . 6 )" + }, + { + "category_id": 13, + "poly": [ + 545, + 732, + 663, + 732, + 663, + 762, + 545, + 762 + ], + "score": 0.72, + "latex": "n _ { h e a d } = 1" + }, + { + "category_id": 13, + "poly": [ + 1118, + 1752, + 1156, + 1752, + 1156, + 1780, + 1118, + 1780 + ], + "score": 0.53, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 451, + 826, + 469, + 826, + 469, + 848, + 451, + 848 + ], + "score": 0.51, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1276, + 1924, + 1294, + 1924, + 1294, + 1948, + 1276, + 1948 + ], + "score": 0.42, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 981, + 1081, + 999, + 1081, + 999, + 1104, + 981, + 1104 + ], + "score": 0.39, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1369, + 1171, + 1400, + 1171, + 1400, + 1201, + 1369, + 1201 + ], + "score": 0.34, + "latex": "\\mathbf { \\Delta } \\mathbf { u } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 673, + 885, + 727, + 885, + 727, + 914, + 673, + 914 + ], + "score": 0.32, + "latex": "_ { : = 0 . 6 }" + }, + { + "category_id": 13, + "poly": [ + 674, + 854, + 728, + 854, + 728, + 882, + 674, + 882 + ], + "score": 0.31, + "latex": "_ { = 0 . 4 }" + }, + { + "category_id": 13, + "poly": [ + 583, + 886, + 733, + 886, + 733, + 915, + 583, + 915 + ], + "score": 0.3, + "latex": "p _ { d r o p o u t } { = } 0 . 6 )" + }, + { + "category_id": 13, + "poly": [ + 434, + 407, + 451, + 407, + 451, + 429, + 434, + 429 + ], + "score": 0.28, + "latex": "+" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1969.0, + 707.0, + 1969.0, + 707.0, + 2008.0, + 331.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1998.0, + 1321.0, + 1998.0, + 1321.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 223.0, + 1218.0, + 223.0, + 1218.0, + 259.0, + 481.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 898.0, + 1481.0, + 898.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 898.0, + 1481.0, + 898.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1016.0, + 1406.0, + 1016.0, + 1406.0, + 1052.0, + 295.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 1405.0, + 1048.0, + 1405.0, + 1082.0, + 294.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1077.0, + 980.0, + 1077.0, + 980.0, + 1112.0, + 295.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1077.0, + 1405.0, + 1077.0, + 1405.0, + 1112.0, + 1000.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1109.0, + 1404.0, + 1109.0, + 1404.0, + 1144.0, + 295.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1139.0, + 1401.0, + 1139.0, + 1401.0, + 1170.0, + 296.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1169.0, + 1209.0, + 1169.0, + 1209.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1169.0, + 1405.0, + 1169.0, + 1405.0, + 1208.0, + 1401.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1199.0, + 1404.0, + 1199.0, + 1404.0, + 1235.0, + 295.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1231.0, + 1405.0, + 1231.0, + 1405.0, + 1266.0, + 295.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 1406.0, + 1260.0, + 1406.0, + 1295.0, + 294.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1290.0, + 1406.0, + 1290.0, + 1406.0, + 1327.0, + 292.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1323.0, + 1405.0, + 1323.0, + 1405.0, + 1355.0, + 296.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1352.0, + 1406.0, + 1352.0, + 1406.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1382.0, + 1406.0, + 1382.0, + 1406.0, + 1418.0, + 295.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1414.0, + 916.0, + 1414.0, + 916.0, + 1446.0, + 296.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1538.0, + 1404.0, + 1538.0, + 1404.0, + 1573.0, + 296.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1569.0, + 1406.0, + 1569.0, + 1406.0, + 1605.0, + 293.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 1404.0, + 1599.0, + 1404.0, + 1633.0, + 293.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1630.0, + 1404.0, + 1630.0, + 1404.0, + 1666.0, + 294.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1662.0, + 1402.0, + 1662.0, + 1402.0, + 1694.0, + 296.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1693.0, + 1404.0, + 1693.0, + 1404.0, + 1725.0, + 292.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1719.0, + 1406.0, + 1719.0, + 1406.0, + 1757.0, + 293.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1752.0, + 877.0, + 1752.0, + 877.0, + 1785.0, + 293.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1752.0, + 1117.0, + 1752.0, + 1117.0, + 1785.0, + 911.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1752.0, + 1406.0, + 1752.0, + 1406.0, + 1785.0, + 1157.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1781.0, + 1405.0, + 1781.0, + 1405.0, + 1817.0, + 294.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1814.0, + 636.0, + 1814.0, + 636.0, + 1846.0, + 294.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1857.0, + 1404.0, + 1857.0, + 1404.0, + 1895.0, + 296.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1891.0, + 1404.0, + 1891.0, + 1404.0, + 1925.0, + 296.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1920.0, + 1275.0, + 1920.0, + 1275.0, + 1954.0, + 295.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1920.0, + 1404.0, + 1920.0, + 1404.0, + 1954.0, + 1295.0, + 1954.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 399, + 1404, + 399, + 1404, + 765, + 298, + 765 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 997, + 1405, + 997, + 1405, + 1302, + 298, + 1302 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 859, + 1390, + 1406, + 1390, + 1406, + 1745, + 859, + 1745 + ], + "score": 0.978, + "html": "
ModelARMINLSTM
Setup1212
Hidden sizenparam(M)nmemTtruncbatch sizeMemory(GB)Speed(chars/s)BPC5005501k1k
4.0254.814.794.791
101
5050100150
384300128128
3.493.562.363.27
98k75k71k70k
1.2381.2261.271.24
" + }, + { + "category_id": 1, + "poly": [ + 300, + 781, + 1402, + 781, + 1402, + 903, + 300, + 903 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 383, + 298, + 383 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1771, + 1404, + 1771, + 1404, + 1957, + 297, + 1957 + ], + "score": 0.964 + }, + { + "category_id": 5, + "poly": [ + 300, + 1545, + 845, + 1545, + 845, + 1689, + 300, + 1689 + ], + "score": 0.95, + "html": "
ModeltIoU=0.6tIoU=0.8
SST (Paper)0.9200.672
SST (We implement)0.8970.732
MA-SST(Ours)0.9110.759
" + }, + { + "category_id": 6, + "poly": [ + 296, + 1377, + 851, + 1377, + 851, + 1530, + 296, + 1530 + ], + "score": 0.949 + }, + { + "category_id": 6, + "poly": [ + 851, + 1323, + 1408, + 1323, + 1408, + 1384, + 851, + 1384 + ], + "score": 0.945 + }, + { + "category_id": 2, + "poly": [ + 297, + 1978, + 1406, + 1978, + 1406, + 2035, + 297, + 2035 + ], + "score": 0.938 + }, + { + "category_id": 0, + "poly": [ + 300, + 941, + 1165, + 941, + 1165, + 971, + 300, + 971 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.7 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 857, + 76, + 857, + 104, + 299, + 104 + ], + "score": 0.656 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.546 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.146 + }, + { + "category_id": 13, + "poly": [ + 458, + 552, + 490, + 552, + 490, + 582, + 458, + 582 + ], + "score": 0.89, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 555, + 1339, + 555, + 1339, + 582, + 1312, + 582 + ], + "score": 0.86, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 541, + 555, + 569, + 555, + 569, + 582, + 541, + 582 + ], + "score": 0.86, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 587, + 326, + 587, + 326, + 613, + 298, + 613 + ], + "score": 0.82, + "latex": "\\mathbf { } _ { o _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1120, + 1301, + 1120, + 1301, + 1149, + 1252, + 1149 + ], + "score": 0.82, + "latex": "^ { 2 0 + }" + }, + { + "category_id": 13, + "poly": [ + 581, + 1500, + 685, + 1500, + 685, + 1528, + 581, + 1528 + ], + "score": 0.52, + "latex": "\\mathrm { \\ t I o U { = } } 0 . 8" + }, + { + "category_id": 13, + "poly": [ + 853, + 1834, + 959, + 1834, + 959, + 1864, + 853, + 1864 + ], + "score": 0.47, + "latex": "\\mathrm { \\ t I o U { = } } 0 . 8" + }, + { + "category_id": 13, + "poly": [ + 519, + 266, + 536, + 266, + 536, + 288, + 519, + 288 + ], + "score": 0.36, + "latex": "+" + }, + { + "category_id": 13, + "poly": [ + 426, + 1438, + 532, + 1438, + 532, + 1467, + 426, + 1467 + ], + "score": 0.35, + "latex": "\\mathrm { \\ t I o U { = } 0 } . 6" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 854.0, + 1374.0, + 854.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1406.0, + 853.0, + 1406.0, + 853.0, + 1441.0, + 295.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1437.0, + 425.0, + 1437.0, + 425.0, + 1469.0, + 295.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1437.0, + 855.0, + 1437.0, + 855.0, + 1469.0, + 533.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1468.0, + 856.0, + 1468.0, + 856.0, + 1500.0, + 294.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1499.0, + 580.0, + 1499.0, + 580.0, + 1530.0, + 295.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1499.0, + 695.0, + 1499.0, + 695.0, + 1530.0, + 686.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1321.0, + 1411.0, + 1321.0, + 1411.0, + 1356.0, + 852.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1353.0, + 1327.0, + 1353.0, + 1327.0, + 1385.0, + 853.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2014.0, + 328.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 939.0, + 1171.0, + 939.0, + 1171.0, + 976.0, + 293.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 398.0, + 1403.0, + 398.0, + 1403.0, + 435.0, + 295.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 429.0, + 1402.0, + 429.0, + 1402.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1404.0, + 461.0, + 1404.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 1404.0, + 491.0, + 1404.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 521.0, + 1404.0, + 521.0, + 1404.0, + 555.0, + 294.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 550.0, + 457.0, + 550.0, + 457.0, + 584.0, + 294.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 550.0, + 540.0, + 550.0, + 540.0, + 584.0, + 491.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 550.0, + 1311.0, + 550.0, + 1311.0, + 584.0, + 570.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 550.0, + 1404.0, + 550.0, + 1404.0, + 584.0, + 1340.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 582.0, + 297.0, + 582.0, + 297.0, + 618.0, + 292.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 582.0, + 1405.0, + 582.0, + 1405.0, + 618.0, + 327.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 611.0, + 1405.0, + 611.0, + 1405.0, + 648.0, + 293.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 637.0, + 1406.0, + 637.0, + 1406.0, + 683.0, + 291.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 674.0, + 1404.0, + 674.0, + 1404.0, + 706.0, + 293.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 703.0, + 1406.0, + 703.0, + 1406.0, + 737.0, + 293.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 1056.0, + 731.0, + 1056.0, + 771.0, + 295.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 1405.0, + 995.0, + 1405.0, + 1032.0, + 293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1029.0, + 1406.0, + 1029.0, + 1406.0, + 1063.0, + 295.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1058.0, + 1405.0, + 1058.0, + 1405.0, + 1094.0, + 295.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1087.0, + 1407.0, + 1087.0, + 1407.0, + 1124.0, + 293.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1119.0, + 1251.0, + 1119.0, + 1251.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1119.0, + 1407.0, + 1119.0, + 1407.0, + 1155.0, + 1302.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1152.0, + 1405.0, + 1152.0, + 1405.0, + 1184.0, + 296.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1180.0, + 1407.0, + 1180.0, + 1407.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1212.0, + 1407.0, + 1212.0, + 1407.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1243.0, + 1402.0, + 1243.0, + 1402.0, + 1275.0, + 296.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1271.0, + 991.0, + 1271.0, + 991.0, + 1305.0, + 293.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 778.0, + 1402.0, + 778.0, + 1402.0, + 817.0, + 295.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 811.0, + 1406.0, + 811.0, + 1406.0, + 847.0, + 294.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 843.0, + 1404.0, + 843.0, + 1404.0, + 876.0, + 295.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 873.0, + 1044.0, + 873.0, + 1044.0, + 906.0, + 296.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 518.0, + 262.0, + 518.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 295.0, + 537.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 352.0, + 859.0, + 352.0, + 859.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 1406.0, + 1771.0, + 1406.0, + 1805.0, + 294.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1406.0, + 1803.0, + 1406.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1832.0, + 852.0, + 1832.0, + 852.0, + 1868.0, + 295.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1832.0, + 1406.0, + 1832.0, + 1406.0, + 1868.0, + 960.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1892.0, + 1404.0, + 1892.0, + 1404.0, + 1930.0, + 294.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 1404.0, + 1924.0, + 1404.0, + 1960.0, + 295.0, + 1960.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1076, + 1404, + 1076, + 1404, + 1834, + 298, + 1834 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1850, + 1403, + 1850, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 313, + 219, + 1380, + 219, + 1380, + 617, + 313, + 617 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 296, + 641, + 1405, + 641, + 1405, + 825, + 296, + 825 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 293, + 888, + 1401, + 888, + 1401, + 952, + 293, + 952 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 295, + 1003, + 1291, + 1003, + 1291, + 1036, + 295, + 1036 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.789 + }, + { + "category_id": 13, + "poly": [ + 1051, + 1622, + 1112, + 1622, + 1112, + 1654, + 1051, + 1654 + ], + "score": 0.92, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 2003, + 459, + 2003, + 459, + 2035, + 298, + 2035 + ], + "score": 0.91, + "latex": "T _ { t r u n c } = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 402, + 1317, + 497, + 1317, + 497, + 1347, + 402, + 1347 + ], + "score": 0.91, + "latex": "2 0 \\times d _ { h }" + }, + { + "category_id": 13, + "poly": [ + 994, + 794, + 1054, + 794, + 1054, + 823, + 994, + 823 + ], + "score": 0.87, + "latex": "2 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 404, + 703, + 530, + 703, + 530, + 733, + 404, + 733 + ], + "score": 0.87, + "latex": "\\mathrm { t I o U } \\geq 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 966, + 1625, + 998, + 1625, + 998, + 1652, + 966, + 1652 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 538, + 1686, + 556, + 1686, + 556, + 1708, + 538, + 1708 + ], + "score": 0.45, + "latex": "+ { \\bf \\Phi }" + }, + { + "category_id": 13, + "poly": [ + 547, + 1594, + 566, + 1594, + 566, + 1617, + 547, + 1617 + ], + "score": 0.42, + "latex": "\\dotplus" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 220.0, + 481.0, + 220.0, + 481.0, + 247.0, + 376.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 220.0, + 811.0, + 220.0, + 811.0, + 247.0, + 509.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 225.0, + 1359.0, + 225.0, + 1359.0, + 248.0, + 972.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 240.0, + 360.0, + 240.0, + 360.0, + 258.0, + 336.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 299.0, + 911.0, + 299.0, + 911.0, + 513.0, + 879.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 301.0, + 935.0, + 301.0, + 935.0, + 323.0, + 904.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 320.0, + 360.0, + 320.0, + 360.0, + 339.0, + 336.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 347.0, + 343.0, + 347.0, + 343.0, + 475.0, + 312.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 364.0, + 935.0, + 364.0, + 935.0, + 386.0, + 906.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 401.0, + 359.0, + 401.0, + 359.0, + 420.0, + 336.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 428.0, + 935.0, + 428.0, + 935.0, + 450.0, + 906.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 484.0, + 359.0, + 484.0, + 359.0, + 502.0, + 337.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 490.0, + 935.0, + 490.0, + 935.0, + 513.0, + 905.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 556.0, + 938.0, + 556.0, + 938.0, + 576.0, + 918.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 569.0, + 380.0, + 569.0, + 380.0, + 592.0, + 347.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 575.0, + 529.0, + 575.0, + 529.0, + 592.0, + 498.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 575.0, + 678.0, + 575.0, + 678.0, + 592.0, + 641.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 575.0, + 825.0, + 575.0, + 825.0, + 592.0, + 783.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 566.0, + 962.0, + 566.0, + 962.0, + 592.0, + 923.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 566.0, + 1048.0, + 566.0, + 1048.0, + 592.0, + 1008.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 566.0, + 1135.0, + 566.0, + 1135.0, + 592.0, + 1093.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 566.0, + 1218.0, + 566.0, + 1218.0, + 592.0, + 1177.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 566.0, + 1304.0, + 566.0, + 1304.0, + 592.0, + 1263.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 569.0, + 1384.0, + 569.0, + 1384.0, + 590.0, + 1353.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 592.0, + 712.0, + 592.0, + 712.0, + 622.0, + 459.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 590.0, + 1181.0, + 590.0, + 1181.0, + 620.0, + 1131.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 394.5, + 610.0, + 394.5, + 610.0, + 409.0, + 537.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 640.0, + 1405.0, + 640.0, + 1405.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 667.0, + 1406.0, + 667.0, + 1406.0, + 712.0, + 291.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 403.0, + 701.0, + 403.0, + 737.0, + 294.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 701.0, + 1405.0, + 701.0, + 1405.0, + 737.0, + 531.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 732.0, + 1405.0, + 732.0, + 1405.0, + 771.0, + 293.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 762.0, + 1403.0, + 762.0, + 1403.0, + 798.0, + 295.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 794.0, + 993.0, + 794.0, + 993.0, + 827.0, + 294.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 794.0, + 1066.0, + 794.0, + 1066.0, + 827.0, + 1055.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 999.0, + 1298.0, + 999.0, + 1298.0, + 1044.0, + 291.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1073.0, + 1405.0, + 1073.0, + 1405.0, + 1109.0, + 294.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1105.0, + 1405.0, + 1105.0, + 1405.0, + 1137.0, + 296.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1133.0, + 1406.0, + 1133.0, + 1406.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1163.0, + 1404.0, + 1163.0, + 1404.0, + 1200.0, + 293.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 1404.0, + 1195.0, + 1404.0, + 1232.0, + 293.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 1405.0, + 1224.0, + 1405.0, + 1262.0, + 293.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1295.0, + 292.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1322.0, + 294.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1316.0, + 401.0, + 1316.0, + 401.0, + 1355.0, + 292.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1316.0, + 1406.0, + 1316.0, + 1406.0, + 1355.0, + 498.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 1404.0, + 1345.0, + 1404.0, + 1384.0, + 292.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1404.0, + 1376.0, + 1404.0, + 1414.0, + 293.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1444.0, + 294.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1437.0, + 1402.0, + 1437.0, + 1402.0, + 1473.0, + 294.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1465.0, + 1404.0, + 1465.0, + 1404.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1404.0, + 1498.0, + 1404.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1528.0, + 1406.0, + 1528.0, + 1406.0, + 1566.0, + 292.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1599.0, + 293.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1586.0, + 546.0, + 1586.0, + 546.0, + 1630.0, + 292.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1586.0, + 1405.0, + 1586.0, + 1405.0, + 1630.0, + 567.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1620.0, + 965.0, + 1620.0, + 965.0, + 1659.0, + 293.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1620.0, + 1050.0, + 1620.0, + 1050.0, + 1659.0, + 999.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1659.0, + 1113.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1652.0, + 1404.0, + 1652.0, + 1404.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 537.0, + 1680.0, + 537.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1718.0, + 557.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1404.0, + 1710.0, + 1404.0, + 1749.0, + 293.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1403.0, + 1740.0, + 1403.0, + 1778.0, + 293.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1770.0, + 1405.0, + 1770.0, + 1405.0, + 1811.0, + 293.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1805.0, + 695.0, + 1805.0, + 695.0, + 1837.0, + 296.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 1405.0, + 1847.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1879.0, + 1407.0, + 1879.0, + 1407.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 2002.0, + 297.0, + 2002.0, + 297.0, + 2038.0, + 290.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 2002.0, + 1408.0, + 2002.0, + 1408.0, + 2038.0, + 460.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 887.0, + 1404.0, + 887.0, + 1404.0, + 924.0, + 293.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 919.0, + 1274.0, + 919.0, + 1274.0, + 956.0, + 293.0, + 956.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1106, + 1404, + 1106, + 1404, + 1382, + 297, + 1382 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 385, + 298, + 385 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 295, + 825, + 1405, + 825, + 1405, + 980, + 295, + 980 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 315, + 435, + 1384, + 435, + 1384, + 815, + 315, + 815 + ], + "score": 0.964 + }, + { + "category_id": 0, + "poly": [ + 299, + 1036, + 543, + 1036, + 543, + 1071, + 299, + 1071 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 300, + 1429, + 487, + 1429, + 487, + 1462, + 300, + 1462 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.767 + }, + { + "category_id": 1, + "poly": [ + 295, + 1480, + 1402, + 1480, + 1402, + 1543, + 295, + 1543 + ], + "score": 0.588 + }, + { + "category_id": 1, + "poly": [ + 310, + 1794, + 1399, + 1794, + 1399, + 1920, + 310, + 1920 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 300, + 1565, + 1400, + 1565, + 1400, + 1659, + 300, + 1659 + ], + "score": 0.483 + }, + { + "category_id": 1, + "poly": [ + 298, + 1680, + 1400, + 1680, + 1400, + 1775, + 298, + 1775 + ], + "score": 0.439 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1403, + 1942, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.234 + }, + { + "category_id": 13, + "poly": [ + 344, + 291, + 399, + 291, + 399, + 321, + 344, + 321 + ], + "score": 0.89, + "latex": "1 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 572, + 261, + 628, + 261, + 628, + 291, + 572, + 291 + ], + "score": 0.89, + "latex": "4 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 412, + 321, + 474, + 321, + 474, + 352, + 412, + 352 + ], + "score": 0.87, + "latex": "8 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 344, + 352, + 407, + 352, + 407, + 382, + 344, + 382 + ], + "score": 0.87, + "latex": "7 . 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 301, + 857, + 454, + 857, + 454, + 887, + 301, + 887 + ], + "score": 0.87, + "latex": "( n _ { m e m } = 2 0 )" + }, + { + "category_id": 13, + "poly": [ + 1175, + 230, + 1238, + 230, + 1238, + 261, + 1175, + 261 + ], + "score": 0.86, + "latex": "6 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 659, + 789, + 712, + 789, + 712, + 814, + 659, + 814 + ], + "score": 0.58, + "latex": "{ \\ = } 1 2 8" + }, + { + "category_id": 13, + "poly": [ + 1231, + 790, + 1261, + 790, + 1261, + 814, + 1231, + 814 + ], + "score": 0.52, + "latex": "{ \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } } { \\boldsymbol { \\mathbf { \\rho } } }" + }, + { + "category_id": 13, + "poly": [ + 440, + 920, + 458, + 920, + 458, + 944, + 440, + 944 + ], + "score": 0.38, + "latex": "+" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 823.0, + 1406.0, + 823.0, + 1406.0, + 864.0, + 293.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 853.0, + 300.0, + 853.0, + 300.0, + 892.0, + 290.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 853.0, + 1406.0, + 853.0, + 1406.0, + 892.0, + 455.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 884.0, + 1403.0, + 884.0, + 1403.0, + 921.0, + 293.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 913.0, + 439.0, + 913.0, + 439.0, + 953.0, + 293.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 913.0, + 1406.0, + 913.0, + 1406.0, + 953.0, + 459.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 531.0, + 948.0, + 531.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 437.0, + 378.0, + 437.0, + 378.0, + 455.0, + 340.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 439.0, + 801.0, + 439.0, + 801.0, + 454.0, + 785.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 436.0, + 929.0, + 436.0, + 929.0, + 458.0, + 895.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 454.0, + 433.0, + 454.0, + 433.0, + 477.0, + 359.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 462.0, + 360.0, + 462.0, + 360.0, + 471.0, + 350.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 475.0, + 347.0, + 475.0, + 347.0, + 685.0, + 311.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 466.0, + 512.0, + 466.0, + 512.0, + 499.0, + 361.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 470.0, + 835.0, + 470.0, + 835.0, + 696.0, + 800.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 477.0, + 904.0, + 477.0, + 904.0, + 689.0, + 869.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 473.0, + 1382.0, + 473.0, + 1382.0, + 696.0, + 1351.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 489.0, + 442.0, + 489.0, + 442.0, + 509.0, + 360.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 501.0, + 438.0, + 501.0, + 438.0, + 521.0, + 361.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 505.0, + 803.0, + 505.0, + 803.0, + 523.0, + 774.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 505.0, + 927.0, + 505.0, + 927.0, + 524.0, + 898.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 504.0, + 1354.0, + 504.0, + 1354.0, + 526.0, + 1322.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 514.0, + 445.0, + 514.0, + 445.0, + 534.0, + 349.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 524.0, + 453.0, + 524.0, + 453.0, + 547.0, + 348.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 538.0, + 440.0, + 538.0, + 440.0, + 557.0, + 361.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 548.0, + 520.0, + 548.0, + 520.0, + 577.0, + 349.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 571.0, + 381.0, + 571.0, + 381.0, + 596.0, + 340.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 574.0, + 803.0, + 574.0, + 803.0, + 592.0, + 779.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 575.0, + 927.0, + 575.0, + 927.0, + 594.0, + 898.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 575.0, + 1352.0, + 575.0, + 1352.0, + 594.0, + 1329.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 641.0, + 378.0, + 641.0, + 378.0, + 663.0, + 340.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 642.0, + 803.0, + 642.0, + 803.0, + 661.0, + 774.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 644.0, + 926.0, + 644.0, + 926.0, + 663.0, + 901.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 644.0, + 1351.0, + 644.0, + 1351.0, + 663.0, + 1323.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 714.0, + 378.0, + 714.0, + 378.0, + 729.0, + 364.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 714.0, + 803.0, + 714.0, + 803.0, + 729.0, + 789.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 714.0, + 929.0, + 714.0, + 929.0, + 730.0, + 914.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 715.0, + 1352.0, + 715.0, + 1352.0, + 730.0, + 1338.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 723.0, + 399.0, + 723.0, + 399.0, + 740.0, + 372.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 721.0, + 437.0, + 721.0, + 437.0, + 740.0, + 409.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 723.0, + 474.0, + 723.0, + 474.0, + 740.0, + 448.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 723.0, + 513.0, + 723.0, + 513.0, + 740.0, + 487.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 723.0, + 552.0, + 723.0, + 552.0, + 740.0, + 524.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 723.0, + 590.0, + 723.0, + 590.0, + 740.0, + 562.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 723.0, + 629.0, + 723.0, + 629.0, + 740.0, + 601.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 723.0, + 666.0, + 723.0, + 666.0, + 740.0, + 638.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 723.0, + 705.0, + 723.0, + 705.0, + 740.0, + 677.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 723.0, + 746.0, + 723.0, + 746.0, + 740.0, + 713.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 723.0, + 783.0, + 723.0, + 783.0, + 740.0, + 751.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 723.0, + 948.0, + 723.0, + 948.0, + 741.0, + 923.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 723.0, + 986.0, + 723.0, + 986.0, + 741.0, + 958.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 723.0, + 1024.0, + 723.0, + 1024.0, + 741.0, + 997.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 721.0, + 1064.0, + 721.0, + 1064.0, + 744.0, + 1035.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 723.0, + 1102.0, + 723.0, + 1102.0, + 741.0, + 1075.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 723.0, + 1139.0, + 723.0, + 1139.0, + 741.0, + 1111.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 724.0, + 1177.0, + 724.0, + 1177.0, + 741.0, + 1150.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 723.0, + 1216.0, + 723.0, + 1216.0, + 741.0, + 1189.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 723.0, + 1254.0, + 723.0, + 1254.0, + 741.0, + 1226.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 724.0, + 1295.0, + 724.0, + 1295.0, + 741.0, + 1262.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 721.0, + 1333.0, + 721.0, + 1333.0, + 743.0, + 1296.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 739.0, + 630.0, + 739.0, + 630.0, + 767.0, + 523.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 740.0, + 1179.0, + 740.0, + 1179.0, + 769.0, + 1072.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 785.0, + 658.0, + 785.0, + 658.0, + 820.0, + 425.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 785.0, + 725.0, + 785.0, + 725.0, + 820.0, + 713.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 785.0, + 1230.0, + 785.0, + 1230.0, + 820.0, + 987.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 785.0, + 1274.0, + 785.0, + 1274.0, + 820.0, + 1262.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1032.0, + 548.0, + 1032.0, + 548.0, + 1080.0, + 292.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1431.0, + 489.0, + 1431.0, + 489.0, + 1464.0, + 297.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 1402.0, + 1108.0, + 1402.0, + 1141.0, + 295.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1136.0, + 1402.0, + 1136.0, + 1402.0, + 1171.0, + 295.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1166.0, + 1406.0, + 1166.0, + 1406.0, + 1206.0, + 294.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1193.0, + 1405.0, + 1193.0, + 1405.0, + 1236.0, + 290.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 1405.0, + 1228.0, + 1405.0, + 1265.0, + 294.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1261.0, + 1405.0, + 1261.0, + 1405.0, + 1294.0, + 296.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1291.0, + 1404.0, + 1291.0, + 1404.0, + 1322.0, + 295.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1320.0, + 1405.0, + 1320.0, + 1405.0, + 1356.0, + 295.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1352.0, + 622.0, + 1352.0, + 622.0, + 1385.0, + 296.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1174.0, + 228.0, + 1174.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 228.0, + 1403.0, + 228.0, + 1403.0, + 267.0, + 1239.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 571.0, + 260.0, + 571.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 298.0, + 629.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 343.0, + 291.0, + 343.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 328.0, + 400.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 316.0, + 411.0, + 316.0, + 411.0, + 360.0, + 293.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 316.0, + 1407.0, + 316.0, + 1407.0, + 360.0, + 475.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 350.0, + 343.0, + 350.0, + 343.0, + 390.0, + 293.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 350.0, + 974.0, + 350.0, + 974.0, + 390.0, + 408.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1405.0, + 1477.0, + 1405.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1512.0, + 612.0, + 1512.0, + 612.0, + 1545.0, + 322.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1793.0, + 1404.0, + 1793.0, + 1404.0, + 1832.0, + 302.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1828.0, + 1406.0, + 1828.0, + 1406.0, + 1861.0, + 323.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1856.0, + 1406.0, + 1856.0, + 1406.0, + 1894.0, + 322.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1889.0, + 1019.0, + 1889.0, + 1019.0, + 1922.0, + 324.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1565.0, + 1405.0, + 1565.0, + 1405.0, + 1602.0, + 295.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1597.0, + 1405.0, + 1597.0, + 1405.0, + 1631.0, + 324.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1627.0, + 977.0, + 1627.0, + 977.0, + 1660.0, + 322.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1678.0, + 1404.0, + 1678.0, + 1404.0, + 1718.0, + 292.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1709.0, + 1406.0, + 1709.0, + 1406.0, + 1746.0, + 321.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1741.0, + 1197.0, + 1741.0, + 1197.0, + 1777.0, + 323.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1403.0, + 1940.0, + 1403.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2007.0, + 320.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2000.0, + 1147.0, + 2000.0, + 1147.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 296, + 1278, + 1401, + 1278, + 1401, + 1372, + 296, + 1372 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 297, + 839, + 1404, + 839, + 1404, + 935, + 297, + 935 + ], + "score": 0.865 + }, + { + "category_id": 1, + "poly": [ + 298, + 1129, + 1404, + 1129, + 1404, + 1255, + 298, + 1255 + ], + "score": 0.864 + }, + { + "category_id": 1, + "poly": [ + 296, + 1044, + 1404, + 1044, + 1404, + 1107, + 296, + 1107 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 292, + 1711, + 1400, + 1711, + 1400, + 1776, + 292, + 1776 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 297, + 1395, + 1399, + 1395, + 1399, + 1459, + 297, + 1459 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 299, + 1482, + 1401, + 1482, + 1401, + 1546, + 299, + 1546 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 299, + 722, + 1402, + 722, + 1402, + 817, + 299, + 817 + ], + "score": 0.843 + }, + { + "category_id": 1, + "poly": [ + 294, + 957, + 1401, + 957, + 1401, + 1022, + 294, + 1022 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 298, + 1568, + 1401, + 1568, + 1401, + 1633, + 298, + 1633 + ], + "score": 0.838 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 297, + 1654, + 1388, + 1654, + 1388, + 1690, + 297, + 1690 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1397, + 1971, + 1397, + 2035, + 298, + 2035 + ], + "score": 0.793 + }, + { + "category_id": 1, + "poly": [ + 298, + 519, + 1405, + 519, + 1405, + 645, + 298, + 645 + ], + "score": 0.791 + }, + { + "category_id": 1, + "poly": [ + 295, + 666, + 1239, + 666, + 1239, + 701, + 295, + 701 + ], + "score": 0.783 + }, + { + "category_id": 1, + "poly": [ + 295, + 1798, + 1402, + 1798, + 1402, + 1862, + 295, + 1862 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 292, + 315, + 1400, + 315, + 1400, + 380, + 292, + 380 + ], + "score": 0.74 + }, + { + "category_id": 1, + "poly": [ + 291, + 228, + 1402, + 228, + 1402, + 295, + 291, + 295 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 297, + 401, + 1405, + 401, + 1405, + 498, + 297, + 498 + ], + "score": 0.734 + }, + { + "category_id": 1, + "poly": [ + 294, + 1884, + 1400, + 1884, + 1400, + 1949, + 294, + 1949 + ], + "score": 0.734 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1276.0, + 1405.0, + 1276.0, + 1405.0, + 1314.0, + 294.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1304.0, + 1406.0, + 1304.0, + 1406.0, + 1348.0, + 320.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1341.0, + 532.0, + 1341.0, + 532.0, + 1371.0, + 325.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 838.0, + 1406.0, + 838.0, + 1406.0, + 878.0, + 294.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 907.0, + 321.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 903.0, + 922.0, + 903.0, + 922.0, + 937.0, + 324.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1404.0, + 1130.0, + 1404.0, + 1166.0, + 294.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1160.0, + 1405.0, + 1160.0, + 1405.0, + 1198.0, + 320.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1192.0, + 1406.0, + 1192.0, + 1406.0, + 1228.0, + 320.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1222.0, + 533.0, + 1222.0, + 533.0, + 1257.0, + 321.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1406.0, + 1039.0, + 1406.0, + 1082.0, + 294.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1074.0, + 596.0, + 1074.0, + 596.0, + 1105.0, + 323.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1709.0, + 1405.0, + 1709.0, + 1405.0, + 1750.0, + 294.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1742.0, + 533.0, + 1742.0, + 533.0, + 1776.0, + 324.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1394.0, + 1403.0, + 1394.0, + 1403.0, + 1430.0, + 296.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1427.0, + 1045.0, + 1427.0, + 1045.0, + 1458.0, + 323.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1519.0, + 294.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1513.0, + 926.0, + 1513.0, + 926.0, + 1546.0, + 323.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 718.0, + 1406.0, + 718.0, + 1406.0, + 763.0, + 293.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 751.0, + 1404.0, + 751.0, + 1404.0, + 796.0, + 319.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 785.0, + 614.0, + 785.0, + 614.0, + 819.0, + 324.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 953.0, + 1407.0, + 953.0, + 1407.0, + 996.0, + 293.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 990.0, + 982.0, + 990.0, + 982.0, + 1023.0, + 323.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1568.0, + 1403.0, + 1568.0, + 1403.0, + 1604.0, + 296.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1601.0, + 982.0, + 1601.0, + 982.0, + 1633.0, + 324.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1392.0, + 1651.0, + 1392.0, + 1695.0, + 294.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 1221.0, + 2002.0, + 1221.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 518.0, + 1406.0, + 518.0, + 1406.0, + 555.0, + 293.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 553.0, + 1403.0, + 553.0, + 1403.0, + 586.0, + 323.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 582.0, + 1406.0, + 582.0, + 1406.0, + 617.0, + 321.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 613.0, + 1329.0, + 613.0, + 1329.0, + 647.0, + 320.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 1241.0, + 666.0, + 1241.0, + 706.0, + 294.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1797.0, + 1405.0, + 1797.0, + 1405.0, + 1836.0, + 294.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1830.0, + 706.0, + 1830.0, + 706.0, + 1863.0, + 320.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1403.0, + 317.0, + 1403.0, + 353.0, + 295.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 349.0, + 858.0, + 349.0, + 858.0, + 378.0, + 324.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1338.0, + 260.0, + 1338.0, + 295.0, + 322.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 399.0, + 1406.0, + 399.0, + 1406.0, + 441.0, + 292.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 435.0, + 1403.0, + 435.0, + 1403.0, + 469.0, + 323.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 464.0, + 1176.0, + 464.0, + 1176.0, + 499.0, + 320.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1886.0, + 1402.0, + 1886.0, + 1402.0, + 1918.0, + 297.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1915.0, + 1179.0, + 1915.0, + 1179.0, + 1950.0, + 324.0, + 1950.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.845 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 295, + 1615, + 1401, + 1615, + 1401, + 1680, + 295, + 1680 + ], + "score": 0.797 + }, + { + "category_id": 1, + "poly": [ + 295, + 1398, + 1403, + 1398, + 1403, + 1464, + 295, + 1464 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 289, + 999, + 1402, + 999, + 1402, + 1064, + 289, + 1064 + ], + "score": 0.77 + }, + { + "category_id": 1, + "poly": [ + 295, + 1954, + 1400, + 1954, + 1400, + 2017, + 295, + 2017 + ], + "score": 0.769 + }, + { + "category_id": 1, + "poly": [ + 301, + 1492, + 1402, + 1492, + 1402, + 1587, + 301, + 1587 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 296, + 1091, + 1398, + 1091, + 1398, + 1155, + 296, + 1155 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 298, + 352, + 1400, + 352, + 1400, + 415, + 298, + 415 + ], + "score": 0.729 + }, + { + "category_id": 1, + "poly": [ + 290, + 1306, + 1400, + 1306, + 1400, + 1371, + 290, + 1371 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 297, + 1708, + 1401, + 1708, + 1401, + 1801, + 297, + 1801 + ], + "score": 0.721 + }, + { + "category_id": 1, + "poly": [ + 292, + 782, + 1401, + 782, + 1401, + 848, + 292, + 848 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 296, + 567, + 1400, + 567, + 1400, + 631, + 296, + 631 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1398, + 229, + 1398, + 323, + 299, + 323 + ], + "score": 0.665 + }, + { + "category_id": 1, + "poly": [ + 296, + 1832, + 1398, + 1832, + 1398, + 1925, + 296, + 1925 + ], + "score": 0.662 + }, + { + "category_id": 1, + "poly": [ + 298, + 875, + 1403, + 875, + 1403, + 971, + 298, + 971 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 298, + 445, + 1399, + 445, + 1399, + 540, + 298, + 540 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 296, + 659, + 1408, + 659, + 1408, + 755, + 296, + 755 + ], + "score": 0.609 + }, + { + "category_id": 1, + "poly": [ + 296, + 1183, + 1404, + 1183, + 1404, + 1280, + 296, + 1280 + ], + "score": 0.581 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 1406.0, + 1615.0, + 1406.0, + 1651.0, + 294.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1645.0, + 946.0, + 1645.0, + 946.0, + 1681.0, + 322.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1435.0, + 297.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1431.0, + 1405.0, + 1431.0, + 1405.0, + 1467.0, + 323.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 998.0, + 1405.0, + 998.0, + 1405.0, + 1035.0, + 295.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1029.0, + 1056.0, + 1029.0, + 1056.0, + 1065.0, + 321.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1950.0, + 1405.0, + 1950.0, + 1405.0, + 1993.0, + 293.0, + 1993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1984.0, + 609.0, + 1984.0, + 609.0, + 2017.0, + 323.0, + 2017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1493.0, + 1403.0, + 1493.0, + 1403.0, + 1527.0, + 296.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1524.0, + 1404.0, + 1524.0, + 1404.0, + 1558.0, + 322.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1554.0, + 825.0, + 1554.0, + 825.0, + 1587.0, + 322.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1087.0, + 1403.0, + 1087.0, + 1403.0, + 1129.0, + 292.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1122.0, + 466.0, + 1122.0, + 466.0, + 1155.0, + 320.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 348.0, + 1405.0, + 348.0, + 1405.0, + 390.0, + 292.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 385.0, + 596.0, + 385.0, + 596.0, + 417.0, + 324.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1308.0, + 1402.0, + 1308.0, + 1402.0, + 1344.0, + 294.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1340.0, + 878.0, + 1340.0, + 878.0, + 1372.0, + 321.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1705.0, + 1405.0, + 1705.0, + 1405.0, + 1746.0, + 293.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1774.0, + 322.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1768.0, + 439.0, + 1768.0, + 439.0, + 1803.0, + 321.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 781.0, + 1403.0, + 781.0, + 1403.0, + 818.0, + 296.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 814.0, + 1223.0, + 814.0, + 1223.0, + 849.0, + 321.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 570.0, + 1404.0, + 570.0, + 1404.0, + 602.0, + 296.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 599.0, + 707.0, + 599.0, + 707.0, + 632.0, + 320.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1403.0, + 227.0, + 1403.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 259.0, + 1402.0, + 259.0, + 1402.0, + 298.0, + 323.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 827.0, + 292.0, + 827.0, + 326.0, + 323.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1828.0, + 1403.0, + 1828.0, + 1403.0, + 1869.0, + 292.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1863.0, + 1404.0, + 1863.0, + 1404.0, + 1897.0, + 324.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1894.0, + 926.0, + 1894.0, + 926.0, + 1925.0, + 322.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 876.0, + 1403.0, + 876.0, + 1403.0, + 911.0, + 296.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 908.0, + 1402.0, + 908.0, + 1402.0, + 943.0, + 325.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 938.0, + 749.0, + 938.0, + 749.0, + 972.0, + 325.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 443.0, + 1404.0, + 443.0, + 1404.0, + 484.0, + 293.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 473.0, + 1404.0, + 473.0, + 1404.0, + 514.0, + 322.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 508.0, + 937.0, + 508.0, + 937.0, + 542.0, + 324.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 660.0, + 1405.0, + 660.0, + 1405.0, + 695.0, + 297.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 691.0, + 1406.0, + 691.0, + 1406.0, + 728.0, + 322.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 723.0, + 932.0, + 723.0, + 932.0, + 757.0, + 327.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1181.0, + 1405.0, + 1181.0, + 1405.0, + 1223.0, + 291.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1213.0, + 1405.0, + 1213.0, + 1405.0, + 1252.0, + 322.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1246.0, + 1328.0, + 1246.0, + 1328.0, + 1280.0, + 321.0, + 1280.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 296, + 1405, + 296, + 1405, + 422, + 297, + 422 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 602, + 1405, + 602, + 1405, + 696, + 297, + 696 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 993, + 1404, + 993, + 1404, + 1090, + 298, + 1090 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1857, + 1402, + 1857, + 1402, + 1958, + 297, + 1958 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 626, + 1449, + 1080, + 1449, + 1080, + 1562, + 626, + 1562 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 615, + 718, + 1092, + 718, + 1092, + 868, + 615, + 868 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 297, + 1209, + 1404, + 1209, + 1404, + 1274, + 297, + 1274 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 525, + 1407, + 525, + 1407, + 589, + 295, + 589 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1970, + 1402, + 1970, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 541, + 1582, + 1163, + 1582, + 1163, + 1672, + 541, + 1672 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 301, + 1138, + 1192, + 1138, + 1192, + 1176, + 301, + 1176 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 298, + 224, + 1098, + 224, + 1098, + 264, + 298, + 264 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 1372, + 903, + 1372, + 903, + 1407, + 298, + 1407 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 674, + 425, + 1025, + 425, + 1025, + 461, + 674, + 461 + ], + "score": 0.929 + }, + { + "category_id": 8, + "poly": [ + 620, + 1689, + 1079, + 1689, + 1079, + 1728, + 620, + 1728 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 299, + 1314, + 629, + 1314, + 629, + 1346, + 299, + 1346 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1351, + 776, + 1400, + 776, + 1400, + 808, + 1351, + 808 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1352, + 947, + 1400, + 947, + 1400, + 978, + 1352, + 978 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1352, + 480, + 1400, + 480, + 1400, + 510, + 1352, + 510 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1352, + 894, + 1400, + 894, + 1400, + 924, + 1352, + 924 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1798, + 1400, + 1798, + 1400, + 1829, + 1352, + 1829 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1491, + 1400, + 1491, + 1400, + 1522, + 1352, + 1522 + ], + "score": 0.89 + }, + { + "category_id": 8, + "poly": [ + 697, + 1744, + 1001, + 1744, + 1001, + 1780, + 697, + 1780 + ], + "score": 0.889 + }, + { + "category_id": 8, + "poly": [ + 700, + 893, + 996, + 893, + 996, + 930, + 700, + 930 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1351, + 427, + 1400, + 427, + 1400, + 458, + 1351, + 458 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1692, + 1400, + 1692, + 1400, + 1723, + 1352, + 1723 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1745, + 1400, + 1745, + 1400, + 1776, + 1352, + 1776 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1611, + 1401, + 1611, + 1401, + 1643, + 1351, + 1643 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.865 + }, + { + "category_id": 8, + "poly": [ + 755, + 479, + 942, + 479, + 942, + 513, + 755, + 513 + ], + "score": 0.862 + }, + { + "category_id": 8, + "poly": [ + 732, + 945, + 964, + 945, + 964, + 980, + 732, + 980 + ], + "score": 0.852 + }, + { + "category_id": 8, + "poly": [ + 729, + 1796, + 968, + 1796, + 968, + 1833, + 729, + 1833 + ], + "score": 0.833 + }, + { + "category_id": 14, + "poly": [ + 610, + 714, + 1091, + 714, + 1091, + 869, + 610, + 869 + ], + "score": 0.94, + "latex": "\\left\\{ \\begin{array} { c } { i _ { t } } \\\\ { f _ { t } } \\\\ { g _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { c } { \\sigma } \\\\ { \\sigma } \\\\ { \\operatorname { t a n h } } \\\\ { \\sigma } \\end{array} \\right\\} W _ { 0 } [ \\ : x _ { t } , h _ { t - 1 } \\ : ] + b _ { 0 } ," + }, + { + "category_id": 14, + "poly": [ + 622, + 1449, + 1079, + 1449, + 1079, + 1563, + 622, + 1563 + ], + "score": 0.93, + "latex": "\\left\\{ \\begin{array} { l } { i _ { t } } \\\\ { f _ { t } } \\\\ { o _ { t } } \\end{array} \\right\\} = \\left\\{ \\begin{array} { l } { \\sigma } \\\\ { \\sigma } \\\\ { \\sigma } \\end{array} \\right\\} W [ \\ : x _ { t } , h _ { t - 1 } , r _ { t } \\ : ] + b \\ : ," + }, + { + "category_id": 13, + "poly": [ + 568, + 523, + 693, + 523, + 693, + 557, + 568, + 557 + ], + "score": 0.93, + "latex": "\\pmb { h } _ { t } ~ \\in ~ \\mathbb { R } ^ { d _ { h } }" + }, + { + "category_id": 13, + "poly": [ + 1313, + 634, + 1400, + 634, + 1400, + 668, + 1313, + 668 + ], + "score": 0.93, + "latex": "\\left( { { h _ { t } } , { c _ { t } } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1008, + 330, + 1190, + 330, + 1190, + 363, + 1008, + 363 + ], + "score": 0.92, + "latex": "\\left\\{ y _ { 1 } , y _ { 2 } , . . . , y _ { T } \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 410, + 330, + 595, + 330, + 595, + 362, + 410, + 362 + ], + "score": 0.92, + "latex": "\\{ \\pmb { x } _ { 1 } , \\pmb { x } _ { 2 } , . . . , \\pmb { x } _ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 1274, + 327, + 1392, + 327, + 1392, + 360, + 1274, + 360 + ], + "score": 0.92, + "latex": "\\pmb { x } _ { t } \\in \\mathbb { R } ^ { d _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 358, + 403, + 358, + 403, + 391, + 298, + 391 + ], + "score": 0.91, + "latex": "\\pmb { y } _ { t } \\in \\mathbb { R } ^ { d _ { o } }" + }, + { + "category_id": 13, + "poly": [ + 725, + 993, + 962, + 993, + 962, + 1029, + 725, + 1029 + ], + "score": 0.91, + "latex": "W _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } \\times ( d _ { i } + d _ { h } ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 556, + 466, + 556, + 466, + 589, + 298, + 589 + ], + "score": 0.91, + "latex": "U , W , V , b , c" + }, + { + "category_id": 14, + "poly": [ + 537, + 1579, + 1161, + 1579, + 1161, + 1674, + 537, + 1674 + ], + "score": 0.91, + "latex": "\\left\\{ \\begin{array} { l } \\alpha _ { t } \\right\\} = \\left\\{ \\begin{array} { l l } { \\mathrm { g u m b e l - s i g m o i d } } \\\\ { \\mathrm { g u m b e l - s i g m o i d } } \\end{array} \\right\\} \\left\\{ \\begin{array} { l } { \\pmb { w } ^ { \\alpha \\top } } \\\\ { \\pmb { w } ^ { \\beta \\top } } \\end{array} \\right\\} \\left[ \\begin{array} { l } { \\pmb { x } _ { t } , \\pmb { h } _ { t - 1 } , \\pmb { r } _ { t } } \\end{array} \\right] , \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1095, + 528, + 1127, + 528, + 1127, + 557, + 1095, + 557 + ], + "score": 0.89, + "latex": "d _ { h }" + }, + { + "category_id": 13, + "poly": [ + 493, + 1893, + 1039, + 1893, + 1039, + 1927, + 493, + 1927 + ], + "score": 0.88, + "latex": ", \\pmb { W } ^ { g } \\in \\mathbb { R } ^ { d _ { h } \\times ( d _ { i } + d _ { h } + d _ { r } ) } , \\pmb { b } \\in \\mathbb { R } ^ { 3 d _ { h } } , \\pmb { b } ^ { g } \\in \\mathbb { R } ^ { d _ { h } }" + }, + { + "category_id": 13, + "poly": [ + 488, + 1927, + 519, + 1927, + 519, + 1956, + 488, + 1956 + ], + "score": 0.88, + "latex": "\\boldsymbol { h } _ { t }" + }, + { + "category_id": 14, + "poly": [ + 672, + 423, + 1024, + 423, + 1024, + 462, + 672, + 462 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\pmb { h } _ { t } = \\sigma ( \\pmb { U } \\pmb { x } _ { t } + \\pmb { W } \\pmb { h } _ { t - 1 } + \\pmb { b } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1332, + 1861, + 1404, + 1861, + 1404, + 1892, + 1332, + 1892 + ], + "score": 0.86, + "latex": "W \\in" + }, + { + "category_id": 13, + "poly": [ + 1244, + 364, + 1272, + 364, + 1272, + 391, + 1244, + 391 + ], + "score": 0.86, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { y } } _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 463, + 362, + 528, + 362, + 528, + 391, + 463, + 391 + ], + "score": 0.84, + "latex": "d _ { i } , d _ { o }" + }, + { + "category_id": 13, + "poly": [ + 570, + 1930, + 597, + 1930, + 597, + 1956, + 570, + 1956 + ], + "score": 0.83, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1859, + 1266, + 1859, + 1266, + 1892, + 373, + 1892 + ], + "score": 0.81, + "latex": "h _ { t - 1 } , h _ { t } , c _ { t } , i _ { t } , f _ { t } , o _ { t } , g _ { t } \\in \\mathbb { R } ^ { d _ { h } } , { r } _ { t } \\in \\mathbb { R } ^ { d _ { r } } , { \\pmb x } _ { t } \\in \\mathbb { R } ^ { d _ { i } } , { \\pmb w } ^ { \\alpha } , { \\pmb w } ^ { \\beta } \\in \\mathbb { R } ^ { d _ { i } + d _ { h } + d _ { r } }" + }, + { + "category_id": 14, + "poly": [ + 755, + 477, + 943, + 477, + 943, + 513, + 755, + 513 + ], + "score": 0.81, + "latex": "\\pmb { y } _ { t } = \\pmb { V } \\pmb { h } _ { t } + \\pmb { c } ," + }, + { + "category_id": 13, + "poly": [ + 978, + 995, + 1105, + 995, + 1105, + 1029, + 978, + 1029 + ], + "score": 0.81, + "latex": "b _ { 0 } \\in \\mathbb { R } ^ { 4 d _ { h } }" + }, + { + "category_id": 13, + "poly": [ + 815, + 1033, + 829, + 1033, + 829, + 1057, + 815, + 1057 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1121, + 997, + 1249, + 997, + 1249, + 1030, + 1121, + 1030 + ], + "score": 0.69, + "latex": "i _ { t } , f _ { t } , g _ { t } o _ { t }" + }, + { + "category_id": 14, + "poly": [ + 616, + 1687, + 1078, + 1687, + 1078, + 1728, + 616, + 1728 + ], + "score": 0.64, + "latex": "g _ { t } = \\operatorname { t a n h } ( W ^ { g } [ x _ { t } , \\alpha _ { t } h _ { t - 1 } , \\beta _ { t } { r } _ { t } ] + b ^ { g } ) ," + }, + { + "category_id": 14, + "poly": [ + 698, + 890, + 998, + 890, + 998, + 930, + 698, + 930 + ], + "score": 0.64, + "latex": "\\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + \\pmb { i } _ { t } \\circ \\pmb { g } _ { t } ," + }, + { + "category_id": 14, + "poly": [ + 732, + 943, + 964, + 943, + 964, + 981, + 732, + 981 + ], + "score": 0.62, + "latex": "h _ { t } = o _ { t } \\circ \\operatorname { t a n h } ( c _ { t } ) ," + }, + { + "category_id": 14, + "poly": [ + 693, + 1742, + 1002, + 1742, + 1002, + 1781, + 693, + 1781 + ], + "score": 0.56, + "latex": "\\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + i _ { t } \\circ \\pmb { g } _ { t } ," + }, + { + "category_id": 13, + "poly": [ + 372, + 1003, + 393, + 1003, + 393, + 1025, + 372, + 1025 + ], + "score": 0.49, + "latex": "\\circ" + }, + { + "category_id": 14, + "poly": [ + 729, + 1794, + 968, + 1794, + 968, + 1832, + 729, + 1832 + ], + "score": 0.33, + "latex": "h _ { t } = o _ { t } \\circ \\operatorname { t a n h } ( c _ { t } ) ," + }, + { + "category_id": 14, + "poly": [ + 702, + 886, + 999, + 886, + 999, + 984, + 702, + 984 + ], + "score": 0.29, + "latex": "\\begin{array} { r } { \\pmb { c } _ { t } = \\pmb { f } _ { t } \\circ \\pmb { c } _ { t - 1 } + \\pmb { i } _ { t } \\circ \\pmb { g } _ { t } , } \\\\ { \\pmb { h } _ { t } = \\pmb { o } _ { t } \\circ \\operatorname { t a n h } ( \\pmb { c } _ { t } ) , } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1137.0, + 1195.0, + 1137.0, + 1195.0, + 1181.0, + 296.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 223.0, + 1101.0, + 223.0, + 1101.0, + 268.0, + 297.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1312.0, + 633.0, + 1312.0, + 633.0, + 1351.0, + 293.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 297.0, + 1405.0, + 297.0, + 1405.0, + 332.0, + 294.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 324.0, + 409.0, + 324.0, + 409.0, + 368.0, + 288.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 324.0, + 1007.0, + 324.0, + 1007.0, + 368.0, + 596.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 324.0, + 1273.0, + 324.0, + 1273.0, + 368.0, + 1191.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 324.0, + 1407.0, + 324.0, + 1407.0, + 368.0, + 1393.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 357.0, + 297.0, + 357.0, + 297.0, + 397.0, + 292.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 357.0, + 462.0, + 357.0, + 462.0, + 397.0, + 404.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 357.0, + 1243.0, + 357.0, + 1243.0, + 397.0, + 529.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 357.0, + 1405.0, + 357.0, + 1405.0, + 397.0, + 1273.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 388.0, + 532.0, + 388.0, + 532.0, + 425.0, + 294.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 603.0, + 1405.0, + 603.0, + 1405.0, + 637.0, + 296.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 632.0, + 1312.0, + 632.0, + 1312.0, + 672.0, + 292.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 665.0, + 1065.0, + 665.0, + 1065.0, + 699.0, + 294.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 986.0, + 371.0, + 986.0, + 371.0, + 1035.0, + 291.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 986.0, + 724.0, + 986.0, + 724.0, + 1035.0, + 394.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 986.0, + 977.0, + 986.0, + 977.0, + 1035.0, + 963.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 986.0, + 1120.0, + 986.0, + 1120.0, + 1035.0, + 1106.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 986.0, + 1409.0, + 986.0, + 1409.0, + 1035.0, + 1250.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 814.0, + 1029.0, + 814.0, + 1061.0, + 293.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1061.0, + 830.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1059.0, + 454.0, + 1059.0, + 454.0, + 1088.0, + 292.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1851.0, + 372.0, + 1851.0, + 372.0, + 1899.0, + 291.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1851.0, + 1331.0, + 1851.0, + 1331.0, + 1899.0, + 1267.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1405.0, + 1851.0, + 1408.0, + 1851.0, + 1408.0, + 1899.0, + 1405.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1884.0, + 492.0, + 1884.0, + 492.0, + 1930.0, + 291.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1884.0, + 1408.0, + 1884.0, + 1408.0, + 1930.0, + 1040.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 487.0, + 1924.0, + 487.0, + 1960.0, + 295.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1924.0, + 569.0, + 1924.0, + 569.0, + 1960.0, + 520.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1924.0, + 611.0, + 1924.0, + 611.0, + 1960.0, + 598.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1209.0, + 1404.0, + 1209.0, + 1404.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 1204.0, + 1239.0, + 1204.0, + 1278.0, + 294.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 521.0, + 567.0, + 521.0, + 567.0, + 561.0, + 294.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 521.0, + 1094.0, + 521.0, + 1094.0, + 561.0, + 694.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 561.0, + 1128.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 553.0, + 754.0, + 553.0, + 754.0, + 592.0, + 467.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1967.0, + 1407.0, + 1967.0, + 1407.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1998.0, + 396.0, + 1998.0, + 396.0, + 2038.0, + 291.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1371.0, + 907.0, + 1371.0, + 907.0, + 1411.0, + 293.0, + 1411.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1580, + 1404, + 1580, + 1404, + 1827, + 298, + 1827 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 898, + 1403, + 898, + 1403, + 992, + 299, + 992 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 709, + 1403, + 709, + 1403, + 802, + 298, + 802 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 301, + 1942, + 1402, + 1942, + 1402, + 2034, + 301, + 2034 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1447, + 1403, + 1447, + 1403, + 1569, + 298, + 1569 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1006, + 1402, + 1006, + 1402, + 1100, + 299, + 1100 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 424, + 1403, + 424, + 1403, + 578, + 298, + 578 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 414, + 298, + 414 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 301, + 1247, + 1402, + 1247, + 1402, + 1341, + 301, + 1341 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 296, + 588, + 1399, + 588, + 1399, + 681, + 296, + 681 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1355, + 1398, + 1355, + 1398, + 1418, + 296, + 1418 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 300, + 840, + 588, + 840, + 588, + 872, + 300, + 872 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 303, + 1870, + 821, + 1870, + 821, + 1907, + 303, + 1907 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1190, + 1399, + 1190, + 1399, + 1220, + 1353, + 1220 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1147, + 1399, + 1147, + 1399, + 1177, + 1352, + 1177 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 8, + "poly": [ + 583, + 1141, + 1114, + 1141, + 1114, + 1223, + 583, + 1223 + ], + "score": 0.497 + }, + { + "category_id": 8, + "poly": [ + 585, + 1141, + 1113, + 1141, + 1113, + 1182, + 585, + 1182 + ], + "score": 0.422 + }, + { + "category_id": 8, + "poly": [ + 699, + 1189, + 997, + 1189, + 997, + 1223, + 699, + 1223 + ], + "score": 0.278 + }, + { + "category_id": 13, + "poly": [ + 375, + 1248, + 524, + 1248, + 524, + 1282, + 375, + 1282 + ], + "score": 0.93, + "latex": "\\{ a , W ^ { \\gamma } , b ^ { \\gamma } \\}" + }, + { + "category_id": 13, + "poly": [ + 1162, + 2001, + 1347, + 2001, + 1347, + 2036, + 1162, + 2036 + ], + "score": 0.93, + "latex": "\\tau \\leq ( n - 1 ) ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1448, + 1307, + 1448, + 1307, + 1481, + 1239, + 1481 + ], + "score": 0.92, + "latex": "M _ { t } [ i ]" + }, + { + "category_id": 13, + "poly": [ + 803, + 232, + 871, + 232, + 871, + 262, + 803, + 262 + ], + "score": 0.91, + "latex": "\\alpha _ { t } , \\beta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1054, + 897, + 1257, + 897, + 1257, + 930, + 1054, + 930 + ], + "score": 0.91, + "latex": "\\mathbf { \\boldsymbol { A } } _ { t } ~ \\in ~ \\mathbb { R } ^ { n _ { m e m } \\times d _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 595, + 929, + 777, + 929, + 777, + 963, + 595, + 963 + ], + "score": 0.91, + "latex": "M _ { t } = [ A _ { t } ; C _ { t } ]" + }, + { + "category_id": 14, + "poly": [ + 581, + 1137, + 1116, + 1137, + 1116, + 1227, + 581, + 1227 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { \\pi _ { t } [ i ] = a ^ { \\top } \\operatorname { t a n h } ( W ^ { \\gamma } [ h _ { t } , x _ { t } , M _ { t } [ i ] , { \\boldsymbol u } _ { t } ] + b ^ { \\gamma } ) , } } \\\\ { { s _ { t } = \\mathrm { g u m b e l - s o f t m a x } ( \\pi _ { t } ) . } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 943, + 1450, + 974, + 1450, + 974, + 1479, + 943, + 1479 + ], + "score": 0.89, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 963, + 487, + 995, + 487, + 995, + 516, + 963, + 516 + ], + "score": 0.89, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 950, + 1583, + 986, + 1583, + 986, + 1613, + 950, + 1613 + ], + "score": 0.89, + "latex": "\\pmb { A } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 690, + 590, + 722, + 590, + 722, + 621, + 690, + 621 + ], + "score": 0.88, + "latex": "\\mathbf { \\delta } _ { o _ { t } ^ { r } }" + }, + { + "category_id": 13, + "poly": [ + 860, + 1283, + 891, + 1283, + 891, + 1310, + 860, + 1310 + ], + "score": 0.87, + "latex": "\\mathbf { \\Delta } \\mathbf { u } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 386, + 929, + 579, + 929, + 579, + 960, + 386, + 960 + ], + "score": 0.86, + "latex": "C _ { t } \\ \\stackrel { \\cdot } { \\in } \\ \\mathbb { R } ^ { n _ { m e m } \\times d _ { r } }" + }, + { + "category_id": 13, + "poly": [ + 1043, + 595, + 1072, + 595, + 1072, + 619, + 1043, + 619 + ], + "score": 0.86, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 537, + 901, + 581, + 901, + 581, + 929, + 537, + 929 + ], + "score": 0.86, + "latex": "M _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1184, + 1586, + 1216, + 1586, + 1216, + 1613, + 1184, + 1613 + ], + "score": 0.85, + "latex": "\\mathbf { \\pmb { u } } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 859, + 1454, + 890, + 1454, + 890, + 1480, + 859, + 1480 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1131, + 429, + 1160, + 429, + 1160, + 457, + 1131, + 457 + ], + "score": 0.85, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { g } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1026, + 624, + 1054, + 624, + 1054, + 650, + 1026, + 650 + ], + "score": 0.85, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 884, + 490, + 912, + 490, + 912, + 516, + 884, + 516 + ], + "score": 0.84, + "latex": "\\mathbf { \\nabla } _ { \\mathbf { r } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 839, + 551, + 867, + 551, + 867, + 577, + 839, + 577 + ], + "score": 0.84, + "latex": "\\mathbf { } _ { \\mathbf { } } ^ { \\mathbf { } } \\mathbf { \\Delta } \\mathbf { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 430, + 1286, + 451, + 1286, + 451, + 1307, + 430, + 1307 + ], + "score": 0.77, + "latex": "\\textbf { \\em a }" + }, + { + "category_id": 13, + "poly": [ + 686, + 1619, + 708, + 1619, + 708, + 1640, + 686, + 1640 + ], + "score": 0.65, + "latex": "^ { a }" + }, + { + "category_id": 13, + "poly": [ + 1039, + 1769, + 1057, + 1769, + 1057, + 1792, + 1039, + 1792 + ], + "score": 0.48, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 297, + 742, + 401, + 742, + 401, + 773, + 297, + 773 + ], + "score": 0.41, + "latex": "p \\mathrm { M N I S T }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 837.0, + 592.0, + 837.0, + 592.0, + 877.0, + 294.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1868.0, + 825.0, + 1868.0, + 825.0, + 1912.0, + 298.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1579.0, + 949.0, + 1579.0, + 949.0, + 1616.0, + 295.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1579.0, + 1183.0, + 1579.0, + 1183.0, + 1616.0, + 987.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1579.0, + 1404.0, + 1579.0, + 1404.0, + 1616.0, + 1217.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1612.0, + 685.0, + 1612.0, + 685.0, + 1644.0, + 293.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1612.0, + 1405.0, + 1612.0, + 1405.0, + 1644.0, + 709.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1643.0, + 1405.0, + 1643.0, + 1405.0, + 1677.0, + 294.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1673.0, + 1405.0, + 1673.0, + 1405.0, + 1707.0, + 293.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 1406.0, + 1702.0, + 1406.0, + 1738.0, + 293.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1734.0, + 1405.0, + 1734.0, + 1405.0, + 1767.0, + 294.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1764.0, + 1038.0, + 1764.0, + 1038.0, + 1797.0, + 293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1764.0, + 1404.0, + 1764.0, + 1404.0, + 1797.0, + 1058.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1792.0, + 487.0, + 1792.0, + 487.0, + 1831.0, + 294.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 894.0, + 536.0, + 894.0, + 536.0, + 933.0, + 292.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 894.0, + 1053.0, + 894.0, + 1053.0, + 933.0, + 582.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 894.0, + 1408.0, + 894.0, + 1408.0, + 933.0, + 1258.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 924.0, + 385.0, + 924.0, + 385.0, + 964.0, + 292.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 924.0, + 594.0, + 924.0, + 594.0, + 964.0, + 580.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 924.0, + 1408.0, + 924.0, + 1408.0, + 964.0, + 778.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 958.0, + 1208.0, + 958.0, + 1208.0, + 998.0, + 292.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1404.0, + 710.0, + 1404.0, + 743.0, + 294.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 741.0, + 1404.0, + 741.0, + 1404.0, + 775.0, + 402.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 772.0, + 586.0, + 772.0, + 586.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 1161.0, + 2001.0, + 1161.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2038.0, + 1348.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1448.0, + 858.0, + 1448.0, + 858.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1448.0, + 942.0, + 1448.0, + 942.0, + 1484.0, + 891.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1448.0, + 1238.0, + 1448.0, + 1238.0, + 1484.0, + 975.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1448.0, + 1404.0, + 1448.0, + 1404.0, + 1484.0, + 1308.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1478.0, + 1405.0, + 1478.0, + 1405.0, + 1514.0, + 293.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1508.0, + 1404.0, + 1508.0, + 1404.0, + 1546.0, + 292.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1539.0, + 358.0, + 1539.0, + 358.0, + 1572.0, + 293.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1003.0, + 1404.0, + 1003.0, + 1404.0, + 1046.0, + 293.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1072.0, + 293.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1069.0, + 819.0, + 1069.0, + 819.0, + 1103.0, + 297.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 423.0, + 1130.0, + 423.0, + 1130.0, + 460.0, + 294.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 423.0, + 1405.0, + 423.0, + 1405.0, + 460.0, + 1161.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 453.0, + 1404.0, + 453.0, + 1404.0, + 490.0, + 296.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 485.0, + 883.0, + 485.0, + 883.0, + 518.0, + 294.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 485.0, + 962.0, + 485.0, + 962.0, + 518.0, + 913.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 485.0, + 1403.0, + 485.0, + 1403.0, + 518.0, + 996.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 512.0, + 1407.0, + 512.0, + 1407.0, + 557.0, + 292.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 546.0, + 838.0, + 546.0, + 838.0, + 582.0, + 292.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 546.0, + 1060.0, + 546.0, + 1060.0, + 582.0, + 868.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 230.0, + 802.0, + 230.0, + 802.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 266.0, + 872.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1406.0, + 262.0, + 1406.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 323.0, + 1403.0, + 323.0, + 1403.0, + 355.0, + 296.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1406.0, + 352.0, + 1406.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 1342.0, + 381.0, + 1342.0, + 417.0, + 293.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1248.0, + 374.0, + 1248.0, + 374.0, + 1282.0, + 296.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1248.0, + 1402.0, + 1248.0, + 1402.0, + 1282.0, + 525.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1278.0, + 429.0, + 1278.0, + 429.0, + 1313.0, + 294.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1278.0, + 859.0, + 1278.0, + 859.0, + 1313.0, + 452.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1278.0, + 1406.0, + 1278.0, + 1406.0, + 1313.0, + 892.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1307.0, + 893.0, + 1307.0, + 893.0, + 1345.0, + 296.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 585.0, + 689.0, + 585.0, + 689.0, + 624.0, + 294.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 585.0, + 1042.0, + 585.0, + 1042.0, + 624.0, + 723.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 585.0, + 1404.0, + 585.0, + 1404.0, + 624.0, + 1073.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 618.0, + 1025.0, + 618.0, + 1025.0, + 654.0, + 295.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 618.0, + 1404.0, + 618.0, + 1404.0, + 654.0, + 1055.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 649.0, + 1157.0, + 649.0, + 1157.0, + 686.0, + 295.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1353.0, + 1402.0, + 1353.0, + 1402.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1387.0, + 564.0, + 1387.0, + 564.0, + 1422.0, + 296.0, + 1422.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1444, + 1404, + 1444, + 1404, + 1660, + 298, + 1660 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1096, + 1404, + 1096, + 1404, + 1343, + 298, + 1343 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1754, + 1403, + 1754, + 1403, + 1878, + 299, + 1878 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 959, + 1403, + 959, + 1403, + 1082, + 300, + 1082 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 563, + 644, + 1135, + 644, + 1135, + 885, + 563, + 885 + ], + "score": 0.974, + "html": "
TaskSetup
Copycopy length =1~50
Repeat Copycopyl length=1~10 number of repeats = 1~10
Associative Recallitem size=3×6 item number = 2~6
Priority SortSort length = 40
" + }, + { + "category_id": 5, + "poly": [ + 547, + 292, + 1154, + 292, + 1154, + 530, + 547, + 530 + ], + "score": 0.972, + "html": "
ModelHidden Sizememory SizeParam Count
LSTM300一376k
NTM120128×2088k
TARDIS12050×3290k
ARMIN+TARDIS10050×3291k
ARMIN10050×3290k
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1402, + 1972, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.946 + }, + { + "category_id": 6, + "poly": [ + 565, + 575, + 1134, + 575, + 1134, + 609, + 565, + 609 + ], + "score": 0.916 + }, + { + "category_id": 6, + "poly": [ + 503, + 224, + 1197, + 224, + 1197, + 258, + 503, + 258 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 299, + 1916, + 563, + 1916, + 563, + 1948, + 299, + 1948 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 300, + 1386, + 627, + 1386, + 627, + 1417, + 300, + 1417 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 299, + 1698, + 462, + 1698, + 462, + 1730, + 299, + 1730 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 837, + 2089, + 864, + 2089, + 864, + 2112, + 837, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 13, + "poly": [ + 917, + 1251, + 946, + 1251, + 946, + 1283, + 917, + 1283 + ], + "score": 0.87, + "latex": "\\pmb { f } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 618, + 1053, + 683, + 1053, + 683, + 1080, + 618, + 1080 + ], + "score": 0.85, + "latex": "n - 1" + }, + { + "category_id": 13, + "poly": [ + 436, + 1602, + 456, + 1602, + 456, + 1625, + 436, + 1625 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 652, + 996, + 671, + 996, + 671, + 1018, + 652, + 1018 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 891, + 1026, + 911, + 1026, + 911, + 1049, + 891, + 1049 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1003, + 851, + 1061, + 851, + 1061, + 877, + 1003, + 877 + ], + "score": 0.74, + "latex": "= 4 0" + }, + { + "category_id": 13, + "poly": [ + 1025, + 756, + 1118, + 756, + 1118, + 784, + 1025, + 784 + ], + "score": 0.69, + "latex": "= 1 { \\sim } 1 0" + }, + { + "category_id": 13, + "poly": [ + 1003, + 819, + 1082, + 819, + 1082, + 846, + 1003, + 846 + ], + "score": 0.68, + "latex": "= 2 { \\sim } 6" + }, + { + "category_id": 13, + "poly": [ + 990, + 695, + 1082, + 695, + 1082, + 722, + 990, + 722 + ], + "score": 0.53, + "latex": "= 1 { \\sim } 5 0" + }, + { + "category_id": 13, + "poly": [ + 761, + 1572, + 779, + 1572, + 779, + 1594, + 761, + 1594 + ], + "score": 0.46, + "latex": "+ { \\bf \\Phi }" + }, + { + "category_id": 13, + "poly": [ + 990, + 727, + 1082, + 727, + 1082, + 753, + 990, + 753 + ], + "score": 0.43, + "latex": "= 1 { \\sim } 1 0" + }, + { + "category_id": 13, + "poly": [ + 972, + 788, + 1060, + 788, + 1060, + 815, + 972, + 815 + ], + "score": 0.29, + "latex": "\\stackrel { \\_ } { \\mathrm { 1 } } = 3 \\times 6" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 569.0, + 1135.0, + 569.0, + 1135.0, + 614.0, + 563.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 222.0, + 1197.0, + 222.0, + 1197.0, + 262.0, + 501.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 569.0, + 1913.0, + 569.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1385.0, + 632.0, + 1385.0, + 632.0, + 1421.0, + 296.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 469.0, + 1695.0, + 469.0, + 1736.0, + 294.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1443.0, + 1405.0, + 1443.0, + 1405.0, + 1480.0, + 293.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 1406.0, + 1477.0, + 1406.0, + 1511.0, + 296.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1503.0, + 1405.0, + 1503.0, + 1405.0, + 1545.0, + 291.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1569.0, + 296.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1567.0, + 760.0, + 1567.0, + 760.0, + 1600.0, + 293.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1567.0, + 1406.0, + 1567.0, + 1406.0, + 1600.0, + 780.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 435.0, + 1599.0, + 435.0, + 1630.0, + 293.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1599.0, + 1406.0, + 1599.0, + 1406.0, + 1630.0, + 457.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1629.0, + 1273.0, + 1629.0, + 1273.0, + 1663.0, + 296.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1406.0, + 1097.0, + 1406.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1130.0, + 1405.0, + 1130.0, + 1405.0, + 1164.0, + 293.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1159.0, + 1406.0, + 1159.0, + 1406.0, + 1195.0, + 292.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1221.0, + 1406.0, + 1221.0, + 1406.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1247.0, + 916.0, + 1247.0, + 916.0, + 1288.0, + 292.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1247.0, + 1405.0, + 1247.0, + 1405.0, + 1288.0, + 947.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1279.0, + 1408.0, + 1279.0, + 1408.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1311.0, + 358.0, + 1311.0, + 358.0, + 1347.0, + 293.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1752.0, + 1407.0, + 1752.0, + 1407.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1783.0, + 1402.0, + 1783.0, + 1402.0, + 1821.0, + 293.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1816.0, + 1405.0, + 1816.0, + 1405.0, + 1850.0, + 293.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1850.0, + 892.0, + 1850.0, + 892.0, + 1879.0, + 297.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 959.0, + 1404.0, + 959.0, + 1404.0, + 994.0, + 295.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 990.0, + 651.0, + 990.0, + 651.0, + 1023.0, + 296.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 990.0, + 1404.0, + 990.0, + 1404.0, + 1023.0, + 672.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 890.0, + 1022.0, + 890.0, + 1054.0, + 295.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1022.0, + 1404.0, + 1022.0, + 1404.0, + 1054.0, + 912.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1052.0, + 617.0, + 1052.0, + 617.0, + 1084.0, + 295.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1052.0, + 1404.0, + 1052.0, + 1404.0, + 1084.0, + 684.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 964, + 1404, + 964, + 1404, + 1210, + 298, + 1210 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1318, + 1404, + 1318, + 1404, + 1533, + 298, + 1533 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 402, + 1403, + 402, + 1403, + 524, + 299, + 524 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 635, + 1403, + 635, + 1403, + 788, + 299, + 788 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1650, + 1403, + 1650, + 1403, + 1804, + 299, + 1804 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 295, + 230, + 1400, + 230, + 1400, + 293, + 295, + 293 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 299, + 574, + 582, + 574, + 582, + 606, + 299, + 606 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 299, + 341, + 661, + 341, + 661, + 374, + 299, + 374 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 301, + 1260, + 816, + 1260, + 816, + 1291, + 301, + 1291 + ], + "score": 0.874 + }, + { + "category_id": 0, + "poly": [ + 302, + 904, + 711, + 904, + 711, + 935, + 302, + 935 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.826 + }, + { + "category_id": 0, + "poly": [ + 301, + 1588, + 1169, + 1588, + 1169, + 1619, + 301, + 1619 + ], + "score": 0.734 + }, + { + "category_id": 1, + "poly": [ + 301, + 842, + 1279, + 842, + 1279, + 873, + 301, + 873 + ], + "score": 0.685 + }, + { + "category_id": 0, + "poly": [ + 301, + 842, + 1279, + 842, + 1279, + 873, + 301, + 873 + ], + "score": 0.27 + }, + { + "category_id": 1, + "poly": [ + 301, + 1588, + 1169, + 1588, + 1169, + 1619, + 301, + 1619 + ], + "score": 0.162 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 987, + 1121, + 1058, + 1121, + 1058, + 1149, + 987, + 1149 + ], + "score": 0.89, + "latex": "n _ { m e m }" + }, + { + "category_id": 13, + "poly": [ + 477, + 1119, + 504, + 1119, + 504, + 1148, + 477, + 1148 + ], + "score": 0.87, + "latex": "d _ { i }" + }, + { + "category_id": 13, + "poly": [ + 790, + 1445, + 861, + 1445, + 861, + 1473, + 790, + 1473 + ], + "score": 0.86, + "latex": "n _ { m e m }" + }, + { + "category_id": 13, + "poly": [ + 900, + 1946, + 971, + 1946, + 971, + 1973, + 900, + 1973 + ], + "score": 0.85, + "latex": "n _ { m e m }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 569.0, + 587.0, + 569.0, + 587.0, + 611.0, + 294.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 339.0, + 666.0, + 339.0, + 666.0, + 377.0, + 295.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1259.0, + 821.0, + 1259.0, + 821.0, + 1294.0, + 297.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 903.0, + 716.0, + 903.0, + 716.0, + 939.0, + 297.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 833.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1586.0, + 1175.0, + 1586.0, + 1175.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 841.0, + 1284.0, + 841.0, + 1284.0, + 876.0, + 295.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 964.0, + 1405.0, + 964.0, + 1405.0, + 997.0, + 294.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 994.0, + 1406.0, + 994.0, + 1406.0, + 1032.0, + 293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1025.0, + 1405.0, + 1025.0, + 1405.0, + 1061.0, + 293.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1085.0, + 1408.0, + 1085.0, + 1408.0, + 1123.0, + 292.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1117.0, + 476.0, + 1117.0, + 476.0, + 1154.0, + 293.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1117.0, + 986.0, + 1117.0, + 986.0, + 1154.0, + 505.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1117.0, + 1405.0, + 1117.0, + 1405.0, + 1154.0, + 1059.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1147.0, + 1405.0, + 1147.0, + 1405.0, + 1183.0, + 292.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1178.0, + 1152.0, + 1178.0, + 1152.0, + 1211.0, + 293.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1318.0, + 1405.0, + 1318.0, + 1405.0, + 1353.0, + 296.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1385.0, + 294.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1379.0, + 1405.0, + 1379.0, + 1405.0, + 1417.0, + 293.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1413.0, + 1404.0, + 1413.0, + 1404.0, + 1444.0, + 294.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1440.0, + 789.0, + 1440.0, + 789.0, + 1478.0, + 293.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1440.0, + 1405.0, + 1440.0, + 1405.0, + 1478.0, + 862.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1472.0, + 1405.0, + 1472.0, + 1405.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1502.0, + 804.0, + 1502.0, + 804.0, + 1536.0, + 294.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 403.0, + 1404.0, + 403.0, + 1404.0, + 435.0, + 295.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 433.0, + 1405.0, + 433.0, + 1405.0, + 465.0, + 294.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 465.0, + 1405.0, + 465.0, + 1405.0, + 498.0, + 295.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 493.0, + 842.0, + 493.0, + 842.0, + 525.0, + 295.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 634.0, + 1404.0, + 634.0, + 1404.0, + 669.0, + 295.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 702.0, + 294.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 696.0, + 1404.0, + 696.0, + 1404.0, + 731.0, + 294.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 726.0, + 1407.0, + 726.0, + 1407.0, + 763.0, + 295.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 755.0, + 1200.0, + 755.0, + 1200.0, + 795.0, + 294.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1649.0, + 1405.0, + 1649.0, + 1405.0, + 1686.0, + 295.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1681.0, + 1403.0, + 1681.0, + 1403.0, + 1714.0, + 297.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1708.0, + 1404.0, + 1708.0, + 1404.0, + 1751.0, + 293.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 660.0, + 1773.0, + 660.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1821.0, + 1404.0, + 1821.0, + 1404.0, + 1855.0, + 296.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1407.0, + 1912.0, + 1407.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 899.0, + 1939.0, + 899.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1979.0, + 972.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 469.0, + 2002.0, + 469.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1402.0, + 228.0, + 1402.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 968.0, + 262.0, + 968.0, + 297.0, + 296.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 841.0, + 1284.0, + 841.0, + 1284.0, + 876.0, + 295.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1586.0, + 1175.0, + 1586.0, + 1175.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 312, + 828, + 1379, + 828, + 1379, + 1294, + 312, + 1294 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 311, + 1532, + 1372, + 1532, + 1372, + 1994, + 311, + 1994 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 311, + 244, + 1390, + 244, + 1390, + 708, + 311, + 708 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 695, + 1302, + 1003, + 1302, + 1003, + 1336, + 695, + 1336 + ], + "score": 0.908 + }, + { + "category_id": 4, + "poly": [ + 665, + 2002, + 1034, + 2002, + 1034, + 2034, + 665, + 2034 + ], + "score": 0.905 + }, + { + "category_id": 4, + "poly": [ + 732, + 716, + 965, + 716, + 965, + 750, + 732, + 750 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 855, + 76, + 855, + 104, + 299, + 104 + ], + "score": 0.097 + }, + { + "category_id": 15, + "poly": [ + 895.0, + 832.0, + 927.0, + 832.0, + 927.0, + 855.0, + 895.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 842.0, + 698.0, + 842.0, + 698.0, + 870.0, + 639.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 842.0, + 1248.0, + 842.0, + 1248.0, + 870.0, + 1190.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 861.0, + 694.0, + 861.0, + 694.0, + 890.0, + 641.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 861.0, + 1244.0, + 861.0, + 1244.0, + 890.0, + 1191.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 882.0, + 716.0, + 882.0, + 716.0, + 911.0, + 640.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 882.0, + 1265.0, + 882.0, + 1265.0, + 911.0, + 1189.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 897.0, + 376.0, + 897.0, + 376.0, + 1123.0, + 313.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 903.0, + 821.0, + 903.0, + 821.0, + 951.0, + 635.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 897.0, + 900.0, + 897.0, + 900.0, + 1124.0, + 870.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 920.0, + 926.0, + 920.0, + 926.0, + 938.0, + 904.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 903.0, + 1368.0, + 903.0, + 1368.0, + 953.0, + 1187.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1000.0, + 376.0, + 1000.0, + 376.0, + 1022.0, + 346.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1001.0, + 926.0, + 1001.0, + 926.0, + 1021.0, + 904.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1083.0, + 374.0, + 1083.0, + 374.0, + 1105.0, + 339.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1085.0, + 926.0, + 1085.0, + 926.0, + 1103.0, + 904.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1173.0, + 371.0, + 1173.0, + 371.0, + 1185.0, + 362.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1171.0, + 927.0, + 1171.0, + 927.0, + 1186.0, + 911.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1185.0, + 391.0, + 1185.0, + 391.0, + 1200.0, + 375.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1181.0, + 495.0, + 1181.0, + 495.0, + 1205.0, + 447.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1180.0, + 584.0, + 1180.0, + 584.0, + 1205.0, + 537.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1180.0, + 674.0, + 1180.0, + 674.0, + 1205.0, + 624.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1181.0, + 762.0, + 1181.0, + 762.0, + 1205.0, + 713.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1181.0, + 950.0, + 1181.0, + 950.0, + 1204.0, + 917.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1181.0, + 1047.0, + 1181.0, + 1047.0, + 1205.0, + 997.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1180.0, + 1135.0, + 1180.0, + 1135.0, + 1205.0, + 1086.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1180.0, + 1224.0, + 1180.0, + 1224.0, + 1205.0, + 1175.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1181.0, + 1311.0, + 1181.0, + 1311.0, + 1205.0, + 1262.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1205.0, + 648.0, + 1205.0, + 648.0, + 1235.0, + 562.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1206.0, + 1197.0, + 1206.0, + 1197.0, + 1233.0, + 1112.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1262.0, + 649.0, + 1262.0, + 649.0, + 1293.0, + 500.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1261.0, + 1284.0, + 1261.0, + 1284.0, + 1296.0, + 976.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1538.0, + 699.0, + 1538.0, + 699.0, + 1568.0, + 640.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1533.0, + 923.0, + 1533.0, + 923.0, + 1554.0, + 900.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1541.0, + 1242.0, + 1541.0, + 1242.0, + 1567.0, + 1186.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1562.0, + 694.0, + 1562.0, + 694.0, + 1589.0, + 643.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1562.0, + 1237.0, + 1562.0, + 1237.0, + 1589.0, + 1187.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1584.0, + 713.0, + 1584.0, + 713.0, + 1609.0, + 644.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1584.0, + 1258.0, + 1584.0, + 1258.0, + 1609.0, + 1189.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1596.0, + 376.0, + 1596.0, + 376.0, + 1824.0, + 313.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1605.0, + 820.0, + 1605.0, + 820.0, + 1629.0, + 642.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1596.0, + 924.0, + 1596.0, + 924.0, + 1823.0, + 867.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1605.0, + 1362.0, + 1605.0, + 1362.0, + 1629.0, + 1185.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1626.0, + 708.0, + 1626.0, + 708.0, + 1651.0, + 641.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1626.0, + 1250.0, + 1626.0, + 1250.0, + 1651.0, + 1185.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1699.0, + 376.0, + 1699.0, + 376.0, + 1723.0, + 347.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1703.0, + 921.0, + 1703.0, + 921.0, + 1720.0, + 907.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1708.0, + 1046.0, + 1708.0, + 1046.0, + 1723.0, + 1028.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1719.0, + 544.0, + 1719.0, + 544.0, + 1734.0, + 527.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1735.0, + 618.0, + 1735.0, + 618.0, + 1749.0, + 597.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1739.0, + 675.0, + 1739.0, + 675.0, + 1762.0, + 633.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1782.0, + 375.0, + 1782.0, + 375.0, + 1804.0, + 340.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1782.0, + 923.0, + 1782.0, + 923.0, + 1804.0, + 894.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1798.0, + 1035.0, + 1798.0, + 1035.0, + 1817.0, + 1014.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1871.0, + 375.0, + 1871.0, + 375.0, + 1887.0, + 359.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1870.0, + 923.0, + 1870.0, + 923.0, + 1887.0, + 907.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1883.0, + 390.0, + 1883.0, + 390.0, + 1901.0, + 375.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1880.0, + 497.0, + 1880.0, + 497.0, + 1905.0, + 447.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1880.0, + 586.0, + 1880.0, + 586.0, + 1905.0, + 537.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1881.0, + 673.0, + 1881.0, + 673.0, + 1902.0, + 627.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1881.0, + 762.0, + 1881.0, + 762.0, + 1902.0, + 715.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1881.0, + 947.0, + 1881.0, + 947.0, + 1902.0, + 913.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1880.0, + 1042.0, + 1880.0, + 1042.0, + 1905.0, + 994.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1880.0, + 1131.0, + 1880.0, + 1131.0, + 1905.0, + 1081.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1881.0, + 1218.0, + 1881.0, + 1218.0, + 1902.0, + 1170.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1881.0, + 1305.0, + 1881.0, + 1305.0, + 1902.0, + 1258.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1904.0, + 647.0, + 1904.0, + 647.0, + 1933.0, + 564.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1906.0, + 1190.0, + 1906.0, + 1190.0, + 1931.0, + 1108.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1962.0, + 648.0, + 1962.0, + 648.0, + 1995.0, + 503.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1961.0, + 1285.0, + 1961.0, + 1285.0, + 1995.0, + 976.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1646.0, + 1016.0, + 1646.0, + 1016.0, + 1689.5, + 932.0, + 1689.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.75, + 1670.0, + 539.75, + 1670.0, + 539.75, + 1710.5, + 382.75, + 1710.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 246.0, + 370.0, + 246.0, + 370.0, + 270.0, + 340.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 247.0, + 928.0, + 247.0, + 928.0, + 270.0, + 895.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 258.0, + 702.0, + 258.0, + 702.0, + 283.0, + 644.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 258.0, + 1256.0, + 258.0, + 1256.0, + 283.0, + 1197.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 275.0, + 698.0, + 275.0, + 698.0, + 305.0, + 644.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 275.0, + 1252.0, + 275.0, + 1252.0, + 305.0, + 1197.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 297.0, + 719.0, + 297.0, + 719.0, + 326.0, + 645.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 297.0, + 1272.0, + 297.0, + 1272.0, + 326.0, + 1199.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 311.0, + 345.0, + 311.0, + 345.0, + 539.0, + 314.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 320.0, + 827.0, + 320.0, + 827.0, + 344.0, + 643.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 312.0, + 900.0, + 312.0, + 900.0, + 536.0, + 869.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 320.0, + 1377.0, + 320.0, + 1377.0, + 344.0, + 1197.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 330.0, + 370.0, + 330.0, + 370.0, + 355.0, + 341.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 332.0, + 927.0, + 332.0, + 927.0, + 354.0, + 903.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 341.0, + 712.0, + 341.0, + 712.0, + 366.0, + 643.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 341.0, + 1264.0, + 341.0, + 1264.0, + 366.0, + 1198.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 414.0, + 369.0, + 414.0, + 369.0, + 437.0, + 340.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 416.0, + 927.0, + 416.0, + 927.0, + 436.0, + 903.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 500.0, + 368.0, + 500.0, + 368.0, + 519.0, + 341.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 500.0, + 927.0, + 500.0, + 927.0, + 520.0, + 903.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 588.0, + 367.0, + 588.0, + 367.0, + 599.0, + 356.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 586.0, + 928.0, + 586.0, + 928.0, + 602.0, + 912.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 599.0, + 385.0, + 599.0, + 385.0, + 616.0, + 370.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 594.0, + 494.0, + 594.0, + 494.0, + 620.0, + 444.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 594.0, + 585.0, + 594.0, + 585.0, + 620.0, + 536.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 594.0, + 676.0, + 594.0, + 676.0, + 620.0, + 627.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 594.0, + 768.0, + 594.0, + 768.0, + 620.0, + 718.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 596.0, + 951.0, + 596.0, + 951.0, + 618.0, + 918.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 594.0, + 1051.0, + 594.0, + 1051.0, + 620.0, + 1000.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 594.0, + 1142.0, + 594.0, + 1142.0, + 620.0, + 1091.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 594.0, + 1231.0, + 594.0, + 1231.0, + 620.0, + 1181.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 594.0, + 1321.0, + 594.0, + 1321.0, + 620.0, + 1271.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 622.0, + 648.0, + 622.0, + 648.0, + 647.0, + 563.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 621.0, + 1201.0, + 621.0, + 1201.0, + 646.0, + 1116.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 675.0, + 650.0, + 675.0, + 650.0, + 709.0, + 500.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 676.0, + 1285.0, + 676.0, + 1285.0, + 711.0, + 976.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 446.0, + 462.0, + 446.0, + 462.0, + 475.5, + 378.0, + 475.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1297.0, + 1005.0, + 1297.0, + 1005.0, + 1342.0, + 691.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1998.0, + 1035.0, + 1998.0, + 1035.0, + 2035.0, + 662.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 712.0, + 967.0, + 712.0, + 967.0, + 755.0, + 730.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1472, + 1404, + 1472, + 1404, + 1779, + 297, + 1779 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1788, + 1402, + 1788, + 1402, + 1943, + 299, + 1943 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1290, + 1403, + 1290, + 1403, + 1443, + 298, + 1443 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 312, + 246, + 1390, + 246, + 1390, + 708, + 312, + 708 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1397, + 1972, + 1397, + 2034, + 298, + 2034 + ], + "score": 0.944 + }, + { + "category_id": 5, + "poly": [ + 620, + 826, + 1077, + 826, + 1077, + 1094, + 620, + 1094 + ], + "score": 0.94, + "html": "
Abaltion ModelBPC
No Control Gates1.354
LSTM (TBPTT length=50) ARMIN(TBPTT length=50)1.390 1.220
Random Read1.224
Queue-style Write1.352
Independent Read/write1.245
SmoothRead/write1.212
" + }, + { + "category_id": 4, + "poly": [ + 696, + 716, + 1002, + 716, + 1002, + 749, + 696, + 749 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 6, + "poly": [ + 586, + 772, + 1113, + 772, + 1113, + 805, + 586, + 805 + ], + "score": 0.76 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.731 + }, + { + "category_id": 0, + "poly": [ + 300, + 1223, + 1399, + 1223, + 1399, + 1256, + 300, + 1256 + ], + "score": 0.609 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 855, + 76, + 855, + 104, + 299, + 104 + ], + "score": 0.487 + }, + { + "category_id": 1, + "poly": [ + 300, + 1148, + 1143, + 1148, + 1143, + 1182, + 300, + 1182 + ], + "score": 0.451 + }, + { + "category_id": 0, + "poly": [ + 300, + 1148, + 1143, + 1148, + 1143, + 1182, + 300, + 1182 + ], + "score": 0.369 + }, + { + "category_id": 1, + "poly": [ + 300, + 1223, + 1399, + 1223, + 1399, + 1256, + 300, + 1256 + ], + "score": 0.369 + }, + { + "category_id": 4, + "poly": [ + 978, + 679, + 1282, + 679, + 1282, + 708, + 978, + 708 + ], + "score": 0.25 + }, + { + "category_id": 1, + "poly": [ + 586, + 772, + 1113, + 772, + 1113, + 805, + 586, + 805 + ], + "score": 0.187 + }, + { + "category_id": 13, + "poly": [ + 584, + 1821, + 644, + 1821, + 644, + 1852, + 584, + 1852 + ], + "score": 0.91, + "latex": "\\boldsymbol { h } _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 849, + 1787, + 883, + 1787, + 883, + 1823, + 849, + 1823 + ], + "score": 0.9, + "latex": "\\pmb { g } _ { t } ^ { h }" + }, + { + "category_id": 13, + "poly": [ + 933, + 1790, + 964, + 1790, + 964, + 1823, + 933, + 1823 + ], + "score": 0.89, + "latex": "\\pmb { g } _ { t } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 504, + 1824, + 535, + 1824, + 535, + 1852, + 504, + 1852 + ], + "score": 0.86, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 505, + 1916, + 533, + 1916, + 533, + 1942, + 505, + 1942 + ], + "score": 0.84, + "latex": "\\mathbf { } _ { \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } r _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1004, + 773, + 1099, + 773, + 1099, + 804, + 1004, + 804 + ], + "score": 0.81, + "latex": "\\scriptstyle ( d _ { h } = 8 0 0 )" + }, + { + "category_id": 13, + "poly": [ + 889, + 908, + 937, + 908, + 937, + 934, + 889, + 934 + ], + "score": 0.33, + "latex": "\\scriptstyle 1 = 5 0" + }, + { + "category_id": 13, + "poly": [ + 908, + 938, + 955, + 938, + 955, + 965, + 908, + 965 + ], + "score": 0.26, + "latex": "\\scriptstyle 1 = 5 0" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 245.0, + 372.0, + 245.0, + 372.0, + 273.0, + 337.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 257.0, + 701.0, + 257.0, + 701.0, + 284.0, + 644.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 246.0, + 922.0, + 246.0, + 922.0, + 270.0, + 894.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 254.0, + 1255.0, + 254.0, + 1255.0, + 284.0, + 1195.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 278.0, + 698.0, + 278.0, + 698.0, + 304.0, + 646.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 276.0, + 1249.0, + 276.0, + 1249.0, + 303.0, + 1196.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 307.0, + 347.0, + 307.0, + 347.0, + 542.0, + 311.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 296.0, + 828.0, + 296.0, + 828.0, + 347.0, + 641.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 311.0, + 902.0, + 311.0, + 902.0, + 540.0, + 870.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 296.0, + 1379.0, + 296.0, + 1379.0, + 344.0, + 1196.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 330.0, + 370.0, + 330.0, + 370.0, + 355.0, + 339.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 342.0, + 711.0, + 342.0, + 711.0, + 367.0, + 641.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 331.0, + 920.0, + 331.0, + 920.0, + 352.0, + 896.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 339.0, + 1266.0, + 339.0, + 1266.0, + 368.0, + 1193.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 414.0, + 370.0, + 414.0, + 370.0, + 438.0, + 339.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 413.0, + 921.0, + 413.0, + 921.0, + 437.0, + 894.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 497.0, + 369.0, + 497.0, + 369.0, + 521.0, + 339.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 497.0, + 920.0, + 497.0, + 920.0, + 518.0, + 896.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 500.0, + 953.0, + 500.0, + 953.0, + 520.0, + 928.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 586.0, + 369.0, + 586.0, + 369.0, + 602.0, + 353.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 584.0, + 923.0, + 584.0, + 923.0, + 604.0, + 904.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 599.0, + 386.0, + 599.0, + 386.0, + 616.0, + 370.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 595.0, + 494.0, + 595.0, + 494.0, + 620.0, + 444.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 594.0, + 586.0, + 594.0, + 586.0, + 620.0, + 534.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 594.0, + 678.0, + 594.0, + 678.0, + 620.0, + 625.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 595.0, + 769.0, + 595.0, + 769.0, + 620.0, + 717.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 595.0, + 947.0, + 595.0, + 947.0, + 619.0, + 911.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 594.0, + 1046.0, + 594.0, + 1046.0, + 620.0, + 995.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 594.0, + 1138.0, + 594.0, + 1138.0, + 620.0, + 1087.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 594.0, + 1230.0, + 594.0, + 1230.0, + 619.0, + 1179.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 594.0, + 1321.0, + 594.0, + 1321.0, + 619.0, + 1269.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 620.0, + 649.0, + 620.0, + 649.0, + 649.0, + 562.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 617.0, + 1202.0, + 617.0, + 1202.0, + 649.0, + 1113.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 675.0, + 651.0, + 675.0, + 651.0, + 709.0, + 500.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 676.0, + 1285.0, + 676.0, + 1285.0, + 710.0, + 975.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 509.5, + 997.0, + 509.5, + 997.0, + 553.5, + 934.0, + 553.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 548.0, + 980.0, + 548.0, + 980.0, + 558.0, + 969.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 572.0, + 1127.0, + 572.0, + 1127.0, + 583.0, + 1114.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 714.0, + 1003.0, + 714.0, + 1003.0, + 753.0, + 695.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 766.0, + 1003.0, + 766.0, + 1003.0, + 809.0, + 584.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 766.0, + 1113.0, + 766.0, + 1113.0, + 809.0, + 1100.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1221.0, + 1404.0, + 1221.0, + 1404.0, + 1260.0, + 297.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1145.0, + 1149.0, + 1145.0, + 1149.0, + 1188.0, + 297.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 677.0, + 1284.0, + 677.0, + 1284.0, + 713.0, + 978.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1474.0, + 1404.0, + 1474.0, + 1404.0, + 1506.0, + 296.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1506.0, + 1405.0, + 1506.0, + 1405.0, + 1538.0, + 295.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1534.0, + 1404.0, + 1534.0, + 1404.0, + 1568.0, + 294.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1562.0, + 1404.0, + 1562.0, + 1404.0, + 1603.0, + 292.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1596.0, + 1404.0, + 1596.0, + 1404.0, + 1628.0, + 296.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1625.0, + 1404.0, + 1625.0, + 1404.0, + 1662.0, + 294.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1656.0, + 1404.0, + 1656.0, + 1404.0, + 1692.0, + 294.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1689.0, + 1404.0, + 1689.0, + 1404.0, + 1721.0, + 296.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1720.0, + 1405.0, + 1720.0, + 1405.0, + 1751.0, + 296.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1749.0, + 1263.0, + 1749.0, + 1263.0, + 1782.0, + 293.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1786.0, + 848.0, + 1786.0, + 848.0, + 1826.0, + 293.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1786.0, + 932.0, + 1786.0, + 932.0, + 1826.0, + 884.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1786.0, + 1406.0, + 1786.0, + 1406.0, + 1826.0, + 965.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 503.0, + 1819.0, + 503.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1819.0, + 583.0, + 1819.0, + 583.0, + 1856.0, + 536.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1819.0, + 1406.0, + 1819.0, + 1406.0, + 1856.0, + 645.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1407.0, + 1850.0, + 1407.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1876.0, + 1406.0, + 1876.0, + 1406.0, + 1919.0, + 293.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 504.0, + 1912.0, + 504.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1912.0, + 1137.0, + 1912.0, + 1137.0, + 1945.0, + 534.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1289.0, + 1403.0, + 1289.0, + 1403.0, + 1327.0, + 293.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1322.0, + 1404.0, + 1322.0, + 1404.0, + 1355.0, + 296.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1352.0, + 1404.0, + 1352.0, + 1404.0, + 1385.0, + 294.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1384.0, + 1404.0, + 1384.0, + 1404.0, + 1417.0, + 296.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1414.0, + 983.0, + 1414.0, + 983.0, + 1444.0, + 294.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1402.0, + 1971.0, + 1402.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1322.0, + 2002.0, + 1322.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1145.0, + 1149.0, + 1145.0, + 1149.0, + 1188.0, + 297.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1221.0, + 1404.0, + 1221.0, + 1404.0, + 1260.0, + 297.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 766.0, + 1003.0, + 766.0, + 1003.0, + 809.0, + 584.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 766.0, + 1113.0, + 766.0, + 1113.0, + 809.0, + 1100.0, + 809.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 629, + 1403, + 629, + 1403, + 752, + 299, + 752 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 464, + 1403, + 464, + 1403, + 618, + 298, + 618 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 331, + 1404, + 331, + 1404, + 455, + 298, + 455 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1401, + 229, + 1401, + 322, + 296, + 322 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 864, + 2089, + 864, + 2112, + 836, + 2112 + ], + "score": 0.837 + }, + { + "category_id": 13, + "poly": [ + 1136, + 497, + 1167, + 497, + 1167, + 527, + 1136, + 527 + ], + "score": 0.87, + "latex": "h _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 501, + 1083, + 501, + 1083, + 527, + 1051, + 527 + ], + "score": 0.85, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 989, + 364, + 1026, + 364, + 1026, + 392, + 989, + 392 + ], + "score": 0.65, + "latex": "M" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 1404.0, + 630.0, + 1404.0, + 663.0, + 295.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 660.0, + 1405.0, + 660.0, + 1405.0, + 693.0, + 294.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 690.0, + 1405.0, + 690.0, + 1405.0, + 723.0, + 295.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 718.0, + 1326.0, + 718.0, + 1326.0, + 758.0, + 293.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 464.0, + 1403.0, + 464.0, + 1403.0, + 497.0, + 296.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 497.0, + 1050.0, + 497.0, + 1050.0, + 530.0, + 296.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 497.0, + 1135.0, + 497.0, + 1135.0, + 530.0, + 1084.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 497.0, + 1404.0, + 497.0, + 1404.0, + 530.0, + 1168.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 525.0, + 1405.0, + 525.0, + 1405.0, + 562.0, + 294.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 556.0, + 1405.0, + 556.0, + 1405.0, + 594.0, + 293.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 589.0, + 871.0, + 589.0, + 871.0, + 622.0, + 294.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 330.0, + 1405.0, + 330.0, + 1405.0, + 367.0, + 295.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 364.0, + 988.0, + 364.0, + 988.0, + 397.0, + 296.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 397.0, + 1027.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 394.0, + 1402.0, + 394.0, + 1402.0, + 427.0, + 294.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 426.0, + 1091.0, + 426.0, + 1091.0, + 458.0, + 294.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 291.0, + 1379.0, + 291.0, + 1379.0, + 325.0, + 292.0, + 325.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SJlsFpVtDB/images/003fdba39a3e5e0f89d17f8fed9ace36999e5baff3c5c0ac7eb3a615112ac523.jpg b/parse/train/SJlsFpVtDB/images/003fdba39a3e5e0f89d17f8fed9ace36999e5baff3c5c0ac7eb3a615112ac523.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9e5434633fcfe555a51b857f959583786aa798f1 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/003fdba39a3e5e0f89d17f8fed9ace36999e5baff3c5c0ac7eb3a615112ac523.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2943ae26cdda4c8224ad340dea818b265d82e64dc921db5621e8caaf1dd329fc +size 74063 diff --git a/parse/train/SJlsFpVtDB/images/013ec60420b5fea41530f6a4ef26ce9838a186c8b9b48f645b32e939f4ac77d8.jpg b/parse/train/SJlsFpVtDB/images/013ec60420b5fea41530f6a4ef26ce9838a186c8b9b48f645b32e939f4ac77d8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..cbb925eb405d8a756687c2dd0f55b59391d3141d --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/013ec60420b5fea41530f6a4ef26ce9838a186c8b9b48f645b32e939f4ac77d8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6975521ef1769b83a94e797f4eff5949b260954b6e5061ca728e35b5d0a32701 +size 4760 diff --git a/parse/train/SJlsFpVtDB/images/03b80f4720bd8b39e45b9244e7bcb2fbd2e811f197b1a803381a86dc9b7e2c03.jpg b/parse/train/SJlsFpVtDB/images/03b80f4720bd8b39e45b9244e7bcb2fbd2e811f197b1a803381a86dc9b7e2c03.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7c31219464395f539b157d2a6fbc7039cd57f1e0 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/03b80f4720bd8b39e45b9244e7bcb2fbd2e811f197b1a803381a86dc9b7e2c03.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:badede3d8810a355bd168462316b16bf8a4db643dc4df7a12313f9add135e206 +size 9431 diff --git a/parse/train/SJlsFpVtDB/images/03c7cc2b4d70135753af7fa7160da93d37ead18780071b050246ca65420f1538.jpg b/parse/train/SJlsFpVtDB/images/03c7cc2b4d70135753af7fa7160da93d37ead18780071b050246ca65420f1538.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4382be97d1c115fa891cc17c3cea5669880d17a0 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/03c7cc2b4d70135753af7fa7160da93d37ead18780071b050246ca65420f1538.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6e51e29719f6d1701733a06dbb30ebcf92f2fc8630f84c8a40e5cceb26bdfb +size 5695 diff --git a/parse/train/SJlsFpVtDB/images/16f5d7cf232ebae48e54e525bf4c84cc47e33589bfd225ce34c823a5f168f6a8.jpg b/parse/train/SJlsFpVtDB/images/16f5d7cf232ebae48e54e525bf4c84cc47e33589bfd225ce34c823a5f168f6a8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c53979269b2fd62cbd59bfcd374f05d8d1548b85 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/16f5d7cf232ebae48e54e525bf4c84cc47e33589bfd225ce34c823a5f168f6a8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e4d4aaa4480d08bb587b6f5bb6be2483cf8b3920a6d7c4ac9d047c097680c4c +size 74392 diff --git a/parse/train/SJlsFpVtDB/images/172f1dd21c799b7aa311e0cce2c7b8cfb5643ec255d3254b45caeb1337a4aab0.jpg b/parse/train/SJlsFpVtDB/images/172f1dd21c799b7aa311e0cce2c7b8cfb5643ec255d3254b45caeb1337a4aab0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6212139ae8b34047bd26955209dfb6f9bad12c9d --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/172f1dd21c799b7aa311e0cce2c7b8cfb5643ec255d3254b45caeb1337a4aab0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:853d84eea94afacbd9e25aceb999a0a56172a9592b59962b5bc8dc2c51861b35 +size 50210 diff --git a/parse/train/SJlsFpVtDB/images/1925d01ac93611dd48a3230a7a968e5585d14212e66a3eeacd01c33c8d5a4b90.jpg b/parse/train/SJlsFpVtDB/images/1925d01ac93611dd48a3230a7a968e5585d14212e66a3eeacd01c33c8d5a4b90.jpg new file mode 100644 index 0000000000000000000000000000000000000000..450348162d12e16af08d57fd983ef21f44e6733a --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/1925d01ac93611dd48a3230a7a968e5585d14212e66a3eeacd01c33c8d5a4b90.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c45976c8b10338b0fe17c96b9fe2a2c823211b01f48636496ef9d26932cd9f4e +size 3588 diff --git a/parse/train/SJlsFpVtDB/images/2d81b6982b4e4a3025110baef55b65736824e0ed2884a4ca288893673761f235.jpg b/parse/train/SJlsFpVtDB/images/2d81b6982b4e4a3025110baef55b65736824e0ed2884a4ca288893673761f235.jpg new file mode 100644 index 0000000000000000000000000000000000000000..473f63dc0ec21d3b5878bb8465b1663afa163052 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/2d81b6982b4e4a3025110baef55b65736824e0ed2884a4ca288893673761f235.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:329ced7f1bdd803691622b03431a060b9c64fde56250a0fa781029b8dbae6195 +size 4089 diff --git a/parse/train/SJlsFpVtDB/images/2dfc3625247d14f347e951bad518cf823cd1dbde1dc7e3105dcde682e2fb3bf2.jpg b/parse/train/SJlsFpVtDB/images/2dfc3625247d14f347e951bad518cf823cd1dbde1dc7e3105dcde682e2fb3bf2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..578b93584478ff014d43164103d0cfdebdf82660 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/2dfc3625247d14f347e951bad518cf823cd1dbde1dc7e3105dcde682e2fb3bf2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4b7cc06ecffaf652516c0b810897f3eac30d0eb6f15e6d8f7697300fbc04905 +size 15653 diff --git a/parse/train/SJlsFpVtDB/images/37accf44d193790fab10a00d7862fcc3aa4ac92ae9773d29ab311e5f9ab7517d.jpg b/parse/train/SJlsFpVtDB/images/37accf44d193790fab10a00d7862fcc3aa4ac92ae9773d29ab311e5f9ab7517d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b17a466a5b11426eb46db91dac351de247306b3e --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/37accf44d193790fab10a00d7862fcc3aa4ac92ae9773d29ab311e5f9ab7517d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45b3bec2d81660665e849ae1044eb3fcf24ae83b3162b6f72ef2618a4da27cc2 +size 20098 diff --git a/parse/train/SJlsFpVtDB/images/4680dad2dc5f8fd65aa642aefd19cab6a05e82da559fd64bfebcb2b69efa59df.jpg b/parse/train/SJlsFpVtDB/images/4680dad2dc5f8fd65aa642aefd19cab6a05e82da559fd64bfebcb2b69efa59df.jpg new file mode 100644 index 0000000000000000000000000000000000000000..68c2bac9e7e98bb0b6e7a438349bec1e5b82a713 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/4680dad2dc5f8fd65aa642aefd19cab6a05e82da559fd64bfebcb2b69efa59df.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7c25bec7b01aa79775d62ffc1575c78c420965846a581043c42959737c6937e +size 124349 diff --git a/parse/train/SJlsFpVtDB/images/4a444f3b3d39e4496da6f3b35272ea6b80f74f1a5f882bfb136fda959a981b57.jpg b/parse/train/SJlsFpVtDB/images/4a444f3b3d39e4496da6f3b35272ea6b80f74f1a5f882bfb136fda959a981b57.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4f0a1bdf1f6048b684f9944791125903eab833c6 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/4a444f3b3d39e4496da6f3b35272ea6b80f74f1a5f882bfb136fda959a981b57.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e0b11b2961d87a5dce23f9fc2aca65b1afd04e305335179cbd78d0d825bc28d +size 17508 diff --git a/parse/train/SJlsFpVtDB/images/5122d6d65c8f312200ff1b748ae5149879ee67f6a7306e2a9eefbf4b1095a553.jpg b/parse/train/SJlsFpVtDB/images/5122d6d65c8f312200ff1b748ae5149879ee67f6a7306e2a9eefbf4b1095a553.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bb650f743a578ee651c2e216d2e77dfc89358d8e --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/5122d6d65c8f312200ff1b748ae5149879ee67f6a7306e2a9eefbf4b1095a553.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37b7e445906316264b780bf1dafda71674b1951e1efb7d17483c5c5ff8edf146 +size 61380 diff --git a/parse/train/SJlsFpVtDB/images/52c2925b5cbb264963e102ecfdc332861bcba999e1521612f0bc1a4499497f72.jpg b/parse/train/SJlsFpVtDB/images/52c2925b5cbb264963e102ecfdc332861bcba999e1521612f0bc1a4499497f72.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ff961df2876e4cc0aa4991e40540d61e6a5c97a3 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/52c2925b5cbb264963e102ecfdc332861bcba999e1521612f0bc1a4499497f72.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5b55518e3d7ba293ae2f6dc57dbfe1513e01eee8124d7a3dd5ca17246a5a173 +size 77543 diff --git a/parse/train/SJlsFpVtDB/images/5c72e66486b9b3258baa3a14c3ca513dc000e23a0ddcc343732ab273f47259be.jpg b/parse/train/SJlsFpVtDB/images/5c72e66486b9b3258baa3a14c3ca513dc000e23a0ddcc343732ab273f47259be.jpg new file mode 100644 index 0000000000000000000000000000000000000000..417a240ca2a5a8a838e89c3798916f344eeabf2b --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/5c72e66486b9b3258baa3a14c3ca513dc000e23a0ddcc343732ab273f47259be.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa6921ead8d508111d78b0a5d3404802eddcdfd77c4315ad1c2c2244c8909c6d +size 8880 diff --git a/parse/train/SJlsFpVtDB/images/5ca3594957d20495ff7979ec0851c16522a556987e00b214248f05ea5922d203.jpg b/parse/train/SJlsFpVtDB/images/5ca3594957d20495ff7979ec0851c16522a556987e00b214248f05ea5922d203.jpg new file mode 100644 index 0000000000000000000000000000000000000000..36cedef0c42187f4ac9cd785d3ec6ad2fa9d284a --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/5ca3594957d20495ff7979ec0851c16522a556987e00b214248f05ea5922d203.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c0680da9cdcdbab6832871807dc768af06dfe1cc5ffafc5d3828dc108653a6e +size 70977 diff --git a/parse/train/SJlsFpVtDB/images/626f88640c08cca8585f7dca9c9fe43b7943f4dffce1a34c78600c3aacf06f7c.jpg b/parse/train/SJlsFpVtDB/images/626f88640c08cca8585f7dca9c9fe43b7943f4dffce1a34c78600c3aacf06f7c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c6aa1f63a7014ed997fb97e5808a55a147c334d5 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/626f88640c08cca8585f7dca9c9fe43b7943f4dffce1a34c78600c3aacf06f7c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90c26d84aa4a8d595d4848a37a48749321e58d424e5595155fed032ea8262889 +size 10765 diff --git a/parse/train/SJlsFpVtDB/images/62f73fa4e249d8507fab383797038663c1102e1a6452c52492b30ca977197adf.jpg b/parse/train/SJlsFpVtDB/images/62f73fa4e249d8507fab383797038663c1102e1a6452c52492b30ca977197adf.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9e27071e40cf4e237b824430e3ad21d8103763ea --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/62f73fa4e249d8507fab383797038663c1102e1a6452c52492b30ca977197adf.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76b4331fd18004b4111d634560e94e75b48e96da3b0a6c3a2a21b2470a3f6f27 +size 5145 diff --git a/parse/train/SJlsFpVtDB/images/6437d8f820db43004ec97676e2dd53db03f4c80e131e8748b05958aca538857f.jpg b/parse/train/SJlsFpVtDB/images/6437d8f820db43004ec97676e2dd53db03f4c80e131e8748b05958aca538857f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1a71bf99455972e3a8db3beb12ec4f6b2d611825 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/6437d8f820db43004ec97676e2dd53db03f4c80e131e8748b05958aca538857f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2db133fc460b546767619edab99877ded18e7524366f058e809ef079cb524e2d +size 8708 diff --git a/parse/train/SJlsFpVtDB/images/648efd68d3e9639f0936675a3a85f2c39b9f2c41f9904e11ca4e6598a30ecce5.jpg b/parse/train/SJlsFpVtDB/images/648efd68d3e9639f0936675a3a85f2c39b9f2c41f9904e11ca4e6598a30ecce5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a6dcbcd1eba4e3ecd2298821e7cf89e007451def --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/648efd68d3e9639f0936675a3a85f2c39b9f2c41f9904e11ca4e6598a30ecce5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c1a8dc5c11bc7dc260e23de58cb49edfd65fae361a5aa68e28df5cd84926a13 +size 4630 diff --git a/parse/train/SJlsFpVtDB/images/700d39069542b238a465c987a90658987b7600a64bbb7634c3d569f76aa03d3e.jpg b/parse/train/SJlsFpVtDB/images/700d39069542b238a465c987a90658987b7600a64bbb7634c3d569f76aa03d3e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9ec809f017900b608820e0cd1babac694402f91a --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/700d39069542b238a465c987a90658987b7600a64bbb7634c3d569f76aa03d3e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7172d90252561ef38c30de5024c455d16582bbe6ba323856e2e091b21e8199 +size 1392 diff --git a/parse/train/SJlsFpVtDB/images/73cb549e55cd48fb9ce20b88c7b972e48cd20f18e5fcaab85d56675aac89d988.jpg b/parse/train/SJlsFpVtDB/images/73cb549e55cd48fb9ce20b88c7b972e48cd20f18e5fcaab85d56675aac89d988.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4328bdbd2c51335b20c91b73f9e024a71e9b2836 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/73cb549e55cd48fb9ce20b88c7b972e48cd20f18e5fcaab85d56675aac89d988.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32ed31cbe1915cf5410191ff2fccb09c70598d61fc65acec937c154b58ebe5c3 +size 3768 diff --git a/parse/train/SJlsFpVtDB/images/7848ea32e6864dfc744b7a6d6797d932dea0c08e741d917391a74a8eec7e9b55.jpg b/parse/train/SJlsFpVtDB/images/7848ea32e6864dfc744b7a6d6797d932dea0c08e741d917391a74a8eec7e9b55.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b4fae7bf171026073266d3e70932e35ffa48946d --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/7848ea32e6864dfc744b7a6d6797d932dea0c08e741d917391a74a8eec7e9b55.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e126c83b310b7f1fcd00cf43921f6d8ebba2448bc92a2041ee574a657e40034d +size 43215 diff --git a/parse/train/SJlsFpVtDB/images/7a96a4dc534b4a666bdaa365eed053b6af296ad5e7ef9852a99d32096c1f6d9d.jpg b/parse/train/SJlsFpVtDB/images/7a96a4dc534b4a666bdaa365eed053b6af296ad5e7ef9852a99d32096c1f6d9d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fe1d42ce591e2db7ef7af7cb05e9c9065e385e33 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/7a96a4dc534b4a666bdaa365eed053b6af296ad5e7ef9852a99d32096c1f6d9d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06dbe2c30337893c141f417f0ae7b1b68c98b248c88f5c69b85f889ae4f9100b +size 14415 diff --git a/parse/train/SJlsFpVtDB/images/8848695cccd6a1037843913e048157c2f29f197acbd1ebef79c76febf8410914.jpg b/parse/train/SJlsFpVtDB/images/8848695cccd6a1037843913e048157c2f29f197acbd1ebef79c76febf8410914.jpg new file mode 100644 index 0000000000000000000000000000000000000000..44e952246e178966ed1a24aeb3032cfab722e57b --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/8848695cccd6a1037843913e048157c2f29f197acbd1ebef79c76febf8410914.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c55c5e41e750584eecbcd269b0e1ec3d4d32810bb1178d148574ce5affe7b727 +size 19145 diff --git a/parse/train/SJlsFpVtDB/images/a12edd848b40cd9d11b07463d5162ba9ad409a930c440f4a27529db9229b9581.jpg b/parse/train/SJlsFpVtDB/images/a12edd848b40cd9d11b07463d5162ba9ad409a930c440f4a27529db9229b9581.jpg new file mode 100644 index 0000000000000000000000000000000000000000..90a67dbd01c2e72b1c86d094ca7139af942647de --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/a12edd848b40cd9d11b07463d5162ba9ad409a930c440f4a27529db9229b9581.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dfed92d892dad3104e5cb5f820510356f6e5e58a1d208e578d55f96429b8af4 +size 7670 diff --git a/parse/train/SJlsFpVtDB/images/aeb9d0e1ef64ef36cded38ef6f8dcfba661ad46753ff035874375cf824d7663e.jpg b/parse/train/SJlsFpVtDB/images/aeb9d0e1ef64ef36cded38ef6f8dcfba661ad46753ff035874375cf824d7663e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..77f469c16ae82b71df23acd10e56ba0cfac8f4f3 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/aeb9d0e1ef64ef36cded38ef6f8dcfba661ad46753ff035874375cf824d7663e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:976ff44ca60cde4297c1675269007647aea71289e3c103cc5f86fa3e1c372cab +size 8832 diff --git a/parse/train/SJlsFpVtDB/images/b690aec51674d4f1655286a2f80ff1e8e5f752581b71bf80da7bf1f88d5dfb52.jpg b/parse/train/SJlsFpVtDB/images/b690aec51674d4f1655286a2f80ff1e8e5f752581b71bf80da7bf1f88d5dfb52.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3594c99a8ce71472495147701c186caec7610e5a --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/b690aec51674d4f1655286a2f80ff1e8e5f752581b71bf80da7bf1f88d5dfb52.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ce77688550ce2b2fa14fdd9868751d900b94b9f963397b04cbca4101a4f348b +size 8316 diff --git a/parse/train/SJlsFpVtDB/images/b73d0ec3ccfbba3fce51050668d0ed35df606906f8a900f489b5d6bf3add8951.jpg b/parse/train/SJlsFpVtDB/images/b73d0ec3ccfbba3fce51050668d0ed35df606906f8a900f489b5d6bf3add8951.jpg new file mode 100644 index 0000000000000000000000000000000000000000..093176ae58315f7bf9c9582165207dfb1ece7c0b --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/b73d0ec3ccfbba3fce51050668d0ed35df606906f8a900f489b5d6bf3add8951.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32d5622f4c66d070b43a353790a07c33d708d2e23df87c84029882ee8ccb4c3e +size 8147 diff --git a/parse/train/SJlsFpVtDB/images/bae05a0fb30e6dfa9c1b9fb5c1726fd1d2e6dc0783065d5b8a7eaca2ee19bfef.jpg b/parse/train/SJlsFpVtDB/images/bae05a0fb30e6dfa9c1b9fb5c1726fd1d2e6dc0783065d5b8a7eaca2ee19bfef.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a32a1051921d4b5b7f9556b80ccb9fd86f9f8237 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/bae05a0fb30e6dfa9c1b9fb5c1726fd1d2e6dc0783065d5b8a7eaca2ee19bfef.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9ed190b17e07fc22929e769ac4f55c381645fa65978de924327b00f293266d6 +size 3143 diff --git a/parse/train/SJlsFpVtDB/images/c66b9cbcf1160c6e9101d320acdb6ac82b33c6eeef8ecf4dd01f60a3033f294a.jpg b/parse/train/SJlsFpVtDB/images/c66b9cbcf1160c6e9101d320acdb6ac82b33c6eeef8ecf4dd01f60a3033f294a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7e5cfb438e1d7c41d109596ea7a9e52ef44e1f8e --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/c66b9cbcf1160c6e9101d320acdb6ac82b33c6eeef8ecf4dd01f60a3033f294a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e83b0d3f656b394cad3f1b6b3bee7b0890a5760d9a9ed4638ebcbe24f9efc98 +size 11666 diff --git a/parse/train/SJlsFpVtDB/images/cf4cd00e1c79b1b58eac4d1d7cfe430808ce3e62a85d5e6772689967b980fa9a.jpg b/parse/train/SJlsFpVtDB/images/cf4cd00e1c79b1b58eac4d1d7cfe430808ce3e62a85d5e6772689967b980fa9a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f2f22544ff73d694253ef8844878cba0a7dfa6fc --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/cf4cd00e1c79b1b58eac4d1d7cfe430808ce3e62a85d5e6772689967b980fa9a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1b375f729047b13f10fb558e9311a0569933c37f751d20fde8ea6ee12a32a00 +size 5955 diff --git a/parse/train/SJlsFpVtDB/images/d158a95e42dff4b1d4e39d17a981866590986155ba28ecaf964164c2b64af5d2.jpg b/parse/train/SJlsFpVtDB/images/d158a95e42dff4b1d4e39d17a981866590986155ba28ecaf964164c2b64af5d2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f388b6369234bc6439ff40400105c04233c59dbb --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/d158a95e42dff4b1d4e39d17a981866590986155ba28ecaf964164c2b64af5d2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57db6ad0b32f828f50045e7ec2ffd59265447c7589898fe9558d97e6e0a1916d +size 68229 diff --git a/parse/train/SJlsFpVtDB/images/d236f35056b691df98c5af718d074e6a0def6a69756636a5948d07fe29f30a73.jpg b/parse/train/SJlsFpVtDB/images/d236f35056b691df98c5af718d074e6a0def6a69756636a5948d07fe29f30a73.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ab0c8c5e838e8f13a3093cb2627eca470cdecb08 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/d236f35056b691df98c5af718d074e6a0def6a69756636a5948d07fe29f30a73.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:611205689d52bfb0da026ba66b5c33a90b28cc57bed97d02485d8f304ebd8778 +size 110690 diff --git a/parse/train/SJlsFpVtDB/images/d4008801a0bdc5f8578b7fc2ab18108687c77ce0b26c80d88de296dc9e779f2c.jpg b/parse/train/SJlsFpVtDB/images/d4008801a0bdc5f8578b7fc2ab18108687c77ce0b26c80d88de296dc9e779f2c.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a777dba655944a8efdaaa51179d6adb5ba4e0888 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/d4008801a0bdc5f8578b7fc2ab18108687c77ce0b26c80d88de296dc9e779f2c.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8222425af45549e96fee661a10a3d133506ef4b0d89c19d20ac3a041b7cc4981 +size 29132 diff --git a/parse/train/SJlsFpVtDB/images/d8c5dc7421a638c61fb979a033d4cdd61f598776840d2e3886602f960e75f701.jpg b/parse/train/SJlsFpVtDB/images/d8c5dc7421a638c61fb979a033d4cdd61f598776840d2e3886602f960e75f701.jpg new file mode 100644 index 0000000000000000000000000000000000000000..99433f09bcfe3e286990e46b74b062406db378f6 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/d8c5dc7421a638c61fb979a033d4cdd61f598776840d2e3886602f960e75f701.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7c3c0230531421bf79d0074710b3bb86019164797a3bebb9bc9c81894d1c48e +size 16167 diff --git a/parse/train/SJlsFpVtDB/images/e3ae29457ffa3d9b8d83883d7440b05ef18b66a90b06c48534eefd45c8472573.jpg b/parse/train/SJlsFpVtDB/images/e3ae29457ffa3d9b8d83883d7440b05ef18b66a90b06c48534eefd45c8472573.jpg new file mode 100644 index 0000000000000000000000000000000000000000..61c0d4484332ff1d0b8203385cb3313cf33a556a --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/e3ae29457ffa3d9b8d83883d7440b05ef18b66a90b06c48534eefd45c8472573.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dab0f101a2a2ada35e824a362a9e80cb08253225bed80703801beb3fcbda32f +size 26740 diff --git a/parse/train/SJlsFpVtDB/images/ecb9740c1544cdc40740bd5a288f8c21dd6edf8e5e0f8f0e296de559dc76e201.jpg b/parse/train/SJlsFpVtDB/images/ecb9740c1544cdc40740bd5a288f8c21dd6edf8e5e0f8f0e296de559dc76e201.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5c181af7eb0002d7d1da79a71a696dfa9882dd82 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/ecb9740c1544cdc40740bd5a288f8c21dd6edf8e5e0f8f0e296de559dc76e201.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc9445856f69856f6729581242bccfbe68929eea6738d46ae2647ea2b858f57d +size 38701 diff --git a/parse/train/SJlsFpVtDB/images/fe2237c49a32693991f8e400f9c94232bbebb2e0a74bdd130449be37c88961ec.jpg b/parse/train/SJlsFpVtDB/images/fe2237c49a32693991f8e400f9c94232bbebb2e0a74bdd130449be37c88961ec.jpg new file mode 100644 index 0000000000000000000000000000000000000000..75c823a0c340ce26eaba455947b5713d9d93f348 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/fe2237c49a32693991f8e400f9c94232bbebb2e0a74bdd130449be37c88961ec.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85e0936cefb7a4605a3958de298a335d7c893713173d65ad46447c750f16abe1 +size 4480 diff --git a/parse/train/SJlsFpVtDB/images/fed75f3a20157c1133e9955d154f044506d9848daacab63f767f921020c6c27d.jpg b/parse/train/SJlsFpVtDB/images/fed75f3a20157c1133e9955d154f044506d9848daacab63f767f921020c6c27d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..776efb0c839c8cd1aedbf86f5019211ba2a3e198 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/fed75f3a20157c1133e9955d154f044506d9848daacab63f767f921020c6c27d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:591745ef8865d31e2b5e430c4a52cc6cb774bf010536e78bf44892f3ab8eb8a9 +size 9915 diff --git a/parse/train/SJlsFpVtDB/images/ff992f530a366808d275b4398b58946860b7f2e5e0d24bf2f83063bba5a7faca.jpg b/parse/train/SJlsFpVtDB/images/ff992f530a366808d275b4398b58946860b7f2e5e0d24bf2f83063bba5a7faca.jpg new file mode 100644 index 0000000000000000000000000000000000000000..679792f793f9d1bf191ab7c62d0812bcc725c2d1 --- /dev/null +++ b/parse/train/SJlsFpVtDB/images/ff992f530a366808d275b4398b58946860b7f2e5e0d24bf2f83063bba5a7faca.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fdb261899ec5a995347f4dcc21da4bf58457c602a679f9e473f3dfa47e8c48f +size 9668 diff --git a/parse/train/SkNSehA9FQ/SkNSehA9FQ.md b/parse/train/SkNSehA9FQ/SkNSehA9FQ.md new file mode 100644 index 0000000000000000000000000000000000000000..611b78061471f8beaa35676a21b94b41beb3b91f --- /dev/null +++ b/parse/train/SkNSehA9FQ/SkNSehA9FQ.md @@ -0,0 +1,301 @@ +# OPEN VOCABULARY LEARNING ON SOURCE CODE WITH A GRAPH–STRUCTURED CACHE + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Machine learning models that take computer program source code as input typically use Natural Language Processing (NLP) techniques. However, a major challenge is that code is written using an open, rapidly changing vocabulary due to, e.g., the coinage of new variable and method names. Reasoning over such a vocabulary is not something for which most NLP methods are designed. We introduce a Graph–Structured Cache to address this problem; this cache contains a node for each new word the model encounters with edges connecting each word to its occurrences in the code. We find that combining this graph–structured cache strategy with recent Graph–Neural–Network–based models for supervised learning on code improves the models’ performance on a code completion task and a variable naming task — with over $100 \%$ relative improvement on the latter — at the cost of a moderate increase in computation time. + +# 1 INTRODUCTION + +Computer program source code is an abundant and accessible form of data from which machine learning algorithms could learn to perform many useful software development tasks, including variable name suggestion, code completion, bug finding, security vulnerability identification, code quality assessment, or automated test generation. But despite the similarities between natural language and source code, deep learning methods for Natural Language Processing (NLP) have not been straightforward to apply to learning problems on source code (Allamanis et al., 2017). + +There are many reasons for this, but two central ones are: + +1. Code’s syntactic structure is unlike natural language. While code contains natural language words and phrases in order to be human–readable, code is not meant to be a read like a natural language text. Code is written in a rigid syntax with delimiters that may open and close dozens of lines apart; it consists in great part of references to faraway lines and different files; and it describes computations that proceed in an order often quite distinct from its written order. + +2. Code is written using an open vocabulary. Natural language is mostly composed of words from a large but closed (a.k.a. fixed–size and unchanging) vocabulary. Standard NLP methods can thus perform well by fixing a large vocabulary of words before training, and labeling the few words they encounter outside this vocabulary as “unknown”. But in code every new variable, class, or method declared requires a name, and this abundance of names leads to the use of many obscure words: abbreviations, brand names, technical terms, etc.1 A model must be able to reason about these newly–coined words to understand code. + +The second of these issues is significant. To give one indication: $28 \%$ of variable names contain out–of–vocabulary words in the test set we use in our experiments below. But more broadly, the open vocabulary issue in code is an acute example of a fundamental challenge in machine learning: + +how to build models that can reason over unbounded domains of entities, sometimes called “open– set” learning. Despite this, the open vocabulary issue in source code has received relatively little attention in prior work. + +The first issue, in contrast, has been the focus of much prior work. A common strategy in these works is to represent source code as an Abstract Syntax Tree (AST) rather than as linear text. Once in this graph–structured format, code can be passed as input to models like Recursive Neural Networks or Graph Neural Networks (GNNs) that can, in principle, exploit the relational structure of their inputs and avoid the difficulties of reading code in linear order (Allamanis et al., 2018). + +Our contribution: In this paper we extend such AST–based models for source code in order to address the open vocabulary issue. We do so by introducing a Graph–Structured Cache (GSC) to handle out–of–vocabulary words. The GSC represents vocabulary words as additional nodes in the AST as they are encountered and connects them with the edges to where they are used in the code. We then process the AST+GSC with a GNN to produce outputs. See Figure 1. + +We empirically evaluated the utility of a Graph–Structured Cache on two tasks: a code completion (a.k.a. fill–in–the–blank) task and a variable naming task. We found that using a GSC improved performance on both tasks at the cost of an approximately $30 \%$ increase in training time. More precisely: even when using hyperparameters optimized for the baseline model, adding a GSC to a baseline model improved its accuracy by at least $7 \%$ on the fill–in–the–blank task and $103 \%$ on the variable naming task. We also report a number of ablation results in which we carefully demonstrate the relative importance of each model component to a model’s performance. + +# 2 PRIOR WORK + +# REPRESENTING CODE AS A GRAPH + +Given their prominence in the study of programming languages, Abstract Syntax Trees (ASTs) and parse trees are a natural choice for representing code and have been used extensively. Often models that operate on source code consume ASTs by linearizing them (usually with a depth–first traversal) as in Amodio et al. (2017), Liu et al. (2017), or Li et al. (2017) or by using AST paths as input features as in Alon et al. (2018), but they can also be processed by deep learning models that take graphs as input, as in White et al. (2016) and Chen et al. (2018) who use Recursive Neural Networks (RveNNs) (Goller & Kuchler, 1996) on ASTs. RveNNs are models that operate on tree– topology graphs, and have been used extensively for language modeling (Socher et al., 2013) and on domains similar to source code, like mathematical expressions (Zaremba et al., 2014; Arabshahi et al., 2018). They can be considered a special case of Message Passing Neural Networks (MPNNs) in the framework of Gilmer et al. (2017): in this analogy RveNNs are to Belief Propagation as MPNNs are to Loopy Belief Propagation. They can also be considered a special case of Graph Networks in the framework of Battaglia et al. (2018). ASTs also serve as a natural basis for models that generate code as output, as in Maddison & Tarlow (2014), Yin & Neubig (2017), Rabinovich et al. (2017), Chen et al. (2018), and Brockschmidt et al. (2018). + +Data–flow graphs are another type of graphical representation of source code with a long history (Krinke, 2001), and they have occasionally been used to featurize source code for machine learning (Chae et al., 2017). + +Most closely related to our work is the work of Allamanis et al. (2018), on which our model is heavily based. Allamanis et al. (2018) combine the data–flow graph and AST representation strategies for source code by representing code as an AST augmented with extra labeled edges indicating semantic information like data– and control–flow between variables. These augmentations yield a directed multigraph rather than just a tree,2 so in Allamanis et al. (2018) a variety of MPNN called a Gated Graph Neural Network (GGNN) (Li et al., 2016) is used to consume the Augmented AST and produce an output for a supervised learning task. + +Graph-based models that are not based on ASTs are also sometimes used for analyzing source code, like Conditional Random Fields for joint variable name prediction in (Raychev et al., 2015). + +# REASONING ABOUT OPEN SETS + +The question of how to gracefully reason over an open vocabulary is longstanding in NLP. Character–level embeddings are a typical way deep learning models handle this issue, whether used on their own as in Kim et al. (2016), or in conjunction with word–level embedding Recurrent Neural Networks (RNNs) as in Luong & Manning (2016), or in conjunction with an $n _ { \mathrm { ~ } }$ –gram model as in Bojanowski et al. (2017). Another approach is to learn new word embeddings on–the–fly from context (Kobayashi et al., 2016). Caching novel words, as we do in our model, is yet another strategy (Grave et al., 2017) and has been used to augment N–gram models for analyzing source code (Hellendoorn & Devanbu, 2017). + +In terms of producing outputs over variable–sized input and outputs, also known as open–set learning, attention-based pointer mechanisms were introduced in Vinyals et al. (2015) and have been used for tasks on code, e.g. in Bhoopchand et al. (2016). Such methods have been used to great effect in NLP in e.g. Gulcehre et al. (2016) and Merity et al. (2017). The latter’s pointer sentinel mixture model is the direct inspiration for the readout function we use in the Variable Naming task below. + +Using graphs to represent arbitrary collections of entities and their relationships for processing by deep networks has been widely used (Johnson, 2017; Bansal et al., 2017; Pham et al., 2018; Lu et al., 2017), but to our knowledge we are the first to use a graph–building strategy for reasoning (at train and test time) about an open vocabulary of words. + +# 3 PRELIMINARIES + +# 3.1 ABSTRACT SYNTAX TREES + +An Abstract Syntax Tree (AST) is a graph — specifically an ordered tree with labeled nodes — that is a representation of some written computer source code. There is a 1–to–1 relationship between source code and an AST of that source code, modulo comments and whitespace in the written source code. + +Typically the leaves of an AST correspond to the tokens written in the source code, like variable and method names, while the non–leaf nodes represent syntactic language constructs like function calls or class definitions. The specific node labels and construction rules of ASTs can differ between or within languages. The first step in Figure 1 shows an example. + +# 3.2 GRAPH NEURAL NETWORKS + +The term Graph Neural Network (GNN) refers to any deep, differentiable model that takes graphs as input. Many GNNs have been presented in the literature, and several nomenclatures have been proposed for describing the computations they perform, in particular in Gilmer et al. (2017) and Battaglia et al. (2018). Here we give a brief recapitulation of supervised learning with GNNs using the Message Passing Neural Network framework from Gilmer et al. (2017). + +A GNN is trained using pairs $( G , y )$ where $G = ( V , E )$ is a graph defined by its vertices $V$ and edges $E$ , and $y$ is a label. $y$ can be any sort of mathematical object: scalar, vector, another graph, etc. In the most general case, each graph in the dataset can be a directed multigraph, each with a different number of nodes and different connectivity. In each graph, each vertex $v \in V$ has associated features $\mathbf { \boldsymbol { x } } _ { v }$ , and each edge $( v , w ) \in E$ has features $e _ { v w }$ . + +A GNN produces a prediction $\hat { y }$ for the label $y$ of a graph $G = ( V , E )$ by the following procedure: + +1. A function $S$ is used to initialize a hidden state vector $h _ { v } ^ { 0 }$ for each vertex $v \in V$ as a function of the vertex’s features (e.g., if the $\mathbf { \Delta } _ { \mathbf { \mathcal { X } } _ { v } }$ are words, $S$ could be a word embedding function): + +$$ +\begin{array} { r } { h _ { v } ^ { 0 } = S ( \pmb { x } _ { v } ) } \end{array} +$$ + +2. For each round $t$ out of $T$ total rounds: + +(a) Each vertex $v \in V$ receives the vector $\boldsymbol { m } _ { v } ^ { t + 1 }$ , which is the sum of “messages” from its neighbors, each produced by a function $M _ { t }$ : + +$$ +m _ { v } ^ { t + 1 } = \sum _ { w \in \mathrm { n e i g h b o r s ~ o f } v } M _ { t } ( h _ { v } ^ { t } , h _ { w } ^ { t } , e _ { v w } ) . +$$ + +(b) Each vertex $v \in V$ updates its hidden state based on the message it received via a function $U _ { t }$ : + +$$ +h _ { v } ^ { t + 1 } = U _ { t } ( h _ { v } ^ { t } , m _ { v } ^ { t + 1 } ) . +$$ + +3. A function $R$ , the “readout function”, produces a prediction based on the hidden states generated during the message passing (usually just those at from time $T$ ): + +$$ +\hat { y } = R ( \{ h _ { v } ^ { t } | v \in V , t \in 1 , . . . , T \} ) . +$$ + +GNNs differ in how they implement $S , M _ { t } , U _ { t }$ , and $R$ . But all these functions are differentiable and most are parameterized, so the model is trainable via stochastic gradient descent of a loss function on $y$ and $\hat { y }$ . + +# 4 MODEL + +Our model consumes an input instance of source code and produces an output for a supervised learning task via the following five steps, sketched in Figure 1: + +1. Parse the source code (snippet, file, repository, version control history, etc.) into an Abstract Syntax Tree. +2. Add edges of varying types (details in Appendix Table 8) to this AST that represent semantic information like data– and control– flow, in the spirit of Allamanis et al. (2018). Also add the reversed version of all edges with their own edge type. This results in a directed multigraph called an Augmented AST. +3. Further augment the Augmented AST by adding a Graph–Structured Cache. That is, add a node to the Augmented AST for each vocabulary word encountered in the input instance. Then connect each such “cache node” with an edge (of edge type WORD USE) to all variables whose names contain its word. +4. Vectorize the Augmented $\mathrm { A S T } + \mathrm { G S C }$ graph into a form suitable for a GNN. (I.e. perform Step 1 from Section 3.2.) Each AST node that doesn’t represent a variable is vectorized as a learned embedding of the language construct it represents, e.g. Parameter, Method Declaration, etc. Each cache node and each node that represents a variable is vectorized as a learned linear map of the concatenation of a type embedding and a name embedding. The name embedding is a Character–Level Convolutional Neural Network (CharCNN) (Zhang et al., 2015) embedding of the word/name the node contains. The type embedding is a learned embedding of the name of the Java type of the token it contains, e.g. int, a user–defined class, etc., with cache nodes having their own unique Cacher Node type. +5. Process the graph with a GNN, as per Section 3.2. (I.e. perform Steps 2 and 3 from Section +3.2.) The readout functions differ depending on the task and are described in the Experiments section below. + +Our main contribution to previous works is the addition of Step 3, the Graph–Structured Cache step. The combination of relational information from the cache nodes’ connections and lexical information from these nodes’ CharCNN embeddings allows the model to, in principle, flexibly reason about words it never saw during training, but also recognize words it did. E.g. it could potentially see a class named “getGuavaDictionary” and a variable named “guava dict” and both (a) utilize the fact that the word “guava” is common to both names despite having never seen this word before, and (b) exploit learned representations for words like “get”, “dictionary”, and “dict” that it has seen during training. + +# 5 EXPERIMENTS + +We evaluated our model, described in Section 4, on two supervised tasks: a Fill–In–The–Blank task and a Variable Naming task. For each task, we compare our model to others that differ in how they parse the code and how they treat the words they encounter. Table 1 details the different variations of the procedure in Section 4 against which we compare our model. + +![](images/fa59b4133b5607587cc70ae7b40b4e6a20f3afab3cf1561f1e3c0ebf601145cb.jpg) +Figure 1: Our model’s procedure for consuming a single input instance of source code and producing an output for a supervised learning task. + +Code to reproduce all experiments is available online.3 4 + +# 5.1 DATA AND IMPLEMENTATION DETAILS + +We chose to use Java source code as the data for our experiments as it is among the most popular programming languages in use today (TIOBE, 2018; Github, 2017). To construct our dataset, we randomly selected 18 of the 100 most popular Java repos from the Maven repository5 to serve as training data. (See Appendix Table 7 for the list.) Together these repositories contain about 500,000 non–empty, non–comment lines of code. We checked for excessive code duplication in our dataset (Lopes et al., 2017) using $\mathrm { C P D ^ { 6 } }$ and found only about $7 \%$ of the lines to be contiguous, duplicated code blocks containing more than 150 tokens. + +Table 1: Nomenclature used in Experiments section. Each abbreviation describes a tweak/ablation to our full model as presented in Section 4. Using this nomenclature, our full model as described in Section 4 and shown in Figure 1 would be a “AugAST–GSC” model. + +
AbbreviationMeaning
Code RepresentationASTSkips Step 2 in Section 4.
AugASTPerforms Step 2 in Section 4.
Vocab StrategiesClosed VocabSkips Step 3 in Section 4,and instead maintains word-embedding vectors for words in a closed vo- cabulary. In Step 4, name embeddings for nodes representing variables are produced by taking the mean of the embeddings of the words in the vari- able's name.Words outside this model's closed vocabulary are labeled as <UNK>. This is the
CharCNNstrategy used in Allamanis et al. (2018). Skips Step 3 in Section 4.
Pointer SentinelFollows Steps 3 and 4 as described in Section 4,except it doesn't add edges connecting cache nodes to the nodes where their word is used. In the Variable Naming task, this is equivalent to us- ing the Pointer Sentinel Mixture Model of Merity
GSCet al. (2017) to produce outputs. Follows Steps 3 and 4 as described in Section 4.
Graph Neural NetworkGGNNPerforms Step 5 in Section 4 using the Gated Graph Neural Network of Li et al. (2016).
DTNNPerforms Step 5 in Section 4 using the Deep Ten- sor Neural Network of Schuitt et al. (2017).
RGCNPerforms Step 5 in Section 4 using the Relational Graph Convolutional Network of Schlichtkrull et al. (2017).
+ +We randomly chose 3 of these repositories to sequester as an “Unseen Repos” test set. We then separated out $15 \%$ of the files in the remaining 15 repositories to serve as our “Seen Repos” test set. The remaining files served as our training set, from which we separated $15 \%$ of the datapoints to act as a validation set. + +Our data preprocessor builds on top of the open–source Javaparser7 library to generate ASTs of our source code and then augment the ASTs with the edges described in Appendix Table 8. We used Apache $\mathrm { { \bf M X N e t } } ^ { 8 }$ as our deep learning framework. All hidden states in the GNN contained 64 units; all GNNs ran for 8 rounds of message passing; all models were optimized using the Adam optimizer (Kingma & Ba, 2015); all inputs to the GNNs were truncated to a maximum size of 500 nodes centered on the or $\mathrm { \ < N A M E - M E > }$ tokens. About $53 \%$ of input graphs were larger than 500 nodes before truncation. The only regularization we used was early stopping — early in our experiments we briefly tried $L _ { 2 }$ and dropout regularization, but saw no effects. + +We performed only a moderate amount of hyperparameter optimization, but all of it was done on the baseline models to avoid biasing our results in favor of our model. Specifically, we tuned all hyperparameters on the Closed Vocab baseline model, and also did a small amount of extra learning rate exploration for the Pointer Sentinel baseline model to try to maximize its performance. + +# 5.2 THE FILL–IN–THE–BLANK TASK + +In this task we randomly selected a single usage of a variable in some source code, replaced it with a token, and then asked the model to predict what variable should have been there. An example instance from our dataset is shown in Figure 2. + +![](images/69398b3fef9fe3f5b523898c2773311a427c29f26873c4f47ff60001f36d5b3c.jpg) +Figure 2: Example of a model’s procedure for completing the Fill–In–The–Blank task. Each Fill– In–The–Blank instance is created by replacing a single usage of a variable (n, in this example) with the special token . The model then processes the code as depicted in Figure 1. To produce outputs, the model’s readout function computes a soft–attention weighting over all nodes in the graph; the model’s output is the variable at the node on which it places maximal attention. In this example, if the model put maximal attention weighting on any of the green– highlighted variables, this would be a correct output. If maximal attention is placed on any other node, it would be an incorrect output. Only in–scope usages of a variable are counted as correct. + +The models indicate their prediction for what variable should go in the blank by pointing with neural attention over all the nodes in the AugAST. This means all training and test instances only considered cases where the obfuscated variable appears somewhere else in the code. Single uses are rare however, since in Java variables must be declared before they are used. It also means there are sometimes multiple usages of the same, correct variable to which a model can point to get the right answer. In our dataset $78 \%$ of variables were used more two times, and $33 \%$ were used more than four times. + +The models compute the attention weightings $y _ { i }$ for each Augmented AST node $i$ differently depending on the readout function of the GNN they use. Models using a GGNN as their GNN component, as all those in Table $2 \ : \mathrm { d o }$ , compute the attention weightings as per Li et al. (2016): + +$$ +\begin{array} { r } { \hat { y } _ { i } = \sigma \left( f _ { 1 } ( { h } _ { v } ^ { T } , { h } _ { v } ^ { 0 } ) \right) \odot f _ { 2 } ( { h } _ { v } ^ { T } ) , } \end{array} +$$ + +where the $f$ s are MLPs, $\boldsymbol { h } _ { v } ^ { t }$ is the hidden state of node $v$ after $t$ message passing iterations, $\sigma$ is the sigmoid function, and $\odot$ is elementwise multiplication. The DTNN and RGCN GNNs compute the attention weightings as per Schutt et al. (2017): ¨ + +$$ +\hat { y } _ { i } = f ( h _ { v } ^ { T } ) , +$$ + +where $f$ is a single hidden layer MLP. The models were trained using a binary cross entropy loss computed across the nodes in the graph. + +The performance of models using our GSC versus those using other methods is reported in Table 2. For context, a baseline strategy of random guessing among all variable nodes within an edge radius of 8 of the $ token achieves an accuracy of 0.22. We also compare the performance of different GNNs in Table 3. + +# 5.3 THE VARIABLE NAMING TASK + +In this task we replaced all usages of a name of a particular variable, method, class, or parameter in the code with the special token $ , and asked the model to produce the obfuscated name (in the form of the sequence of words that compose the name). An example instance from our dataset is shown in Figure 3. + +Table 2: Accuracy on the Fill–In–The–Blank task. Our model is the AugAST–GSC. The first number in each cell is the accuracy of the model, where a correct prediction is one in which the graph node that received the maximum attention weighting by the model contained the variable that was originally in the $ spot. The second, parenthetical numbers are the top– 5 accuracies, i.e. whether the correct node was among those that received the 5 largest attentions weightings from the model. See Table 1 for explanations of the abbreviations. All models use Gated Graph Neural Networks as their GNN component. + +
Closed VocabCharCNNGSC
Seen reposAST0.57 (0.83)0.60 (0.84)0.89 (0.96)
AugAST0.80 (0.90)0.90 (0.94)0.97 (0.99)
Unseen reposAST AugAST0.36 (0.68)0.48 (0.80)0.80 (0.93)
0.59 (0.78)0.84 (0.92)0.92 (0.96)
+ +Table 3: Accuracy (and top–5 accuracy) on the Fill–In–The–Blank task, depending on which type of GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST as their code representation. + +
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.80 (0.90) 0.97 (0.99)0.72 (0.84) 0.89 (0.95)0.80 (0.90) 0.95 (0.98)
Unseen reposClosed Vocab GSC0.59 (0.78) 0.92 (0.96)0.46 (0.68) 0.80 (0.89)0.62 (0.79) 0.88 (0.95)
+ +![](images/b21bc85d57d3da138b5330d252092f70b403cbab07c31d848435d51028931d7a.jpg) +Figure 3: Example of a model’s procedure for completing the Variable Naming task. Each Variable Naming instance is created by replacing all uses of some variable (expectedLength, in this example) with a special $\mathrm { \ < N A M E - M E > }$ token. The model then processes the code as depicted in Figure 1. To produce outputs, the model takes the mean of the nodes’ hidden states (depicted here in orange), uses them as the initial hidden state of a Recurrent Neural Network, and unrolls this RNN to produce a name as a sequence of words. + +To produce a name from the output of the GNN, our models used the readout function of Allamanis et al. (2018). This readout function computes the mean of the hidden states of the $\mathrm { < N A M E - M E > }$ nodes and passing it as the initial hidden state to a 1–layer Gated Recurrent Unit (GRU) RNN (Cho et al., 2014). This GRU is then unrolled to produce words in its predicted name, in the style of a traditional NLP decoder. We used a fixed length unrolling of 8 words, as $9 9 . 8 \%$ of names in our training set were 8 or fewer words long. The models were trained by cross entropy loss over the sequence of words in the name. + +To decode each hidden state output of the GRU $^ { h }$ into a probability distribution $P _ { \mathrm { v o c a b } } ( \mathbf { w } | h )$ over words w, the Closed Vocab and CharCNN models pass $^ { h }$ through a linear layer and a softmax layer with output dimension equal to the number of words in their closed vocabularies (i.e. a traditional decoder output for NLP). In contrast, the GSC model not only has access to a fixed–size vocabulary but can also produce words by pointing to cache nodes in its Graph–Structured Cache. Specifically, it uses a decoder architecture inspired by the Pointer Sentinel Mixture Model of Merity et al. (2017): the probability of a word w being the GSC decoder’s output given that the GRU’s hidden state was $^ { h }$ is + +$$ +P ( \mathbf { w } | h ) = P _ { \mathrm { g r a p h } } ( \mathbf { s } | h ) P _ { \mathrm { g r a p h } } ( \mathbf { w } | h ) + ( 1 - P _ { \mathrm { g r a p h } } ( \mathbf { s } | h ) ) P _ { \mathrm { v o c a b } } ( \mathbf { w } | h ) +$$ + +where $P _ { \mathrm { g r a p h } } ( \cdot | h )$ is a conditional probability distribution over cache nodes in the GSC and the sentinel s, and $P _ { \mathrm { v o c a b } } ( \cdot | h )$ is a conditional probability distribution over words in a closed vocabulary. $P _ { \mathrm { g r a p h } } ( \cdot | h )$ is computed by passing the hidden states of all cache nodes and the sentinel node through a single linear layer and then computing the softmax dot–product attention of these values with $^ { h }$ . $P _ { \mathrm { v o c a b } } ( \cdot | h )$ is computed as the softmax of a linear mapping of $^ { h }$ to indices in a closed vocabulary, as in the Closed Vocab and CharCNN models. If there is no cache node for w in the Augmented AST or if w is not in the model’s closed dictionary then $P _ { \mathrm { g r a p h } } ( \mathbf { w } | h )$ and $P _ { \mathrm { v o c a b } } ( \mathbf { w } | h )$ are 0, respectively. + +The performance of our GSC versus other methods is reported in Table 4. More granular performance statistics are reported in Appendix Table 6. We also compare the performance of different GNNs in Table 5. + +Table 4: Accuracy on the Variable Naming task. Our model is the AugAST–GSC. The first number in each cell is the accuracy of the model, where we consider a correct output to be exact reproduction of the full name of the obfuscated variable (i.e. all the words in the name and then a EOS token). The second, parenthetical numbers are the top–5 accuracies, i.e. whether the correct full name was among the 5 most probable sequences output by the model. See Table 1 for explanations of the abbreviations. All models use Gated Graph Neural Networks as their GNN component. + +
Closed VocabCharCNNPointer SentinelGSC
Seen reposAST0.23 (0.31)0.22 (0.28)0.19 (0.33)0.49 (0.67)
AugAST0.19 (0.26)0.20 (0.27)0.26 (0.40)0.53 (0.69)
Unseen reposAST0.05 (0.07)0.06 (0.09)0.06 (0.11)0.38 (0.53)
AugAST0.04 (0.07)0.06 (0.08)0.08 (0.14)0.41 (0.57)
+ +Table 5: Accuracy (and top–5 accuracy) on the Variable Naming task, depending on which type of GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST as their code representation. + +
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.19 (0.26)0.23 (0.31)0.27 (0.34)
0.53 (0.69)0.33 (0.48)0.46 (0.63)
Unseen reposClosed Vocab GSC0.04 (0.07) 0.41 (0.57)0.06 (0.08) 0.25 (0.40)0.06 (0.09) 0.35 (0.49)
+ +# 6 DISCUSSION + +As can be seen in Tables 2 and 4, the addition of a GSC improved performance on all tasks. Our full model, the AugAST–GSC model, outperforms the other models tested and does comparatively well at maintaining accuracy between the seen and unseen test repos on the Variable Naming task. + +To some degree the improved performance from adding the GSC is unsurprising: its addition to a graph–based model is essentially just adding extra features and doesn’t remove any information or flexibility. Under a satisfactory training regime, a model could simply learn to ignore it if it is unhelpful, so its inclusion should never hurt performance. The degree to which it helps, though, especially on the Variable Naming task, suggests that a GSC is well worth using for some tasks. Moreover, the fact that the Pointer Sentinel approach shown in Table 4 performs noticeably less well than the full GSC approach suggests that the relational aspect of the GSC is key: simply having the ability to output out–of–vocabulary words without relational information about their usage appears to be much less helpful. + +The downsides of using a GSC are thus primarily computational. Our GSC models ran about $30 \%$ slower than the Closed Vocab models. Since we capped the graph size at 500 nodes, the slowdown is presumably due to the large number of edges to and from the graph cache nodes. Better support for sparse operations on GPU in deep learning frameworks would be useful for alleviating this downside. + +In the near term, there remain a number of design choices to explore regarding AST– and GNN– models for processing source code. Adding information about word order to the GSC might improve performance, as might constructing the vocabulary out of subwords rather than words. It also might help to treat variable types as the GSC treats words: storing them in a GSC and connecting them with edges to the variables of those types; this could be particularly useful when working with code snippets rather than fully compilable code. For the Variable Naming task, there are also many architecture choices to be explored in how to produce a sequence of words for a name: how to unroll the RNN, what to use as the initial hidden state, etc. + +In the longer term, given that all results above show that augmenting ASTs with data– and control– flow edges improves performance, it would be worth exploring other static analysis concepts from the Programming Language and Software Verification literatures and seeing whether they could be usefully incorporated into Augmented ASTs. Better understanding of how Graph Neural Networks learn is also crucial, since they are central to the performance of our model and many others. Additionally, the entire domain of machine learning on source code faces the practical issue that many of the best data for supervised learning on source code — things like high–quality code reviews, integration test results, code with high test coverage, etc. — are not available outside private organizations. + +# REFERENCES + +Miltiadis Allamanis, Earl T. Barr, Premkumar Devanbu, and Charles Sutton. A survey of machine learning for big code and naturalness. arXiv:1709.06182 [cs], 2017. URL https://arxiv. org/abs/1709.06182. +Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs with graphs. In International Conference on Learning Representations, 2018. URL https: //openreview.net/forum?id $=$ BJOFETxR-. +Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representation for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New York, NY, USA, 2018. ACM. ISBN 978-1-4503-5698-5. doi: 10.1145/3192366.3192412. URL http: //doi.acm.org/10.1145/3192366.3192412. +Matthew Amodio, Swarat Chaudhuri, and Thomas Reps. Neural Attribute Machines for Program Generation. arXiv:1705.09231 [cs], May 2017. URL http://arxiv.org/abs/1705. 09231. +Forough Arabshahi, Sameer Singh, and Animashree Anandkumar. Combining symbolic expressions and black-box function evaluations for training neural programs. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ Hksj2WWAW. +Trapit Bansal, Arvind Neelakantan, and Andrew McCallum. Relnet: End-to-end modeling of entities & relations. arXiv:1706.07179 [cs], 2017. URL http://arxiv.org/abs/1706. 07179. +Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vin´ıcius Flores Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, aglar Gulehre, Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, ¨ Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matthew Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. abs/1806.01261, 2018. URL https: //arxiv.org/abs/1806.01261. + +Avishkar Bhoopchand, Tim Rocktaschel, Earl T. Barr, and Sebastian Riedel. Learning python code ¨ suggestion with a sparse pointer network. abs/1611.08307, 2016. URL https://arxiv. org/abs/1611.08307. + +Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 5:135–146, 2017. + +Marc Brockschmidt, Miltiadis Allamanis, Alexander L. Gaunt, and Oleksandr Polozov. Generative code modeling with graphs. abs/1805.08490, 2018. URL https://arxiv.org/abs/ 1805.08490. + +Kwonsoo Chae, Hakjoo Oh, Kihong Heo, and Hongseok Yang. Automatically generating features for learning program analysis heuristics for c-like languages. Proc. ACM Program. Lang., 1 (OOPSLA):101:1–101:25, October 2017. ISSN 2475-1421. doi: 10.1145/3133925. URL http: //doi.acm.org/10.1145/3133925. + +Xinyun Chen, Chang Liu, and Dawn Song. Tree-to-tree neural networks for program translation, 2018. URL https://openreview.net/forum?id ${ . } = { }$ rkxY-sl0W. + +Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1724–1734. Association for Computational Linguistics, 2014. doi: 10.3115/v1/D14-1179. URL http://www.aclweb.org/anthology/ D14-1179. + +Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp. 1263–1272, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR. URL http://proceedings.mlr.press/v70/gilmer17a.html. + +Github. The State of the Octoverse 2017, 2017. URL https://octoverse.github.com/. + +C. Goller and A. Kuchler. Learning task-dependent distributed representations by backpropagation through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pp. 347–352 vol.1, Jun 1996. doi: 10.1109/ICNN.1996.548916. + +Edouard Grave, Moustapha Cisse, and Armand Joulin. Unbounded cache model for online language ´ modeling with open vocabulary. In NIPS, 2017. + +Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 140–149. Association for Computational Linguistics, 2016. doi: 10.18653/v1/P16-1014. URL http://www.aclweb.org/anthology/ P16-1014. + +Vincent J. Hellendoorn and Premkumar T. Devanbu. Are deep neural networks the best choice for modeling source code? In ESEC/SIGSOFT FSE, 2017. + +Daniel D. Johnson. Learning Graphical State Transitions. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id ${ . } =$ HJ0NvFzxl. + +Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16, pp. 2741–2749. AAAI Press, 2016. URL http://dl.acm.org/citation.cfm?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ 3016100.3016285. + +Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. URL https://arxiv.org/abs/1412. 6980. + +Sosuke Kobayashi, Ran Tian, Naoaki Okazaki, and Kentaro Inui. Dynamic Entity Representation with Max-pooling Improves Machine Reading. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego, California, June 2016. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/N16-1099. + +Jens Krinke. Identifying similar code with program dependence graphs. In Proceedings of the Eighth Working Conference on Reverse Engineering (WCRE’01), WCRE ’01, pp. 301–, Washington, DC, USA, 2001. IEEE Computer Society. ISBN 0-7695-1303-4. URL http://dl.acm. org/citation.cfm?id $= \mathfrak { C }$ 832308.837142. + +Jian Li, Yue Wang, Irwin King, and Michael R. Lyu. Code Completion with Neural Attention and Pointer Networks. arXiv:1711.09573 [cs], November 2017. URL http://arxiv.org/abs/ 1711.09573. + +Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated Graph Sequence Neural Networks. In International Conference on Learning Representations, 2016. URL https:// arxiv.org/abs/1511.05493. + +Chang Liu, Xin Wang, Richard Shin, Joseph E. Gonzalez, and Dawn Xiaodong Song. Neural code completion. 2017. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ rJbPBt9lg& noteId $\underline { { \underline { { \mathbf { \Pi } } } } }$ rJbPBt9lg. + +Cristina V. Lopes, Petr Maj, Pedro Henrique Martins, Vaibhav Saini, Di Yang, Jakub Zitny, Hitesh Sajnani, and Jan Vitek. Dej´ avu: a map of code duplicates on github. \` PACMPL, 1:84:1–84:28, 2017. + +Zhengdong Lu, Haotian Cui, Xianggen Liu, Yukun Yan, and Daqi Zheng. Object-oriented Neural Programming (OONP) for Document Understanding. arXiv:1709.08853 [cs], September 2017. URL http://arxiv.org/abs/1709.08853. arXiv: 1709.08853. + +Minh-Thang Luong and Christopher D. Manning. Achieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models. arXiv:1604.00788 [cs], April 2016. URL http://arxiv.org/abs/1604.00788. + +Chris J. Maddison and Daniel Tarlow. Structured generative models of natural source code. pp. II–649–II–657, 2014. URL http://dl.acm.org/citation.cfm?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ 3044805. 3044965. + +Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer Sentinel Mixture Models. In International Conference on Learning Representations, 2017. URL https:// openreview.net/pdf?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ Byj72udxe. + +Trang Pham, Truyen Tran, and Svetha Venkatesh. Graph Memory Networks for Molecular Activity Prediction. arXiv:1801.02622 [cs], 2018. URL http://arxiv.org/abs/1801.02622. + +Maxim Rabinovich, Mitchell Stern, and Dan Klein. Abstract syntax networks for code generation and semantic parsing. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1139–1149. Association for Computational Linguistics, 2017. doi: 10.18653/v1/P17-1105. URL http://www.aclweb.org/ anthology/P17-1105. + +Veselin Raychev, Martin Vechev, and Andreas Krause. Predicting program properties from ”big code”. In Proceedings of the 42Nd Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’15, pp. 111–124, New York, NY, USA, 2015. ACM. ISBN 978- 1-4503-3300-9. doi: 10.1145/2676726.2677009. URL http://doi.acm.org/10.1145/ 2676726.2677009. + +Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling Relational Data with Graph Convolutional Networks. arXiv:1703.06103 [cs, stat], March 2017. URL http://arxiv.org/abs/1703.06103. + +Kristof T. Schutt, Farhad Arbabzadah, Stefan Chmiela, Klaus R. M ¨ uller, and Alexandre Tkatchenko.¨ Quantum-chemical insights from deep tensor neural networks. In Nature communications, 2017. + +Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew $\mathrm { N g }$ , and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1631–1642, Seattle, Washington, USA, October 2013. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/D13-1170. + +TIOBE. TIOBE Index for September 2018, 2018. URL https://www.tiobe.com/ tiobe-index/. + +Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 2692–2700. Curran Associates, Inc., 2015. URL http://papers.nips. cc/paper/5866-pointer-networks.pdf. + +M. White, M. Tufano, C. Vendome, and D. Poshyvanyk. Deep learning code fragments for code clone detection. In 2016 31st IEEE/ACM International Conference on Automated Software Engineering (ASE), pp. 87–98, Sept 2016. URL http://www.cs.wm.edu/˜mtufano/ publications/C5.pdf. + +Pengcheng Yin and Graham Neubig. A syntactic neural model for general-purpose code generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 440–450. Association for Computational Linguistics, 2017. doi: 10.18653/v1/P17-1041. URL http://www.aclweb.org/anthology/P17-1041. + +Wojciech Zaremba, Karol Kurach, and Rob Fergus. Learning to discover efficient mathematical identities. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp. 1278–1286. Curran Associates, Inc., 2014. URL http://papers.nips.cc/paper/ 5350-learning-to-discover-efficient-mathematical-identities.pdf. + +Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 649– 657. Curran Associates, Inc., 2015. URL http://papers.nips.cc/paper/ 5782-character-level-convolutional-networks-for-text-classification. pdf. + +# APPENDIX + +ADDITIONAL EXPERIMENT INFORMATION + +Table 6: Extra information about performance on the Variable Naming task. Entries in this table are of the form “subword accuracy, edit distance, edit distance divided by real name length”. The edit distance is the mean of the character–wise Levenshtein distance between the produced name and the real name. + +
Closed VocabCharCNNPointer SentinelGSC (ours)
Seen reposAST AugAST0.30, 7.22, 0.94 0.26, 7.64, 0.940.28,8.67, 1.08 0.28, 7.46, 0.990.32, 8.00, 1.07 0.35, 6.51, 0.770.56, 3.87, 0.39 0.60, 3.68, 0.37
Unseen reposAST0.09,8.66,1.230.10, 8.82, 1.120.13, 9,39, 1.370.42, 4.81, 0.59
AugAST0.09, 8.34, 1.140.10, 8.16, 1.120.14, 8.03, 1.070.48, 4.28, 0.49
+ +DATASET INFORMATION + +Table 7: Repositories used in experiments. All were taken from the Maven repository (https://mvnrepository.com/). Entries are in the form “group/repository name/version”. + +
Seen Repos
com.fasterxml.jackson.core/jackson-core/2.9.5 com.h2database/h2/1.4.195 javax.enterprise/cdi-api/2.0 junit/junit/4.12 mysql/mysql-connector-java/6.0.6 org.apache.commons/commons-collections4/4.1 org.apache.commons/commons-math3/3.6.1 org.apache.commons/commons-pool2/2.5.0 org.apache.maven/maven-project/2.2.1 org.codehaus.plexus/plexus-utils/3.1.0 org.eclipse.jetty/jetty-server/9.4.9.v20180320
org.scalacheck/scalacheck_2.12/1.13.5 org.slf4j/slf4j-api/1.7.25 org.slf4j/slf4j-log4j12/1.7.25 Unseen Repos
org.javassist/javassist/3.22.0-GA joda-time/joda-time/2.9.9 org.mockito/mockito-core/2.17.0
+ +Table 8: Edge types used in Augmented ASTs. The initial AST is constructed using the AST and NEXT TOKEN edges, and then the remaining edges are added. In other words, the the “AST” model from Table 1 uses a graph that contains only the AST and NEXT TOKEN edge types (and WORD USE if it also uses a GSC), while the “AugAST” model contains all the edge types below. The reversed version of every edge is also added as its own type (e.g. reverse AST, reverse LAST READ) to let the GNN message passing occur in both directions. + +
Edge NameDescription
AST NEXT_TOKENThe edges used to construct the original AST. Edges added to the original AST that specify the left-to-
right ordering of the children of a node in the AST.These edges are necessary since ASTs have ordered children, but we are representing the AST as a directed multigraph.
COMPUTED_FROMConnects a node representing a variable on the left of an equality to those on the right. (E.g. edges from y to x and Z to X in x = y + z.) The same as in Allamanis et al. (2018).
LAST_READConnects a node representing a usage of a variable to all nodes in the AST at which that variable's value could have been last read from memory. The same as in Allamanis et al. (2018).
LAST_WRITEConnects a node representing a usage of a variable to all nodesin the AST at which that variable's value could have been last written to memory. The same as in Allamanis et al.
RETURNS_TO(2018). Points a node in a return statement to the node contain- ing the return type of the method. (E.g. x in return X gets an edge pointing to int in public static int
LAST_SCOPE_USEgetx(x).) Connects a node representing a variable to the node repre- senting the last time this variable's name was used in the text of the code (i.e. capturing information about the text, not the control flow),but only within lexical scope. This edge exists to try and give the non-GSC models as much
LAST_FIELD_LEXlexical information as possible to make them as comparable with the GSC model. Connects a field access (e.g. this.whatever or Foo .whatever) node to the last use of this.whatever (or to the variable's initialization,if it's the first use).This is not lexical-scope aware (and,in fact,can't be in Java,in
FIELDgeneral). Pointseach node representing a field access(e.g. this.whatever) to the node where that field was de-
WORD_USEclared. Points cache nodes to nodes representing variables in which the vocab word was used in the variable's name.
\ No newline at end of file diff --git a/parse/train/SkNSehA9FQ/SkNSehA9FQ_content_list.json b/parse/train/SkNSehA9FQ/SkNSehA9FQ_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2f91d7b857247e65fe9a3ad293c79e9d8917ce79 --- /dev/null +++ b/parse/train/SkNSehA9FQ/SkNSehA9FQ_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "OPEN VOCABULARY LEARNING ON SOURCE CODE WITH A GRAPH–STRUCTURED CACHE ", + "text_level": 1, + "bbox": [ + 176, + 99, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 236, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning models that take computer program source code as input typically use Natural Language Processing (NLP) techniques. However, a major challenge is that code is written using an open, rapidly changing vocabulary due to, e.g., the coinage of new variable and method names. Reasoning over such a vocabulary is not something for which most NLP methods are designed. We introduce a Graph–Structured Cache to address this problem; this cache contains a node for each new word the model encounters with edges connecting each word to its occurrences in the code. We find that combining this graph–structured cache strategy with recent Graph–Neural–Network–based models for supervised learning on code improves the models’ performance on a code completion task and a variable naming task — with over $100 \\%$ relative improvement on the latter — at the cost of a moderate increase in computation time. ", + "bbox": [ + 233, + 268, + 764, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 465, + 336, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer program source code is an abundant and accessible form of data from which machine learning algorithms could learn to perform many useful software development tasks, including variable name suggestion, code completion, bug finding, security vulnerability identification, code quality assessment, or automated test generation. But despite the similarities between natural language and source code, deep learning methods for Natural Language Processing (NLP) have not been straightforward to apply to learning problems on source code (Allamanis et al., 2017). ", + "bbox": [ + 174, + 498, + 825, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are many reasons for this, but two central ones are: ", + "bbox": [ + 174, + 589, + 549, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Code’s syntactic structure is unlike natural language. While code contains natural language words and phrases in order to be human–readable, code is not meant to be a read like a natural language text. Code is written in a rigid syntax with delimiters that may open and close dozens of lines apart; it consists in great part of references to faraway lines and different files; and it describes computations that proceed in an order often quite distinct from its written order. ", + "bbox": [ + 174, + 616, + 825, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Code is written using an open vocabulary. Natural language is mostly composed of words from a large but closed (a.k.a. fixed–size and unchanging) vocabulary. Standard NLP methods can thus perform well by fixing a large vocabulary of words before training, and labeling the few words they encounter outside this vocabulary as “unknown”. But in code every new variable, class, or method declared requires a name, and this abundance of names leads to the use of many obscure words: abbreviations, brand names, technical terms, etc.1 A model must be able to reason about these newly–coined words to understand code. ", + "bbox": [ + 176, + 693, + 825, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The second of these issues is significant. To give one indication: $28 \\%$ of variable names contain out–of–vocabulary words in the test set we use in our experiments below. But more broadly, the open vocabulary issue in code is an acute example of a fundamental challenge in machine learning: ", + "bbox": [ + 176, + 803, + 821, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "how to build models that can reason over unbounded domains of entities, sometimes called “open– set” learning. Despite this, the open vocabulary issue in source code has received relatively little attention in prior work. ", + "bbox": [ + 176, + 103, + 821, + 145 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first issue, in contrast, has been the focus of much prior work. A common strategy in these works is to represent source code as an Abstract Syntax Tree (AST) rather than as linear text. Once in this graph–structured format, code can be passed as input to models like Recursive Neural Networks or Graph Neural Networks (GNNs) that can, in principle, exploit the relational structure of their inputs and avoid the difficulties of reading code in linear order (Allamanis et al., 2018). ", + "bbox": [ + 174, + 152, + 825, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution: In this paper we extend such AST–based models for source code in order to address the open vocabulary issue. We do so by introducing a Graph–Structured Cache (GSC) to handle out–of–vocabulary words. The GSC represents vocabulary words as additional nodes in the AST as they are encountered and connects them with the edges to where they are used in the code. We then process the AST+GSC with a GNN to produce outputs. See Figure 1. ", + "bbox": [ + 174, + 229, + 825, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We empirically evaluated the utility of a Graph–Structured Cache on two tasks: a code completion (a.k.a. fill–in–the–blank) task and a variable naming task. We found that using a GSC improved performance on both tasks at the cost of an approximately $30 \\%$ increase in training time. More precisely: even when using hyperparameters optimized for the baseline model, adding a GSC to a baseline model improved its accuracy by at least $7 \\%$ on the fill–in–the–blank task and $103 \\%$ on the variable naming task. We also report a number of ablation results in which we carefully demonstrate the relative importance of each model component to a model’s performance. ", + "bbox": [ + 174, + 306, + 825, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 PRIOR WORK ", + "text_level": 1, + "bbox": [ + 176, + 426, + 318, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "REPRESENTING CODE AS A GRAPH ", + "text_level": 1, + "bbox": [ + 178, + 459, + 419, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given their prominence in the study of programming languages, Abstract Syntax Trees (ASTs) and parse trees are a natural choice for representing code and have been used extensively. Often models that operate on source code consume ASTs by linearizing them (usually with a depth–first traversal) as in Amodio et al. (2017), Liu et al. (2017), or Li et al. (2017) or by using AST paths as input features as in Alon et al. (2018), but they can also be processed by deep learning models that take graphs as input, as in White et al. (2016) and Chen et al. (2018) who use Recursive Neural Networks (RveNNs) (Goller & Kuchler, 1996) on ASTs. RveNNs are models that operate on tree– topology graphs, and have been used extensively for language modeling (Socher et al., 2013) and on domains similar to source code, like mathematical expressions (Zaremba et al., 2014; Arabshahi et al., 2018). They can be considered a special case of Message Passing Neural Networks (MPNNs) in the framework of Gilmer et al. (2017): in this analogy RveNNs are to Belief Propagation as MPNNs are to Loopy Belief Propagation. They can also be considered a special case of Graph Networks in the framework of Battaglia et al. (2018). ASTs also serve as a natural basis for models that generate code as output, as in Maddison & Tarlow (2014), Yin & Neubig (2017), Rabinovich et al. (2017), Chen et al. (2018), and Brockschmidt et al. (2018). ", + "bbox": [ + 174, + 486, + 825, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data–flow graphs are another type of graphical representation of source code with a long history (Krinke, 2001), and they have occasionally been used to featurize source code for machine learning (Chae et al., 2017). ", + "bbox": [ + 176, + 702, + 820, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most closely related to our work is the work of Allamanis et al. (2018), on which our model is heavily based. Allamanis et al. (2018) combine the data–flow graph and AST representation strategies for source code by representing code as an AST augmented with extra labeled edges indicating semantic information like data– and control–flow between variables. These augmentations yield a directed multigraph rather than just a tree,2 so in Allamanis et al. (2018) a variety of MPNN called a Gated Graph Neural Network (GGNN) (Li et al., 2016) is used to consume the Augmented AST and produce an output for a supervised learning task. ", + "bbox": [ + 174, + 750, + 825, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph-based models that are not based on ASTs are also sometimes used for analyzing source code, like Conditional Random Fields for joint variable name prediction in (Raychev et al., 2015). ", + "bbox": [ + 174, + 854, + 823, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "REASONING ABOUT OPEN SETS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 397, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The question of how to gracefully reason over an open vocabulary is longstanding in NLP. Character–level embeddings are a typical way deep learning models handle this issue, whether used on their own as in Kim et al. (2016), or in conjunction with word–level embedding Recurrent Neural Networks (RNNs) as in Luong & Manning (2016), or in conjunction with an $n _ { \\mathrm { ~ } }$ –gram model as in Bojanowski et al. (2017). Another approach is to learn new word embeddings on–the–fly from context (Kobayashi et al., 2016). Caching novel words, as we do in our model, is yet another strategy (Grave et al., 2017) and has been used to augment N–gram models for analyzing source code (Hellendoorn & Devanbu, 2017). ", + "bbox": [ + 173, + 130, + 825, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In terms of producing outputs over variable–sized input and outputs, also known as open–set learning, attention-based pointer mechanisms were introduced in Vinyals et al. (2015) and have been used for tasks on code, e.g. in Bhoopchand et al. (2016). Such methods have been used to great effect in NLP in e.g. Gulcehre et al. (2016) and Merity et al. (2017). The latter’s pointer sentinel mixture model is the direct inspiration for the readout function we use in the Variable Naming task below. ", + "bbox": [ + 174, + 247, + 825, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using graphs to represent arbitrary collections of entities and their relationships for processing by deep networks has been widely used (Johnson, 2017; Bansal et al., 2017; Pham et al., 2018; Lu et al., 2017), but to our knowledge we are the first to use a graph–building strategy for reasoning (at train and test time) about an open vocabulary of words. ", + "bbox": [ + 174, + 324, + 825, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 400, + 339, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 ABSTRACT SYNTAX TREES ", + "text_level": 1, + "bbox": [ + 176, + 430, + 403, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An Abstract Syntax Tree (AST) is a graph — specifically an ordered tree with labeled nodes — that is a representation of some written computer source code. There is a 1–to–1 relationship between source code and an AST of that source code, modulo comments and whitespace in the written source code. ", + "bbox": [ + 174, + 455, + 825, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically the leaves of an AST correspond to the tokens written in the source code, like variable and method names, while the non–leaf nodes represent syntactic language constructs like function calls or class definitions. The specific node labels and construction rules of ASTs can differ between or within languages. The first step in Figure 1 shows an example. ", + "bbox": [ + 174, + 518, + 825, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 GRAPH NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 590, + 415, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The term Graph Neural Network (GNN) refers to any deep, differentiable model that takes graphs as input. Many GNNs have been presented in the literature, and several nomenclatures have been proposed for describing the computations they perform, in particular in Gilmer et al. (2017) and Battaglia et al. (2018). Here we give a brief recapitulation of supervised learning with GNNs using the Message Passing Neural Network framework from Gilmer et al. (2017). ", + "bbox": [ + 174, + 617, + 825, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A GNN is trained using pairs $( G , y )$ where $G = ( V , E )$ is a graph defined by its vertices $V$ and edges $E$ , and $y$ is a label. $y$ can be any sort of mathematical object: scalar, vector, another graph, etc. In the most general case, each graph in the dataset can be a directed multigraph, each with a different number of nodes and different connectivity. In each graph, each vertex $v \\in V$ has associated features $\\mathbf { \\boldsymbol { x } } _ { v }$ , and each edge $( v , w ) \\in E$ has features $e _ { v w }$ . ", + "bbox": [ + 174, + 693, + 825, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A GNN produces a prediction $\\hat { y }$ for the label $y$ of a graph $G = ( V , E )$ by the following procedure: ", + "bbox": [ + 174, + 770, + 816, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. A function $S$ is used to initialize a hidden state vector $h _ { v } ^ { 0 }$ for each vertex $v \\in V$ as a function of the vertex’s features (e.g., if the $\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { v } }$ are words, $S$ could be a word embedding function): ", + "bbox": [ + 183, + 795, + 820, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/94e458fc91a8fca3fbeee77f4db187895e70db6ba937f66e287ccb86e7029472.jpg", + "text": "$$\n\\begin{array} { r } { h _ { v } ^ { 0 } = S ( \\pmb { x } _ { v } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 472, + 824, + 558, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. For each round $t$ out of $T$ total rounds: ", + "bbox": [ + 186, + 845, + 460, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(a) Each vertex $v \\in V$ receives the vector $\\boldsymbol { m } _ { v } ^ { t + 1 }$ , which is the sum of “messages” from its neighbors, each produced by a function $M _ { t }$ : ", + "bbox": [ + 210, + 863, + 825, + 891 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b7a1c5f247ce69866b2fcc30405e0a9b6fd08d99af662eb1e4578dfa11bca720.jpg", + "text": "$$\nm _ { v } ^ { t + 1 } = \\sum _ { w \\in \\mathrm { n e i g h b o r s ~ o f } v } M _ { t } ( h _ { v } ^ { t } , h _ { w } ^ { t } , e _ { v w } ) .\n$$", + "text_format": "latex", + "bbox": [ + 395, + 893, + 666, + 926 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(b) Each vertex $v \\in V$ updates its hidden state based on the message it received via a function $U _ { t }$ : ", + "bbox": [ + 205, + 103, + 825, + 131 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/3dbc90bc50fd06202e01f73e8218ed391fc7b8275fcd2ebea9880febecb6a67a.jpg", + "text": "$$\nh _ { v } ^ { t + 1 } = U _ { t } ( h _ { v } ^ { t } , m _ { v } ^ { t + 1 } ) .\n$$", + "text_format": "latex", + "bbox": [ + 452, + 131, + 611, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. A function $R$ , the “readout function”, produces a prediction based on the hidden states generated during the message passing (usually just those at from time $T$ ): ", + "bbox": [ + 181, + 155, + 823, + 184 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/4a007b03cf24477ac91e4a1b0db3625e7401553f5af46fc8f711e359018c39a2.jpg", + "text": "$$\n\\hat { y } = R ( \\{ h _ { v } ^ { t } | v \\in V , t \\in 1 , . . . , T \\} ) .\n$$", + "text_format": "latex", + "bbox": [ + 403, + 190, + 627, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GNNs differ in how they implement $S , M _ { t } , U _ { t }$ , and $R$ . But all these functions are differentiable and most are parameterized, so the model is trainable via stochastic gradient descent of a loss function on $y$ and $\\hat { y }$ . ", + "bbox": [ + 174, + 222, + 825, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 MODEL ", + "text_level": 1, + "bbox": [ + 174, + 285, + 269, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model consumes an input instance of source code and produces an output for a supervised learning task via the following five steps, sketched in Figure 1: ", + "bbox": [ + 176, + 319, + 823, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. Parse the source code (snippet, file, repository, version control history, etc.) into an Abstract Syntax Tree. \n2. Add edges of varying types (details in Appendix Table 8) to this AST that represent semantic information like data– and control– flow, in the spirit of Allamanis et al. (2018). Also add the reversed version of all edges with their own edge type. This results in a directed multigraph called an Augmented AST. \n3. Further augment the Augmented AST by adding a Graph–Structured Cache. That is, add a node to the Augmented AST for each vocabulary word encountered in the input instance. Then connect each such “cache node” with an edge (of edge type WORD USE) to all variables whose names contain its word. \n4. Vectorize the Augmented $\\mathrm { A S T } + \\mathrm { G S C }$ graph into a form suitable for a GNN. (I.e. perform Step 1 from Section 3.2.) Each AST node that doesn’t represent a variable is vectorized as a learned embedding of the language construct it represents, e.g. Parameter, Method Declaration, etc. Each cache node and each node that represents a variable is vectorized as a learned linear map of the concatenation of a type embedding and a name embedding. The name embedding is a Character–Level Convolutional Neural Network (CharCNN) (Zhang et al., 2015) embedding of the word/name the node contains. The type embedding is a learned embedding of the name of the Java type of the token it contains, e.g. int, a user–defined class, etc., with cache nodes having their own unique Cacher Node type. \n5. Process the graph with a GNN, as per Section 3.2. (I.e. perform Steps 2 and 3 from Section \n3.2.) The readout functions differ depending on the task and are described in the Experiments section below. ", + "bbox": [ + 184, + 358, + 826, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our main contribution to previous works is the addition of Step 3, the Graph–Structured Cache step. The combination of relational information from the cache nodes’ connections and lexical information from these nodes’ CharCNN embeddings allows the model to, in principle, flexibly reason about words it never saw during training, but also recognize words it did. E.g. it could potentially see a class named “getGuavaDictionary” and a variable named “guava dict” and both (a) utilize the fact that the word “guava” is common to both names despite having never seen this word before, and (b) exploit learned representations for words like “get”, “dictionary”, and “dict” that it has seen during training. ", + "bbox": [ + 173, + 702, + 825, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 835, + 326, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluated our model, described in Section 4, on two supervised tasks: a Fill–In–The–Blank task and a Variable Naming task. For each task, we compare our model to others that differ in how they parse the code and how they treat the words they encounter. Table 1 details the different variations of the procedure in Section 4 against which we compare our model. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa59b4133b5607587cc70ae7b40b4e6a20f3afab3cf1561f1e3c0ebf601145cb.jpg", + "image_caption": [ + "Figure 1: Our model’s procedure for consuming a single input instance of source code and producing an output for a supervised learning task. " + ], + "image_footnote": [], + "bbox": [ + 179, + 103, + 831, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Code to reproduce all experiments is available online.3 4 ", + "bbox": [ + 173, + 797, + 544, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 DATA AND IMPLEMENTATION DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 830, + 482, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We chose to use Java source code as the data for our experiments as it is among the most popular programming languages in use today (TIOBE, 2018; Github, 2017). To construct our dataset, we randomly selected 18 of the 100 most popular Java repos from the Maven repository5 to serve as training data. (See Appendix Table 7 for the list.) Together these repositories contain about 500,000 non–empty, non–comment lines of code. We checked for excessive code duplication in our dataset (Lopes et al., 2017) using $\\mathrm { C P D ^ { 6 } }$ and found only about $7 \\%$ of the lines to be contiguous, duplicated code blocks containing more than 150 tokens. ", + "bbox": [ + 176, + 857, + 825, + 886 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c127588b501d99382c3ea3beceafe747f6da712578b6b15fbdf67ad98a5f51cb.jpg", + "table_caption": [ + "Table 1: Nomenclature used in Experiments section. Each abbreviation describes a tweak/ablation to our full model as presented in Section 4. Using this nomenclature, our full model as described in Section 4 and shown in Figure 1 would be a “AugAST–GSC” model. " + ], + "table_footnote": [], + "table_body": "
AbbreviationMeaning
Code RepresentationASTSkips Step 2 in Section 4.
AugASTPerforms Step 2 in Section 4.
Vocab StrategiesClosed VocabSkips Step 3 in Section 4,and instead maintains word-embedding vectors for words in a closed vo- cabulary. In Step 4, name embeddings for nodes representing variables are produced by taking the mean of the embeddings of the words in the vari- able's name.Words outside this model's closed vocabulary are labeled as <UNK>. This is the
CharCNNstrategy used in Allamanis et al. (2018). Skips Step 3 in Section 4.
Pointer SentinelFollows Steps 3 and 4 as described in Section 4,except it doesn't add edges connecting cache nodes to the nodes where their word is used. In the Variable Naming task, this is equivalent to us- ing the Pointer Sentinel Mixture Model of Merity
GSCet al. (2017) to produce outputs. Follows Steps 3 and 4 as described in Section 4.
Graph Neural NetworkGGNNPerforms Step 5 in Section 4 using the Gated Graph Neural Network of Li et al. (2016).
DTNNPerforms Step 5 in Section 4 using the Deep Ten- sor Neural Network of Schuitt et al. (2017).
RGCNPerforms Step 5 in Section 4 using the Relational Graph Convolutional Network of Schlichtkrull et al. (2017).
", + "bbox": [ + 176, + 155, + 823, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 593, + 825, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We randomly chose 3 of these repositories to sequester as an “Unseen Repos” test set. We then separated out $15 \\%$ of the files in the remaining 15 repositories to serve as our “Seen Repos” test set. The remaining files served as our training set, from which we separated $15 \\%$ of the datapoints to act as a validation set. ", + "bbox": [ + 174, + 670, + 825, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our data preprocessor builds on top of the open–source Javaparser7 library to generate ASTs of our source code and then augment the ASTs with the edges described in Appendix Table 8. We used Apache $\\mathrm { { \\bf M X N e t } } ^ { 8 }$ as our deep learning framework. All hidden states in the GNN contained 64 units; all GNNs ran for 8 rounds of message passing; all models were optimized using the Adam optimizer (Kingma & Ba, 2015); all inputs to the GNNs were truncated to a maximum size of 500 nodes centered on the or $\\mathrm { \\ < N A M E - M E > }$ tokens. About $53 \\%$ of input graphs were larger than 500 nodes before truncation. The only regularization we used was early stopping — early in our experiments we briefly tried $L _ { 2 }$ and dropout regularization, but saw no effects. ", + "bbox": [ + 174, + 733, + 826, + 858 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We performed only a moderate amount of hyperparameter optimization, but all of it was done on the baseline models to avoid biasing our results in favor of our model. Specifically, we tuned all hyperparameters on the Closed Vocab baseline model, and also did a small amount of extra learning rate exploration for the Pointer Sentinel baseline model to try to maximize its performance. ", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 THE FILL–IN–THE–BLANK TASK ", + "text_level": 1, + "bbox": [ + 176, + 176, + 447, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this task we randomly selected a single usage of a variable in some source code, replaced it with a token, and then asked the model to predict what variable should have been there. An example instance from our dataset is shown in Figure 2. ", + "bbox": [ + 174, + 202, + 826, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69398b3fef9fe3f5b523898c2773311a427c29f26873c4f47ff60001f36d5b3c.jpg", + "image_caption": [ + "Figure 2: Example of a model’s procedure for completing the Fill–In–The–Blank task. Each Fill– In–The–Blank instance is created by replacing a single usage of a variable (n, in this example) with the special token . The model then processes the code as depicted in Figure 1. To produce outputs, the model’s readout function computes a soft–attention weighting over all nodes in the graph; the model’s output is the variable at the node on which it places maximal attention. In this example, if the model put maximal attention weighting on any of the green– highlighted variables, this would be a correct output. If maximal attention is placed on any other node, it would be an incorrect output. Only in–scope usages of a variable are counted as correct. " + ], + "image_footnote": [], + "bbox": [ + 179, + 257, + 834, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The models indicate their prediction for what variable should go in the blank by pointing with neural attention over all the nodes in the AugAST. This means all training and test instances only considered cases where the obfuscated variable appears somewhere else in the code. Single uses are rare however, since in Java variables must be declared before they are used. It also means there are sometimes multiple usages of the same, correct variable to which a model can point to get the right answer. In our dataset $78 \\%$ of variables were used more two times, and $33 \\%$ were used more than four times. ", + "bbox": [ + 173, + 525, + 825, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The models compute the attention weightings $y _ { i }$ for each Augmented AST node $i$ differently depending on the readout function of the GNN they use. Models using a GGNN as their GNN component, as all those in Table $2 \\ : \\mathrm { d o }$ , compute the attention weightings as per Li et al. (2016): ", + "bbox": [ + 173, + 630, + 821, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/a925b395ee215cf84d388e8c4326652c58f7e2ee0e0385f03d8bdf851612fd8a.jpg", + "text": "$$\n\\begin{array} { r } { \\hat { y } _ { i } = \\sigma \\left( f _ { 1 } ( { h } _ { v } ^ { T } , { h } _ { v } ^ { 0 } ) \\right) \\odot f _ { 2 } ( { h } _ { v } ^ { T } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 390, + 674, + 604, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the $f$ s are MLPs, $\\boldsymbol { h } _ { v } ^ { t }$ is the hidden state of node $v$ after $t$ message passing iterations, $\\sigma$ is the sigmoid function, and $\\odot$ is elementwise multiplication. The DTNN and RGCN GNNs compute the attention weightings as per Schutt et al. (2017): ¨ ", + "bbox": [ + 174, + 696, + 825, + 738 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/6c6afc591cf1bccd6c9f965cb2e5cf87ed755236782ed9c75f2680482e78247c.jpg", + "text": "$$\n\\hat { y } _ { i } = f ( h _ { v } ^ { T } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 454, + 741, + 540, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $f$ is a single hidden layer MLP. The models were trained using a binary cross entropy loss computed across the nodes in the graph. ", + "bbox": [ + 174, + 762, + 823, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance of models using our GSC versus those using other methods is reported in Table 2. For context, a baseline strategy of random guessing among all variable nodes within an edge radius of 8 of the $ token achieves an accuracy of 0.22. We also compare the performance of different GNNs in Table 3. ", + "bbox": [ + 173, + 796, + 825, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 THE VARIABLE NAMING TASK ", + "text_level": 1, + "bbox": [ + 176, + 869, + 428, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this task we replaced all usages of a name of a particular variable, method, class, or parameter in the code with the special token $ , and asked the model to produce the obfuscated name (in the form of the sequence of words that compose the name). An example instance from our dataset is shown in Figure 3. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64e9ca199b864180f85246fd6d2fa82eed5b992d8efb9255253967146629d494.jpg", + "table_caption": [ + "Table 2: Accuracy on the Fill–In–The–Blank task. Our model is the AugAST–GSC. The first number in each cell is the accuracy of the model, where a correct prediction is one in which the graph node that received the maximum attention weighting by the model contained the variable that was originally in the $ spot. The second, parenthetical numbers are the top– 5 accuracies, i.e. whether the correct node was among those that received the 5 largest attentions weightings from the model. See Table 1 for explanations of the abbreviations. All models use Gated Graph Neural Networks as their GNN component. " + ], + "table_footnote": [], + "table_body": "
Closed VocabCharCNNGSC
Seen reposAST0.57 (0.83)0.60 (0.84)0.89 (0.96)
AugAST0.80 (0.90)0.90 (0.94)0.97 (0.99)
Unseen reposAST AugAST0.36 (0.68)0.48 (0.80)0.80 (0.93)
0.59 (0.78)0.84 (0.92)0.92 (0.96)
", + "bbox": [ + 263, + 210, + 733, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/407aa777975ba994adc29c62d255d66c74058346b9d028a55509f55635c5f2cf.jpg", + "table_caption": [ + "Table 3: Accuracy (and top–5 accuracy) on the Fill–In–The–Blank task, depending on which type of GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST as their code representation. " + ], + "table_footnote": [], + "table_body": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.80 (0.90) 0.97 (0.99)0.72 (0.84) 0.89 (0.95)0.80 (0.90) 0.95 (0.98)
Unseen reposClosed Vocab GSC0.59 (0.78) 0.92 (0.96)0.46 (0.68) 0.80 (0.89)0.62 (0.79) 0.88 (0.95)
", + "bbox": [ + 256, + 366, + 741, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 481, + 825, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b21bc85d57d3da138b5330d252092f70b403cbab07c31d848435d51028931d7a.jpg", + "image_caption": [ + "Figure 3: Example of a model’s procedure for completing the Variable Naming task. Each Variable Naming instance is created by replacing all uses of some variable (expectedLength, in this example) with a special $\\mathrm { \\ < N A M E - M E > }$ token. The model then processes the code as depicted in Figure 1. To produce outputs, the model takes the mean of the nodes’ hidden states (depicted here in orange), uses them as the initial hidden state of a Recurrent Neural Network, and unrolls this RNN to produce a name as a sequence of words. " + ], + "image_footnote": [], + "bbox": [ + 176, + 525, + 836, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To produce a name from the output of the GNN, our models used the readout function of Allamanis et al. (2018). This readout function computes the mean of the hidden states of the $\\mathrm { < N A M E - M E > }$ nodes and passing it as the initial hidden state to a 1–layer Gated Recurrent Unit (GRU) RNN (Cho et al., 2014). This GRU is then unrolled to produce words in its predicted name, in the style of a traditional NLP decoder. We used a fixed length unrolling of 8 words, as $9 9 . 8 \\%$ of names in our training set were 8 or fewer words long. The models were trained by cross entropy loss over the sequence of words in the name. ", + "bbox": [ + 173, + 762, + 825, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To decode each hidden state output of the GRU $^ { h }$ into a probability distribution $P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )$ over words w, the Closed Vocab and CharCNN models pass $^ { h }$ through a linear layer and a softmax layer with output dimension equal to the number of words in their closed vocabularies (i.e. a traditional decoder output for NLP). In contrast, the GSC model not only has access to a fixed–size vocabulary but can also produce words by pointing to cache nodes in its Graph–Structured Cache. Specifically, it uses a decoder architecture inspired by the Pointer Sentinel Mixture Model of Merity et al. (2017): the probability of a word w being the GSC decoder’s output given that the GRU’s hidden state was $^ { h }$ is ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 159 + ], + "page_idx": 8 + }, + { + "type": "equation", + "img_path": "images/e033b845db19ed064700348954e36074ceecf2517c8ee2467099c2046acd4d6e.jpg", + "text": "$$\nP ( \\mathbf { w } | h ) = P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h ) + ( 1 - P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) ) P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )\n$$", + "text_format": "latex", + "bbox": [ + 276, + 159, + 722, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $P _ { \\mathrm { g r a p h } } ( \\cdot | h )$ is a conditional probability distribution over cache nodes in the GSC and the sentinel s, and $P _ { \\mathrm { v o c a b } } ( \\cdot | h )$ is a conditional probability distribution over words in a closed vocabulary. $P _ { \\mathrm { g r a p h } } ( \\cdot | h )$ is computed by passing the hidden states of all cache nodes and the sentinel node through a single linear layer and then computing the softmax dot–product attention of these values with $^ { h }$ . $P _ { \\mathrm { v o c a b } } ( \\cdot | h )$ is computed as the softmax of a linear mapping of $^ { h }$ to indices in a closed vocabulary, as in the Closed Vocab and CharCNN models. If there is no cache node for w in the Augmented AST or if w is not in the model’s closed dictionary then $P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h )$ and $P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )$ are 0, respectively. ", + "bbox": [ + 173, + 180, + 825, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The performance of our GSC versus other methods is reported in Table 4. More granular performance statistics are reported in Appendix Table 6. We also compare the performance of different GNNs in Table 5. ", + "bbox": [ + 174, + 285, + 825, + 327 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5af4ae8799d6bb4f9a24558aaf1823daccb2849b869291d358fa6d0540c34964.jpg", + "table_caption": [ + "Table 4: Accuracy on the Variable Naming task. Our model is the AugAST–GSC. The first number in each cell is the accuracy of the model, where we consider a correct output to be exact reproduction of the full name of the obfuscated variable (i.e. all the words in the name and then a EOS token). The second, parenthetical numbers are the top–5 accuracies, i.e. whether the correct full name was among the 5 most probable sequences output by the model. See Table 1 for explanations of the abbreviations. All models use Gated Graph Neural Networks as their GNN component. " + ], + "table_footnote": [], + "table_body": "
Closed VocabCharCNNPointer SentinelGSC
Seen reposAST0.23 (0.31)0.22 (0.28)0.19 (0.33)0.49 (0.67)
AugAST0.19 (0.26)0.20 (0.27)0.26 (0.40)0.53 (0.69)
Unseen reposAST0.05 (0.07)0.06 (0.09)0.06 (0.11)0.38 (0.53)
AugAST0.04 (0.07)0.06 (0.08)0.08 (0.14)0.41 (0.57)
", + "bbox": [ + 200, + 434, + 797, + 522 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/373973edcaadd5d2a832f2390824e23ed065f27ac06f83ad1f85a9edf60c7002.jpg", + "table_caption": [ + "Table 5: Accuracy (and top–5 accuracy) on the Variable Naming task, depending on which type of GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST as their code representation. " + ], + "table_footnote": [], + "table_body": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.19 (0.26)0.23 (0.31)0.27 (0.34)
0.53 (0.69)0.33 (0.48)0.46 (0.63)
Unseen reposClosed Vocab GSC0.04 (0.07) 0.41 (0.57)0.06 (0.08) 0.25 (0.40)0.06 (0.09) 0.35 (0.49)
", + "bbox": [ + 254, + 598, + 741, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 DISCUSSION ", + "text_level": 1, + "bbox": [ + 174, + 717, + 310, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As can be seen in Tables 2 and 4, the addition of a GSC improved performance on all tasks. Our full model, the AugAST–GSC model, outperforms the other models tested and does comparatively well at maintaining accuracy between the seen and unseen test repos on the Variable Naming task. ", + "bbox": [ + 176, + 750, + 823, + 791 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To some degree the improved performance from adding the GSC is unsurprising: its addition to a graph–based model is essentially just adding extra features and doesn’t remove any information or flexibility. Under a satisfactory training regime, a model could simply learn to ignore it if it is unhelpful, so its inclusion should never hurt performance. The degree to which it helps, though, especially on the Variable Naming task, suggests that a GSC is well worth using for some tasks. Moreover, the fact that the Pointer Sentinel approach shown in Table 4 performs noticeably less well than the full GSC approach suggests that the relational aspect of the GSC is key: simply having the ability to output out–of–vocabulary words without relational information about their usage appears to be much less helpful. ", + "bbox": [ + 173, + 797, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The downsides of using a GSC are thus primarily computational. Our GSC models ran about $30 \\%$ slower than the Closed Vocab models. Since we capped the graph size at 500 nodes, the slowdown is presumably due to the large number of edges to and from the graph cache nodes. Better support for sparse operations on GPU in deep learning frameworks would be useful for alleviating this downside. ", + "bbox": [ + 174, + 103, + 823, + 172 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the near term, there remain a number of design choices to explore regarding AST– and GNN– models for processing source code. Adding information about word order to the GSC might improve performance, as might constructing the vocabulary out of subwords rather than words. It also might help to treat variable types as the GSC treats words: storing them in a GSC and connecting them with edges to the variables of those types; this could be particularly useful when working with code snippets rather than fully compilable code. For the Variable Naming task, there are also many architecture choices to be explored in how to produce a sequence of words for a name: how to unroll the RNN, what to use as the initial hidden state, etc. ", + "bbox": [ + 174, + 180, + 825, + 291 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the longer term, given that all results above show that augmenting ASTs with data– and control– flow edges improves performance, it would be worth exploring other static analysis concepts from the Programming Language and Software Verification literatures and seeing whether they could be usefully incorporated into Augmented ASTs. Better understanding of how Graph Neural Networks learn is also crucial, since they are central to the performance of our model and many others. Additionally, the entire domain of machine learning on source code faces the practical issue that many of the best data for supervised learning on source code — things like high–quality code reviews, integration test results, code with high test coverage, etc. — are not available outside private organizations. ", + "bbox": [ + 174, + 299, + 825, + 424 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 445, + 285, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Miltiadis Allamanis, Earl T. Barr, Premkumar Devanbu, and Charles Sutton. A survey of machine learning for big code and naturalness. arXiv:1709.06182 [cs], 2017. URL https://arxiv. org/abs/1709.06182. \nMiltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs with graphs. In International Conference on Learning Representations, 2018. URL https: //openreview.net/forum?id $=$ BJOFETxR-. \nUri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representation for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New York, NY, USA, 2018. ACM. ISBN 978-1-4503-5698-5. doi: 10.1145/3192366.3192412. URL http: //doi.acm.org/10.1145/3192366.3192412. \nMatthew Amodio, Swarat Chaudhuri, and Thomas Reps. Neural Attribute Machines for Program Generation. arXiv:1705.09231 [cs], May 2017. URL http://arxiv.org/abs/1705. 09231. \nForough Arabshahi, Sameer Singh, and Animashree Anandkumar. Combining symbolic expressions and black-box function evaluations for training neural programs. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ Hksj2WWAW. \nTrapit Bansal, Arvind Neelakantan, and Andrew McCallum. Relnet: End-to-end modeling of entities & relations. arXiv:1706.07179 [cs], 2017. URL http://arxiv.org/abs/1706. 07179. \nPeter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vin´ıcius Flores Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, aglar Gulehre, Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, ¨ Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matthew Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational inductive biases, deep learning, and graph networks. abs/1806.01261, 2018. URL https: //arxiv.org/abs/1806.01261. ", + "bbox": [ + 171, + 460, + 825, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Avishkar Bhoopchand, Tim Rocktaschel, Earl T. Barr, and Sebastian Riedel. Learning python code ¨ suggestion with a sparse pointer network. abs/1611.08307, 2016. URL https://arxiv. org/abs/1611.08307. ", + "bbox": [ + 174, + 103, + 821, + 145 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 5:135–146, 2017. ", + "bbox": [ + 174, + 156, + 821, + 185 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Marc Brockschmidt, Miltiadis Allamanis, Alexander L. Gaunt, and Oleksandr Polozov. Generative code modeling with graphs. abs/1805.08490, 2018. URL https://arxiv.org/abs/ 1805.08490. ", + "bbox": [ + 173, + 195, + 825, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kwonsoo Chae, Hakjoo Oh, Kihong Heo, and Hongseok Yang. Automatically generating features for learning program analysis heuristics for c-like languages. Proc. ACM Program. Lang., 1 (OOPSLA):101:1–101:25, October 2017. ISSN 2475-1421. doi: 10.1145/3133925. URL http: //doi.acm.org/10.1145/3133925. ", + "bbox": [ + 173, + 250, + 825, + 306 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xinyun Chen, Chang Liu, and Dawn Song. Tree-to-tree neural networks for program translation, 2018. URL https://openreview.net/forum?id ${ . } = { }$ rkxY-sl0W. ", + "bbox": [ + 171, + 318, + 823, + 347 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1724–1734. Association for Computational Linguistics, 2014. doi: 10.3115/v1/D14-1179. URL http://www.aclweb.org/anthology/ D14-1179. ", + "bbox": [ + 173, + 357, + 825, + 440 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp. 1263–1272, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR. URL http://proceedings.mlr.press/v70/gilmer17a.html. ", + "bbox": [ + 173, + 452, + 825, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Github. The State of the Octoverse 2017, 2017. URL https://octoverse.github.com/. ", + "bbox": [ + 171, + 534, + 816, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C. Goller and A. Kuchler. Learning task-dependent distributed representations by backpropagation through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pp. 347–352 vol.1, Jun 1996. doi: 10.1109/ICNN.1996.548916. ", + "bbox": [ + 174, + 559, + 825, + 602 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Edouard Grave, Moustapha Cisse, and Armand Joulin. Unbounded cache model for online language ´ modeling with open vocabulary. In NIPS, 2017. ", + "bbox": [ + 173, + 613, + 823, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 140–149. Association for Computational Linguistics, 2016. doi: 10.18653/v1/P16-1014. URL http://www.aclweb.org/anthology/ P16-1014. ", + "bbox": [ + 173, + 652, + 825, + 723 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Vincent J. Hellendoorn and Premkumar T. Devanbu. Are deep neural networks the best choice for modeling source code? In ESEC/SIGSOFT FSE, 2017. ", + "bbox": [ + 173, + 734, + 823, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Daniel D. Johnson. Learning Graphical State Transitions. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id ${ . } =$ HJ0NvFzxl. ", + "bbox": [ + 174, + 773, + 823, + 803 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16, pp. 2741–2749. AAAI Press, 2016. URL http://dl.acm.org/citation.cfm?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ 3016100.3016285. ", + "bbox": [ + 174, + 814, + 825, + 869 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. URL https://arxiv.org/abs/1412. 6980. ", + "bbox": [ + 174, + 882, + 823, + 922 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sosuke Kobayashi, Ran Tian, Naoaki Okazaki, and Kentaro Inui. Dynamic Entity Representation with Max-pooling Improves Machine Reading. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego, California, June 2016. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/N16-1099. ", + "bbox": [ + 174, + 103, + 826, + 174 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jens Krinke. Identifying similar code with program dependence graphs. In Proceedings of the Eighth Working Conference on Reverse Engineering (WCRE’01), WCRE ’01, pp. 301–, Washington, DC, USA, 2001. IEEE Computer Society. ISBN 0-7695-1303-4. URL http://dl.acm. org/citation.cfm?id $= \\mathfrak { C }$ 832308.837142. ", + "bbox": [ + 173, + 181, + 825, + 238 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jian Li, Yue Wang, Irwin King, and Michael R. Lyu. Code Completion with Neural Attention and Pointer Networks. arXiv:1711.09573 [cs], November 2017. URL http://arxiv.org/abs/ 1711.09573. ", + "bbox": [ + 173, + 246, + 825, + 289 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated Graph Sequence Neural Networks. In International Conference on Learning Representations, 2016. URL https:// arxiv.org/abs/1511.05493. ", + "bbox": [ + 174, + 296, + 825, + 339 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chang Liu, Xin Wang, Richard Shin, Joseph E. Gonzalez, and Dawn Xiaodong Song. Neural code completion. 2017. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ rJbPBt9lg& noteId $\\underline { { \\underline { { \\mathbf { \\Pi } } } } }$ rJbPBt9lg. ", + "bbox": [ + 173, + 347, + 825, + 390 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cristina V. Lopes, Petr Maj, Pedro Henrique Martins, Vaibhav Saini, Di Yang, Jakub Zitny, Hitesh Sajnani, and Jan Vitek. Dej´ avu: a map of code duplicates on github. \\` PACMPL, 1:84:1–84:28, 2017. ", + "bbox": [ + 174, + 397, + 825, + 440 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Zhengdong Lu, Haotian Cui, Xianggen Liu, Yukun Yan, and Daqi Zheng. Object-oriented Neural Programming (OONP) for Document Understanding. arXiv:1709.08853 [cs], September 2017. URL http://arxiv.org/abs/1709.08853. arXiv: 1709.08853. ", + "bbox": [ + 173, + 448, + 825, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Minh-Thang Luong and Christopher D. Manning. Achieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models. arXiv:1604.00788 [cs], April 2016. URL http://arxiv.org/abs/1604.00788. ", + "bbox": [ + 173, + 498, + 825, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chris J. Maddison and Daniel Tarlow. Structured generative models of natural source code. pp. II–649–II–657, 2014. URL http://dl.acm.org/citation.cfm?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ 3044805. 3044965. ", + "bbox": [ + 173, + 549, + 823, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer Sentinel Mixture Models. In International Conference on Learning Representations, 2017. URL https:// openreview.net/pdf?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ Byj72udxe. ", + "bbox": [ + 174, + 601, + 823, + 643 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Trang Pham, Truyen Tran, and Svetha Venkatesh. Graph Memory Networks for Molecular Activity Prediction. arXiv:1801.02622 [cs], 2018. URL http://arxiv.org/abs/1801.02622. ", + "bbox": [ + 173, + 651, + 823, + 681 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Maxim Rabinovich, Mitchell Stern, and Dan Klein. Abstract syntax networks for code generation and semantic parsing. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1139–1149. Association for Computational Linguistics, 2017. doi: 10.18653/v1/P17-1105. URL http://www.aclweb.org/ anthology/P17-1105. ", + "bbox": [ + 173, + 688, + 825, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Veselin Raychev, Martin Vechev, and Andreas Krause. Predicting program properties from ”big code”. In Proceedings of the 42Nd Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’15, pp. 111–124, New York, NY, USA, 2015. ACM. ISBN 978- 1-4503-3300-9. doi: 10.1145/2676726.2677009. URL http://doi.acm.org/10.1145/ 2676726.2677009. ", + "bbox": [ + 173, + 766, + 825, + 837 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling Relational Data with Graph Convolutional Networks. arXiv:1703.06103 [cs, stat], March 2017. URL http://arxiv.org/abs/1703.06103. ", + "bbox": [ + 174, + 844, + 823, + 887 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kristof T. Schutt, Farhad Arbabzadah, Stefan Chmiela, Klaus R. M ¨ uller, and Alexandre Tkatchenko.¨ Quantum-chemical insights from deep tensor neural networks. In Nature communications, 2017. ", + "bbox": [ + 176, + 895, + 821, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew $\\mathrm { N g }$ , and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1631–1642, Seattle, Washington, USA, October 2013. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/D13-1170. ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "TIOBE. TIOBE Index for September 2018, 2018. URL https://www.tiobe.com/ tiobe-index/. ", + "bbox": [ + 174, + 181, + 821, + 212 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 2692–2700. Curran Associates, Inc., 2015. URL http://papers.nips. cc/paper/5866-pointer-networks.pdf. ", + "bbox": [ + 174, + 219, + 825, + 277 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "M. White, M. Tufano, C. Vendome, and D. Poshyvanyk. Deep learning code fragments for code clone detection. In 2016 31st IEEE/ACM International Conference on Automated Software Engineering (ASE), pp. 87–98, Sept 2016. URL http://www.cs.wm.edu/˜mtufano/ publications/C5.pdf. ", + "bbox": [ + 173, + 285, + 825, + 342 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Pengcheng Yin and Graham Neubig. A syntactic neural model for general-purpose code generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 440–450. Association for Computational Linguistics, 2017. doi: 10.18653/v1/P17-1041. URL http://www.aclweb.org/anthology/P17-1041. ", + "bbox": [ + 173, + 351, + 825, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Wojciech Zaremba, Karol Kurach, and Rob Fergus. Learning to discover efficient mathematical identities. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp. 1278–1286. Curran Associates, Inc., 2014. URL http://papers.nips.cc/paper/ 5350-learning-to-discover-efficient-mathematical-identities.pdf. ", + "bbox": [ + 174, + 416, + 825, + 486 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 649– 657. Curran Associates, Inc., 2015. URL http://papers.nips.cc/paper/ 5782-character-level-convolutional-networks-for-text-classification. pdf. ", + "bbox": [ + 173, + 496, + 861, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 103, + 264, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ADDITIONAL EXPERIMENT INFORMATION ", + "bbox": [ + 176, + 133, + 467, + 148 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/eede5588c6566c88a1a2b4aaa322127dac62da0a73f053e503ae0a31e2d94f9c.jpg", + "table_caption": [ + "Table 6: Extra information about performance on the Variable Naming task. Entries in this table are of the form “subword accuracy, edit distance, edit distance divided by real name length”. The edit distance is the mean of the character–wise Levenshtein distance between the produced name and the real name. " + ], + "table_footnote": [], + "table_body": "
Closed VocabCharCNNPointer SentinelGSC (ours)
Seen reposAST AugAST0.30, 7.22, 0.94 0.26, 7.64, 0.940.28,8.67, 1.08 0.28, 7.46, 0.990.32, 8.00, 1.07 0.35, 6.51, 0.770.56, 3.87, 0.39 0.60, 3.68, 0.37
Unseen reposAST0.09,8.66,1.230.10, 8.82, 1.120.13, 9,39, 1.370.42, 4.81, 0.59
AugAST0.09, 8.34, 1.140.10, 8.16, 1.120.14, 8.03, 1.070.48, 4.28, 0.49
", + "bbox": [ + 173, + 228, + 838, + 316 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "DATASET INFORMATION ", + "bbox": [ + 174, + 343, + 343, + 356 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/3cdbb6558a551b95d720af6941b5a58667c04241856332d0fca4e455b35250a5.jpg", + "table_caption": [ + "Table 7: Repositories used in experiments. All were taken from the Maven repository (https://mvnrepository.com/). Entries are in the form “group/repository name/version”. " + ], + "table_footnote": [], + "table_body": "
Seen Repos
com.fasterxml.jackson.core/jackson-core/2.9.5 com.h2database/h2/1.4.195 javax.enterprise/cdi-api/2.0 junit/junit/4.12 mysql/mysql-connector-java/6.0.6 org.apache.commons/commons-collections4/4.1 org.apache.commons/commons-math3/3.6.1 org.apache.commons/commons-pool2/2.5.0 org.apache.maven/maven-project/2.2.1 org.codehaus.plexus/plexus-utils/3.1.0 org.eclipse.jetty/jetty-server/9.4.9.v20180320
org.scalacheck/scalacheck_2.12/1.13.5 org.slf4j/slf4j-api/1.7.25 org.slf4j/slf4j-log4j12/1.7.25 Unseen Repos
org.javassist/javassist/3.22.0-GA joda-time/joda-time/2.9.9 org.mockito/mockito-core/2.17.0
", + "bbox": [ + 328, + 411, + 666, + 694 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/3be36687fd308695ae5ec3a990a3322b5d2dc3db17b8b2b431ec39007d2d0335.jpg", + "table_caption": [ + "Table 8: Edge types used in Augmented ASTs. The initial AST is constructed using the AST and NEXT TOKEN edges, and then the remaining edges are added. In other words, the the “AST” model from Table 1 uses a graph that contains only the AST and NEXT TOKEN edge types (and WORD USE if it also uses a GSC), while the “AugAST” model contains all the edge types below. The reversed version of every edge is also added as its own type (e.g. reverse AST, reverse LAST READ) to let the GNN message passing occur in both directions. " + ], + "table_footnote": [], + "table_body": "
Edge NameDescription
AST NEXT_TOKENThe edges used to construct the original AST. Edges added to the original AST that specify the left-to-
right ordering of the children of a node in the AST.These edges are necessary since ASTs have ordered children, but we are representing the AST as a directed multigraph.
COMPUTED_FROMConnects a node representing a variable on the left of an equality to those on the right. (E.g. edges from y to x and Z to X in x = y + z.) The same as in Allamanis et al. (2018).
LAST_READConnects a node representing a usage of a variable to all nodes in the AST at which that variable's value could have been last read from memory. The same as in Allamanis et al. (2018).
LAST_WRITEConnects a node representing a usage of a variable to all nodesin the AST at which that variable's value could have been last written to memory. The same as in Allamanis et al.
RETURNS_TO(2018). Points a node in a return statement to the node contain- ing the return type of the method. (E.g. x in return X gets an edge pointing to int in public static int
LAST_SCOPE_USEgetx(x).) Connects a node representing a variable to the node repre- senting the last time this variable's name was used in the text of the code (i.e. capturing information about the text, not the control flow),but only within lexical scope. This edge exists to try and give the non-GSC models as much
LAST_FIELD_LEXlexical information as possible to make them as comparable with the GSC model. Connects a field access (e.g. this.whatever or Foo .whatever) node to the last use of this.whatever (or to the variable's initialization,if it's the first use).This is not lexical-scope aware (and,in fact,can't be in Java,in
FIELDgeneral). Pointseach node representing a field access(e.g. this.whatever) to the node where that field was de-
WORD_USEclared. Points cache nodes to nodes representing variables in which the vocab word was used in the variable's name.
", + "bbox": [ + 212, + 281, + 769, + 834 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/parse/train/SkNSehA9FQ/SkNSehA9FQ_middle.json b/parse/train/SkNSehA9FQ/SkNSehA9FQ_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..61401f506998ec2319af4542363b9ffa00b820fe --- /dev/null +++ b/parse/train/SkNSehA9FQ/SkNSehA9FQ_middle.json @@ -0,0 +1,33578 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "OPEN VOCABULARY LEARNING ON SOURCE CODE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 388, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 388, + 117 + ], + "score": 1.0, + "content": "WITH A GRAPH–STRUCTURED CACHE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 185, + 336, + 201 + ], + "spans": [ + { + "bbox": [ + 276, + 185, + 336, + 201 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 213, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 226 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 226 + ], + "score": 1.0, + "content": "Machine learning models that take computer program source code as input typi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 470, + 237 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 470, + 237 + ], + "score": 1.0, + "content": "cally use Natural Language Processing (NLP) techniques. However, a major chal-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "score": 1.0, + "content": "lenge is that code is written using an open, rapidly changing vocabulary due to,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 258 + ], + "score": 1.0, + "content": "e.g., the coinage of new variable and method names. Reasoning over such a vocab-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 270 + ], + "score": 1.0, + "content": "ulary is not something for which most NLP methods are designed. We introduce", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "score": 1.0, + "content": "a Graph–Structured Cache to address this problem; this cache contains a node", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 470, + 290 + ], + "score": 1.0, + "content": "for each new word the model encounters with edges connecting each word to its", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "score": 1.0, + "content": "occurrences in the code. We find that combining this graph–structured cache strat-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 314 + ], + "score": 1.0, + "content": "egy with recent Graph–Neural–Network–based models for supervised learning on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 312, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 323 + ], + "score": 1.0, + "content": "code improves the models’ performance on a code completion task and a variable", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 323, + 248, + 335 + ], + "score": 1.0, + "content": "naming task — with over", + "type": "text" + }, + { + "bbox": [ + 248, + 323, + 272, + 333 + ], + "score": 0.85, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "relative improvement on the latter — at the cost", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 333, + 320, + 346 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 320, + 346 + ], + "score": 1.0, + "content": "of a moderate increase in computation time.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 369, + 206, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 208, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 208, + 385 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "Computer program source code is an abundant and accessible form of data from which machine", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 504, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 504, + 418 + ], + "score": 1.0, + "content": "learning algorithms could learn to perform many useful software development tasks, including vari-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "able name suggestion, code completion, bug finding, security vulnerability identification, code qual-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 427, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 442 + ], + "score": 1.0, + "content": "ity assessment, or automated test generation. But despite the similarities between natural language", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "and source code, deep learning methods for Natural Language Processing (NLP) have not been", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 450, + 452, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 452, + 462 + ], + "score": 1.0, + "content": "straightforward to apply to learning problems on source code (Allamanis et al., 2017).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 336, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 338, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 338, + 480 + ], + "score": 1.0, + "content": "There are many reasons for this, but two central ones are:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "score": 1.0, + "content": "1. Code’s syntactic structure is unlike natural language. While code contains natural language", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 119, + 500, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 119, + 500, + 504, + 511 + ], + "score": 1.0, + "content": "words and phrases in order to be human–readable, code is not meant to be a read like a natural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 118, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 118, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "language text. Code is written in a rigid syntax with delimiters that may open and close dozens", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 118, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 118, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "of lines apart; it consists in great part of references to faraway lines and different files; and it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 119, + 533, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 119, + 533, + 479, + 545 + ], + "score": 1.0, + "content": "describes computations that proceed in an order often quite distinct from its written order.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 549, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 104, + 548, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 104, + 548, + 506, + 563 + ], + "score": 1.0, + "content": "2. Code is written using an open vocabulary. Natural language is mostly composed of words from a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 118, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 118, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "large but closed (a.k.a. fixed–size and unchanging) vocabulary. Standard NLP methods can thus", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 118, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 118, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "perform well by fixing a large vocabulary of words before training, and labeling the few words", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 118, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 118, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "they encounter outside this vocabulary as “unknown”. But in code every new variable, class, or", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 118, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 118, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "method declared requires a name, and this abundance of names leads to the use of many obscure", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 119, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 119, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "words: abbreviations, brand names, technical terms, etc.1 A model must be able to reason about", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 119, + 616, + 307, + 627 + ], + "spans": [ + { + "bbox": [ + 119, + 616, + 307, + 627 + ], + "score": 1.0, + "content": "these newly–coined words to understand code.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 636, + 503, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 376, + 649 + ], + "score": 1.0, + "content": "The second of these issues is significant. To give one indication:", + "type": "text" + }, + { + "bbox": [ + 376, + 637, + 396, + 647 + ], + "score": 0.86, + "content": "28 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "of variable names contain", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "out–of–vocabulary words in the test set we use in our experiments below. But more broadly, the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 658, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 673 + ], + "score": 1.0, + "content": "open vocabulary issue in code is an acute example of a fundamental challenge in machine learning:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 681, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 119, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "1We use the terminology that a name in source code is a sequence of words, split on CamelCase or", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 691, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 702 + ], + "score": 1.0, + "content": "snake case. E.g. the method name addItemToList is composed of the words add, item, to, and", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 702, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 133, + 712 + ], + "score": 1.0, + "content": "list.", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "We also use the term variable in its slightly broader sense to refer to any user–named language construct,", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 470, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 470, + 732 + ], + "score": 1.0, + "content": "including function parameter, method, class, and field names, in addition to declared variable names.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "OPEN VOCABULARY LEARNING ON SOURCE CODE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 388, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 388, + 117 + ], + "score": 1.0, + "content": "WITH A GRAPH–STRUCTURED CACHE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 185, + 336, + 201 + ], + "spans": [ + { + "bbox": [ + 276, + 185, + 336, + 201 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 213, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 226 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 226 + ], + "score": 1.0, + "content": "Machine learning models that take computer program source code as input typi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 470, + 237 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 470, + 237 + ], + "score": 1.0, + "content": "cally use Natural Language Processing (NLP) techniques. However, a major chal-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "score": 1.0, + "content": "lenge is that code is written using an open, rapidly changing vocabulary due to,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 258 + ], + "score": 1.0, + "content": "e.g., the coinage of new variable and method names. Reasoning over such a vocab-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 270 + ], + "score": 1.0, + "content": "ulary is not something for which most NLP methods are designed. We introduce", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 469, + 280 + ], + "score": 1.0, + "content": "a Graph–Structured Cache to address this problem; this cache contains a node", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 470, + 290 + ], + "score": 1.0, + "content": "for each new word the model encounters with edges connecting each word to its", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "score": 1.0, + "content": "occurrences in the code. We find that combining this graph–structured cache strat-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 314 + ], + "score": 1.0, + "content": "egy with recent Graph–Neural–Network–based models for supervised learning on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 312, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 323 + ], + "score": 1.0, + "content": "code improves the models’ performance on a code completion task and a variable", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 323, + 248, + 335 + ], + "score": 1.0, + "content": "naming task — with over", + "type": "text" + }, + { + "bbox": [ + 248, + 323, + 272, + 333 + ], + "score": 0.85, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "relative improvement on the latter — at the cost", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 333, + 320, + 346 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 320, + 346 + ], + "score": 1.0, + "content": "of a moderate increase in computation time.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5, + "bbox_fs": [ + 141, + 212, + 470, + 346 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 369, + 206, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 208, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 208, + 385 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "Computer program source code is an abundant and accessible form of data from which machine", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 504, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 504, + 418 + ], + "score": 1.0, + "content": "learning algorithms could learn to perform many useful software development tasks, including vari-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "able name suggestion, code completion, bug finding, security vulnerability identification, code qual-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 427, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 442 + ], + "score": 1.0, + "content": "ity assessment, or automated test generation. But despite the similarities between natural language", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "and source code, deep learning methods for Natural Language Processing (NLP) have not been", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 450, + 452, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 452, + 462 + ], + "score": 1.0, + "content": "straightforward to apply to learning problems on source code (Allamanis et al., 2017).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 395, + 505, + 462 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 336, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 338, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 338, + 480 + ], + "score": 1.0, + "content": "There are many reasons for this, but two central ones are:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 466, + 338, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "score": 1.0, + "content": "1. Code’s syntactic structure is unlike natural language. While code contains natural language", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 119, + 500, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 119, + 500, + 504, + 511 + ], + "score": 1.0, + "content": "words and phrases in order to be human–readable, code is not meant to be a read like a natural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 118, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 118, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "language text. Code is written in a rigid syntax with delimiters that may open and close dozens", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 118, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 118, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "of lines apart; it consists in great part of references to faraway lines and different files; and it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 119, + 533, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 119, + 533, + 479, + 545 + ], + "score": 1.0, + "content": "describes computations that proceed in an order often quite distinct from its written order.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 488, + 506, + 545 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 549, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 104, + 548, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 104, + 548, + 506, + 563 + ], + "score": 1.0, + "content": "2. Code is written using an open vocabulary. Natural language is mostly composed of words from a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 118, + 560, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 118, + 560, + 505, + 572 + ], + "score": 1.0, + "content": "large but closed (a.k.a. fixed–size and unchanging) vocabulary. Standard NLP methods can thus", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 118, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 118, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "perform well by fixing a large vocabulary of words before training, and labeling the few words", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 118, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 118, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "they encounter outside this vocabulary as “unknown”. But in code every new variable, class, or", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 118, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 118, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "method declared requires a name, and this abundance of names leads to the use of many obscure", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 119, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 119, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "words: abbreviations, brand names, technical terms, etc.1 A model must be able to reason about", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 119, + 616, + 307, + 627 + ], + "spans": [ + { + "bbox": [ + 119, + 616, + 307, + 627 + ], + "score": 1.0, + "content": "these newly–coined words to understand code.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 548, + 506, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 636, + 503, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 376, + 649 + ], + "score": 1.0, + "content": "The second of these issues is significant. To give one indication:", + "type": "text" + }, + { + "bbox": [ + 376, + 637, + 396, + 647 + ], + "score": 0.86, + "content": "28 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "of variable names contain", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "out–of–vocabulary words in the test set we use in our experiments below. But more broadly, the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 658, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 673 + ], + "score": 1.0, + "content": "open vocabulary issue in code is an acute example of a fundamental challenge in machine learning:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 637, + 505, + 673 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "score": 1.0, + "content": "how to build models that can reason over unbounded domains of entities, sometimes called “open–", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "set” learning. Despite this, the open vocabulary issue in source code has received relatively little", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 201, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 201, + 117 + ], + "score": 1.0, + "content": "attention in prior work.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "The first issue, in contrast, has been the focus of much prior work. A common strategy in these works", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "is to represent source code as an Abstract Syntax Tree (AST) rather than as linear text. Once in this", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "graph–structured format, code can be passed as input to models like Recursive Neural Networks or", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "Graph Neural Networks (GNNs) that can, in principle, exploit the relational structure of their inputs", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 430, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 430, + 178 + ], + "score": 1.0, + "content": "and avoid the difficulties of reading code in linear order (Allamanis et al., 2018).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "score": 1.0, + "content": "Our contribution: In this paper we extend such AST–based models for source code in order to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "address the open vocabulary issue. We do so by introducing a Graph–Structured Cache (GSC) to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "handle out–of–vocabulary words. The GSC represents vocabulary words as additional nodes in the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "AST as they are encountered and connects them with the edges to where they are used in the code.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 422, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 422, + 239 + ], + "score": 1.0, + "content": "We then process the AST+GSC with a GNN to produce outputs. See Figure 1.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "We empirically evaluated the utility of a Graph–Structured Cache on two tasks: a code completion", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "(a.k.a. fill–in–the–blank) task and a variable naming task. We found that using a GSC improved", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 350, + 277 + ], + "score": 1.0, + "content": "performance on both tasks at the cost of an approximately", + "type": "text" + }, + { + "bbox": [ + 350, + 265, + 370, + 276 + ], + "score": 0.87, + "content": "30 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "increase in training time. More", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "score": 1.0, + "content": "precisely: even when using hyperparameters optimized for the baseline model, adding a GSC to a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 302, + 298 + ], + "score": 1.0, + "content": "baseline model improved its accuracy by at least", + "type": "text" + }, + { + "bbox": [ + 302, + 287, + 317, + 297 + ], + "score": 0.85, + "content": "7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 286, + 452, + 298 + ], + "score": 1.0, + "content": "on the fill–in–the–blank task and", + "type": "text" + }, + { + "bbox": [ + 452, + 287, + 477, + 297 + ], + "score": 0.91, + "content": "103 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "on the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "variable naming task. We also report a number of ablation results in which we carefully demonstrate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 412, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 412, + 321 + ], + "score": 1.0, + "content": "the relative importance of each model component to a model’s performance.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 338, + 195, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 197, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 197, + 353 + ], + "score": 1.0, + "content": "2 PRIOR WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 109, + 364, + 257, + 375 + ], + "lines": [ + { + "bbox": [ + 107, + 364, + 258, + 376 + ], + "spans": [ + { + "bbox": [ + 107, + 364, + 258, + 376 + ], + "score": 1.0, + "content": "REPRESENTING CODE AS A GRAPH", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "Given their prominence in the study of programming languages, Abstract Syntax Trees (ASTs)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "and parse trees are a natural choice for representing code and have been used extensively. Often", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "models that operate on source code consume ASTs by linearizing them (usually with a depth–first", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "traversal) as in Amodio et al. (2017), Liu et al. (2017), or Li et al. (2017) or by using AST paths", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "as input features as in Alon et al. (2018), but they can also be processed by deep learning models", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "that take graphs as input, as in White et al. (2016) and Chen et al. (2018) who use Recursive Neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "Networks (RveNNs) (Goller & Kuchler, 1996) on ASTs. RveNNs are models that operate on tree–", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "topology graphs, and have been used extensively for language modeling (Socher et al., 2013) and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "on domains similar to source code, like mathematical expressions (Zaremba et al., 2014; Arabshahi", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "et al., 2018). They can be considered a special case of Message Passing Neural Networks (MPNNs)", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "in the framework of Gilmer et al. (2017): in this analogy RveNNs are to Belief Propagation as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "MPNNs are to Loopy Belief Propagation. They can also be considered a special case of Graph", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Networks in the framework of Battaglia et al. (2018). ASTs also serve as a natural basis for models", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "that generate code as output, as in Maddison & Tarlow (2014), Yin & Neubig (2017), Rabinovich", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 539, + 365, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 365, + 551 + ], + "score": 1.0, + "content": "et al. (2017), Chen et al. (2018), and Brockschmidt et al. (2018).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 556, + 502, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 556, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 504, + 568 + ], + "score": 1.0, + "content": "Data–flow graphs are another type of graphical representation of source code with a long his-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "tory (Krinke, 2001), and they have occasionally been used to featurize source code for machine", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 578, + 221, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 221, + 590 + ], + "score": 1.0, + "content": "learning (Chae et al., 2017).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "Most closely related to our work is the work of Allamanis et al. (2018), on which our model is heavily", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "based. Allamanis et al. (2018) combine the data–flow graph and AST representation strategies for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "source code by representing code as an AST augmented with extra labeled edges indicating semantic", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "information like data– and control–flow between variables. These augmentations yield a directed", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "multigraph rather than just a tree,2 so in Allamanis et al. (2018) a variety of MPNN called a Gated", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 649, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 663 + ], + "score": 1.0, + "content": "Graph Neural Network (GGNN) (Li et al., 2016) is used to consume the Augmented AST and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 661, + 303, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 303, + 673 + ], + "score": 1.0, + "content": "produce an output for a supervised learning task.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "Graph-based models that are not based on ASTs are also sometimes used for analyzing source code,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 476, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 476, + 701 + ], + "score": 1.0, + "content": "like Conditional Random Fields for joint variable name prediction in (Raychev et al., 2015).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 711, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 117, + 708, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 506, + 725 + ], + "score": 1.0, + "content": "2This multigraph was referred to as a Program Graph in Allamanis et al. (2017) and is called an Augmented", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 719, + 154, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 154, + 733 + ], + "score": 1.0, + "content": "AST herein.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "score": 1.0, + "content": "how to build models that can reason over unbounded domains of entities, sometimes called “open–", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "set” learning. Despite this, the open vocabulary issue in source code has received relatively little", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 201, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 201, + 117 + ], + "score": 1.0, + "content": "attention in prior work.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "The first issue, in contrast, has been the focus of much prior work. A common strategy in these works", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "is to represent source code as an Abstract Syntax Tree (AST) rather than as linear text. Once in this", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "graph–structured format, code can be passed as input to models like Recursive Neural Networks or", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "Graph Neural Networks (GNNs) that can, in principle, exploit the relational structure of their inputs", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 430, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 430, + 178 + ], + "score": 1.0, + "content": "and avoid the difficulties of reading code in linear order (Allamanis et al., 2018).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 121, + 506, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "score": 1.0, + "content": "Our contribution: In this paper we extend such AST–based models for source code in order to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "address the open vocabulary issue. We do so by introducing a Graph–Structured Cache (GSC) to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "handle out–of–vocabulary words. The GSC represents vocabulary words as additional nodes in the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "AST as they are encountered and connects them with the edges to where they are used in the code.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 422, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 422, + 239 + ], + "score": 1.0, + "content": "We then process the AST+GSC with a GNN to produce outputs. See Figure 1.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 181, + 506, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "We empirically evaluated the utility of a Graph–Structured Cache on two tasks: a code completion", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "(a.k.a. fill–in–the–blank) task and a variable naming task. We found that using a GSC improved", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 350, + 277 + ], + "score": 1.0, + "content": "performance on both tasks at the cost of an approximately", + "type": "text" + }, + { + "bbox": [ + 350, + 265, + 370, + 276 + ], + "score": 0.87, + "content": "30 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "increase in training time. More", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 288 + ], + "score": 1.0, + "content": "precisely: even when using hyperparameters optimized for the baseline model, adding a GSC to a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 302, + 298 + ], + "score": 1.0, + "content": "baseline model improved its accuracy by at least", + "type": "text" + }, + { + "bbox": [ + 302, + 287, + 317, + 297 + ], + "score": 0.85, + "content": "7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 286, + 452, + 298 + ], + "score": 1.0, + "content": "on the fill–in–the–blank task and", + "type": "text" + }, + { + "bbox": [ + 452, + 287, + 477, + 297 + ], + "score": 0.91, + "content": "103 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "on the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "variable naming task. We also report a number of ablation results in which we carefully demonstrate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 412, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 412, + 321 + ], + "score": 1.0, + "content": "the relative importance of each model component to a model’s performance.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 243, + 506, + 321 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 338, + 195, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 197, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 197, + 353 + ], + "score": 1.0, + "content": "2 PRIOR WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 109, + 364, + 257, + 375 + ], + "lines": [ + { + "bbox": [ + 107, + 364, + 258, + 376 + ], + "spans": [ + { + "bbox": [ + 107, + 364, + 258, + 376 + ], + "score": 1.0, + "content": "REPRESENTING CODE AS A GRAPH", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "Given their prominence in the study of programming languages, Abstract Syntax Trees (ASTs)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "and parse trees are a natural choice for representing code and have been used extensively. Often", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "models that operate on source code consume ASTs by linearizing them (usually with a depth–first", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "traversal) as in Amodio et al. (2017), Liu et al. (2017), or Li et al. (2017) or by using AST paths", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "as input features as in Alon et al. (2018), but they can also be processed by deep learning models", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "that take graphs as input, as in White et al. (2016) and Chen et al. (2018) who use Recursive Neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "Networks (RveNNs) (Goller & Kuchler, 1996) on ASTs. RveNNs are models that operate on tree–", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "topology graphs, and have been used extensively for language modeling (Socher et al., 2013) and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "on domains similar to source code, like mathematical expressions (Zaremba et al., 2014; Arabshahi", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "et al., 2018). They can be considered a special case of Message Passing Neural Networks (MPNNs)", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "in the framework of Gilmer et al. (2017): in this analogy RveNNs are to Belief Propagation as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "MPNNs are to Loopy Belief Propagation. They can also be considered a special case of Graph", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "Networks in the framework of Battaglia et al. (2018). ASTs also serve as a natural basis for models", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "that generate code as output, as in Maddison & Tarlow (2014), Yin & Neubig (2017), Rabinovich", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 539, + 365, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 365, + 551 + ], + "score": 1.0, + "content": "et al. (2017), Chen et al. (2018), and Brockschmidt et al. (2018).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29, + "bbox_fs": [ + 104, + 385, + 506, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 556, + 502, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 556, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 504, + 568 + ], + "score": 1.0, + "content": "Data–flow graphs are another type of graphical representation of source code with a long his-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "tory (Krinke, 2001), and they have occasionally been used to featurize source code for machine", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 578, + 221, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 221, + 590 + ], + "score": 1.0, + "content": "learning (Chae et al., 2017).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 556, + 505, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "Most closely related to our work is the work of Allamanis et al. (2018), on which our model is heavily", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "based. Allamanis et al. (2018) combine the data–flow graph and AST representation strategies for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "source code by representing code as an AST augmented with extra labeled edges indicating semantic", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "information like data– and control–flow between variables. These augmentations yield a directed", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "multigraph rather than just a tree,2 so in Allamanis et al. (2018) a variety of MPNN called a Gated", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 649, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 663 + ], + "score": 1.0, + "content": "Graph Neural Network (GGNN) (Li et al., 2016) is used to consume the Augmented AST and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 661, + 303, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 303, + 673 + ], + "score": 1.0, + "content": "produce an output for a supervised learning task.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 594, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "Graph-based models that are not based on ASTs are also sometimes used for analyzing source code,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 476, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 476, + 701 + ], + "score": 1.0, + "content": "like Conditional Random Fields for joint variable name prediction in (Raychev et al., 2015).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 676, + 505, + 701 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 243, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 244, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 244, + 94 + ], + "score": 1.0, + "content": "REASONING ABOUT OPEN SETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 103, + 505, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 117 + ], + "score": 1.0, + "content": "The question of how to gracefully reason over an open vocabulary is longstanding in NLP.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "Character–level embeddings are a typical way deep learning models handle this issue, whether used", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "on their own as in Kim et al. (2016), or in conjunction with word–level embedding Recurrent Neu-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 442, + 148 + ], + "score": 1.0, + "content": "ral Networks (RNNs) as in Luong & Manning (2016), or in conjunction with an", + "type": "text" + }, + { + "bbox": [ + 443, + 138, + 450, + 146 + ], + "score": 0.7, + "content": "n _ { \\mathrm { ~ } }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 136, + 505, + 148 + ], + "score": 1.0, + "content": "–gram model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "score": 1.0, + "content": "as in Bojanowski et al. (2017). Another approach is to learn new word embeddings on–the–fly", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "score": 1.0, + "content": "from context (Kobayashi et al., 2016). Caching novel words, as we do in our model, is yet another", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 168, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 182 + ], + "score": 1.0, + "content": "strategy (Grave et al., 2017) and has been used to augment N–gram models for analyzing source", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 180, + 262, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 262, + 192 + ], + "score": 1.0, + "content": "code (Hellendoorn & Devanbu, 2017).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 196, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "In terms of producing outputs over variable–sized input and outputs, also known as open–set learn-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "ing, attention-based pointer mechanisms were introduced in Vinyals et al. (2015) and have been used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "for tasks on code, e.g. in Bhoopchand et al. (2016). Such methods have been used to great effect", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "in NLP in e.g. Gulcehre et al. (2016) and Merity et al. (2017). The latter’s pointer sentinel mixture", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 497, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 497, + 254 + ], + "score": 1.0, + "content": "model is the direct inspiration for the readout function we use in the Variable Naming task below.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Using graphs to represent arbitrary collections of entities and their relationships for processing by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 268, + 504, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 504, + 280 + ], + "score": 1.0, + "content": "deep networks has been widely used (Johnson, 2017; Bansal et al., 2017; Pham et al., 2018; Lu et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "2017), but to our knowledge we are the first to use a graph–building strategy for reasoning (at train", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 308, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 308, + 302 + ], + "score": 1.0, + "content": "and test time) about an open vocabulary of words.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 317, + 208, + 329 + ], + "lines": [ + { + "bbox": [ + 104, + 315, + 209, + 332 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 209, + 332 + ], + "score": 1.0, + "content": "3 PRELIMINARIES", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 247, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 248, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 248, + 354 + ], + "score": 1.0, + "content": "3.1 ABSTRACT SYNTAX TREES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "An Abstract Syntax Tree (AST) is a graph — specifically an ordered tree with labeled nodes — that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "is a representation of some written computer source code. There is a 1–to–1 relationship between", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "source code and an AST of that source code, modulo comments and whitespace in the written source", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 131, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 131, + 406 + ], + "score": 1.0, + "content": "code.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "Typically the leaves of an AST correspond to the tokens written in the source code, like variable and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "method names, while the non–leaf nodes represent syntactic language constructs like function calls", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "or class definitions. The specific node labels and construction rules of ASTs can differ between or", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 445, + 358, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 358, + 457 + ], + "score": 1.0, + "content": "within languages. The first step in Figure 1 shows an example.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 468, + 254, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 255, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 255, + 482 + ], + "score": 1.0, + "content": "3.2 GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "The term Graph Neural Network (GNN) refers to any deep, differentiable model that takes graphs", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "as input. Many GNNs have been presented in the literature, and several nomenclatures have", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "been proposed for describing the computations they perform, in particular in Gilmer et al. (2017)", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "score": 1.0, + "content": "and Battaglia et al. (2018). Here we give a brief recapitulation of supervised learning with GNNs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 532, + 433, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 433, + 546 + ], + "score": 1.0, + "content": "using the Message Passing Neural Network framework from Gilmer et al. (2017).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 549, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 223, + 563 + ], + "score": 1.0, + "content": "A GNN is trained using pairs", + "type": "text" + }, + { + "bbox": [ + 223, + 550, + 249, + 562 + ], + "score": 0.91, + "content": "( G , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 550, + 276, + 563 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 277, + 550, + 325, + 562 + ], + "score": 0.92, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 550, + 453, + 563 + ], + "score": 1.0, + "content": "is a graph defined by its vertices", + "type": "text" + }, + { + "bbox": [ + 453, + 550, + 463, + 560 + ], + "score": 0.65, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "and edges", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 561, + 115, + 571 + ], + "score": 0.78, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 561, + 136, + 573 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 137, + 563, + 144, + 573 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 561, + 187, + 573 + ], + "score": 1.0, + "content": "is a label.", + "type": "text" + }, + { + "bbox": [ + 187, + 563, + 194, + 573 + ], + "score": 0.73, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "can be any sort of mathematical object: scalar, vector, another graph, etc. In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "the most general case, each graph in the dataset can be a directed multigraph, each with a different", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 385, + 596 + ], + "score": 1.0, + "content": "number of nodes and different connectivity. In each graph, each vertex", + "type": "text" + }, + { + "bbox": [ + 385, + 583, + 412, + 593 + ], + "score": 0.92, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "has associated features", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 592, + 301, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 119, + 605 + ], + "score": 0.84, + "content": "\\mathbf { \\boldsymbol { x } } _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 592, + 182, + 608 + ], + "score": 1.0, + "content": ", and each edge", + "type": "text" + }, + { + "bbox": [ + 182, + 594, + 228, + 606 + ], + "score": 0.93, + "content": "( v , w ) \\in E", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 592, + 279, + 608 + ], + "score": 1.0, + "content": "has features", + "type": "text" + }, + { + "bbox": [ + 279, + 596, + 296, + 605 + ], + "score": 0.88, + "content": "e _ { v w }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 592, + 301, + 608 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 500, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 501, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 228, + 625 + ], + "score": 1.0, + "content": "A GNN produces a prediction", + "type": "text" + }, + { + "bbox": [ + 228, + 611, + 235, + 622 + ], + "score": 0.85, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 609, + 286, + 625 + ], + "score": 1.0, + "content": "for the label", + "type": "text" + }, + { + "bbox": [ + 287, + 613, + 294, + 622 + ], + "score": 0.82, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 609, + 337, + 625 + ], + "score": 1.0, + "content": "of a graph", + "type": "text" + }, + { + "bbox": [ + 338, + 610, + 386, + 623 + ], + "score": 0.93, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 609, + 501, + 625 + ], + "score": 1.0, + "content": "by the following procedure:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 112, + 630, + 502, + 653 + ], + "lines": [ + { + "bbox": [ + 114, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 114, + 630, + 172, + 642 + ], + "score": 1.0, + "content": "1. A function", + "type": "text" + }, + { + "bbox": [ + 172, + 631, + 180, + 640 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 630, + 345, + 642 + ], + "score": 1.0, + "content": "is used to initialize a hidden state vector", + "type": "text" + }, + { + "bbox": [ + 345, + 630, + 358, + 642 + ], + "score": 0.9, + "content": "h _ { v } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 630, + 422, + 642 + ], + "score": 1.0, + "content": "for each vertex", + "type": "text" + }, + { + "bbox": [ + 422, + 631, + 450, + 641 + ], + "score": 0.91, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "as a function", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 641, + 487, + 654 + ], + "spans": [ + { + "bbox": [ + 126, + 641, + 265, + 654 + ], + "score": 1.0, + "content": "of the vertex’s features (e.g., if the", + "type": "text" + }, + { + "bbox": [ + 266, + 643, + 278, + 653 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 641, + 323, + 654 + ], + "score": 1.0, + "content": "are words,", + "type": "text" + }, + { + "bbox": [ + 323, + 642, + 331, + 651 + ], + "score": 0.84, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 641, + 487, + 654 + ], + "score": 1.0, + "content": "could be a word embedding function):", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "interline_equation", + "bbox": [ + 289, + 653, + 342, + 667 + ], + "lines": [ + { + "bbox": [ + 289, + 653, + 342, + 667 + ], + "spans": [ + { + "bbox": [ + 289, + 653, + 342, + 667 + ], + "score": 0.91, + "content": "\\begin{array} { r } { h _ { v } ^ { 0 } = S ( \\pmb { x } _ { v } ) } \\end{array}", + "type": "interline_equation", + "image_path": "94e458fc91a8fca3fbeee77f4db187895e70db6ba937f66e287ccb86e7029472.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 289, + 653, + 342, + 667 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 114, + 670, + 282, + 681 + ], + "lines": [ + { + "bbox": [ + 113, + 669, + 283, + 682 + ], + "spans": [ + { + "bbox": [ + 113, + 669, + 189, + 682 + ], + "score": 1.0, + "content": "2. For each round", + "type": "text" + }, + { + "bbox": [ + 189, + 671, + 194, + 680 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 669, + 221, + 682 + ], + "score": 1.0, + "content": "out of", + "type": "text" + }, + { + "bbox": [ + 221, + 671, + 230, + 680 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 669, + 283, + 682 + ], + "score": 1.0, + "content": "total rounds:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 129, + 684, + 505, + 706 + ], + "lines": [ + { + "bbox": [ + 126, + 680, + 508, + 700 + ], + "spans": [ + { + "bbox": [ + 126, + 680, + 197, + 700 + ], + "score": 1.0, + "content": "(a) Each vertex", + "type": "text" + }, + { + "bbox": [ + 197, + 684, + 226, + 694 + ], + "score": 0.9, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 680, + 307, + 700 + ], + "score": 1.0, + "content": "receives the vector", + "type": "text" + }, + { + "bbox": [ + 307, + 684, + 332, + 696 + ], + "score": 0.92, + "content": "\\boldsymbol { m } _ { v } ^ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 680, + 508, + 700 + ], + "score": 1.0, + "content": ", which is the sum of “messages” from its", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 146, + 695, + 325, + 708 + ], + "spans": [ + { + "bbox": [ + 146, + 695, + 306, + 708 + ], + "score": 1.0, + "content": "neighbors, each produced by a function", + "type": "text" + }, + { + "bbox": [ + 306, + 696, + 320, + 706 + ], + "score": 0.89, + "content": "M _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 695, + 325, + 708 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 708, + 408, + 734 + ], + "lines": [ + { + "bbox": [ + 242, + 708, + 408, + 734 + ], + "spans": [ + { + "bbox": [ + 242, + 708, + 408, + 734 + ], + "score": 0.94, + "content": "m _ { v } ^ { t + 1 } = \\sum _ { w \\in \\mathrm { n e i g h b o r s ~ o f } v } M _ { t } ( h _ { v } ^ { t } , h _ { w } ^ { t } , e _ { v w } ) .", + "type": "interline_equation", + "image_path": "b7a1c5f247ce69866b2fcc30405e0a9b6fd08d99af662eb1e4578dfa11bca720.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 242, + 708, + 408, + 734 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 243, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 244, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 244, + 94 + ], + "score": 1.0, + "content": "REASONING ABOUT OPEN SETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 103, + 505, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 117 + ], + "score": 1.0, + "content": "The question of how to gracefully reason over an open vocabulary is longstanding in NLP.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "Character–level embeddings are a typical way deep learning models handle this issue, whether used", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "on their own as in Kim et al. (2016), or in conjunction with word–level embedding Recurrent Neu-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 442, + 148 + ], + "score": 1.0, + "content": "ral Networks (RNNs) as in Luong & Manning (2016), or in conjunction with an", + "type": "text" + }, + { + "bbox": [ + 443, + 138, + 450, + 146 + ], + "score": 0.7, + "content": "n _ { \\mathrm { ~ } }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 136, + 505, + 148 + ], + "score": 1.0, + "content": "–gram model", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "score": 1.0, + "content": "as in Bojanowski et al. (2017). Another approach is to learn new word embeddings on–the–fly", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "score": 1.0, + "content": "from context (Kobayashi et al., 2016). Caching novel words, as we do in our model, is yet another", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 168, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 182 + ], + "score": 1.0, + "content": "strategy (Grave et al., 2017) and has been used to augment N–gram models for analyzing source", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 180, + 262, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 262, + 192 + ], + "score": 1.0, + "content": "code (Hellendoorn & Devanbu, 2017).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 102, + 506, + 192 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 196, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "In terms of producing outputs over variable–sized input and outputs, also known as open–set learn-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "ing, attention-based pointer mechanisms were introduced in Vinyals et al. (2015) and have been used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "for tasks on code, e.g. in Bhoopchand et al. (2016). Such methods have been used to great effect", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "in NLP in e.g. Gulcehre et al. (2016) and Merity et al. (2017). The latter’s pointer sentinel mixture", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 497, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 497, + 254 + ], + "score": 1.0, + "content": "model is the direct inspiration for the readout function we use in the Variable Naming task below.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 196, + 505, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Using graphs to represent arbitrary collections of entities and their relationships for processing by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 268, + 504, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 504, + 280 + ], + "score": 1.0, + "content": "deep networks has been widely used (Johnson, 2017; Bansal et al., 2017; Pham et al., 2018; Lu et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "2017), but to our knowledge we are the first to use a graph–building strategy for reasoning (at train", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 308, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 308, + 302 + ], + "score": 1.0, + "content": "and test time) about an open vocabulary of words.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 257, + 505, + 302 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 317, + 208, + 329 + ], + "lines": [ + { + "bbox": [ + 104, + 315, + 209, + 332 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 209, + 332 + ], + "score": 1.0, + "content": "3 PRELIMINARIES", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 247, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 248, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 248, + 354 + ], + "score": 1.0, + "content": "3.1 ABSTRACT SYNTAX TREES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "An Abstract Syntax Tree (AST) is a graph — specifically an ordered tree with labeled nodes — that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "is a representation of some written computer source code. There is a 1–to–1 relationship between", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "source code and an AST of that source code, modulo comments and whitespace in the written source", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 131, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 131, + 406 + ], + "score": 1.0, + "content": "code.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 361, + 506, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "Typically the leaves of an AST correspond to the tokens written in the source code, like variable and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "method names, while the non–leaf nodes represent syntactic language constructs like function calls", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "or class definitions. The specific node labels and construction rules of ASTs can differ between or", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 445, + 358, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 358, + 457 + ], + "score": 1.0, + "content": "within languages. The first step in Figure 1 shows an example.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 411, + 505, + 457 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 468, + 254, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 255, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 255, + 482 + ], + "score": 1.0, + "content": "3.2 GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "The term Graph Neural Network (GNN) refers to any deep, differentiable model that takes graphs", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "as input. Many GNNs have been presented in the literature, and several nomenclatures have", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "been proposed for describing the computations they perform, in particular in Gilmer et al. (2017)", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "score": 1.0, + "content": "and Battaglia et al. (2018). Here we give a brief recapitulation of supervised learning with GNNs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 532, + 433, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 433, + 546 + ], + "score": 1.0, + "content": "using the Message Passing Neural Network framework from Gilmer et al. (2017).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 489, + 506, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 549, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 223, + 563 + ], + "score": 1.0, + "content": "A GNN is trained using pairs", + "type": "text" + }, + { + "bbox": [ + 223, + 550, + 249, + 562 + ], + "score": 0.91, + "content": "( G , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 550, + 276, + 563 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 277, + 550, + 325, + 562 + ], + "score": 0.92, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 550, + 453, + 563 + ], + "score": 1.0, + "content": "is a graph defined by its vertices", + "type": "text" + }, + { + "bbox": [ + 453, + 550, + 463, + 560 + ], + "score": 0.65, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "and edges", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 561, + 115, + 571 + ], + "score": 0.78, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 561, + 136, + 573 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 137, + 563, + 144, + 573 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 561, + 187, + 573 + ], + "score": 1.0, + "content": "is a label.", + "type": "text" + }, + { + "bbox": [ + 187, + 563, + 194, + 573 + ], + "score": 0.73, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "can be any sort of mathematical object: scalar, vector, another graph, etc. In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "the most general case, each graph in the dataset can be a directed multigraph, each with a different", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 385, + 596 + ], + "score": 1.0, + "content": "number of nodes and different connectivity. In each graph, each vertex", + "type": "text" + }, + { + "bbox": [ + 385, + 583, + 412, + 593 + ], + "score": 0.92, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "has associated features", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 592, + 301, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 119, + 605 + ], + "score": 0.84, + "content": "\\mathbf { \\boldsymbol { x } } _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 592, + 182, + 608 + ], + "score": 1.0, + "content": ", and each edge", + "type": "text" + }, + { + "bbox": [ + 182, + 594, + 228, + 606 + ], + "score": 0.93, + "content": "( v , w ) \\in E", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 592, + 279, + 608 + ], + "score": 1.0, + "content": "has features", + "type": "text" + }, + { + "bbox": [ + 279, + 596, + 296, + 605 + ], + "score": 0.88, + "content": "e _ { v w }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 592, + 301, + 608 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 550, + 505, + 608 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 500, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 501, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 228, + 625 + ], + "score": 1.0, + "content": "A GNN produces a prediction", + "type": "text" + }, + { + "bbox": [ + 228, + 611, + 235, + 622 + ], + "score": 0.85, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 609, + 286, + 625 + ], + "score": 1.0, + "content": "for the label", + "type": "text" + }, + { + "bbox": [ + 287, + 613, + 294, + 622 + ], + "score": 0.82, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 609, + 337, + 625 + ], + "score": 1.0, + "content": "of a graph", + "type": "text" + }, + { + "bbox": [ + 338, + 610, + 386, + 623 + ], + "score": 0.93, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 609, + 501, + 625 + ], + "score": 1.0, + "content": "by the following procedure:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 609, + 501, + 625 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 630, + 502, + 653 + ], + "lines": [ + { + "bbox": [ + 114, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 114, + 630, + 172, + 642 + ], + "score": 1.0, + "content": "1. A function", + "type": "text" + }, + { + "bbox": [ + 172, + 631, + 180, + 640 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 630, + 345, + 642 + ], + "score": 1.0, + "content": "is used to initialize a hidden state vector", + "type": "text" + }, + { + "bbox": [ + 345, + 630, + 358, + 642 + ], + "score": 0.9, + "content": "h _ { v } ^ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 630, + 422, + 642 + ], + "score": 1.0, + "content": "for each vertex", + "type": "text" + }, + { + "bbox": [ + 422, + 631, + 450, + 641 + ], + "score": 0.91, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "as a function", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 641, + 487, + 654 + ], + "spans": [ + { + "bbox": [ + 126, + 641, + 265, + 654 + ], + "score": 1.0, + "content": "of the vertex’s features (e.g., if the", + "type": "text" + }, + { + "bbox": [ + 266, + 643, + 278, + 653 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 641, + 323, + 654 + ], + "score": 1.0, + "content": "are words,", + "type": "text" + }, + { + "bbox": [ + 323, + 642, + 331, + 651 + ], + "score": 0.84, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 641, + 487, + 654 + ], + "score": 1.0, + "content": "could be a word embedding function):", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 114, + 630, + 505, + 654 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 289, + 653, + 342, + 667 + ], + "lines": [ + { + "bbox": [ + 289, + 653, + 342, + 667 + ], + "spans": [ + { + "bbox": [ + 289, + 653, + 342, + 667 + ], + "score": 0.91, + "content": "\\begin{array} { r } { h _ { v } ^ { 0 } = S ( \\pmb { x } _ { v } ) } \\end{array}", + "type": "interline_equation", + "image_path": "94e458fc91a8fca3fbeee77f4db187895e70db6ba937f66e287ccb86e7029472.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 289, + 653, + 342, + 667 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 114, + 670, + 282, + 681 + ], + "lines": [ + { + "bbox": [ + 113, + 669, + 283, + 682 + ], + "spans": [ + { + "bbox": [ + 113, + 669, + 189, + 682 + ], + "score": 1.0, + "content": "2. For each round", + "type": "text" + }, + { + "bbox": [ + 189, + 671, + 194, + 680 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 669, + 221, + 682 + ], + "score": 1.0, + "content": "out of", + "type": "text" + }, + { + "bbox": [ + 221, + 671, + 230, + 680 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 669, + 283, + 682 + ], + "score": 1.0, + "content": "total rounds:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43, + "bbox_fs": [ + 113, + 669, + 283, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 684, + 505, + 706 + ], + "lines": [ + { + "bbox": [ + 126, + 680, + 508, + 700 + ], + "spans": [ + { + "bbox": [ + 126, + 680, + 197, + 700 + ], + "score": 1.0, + "content": "(a) Each vertex", + "type": "text" + }, + { + "bbox": [ + 197, + 684, + 226, + 694 + ], + "score": 0.9, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 680, + 307, + 700 + ], + "score": 1.0, + "content": "receives the vector", + "type": "text" + }, + { + "bbox": [ + 307, + 684, + 332, + 696 + ], + "score": 0.92, + "content": "\\boldsymbol { m } _ { v } ^ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 680, + 508, + 700 + ], + "score": 1.0, + "content": ", which is the sum of “messages” from its", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 146, + 695, + 325, + 708 + ], + "spans": [ + { + "bbox": [ + 146, + 695, + 306, + 708 + ], + "score": 1.0, + "content": "neighbors, each produced by a function", + "type": "text" + }, + { + "bbox": [ + 306, + 696, + 320, + 706 + ], + "score": 0.89, + "content": "M _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 695, + 325, + 708 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 126, + 680, + 508, + 708 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 708, + 408, + 734 + ], + "lines": [ + { + "bbox": [ + 242, + 708, + 408, + 734 + ], + "spans": [ + { + "bbox": [ + 242, + 708, + 408, + 734 + ], + "score": 0.94, + "content": "m _ { v } ^ { t + 1 } = \\sum _ { w \\in \\mathrm { n e i g h b o r s ~ o f } v } M _ { t } ( h _ { v } ^ { t } , h _ { w } ^ { t } , e _ { v w } ) .", + "type": "interline_equation", + "image_path": "b7a1c5f247ce69866b2fcc30405e0a9b6fd08d99af662eb1e4578dfa11bca720.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 242, + 708, + 408, + 734 + ], + "spans": [], + "index": 46 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 126, + 82, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 129, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 129, + 81, + 195, + 96 + ], + "score": 1.0, + "content": "(b) Each vertex", + "type": "text" + }, + { + "bbox": [ + 195, + 83, + 222, + 93 + ], + "score": 0.9, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "updates its hidden state based on the message it received via a function", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 92, + 164, + 108 + ], + "spans": [ + { + "bbox": [ + 146, + 94, + 158, + 105 + ], + "score": 0.85, + "content": "U _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 92, + 164, + 108 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 277, + 104, + 374, + 119 + ], + "lines": [ + { + "bbox": [ + 277, + 104, + 374, + 119 + ], + "spans": [ + { + "bbox": [ + 277, + 104, + 374, + 119 + ], + "score": 0.93, + "content": "h _ { v } ^ { t + 1 } = U _ { t } ( h _ { v } ^ { t } , m _ { v } ^ { t + 1 } ) .", + "type": "interline_equation", + "image_path": "3dbc90bc50fd06202e01f73e8218ed391fc7b8275fcd2ebea9880febecb6a67a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 277, + 104, + 374, + 119 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 111, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 113, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 113, + 122, + 172, + 136 + ], + "score": 1.0, + "content": "3. A function", + "type": "text" + }, + { + "bbox": [ + 173, + 124, + 181, + 134 + ], + "score": 0.81, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 122, + 505, + 136 + ], + "score": 1.0, + "content": ", the “readout function”, produces a prediction based on the hidden states gener-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 135, + 399, + 147 + ], + "spans": [ + { + "bbox": [ + 126, + 135, + 384, + 147 + ], + "score": 1.0, + "content": "ated during the message passing (usually just those at from time", + "type": "text" + }, + { + "bbox": [ + 384, + 135, + 392, + 145 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 135, + 399, + 147 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 151, + 384, + 167 + ], + "lines": [ + { + "bbox": [ + 247, + 151, + 384, + 167 + ], + "spans": [ + { + "bbox": [ + 247, + 151, + 384, + 167 + ], + "score": 0.91, + "content": "\\hat { y } = R ( \\{ h _ { v } ^ { t } | v \\in V , t \\in 1 , . . . , T \\} ) .", + "type": "interline_equation", + "image_path": "4a007b03cf24477ac91e4a1b0db3625e7401553f5af46fc8f711e359018c39a2.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 247, + 151, + 384, + 167 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 176, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 252, + 188 + ], + "score": 1.0, + "content": "GNNs differ in how they implement", + "type": "text" + }, + { + "bbox": [ + 253, + 177, + 293, + 187 + ], + "score": 0.27, + "content": "S , M _ { t } , U _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 176, + 313, + 188 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 314, + 177, + 322, + 186 + ], + "score": 0.81, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 176, + 505, + 188 + ], + "score": 1.0, + "content": ". But all these functions are differentiable and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "most are parameterized, so the model is trainable via stochastic gradient descent of a loss function", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 154, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 119, + 210 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 200, + 126, + 210 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 198, + 144, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 199, + 150, + 210 + ], + "score": 0.83, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 198, + 154, + 210 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 226, + 165, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 167, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 167, + 241 + ], + "score": 1.0, + "content": "4 MODEL", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 253, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "score": 1.0, + "content": "Our model consumes an input instance of source code and produces an output for a supervised", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 263, + 359, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 359, + 276 + ], + "score": 1.0, + "content": "learning task via the following five steps, sketched in Figure 1:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 113, + 284, + 506, + 546 + ], + "lines": [ + { + "bbox": [ + 113, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 113, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "1. Parse the source code (snippet, file, repository, version control history, etc.) into an Abstract", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 296, + 179, + 307 + ], + "spans": [ + { + "bbox": [ + 126, + 296, + 179, + 307 + ], + "score": 1.0, + "content": "Syntax Tree.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 111, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 111, + 310, + 506, + 325 + ], + "score": 1.0, + "content": "2. Add edges of varying types (details in Appendix Table 8) to this AST that represent semantic", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 126, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "information like data– and control– flow, in the spirit of Allamanis et al. (2018). Also add the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 125, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "reversed version of all edges with their own edge type. This results in a directed multigraph", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 345, + 236, + 357 + ], + "spans": [ + { + "bbox": [ + 126, + 345, + 236, + 357 + ], + "score": 1.0, + "content": "called an Augmented AST.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 113, + 361, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 113, + 361, + 506, + 373 + ], + "score": 1.0, + "content": "3. Further augment the Augmented AST by adding a Graph–Structured Cache. That is, add a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 372, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 126, + 372, + 505, + 383 + ], + "score": 1.0, + "content": "node to the Augmented AST for each vocabulary word encountered in the input instance. Then", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 383, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 126, + 383, + 505, + 395 + ], + "score": 1.0, + "content": "connect each such “cache node” with an edge (of edge type WORD USE) to all variables whose", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 394, + 223, + 405 + ], + "spans": [ + { + "bbox": [ + 125, + 394, + 223, + 405 + ], + "score": 1.0, + "content": "names contain its word.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 112, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 112, + 409, + 233, + 422 + ], + "score": 1.0, + "content": "4. Vectorize the Augmented", + "type": "text" + }, + { + "bbox": [ + 233, + 410, + 286, + 420 + ], + "score": 0.4, + "content": "\\mathrm { A S T } + \\mathrm { G S C }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "graph into a form suitable for a GNN. (I.e. perform", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 126, + 420, + 505, + 433 + ], + "score": 1.0, + "content": "Step 1 from Section 3.2.) Each AST node that doesn’t represent a variable is vectorized as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 431, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 125, + 431, + 506, + 444 + ], + "score": 1.0, + "content": "a learned embedding of the language construct it represents, e.g. Parameter, Method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 126, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "Declaration, etc. Each cache node and each node that represents a variable is vector-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 124, + 452, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 124, + 452, + 505, + 467 + ], + "score": 1.0, + "content": "ized as a learned linear map of the concatenation of a type embedding and a name embedding.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 463, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 125, + 463, + 506, + 478 + ], + "score": 1.0, + "content": "The name embedding is a Character–Level Convolutional Neural Network (CharCNN) (Zhang", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 126, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "et al., 2015) embedding of the word/name the node contains. The type embedding is a learned", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 125, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "embedding of the name of the Java type of the token it contains, e.g. int, a user–defined class,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 497, + 406, + 510 + ], + "spans": [ + { + "bbox": [ + 125, + 497, + 406, + 510 + ], + "score": 1.0, + "content": "etc., with cache nodes having their own unique Cacher Node type.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 112, + 512, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 112, + 512, + 505, + 526 + ], + "score": 1.0, + "content": "5. Process the graph with a GNN, as per Section 3.2. (I.e. perform Steps 2 and 3 from Section", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 125, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "3.2.) The readout functions differ depending on the task and are described in the Experiments", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 536, + 185, + 547 + ], + "spans": [ + { + "bbox": [ + 126, + 536, + 185, + 547 + ], + "score": 1.0, + "content": "section below.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 556, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 557, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 504, + 568 + ], + "score": 1.0, + "content": "Our main contribution to previous works is the addition of Step 3, the Graph–Structured Cache", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "step. The combination of relational information from the cache nodes’ connections and lexical", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "information from these nodes’ CharCNN embeddings allows the model to, in principle, flexibly", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 590, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 601 + ], + "score": 1.0, + "content": "reason about words it never saw during training, but also recognize words it did. E.g. it could", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "potentially see a class named “getGuavaDictionary” and a variable named “guava dict”", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "and both (a) utilize the fact that the word “guava” is common to both names despite having never", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "score": 1.0, + "content": "seen this word before, and (b) exploit learned representations for words like “get”, “dictionary”, and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 258, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 258, + 647 + ], + "score": 1.0, + "content": "“dict” that it has seen during training.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 108, + 662, + 200, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 660, + 201, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 201, + 676 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "We evaluated our model, described in Section 4, on two supervised tasks: a Fill–In–The–Blank task", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "and a Variable Naming task. For each task, we compare our model to others that differ in how they", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "parse the code and how they treat the words they encounter. Table 1 details the different variations", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 379, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 379, + 733 + ], + "score": 1.0, + "content": "of the procedure in Section 4 against which we compare our model.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 126, + 82, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 129, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 129, + 81, + 195, + 96 + ], + "score": 1.0, + "content": "(b) Each vertex", + "type": "text" + }, + { + "bbox": [ + 195, + 83, + 222, + 93 + ], + "score": 0.9, + "content": "v \\in V", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "updates its hidden state based on the message it received via a function", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 92, + 164, + 108 + ], + "spans": [ + { + "bbox": [ + 146, + 94, + 158, + 105 + ], + "score": 0.85, + "content": "U _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 92, + 164, + 108 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 129, + 81, + 505, + 108 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 277, + 104, + 374, + 119 + ], + "lines": [ + { + "bbox": [ + 277, + 104, + 374, + 119 + ], + "spans": [ + { + "bbox": [ + 277, + 104, + 374, + 119 + ], + "score": 0.93, + "content": "h _ { v } ^ { t + 1 } = U _ { t } ( h _ { v } ^ { t } , m _ { v } ^ { t + 1 } ) .", + "type": "interline_equation", + "image_path": "3dbc90bc50fd06202e01f73e8218ed391fc7b8275fcd2ebea9880febecb6a67a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 277, + 104, + 374, + 119 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 111, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 113, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 113, + 122, + 172, + 136 + ], + "score": 1.0, + "content": "3. A function", + "type": "text" + }, + { + "bbox": [ + 173, + 124, + 181, + 134 + ], + "score": 0.81, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 122, + 505, + 136 + ], + "score": 1.0, + "content": ", the “readout function”, produces a prediction based on the hidden states gener-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 135, + 399, + 147 + ], + "spans": [ + { + "bbox": [ + 126, + 135, + 384, + 147 + ], + "score": 1.0, + "content": "ated during the message passing (usually just those at from time", + "type": "text" + }, + { + "bbox": [ + 384, + 135, + 392, + 145 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 135, + 399, + 147 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 113, + 122, + 505, + 147 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 151, + 384, + 167 + ], + "lines": [ + { + "bbox": [ + 247, + 151, + 384, + 167 + ], + "spans": [ + { + "bbox": [ + 247, + 151, + 384, + 167 + ], + "score": 0.91, + "content": "\\hat { y } = R ( \\{ h _ { v } ^ { t } | v \\in V , t \\in 1 , . . . , T \\} ) .", + "type": "interline_equation", + "image_path": "4a007b03cf24477ac91e4a1b0db3625e7401553f5af46fc8f711e359018c39a2.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 247, + 151, + 384, + 167 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 176, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 252, + 188 + ], + "score": 1.0, + "content": "GNNs differ in how they implement", + "type": "text" + }, + { + "bbox": [ + 253, + 177, + 293, + 187 + ], + "score": 0.27, + "content": "S , M _ { t } , U _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 176, + 313, + 188 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 314, + 177, + 322, + 186 + ], + "score": 0.81, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 176, + 505, + 188 + ], + "score": 1.0, + "content": ". But all these functions are differentiable and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "most are parameterized, so the model is trainable via stochastic gradient descent of a loss function", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 154, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 119, + 210 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 200, + 126, + 210 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 198, + 144, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 199, + 150, + 210 + ], + "score": 0.83, + "content": "\\hat { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 198, + 154, + 210 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 176, + 505, + 210 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 226, + 165, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 167, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 167, + 241 + ], + "score": 1.0, + "content": "4 MODEL", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 253, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "score": 1.0, + "content": "Our model consumes an input instance of source code and produces an output for a supervised", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 263, + 359, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 359, + 276 + ], + "score": 1.0, + "content": "learning task via the following five steps, sketched in Figure 1:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 251, + 506, + 276 + ] + }, + { + "type": "list", + "bbox": [ + 113, + 284, + 506, + 546 + ], + "lines": [ + { + "bbox": [ + 113, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 113, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "1. Parse the source code (snippet, file, repository, version control history, etc.) into an Abstract", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 296, + 179, + 307 + ], + "spans": [ + { + "bbox": [ + 126, + 296, + 179, + 307 + ], + "score": 1.0, + "content": "Syntax Tree.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 111, + 310, + 506, + 325 + ], + "score": 1.0, + "content": "2. Add edges of varying types (details in Appendix Table 8) to this AST that represent semantic", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 126, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "information like data– and control– flow, in the spirit of Allamanis et al. (2018). Also add the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 125, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "reversed version of all edges with their own edge type. This results in a directed multigraph", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 345, + 236, + 357 + ], + "spans": [ + { + "bbox": [ + 126, + 345, + 236, + 357 + ], + "score": 1.0, + "content": "called an Augmented AST.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 113, + 361, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 113, + 361, + 506, + 373 + ], + "score": 1.0, + "content": "3. Further augment the Augmented AST by adding a Graph–Structured Cache. That is, add a", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 372, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 126, + 372, + 505, + 383 + ], + "score": 1.0, + "content": "node to the Augmented AST for each vocabulary word encountered in the input instance. Then", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 383, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 126, + 383, + 505, + 395 + ], + "score": 1.0, + "content": "connect each such “cache node” with an edge (of edge type WORD USE) to all variables whose", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 394, + 223, + 405 + ], + "spans": [ + { + "bbox": [ + 125, + 394, + 223, + 405 + ], + "score": 1.0, + "content": "names contain its word.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 112, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 112, + 409, + 233, + 422 + ], + "score": 1.0, + "content": "4. Vectorize the Augmented", + "type": "text" + }, + { + "bbox": [ + 233, + 410, + 286, + 420 + ], + "score": 0.4, + "content": "\\mathrm { A S T } + \\mathrm { G S C }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "graph into a form suitable for a GNN. (I.e. perform", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 126, + 420, + 505, + 433 + ], + "score": 1.0, + "content": "Step 1 from Section 3.2.) Each AST node that doesn’t represent a variable is vectorized as", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 431, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 125, + 431, + 506, + 444 + ], + "score": 1.0, + "content": "a learned embedding of the language construct it represents, e.g. Parameter, Method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 126, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "Declaration, etc. Each cache node and each node that represents a variable is vector-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 124, + 452, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 124, + 452, + 505, + 467 + ], + "score": 1.0, + "content": "ized as a learned linear map of the concatenation of a type embedding and a name embedding.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 125, + 463, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 125, + 463, + 506, + 478 + ], + "score": 1.0, + "content": "The name embedding is a Character–Level Convolutional Neural Network (CharCNN) (Zhang", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 126, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "et al., 2015) embedding of the word/name the node contains. The type embedding is a learned", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 125, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "embedding of the name of the Java type of the token it contains, e.g. int, a user–defined class,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 497, + 406, + 510 + ], + "spans": [ + { + "bbox": [ + 125, + 497, + 406, + 510 + ], + "score": 1.0, + "content": "etc., with cache nodes having their own unique Cacher Node type.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 112, + 512, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 112, + 512, + 505, + 526 + ], + "score": 1.0, + "content": "5. Process the graph with a GNN, as per Section 3.2. (I.e. perform Steps 2 and 3 from Section", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 125, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "3.2.) The readout functions differ depending on the task and are described in the Experiments", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 536, + 185, + 547 + ], + "spans": [ + { + "bbox": [ + 126, + 536, + 185, + 547 + ], + "score": 1.0, + "content": "section below.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + } + ], + "index": 22.5, + "bbox_fs": [ + 111, + 284, + 506, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 556, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 557, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 504, + 568 + ], + "score": 1.0, + "content": "Our main contribution to previous works is the addition of Step 3, the Graph–Structured Cache", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "step. The combination of relational information from the cache nodes’ connections and lexical", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "information from these nodes’ CharCNN embeddings allows the model to, in principle, flexibly", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 590, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 601 + ], + "score": 1.0, + "content": "reason about words it never saw during training, but also recognize words it did. E.g. it could", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "potentially see a class named “getGuavaDictionary” and a variable named “guava dict”", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "and both (a) utilize the fact that the word “guava” is common to both names despite having never", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "score": 1.0, + "content": "seen this word before, and (b) exploit learned representations for words like “get”, “dictionary”, and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 258, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 258, + 647 + ], + "score": 1.0, + "content": "“dict” that it has seen during training.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 557, + 506, + 647 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 662, + 200, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 660, + 201, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 201, + 676 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "We evaluated our model, described in Section 4, on two supervised tasks: a Fill–In–The–Blank task", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "and a Variable Naming task. For each task, we compare our model to others that differ in how they", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "parse the code and how they treat the words they encounter. Table 1 details the different variations", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 379, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 379, + 733 + ], + "score": 1.0, + "content": "of the procedure in Section 4 against which we compare our model.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 82, + 509, + 582 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 82, + 509, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 82, + 509, + 582 + ], + "spans": [ + { + "bbox": [ + 110, + 82, + 509, + 582 + ], + "score": 0.829, + "type": "image", + "image_path": "fa59b4133b5607587cc70ae7b40b4e6a20f3afab3cf1561f1e3c0ebf601145cb.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 82, + 509, + 248.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 248.66666666666666, + 509, + 415.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 415.3333333333333, + 509, + 582.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 589, + 505, + 612 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 604 + ], + "score": 1.0, + "content": "Figure 1: Our model’s procedure for consuming a single input instance of source code and producing", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 601, + 268, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 268, + 612 + ], + "score": 1.0, + "content": "an output for a supervised learning task.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 333, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 335, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 335, + 645 + ], + "score": 1.0, + "content": "Code to reproduce all experiments is available online.3 4", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 658, + 295, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 297, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 297, + 671 + ], + "score": 1.0, + "content": "5.1 DATA AND IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 108, + 679, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "We chose to use Java source code as the data for our experiments as it is among the most popular", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "score": 1.0, + "content": "programming languages in use today (TIOBE, 2018; Github, 2017). To construct our dataset, we", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 119, + 710, + 264, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 708, + 264, + 722 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 264, + 722 + ], + "score": 1.0, + "content": "3URL redacted to preserve review blind", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 264, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 264, + 734 + ], + "score": 1.0, + "content": "4URL redacted to preserve review blind", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 82, + 509, + 582 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 82, + 509, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 82, + 509, + 582 + ], + "spans": [ + { + "bbox": [ + 110, + 82, + 509, + 582 + ], + "score": 0.829, + "type": "image", + "image_path": "fa59b4133b5607587cc70ae7b40b4e6a20f3afab3cf1561f1e3c0ebf601145cb.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 82, + 509, + 248.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 248.66666666666666, + 509, + 415.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 415.3333333333333, + 509, + 582.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 589, + 505, + 612 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 604 + ], + "score": 1.0, + "content": "Figure 1: Our model’s procedure for consuming a single input instance of source code and producing", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 601, + 268, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 268, + 612 + ], + "score": 1.0, + "content": "an output for a supervised learning task.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 333, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 335, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 335, + 645 + ], + "score": 1.0, + "content": "Code to reproduce all experiments is available online.3 4", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 631, + 335, + 645 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 658, + 295, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 297, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 297, + 671 + ], + "score": 1.0, + "content": "5.1 DATA AND IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 108, + 679, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "We chose to use Java source code as the data for our experiments as it is among the most popular", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "score": 1.0, + "content": "programming languages in use today (TIOBE, 2018; Github, 2017). To construct our dataset, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "randomly selected 18 of the 100 most popular Java repos from the Maven repository5 to serve as", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 482, + 504, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 504, + 494 + ], + "score": 1.0, + "content": "training data. (See Appendix Table 7 for the list.) Together these repositories contain about 500,000", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 506, + 505 + ], + "score": 1.0, + "content": "non–empty, non–comment lines of code. We checked for excessive code duplication in our dataset", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 503, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 211, + 517 + ], + "score": 1.0, + "content": "(Lopes et al., 2017) using", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 212, + 503, + 237, + 514 + ], + "score": 0.85, + "content": "\\mathrm { C P D ^ { 6 } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 237, + 503, + 326, + 517 + ], + "score": 1.0, + "content": "and found only about", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 326, + 504, + 341, + 514 + ], + "score": 0.85, + "content": "7 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 341, + 503, + 506, + 517 + ], + "score": 1.0, + "content": "of the lines to be contiguous, duplicated", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 514, + 291, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 291, + 528 + ], + "score": 1.0, + "content": "code blocks containing more than 150 tokens.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 678, + 506, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 123, + 504, + 449 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 123 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "Table 1: Nomenclature used in Experiments section. Each abbreviation describes a tweak/ablation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "to our full model as presented in Section 4. Using this nomenclature, our full model as described in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 384, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 384, + 124 + ], + "score": 1.0, + "content": "Section 4 and shown in Figure 1 would be a “AugAST–GSC” model.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 108, + 123, + 504, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 123, + 504, + 449 + ], + "spans": [ + { + "bbox": [ + 108, + 123, + 504, + 449 + ], + "score": 0.984, + "html": "
AbbreviationMeaning
Code RepresentationASTSkips Step 2 in Section 4.
AugASTPerforms Step 2 in Section 4.
Vocab StrategiesClosed VocabSkips Step 3 in Section 4,and instead maintains word-embedding vectors for words in a closed vo- cabulary. In Step 4, name embeddings for nodes representing variables are produced by taking the mean of the embeddings of the words in the vari- able's name.Words outside this model's closed vocabulary are labeled as <UNK>. This is the
CharCNNstrategy used in Allamanis et al. (2018). Skips Step 3 in Section 4.
Pointer SentinelFollows Steps 3 and 4 as described in Section 4,except it doesn't add edges connecting cache nodes to the nodes where their word is used. In the Variable Naming task, this is equivalent to us- ing the Pointer Sentinel Mixture Model of Merity
GSCet al. (2017) to produce outputs. Follows Steps 3 and 4 as described in Section 4.
Graph Neural NetworkGGNNPerforms Step 5 in Section 4 using the Gated Graph Neural Network of Li et al. (2016).
DTNNPerforms Step 5 in Section 4 using the Deep Ten- sor Neural Network of Schuitt et al. (2017).
RGCNPerforms Step 5 in Section 4 using the Relational Graph Convolutional Network of Schlichtkrull et al. (2017).
", + "type": "table", + "image_path": "c127588b501d99382c3ea3beceafe747f6da712578b6b15fbdf67ad98a5f51cb.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 108, + 123, + 504, + 231.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 231.66666666666669, + 504, + 340.33333333333337 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 340.33333333333337, + 504, + 449.00000000000006 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "randomly selected 18 of the 100 most popular Java repos from the Maven repository5 to serve as", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 482, + 504, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 504, + 494 + ], + "score": 1.0, + "content": "training data. (See Appendix Table 7 for the list.) Together these repositories contain about 500,000", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 506, + 505 + ], + "score": 1.0, + "content": "non–empty, non–comment lines of code. We checked for excessive code duplication in our dataset", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 503, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 211, + 517 + ], + "score": 1.0, + "content": "(Lopes et al., 2017) using", + "type": "text" + }, + { + "bbox": [ + 212, + 503, + 237, + 514 + ], + "score": 0.85, + "content": "\\mathrm { C P D ^ { 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 503, + 326, + 517 + ], + "score": 1.0, + "content": "and found only about", + "type": "text" + }, + { + "bbox": [ + 326, + 504, + 341, + 514 + ], + "score": 0.85, + "content": "7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 503, + 506, + 517 + ], + "score": 1.0, + "content": "of the lines to be contiguous, duplicated", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 514, + 291, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 291, + 528 + ], + "score": 1.0, + "content": "code blocks containing more than 150 tokens.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 531, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "We randomly chose 3 of these repositories to sequester as an “Unseen Repos” test set. We then", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 162, + 555 + ], + "score": 1.0, + "content": "separated out", + "type": "text" + }, + { + "bbox": [ + 162, + 543, + 181, + 553 + ], + "score": 0.87, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "of the files in the remaining 15 repositories to serve as our “Seen Repos” test set.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 553, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 392, + 566 + ], + "score": 1.0, + "content": "The remaining files served as our training set, from which we separated", + "type": "text" + }, + { + "bbox": [ + 392, + 554, + 411, + 564 + ], + "score": 0.86, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 553, + 506, + 566 + ], + "score": 1.0, + "content": "of the datapoints to act", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 564, + 182, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 182, + 576 + ], + "score": 1.0, + "content": "as a validation set.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 506, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "Our data preprocessor builds on top of the open–source Javaparser7 library to generate ASTs of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "our source code and then augment the ASTs with the edges described in Appendix Table 8. We", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 160, + 616 + ], + "score": 1.0, + "content": "used Apache", + "type": "text" + }, + { + "bbox": [ + 160, + 603, + 196, + 614 + ], + "score": 0.51, + "content": "\\mathrm { { \\bf M X N e t } } ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "as our deep learning framework. All hidden states in the GNN contained 64", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "score": 1.0, + "content": "units; all GNNs ran for 8 rounds of message passing; all models were optimized using the Adam", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "optimizer (Kingma & Ba, 2015); all inputs to the GNNs were truncated to a maximum size of 500", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 329, + 648 + ], + "score": 1.0, + "content": "nodes centered on the or", + "type": "text" + }, + { + "bbox": [ + 329, + 637, + 384, + 647 + ], + "score": 0.64, + "content": "\\mathrm { \\ < N A M E - M E > }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 636, + 448, + 648 + ], + "score": 1.0, + "content": "tokens. About", + "type": "text" + }, + { + "bbox": [ + 448, + 636, + 468, + 647 + ], + "score": 0.85, + "content": "53 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "of input", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "graphs were larger than 500 nodes before truncation. The only regularization we used was early", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 328, + 672 + ], + "score": 1.0, + "content": "stopping — early in our experiments we briefly tried", + "type": "text" + }, + { + "bbox": [ + 329, + 659, + 341, + 669 + ], + "score": 0.88, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "and dropout regularization, but saw no", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 668, + 138, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 138, + 681 + ], + "score": 1.0, + "content": "effects.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 688, + 393, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 686, + 265, + 702 + ], + "spans": [ + { + "bbox": [ + 118, + 686, + 265, + 702 + ], + "score": 1.0, + "content": "5https://mvnrepository.com/", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 696, + 394, + 713 + ], + "spans": [ + { + "bbox": [ + 117, + 696, + 394, + 713 + ], + "score": 1.0, + "content": "6https://pmd.github.io/latest/pmd_userdocs_cpd.html", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 708, + 249, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 249, + 724 + ], + "score": 1.0, + "content": "7https://javaparser.org/", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 718, + 259, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 718, + 259, + 734 + ], + "score": 1.0, + "content": "8https://mxnet.apache.org/", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 123, + 504, + 449 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 123 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "Table 1: Nomenclature used in Experiments section. Each abbreviation describes a tweak/ablation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "to our full model as presented in Section 4. Using this nomenclature, our full model as described in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 384, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 384, + 124 + ], + "score": 1.0, + "content": "Section 4 and shown in Figure 1 would be a “AugAST–GSC” model.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 108, + 123, + 504, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 123, + 504, + 449 + ], + "spans": [ + { + "bbox": [ + 108, + 123, + 504, + 449 + ], + "score": 0.984, + "html": "
AbbreviationMeaning
Code RepresentationASTSkips Step 2 in Section 4.
AugASTPerforms Step 2 in Section 4.
Vocab StrategiesClosed VocabSkips Step 3 in Section 4,and instead maintains word-embedding vectors for words in a closed vo- cabulary. In Step 4, name embeddings for nodes representing variables are produced by taking the mean of the embeddings of the words in the vari- able's name.Words outside this model's closed vocabulary are labeled as <UNK>. This is the
CharCNNstrategy used in Allamanis et al. (2018). Skips Step 3 in Section 4.
Pointer SentinelFollows Steps 3 and 4 as described in Section 4,except it doesn't add edges connecting cache nodes to the nodes where their word is used. In the Variable Naming task, this is equivalent to us- ing the Pointer Sentinel Mixture Model of Merity
GSCet al. (2017) to produce outputs. Follows Steps 3 and 4 as described in Section 4.
Graph Neural NetworkGGNNPerforms Step 5 in Section 4 using the Gated Graph Neural Network of Li et al. (2016).
DTNNPerforms Step 5 in Section 4 using the Deep Ten- sor Neural Network of Schuitt et al. (2017).
RGCNPerforms Step 5 in Section 4 using the Relational Graph Convolutional Network of Schlichtkrull et al. (2017).
", + "type": "table", + "image_path": "c127588b501d99382c3ea3beceafe747f6da712578b6b15fbdf67ad98a5f51cb.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 108, + 123, + 504, + 231.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 231.66666666666669, + 504, + 340.33333333333337 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 340.33333333333337, + 504, + 449.00000000000006 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 526 + ], + "lines": [], + "index": 8, + "bbox_fs": [ + 104, + 470, + 506, + 528 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 531, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "We randomly chose 3 of these repositories to sequester as an “Unseen Repos” test set. We then", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 162, + 555 + ], + "score": 1.0, + "content": "separated out", + "type": "text" + }, + { + "bbox": [ + 162, + 543, + 181, + 553 + ], + "score": 0.87, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "of the files in the remaining 15 repositories to serve as our “Seen Repos” test set.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 553, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 392, + 566 + ], + "score": 1.0, + "content": "The remaining files served as our training set, from which we separated", + "type": "text" + }, + { + "bbox": [ + 392, + 554, + 411, + 564 + ], + "score": 0.86, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 553, + 506, + 566 + ], + "score": 1.0, + "content": "of the datapoints to act", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 564, + 182, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 182, + 576 + ], + "score": 1.0, + "content": "as a validation set.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 531, + 506, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 506, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "Our data preprocessor builds on top of the open–source Javaparser7 library to generate ASTs of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "our source code and then augment the ASTs with the edges described in Appendix Table 8. We", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 160, + 616 + ], + "score": 1.0, + "content": "used Apache", + "type": "text" + }, + { + "bbox": [ + 160, + 603, + 196, + 614 + ], + "score": 0.51, + "content": "\\mathrm { { \\bf M X N e t } } ^ { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "as our deep learning framework. All hidden states in the GNN contained 64", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "score": 1.0, + "content": "units; all GNNs ran for 8 rounds of message passing; all models were optimized using the Adam", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "optimizer (Kingma & Ba, 2015); all inputs to the GNNs were truncated to a maximum size of 500", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 329, + 648 + ], + "score": 1.0, + "content": "nodes centered on the or", + "type": "text" + }, + { + "bbox": [ + 329, + 637, + 384, + 647 + ], + "score": 0.64, + "content": "\\mathrm { \\ < N A M E - M E > }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 636, + 448, + 648 + ], + "score": 1.0, + "content": "tokens. About", + "type": "text" + }, + { + "bbox": [ + 448, + 636, + 468, + 647 + ], + "score": 0.85, + "content": "53 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "of input", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "graphs were larger than 500 nodes before truncation. The only regularization we used was early", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 328, + 672 + ], + "score": 1.0, + "content": "stopping — early in our experiments we briefly tried", + "type": "text" + }, + { + "bbox": [ + 329, + 659, + 341, + 669 + ], + "score": 0.88, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "and dropout regularization, but saw no", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 668, + 138, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 138, + 681 + ], + "score": 1.0, + "content": "effects.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 581, + 506, + 681 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "We performed only a moderate amount of hyperparameter optimization, but all of it was done on", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "the baseline models to avoid biasing our results in favor of our model. Specifically, we tuned all", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "hyperparameters on the Closed Vocab baseline model, and also did a small amount of extra learning", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 473, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 473, + 128 + ], + "score": 1.0, + "content": "rate exploration for the Pointer Sentinel baseline model to try to maximize its performance.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 108, + 140, + 274, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 275, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 275, + 153 + ], + "score": 1.0, + "content": "5.2 THE FILL–IN–THE–BLANK TASK", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 160, + 506, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "In this task we randomly selected a single usage of a variable in some source code, replaced it with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "a token, and then asked the model to predict what variable should have", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 182, + 393, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 393, + 195 + ], + "score": 1.0, + "content": "been there. An example instance from our dataset is shown in Figure 2.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 110, + 204, + 511, + 304 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 204, + 511, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 204, + 511, + 304 + ], + "spans": [ + { + "bbox": [ + 110, + 204, + 511, + 304 + ], + "score": 0.959, + "type": "image", + "image_path": "69398b3fef9fe3f5b523898c2773311a427c29f26873c4f47ff60001f36d5b3c.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 204, + 511, + 237.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 237.33333333333334, + 511, + 270.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 270.6666666666667, + 511, + 304.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 318, + 506, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "Figure 2: Example of a model’s procedure for completing the Fill–In–The–Blank task. Each Fill–", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "In–The–Blank instance is created by replacing a single usage of a variable (n, in this example) with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "the special token . The model then processes the code as depicted in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "Figure 1. To produce outputs, the model’s readout function computes a soft–attention weighting", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 363, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 374 + ], + "score": 1.0, + "content": "over all nodes in the graph; the model’s output is the variable at the node on which it places maximal", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "attention. In this example, if the model put maximal attention weighting on any of the green–", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "score": 1.0, + "content": "highlighted variables, this would be a correct output. If maximal attention is placed on any other", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 396, + 492, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 492, + 408 + ], + "score": 1.0, + "content": "node, it would be an incorrect output. Only in–scope usages of a variable are counted as correct.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + } + ], + "index": 11.75 + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "The models indicate their prediction for what variable should go in the blank by pointing with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "neural attention over all the nodes in the AugAST. This means all training and test instances only", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 437, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 452 + ], + "score": 1.0, + "content": "considered cases where the obfuscated variable appears somewhere else in the code. Single uses are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 450, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 462 + ], + "score": 1.0, + "content": "rare however, since in Java variables must be declared before they are used. It also means there are", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "sometimes multiple usages of the same, correct variable to which a model can point to get the right", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 472, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 198, + 483 + ], + "score": 1.0, + "content": "answer. In our dataset", + "type": "text" + }, + { + "bbox": [ + 198, + 472, + 218, + 482 + ], + "score": 0.86, + "content": "78 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 472, + 397, + 483 + ], + "score": 1.0, + "content": "of variables were used more two times, and", + "type": "text" + }, + { + "bbox": [ + 398, + 472, + 417, + 482 + ], + "score": 0.86, + "content": "33 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 472, + 505, + 483 + ], + "score": 1.0, + "content": "were used more than", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 153, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 153, + 494 + ], + "score": 1.0, + "content": "four times.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 503, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 287, + 513 + ], + "score": 1.0, + "content": "The models compute the attention weightings", + "type": "text" + }, + { + "bbox": [ + 287, + 501, + 297, + 511 + ], + "score": 0.86, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 498, + 421, + 513 + ], + "score": 1.0, + "content": "for each Augmented AST node", + "type": "text" + }, + { + "bbox": [ + 421, + 500, + 426, + 509 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 498, + 505, + 513 + ], + "score": 1.0, + "content": "differently depend-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "ing on the readout function of the GNN they use. Models using a GGNN as their GNN component,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 521, + 437, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 187, + 534 + ], + "score": 1.0, + "content": "as all those in Table", + "type": "text" + }, + { + "bbox": [ + 188, + 522, + 207, + 532 + ], + "score": 0.6, + "content": "2 \\ : \\mathrm { d o }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 521, + 437, + 534 + ], + "score": 1.0, + "content": ", compute the attention weightings as per Li et al. (2016):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 534, + 370, + 550 + ], + "lines": [ + { + "bbox": [ + 239, + 534, + 370, + 550 + ], + "spans": [ + { + "bbox": [ + 239, + 534, + 370, + 550 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\hat { y } _ { i } = \\sigma \\left( f _ { 1 } ( { h } _ { v } ^ { T } , { h } _ { v } ^ { 0 } ) \\right) \\odot f _ { 2 } ( { h } _ { v } ^ { T } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "a925b395ee215cf84d388e8c4326652c58f7e2ee0e0385f03d8bdf851612fd8a.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 239, + 534, + 370, + 550 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 148, + 565 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 148, + 553, + 155, + 564 + ], + "score": 0.61, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 551, + 204, + 565 + ], + "score": 1.0, + "content": "s are MLPs,", + "type": "text" + }, + { + "bbox": [ + 205, + 552, + 217, + 564 + ], + "score": 0.9, + "content": "\\boldsymbol { h } _ { v } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 551, + 326, + 565 + ], + "score": 1.0, + "content": "is the hidden state of node", + "type": "text" + }, + { + "bbox": [ + 326, + 555, + 333, + 562 + ], + "score": 0.78, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 551, + 354, + 565 + ], + "score": 1.0, + "content": "after", + "type": "text" + }, + { + "bbox": [ + 355, + 554, + 360, + 562 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 551, + 472, + 565 + ], + "score": 1.0, + "content": "message passing iterations,", + "type": "text" + }, + { + "bbox": [ + 473, + 555, + 480, + 562 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 196, + 576 + ], + "score": 1.0, + "content": "sigmoid function, and", + "type": "text" + }, + { + "bbox": [ + 197, + 565, + 205, + 574 + ], + "score": 0.85, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "is elementwise multiplication. The DTNN and RGCN GNNs compute the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 574, + 298, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 298, + 586 + ], + "score": 1.0, + "content": "attention weightings as per Schutt et al. (2017): ¨", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 278, + 587, + 331, + 602 + ], + "lines": [ + { + "bbox": [ + 278, + 587, + 331, + 602 + ], + "spans": [ + { + "bbox": [ + 278, + 587, + 331, + 602 + ], + "score": 0.91, + "content": "\\hat { y } _ { i } = f ( h _ { v } ^ { T } ) ,", + "type": "interline_equation", + "image_path": "6c6afc591cf1bccd6c9f965cb2e5cf87ed755236782ed9c75f2680482e78247c.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 278, + 587, + 331, + 602 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 133, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 604, + 141, + 615 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 601, + 506, + 617 + ], + "score": 1.0, + "content": "is a single hidden layer MLP. The models were trained using a binary cross entropy loss", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 615, + 268, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 268, + 628 + ], + "score": 1.0, + "content": "computed across the nodes in the graph.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "The performance of models using our GSC versus those using other methods is reported in Table 2.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "For context, a baseline strategy of random guessing among all variable nodes within an edge radius", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 260, + 666 + ], + "score": 1.0, + "content": "of 8 of the ", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "token achieves an accuracy of 0.22. We also compare the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 665, + 280, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 280, + 676 + ], + "score": 1.0, + "content": "performance of different GNNs in Table 3.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 262, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 263, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 263, + 701 + ], + "score": 1.0, + "content": "5.3 THE VARIABLE NAMING TASK", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "In this task we replaced all usages of a name of a particular variable, method, class, or parameter", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 297, + 733 + ], + "score": 1.0, + "content": "in the code with the special token ", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ", and asked the model to produce the obfuscated", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "We performed only a moderate amount of hyperparameter optimization, but all of it was done on", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "the baseline models to avoid biasing our results in favor of our model. Specifically, we tuned all", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "hyperparameters on the Closed Vocab baseline model, and also did a small amount of extra learning", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 473, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 473, + 128 + ], + "score": 1.0, + "content": "rate exploration for the Pointer Sentinel baseline model to try to maximize its performance.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 128 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 140, + 274, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 275, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 275, + 153 + ], + "score": 1.0, + "content": "5.2 THE FILL–IN–THE–BLANK TASK", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 160, + 506, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "In this task we randomly selected a single usage of a variable in some source code, replaced it with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "a token, and then asked the model to predict what variable should have", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 182, + 393, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 393, + 195 + ], + "score": 1.0, + "content": "been there. An example instance from our dataset is shown in Figure 2.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 160, + 506, + 195 + ] + }, + { + "type": "image", + "bbox": [ + 110, + 204, + 511, + 304 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 204, + 511, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 204, + 511, + 304 + ], + "spans": [ + { + "bbox": [ + 110, + 204, + 511, + 304 + ], + "score": 0.959, + "type": "image", + "image_path": "69398b3fef9fe3f5b523898c2773311a427c29f26873c4f47ff60001f36d5b3c.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 204, + 511, + 237.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 237.33333333333334, + 511, + 270.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 270.6666666666667, + 511, + 304.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 318, + 506, + 407 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "Figure 2: Example of a model’s procedure for completing the Fill–In–The–Blank task. Each Fill–", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "In–The–Blank instance is created by replacing a single usage of a variable (n, in this example) with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "the special token . The model then processes the code as depicted in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "Figure 1. To produce outputs, the model’s readout function computes a soft–attention weighting", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 363, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 374 + ], + "score": 1.0, + "content": "over all nodes in the graph; the model’s output is the variable at the node on which it places maximal", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "attention. In this example, if the model put maximal attention weighting on any of the green–", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "score": 1.0, + "content": "highlighted variables, this would be a correct output. If maximal attention is placed on any other", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 396, + 492, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 492, + 408 + ], + "score": 1.0, + "content": "node, it would be an incorrect output. Only in–scope usages of a variable are counted as correct.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + } + ], + "index": 11.75 + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "The models indicate their prediction for what variable should go in the blank by pointing with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "neural attention over all the nodes in the AugAST. This means all training and test instances only", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 437, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 452 + ], + "score": 1.0, + "content": "considered cases where the obfuscated variable appears somewhere else in the code. Single uses are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 450, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 462 + ], + "score": 1.0, + "content": "rare however, since in Java variables must be declared before they are used. It also means there are", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "sometimes multiple usages of the same, correct variable to which a model can point to get the right", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 472, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 198, + 483 + ], + "score": 1.0, + "content": "answer. In our dataset", + "type": "text" + }, + { + "bbox": [ + 198, + 472, + 218, + 482 + ], + "score": 0.86, + "content": "78 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 472, + 397, + 483 + ], + "score": 1.0, + "content": "of variables were used more two times, and", + "type": "text" + }, + { + "bbox": [ + 398, + 472, + 417, + 482 + ], + "score": 0.86, + "content": "33 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 472, + 505, + 483 + ], + "score": 1.0, + "content": "were used more than", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 153, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 153, + 494 + ], + "score": 1.0, + "content": "four times.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 416, + 506, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 503, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 287, + 513 + ], + "score": 1.0, + "content": "The models compute the attention weightings", + "type": "text" + }, + { + "bbox": [ + 287, + 501, + 297, + 511 + ], + "score": 0.86, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 498, + 421, + 513 + ], + "score": 1.0, + "content": "for each Augmented AST node", + "type": "text" + }, + { + "bbox": [ + 421, + 500, + 426, + 509 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 498, + 505, + 513 + ], + "score": 1.0, + "content": "differently depend-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "ing on the readout function of the GNN they use. Models using a GGNN as their GNN component,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 521, + 437, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 187, + 534 + ], + "score": 1.0, + "content": "as all those in Table", + "type": "text" + }, + { + "bbox": [ + 188, + 522, + 207, + 532 + ], + "score": 0.6, + "content": "2 \\ : \\mathrm { d o }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 521, + 437, + 534 + ], + "score": 1.0, + "content": ", compute the attention weightings as per Li et al. (2016):", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 498, + 505, + 534 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 534, + 370, + 550 + ], + "lines": [ + { + "bbox": [ + 239, + 534, + 370, + 550 + ], + "spans": [ + { + "bbox": [ + 239, + 534, + 370, + 550 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\hat { y } _ { i } = \\sigma \\left( f _ { 1 } ( { h } _ { v } ^ { T } , { h } _ { v } ^ { 0 } ) \\right) \\odot f _ { 2 } ( { h } _ { v } ^ { T } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "a925b395ee215cf84d388e8c4326652c58f7e2ee0e0385f03d8bdf851612fd8a.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 239, + 534, + 370, + 550 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 148, + 565 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 148, + 553, + 155, + 564 + ], + "score": 0.61, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 551, + 204, + 565 + ], + "score": 1.0, + "content": "s are MLPs,", + "type": "text" + }, + { + "bbox": [ + 205, + 552, + 217, + 564 + ], + "score": 0.9, + "content": "\\boldsymbol { h } _ { v } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 551, + 326, + 565 + ], + "score": 1.0, + "content": "is the hidden state of node", + "type": "text" + }, + { + "bbox": [ + 326, + 555, + 333, + 562 + ], + "score": 0.78, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 551, + 354, + 565 + ], + "score": 1.0, + "content": "after", + "type": "text" + }, + { + "bbox": [ + 355, + 554, + 360, + 562 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 551, + 472, + 565 + ], + "score": 1.0, + "content": "message passing iterations,", + "type": "text" + }, + { + "bbox": [ + 473, + 555, + 480, + 562 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 196, + 576 + ], + "score": 1.0, + "content": "sigmoid function, and", + "type": "text" + }, + { + "bbox": [ + 197, + 565, + 205, + 574 + ], + "score": 0.85, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "is elementwise multiplication. The DTNN and RGCN GNNs compute the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 574, + 298, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 298, + 586 + ], + "score": 1.0, + "content": "attention weightings as per Schutt et al. (2017): ¨", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 551, + 505, + 586 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 278, + 587, + 331, + 602 + ], + "lines": [ + { + "bbox": [ + 278, + 587, + 331, + 602 + ], + "spans": [ + { + "bbox": [ + 278, + 587, + 331, + 602 + ], + "score": 0.91, + "content": "\\hat { y } _ { i } = f ( h _ { v } ^ { T } ) ,", + "type": "interline_equation", + "image_path": "6c6afc591cf1bccd6c9f965cb2e5cf87ed755236782ed9c75f2680482e78247c.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 278, + 587, + 331, + 602 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 133, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 604, + 141, + 615 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 601, + 506, + 617 + ], + "score": 1.0, + "content": "is a single hidden layer MLP. The models were trained using a binary cross entropy loss", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 615, + 268, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 268, + 628 + ], + "score": 1.0, + "content": "computed across the nodes in the graph.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 601, + 506, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "The performance of models using our GSC versus those using other methods is reported in Table 2.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "score": 1.0, + "content": "For context, a baseline strategy of random guessing among all variable nodes within an edge radius", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 260, + 666 + ], + "score": 1.0, + "content": "of 8 of the ", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "token achieves an accuracy of 0.22. We also compare the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 665, + 280, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 280, + 676 + ], + "score": 1.0, + "content": "performance of different GNNs in Table 3.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 631, + 505, + 676 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 262, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 263, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 263, + 701 + ], + "score": 1.0, + "content": "5.3 THE VARIABLE NAMING TASK", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "In this task we replaced all usages of a name of a particular variable, method, class, or parameter", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 297, + 733 + ], + "score": 1.0, + "content": "in the code with the special token ", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ", and asked the model to produce the obfuscated", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 379, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 394 + ], + "score": 1.0, + "content": "name (in the form of the sequence of words that compose the name). An example instance from our", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 391, + 223, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 223, + 404 + ], + "score": 1.0, + "content": "dataset is shown in Figure 3.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 709, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 167, + 449, + 236 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 88, + 506, + 166 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "score": 1.0, + "content": "Table 2: Accuracy on the Fill–In–The–Blank task. Our model is the AugAST–GSC. The first number", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "in each cell is the accuracy of the model, where a correct prediction is one in which the graph", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 112, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 123 + ], + "score": 1.0, + "content": "node that received the maximum attention weighting by the model contained the variable that was", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 281, + 135 + ], + "score": 1.0, + "content": "originally in the ", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "spot. The second, parenthetical numbers are the top–", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "5 accuracies, i.e. whether the correct node was among those that received the 5 largest attentions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "weightings from the model. See Table 1 for explanations of the abbreviations. All models use Gated", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 309, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 309, + 168 + ], + "score": 1.0, + "content": "Graph Neural Networks as their GNN component.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 161, + 167, + 449, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 161, + 167, + 449, + 236 + ], + "spans": [ + { + "bbox": [ + 161, + 167, + 449, + 236 + ], + "score": 0.976, + "html": "
Closed VocabCharCNNGSC
Seen reposAST0.57 (0.83)0.60 (0.84)0.89 (0.96)
AugAST0.80 (0.90)0.90 (0.94)0.97 (0.99)
Unseen reposAST AugAST0.36 (0.68)0.48 (0.80)0.80 (0.93)
0.59 (0.78)0.84 (0.92)0.92 (0.96)
", + "type": "table", + "image_path": "64e9ca199b864180f85246fd6d2fa82eed5b992d8efb9255253967146629d494.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 161, + 167, + 449, + 190.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 190.0, + 449, + 213.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 161, + 213.0, + 449, + 236.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.5 + }, + { + "type": "table", + "bbox": [ + 157, + 290, + 454, + 360 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 256, + 504, + 289 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 255, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 270 + ], + "score": 1.0, + "content": "Table 3: Accuracy (and top–5 accuracy) on the Fill–In–The–Blank task, depending on which type", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "of GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 279, + 221, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 221, + 291 + ], + "score": 1.0, + "content": "as their code representation.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "table_body", + "bbox": [ + 157, + 290, + 454, + 360 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 157, + 290, + 454, + 360 + ], + "spans": [ + { + "bbox": [ + 157, + 290, + 454, + 360 + ], + "score": 0.972, + "html": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.80 (0.90) 0.97 (0.99)0.72 (0.84) 0.89 (0.95)0.80 (0.90) 0.95 (0.98)
Unseen reposClosed Vocab GSC0.59 (0.78) 0.92 (0.96)0.46 (0.68) 0.80 (0.89)0.62 (0.79) 0.88 (0.95)
", + "type": "table", + "image_path": "407aa777975ba994adc29c62d255d66c74058346b9d028a55509f55635c5f2cf.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 157, + 290, + 454, + 313.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 157, + 313.3333333333333, + 454, + 336.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 157, + 336.66666666666663, + 454, + 359.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 381, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 394 + ], + "score": 1.0, + "content": "name (in the form of the sequence of words that compose the name). An example instance from our", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 391, + 223, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 223, + 404 + ], + "score": 1.0, + "content": "dataset is shown in Figure 3.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "image", + "bbox": [ + 108, + 416, + 512, + 513 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 416, + 512, + 513 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 416, + 512, + 513 + ], + "spans": [ + { + "bbox": [ + 108, + 416, + 512, + 513 + ], + "score": 0.962, + "type": "image", + "image_path": "b21bc85d57d3da138b5330d252092f70b403cbab07c31d848435d51028931d7a.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 108, + 416, + 512, + 448.3333333333333 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 108, + 448.3333333333333, + 512, + 480.66666666666663 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 108, + 480.66666666666663, + 512, + 513.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 525, + 506, + 592 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Figure 3: Example of a model’s procedure for completing the Variable Naming task. Each Variable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 537, + 504, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 504, + 548 + ], + "score": 1.0, + "content": "Naming instance is created by replacing all uses of some variable (expectedLength, in this", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 208, + 560 + ], + "score": 1.0, + "content": "example) with a special", + "type": "text" + }, + { + "bbox": [ + 208, + 549, + 263, + 559 + ], + "score": 0.53, + "content": "\\mathrm { \\ < N A M E - M E > }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 548, + 506, + 560 + ], + "score": 1.0, + "content": "token. The model then processes the code as depicted in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Figure 1. To produce outputs, the model takes the mean of the nodes’ hidden states", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 570, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 506, + 583 + ], + "score": 1.0, + "content": "(depicted here in orange), uses them as the initial hidden state of a Recurrent Neural Network, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 581, + 350, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 350, + 594 + ], + "score": 1.0, + "content": "unrolls this RNN to produce a name as a sequence of words.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + } + ], + "index": 21.25 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "To produce a name from the output of the GNN, our models used the readout function of Allamanis", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 616, + 504, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 448, + 628 + ], + "score": 1.0, + "content": "et al. (2018). This readout function computes the mean of the hidden states of the", + "type": "text" + }, + { + "bbox": [ + 449, + 617, + 504, + 627 + ], + "score": 0.42, + "content": "\\mathrm { < N A M E - M E > }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "nodes and passing it as the initial hidden state to a 1–layer Gated Recurrent Unit (GRU) RNN (Cho", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "et al., 2014). This GRU is then unrolled to produce words in its predicted name, in the style of a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 408, + 662 + ], + "score": 1.0, + "content": "traditional NLP decoder. We used a fixed length unrolling of 8 words, as", + "type": "text" + }, + { + "bbox": [ + 409, + 649, + 436, + 659 + ], + "score": 0.86, + "content": "9 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "of names in our", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "score": 1.0, + "content": "training set were 8 or fewer words long. The models were trained by cross entropy loss over the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 671, + 235, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 235, + 684 + ], + "score": 1.0, + "content": "sequence of words in the name.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 302, + 700 + ], + "score": 1.0, + "content": "To decode each hidden state output of the GRU", + "type": "text" + }, + { + "bbox": [ + 302, + 688, + 310, + 698 + ], + "score": 0.75, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 687, + 433, + 700 + ], + "score": 1.0, + "content": "into a probability distribution", + "type": "text" + }, + { + "bbox": [ + 433, + 687, + 483, + 700 + ], + "score": 0.93, + "content": "P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "over", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 328, + 711 + ], + "score": 1.0, + "content": "words w, the Closed Vocab and CharCNN models pass", + "type": "text" + }, + { + "bbox": [ + 329, + 699, + 336, + 709 + ], + "score": 0.81, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "through a linear layer and a softmax layer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "with output dimension equal to the number of words in their closed vocabularies (i.e. a traditional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "decoder output for NLP). In contrast, the GSC model not only has access to a fixed–size vocabulary", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 167, + 449, + 236 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 88, + 506, + 166 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "score": 1.0, + "content": "Table 2: Accuracy on the Fill–In–The–Blank task. Our model is the AugAST–GSC. The first number", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "in each cell is the accuracy of the model, where a correct prediction is one in which the graph", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 112, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 123 + ], + "score": 1.0, + "content": "node that received the maximum attention weighting by the model contained the variable that was", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 281, + 135 + ], + "score": 1.0, + "content": "originally in the ", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "spot. The second, parenthetical numbers are the top–", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "5 accuracies, i.e. whether the correct node was among those that received the 5 largest attentions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "weightings from the model. See Table 1 for explanations of the abbreviations. All models use Gated", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 309, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 309, + 168 + ], + "score": 1.0, + "content": "Graph Neural Networks as their GNN component.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 161, + 167, + 449, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 161, + 167, + 449, + 236 + ], + "spans": [ + { + "bbox": [ + 161, + 167, + 449, + 236 + ], + "score": 0.976, + "html": "
Closed VocabCharCNNGSC
Seen reposAST0.57 (0.83)0.60 (0.84)0.89 (0.96)
AugAST0.80 (0.90)0.90 (0.94)0.97 (0.99)
Unseen reposAST AugAST0.36 (0.68)0.48 (0.80)0.80 (0.93)
0.59 (0.78)0.84 (0.92)0.92 (0.96)
", + "type": "table", + "image_path": "64e9ca199b864180f85246fd6d2fa82eed5b992d8efb9255253967146629d494.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 161, + 167, + 449, + 190.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 161, + 190.0, + 449, + 213.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 161, + 213.0, + 449, + 236.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.5 + }, + { + "type": "table", + "bbox": [ + 157, + 290, + 454, + 360 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 256, + 504, + 289 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 255, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 270 + ], + "score": 1.0, + "content": "Table 3: Accuracy (and top–5 accuracy) on the Fill–In–The–Blank task, depending on which type", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "of GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 279, + 221, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 221, + 291 + ], + "score": 1.0, + "content": "as their code representation.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "table_body", + "bbox": [ + 157, + 290, + 454, + 360 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 157, + 290, + 454, + 360 + ], + "spans": [ + { + "bbox": [ + 157, + 290, + 454, + 360 + ], + "score": 0.972, + "html": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.80 (0.90) 0.97 (0.99)0.72 (0.84) 0.89 (0.95)0.80 (0.90) 0.95 (0.98)
Unseen reposClosed Vocab GSC0.59 (0.78) 0.92 (0.96)0.46 (0.68) 0.80 (0.89)0.62 (0.79) 0.88 (0.95)
", + "type": "table", + "image_path": "407aa777975ba994adc29c62d255d66c74058346b9d028a55509f55635c5f2cf.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 157, + 290, + 454, + 313.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 157, + 313.3333333333333, + 454, + 336.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 157, + 336.66666666666663, + 454, + 359.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 381, + 505, + 403 + ], + "lines": [], + "index": 16.5, + "bbox_fs": [ + 105, + 379, + 506, + 404 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 108, + 416, + 512, + 513 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 416, + 512, + 513 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 416, + 512, + 513 + ], + "spans": [ + { + "bbox": [ + 108, + 416, + 512, + 513 + ], + "score": 0.962, + "type": "image", + "image_path": "b21bc85d57d3da138b5330d252092f70b403cbab07c31d848435d51028931d7a.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 108, + 416, + 512, + 448.3333333333333 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 108, + 448.3333333333333, + 512, + 480.66666666666663 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 108, + 480.66666666666663, + 512, + 513.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 525, + 506, + 592 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Figure 3: Example of a model’s procedure for completing the Variable Naming task. Each Variable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 537, + 504, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 504, + 548 + ], + "score": 1.0, + "content": "Naming instance is created by replacing all uses of some variable (expectedLength, in this", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 548, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 208, + 560 + ], + "score": 1.0, + "content": "example) with a special", + "type": "text" + }, + { + "bbox": [ + 208, + 549, + 263, + 559 + ], + "score": 0.53, + "content": "\\mathrm { \\ < N A M E - M E > }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 548, + 506, + 560 + ], + "score": 1.0, + "content": "token. The model then processes the code as depicted in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Figure 1. To produce outputs, the model takes the mean of the nodes’ hidden states", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 570, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 506, + 583 + ], + "score": 1.0, + "content": "(depicted here in orange), uses them as the initial hidden state of a Recurrent Neural Network, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 581, + 350, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 350, + 594 + ], + "score": 1.0, + "content": "unrolls this RNN to produce a name as a sequence of words.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + } + ], + "index": 21.25 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "To produce a name from the output of the GNN, our models used the readout function of Allamanis", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 616, + 504, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 448, + 628 + ], + "score": 1.0, + "content": "et al. (2018). This readout function computes the mean of the hidden states of the", + "type": "text" + }, + { + "bbox": [ + 449, + 617, + 504, + 627 + ], + "score": 0.42, + "content": "\\mathrm { < N A M E - M E > }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "nodes and passing it as the initial hidden state to a 1–layer Gated Recurrent Unit (GRU) RNN (Cho", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "et al., 2014). This GRU is then unrolled to produce words in its predicted name, in the style of a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 408, + 662 + ], + "score": 1.0, + "content": "traditional NLP decoder. We used a fixed length unrolling of 8 words, as", + "type": "text" + }, + { + "bbox": [ + 409, + 649, + 436, + 659 + ], + "score": 0.86, + "content": "9 9 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "of names in our", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "score": 1.0, + "content": "training set were 8 or fewer words long. The models were trained by cross entropy loss over the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 671, + 235, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 235, + 684 + ], + "score": 1.0, + "content": "sequence of words in the name.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 604, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 302, + 700 + ], + "score": 1.0, + "content": "To decode each hidden state output of the GRU", + "type": "text" + }, + { + "bbox": [ + 302, + 688, + 310, + 698 + ], + "score": 0.75, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 687, + 433, + 700 + ], + "score": 1.0, + "content": "into a probability distribution", + "type": "text" + }, + { + "bbox": [ + 433, + 687, + 483, + 700 + ], + "score": 0.93, + "content": "P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "over", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 328, + 711 + ], + "score": 1.0, + "content": "words w, the Closed Vocab and CharCNN models pass", + "type": "text" + }, + { + "bbox": [ + 329, + 699, + 336, + 709 + ], + "score": 0.81, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "through a linear layer and a softmax layer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "with output dimension equal to the number of words in their closed vocabularies (i.e. a traditional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "decoder output for NLP). In contrast, the GSC model not only has access to a fixed–size vocabulary", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "but can also produce words by pointing to cache nodes in its Graph–Structured Cache. Specifically,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "it uses a decoder architecture inspired by the Pointer Sentinel Mixture Model of Merity et al. (2017):", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "the probability of a word w being the GSC decoder’s output given that the GRU’s hidden state was", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 113, + 127, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 114, + 125 + ], + "score": 0.79, + "content": "^ { h }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 115, + 113, + 127, + 128 + ], + "score": 1.0, + "content": "is", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 687, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "but can also produce words by pointing to cache nodes in its Graph–Structured Cache. Specifically,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "it uses a decoder architecture inspired by the Pointer Sentinel Mixture Model of Merity et al. (2017):", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "the probability of a word w being the GSC decoder’s output given that the GRU’s hidden state was", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 113, + 127, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 114, + 125 + ], + "score": 0.79, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 113, + 127, + 128 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 126, + 442, + 140 + ], + "lines": [ + { + "bbox": [ + 169, + 126, + 442, + 140 + ], + "spans": [ + { + "bbox": [ + 169, + 126, + 442, + 140 + ], + "score": 0.83, + "content": "P ( \\mathbf { w } | h ) = P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h ) + ( 1 - P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) ) P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )", + "type": "interline_equation", + "image_path": "e033b845db19ed064700348954e36074ceecf2517c8ee2467099c2046acd4d6e.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 169, + 126, + 442, + 140 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 135, + 157 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 144, + 178, + 156 + ], + "score": 0.94, + "content": "P _ { \\mathrm { g r a p h } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "is a conditional probability distribution over cache nodes in the GSC and the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 165, + 168 + ], + "score": 1.0, + "content": "sentinel s, and", + "type": "text" + }, + { + "bbox": [ + 166, + 155, + 209, + 167 + ], + "score": 0.92, + "content": "P _ { \\mathrm { v o c a b } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "is a conditional probability distribution over words in a closed vocabulary.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 164, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 107, + 165, + 150, + 178 + ], + "score": 0.91, + "content": "P _ { \\mathrm { g r a p h } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 164, + 506, + 181 + ], + "score": 1.0, + "content": "is computed by passing the hidden states of all cache nodes and the sentinel node through", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 493, + 189 + ], + "score": 1.0, + "content": "a single linear layer and then computing the softmax dot–product attention of these values with", + "type": "text" + }, + { + "bbox": [ + 493, + 177, + 501, + 186 + ], + "score": 0.76, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 176, + 505, + 189 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 187, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 107, + 187, + 151, + 200 + ], + "score": 0.91, + "content": "P _ { \\mathrm { v o c a b } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 187, + 352, + 201 + ], + "score": 1.0, + "content": "is computed as the softmax of a linear mapping of", + "type": "text" + }, + { + "bbox": [ + 352, + 188, + 360, + 198 + ], + "score": 0.81, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 187, + 506, + 201 + ], + "score": 1.0, + "content": "to indices in a closed vocabulary, as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 104, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "in the Closed Vocab and CharCNN models. If there is no cache node for w in the Augmented AST", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 309, + 223 + ], + "score": 1.0, + "content": "or if w is not in the model’s closed dictionary then", + "type": "text" + }, + { + "bbox": [ + 310, + 209, + 359, + 222 + ], + "score": 0.92, + "content": "P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 209, + 377, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 209, + 427, + 221 + ], + "score": 0.93, + "content": "P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 209, + 505, + 223 + ], + "score": 1.0, + "content": "are 0, respectively.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "The performance of our GSC versus other methods is reported in Table 4. More granular perfor-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "mance statistics are reported in Appendix Table 6. We also compare the performance of different", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 180, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 180, + 259 + ], + "score": 1.0, + "content": "GNNs in Table 5.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 123, + 344, + 488, + 414 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 277, + 505, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "Table 4: Accuracy on the Variable Naming task. Our model is the AugAST–GSC. The first number", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 300 + ], + "score": 1.0, + "content": "in each cell is the accuracy of the model, where we consider a correct output to be exact reproduction", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "of the full name of the obfuscated variable (i.e. all the words in the name and then a EOS token).", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "score": 1.0, + "content": "The second, parenthetical numbers are the top–5 accuracies, i.e. whether the correct full name was", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "among the 5 most probable sequences output by the model. See Table 1 for explanations of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 457, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 457, + 345 + ], + "score": 1.0, + "content": "abbreviations. All models use Gated Graph Neural Networks as their GNN component.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "table_body", + "bbox": [ + 123, + 344, + 488, + 414 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 344, + 488, + 414 + ], + "spans": [ + { + "bbox": [ + 123, + 344, + 488, + 414 + ], + "score": 0.978, + "html": "
Closed VocabCharCNNPointer SentinelGSC
Seen reposAST0.23 (0.31)0.22 (0.28)0.19 (0.33)0.49 (0.67)
AugAST0.19 (0.26)0.20 (0.27)0.26 (0.40)0.53 (0.69)
Unseen reposAST0.05 (0.07)0.06 (0.09)0.06 (0.11)0.38 (0.53)
AugAST0.04 (0.07)0.06 (0.08)0.08 (0.14)0.41 (0.57)
", + "type": "table", + "image_path": "5af4ae8799d6bb4f9a24558aaf1823daccb2849b869291d358fa6d0540c34964.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 123, + 344, + 488, + 367.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 123, + 367.3333333333333, + 488, + 390.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 123, + 390.66666666666663, + 488, + 413.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 19.75 + }, + { + "type": "table", + "bbox": [ + 156, + 474, + 454, + 544 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 440, + 505, + 473 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "Table 5: Accuracy (and top–5 accuracy) on the Variable Naming task, depending on which type of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 462, + 210, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 210, + 474 + ], + "score": 1.0, + "content": "their code representation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "table_body", + "bbox": [ + 156, + 474, + 454, + 544 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 156, + 474, + 454, + 544 + ], + "spans": [ + { + "bbox": [ + 156, + 474, + 454, + 544 + ], + "score": 0.977, + "html": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.19 (0.26)0.23 (0.31)0.27 (0.34)
0.53 (0.69)0.33 (0.48)0.46 (0.63)
Unseen reposClosed Vocab GSC0.04 (0.07) 0.41 (0.57)0.06 (0.08) 0.25 (0.40)0.06 (0.09) 0.35 (0.49)
", + "type": "table", + "image_path": "373973edcaadd5d2a832f2390824e23ed065f27ac06f83ad1f85a9edf60c7002.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 156, + 474, + 454, + 497.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 156, + 497.3333333333333, + 454, + 520.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 156, + 520.6666666666666, + 454, + 544.0 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 107, + 568, + 190, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 192, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 192, + 583 + ], + "score": 1.0, + "content": "6 DISCUSSION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 594, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "As can be seen in Tables 2 and 4, the addition of a GSC improved performance on all tasks. Our full", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "model, the AugAST–GSC model, outperforms the other models tested and does comparatively well", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 616, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 481, + 630 + ], + "score": 1.0, + "content": "at maintaining accuracy between the seen and unseen test repos on the Variable Naming task.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "To some degree the improved performance from adding the GSC is unsurprising: its addition to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "a graph–based model is essentially just adding extra features and doesn’t remove any information", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "or flexibility. Under a satisfactory training regime, a model could simply learn to ignore it if it is", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 664, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 680 + ], + "score": 1.0, + "content": "unhelpful, so its inclusion should never hurt performance. The degree to which it helps, though,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "especially on the Variable Naming task, suggests that a GSC is well worth using for some tasks.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "score": 1.0, + "content": "Moreover, the fact that the Pointer Sentinel approach shown in Table 4 performs noticeably less well", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "than the full GSC approach suggests that the relational aspect of the GSC is key: simply having the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "ability to output out–of–vocabulary words without relational information about their usage appears", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 204, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 204, + 732 + ], + "score": 1.0, + "content": "to be much less helpful.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 505, + 128 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 126, + 442, + 140 + ], + "lines": [ + { + "bbox": [ + 169, + 126, + 442, + 140 + ], + "spans": [ + { + "bbox": [ + 169, + 126, + 442, + 140 + ], + "score": 0.83, + "content": "P ( \\mathbf { w } | h ) = P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h ) + ( 1 - P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) ) P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )", + "type": "interline_equation", + "image_path": "e033b845db19ed064700348954e36074ceecf2517c8ee2467099c2046acd4d6e.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 169, + 126, + 442, + 140 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 135, + 157 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 144, + 178, + 156 + ], + "score": 0.94, + "content": "P _ { \\mathrm { g r a p h } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "is a conditional probability distribution over cache nodes in the GSC and the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 165, + 168 + ], + "score": 1.0, + "content": "sentinel s, and", + "type": "text" + }, + { + "bbox": [ + 166, + 155, + 209, + 167 + ], + "score": 0.92, + "content": "P _ { \\mathrm { v o c a b } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "is a conditional probability distribution over words in a closed vocabulary.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 164, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 107, + 165, + 150, + 178 + ], + "score": 0.91, + "content": "P _ { \\mathrm { g r a p h } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 164, + 506, + 181 + ], + "score": 1.0, + "content": "is computed by passing the hidden states of all cache nodes and the sentinel node through", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 176, + 493, + 189 + ], + "score": 1.0, + "content": "a single linear layer and then computing the softmax dot–product attention of these values with", + "type": "text" + }, + { + "bbox": [ + 493, + 177, + 501, + 186 + ], + "score": 0.76, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 176, + 505, + 189 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 187, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 107, + 187, + 151, + 200 + ], + "score": 0.91, + "content": "P _ { \\mathrm { v o c a b } } ( \\cdot | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 187, + 352, + 201 + ], + "score": 1.0, + "content": "is computed as the softmax of a linear mapping of", + "type": "text" + }, + { + "bbox": [ + 352, + 188, + 360, + 198 + ], + "score": 0.81, + "content": "^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 187, + 506, + 201 + ], + "score": 1.0, + "content": "to indices in a closed vocabulary, as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 104, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "in the Closed Vocab and CharCNN models. If there is no cache node for w in the Augmented AST", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 309, + 223 + ], + "score": 1.0, + "content": "or if w is not in the model’s closed dictionary then", + "type": "text" + }, + { + "bbox": [ + 310, + 209, + 359, + 222 + ], + "score": 0.92, + "content": "P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 209, + 377, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 378, + 209, + 427, + 221 + ], + "score": 0.93, + "content": "P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 209, + 505, + 223 + ], + "score": 1.0, + "content": "are 0, respectively.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 143, + 506, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "The performance of our GSC versus other methods is reported in Table 4. More granular perfor-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "mance statistics are reported in Appendix Table 6. We also compare the performance of different", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 180, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 180, + 259 + ], + "score": 1.0, + "content": "GNNs in Table 5.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 225, + 505, + 259 + ] + }, + { + "type": "table", + "bbox": [ + 123, + 344, + 488, + 414 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 277, + 505, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "Table 4: Accuracy on the Variable Naming task. Our model is the AugAST–GSC. The first number", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 300 + ], + "score": 1.0, + "content": "in each cell is the accuracy of the model, where we consider a correct output to be exact reproduction", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "of the full name of the obfuscated variable (i.e. all the words in the name and then a EOS token).", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "score": 1.0, + "content": "The second, parenthetical numbers are the top–5 accuracies, i.e. whether the correct full name was", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "among the 5 most probable sequences output by the model. See Table 1 for explanations of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 457, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 457, + 345 + ], + "score": 1.0, + "content": "abbreviations. All models use Gated Graph Neural Networks as their GNN component.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "table_body", + "bbox": [ + 123, + 344, + 488, + 414 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 344, + 488, + 414 + ], + "spans": [ + { + "bbox": [ + 123, + 344, + 488, + 414 + ], + "score": 0.978, + "html": "
Closed VocabCharCNNPointer SentinelGSC
Seen reposAST0.23 (0.31)0.22 (0.28)0.19 (0.33)0.49 (0.67)
AugAST0.19 (0.26)0.20 (0.27)0.26 (0.40)0.53 (0.69)
Unseen reposAST0.05 (0.07)0.06 (0.09)0.06 (0.11)0.38 (0.53)
AugAST0.04 (0.07)0.06 (0.08)0.08 (0.14)0.41 (0.57)
", + "type": "table", + "image_path": "5af4ae8799d6bb4f9a24558aaf1823daccb2849b869291d358fa6d0540c34964.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 123, + 344, + 488, + 367.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 123, + 367.3333333333333, + 488, + 390.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 123, + 390.66666666666663, + 488, + 413.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 19.75 + }, + { + "type": "table", + "bbox": [ + 156, + 474, + 454, + 544 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 440, + 505, + 473 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "Table 5: Accuracy (and top–5 accuracy) on the Variable Naming task, depending on which type of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "GNN the model uses. See Table 1 for explanations of the abbreviations. All models use AugAST as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 462, + 210, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 210, + 474 + ], + "score": 1.0, + "content": "their code representation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "table_body", + "bbox": [ + 156, + 474, + 454, + 544 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 156, + 474, + 454, + 544 + ], + "spans": [ + { + "bbox": [ + 156, + 474, + 454, + 544 + ], + "score": 0.977, + "html": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.19 (0.26)0.23 (0.31)0.27 (0.34)
0.53 (0.69)0.33 (0.48)0.46 (0.63)
Unseen reposClosed Vocab GSC0.04 (0.07) 0.41 (0.57)0.06 (0.08) 0.25 (0.40)0.06 (0.09) 0.35 (0.49)
", + "type": "table", + "image_path": "373973edcaadd5d2a832f2390824e23ed065f27ac06f83ad1f85a9edf60c7002.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 156, + 474, + 454, + 497.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 156, + 497.3333333333333, + 454, + 520.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 156, + 520.6666666666666, + 454, + 544.0 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 107, + 568, + 190, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 192, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 192, + 583 + ], + "score": 1.0, + "content": "6 DISCUSSION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 594, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "As can be seen in Tables 2 and 4, the addition of a GSC improved performance on all tasks. Our full", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "model, the AugAST–GSC model, outperforms the other models tested and does comparatively well", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 616, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 481, + 630 + ], + "score": 1.0, + "content": "at maintaining accuracy between the seen and unseen test repos on the Variable Naming task.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 593, + 506, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "To some degree the improved performance from adding the GSC is unsurprising: its addition to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "a graph–based model is essentially just adding extra features and doesn’t remove any information", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "or flexibility. Under a satisfactory training regime, a model could simply learn to ignore it if it is", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 664, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 680 + ], + "score": 1.0, + "content": "unhelpful, so its inclusion should never hurt performance. The degree to which it helps, though,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "especially on the Variable Naming task, suggests that a GSC is well worth using for some tasks.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 700 + ], + "score": 1.0, + "content": "Moreover, the fact that the Pointer Sentinel approach shown in Table 4 performs noticeably less well", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "than the full GSC approach suggests that the relational aspect of the GSC is key: simply having the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "ability to output out–of–vocabulary words without relational information about their usage appears", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 204, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 204, + 732 + ], + "score": 1.0, + "content": "to be much less helpful.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 632, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 484, + 95 + ], + "score": 1.0, + "content": "The downsides of using a GSC are thus primarily computational. Our GSC models ran about", + "type": "text" + }, + { + "bbox": [ + 484, + 82, + 504, + 93 + ], + "score": 0.84, + "content": "30 \\%", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "slower than the Closed Vocab models. Since we capped the graph size at 500 nodes, the slowdown", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "is presumably due to the large number of edges to and from the graph cache nodes. Better support", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "score": 1.0, + "content": "for sparse operations on GPU in deep learning frameworks would be useful for alleviating this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 149, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 149, + 138 + ], + "score": 1.0, + "content": "downside.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "In the near term, there remain a number of design choices to explore regarding AST– and GNN–", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "models for processing source code. Adding information about word order to the GSC might improve", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "performance, as might constructing the vocabulary out of subwords rather than words. It also might", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "help to treat variable types as the GSC treats words: storing them in a GSC and connecting them", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "with edges to the variables of those types; this could be particularly useful when working with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "score": 1.0, + "content": "code snippets rather than fully compilable code. For the Variable Naming task, there are also many", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "architecture choices to be explored in how to produce a sequence of words for a name: how to unroll", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 219, + 316, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 316, + 232 + ], + "score": 1.0, + "content": "the RNN, what to use as the initial hidden state, etc.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "In the longer term, given that all results above show that augmenting ASTs with data– and control–", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "flow edges improves performance, it would be worth exploring other static analysis concepts from", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 504, + 271 + ], + "score": 1.0, + "content": "the Programming Language and Software Verification literatures and seeing whether they could be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "usefully incorporated into Augmented ASTs. Better understanding of how Graph Neural Networks", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "learn is also crucial, since they are central to the performance of our model and many others. Ad-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "score": 1.0, + "content": "ditionally, the entire domain of machine learning on source code faces the practical issue that many", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "of the best data for supervised learning on source code — things like high–quality code reviews,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 312, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 104, + 312, + 505, + 328 + ], + "score": 1.0, + "content": "integration test results, code with high test coverage, etc. — are not available outside private orga-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 148, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 148, + 338 + ], + "score": 1.0, + "content": "nizations.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 353, + 175, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 176, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 176, + 366 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 365, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Earl T. Barr, Premkumar Devanbu, and Charles Sutton. A survey of machine", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 382, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 504, + 396 + ], + "score": 1.0, + "content": "learning for big code and naturalness. arXiv:1709.06182 [cs], 2017. URL https://arxiv.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 394, + 229, + 406 + ], + "spans": [ + { + "bbox": [ + 117, + 394, + 229, + 406 + ], + "score": 1.0, + "content": "org/abs/1709.06182.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 412, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 428 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "with graphs. In International Conference on Learning Representations, 2018. URL https:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 436, + 331, + 449 + ], + "spans": [ + { + "bbox": [ + 117, + 436, + 266, + 449 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 437, + 273, + 445 + ], + "score": 0.54, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 436, + 331, + 449 + ], + "score": 1.0, + "content": "BJOFETxR-.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 454, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 469 + ], + "score": 1.0, + "content": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 465, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 114, + 465, + 506, + 480 + ], + "score": 1.0, + "content": "for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Conference on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 115, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New York, NY,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 486, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 486, + 505, + 502 + ], + "score": 1.0, + "content": "USA, 2018. ACM. ISBN 978-1-4503-5698-5. doi: 10.1145/3192366.3192412. URL http:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 500, + 343, + 511 + ], + "spans": [ + { + "bbox": [ + 117, + 500, + 343, + 511 + ], + "score": 1.0, + "content": "//doi.acm.org/10.1145/3192366.3192412.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "Matthew Amodio, Swarat Chaudhuri, and Thomas Reps. Neural Attribute Machines for Program", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 530, + 504, + 542 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 504, + 542 + ], + "score": 1.0, + "content": "Generation. arXiv:1705.09231 [cs], May 2017. URL http://arxiv.org/abs/1705.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 540, + 152, + 553 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 152, + 553 + ], + "score": 1.0, + "content": "09231.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 574 + ], + "score": 1.0, + "content": "Forough Arabshahi, Sameer Singh, and Animashree Anandkumar. Combining symbolic expres-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 115, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "sions and black-box function evaluations for training neural programs. In International Confer-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 116, + 583, + 497, + 595 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 497, + 584, + 505, + 592 + ], + "score": 0.31, + "content": "=", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 594, + 175, + 606 + ], + "spans": [ + { + "bbox": [ + 116, + 594, + 175, + 606 + ], + "score": 1.0, + "content": "Hksj2WWAW.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "Trapit Bansal, Arvind Neelakantan, and Andrew McCallum. Relnet: End-to-end modeling of en-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 623, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 115, + 623, + 504, + 636 + ], + "score": 1.0, + "content": "tities & relations. arXiv:1706.07179 [cs], 2017. URL http://arxiv.org/abs/1706.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 635, + 151, + 646 + ], + "spans": [ + { + "bbox": [ + 116, + 635, + 151, + 646 + ], + "score": 1.0, + "content": "07179.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vin´ıcius Flores", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 115, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 116, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "aglar Gulehre, Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, ¨", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Pushmeet Kohli, Matthew Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "inductive biases, deep learning, and graph networks. abs/1806.01261, 2018. URL https:", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 117, + 721, + 277, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 277, + 732 + ], + "score": 1.0, + "content": "//arxiv.org/abs/1806.01261.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 36.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 484, + 95 + ], + "score": 1.0, + "content": "The downsides of using a GSC are thus primarily computational. Our GSC models ran about", + "type": "text" + }, + { + "bbox": [ + 484, + 82, + 504, + 93 + ], + "score": 0.84, + "content": "30 \\%", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "slower than the Closed Vocab models. Since we capped the graph size at 500 nodes, the slowdown", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "is presumably due to the large number of edges to and from the graph cache nodes. Better support", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 129 + ], + "score": 1.0, + "content": "for sparse operations on GPU in deep learning frameworks would be useful for alleviating this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 149, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 149, + 138 + ], + "score": 1.0, + "content": "downside.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "In the near term, there remain a number of design choices to explore regarding AST– and GNN–", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "models for processing source code. Adding information about word order to the GSC might improve", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "performance, as might constructing the vocabulary out of subwords rather than words. It also might", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "help to treat variable types as the GSC treats words: storing them in a GSC and connecting them", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "with edges to the variables of those types; this could be particularly useful when working with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "score": 1.0, + "content": "code snippets rather than fully compilable code. For the Variable Naming task, there are also many", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "architecture choices to be explored in how to produce a sequence of words for a name: how to unroll", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 219, + 316, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 316, + 232 + ], + "score": 1.0, + "content": "the RNN, what to use as the initial hidden state, etc.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 143, + 506, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "In the longer term, given that all results above show that augmenting ASTs with data– and control–", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "flow edges improves performance, it would be worth exploring other static analysis concepts from", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 504, + 271 + ], + "score": 1.0, + "content": "the Programming Language and Software Verification literatures and seeing whether they could be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "usefully incorporated into Augmented ASTs. Better understanding of how Graph Neural Networks", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "learn is also crucial, since they are central to the performance of our model and many others. Ad-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "score": 1.0, + "content": "ditionally, the entire domain of machine learning on source code faces the practical issue that many", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "of the best data for supervised learning on source code — things like high–quality code reviews,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 312, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 104, + 312, + 505, + 328 + ], + "score": 1.0, + "content": "integration test results, code with high test coverage, etc. — are not available outside private orga-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 148, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 148, + 338 + ], + "score": 1.0, + "content": "nizations.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 236, + 506, + 338 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 353, + 175, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 176, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 176, + 366 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "list", + "bbox": [ + 105, + 365, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Earl T. Barr, Premkumar Devanbu, and Charles Sutton. A survey of machine", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 382, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 504, + 396 + ], + "score": 1.0, + "content": "learning for big code and naturalness. arXiv:1709.06182 [cs], 2017. URL https://arxiv.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 394, + 229, + 406 + ], + "spans": [ + { + "bbox": [ + 117, + 394, + 229, + 406 + ], + "score": 1.0, + "content": "org/abs/1709.06182.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 412, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 428 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "with graphs. In International Conference on Learning Representations, 2018. URL https:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 436, + 331, + 449 + ], + "spans": [ + { + "bbox": [ + 117, + 436, + 266, + 449 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 437, + 273, + 445 + ], + "score": 0.54, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 436, + 331, + 449 + ], + "score": 1.0, + "content": "BJOFETxR-.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 454, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 469 + ], + "score": 1.0, + "content": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representation", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 465, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 114, + 465, + 506, + 480 + ], + "score": 1.0, + "content": "for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Conference on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 115, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New York, NY,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 486, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 486, + 505, + 502 + ], + "score": 1.0, + "content": "USA, 2018. ACM. ISBN 978-1-4503-5698-5. doi: 10.1145/3192366.3192412. URL http:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 500, + 343, + 511 + ], + "spans": [ + { + "bbox": [ + 117, + 500, + 343, + 511 + ], + "score": 1.0, + "content": "//doi.acm.org/10.1145/3192366.3192412.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "Matthew Amodio, Swarat Chaudhuri, and Thomas Reps. Neural Attribute Machines for Program", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 530, + 504, + 542 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 504, + 542 + ], + "score": 1.0, + "content": "Generation. arXiv:1705.09231 [cs], May 2017. URL http://arxiv.org/abs/1705.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 540, + 152, + 553 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 152, + 553 + ], + "score": 1.0, + "content": "09231.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 559, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 574 + ], + "score": 1.0, + "content": "Forough Arabshahi, Sameer Singh, and Animashree Anandkumar. Combining symbolic expres-", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 115, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "sions and black-box function evaluations for training neural programs. In International Confer-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 116, + 583, + 497, + 595 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 497, + 584, + 505, + 592 + ], + "score": 0.31, + "content": "=", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 594, + 175, + 606 + ], + "spans": [ + { + "bbox": [ + 116, + 594, + 175, + 606 + ], + "score": 1.0, + "content": "Hksj2WWAW.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "Trapit Bansal, Arvind Neelakantan, and Andrew McCallum. Relnet: End-to-end modeling of en-", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 623, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 115, + 623, + 504, + 636 + ], + "score": 1.0, + "content": "tities & relations. arXiv:1706.07179 [cs], 2017. URL http://arxiv.org/abs/1706.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 635, + 151, + 646 + ], + "spans": [ + { + "bbox": [ + 116, + 635, + 151, + 646 + ], + "score": 1.0, + "content": "07179.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vin´ıcius Flores", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 115, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 116, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "aglar Gulehre, Francis Song, Andrew J. Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, ¨", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "Kelsey R. Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Pushmeet Kohli, Matthew Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. Relational", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "inductive biases, deep learning, and graph networks. abs/1806.01261, 2018. URL https:", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 117, + 721, + 277, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 277, + 732 + ], + "score": 1.0, + "content": "//arxiv.org/abs/1806.01261.", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 371, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Avishkar Bhoopchand, Tim Rocktaschel, Earl T. Barr, and Sebastian Riedel. Learning python code ¨", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "suggestion with a sparse pointer network. abs/1611.08307, 2016. URL https://arxiv.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 229, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 229, + 116 + ], + "score": 1.0, + "content": "org/abs/1611.08307.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 138 + ], + "score": 1.0, + "content": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 136, + 327, + 147 + ], + "spans": [ + { + "bbox": [ + 117, + 136, + 327, + 147 + ], + "score": 1.0, + "content": "with subword information. TACL, 5:135–146, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "score": 1.0, + "content": "Marc Brockschmidt, Miltiadis Allamanis, Alexander L. Gaunt, and Oleksandr Polozov. Genera-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 167, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 504, + 180 + ], + "score": 1.0, + "content": "tive code modeling with graphs. abs/1805.08490, 2018. URL https://arxiv.org/abs/", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 178, + 182, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 178, + 182, + 189 + ], + "score": 1.0, + "content": "1805.08490.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "score": 1.0, + "content": "Kwonsoo Chae, Hakjoo Oh, Kihong Heo, and Hongseok Yang. Automatically generating features", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "for learning program analysis heuristics for c-like languages. Proc. ACM Program. Lang., 1", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "(OOPSLA):101:1–101:25, October 2017. ISSN 2475-1421. doi: 10.1145/3133925. URL http:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 232, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 117, + 232, + 295, + 244 + ], + "score": 1.0, + "content": "//doi.acm.org/10.1145/3133925.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 105, + 252, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 266 + ], + "score": 1.0, + "content": "Xinyun Chen, Chang Liu, and Dawn Song. Tree-to-tree neural networks for program translation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 262, + 415, + 275 + ], + "spans": [ + { + "bbox": [ + 116, + 262, + 350, + 275 + ], + "score": 1.0, + "content": "2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 351, + 265, + 357, + 272 + ], + "score": 0.62, + "content": "{ . } = { }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 262, + 415, + 275 + ], + "score": 1.0, + "content": "rkxY-sl0W.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder–decoder", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "in Natural Language Processing (EMNLP), pp. 1724–1734. Association for Computational Lin-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 326, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 505, + 342 + ], + "score": 1.0, + "content": "guistics, 2014. doi: 10.3115/v1/D14-1179. URL http://www.aclweb.org/anthology/", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 339, + 170, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 170, + 350 + ], + "score": 1.0, + "content": "D14-1179.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 369, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 115, + 369, + 505, + 384 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In Doina Precup and Yee Whye Teh (eds.), Proceedings", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 117, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 504, + 405 + ], + "score": 1.0, + "content": "Learning Research, pp. 1263–1272, International Convention Centre, Sydney, Australia, 06–11", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 403, + 503, + 416 + ], + "spans": [ + { + "bbox": [ + 116, + 403, + 503, + 416 + ], + "score": 1.0, + "content": "Aug 2017. PMLR. URL http://proceedings.mlr.press/v70/gilmer17a.html.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 423, + 500, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 500, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 500, + 437 + ], + "score": 1.0, + "content": "Github. The State of the Octoverse 2017, 2017. URL https://octoverse.github.com/.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "C. Goller and A. Kuchler. Learning task-dependent distributed representations by backpropagation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 453, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 505, + 469 + ], + "score": 1.0, + "content": "through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pp.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 466, + 359, + 477 + ], + "spans": [ + { + "bbox": [ + 117, + 466, + 359, + 477 + ], + "score": 1.0, + "content": "347–352 vol.1, Jun 1996. doi: 10.1109/ICNN.1996.548916.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 504, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 500 + ], + "score": 1.0, + "content": "Edouard Grave, Moustapha Cisse, and Armand Joulin. Unbounded cache model for online language ´", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 497, + 309, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 497, + 309, + 509 + ], + "score": 1.0, + "content": "modeling with open vocabulary. In NIPS, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 532 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computa-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 115, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "tional Linguistics (Volume 1: Long Papers), pp. 140–149. Association for Computational Linguis-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 549, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 549, + 504, + 564 + ], + "score": 1.0, + "content": "tics, 2016. doi: 10.18653/v1/P16-1014. URL http://www.aclweb.org/anthology/", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 562, + 170, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 562, + 170, + 573 + ], + "score": 1.0, + "content": "P16-1014.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 504, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "Vincent J. Hellendoorn and Premkumar T. Devanbu. Are deep neural networks the best choice for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 593, + 339, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 593, + 339, + 605 + ], + "score": 1.0, + "content": "modeling source code? In ESEC/SIGSOFT FSE, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 504, + 636 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 628 + ], + "score": 1.0, + "content": "Daniel D. Johnson. Learning Graphical State Transitions. In International Conference on Learning", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 624, + 483, + 637 + ], + "spans": [ + { + "bbox": [ + 116, + 624, + 419, + 637 + ], + "score": 1.0, + "content": "Representations, 2017. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 419, + 626, + 426, + 634 + ], + "score": 0.5, + "content": "{ . } =", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 624, + 483, + 637 + ], + "score": 1.0, + "content": "HJ0NvFzxl.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 659 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 116, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 666, + 497, + 680 + ], + "score": 1.0, + "content": "pp. 2741–2749. AAAI Press, 2016. URL http://dl.acm.org/citation.cfm?id", + "type": "text" + }, + { + "bbox": [ + 497, + 668, + 505, + 677 + ], + "score": 0.59, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 677, + 212, + 689 + ], + "spans": [ + { + "bbox": [ + 116, + 677, + 212, + 689 + ], + "score": 1.0, + "content": "3016100.3016285.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 709, + 503, + 721 + ], + "spans": [ + { + "bbox": [ + 116, + 709, + 503, + 721 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2015. URL https://arxiv.org/abs/1412.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 146, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 146, + 732 + ], + "score": 1.0, + "content": "6980.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Avishkar Bhoopchand, Tim Rocktaschel, Earl T. Barr, and Sebastian Riedel. Learning python code ¨", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "suggestion with a sparse pointer network. abs/1611.08307, 2016. URL https://arxiv.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 229, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 229, + 116 + ], + "score": 1.0, + "content": "org/abs/1611.08307.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 503, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 138 + ], + "score": 1.0, + "content": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 136, + 327, + 147 + ], + "spans": [ + { + "bbox": [ + 117, + 136, + 327, + 147 + ], + "score": 1.0, + "content": "with subword information. TACL, 5:135–146, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 123, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 504, + 168 + ], + "score": 1.0, + "content": "Marc Brockschmidt, Miltiadis Allamanis, Alexander L. Gaunt, and Oleksandr Polozov. Genera-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 167, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 504, + 180 + ], + "score": 1.0, + "content": "tive code modeling with graphs. abs/1805.08490, 2018. URL https://arxiv.org/abs/", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 178, + 182, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 178, + 182, + 189 + ], + "score": 1.0, + "content": "1805.08490.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 155, + 504, + 189 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "score": 1.0, + "content": "Kwonsoo Chae, Hakjoo Oh, Kihong Heo, and Hongseok Yang. Automatically generating features", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "for learning program analysis heuristics for c-like languages. Proc. ACM Program. Lang., 1", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "(OOPSLA):101:1–101:25, October 2017. ISSN 2475-1421. doi: 10.1145/3133925. URL http:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 232, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 117, + 232, + 295, + 244 + ], + "score": 1.0, + "content": "//doi.acm.org/10.1145/3133925.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 198, + 506, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 252, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 266 + ], + "score": 1.0, + "content": "Xinyun Chen, Chang Liu, and Dawn Song. Tree-to-tree neural networks for program translation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 262, + 415, + 275 + ], + "spans": [ + { + "bbox": [ + 116, + 262, + 350, + 275 + ], + "score": 1.0, + "content": "2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 351, + 265, + 357, + 272 + ], + "score": 0.62, + "content": "{ . } = { }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 262, + 415, + 275 + ], + "score": 1.0, + "content": "rkxY-sl0W.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 251, + 505, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder–decoder", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 316, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 505, + 329 + ], + "score": 1.0, + "content": "in Natural Language Processing (EMNLP), pp. 1724–1734. Association for Computational Lin-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 326, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 505, + 342 + ], + "score": 1.0, + "content": "guistics, 2014. doi: 10.3115/v1/D14-1179. URL http://www.aclweb.org/anthology/", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 339, + 170, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 170, + 350 + ], + "score": 1.0, + "content": "D14-1179.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 284, + 506, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 369, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 115, + 369, + 505, + 384 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In Doina Precup and Yee Whye Teh (eds.), Proceedings", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 117, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 504, + 405 + ], + "score": 1.0, + "content": "Learning Research, pp. 1263–1272, International Convention Centre, Sydney, Australia, 06–11", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 403, + 503, + 416 + ], + "spans": [ + { + "bbox": [ + 116, + 403, + 503, + 416 + ], + "score": 1.0, + "content": "Aug 2017. PMLR. URL http://proceedings.mlr.press/v70/gilmer17a.html.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 359, + 505, + 416 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 423, + 500, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 500, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 500, + 437 + ], + "score": 1.0, + "content": "Github. The State of the Octoverse 2017, 2017. URL https://octoverse.github.com/.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 421, + 500, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "C. Goller and A. Kuchler. Learning task-dependent distributed representations by backpropagation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 453, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 453, + 505, + 469 + ], + "score": 1.0, + "content": "through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1, pp.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 466, + 359, + 477 + ], + "spans": [ + { + "bbox": [ + 117, + 466, + 359, + 477 + ], + "score": 1.0, + "content": "347–352 vol.1, Jun 1996. doi: 10.1109/ICNN.1996.548916.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 443, + 505, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 504, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 500 + ], + "score": 1.0, + "content": "Edouard Grave, Moustapha Cisse, and Armand Joulin. Unbounded cache model for online language ´", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 497, + 309, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 497, + 309, + 509 + ], + "score": 1.0, + "content": "modeling with open vocabulary. In NIPS, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 484, + 506, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 532 + ], + "score": 1.0, + "content": "Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. Pointing", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computa-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 115, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "tional Linguistics (Volume 1: Long Papers), pp. 140–149. Association for Computational Linguis-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 549, + 504, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 549, + 504, + 564 + ], + "score": 1.0, + "content": "tics, 2016. doi: 10.18653/v1/P16-1014. URL http://www.aclweb.org/anthology/", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 562, + 170, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 562, + 170, + 573 + ], + "score": 1.0, + "content": "P16-1014.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 516, + 506, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 504, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "Vincent J. Hellendoorn and Premkumar T. Devanbu. Are deep neural networks the best choice for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 593, + 339, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 593, + 339, + 605 + ], + "score": 1.0, + "content": "modeling source code? In ESEC/SIGSOFT FSE, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 581, + 505, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 504, + 636 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 628 + ], + "score": 1.0, + "content": "Daniel D. Johnson. Learning Graphical State Transitions. In International Conference on Learning", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 624, + 483, + 637 + ], + "spans": [ + { + "bbox": [ + 116, + 624, + 419, + 637 + ], + "score": 1.0, + "content": "Representations, 2017. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 419, + 626, + 426, + 634 + ], + "score": 0.5, + "content": "{ . } =", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 624, + 483, + 637 + ], + "score": 1.0, + "content": "HJ0NvFzxl.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 612, + 506, + 637 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 659 + ], + "score": 1.0, + "content": "Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. Character-aware neural language", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 116, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 666, + 497, + 680 + ], + "score": 1.0, + "content": "pp. 2741–2749. AAAI Press, 2016. URL http://dl.acm.org/citation.cfm?id", + "type": "text" + }, + { + "bbox": [ + 497, + 668, + 505, + 677 + ], + "score": 0.59, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 677, + 212, + 689 + ], + "spans": [ + { + "bbox": [ + 116, + 677, + 212, + 689 + ], + "score": 1.0, + "content": "3016100.3016285.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 643, + 505, + 689 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 709, + 503, + 721 + ], + "spans": [ + { + "bbox": [ + 116, + 709, + 503, + 721 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2015. URL https://arxiv.org/abs/1412.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 146, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 146, + 732 + ], + "score": 1.0, + "content": "6980.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 699, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Sosuke Kobayashi, Ran Tian, Naoaki Okazaki, and Kentaro Inui. Dynamic Entity Representation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "with Max-pooling Improves Machine Reading. In Proceedings of the 2016 Conference of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 506, + 119 + ], + "score": 1.0, + "content": "North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "Technologies, San Diego, California, June 2016. Association for Computational Linguistics. URL", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 362, + 140 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 362, + 140 + ], + "score": 1.0, + "content": "http://www.aclweb.org/anthology/N16-1099.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 144, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 157 + ], + "score": 1.0, + "content": "Jens Krinke. Identifying similar code with program dependence graphs. In Proceedings of the Eighth", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 154, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 115, + 154, + 506, + 169 + ], + "score": 1.0, + "content": "Working Conference on Reverse Engineering (WCRE’01), WCRE ’01, pp. 301–, Washington,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 165, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 504, + 180 + ], + "score": 1.0, + "content": "DC, USA, 2001. IEEE Computer Society. ISBN 0-7695-1303-4. URL http://dl.acm.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 178, + 319, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 178, + 231, + 190 + ], + "score": 1.0, + "content": "org/citation.cfm?id", + "type": "text" + }, + { + "bbox": [ + 231, + 178, + 239, + 187 + ], + "score": 0.28, + "content": "= \\mathfrak { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 178, + 319, + 190 + ], + "score": 1.0, + "content": "832308.837142.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 195, + 505, + 229 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "Jian Li, Yue Wang, Irwin King, and Michael R. Lyu. Code Completion with Neural Attention and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 207, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 116, + 207, + 504, + 219 + ], + "score": 1.0, + "content": "Pointer Networks. arXiv:1711.09573 [cs], November 2017. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 217, + 182, + 229 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 182, + 229 + ], + "score": 1.0, + "content": "1711.09573.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated Graph Sequence Neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 246, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 504, + 259 + ], + "score": 1.0, + "content": "Networks. In International Conference on Learning Representations, 2016. URL https://", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 258, + 266, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 266, + 270 + ], + "score": 1.0, + "content": "arxiv.org/abs/1511.05493.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "Chang Liu, Xin Wang, Richard Shin, Joseph E. Gonzalez, and Dawn Xiaodong Song. Neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 286, + 437, + 299 + ], + "score": 1.0, + "content": "code completion. 2017. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 437, + 288, + 444, + 296 + ], + "score": 0.39, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "rJbPBt9lg&", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 297, + 218, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 297, + 153, + 311 + ], + "score": 1.0, + "content": "noteId", + "type": "text" + }, + { + "bbox": [ + 153, + 299, + 159, + 307 + ], + "score": 0.46, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 297, + 218, + 311 + ], + "score": 1.0, + "content": "rJbPBt9lg.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Cristina V. Lopes, Petr Maj, Pedro Henrique Martins, Vaibhav Saini, Di Yang, Jakub Zitny, Hitesh", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "Sajnani, and Jan Vitek. Dej´ avu: a map of code duplicates on github. ` PACMPL, 1:84:1–84:28,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 336, + 144, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 144, + 349 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Zhengdong Lu, Haotian Cui, Xianggen Liu, Yukun Yan, and Daqi Zheng. Object-oriented Neural", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "Programming (OONP) for Document Understanding. arXiv:1709.08853 [cs], September 2017.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 377, + 411, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 377, + 411, + 391 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1709.08853. arXiv: 1709.08853.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "Minh-Thang Luong and Christopher D. Manning. Achieving Open Vocabulary Neural Machine", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "Translation with Hybrid Word-Character Models. arXiv:1604.00788 [cs], April 2016. URL", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 418, + 308, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 418, + 308, + 430 + ], + "score": 1.0, + "content": "http://arxiv.org/abs/1604.00788.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 504, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 436, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 504, + 447 + ], + "score": 1.0, + "content": "Chris J. Maddison and Daniel Tarlow. Structured generative models of natural source code.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 447, + 504, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 447, + 449, + 459 + ], + "score": 1.0, + "content": "pp. II–649–II–657, 2014. URL http://dl.acm.org/citation.cfm?id", + "type": "text" + }, + { + "bbox": [ + 449, + 448, + 456, + 457 + ], + "score": 0.32, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 447, + 504, + 459 + ], + "score": 1.0, + "content": "3044805.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 458, + 164, + 469 + ], + "spans": [ + { + "bbox": [ + 116, + 458, + 164, + 469 + ], + "score": 1.0, + "content": "3044965.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer Sentinel Mixture", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 487, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 504, + 500 + ], + "score": 1.0, + "content": "Models. In International Conference on Learning Representations, 2017. URL https://", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 117, + 499, + 307, + 511 + ], + "spans": [ + { + "bbox": [ + 117, + 499, + 242, + 511 + ], + "score": 1.0, + "content": "openreview.net/pdf?id", + "type": "text" + }, + { + "bbox": [ + 243, + 500, + 249, + 507 + ], + "score": 0.28, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 499, + 307, + 511 + ], + "score": 1.0, + "content": "Byj72udxe.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 504, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "Trang Pham, Truyen Tran, and Svetha Venkatesh. Graph Memory Networks for Molecular Activity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 527, + 500, + 540 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 500, + 540 + ], + "score": 1.0, + "content": "Prediction. arXiv:1801.02622 [cs], 2018. URL http://arxiv.org/abs/1801.02622.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Maxim Rabinovich, Mitchell Stern, and Dan Klein. Abstract syntax networks for code genera-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 116, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "tion and semantic parsing. In Proceedings of the 55th Annual Meeting of the Association for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "Computational Linguistics (Volume 1: Long Papers), pp. 1139–1149. Association for Compu-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 577, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 504, + 591 + ], + "score": 1.0, + "content": "tational Linguistics, 2017. doi: 10.18653/v1/P17-1105. URL http://www.aclweb.org/", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 589, + 230, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 589, + 230, + 603 + ], + "score": 1.0, + "content": "anthology/P17-1105.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 622 + ], + "score": 1.0, + "content": "Veselin Raychev, Martin Vechev, and Andreas Krause. Predicting program properties from ”big", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "code”. In Proceedings of the 42Nd Annual ACM SIGPLAN-SIGACT Symposium on Principles of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 505, + 642 + ], + "score": 1.0, + "content": "Programming Languages, POPL ’15, pp. 111–124, New York, NY, USA, 2015. ACM. ISBN 978-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "1-4503-3300-9. doi: 10.1145/2676726.2677009. URL http://doi.acm.org/10.1145/", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 651, + 212, + 663 + ], + "spans": [ + { + "bbox": [ + 116, + 651, + 212, + 663 + ], + "score": 1.0, + "content": "2676726.2677009.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 669, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "Welling. Modeling Relational Data with Graph Convolutional Networks. arXiv:1703.06103 [cs,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 691, + 408, + 704 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 408, + 704 + ], + "score": 1.0, + "content": "stat], March 2017. URL http://arxiv.org/abs/1703.06103.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 721 + ], + "score": 1.0, + "content": "Kristof T. Schutt, Farhad Arbabzadah, Stefan Chmiela, Klaus R. M ¨ uller, and Alexandre Tkatchenko.¨", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 117, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "Quantum-chemical insights from deep tensor neural networks. In Nature communications, 2017.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Sosuke Kobayashi, Ran Tian, Naoaki Okazaki, and Kentaro Inui. Dynamic Entity Representation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "with Max-pooling Improves Machine Reading. In Proceedings of the 2016 Conference of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 506, + 119 + ], + "score": 1.0, + "content": "North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "Technologies, San Diego, California, June 2016. Association for Computational Linguistics. URL", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 362, + 140 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 362, + 140 + ], + "score": 1.0, + "content": "http://www.aclweb.org/anthology/N16-1099.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 83, + 506, + 140 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 144, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 157 + ], + "score": 1.0, + "content": "Jens Krinke. Identifying similar code with program dependence graphs. In Proceedings of the Eighth", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 154, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 115, + 154, + 506, + 169 + ], + "score": 1.0, + "content": "Working Conference on Reverse Engineering (WCRE’01), WCRE ’01, pp. 301–, Washington,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 165, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 504, + 180 + ], + "score": 1.0, + "content": "DC, USA, 2001. IEEE Computer Society. ISBN 0-7695-1303-4. URL http://dl.acm.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 178, + 319, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 178, + 231, + 190 + ], + "score": 1.0, + "content": "org/citation.cfm?id", + "type": "text" + }, + { + "bbox": [ + 231, + 178, + 239, + 187 + ], + "score": 0.28, + "content": "= \\mathfrak { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 178, + 319, + 190 + ], + "score": 1.0, + "content": "832308.837142.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 144, + 506, + 190 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 195, + 505, + 229 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "Jian Li, Yue Wang, Irwin King, and Michael R. Lyu. Code Completion with Neural Attention and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 207, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 116, + 207, + 504, + 219 + ], + "score": 1.0, + "content": "Pointer Networks. arXiv:1711.09573 [cs], November 2017. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 217, + 182, + 229 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 182, + 229 + ], + "score": 1.0, + "content": "1711.09573.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 195, + 506, + 229 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated Graph Sequence Neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 246, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 504, + 259 + ], + "score": 1.0, + "content": "Networks. In International Conference on Learning Representations, 2016. URL https://", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 258, + 266, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 266, + 270 + ], + "score": 1.0, + "content": "arxiv.org/abs/1511.05493.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 236, + 505, + 270 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "Chang Liu, Xin Wang, Richard Shin, Joseph E. Gonzalez, and Dawn Xiaodong Song. Neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 286, + 437, + 299 + ], + "score": 1.0, + "content": "code completion. 2017. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 437, + 288, + 444, + 296 + ], + "score": 0.39, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "rJbPBt9lg&", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 297, + 218, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 297, + 153, + 311 + ], + "score": 1.0, + "content": "noteId", + "type": "text" + }, + { + "bbox": [ + 153, + 299, + 159, + 307 + ], + "score": 0.46, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 297, + 218, + 311 + ], + "score": 1.0, + "content": "rJbPBt9lg.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 275, + 505, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Cristina V. Lopes, Petr Maj, Pedro Henrique Martins, Vaibhav Saini, Di Yang, Jakub Zitny, Hitesh", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 326, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 505, + 340 + ], + "score": 1.0, + "content": "Sajnani, and Jan Vitek. Dej´ avu: a map of code duplicates on github. ` PACMPL, 1:84:1–84:28,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 336, + 144, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 144, + 349 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 315, + 505, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Zhengdong Lu, Haotian Cui, Xianggen Liu, Yukun Yan, and Daqi Zheng. Object-oriented Neural", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "Programming (OONP) for Document Understanding. arXiv:1709.08853 [cs], September 2017.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 377, + 411, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 377, + 411, + 391 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1709.08853. arXiv: 1709.08853.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 356, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "Minh-Thang Luong and Christopher D. Manning. Achieving Open Vocabulary Neural Machine", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "Translation with Hybrid Word-Character Models. arXiv:1604.00788 [cs], April 2016. URL", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 418, + 308, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 418, + 308, + 430 + ], + "score": 1.0, + "content": "http://arxiv.org/abs/1604.00788.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 395, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 504, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 436, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 504, + 447 + ], + "score": 1.0, + "content": "Chris J. Maddison and Daniel Tarlow. Structured generative models of natural source code.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 447, + 504, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 447, + 449, + 459 + ], + "score": 1.0, + "content": "pp. II–649–II–657, 2014. URL http://dl.acm.org/citation.cfm?id", + "type": "text" + }, + { + "bbox": [ + 449, + 448, + 456, + 457 + ], + "score": 0.32, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 447, + 504, + 459 + ], + "score": 1.0, + "content": "3044805.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 458, + 164, + 469 + ], + "spans": [ + { + "bbox": [ + 116, + 458, + 164, + 469 + ], + "score": 1.0, + "content": "3044965.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 436, + 504, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer Sentinel Mixture", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 487, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 504, + 500 + ], + "score": 1.0, + "content": "Models. In International Conference on Learning Representations, 2017. URL https://", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 117, + 499, + 307, + 511 + ], + "spans": [ + { + "bbox": [ + 117, + 499, + 242, + 511 + ], + "score": 1.0, + "content": "openreview.net/pdf?id", + "type": "text" + }, + { + "bbox": [ + 243, + 500, + 249, + 507 + ], + "score": 0.28, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 499, + 307, + 511 + ], + "score": 1.0, + "content": "Byj72udxe.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 476, + 506, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 504, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "Trang Pham, Truyen Tran, and Svetha Venkatesh. Graph Memory Networks for Molecular Activity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 527, + 500, + 540 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 500, + 540 + ], + "score": 1.0, + "content": "Prediction. arXiv:1801.02622 [cs], 2018. URL http://arxiv.org/abs/1801.02622.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 515, + 505, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Maxim Rabinovich, Mitchell Stern, and Dan Klein. Abstract syntax networks for code genera-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 116, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "tion and semantic parsing. In Proceedings of the 55th Annual Meeting of the Association for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "Computational Linguistics (Volume 1: Long Papers), pp. 1139–1149. Association for Compu-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 577, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 504, + 591 + ], + "score": 1.0, + "content": "tational Linguistics, 2017. doi: 10.18653/v1/P17-1105. URL http://www.aclweb.org/", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 589, + 230, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 589, + 230, + 603 + ], + "score": 1.0, + "content": "anthology/P17-1105.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 545, + 505, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 622 + ], + "score": 1.0, + "content": "Veselin Raychev, Martin Vechev, and Andreas Krause. Predicting program properties from ”big", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "code”. In Proceedings of the 42Nd Annual ACM SIGPLAN-SIGACT Symposium on Principles of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 505, + 642 + ], + "score": 1.0, + "content": "Programming Languages, POPL ’15, pp. 111–124, New York, NY, USA, 2015. ACM. ISBN 978-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "1-4503-3300-9. doi: 10.1145/2676726.2677009. URL http://doi.acm.org/10.1145/", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 651, + 212, + 663 + ], + "spans": [ + { + "bbox": [ + 116, + 651, + 212, + 663 + ], + "score": 1.0, + "content": "2676726.2677009.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 605, + 506, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 669, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "Welling. Modeling Relational Data with Graph Convolutional Networks. arXiv:1703.06103 [cs,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 691, + 408, + 704 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 408, + 704 + ], + "score": 1.0, + "content": "stat], March 2017. URL http://arxiv.org/abs/1703.06103.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 669, + 505, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 721 + ], + "score": 1.0, + "content": "Kristof T. Schutt, Farhad Arbabzadah, Stefan Chmiela, Klaus R. M ¨ uller, and Alexandre Tkatchenko.¨", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 117, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "Quantum-chemical insights from deep tensor neural networks. In Nature communications, 2017.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 487, + 96 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew", + "type": "text" + }, + { + "bbox": [ + 487, + 83, + 502, + 94 + ], + "score": 0.38, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 80, + 505, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 506, + 118 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "Processing, pp. 1631–1642, Seattle, Washington, USA, October 2013. Association for Computa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 127, + 459, + 138 + ], + "spans": [ + { + "bbox": [ + 116, + 127, + 459, + 138 + ], + "score": 1.0, + "content": "tional Linguistics. URL http://www.aclweb.org/anthology/D13-1170.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 504, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 504, + 158 + ], + "score": 1.0, + "content": "TIOBE. TIOBE Index for September 2018, 2018. URL https://www.tiobe.com/", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 156, + 194, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 194, + 168 + ], + "score": 1.0, + "content": "tiobe-index/.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In C. Cortes, N. D. Lawrence,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 197, + 504, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 504, + 209 + ], + "score": 1.0, + "content": "Systems 28, pp. 2692–2700. Curran Associates, Inc., 2015. URL http://papers.nips.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 209, + 325, + 221 + ], + "spans": [ + { + "bbox": [ + 117, + 209, + 325, + 221 + ], + "score": 1.0, + "content": "cc/paper/5866-pointer-networks.pdf.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 505, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "M. White, M. Tufano, C. Vendome, and D. Poshyvanyk. Deep learning code fragments for code", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "clone detection. In 2016 31st IEEE/ACM International Conference on Automated Software", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 249, + 504, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 504, + 261 + ], + "score": 1.0, + "content": "Engineering (ASE), pp. 87–98, Sept 2016. URL http://www.cs.wm.edu/˜mtufano/", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 260, + 236, + 272 + ], + "spans": [ + { + "bbox": [ + 115, + 260, + 236, + 272 + ], + "score": 1.0, + "content": "publications/C5.pdf.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 323 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Pengcheng Yin and Graham Neubig. A syntactic neural model for general-purpose code generation.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 289, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 115, + 289, + 505, + 303 + ], + "score": 1.0, + "content": "In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 300, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 115, + 300, + 505, + 314 + ], + "score": 1.0, + "content": "(Volume 1: Long Papers), pp. 440–450. Association for Computational Linguistics, 2017. doi:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 311, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 482, + 325 + ], + "score": 1.0, + "content": "10.18653/v1/P17-1041. URL http://www.aclweb.org/anthology/P17-1041.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "Wojciech Zaremba, Karol Kurach, and Rob Fergus. Learning to discover efficient math-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 342, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 505, + 353 + ], + "score": 1.0, + "content": "ematical identities. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 362, + 504, + 376 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 504, + 376 + ], + "score": 1.0, + "content": "1278–1286. Curran Associates, Inc., 2014. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 374, + 502, + 386 + ], + "spans": [ + { + "bbox": [ + 116, + 374, + 502, + 386 + ], + "score": 1.0, + "content": "5350-learning-to-discover-efficient-mathematical-identities.pdf.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 527, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 321, + 406 + ], + "score": 1.0, + "content": "Xiang Zhang, Junbo Zhao, and Yann LeCun.", + "type": "text" + }, + { + "bbox": [ + 335, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "Character-level convolutional networks", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "for text classification. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 114, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 649–", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 425, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 289, + 438 + ], + "score": 1.0, + "content": "657. Curran Associates, Inc., 2015.", + "type": "text" + }, + { + "bbox": [ + 305, + 425, + 505, + 439 + ], + "score": 1.0, + "content": "URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 436, + 528, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 528, + 450 + ], + "score": 1.0, + "content": "5782-character-level-convolutional-networks-for-text-classification.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 113, + 448, + 141, + 460 + ], + "spans": [ + { + "bbox": [ + 113, + 448, + 141, + 460 + ], + "score": 1.0, + "content": "pdf.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 487, + 96 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew", + "type": "text" + }, + { + "bbox": [ + 487, + 83, + 502, + 94 + ], + "score": 0.38, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 80, + 505, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 506, + 118 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "Processing, pp. 1631–1642, Seattle, Washington, USA, October 2013. Association for Computa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 127, + 459, + 138 + ], + "spans": [ + { + "bbox": [ + 116, + 127, + 459, + 138 + ], + "score": 1.0, + "content": "tional Linguistics. URL http://www.aclweb.org/anthology/D13-1170.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 80, + 506, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 504, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 504, + 158 + ], + "score": 1.0, + "content": "TIOBE. TIOBE Index for September 2018, 2018. URL https://www.tiobe.com/", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 156, + 194, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 194, + 168 + ], + "score": 1.0, + "content": "tiobe-index/.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 143, + 504, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 506, + 188 + ], + "score": 1.0, + "content": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In C. Cortes, N. D. Lawrence,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 197, + 504, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 504, + 209 + ], + "score": 1.0, + "content": "Systems 28, pp. 2692–2700. Curran Associates, Inc., 2015. URL http://papers.nips.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 209, + 325, + 221 + ], + "spans": [ + { + "bbox": [ + 117, + 209, + 325, + 221 + ], + "score": 1.0, + "content": "cc/paper/5866-pointer-networks.pdf.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 174, + 506, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 505, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "M. White, M. Tufano, C. Vendome, and D. Poshyvanyk. Deep learning code fragments for code", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "clone detection. In 2016 31st IEEE/ACM International Conference on Automated Software", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 249, + 504, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 504, + 261 + ], + "score": 1.0, + "content": "Engineering (ASE), pp. 87–98, Sept 2016. URL http://www.cs.wm.edu/˜mtufano/", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 260, + 236, + 272 + ], + "spans": [ + { + "bbox": [ + 115, + 260, + 236, + 272 + ], + "score": 1.0, + "content": "publications/C5.pdf.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 226, + 505, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 323 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Pengcheng Yin and Graham Neubig. A syntactic neural model for general-purpose code generation.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 289, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 115, + 289, + 505, + 303 + ], + "score": 1.0, + "content": "In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 300, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 115, + 300, + 505, + 314 + ], + "score": 1.0, + "content": "(Volume 1: Long Papers), pp. 440–450. Association for Computational Linguistics, 2017. doi:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 311, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 482, + 325 + ], + "score": 1.0, + "content": "10.18653/v1/P17-1041. URL http://www.aclweb.org/anthology/P17-1041.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 278, + 505, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "Wojciech Zaremba, Karol Kurach, and Rob Fergus. Learning to discover efficient math-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 342, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 115, + 342, + 505, + 353 + ], + "score": 1.0, + "content": "ematical identities. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 362, + 504, + 376 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 504, + 376 + ], + "score": 1.0, + "content": "1278–1286. Curran Associates, Inc., 2014. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 374, + 502, + 386 + ], + "spans": [ + { + "bbox": [ + 116, + 374, + 502, + 386 + ], + "score": 1.0, + "content": "5350-learning-to-discover-efficient-mathematical-identities.pdf.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 330, + 505, + 386 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 527, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 321, + 406 + ], + "score": 1.0, + "content": "Xiang Zhang, Junbo Zhao, and Yann LeCun.", + "type": "text" + }, + { + "bbox": [ + 335, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "Character-level convolutional networks", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "for text classification. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 114, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "R. Garnett (eds.), Advances in Neural Information Processing Systems 28, pp. 649–", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 425, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 289, + 438 + ], + "score": 1.0, + "content": "657. Curran Associates, Inc., 2015.", + "type": "text" + }, + { + "bbox": [ + 305, + 425, + 505, + 439 + ], + "score": 1.0, + "content": "URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 436, + 528, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 528, + 450 + ], + "score": 1.0, + "content": "5782-character-level-convolutional-networks-for-text-classification.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 113, + 448, + 141, + 460 + ], + "spans": [ + { + "bbox": [ + 113, + 448, + 141, + 460 + ], + "score": 1.0, + "content": "pdf.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 393, + 528, + 460 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 162, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 106, + 286, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 287, + 119 + ], + "score": 1.0, + "content": "ADDITIONAL EXPERIMENT INFORMATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 106, + 181, + 513, + 251 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 138, + 507, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "Table 6: Extra information about performance on the Variable Naming task. Entries in this table are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "of the form “subword accuracy, edit distance, edit distance divided by real name length”. The edit", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "distance is the mean of the character–wise Levenshtein distance between the produced name and the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 172, + 151, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 151, + 183 + ], + "score": 1.0, + "content": "real name.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 181, + 513, + 251 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 181, + 513, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 513, + 251 + ], + "score": 0.976, + "html": "
Closed VocabCharCNNPointer SentinelGSC (ours)
Seen reposAST AugAST0.30, 7.22, 0.94 0.26, 7.64, 0.940.28,8.67, 1.08 0.28, 7.46, 0.990.32, 8.00, 1.07 0.35, 6.51, 0.770.56, 3.87, 0.39 0.60, 3.68, 0.37
Unseen reposAST0.09,8.66,1.230.10, 8.82, 1.120.13, 9,39, 1.370.42, 4.81, 0.59
AugAST0.09, 8.34, 1.140.10, 8.16, 1.120.14, 8.03, 1.070.48, 4.28, 0.49
", + "type": "table", + "image_path": "eede5588c6566c88a1a2b4aaa322127dac62da0a73f053e503ae0a31e2d94f9c.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 106, + 181, + 513, + 204.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 204.33333333333334, + 513, + 227.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 227.66666666666669, + 513, + 251.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 5.25 + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 210, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 212, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 212, + 283 + ], + "score": 1.0, + "content": "DATASET INFORMATION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 201, + 326, + 408, + 550 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 303, + 504, + 326 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "score": 1.0, + "content": "Table 7: Repositories used in experiments. All were taken from the Maven repository", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 315, + 455, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 455, + 328 + ], + "score": 1.0, + "content": "(https://mvnrepository.com/). Entries are in the form “group/repository name/version”.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 201, + 326, + 408, + 550 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 201, + 326, + 408, + 550 + ], + "spans": [ + { + "bbox": [ + 201, + 326, + 408, + 550 + ], + "score": 0.117, + "html": "
Seen Repos
com.fasterxml.jackson.core/jackson-core/2.9.5 com.h2database/h2/1.4.195 javax.enterprise/cdi-api/2.0 junit/junit/4.12 mysql/mysql-connector-java/6.0.6 org.apache.commons/commons-collections4/4.1 org.apache.commons/commons-math3/3.6.1 org.apache.commons/commons-pool2/2.5.0 org.apache.maven/maven-project/2.2.1 org.codehaus.plexus/plexus-utils/3.1.0 org.eclipse.jetty/jetty-server/9.4.9.v20180320
org.scalacheck/scalacheck_2.12/1.13.5 org.slf4j/slf4j-api/1.7.25 org.slf4j/slf4j-log4j12/1.7.25 Unseen Repos
org.javassist/javassist/3.22.0-GA joda-time/joda-time/2.9.9 org.mockito/mockito-core/2.17.0
", + "type": "table", + "image_path": "3cdbb6558a551b95d720af6941b5a58667c04241856332d0fca4e455b35250a5.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 201, + 326, + 408, + 339.1764705882353 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 201, + 339.1764705882353, + 408, + 352.3529411764706 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 201, + 352.3529411764706, + 408, + 365.5294117647059 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 201, + 365.5294117647059, + 408, + 378.7058823529412 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 201, + 378.7058823529412, + 408, + 391.8823529411765 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 201, + 391.8823529411765, + 408, + 405.0588235294118 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 201, + 405.0588235294118, + 408, + 418.23529411764713 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 201, + 418.23529411764713, + 408, + 431.41176470588243 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 201, + 431.41176470588243, + 408, + 444.58823529411774 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 201, + 444.58823529411774, + 408, + 457.76470588235304 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 201, + 457.76470588235304, + 408, + 470.94117647058835 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 201, + 470.94117647058835, + 408, + 484.11764705882365 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 201, + 484.11764705882365, + 408, + 497.29411764705895 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 201, + 497.29411764705895, + 408, + 510.47058823529426 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 201, + 510.47058823529426, + 408, + 523.6470588235295 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 201, + 523.6470588235295, + 408, + 536.8235294117648 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 201, + 536.8235294117648, + 408, + 550.0 + ], + "spans": [], + "index": 28 + } + ] + } + ], + "index": 15.25 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 570, + 204, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 205, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 205, + 582 + ], + "score": 1.0, + "content": "MODEL INFORMATION", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 162, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 106, + 286, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 287, + 119 + ], + "score": 1.0, + "content": "ADDITIONAL EXPERIMENT INFORMATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 105, + 287, + 119 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 181, + 513, + 251 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 138, + 507, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "Table 6: Extra information about performance on the Variable Naming task. Entries in this table are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "of the form “subword accuracy, edit distance, edit distance divided by real name length”. The edit", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "distance is the mean of the character–wise Levenshtein distance between the produced name and the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 172, + 151, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 151, + 183 + ], + "score": 1.0, + "content": "real name.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 181, + 513, + 251 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 181, + 513, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 513, + 251 + ], + "score": 0.976, + "html": "
Closed VocabCharCNNPointer SentinelGSC (ours)
Seen reposAST AugAST0.30, 7.22, 0.94 0.26, 7.64, 0.940.28,8.67, 1.08 0.28, 7.46, 0.990.32, 8.00, 1.07 0.35, 6.51, 0.770.56, 3.87, 0.39 0.60, 3.68, 0.37
Unseen reposAST0.09,8.66,1.230.10, 8.82, 1.120.13, 9,39, 1.370.42, 4.81, 0.59
AugAST0.09, 8.34, 1.140.10, 8.16, 1.120.14, 8.03, 1.070.48, 4.28, 0.49
", + "type": "table", + "image_path": "eede5588c6566c88a1a2b4aaa322127dac62da0a73f053e503ae0a31e2d94f9c.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 106, + 181, + 513, + 204.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 204.33333333333334, + 513, + 227.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 227.66666666666669, + 513, + 251.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 5.25 + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 210, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 212, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 212, + 283 + ], + "score": 1.0, + "content": "DATASET INFORMATION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 271, + 212, + 283 + ] + }, + { + "type": "table", + "bbox": [ + 201, + 326, + 408, + 550 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 303, + 504, + 326 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "score": 1.0, + "content": "Table 7: Repositories used in experiments. All were taken from the Maven repository", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 315, + 455, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 455, + 328 + ], + "score": 1.0, + "content": "(https://mvnrepository.com/). Entries are in the form “group/repository name/version”.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 201, + 326, + 408, + 550 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 201, + 326, + 408, + 550 + ], + "spans": [ + { + "bbox": [ + 201, + 326, + 408, + 550 + ], + "score": 0.117, + "html": "
Seen Repos
com.fasterxml.jackson.core/jackson-core/2.9.5 com.h2database/h2/1.4.195 javax.enterprise/cdi-api/2.0 junit/junit/4.12 mysql/mysql-connector-java/6.0.6 org.apache.commons/commons-collections4/4.1 org.apache.commons/commons-math3/3.6.1 org.apache.commons/commons-pool2/2.5.0 org.apache.maven/maven-project/2.2.1 org.codehaus.plexus/plexus-utils/3.1.0 org.eclipse.jetty/jetty-server/9.4.9.v20180320
org.scalacheck/scalacheck_2.12/1.13.5 org.slf4j/slf4j-api/1.7.25 org.slf4j/slf4j-log4j12/1.7.25 Unseen Repos
org.javassist/javassist/3.22.0-GA joda-time/joda-time/2.9.9 org.mockito/mockito-core/2.17.0
", + "type": "table", + "image_path": "3cdbb6558a551b95d720af6941b5a58667c04241856332d0fca4e455b35250a5.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 201, + 326, + 408, + 339.1764705882353 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 201, + 339.1764705882353, + 408, + 352.3529411764706 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 201, + 352.3529411764706, + 408, + 365.5294117647059 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 201, + 365.5294117647059, + 408, + 378.7058823529412 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 201, + 378.7058823529412, + 408, + 391.8823529411765 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 201, + 391.8823529411765, + 408, + 405.0588235294118 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 201, + 405.0588235294118, + 408, + 418.23529411764713 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 201, + 418.23529411764713, + 408, + 431.41176470588243 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 201, + 431.41176470588243, + 408, + 444.58823529411774 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 201, + 444.58823529411774, + 408, + 457.76470588235304 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 201, + 457.76470588235304, + 408, + 470.94117647058835 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 201, + 470.94117647058835, + 408, + 484.11764705882365 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 201, + 484.11764705882365, + 408, + 497.29411764705895 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 201, + 497.29411764705895, + 408, + 510.47058823529426 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 201, + 510.47058823529426, + 408, + 523.6470588235295 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 201, + 523.6470588235295, + 408, + 536.8235294117648 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 201, + 536.8235294117648, + 408, + 550.0 + ], + "spans": [], + "index": 28 + } + ] + } + ], + "index": 15.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 223, + 471, + 661 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 156, + 506, + 223 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "Table 8: Edge types used in Augmented ASTs. The initial AST is constructed using the AST and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "NEXT TOKEN edges, and then the remaining edges are added. In other words, the the “AST” model", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 178, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 506, + 191 + ], + "score": 1.0, + "content": "from Table 1 uses a graph that contains only the AST and NEXT TOKEN edge types (and WORD USE", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "score": 1.0, + "content": "if it also uses a GSC), while the “AugAST” model contains all the edge types below. The reversed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "version of every edge is also added as its own type (e.g. reverse AST, reverse LAST READ)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 211, + 336, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 336, + 224 + ], + "score": 1.0, + "content": "to let the GNN message passing occur in both directions.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 130, + 223, + 471, + 661 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 471, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 471, + 661 + ], + "score": 0.981, + "html": "
Edge NameDescription
AST NEXT_TOKENThe edges used to construct the original AST. Edges added to the original AST that specify the left-to-
right ordering of the children of a node in the AST.These edges are necessary since ASTs have ordered children, but we are representing the AST as a directed multigraph.
COMPUTED_FROMConnects a node representing a variable on the left of an equality to those on the right. (E.g. edges from y to x and Z to X in x = y + z.) The same as in Allamanis et al. (2018).
LAST_READConnects a node representing a usage of a variable to all nodes in the AST at which that variable's value could have been last read from memory. The same as in Allamanis et al. (2018).
LAST_WRITEConnects a node representing a usage of a variable to all nodesin the AST at which that variable's value could have been last written to memory. The same as in Allamanis et al.
RETURNS_TO(2018). Points a node in a return statement to the node contain- ing the return type of the method. (E.g. x in return X gets an edge pointing to int in public static int
LAST_SCOPE_USEgetx(x).) Connects a node representing a variable to the node repre- senting the last time this variable's name was used in the text of the code (i.e. capturing information about the text, not the control flow),but only within lexical scope. This edge exists to try and give the non-GSC models as much
LAST_FIELD_LEXlexical information as possible to make them as comparable with the GSC model. Connects a field access (e.g. this.whatever or Foo .whatever) node to the last use of this.whatever (or to the variable's initialization,if it's the first use).This is not lexical-scope aware (and,in fact,can't be in Java,in
FIELDgeneral). Pointseach node representing a field access(e.g. this.whatever) to the node where that field was de-
WORD_USEclared. Points cache nodes to nodes representing variables in which the vocab word was used in the variable's name.
", + "type": "table", + "image_path": "3be36687fd308695ae5ec3a990a3322b5d2dc3db17b8b2b431ec39007d2d0335.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 130, + 223, + 471, + 369.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 130, + 369.0, + 471, + 515.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 130, + 515.0, + 471, + 661.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4.75 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 223, + 471, + 661 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 156, + 506, + 223 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "Table 8: Edge types used in Augmented ASTs. The initial AST is constructed using the AST and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "NEXT TOKEN edges, and then the remaining edges are added. In other words, the the “AST” model", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 178, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 506, + 191 + ], + "score": 1.0, + "content": "from Table 1 uses a graph that contains only the AST and NEXT TOKEN edge types (and WORD USE", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 201 + ], + "score": 1.0, + "content": "if it also uses a GSC), while the “AugAST” model contains all the edge types below. The reversed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 201, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 506, + 214 + ], + "score": 1.0, + "content": "version of every edge is also added as its own type (e.g. reverse AST, reverse LAST READ)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 211, + 336, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 336, + 224 + ], + "score": 1.0, + "content": "to let the GNN message passing occur in both directions.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 130, + 223, + 471, + 661 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 471, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 471, + 661 + ], + "score": 0.981, + "html": "
Edge NameDescription
AST NEXT_TOKENThe edges used to construct the original AST. Edges added to the original AST that specify the left-to-
right ordering of the children of a node in the AST.These edges are necessary since ASTs have ordered children, but we are representing the AST as a directed multigraph.
COMPUTED_FROMConnects a node representing a variable on the left of an equality to those on the right. (E.g. edges from y to x and Z to X in x = y + z.) The same as in Allamanis et al. (2018).
LAST_READConnects a node representing a usage of a variable to all nodes in the AST at which that variable's value could have been last read from memory. The same as in Allamanis et al. (2018).
LAST_WRITEConnects a node representing a usage of a variable to all nodesin the AST at which that variable's value could have been last written to memory. The same as in Allamanis et al.
RETURNS_TO(2018). Points a node in a return statement to the node contain- ing the return type of the method. (E.g. x in return X gets an edge pointing to int in public static int
LAST_SCOPE_USEgetx(x).) Connects a node representing a variable to the node repre- senting the last time this variable's name was used in the text of the code (i.e. capturing information about the text, not the control flow),but only within lexical scope. This edge exists to try and give the non-GSC models as much
LAST_FIELD_LEXlexical information as possible to make them as comparable with the GSC model. Connects a field access (e.g. this.whatever or Foo .whatever) node to the last use of this.whatever (or to the variable's initialization,if it's the first use).This is not lexical-scope aware (and,in fact,can't be in Java,in
FIELDgeneral). Pointseach node representing a field access(e.g. this.whatever) to the node where that field was de-
WORD_USEclared. Points cache nodes to nodes representing variables in which the vocab word was used in the variable's name.
", + "type": "table", + "image_path": "3be36687fd308695ae5ec3a990a3322b5d2dc3db17b8b2b431ec39007d2d0335.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 130, + 223, + 471, + 369.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 130, + 369.0, + 471, + 515.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 130, + 515.0, + 471, + 661.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4.75 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/SkNSehA9FQ/SkNSehA9FQ_model.json b/parse/train/SkNSehA9FQ/SkNSehA9FQ_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40a11a3da5f7498c428403f1679693eb3869bb8d --- /dev/null +++ b/parse/train/SkNSehA9FQ/SkNSehA9FQ_model.json @@ -0,0 +1,17203 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 592, + 1302, + 592, + 1302, + 958, + 398, + 958 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1098, + 1403, + 1098, + 1403, + 1283, + 298, + 1283 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 301, + 220, + 1401, + 220, + 1401, + 324, + 301, + 324 + ], + "score": 0.964 + }, + { + "category_id": 2, + "poly": [ + 298, + 1894, + 1404, + 1894, + 1404, + 2033, + 298, + 2033 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 300, + 1769, + 1399, + 1769, + 1399, + 1862, + 300, + 1862 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 298, + 1298, + 934, + 1298, + 934, + 1329, + 298, + 1329 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 302, + 1026, + 573, + 1026, + 573, + 1062, + 302, + 1062 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 773, + 520, + 927, + 520, + 927, + 553, + 773, + 553 + ], + "score": 0.857 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 299, + 1358, + 1405, + 1358, + 1405, + 1511, + 299, + 1511 + ], + "score": 0.55 + }, + { + "category_id": 1, + "poly": [ + 300, + 1526, + 1403, + 1526, + 1403, + 1740, + 300, + 1740 + ], + "score": 0.439 + }, + { + "category_id": 13, + "poly": [ + 1046, + 1770, + 1101, + 1770, + 1101, + 1799, + 1046, + 1799 + ], + "score": 0.86, + "latex": "28 \\%" + }, + { + "category_id": 13, + "poly": [ + 690, + 898, + 758, + 898, + 758, + 927, + 690, + 927 + ], + "score": 0.85, + "latex": "100 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 215.0, + 1405.0, + 215.0, + 1405.0, + 273.0, + 295.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 275.0, + 1080.0, + 275.0, + 1080.0, + 327.0, + 297.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1890.0, + 1406.0, + 1890.0, + 1406.0, + 1929.0, + 331.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1922.0, + 1404.0, + 1922.0, + 1404.0, + 1952.0, + 294.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1951.0, + 372.0, + 1951.0, + 372.0, + 1980.0, + 296.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2005.0, + 1307.0, + 2005.0, + 1307.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1023.0, + 579.0, + 1023.0, + 579.0, + 1070.0, + 294.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 516.0, + 934.0, + 516.0, + 934.0, + 559.0, + 768.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 590.0, + 1305.0, + 590.0, + 1305.0, + 630.0, + 392.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 622.0, + 1306.0, + 622.0, + 1306.0, + 661.0, + 394.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 652.0, + 1306.0, + 652.0, + 1306.0, + 690.0, + 392.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 685.0, + 1306.0, + 685.0, + 1306.0, + 717.0, + 393.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 715.0, + 1306.0, + 715.0, + 1306.0, + 750.0, + 394.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 746.0, + 1305.0, + 746.0, + 1305.0, + 778.0, + 393.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 775.0, + 1306.0, + 775.0, + 1306.0, + 808.0, + 393.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 806.0, + 1305.0, + 806.0, + 1305.0, + 838.0, + 394.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 835.0, + 1307.0, + 835.0, + 1307.0, + 874.0, + 393.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 868.0, + 1306.0, + 868.0, + 1306.0, + 899.0, + 394.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 899.0, + 689.0, + 899.0, + 689.0, + 931.0, + 395.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 899.0, + 1306.0, + 899.0, + 1306.0, + 931.0, + 759.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 927.0, + 890.0, + 927.0, + 890.0, + 962.0, + 395.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1098.0, + 1405.0, + 1098.0, + 1405.0, + 1133.0, + 294.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 1402.0, + 1131.0, + 1402.0, + 1163.0, + 294.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1157.0, + 1404.0, + 1157.0, + 1404.0, + 1197.0, + 292.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1188.0, + 1405.0, + 1188.0, + 1405.0, + 1229.0, + 292.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1222.0, + 1404.0, + 1222.0, + 1404.0, + 1254.0, + 296.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1250.0, + 1256.0, + 1250.0, + 1256.0, + 1286.0, + 293.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1770.0, + 1045.0, + 1770.0, + 1045.0, + 1803.0, + 296.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1770.0, + 1404.0, + 1770.0, + 1404.0, + 1803.0, + 1102.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1800.0, + 1404.0, + 1800.0, + 1404.0, + 1834.0, + 295.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1829.0, + 1404.0, + 1829.0, + 1404.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1296.0, + 939.0, + 1296.0, + 939.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1358.0, + 1402.0, + 1358.0, + 1402.0, + 1391.0, + 297.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1391.0, + 1402.0, + 1391.0, + 1402.0, + 1420.0, + 332.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1454.0, + 328.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1446.0, + 1407.0, + 1446.0, + 1407.0, + 1485.0, + 328.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1482.0, + 1331.0, + 1482.0, + 1331.0, + 1515.0, + 331.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1523.0, + 1408.0, + 1523.0, + 1408.0, + 1564.0, + 291.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1557.0, + 1405.0, + 1557.0, + 1405.0, + 1591.0, + 329.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1588.0, + 1404.0, + 1588.0, + 1404.0, + 1622.0, + 330.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1617.0, + 1405.0, + 1617.0, + 1405.0, + 1653.0, + 328.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1648.0, + 1405.0, + 1648.0, + 1405.0, + 1683.0, + 329.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1710.0, + 331.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1712.0, + 854.0, + 1712.0, + 854.0, + 1742.0, + 331.0, + 1742.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1072, + 1404, + 1072, + 1404, + 1529, + 298, + 1529 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 676, + 1404, + 676, + 1404, + 889, + 298, + 889 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1652, + 1403, + 1652, + 1403, + 1867, + 298, + 1867 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 338, + 1403, + 338, + 1403, + 491, + 299, + 491 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 507, + 1403, + 507, + 1403, + 660, + 298, + 660 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 230, + 1398, + 230, + 1398, + 322, + 300, + 322 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 302, + 1545, + 1397, + 1545, + 1397, + 1637, + 302, + 1637 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 1882, + 1400, + 1882, + 1400, + 1945, + 299, + 1945 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 300, + 939, + 543, + 939, + 543, + 976, + 300, + 976 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 857, + 76, + 857, + 104, + 299, + 104 + ], + "score": 0.819 + }, + { + "category_id": 2, + "poly": [ + 302, + 1977, + 1403, + 1977, + 1403, + 2033, + 302, + 2033 + ], + "score": 0.748 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.74 + }, + { + "category_id": 0, + "poly": [ + 303, + 1012, + 715, + 1012, + 715, + 1042, + 303, + 1042 + ], + "score": 0.577 + }, + { + "category_id": 2, + "poly": [ + 299, + 1978, + 1404, + 1978, + 1404, + 2033, + 299, + 2033 + ], + "score": 0.477 + }, + { + "category_id": 1, + "poly": [ + 303, + 1012, + 715, + 1012, + 715, + 1042, + 303, + 1042 + ], + "score": 0.319 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.26 + }, + { + "category_id": 13, + "poly": [ + 1258, + 798, + 1325, + 798, + 1325, + 827, + 1258, + 827 + ], + "score": 0.91, + "latex": "103 \\%" + }, + { + "category_id": 13, + "poly": [ + 974, + 738, + 1030, + 738, + 1030, + 767, + 974, + 767 + ], + "score": 0.87, + "latex": "30 \\%" + }, + { + "category_id": 13, + "poly": [ + 841, + 798, + 882, + 798, + 882, + 826, + 841, + 826 + ], + "score": 0.85, + "latex": "7 \\%" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 936.0, + 548.0, + 936.0, + 548.0, + 981.0, + 292.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1968.0, + 1407.0, + 1968.0, + 1407.0, + 2015.0, + 325.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2000.0, + 426.0, + 2000.0, + 426.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1013.0, + 719.0, + 1013.0, + 719.0, + 1045.0, + 298.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2014.0, + 326.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 428.0, + 1999.0, + 428.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1070.0, + 1406.0, + 1070.0, + 1406.0, + 1109.0, + 293.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1103.0, + 1404.0, + 1103.0, + 1404.0, + 1138.0, + 296.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1133.0, + 1406.0, + 1133.0, + 1406.0, + 1167.0, + 294.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1164.0, + 1404.0, + 1164.0, + 1404.0, + 1199.0, + 294.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1192.0, + 1405.0, + 1192.0, + 1405.0, + 1227.0, + 293.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1225.0, + 1405.0, + 1225.0, + 1405.0, + 1259.0, + 293.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1251.0, + 1406.0, + 1251.0, + 1406.0, + 1292.0, + 291.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1284.0, + 1406.0, + 1284.0, + 1406.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1315.0, + 1405.0, + 1315.0, + 1405.0, + 1350.0, + 293.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1344.0, + 1406.0, + 1344.0, + 1406.0, + 1383.0, + 292.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1411.0, + 293.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1405.0, + 1404.0, + 1405.0, + 1404.0, + 1442.0, + 292.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1437.0, + 1404.0, + 1437.0, + 1404.0, + 1472.0, + 294.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 1405.0, + 1467.0, + 1405.0, + 1502.0, + 294.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1498.0, + 1016.0, + 1498.0, + 1016.0, + 1533.0, + 293.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 676.0, + 1404.0, + 676.0, + 1404.0, + 710.0, + 296.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 707.0, + 1405.0, + 707.0, + 1405.0, + 741.0, + 294.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 738.0, + 973.0, + 738.0, + 973.0, + 771.0, + 293.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 738.0, + 1404.0, + 738.0, + 1404.0, + 771.0, + 1031.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 768.0, + 1407.0, + 768.0, + 1407.0, + 802.0, + 293.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 796.0, + 840.0, + 796.0, + 840.0, + 830.0, + 294.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 796.0, + 1257.0, + 796.0, + 1257.0, + 830.0, + 883.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 796.0, + 1404.0, + 796.0, + 1404.0, + 830.0, + 1326.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 863.0, + 294.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 858.0, + 1147.0, + 858.0, + 1147.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1651.0, + 1403.0, + 1651.0, + 1403.0, + 1688.0, + 293.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1716.0, + 1403.0, + 1716.0, + 1403.0, + 1747.0, + 296.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1774.0, + 1405.0, + 1774.0, + 1405.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1803.0, + 1407.0, + 1803.0, + 1407.0, + 1842.0, + 293.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1837.0, + 842.0, + 1837.0, + 842.0, + 1871.0, + 294.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 337.0, + 1404.0, + 337.0, + 1404.0, + 374.0, + 295.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 368.0, + 1404.0, + 368.0, + 1404.0, + 401.0, + 294.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 397.0, + 1408.0, + 397.0, + 1408.0, + 434.0, + 290.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 426.0, + 1405.0, + 426.0, + 1405.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 459.0, + 1195.0, + 459.0, + 1195.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 504.0, + 1407.0, + 504.0, + 1407.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 535.0, + 1407.0, + 535.0, + 1407.0, + 573.0, + 293.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 568.0, + 1405.0, + 568.0, + 1405.0, + 605.0, + 294.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 599.0, + 1403.0, + 599.0, + 1403.0, + 632.0, + 296.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 628.0, + 1173.0, + 628.0, + 1173.0, + 665.0, + 294.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 1402.0, + 227.0, + 1402.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 259.0, + 1403.0, + 259.0, + 1403.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 561.0, + 292.0, + 561.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1545.0, + 1401.0, + 1545.0, + 1401.0, + 1579.0, + 297.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1574.0, + 1403.0, + 1574.0, + 1403.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1607.0, + 614.0, + 1607.0, + 614.0, + 1641.0, + 297.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1879.0, + 1404.0, + 1879.0, + 1404.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1324.0, + 1911.0, + 1324.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1013.0, + 719.0, + 1013.0, + 719.0, + 1045.0, + 298.0, + 1045.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 287, + 1403, + 287, + 1403, + 532, + 297, + 532 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 546, + 1403, + 546, + 1403, + 701, + 298, + 701 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1359, + 1403, + 1359, + 1403, + 1514, + 298, + 1514 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 715, + 1404, + 715, + 1404, + 839, + 298, + 839 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1527, + 1403, + 1527, + 1403, + 1684, + 298, + 1684 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1005, + 1403, + 1005, + 1403, + 1128, + 299, + 1128 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1143, + 1404, + 1143, + 1404, + 1268, + 299, + 1268 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 673, + 1969, + 1133, + 1969, + 1133, + 2041, + 673, + 2041 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 312, + 1750, + 1397, + 1750, + 1397, + 1815, + 312, + 1815 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 361, + 1900, + 1403, + 1900, + 1403, + 1963, + 361, + 1963 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 803, + 1818, + 949, + 1818, + 949, + 1855, + 803, + 1855 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 299, + 1695, + 1390, + 1695, + 1390, + 1731, + 299, + 1731 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 317, + 1862, + 786, + 1862, + 786, + 1894, + 317, + 1894 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 301, + 948, + 688, + 948, + 688, + 980, + 301, + 980 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 301, + 881, + 578, + 881, + 578, + 916, + 301, + 916 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 302, + 1302, + 706, + 1302, + 706, + 1334, + 302, + 1334 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.683 + }, + { + "category_id": 0, + "poly": [ + 300, + 230, + 676, + 230, + 676, + 261, + 300, + 261 + ], + "score": 0.675 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.382 + }, + { + "category_id": 1, + "poly": [ + 300, + 230, + 676, + 230, + 676, + 261, + 300, + 261 + ], + "score": 0.249 + }, + { + "category_id": 14, + "poly": [ + 673, + 1967, + 1135, + 1967, + 1135, + 2041, + 673, + 2041 + ], + "score": 0.94, + "latex": "m _ { v } ^ { t + 1 } = \\sum _ { w \\in \\mathrm { n e i g h b o r s ~ o f } v } M _ { t } ( h _ { v } ^ { t } , h _ { w } ^ { t } , e _ { v w } ) ." + }, + { + "category_id": 13, + "poly": [ + 939, + 1697, + 1074, + 1697, + 1074, + 1731, + 939, + 1731 + ], + "score": 0.93, + "latex": "G = ( V , E )" + }, + { + "category_id": 13, + "poly": [ + 508, + 1650, + 634, + 1650, + 634, + 1684, + 508, + 1684 + ], + "score": 0.93, + "latex": "( v , w ) \\in E" + }, + { + "category_id": 13, + "poly": [ + 770, + 1529, + 904, + 1529, + 904, + 1563, + 770, + 1563 + ], + "score": 0.92, + "latex": "G = ( V , E )" + }, + { + "category_id": 13, + "poly": [ + 1072, + 1622, + 1146, + 1622, + 1146, + 1649, + 1072, + 1649 + ], + "score": 0.92, + "latex": "v \\in V" + }, + { + "category_id": 13, + "poly": [ + 855, + 1900, + 923, + 1900, + 923, + 1934, + 855, + 1934 + ], + "score": 0.92, + "latex": "\\boldsymbol { m } _ { v } ^ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 622, + 1529, + 693, + 1529, + 693, + 1563, + 622, + 1563 + ], + "score": 0.91, + "latex": "( G , y )" + }, + { + "category_id": 14, + "poly": [ + 803, + 1816, + 950, + 1816, + 950, + 1855, + 803, + 1855 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { h _ { v } ^ { 0 } = S ( \\pmb { x } _ { v } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1174, + 1753, + 1251, + 1753, + 1251, + 1781, + 1174, + 1781 + ], + "score": 0.91, + "latex": "v \\in V" + }, + { + "category_id": 13, + "poly": [ + 961, + 1751, + 996, + 1751, + 996, + 1785, + 961, + 1785 + ], + "score": 0.9, + "latex": "h _ { v } ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 549, + 1902, + 630, + 1902, + 630, + 1930, + 549, + 1930 + ], + "score": 0.9, + "latex": "v \\in V" + }, + { + "category_id": 13, + "poly": [ + 852, + 1935, + 890, + 1935, + 890, + 1963, + 852, + 1963 + ], + "score": 0.89, + "latex": "M _ { t }" + }, + { + "category_id": 13, + "poly": [ + 776, + 1656, + 823, + 1656, + 823, + 1682, + 776, + 1682 + ], + "score": 0.88, + "latex": "e _ { v w }" + }, + { + "category_id": 13, + "poly": [ + 739, + 1787, + 773, + 1787, + 773, + 1814, + 739, + 1814 + ], + "score": 0.87, + "latex": "\\mathbf { \\Delta } _ { \\mathbf { \\mathcal { X } } _ { v } }" + }, + { + "category_id": 13, + "poly": [ + 636, + 1699, + 655, + 1699, + 655, + 1729, + 636, + 1729 + ], + "score": 0.85, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1655, + 331, + 1655, + 331, + 1682, + 297, + 1682 + ], + "score": 0.84, + "latex": "\\mathbf { \\boldsymbol { x } } _ { v }" + }, + { + "category_id": 13, + "poly": [ + 899, + 1785, + 922, + 1785, + 922, + 1811, + 899, + 1811 + ], + "score": 0.84, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 616, + 1865, + 639, + 1865, + 639, + 1889, + 616, + 1889 + ], + "score": 0.83, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 798, + 1703, + 817, + 1703, + 817, + 1730, + 798, + 1730 + ], + "score": 0.82, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 480, + 1754, + 502, + 1754, + 502, + 1780, + 480, + 1780 + ], + "score": 0.81, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 381, + 1564, + 400, + 1564, + 400, + 1592, + 381, + 1592 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 298, + 1561, + 322, + 1561, + 322, + 1587, + 298, + 1587 + ], + "score": 0.78, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 526, + 1866, + 541, + 1866, + 541, + 1890, + 526, + 1890 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 522, + 1565, + 541, + 1565, + 541, + 1592, + 522, + 1592 + ], + "score": 0.73, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1231, + 384, + 1251, + 384, + 1251, + 407, + 1231, + 407 + ], + "score": 0.7, + "latex": "n _ { \\mathrm { ~ } }" + }, + { + "category_id": 13, + "poly": [ + 1261, + 1530, + 1287, + 1530, + 1287, + 1558, + 1261, + 1558 + ], + "score": 0.65, + "latex": "V" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 690.0, + 948.0, + 690.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 877.0, + 583.0, + 877.0, + 583.0, + 923.0, + 291.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1299.0, + 711.0, + 1299.0, + 711.0, + 1339.0, + 293.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 678.0, + 230.0, + 678.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 285.0, + 1405.0, + 285.0, + 1405.0, + 326.0, + 292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1407.0, + 317.0, + 1407.0, + 354.0, + 295.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 350.0, + 1403.0, + 350.0, + 1403.0, + 383.0, + 295.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 380.0, + 1230.0, + 380.0, + 1230.0, + 413.0, + 293.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 380.0, + 1405.0, + 380.0, + 1405.0, + 413.0, + 1252.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 411.0, + 1403.0, + 411.0, + 1403.0, + 444.0, + 295.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 437.0, + 1405.0, + 437.0, + 1405.0, + 475.0, + 294.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 468.0, + 1407.0, + 468.0, + 1407.0, + 508.0, + 292.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 501.0, + 729.0, + 501.0, + 729.0, + 534.0, + 295.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 545.0, + 1405.0, + 545.0, + 1405.0, + 584.0, + 293.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 1404.0, + 580.0, + 1404.0, + 612.0, + 294.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 608.0, + 1405.0, + 608.0, + 1405.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 637.0, + 1405.0, + 637.0, + 1405.0, + 677.0, + 293.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 669.0, + 1381.0, + 669.0, + 1381.0, + 706.0, + 294.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1361.0, + 1404.0, + 1361.0, + 1404.0, + 1394.0, + 297.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1391.0, + 1405.0, + 1391.0, + 1405.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1458.0, + 293.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1451.0, + 1407.0, + 1451.0, + 1407.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1480.0, + 1204.0, + 1480.0, + 1204.0, + 1518.0, + 296.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 715.0, + 1404.0, + 715.0, + 1404.0, + 751.0, + 294.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 745.0, + 1402.0, + 745.0, + 1402.0, + 778.0, + 296.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 775.0, + 1405.0, + 775.0, + 1405.0, + 814.0, + 293.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 807.0, + 857.0, + 807.0, + 857.0, + 840.0, + 294.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 621.0, + 1528.0, + 621.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1528.0, + 769.0, + 1528.0, + 769.0, + 1565.0, + 694.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1528.0, + 1260.0, + 1528.0, + 1260.0, + 1565.0, + 905.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1565.0, + 1288.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1560.0, + 380.0, + 1560.0, + 380.0, + 1593.0, + 323.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1560.0, + 521.0, + 1560.0, + 521.0, + 1593.0, + 401.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1560.0, + 1404.0, + 1560.0, + 1404.0, + 1593.0, + 542.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1624.0, + 296.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1619.0, + 1071.0, + 1619.0, + 1071.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1619.0, + 1405.0, + 1619.0, + 1405.0, + 1656.0, + 1147.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1646.0, + 296.0, + 1646.0, + 296.0, + 1691.0, + 292.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1646.0, + 507.0, + 1646.0, + 507.0, + 1691.0, + 332.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1646.0, + 775.0, + 1646.0, + 775.0, + 1691.0, + 635.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1646.0, + 837.0, + 1646.0, + 837.0, + 1691.0, + 824.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1005.0, + 1407.0, + 1005.0, + 1407.0, + 1040.0, + 293.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1037.0, + 1404.0, + 1037.0, + 1404.0, + 1070.0, + 293.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1066.0, + 1405.0, + 1066.0, + 1405.0, + 1101.0, + 293.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 366.0, + 1097.0, + 366.0, + 1129.0, + 293.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1143.0, + 1405.0, + 1143.0, + 1405.0, + 1177.0, + 294.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 1404.0, + 1176.0, + 1404.0, + 1209.0, + 295.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1239.0, + 295.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1238.0, + 996.0, + 1238.0, + 996.0, + 1271.0, + 297.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1750.0, + 479.0, + 1750.0, + 479.0, + 1786.0, + 318.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1750.0, + 960.0, + 1750.0, + 960.0, + 1786.0, + 503.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1750.0, + 1173.0, + 1750.0, + 1173.0, + 1786.0, + 997.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1750.0, + 1404.0, + 1750.0, + 1404.0, + 1786.0, + 1252.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1781.0, + 738.0, + 1781.0, + 738.0, + 1819.0, + 350.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1781.0, + 898.0, + 1781.0, + 898.0, + 1819.0, + 774.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1781.0, + 1353.0, + 1781.0, + 1353.0, + 1819.0, + 923.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1891.0, + 548.0, + 1891.0, + 548.0, + 1947.0, + 351.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1891.0, + 854.0, + 1891.0, + 854.0, + 1947.0, + 631.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1891.0, + 1412.0, + 1891.0, + 1412.0, + 1947.0, + 924.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1932.0, + 851.0, + 1932.0, + 851.0, + 1967.0, + 406.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1932.0, + 903.0, + 1932.0, + 903.0, + 1967.0, + 891.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1693.0, + 635.0, + 1693.0, + 635.0, + 1737.0, + 294.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1693.0, + 797.0, + 1693.0, + 797.0, + 1737.0, + 656.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1693.0, + 938.0, + 1693.0, + 938.0, + 1737.0, + 818.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1693.0, + 1393.0, + 1693.0, + 1393.0, + 1737.0, + 1075.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1860.0, + 525.0, + 1860.0, + 525.0, + 1896.0, + 314.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1860.0, + 615.0, + 1860.0, + 615.0, + 1896.0, + 542.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1860.0, + 788.0, + 1860.0, + 788.0, + 1896.0, + 640.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 678.0, + 230.0, + 678.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 489, + 1404, + 489, + 1404, + 584, + 299, + 584 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1547, + 1405, + 1547, + 1405, + 1793, + 297, + 1793 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 316, + 791, + 1406, + 791, + 1406, + 1518, + 316, + 1518 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 309, + 344, + 1400, + 344, + 1400, + 408, + 309, + 408 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 703, + 1402, + 703, + 1402, + 765, + 300, + 765 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 687, + 423, + 1064, + 423, + 1064, + 465, + 687, + 465 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 350, + 228, + 1403, + 228, + 1403, + 290, + 350, + 290 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 770, + 291, + 1036, + 291, + 1036, + 332, + 770, + 332 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 106, + 298, + 106 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 630, + 459, + 630, + 459, + 666, + 299, + 666 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 1839, + 557, + 1839, + 557, + 1874, + 300, + 1874 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.761 + }, + { + "category_id": 14, + "poly": [ + 770, + 289, + 1039, + 289, + 1039, + 331, + 770, + 331 + ], + "score": 0.93, + "latex": "h _ { v } ^ { t + 1 } = U _ { t } ( h _ { v } ^ { t } , m _ { v } ^ { t + 1 } ) ." + }, + { + "category_id": 14, + "poly": [ + 687, + 422, + 1068, + 422, + 1068, + 464, + 687, + 464 + ], + "score": 0.91, + "latex": "\\hat { y } = R ( \\{ h _ { v } ^ { t } | v \\in V , t \\in 1 , . . . , T \\} ) ." + }, + { + "category_id": 13, + "poly": [ + 544, + 232, + 618, + 232, + 618, + 259, + 544, + 259 + ], + "score": 0.9, + "latex": "v \\in V" + }, + { + "category_id": 13, + "poly": [ + 408, + 262, + 439, + 262, + 439, + 292, + 408, + 292 + ], + "score": 0.85, + "latex": "U _ { t }" + }, + { + "category_id": 13, + "poly": [ + 401, + 554, + 417, + 554, + 417, + 584, + 401, + 584 + ], + "score": 0.83, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 873, + 493, + 897, + 493, + 897, + 519, + 873, + 519 + ], + "score": 0.81, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 481, + 346, + 505, + 346, + 505, + 373, + 481, + 373 + ], + "score": 0.81, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 332, + 557, + 350, + 557, + 350, + 584, + 332, + 584 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1068, + 377, + 1091, + 377, + 1091, + 403, + 1068, + 403 + ], + "score": 0.8, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 783, + 492, + 815, + 492, + 815, + 522, + 783, + 522 + ], + "score": 0.56, + "latex": "U _ { t }" + }, + { + "category_id": 13, + "poly": [ + 702, + 493, + 723, + 493, + 723, + 520, + 702, + 520 + ], + "score": 0.51, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 734, + 492, + 772, + 492, + 772, + 522, + 734, + 522 + ], + "score": 0.49, + "latex": "M _ { t }" + }, + { + "category_id": 13, + "poly": [ + 708, + 1143, + 733, + 1143, + 733, + 1168, + 708, + 1168 + ], + "score": 0.47, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 649, + 1140, + 795, + 1140, + 795, + 1169, + 649, + 1169 + ], + "score": 0.4, + "latex": "\\mathrm { A S T } + \\mathrm { G S C }" + }, + { + "category_id": 13, + "poly": [ + 705, + 492, + 816, + 492, + 816, + 522, + 705, + 522 + ], + "score": 0.27, + "latex": "S , M _ { t } , U _ { t }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 298.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 628.0, + 466.0, + 628.0, + 466.0, + 672.0, + 292.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1836.0, + 561.0, + 1836.0, + 561.0, + 1880.0, + 292.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1407.0, + 1973.0, + 1407.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1055.0, + 2002.0, + 1055.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 490.0, + 701.0, + 490.0, + 701.0, + 524.0, + 295.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 490.0, + 872.0, + 490.0, + 872.0, + 524.0, + 817.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 490.0, + 1405.0, + 490.0, + 1405.0, + 524.0, + 898.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 555.0, + 293.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 551.0, + 331.0, + 551.0, + 331.0, + 585.0, + 293.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 551.0, + 400.0, + 551.0, + 400.0, + 585.0, + 351.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 551.0, + 430.0, + 551.0, + 430.0, + 585.0, + 418.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1549.0, + 1402.0, + 1549.0, + 1402.0, + 1579.0, + 296.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1579.0, + 1403.0, + 1579.0, + 1403.0, + 1609.0, + 294.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1606.0, + 1403.0, + 1606.0, + 1403.0, + 1642.0, + 294.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 1405.0, + 1639.0, + 1405.0, + 1672.0, + 294.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1704.0, + 291.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1734.0, + 295.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1731.0, + 1405.0, + 1731.0, + 1405.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1758.0, + 719.0, + 1758.0, + 719.0, + 1799.0, + 294.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 789.0, + 1406.0, + 789.0, + 1406.0, + 828.0, + 314.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 823.0, + 498.0, + 823.0, + 498.0, + 855.0, + 350.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 863.0, + 1406.0, + 863.0, + 1406.0, + 905.0, + 311.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 900.0, + 1404.0, + 900.0, + 1404.0, + 931.0, + 352.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 929.0, + 1404.0, + 929.0, + 1404.0, + 964.0, + 349.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 961.0, + 658.0, + 961.0, + 658.0, + 992.0, + 350.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1003.0, + 1406.0, + 1003.0, + 1406.0, + 1038.0, + 314.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1035.0, + 1404.0, + 1035.0, + 1404.0, + 1066.0, + 352.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1065.0, + 1405.0, + 1065.0, + 1405.0, + 1099.0, + 350.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1097.0, + 620.0, + 1097.0, + 620.0, + 1125.0, + 349.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1137.0, + 648.0, + 1137.0, + 648.0, + 1174.0, + 312.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1137.0, + 1405.0, + 1137.0, + 1405.0, + 1174.0, + 796.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1169.0, + 1405.0, + 1169.0, + 1405.0, + 1204.0, + 350.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1199.0, + 1406.0, + 1199.0, + 1406.0, + 1235.0, + 348.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1230.0, + 1404.0, + 1230.0, + 1404.0, + 1265.0, + 350.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1257.0, + 1405.0, + 1257.0, + 1405.0, + 1299.0, + 347.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1287.0, + 1406.0, + 1287.0, + 1406.0, + 1329.0, + 348.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1322.0, + 1405.0, + 1322.0, + 1405.0, + 1356.0, + 350.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1351.0, + 1406.0, + 1351.0, + 1406.0, + 1390.0, + 349.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1382.0, + 1129.0, + 1382.0, + 1129.0, + 1419.0, + 349.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1462.0, + 312.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1456.0, + 1405.0, + 1456.0, + 1405.0, + 1492.0, + 348.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1489.0, + 516.0, + 1489.0, + 516.0, + 1520.0, + 350.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 341.0, + 480.0, + 341.0, + 480.0, + 380.0, + 314.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 341.0, + 1403.0, + 341.0, + 1403.0, + 380.0, + 506.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 377.0, + 1067.0, + 377.0, + 1067.0, + 409.0, + 351.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 377.0, + 1109.0, + 377.0, + 1109.0, + 409.0, + 1092.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1406.0, + 699.0, + 1406.0, + 740.0, + 294.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 733.0, + 999.0, + 733.0, + 999.0, + 768.0, + 295.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 227.0, + 543.0, + 227.0, + 543.0, + 267.0, + 359.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 619.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 258.0, + 407.0, + 258.0, + 407.0, + 300.0, + 402.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 258.0, + 457.0, + 258.0, + 457.0, + 300.0, + 440.0, + 300.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 301, + 1887, + 1403, + 1887, + 1403, + 1951, + 301, + 1951 + ], + "score": 0.942 + }, + { + "category_id": 4, + "poly": [ + 293, + 1638, + 1403, + 1638, + 1403, + 1702, + 293, + 1702 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 297, + 1757, + 927, + 1757, + 927, + 1792, + 297, + 1792 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 301, + 1829, + 822, + 1829, + 822, + 1862, + 301, + 1862 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.866 + }, + { + "category_id": 3, + "poly": [ + 306, + 229, + 1415, + 229, + 1415, + 1619, + 306, + 1619 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 333, + 1974, + 734, + 1974, + 734, + 2034, + 333, + 2034 + ], + "score": 0.766 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.73 + }, + { + "category_id": 2, + "poly": [ + 334, + 1974, + 733, + 1974, + 733, + 2034, + 334, + 2034 + ], + "score": 0.436 + }, + { + "category_id": 3, + "poly": [ + 305, + 225, + 1413, + 225, + 1413, + 1621, + 305, + 1621 + ], + "score": 0.363 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1634.0, + 1405.0, + 1634.0, + 1405.0, + 1678.0, + 292.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1670.0, + 746.0, + 1670.0, + 746.0, + 1702.0, + 295.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1828.0, + 827.0, + 1828.0, + 827.0, + 1865.0, + 294.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 858.0, + 70.0, + 858.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 274.0, + 563.0, + 274.0, + 563.0, + 311.0, + 491.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 320.0, + 563.0, + 320.0, + 563.0, + 350.0, + 377.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 316.0, + 1224.0, + 316.0, + 1224.0, + 345.0, + 1093.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 392.0, + 1108.0, + 392.0, + 1108.0, + 418.0, + 1035.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 392.0, + 1201.0, + 392.0, + 1201.0, + 418.0, + 1123.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 384.0, + 1284.0, + 384.0, + 1284.0, + 424.0, + 1235.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 451.0, + 678.0, + 451.0, + 678.0, + 480.0, + 393.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 438.0, + 1026.0, + 438.0, + 1026.0, + 492.0, + 985.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 452.0, + 1180.0, + 452.0, + 1180.0, + 479.0, + 1134.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 454.0, + 1303.0, + 454.0, + 1303.0, + 480.0, + 1218.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 471.0, + 618.0, + 471.0, + 618.0, + 500.0, + 412.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 491.0, + 415.0, + 491.0, + 415.0, + 517.0, + 391.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 489.0, + 953.0, + 489.0, + 953.0, + 525.0, + 820.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 517.0, + 942.0, + 517.0, + 942.0, + 553.0, + 835.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 508.0, + 1199.0, + 508.0, + 1199.0, + 548.0, + 1149.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 517.0, + 1276.0, + 517.0, + 1276.0, + 541.0, + 1241.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 516.0, + 1380.0, + 516.0, + 1380.0, + 542.0, + 1300.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 564.0, + 1369.0, + 564.0, + 1369.0, + 595.0, + 1309.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 638.0, + 1205.0, + 638.0, + 1205.0, + 673.0, + 999.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 666.0, + 1216.0, + 666.0, + 1216.0, + 700.0, + 988.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 696.0, + 1276.0, + 696.0, + 1276.0, + 708.0, + 1263.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 697.0, + 1251.0, + 697.0, + 1251.0, + 708.0, + 1237.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 794.0, + 695.0, + 794.0, + 695.0, + 820.0, + 648.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 794.0, + 1349.0, + 794.0, + 1349.0, + 820.0, + 1303.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 809.0, + 576.0, + 809.0, + 576.0, + 829.0, + 456.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 812.0, + 711.0, + 812.0, + 711.0, + 837.0, + 632.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 809.0, + 1229.0, + 809.0, + 1229.0, + 829.0, + 1109.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 809.0, + 1366.0, + 809.0, + 1366.0, + 839.0, + 1285.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 887.0, + 361.0, + 887.0, + 361.0, + 912.0, + 328.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 879.0, + 456.0, + 879.0, + 456.0, + 904.0, + 399.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 882.0, + 550.0, + 882.0, + 550.0, + 902.0, + 483.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 871.0, + 639.0, + 871.0, + 639.0, + 912.0, + 590.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 881.0, + 1114.0, + 881.0, + 1114.0, + 905.0, + 1046.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 882.0, + 1205.0, + 882.0, + 1205.0, + 902.0, + 1136.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 871.0, + 1292.0, + 871.0, + 1292.0, + 912.0, + 1243.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 929.0, + 366.0, + 929.0, + 366.0, + 957.0, + 324.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 915.0, + 495.0, + 915.0, + 495.0, + 930.0, + 465.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 947.0, + 524.0, + 947.0, + 524.0, + 960.0, + 505.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 941.0, + 656.0, + 941.0, + 656.0, + 966.0, + 572.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 918.0, + 1007.0, + 918.0, + 1007.0, + 980.0, + 766.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 910.0, + 1188.0, + 910.0, + 1188.0, + 964.0, + 1075.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 941.0, + 1309.0, + 941.0, + 1309.0, + 966.0, + 1226.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 972.0, + 363.0, + 972.0, + 363.0, + 1002.0, + 327.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 971.0, + 1177.0, + 971.0, + 1177.0, + 983.0, + 1166.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1016.0, + 365.0, + 1016.0, + 365.0, + 1044.0, + 325.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 997.0, + 547.0, + 997.0, + 547.0, + 1034.0, + 502.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 997.0, + 631.0, + 997.0, + 631.0, + 1028.0, + 579.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1000.0, + 732.0, + 1000.0, + 732.0, + 1030.0, + 654.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1002.0, + 1193.0, + 1002.0, + 1193.0, + 1034.0, + 1156.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1005.0, + 1286.0, + 1005.0, + 1286.0, + 1028.0, + 1251.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1003.0, + 1385.0, + 1003.0, + 1385.0, + 1028.0, + 1309.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1056.0, + 453.0, + 1056.0, + 453.0, + 1081.0, + 374.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1057.0, + 498.0, + 1057.0, + 498.0, + 1070.0, + 480.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1054.0, + 555.0, + 1054.0, + 555.0, + 1081.0, + 500.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1061.0, + 667.0, + 1061.0, + 667.0, + 1087.0, + 596.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1059.0, + 708.0, + 1059.0, + 708.0, + 1076.0, + 678.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1054.0, + 1208.0, + 1054.0, + 1208.0, + 1081.0, + 1155.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1061.0, + 1320.0, + 1061.0, + 1320.0, + 1087.0, + 1249.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 1059.0, + 1363.0, + 1059.0, + 1363.0, + 1076.0, + 1334.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1130.0, + 875.0, + 1130.0, + 875.0, + 1168.0, + 569.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1200.0, + 1284.0, + 1200.0, + 1284.0, + 1238.0, + 1194.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1228.0, + 1333.0, + 1228.0, + 1333.0, + 1262.0, + 1144.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 1275.0, + 1327.0, + 1275.0, + 1327.0, + 1314.0, + 1151.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1301.0, + 607.0, + 1301.0, + 607.0, + 1315.0, + 593.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1299.0, + 1391.0, + 1299.0, + 1391.0, + 1340.0, + 1083.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1329.0, + 1383.0, + 1329.0, + 1383.0, + 1365.0, + 1092.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1357.0, + 1300.0, + 1357.0, + 1300.0, + 1388.0, + 1178.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1425.0, + 687.0, + 1425.0, + 687.0, + 1446.0, + 667.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1417.0, + 995.0, + 1417.0, + 995.0, + 1481.0, + 776.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1427.0, + 1400.0, + 1427.0, + 1400.0, + 1491.0, + 1078.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1498.0, + 684.0, + 1498.0, + 684.0, + 1512.0, + 668.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1481.0, + 1319.0, + 1481.0, + 1319.0, + 1543.0, + 1150.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 564.5, + 1205.0, + 564.5, + 1205.0, + 596.0, + 1132.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1466.0, + 690.0, + 1466.0, + 690.0, + 1482.5, + 664.0, + 1482.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1968.0, + 736.0, + 1968.0, + 736.0, + 2008.0, + 329.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1999.0, + 736.0, + 1999.0, + 736.0, + 2039.0, + 329.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1968.0, + 736.0, + 1968.0, + 736.0, + 2008.0, + 329.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1999.0, + 736.0, + 1999.0, + 736.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 275.0, + 563.0, + 275.0, + 563.0, + 312.0, + 491.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 321.0, + 563.0, + 321.0, + 563.0, + 351.0, + 378.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 320.0, + 1225.0, + 320.0, + 1225.0, + 345.0, + 1098.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 393.0, + 1104.0, + 393.0, + 1104.0, + 418.0, + 1027.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 393.0, + 1200.0, + 393.0, + 1200.0, + 418.0, + 1123.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 384.0, + 1284.0, + 384.0, + 1284.0, + 424.0, + 1235.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 449.0, + 678.0, + 449.0, + 678.0, + 479.0, + 392.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 437.0, + 1024.0, + 437.0, + 1024.0, + 488.0, + 987.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 452.0, + 1189.0, + 452.0, + 1189.0, + 480.0, + 1129.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 454.0, + 1304.0, + 454.0, + 1304.0, + 480.0, + 1216.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 471.0, + 618.0, + 471.0, + 618.0, + 501.0, + 411.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 496.0, + 409.0, + 496.0, + 409.0, + 512.0, + 397.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 491.0, + 952.0, + 491.0, + 952.0, + 523.0, + 821.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 518.0, + 939.0, + 518.0, + 939.0, + 554.0, + 834.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 512.0, + 1200.0, + 512.0, + 1200.0, + 547.0, + 1147.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 516.0, + 1288.0, + 516.0, + 1288.0, + 543.0, + 1232.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 516.0, + 1381.0, + 516.0, + 1381.0, + 543.0, + 1298.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 552.0, + 1181.0, + 552.0, + 1181.0, + 565.0, + 1167.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 568.0, + 1207.0, + 568.0, + 1207.0, + 594.0, + 1144.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 566.0, + 1365.0, + 566.0, + 1365.0, + 593.0, + 1312.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 638.0, + 1206.0, + 638.0, + 1206.0, + 670.0, + 997.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 667.0, + 1216.0, + 667.0, + 1216.0, + 702.0, + 986.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 697.0, + 1252.0, + 697.0, + 1252.0, + 710.0, + 1240.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 794.0, + 695.0, + 794.0, + 695.0, + 820.0, + 648.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 790.0, + 1349.0, + 790.0, + 1349.0, + 822.0, + 1303.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 808.0, + 577.0, + 808.0, + 577.0, + 833.0, + 453.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 811.0, + 711.0, + 811.0, + 711.0, + 836.0, + 633.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 809.0, + 1230.0, + 809.0, + 1230.0, + 829.0, + 1109.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 811.0, + 1365.0, + 811.0, + 1365.0, + 836.0, + 1285.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 887.0, + 362.0, + 887.0, + 362.0, + 912.0, + 331.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 877.0, + 454.0, + 877.0, + 454.0, + 903.0, + 399.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 879.0, + 552.0, + 879.0, + 552.0, + 904.0, + 480.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 873.0, + 639.0, + 873.0, + 639.0, + 912.0, + 590.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 879.0, + 1112.0, + 879.0, + 1112.0, + 904.0, + 1046.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 879.0, + 1207.0, + 879.0, + 1207.0, + 904.0, + 1133.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 870.0, + 1292.0, + 870.0, + 1292.0, + 912.0, + 1243.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 928.0, + 367.0, + 928.0, + 367.0, + 956.0, + 326.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 912.0, + 499.0, + 912.0, + 499.0, + 934.0, + 456.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 918.0, + 1005.0, + 918.0, + 1005.0, + 953.0, + 766.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 911.0, + 1155.0, + 911.0, + 1155.0, + 935.0, + 1074.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 971.0, + 364.0, + 971.0, + 364.0, + 1002.0, + 329.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 942.0, + 537.0, + 942.0, + 537.0, + 965.0, + 500.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 940.0, + 656.0, + 940.0, + 656.0, + 967.0, + 571.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 943.0, + 990.0, + 943.0, + 990.0, + 982.0, + 781.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 940.0, + 1196.0, + 940.0, + 1196.0, + 982.0, + 1142.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 942.0, + 1310.0, + 942.0, + 1310.0, + 967.0, + 1225.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1015.0, + 367.0, + 1015.0, + 367.0, + 1045.0, + 326.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 995.0, + 554.0, + 995.0, + 554.0, + 1034.0, + 504.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1006.0, + 633.0, + 1006.0, + 633.0, + 1027.0, + 601.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1004.0, + 735.0, + 1004.0, + 735.0, + 1029.0, + 655.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1001.0, + 1192.0, + 1001.0, + 1192.0, + 1034.0, + 1156.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1004.0, + 1285.0, + 1004.0, + 1285.0, + 1027.0, + 1249.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1004.0, + 1386.0, + 1004.0, + 1386.0, + 1029.0, + 1309.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1040.0, + 535.0, + 1040.0, + 535.0, + 1052.0, + 522.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1055.0, + 453.0, + 1055.0, + 453.0, + 1080.0, + 375.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1054.0, + 557.0, + 1054.0, + 557.0, + 1080.0, + 478.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1057.0, + 666.0, + 1057.0, + 666.0, + 1087.0, + 594.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1059.0, + 706.0, + 1059.0, + 706.0, + 1076.0, + 678.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1057.0, + 1208.0, + 1057.0, + 1208.0, + 1079.0, + 1155.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1060.0, + 1320.0, + 1060.0, + 1320.0, + 1087.0, + 1249.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 1062.0, + 1359.0, + 1062.0, + 1359.0, + 1074.0, + 1340.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1130.0, + 876.0, + 1130.0, + 876.0, + 1169.0, + 570.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1200.0, + 1284.0, + 1200.0, + 1284.0, + 1236.0, + 1196.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1228.0, + 1334.0, + 1228.0, + 1334.0, + 1263.0, + 1144.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1278.0, + 1326.0, + 1278.0, + 1326.0, + 1313.0, + 1152.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1300.0, + 1392.0, + 1300.0, + 1392.0, + 1341.0, + 1083.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1331.0, + 1381.0, + 1331.0, + 1381.0, + 1361.0, + 1095.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1369.0, + 1010.0, + 1369.0, + 1010.0, + 1380.0, + 997.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1358.0, + 1299.0, + 1358.0, + 1299.0, + 1387.0, + 1178.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1429.0, + 684.0, + 1429.0, + 684.0, + 1440.0, + 670.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1417.0, + 994.0, + 1417.0, + 994.0, + 1456.0, + 776.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1428.0, + 1327.0, + 1428.0, + 1327.0, + 1469.0, + 1149.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1450.0, + 371.0, + 1450.0, + 371.0, + 1462.0, + 359.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1447.0, + 974.0, + 1447.0, + 974.0, + 1481.0, + 798.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1467.0, + 376.0, + 1467.0, + 376.0, + 1509.0, + 356.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1459.0, + 1400.0, + 1459.0, + 1400.0, + 1520.0, + 1079.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1504.0, + 375.0, + 1504.0, + 375.0, + 1528.0, + 357.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1503.0, + 1318.0, + 1503.0, + 1318.0, + 1543.0, + 1148.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1385.0, + 1029.0, + 1385.0, + 1029.0, + 1400.5, + 986.0, + 1400.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1884.0, + 1405.0, + 1884.0, + 1405.0, + 1922.0, + 296.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1916.0, + 1407.0, + 1916.0, + 1407.0, + 1955.0, + 294.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1755.0, + 933.0, + 1755.0, + 933.0, + 1794.0, + 292.0, + 1794.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 302, + 344, + 1400, + 344, + 1400, + 1249, + 302, + 1249 + ], + "score": 0.984, + "html": "
AbbreviationMeaning
Code RepresentationASTSkips Step 2 in Section 4.
AugASTPerforms Step 2 in Section 4.
Vocab StrategiesClosed VocabSkips Step 3 in Section 4,and instead maintains word-embedding vectors for words in a closed vo- cabulary. In Step 4, name embeddings for nodes representing variables are produced by taking the mean of the embeddings of the words in the vari- able's name.Words outside this model's closed vocabulary are labeled as <UNK>. This is the
CharCNNstrategy used in Allamanis et al. (2018). Skips Step 3 in Section 4.
Pointer SentinelFollows Steps 3 and 4 as described in Section 4,except it doesn't add edges connecting cache nodes to the nodes where their word is used. In the Variable Naming task, this is equivalent to us- ing the Pointer Sentinel Mixture Model of Merity
GSCet al. (2017) to produce outputs. Follows Steps 3 and 4 as described in Section 4.
Graph Neural NetworkGGNNPerforms Step 5 in Section 4 using the Gated Graph Neural Network of Li et al. (2016).
DTNNPerforms Step 5 in Section 4 using the Deep Ten- sor Neural Network of Schuitt et al. (2017).
RGCNPerforms Step 5 in Section 4 using the Relational Graph Convolutional Network of Schlichtkrull et al. (2017).
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1477, + 1403, + 1477, + 1403, + 1601, + 298, + 1601 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1615, + 1406, + 1615, + 1406, + 1890, + 298, + 1890 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1307, + 1404, + 1307, + 1404, + 1462, + 298, + 1462 + ], + "score": 0.969 + }, + { + "category_id": 2, + "poly": [ + 329, + 1913, + 1093, + 1913, + 1093, + 2036, + 329, + 2036 + ], + "score": 0.959 + }, + { + "category_id": 6, + "poly": [ + 297, + 248, + 1404, + 248, + 1404, + 342, + 297, + 342 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.786 + }, + { + "category_id": 13, + "poly": [ + 914, + 1831, + 949, + 1831, + 949, + 1860, + 914, + 1860 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 451, + 1509, + 505, + 1509, + 505, + 1538, + 451, + 1538 + ], + "score": 0.87, + "latex": "15 \\%" + }, + { + "category_id": 13, + "poly": [ + 1091, + 1539, + 1144, + 1539, + 1144, + 1568, + 1091, + 1568 + ], + "score": 0.86, + "latex": "15 \\%" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1769, + 1302, + 1769, + 1302, + 1798, + 1247, + 1798 + ], + "score": 0.85, + "latex": "53 \\%" + }, + { + "category_id": 13, + "poly": [ + 589, + 1399, + 659, + 1399, + 659, + 1430, + 589, + 1430 + ], + "score": 0.85, + "latex": "\\mathrm { C P D ^ { 6 } }" + }, + { + "category_id": 13, + "poly": [ + 907, + 1401, + 948, + 1401, + 948, + 1429, + 907, + 1429 + ], + "score": 0.85, + "latex": "7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1150, + 677, + 1235, + 677, + 1235, + 704, + 1150, + 704 + ], + "score": 0.8, + "latex": "{ \\mathrm { \\ k U N K { > } } }" + }, + { + "category_id": 13, + "poly": [ + 915, + 1771, + 1068, + 1771, + 1068, + 1798, + 915, + 1798 + ], + "score": 0.64, + "latex": "\\mathrm { \\ < N A M E - M E > }" + }, + { + "category_id": 13, + "poly": [ + 447, + 1675, + 547, + 1675, + 547, + 1707, + 447, + 1707 + ], + "score": 0.51, + "latex": "\\mathrm { { \\bf M X N e t } } ^ { 8 }" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1907.0, + 737.0, + 1907.0, + 737.0, + 1950.0, + 328.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1936.0, + 1097.0, + 1936.0, + 1097.0, + 1982.0, + 327.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1967.0, + 693.0, + 1967.0, + 693.0, + 2012.0, + 328.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1997.0, + 721.0, + 1997.0, + 721.0, + 2041.0, + 329.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 248.0, + 1404.0, + 248.0, + 1404.0, + 283.0, + 294.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 1404.0, + 279.0, + 1404.0, + 312.0, + 294.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 308.0, + 1068.0, + 308.0, + 1068.0, + 346.0, + 294.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1476.0, + 1404.0, + 1476.0, + 1404.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1508.0, + 450.0, + 1508.0, + 450.0, + 1542.0, + 292.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1508.0, + 1405.0, + 1508.0, + 1405.0, + 1542.0, + 506.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1537.0, + 1090.0, + 1537.0, + 1090.0, + 1574.0, + 293.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1537.0, + 1407.0, + 1537.0, + 1407.0, + 1574.0, + 1145.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1569.0, + 508.0, + 1569.0, + 508.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1614.0, + 1406.0, + 1614.0, + 1406.0, + 1651.0, + 296.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1646.0, + 1406.0, + 1646.0, + 1406.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1676.0, + 446.0, + 1676.0, + 446.0, + 1712.0, + 295.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1676.0, + 1406.0, + 1676.0, + 1406.0, + 1712.0, + 548.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1707.0, + 1406.0, + 1707.0, + 1406.0, + 1743.0, + 295.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1737.0, + 1406.0, + 1737.0, + 1406.0, + 1773.0, + 295.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1769.0, + 914.0, + 1769.0, + 914.0, + 1802.0, + 296.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1769.0, + 1246.0, + 1769.0, + 1246.0, + 1802.0, + 1069.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 1769.0, + 1404.0, + 1769.0, + 1404.0, + 1802.0, + 1303.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1797.0, + 1404.0, + 1797.0, + 1404.0, + 1836.0, + 292.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1828.0, + 913.0, + 1828.0, + 913.0, + 1868.0, + 293.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1828.0, + 1407.0, + 1828.0, + 1407.0, + 1868.0, + 950.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1856.0, + 386.0, + 1856.0, + 386.0, + 1894.0, + 293.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1307.0, + 1406.0, + 1307.0, + 1406.0, + 1345.0, + 294.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1402.0, + 1339.0, + 1402.0, + 1373.0, + 294.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1369.0, + 1406.0, + 1369.0, + 1406.0, + 1404.0, + 291.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1398.0, + 588.0, + 1398.0, + 588.0, + 1438.0, + 292.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1398.0, + 906.0, + 1398.0, + 906.0, + 1438.0, + 660.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1398.0, + 1406.0, + 1398.0, + 1406.0, + 1438.0, + 949.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1430.0, + 811.0, + 1430.0, + 811.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1156, + 1404, + 1156, + 1404, + 1372, + 297, + 1372 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 353, + 298, + 353 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1755, + 1404, + 1755, + 1404, + 1879, + 297, + 1879 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1535, + 1405, + 1535, + 1405, + 1627, + 298, + 1627 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 445, + 1406, + 445, + 1406, + 539, + 298, + 539 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 1387, + 1398, + 1387, + 1398, + 1481, + 297, + 1481 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 308, + 567, + 1421, + 567, + 1421, + 847, + 308, + 847 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1402, + 1972, + 1402, + 2034, + 297, + 2034 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 665, + 1486, + 1032, + 1486, + 1032, + 1529, + 665, + 1529 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 775, + 1635, + 921, + 1635, + 921, + 1674, + 775, + 1674 + ], + "score": 0.943 + }, + { + "category_id": 4, + "poly": [ + 295, + 884, + 1406, + 884, + 1406, + 1132, + 295, + 1132 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 298, + 1678, + 1402, + 1678, + 1402, + 1742, + 298, + 1742 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 300, + 1914, + 728, + 1914, + 728, + 1947, + 300, + 1947 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 389, + 762, + 389, + 762, + 420, + 300, + 420 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.653 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.278 + }, + { + "category_id": 14, + "poly": [ + 666, + 1485, + 1029, + 1485, + 1029, + 1529, + 666, + 1529 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\hat { y } _ { i } = \\sigma \\left( f _ { 1 } ( { h } _ { v } ^ { T } , { h } _ { v } ^ { 0 } ) \\right) \\odot f _ { 2 } ( { h } _ { v } ^ { T } ) , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 774, + 1632, + 922, + 1632, + 922, + 1674, + 774, + 1674 + ], + "score": 0.91, + "latex": "\\hat { y } _ { i } = f ( h _ { v } ^ { T } ) ," + }, + { + "category_id": 13, + "poly": [ + 570, + 1535, + 605, + 1535, + 605, + 1569, + 570, + 1569 + ], + "score": 0.9, + "latex": "\\boldsymbol { h } _ { v } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1106, + 1312, + 1161, + 1312, + 1161, + 1341, + 1106, + 1341 + ], + "score": 0.86, + "latex": "33 \\%" + }, + { + "category_id": 13, + "poly": [ + 552, + 1312, + 607, + 1312, + 607, + 1341, + 552, + 1341 + ], + "score": 0.86, + "latex": "78 \\%" + }, + { + "category_id": 13, + "poly": [ + 799, + 1394, + 825, + 1394, + 825, + 1421, + 799, + 1421 + ], + "score": 0.86, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1680, + 394, + 1680, + 394, + 1711, + 373, + 1711 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 548, + 1570, + 572, + 1570, + 572, + 1595, + 548, + 1595 + ], + "score": 0.85, + "latex": "\\odot" + }, + { + "category_id": 13, + "poly": [ + 987, + 1539, + 1002, + 1539, + 1002, + 1563, + 987, + 1563 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 908, + 1542, + 927, + 1542, + 927, + 1563, + 908, + 1563 + ], + "score": 0.78, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 1314, + 1542, + 1334, + 1542, + 1334, + 1563, + 1314, + 1563 + ], + "score": 0.76, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1172, + 1391, + 1185, + 1391, + 1185, + 1416, + 1172, + 1416 + ], + "score": 0.73, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 413, + 1537, + 431, + 1537, + 431, + 1568, + 413, + 1568 + ], + "score": 0.61, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 523, + 1450, + 575, + 1450, + 575, + 1478, + 523, + 1478 + ], + "score": 0.6, + "latex": "2 \\ : \\mathrm { d o }" + }, + { + "category_id": 13, + "poly": [ + 725, + 1822, + 744, + 1822, + 744, + 1845, + 725, + 1845 + ], + "score": 0.46, + "latex": ">" + }, + { + "category_id": 13, + "poly": [ + 827, + 2008, + 845, + 2008, + 845, + 2030, + 827, + 2030 + ], + "score": 0.3, + "latex": ">" + }, + { + "category_id": 13, + "poly": [ + 377, + 657, + 429, + 657, + 429, + 675, + 377, + 675 + ], + "score": 0.25, + "latex": "\\left( \\mathsf { n } \\ < \\ 2 \\right)" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 609.0, + 522.0, + 609.0, + 522.0, + 636.0, + 470.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 637.0, + 648.0, + 637.0, + 648.0, + 662.0, + 331.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 638.0, + 900.0, + 638.0, + 900.0, + 649.0, + 890.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 640.0, + 1030.0, + 640.0, + 1030.0, + 657.0, + 1016.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 634.0, + 1346.0, + 634.0, + 1346.0, + 665.0, + 1284.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 656.0, + 376.0, + 656.0, + 376.0, + 677.0, + 346.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 656.0, + 455.0, + 656.0, + 455.0, + 677.0, + 430.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 652.0, + 1033.0, + 652.0, + 1033.0, + 687.0, + 1016.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 667.0, + 629.0, + 667.0, + 629.0, + 760.0, + 346.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 727.0, + 803.0, + 727.0, + 803.0, + 753.0, + 705.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 681.0, + 831.0, + 681.0, + 831.0, + 717.0, + 807.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 670.0, + 863.0, + 670.0, + 863.0, + 718.0, + 846.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 708.0, + 959.0, + 708.0, + 959.0, + 755.0, + 942.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 727.0, + 1156.0, + 727.0, + 1156.0, + 752.0, + 1078.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 682.0, + 1198.0, + 682.0, + 1198.0, + 719.0, + 1174.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 670.0, + 1374.0, + 670.0, + 1374.0, + 736.0, + 1254.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 753.0, + 379.0, + 753.0, + 379.0, + 770.0, + 365.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 749.0, + 791.0, + 749.0, + 791.0, + 776.0, + 719.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 749.0, + 1154.0, + 749.0, + 1154.0, + 772.0, + 1081.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 772.0, + 362.0, + 772.0, + 362.0, + 784.0, + 352.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 768.0, + 796.0, + 768.0, + 796.0, + 794.0, + 714.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 774.0, + 1029.0, + 774.0, + 1029.0, + 789.0, + 1019.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 768.0, + 1155.0, + 768.0, + 1155.0, + 793.0, + 1077.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 802.0, + 345.0, + 802.0, + 345.0, + 815.0, + 336.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 782.0, + 652.0, + 782.0, + 652.0, + 803.0, + 348.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 781.0, + 987.0, + 781.0, + 987.0, + 834.0, + 971.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 787.0, + 1164.0, + 787.0, + 1164.0, + 813.0, + 1070.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 886.0, + 1403.0, + 886.0, + 1403.0, + 920.0, + 295.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 915.0, + 1406.0, + 915.0, + 1406.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 1406.0, + 948.0, + 1406.0, + 981.0, + 295.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 975.0, + 1406.0, + 975.0, + 1406.0, + 1015.0, + 293.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1010.0, + 1403.0, + 1010.0, + 1403.0, + 1041.0, + 295.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1404.0, + 1039.0, + 1404.0, + 1073.0, + 294.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1071.0, + 1404.0, + 1071.0, + 1404.0, + 1101.0, + 297.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 1368.0, + 1100.0, + 1368.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 733.0, + 1913.0, + 733.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 386.0, + 766.0, + 386.0, + 766.0, + 425.0, + 295.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1192.0, + 296.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1189.0, + 1404.0, + 1189.0, + 1404.0, + 1225.0, + 292.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1406.0, + 1216.0, + 1406.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1250.0, + 1406.0, + 1250.0, + 1406.0, + 1285.0, + 294.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1279.0, + 1404.0, + 1279.0, + 1404.0, + 1313.0, + 295.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1312.0, + 551.0, + 1312.0, + 551.0, + 1343.0, + 296.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1312.0, + 1105.0, + 1312.0, + 1105.0, + 1343.0, + 608.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1312.0, + 1404.0, + 1312.0, + 1404.0, + 1343.0, + 1162.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1340.0, + 425.0, + 1340.0, + 425.0, + 1374.0, + 294.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1407.0, + 228.0, + 1407.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 1407.0, + 261.0, + 1407.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 291.0, + 1407.0, + 291.0, + 1407.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1314.0, + 321.0, + 1314.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1755.0, + 1405.0, + 1755.0, + 1405.0, + 1791.0, + 294.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1784.0, + 1405.0, + 1784.0, + 1405.0, + 1822.0, + 292.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 724.0, + 1817.0, + 724.0, + 1852.0, + 292.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1817.0, + 1404.0, + 1817.0, + 1404.0, + 1852.0, + 745.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 779.0, + 1849.0, + 779.0, + 1879.0, + 293.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1532.0, + 412.0, + 1532.0, + 412.0, + 1571.0, + 293.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1532.0, + 569.0, + 1532.0, + 569.0, + 1571.0, + 432.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1532.0, + 907.0, + 1532.0, + 907.0, + 1571.0, + 606.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1532.0, + 986.0, + 1532.0, + 986.0, + 1571.0, + 928.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1532.0, + 1313.0, + 1532.0, + 1313.0, + 1571.0, + 1003.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1532.0, + 1405.0, + 1532.0, + 1405.0, + 1571.0, + 1335.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1565.0, + 547.0, + 1565.0, + 547.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1565.0, + 1405.0, + 1565.0, + 1405.0, + 1601.0, + 573.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1597.0, + 829.0, + 1597.0, + 829.0, + 1630.0, + 296.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 445.0, + 1406.0, + 445.0, + 1406.0, + 482.0, + 292.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 476.0, + 1406.0, + 476.0, + 1406.0, + 510.0, + 292.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 508.0, + 1092.0, + 508.0, + 1092.0, + 542.0, + 296.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1385.0, + 798.0, + 1385.0, + 798.0, + 1427.0, + 293.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1385.0, + 1171.0, + 1385.0, + 1171.0, + 1427.0, + 826.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1385.0, + 1404.0, + 1385.0, + 1404.0, + 1427.0, + 1186.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1418.0, + 1403.0, + 1418.0, + 1403.0, + 1455.0, + 295.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1449.0, + 522.0, + 1449.0, + 522.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1449.0, + 1214.0, + 1449.0, + 1214.0, + 1486.0, + 576.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 826.0, + 2002.0, + 826.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2037.0, + 846.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1672.0, + 372.0, + 1672.0, + 372.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1672.0, + 1406.0, + 1672.0, + 1406.0, + 1716.0, + 395.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1709.0, + 747.0, + 1709.0, + 747.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1680, + 1404, + 1680, + 1404, + 1896, + 297, + 1896 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 449, + 464, + 1249, + 464, + 1249, + 658, + 449, + 658 + ], + "score": 0.976, + "html": "
Closed VocabCharCNNGSC
Seen reposAST0.57 (0.83)0.60 (0.84)0.89 (0.96)
AugAST0.80 (0.90)0.90 (0.94)0.97 (0.99)
Unseen reposAST AugAST0.36 (0.68)0.48 (0.80)0.80 (0.93)
0.59 (0.78)0.84 (0.92)0.92 (0.96)
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 438, + 806, + 1262, + 806, + 1262, + 1000, + 438, + 1000 + ], + "score": 0.972, + "html": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.80 (0.90) 0.97 (0.99)0.72 (0.84) 0.89 (0.95)0.80 (0.90) 0.95 (0.98)
Unseen reposClosed Vocab GSC0.59 (0.78) 0.92 (0.96)0.46 (0.68) 0.80 (0.89)0.62 (0.79) 0.88 (0.95)
" + }, + { + "category_id": 6, + "poly": [ + 294, + 247, + 1406, + 247, + 1406, + 463, + 294, + 463 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 300, + 1157, + 1423, + 1157, + 1423, + 1425, + 300, + 1425 + ], + "score": 0.962 + }, + { + "category_id": 6, + "poly": [ + 303, + 713, + 1401, + 713, + 1401, + 804, + 303, + 804 + ], + "score": 0.959 + }, + { + "category_id": 4, + "poly": [ + 295, + 1461, + 1406, + 1461, + 1406, + 1646, + 295, + 1646 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 294, + 1059, + 1403, + 1059, + 1403, + 1122, + 294, + 1122 + ], + "score": 0.933 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.793 + }, + { + "category_id": 13, + "poly": [ + 1205, + 1911, + 1343, + 1911, + 1343, + 1945, + 1205, + 1945 + ], + "score": 0.93, + "latex": "P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1804, + 1213, + 1804, + 1213, + 1833, + 1137, + 1833 + ], + "score": 0.86, + "latex": "9 9 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 914, + 1944, + 935, + 1944, + 935, + 1970, + 914, + 1970 + ], + "score": 0.81, + "latex": "^ { h }" + }, + { + "category_id": 13, + "poly": [ + 840, + 1913, + 862, + 1913, + 862, + 1939, + 840, + 1939 + ], + "score": 0.75, + "latex": "^ { h }" + }, + { + "category_id": 13, + "poly": [ + 579, + 1526, + 733, + 1526, + 733, + 1553, + 579, + 1553 + ], + "score": 0.53, + "latex": "\\mathrm { \\ < N A M E - M E > }" + }, + { + "category_id": 13, + "poly": [ + 1248, + 1715, + 1402, + 1715, + 1402, + 1742, + 1248, + 1742 + ], + "score": 0.42, + "latex": "\\mathrm { < N A M E - M E > }" + }, + { + "category_id": 13, + "poly": [ + 783, + 345, + 802, + 345, + 802, + 368, + 783, + 368 + ], + "score": 0.3, + "latex": ">" + }, + { + "category_id": 13, + "poly": [ + 401, + 1299, + 444, + 1299, + 444, + 1316, + 401, + 1316 + ], + "score": 0.28, + "latex": "\\dot { \\textbf { 1 } } = \\textbf { \\theta }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 250.0, + 1403.0, + 250.0, + 1403.0, + 281.0, + 296.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 280.0, + 1404.0, + 280.0, + 1404.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 312.0, + 1404.0, + 312.0, + 1404.0, + 343.0, + 296.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 339.0, + 782.0, + 339.0, + 782.0, + 376.0, + 293.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 339.0, + 1404.0, + 339.0, + 1404.0, + 376.0, + 803.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 371.0, + 1404.0, + 371.0, + 1404.0, + 406.0, + 293.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 402.0, + 1403.0, + 402.0, + 1403.0, + 433.0, + 296.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 433.0, + 860.0, + 433.0, + 860.0, + 467.0, + 296.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1209.0, + 509.0, + 1209.0, + 509.0, + 1235.0, + 458.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1227.0, + 980.0, + 1227.0, + 980.0, + 1233.0, + 974.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1216.0, + 1318.0, + 1216.0, + 1318.0, + 1247.0, + 1256.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1236.0, + 633.0, + 1236.0, + 633.0, + 1273.0, + 329.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1240.0, + 986.0, + 1240.0, + 986.0, + 1269.0, + 969.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1254.0, + 1250.0, + 1254.0, + 1250.0, + 1279.0, + 1167.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1254.0, + 1325.0, + 1254.0, + 1325.0, + 1279.0, + 1255.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1255.0, + 1400.0, + 1255.0, + 1400.0, + 1277.0, + 1342.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1264.0, + 515.0, + 1264.0, + 515.0, + 1291.0, + 374.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1264.0, + 918.0, + 1264.0, + 918.0, + 1337.0, + 897.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1282.0, + 1217.0, + 1282.0, + 1217.0, + 1306.0, + 1198.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1282.0, + 1300.0, + 1282.0, + 1300.0, + 1308.0, + 1281.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1283.0, + 1380.0, + 1283.0, + 1380.0, + 1300.0, + 1365.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1296.0, + 400.0, + 1296.0, + 400.0, + 1349.0, + 332.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1296.0, + 625.0, + 1296.0, + 625.0, + 1349.0, + 445.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1308.0, + 772.0, + 1308.0, + 772.0, + 1351.0, + 678.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1297.0, + 825.0, + 1297.0, + 825.0, + 1347.0, + 805.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1340.0, + 940.0, + 1340.0, + 940.0, + 1356.0, + 924.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1303.0, + 1116.0, + 1303.0, + 1116.0, + 1346.0, + 1027.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1318.0, + 1230.0, + 1318.0, + 1230.0, + 1345.0, + 1183.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1318.0, + 1312.0, + 1318.0, + 1312.0, + 1347.0, + 1263.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1318.0, + 1393.0, + 1318.0, + 1393.0, + 1346.0, + 1346.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1347.0, + 767.0, + 1347.0, + 767.0, + 1373.0, + 686.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1363.0, + 944.0, + 1363.0, + 944.0, + 1412.0, + 923.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1405.0, + 710.0, + 1405.0, + 750.0, + 294.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 744.0, + 1404.0, + 744.0, + 1404.0, + 778.0, + 297.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 775.0, + 615.0, + 775.0, + 615.0, + 809.0, + 297.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1463.0, + 1404.0, + 1463.0, + 1404.0, + 1495.0, + 295.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1492.0, + 1402.0, + 1492.0, + 1402.0, + 1524.0, + 295.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1523.0, + 578.0, + 1523.0, + 578.0, + 1557.0, + 293.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1523.0, + 1406.0, + 1523.0, + 1406.0, + 1557.0, + 734.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1554.0, + 1405.0, + 1554.0, + 1405.0, + 1587.0, + 293.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1584.0, + 1406.0, + 1584.0, + 1406.0, + 1620.0, + 295.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 974.0, + 1615.0, + 974.0, + 1650.0, + 294.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1718.0, + 294.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1247.0, + 1713.0, + 1247.0, + 1747.0, + 294.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1404.0, + 1742.0, + 1404.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1807.0, + 294.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1802.0, + 1136.0, + 1802.0, + 1136.0, + 1839.0, + 292.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1802.0, + 1406.0, + 1802.0, + 1406.0, + 1839.0, + 1214.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1837.0, + 1404.0, + 1837.0, + 1404.0, + 1868.0, + 295.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1866.0, + 654.0, + 1866.0, + 654.0, + 1900.0, + 295.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 839.0, + 1909.0, + 839.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1909.0, + 1204.0, + 1909.0, + 1204.0, + 1947.0, + 863.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1947.0, + 1344.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 913.0, + 1942.0, + 913.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1975.0, + 936.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2004.0, + 294.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1403.0, + 2002.0, + 1403.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1055.0, + 1406.0, + 1055.0, + 1406.0, + 1097.0, + 292.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1088.0, + 622.0, + 1088.0, + 622.0, + 1124.0, + 294.0, + 1124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1404, + 1758, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 343, + 958, + 1356, + 958, + 1356, + 1152, + 343, + 1152 + ], + "score": 0.978, + "html": "
Closed VocabCharCNNPointer SentinelGSC
Seen reposAST0.23 (0.31)0.22 (0.28)0.19 (0.33)0.49 (0.67)
AugAST0.19 (0.26)0.20 (0.27)0.26 (0.40)0.53 (0.69)
Unseen reposAST0.05 (0.07)0.06 (0.09)0.06 (0.11)0.38 (0.53)
AugAST0.04 (0.07)0.06 (0.08)0.08 (0.14)0.41 (0.57)
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1650, + 1402, + 1650, + 1402, + 1743, + 300, + 1743 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 399, + 1405, + 399, + 1405, + 617, + 297, + 617 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 436, + 1317, + 1262, + 1317, + 1262, + 1512, + 436, + 1512 + ], + "score": 0.977, + "html": "
GGNNDTNNRGCN
Seen reposClosed Vocab GSC0.19 (0.26)0.23 (0.31)0.27 (0.34)
0.53 (0.69)0.33 (0.48)0.46 (0.63)
Unseen reposClosed Vocab GSC0.04 (0.07) 0.41 (0.57)0.06 (0.08) 0.25 (0.40)0.06 (0.09) 0.35 (0.49)
" + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 350, + 298, + 350 + ], + "score": 0.975 + }, + { + "category_id": 6, + "poly": [ + 298, + 770, + 1403, + 770, + 1403, + 957, + 298, + 957 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 299, + 628, + 1404, + 628, + 1404, + 722, + 299, + 722 + ], + "score": 0.967 + }, + { + "category_id": 6, + "poly": [ + 299, + 1223, + 1403, + 1223, + 1403, + 1315, + 299, + 1315 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 469, + 353, + 1227, + 353, + 1227, + 391, + 469, + 391 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 298, + 1579, + 530, + 1579, + 530, + 1616, + 298, + 1616 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.796 + }, + { + "category_id": 13, + "poly": [ + 376, + 400, + 497, + 400, + 497, + 434, + 376, + 434 + ], + "score": 0.94, + "latex": "P _ { \\mathrm { g r a p h } } ( \\cdot | h )" + }, + { + "category_id": 13, + "poly": [ + 1050, + 582, + 1187, + 582, + 1187, + 616, + 1050, + 616 + ], + "score": 0.93, + "latex": "P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )" + }, + { + "category_id": 13, + "poly": [ + 862, + 583, + 998, + 583, + 998, + 618, + 862, + 618 + ], + "score": 0.92, + "latex": "P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h )" + }, + { + "category_id": 13, + "poly": [ + 462, + 431, + 583, + 431, + 583, + 464, + 462, + 464 + ], + "score": 0.92, + "latex": "P _ { \\mathrm { v o c a b } } ( \\cdot | h )" + }, + { + "category_id": 13, + "poly": [ + 298, + 522, + 420, + 522, + 420, + 556, + 298, + 556 + ], + "score": 0.91, + "latex": "P _ { \\mathrm { v o c a b } } ( \\cdot | h )" + }, + { + "category_id": 13, + "poly": [ + 298, + 461, + 418, + 461, + 418, + 495, + 298, + 495 + ], + "score": 0.91, + "latex": "P _ { \\mathrm { g r a p h } } ( \\cdot | h )" + }, + { + "category_id": 14, + "poly": [ + 471, + 352, + 1229, + 352, + 1229, + 391, + 471, + 391 + ], + "score": 0.83, + "latex": "P ( \\mathbf { w } | h ) = P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) P _ { \\mathrm { g r a p h } } ( \\mathbf { w } | h ) + ( 1 - P _ { \\mathrm { g r a p h } } ( \\mathbf { s } | h ) ) P _ { \\mathrm { v o c a b } } ( \\mathbf { w } | h )" + }, + { + "category_id": 13, + "poly": [ + 980, + 524, + 1002, + 524, + 1002, + 551, + 980, + 551 + ], + "score": 0.81, + "latex": "^ { h }" + }, + { + "category_id": 13, + "poly": [ + 297, + 323, + 319, + 323, + 319, + 348, + 297, + 348 + ], + "score": 0.79, + "latex": "^ { h }" + }, + { + "category_id": 13, + "poly": [ + 1372, + 493, + 1393, + 493, + 1393, + 519, + 1372, + 519 + ], + "score": 0.76, + "latex": "^ { h }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 774.0, + 1404.0, + 774.0, + 1404.0, + 806.0, + 296.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 803.0, + 1403.0, + 803.0, + 1403.0, + 835.0, + 294.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 833.0, + 1405.0, + 833.0, + 1405.0, + 869.0, + 293.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 866.0, + 1405.0, + 866.0, + 1405.0, + 898.0, + 296.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 894.0, + 1405.0, + 894.0, + 1405.0, + 930.0, + 293.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 922.0, + 1270.0, + 922.0, + 1270.0, + 961.0, + 293.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1220.0, + 1408.0, + 1220.0, + 1408.0, + 1259.0, + 293.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1252.0, + 1404.0, + 1252.0, + 1404.0, + 1288.0, + 294.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1286.0, + 584.0, + 1286.0, + 584.0, + 1319.0, + 295.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1577.0, + 534.0, + 1577.0, + 534.0, + 1622.0, + 293.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1796.0, + 294.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1788.0, + 1406.0, + 1788.0, + 1406.0, + 1826.0, + 291.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1822.0, + 1405.0, + 1822.0, + 1405.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1847.0, + 1405.0, + 1847.0, + 1405.0, + 1889.0, + 292.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 292.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1908.0, + 1404.0, + 1908.0, + 1404.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2012.0, + 291.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 567.0, + 2001.0, + 567.0, + 2036.0, + 292.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1649.0, + 1406.0, + 1649.0, + 1406.0, + 1687.0, + 295.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1681.0, + 1403.0, + 1681.0, + 1403.0, + 1715.0, + 295.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1337.0, + 1712.0, + 1337.0, + 1750.0, + 294.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 399.0, + 375.0, + 399.0, + 375.0, + 437.0, + 294.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 399.0, + 1407.0, + 399.0, + 1407.0, + 437.0, + 498.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 429.0, + 461.0, + 429.0, + 461.0, + 467.0, + 292.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 429.0, + 1405.0, + 429.0, + 1405.0, + 467.0, + 584.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 457.0, + 297.0, + 457.0, + 297.0, + 503.0, + 292.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 457.0, + 1407.0, + 457.0, + 1407.0, + 503.0, + 419.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 491.0, + 1371.0, + 491.0, + 1371.0, + 525.0, + 291.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 525.0, + 1394.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 297.0, + 520.0, + 297.0, + 559.0, + 294.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 520.0, + 979.0, + 520.0, + 979.0, + 559.0, + 421.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 520.0, + 1407.0, + 520.0, + 1407.0, + 559.0, + 1003.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 549.0, + 1406.0, + 549.0, + 1406.0, + 586.0, + 291.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 581.0, + 861.0, + 581.0, + 861.0, + 620.0, + 292.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 581.0, + 1049.0, + 581.0, + 1049.0, + 620.0, + 999.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 581.0, + 1403.0, + 581.0, + 1403.0, + 620.0, + 1188.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 316.0, + 296.0, + 316.0, + 296.0, + 357.0, + 290.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 316.0, + 353.0, + 316.0, + 353.0, + 357.0, + 320.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 626.0, + 1404.0, + 626.0, + 1404.0, + 665.0, + 293.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 694.0, + 294.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 691.0, + 500.0, + 691.0, + 500.0, + 721.0, + 295.0, + 721.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 399, + 1403, + 399, + 1403, + 643, + 298, + 643 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 659, + 1404, + 659, + 1404, + 934, + 298, + 934 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 382, + 298, + 382 + ], + "score": 0.965 + }, + { + "category_id": 0, + "poly": [ + 300, + 982, + 487, + 982, + 487, + 1015, + 300, + 1015 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 293, + 1016, + 1405, + 1016, + 1405, + 2033, + 293, + 2033 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.833 + }, + { + "category_id": 13, + "poly": [ + 1346, + 230, + 1401, + 230, + 1401, + 260, + 1346, + 260 + ], + "score": 0.84, + "latex": "30 \\%" + }, + { + "category_id": 13, + "poly": [ + 742, + 1216, + 760, + 1216, + 760, + 1237, + 742, + 1237 + ], + "score": 0.54, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1623, + 1403, + 1623, + 1403, + 1646, + 1383, + 1646 + ], + "score": 0.31, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 984.0, + 489.0, + 984.0, + 489.0, + 1017.0, + 297.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 398.0, + 1407.0, + 398.0, + 1407.0, + 434.0, + 292.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 429.0, + 1407.0, + 429.0, + 1407.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1404.0, + 461.0, + 1404.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 521.0, + 1404.0, + 521.0, + 1404.0, + 554.0, + 296.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 546.0, + 1405.0, + 546.0, + 1405.0, + 590.0, + 291.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 583.0, + 1405.0, + 583.0, + 1405.0, + 617.0, + 294.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 611.0, + 878.0, + 611.0, + 878.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1404.0, + 658.0, + 1404.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 691.0, + 1404.0, + 691.0, + 1404.0, + 724.0, + 296.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 721.0, + 1402.0, + 721.0, + 1402.0, + 754.0, + 296.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 1405.0, + 750.0, + 1405.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 1404.0, + 781.0, + 1404.0, + 817.0, + 293.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 809.0, + 1404.0, + 809.0, + 1404.0, + 849.0, + 293.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 839.0, + 1406.0, + 839.0, + 1406.0, + 879.0, + 293.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 912.0, + 291.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 902.0, + 413.0, + 902.0, + 413.0, + 939.0, + 294.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1345.0, + 229.0, + 1345.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 1406.0, + 258.0, + 1406.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 319.0, + 1407.0, + 319.0, + 1407.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 416.0, + 352.0, + 416.0, + 386.0, + 294.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1031.0, + 1405.0, + 1031.0, + 1405.0, + 1072.0, + 292.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1063.0, + 1401.0, + 1063.0, + 1401.0, + 1102.0, + 321.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1097.0, + 637.0, + 1097.0, + 637.0, + 1129.0, + 325.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1145.0, + 1405.0, + 1145.0, + 1405.0, + 1189.0, + 292.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1178.0, + 1403.0, + 1178.0, + 1403.0, + 1219.0, + 321.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1213.0, + 741.0, + 1213.0, + 741.0, + 1248.0, + 325.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1213.0, + 922.0, + 1213.0, + 922.0, + 1248.0, + 761.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1263.0, + 1405.0, + 1263.0, + 1405.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1294.0, + 1407.0, + 1294.0, + 1407.0, + 1336.0, + 319.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1328.0, + 1404.0, + 1328.0, + 1404.0, + 1362.0, + 321.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1352.0, + 1403.0, + 1352.0, + 1403.0, + 1396.0, + 320.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1389.0, + 953.0, + 1389.0, + 953.0, + 1420.0, + 326.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1441.0, + 1403.0, + 1441.0, + 1403.0, + 1478.0, + 294.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1473.0, + 1402.0, + 1473.0, + 1402.0, + 1508.0, + 324.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1502.0, + 424.0, + 1502.0, + 424.0, + 1537.0, + 323.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1554.0, + 1405.0, + 1554.0, + 1405.0, + 1596.0, + 292.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1589.0, + 1403.0, + 1589.0, + 1403.0, + 1624.0, + 321.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1620.0, + 1382.0, + 1620.0, + 1382.0, + 1655.0, + 323.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1652.0, + 488.0, + 1652.0, + 488.0, + 1684.0, + 324.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1703.0, + 1403.0, + 1703.0, + 1403.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1733.0, + 1402.0, + 1733.0, + 1402.0, + 1769.0, + 321.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1766.0, + 421.0, + 1766.0, + 421.0, + 1795.0, + 324.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1403.0, + 1819.0, + 1403.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1886.0, + 320.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1917.0, + 323.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1950.0, + 320.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1977.0, + 321.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2008.0, + 321.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 2004.0, + 772.0, + 2004.0, + 772.0, + 2036.0, + 326.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 552, + 1403, + 552, + 1403, + 677, + 297, + 677 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 299, + 1792, + 1404, + 1792, + 1404, + 1916, + 299, + 1916 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1402, + 1942, + 1402, + 2033, + 299, + 2033 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 297, + 788, + 1404, + 788, + 1404, + 972, + 297, + 972 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 297, + 1438, + 1404, + 1438, + 1404, + 1593, + 297, + 1593 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 297, + 997, + 1404, + 997, + 1404, + 1154, + 297, + 1154 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 297, + 433, + 1404, + 433, + 1404, + 527, + 297, + 527 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 298, + 1233, + 1404, + 1233, + 1404, + 1327, + 298, + 1327 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1399, + 229, + 1399, + 322, + 298, + 322 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 292, + 702, + 1401, + 702, + 1401, + 765, + 292, + 765 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 299, + 1704, + 1402, + 1704, + 1402, + 1769, + 299, + 1769 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 295, + 1351, + 1402, + 1351, + 1402, + 1415, + 295, + 1415 + ], + "score": 0.864 + }, + { + "category_id": 1, + "poly": [ + 299, + 347, + 1398, + 347, + 1398, + 410, + 299, + 410 + ], + "score": 0.864 + }, + { + "category_id": 1, + "poly": [ + 296, + 1617, + 1400, + 1617, + 1400, + 1682, + 296, + 1682 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 294, + 1175, + 1389, + 1175, + 1389, + 1211, + 294, + 1211 + ], + "score": 0.828 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.82 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.774 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.133 + }, + { + "category_id": 13, + "poly": [ + 975, + 737, + 994, + 737, + 994, + 758, + 975, + 758 + ], + "score": 0.62, + "latex": "{ . } = { }" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1858, + 1403, + 1858, + 1403, + 1881, + 1382, + 1881 + ], + "score": 0.59, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1740, + 1185, + 1740, + 1185, + 1763, + 1166, + 1763 + ], + "score": 0.5, + "latex": "{ . } =" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 550.0, + 1407.0, + 550.0, + 1407.0, + 590.0, + 294.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 583.0, + 1403.0, + 583.0, + 1403.0, + 619.0, + 321.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 611.0, + 1403.0, + 611.0, + 1403.0, + 651.0, + 321.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 646.0, + 820.0, + 646.0, + 820.0, + 679.0, + 325.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1832.0, + 294.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1824.0, + 1403.0, + 1824.0, + 1403.0, + 1856.0, + 323.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1852.0, + 1381.0, + 1852.0, + 1381.0, + 1890.0, + 321.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1852.0, + 1408.0, + 1852.0, + 1408.0, + 1890.0, + 1404.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1883.0, + 589.0, + 1883.0, + 589.0, + 1916.0, + 323.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1972.0, + 1399.0, + 1972.0, + 1399.0, + 2005.0, + 324.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2002.0, + 408.0, + 2002.0, + 408.0, + 2034.0, + 324.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 789.0, + 1404.0, + 789.0, + 1404.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 819.0, + 1406.0, + 819.0, + 1406.0, + 856.0, + 320.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 848.0, + 1404.0, + 848.0, + 1404.0, + 885.0, + 322.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 880.0, + 1404.0, + 880.0, + 1404.0, + 916.0, + 322.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 908.0, + 1404.0, + 908.0, + 1404.0, + 950.0, + 320.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 943.0, + 473.0, + 943.0, + 473.0, + 973.0, + 322.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1434.0, + 1406.0, + 1434.0, + 1406.0, + 1478.0, + 292.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1466.0, + 1404.0, + 1466.0, + 1404.0, + 1508.0, + 321.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1537.0, + 320.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1527.0, + 1402.0, + 1527.0, + 1402.0, + 1569.0, + 320.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1563.0, + 473.0, + 1563.0, + 473.0, + 1593.0, + 322.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 999.0, + 1404.0, + 999.0, + 1404.0, + 1033.0, + 295.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1026.0, + 1405.0, + 1026.0, + 1405.0, + 1069.0, + 320.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1061.0, + 1404.0, + 1061.0, + 1404.0, + 1094.0, + 325.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1089.0, + 1402.0, + 1089.0, + 1402.0, + 1125.0, + 321.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1120.0, + 1399.0, + 1120.0, + 1399.0, + 1157.0, + 324.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 433.0, + 1402.0, + 433.0, + 1402.0, + 469.0, + 294.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 464.0, + 1401.0, + 464.0, + 1401.0, + 501.0, + 322.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 497.0, + 506.0, + 497.0, + 506.0, + 527.0, + 324.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1233.0, + 1403.0, + 1233.0, + 1403.0, + 1269.0, + 295.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1261.0, + 1404.0, + 1261.0, + 1404.0, + 1304.0, + 322.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1296.0, + 998.0, + 1296.0, + 998.0, + 1326.0, + 325.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1403.0, + 260.0, + 1403.0, + 297.0, + 322.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 638.0, + 292.0, + 638.0, + 324.0, + 323.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1404.0, + 699.0, + 1404.0, + 739.0, + 294.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 730.0, + 974.0, + 730.0, + 974.0, + 765.0, + 324.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 730.0, + 1155.0, + 730.0, + 1155.0, + 765.0, + 995.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1700.0, + 1407.0, + 1700.0, + 1407.0, + 1746.0, + 293.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1734.0, + 1165.0, + 1734.0, + 1165.0, + 1770.0, + 324.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1734.0, + 1344.0, + 1734.0, + 1344.0, + 1770.0, + 1186.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1346.0, + 1406.0, + 1346.0, + 1406.0, + 1391.0, + 293.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1382.0, + 861.0, + 1382.0, + 861.0, + 1416.0, + 322.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 344.0, + 1404.0, + 344.0, + 1404.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 379.0, + 911.0, + 379.0, + 911.0, + 410.0, + 325.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1615.0, + 1405.0, + 1615.0, + 1405.0, + 1653.0, + 296.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1648.0, + 942.0, + 1648.0, + 942.0, + 1682.0, + 322.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1172.0, + 1389.0, + 1172.0, + 1389.0, + 1214.0, + 294.0, + 1214.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 300, + 1971, + 1399, + 1971, + 1399, + 2034, + 300, + 2034 + ], + "score": 0.74 + }, + { + "category_id": 1, + "poly": [ + 295, + 1434, + 1400, + 1434, + 1400, + 1500, + 295, + 1500 + ], + "score": 0.731 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.728 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1406, + 229, + 1406, + 385, + 299, + 385 + ], + "score": 0.698 + }, + { + "category_id": 1, + "poly": [ + 298, + 876, + 1404, + 876, + 1404, + 971, + 298, + 971 + ], + "score": 0.686 + }, + { + "category_id": 1, + "poly": [ + 297, + 766, + 1405, + 766, + 1405, + 860, + 297, + 860 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 299, + 1323, + 1402, + 1323, + 1402, + 1418, + 299, + 1418 + ], + "score": 0.68 + }, + { + "category_id": 1, + "poly": [ + 297, + 401, + 1404, + 401, + 1404, + 526, + 297, + 526 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 297, + 1099, + 1404, + 1099, + 1404, + 1195, + 297, + 1195 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 297, + 1514, + 1405, + 1514, + 1405, + 1670, + 297, + 1670 + ], + "score": 0.645 + }, + { + "category_id": 1, + "poly": [ + 297, + 543, + 1405, + 543, + 1405, + 637, + 297, + 637 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 298, + 1860, + 1401, + 1860, + 1401, + 1954, + 298, + 1954 + ], + "score": 0.61 + }, + { + "category_id": 1, + "poly": [ + 297, + 1687, + 1405, + 1687, + 1405, + 1842, + 297, + 1842 + ], + "score": 0.599 + }, + { + "category_id": 1, + "poly": [ + 296, + 1211, + 1400, + 1211, + 1400, + 1305, + 296, + 1305 + ], + "score": 0.541 + }, + { + "category_id": 1, + "poly": [ + 299, + 655, + 1403, + 655, + 1403, + 748, + 299, + 748 + ], + "score": 0.519 + }, + { + "category_id": 1, + "poly": [ + 296, + 988, + 1405, + 988, + 1405, + 1084, + 296, + 1084 + ], + "score": 0.306 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.196 + }, + { + "category_id": 13, + "poly": [ + 426, + 833, + 444, + 833, + 444, + 855, + 426, + 855 + ], + "score": 0.46, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } }" + }, + { + "category_id": 13, + "poly": [ + 1216, + 802, + 1235, + 802, + 1235, + 823, + 1216, + 823 + ], + "score": 0.39, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 13, + "poly": [ + 1249, + 1247, + 1268, + 1247, + 1268, + 1270, + 1249, + 1270 + ], + "score": 0.32, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 13, + "poly": [ + 675, + 1389, + 693, + 1389, + 693, + 1411, + 675, + 1411 + ], + "score": 0.28, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 13, + "poly": [ + 643, + 497, + 666, + 497, + 666, + 522, + 643, + 522 + ], + "score": 0.28, + "latex": "= \\mathfrak { C }" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1972.0, + 1402.0, + 1972.0, + 1402.0, + 2004.0, + 297.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2036.0, + 325.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1433.0, + 1404.0, + 1433.0, + 1404.0, + 1471.0, + 295.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1464.0, + 1389.0, + 1464.0, + 1389.0, + 1502.0, + 322.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 298.0, + 324.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 287.0, + 1406.0, + 287.0, + 1406.0, + 331.0, + 319.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 320.0, + 1406.0, + 320.0, + 1406.0, + 358.0, + 322.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 352.0, + 1006.0, + 352.0, + 1006.0, + 390.0, + 322.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 877.0, + 1404.0, + 877.0, + 1404.0, + 911.0, + 296.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 908.0, + 1405.0, + 908.0, + 1405.0, + 945.0, + 322.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 936.0, + 401.0, + 936.0, + 401.0, + 972.0, + 320.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 766.0, + 1405.0, + 766.0, + 1405.0, + 803.0, + 296.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 796.0, + 1215.0, + 796.0, + 1215.0, + 832.0, + 323.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 796.0, + 1403.0, + 796.0, + 1403.0, + 832.0, + 1236.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 826.0, + 425.0, + 826.0, + 425.0, + 864.0, + 323.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 826.0, + 607.0, + 826.0, + 607.0, + 864.0, + 445.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1323.0, + 1406.0, + 1323.0, + 1406.0, + 1361.0, + 295.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1355.0, + 1401.0, + 1355.0, + 1401.0, + 1389.0, + 323.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1388.0, + 674.0, + 1388.0, + 674.0, + 1422.0, + 325.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1388.0, + 854.0, + 1388.0, + 854.0, + 1422.0, + 694.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 438.0, + 292.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 429.0, + 1406.0, + 429.0, + 1406.0, + 472.0, + 321.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 461.0, + 1402.0, + 461.0, + 1402.0, + 500.0, + 321.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 495.0, + 642.0, + 495.0, + 642.0, + 528.0, + 322.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 495.0, + 888.0, + 495.0, + 888.0, + 528.0, + 667.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1099.0, + 1404.0, + 1099.0, + 1404.0, + 1137.0, + 294.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1165.0, + 324.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1163.0, + 856.0, + 1163.0, + 856.0, + 1197.0, + 324.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1515.0, + 1403.0, + 1515.0, + 1403.0, + 1551.0, + 295.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1548.0, + 1403.0, + 1548.0, + 1403.0, + 1581.0, + 324.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1575.0, + 1403.0, + 1575.0, + 1403.0, + 1612.0, + 324.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1604.0, + 1402.0, + 1604.0, + 1402.0, + 1644.0, + 321.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1637.0, + 640.0, + 1637.0, + 640.0, + 1675.0, + 323.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 1406.0, + 543.0, + 1406.0, + 580.0, + 294.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 575.0, + 1402.0, + 575.0, + 1402.0, + 609.0, + 323.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 603.0, + 506.0, + 603.0, + 506.0, + 637.0, + 323.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1861.0, + 1405.0, + 1861.0, + 1405.0, + 1895.0, + 294.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1889.0, + 1405.0, + 1889.0, + 1405.0, + 1928.0, + 323.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1922.0, + 1134.0, + 1922.0, + 1134.0, + 1956.0, + 324.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1408.0, + 1682.0, + 1408.0, + 1729.0, + 294.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1717.0, + 1407.0, + 1717.0, + 1407.0, + 1757.0, + 321.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1749.0, + 1405.0, + 1749.0, + 1405.0, + 1784.0, + 320.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1776.0, + 1403.0, + 1776.0, + 1403.0, + 1816.0, + 321.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1811.0, + 589.0, + 1811.0, + 589.0, + 1844.0, + 323.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1213.0, + 1402.0, + 1213.0, + 1402.0, + 1243.0, + 297.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1242.0, + 1248.0, + 1242.0, + 1248.0, + 1277.0, + 320.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1242.0, + 1401.0, + 1242.0, + 1401.0, + 1277.0, + 1269.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1274.0, + 457.0, + 1274.0, + 457.0, + 1305.0, + 323.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 656.0, + 1403.0, + 656.0, + 1403.0, + 689.0, + 297.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 685.0, + 1402.0, + 685.0, + 1402.0, + 721.0, + 321.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 717.0, + 739.0, + 717.0, + 739.0, + 750.0, + 324.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 991.0, + 1403.0, + 991.0, + 1403.0, + 1025.0, + 296.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1020.0, + 1403.0, + 1020.0, + 1403.0, + 1054.0, + 322.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1049.0, + 1144.0, + 1049.0, + 1144.0, + 1087.0, + 323.0, + 1087.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1092, + 1465, + 1092, + 1465, + 1279, + 296, + 1279 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 384, + 298, + 384 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 298, + 918, + 1405, + 918, + 1405, + 1072, + 298, + 1072 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 297, + 774, + 1403, + 774, + 1403, + 899, + 297, + 899 + ], + "score": 0.801 + }, + { + "category_id": 1, + "poly": [ + 297, + 630, + 1405, + 630, + 1405, + 755, + 297, + 755 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 298, + 486, + 1404, + 486, + 1404, + 612, + 298, + 612 + ], + "score": 0.754 + }, + { + "category_id": 1, + "poly": [ + 298, + 402, + 1398, + 402, + 1398, + 467, + 298, + 467 + ], + "score": 0.744 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.508 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.396 + }, + { + "category_id": 13, + "poly": [ + 1355, + 231, + 1395, + 231, + 1395, + 262, + 1355, + 262 + ], + "score": 0.38, + "latex": "\\mathrm { N g }" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1092.0, + 894.0, + 1092.0, + 894.0, + 1128.0, + 295.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1093.0, + 1403.0, + 1093.0, + 1403.0, + 1125.0, + 933.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1123.0, + 1404.0, + 1123.0, + 1404.0, + 1157.0, + 320.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1151.0, + 1404.0, + 1151.0, + 1404.0, + 1189.0, + 319.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1183.0, + 804.0, + 1183.0, + 804.0, + 1219.0, + 321.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1183.0, + 1403.0, + 1183.0, + 1403.0, + 1220.0, + 848.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1212.0, + 1467.0, + 1212.0, + 1467.0, + 1250.0, + 320.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1245.0, + 392.0, + 1245.0, + 392.0, + 1279.0, + 315.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 1354.0, + 224.0, + 1354.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 224.0, + 1405.0, + 224.0, + 1405.0, + 269.0, + 1396.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 296.0, + 322.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 285.0, + 1406.0, + 285.0, + 1406.0, + 330.0, + 320.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 357.0, + 322.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 353.0, + 1276.0, + 353.0, + 1276.0, + 386.0, + 323.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 917.0, + 1403.0, + 917.0, + 1403.0, + 954.0, + 296.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 950.0, + 1405.0, + 950.0, + 1405.0, + 983.0, + 322.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 976.0, + 1405.0, + 976.0, + 1405.0, + 1018.0, + 321.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1008.0, + 1401.0, + 1008.0, + 1401.0, + 1045.0, + 323.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1041.0, + 1395.0, + 1041.0, + 1395.0, + 1074.0, + 323.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 774.0, + 1403.0, + 774.0, + 1403.0, + 811.0, + 293.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 805.0, + 1405.0, + 805.0, + 1405.0, + 842.0, + 320.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 836.0, + 1404.0, + 836.0, + 1404.0, + 873.0, + 322.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 866.0, + 1341.0, + 866.0, + 1341.0, + 903.0, + 321.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 628.0, + 1405.0, + 628.0, + 1405.0, + 667.0, + 292.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 660.0, + 1403.0, + 660.0, + 1403.0, + 697.0, + 323.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 694.0, + 1401.0, + 694.0, + 1401.0, + 726.0, + 324.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 724.0, + 656.0, + 724.0, + 656.0, + 757.0, + 322.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 484.0, + 1408.0, + 484.0, + 1408.0, + 523.0, + 293.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 516.0, + 1405.0, + 516.0, + 1405.0, + 555.0, + 320.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 548.0, + 1401.0, + 548.0, + 1401.0, + 583.0, + 322.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 581.0, + 905.0, + 581.0, + 905.0, + 614.0, + 325.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 1402.0, + 398.0, + 1402.0, + 441.0, + 293.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 436.0, + 541.0, + 436.0, + 541.0, + 468.0, + 324.0, + 468.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 296, + 504, + 1425, + 504, + 1425, + 699, + 296, + 699 + ], + "score": 0.976, + "html": "
Closed VocabCharCNNPointer SentinelGSC (ours)
Seen reposAST AugAST0.30, 7.22, 0.94 0.26, 7.64, 0.940.28,8.67, 1.08 0.28, 7.46, 0.990.32, 8.00, 1.07 0.35, 6.51, 0.770.56, 3.87, 0.39 0.60, 3.68, 0.37
Unseen reposAST0.09,8.66,1.230.10, 8.82, 1.120.13, 9,39, 1.370.42, 4.81, 0.59
AugAST0.09, 8.34, 1.140.10, 8.16, 1.120.14, 8.03, 1.070.48, 4.28, 0.49
" + }, + { + "category_id": 6, + "poly": [ + 295, + 385, + 1410, + 385, + 1410, + 503, + 295, + 503 + ], + "score": 0.965 + }, + { + "category_id": 6, + "poly": [ + 296, + 844, + 1402, + 844, + 1402, + 908, + 296, + 908 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 301, + 296, + 795, + 296, + 795, + 328, + 301, + 328 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.836 + }, + { + "category_id": 0, + "poly": [ + 301, + 228, + 450, + 228, + 450, + 262, + 301, + 262 + ], + "score": 0.816 + }, + { + "category_id": 2, + "poly": [ + 298, + 1586, + 568, + 1586, + 568, + 1615, + 298, + 1615 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 299, + 756, + 586, + 756, + 586, + 786, + 299, + 786 + ], + "score": 0.656 + }, + { + "category_id": 7, + "poly": [ + 572, + 1431, + 952, + 1431, + 952, + 1522, + 572, + 1522 + ], + "score": 0.369 + }, + { + "category_id": 1, + "poly": [ + 298, + 1586, + 568, + 1586, + 568, + 1615, + 298, + 1615 + ], + "score": 0.239 + }, + { + "category_id": 0, + "poly": [ + 299, + 756, + 586, + 756, + 586, + 786, + 299, + 786 + ], + "score": 0.225 + }, + { + "category_id": 5, + "poly": [ + 559, + 907, + 1136, + 907, + 1136, + 1529, + 559, + 1529 + ], + "score": 0.117, + "html": "
Seen Repos
com.fasterxml.jackson.core/jackson-core/2.9.5 com.h2database/h2/1.4.195 javax.enterprise/cdi-api/2.0 junit/junit/4.12 mysql/mysql-connector-java/6.0.6 org.apache.commons/commons-collections4/4.1 org.apache.commons/commons-math3/3.6.1 org.apache.commons/commons-pool2/2.5.0 org.apache.maven/maven-project/2.2.1 org.codehaus.plexus/plexus-utils/3.1.0 org.eclipse.jetty/jetty-server/9.4.9.v20180320
org.scalacheck/scalacheck_2.12/1.13.5 org.slf4j/slf4j-api/1.7.25 org.slf4j/slf4j-log4j12/1.7.25 Unseen Repos
org.javassist/javassist/3.22.0-GA joda-time/joda-time/2.9.9 org.mockito/mockito-core/2.17.0
" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 383.0, + 1406.0, + 383.0, + 1406.0, + 422.0, + 296.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 417.0, + 1404.0, + 417.0, + 1404.0, + 451.0, + 296.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 445.0, + 1403.0, + 445.0, + 1403.0, + 482.0, + 294.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 479.0, + 420.0, + 479.0, + 420.0, + 511.0, + 294.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 840.0, + 1403.0, + 840.0, + 1403.0, + 882.0, + 294.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 876.0, + 1266.0, + 876.0, + 1266.0, + 912.0, + 296.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 223.0, + 454.0, + 223.0, + 454.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1581.0, + 571.0, + 1581.0, + 571.0, + 1619.0, + 295.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1430.0, + 950.0, + 1430.0, + 950.0, + 1460.0, + 576.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1459.0, + 871.0, + 1459.0, + 871.0, + 1490.0, + 575.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1490.0, + 953.0, + 1490.0, + 953.0, + 1522.0, + 576.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 754.0, + 589.0, + 754.0, + 589.0, + 788.0, + 297.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 294.0, + 799.0, + 294.0, + 799.0, + 332.0, + 296.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 754.0, + 589.0, + 754.0, + 589.0, + 788.0, + 297.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1581.0, + 571.0, + 1581.0, + 571.0, + 1619.0, + 295.0, + 1619.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 363, + 620, + 1311, + 620, + 1311, + 1837, + 363, + 1837 + ], + "score": 0.981, + "html": "
Edge NameDescription
AST NEXT_TOKENThe edges used to construct the original AST. Edges added to the original AST that specify the left-to-
right ordering of the children of a node in the AST.These edges are necessary since ASTs have ordered children, but we are representing the AST as a directed multigraph.
COMPUTED_FROMConnects a node representing a variable on the left of an equality to those on the right. (E.g. edges from y to x and Z to X in x = y + z.) The same as in Allamanis et al. (2018).
LAST_READConnects a node representing a usage of a variable to all nodes in the AST at which that variable's value could have been last read from memory. The same as in Allamanis et al. (2018).
LAST_WRITEConnects a node representing a usage of a variable to all nodesin the AST at which that variable's value could have been last written to memory. The same as in Allamanis et al.
RETURNS_TO(2018). Points a node in a return statement to the node contain- ing the return type of the method. (E.g. x in return X gets an edge pointing to int in public static int
LAST_SCOPE_USEgetx(x).) Connects a node representing a variable to the node repre- senting the last time this variable's name was used in the text of the code (i.e. capturing information about the text, not the control flow),but only within lexical scope. This edge exists to try and give the non-GSC models as much
LAST_FIELD_LEXlexical information as possible to make them as comparable with the GSC model. Connects a field access (e.g. this.whatever or Foo .whatever) node to the last use of this.whatever (or to the variable's initialization,if it's the first use).This is not lexical-scope aware (and,in fact,can't be in Java,in
FIELDgeneral). Pointseach node representing a field access(e.g. this.whatever) to the node where that field was de-
WORD_USEclared. Points cache nodes to nodes representing variables in which the vocab word was used in the variable's name.
" + }, + { + "category_id": 6, + "poly": [ + 295, + 434, + 1406, + 434, + 1406, + 620, + 295, + 620 + ], + "score": 0.975 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 13, + "poly": [ + 759, + 888, + 925, + 888, + 925, + 919, + 759, + 919 + ], + "score": 0.85, + "latex": "\\mathrm { ~ ~ x ~ } = \\mathrm { ~ ~ y ~ } + \\mathrm { ~ ~ z . ~ } )" + }, + { + "category_id": 13, + "poly": [ + 700, + 891, + 722, + 891, + 722, + 915, + 700, + 915 + ], + "score": 0.52, + "latex": "_ \\textrm { x }" + }, + { + "category_id": 13, + "poly": [ + 1238, + 860, + 1261, + 860, + 1261, + 884, + 1238, + 884 + ], + "score": 0.5, + "latex": "_ \\textrm { x }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 436.0, + 1403.0, + 436.0, + 1403.0, + 468.0, + 295.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 467.0, + 1406.0, + 467.0, + 1406.0, + 503.0, + 294.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 495.0, + 1406.0, + 495.0, + 1406.0, + 532.0, + 293.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 528.0, + 1404.0, + 528.0, + 1404.0, + 560.0, + 294.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 559.0, + 1406.0, + 559.0, + 1406.0, + 595.0, + 295.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 588.0, + 935.0, + 588.0, + 935.0, + 624.0, + 294.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SkzK4iC5Ym/images/00a9bdecebeb17bac6895b304252ade06d7fa501f7f20b54dec61fbf0d5a0153.jpg b/parse/train/SkzK4iC5Ym/images/00a9bdecebeb17bac6895b304252ade06d7fa501f7f20b54dec61fbf0d5a0153.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6140a985de6c804c0614957e49e459bc012218f5 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/00a9bdecebeb17bac6895b304252ade06d7fa501f7f20b54dec61fbf0d5a0153.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32e16bdbd5cb89d708ad2ee82e40e28d7bbec1462b6221bb51c3a980ad769aa6 +size 6667 diff --git a/parse/train/SkzK4iC5Ym/images/02bd2a93a635bd7ce79b262523dbc96947efc3ce351c270569613d7dfdb2cd86.jpg b/parse/train/SkzK4iC5Ym/images/02bd2a93a635bd7ce79b262523dbc96947efc3ce351c270569613d7dfdb2cd86.jpg new file mode 100644 index 0000000000000000000000000000000000000000..29e999117fea4722b9554e2aac4a884234a812f7 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/02bd2a93a635bd7ce79b262523dbc96947efc3ce351c270569613d7dfdb2cd86.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a95fc2f5665a5011792756a96c58dfcaec31aab9dcb0c61479dd91a7eb7957ba +size 7492 diff --git a/parse/train/SkzK4iC5Ym/images/060901203410993def2b3a46a36f7ef4b16cfb2f7cf73fcef790b1f53ec3647e.jpg b/parse/train/SkzK4iC5Ym/images/060901203410993def2b3a46a36f7ef4b16cfb2f7cf73fcef790b1f53ec3647e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..58fddab9123fbc743d1d216eea4e8383224a2784 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/060901203410993def2b3a46a36f7ef4b16cfb2f7cf73fcef790b1f53ec3647e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b91ec0dfec563c4ca1b8424d1d6b2bc8bf9cafc3fc0c09cd316cfabfd38d79f1 +size 13299 diff --git a/parse/train/SkzK4iC5Ym/images/0889f958adfa68466caf49663390c715db916114d182679b1cbcb25da6335e23.jpg b/parse/train/SkzK4iC5Ym/images/0889f958adfa68466caf49663390c715db916114d182679b1cbcb25da6335e23.jpg new file mode 100644 index 0000000000000000000000000000000000000000..81395a06cf0c1a96fd6625dcfb600a87296940cc --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/0889f958adfa68466caf49663390c715db916114d182679b1cbcb25da6335e23.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96cdb3184ccf2920de8c7be1d995a3632bba4bda2a3c0d44443dbab5dfacef3f +size 5234 diff --git a/parse/train/SkzK4iC5Ym/images/09b734e1314b4b16536c33b136a1711079c68c6fb9238bb0ccc40a57f2f8f244.jpg b/parse/train/SkzK4iC5Ym/images/09b734e1314b4b16536c33b136a1711079c68c6fb9238bb0ccc40a57f2f8f244.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ada1ec9009892995b8149051954266bf0d2a0907 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/09b734e1314b4b16536c33b136a1711079c68c6fb9238bb0ccc40a57f2f8f244.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cc9d389a44b9cb011886f02f543e2b900a578ce31cb9e39e1f8af6cba0afc4d +size 12004 diff --git a/parse/train/SkzK4iC5Ym/images/0a76fdc18060aa82c48890e1fe613d676c4ee2bd556eb5ae102efda96c9723c1.jpg b/parse/train/SkzK4iC5Ym/images/0a76fdc18060aa82c48890e1fe613d676c4ee2bd556eb5ae102efda96c9723c1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..683ee318888550489fa1e0b75626ef3f34ed03e7 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/0a76fdc18060aa82c48890e1fe613d676c4ee2bd556eb5ae102efda96c9723c1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4033e9395def8eedfc897c75f9ca771f73eb45522a65ab0fa9ba7925e6d1b488 +size 6080 diff --git a/parse/train/SkzK4iC5Ym/images/0ad8b51e602c103b6f60cc162b4532edeadfdd06e553380988f698675c40e5c8.jpg b/parse/train/SkzK4iC5Ym/images/0ad8b51e602c103b6f60cc162b4532edeadfdd06e553380988f698675c40e5c8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8a3ba90a18e75d4ff894718f0bada92c2da45ff2 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/0ad8b51e602c103b6f60cc162b4532edeadfdd06e553380988f698675c40e5c8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1da9a8bc2e6b64cfcdf23a0422c5d7c7aa029ae310d0309c79e8757fc9a7a13b +size 20480 diff --git a/parse/train/SkzK4iC5Ym/images/10289fb7b1ae8d49007e058135592213b3738348aca9a26361fad74f14ea0066.jpg b/parse/train/SkzK4iC5Ym/images/10289fb7b1ae8d49007e058135592213b3738348aca9a26361fad74f14ea0066.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f9c98906cfb33d9288d39b7161efac1165640db0 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/10289fb7b1ae8d49007e058135592213b3738348aca9a26361fad74f14ea0066.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:735b933e8dfbc313d823395d1ad13fec558e35add51f04eafce681ca788a3e52 +size 34415 diff --git a/parse/train/SkzK4iC5Ym/images/103a870addb9578c22ef6cb7cfc81a3bca2c6dd9abd7895ba4d9f01efad91cf8.jpg b/parse/train/SkzK4iC5Ym/images/103a870addb9578c22ef6cb7cfc81a3bca2c6dd9abd7895ba4d9f01efad91cf8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..37a865d60a1adf006ca07945dbb23136f9d80d4e --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/103a870addb9578c22ef6cb7cfc81a3bca2c6dd9abd7895ba4d9f01efad91cf8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74d96385a1968e2c5ce15dd90d3a3cb1ead5c91fc3098340fddb667bb80c5cd5 +size 18274 diff --git a/parse/train/SkzK4iC5Ym/images/126f52270d3c87f25a8a7ec037f755775e267e1afc55a94bedc48fcf4ca2be44.jpg b/parse/train/SkzK4iC5Ym/images/126f52270d3c87f25a8a7ec037f755775e267e1afc55a94bedc48fcf4ca2be44.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9ac8c81b2a481ca02b78d5f932cb57aace485f34 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/126f52270d3c87f25a8a7ec037f755775e267e1afc55a94bedc48fcf4ca2be44.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ee9a2189a07dd4d51ea27aad01f011ff19b9506d817bcf5e18a051bc1832c03 +size 28951 diff --git a/parse/train/SkzK4iC5Ym/images/14a58d3a31634abc0bbbdb52565dc56f0dab98b67d90dfe99f2bf6384c0abc21.jpg b/parse/train/SkzK4iC5Ym/images/14a58d3a31634abc0bbbdb52565dc56f0dab98b67d90dfe99f2bf6384c0abc21.jpg new file mode 100644 index 0000000000000000000000000000000000000000..05dc6604a7192c80a2eae917989f7df87cfd1f3a --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/14a58d3a31634abc0bbbdb52565dc56f0dab98b67d90dfe99f2bf6384c0abc21.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:278f38c30c9eea7e6d39bd932d014e1e6b09527aa3bb6ce1df6633859886269e +size 6658 diff --git a/parse/train/SkzK4iC5Ym/images/17e61296d5597334c53b5b06bbfb9c6a3d3193155f3f6318dab28e651eee4fd4.jpg b/parse/train/SkzK4iC5Ym/images/17e61296d5597334c53b5b06bbfb9c6a3d3193155f3f6318dab28e651eee4fd4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..22fce9742ce4013959c7fd8693b9ca206d5338c4 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/17e61296d5597334c53b5b06bbfb9c6a3d3193155f3f6318dab28e651eee4fd4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6188c1342f606e06321898182a8f9d87694eccbf5510d69372aaa7cbce47d552 +size 16973 diff --git a/parse/train/SkzK4iC5Ym/images/17ea873f5a828a3eb0afb5531ecad760ab3f9f89d1870ea3f0b7012823a6a122.jpg b/parse/train/SkzK4iC5Ym/images/17ea873f5a828a3eb0afb5531ecad760ab3f9f89d1870ea3f0b7012823a6a122.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9230e7b59f6c50406640d9d262eed7bc88dcc40c --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/17ea873f5a828a3eb0afb5531ecad760ab3f9f89d1870ea3f0b7012823a6a122.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d8e6dc9affd0735111dbd318d08b6c8bad3c797e328ec99ed0fb9df226f98a4 +size 5527 diff --git a/parse/train/SkzK4iC5Ym/images/187049602c7266439a9ef47a2f0ef463df43831c5ba49d09046fbbe0f9fb4954.jpg b/parse/train/SkzK4iC5Ym/images/187049602c7266439a9ef47a2f0ef463df43831c5ba49d09046fbbe0f9fb4954.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a4caf7a8f7e3447fd0c127d4463aaf0bbafd57b8 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/187049602c7266439a9ef47a2f0ef463df43831c5ba49d09046fbbe0f9fb4954.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53eb9bbc3c72689267585000244b089626e3f40c5a240cabf2ff76697187e3c1 +size 11436 diff --git a/parse/train/SkzK4iC5Ym/images/1ace7732f08da7055d67784d3295ce81394c22c5b4aa34fdf62314669059181d.jpg b/parse/train/SkzK4iC5Ym/images/1ace7732f08da7055d67784d3295ce81394c22c5b4aa34fdf62314669059181d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c4d98f4aa0a7b2fd0da55d4c6f9cb0041041a0a3 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/1ace7732f08da7055d67784d3295ce81394c22c5b4aa34fdf62314669059181d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85d9ca263815e99a93c705a96a28ef021a381043ffeb00774233977ae306a6a5 +size 9402 diff --git a/parse/train/SkzK4iC5Ym/images/1d48ff5341c983cc03fcb5a441e6c8918b57841fc3df93746024e113c214f816.jpg b/parse/train/SkzK4iC5Ym/images/1d48ff5341c983cc03fcb5a441e6c8918b57841fc3df93746024e113c214f816.jpg new file mode 100644 index 0000000000000000000000000000000000000000..20a5294fe7663260dd0d905ec6e0f21f3b6f1cd7 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/1d48ff5341c983cc03fcb5a441e6c8918b57841fc3df93746024e113c214f816.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0aae5398dce71db6f73743e64eadd7a14f38da619192df8c281e14416b85a93e +size 22463 diff --git a/parse/train/SkzK4iC5Ym/images/2212bf01ddc09fe31901a35d1f989fe5cd2e6f250ae731d633292a2c0d41c3c0.jpg b/parse/train/SkzK4iC5Ym/images/2212bf01ddc09fe31901a35d1f989fe5cd2e6f250ae731d633292a2c0d41c3c0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..510abc5dedaa4eb4bd6ee21c299f7743f638d27f --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/2212bf01ddc09fe31901a35d1f989fe5cd2e6f250ae731d633292a2c0d41c3c0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d7b188d954a1077cd503d1a637dc2795f03720194e6656872fc730498c35f93 +size 10010 diff --git a/parse/train/SkzK4iC5Ym/images/22f6e2e9ad27d26d43466104627fc62e697481ead2c3a782df045d849f3d5003.jpg b/parse/train/SkzK4iC5Ym/images/22f6e2e9ad27d26d43466104627fc62e697481ead2c3a782df045d849f3d5003.jpg new file mode 100644 index 0000000000000000000000000000000000000000..08277738c856ba1489b782b4a4ef439e5e1a2715 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/22f6e2e9ad27d26d43466104627fc62e697481ead2c3a782df045d849f3d5003.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17d0f8a074e35c55fd0f59ca724796653ae7fd0f735035b83334e0d92683f5e1 +size 4300 diff --git a/parse/train/SkzK4iC5Ym/images/231d8d0a1f73001579c3109540701fff94d167321b8bf2395d91fa597a0207c7.jpg b/parse/train/SkzK4iC5Ym/images/231d8d0a1f73001579c3109540701fff94d167321b8bf2395d91fa597a0207c7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fe1ec22fa69d7c5f3c13e8fef8c99ad3e5125cba --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/231d8d0a1f73001579c3109540701fff94d167321b8bf2395d91fa597a0207c7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfdbf29f5c0996d5357560625dc60d55bb783ed2d1a7d21e74d6d7b6815da7f5 +size 9986 diff --git a/parse/train/SkzK4iC5Ym/images/234f312a787361eae64be417c97f12f95247e054faf87e76a684485f42ee18e0.jpg b/parse/train/SkzK4iC5Ym/images/234f312a787361eae64be417c97f12f95247e054faf87e76a684485f42ee18e0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0cb6f764191d78cb617aaaf0fce7b5335776aa79 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/234f312a787361eae64be417c97f12f95247e054faf87e76a684485f42ee18e0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be291f73c6602c2f235ef2a64795305ed12aa76d4d3e4a61c4bc040abcffd3fa +size 8084 diff --git a/parse/train/SkzK4iC5Ym/images/24b37c6caae308efa4b42610c1686831eeac6104269358cd86eb480ffddf4a6b.jpg b/parse/train/SkzK4iC5Ym/images/24b37c6caae308efa4b42610c1686831eeac6104269358cd86eb480ffddf4a6b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8196dcfb7c979a0937d48516b7c8975ad0a043e3 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/24b37c6caae308efa4b42610c1686831eeac6104269358cd86eb480ffddf4a6b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1c730711c94be80bd788ad36f0d56ca381857c53adb4651b65aadf0b870e2ae +size 8662 diff --git a/parse/train/SkzK4iC5Ym/images/27a9bfb1105e65fdc2870f450ea19fa28730642ace7262ff149f2ee6068fb07d.jpg b/parse/train/SkzK4iC5Ym/images/27a9bfb1105e65fdc2870f450ea19fa28730642ace7262ff149f2ee6068fb07d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..607412787c0fdd28f5c491e419c267503420833a --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/27a9bfb1105e65fdc2870f450ea19fa28730642ace7262ff149f2ee6068fb07d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f13e42e15b23a4e83920ef3c843af7ce2603f9681b58b00c10dca8a330f03eaa +size 3765 diff --git a/parse/train/SkzK4iC5Ym/images/2be5ea480c57575a0932283c983ceb2d775dea43d9d6b47564cb6d8545690214.jpg b/parse/train/SkzK4iC5Ym/images/2be5ea480c57575a0932283c983ceb2d775dea43d9d6b47564cb6d8545690214.jpg new file mode 100644 index 0000000000000000000000000000000000000000..00a38254433d6fdd4fc57ad74f4ee6b0d4054090 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/2be5ea480c57575a0932283c983ceb2d775dea43d9d6b47564cb6d8545690214.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33eaa6bb36cf8feb3a074576b78bd4ffd52ac0736dfb1b8639ba8ed2a3b139cb +size 8453 diff --git a/parse/train/SkzK4iC5Ym/images/2cf7a6ab6c6c2f5f39e92d61d4c4a6861f1c6a509b74eff15bd914b938b0f996.jpg b/parse/train/SkzK4iC5Ym/images/2cf7a6ab6c6c2f5f39e92d61d4c4a6861f1c6a509b74eff15bd914b938b0f996.jpg new file mode 100644 index 0000000000000000000000000000000000000000..038c45dc66131d5b84fb68def4a39ee078bc2de6 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/2cf7a6ab6c6c2f5f39e92d61d4c4a6861f1c6a509b74eff15bd914b938b0f996.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:891ba948c6a751c745068695fdd91bf79da473aec6a47feffcc848cd82552b73 +size 16083 diff --git a/parse/train/SkzK4iC5Ym/images/2da6d63fb725c99f270d829d1f4a895a4c6108f9b832edd8fce5cb6dcb059a96.jpg b/parse/train/SkzK4iC5Ym/images/2da6d63fb725c99f270d829d1f4a895a4c6108f9b832edd8fce5cb6dcb059a96.jpg new file mode 100644 index 0000000000000000000000000000000000000000..09e3227a7008d7351152f89d3cd4268bef876add --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/2da6d63fb725c99f270d829d1f4a895a4c6108f9b832edd8fce5cb6dcb059a96.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:034107ae3393104076ef6dc90074d27c0a3fb1918ad01700fed4ee3f2cf0ad90 +size 2747 diff --git a/parse/train/SkzK4iC5Ym/images/316145e4e86f218081a8ce25df49b60be1ca6eddae8c899e55729e86ad4bc071.jpg b/parse/train/SkzK4iC5Ym/images/316145e4e86f218081a8ce25df49b60be1ca6eddae8c899e55729e86ad4bc071.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7ba67d150f9d8ea6789151aad2260996b0428834 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/316145e4e86f218081a8ce25df49b60be1ca6eddae8c899e55729e86ad4bc071.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51a20b1427d7793ebf6873de68476bd62ac776254024732476f7f27f27fc6062 +size 6354 diff --git a/parse/train/SkzK4iC5Ym/images/33818be810b1e5d336776c9aa6ce02b12f7c2a6afeedecbf082ca59f7d0ccb1f.jpg b/parse/train/SkzK4iC5Ym/images/33818be810b1e5d336776c9aa6ce02b12f7c2a6afeedecbf082ca59f7d0ccb1f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..971b589c267aa1c62eaa2c90f396b1019de3c5c2 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/33818be810b1e5d336776c9aa6ce02b12f7c2a6afeedecbf082ca59f7d0ccb1f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62fb2b94d1f8389cbb47e4fd5dc6b268919a80c102ea5f64369c3b859c065443 +size 4397 diff --git a/parse/train/SkzK4iC5Ym/images/3395b4e25e238be78628b7694d0136e42bf182fbc538be547c20b407de5d84e1.jpg b/parse/train/SkzK4iC5Ym/images/3395b4e25e238be78628b7694d0136e42bf182fbc538be547c20b407de5d84e1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..96116603c6bb05aa792ae2ee85aacf9a30e4c6cd --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/3395b4e25e238be78628b7694d0136e42bf182fbc538be547c20b407de5d84e1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5c572cad84fe9ae8d32eceb86ece8d18d08f2c68e6a9a41de3780fb2f760330 +size 8085 diff --git a/parse/train/SkzK4iC5Ym/images/345892dd6db58c30fc3da63a6aceead7ff1df12c568a317ec3cf3db19b679aa8.jpg b/parse/train/SkzK4iC5Ym/images/345892dd6db58c30fc3da63a6aceead7ff1df12c568a317ec3cf3db19b679aa8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6007c4de89faeb27a9af87948bb82e7c94bcc86f --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/345892dd6db58c30fc3da63a6aceead7ff1df12c568a317ec3cf3db19b679aa8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70373479f39d013014eb757d8ee18caefc9d8e7647fdeae8a97c82eadb342495 +size 7036 diff --git a/parse/train/SkzK4iC5Ym/images/3517162a0f979718c7d33e81fc3f0ff61187473098ff3cf34264751fb30219d6.jpg b/parse/train/SkzK4iC5Ym/images/3517162a0f979718c7d33e81fc3f0ff61187473098ff3cf34264751fb30219d6.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bb1c3fd32f17a2c6245ba44ffe8452368f1d1da0 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/3517162a0f979718c7d33e81fc3f0ff61187473098ff3cf34264751fb30219d6.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:067fc2c66c17174fbe0f55479d691369f549e259b3faa3708f95e28084d24732 +size 36268 diff --git a/parse/train/SkzK4iC5Ym/images/3653c8b21c801d9109db762eb6b555481f06eb3409d8901201289f45a7a7b4a3.jpg b/parse/train/SkzK4iC5Ym/images/3653c8b21c801d9109db762eb6b555481f06eb3409d8901201289f45a7a7b4a3.jpg new file mode 100644 index 0000000000000000000000000000000000000000..834865efd04e1fe83779523a587c48eab3203769 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/3653c8b21c801d9109db762eb6b555481f06eb3409d8901201289f45a7a7b4a3.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e42dcc8022607d91ff9575f1e89b011c292f2c2481e4ebf7c747b215afc4e9bc +size 44834 diff --git a/parse/train/SkzK4iC5Ym/images/3accad6fb4cc3cf1cdb64a748249cd13cc1754ef8c53987e9407c01dbb451ed8.jpg b/parse/train/SkzK4iC5Ym/images/3accad6fb4cc3cf1cdb64a748249cd13cc1754ef8c53987e9407c01dbb451ed8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7451edf333f6239e8e896c9345e5edaf6b2d5c09 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/3accad6fb4cc3cf1cdb64a748249cd13cc1754ef8c53987e9407c01dbb451ed8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a941e77ae78af3236d862aef9502c0714f0aea6ebebba563e71fa5cafec9f0fe +size 6079 diff --git a/parse/train/SkzK4iC5Ym/images/3b4ece42648b5d713c7b5c20e2c385437e96c25f88e965cd5b71a4533c2fb2cd.jpg b/parse/train/SkzK4iC5Ym/images/3b4ece42648b5d713c7b5c20e2c385437e96c25f88e965cd5b71a4533c2fb2cd.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8b7d1fa60b0f3262d9591d6ca4ce22e0003015f7 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/3b4ece42648b5d713c7b5c20e2c385437e96c25f88e965cd5b71a4533c2fb2cd.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49d2c512befbf11db57a513dabd9ca758390a7179d6a372c092aacd2d05a6a1e +size 5991 diff --git a/parse/train/SkzK4iC5Ym/images/408438ed6b060211a3d9634dbd251d9915706e7cd478384eea26f1f953020ecd.jpg b/parse/train/SkzK4iC5Ym/images/408438ed6b060211a3d9634dbd251d9915706e7cd478384eea26f1f953020ecd.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c78ace1dd963b5cb2ba189a66682ccd342850224 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/408438ed6b060211a3d9634dbd251d9915706e7cd478384eea26f1f953020ecd.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e98ec1dfc6feaf430869d2cfb75eb12ec763336dcf0e85685bf92c794324b097 +size 17389 diff --git a/parse/train/SkzK4iC5Ym/images/40d9edd6ade32920a30747185f5cd0c742590fd44eb77da89842494244056748.jpg b/parse/train/SkzK4iC5Ym/images/40d9edd6ade32920a30747185f5cd0c742590fd44eb77da89842494244056748.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2b4427f420d7eb2eee953963c285b913c7ced912 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/40d9edd6ade32920a30747185f5cd0c742590fd44eb77da89842494244056748.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59a8fb36e0fcb611f68ec9732a2f0500aa2c42cf67ec869a07c494864873a041 +size 8540 diff --git a/parse/train/SkzK4iC5Ym/images/41019f22345a37995b517b6424f5397752f19663968fa3df32959148377cda1e.jpg b/parse/train/SkzK4iC5Ym/images/41019f22345a37995b517b6424f5397752f19663968fa3df32959148377cda1e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5d049a890e9ba21fe93ef140bccd3eba4522228d --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/41019f22345a37995b517b6424f5397752f19663968fa3df32959148377cda1e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89c9186f288367fa5f65e98d6bd1909884c2edaaa9bd6f6ea865a31a833270de +size 10235 diff --git a/parse/train/SkzK4iC5Ym/images/416c7a15c55d33f46cdf5c1a07de4f54e54c16dfd277cefa04f32e7a46a45773.jpg b/parse/train/SkzK4iC5Ym/images/416c7a15c55d33f46cdf5c1a07de4f54e54c16dfd277cefa04f32e7a46a45773.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8c56d3f520d1c4eb2814e87607c5a2b0c52b4b83 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/416c7a15c55d33f46cdf5c1a07de4f54e54c16dfd277cefa04f32e7a46a45773.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a00209553ecb7cffd547ca1ba00c17d0f99338db66487405e037c69de22d19b +size 6328 diff --git a/parse/train/SkzK4iC5Ym/images/45142e12c48a2a7306c3417cd50ea4b1e079f8cc62da31225ca31a2d17b4f073.jpg b/parse/train/SkzK4iC5Ym/images/45142e12c48a2a7306c3417cd50ea4b1e079f8cc62da31225ca31a2d17b4f073.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e082b9610982fe3f18e80a1161afb6a86c5903cf --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/45142e12c48a2a7306c3417cd50ea4b1e079f8cc62da31225ca31a2d17b4f073.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a053b6aaefbc46fe3a2f2f7a891361d2a4e38b67d21bd4659372ae70213b3eb +size 6833 diff --git a/parse/train/SkzK4iC5Ym/images/45a0aae30024041eaa8edc2c3a6253534faa9d6c84e8f2c4361eb732262aa046.jpg b/parse/train/SkzK4iC5Ym/images/45a0aae30024041eaa8edc2c3a6253534faa9d6c84e8f2c4361eb732262aa046.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8e884865c4662d9bda91140a033545b2ef32a9cd --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/45a0aae30024041eaa8edc2c3a6253534faa9d6c84e8f2c4361eb732262aa046.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56836fee787d0c8272e52cc47305084a003bbfb42c9bdc65da84eebc849f86d8 +size 6690 diff --git a/parse/train/SkzK4iC5Ym/images/461ade00b56b40822e04c9ad3b431148dd28020e20850a58d10f234f786a88ed.jpg b/parse/train/SkzK4iC5Ym/images/461ade00b56b40822e04c9ad3b431148dd28020e20850a58d10f234f786a88ed.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6a1043428ec154ba37038202d0101b5d63ab6781 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/461ade00b56b40822e04c9ad3b431148dd28020e20850a58d10f234f786a88ed.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bbff31c62d838982f55035e7a149af9561520136c9ef70b92a1045af9595aca +size 6064 diff --git a/parse/train/SkzK4iC5Ym/images/471686cb274350d856acdee4a2be1b74b1bce4ea04389c89660e470684162671.jpg b/parse/train/SkzK4iC5Ym/images/471686cb274350d856acdee4a2be1b74b1bce4ea04389c89660e470684162671.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e12417b0fb2dbdb969737e328cce9b6ebf273fe2 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/471686cb274350d856acdee4a2be1b74b1bce4ea04389c89660e470684162671.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b922d8fd37a14c7152ad69ca252d97f7bf4200dbc5cafd1c460af7324074d691 +size 35526 diff --git a/parse/train/SkzK4iC5Ym/images/4934b2db8e7a82cdec337018064237ef67ed57bcf89f4d94213e9e6039b26311.jpg b/parse/train/SkzK4iC5Ym/images/4934b2db8e7a82cdec337018064237ef67ed57bcf89f4d94213e9e6039b26311.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f2ba4b731a203690c36d0312a1f006d2bd6a0047 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/4934b2db8e7a82cdec337018064237ef67ed57bcf89f4d94213e9e6039b26311.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c12276e3b63b4e1ced02dec31dd2200a166d20c8879bbc735fb332594db7018b +size 8793 diff --git a/parse/train/SkzK4iC5Ym/images/4a766940e121087b27861abe3b7bd3076b69ae19e3b649211c2a593d5b1b0c68.jpg b/parse/train/SkzK4iC5Ym/images/4a766940e121087b27861abe3b7bd3076b69ae19e3b649211c2a593d5b1b0c68.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6d72affde19fa67fcd41d904cee38f988a6b34e7 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/4a766940e121087b27861abe3b7bd3076b69ae19e3b649211c2a593d5b1b0c68.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a13dcc7345924e687e283b904058b4dda70ea6a260f5d68159ac2b3f64bf6668 +size 14696 diff --git a/parse/train/SkzK4iC5Ym/images/4c0ca7c1e955e64abab117765815e22bdb923220b0870f93b8158fd21bce55d7.jpg b/parse/train/SkzK4iC5Ym/images/4c0ca7c1e955e64abab117765815e22bdb923220b0870f93b8158fd21bce55d7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9a47884d84bfda0de80440cb16022775f6cacb2a --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/4c0ca7c1e955e64abab117765815e22bdb923220b0870f93b8158fd21bce55d7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cbbb0d6c599b635e84a97e1acb088033f33f37b1c52f4d778100a4f84c7b2ca +size 10985 diff --git a/parse/train/SkzK4iC5Ym/images/4ccefc187b7df151b8fc4537f5ee3c2881e298836ae83a7754d481e02ffdbd4e.jpg b/parse/train/SkzK4iC5Ym/images/4ccefc187b7df151b8fc4537f5ee3c2881e298836ae83a7754d481e02ffdbd4e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..dee0465da021e827a942be9f4e80f9cc6fe04407 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/4ccefc187b7df151b8fc4537f5ee3c2881e298836ae83a7754d481e02ffdbd4e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d638a32427fd19a80563a61174b59a766037db3fdf0c513a9f107d5e489fa2a +size 9453 diff --git a/parse/train/SkzK4iC5Ym/images/506015a69c360e84120eb7581529edc79ac70335f37e5378a238e8fe440bdcba.jpg b/parse/train/SkzK4iC5Ym/images/506015a69c360e84120eb7581529edc79ac70335f37e5378a238e8fe440bdcba.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ccdc3a3f801522cde2ab9c0428a0c1be041e3d42 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/506015a69c360e84120eb7581529edc79ac70335f37e5378a238e8fe440bdcba.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:479080063371065a9eb0342c167d2a309c4e42901adde962a429e38dc3e53cee +size 6236 diff --git a/parse/train/SkzK4iC5Ym/images/5a19228d1ccb28a17e6b9c86252c654b329a04e541f24302bda81cbbc985074e.jpg b/parse/train/SkzK4iC5Ym/images/5a19228d1ccb28a17e6b9c86252c654b329a04e541f24302bda81cbbc985074e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..8783bd861b4e09b8ca39d6212795e3c93455d686 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/5a19228d1ccb28a17e6b9c86252c654b329a04e541f24302bda81cbbc985074e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a332836fc0494849d3f3b9aabd0b8e270b1fe34a141a937435891701ef805f66 +size 8267 diff --git a/parse/train/SkzK4iC5Ym/images/5b9af4f82f423545709022ac7cb63d3cf1221b93d4692a28dba882a58e416a1f.jpg b/parse/train/SkzK4iC5Ym/images/5b9af4f82f423545709022ac7cb63d3cf1221b93d4692a28dba882a58e416a1f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..acae05dfd99a3f4ac8ed92645dbc46ce9a97640b --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/5b9af4f82f423545709022ac7cb63d3cf1221b93d4692a28dba882a58e416a1f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7054466dea6783018f87db046cf511a15717c9b0beab610db293313f7e043fd3 +size 7326 diff --git a/parse/train/SkzK4iC5Ym/images/5e9045f5960490c53ab6d7a1bb13f1a603862f5143bd81b256d9d55f740283ca.jpg b/parse/train/SkzK4iC5Ym/images/5e9045f5960490c53ab6d7a1bb13f1a603862f5143bd81b256d9d55f740283ca.jpg new file mode 100644 index 0000000000000000000000000000000000000000..bbc883ac58b9e40006e0600c63d41be8925c352c --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/5e9045f5960490c53ab6d7a1bb13f1a603862f5143bd81b256d9d55f740283ca.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e579a4fe09435a94ba3a6cfae0449914cd54cef7cd5b71110e688fe8a0123002 +size 10091 diff --git a/parse/train/SkzK4iC5Ym/images/60052986dc92139bb28334f738603cb1241691a4c765065a21aa687a97ac56f1.jpg b/parse/train/SkzK4iC5Ym/images/60052986dc92139bb28334f738603cb1241691a4c765065a21aa687a97ac56f1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b0fe3702a7813678ae187bab618413108073cfe0 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/60052986dc92139bb28334f738603cb1241691a4c765065a21aa687a97ac56f1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a67a2392e72d9f01d2ecc1061078dfd08c08d45bc456d37ef46bbe778abf83f6 +size 28029 diff --git a/parse/train/SkzK4iC5Ym/images/68381a816495bbad4379f37319a2f43ca276149c1f7d39726322249eac977a96.jpg b/parse/train/SkzK4iC5Ym/images/68381a816495bbad4379f37319a2f43ca276149c1f7d39726322249eac977a96.jpg new file mode 100644 index 0000000000000000000000000000000000000000..291d9d48de9750eb1dc66d4a7fc0b7ea91aba611 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/68381a816495bbad4379f37319a2f43ca276149c1f7d39726322249eac977a96.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5c335a1a96bd634611fb752028550afb8697d2807446d2328569ba5bec611b9 +size 10956 diff --git a/parse/train/SkzK4iC5Ym/images/699264feca8460cea40bf88ae40cf034c42629cba3c86d448590a35522301cc7.jpg b/parse/train/SkzK4iC5Ym/images/699264feca8460cea40bf88ae40cf034c42629cba3c86d448590a35522301cc7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..11744372be767e4efc182ae0de08087514a72c2c --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/699264feca8460cea40bf88ae40cf034c42629cba3c86d448590a35522301cc7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14e752be740411c3cc186cef9cc4169cbbe09e3f42a2aeabfe999a9ee20fdc9b +size 22058 diff --git a/parse/train/SkzK4iC5Ym/images/6e81609839f29575b1ba848ea8d9a6fef96936f5163e6621377248f838ad4ff4.jpg b/parse/train/SkzK4iC5Ym/images/6e81609839f29575b1ba848ea8d9a6fef96936f5163e6621377248f838ad4ff4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..48152f8716fe1a6b051cc65c6fa8c44419023878 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/6e81609839f29575b1ba848ea8d9a6fef96936f5163e6621377248f838ad4ff4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:210e4cbb4f1e153a1f6d5dcb7f021fa13ddda7fd3891693449e0b4f98d79b0ac +size 45658 diff --git a/parse/train/SkzK4iC5Ym/images/6fff7201f3bdf29680d797fb9505c520a61d7f0f670128ac480bb32c8d49b225.jpg b/parse/train/SkzK4iC5Ym/images/6fff7201f3bdf29680d797fb9505c520a61d7f0f670128ac480bb32c8d49b225.jpg new file mode 100644 index 0000000000000000000000000000000000000000..0fa3a2d7c848ff7eeb73fa298327a2ad6aef699b --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/6fff7201f3bdf29680d797fb9505c520a61d7f0f670128ac480bb32c8d49b225.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db7d16539ded0b54a2fb1eb465fb2a31a4b2f9cf7dedd3dab31f0f23104da785 +size 6704 diff --git a/parse/train/SkzK4iC5Ym/images/736cdee43b0b6a8f2cb932fc8e80cf835c1f7216a71a9846b3c78771a9cc9d05.jpg b/parse/train/SkzK4iC5Ym/images/736cdee43b0b6a8f2cb932fc8e80cf835c1f7216a71a9846b3c78771a9cc9d05.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a09d7a5ba690e78ef3b80a0a0f0d7dbfce7dbb90 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/736cdee43b0b6a8f2cb932fc8e80cf835c1f7216a71a9846b3c78771a9cc9d05.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10ea983905ff990bc7ccd0a98f08b6671f1bc9c25daa84a28278264c478619cd +size 5179 diff --git a/parse/train/SkzK4iC5Ym/images/74276f2cb267e20add275f3c4f45724b2cb05ddcf891f332e96e31188aacc229.jpg b/parse/train/SkzK4iC5Ym/images/74276f2cb267e20add275f3c4f45724b2cb05ddcf891f332e96e31188aacc229.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6c7b50b4f6421208324abaa3603e1cc5057f12e1 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/74276f2cb267e20add275f3c4f45724b2cb05ddcf891f332e96e31188aacc229.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aea2d630c3a82786aa2fe600be4726714a090617b71d3f8bf03c608353dab657 +size 4010 diff --git a/parse/train/SkzK4iC5Ym/images/74698f18dd6ece603d51da6e48284f37c14d4a99df60fc8f57f8200c991729d2.jpg b/parse/train/SkzK4iC5Ym/images/74698f18dd6ece603d51da6e48284f37c14d4a99df60fc8f57f8200c991729d2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..670ee303ec1d5700d0f54b2465ef50f2b8c50ee3 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/74698f18dd6ece603d51da6e48284f37c14d4a99df60fc8f57f8200c991729d2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98e103292b1dc2991d49bf646fc9828a685c79ffdbe4ef2e18fdfed9f811f042 +size 9920 diff --git a/parse/train/SkzK4iC5Ym/images/77d8c45e584afba3f4b6e4c86ac10c2ce0dc55cbd147a6161eea534f7547788a.jpg b/parse/train/SkzK4iC5Ym/images/77d8c45e584afba3f4b6e4c86ac10c2ce0dc55cbd147a6161eea534f7547788a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..be8d8c554bd2253af60bff808ebdb32753b14744 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/77d8c45e584afba3f4b6e4c86ac10c2ce0dc55cbd147a6161eea534f7547788a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52c9e54d1997685e9992703e292fdbb78860b7700f794ad5a050d39dbfc7d57d +size 7931 diff --git a/parse/train/SkzK4iC5Ym/images/783fb10bf5e854faf858a9dc4c02bc632682bbe4599ee59f37147477a9d9ca6a.jpg b/parse/train/SkzK4iC5Ym/images/783fb10bf5e854faf858a9dc4c02bc632682bbe4599ee59f37147477a9d9ca6a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..19132635e503ab5ee7ce7f0fe64ff53d70dd3569 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/783fb10bf5e854faf858a9dc4c02bc632682bbe4599ee59f37147477a9d9ca6a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f82a9aded2822cd19866056908592cbfde35521aff07533b5c4205345c3a479 +size 29465 diff --git a/parse/train/SkzK4iC5Ym/images/79692f81ae2ce516c63af3aff3477472121107ce40277e75826cc81b5008cd13.jpg b/parse/train/SkzK4iC5Ym/images/79692f81ae2ce516c63af3aff3477472121107ce40277e75826cc81b5008cd13.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5c51ed3eeb048d18fb86a7cbf8d071446fb9f3ba --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/79692f81ae2ce516c63af3aff3477472121107ce40277e75826cc81b5008cd13.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d232b0047f01e7bcf608d72366c78e17ebea718c1bda649cdbe5a7b8afcea80c +size 8566 diff --git a/parse/train/SkzK4iC5Ym/images/7c0616478ba11bd94c250263ececd82a0117227d0645dfd71a52ef8b53b8edef.jpg b/parse/train/SkzK4iC5Ym/images/7c0616478ba11bd94c250263ececd82a0117227d0645dfd71a52ef8b53b8edef.jpg new file mode 100644 index 0000000000000000000000000000000000000000..52176a46cf6a92ded90738e08226ada399165cd2 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/7c0616478ba11bd94c250263ececd82a0117227d0645dfd71a52ef8b53b8edef.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b82edcafbc71fddfa872edfd6333cc7dfaa84bf2a29795d4676e2b04a37631b +size 13137 diff --git a/parse/train/SkzK4iC5Ym/images/7e8c11b4e436588a109217406e3c6fad5ad073c6692c074b41e685319401adbe.jpg b/parse/train/SkzK4iC5Ym/images/7e8c11b4e436588a109217406e3c6fad5ad073c6692c074b41e685319401adbe.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ccd26d253af6caa0072b10231a38f72ef8f380bc --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/7e8c11b4e436588a109217406e3c6fad5ad073c6692c074b41e685319401adbe.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86f15ec085ae085d20091290b7be965d04a073224bff42092af37d896a03d671 +size 9878 diff --git a/parse/train/SkzK4iC5Ym/images/7e9322791de21755ec92429aaf7e8a6747759b4fcbe2e9d01c6117a8bb8f3309.jpg b/parse/train/SkzK4iC5Ym/images/7e9322791de21755ec92429aaf7e8a6747759b4fcbe2e9d01c6117a8bb8f3309.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a72a32b75f93de08aa217e82fd49a0264a567a5a --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/7e9322791de21755ec92429aaf7e8a6747759b4fcbe2e9d01c6117a8bb8f3309.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:368f7582558040979845b84b6adaa5e4442564bcc14f5f3e767e1c191216dd12 +size 6412 diff --git a/parse/train/SkzK4iC5Ym/images/7ec9fbb5d97faf3ed2057546818cd097878ab7627712bf9e293aa623efd69b13.jpg b/parse/train/SkzK4iC5Ym/images/7ec9fbb5d97faf3ed2057546818cd097878ab7627712bf9e293aa623efd69b13.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d5825ca0c95d63f5a5a2441638b2d332ea9aafdd --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/7ec9fbb5d97faf3ed2057546818cd097878ab7627712bf9e293aa623efd69b13.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6a60a151e3a16101022e3b184f0314785584bc02d2603a470bddf17829386e0 +size 4337 diff --git a/parse/train/SkzK4iC5Ym/images/818a5af76c61e81b43cc671b05e6345c81d5fb4b65dcd2a915c38af3cd7e16d9.jpg b/parse/train/SkzK4iC5Ym/images/818a5af76c61e81b43cc671b05e6345c81d5fb4b65dcd2a915c38af3cd7e16d9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ce631b294efada1d5736a75ca87a8b9757633ea2 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/818a5af76c61e81b43cc671b05e6345c81d5fb4b65dcd2a915c38af3cd7e16d9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd65a0fdbf1331da1d82845efc985d72fa2e6d66ea437679195802461a663d8e +size 7343 diff --git a/parse/train/SkzK4iC5Ym/images/8574199601e6ab7113b10cd4ce392e3ecbe22f9683aa537432eb8b3d47cf1fe1.jpg b/parse/train/SkzK4iC5Ym/images/8574199601e6ab7113b10cd4ce392e3ecbe22f9683aa537432eb8b3d47cf1fe1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a705a8c86a6db57daecfb62ee3708aa8058e44a8 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/8574199601e6ab7113b10cd4ce392e3ecbe22f9683aa537432eb8b3d47cf1fe1.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34afde133c5065bd6dda935a41b556cb1357f07975f8fc3d7c247d802b1e5e36 +size 8119 diff --git a/parse/train/SkzK4iC5Ym/images/86664f4721e798fc9e44ba7a848f8c1f4068ebc2a778224b71207a0dc959c67d.jpg b/parse/train/SkzK4iC5Ym/images/86664f4721e798fc9e44ba7a848f8c1f4068ebc2a778224b71207a0dc959c67d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..149c850c9acf48d638d5d8d720c85865966467ca --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/86664f4721e798fc9e44ba7a848f8c1f4068ebc2a778224b71207a0dc959c67d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07707c610b1cda2641e09d906669cffd4226d58aac87844d1d8f1ac0ea518d90 +size 2687 diff --git a/parse/train/SkzK4iC5Ym/images/872dc221880dd5765d45ac079d2d497d0ae93c90cc65b05f1f959382c5cb8ff9.jpg b/parse/train/SkzK4iC5Ym/images/872dc221880dd5765d45ac079d2d497d0ae93c90cc65b05f1f959382c5cb8ff9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..52ea32384d2adb94b352adbfb21087b395f71877 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/872dc221880dd5765d45ac079d2d497d0ae93c90cc65b05f1f959382c5cb8ff9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:364493ac072acb9dd4839e4f1eb8a18d9ab65237995c32bb9b4d17181d01e675 +size 9401 diff --git a/parse/train/SkzK4iC5Ym/images/89cff3b919a455812092d6202dd2446d50fa47be0580b6df334cc63ea0b70228.jpg b/parse/train/SkzK4iC5Ym/images/89cff3b919a455812092d6202dd2446d50fa47be0580b6df334cc63ea0b70228.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b4e791ca94ca29547e22718eccbda049b2d43dfd --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/89cff3b919a455812092d6202dd2446d50fa47be0580b6df334cc63ea0b70228.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6445061eab3f77d7f7741d70a5e7e95f6e08505cf42136329986d846c9650f3 +size 2193 diff --git a/parse/train/SkzK4iC5Ym/images/8aa0ac517983cc3cee890981cac34d97a9bbdf95927a7911d1e0d205d6259b66.jpg b/parse/train/SkzK4iC5Ym/images/8aa0ac517983cc3cee890981cac34d97a9bbdf95927a7911d1e0d205d6259b66.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5b5e1b5ba5648e4667316f2bd473e2cff7d7794e --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/8aa0ac517983cc3cee890981cac34d97a9bbdf95927a7911d1e0d205d6259b66.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4463ba1edc45a3e1748b6c43284329f0357c0c37dd38f736350fec95b804124c +size 7437 diff --git a/parse/train/SkzK4iC5Ym/images/8d564e1c0ee08224e870e45977df74fe907676728b96ef18b7407fd5e21f9ed8.jpg b/parse/train/SkzK4iC5Ym/images/8d564e1c0ee08224e870e45977df74fe907676728b96ef18b7407fd5e21f9ed8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7c445f8a236c7f7e791ca422a1c8813df4335042 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/8d564e1c0ee08224e870e45977df74fe907676728b96ef18b7407fd5e21f9ed8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbc387f333cbb2666ceadd66a4515b9e8c40a762db5108fff616b6951c30cd72 +size 5861 diff --git a/parse/train/SkzK4iC5Ym/images/91f25ddb01e2bf83a2454b6d3bae0efd3831a9abdaa59642eeea3b0cc7ade01e.jpg b/parse/train/SkzK4iC5Ym/images/91f25ddb01e2bf83a2454b6d3bae0efd3831a9abdaa59642eeea3b0cc7ade01e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..79bf34f1bdfec07b1b11fd7d74c7105c147467b9 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/91f25ddb01e2bf83a2454b6d3bae0efd3831a9abdaa59642eeea3b0cc7ade01e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8afdc7ee16d74edd6acf79b36f67fa20d042fc0e8bf76ae5975a5727569b00cb +size 3319 diff --git a/parse/train/SkzK4iC5Ym/images/95e4070c0c79b77db2d213bf6b616009e1f3ded187d0e747a0e82e292b475de0.jpg b/parse/train/SkzK4iC5Ym/images/95e4070c0c79b77db2d213bf6b616009e1f3ded187d0e747a0e82e292b475de0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9e1fa6cf499f8a84ce9c85a5dc24ec3501ab8c00 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/95e4070c0c79b77db2d213bf6b616009e1f3ded187d0e747a0e82e292b475de0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f39e2e10ae292bdec8206713f132a6985607a92b25cd2ccb1299a09b7abaecde +size 3670 diff --git a/parse/train/SkzK4iC5Ym/images/9810c4eb7f8708aaa19af07f873d3f11fe576ac0f448f70ef858f11cbe95f274.jpg b/parse/train/SkzK4iC5Ym/images/9810c4eb7f8708aaa19af07f873d3f11fe576ac0f448f70ef858f11cbe95f274.jpg new file mode 100644 index 0000000000000000000000000000000000000000..c8be0e9a642ce94dd8f935333644b7829bff06eb --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/9810c4eb7f8708aaa19af07f873d3f11fe576ac0f448f70ef858f11cbe95f274.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ac774244a59c947f3d3ba9976350dab411d3eab4342ab7137026b9015980d9d +size 5992 diff --git a/parse/train/SkzK4iC5Ym/images/9cbf4f9acc70b1cc743126e92920856e397d39f4f981ebd380986c6ce1bb5c7f.jpg b/parse/train/SkzK4iC5Ym/images/9cbf4f9acc70b1cc743126e92920856e397d39f4f981ebd380986c6ce1bb5c7f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..36b68a3a940c9688bedd5d5b9672cc0d654384ac --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/9cbf4f9acc70b1cc743126e92920856e397d39f4f981ebd380986c6ce1bb5c7f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a65d9e0d047a75b47e303344daca16c38f6fe21c56b93f743eb7dfb72af10bb +size 5378 diff --git a/parse/train/SkzK4iC5Ym/images/9cd15ff67e2f06613dd3860723a7fa21e5ca2280eddb642640f274a2ca23e4f8.jpg b/parse/train/SkzK4iC5Ym/images/9cd15ff67e2f06613dd3860723a7fa21e5ca2280eddb642640f274a2ca23e4f8.jpg new file mode 100644 index 0000000000000000000000000000000000000000..90feed4b18f512da4310237f4a4edf27be495336 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/9cd15ff67e2f06613dd3860723a7fa21e5ca2280eddb642640f274a2ca23e4f8.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6bb8b8163674614548cd930220d2894dd4ffb717aa9cff7396490ce0e30043a +size 4498 diff --git a/parse/train/SkzK4iC5Ym/images/9cd6122f957063481b7381704c330d7610acbdd5ddb8bc590d51f6b19907da22.jpg b/parse/train/SkzK4iC5Ym/images/9cd6122f957063481b7381704c330d7610acbdd5ddb8bc590d51f6b19907da22.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3e5ce23d5acb659bcdc8baec2e122b1140f3a0ab --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/9cd6122f957063481b7381704c330d7610acbdd5ddb8bc590d51f6b19907da22.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b23c3059739cc0704f27294ab872dcda3eec32554993957a8aed4ea8677392d0 +size 25371 diff --git a/parse/train/SkzK4iC5Ym/images/9d5cab8cfff14c7defe5df2092b7336be6be1e6b75d14b5636180794a8dfbe9e.jpg b/parse/train/SkzK4iC5Ym/images/9d5cab8cfff14c7defe5df2092b7336be6be1e6b75d14b5636180794a8dfbe9e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1a29966d67a3c4307e61648c049fd6127bb71b19 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/9d5cab8cfff14c7defe5df2092b7336be6be1e6b75d14b5636180794a8dfbe9e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfb48f2abd93d04b501890aaa608c2b6dd2e0d940addc4d597cfd7afc2cdf7c8 +size 3743 diff --git a/parse/train/SkzK4iC5Ym/images/a05c2ab3dbb4fd6d97b3a414c89a901024f082bc8893f3cb0f52aa9ad0b73c6f.jpg b/parse/train/SkzK4iC5Ym/images/a05c2ab3dbb4fd6d97b3a414c89a901024f082bc8893f3cb0f52aa9ad0b73c6f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7aeb0e3f503eba513177bf636ba9d9c664ad0cf8 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a05c2ab3dbb4fd6d97b3a414c89a901024f082bc8893f3cb0f52aa9ad0b73c6f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89924cd7341d8dad1343d1b255db2f3f88170746468088964005e53a7153eabe +size 5690 diff --git a/parse/train/SkzK4iC5Ym/images/a07d66f2c888577328d166d16f99d90566e57de18071d611565eef00d5ff35ca.jpg b/parse/train/SkzK4iC5Ym/images/a07d66f2c888577328d166d16f99d90566e57de18071d611565eef00d5ff35ca.jpg new file mode 100644 index 0000000000000000000000000000000000000000..f60d20b49bad43201d5e41278fab4277fd6d9fa3 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a07d66f2c888577328d166d16f99d90566e57de18071d611565eef00d5ff35ca.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab70993886eac11616991d211b232b4080a9ad8973bf24d6797882e580d13ea3 +size 14336 diff --git a/parse/train/SkzK4iC5Ym/images/a2582ab2712fdd4c9d1ca6f70d7fb37b42f62328726cafa4850dbb06890cc3b7.jpg b/parse/train/SkzK4iC5Ym/images/a2582ab2712fdd4c9d1ca6f70d7fb37b42f62328726cafa4850dbb06890cc3b7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2bd74fa64f8ea141bdf3157ded94ef6af399fbae --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a2582ab2712fdd4c9d1ca6f70d7fb37b42f62328726cafa4850dbb06890cc3b7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97770337d89450e2fc39d211d732ca06be0735ba601667c8d3f96594f0eb61b5 +size 8264 diff --git a/parse/train/SkzK4iC5Ym/images/a30952e6b31b3e6e37fe2ad5ae2cb3fc29be09b982e7e1bb6fde94d6a48ab93b.jpg b/parse/train/SkzK4iC5Ym/images/a30952e6b31b3e6e37fe2ad5ae2cb3fc29be09b982e7e1bb6fde94d6a48ab93b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..041147b5008a9e82c78a2b148d3221b14ec257dc --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a30952e6b31b3e6e37fe2ad5ae2cb3fc29be09b982e7e1bb6fde94d6a48ab93b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d96f63f0b8cb798a263629c32004f820b907e75a8a163b1b3ebdb65cc801ed9 +size 10228 diff --git a/parse/train/SkzK4iC5Ym/images/a31a5615f6fd5a8a7a5c2cfd9e4b5e212162b1afd1e4b5c4e4102e58e4a81cfc.jpg b/parse/train/SkzK4iC5Ym/images/a31a5615f6fd5a8a7a5c2cfd9e4b5e212162b1afd1e4b5c4e4102e58e4a81cfc.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1a9911cfdc77beff96d1dc0d74a5f378cf4c6b11 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a31a5615f6fd5a8a7a5c2cfd9e4b5e212162b1afd1e4b5c4e4102e58e4a81cfc.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e3de5f7f158f6ee041a900009776c01b25d7206c42946624c24a7a9e130ec53 +size 6272 diff --git a/parse/train/SkzK4iC5Ym/images/a65bceb203a3d5ea26ce25fb45b6f127149458d08e6449aed60735468873b373.jpg b/parse/train/SkzK4iC5Ym/images/a65bceb203a3d5ea26ce25fb45b6f127149458d08e6449aed60735468873b373.jpg new file mode 100644 index 0000000000000000000000000000000000000000..99554956264be29e5053069f69c8cd579136a4a9 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a65bceb203a3d5ea26ce25fb45b6f127149458d08e6449aed60735468873b373.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7867ceea6919f48ce2a18edf46d78bfd248ba4eb001bb5a6e0515fe0c312207 +size 19533 diff --git a/parse/train/SkzK4iC5Ym/images/a74353f07d5410bb8651cf5432f0275a38b7c6676ecf177aacf874db67d5cbbc.jpg b/parse/train/SkzK4iC5Ym/images/a74353f07d5410bb8651cf5432f0275a38b7c6676ecf177aacf874db67d5cbbc.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1c46e5e472e401dc803bc58b39091d72ad07ce95 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/a74353f07d5410bb8651cf5432f0275a38b7c6676ecf177aacf874db67d5cbbc.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd936e826118a79d640324764d91306f984eb51bd610404c744a94fe408d0bc7 +size 6904 diff --git a/parse/train/SkzK4iC5Ym/images/aa44c3fe50e858f4b133e11b96dbabea330effc54500e535d4adfe6d934d6a68.jpg b/parse/train/SkzK4iC5Ym/images/aa44c3fe50e858f4b133e11b96dbabea330effc54500e535d4adfe6d934d6a68.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3b1e576899c21c8865b92d408e6a74f7ef99d07e --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/aa44c3fe50e858f4b133e11b96dbabea330effc54500e535d4adfe6d934d6a68.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3010a8e9e29bf1f9c62b793cdc651c0e120c54251938dfb78db333c33dd1be5d +size 8592 diff --git a/parse/train/SkzK4iC5Ym/images/b0696d7097b02f77fa1b85401928394e19098814bc7387a3961c84a164179055.jpg b/parse/train/SkzK4iC5Ym/images/b0696d7097b02f77fa1b85401928394e19098814bc7387a3961c84a164179055.jpg new file mode 100644 index 0000000000000000000000000000000000000000..57d7ff4bfc7029d9cbf394aef8bbe577c09922a5 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/b0696d7097b02f77fa1b85401928394e19098814bc7387a3961c84a164179055.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee549bd98fdff8d67ba4ed627e54f55a200a52c170ade44813651bde957d187e +size 6131 diff --git a/parse/train/SkzK4iC5Ym/images/b62918fe4f8358344e4d7ffabd980bd18f33eab1657a432e1d122cd4126510f0.jpg b/parse/train/SkzK4iC5Ym/images/b62918fe4f8358344e4d7ffabd980bd18f33eab1657a432e1d122cd4126510f0.jpg new file mode 100644 index 0000000000000000000000000000000000000000..fb06b5500ead845b356229f642418d3990f31311 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/b62918fe4f8358344e4d7ffabd980bd18f33eab1657a432e1d122cd4126510f0.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a37f16523423a2af940294193a277eaefbef72b8173164bb7c6b1baabcf0aa0 +size 8674 diff --git a/parse/train/SkzK4iC5Ym/images/b89941236fe04b1eedcac734d678929f14d0404d7e31e8be40e7012ba04fe935.jpg b/parse/train/SkzK4iC5Ym/images/b89941236fe04b1eedcac734d678929f14d0404d7e31e8be40e7012ba04fe935.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2d892783983c4519ccfb83ca5ba8ce8046958df0 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/b89941236fe04b1eedcac734d678929f14d0404d7e31e8be40e7012ba04fe935.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2497eb6f03f601312db2acf40d160b96f4e9a4fe79e71eabf0c6fbe63e02cec +size 5311 diff --git a/parse/train/SkzK4iC5Ym/images/b9a86a8beff1a00eed4d558fc552782affd8651b0d4772907560cf5bef3e098a.jpg b/parse/train/SkzK4iC5Ym/images/b9a86a8beff1a00eed4d558fc552782affd8651b0d4772907560cf5bef3e098a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b579c2d5dfa9b12527b733a6a10637fb2c56e3d4 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/b9a86a8beff1a00eed4d558fc552782affd8651b0d4772907560cf5bef3e098a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbc1379bb6ca2de0d3addd12c1d427b78d4351c21c46d203549638eee41a49bd +size 9898 diff --git a/parse/train/SkzK4iC5Ym/images/ba5940fd74359d24d48af4e835036feea20df96d80f661367a006265b1f0484d.jpg b/parse/train/SkzK4iC5Ym/images/ba5940fd74359d24d48af4e835036feea20df96d80f661367a006265b1f0484d.jpg new file mode 100644 index 0000000000000000000000000000000000000000..36c6b0346e6802d3d2c3a30ddec3f3aeb537dbe5 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/ba5940fd74359d24d48af4e835036feea20df96d80f661367a006265b1f0484d.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9ee73b53b31b8f904ae9d2c8fef8288a30bf6467e22c7940a336a49386dd034 +size 6758 diff --git a/parse/train/SkzK4iC5Ym/images/c00add28b80bcde3b0a06480e4fabef5392387eb41434187aec7e14267dbb809.jpg b/parse/train/SkzK4iC5Ym/images/c00add28b80bcde3b0a06480e4fabef5392387eb41434187aec7e14267dbb809.jpg new file mode 100644 index 0000000000000000000000000000000000000000..107cd502c7f6eaa7de6bf7a238ff320b8c1c383b --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/c00add28b80bcde3b0a06480e4fabef5392387eb41434187aec7e14267dbb809.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ccd426dfced1af31ded279377524363c4cd6f8971155bd752567301c395f10f +size 5614 diff --git a/parse/train/SkzK4iC5Ym/images/c22f70867db6c130b36bb3e51edc50f8c7d11158f728352fb8fe49a0d27f8875.jpg b/parse/train/SkzK4iC5Ym/images/c22f70867db6c130b36bb3e51edc50f8c7d11158f728352fb8fe49a0d27f8875.jpg new file mode 100644 index 0000000000000000000000000000000000000000..5e27a0078bd4d6aebe009ffd1745173140241b5c --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/c22f70867db6c130b36bb3e51edc50f8c7d11158f728352fb8fe49a0d27f8875.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d032537392c2d7a474f7d2c2c4a02619f0142f38956faec1a703da3b9c6e04a7 +size 26351 diff --git a/parse/train/SkzK4iC5Ym/images/c24f6a8df8cabcb5f393a80bed7a62df55eff3e7401b4dfc0c75b6813d122355.jpg b/parse/train/SkzK4iC5Ym/images/c24f6a8df8cabcb5f393a80bed7a62df55eff3e7401b4dfc0c75b6813d122355.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4dbc42f80582b0cbf338cf1ab36a9f2f088144cf --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/c24f6a8df8cabcb5f393a80bed7a62df55eff3e7401b4dfc0c75b6813d122355.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97bb5d05aad619d3e4f4f41f15b45f2bdc27e7d7e7b7d6ef23c18c95600bdf45 +size 4416 diff --git a/parse/train/SkzK4iC5Ym/images/c2de0a222cf6c3cc0519d61d6f00da5134b183c14be202fd48e8def0c9c4bb6f.jpg b/parse/train/SkzK4iC5Ym/images/c2de0a222cf6c3cc0519d61d6f00da5134b183c14be202fd48e8def0c9c4bb6f.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b9ac364f512c788c0e634e9302293ad23ad5a265 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/c2de0a222cf6c3cc0519d61d6f00da5134b183c14be202fd48e8def0c9c4bb6f.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56ee6c81a0da6f5d02b15d836e9defcd5c54d9a5e0e363e37e6cfee5d056ec9f +size 29271 diff --git a/parse/train/SkzK4iC5Ym/images/c76edb54f96a8bd3f77c18417e1edc7feb1847caef55144c2435926eccc8b207.jpg b/parse/train/SkzK4iC5Ym/images/c76edb54f96a8bd3f77c18417e1edc7feb1847caef55144c2435926eccc8b207.jpg new file mode 100644 index 0000000000000000000000000000000000000000..1909ef46f21aa2e646b37d5bfb7307bd9bb72624 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/c76edb54f96a8bd3f77c18417e1edc7feb1847caef55144c2435926eccc8b207.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1335d5a214287a1a2a0bf80b57f062b4f88a338444bc65c191f3a81c5fc1a956 +size 15993 diff --git a/parse/train/SkzK4iC5Ym/images/cc7153013bc987c5a6692ee4b5ddc3478f225b071b081751f5ad3149214421a7.jpg b/parse/train/SkzK4iC5Ym/images/cc7153013bc987c5a6692ee4b5ddc3478f225b071b081751f5ad3149214421a7.jpg new file mode 100644 index 0000000000000000000000000000000000000000..4ff49e6d169d69b7b56f7a374ced67c3f5599d31 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/cc7153013bc987c5a6692ee4b5ddc3478f225b071b081751f5ad3149214421a7.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00983ee0adf8a7d6591f4dfec70808db86f2b16546e6a52bd1ad53e4c63e7ec7 +size 7274 diff --git a/parse/train/SkzK4iC5Ym/images/ccd68674a6efd440425ef4a2c4db7f00e57796d60bdaf9cc3f37cf5a13d69784.jpg b/parse/train/SkzK4iC5Ym/images/ccd68674a6efd440425ef4a2c4db7f00e57796d60bdaf9cc3f37cf5a13d69784.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9a631e5b1d6ad2639f508caf3517b6a5b16178e5 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/ccd68674a6efd440425ef4a2c4db7f00e57796d60bdaf9cc3f37cf5a13d69784.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b17f756fef3f109256b0cc044d4d39775068a26c63e85c9b886da2a70f455c9 +size 5258 diff --git a/parse/train/SkzK4iC5Ym/images/cec1045d7fdeb92f6517727ceeaedf1ddae2adb3f59af2a689f2cf7bb0aa12bf.jpg b/parse/train/SkzK4iC5Ym/images/cec1045d7fdeb92f6517727ceeaedf1ddae2adb3f59af2a689f2cf7bb0aa12bf.jpg new file mode 100644 index 0000000000000000000000000000000000000000..d552d6ec51f322a230d378c8ad57d9af31beeae8 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/cec1045d7fdeb92f6517727ceeaedf1ddae2adb3f59af2a689f2cf7bb0aa12bf.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be32cedb44adf7d781ed953c3100e57d0688f5477f47b47ab3383293add28709 +size 6030 diff --git a/parse/train/SkzK4iC5Ym/images/d0a7189dbbdfd97ffdf67ca904f9aa766e1aa93a09f6d4619ff461c99e6cafa5.jpg b/parse/train/SkzK4iC5Ym/images/d0a7189dbbdfd97ffdf67ca904f9aa766e1aa93a09f6d4619ff461c99e6cafa5.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a9a349dddd6067e04c7c86bb6b57b9086b688b86 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/d0a7189dbbdfd97ffdf67ca904f9aa766e1aa93a09f6d4619ff461c99e6cafa5.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2f9d01ae928a14e3d9e860ba2436ba7d051f0b5f6fb3712400c2cd2833b098f +size 7969 diff --git a/parse/train/SkzK4iC5Ym/images/d0d4059821c8c00115333e7f5118e8a632a9583979aa359099c87150e0bda16b.jpg b/parse/train/SkzK4iC5Ym/images/d0d4059821c8c00115333e7f5118e8a632a9583979aa359099c87150e0bda16b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..eb74961691bf68d7644d84a3dc1822eca299e76a --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/d0d4059821c8c00115333e7f5118e8a632a9583979aa359099c87150e0bda16b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21ba6de3e6cb9de16d789115e9cb135eaf15a62d2e8cdedc3802d72adb136e78 +size 19395 diff --git a/parse/train/SkzK4iC5Ym/images/d2a642059161a5e357448569e34dbe7b443b763294d01fa9aa1d49d4845bdbbb.jpg b/parse/train/SkzK4iC5Ym/images/d2a642059161a5e357448569e34dbe7b443b763294d01fa9aa1d49d4845bdbbb.jpg new file mode 100644 index 0000000000000000000000000000000000000000..7e6f44c4b92ef1aa2d352b738d3a5507c945df80 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/d2a642059161a5e357448569e34dbe7b443b763294d01fa9aa1d49d4845bdbbb.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86cc306e326b0986b37dbbdf84588d3af6e3115040890b7418bcffbdeeacfc93 +size 9306 diff --git a/parse/train/SkzK4iC5Ym/images/d5294125968c2617a36297a9083d6a36bf5c6b90265baf563787d305312a7895.jpg b/parse/train/SkzK4iC5Ym/images/d5294125968c2617a36297a9083d6a36bf5c6b90265baf563787d305312a7895.jpg new file mode 100644 index 0000000000000000000000000000000000000000..31450c1b7d4ac6752c5320aeae93d6c75269bace --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/d5294125968c2617a36297a9083d6a36bf5c6b90265baf563787d305312a7895.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eae7365d3310e91a64a6afb6ef073d7a2c625623c488d12353ffffd2ef18052 +size 7749 diff --git a/parse/train/SkzK4iC5Ym/images/da9b724a344b0488a7b3a9d92e6858ed428c1a210ef42c46bca9d59ad039be28.jpg b/parse/train/SkzK4iC5Ym/images/da9b724a344b0488a7b3a9d92e6858ed428c1a210ef42c46bca9d59ad039be28.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2e879a9387182858faf9c2a0277edf86e74c9115 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/da9b724a344b0488a7b3a9d92e6858ed428c1a210ef42c46bca9d59ad039be28.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88a84bff7e1e16faadd2a0a8a0269ffc26601a035046c849abadef988b28b262 +size 58122 diff --git a/parse/train/SkzK4iC5Ym/images/dec525ae42cf80c68b243d76c8dc79e8ac903c82706fd4eb2a97422b68d7cd1e.jpg b/parse/train/SkzK4iC5Ym/images/dec525ae42cf80c68b243d76c8dc79e8ac903c82706fd4eb2a97422b68d7cd1e.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ad7554c3bc17630fdb5015c269ea133d31d80d72 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/dec525ae42cf80c68b243d76c8dc79e8ac903c82706fd4eb2a97422b68d7cd1e.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c1c2919cd918b304dfbae79199b2a77f8afc35100bf9ebe082d8332547a223d +size 6249 diff --git a/parse/train/SkzK4iC5Ym/images/e16dd171daf9583a2c31e3d9becf8e0f9eaa9a815ac92ef83ff3961d2e61c5e9.jpg b/parse/train/SkzK4iC5Ym/images/e16dd171daf9583a2c31e3d9becf8e0f9eaa9a815ac92ef83ff3961d2e61c5e9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3e2c8e172187e0c6bf2c91bcffeb68b28f8fb1cc --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/e16dd171daf9583a2c31e3d9becf8e0f9eaa9a815ac92ef83ff3961d2e61c5e9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b9f08c4c1b44ebf78242f1bc9a4d7b68f87d1227acf928c2253e24bd27e3765 +size 3999 diff --git a/parse/train/SkzK4iC5Ym/images/e40141545c4a6ca0b259332ceeae6604c775382234d9d124c6316e990db0ee19.jpg b/parse/train/SkzK4iC5Ym/images/e40141545c4a6ca0b259332ceeae6604c775382234d9d124c6316e990db0ee19.jpg new file mode 100644 index 0000000000000000000000000000000000000000..2e475832112c6f591dc930ff7c34ed2c375a5110 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/e40141545c4a6ca0b259332ceeae6604c775382234d9d124c6316e990db0ee19.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c7a8fa89c63f7a1f23d73208ba0b68e92cb61216fe6784c58427dfccf45bee1 +size 7679 diff --git a/parse/train/SkzK4iC5Ym/images/e533b80e0e4c4120e695c4fbac7b311bf5ca9f36e3ac79ee87514515bebbf61b.jpg b/parse/train/SkzK4iC5Ym/images/e533b80e0e4c4120e695c4fbac7b311bf5ca9f36e3ac79ee87514515bebbf61b.jpg new file mode 100644 index 0000000000000000000000000000000000000000..70c1a968ca28b9b5fe46bb2bf4250f56bf956c87 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/e533b80e0e4c4120e695c4fbac7b311bf5ca9f36e3ac79ee87514515bebbf61b.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d9c2c2511ddcd721d39fc9f8e3afc396af58b18c962e8073f27bfbb9fe7706e +size 9178 diff --git a/parse/train/SkzK4iC5Ym/images/e598501c5987f035588414a60c20f2382eb3b945592c9cc8ad8fd94141bf33f4.jpg b/parse/train/SkzK4iC5Ym/images/e598501c5987f035588414a60c20f2382eb3b945592c9cc8ad8fd94141bf33f4.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a0d8d37434defc521335e3fef476a18be1790812 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/e598501c5987f035588414a60c20f2382eb3b945592c9cc8ad8fd94141bf33f4.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4afe11761e20d87e6b9a4c364899392cdbf718313822932b345a31990ea2649a +size 5894 diff --git a/parse/train/SkzK4iC5Ym/images/e67fc942ff6de252b124ac95342b30959a7df47643a421312bfbaa57ee0aaf68.jpg b/parse/train/SkzK4iC5Ym/images/e67fc942ff6de252b124ac95342b30959a7df47643a421312bfbaa57ee0aaf68.jpg new file mode 100644 index 0000000000000000000000000000000000000000..9703a7d6b9f8e39b5e424bae9c94fe95d567a7d5 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/e67fc942ff6de252b124ac95342b30959a7df47643a421312bfbaa57ee0aaf68.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92ecbe241ad081ab308bbf85820d1a6c7e3bf876dab18ac2498ffd04625db369 +size 3246 diff --git a/parse/train/SkzK4iC5Ym/images/e75fc55c58174dcd65a2c24033c40cf8cb0df28d507077b3ae27c5a1f9854238.jpg b/parse/train/SkzK4iC5Ym/images/e75fc55c58174dcd65a2c24033c40cf8cb0df28d507077b3ae27c5a1f9854238.jpg new file mode 100644 index 0000000000000000000000000000000000000000..feedc53dcabda404f91ddd1375de1e33ba040906 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/e75fc55c58174dcd65a2c24033c40cf8cb0df28d507077b3ae27c5a1f9854238.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85a7bba976f798d0a28a2e102088937ddd7acdca29738fdc83f637ca158231a1 +size 34966 diff --git a/parse/train/SkzK4iC5Ym/images/eaded081b85ff54c62a93b7c6d3abfa341179584d1cf044503b2d015e552bfd9.jpg b/parse/train/SkzK4iC5Ym/images/eaded081b85ff54c62a93b7c6d3abfa341179584d1cf044503b2d015e552bfd9.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3209e5eb2dff8662f03ab8f53bf1f1f303a3a720 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/eaded081b85ff54c62a93b7c6d3abfa341179584d1cf044503b2d015e552bfd9.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e625bf19b27c73dee972ecfb987341a4b7bce5411f52f7dffd88bcd934b3c3d +size 9448 diff --git a/parse/train/SkzK4iC5Ym/images/f0c1609ef0781fe33f30ef921ce74b0d8700669caf110b605b6904bee0abdc27.jpg b/parse/train/SkzK4iC5Ym/images/f0c1609ef0781fe33f30ef921ce74b0d8700669caf110b605b6904bee0abdc27.jpg new file mode 100644 index 0000000000000000000000000000000000000000..19e565544f4304c4f04d319616e37345e90183ce --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/f0c1609ef0781fe33f30ef921ce74b0d8700669caf110b605b6904bee0abdc27.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89f17142d7acb93ed9009a395b38b90f7d0f3c91623f0476ce44fee8b844661f +size 10954 diff --git a/parse/train/SkzK4iC5Ym/images/f370c67fa605508301969974e93b886926b30226ef87b90287803fb1229809ec.jpg b/parse/train/SkzK4iC5Ym/images/f370c67fa605508301969974e93b886926b30226ef87b90287803fb1229809ec.jpg new file mode 100644 index 0000000000000000000000000000000000000000..e8a320c9485df9815283cc37470dd8dde2a24f1c --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/f370c67fa605508301969974e93b886926b30226ef87b90287803fb1229809ec.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9ca2dc75fe5a6c26f59f8832952f7b975b2ffa5e57421c7c0ed5187a60d183a +size 4237 diff --git a/parse/train/SkzK4iC5Ym/images/f3900582c3bb1567acdb716ab19d6b86cb04e19deb6ec5f757d4a7abb6ed29b2.jpg b/parse/train/SkzK4iC5Ym/images/f3900582c3bb1567acdb716ab19d6b86cb04e19deb6ec5f757d4a7abb6ed29b2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..70bfd9f374e996bf0f22b46950ed361db6e5be8a --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/f3900582c3bb1567acdb716ab19d6b86cb04e19deb6ec5f757d4a7abb6ed29b2.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3de0566868a28dcca40f348667dee8dbadebc806ae2bd512c06ba3ad0bfc5808 +size 54955 diff --git a/parse/train/SkzK4iC5Ym/images/f8af08eba22728fa1e7fac8a1c62e74325217ff6d94062cfff6d1d9cc49880ba.jpg b/parse/train/SkzK4iC5Ym/images/f8af08eba22728fa1e7fac8a1c62e74325217ff6d94062cfff6d1d9cc49880ba.jpg new file mode 100644 index 0000000000000000000000000000000000000000..05bba0f187247ae8a1e1d996da8bcdc5010a19ac --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/f8af08eba22728fa1e7fac8a1c62e74325217ff6d94062cfff6d1d9cc49880ba.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d41c640a68db31582853216bfdd35d2fd9f8e7f2e4845c46b678c404fa6de6d +size 5947 diff --git a/parse/train/SkzK4iC5Ym/images/fa2e3c616c76b18fb4b4b77773de7a77746808f91379ae6968a5d92d5c954f8a.jpg b/parse/train/SkzK4iC5Ym/images/fa2e3c616c76b18fb4b4b77773de7a77746808f91379ae6968a5d92d5c954f8a.jpg new file mode 100644 index 0000000000000000000000000000000000000000..449c968e3ccb5b6e2ab513f9d89392ece8706e23 --- /dev/null +++ b/parse/train/SkzK4iC5Ym/images/fa2e3c616c76b18fb4b4b77773de7a77746808f91379ae6968a5d92d5c954f8a.jpg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1b429194e56d058e441e260824058f9e23c876248afcbf6cf9f9982bda3a4e4 +size 2310 diff --git a/parse/train/SyCSsUDee/SyCSsUDee.md b/parse/train/SyCSsUDee/SyCSsUDee.md new file mode 100644 index 0000000000000000000000000000000000000000..fd1280be8813bbfd0071f0160c61d5c152c85847 --- /dev/null +++ b/parse/train/SyCSsUDee/SyCSsUDee.md @@ -0,0 +1,323 @@ +# SEMANTIC NOISE MODELING FORBETTER REPRESENTATION LEARNING + +Hyo-Eun $\mathbf { K i m } ^ { * }$ and Sangheum Hwang + +Lunit Inc. +Seoul, South Korea +{hekim, shwang}@lunit.io +Kyunghyun Cho +Courant Institute of Mathematical Sciences and Centre for Data Science +New York University +New York, NY 10012, USA +kyunghyun.cho@nyu.edu + +# ABSTRACT + +Latent representation learned from multi-layered neural networks via hierarchical feature abstraction enables recent success of deep learning. Under the deep learning framework, generalization performance highly depends on the learned latent representation. In this work, we propose a novel latent space modeling method to learn better latent representation. We designed a neural network model based on the assumption that good base representation for supervised tasks can be attained by maximizing the sum of hierarchical mutual informations between the input, latent, and output variables. From this base model, we introduce a semantic noise modeling method which enables semantic perturbation on the latent space to enhance the representational power of learned latent feature. During training, latent vector representation can be stochastically perturbed by a modeled additive noise while preserving its original semantics. It implicitly brings the effect of semantic augmentation on the latent space. The proposed model can be easily learned by back-propagation with common gradient-based optimization algorithms. Experimental results show that the proposed method helps to achieve performance benefits against various previous approaches. We also provide the empirical analyses for the proposed latent space modeling method including t-SNE visualization. + +# 1 INTRODUCTION + +Enhancing the generalization performance against unseen data given some sample data is the main objective in machine learning. Under that point of view, deep learning has been achieved many breakthroughs in several domains such as computer vision (Krizhevsky et al., 2012; Simonyan & Zisserman, 2015; He et al., 2016), natural language processing (Collobert & Weston, 2008; Bahdanau et al., 2015), and speech recognition (Hinton et al., 2012; Graves et al., 2013). Deep learning is basically realized on deep layered neural network architecture, and it learns appropriate taskspecific latent representation based on given training data. Better latent representation learned from training data results in better generalization over the future unseen data. Representation learning or latent space modeling becomes one of the key research topics in deep learning. During the past decade, researchers focused on unsupervised representation learning and achieved several remarkable landmarks on deep learning history (Vincent et al., 2010; Hinton et al., 2006; Salakhutdinov & Hinton, 2009). In terms of utilizing good base features for supervised learning, the base representation learned from unsupervised learning can be a good solution for supervised tasks (Bengio et al., 2007; Masci et al., 2011). + +The definition of ‘good’ representation is, however, different according to target tasks. In unsupervised learning, a model is learned from unlabelled examples. Its main objective is to build a model to estimate true data distribution given examples available for training, so the learned latent representation normally includes broadly-informative components of the raw input data (e.g., mutual information between the input and the latent variable can be maximized for this objective). In supervised learning, however, a model is learned from labelled examples. In the case of classification, a supervised model learns to discriminate input data in terms of the target task using corresponding labels. Latent representation is therefore obtained to maximize the performance on the target supervised tasks. + +Since the meaning of good representations vary according to target tasks (unsupervised or supervised), pre-trained features from the unsupervised model are not be guaranteed to be useful for subsequent supervised tasks. Instead of the two stage learning strategy (unsupervised pre-training followed by supervised fine-tuning), several works focused on a joint learning model which optimizes unsupervised and supervised objectives concurrently, resulting in better generalization performance (Goodfellow et al., 2013; Larochelle & Bengio, 2008a; Rasmus et al., 2015; Zhao et al., 2015; Zhang et al., 2016; Cho & Chen, 2014). + +In this work, we propose a novel latent space modeling method for supervised learning as an extension of the joint learning approach. We define a good latent representation of standard feed-forward neural networks under the basis of information theory. Then, we introduce a semantic noise modeling method in order to enhance the generalization performance. The proposed method stochastically perturbs the latent representation of a training example by injecting a modeled semantic additive noise. Since the additive noise is randomly sampled from a pre-defined probability distribution every training iteration, different latent vectors from a single training example can be fully utilized during training. The multiple different latent vectors produced from a single training example are semantically similar under the proposed latent space modeling method, so we can expect semantic augmentation effect on the latent space. + +Experiments are performed on two datasets; MNIST and CIFAR-10. The proposed model results in better classification performance compared to previous approaches through notable generalization effect (stochastically perturbed training examples well cover the distribution of unseen data). + +# 2 METHODOLOGY + +The proposed method starts from the existing joint learning viewpoint. This section first explains the process of obtaining a good base representation for supervised learning which is the basis of the proposed latent space modeling method. And then, we will describe how the proposed semantic noise modeling method perturbs the latent space while maintaining the original semantics. + +# 2.1 BASE JOINT LEARNING MODEL + +In a traditional feed-forward neural network model (Figure 1(a)), output $Y$ of input data $X$ is compared with its true label, and the error is propagated backward from top to bottom, which implicitly learns a task-specific latent representation $Z$ of the input $X$ . As an extension of a joint learning approach, an objective to be optimized can be described in general as below (Larochelle & Bengio, 2008b): + +$$ +\operatorname* { m i n } _ { \theta } \lambda L _ { u n s u p } + L _ { s u p } +$$ + +where $L _ { u n s u p }$ and $L _ { s u p }$ are respectively an unsupervised loss and a supervised loss, and $\theta$ and $\lambda$ are model parameters to be optimized during training and a loss weighting coefficient, respectively. In terms of modeling $L _ { u n s u p }$ in Eq. (1), we assume that good latent representation $Z$ is attained by maximizing the sum of hierarchical mutual informations between the input, latent, and output variables; i.e. the sum of the mutual information between the input $X$ and the $Z$ and the mutual information between the $Z$ and the output $Y$ . Each mutual information is decomposed into an entropy and a conditional entropy terms, so the sum of hierarchical mutual informations is expressed as follows: + +$$ +{ \bf { I } } ( X ; Z ) + { \bf { I } } ( Z ; Y ) = { \bf { H } } ( X ) - { \bf { H } } ( X \vert Z ) + { \bf { H } } ( Z ) - { \bf { H } } ( Z \vert Y ) +$$ + +![](images/1137fde88fa95773239a9891f73a9430060b5253fa34e7cd20bedd7afffb1039.jpg) +Figure 1: (a) Standard feed-forward neural network model, (b) feed-forward neural network model with reconstruction paths, and (c) feed-forward neural network model with reconstruction and stochastic perturbation paths. + +where $\mathbf { I } ( \cdot ; \cdot )$ is the mutual information between random variables, and $\mathbf { H } ( \cdot )$ and $\mathbf { H } ( \cdot | \cdot )$ are the entropy and the conditional entropy of random variables, respectively. Note that the sum of those mutual informations becomes equivalent to the total correlation of $X , Z$ , and $Y$ under the graphical structure of the general feed-forward model described in Figure 1(a); $P ( X , Z , Y ) = P ( Y | \bar { Z } ) \bar { P } ( Z | X ) P ( X )$ . The total correlation is equal to the sum of all pairwise mutual informations (Watanabe, 1960). + +Our objective is to find the model parameters which maximize $\mathbf { I } ( X ; Z ) + \mathbf { I } ( Z ; Y )$ . Since $\mathbf { H } ( X )$ and $\mathbf { H } ( Z )$ are non-negative, and $\mathbf { H } ( X )$ is constant in this case, the lower bound on $\mathbf { I } ( X ; Z ) + \mathbf { I } ( Z ; Y )$ can be reduced to 1: + +$$ +\mathbf { I } ( X ; Z ) + \mathbf { I } ( Z ; Y ) \geq - \mathbf { H } ( X | Z ) - \mathbf { H } ( Z | Y ) . +$$ + +It is known that maximizing $- \mathbf { H } ( X | Z )$ can be formulated as minimizing the reconstruction error between the input $x ^ { ( i ) }$ ( $i$ -th example sampled from $X$ ) and its reconstruction $x _ { R } ^ { ( i ) }$ under the general audo-encoder framework (Vincent et al., 2010). Since $\mathbf { H } ( X | Z ) + \mathbf { H } ( Z | Y )$ is proportional to the sum of reconstruction errors of $x ^ { ( i ) }$ (with its reconstruction $x _ { R } ^ { ( i ) }$ ) and $z ^ { ( i ) }$ (with its reconstruction $z _ { R } ^ { ( i ) } )$ ), the target objective can be expressed as follows (refer to Appendix (A1) for the details ofhematical derivations): + +$$ +\operatorname* { m i n } _ { \theta } \sum _ { i } L _ { r e c } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { r e c } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) +$$ + +where $L _ { r e c }$ is a reconstruction loss. + +Figure 1(b) shows the target model obtained from the assumption that good latent representation $Z$ can be obtained by maximizing the sum of hierarchical mutual informations. Given an input sample $x$ , feed-forward vectors and their reconstructions are attained deterministically by: + +$$ +\begin{array} { r l } & { z = f _ { \theta _ { 1 } } ( x ) } \\ & { y = f _ { \theta _ { 2 } } ( f _ { \theta _ { 1 } } ( x ) ) } \\ & { x _ { R } = g _ { \theta _ { 1 } ^ { \prime } } ( z ) = g _ { \theta _ { 1 } ^ { \prime } } ( f _ { \theta _ { 1 } } ( x ) ) } \\ & { z _ { R } = g _ { \theta _ { 2 } ^ { \prime } } ( y ) = g _ { \theta _ { 2 } ^ { \prime } } ( f _ { \theta _ { 2 } } ( f _ { \theta _ { 1 } } ( x ) ) . } \end{array} +$$ + +Given a set of training pairs $( x ^ { ( i ) } , t ^ { ( i ) } )$ where $x ^ { ( i ) }$ and $t ^ { ( i ) }$ are the $i$ -th input example and its label, target objective in Eq. (1) under the model described in Figure 1(b) can be organized as below (with real-valued input samples, L2 loss $L _ { L 2 }$ is a proper choice for the reconstruction loss $L _ { r e c , }$ ): + +$$ +\operatorname* { m i n } _ { \theta : \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } , \theta _ { 2 } ^ { \prime } \} } \sum _ { i } \lambda \left( L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \right) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) +$$ + +where $L _ { N L L }$ is a negative log-likelihood loss for the target supervised task. Note that Eq. (6) represents the ‘proposed-base’ in our experiment (see Section 4.3). + +# 2.2 SEMANTIC NOISE MODELING + +Based on the architecture shown in Figure 1(b) with the target objective in Eq. (6), we conjecture that stochastic perturbation on the latent space during training helps to achieve better generalization performance for supervised tasks. Figure 1(c) shows this strategy which integrates the stochastic perturbation process during training. Suppose that $Z _ { P }$ is a perturbed version of $Z$ , and $Y _ { P }$ is an output which is feed-forwarded from $Z _ { P }$ . Given a latent vector $z = f _ { \boldsymbol { \theta } _ { 1 } } ( x )$ from an input sample $x$ , + +$$ +z ^ { \prime } = z + z _ { e } { \mathrm { ~ } } \mathrm { a n d } { \hat { y } } = f _ { \theta _ { 2 } } ( z ^ { \prime } ) +$$ + +where $z ^ { \prime }$ and $\hat { y }$ are a perturbed latent vector and its output respectively, and $z _ { e }$ is an additive noise used in the perturbation process of $z$ . Based on the architecture shown in Figure 1(c), target objective can be modified as: + +$$ +\operatorname* { m i n } _ { \substack { \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } , \theta _ { 2 } ^ { \prime } \} } } \sum _ { i } \lambda _ { 1 } \left( L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \right) + \lambda _ { 2 } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) + L _ { N L L } ( \hat { y } ^ { ( i ) } , t ^ { ( i ) } ) . +$$ + +Using random additive noise directly on $z _ { e }$ is the most intuitive approach (‘proposed-perturb (random)’ in Section 4.3). However, preserving the semantics of the original latent representation $z$ cannot be guaranteed under the direct random perturbation on the latent space. While the latent space is not directly interpretable in general, the output logit $y$ of the latent representation $z$ is interpretable, because the output logit is tightly coupled to the prediction of the target label. In order to preserve the semantics of the original latent representation after perturbation, we indirectly model a semantic noise on the latent space by adding small random noise directly on the output space. + +Based on the output (pre-softmax) logit $y$ , the semantic-preserving variation of $y$ (i.e. $y ^ { \prime }$ ) can be modeled by $y ^ { \prime } = y + y _ { e }$ , where $y _ { e }$ is a random noise vector stochastically sampled from a zeromean Gaussian with small standard deviation $\sigma ; \mathcal { N } ( \mathbf { 0 } , \sigma ^ { 2 } \mathbf { I } )$ . Now, the semantic perturbation $z ^ { \prime }$ can be reconstructed from the random perturbation $y ^ { \prime }$ through the decoding path $g _ { \theta _ { 2 } ^ { \prime } }$ in Figure 1(c). From the original output logit $y$ and the randomly perturbed output logit $y ^ { \prime }$ , semantic additive noise $z _ { e }$ on the latent space can be approximately modeled as below: + +$$ +\begin{array} { r l } & { z _ { R } = g _ { \theta _ { 2 } ^ { \prime } } ( y ) } \\ & { z _ { R } ^ { \prime } = g _ { \theta _ { 2 } ^ { \prime } } ( y ^ { \prime } ) = g _ { \theta _ { 2 } ^ { \prime } } ( y + y _ { e } ) } \\ & { z _ { e } \simeq z _ { R } ^ { \prime } - z _ { R } = g _ { \theta _ { 2 } ^ { \prime } } ( y + y _ { e } ) - g _ { \theta _ { 2 } ^ { \prime } } ( y ) } \end{array} +$$ + +By using the modeled semantic additive noise $z _ { e }$ and the original latent representation $z$ , we can obtain the semantic perturbation $z ^ { \prime }$ as well as its output $\hat { y }$ via Eq. (7) for our target objective Eq. (8). + +From the described semantic noise modeling process (‘proposed-perturb (semantic)’ in Section 4.3), we expect to achieve better representation on the latent space. The effect of the proposed model in terms of learned latent representation will be explained in more detail in Section 4.4. + +![](images/687514226030a4d7c6ee1ffba4053d410f23d742a7430240e9b83e42fb27c8fd.jpg) +Figure 2: Previous works for supervised learning; (a) traditional feed-forward model, and (b) joint learning model with both supervised and unsupervised losses. + +# 3 RELATED WORKS + +Previous works on deep neural networks for supervised learning can be categorized into two types as shown in Figure 2; (a) a general feed-forward neural network model (LeCun et al., 1998; Krizhevsky et al., 2012; Simonyan & Zisserman, 2015; He et al., 2016), and (b) a joint learning model which optimizes unsupervised and supervised objectives at the same time (Zhao et al., 2015; Zhang et al., 2016; Cho & Chen, 2014). Here are the corresponding objective functions: + +$$ +\begin{array} { c } { { \displaystyle { \operatorname* { m i n } _ { \theta : \{ \theta _ { 1 } , \theta _ { 2 } \} } \sum _ { i } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) } } } \\ { { \displaystyle { \operatorname* { m i n } _ { \theta : \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } \} } \sum _ { i } \lambda L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) } } } \end{array} +$$ + +where $\lambda$ is a loss weighting coefficient between unsupervised and supervised losses. + +Since the feed-forward neural network model is normally implemented with multiple layers in a deep learning framework, the joint learning model can be sub-classified into two types according to the type of reconstruction; reconstruction only with the input data $x$ (Eq. (11)) and reconstruction with all the intermediate features including the input data $x$ as follows: + +$$ +\operatorname* { m i n } _ { \theta } \sum _ { i } \left( \lambda _ { 0 } L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + \sum _ { j } \lambda _ { j } L _ { L 2 } ( h _ { j } ^ { ( i ) } , h _ { j R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) \right) \ . +$$ + +where h(i)j and $h _ { j _ { R } } ^ { ( i ) }$ are the $j$ -th hidden representation of the $i$ -th training example and its reconstruction. + +Another type of the joint learning model, a ladder network (Figure 3), was introduced for semisupervised learning (Rasmus et al., 2015). The key concept of the ladder network is to obtain robust features by learning de-noising functions $( g _ { \theta ^ { \prime } } )$ of the representations at every layer of the model via reconstruction losses, and the supervised loss is combined with the reconstruction losses in order to build the semi-supervised model. The ladder network achieved the best performance in semi-supervised tasks, but it is not appropriate for supervised tasks with small-scale training set (experimental analysis for supervised learning on permutation-invariant MNIST is briefly summarized in Appendix (A2)). The proposed model in this work can be extended to semi-supervised learning, but our main focus is to enhance the representational power on latent space given labelled data for supervised learning. We leave the study for semi-supervised learning scenario based on the proposed methodology as our future research. + +![](images/0adf99802513d294306f73fcbe86e7d7c12a66294e544f6d889b4fc2e8016158.jpg) +Figure 3: Ladder network; a representative model for semi-supervised learning (Rasmus et al., 2015). + +# 4 EXPERIMENTS + +For quantitative analysis, we compare the proposed methodology with previous approaches described in Section 3; a traditional feed-forward supervised learning model and a joint learning model with two different types of reconstruction losses (reconstruction only with the first layer or with all the intermediate layers including the first layer). The proposed methodology includes a baseline model in Figure 1(b) as well as a stochastic perturbation model in Figure 1(c). Especially in the stochastic perturbation model, we compare the random and semantic perturbations and present some qualitative analysis on the meaning of the proposed perturbation methodology. + +# 4.1 DATASETS + +We experiment with two public datasets; MNIST (including a permutation-invariant MNIST case) and CIFAR-10. MNIST (10 classes) consists of $5 0 \mathrm { k }$ , $1 0 \mathrm { k }$ , and 10k $2 8 \times 2 8$ gray-scale images for training, validation, and test datasets, respectively. CIFAR-10 (10 classes) consists of $5 0 \mathrm { k }$ and 10k $3 2 \times 3 2$ 3-channel images for training and test sets, respectively. We split the $5 0 \mathrm { k }$ CIFAR-10 training images into 40k and 10k for training and validation. Experiments are performed with different sizes of training set (from 10 examples per class to the entire training set) in order to verify the effectiveness of the proposed model in terms of generalization performance under varying sizes of training set. + +# 4.2 IMPLEMENTATION + +Figure 4 shows the architecture of the neural network model used in this experiment. W ’s are convolution or fully-connected weights (biases are excluded for visual brevity). Three convolution $3 { \times } 3$ (2) 32, $3 \times 3$ (2) 64, $3 \times 3$ (2) 96, where each item means the filter kernel size and (stride) with the number of filters) and two fully-connected (the numbers of output nodes are 128 and 10, respectively) layers are used for MNIST. For the permutation-invariant MNIST setting, 784-512- 256-256-128-10 nodes of fully-connected layers are used. Four convolution $( 5 \times 5$ (1) 64, $3 \times 3$ (2) 64, $3 \times 3$ (2) 64, and $3 \times 3$ (2) 96) and three fully-connected (128, 128, and 10 nodes) layers are used for CIFAR-10. Weights on the decoding (reconstruction) path are tied with corresponding weights on the encoding path as shown in Figure 4 (transposed convolution for the tied convolution layer and transposed matrix multiplication for the tied fully-connected layer). + +In Figure 4, $z ^ { \prime }$ is perturbed directly from $z$ by adding Gaussian random noise for random perturbation. For semantic perturbation, $z ^ { \prime }$ is indirectly generated from $y ^ { \prime }$ which is perturbed by adding Gaussian random noise on $y$ based on Eq. (9). For perturbation, base activation vector ( $z$ is the base + +$$ +\begin{array} { r l r } & { } & { z ^ { \prime } \xrightarrow { W _ { N } } \hat { y } \cdots t } \\ & { } & { x \xrightarrow [ ] { W _ { 1 } } \gamma _ { h _ { 1 } } \xrightarrow [ ] { W _ { 2 } } h _ { 2 } \quad \bullet \bullet \quad \xrightarrow [ ] { W _ { N - 1 } } \overbrace { \overset { \underset { \mathrm { A } } { \mathrm { A } } } { \overset { \mathrm { A } } { \sum } } \frac { W _ { N } } { y } } ^ { \overset { \mathrm { A } } { \sum } } y \cdots t } \\ & { } & { x \xrightarrow [ ] { W _ { 1 } ^ { T } } h _ { 1 _ { R } } \overbrace { \Longleftarrow } ^ { W _ { 2 } ^ { T } } h _ { 2 _ { R } } \bullet \bullet \quad \bullet \overbrace { \overset { w _ { N - 1 } ^ { T } } { \Longleftarrow } } ^ { W _ { N - 1 } ^ { T } } \overbrace { \overset { z } { z } _ { R } \overset { w _ { N } ^ { T } } { \llcorner } } ^ { \overset { W _ { N } ^ { T } } { \longrightarrow } \overset { \mathcal { Y } } } \overset { \cdot } { y ^ { \prime } } } \\ & { } & { z _ { R } ^ { \prime } \xrightarrow [ ] { w _ { N } ^ { T } } } \end{array} +$$ + +Figure 4: Target network architecture; 3 convolution and 2 fully-connected layers were used for MNIST, 5 fully-connected layers were used for permutation-invariant MNIST, and 4 convolution and 3 fully-connected layers were used for CIFAR-10. + +Table 1: Error rate $( \% )$ on the test set using the model with the best performance on the validation set. Numbers on the first row of each sub-table are the number of randomly chosen per-class training examples. The average performance and the standard deviation of three different random-split datasets (except for the case using the entire training set in the last column) are described in this table (error rate on each random set is summarized in Appendix (A3)). Performance of three previous approaches (with gray background; previous-1, 2, 3 are feed-forward model Figure 2(a), joint learning model with recon-one Figure 2(b), joint learning model with recon-all Figure 2(b), respectively) and the proposed methods (proposed-1, 2, 3 are baseline Figure 1(b), random perturbation Figure 1(c), semantic perturbation Figure 1(c), respectively) is summarized. + +
datasetnumber of per-class examples chosen from 5Ok entire MNIST training examplesentire set
MNIST1020501002005001k2k50k
previous-124.55 (3.04)16.00 (1.33)10.35 (0.66)6.58 (0.42)4.71 (0.28)2.94 (0.23)1.90 (0.27)1.45 (0.08)1.04
previous-221.67 (3.19)13.60 (0.99)7.85 (0.10)5.44 (0.37)4.14 (0.08)2.50 (0.15)1.84 (0.07)1.45 (0.07)1.12
previous-320.11 (2.81)13.69 (0.62)9.15 (0.15)6.77 (0.25)5.39 (0.11)3.89 (0.27)2.91 (0.17)2.28 (0.10)1.87
proposed-121.35 (1.16)11.65 (1.15)6.33 (0.10)4.32 (0.31)3.07 (0.11)1.98 (0.11)1.29 (0.09)0.94 (0.02)0.80
proposed-220.17 (1.52)11.68 (0.81)6.24 (0.29)4.12 (0.24)3.04 (0.13)1.88 (0.05)1.24 (0.03)0.96 (0.08)0.65
proposed-320.11 (0.81)10.59 (0.74)5.92 (0.12)3.79 (0.23)2.72 (0.09)1.78 (0.05)1.15 (0.01)0.88 (0.03)0.62
datasetnumber of per-class examples chosen from 4Ok entire CIFAR-10 training examplesentire set
CIFAR-101020501002005001k2k40k
previous-173.82 (1.43)68.99 (0.54)61.30 (0.83)54.93 (0.56)46.97 (0.59)33.69 (0.43)26.63 (0.39)20.97 (0.09)17.80
previous-275.68 (1.56)69.05 (1.13)61.44 (0.63)55.02 (0.34)46.18 (0.51)33.62 (0.38)26.78 (0.48)21.25 (0.40)17.68
previous-373.33 (1.06)67.63 (0.56)62.59 (0.76)56.37 (0.20)50.51 (0.61)41.26 (0.73)32.55 (1.20)26.38 (0.08)22.71
proposed-171.63 (0.69)66.17 (0.40)58.91 (0.86)52.65 (0.28)43.46 (0.30)31.86 (0.54)25.76 (0.31)21.06 (0.18)17.45
proposed-271.69 (0.25)66.75 (0.54)58.95 (0.63)53.01 (0.26)43.71 (0.19)31.80 (0.18)25.50 (0.33)20.81 (0.27)17.43
proposed-371.50 (1.14)66.87 (0.17)58.30 (0.62)52.32 (0.08)42.98 (0.34)30.91 (0.23)24.81 (0.26)20.19 (0.25)16.16
+ +vector for the random perturbation and $y$ is the base vector for the semantic perturbation) is scaled to [0.0, 1.0], and the zero-mean Gaussian noise with 0.2 of standard deviation is added (via elementwise addition) on the normalized base activation. This perturbed scaled activation is de-scaled with the original min and max activations of the base vector. + +Initial learning rates are 0.005 and 0.001 for MNIST and permutation-invariant MNIST, and 0.002 for CIFAR-10, respectively. The learning rates are decayed by a factor of 5 every 40 epochs until the 120-th epoch. For both datasets, the minibatch size is set to 100, and the target objective is optimized using Adam optimizer (Kingma & Ba, 2015) with a momentum 0.9. All the $\lambda$ ’s for reconstruction losses in Eq. (11) and Eq. (12) are 0.03 and 0.01 for MNIST and CIFAR-10, respectively. The same weighting factors for reconstruction losses (0.03 for MNIST and 0.01 for CIFAR-10) are used for $\lambda _ { 1 }$ in Eq (8), and 1.0 is used for $\lambda _ { 2 }$ . + +Input data is first scaled to [0.0, 1.0] and then whitened by the average across all the training examples. In CIFAR-10, random cropping $2 4 \times 2 4$ image is randomly cropped from the original $3 2 \times 3 2$ image) and random horizontal flipping (mirroring) are used for data augmentation. We selected the network that performed best on the validation dataset for evaluation on the test dataset. All the experiments are performed with TensorFlow (Abadi et al., 2015). + +# 4.3 QUANTITATIVE ANALYSIS + +Three previous approaches (a traditional feed-forward model, a joint learning model with the input reconstruction loss, and a joint learning model with reconstruction losses of all the intermediate layers including the input layer) are compared with the proposed methods (the baseline model in Figure 1(b), and the stochastic perturbation model in Figure 1(c) with two different perturbation methods; random and semantic). We measure the classification performance according to varying sizes of training set (examples randomly chosen from the original training dataset). Performance is averaged over three different random trials. + +![](images/fd063f072e83bb12f18e90c57656142a6c76f9eaf9710d6764d7b195a4663837.jpg) +Figure 5: Examples reconstructed from the perturbed latent vectors via (a) random perturbation, and (b) semantic perturbation (top row shows the original training examples). More examples are summarized in Appendix (A4.1). + +Table 1 summarizes the classification performance for MNIST and CIFAR-10. As we expected, the base model obtained by maximizing the sum of mutual informations (proposed-base) mostly performs better than previous approaches, and the model with the semantic perturbation (proposedperturb (semantic)) performs best among all the comparison targets. Especially in MNIST, the error rate of ‘proposed-perturb (semantic)’ with $2 \mathrm { k }$ per-class training examples is less than the error rate of all types of previous works with the entire training set (approximately 5k per-class examples). + +We further verify the proposed method on the permutation-invariant MNIST task with a standard feed-forward neural network. Classification performance is measured against three different sizes of training set (1k, 2k, and 5k per-class training examples). ‘Proposed-perturb (semantic)’ achieves the best performance among all the configurations; $2 . 5 7 \%$ , $1 . 8 2 \%$ , and $1 . 2 8 \%$ error rates for 1k, 2k, and 5k per-class training examples, respectively. The joint learning model with the input reconstruction loss performs best among three previous approaches; $2 . 7 2 \%$ , $1 . 9 7 \%$ , and $1 . 3 8 \%$ error rates for 1k, 2k, and 5k per-class training examples, respectively. + +# 4.4 QUALITATIVE ANALYSIS + +As mentioned before, random perturbation by adding unstructured noise directly to the latent representation cannot guarantee preserving the semantics of the original representation. We compared two different perturbation methods (random and semantic) by visualizing the examples reconstructed from the perturbed latent vectors (Figure 5). Top row is the original examples selected from training set (among 2k per-class training examples), and the rest are the reconstructions of their perturbed latent representations. Based on the architecture described in Figure 1(b), we generated five different perturbed latent representations according to the type of perturbation, and reconstructed the perturbed latent vectors through decoding path for reconstruction. + +Figure 5(a) and (b) show the examples reconstructed from the random and semantic perturbations, respectively. For both cases, zero-mean Gaussian random noise (0.2 standard deviation) is used for perturbation. As shown in Figure 5(a), random perturbation partially destroys the original semantics; for example, semantics of $\cdot _ { 1 } \cdot$ is mostly destroyed under random perturbation, and some examples of $\cdot _ { 3 } \cdot$ are reconstructed as being similar to $\cdot _ { 8 } \cdot $ rather than its original content $\cdot _ { 3 } \cdot$ . Figure 5(b) shows the examples reconstructed from the semantic perturbation. The reconstructed examples show subtle semantic variations while preserving the original semantic contents; for example, thickness difference in $\cdot _ { 3 } \cdot$ (example on the third row) or writing style difference in $\cdot _ { 8 } \cdot $ (openness of the top left corner). + +Figure 6 shows the overall effect of the perturbation. In this analysis, 100 per-class MNIST examples are used for training. From the trained model based on the architecture described in Figure 1(b), latent representations $z$ of all the $5 0 \mathrm { k }$ examples (among 50k examples, only 1k examples were used for training) are visualized by using t-SNE (Maaten & Hinton, 2008). Only the training examples of three classes (0, 1, and 9) among ten classes are depicted as black circles for visual discrimination in + +![](images/762dd002da7354f6b01a3b543c08def476bbdd57561f3a870f21d90555f0e26a.jpg) +Figure 6: Training examples (circles or crosses with colors described below) over the examples not used for training (depicted as background with different colors); (a) training examples (black circles), (b) training examples (yellow circles) with $3 \times$ random-perturbed samples (blue crosses), and (c) training examples (yellow circles) with $3 \times$ semantic-perturbed samples (blue crosses). Best viewed in color. + +Figure 6(a). The rest of the examples which were not used for training (approximately $4 . 9 \mathrm { k }$ examples per class) are depicted as a background with different colors. We treat the colored background examples (not used for training) as a true distribution of unseen data in order to estimate the generalization level of learned representation according to the type of perturbation. Figure 6(b) and (c) show the training examples (100 examples per class with yellow circles) and their perturbed ones $3 \times$ sampled from each example with blue crosses) through random and semantic perturbations, respectively. + +In Figure 6(b), perturbed samples are distributed near the original training examples, but some samples outside the true distribution cannot be identified easily with appropriate classes. This can be explained with Figure 5(a), since some perturbed samples are ambiguous semantically. In Figure 6(c), however, most of the perturbed samples evenly cover the true distribution. As mentioned before, stochastic perturbation with the semantic additive noise during training implicitly incurs the effect of augmentation on the latent space while resulting in better generalization. Per-class t-SNE results are summarized in Appendix (A4.2). + +# 5 DISCUSSION + +We introduced a novel latent space modeling method for supervised tasks based on the standard feed-forward neural network architecture. The presented model simultaneously optimizes both supervised and unsupervised losses based on the assumption that the better latent representation can be obtained by maximizing the sum of hierarchical mutual informations. Especially the stochastic perturbation process which is achieved by modeling the semantic additive noise during training enhances the representational power of the latent space. From the proposed semantic noise modeling process, we can expect improvement of generalization performance in supervised learning with implicit semantic augmentation effect on the latent space. + +The presented model architecture can be intuitively extended to semi-supervised learning because it is implemented as the joint optimization of supervised and unsupervised objectives. For semisupervised learning, however, logical link between features learned from labelled and unlabelled data needs to be considered additionally. We leave the extension of the presented approach to semisupervised learning for the future. + +# REFERENCES + +Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´ Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´ berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations (ICLR), 2015. + +Yoshua Bengio, Pascal Lamblin, Dan Popovici, Hugo Larochelle, et al. Greedy layer-wise training of deep networks. In Advances in Neural Information Processing Systems (NIPS), 2007. + +Kyunghyun Cho and Xi Chen. Classifying and visualizing motion capture sequences using deep neural networks. In International Conference on Computer Vision Theory and Applications, 2014. + +Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In International Conference on Machine Learning (ICML), 2008. + +Ian Goodfellow, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Multi-prediction deep boltzmann machines. In Advances in Neural Information Processing Systems (NIPS), 2013. + +Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep recurrent neural networks. In International conference on acoustics, speech and signal processing, 2013. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. + +Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. Signal Processing Magazine, IEEE, 29(6):82–97, 2012. + +Geoffrey E. Hinton, Simon Osindero, and Yee Whye Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18:1527–1554, 2006. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NIPS), 2012. + +Hugo Larochelle and Yoshua Bengio. Classification using discriminative restricted boltzmann machines. In International Conference on Machine Learning (ICML), 2008a. + +Hugo Larochelle and Yoshua Bengio. Classification using discriminative restricted boltzmann machines. In International Conference on Machine Learning (ICML), 2008b. + +Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. + +Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research (JMLR), 9(Nov):2579–2605, 2008. + +Jonathan Masci, Ueli Meier, Dan Cires¸an, and Jurgen Schmidhuber. Stacked convolutional auto- ¨ encoders for hierarchical feature extraction. In International Conference on Artificial Neural Networks, 2011. + +Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semisupervised learning with ladder networks. In Advances in Neural Information Processing Systems (NIPS), 2015. + +Ruslan Salakhutdinov and Geoffrey E Hinton. Deep boltzmann machines. In Artificial Intelligence and Statistics Conference (AISTATS), 2009. + +Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. + +Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research (JMLR), 11:3371–3408, 2010. + +Satosi Watanabe. Information theoretical analysis of multivariate correlation. IBM Journal of research and development, 4(1):66–82, 1960. + +Yuting Zhang, Kibok Lee, and Honglak Lee. Augmenting supervised neural networks with unsupervised objectives for large-scale image classification. In International Conference on Machine Learning (ICML), 2016. + +Junbo Zhao, Michael Mathieu, Ross Goroshin, and Yann Lecun. Stacked what-where auto-encoders. In International Conference on Learning Representations (ICLR), 2015. + +# APPENDIX + +# (A1) DERIVATION OF RECONSTRUCTION ERRORS FROM CONDITIONAL ENTROPY TERMS + +Extended from Section 2. From the lower bound in Eq. (3), we consider the following optimization problem (refer to ‘Section 2. From mutual information to autoencoders’ in (Vincent et al., 2010)): + +$$ +\operatorname* { m a x } _ { \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } , \theta _ { 2 } ^ { \prime } \} } \mathbb { E } _ { q ( X , Z , Y ) } \left[ \log q ( X | Z ) \right] + \mathbb { E } _ { q ( X , Z , Y ) } \left[ \log q ( Z | Y ) \right] . +$$ + +Here, we denote $q ( X , Z , Y )$ an unknown joint distribution. Note that $Z$ and $Y$ are respectively the variables transformed from parametric mappings $Z = f _ { \boldsymbol { \theta } _ { 1 } } ( { \boldsymbol { X } } )$ and $Y = f _ { \theta _ { 2 } } ( Z )$ (see Fig. 1). $q ( X , Z , Y )$ then can be reduced to $q ( X )$ from $q ( Z | X ; \theta _ { 1 } ) = \delta ( Z - f _ { \theta _ { 1 } } ( X ) )$ and $\overset { \cdot } { q } ( Y | Z ; \theta _ { 2 } ) =$ $\delta ( Y - f _ { \theta _ { 2 } } ( Z ) )$ where $\delta$ denotes Dirac-delta function. + +From the Kullback-Leibler divergence that $D _ { \mathrm { K L } } ( q | | p ) \geq 0$ for any two distributions $p$ and $q$ , the optimization in Eq. (13) corresponds to the following optimization problem where $p ( \cdot )$ denotes a parametric distribution: + +$$ +\operatorname* { m a x } _ { \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } , \theta _ { 2 } ^ { \prime } \} } \mathbb { E } _ { q ( X ) } \left[ \log p ( X | Z ; \theta _ { 1 } ^ { \prime } ) \right] + \mathbb { E } _ { q ( X ) } \left[ \log p ( Z | Y ; \theta _ { 2 } ^ { \prime } ) \right] . +$$ + +By replacing $q ( X )$ with a sample distribution $q ^ { 0 } ( X )$ and putting all parametric dependencies between $X , Z$ and $Y$ , we will have + +$$ +\operatorname* { m a x } _ { \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } , \theta _ { 2 } ^ { \prime } \} } \mathbb { E } _ { q ^ { 0 } ( X ) } \left[ \log p ( X | Z = f _ { \theta _ { 1 } } ( X ) ; \theta _ { 1 } ^ { \prime } ) \right] + \mathbb { E } _ { q ^ { 0 } ( X ) } \left[ \log p ( Z | Y = f _ { \theta _ { 2 } } ( f _ { \theta _ { 1 } } ( X ) ) ; \theta _ { 2 } ^ { \prime } ) \right] ~ . +$$ + +For a given input sample $x$ of $X$ , it is general to interpret $x _ { R }$ and $z _ { R }$ as the parameters of distributions $p ( X | X _ { R } = x _ { R } )$ and $p ( Z | Z _ { R } = z _ { R } )$ which reconstruct $x$ and $z$ with high probability (i.e. $x _ { R }$ and $z _ { R }$ are not exact reconstructions of $x$ and $z$ ). Since $x _ { R }$ and $z _ { R }$ are real-valued, we assume Gaussian distribution for these conditional distributions, that is, + +$$ +\begin{array} { r } { p ( { \cal X } | { \cal X } _ { R } = x _ { R } ) = N ( x _ { R } , \sigma _ { 0 } ^ { 2 } { \bf I } ) } \\ { p ( { \cal Z } | { \cal Z } _ { R } = z _ { R } ) = N ( z _ { R } , \sigma _ { 0 } ^ { 2 } { \bf I } ) . } \end{array} +$$ + +The assumptions yield $- \log p ( \cdot | \cdot ) \propto L _ { L 2 } ( \cdot , \cdot )$ . + +With the following relations for log terms in Eq. (15), + +$$ +\begin{array} { c } { { p ( X | Z = f _ { \theta _ { 1 } } ( x ) ; \theta _ { 1 } ^ { \prime } ) = p ( X | X _ { R } = g _ { \theta _ { 1 } ^ { \prime } } ( f _ { \theta _ { 1 } } ( x ) ) ) } } \\ { { p ( Z | Y = f _ { \theta _ { 2 } } ( f _ { \theta _ { 1 } } ( x ) ) ; \theta _ { 2 } ^ { \prime } ) = p ( Z | Z _ { R } = g _ { \theta _ { 2 } ^ { \prime } } ( f _ { \theta _ { 2 } } ( f _ { \theta _ { 1 } } ( x ) ) ) , } } \end{array} +$$ + +the optimization problem in Eq. (15) corresponds to the minimization problem of reconstruction errors for input examples $x ^ { ( i ) }$ as below: + +$$ +\operatorname * { m i n } _ { \{ \theta _ { 1 } , \theta _ { 1 } ^ { \prime } , \theta _ { 2 } , \theta _ { 2 } ^ { \prime } \} } \sum _ { i } L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \ : . +$$ + +# (A2) LADDER NETWORK, A REPRESENTATIVE SEMI-SUPERVISED LEARNING MODEL + +Extended from Section 3. We performed experiments with a ladder network model (Rasmus et al., 2015) in order to estimate the performance on pure supervised tasks according to different sizes of training set. We used the code (https://github.com/rinuboney/ladder.git) for this experiment. The network architecture implemented on the source code is used as is; (784-1000-500-250-250-250- 10). Based on the same network architecture, we implemented the proposed stochastic perturbation model described in Figure 1(c) and compared the classification performance with the ladder network as described in Table 2 (we did not focus on searching the optimal hyperparameters for the proposed model in this experiment). As summarized in the bottom of the table (mean over 3 random trials), the proposed semantic noise modeling method shows a fairly large performance gain compared to the ladder network model with small-scale datasets (e.g., in a case of 10 per-class training examples, the proposed method achieves $2 2 . 1 1 \%$ of error rate, while the ladder network shows $2 9 . 6 6 \%$ ). + +Table 2: Classification performance (error rate in $\%$ ) of the ladder network and the proposed model on three different sets of randomly chosen training examples (MNIST). + +
set No.1 (# training examples per class)1020501002005001k2k (all) 5k
ladder network model; Figure 325.8516.489.266.004.663.072.151.260.91
proposed-perturb (semantic); Figure 1(c)19.7612.338.776.064.592.931.871.310.93
set No.2 (# training examples per class)1020501002005001k2k
ladder network model; Figure 333.1417.4610.446.674.432.821.941.37
proposed-perturb (semantic); Figure 1(c)23.3615.359.435.754.432.991.871.39
set No.3 (# training examples per class)1020501002005001k2k
ladder network model; Figure 329.9916.999.737.344.393.002.121.47
proposed-perturb (semantic); Figure 1(c)23.2113.988.836.514.322.942.221.49
mean over 3 random trials1020501002005001k2k (all) 5k
ladder network model; Figure 329.6616.989.816.674.492.962.071.370.91
proposed-perturb (semantic); Figure 1(c)22.1113.899.016.114.452.951.991.400.93
+ +# (A3) QUANTITATIVE ANALYSIS + +Extended from Section 4.3. Among the total $5 0 \mathrm { k }$ and $4 0 \mathrm { k }$ training examples in MNIST and CIFAR10, we randomly select the examples for training. Classification performance according to three different randomly chosen training sets are summarized in Table 3 (MNIST) and Table 4 (CIFAR10). Further experiments with denoising constraints are also included. Zero-mean Gaussian random noise with 0.1 standard deviation is used for noise injection. Denoising function helps to achieve slightly better performance on MNIST, but it results in performance degradation on CIFAR-10 (we did not focus on searching the optimal parameters for noise injection in this experiments). + +Table 3: Classification performance (error rate in $\%$ ) on three different sets of randomly chosen training examples (MNIST). + +
Set No.1(#train examples per class)
feed-forward model; Figure 2(a)10 22.6120 14.2050 11.25100 6.37200 4.34500 2.631k 1.832k 1.56(all) 5k 1.04
joint learning model with recon-one; Figure 2(b)18.6912.217.845.174.022.581.791.471.12
joint learning model with recon-one with denoising constraints20.3911.917.414.643.652.571.971.530.97
joint learning model with recon-all; Figure 2(b)18.8212.829.346.435.234.122.682.421.87
joint learning model with recon-all with denoisingconstraints17.9311.767.324.783.913.042.521.991.36
proposed-base; Figure 1(b)20.2310.186.473.893.041.891.330.910.80
proposed-base with denoising constraints19.8810.896.624.263.402.442.111.541.13
proposed-perturb (random); Figure 1(c)18.3810.586.643.783.141.901.210.890.65
proposed-perturb (semantic); Figure 1(c)19.339.725.983.472.841.841.160.840.62
Set No.2 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)28.8417.3610.146.204.783.021.611.41
joint learning model with recon-one; Figure 2(b)26.0914.407.985.184.172.291.941.52
joint learning model with recon-one with denoising constraints27.6913.116.955.073.542.371.831.28
joint learning model with recon-all; Figure 2(b)24.0114.138.986.845.443.512.982.18
joint learning model with recon-all with denoising constraints23.0513.297.795.123.923.012.271.84
proposed-base; Figure 1(b)22.9512.986.274.433.222.141.370.96
proposed-base with denoising constraints26.9612.216.454.623.132.531.881.49
proposed-perturb (random); Figure 1(c)22.1012.525.974.262.861.941.230.92
proposed-perturb (semantic); Figure 1(c)21.2211.525.753.912.611.731.140.89
Set No.3 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)22.2016.439.677.165.023.172.251.39
joint learning model with recon-one; Figure 2(b)20.2314.197.735.964.222.621.791.35
joint learning model with recon-one with denoising constraints19.3212.257.445.393.582.371.491.56
joint learning model with recon-al; Figure 2(b)17.5114.129.127.045.494.053.082.25
joint learning model with recon-all with denoising constraints17.0712.507.865.484.052.972.021.98
proposed-base; Figure 1(b)20.8611.796.254.632.961.911.160.96
proposed-base with denoising constraints19.8911.306.264.573.502.631.611.47
proposed-perturb (random); Figure 1(c)20.0211.946.124.323.131.811.281.08
proposed-perturb (semantic); Figure 1(c)19.7810.536.034.002.701.761.140.92
+ +Table 4: Classification performance (error rate in $\%$ ) on three different sets of randomly chosen training examples (CIFAR-10). + +
Set No.1(# train examples per class)1020501002005001k2k(all) 4k
feed-forward model; Figure 2(a)73.3069.2562.4255.6547.7134.3027.0421.0617.80
joint learning model with recon-one; Figure 2(b)75.1970.3862.2555.3046.8934.1226.6321.0517.68
joint learning model with recon-one with denoising constraints73.7268.2061.9955.2346.6436.3729.7825.5321.73
joint learning model with recon-all; Figure 2(b)74.7968.3362.9256.2451.3740.3030.9126.4922.71
joint learning model with recon-all with denoising constraints76.5669.6764.5357.8852.7442.2436.9030.9327.41
proposed-base; Figure 1(b)70.7966.5759.9152.9843.2932.2526.1920.9217.45
proposed-base with denoising constraints71.0367.4960.3753.5244.2833.4028.0025.0621.34
proposed-perturb (random); Figure 1(c)71.8967.1259.2252.7943.8731.8225.0420.9717.43
proposed-perturb (semantic); Figure 1(c)71.5966.9058.6452.3442.7430.9424.4520.1016.16
Set No.2 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)72.3969.4960.4554.8546.9133.3926.7321.00
joint learning model with recon-one; Figure 2(b)74.0669.1460.7154.5445.7033.5427.4320.90
joint learning model with recon-one with denoising constraints76.4069.3360.2855.3847.4036.2929.3124.60
joint learning model with recon-all; Figure 2(b)72.2867.6061.5356.6549.9942.0832.9926.33
joint learning model with recon-all with denoising constraints73.9069.2361.9057.9952.3545.1237.2330.14
proposed-base; Figure 1(b)72.4965.6257.8252.6643.2032.2425.6021.32
proposed-base with denoising constraints72.9966.7557.7853.8144.3333.5628.4025.03
proposed-perturb (random); Figure 1(c)71.8465.9858.0853.3743.4431.5625.6921.03
proposed-perturb (semantic); Figure 1(c)72.8566.6557.4452.2142.7431.1724.9920.54
Set No.3(# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)75.7868.2461.0254.2946.2833.3826.1120.85
joint learning model with recon-one; Figure 2(b)77.7967.6261.3755.2245.9633.2126.2921.81
joint learning model with recon-one with denoising constraints76.6069.2761.1355.1047.5037.1229.6324.88
joint learning model with recon-all; Figure 2(b)72.9266.9763.3156.2350.1641.4133.7526.31
joint learning model with recon-all with denoising constraints76.8368.5365.5858.2952.4345.4239.0132.32
proposed-base; Figure 1(b)71.6066.3158.9952.3043.8831.1025.4820.95
proposed-base with denoising constraints72.3967.2060.6052.6444.6233.5228.0125.25
proposed-perturb (random); Figure 1(c)71.3467.1559.5552.8643.8132.0125.7820.42
proposed-perturb (semantic); Figure 1(c)70.0667.0758.8352.4143.4730.6125.0019.94
+ +# (A4.1) QUALITATIVE ANALYSIS + +Extended from Section 4.4. Figure 7 shows reconstructed examples from perturbed (random or semantic) latent representations (refer to Figure 5 and the analysis described in Section 4.4). + +![](images/f6a972dc25f50eb96aab8b6b183f045da818ad62acb9bbdc0a1fa7537bee6201.jpg) +Figure 7: For each example, top row is the original examples selected from the training set, and the rest are reconstructed from the perturbed representations via random (left) and semantic (right) perturbations. + +# (A4.2) QUALITATIVE ANALYSIS + +Extended from Section 4.4. Figure 8 shows the t-SNE results per class on MNIST. The overall tendency is similar to the description in Section 4.4. + +![](images/67539ce30ef40d4cff811931a9fdada3fcb6273cc65b7d415f958e58e94aa3fd.jpg) + +![](images/571139d772703de3df5b815c8ae20f2cb437f0dbde8fae9377558d9641fe2d1c.jpg) +Figure 8: From top to bottom: 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. From left to right: training examples (circle), training examples (circle) $^ +$ random-perturbed samples (cross), and training examples (circle) $^ +$ semantic-perturbed samples (cross). Best viewed in color. \ No newline at end of file diff --git a/parse/train/SyCSsUDee/SyCSsUDee_content_list.json b/parse/train/SyCSsUDee/SyCSsUDee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..041775934911d8b18de06fc94dc07d9c1f6b4040 --- /dev/null +++ b/parse/train/SyCSsUDee/SyCSsUDee_content_list.json @@ -0,0 +1,1622 @@ +[ + { + "type": "text", + "text": "SEMANTIC NOISE MODELING FORBETTER REPRESENTATION LEARNING", + "text_level": 1, + "bbox": [ + 174, + 99, + 632, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyo-Eun $\\mathbf { K i m } ^ { * }$ and Sangheum Hwang ", + "bbox": [ + 184, + 170, + 449, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lunit Inc. \nSeoul, South Korea \n{hekim, shwang}@lunit.io \nKyunghyun Cho \nCourant Institute of Mathematical Sciences and Centre for Data Science \nNew York University \nNew York, NY 10012, USA \nkyunghyun.cho@nyu.edu ", + "bbox": [ + 184, + 185, + 418, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 184, + 247, + 656, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 354, + 544, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Latent representation learned from multi-layered neural networks via hierarchical feature abstraction enables recent success of deep learning. Under the deep learning framework, generalization performance highly depends on the learned latent representation. In this work, we propose a novel latent space modeling method to learn better latent representation. We designed a neural network model based on the assumption that good base representation for supervised tasks can be attained by maximizing the sum of hierarchical mutual informations between the input, latent, and output variables. From this base model, we introduce a semantic noise modeling method which enables semantic perturbation on the latent space to enhance the representational power of learned latent feature. During training, latent vector representation can be stochastically perturbed by a modeled additive noise while preserving its original semantics. It implicitly brings the effect of semantic augmentation on the latent space. The proposed model can be easily learned by back-propagation with common gradient-based optimization algorithms. Experimental results show that the proposed method helps to achieve performance benefits against various previous approaches. We also provide the empirical analyses for the proposed latent space modeling method including t-SNE visualization. ", + "bbox": [ + 233, + 383, + 764, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 643, + 336, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enhancing the generalization performance against unseen data given some sample data is the main objective in machine learning. Under that point of view, deep learning has been achieved many breakthroughs in several domains such as computer vision (Krizhevsky et al., 2012; Simonyan & Zisserman, 2015; He et al., 2016), natural language processing (Collobert & Weston, 2008; Bahdanau et al., 2015), and speech recognition (Hinton et al., 2012; Graves et al., 2013). Deep learning is basically realized on deep layered neural network architecture, and it learns appropriate taskspecific latent representation based on given training data. Better latent representation learned from training data results in better generalization over the future unseen data. Representation learning or latent space modeling becomes one of the key research topics in deep learning. During the past decade, researchers focused on unsupervised representation learning and achieved several remarkable landmarks on deep learning history (Vincent et al., 2010; Hinton et al., 2006; Salakhutdinov & Hinton, 2009). In terms of utilizing good base features for supervised learning, the base representation learned from unsupervised learning can be a good solution for supervised tasks (Bengio et al., 2007; Masci et al., 2011). ", + "bbox": [ + 174, + 674, + 825, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The definition of ‘good’ representation is, however, different according to target tasks. In unsupervised learning, a model is learned from unlabelled examples. Its main objective is to build a model to estimate true data distribution given examples available for training, so the learned latent representation normally includes broadly-informative components of the raw input data (e.g., mutual information between the input and the latent variable can be maximized for this objective). In supervised learning, however, a model is learned from labelled examples. In the case of classification, a supervised model learns to discriminate input data in terms of the target task using corresponding labels. Latent representation is therefore obtained to maximize the performance on the target supervised tasks. ", + "bbox": [ + 176, + 875, + 821, + 904 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the meaning of good representations vary according to target tasks (unsupervised or supervised), pre-trained features from the unsupervised model are not be guaranteed to be useful for subsequent supervised tasks. Instead of the two stage learning strategy (unsupervised pre-training followed by supervised fine-tuning), several works focused on a joint learning model which optimizes unsupervised and supervised objectives concurrently, resulting in better generalization performance (Goodfellow et al., 2013; Larochelle & Bengio, 2008a; Rasmus et al., 2015; Zhao et al., 2015; Zhang et al., 2016; Cho & Chen, 2014). ", + "bbox": [ + 174, + 208, + 825, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel latent space modeling method for supervised learning as an extension of the joint learning approach. We define a good latent representation of standard feed-forward neural networks under the basis of information theory. Then, we introduce a semantic noise modeling method in order to enhance the generalization performance. The proposed method stochastically perturbs the latent representation of a training example by injecting a modeled semantic additive noise. Since the additive noise is randomly sampled from a pre-defined probability distribution every training iteration, different latent vectors from a single training example can be fully utilized during training. The multiple different latent vectors produced from a single training example are semantically similar under the proposed latent space modeling method, so we can expect semantic augmentation effect on the latent space. ", + "bbox": [ + 174, + 311, + 825, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments are performed on two datasets; MNIST and CIFAR-10. The proposed model results in better classification performance compared to previous approaches through notable generalization effect (stochastically perturbed training examples well cover the distribution of unseen data). ", + "bbox": [ + 176, + 458, + 825, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 METHODOLOGY ", + "text_level": 1, + "bbox": [ + 176, + 521, + 341, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method starts from the existing joint learning viewpoint. This section first explains the process of obtaining a good base representation for supervised learning which is the basis of the proposed latent space modeling method. And then, we will describe how the proposed semantic noise modeling method perturbs the latent space while maintaining the original semantics. ", + "bbox": [ + 174, + 553, + 825, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 BASE JOINT LEARNING MODEL ", + "text_level": 1, + "bbox": [ + 176, + 627, + 428, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a traditional feed-forward neural network model (Figure 1(a)), output $Y$ of input data $X$ is compared with its true label, and the error is propagated backward from top to bottom, which implicitly learns a task-specific latent representation $Z$ of the input $X$ . As an extension of a joint learning approach, an objective to be optimized can be described in general as below (Larochelle & Bengio, 2008b): ", + "bbox": [ + 174, + 652, + 825, + 722 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/a16811e86fe508070126d387fac68490c3aec3d0094ff45957654dc74cb2c4ce.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta } \\lambda L _ { u n s u p } + L _ { s u p }\n$$", + "text_format": "latex", + "bbox": [ + 426, + 743, + 570, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $L _ { u n s u p }$ and $L _ { s u p }$ are respectively an unsupervised loss and a supervised loss, and $\\theta$ and $\\lambda$ are model parameters to be optimized during training and a loss weighting coefficient, respectively. In terms of modeling $L _ { u n s u p }$ in Eq. (1), we assume that good latent representation $Z$ is attained by maximizing the sum of hierarchical mutual informations between the input, latent, and output variables; i.e. the sum of the mutual information between the input $X$ and the $Z$ and the mutual information between the $Z$ and the output $Y$ . Each mutual information is decomposed into an entropy and a conditional entropy terms, so the sum of hierarchical mutual informations is expressed as follows: ", + "bbox": [ + 174, + 775, + 825, + 887 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/343bc965dc2f1b3bc27227b35052f1ea18ea5dc951dc11b4efe699b436df2aee.jpg", + "text": "$$\n{ \\bf { I } } ( X ; Z ) + { \\bf { I } } ( Z ; Y ) = { \\bf { H } } ( X ) - { \\bf { H } } ( X \\vert Z ) + { \\bf { H } } ( Z ) - { \\bf { H } } ( Z \\vert Y )\n$$", + "text_format": "latex", + "bbox": [ + 297, + 907, + 699, + 925 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1137fde88fa95773239a9891f73a9430060b5253fa34e7cd20bedd7afffb1039.jpg", + "image_caption": [ + "Figure 1: (a) Standard feed-forward neural network model, (b) feed-forward neural network model with reconstruction paths, and (c) feed-forward neural network model with reconstruction and stochastic perturbation paths. " + ], + "image_footnote": [], + "bbox": [ + 264, + 103, + 732, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf { I } ( \\cdot ; \\cdot )$ is the mutual information between random variables, and $\\mathbf { H } ( \\cdot )$ and $\\mathbf { H } ( \\cdot | \\cdot )$ are the entropy and the conditional entropy of random variables, respectively. Note that the sum of those mutual informations becomes equivalent to the total correlation of $X , Z$ , and $Y$ under the graphical structure of the general feed-forward model described in Figure 1(a); $P ( X , Z , Y ) = P ( Y | \\bar { Z } ) \\bar { P } ( Z | X ) P ( X )$ . The total correlation is equal to the sum of all pairwise mutual informations (Watanabe, 1960). ", + "bbox": [ + 173, + 337, + 825, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to find the model parameters which maximize $\\mathbf { I } ( X ; Z ) + \\mathbf { I } ( Z ; Y )$ . Since $\\mathbf { H } ( X )$ and $\\mathbf { H } ( Z )$ are non-negative, and $\\mathbf { H } ( X )$ is constant in this case, the lower bound on $\\mathbf { I } ( X ; Z ) + \\mathbf { I } ( Z ; Y )$ can be reduced to 1: ", + "bbox": [ + 173, + 415, + 823, + 457 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/1ac02bd2831ba39964173e56ce8060093eb5755a8e44b5fd4efe3ce0d1f8c8a7.jpg", + "text": "$$\n\\mathbf { I } ( X ; Z ) + \\mathbf { I } ( Z ; Y ) \\geq - \\mathbf { H } ( X | Z ) - \\mathbf { H } ( Z | Y ) .\n$$", + "text_format": "latex", + "bbox": [ + 346, + 486, + 651, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is known that maximizing $- \\mathbf { H } ( X | Z )$ can be formulated as minimizing the reconstruction error between the input $x ^ { ( i ) }$ ( $i$ -th example sampled from $X$ ) and its reconstruction $x _ { R } ^ { ( i ) }$ under the general audo-encoder framework (Vincent et al., 2010). Since $\\mathbf { H } ( X | Z ) + \\mathbf { H } ( Z | Y )$ is proportional to the sum of reconstruction errors of $x ^ { ( i ) }$ (with its reconstruction $x _ { R } ^ { ( i ) }$ ) and $z ^ { ( i ) }$ (with its reconstruction $z _ { R } ^ { ( i ) } )$ ), the target objective can be expressed as follows (refer to Appendix (A1) for the details ofhematical derivations): ", + "bbox": [ + 173, + 520, + 826, + 618 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/edc93073d657780891af347fa72a6a4984ccea23c718a6d317b93732a6797771.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta } \\sum _ { i } L _ { r e c } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { r e c } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } )\n$$", + "text_format": "latex", + "bbox": [ + 357, + 642, + 638, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L _ { r e c }$ is a reconstruction loss. ", + "bbox": [ + 176, + 695, + 406, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1(b) shows the target model obtained from the assumption that good latent representation $Z$ can be obtained by maximizing the sum of hierarchical mutual informations. Given an input sample $x$ , feed-forward vectors and their reconstructions are attained deterministically by: ", + "bbox": [ + 171, + 715, + 825, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/18c3c0fc8b371c95119a0fc8bf0de1ef9af09d7f99df5f8ad3941022104c55d6.jpg", + "text": "$$\n\\begin{array} { r l } & { z = f _ { \\theta _ { 1 } } ( x ) } \\\\ & { y = f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) } \\\\ & { x _ { R } = g _ { \\theta _ { 1 } ^ { \\prime } } ( z ) = g _ { \\theta _ { 1 } ^ { \\prime } } ( f _ { \\theta _ { 1 } } ( x ) ) } \\\\ & { z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y ) = g _ { \\theta _ { 2 } ^ { \\prime } } ( f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 387, + 789, + 609, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of training pairs $( x ^ { ( i ) } , t ^ { ( i ) } )$ where $x ^ { ( i ) }$ and $t ^ { ( i ) }$ are the $i$ -th input example and its label, target objective in Eq. (1) under the model described in Figure 1(b) can be organized as below (with real-valued input samples, L2 loss $L _ { L 2 }$ is a proper choice for the reconstruction loss $L _ { r e c , }$ ): ", + "bbox": [ + 173, + 102, + 825, + 147 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/fbcaafc85f59bb7fa6d96e6d5bae5e54abfb8b45a4b8f54e72970c90370d0bfc.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\sum _ { i } \\lambda \\left( L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \\right) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } )\n$$", + "text_format": "latex", + "bbox": [ + 254, + 167, + 743, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L _ { N L L }$ is a negative log-likelihood loss for the target supervised task. Note that Eq. (6) represents the ‘proposed-base’ in our experiment (see Section 4.3). ", + "bbox": [ + 173, + 218, + 825, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 SEMANTIC NOISE MODELING ", + "text_level": 1, + "bbox": [ + 176, + 268, + 416, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the architecture shown in Figure 1(b) with the target objective in Eq. (6), we conjecture that stochastic perturbation on the latent space during training helps to achieve better generalization performance for supervised tasks. Figure 1(c) shows this strategy which integrates the stochastic perturbation process during training. Suppose that $Z _ { P }$ is a perturbed version of $Z$ , and $Y _ { P }$ is an output which is feed-forwarded from $Z _ { P }$ . Given a latent vector $z = f _ { \\boldsymbol { \\theta } _ { 1 } } ( x )$ from an input sample $x$ , ", + "bbox": [ + 173, + 296, + 825, + 367 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f2301060da4cea73931af1b13b0840d02949916e1fc674b6cb7a8f2e6103dff6.jpg", + "text": "$$\nz ^ { \\prime } = z + z _ { e } { \\mathrm { ~ } } \\mathrm { a n d } { \\hat { y } } = f _ { \\theta _ { 2 } } ( z ^ { \\prime } )\n$$", + "text_format": "latex", + "bbox": [ + 398, + 391, + 598, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z ^ { \\prime }$ and $\\hat { y }$ are a perturbed latent vector and its output respectively, and $z _ { e }$ is an additive noise used in the perturbation process of $z$ . Based on the architecture shown in Figure 1(c), target objective can be modified as: ", + "bbox": [ + 173, + 425, + 825, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f81f4fd965a33778677b52a0d993b6fa81a221c156ffdecc13721223ff59540c.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\substack { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } } \\sum _ { i } \\lambda _ { 1 } \\left( L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \\right) + \\lambda _ { 2 } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) + L _ { N L L } ( \\hat { y } ^ { ( i ) } , t ^ { ( i ) } ) .\n$$", + "text_format": "latex", + "bbox": [ + 181, + 497, + 823, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using random additive noise directly on $z _ { e }$ is the most intuitive approach (‘proposed-perturb (random)’ in Section 4.3). However, preserving the semantics of the original latent representation $z$ cannot be guaranteed under the direct random perturbation on the latent space. While the latent space is not directly interpretable in general, the output logit $y$ of the latent representation $z$ is interpretable, because the output logit is tightly coupled to the prediction of the target label. In order to preserve the semantics of the original latent representation after perturbation, we indirectly model a semantic noise on the latent space by adding small random noise directly on the output space. ", + "bbox": [ + 173, + 549, + 825, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the output (pre-softmax) logit $y$ , the semantic-preserving variation of $y$ (i.e. $y ^ { \\prime }$ ) can be modeled by $y ^ { \\prime } = y + y _ { e }$ , where $y _ { e }$ is a random noise vector stochastically sampled from a zeromean Gaussian with small standard deviation $\\sigma ; \\mathcal { N } ( \\mathbf { 0 } , \\sigma ^ { 2 } \\mathbf { I } )$ . Now, the semantic perturbation $z ^ { \\prime }$ can be reconstructed from the random perturbation $y ^ { \\prime }$ through the decoding path $g _ { \\theta _ { 2 } ^ { \\prime } }$ in Figure 1(c). From the original output logit $y$ and the randomly perturbed output logit $y ^ { \\prime }$ , semantic additive noise $z _ { e }$ on the latent space can be approximately modeled as below: ", + "bbox": [ + 173, + 655, + 825, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/4999b6772dd947fa03a33aff78d2e938f1f5434065bdde691ee9cfe8ae75e6f4.jpg", + "text": "$$\n\\begin{array} { r l } & { z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y ) } \\\\ & { z _ { R } ^ { \\prime } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y ^ { \\prime } ) = g _ { \\theta _ { 2 } ^ { \\prime } } ( y + y _ { e } ) } \\\\ & { z _ { e } \\simeq z _ { R } ^ { \\prime } - z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y + y _ { e } ) - g _ { \\theta _ { 2 } ^ { \\prime } } ( y ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 369, + 768, + 629, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By using the modeled semantic additive noise $z _ { e }$ and the original latent representation $z$ , we can obtain the semantic perturbation $z ^ { \\prime }$ as well as its output $\\hat { y }$ via Eq. (7) for our target objective Eq. (8). ", + "bbox": [ + 173, + 845, + 825, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the described semantic noise modeling process (‘proposed-perturb (semantic)’ in Section 4.3), we expect to achieve better representation on the latent space. The effect of the proposed model in terms of learned latent representation will be explained in more detail in Section 4.4. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/687514226030a4d7c6ee1ffba4053d410f23d742a7430240e9b83e42fb27c8fd.jpg", + "image_caption": [ + "Figure 2: Previous works for supervised learning; (a) traditional feed-forward model, and (b) joint learning model with both supervised and unsupervised losses. " + ], + "image_footnote": [], + "bbox": [ + 276, + 103, + 723, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RELATED WORKS ", + "text_level": 1, + "bbox": [ + 176, + 252, + 351, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works on deep neural networks for supervised learning can be categorized into two types as shown in Figure 2; (a) a general feed-forward neural network model (LeCun et al., 1998; Krizhevsky et al., 2012; Simonyan & Zisserman, 2015; He et al., 2016), and (b) a joint learning model which optimizes unsupervised and supervised objectives at the same time (Zhao et al., 2015; Zhang et al., 2016; Cho & Chen, 2014). Here are the corresponding objective functions: ", + "bbox": [ + 173, + 282, + 825, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/6a83b18bd194965c3057b59b11c2ccaf5cb7f914ce1299e645105d6e95a44ffc.jpg", + "text": "$$\n\\begin{array} { c } { { \\displaystyle { \\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 2 } \\} } \\sum _ { i } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) } } } \\\\ { { \\displaystyle { \\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } \\} } \\sum _ { i } \\lambda L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 331, + 368, + 666, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a loss weighting coefficient between unsupervised and supervised losses. ", + "bbox": [ + 176, + 458, + 723, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the feed-forward neural network model is normally implemented with multiple layers in a deep learning framework, the joint learning model can be sub-classified into two types according to the type of reconstruction; reconstruction only with the input data $x$ (Eq. (11)) and reconstruction with all the intermediate features including the input data $x$ as follows: ", + "bbox": [ + 174, + 479, + 826, + 536 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/6f6ca7f8e03858d18c7985d4aa1a5352221fc0881b3a1290ba74fff57473f947.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta } \\sum _ { i } \\left( \\lambda _ { 0 } L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + \\sum _ { j } \\lambda _ { j } L _ { L 2 } ( h _ { j } ^ { ( i ) } , h _ { j R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) \\right) \\ .\n$$", + "text_format": "latex", + "bbox": [ + 246, + 553, + 750, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where h(i)j and $h _ { j _ { R } } ^ { ( i ) }$ are the $j$ -th hidden representation of the $i$ -th training example and its reconstruction. ", + "bbox": [ + 173, + 616, + 823, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another type of the joint learning model, a ladder network (Figure 3), was introduced for semisupervised learning (Rasmus et al., 2015). The key concept of the ladder network is to obtain robust features by learning de-noising functions $( g _ { \\theta ^ { \\prime } } )$ of the representations at every layer of the model via reconstruction losses, and the supervised loss is combined with the reconstruction losses in order to build the semi-supervised model. The ladder network achieved the best performance in semi-supervised tasks, but it is not appropriate for supervised tasks with small-scale training set (experimental analysis for supervised learning on permutation-invariant MNIST is briefly summarized in Appendix (A2)). The proposed model in this work can be extended to semi-supervised learning, but our main focus is to enhance the representational power on latent space given labelled data for supervised learning. We leave the study for semi-supervised learning scenario based on the proposed methodology as our future research. ", + "bbox": [ + 173, + 651, + 825, + 751 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0adf99802513d294306f73fcbe86e7d7c12a66294e544f6d889b4fc2e8016158.jpg", + "image_caption": [ + "Figure 3: Ladder network; a representative model for semi-supervised learning (Rasmus et al., 2015). " + ], + "image_footnote": [], + "bbox": [ + 375, + 777, + 622, + 871 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 180, + 326, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For quantitative analysis, we compare the proposed methodology with previous approaches described in Section 3; a traditional feed-forward supervised learning model and a joint learning model with two different types of reconstruction losses (reconstruction only with the first layer or with all the intermediate layers including the first layer). The proposed methodology includes a baseline model in Figure 1(b) as well as a stochastic perturbation model in Figure 1(c). Especially in the stochastic perturbation model, we compare the random and semantic perturbations and present some qualitative analysis on the meaning of the proposed perturbation methodology. ", + "bbox": [ + 174, + 212, + 825, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 DATASETS ", + "text_level": 1, + "bbox": [ + 174, + 327, + 287, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment with two public datasets; MNIST (including a permutation-invariant MNIST case) and CIFAR-10. MNIST (10 classes) consists of $5 0 \\mathrm { k }$ , $1 0 \\mathrm { k }$ , and 10k $2 8 \\times 2 8$ gray-scale images for training, validation, and test datasets, respectively. CIFAR-10 (10 classes) consists of $5 0 \\mathrm { k }$ and 10k $3 2 \\times 3 2$ 3-channel images for training and test sets, respectively. We split the $5 0 \\mathrm { k }$ CIFAR-10 training images into 40k and 10k for training and validation. Experiments are performed with different sizes of training set (from 10 examples per class to the entire training set) in order to verify the effectiveness of the proposed model in terms of generalization performance under varying sizes of training set. ", + "bbox": [ + 174, + 352, + 825, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 IMPLEMENTATION ", + "text_level": 1, + "bbox": [ + 174, + 481, + 341, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 shows the architecture of the neural network model used in this experiment. W ’s are convolution or fully-connected weights (biases are excluded for visual brevity). Three convolution $3 { \\times } 3$ (2) 32, $3 \\times 3$ (2) 64, $3 \\times 3$ (2) 96, where each item means the filter kernel size and (stride) with the number of filters) and two fully-connected (the numbers of output nodes are 128 and 10, respectively) layers are used for MNIST. For the permutation-invariant MNIST setting, 784-512- 256-256-128-10 nodes of fully-connected layers are used. Four convolution $( 5 \\times 5$ (1) 64, $3 \\times 3$ (2) 64, $3 \\times 3$ (2) 64, and $3 \\times 3$ (2) 96) and three fully-connected (128, 128, and 10 nodes) layers are used for CIFAR-10. Weights on the decoding (reconstruction) path are tied with corresponding weights on the encoding path as shown in Figure 4 (transposed convolution for the tied convolution layer and transposed matrix multiplication for the tied fully-connected layer). ", + "bbox": [ + 173, + 507, + 825, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 4, $z ^ { \\prime }$ is perturbed directly from $z$ by adding Gaussian random noise for random perturbation. For semantic perturbation, $z ^ { \\prime }$ is indirectly generated from $y ^ { \\prime }$ which is perturbed by adding Gaussian random noise on $y$ based on Eq. (9). For perturbation, base activation vector ( $z$ is the base ", + "bbox": [ + 173, + 654, + 823, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/792c9e8a2575298b7031a74598465f11a754decf19dff68c273d14580802e5a3.jpg", + "text": "$$\n\\begin{array} { r l r } & { } & { z ^ { \\prime } \\xrightarrow { W _ { N } } \\hat { y } \\cdots t } \\\\ & { } & { x \\xrightarrow [ ] { W _ { 1 } } \\gamma _ { h _ { 1 } } \\xrightarrow [ ] { W _ { 2 } } h _ { 2 } \\quad \\bullet \\bullet \\quad \\xrightarrow [ ] { W _ { N - 1 } } \\overbrace { \\overset { \\underset { \\mathrm { A } } { \\mathrm { A } } } { \\overset { \\mathrm { A } } { \\sum } } \\frac { W _ { N } } { y } } ^ { \\overset { \\mathrm { A } } { \\sum } } y \\cdots t } \\\\ & { } & { x \\xrightarrow [ ] { W _ { 1 } ^ { T } } h _ { 1 _ { R } } \\overbrace { \\Longleftarrow } ^ { W _ { 2 } ^ { T } } h _ { 2 _ { R } } \\bullet \\bullet \\quad \\bullet \\overbrace { \\overset { w _ { N - 1 } ^ { T } } { \\Longleftarrow } } ^ { W _ { N - 1 } ^ { T } } \\overbrace { \\overset { z } { z } _ { R } \\overset { w _ { N } ^ { T } } { \\llcorner } } ^ { \\overset { W _ { N } ^ { T } } { \\longrightarrow } \\overset { \\mathcal { Y } } } \\overset { \\cdot } { y ^ { \\prime } } } \\\\ & { } & { z _ { R } ^ { \\prime } \\xrightarrow [ ] { w _ { N } ^ { T } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 277, + 722, + 723, + 858 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "", + "image_caption": [ + "Figure 4: Target network architecture; 3 convolution and 2 fully-connected layers were used for MNIST, 5 fully-connected layers were used for permutation-invariant MNIST, and 4 convolution and 3 fully-connected layers were used for CIFAR-10. " + ], + "image_footnote": [], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Error rate $( \\% )$ on the test set using the model with the best performance on the validation set. Numbers on the first row of each sub-table are the number of randomly chosen per-class training examples. The average performance and the standard deviation of three different random-split datasets (except for the case using the entire training set in the last column) are described in this table (error rate on each random set is summarized in Appendix (A3)). Performance of three previous approaches (with gray background; previous-1, 2, 3 are feed-forward model Figure 2(a), joint learning model with recon-one Figure 2(b), joint learning model with recon-all Figure 2(b), respectively) and the proposed methods (proposed-1, 2, 3 are baseline Figure 1(b), random perturbation Figure 1(c), semantic perturbation Figure 1(c), respectively) is summarized. ", + "bbox": [ + 173, + 99, + 825, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/187c7c7dc663dcb652453c1a4c3cb89acb31c31959ed3658ba1f2d510222dcff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetnumber of per-class examples chosen from 5Ok entire MNIST training examplesentire set
MNIST1020501002005001k2k50k
previous-124.55 (3.04)16.00 (1.33)10.35 (0.66)6.58 (0.42)4.71 (0.28)2.94 (0.23)1.90 (0.27)1.45 (0.08)1.04
previous-221.67 (3.19)13.60 (0.99)7.85 (0.10)5.44 (0.37)4.14 (0.08)2.50 (0.15)1.84 (0.07)1.45 (0.07)1.12
previous-320.11 (2.81)13.69 (0.62)9.15 (0.15)6.77 (0.25)5.39 (0.11)3.89 (0.27)2.91 (0.17)2.28 (0.10)1.87
proposed-121.35 (1.16)11.65 (1.15)6.33 (0.10)4.32 (0.31)3.07 (0.11)1.98 (0.11)1.29 (0.09)0.94 (0.02)0.80
proposed-220.17 (1.52)11.68 (0.81)6.24 (0.29)4.12 (0.24)3.04 (0.13)1.88 (0.05)1.24 (0.03)0.96 (0.08)0.65
proposed-320.11 (0.81)10.59 (0.74)5.92 (0.12)3.79 (0.23)2.72 (0.09)1.78 (0.05)1.15 (0.01)0.88 (0.03)0.62
datasetnumber of per-class examples chosen from 4Ok entire CIFAR-10 training examplesentire set
CIFAR-101020501002005001k2k40k
previous-173.82 (1.43)68.99 (0.54)61.30 (0.83)54.93 (0.56)46.97 (0.59)33.69 (0.43)26.63 (0.39)20.97 (0.09)17.80
previous-275.68 (1.56)69.05 (1.13)61.44 (0.63)55.02 (0.34)46.18 (0.51)33.62 (0.38)26.78 (0.48)21.25 (0.40)17.68
previous-373.33 (1.06)67.63 (0.56)62.59 (0.76)56.37 (0.20)50.51 (0.61)41.26 (0.73)32.55 (1.20)26.38 (0.08)22.71
proposed-171.63 (0.69)66.17 (0.40)58.91 (0.86)52.65 (0.28)43.46 (0.30)31.86 (0.54)25.76 (0.31)21.06 (0.18)17.45
proposed-271.69 (0.25)66.75 (0.54)58.95 (0.63)53.01 (0.26)43.71 (0.19)31.80 (0.18)25.50 (0.33)20.81 (0.27)17.43
proposed-371.50 (1.14)66.87 (0.17)58.30 (0.62)52.32 (0.08)42.98 (0.34)30.91 (0.23)24.81 (0.26)20.19 (0.25)16.16
", + "bbox": [ + 173, + 246, + 838, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vector for the random perturbation and $y$ is the base vector for the semantic perturbation) is scaled to [0.0, 1.0], and the zero-mean Gaussian noise with 0.2 of standard deviation is added (via elementwise addition) on the normalized base activation. This perturbed scaled activation is de-scaled with the original min and max activations of the base vector. ", + "bbox": [ + 174, + 545, + 825, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initial learning rates are 0.005 and 0.001 for MNIST and permutation-invariant MNIST, and 0.002 for CIFAR-10, respectively. The learning rates are decayed by a factor of 5 every 40 epochs until the 120-th epoch. For both datasets, the minibatch size is set to 100, and the target objective is optimized using Adam optimizer (Kingma & Ba, 2015) with a momentum 0.9. All the $\\lambda$ ’s for reconstruction losses in Eq. (11) and Eq. (12) are 0.03 and 0.01 for MNIST and CIFAR-10, respectively. The same weighting factors for reconstruction losses (0.03 for MNIST and 0.01 for CIFAR-10) are used for $\\lambda _ { 1 }$ in Eq (8), and 1.0 is used for $\\lambda _ { 2 }$ . ", + "bbox": [ + 174, + 607, + 825, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input data is first scaled to [0.0, 1.0] and then whitened by the average across all the training examples. In CIFAR-10, random cropping $2 4 \\times 2 4$ image is randomly cropped from the original $3 2 \\times 3 2$ image) and random horizontal flipping (mirroring) are used for data augmentation. We selected the network that performed best on the validation dataset for evaluation on the test dataset. All the experiments are performed with TensorFlow (Abadi et al., 2015). ", + "bbox": [ + 174, + 712, + 825, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 QUANTITATIVE ANALYSIS ", + "text_level": 1, + "bbox": [ + 174, + 800, + 393, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Three previous approaches (a traditional feed-forward model, a joint learning model with the input reconstruction loss, and a joint learning model with reconstruction losses of all the intermediate layers including the input layer) are compared with the proposed methods (the baseline model in Figure 1(b), and the stochastic perturbation model in Figure 1(c) with two different perturbation methods; random and semantic). We measure the classification performance according to varying sizes of training set (examples randomly chosen from the original training dataset). Performance is averaged over three different random trials. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd063f072e83bb12f18e90c57656142a6c76f9eaf9710d6764d7b195a4663837.jpg", + "image_caption": [ + "Figure 5: Examples reconstructed from the perturbed latent vectors via (a) random perturbation, and (b) semantic perturbation (top row shows the original training examples). More examples are summarized in Appendix (A4.1). " + ], + "image_footnote": [], + "bbox": [ + 184, + 108, + 813, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 summarizes the classification performance for MNIST and CIFAR-10. As we expected, the base model obtained by maximizing the sum of mutual informations (proposed-base) mostly performs better than previous approaches, and the model with the semantic perturbation (proposedperturb (semantic)) performs best among all the comparison targets. Especially in MNIST, the error rate of ‘proposed-perturb (semantic)’ with $2 \\mathrm { k }$ per-class training examples is less than the error rate of all types of previous works with the entire training set (approximately 5k per-class examples). ", + "bbox": [ + 173, + 372, + 825, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further verify the proposed method on the permutation-invariant MNIST task with a standard feed-forward neural network. Classification performance is measured against three different sizes of training set (1k, 2k, and 5k per-class training examples). ‘Proposed-perturb (semantic)’ achieves the best performance among all the configurations; $2 . 5 7 \\%$ , $1 . 8 2 \\%$ , and $1 . 2 8 \\%$ error rates for 1k, 2k, and 5k per-class training examples, respectively. The joint learning model with the input reconstruction loss performs best among three previous approaches; $2 . 7 2 \\%$ , $1 . 9 7 \\%$ , and $1 . 3 8 \\%$ error rates for 1k, 2k, and 5k per-class training examples, respectively. ", + "bbox": [ + 174, + 463, + 825, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 QUALITATIVE ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 578, + 383, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As mentioned before, random perturbation by adding unstructured noise directly to the latent representation cannot guarantee preserving the semantics of the original representation. We compared two different perturbation methods (random and semantic) by visualizing the examples reconstructed from the perturbed latent vectors (Figure 5). Top row is the original examples selected from training set (among 2k per-class training examples), and the rest are the reconstructions of their perturbed latent representations. Based on the architecture described in Figure 1(b), we generated five different perturbed latent representations according to the type of perturbation, and reconstructed the perturbed latent vectors through decoding path for reconstruction. ", + "bbox": [ + 174, + 603, + 825, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5(a) and (b) show the examples reconstructed from the random and semantic perturbations, respectively. For both cases, zero-mean Gaussian random noise (0.2 standard deviation) is used for perturbation. As shown in Figure 5(a), random perturbation partially destroys the original semantics; for example, semantics of $\\cdot _ { 1 } \\cdot$ is mostly destroyed under random perturbation, and some examples of $\\cdot _ { 3 } \\cdot$ are reconstructed as being similar to $\\cdot _ { 8 } \\cdot $ rather than its original content $\\cdot _ { 3 } \\cdot$ . Figure 5(b) shows the examples reconstructed from the semantic perturbation. The reconstructed examples show subtle semantic variations while preserving the original semantic contents; for example, thickness difference in $\\cdot _ { 3 } \\cdot$ (example on the third row) or writing style difference in $\\cdot _ { 8 } \\cdot $ (openness of the top left corner). ", + "bbox": [ + 174, + 722, + 825, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 shows the overall effect of the perturbation. In this analysis, 100 per-class MNIST examples are used for training. From the trained model based on the architecture described in Figure 1(b), latent representations $z$ of all the $5 0 \\mathrm { k }$ examples (among 50k examples, only 1k examples were used for training) are visualized by using t-SNE (Maaten & Hinton, 2008). Only the training examples of three classes (0, 1, and 9) among ten classes are depicted as black circles for visual discrimination in ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/762dd002da7354f6b01a3b543c08def476bbdd57561f3a870f21d90555f0e26a.jpg", + "image_caption": [ + "Figure 6: Training examples (circles or crosses with colors described below) over the examples not used for training (depicted as background with different colors); (a) training examples (black circles), (b) training examples (yellow circles) with $3 \\times$ random-perturbed samples (blue crosses), and (c) training examples (yellow circles) with $3 \\times$ semantic-perturbed samples (blue crosses). Best viewed in color. " + ], + "image_footnote": [], + "bbox": [ + 181, + 113, + 818, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 6(a). The rest of the examples which were not used for training (approximately $4 . 9 \\mathrm { k }$ examples per class) are depicted as a background with different colors. We treat the colored background examples (not used for training) as a true distribution of unseen data in order to estimate the generalization level of learned representation according to the type of perturbation. Figure 6(b) and (c) show the training examples (100 examples per class with yellow circles) and their perturbed ones $3 \\times$ sampled from each example with blue crosses) through random and semantic perturbations, respectively. ", + "bbox": [ + 173, + 369, + 825, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Figure 6(b), perturbed samples are distributed near the original training examples, but some samples outside the true distribution cannot be identified easily with appropriate classes. This can be explained with Figure 5(a), since some perturbed samples are ambiguous semantically. In Figure 6(c), however, most of the perturbed samples evenly cover the true distribution. As mentioned before, stochastic perturbation with the semantic additive noise during training implicitly incurs the effect of augmentation on the latent space while resulting in better generalization. Per-class t-SNE results are summarized in Appendix (A4.2). ", + "bbox": [ + 174, + 473, + 825, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 DISCUSSION ", + "text_level": 1, + "bbox": [ + 176, + 590, + 310, + 607 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduced a novel latent space modeling method for supervised tasks based on the standard feed-forward neural network architecture. The presented model simultaneously optimizes both supervised and unsupervised losses based on the assumption that the better latent representation can be obtained by maximizing the sum of hierarchical mutual informations. Especially the stochastic perturbation process which is achieved by modeling the semantic additive noise during training enhances the representational power of the latent space. From the proposed semantic noise modeling process, we can expect improvement of generalization performance in supervised learning with implicit semantic augmentation effect on the latent space. ", + "bbox": [ + 174, + 622, + 825, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The presented model architecture can be intuitively extended to semi-supervised learning because it is implemented as the joint optimization of supervised and unsupervised objectives. For semisupervised learning, however, logical link between features learned from labelled and unlabelled data needs to be considered additionally. We leave the extension of the presented approach to semisupervised learning for the future. ", + "bbox": [ + 174, + 741, + 823, + 810 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 830, + 285, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´ Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´ berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. ", + "bbox": [ + 178, + 854, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 189, + 103, + 823, + 160 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations (ICLR), 2015. ", + "bbox": [ + 176, + 167, + 825, + 210 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yoshua Bengio, Pascal Lamblin, Dan Popovici, Hugo Larochelle, et al. Greedy layer-wise training of deep networks. In Advances in Neural Information Processing Systems (NIPS), 2007. ", + "bbox": [ + 173, + 218, + 823, + 247 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kyunghyun Cho and Xi Chen. Classifying and visualizing motion capture sequences using deep neural networks. In International Conference on Computer Vision Theory and Applications, 2014. ", + "bbox": [ + 174, + 256, + 823, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In International Conference on Machine Learning (ICML), 2008. ", + "bbox": [ + 173, + 292, + 825, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ian Goodfellow, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. Multi-prediction deep boltzmann machines. In Advances in Neural Information Processing Systems (NIPS), 2013. ", + "bbox": [ + 173, + 343, + 823, + 373 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep recurrent neural networks. In International conference on acoustics, speech and signal processing, 2013. ", + "bbox": [ + 174, + 381, + 825, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016. ", + "bbox": [ + 174, + 431, + 820, + 462 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. Signal Processing Magazine, IEEE, 29(6):82–97, 2012. ", + "bbox": [ + 173, + 468, + 825, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey E. Hinton, Simon Osindero, and Yee Whye Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18:1527–1554, 2006. ", + "bbox": [ + 173, + 534, + 823, + 563 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. ", + "bbox": [ + 173, + 570, + 821, + 601 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NIPS), 2012. ", + "bbox": [ + 174, + 607, + 821, + 637 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hugo Larochelle and Yoshua Bengio. Classification using discriminative restricted boltzmann machines. In International Conference on Machine Learning (ICML), 2008a. ", + "bbox": [ + 174, + 645, + 821, + 674 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hugo Larochelle and Yoshua Bengio. Classification using discriminative restricted boltzmann machines. In International Conference on Machine Learning (ICML), 2008b. ", + "bbox": [ + 174, + 681, + 821, + 712 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ", + "bbox": [ + 174, + 718, + 821, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research (JMLR), 9(Nov):2579–2605, 2008. ", + "bbox": [ + 173, + 756, + 823, + 785 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jonathan Masci, Ueli Meier, Dan Cires¸an, and Jurgen Schmidhuber. Stacked convolutional auto- ¨ encoders for hierarchical feature extraction. In International Conference on Artificial Neural Networks, 2011. ", + "bbox": [ + 171, + 792, + 825, + 835 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semisupervised learning with ladder networks. In Advances in Neural Information Processing Systems (NIPS), 2015. ", + "bbox": [ + 173, + 844, + 825, + 887 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ruslan Salakhutdinov and Geoffrey E Hinton. Deep boltzmann machines. In Artificial Intelligence and Statistics Conference (AISTATS), 2009. ", + "bbox": [ + 171, + 895, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research (JMLR), 11:3371–3408, 2010. ", + "bbox": [ + 174, + 140, + 821, + 184 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Satosi Watanabe. Information theoretical analysis of multivariate correlation. IBM Journal of research and development, 4(1):66–82, 1960. ", + "bbox": [ + 171, + 193, + 823, + 222 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yuting Zhang, Kibok Lee, and Honglak Lee. Augmenting supervised neural networks with unsupervised objectives for large-scale image classification. In International Conference on Machine Learning (ICML), 2016. ", + "bbox": [ + 174, + 231, + 823, + 272 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Junbo Zhao, Michael Mathieu, Ross Goroshin, and Yann Lecun. Stacked what-where auto-encoders. In International Conference on Learning Representations (ICLR), 2015. ", + "bbox": [ + 174, + 281, + 823, + 310 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 264, + 117 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(A1) DERIVATION OF RECONSTRUCTION ERRORS FROM CONDITIONAL ENTROPY TERMS ", + "text_level": 1, + "bbox": [ + 174, + 133, + 813, + 148 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Extended from Section 2. From the lower bound in Eq. (3), we consider the following optimization problem (refer to ‘Section 2. From mutual information to autoencoders’ in (Vincent et al., 2010)): ", + "bbox": [ + 173, + 159, + 823, + 188 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/1ce3a20666d9c54155e2316a017f5092c59de56777fae78e4bbfc4295265b501.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\mathbb { E } _ { q ( X , Z , Y ) } \\left[ \\log q ( X | Z ) \\right] + \\mathbb { E } _ { q ( X , Z , Y ) } \\left[ \\log q ( Z | Y ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 287, + 207, + 709, + 233 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Here, we denote $q ( X , Z , Y )$ an unknown joint distribution. Note that $Z$ and $Y$ are respectively the variables transformed from parametric mappings $Z = f _ { \\boldsymbol { \\theta } _ { 1 } } ( { \\boldsymbol { X } } )$ and $Y = f _ { \\theta _ { 2 } } ( Z )$ (see Fig. 1). $q ( X , Z , Y )$ then can be reduced to $q ( X )$ from $q ( Z | X ; \\theta _ { 1 } ) = \\delta ( Z - f _ { \\theta _ { 1 } } ( X ) )$ and $\\overset { \\cdot } { q } ( Y | Z ; \\theta _ { 2 } ) =$ $\\delta ( Y - f _ { \\theta _ { 2 } } ( Z ) )$ where $\\delta$ denotes Dirac-delta function. ", + "bbox": [ + 173, + 243, + 825, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "From the Kullback-Leibler divergence that $D _ { \\mathrm { K L } } ( q | | p ) \\geq 0$ for any two distributions $p$ and $q$ , the optimization in Eq. (13) corresponds to the following optimization problem where $p ( \\cdot )$ denotes a parametric distribution: ", + "bbox": [ + 173, + 306, + 825, + 349 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/e4503b9cdec788a400c98465f93b13c7130abd846c0d52b97b0041b4e6c7a4ea.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\mathbb { E } _ { q ( X ) } \\left[ \\log p ( X | Z ; \\theta _ { 1 } ^ { \\prime } ) \\right] + \\mathbb { E } _ { q ( X ) } \\left[ \\log p ( Z | Y ; \\theta _ { 2 } ^ { \\prime } ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 292, + 368, + 704, + 395 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "By replacing $q ( X )$ with a sample distribution $q ^ { 0 } ( X )$ and putting all parametric dependencies between $X , Z$ and $Y$ , we will have ", + "bbox": [ + 176, + 406, + 823, + 435 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/ab9fa16dde9164e5f52d9f1acdc1b79c3f7cadef1e715d2876e65db711694170.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\mathbb { E } _ { q ^ { 0 } ( X ) } \\left[ \\log p ( X | Z = f _ { \\theta _ { 1 } } ( X ) ; \\theta _ { 1 } ^ { \\prime } ) \\right] + \\mathbb { E } _ { q ^ { 0 } ( X ) } \\left[ \\log p ( Z | Y = f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( X ) ) ; \\theta _ { 2 } ^ { \\prime } ) \\right] ~ .\n$$", + "text_format": "latex", + "bbox": [ + 186, + 462, + 784, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For a given input sample $x$ of $X$ , it is general to interpret $x _ { R }$ and $z _ { R }$ as the parameters of distributions $p ( X | X _ { R } = x _ { R } )$ and $p ( Z | Z _ { R } = z _ { R } )$ which reconstruct $x$ and $z$ with high probability (i.e. $x _ { R }$ and $z _ { R }$ are not exact reconstructions of $x$ and $z$ ). Since $x _ { R }$ and $z _ { R }$ are real-valued, we assume Gaussian distribution for these conditional distributions, that is, ", + "bbox": [ + 173, + 503, + 825, + 560 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/ae1c2f64c599c33838857a2a405065c8f927947874e390a5ba5421fa6c354cf9.jpg", + "text": "$$\n\\begin{array} { r } { p ( { \\cal X } | { \\cal X } _ { R } = x _ { R } ) = N ( x _ { R } , \\sigma _ { 0 } ^ { 2 } { \\bf I } ) } \\\\ { p ( { \\cal Z } | { \\cal Z } _ { R } = z _ { R } ) = N ( z _ { R } , \\sigma _ { 0 } ^ { 2 } { \\bf I } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 388, + 575, + 612, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The assumptions yield $- \\log p ( \\cdot | \\cdot ) \\propto L _ { L 2 } ( \\cdot , \\cdot )$ . ", + "bbox": [ + 174, + 626, + 482, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "With the following relations for log terms in Eq. (15), ", + "bbox": [ + 174, + 647, + 527, + 662 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/22516b298a3a32fe1d6a556977a711c3832afa62a01a904d7bd7a61b664548ed.jpg", + "text": "$$\n\\begin{array} { c } { { p ( X | Z = f _ { \\theta _ { 1 } } ( x ) ; \\theta _ { 1 } ^ { \\prime } ) = p ( X | X _ { R } = g _ { \\theta _ { 1 } ^ { \\prime } } ( f _ { \\theta _ { 1 } } ( x ) ) ) } } \\\\ { { p ( Z | Y = f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) ; \\theta _ { 2 } ^ { \\prime } ) = p ( Z | Z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) ) , } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 302, + 676, + 696, + 719 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the optimization problem in Eq. (15) corresponds to the minimization problem of reconstruction errors for input examples $x ^ { ( i ) }$ as below: ", + "bbox": [ + 171, + 727, + 823, + 757 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/448aa14c5f3795823f70febf4cfbf2bd5abc917b7caefa2b369d2ce57ee76f63.jpg", + "text": "$$\n\\operatorname * { m i n } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\sum _ { i } L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \\ : .\n$$", + "text_format": "latex", + "bbox": [ + 334, + 772, + 663, + 806 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(A2) LADDER NETWORK, A REPRESENTATIVE SEMI-SUPERVISED LEARNING MODEL ", + "text_level": 1, + "bbox": [ + 176, + 104, + 781, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Extended from Section 3. We performed experiments with a ladder network model (Rasmus et al., 2015) in order to estimate the performance on pure supervised tasks according to different sizes of training set. We used the code (https://github.com/rinuboney/ladder.git) for this experiment. The network architecture implemented on the source code is used as is; (784-1000-500-250-250-250- 10). Based on the same network architecture, we implemented the proposed stochastic perturbation model described in Figure 1(c) and compared the classification performance with the ladder network as described in Table 2 (we did not focus on searching the optimal hyperparameters for the proposed model in this experiment). As summarized in the bottom of the table (mean over 3 random trials), the proposed semantic noise modeling method shows a fairly large performance gain compared to the ladder network model with small-scale datasets (e.g., in a case of 10 per-class training examples, the proposed method achieves $2 2 . 1 1 \\%$ of error rate, while the ladder network shows $2 9 . 6 6 \\%$ ). ", + "bbox": [ + 173, + 128, + 825, + 282 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/532ba7f67e6ee7b04a26033ba51cc8c91e0470715dd11a9337043eb0ff2b94e3.jpg", + "table_caption": [ + "Table 2: Classification performance (error rate in $\\%$ ) of the ladder network and the proposed model on three different sets of randomly chosen training examples (MNIST). " + ], + "table_footnote": [], + "table_body": "
set No.1 (# training examples per class)1020501002005001k2k (all) 5k
ladder network model; Figure 325.8516.489.266.004.663.072.151.260.91
proposed-perturb (semantic); Figure 1(c)19.7612.338.776.064.592.931.871.310.93
set No.2 (# training examples per class)1020501002005001k2k
ladder network model; Figure 333.1417.4610.446.674.432.821.941.37
proposed-perturb (semantic); Figure 1(c)23.3615.359.435.754.432.991.871.39
set No.3 (# training examples per class)1020501002005001k2k
ladder network model; Figure 329.9916.999.737.344.393.002.121.47
proposed-perturb (semantic); Figure 1(c)23.2113.988.836.514.322.942.221.49
mean over 3 random trials1020501002005001k2k (all) 5k
ladder network model; Figure 329.6616.989.816.674.492.962.071.370.91
proposed-perturb (semantic); Figure 1(c)22.1113.899.016.114.452.951.991.400.93
", + "bbox": [ + 228, + 344, + 767, + 545 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(A3) QUANTITATIVE ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 400, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Extended from Section 4.3. Among the total $5 0 \\mathrm { k }$ and $4 0 \\mathrm { k }$ training examples in MNIST and CIFAR10, we randomly select the examples for training. Classification performance according to three different randomly chosen training sets are summarized in Table 3 (MNIST) and Table 4 (CIFAR10). Further experiments with denoising constraints are also included. Zero-mean Gaussian random noise with 0.1 standard deviation is used for noise injection. Denoising function helps to achieve slightly better performance on MNIST, but it results in performance degradation on CIFAR-10 (we did not focus on searching the optimal parameters for noise injection in this experiments). ", + "bbox": [ + 173, + 128, + 825, + 228 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e1075899c3ae1858441c5aa2bbbfb8e3e61cc044f6be718c9ccd16ebe407a84e.jpg", + "table_caption": [ + "Table 3: Classification performance (error rate in $\\%$ ) on three different sets of randomly chosen training examples (MNIST). " + ], + "table_footnote": [], + "table_body": "
Set No.1(#train examples per class)
feed-forward model; Figure 2(a)10 22.6120 14.2050 11.25100 6.37200 4.34500 2.631k 1.832k 1.56(all) 5k 1.04
joint learning model with recon-one; Figure 2(b)18.6912.217.845.174.022.581.791.471.12
joint learning model with recon-one with denoising constraints20.3911.917.414.643.652.571.971.530.97
joint learning model with recon-all; Figure 2(b)18.8212.829.346.435.234.122.682.421.87
joint learning model with recon-all with denoisingconstraints17.9311.767.324.783.913.042.521.991.36
proposed-base; Figure 1(b)20.2310.186.473.893.041.891.330.910.80
proposed-base with denoising constraints19.8810.896.624.263.402.442.111.541.13
proposed-perturb (random); Figure 1(c)18.3810.586.643.783.141.901.210.890.65
proposed-perturb (semantic); Figure 1(c)19.339.725.983.472.841.841.160.840.62
Set No.2 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)28.8417.3610.146.204.783.021.611.41
joint learning model with recon-one; Figure 2(b)26.0914.407.985.184.172.291.941.52
joint learning model with recon-one with denoising constraints27.6913.116.955.073.542.371.831.28
joint learning model with recon-all; Figure 2(b)24.0114.138.986.845.443.512.982.18
joint learning model with recon-all with denoising constraints23.0513.297.795.123.923.012.271.84
proposed-base; Figure 1(b)22.9512.986.274.433.222.141.370.96
proposed-base with denoising constraints26.9612.216.454.623.132.531.881.49
proposed-perturb (random); Figure 1(c)22.1012.525.974.262.861.941.230.92
proposed-perturb (semantic); Figure 1(c)21.2211.525.753.912.611.731.140.89
Set No.3 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)22.2016.439.677.165.023.172.251.39
joint learning model with recon-one; Figure 2(b)20.2314.197.735.964.222.621.791.35
joint learning model with recon-one with denoising constraints19.3212.257.445.393.582.371.491.56
joint learning model with recon-al; Figure 2(b)17.5114.129.127.045.494.053.082.25
joint learning model with recon-all with denoising constraints17.0712.507.865.484.052.972.021.98
proposed-base; Figure 1(b)20.8611.796.254.632.961.911.160.96
proposed-base with denoising constraints19.8911.306.264.573.502.631.611.47
proposed-perturb (random); Figure 1(c)20.0211.946.124.323.131.811.281.08
proposed-perturb (semantic); Figure 1(c)19.7810.536.034.002.701.761.140.92
", + "bbox": [ + 200, + 286, + 794, + 780 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/a2f876e9d0cab81e2976db980f4a25e50cb0e8e656d85a316bddb74517125b3e.jpg", + "table_caption": [ + "Table 4: Classification performance (error rate in $\\%$ ) on three different sets of randomly chosen training examples (CIFAR-10). " + ], + "table_footnote": [], + "table_body": "
Set No.1(# train examples per class)1020501002005001k2k(all) 4k
feed-forward model; Figure 2(a)73.3069.2562.4255.6547.7134.3027.0421.0617.80
joint learning model with recon-one; Figure 2(b)75.1970.3862.2555.3046.8934.1226.6321.0517.68
joint learning model with recon-one with denoising constraints73.7268.2061.9955.2346.6436.3729.7825.5321.73
joint learning model with recon-all; Figure 2(b)74.7968.3362.9256.2451.3740.3030.9126.4922.71
joint learning model with recon-all with denoising constraints76.5669.6764.5357.8852.7442.2436.9030.9327.41
proposed-base; Figure 1(b)70.7966.5759.9152.9843.2932.2526.1920.9217.45
proposed-base with denoising constraints71.0367.4960.3753.5244.2833.4028.0025.0621.34
proposed-perturb (random); Figure 1(c)71.8967.1259.2252.7943.8731.8225.0420.9717.43
proposed-perturb (semantic); Figure 1(c)71.5966.9058.6452.3442.7430.9424.4520.1016.16
Set No.2 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)72.3969.4960.4554.8546.9133.3926.7321.00
joint learning model with recon-one; Figure 2(b)74.0669.1460.7154.5445.7033.5427.4320.90
joint learning model with recon-one with denoising constraints76.4069.3360.2855.3847.4036.2929.3124.60
joint learning model with recon-all; Figure 2(b)72.2867.6061.5356.6549.9942.0832.9926.33
joint learning model with recon-all with denoising constraints73.9069.2361.9057.9952.3545.1237.2330.14
proposed-base; Figure 1(b)72.4965.6257.8252.6643.2032.2425.6021.32
proposed-base with denoising constraints72.9966.7557.7853.8144.3333.5628.4025.03
proposed-perturb (random); Figure 1(c)71.8465.9858.0853.3743.4431.5625.6921.03
proposed-perturb (semantic); Figure 1(c)72.8566.6557.4452.2142.7431.1724.9920.54
Set No.3(# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)75.7868.2461.0254.2946.2833.3826.1120.85
joint learning model with recon-one; Figure 2(b)77.7967.6261.3755.2245.9633.2126.2921.81
joint learning model with recon-one with denoising constraints76.6069.2761.1355.1047.5037.1229.6324.88
joint learning model with recon-all; Figure 2(b)72.9266.9763.3156.2350.1641.4133.7526.31
joint learning model with recon-all with denoising constraints76.8368.5365.5858.2952.4345.4239.0132.32
proposed-base; Figure 1(b)71.6066.3158.9952.3043.8831.1025.4820.95
proposed-base with denoising constraints72.3967.2060.6052.6444.6233.5228.0125.25
proposed-perturb (random); Figure 1(c)71.3467.1559.5552.8643.8132.0125.7820.42
proposed-perturb (semantic); Figure 1(c)70.0667.0758.8352.4143.4730.6125.0019.94
", + "bbox": [ + 191, + 145, + 807, + 640 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "(A4.1) QUALITATIVE ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 403, + 118 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Extended from Section 4.4. Figure 7 shows reconstructed examples from perturbed (random or semantic) latent representations (refer to Figure 5 and the analysis described in Section 4.4). ", + "bbox": [ + 171, + 128, + 825, + 159 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/f6a972dc25f50eb96aab8b6b183f045da818ad62acb9bbdc0a1fa7537bee6201.jpg", + "image_caption": [ + "Figure 7: For each example, top row is the original examples selected from the training set, and the rest are reconstructed from the perturbed representations via random (left) and semantic (right) perturbations. " + ], + "image_footnote": [], + "bbox": [ + 184, + 181, + 816, + 531 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "(A4.2) QUALITATIVE ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 405, + 118 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Extended from Section 4.4. Figure 8 shows the t-SNE results per class on MNIST. The overall tendency is similar to the description in Section 4.4. ", + "bbox": [ + 173, + 159, + 823, + 188 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/67539ce30ef40d4cff811931a9fdada3fcb6273cc65b7d415f958e58e94aa3fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 256, + 797, + 916 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/571139d772703de3df5b815c8ae20f2cb437f0dbde8fae9377558d9641fe2d1c.jpg", + "image_caption": [ + "Figure 8: From top to bottom: 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. From left to right: training examples (circle), training examples (circle) $^ +$ random-perturbed samples (cross), and training examples (circle) $^ +$ semantic-perturbed samples (cross). Best viewed in color. " + ], + "image_footnote": [], + "bbox": [ + 207, + 90, + 799, + 522 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/parse/train/SyCSsUDee/SyCSsUDee_model.json b/parse/train/SyCSsUDee/SyCSsUDee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..860debf7c12124afdd195efdcd83871fcac8c724 --- /dev/null +++ b/parse/train/SyCSsUDee/SyCSsUDee_model.json @@ -0,0 +1,19947 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 846, + 1302, + 846, + 1302, + 1363, + 398, + 1363 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1484, + 1404, + 1484, + 1404, + 1910, + 298, + 1910 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 315, + 546, + 1118, + 546, + 1118, + 699, + 315, + 699 + ], + "score": 0.966 + }, + { + "category_id": 0, + "poly": [ + 299, + 220, + 1077, + 220, + 1077, + 324, + 299, + 324 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 301, + 1926, + 1399, + 1926, + 1399, + 1989, + 301, + 1989 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 333, + 2007, + 564, + 2007, + 564, + 2033, + 333, + 2033 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 302, + 1417, + 573, + 1417, + 573, + 1451, + 302, + 1451 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 773, + 781, + 927, + 781, + 927, + 814, + 773, + 814 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 315, + 411, + 713, + 411, + 713, + 500, + 315, + 500 + ], + "score": 0.587 + }, + { + "category_id": 1, + "poly": [ + 316, + 376, + 765, + 376, + 765, + 409, + 316, + 409 + ], + "score": 0.378 + }, + { + "category_id": 0, + "poly": [ + 316, + 547, + 513, + 547, + 513, + 578, + 316, + 578 + ], + "score": 0.108 + }, + { + "category_id": 13, + "poly": [ + 427, + 377, + 494, + 377, + 494, + 407, + 427, + 407 + ], + "score": 0.31, + "latex": "\\mathbf { K i m } ^ { * }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 217.0, + 1004.0, + 217.0, + 1004.0, + 270.0, + 296.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 273.0, + 1078.0, + 273.0, + 1078.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2001.0, + 570.0, + 2001.0, + 570.0, + 2039.0, + 331.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1413.0, + 579.0, + 1413.0, + 579.0, + 1459.0, + 294.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 777.0, + 934.0, + 777.0, + 934.0, + 820.0, + 768.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 544.0, + 516.0, + 544.0, + 516.0, + 584.0, + 311.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 846.0, + 1305.0, + 846.0, + 1305.0, + 880.0, + 394.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 876.0, + 1305.0, + 876.0, + 1305.0, + 910.0, + 394.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 907.0, + 1305.0, + 907.0, + 1305.0, + 941.0, + 394.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 937.0, + 1306.0, + 937.0, + 1306.0, + 971.0, + 393.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 968.0, + 1306.0, + 968.0, + 1306.0, + 1002.0, + 394.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 998.0, + 1305.0, + 998.0, + 1305.0, + 1032.0, + 394.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1030.0, + 1305.0, + 1030.0, + 1305.0, + 1063.0, + 394.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1058.0, + 1306.0, + 1058.0, + 1306.0, + 1093.0, + 393.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1088.0, + 1305.0, + 1088.0, + 1305.0, + 1124.0, + 393.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1118.0, + 1307.0, + 1118.0, + 1307.0, + 1155.0, + 393.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1151.0, + 1306.0, + 1151.0, + 1306.0, + 1183.0, + 394.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1181.0, + 1306.0, + 1181.0, + 1306.0, + 1217.0, + 394.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1211.0, + 1304.0, + 1211.0, + 1304.0, + 1247.0, + 393.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1242.0, + 1306.0, + 1242.0, + 1306.0, + 1278.0, + 394.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1273.0, + 1305.0, + 1273.0, + 1305.0, + 1306.0, + 394.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1303.0, + 1304.0, + 1303.0, + 1304.0, + 1338.0, + 394.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1334.0, + 1265.0, + 1334.0, + 1265.0, + 1367.0, + 394.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1514.0, + 1404.0, + 1514.0, + 1404.0, + 1550.0, + 293.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1547.0, + 1404.0, + 1547.0, + 1404.0, + 1578.0, + 296.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1572.0, + 1405.0, + 1572.0, + 1405.0, + 1614.0, + 291.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1644.0, + 292.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1635.0, + 1404.0, + 1635.0, + 1404.0, + 1672.0, + 292.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1667.0, + 1404.0, + 1667.0, + 1404.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1694.0, + 1405.0, + 1694.0, + 1405.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 1404.0, + 1756.0, + 1404.0, + 1793.0, + 293.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 1401.0, + 1788.0, + 1401.0, + 1820.0, + 294.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1817.0, + 1404.0, + 1817.0, + 1404.0, + 1855.0, + 293.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1847.0, + 1406.0, + 1847.0, + 1406.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1878.0, + 587.0, + 1878.0, + 587.0, + 1913.0, + 295.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 547.0, + 516.0, + 547.0, + 516.0, + 579.0, + 310.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 575.0, + 1120.0, + 575.0, + 1120.0, + 611.0, + 310.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 602.0, + 557.0, + 602.0, + 557.0, + 643.0, + 310.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 638.0, + 631.0, + 638.0, + 631.0, + 669.0, + 311.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 671.0, + 668.0, + 671.0, + 668.0, + 702.0, + 311.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1923.0, + 1403.0, + 1923.0, + 1403.0, + 1963.0, + 296.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1955.0, + 1403.0, + 1955.0, + 1403.0, + 1991.0, + 296.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 404.0, + 432.0, + 404.0, + 432.0, + 439.0, + 311.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 437.0, + 537.0, + 437.0, + 537.0, + 468.0, + 312.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 467.0, + 713.0, + 467.0, + 713.0, + 503.0, + 313.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 372.0, + 426.0, + 372.0, + 426.0, + 419.0, + 310.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 372.0, + 771.0, + 372.0, + 771.0, + 419.0, + 495.0, + 419.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 688, + 1404, + 688, + 1404, + 995, + 298, + 995 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1708, + 1403, + 1708, + 1403, + 1955, + 298, + 1955 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 459, + 1403, + 459, + 1403, + 674, + 298, + 674 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1218, + 1403, + 1218, + 1403, + 1342, + 299, + 1342 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 444, + 298, + 444 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1437, + 1403, + 1437, + 1403, + 1591, + 298, + 1591 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1009, + 1403, + 1009, + 1403, + 1103, + 300, + 1103 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 728, + 1637, + 971, + 1637, + 971, + 1685, + 728, + 1685 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 507, + 1998, + 1191, + 1998, + 1191, + 2037, + 507, + 2037 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 300, + 1149, + 581, + 1149, + 581, + 1185, + 300, + 1185 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1641, + 1399, + 1641, + 1399, + 1670, + 1366, + 1670 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 2004, + 1400, + 2004, + 1400, + 2033, + 1366, + 2033 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 301, + 1382, + 729, + 1382, + 729, + 1412, + 301, + 1412 + ], + "score": 0.816 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.717 + }, + { + "category_id": 0, + "poly": [ + 301, + 1382, + 728, + 1382, + 728, + 1412, + 301, + 1412 + ], + "score": 0.276 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.22 + }, + { + "category_id": 13, + "poly": [ + 373, + 1712, + 459, + 1712, + 459, + 1746, + 373, + 1746 + ], + "score": 0.92, + "latex": "L _ { u n s u p }" + }, + { + "category_id": 13, + "poly": [ + 514, + 1712, + 572, + 1712, + 572, + 1745, + 514, + 1745 + ], + "score": 0.92, + "latex": "L _ { s u p }" + }, + { + "category_id": 13, + "poly": [ + 547, + 1773, + 633, + 1773, + 633, + 1807, + 547, + 1807 + ], + "score": 0.92, + "latex": "L _ { u n s u p }" + }, + { + "category_id": 14, + "poly": [ + 725, + 1637, + 972, + 1637, + 972, + 1686, + 725, + 1686 + ], + "score": 0.9, + "latex": "\\operatorname* { m i n } _ { \\theta } \\lambda L _ { u n s u p } + L _ { s u p }" + }, + { + "category_id": 14, + "poly": [ + 508, + 1999, + 1191, + 1999, + 1191, + 2038, + 508, + 2038 + ], + "score": 0.89, + "latex": "{ \\bf { I } } ( X ; Z ) + { \\bf { I } } ( Z ; Y ) = { \\bf { H } } ( X ) - { \\bf { H } } ( X \\vert Z ) + { \\bf { H } } ( Z ) - { \\bf { H } } ( Z \\vert Y )" + }, + { + "category_id": 13, + "poly": [ + 1283, + 1441, + 1311, + 1441, + 1311, + 1466, + 1283, + 1466 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1773, + 1276, + 1773, + 1276, + 1799, + 1252, + 1799 + ], + "score": 0.83, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 1073, + 1834, + 1101, + 1834, + 1101, + 1861, + 1073, + 1861 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1198, + 1834, + 1223, + 1834, + 1223, + 1860, + 1198, + 1860 + ], + "score": 0.81, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 780, + 1501, + 805, + 1501, + 805, + 1527, + 780, + 1527 + ], + "score": 0.81, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 1309, + 1712, + 1327, + 1712, + 1327, + 1738, + 1309, + 1738 + ], + "score": 0.81, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 586, + 1865, + 611, + 1865, + 611, + 1891, + 586, + 1891 + ], + "score": 0.8, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 795, + 1865, + 819, + 1865, + 819, + 1891, + 795, + 1891 + ], + "score": 0.79, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1105, + 1440, + 1131, + 1440, + 1131, + 1466, + 1105, + 1466 + ], + "score": 0.79, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1712, + 1402, + 1712, + 1402, + 1738, + 1381, + 1738 + ], + "score": 0.77, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 953, + 1501, + 980, + 1501, + 980, + 1527, + 953, + 1527 + ], + "score": 0.76, + "latex": "X" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1144.0, + 587.0, + 1144.0, + 587.0, + 1194.0, + 290.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1382.0, + 735.0, + 1382.0, + 735.0, + 1414.0, + 295.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1382.0, + 734.0, + 1382.0, + 734.0, + 1414.0, + 296.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 689.0, + 1405.0, + 689.0, + 1405.0, + 724.0, + 294.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 719.0, + 1404.0, + 719.0, + 1404.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 750.0, + 1404.0, + 750.0, + 1404.0, + 786.0, + 293.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1404.0, + 782.0, + 1404.0, + 816.0, + 293.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 847.0, + 292.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 843.0, + 1405.0, + 843.0, + 1405.0, + 878.0, + 293.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 872.0, + 1406.0, + 872.0, + 1406.0, + 909.0, + 292.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 904.0, + 1405.0, + 904.0, + 1405.0, + 939.0, + 294.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 934.0, + 1404.0, + 934.0, + 1404.0, + 970.0, + 294.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 964.0, + 742.0, + 964.0, + 742.0, + 999.0, + 296.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1707.0, + 372.0, + 1707.0, + 372.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 1707.0, + 513.0, + 1707.0, + 513.0, + 1748.0, + 460.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1707.0, + 1308.0, + 1707.0, + 1308.0, + 1748.0, + 573.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1707.0, + 1380.0, + 1707.0, + 1380.0, + 1748.0, + 1328.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1707.0, + 1408.0, + 1707.0, + 1408.0, + 1748.0, + 1403.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1402.0, + 1741.0, + 1402.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1770.0, + 546.0, + 1770.0, + 546.0, + 1811.0, + 292.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1770.0, + 1251.0, + 1770.0, + 1251.0, + 1811.0, + 634.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1770.0, + 1408.0, + 1770.0, + 1408.0, + 1811.0, + 1277.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 1404.0, + 1801.0, + 1404.0, + 1835.0, + 294.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1834.0, + 1072.0, + 1834.0, + 1072.0, + 1864.0, + 297.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1834.0, + 1197.0, + 1834.0, + 1197.0, + 1864.0, + 1102.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1834.0, + 1404.0, + 1834.0, + 1404.0, + 1864.0, + 1224.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1862.0, + 585.0, + 1862.0, + 585.0, + 1898.0, + 293.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1862.0, + 794.0, + 1862.0, + 794.0, + 1898.0, + 612.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1862.0, + 1405.0, + 1862.0, + 1405.0, + 1898.0, + 820.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1894.0, + 1407.0, + 1894.0, + 1407.0, + 1928.0, + 293.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 426.0, + 1923.0, + 426.0, + 1958.0, + 293.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 458.0, + 1404.0, + 458.0, + 1404.0, + 499.0, + 293.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 526.0, + 296.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 521.0, + 1404.0, + 521.0, + 1404.0, + 558.0, + 294.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 551.0, + 1404.0, + 551.0, + 1404.0, + 589.0, + 293.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 582.0, + 1403.0, + 582.0, + 1403.0, + 617.0, + 294.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 1405.0, + 613.0, + 1405.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 814.0, + 643.0, + 814.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1218.0, + 1404.0, + 1218.0, + 1404.0, + 1254.0, + 295.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 1404.0, + 1251.0, + 1404.0, + 1284.0, + 295.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 1405.0, + 1280.0, + 1405.0, + 1317.0, + 293.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1310.0, + 1304.0, + 1310.0, + 1304.0, + 1345.0, + 293.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1403.0, + 226.0, + 1403.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 357.0, + 292.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 419.0, + 293.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 414.0, + 492.0, + 414.0, + 492.0, + 447.0, + 293.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 1104.0, + 1436.0, + 1104.0, + 1473.0, + 294.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1436.0, + 1282.0, + 1436.0, + 1282.0, + 1473.0, + 1132.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1436.0, + 1404.0, + 1436.0, + 1404.0, + 1473.0, + 1312.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 1404.0, + 1470.0, + 1404.0, + 1503.0, + 296.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1495.0, + 779.0, + 1495.0, + 779.0, + 1537.0, + 293.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1495.0, + 952.0, + 1495.0, + 952.0, + 1537.0, + 806.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1495.0, + 1405.0, + 1495.0, + 1405.0, + 1537.0, + 981.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1526.0, + 1407.0, + 1526.0, + 1407.0, + 1570.0, + 291.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1555.0, + 393.0, + 1555.0, + 393.0, + 1598.0, + 291.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 1405.0, + 1007.0, + 1405.0, + 1046.0, + 294.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1040.0, + 1404.0, + 1040.0, + 1404.0, + 1077.0, + 294.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1072.0, + 1329.0, + 1072.0, + 1329.0, + 1106.0, + 296.0, + 1106.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1147, + 1406, + 1147, + 1406, + 1363, + 296, + 1363 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 744, + 1403, + 744, + 1403, + 901, + 297, + 901 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 294, + 1577, + 1404, + 1577, + 1404, + 1671, + 294, + 1671 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 297, + 914, + 1402, + 914, + 1402, + 1008, + 297, + 1008 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 452, + 229, + 1245, + 229, + 1245, + 536, + 452, + 536 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 612, + 1414, + 1086, + 1414, + 1086, + 1487, + 612, + 1487 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 658, + 1739, + 1036, + 1739, + 1036, + 1906, + 658, + 1906 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 298, + 565, + 1404, + 565, + 1404, + 659, + 298, + 659 + ], + "score": 0.957 + }, + { + "category_id": 2, + "poly": [ + 297, + 1947, + 1404, + 1947, + 1404, + 2036, + 297, + 2036 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 590, + 1068, + 1104, + 1068, + 1104, + 1111, + 590, + 1111 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 300, + 1531, + 692, + 1531, + 692, + 1563, + 300, + 1563 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.915 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1074, + 1400, + 1074, + 1400, + 1104, + 1366, + 1104 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1427, + 1401, + 1427, + 1401, + 1458, + 1366, + 1458 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1805, + 1401, + 1805, + 1401, + 1836, + 1365, + 1836 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.703 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.382 + }, + { + "category_id": 14, + "poly": [ + 660, + 1737, + 1038, + 1737, + 1038, + 1906, + 660, + 1906 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { z = f _ { \\theta _ { 1 } } ( x ) } \\\\ & { y = f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) } \\\\ & { x _ { R } = g _ { \\theta _ { 1 } ^ { \\prime } } ( z ) = g _ { \\theta _ { 1 } ^ { \\prime } } ( f _ { \\theta _ { 1 } } ( x ) ) } \\\\ & { z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y ) = g _ { \\theta _ { 2 } ^ { \\prime } } ( f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 626, + 1151, + 746, + 1151, + 746, + 1185, + 626, + 1185 + ], + "score": 0.94, + "latex": "- \\mathbf { H } ( X | Z )" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1949, + 1338, + 1949, + 1338, + 1980, + 1250, + 1980 + ], + "score": 0.92, + "latex": "\\mathbf { H } ( Z | Y )" + }, + { + "category_id": 13, + "poly": [ + 616, + 946, + 687, + 946, + 687, + 979, + 616, + 979 + ], + "score": 0.92, + "latex": "\\mathbf { H } ( X )" + }, + { + "category_id": 13, + "poly": [ + 1156, + 1182, + 1201, + 1182, + 1201, + 1224, + 1156, + 1224 + ], + "score": 0.92, + "latex": "x _ { R } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 992, + 915, + 1203, + 915, + 1203, + 949, + 992, + 949 + ], + "score": 0.91, + "latex": "\\mathbf { I } ( X ; Z ) + \\mathbf { I } ( Z ; Y )" + }, + { + "category_id": 13, + "poly": [ + 923, + 1220, + 1152, + 1220, + 1152, + 1253, + 923, + 1253 + ], + "score": 0.91, + "latex": "\\mathbf { H } ( X | Z ) + \\mathbf { H } ( Z | Y )" + }, + { + "category_id": 14, + "poly": [ + 610, + 1415, + 1087, + 1415, + 1087, + 1489, + 610, + 1489 + ], + "score": 0.91, + "latex": "\\operatorname* { m i n } _ { \\theta } \\sum _ { i } L _ { r e c } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { r e c } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1533, + 427, + 1533, + 427, + 1563, + 372, + 1563 + ], + "score": 0.91, + "latex": "L _ { r e c }" + }, + { + "category_id": 13, + "poly": [ + 963, + 837, + 1394, + 837, + 1394, + 871, + 963, + 871 + ], + "score": 0.9, + "latex": "P ( X , Z , Y ) = P ( Y | \\bar { Z } ) \\bar { P } ( Z | X ) P ( X )" + }, + { + "category_id": 13, + "poly": [ + 297, + 946, + 364, + 946, + 364, + 979, + 297, + 979 + ], + "score": 0.9, + "latex": "\\mathbf { H } ( Z )" + }, + { + "category_id": 13, + "poly": [ + 657, + 1254, + 702, + 1254, + 702, + 1287, + 657, + 1287 + ], + "score": 0.9, + "latex": "x ^ { ( i ) }" + }, + { + "category_id": 14, + "poly": [ + 589, + 1070, + 1109, + 1070, + 1109, + 1109, + 589, + 1109 + ], + "score": 0.9, + "latex": "\\mathbf { I } ( X ; Z ) + \\mathbf { I } ( Z ; Y ) \\geq - \\mathbf { H } ( X | Z ) - \\mathbf { H } ( Z | Y ) ." + }, + { + "category_id": 13, + "poly": [ + 1164, + 746, + 1232, + 746, + 1232, + 780, + 1164, + 780 + ], + "score": 0.9, + "latex": "\\mathbf { H } ( \\cdot | \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1185, + 947, + 1401, + 947, + 1401, + 979, + 1185, + 979 + ], + "score": 0.89, + "latex": "\\mathbf { I } ( X ; Z ) + \\mathbf { I } ( Z ; Y )" + }, + { + "category_id": 13, + "poly": [ + 1062, + 746, + 1115, + 746, + 1115, + 780, + 1062, + 780 + ], + "score": 0.89, + "latex": "\\mathbf { H } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 978, + 1253, + 1023, + 1253, + 1023, + 1293, + 978, + 1293 + ], + "score": 0.88, + "latex": "x _ { R } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 443, + 1950, + 504, + 1950, + 504, + 1980, + 443, + 1980 + ], + "score": 0.88, + "latex": "\\mathbf { H } ( Z )" + }, + { + "category_id": 13, + "poly": [ + 502, + 1185, + 546, + 1185, + 546, + 1218, + 502, + 1218 + ], + "score": 0.88, + "latex": "x ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 370, + 746, + 433, + 746, + 433, + 780, + 370, + 780 + ], + "score": 0.88, + "latex": "\\mathbf { I } ( \\cdot ; \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1087, + 1257, + 1130, + 1257, + 1130, + 1287, + 1087, + 1287 + ], + "score": 0.87, + "latex": "z ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 915, + 1353, + 915, + 1353, + 947, + 1283, + 947 + ], + "score": 0.86, + "latex": "\\mathbf { H } ( X )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1290, + 346, + 1290, + 346, + 1333, + 297, + 1333 + ], + "score": 0.85, + "latex": "z _ { R } ^ { ( i ) } )" + }, + { + "category_id": 13, + "poly": [ + 719, + 1949, + 807, + 1949, + 807, + 1980, + 719, + 1980 + ], + "score": 0.84, + "latex": "\\mathbf { H } ( Z | Y )" + }, + { + "category_id": 13, + "poly": [ + 869, + 1191, + 897, + 1191, + 897, + 1218, + 869, + 1218 + ], + "score": 0.79, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1378, + 1580, + 1401, + 1580, + 1401, + 1606, + 1378, + 1606 + ], + "score": 0.79, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 1057, + 809, + 1082, + 809, + 1082, + 835, + 1057, + 835 + ], + "score": 0.77, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 298, + 1647, + 316, + 1647, + 316, + 1666, + 298, + 1666 + ], + "score": 0.7, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 817, + 1950, + 878, + 1950, + 878, + 1980, + 817, + 1980 + ], + "score": 0.6, + "latex": "\\mathbf { H } ( Z )" + }, + { + "category_id": 13, + "poly": [ + 560, + 1194, + 572, + 1194, + 572, + 1218, + 560, + 1218 + ], + "score": 0.52, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 941, + 808, + 1003, + 808, + 1003, + 836, + 941, + 836 + ], + "score": 0.49, + "latex": "X , Z" + }, + { + "category_id": 13, + "poly": [ + 978, + 809, + 1002, + 809, + 1002, + 835, + 978, + 835 + ], + "score": 0.47, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 941, + 808, + 968, + 808, + 968, + 835, + 941, + 835 + ], + "score": 0.39, + "latex": "X" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 226.0, + 594.0, + 226.0, + 594.0, + 280.0, + 533.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 225.0, + 767.0, + 225.0, + 767.0, + 283.0, + 705.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 265.0, + 490.0, + 265.0, + 490.0, + 297.0, + 454.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 267.0, + 656.0, + 267.0, + 656.0, + 294.0, + 631.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 261.0, + 830.0, + 261.0, + 830.0, + 298.0, + 762.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 304.0, + 667.0, + 304.0, + 667.0, + 343.0, + 624.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 304.0, + 1086.0, + 304.0, + 1086.0, + 424.0, + 1037.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 305.0, + 1251.0, + 305.0, + 1251.0, + 352.0, + 1177.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 384.0, + 489.0, + 384.0, + 489.0, + 418.0, + 454.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 345.0, + 597.0, + 345.0, + 597.0, + 404.0, + 533.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 386.0, + 658.0, + 386.0, + 658.0, + 416.0, + 630.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 346.0, + 767.0, + 346.0, + 767.0, + 405.0, + 705.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 387.0, + 826.0, + 387.0, + 826.0, + 412.0, + 799.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 384.0, + 897.0, + 384.0, + 897.0, + 418.0, + 863.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 344.0, + 1006.0, + 344.0, + 1006.0, + 406.0, + 944.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 346.0, + 1177.0, + 346.0, + 1177.0, + 406.0, + 1113.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 396.0, + 1189.0, + 396.0, + 1189.0, + 404.0, + 1181.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 385.0, + 1237.0, + 385.0, + 1237.0, + 416.0, + 1208.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 431.0, + 765.0, + 431.0, + 765.0, + 487.0, + 699.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 432.0, + 1004.0, + 432.0, + 1004.0, + 481.0, + 941.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 430.0, + 1175.0, + 430.0, + 1175.0, + 485.0, + 1112.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 461.0, + 500.0, + 461.0, + 500.0, + 507.0, + 453.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 474.0, + 523.0, + 474.0, + 523.0, + 482.0, + 512.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 458.0, + 674.0, + 458.0, + 674.0, + 509.0, + 625.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 474.0, + 929.0, + 474.0, + 929.0, + 482.0, + 922.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 459.0, + 1099.0, + 459.0, + 1099.0, + 509.0, + 1035.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 502.0, + 668.0, + 502.0, + 668.0, + 542.0, + 624.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 503.0, + 1078.0, + 503.0, + 1078.0, + 542.0, + 1035.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.25, + 274.0, + 1168.25, + 274.0, + 1168.25, + 323.5, + 1116.25, + 323.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 431.0, + 589.0, + 431.0, + 589.0, + 481.0, + 536.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 460.0, + 905.0, + 460.0, + 905.0, + 505.5, + 866.0, + 505.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 566.0, + 1405.0, + 566.0, + 1405.0, + 600.0, + 294.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 597.0, + 1402.0, + 597.0, + 1402.0, + 628.0, + 296.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 627.0, + 626.0, + 627.0, + 626.0, + 660.0, + 296.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1943.0, + 442.0, + 1943.0, + 442.0, + 1986.0, + 330.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1943.0, + 718.0, + 1943.0, + 718.0, + 1986.0, + 505.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1943.0, + 816.0, + 1943.0, + 816.0, + 1986.0, + 808.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1943.0, + 1249.0, + 1943.0, + 1249.0, + 1986.0, + 879.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1986.0, + 1339.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1979.0, + 1402.0, + 1979.0, + 1402.0, + 2008.0, + 296.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2006.0, + 771.0, + 2006.0, + 771.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1150.0, + 625.0, + 1150.0, + 625.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1150.0, + 1404.0, + 1150.0, + 1404.0, + 1185.0, + 747.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1175.0, + 501.0, + 1175.0, + 501.0, + 1235.0, + 289.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1175.0, + 559.0, + 1175.0, + 559.0, + 1235.0, + 547.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1175.0, + 868.0, + 1175.0, + 868.0, + 1235.0, + 573.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1175.0, + 1155.0, + 1175.0, + 1155.0, + 1235.0, + 898.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1175.0, + 1407.0, + 1175.0, + 1407.0, + 1235.0, + 1202.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1220.0, + 922.0, + 1220.0, + 922.0, + 1255.0, + 295.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1220.0, + 1404.0, + 1220.0, + 1404.0, + 1255.0, + 1153.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1244.0, + 656.0, + 1244.0, + 656.0, + 1302.0, + 286.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1244.0, + 977.0, + 1244.0, + 977.0, + 1302.0, + 703.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1244.0, + 1086.0, + 1244.0, + 1086.0, + 1302.0, + 1024.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1244.0, + 1412.0, + 1244.0, + 1412.0, + 1302.0, + 1131.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1280.0, + 296.0, + 1280.0, + 296.0, + 1364.0, + 286.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1280.0, + 1418.0, + 1280.0, + 1418.0, + 1364.0, + 347.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 742.0, + 369.0, + 742.0, + 369.0, + 784.0, + 295.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 742.0, + 1061.0, + 742.0, + 1061.0, + 784.0, + 434.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 742.0, + 1163.0, + 742.0, + 1163.0, + 784.0, + 1116.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 742.0, + 1403.0, + 742.0, + 1403.0, + 784.0, + 1233.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 776.0, + 1403.0, + 776.0, + 1403.0, + 809.0, + 296.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 808.0, + 940.0, + 808.0, + 940.0, + 841.0, + 296.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 808.0, + 1056.0, + 808.0, + 1056.0, + 841.0, + 1004.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 841.0, + 1083.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 835.0, + 962.0, + 835.0, + 962.0, + 873.0, + 294.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 835.0, + 1405.0, + 835.0, + 1405.0, + 873.0, + 1395.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 867.0, + 1349.0, + 867.0, + 1349.0, + 900.0, + 296.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1577.0, + 1377.0, + 1577.0, + 1377.0, + 1613.0, + 293.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1577.0, + 1405.0, + 1577.0, + 1405.0, + 1613.0, + 1402.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1609.0, + 1404.0, + 1609.0, + 1404.0, + 1643.0, + 296.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1636.0, + 297.0, + 1636.0, + 297.0, + 1675.0, + 292.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1636.0, + 1217.0, + 1636.0, + 1217.0, + 1675.0, + 317.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 912.0, + 991.0, + 912.0, + 991.0, + 950.0, + 295.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 912.0, + 1282.0, + 912.0, + 1282.0, + 950.0, + 1204.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 912.0, + 1406.0, + 912.0, + 1406.0, + 950.0, + 1354.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 943.0, + 296.0, + 943.0, + 296.0, + 980.0, + 293.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 943.0, + 615.0, + 943.0, + 615.0, + 980.0, + 365.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 943.0, + 1184.0, + 943.0, + 1184.0, + 980.0, + 688.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 974.0, + 525.0, + 974.0, + 525.0, + 1006.0, + 292.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1529.0, + 371.0, + 1529.0, + 371.0, + 1566.0, + 297.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 1529.0, + 697.0, + 1529.0, + 697.0, + 1566.0, + 428.0, + 1566.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1211, + 1404, + 1211, + 1404, + 1430, + 296, + 1430 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1442, + 1404, + 1442, + 1404, + 1633, + 297, + 1633 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 653, + 1405, + 653, + 1405, + 811, + 297, + 811 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 225, + 1404, + 225, + 1404, + 325, + 297, + 325 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 938, + 1405, + 938, + 1405, + 1031, + 297, + 1031 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1404, + 1941, + 1404, + 2035, + 299, + 2035 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 483, + 1403, + 483, + 1403, + 547, + 296, + 547 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 629, + 1698, + 1070, + 1698, + 1070, + 1829, + 629, + 1829 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 297, + 1862, + 1405, + 1862, + 1405, + 1928, + 297, + 1928 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 431, + 371, + 1265, + 371, + 1265, + 444, + 431, + 444 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 682, + 863, + 1018, + 863, + 1018, + 904, + 682, + 904 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 298, + 1095, + 1397, + 1095, + 1397, + 1167, + 298, + 1167 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 857, + 74, + 857, + 106, + 297, + 106 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 300, + 593, + 709, + 593, + 709, + 625, + 300, + 625 + ], + "score": 0.915 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1746, + 1400, + 1746, + 1400, + 1775, + 1366, + 1775 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 868, + 1400, + 868, + 1400, + 898, + 1366, + 898 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 388, + 1400, + 388, + 1400, + 417, + 1366, + 417 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1168, + 1400, + 1168, + 1400, + 1197, + 1366, + 1197 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.786 + }, + { + "category_id": 14, + "poly": [ + 629, + 1693, + 1071, + 1693, + 1071, + 1831, + 629, + 1831 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y ) } \\\\ & { z _ { R } ^ { \\prime } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y ^ { \\prime } ) = g _ { \\theta _ { 2 } ^ { \\prime } } ( y + y _ { e } ) } \\\\ & { z _ { e } \\simeq z _ { R } ^ { \\prime } - z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( y + y _ { e } ) - g _ { \\theta _ { 2 } ^ { \\prime } } ( y ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1001, + 777, + 1128, + 777, + 1128, + 810, + 1001, + 810 + ], + "score": 0.94, + "latex": "z = f _ { \\boldsymbol { \\theta } _ { 1 } } ( x )" + }, + { + "category_id": 13, + "poly": [ + 438, + 1474, + 584, + 1474, + 584, + 1506, + 438, + 1506 + ], + "score": 0.93, + "latex": "y ^ { \\prime } = y + y _ { e }" + }, + { + "category_id": 14, + "poly": [ + 680, + 863, + 1019, + 863, + 1019, + 903, + 680, + 903 + ], + "score": 0.91, + "latex": "z ^ { \\prime } = z + z _ { e } { \\mathrm { ~ } } \\mathrm { a n d } { \\hat { y } } = f _ { \\theta _ { 2 } } ( z ^ { \\prime } )" + }, + { + "category_id": 14, + "poly": [ + 434, + 371, + 1266, + 371, + 1266, + 448, + 434, + 448 + ], + "score": 0.91, + "latex": "\\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\sum _ { i } \\lambda \\left( L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \\right) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } )" + }, + { + "category_id": 13, + "poly": [ + 376, + 486, + 449, + 486, + 449, + 515, + 376, + 515 + ], + "score": 0.91, + "latex": "L _ { N L L }" + }, + { + "category_id": 13, + "poly": [ + 812, + 1503, + 961, + 1503, + 961, + 1537, + 812, + 1537 + ], + "score": 0.91, + "latex": "\\sigma ; \\mathcal { N } ( \\mathbf { 0 } , \\sigma ^ { 2 } \\mathbf { I } )" + }, + { + "category_id": 13, + "poly": [ + 683, + 293, + 732, + 293, + 732, + 323, + 683, + 323 + ], + "score": 0.9, + "latex": "L _ { L 2 }" + }, + { + "category_id": 14, + "poly": [ + 310, + 1095, + 1401, + 1095, + 1401, + 1172, + 310, + 1172 + ], + "score": 0.9, + "latex": "\\operatorname* { m i n } _ { \\substack { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } } \\sum _ { i } \\lambda _ { 1 } \\left( L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \\right) + \\lambda _ { 2 } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) + L _ { N L L } ( \\hat { y } ^ { ( i ) } , t ^ { ( i ) } ) ." + }, + { + "category_id": 13, + "poly": [ + 712, + 778, + 751, + 778, + 751, + 808, + 712, + 808 + ], + "score": 0.9, + "latex": "Z _ { P }" + }, + { + "category_id": 13, + "poly": [ + 814, + 226, + 859, + 226, + 859, + 259, + 814, + 259 + ], + "score": 0.9, + "latex": "x ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 1234, + 293, + 1291, + 293, + 1291, + 323, + 1234, + 323 + ], + "score": 0.9, + "latex": "L _ { r e c , }" + }, + { + "category_id": 13, + "poly": [ + 879, + 747, + 919, + 747, + 919, + 777, + 879, + 777 + ], + "score": 0.9, + "latex": "Z _ { P }" + }, + { + "category_id": 13, + "poly": [ + 912, + 226, + 951, + 226, + 951, + 258, + 912, + 258 + ], + "score": 0.89, + "latex": "t ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 618, + 226, + 734, + 226, + 734, + 262, + 618, + 262 + ], + "score": 0.88, + "latex": "( x ^ { ( i ) } , t ^ { ( i ) } )" + }, + { + "category_id": 13, + "poly": [ + 1299, + 747, + 1337, + 747, + 1337, + 777, + 1299, + 777 + ], + "score": 0.88, + "latex": "Y _ { P }" + }, + { + "category_id": 13, + "poly": [ + 1185, + 1538, + 1223, + 1538, + 1223, + 1572, + 1185, + 1572 + ], + "score": 0.88, + "latex": "g _ { \\theta _ { 2 } ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 1570, + 1128, + 1570, + 1128, + 1601, + 1103, + 1601 + ], + "score": 0.87, + "latex": "y ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 372, + 939, + 398, + 939, + 398, + 967, + 372, + 967 + ], + "score": 0.86, + "latex": "z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 842, + 1536, + 868, + 1536, + 868, + 1567, + 842, + 1567 + ], + "score": 0.86, + "latex": "y ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 912, + 1897, + 929, + 1897, + 929, + 1927, + 912, + 1927 + ], + "score": 0.86, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 674, + 1479, + 703, + 1479, + 703, + 1506, + 674, + 1506 + ], + "score": 0.85, + "latex": "y _ { e }" + }, + { + "category_id": 13, + "poly": [ + 661, + 1895, + 685, + 1895, + 685, + 1922, + 661, + 1922 + ], + "score": 0.85, + "latex": "z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 751, + 1219, + 780, + 1219, + 780, + 1245, + 751, + 1245 + ], + "score": 0.85, + "latex": "z _ { e }" + }, + { + "category_id": 13, + "poly": [ + 450, + 941, + 469, + 941, + 469, + 972, + 450, + 972 + ], + "score": 0.85, + "latex": "\\hat { y }" + }, + { + "category_id": 13, + "poly": [ + 1149, + 945, + 1177, + 945, + 1177, + 970, + 1149, + 970 + ], + "score": 0.85, + "latex": "z _ { e }" + }, + { + "category_id": 13, + "poly": [ + 828, + 1870, + 856, + 1870, + 856, + 1895, + 828, + 1895 + ], + "score": 0.84, + "latex": "z _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1330, + 1505, + 1355, + 1505, + 1355, + 1532, + 1330, + 1532 + ], + "score": 0.84, + "latex": "z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1604, + 325, + 1604, + 325, + 1631, + 297, + 1631 + ], + "score": 0.83, + "latex": "z _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 747, + 1237, + 747, + 1237, + 774, + 1213, + 774 + ], + "score": 0.8, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 633, + 1574, + 652, + 1574, + 652, + 1601, + 633, + 1601 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 755, + 1450, + 772, + 1450, + 772, + 1476, + 755, + 1476 + ], + "score": 0.78, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 969, + 1310, + 988, + 1310, + 988, + 1338, + 969, + 1338 + ], + "score": 0.78, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1201, + 1449, + 1220, + 1449, + 1220, + 1476, + 1201, + 1476 + ], + "score": 0.77, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1311, + 1310, + 1311, + 1310, + 1333, + 1292, + 1333 + ], + "score": 0.77, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 679, + 975, + 695, + 975, + 695, + 996, + 679, + 996 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1375, + 782, + 1393, + 782, + 1393, + 804, + 1375, + 804 + ], + "score": 0.74, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1282, + 1444, + 1308, + 1444, + 1308, + 1476, + 1282, + 1476 + ], + "score": 0.74, + "latex": "y ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1872, + 1304, + 1872, + 1304, + 1892, + 1287, + 1892 + ], + "score": 0.74, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1040, + 233, + 1053, + 233, + 1053, + 258, + 1040, + 258 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1252, + 1401, + 1252, + 1401, + 1272, + 1383, + 1272 + ], + "score": 0.72, + "latex": "z" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 593.0, + 712.0, + 593.0, + 712.0, + 629.0, + 295.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1210.0, + 750.0, + 1210.0, + 750.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1210.0, + 1405.0, + 1210.0, + 1405.0, + 1251.0, + 781.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1244.0, + 1382.0, + 1244.0, + 1382.0, + 1279.0, + 295.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1244.0, + 1406.0, + 1244.0, + 1406.0, + 1279.0, + 1402.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1406.0, + 1274.0, + 1406.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1306.0, + 968.0, + 1306.0, + 968.0, + 1342.0, + 293.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1306.0, + 1291.0, + 1306.0, + 1291.0, + 1342.0, + 989.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1306.0, + 1406.0, + 1306.0, + 1406.0, + 1342.0, + 1311.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 1407.0, + 1334.0, + 1407.0, + 1372.0, + 293.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1368.0, + 1406.0, + 1368.0, + 1406.0, + 1400.0, + 294.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1341.0, + 1396.0, + 1341.0, + 1433.0, + 293.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1443.0, + 754.0, + 1443.0, + 754.0, + 1479.0, + 295.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1443.0, + 1200.0, + 1443.0, + 1200.0, + 1479.0, + 773.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1443.0, + 1281.0, + 1443.0, + 1281.0, + 1479.0, + 1221.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1443.0, + 1405.0, + 1443.0, + 1405.0, + 1479.0, + 1309.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1471.0, + 437.0, + 1471.0, + 437.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1471.0, + 673.0, + 1471.0, + 673.0, + 1512.0, + 585.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1471.0, + 1405.0, + 1471.0, + 1405.0, + 1512.0, + 704.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1501.0, + 811.0, + 1501.0, + 811.0, + 1540.0, + 292.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1501.0, + 1329.0, + 1501.0, + 1329.0, + 1540.0, + 962.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1540.0, + 1356.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1532.0, + 841.0, + 1532.0, + 841.0, + 1571.0, + 292.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1532.0, + 1184.0, + 1532.0, + 1184.0, + 1571.0, + 869.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1532.0, + 1404.0, + 1532.0, + 1404.0, + 1571.0, + 1224.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1569.0, + 632.0, + 1569.0, + 632.0, + 1605.0, + 295.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1569.0, + 1102.0, + 1569.0, + 1102.0, + 1605.0, + 653.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1569.0, + 1406.0, + 1569.0, + 1406.0, + 1605.0, + 1129.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1600.0, + 296.0, + 1600.0, + 296.0, + 1635.0, + 293.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1600.0, + 1000.0, + 1600.0, + 1000.0, + 1635.0, + 326.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 655.0, + 1405.0, + 655.0, + 1405.0, + 689.0, + 296.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 686.0, + 1403.0, + 686.0, + 1403.0, + 720.0, + 296.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 716.0, + 1405.0, + 716.0, + 1405.0, + 752.0, + 292.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 745.0, + 878.0, + 745.0, + 878.0, + 783.0, + 292.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 745.0, + 1212.0, + 745.0, + 1212.0, + 783.0, + 920.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 745.0, + 1298.0, + 745.0, + 1298.0, + 783.0, + 1238.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 745.0, + 1405.0, + 745.0, + 1405.0, + 783.0, + 1338.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 774.0, + 711.0, + 774.0, + 711.0, + 814.0, + 294.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 774.0, + 1000.0, + 774.0, + 1000.0, + 814.0, + 752.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 774.0, + 1374.0, + 774.0, + 1374.0, + 814.0, + 1129.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 774.0, + 1405.0, + 774.0, + 1405.0, + 814.0, + 1394.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 617.0, + 224.0, + 617.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 224.0, + 813.0, + 224.0, + 813.0, + 266.0, + 735.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 224.0, + 911.0, + 224.0, + 911.0, + 266.0, + 860.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 224.0, + 1039.0, + 224.0, + 1039.0, + 266.0, + 952.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 224.0, + 1405.0, + 224.0, + 1405.0, + 266.0, + 1054.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 682.0, + 291.0, + 682.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 291.0, + 1233.0, + 291.0, + 1233.0, + 326.0, + 733.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 291.0, + 1308.0, + 291.0, + 1308.0, + 326.0, + 1292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 936.0, + 371.0, + 936.0, + 371.0, + 975.0, + 296.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 936.0, + 449.0, + 936.0, + 449.0, + 975.0, + 399.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 936.0, + 1148.0, + 936.0, + 1148.0, + 975.0, + 470.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 936.0, + 1403.0, + 936.0, + 1403.0, + 975.0, + 1178.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 970.0, + 678.0, + 970.0, + 678.0, + 1004.0, + 295.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 970.0, + 1403.0, + 970.0, + 1403.0, + 1004.0, + 696.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1000.0, + 518.0, + 1000.0, + 518.0, + 1032.0, + 296.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1242.0, + 2003.0, + 1242.0, + 2035.0, + 294.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 479.0, + 375.0, + 479.0, + 375.0, + 521.0, + 294.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 479.0, + 1404.0, + 479.0, + 1404.0, + 521.0, + 450.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 515.0, + 1048.0, + 515.0, + 1048.0, + 548.0, + 294.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1864.0, + 827.0, + 1864.0, + 827.0, + 1900.0, + 295.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1864.0, + 1286.0, + 1864.0, + 1286.0, + 1900.0, + 857.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1900.0, + 1305.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1892.0, + 660.0, + 1892.0, + 660.0, + 1931.0, + 294.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1892.0, + 911.0, + 1892.0, + 911.0, + 1931.0, + 686.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1892.0, + 1403.0, + 1892.0, + 1403.0, + 1931.0, + 930.0, + 1931.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1435, + 1405, + 1435, + 1405, + 1654, + 296, + 1654 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 624, + 1404, + 624, + 1404, + 782, + 296, + 782 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1056, + 1406, + 1056, + 1406, + 1182, + 298, + 1182 + ], + "score": 0.978 + }, + { + "category_id": 8, + "poly": [ + 423, + 1218, + 1274, + 1218, + 1274, + 1327, + 423, + 1327 + ], + "score": 0.961 + }, + { + "category_id": 3, + "poly": [ + 471, + 230, + 1231, + 230, + 1231, + 395, + 471, + 395 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 563, + 914, + 1136, + 914, + 1136, + 985, + 563, + 985 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 297, + 1356, + 1401, + 1356, + 1401, + 1421, + 297, + 1421 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 686, + 816, + 1012, + 816, + 1012, + 886, + 686, + 886 + ], + "score": 0.937 + }, + { + "category_id": 3, + "poly": [ + 639, + 1713, + 1059, + 1713, + 1059, + 1919, + 639, + 1919 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 301, + 556, + 598, + 556, + 598, + 591, + 301, + 591 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 302, + 1009, + 1232, + 1009, + 1232, + 1043, + 302, + 1043 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 296, + 74, + 857, + 74, + 857, + 105, + 296, + 105 + ], + "score": 0.919 + }, + { + "category_id": 4, + "poly": [ + 293, + 1948, + 1401, + 1948, + 1401, + 2012, + 293, + 2012 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1352, + 826, + 1400, + 826, + 1400, + 857, + 1352, + 857 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1351, + 924, + 1400, + 924, + 1400, + 956, + 1351, + 956 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1257, + 1400, + 1257, + 1400, + 1287, + 1352, + 1287 + ], + "score": 0.885 + }, + { + "category_id": 4, + "poly": [ + 294, + 421, + 1403, + 421, + 1403, + 488, + 294, + 488 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 859, + 2087, + 859, + 2113, + 840, + 2113 + ], + "score": 0.775 + }, + { + "category_id": 14, + "poly": [ + 421, + 1217, + 1275, + 1217, + 1275, + 1328, + 421, + 1328 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { \\theta } \\sum _ { i } \\left( \\lambda _ { 0 } L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + \\sum _ { j } \\lambda _ { j } L _ { L 2 } ( h _ { j } ^ { ( i ) } , h _ { j R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) \\right) \\ ." + }, + { + "category_id": 13, + "poly": [ + 466, + 1352, + 511, + 1352, + 511, + 1398, + 466, + 1398 + ], + "score": 0.92, + "latex": "h _ { j _ { R } } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1352, + 416, + 1352, + 416, + 1398, + 371, + 1398 + ], + "score": 0.92, + "latex": "h _ { j } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 851, + 1499, + 905, + 1499, + 905, + 1530, + 851, + 1530 + ], + "score": 0.85, + "latex": "( g _ { \\theta ^ { \\prime } } )" + }, + { + "category_id": 14, + "poly": [ + 686, + 816, + 1014, + 816, + 1014, + 888, + 686, + 888 + ], + "score": 0.85, + "latex": "\\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 2 } \\} } \\sum _ { i } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1012, + 391, + 1012, + 391, + 1037, + 372, + 1037 + ], + "score": 0.82, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 594, + 1361, + 610, + 1361, + 610, + 1392, + 594, + 1392 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 937, + 1155, + 957, + 1155, + 957, + 1177, + 937, + 1177 + ], + "score": 0.79, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 957, + 1362, + 970, + 1362, + 970, + 1387, + 957, + 1387 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1047, + 1124, + 1067, + 1124, + 1067, + 1146, + 1047, + 1146 + ], + "score": 0.74, + "latex": "x" + }, + { + "category_id": 14, + "poly": [ + 565, + 913, + 1135, + 913, + 1135, + 987, + 565, + 987 + ], + "score": 0.72, + "latex": "\\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } \\} } \\sum _ { i } \\lambda L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } )" + }, + { + "category_id": 14, + "poly": [ + 565, + 812, + 1136, + 812, + 1136, + 988, + 565, + 988 + ], + "score": 0.32, + "latex": "\\begin{array} { c } { { \\displaystyle { \\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 2 } \\} } \\sum _ { i } L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) } } } \\\\ { { \\displaystyle { \\operatorname* { m i n } _ { \\theta : \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } \\} } \\sum _ { i } \\lambda L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { N L L } ( y ^ { ( i ) } , t ^ { ( i ) } ) } } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 259.0, + 509.0, + 259.0, + 509.0, + 291.0, + 473.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 231.0, + 572.0, + 231.0, + 572.0, + 275.0, + 521.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 257.0, + 623.0, + 257.0, + 623.0, + 290.0, + 589.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 244.0, + 796.0, + 244.0, + 796.0, + 305.0, + 694.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 259.0, + 946.0, + 259.0, + 946.0, + 293.0, + 909.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 230.0, + 1011.0, + 230.0, + 1011.0, + 277.0, + 960.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 259.0, + 1057.0, + 259.0, + 1057.0, + 291.0, + 1029.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 243.0, + 1236.0, + 243.0, + 1236.0, + 307.0, + 1132.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 298.0, + 1011.0, + 298.0, + 1011.0, + 344.0, + 958.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 361.0, + 651.0, + 361.0, + 651.0, + 397.0, + 611.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 361.0, + 1089.0, + 361.0, + 1089.0, + 397.0, + 1049.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 232.5, + 686.0, + 232.5, + 686.0, + 276.5, + 635.0, + 276.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 232.5, + 1122.0, + 232.5, + 1122.0, + 275.5, + 1075.0, + 275.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 327.5, + 959.0, + 327.5, + 959.0, + 369.5, + 912.0, + 369.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1713.0, + 802.0, + 1713.0, + 802.0, + 1755.0, + 687.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1740.0, + 839.0, + 1740.0, + 839.0, + 1767.0, + 814.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1712.0, + 1067.0, + 1712.0, + 1067.0, + 1782.0, + 850.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1765.0, + 1046.0, + 1765.0, + 1046.0, + 1817.0, + 974.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1778.0, + 781.0, + 1778.0, + 781.0, + 1825.0, + 725.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1782.0, + 941.0, + 1782.0, + 941.0, + 1823.0, + 891.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1792.0, + 877.0, + 1792.0, + 877.0, + 1925.0, + 794.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1801.0, + 1028.0, + 1801.0, + 1028.0, + 1853.0, + 966.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1811.0, + 680.0, + 1811.0, + 680.0, + 1844.0, + 645.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1821.0, + 700.0, + 1821.0, + 700.0, + 1853.0, + 666.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1849.0, + 666.0, + 1849.0, + 666.0, + 1870.0, + 655.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1845.0, + 744.0, + 1845.0, + 744.0, + 1892.0, + 694.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1847.0, + 908.0, + 1847.0, + 908.0, + 1892.0, + 861.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1873.0, + 679.0, + 1873.0, + 679.0, + 1907.0, + 646.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1869.0, + 1008.0, + 1869.0, + 1008.0, + 1911.0, + 938.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1721.0, + 672.0, + 1721.0, + 672.0, + 1771.5, + 660.0, + 1771.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 553.0, + 603.0, + 553.0, + 603.0, + 597.0, + 292.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1986.0, + 291.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1973.0, + 381.0, + 1973.0, + 381.0, + 2016.0, + 291.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 422.0, + 1404.0, + 422.0, + 1404.0, + 458.0, + 296.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 453.0, + 988.0, + 453.0, + 988.0, + 490.0, + 296.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2119.0, + 839.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1437.0, + 1405.0, + 1437.0, + 1405.0, + 1472.0, + 296.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 1405.0, + 1467.0, + 1405.0, + 1501.0, + 293.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1500.0, + 850.0, + 1500.0, + 850.0, + 1531.0, + 295.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1500.0, + 1403.0, + 1500.0, + 1403.0, + 1531.0, + 906.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1563.0, + 294.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1558.0, + 1405.0, + 1558.0, + 1405.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1592.0, + 1403.0, + 1592.0, + 1403.0, + 1623.0, + 295.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1619.0, + 1406.0, + 1619.0, + 1406.0, + 1654.0, + 294.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 624.0, + 1406.0, + 624.0, + 1406.0, + 663.0, + 293.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 655.0, + 1404.0, + 655.0, + 1404.0, + 694.0, + 294.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 1405.0, + 686.0, + 1405.0, + 722.0, + 293.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 715.0, + 1406.0, + 715.0, + 1406.0, + 757.0, + 293.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 745.0, + 1134.0, + 745.0, + 1134.0, + 786.0, + 293.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 1408.0, + 1054.0, + 1408.0, + 1093.0, + 293.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1087.0, + 1406.0, + 1087.0, + 1406.0, + 1125.0, + 293.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1120.0, + 1046.0, + 1120.0, + 1046.0, + 1153.0, + 295.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1120.0, + 1404.0, + 1120.0, + 1404.0, + 1153.0, + 1068.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 936.0, + 1148.0, + 936.0, + 1184.0, + 295.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1148.0, + 1087.0, + 1148.0, + 1087.0, + 1184.0, + 958.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1337.0, + 465.0, + 1337.0, + 465.0, + 1416.0, + 417.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1337.0, + 593.0, + 1337.0, + 593.0, + 1416.0, + 512.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1337.0, + 956.0, + 1337.0, + 956.0, + 1416.0, + 611.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1337.0, + 1406.0, + 1337.0, + 1406.0, + 1416.0, + 971.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1388.0, + 357.0, + 1388.0, + 357.0, + 1425.0, + 292.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.25, + 1340.0, + 411.25, + 1340.0, + 411.25, + 1406.0, + 291.25, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1004.0, + 371.0, + 1004.0, + 371.0, + 1049.0, + 295.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1004.0, + 1236.0, + 1004.0, + 1236.0, + 1049.0, + 392.0, + 1049.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 468, + 1404, + 468, + 1404, + 684, + 298, + 684 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 777, + 1405, + 777, + 1405, + 1024, + 298, + 1024 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 296, + 1117, + 1403, + 1117, + 1403, + 1424, + 296, + 1424 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 352, + 299, + 352 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1439, + 1402, + 1439, + 1402, + 1534, + 297, + 1534 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 298, + 1918, + 1401, + 1918, + 1401, + 2012, + 298, + 2012 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 299, + 1061, + 581, + 1061, + 581, + 1092, + 299, + 1092 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 300, + 398, + 557, + 398, + 557, + 433, + 300, + 433 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 298, + 721, + 489, + 721, + 489, + 752, + 298, + 752 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.791 + }, + { + "category_id": 3, + "poly": [ + 473, + 1595, + 1228, + 1595, + 1228, + 1888, + 473, + 1888 + ], + "score": 0.607 + }, + { + "category_id": 8, + "poly": [ + 473, + 1595, + 1228, + 1595, + 1228, + 1888, + 473, + 1888 + ], + "score": 0.586 + }, + { + "category_id": 14, + "poly": [ + 478, + 1589, + 1233, + 1589, + 1233, + 1891, + 478, + 1891 + ], + "score": 0.9, + "latex": "\\begin{array} { r l r } & { } & { z ^ { \\prime } \\xrightarrow { W _ { N } } \\hat { y } \\cdots t } \\\\ & { } & { x \\xrightarrow [ ] { W _ { 1 } } \\gamma _ { h _ { 1 } } \\xrightarrow [ ] { W _ { 2 } } h _ { 2 } \\quad \\bullet \\bullet \\quad \\xrightarrow [ ] { W _ { N - 1 } } \\overbrace { \\overset { \\underset { \\mathrm { A } } { \\mathrm { A } } } { \\overset { \\mathrm { A } } { \\sum } } \\frac { W _ { N } } { y } } ^ { \\overset { \\mathrm { A } } { \\sum } } y \\cdots t } \\\\ & { } & { x \\xrightarrow [ ] { W _ { 1 } ^ { T } } h _ { 1 _ { R } } \\overbrace { \\Longleftarrow } ^ { W _ { 2 } ^ { T } } h _ { 2 _ { R } } \\bullet \\bullet \\quad \\bullet \\overbrace { \\overset { w _ { N - 1 } ^ { T } } { \\Longleftarrow } } ^ { W _ { N - 1 } ^ { T } } \\overbrace { \\overset { z } { z } _ { R } \\overset { w _ { N } ^ { T } } { \\llcorner } } ^ { \\overset { W _ { N } ^ { T } } { \\longrightarrow } \\overset { \\mathcal { Y } } } \\overset { \\cdot } { y ^ { \\prime } } } \\\\ & { } & { z _ { R } ^ { \\prime } \\xrightarrow [ ] { w _ { N } ^ { T } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1067, + 809, + 1150, + 809, + 1150, + 839, + 1067, + 839 + ], + "score": 0.87, + "latex": "2 8 \\times 2 8" + }, + { + "category_id": 13, + "poly": [ + 522, + 1302, + 576, + 1302, + 576, + 1332, + 522, + 1332 + ], + "score": 0.87, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1471, + 1065, + 1471, + 1065, + 1502, + 1038, + 1502 + ], + "score": 0.87, + "latex": "y ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 600, + 1180, + 656, + 1180, + 656, + 1210, + 600, + 1210 + ], + "score": 0.86, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 296, + 870, + 379, + 870, + 379, + 900, + 296, + 900 + ], + "score": 0.85, + "latex": "3 2 \\times 3 2" + }, + { + "category_id": 13, + "poly": [ + 689, + 1471, + 715, + 1471, + 715, + 1498, + 689, + 1498 + ], + "score": 0.85, + "latex": "z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 305, + 1180, + 361, + 1180, + 361, + 1210, + 305, + 1210 + ], + "score": 0.85, + "latex": "3 { \\times } 3" + }, + { + "category_id": 13, + "poly": [ + 1163, + 1272, + 1220, + 1272, + 1220, + 1302, + 1163, + 1302 + ], + "score": 0.84, + "latex": "( 5 \\times 5" + }, + { + "category_id": 13, + "poly": [ + 439, + 1439, + 466, + 1439, + 466, + 1468, + 439, + 1468 + ], + "score": 0.83, + "latex": "z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1307, + 1272, + 1362, + 1272, + 1362, + 1302, + 1307, + 1302 + ], + "score": 0.82, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 453, + 1180, + 508, + 1180, + 508, + 1210, + 453, + 1210 + ], + "score": 0.82, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 338, + 1302, + 392, + 1302, + 392, + 1331, + 338, + 1331 + ], + "score": 0.81, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 598, + 1506, + 617, + 1506, + 617, + 1533, + 598, + 1533 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 769, + 1445, + 788, + 1445, + 788, + 1467, + 769, + 1467 + ], + "score": 0.74, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1507, + 1280, + 1507, + 1280, + 1528, + 1263, + 1528 + ], + "score": 0.7, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1141, + 870, + 1187, + 870, + 1187, + 899, + 1141, + 899 + ], + "score": 0.47, + "latex": "5 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1258, + 840, + 1305, + 840, + 1305, + 869, + 1258, + 869 + ], + "score": 0.43, + "latex": "5 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 851, + 809, + 897, + 809, + 897, + 838, + 851, + 838 + ], + "score": 0.35, + "latex": "5 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 911, + 809, + 955, + 809, + 955, + 838, + 911, + 838 + ], + "score": 0.3, + "latex": "1 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1918.0, + 1406.0, + 1918.0, + 1406.0, + 1953.0, + 293.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1949.0, + 1405.0, + 1949.0, + 1405.0, + 1983.0, + 294.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1981.0, + 905.0, + 1981.0, + 905.0, + 2011.0, + 296.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1061.0, + 584.0, + 1061.0, + 584.0, + 1094.0, + 296.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 395.0, + 561.0, + 395.0, + 561.0, + 439.0, + 293.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 719.0, + 495.0, + 719.0, + 495.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1677.0, + 477.0, + 1677.0, + 477.0, + 1811.0, + 470.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1664.0, + 711.0, + 1664.0, + 711.0, + 1702.0, + 669.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1665.0, + 1109.0, + 1665.0, + 1109.0, + 1702.0, + 1061.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1682.5, + 649.0, + 1682.5, + 649.0, + 1726.5, + 607.0, + 1726.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1685.5, + 779.0, + 1685.5, + 779.0, + 1725.5, + 736.0, + 1725.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1762.5, + 668.0, + 1762.5, + 668.0, + 1806.5, + 605.0, + 1806.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1762.0, + 797.0, + 1762.0, + 797.0, + 1805.0, + 737.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 467.0, + 1404.0, + 467.0, + 1404.0, + 506.0, + 293.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 498.0, + 1406.0, + 498.0, + 1406.0, + 536.0, + 293.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 529.0, + 1405.0, + 529.0, + 1405.0, + 564.0, + 296.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 561.0, + 1404.0, + 561.0, + 1404.0, + 596.0, + 294.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 591.0, + 1404.0, + 591.0, + 1404.0, + 625.0, + 294.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 622.0, + 1405.0, + 622.0, + 1405.0, + 656.0, + 294.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 649.0, + 1174.0, + 649.0, + 1174.0, + 690.0, + 293.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 776.0, + 1405.0, + 776.0, + 1405.0, + 813.0, + 295.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 809.0, + 850.0, + 809.0, + 850.0, + 839.0, + 296.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 809.0, + 910.0, + 809.0, + 910.0, + 839.0, + 898.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 809.0, + 1066.0, + 809.0, + 1066.0, + 839.0, + 956.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 809.0, + 1403.0, + 809.0, + 1403.0, + 839.0, + 1151.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 839.0, + 1257.0, + 839.0, + 1257.0, + 873.0, + 293.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 839.0, + 1406.0, + 839.0, + 1406.0, + 873.0, + 1306.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 866.0, + 295.0, + 866.0, + 295.0, + 908.0, + 292.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 866.0, + 1140.0, + 866.0, + 1140.0, + 908.0, + 380.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 866.0, + 1406.0, + 866.0, + 1406.0, + 908.0, + 1188.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 934.0, + 295.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 931.0, + 1405.0, + 931.0, + 1405.0, + 964.0, + 295.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 961.0, + 1408.0, + 961.0, + 1408.0, + 998.0, + 293.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 437.0, + 992.0, + 437.0, + 1027.0, + 295.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1115.0, + 1405.0, + 1115.0, + 1405.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1151.0, + 1405.0, + 1151.0, + 1405.0, + 1183.0, + 295.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1179.0, + 304.0, + 1179.0, + 304.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1179.0, + 452.0, + 1179.0, + 452.0, + 1215.0, + 362.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1179.0, + 599.0, + 1179.0, + 599.0, + 1215.0, + 509.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1215.0, + 657.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1405.0, + 1209.0, + 1405.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1405.0, + 1240.0, + 1405.0, + 1276.0, + 293.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1268.0, + 1162.0, + 1268.0, + 1162.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1268.0, + 1306.0, + 1268.0, + 1306.0, + 1307.0, + 1221.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1307.0, + 1363.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1301.0, + 337.0, + 1301.0, + 337.0, + 1337.0, + 294.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1301.0, + 521.0, + 1301.0, + 521.0, + 1337.0, + 393.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1301.0, + 1406.0, + 1301.0, + 1406.0, + 1337.0, + 577.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1328.0, + 1405.0, + 1328.0, + 1405.0, + 1368.0, + 293.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1399.0, + 294.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1393.0, + 1098.0, + 1393.0, + 1098.0, + 1428.0, + 295.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 292.0, + 1407.0, + 292.0, + 1407.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 703.0, + 322.0, + 703.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 438.0, + 1438.0, + 438.0, + 1476.0, + 293.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1438.0, + 768.0, + 1438.0, + 768.0, + 1476.0, + 467.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1438.0, + 1404.0, + 1438.0, + 1404.0, + 1476.0, + 789.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1468.0, + 688.0, + 1468.0, + 688.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1468.0, + 1037.0, + 1468.0, + 1037.0, + 1506.0, + 716.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1506.0, + 1066.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1501.0, + 597.0, + 1501.0, + 597.0, + 1535.0, + 297.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1501.0, + 1262.0, + 1501.0, + 1262.0, + 1535.0, + 618.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1535.0, + 1281.0, + 1535.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 296, + 544, + 1426, + 544, + 1426, + 1127, + 296, + 1127 + ], + "score": 0.983, + "html": "
datasetnumber of per-class examples chosen from 5Ok entire MNIST training examplesentire set
MNIST1020501002005001k2k50k
previous-124.55 (3.04)16.00 (1.33)10.35 (0.66)6.58 (0.42)4.71 (0.28)2.94 (0.23)1.90 (0.27)1.45 (0.08)1.04
previous-221.67 (3.19)13.60 (0.99)7.85 (0.10)5.44 (0.37)4.14 (0.08)2.50 (0.15)1.84 (0.07)1.45 (0.07)1.12
previous-320.11 (2.81)13.69 (0.62)9.15 (0.15)6.77 (0.25)5.39 (0.11)3.89 (0.27)2.91 (0.17)2.28 (0.10)1.87
proposed-121.35 (1.16)11.65 (1.15)6.33 (0.10)4.32 (0.31)3.07 (0.11)1.98 (0.11)1.29 (0.09)0.94 (0.02)0.80
proposed-220.17 (1.52)11.68 (0.81)6.24 (0.29)4.12 (0.24)3.04 (0.13)1.88 (0.05)1.24 (0.03)0.96 (0.08)0.65
proposed-320.11 (0.81)10.59 (0.74)5.92 (0.12)3.79 (0.23)2.72 (0.09)1.78 (0.05)1.15 (0.01)0.88 (0.03)0.62
datasetnumber of per-class examples chosen from 4Ok entire CIFAR-10 training examplesentire set
CIFAR-101020501002005001k2k40k
previous-173.82 (1.43)68.99 (0.54)61.30 (0.83)54.93 (0.56)46.97 (0.59)33.69 (0.43)26.63 (0.39)20.97 (0.09)17.80
previous-275.68 (1.56)69.05 (1.13)61.44 (0.63)55.02 (0.34)46.18 (0.51)33.62 (0.38)26.78 (0.48)21.25 (0.40)17.68
previous-373.33 (1.06)67.63 (0.56)62.59 (0.76)56.37 (0.20)50.51 (0.61)41.26 (0.73)32.55 (1.20)26.38 (0.08)22.71
proposed-171.63 (0.69)66.17 (0.40)58.91 (0.86)52.65 (0.28)43.46 (0.30)31.86 (0.54)25.76 (0.31)21.06 (0.18)17.45
proposed-271.69 (0.25)66.75 (0.54)58.95 (0.63)53.01 (0.26)43.71 (0.19)31.80 (0.18)25.50 (0.33)20.81 (0.27)17.43
proposed-371.50 (1.14)66.87 (0.17)58.30 (0.62)52.32 (0.08)42.98 (0.34)30.91 (0.23)24.81 (0.26)20.19 (0.25)16.16
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1338, + 1404, + 1338, + 1404, + 1555, + 298, + 1555 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1569, + 1403, + 1569, + 1403, + 1724, + 299, + 1724 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1202, + 1403, + 1202, + 1403, + 1324, + 298, + 1324 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 222, + 1404, + 222, + 1404, + 500, + 297, + 500 + ], + "score": 0.956 + }, + { + "category_id": 0, + "poly": [ + 299, + 1762, + 671, + 1762, + 671, + 1793, + 299, + 1793 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 856, + 75, + 856, + 104, + 299, + 104 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.673 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.186 + }, + { + "category_id": 13, + "poly": [ + 1320, + 1601, + 1402, + 1601, + 1402, + 1631, + 1320, + 1631 + ], + "score": 0.88, + "latex": "3 2 \\times 3 2" + }, + { + "category_id": 13, + "poly": [ + 656, + 1524, + 687, + 1524, + 687, + 1554, + 656, + 1554 + ], + "score": 0.87, + "latex": "\\lambda _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 728, + 1601, + 810, + 1601, + 810, + 1631, + 728, + 1631 + ], + "score": 0.87, + "latex": "2 4 \\times 2 4" + }, + { + "category_id": 13, + "poly": [ + 297, + 1524, + 329, + 1524, + 329, + 1555, + 297, + 1555 + ], + "score": 0.86, + "latex": "\\lambda _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 728, + 1208, + 746, + 1208, + 746, + 1234, + 728, + 1234 + ], + "score": 0.81, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 514, + 226, + 558, + 226, + 558, + 255, + 514, + 255 + ], + "score": 0.77, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1156, + 1434, + 1176, + 1434, + 1176, + 1460, + 1156, + 1460 + ], + "score": 0.64, + "latex": "\\lambda" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1761.0, + 673.0, + 1761.0, + 673.0, + 1797.0, + 295.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1340.0, + 1405.0, + 1340.0, + 1405.0, + 1375.0, + 294.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1370.0, + 1405.0, + 1370.0, + 1405.0, + 1407.0, + 293.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1437.0, + 295.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1433.0, + 1155.0, + 1433.0, + 1155.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1433.0, + 1404.0, + 1433.0, + 1404.0, + 1468.0, + 1177.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1461.0, + 1402.0, + 1461.0, + 1402.0, + 1496.0, + 294.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1492.0, + 1405.0, + 1492.0, + 1405.0, + 1527.0, + 294.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1522.0, + 655.0, + 1522.0, + 655.0, + 1559.0, + 330.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1522.0, + 698.0, + 1522.0, + 698.0, + 1559.0, + 688.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 296.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1853.0, + 1404.0, + 1853.0, + 1404.0, + 1883.0, + 294.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1914.0, + 1404.0, + 1914.0, + 1404.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1938.0, + 1405.0, + 1938.0, + 1405.0, + 1981.0, + 292.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 783.0, + 2003.0, + 783.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1571.0, + 1404.0, + 1571.0, + 1404.0, + 1604.0, + 295.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 727.0, + 1599.0, + 727.0, + 1635.0, + 293.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1599.0, + 1319.0, + 1599.0, + 1319.0, + 1635.0, + 811.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 1405.0, + 1631.0, + 1405.0, + 1668.0, + 293.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1662.0, + 1404.0, + 1662.0, + 1404.0, + 1696.0, + 295.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1692.0, + 1025.0, + 1692.0, + 1025.0, + 1727.0, + 294.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1202.0, + 727.0, + 1202.0, + 727.0, + 1236.0, + 293.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1202.0, + 1407.0, + 1202.0, + 1407.0, + 1236.0, + 747.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1232.0, + 1402.0, + 1232.0, + 1402.0, + 1264.0, + 296.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1264.0, + 1404.0, + 1264.0, + 1404.0, + 1295.0, + 293.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1294.0, + 916.0, + 1294.0, + 916.0, + 1327.0, + 294.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 223.0, + 513.0, + 223.0, + 513.0, + 259.0, + 295.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 223.0, + 1404.0, + 223.0, + 1404.0, + 259.0, + 559.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 253.0, + 1405.0, + 253.0, + 1405.0, + 289.0, + 294.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 285.0, + 1406.0, + 285.0, + 1406.0, + 320.0, + 292.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 316.0, + 1404.0, + 316.0, + 1404.0, + 349.0, + 295.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 343.0, + 1404.0, + 343.0, + 1404.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 374.0, + 1406.0, + 374.0, + 1406.0, + 416.0, + 292.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 407.0, + 1404.0, + 407.0, + 1404.0, + 444.0, + 294.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 434.0, + 1406.0, + 434.0, + 1406.0, + 477.0, + 292.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 467.0, + 1005.0, + 467.0, + 1005.0, + 504.0, + 295.0, + 504.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1329, + 1403, + 1329, + 1403, + 1575, + 298, + 1575 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1589, + 1403, + 1589, + 1403, + 1866, + 298, + 1866 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1020, + 1404, + 1020, + 1404, + 1235, + 299, + 1235 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 820, + 1403, + 820, + 1403, + 1005, + 297, + 1005 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 316, + 239, + 1385, + 239, + 1385, + 616, + 316, + 616 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1881, + 1402, + 1881, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 298, + 654, + 1404, + 654, + 1404, + 746, + 298, + 746 + ], + "score": 0.956 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 856, + 76, + 856, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 300, + 1274, + 653, + 1274, + 653, + 1303, + 300, + 1303 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 13, + "poly": [ + 1041, + 1112, + 1115, + 1112, + 1115, + 1141, + 1041, + 1141 + ], + "score": 0.87, + "latex": "1 . 2 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 987, + 1173, + 1060, + 1173, + 1060, + 1202, + 987, + 1202 + ], + "score": 0.86, + "latex": "1 . 9 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1122, + 1173, + 1196, + 1173, + 1196, + 1202, + 1122, + 1202 + ], + "score": 0.86, + "latex": "1 . 3 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 909, + 1112, + 983, + 1112, + 983, + 1141, + 909, + 1141 + ], + "score": 0.83, + "latex": "1 . 8 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 823, + 1112, + 898, + 1112, + 898, + 1142, + 823, + 1142 + ], + "score": 0.83, + "latex": "2 . 5 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 898, + 1173, + 973, + 1173, + 973, + 1203, + 898, + 1203 + ], + "score": 0.82, + "latex": "2 . 7 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 540, + 1948, + 558, + 1948, + 558, + 1970, + 540, + 1970 + ], + "score": 0.71, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 449, + 1804, + 483, + 1804, + 483, + 1833, + 449, + 1833 + ], + "score": 0.71, + "latex": "\\cdot _ { 3 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 333, + 1713, + 367, + 1713, + 367, + 1741, + 333, + 1741 + ], + "score": 0.67, + "latex": "\\cdot _ { 3 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1804, + 1165, + 1804, + 1165, + 1833, + 1129, + 1833 + ], + "score": 0.62, + "latex": "\\cdot _ { 8 } \\cdot " + }, + { + "category_id": 13, + "poly": [ + 777, + 944, + 810, + 944, + 810, + 973, + 777, + 973 + ], + "score": 0.45, + "latex": "2 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 806, + 1713, + 840, + 1713, + 840, + 1742, + 806, + 1742 + ], + "score": 0.38, + "latex": "\\cdot _ { 8 } \\cdot " + }, + { + "category_id": 13, + "poly": [ + 598, + 1683, + 631, + 1683, + 631, + 1710, + 598, + 1710 + ], + "score": 0.31, + "latex": "\\cdot _ { 1 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 1210, + 1713, + 1242, + 1713, + 1242, + 1742, + 1210, + 1742 + ], + "score": 0.31, + "latex": "\\cdot _ { 3 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 666, + 1943, + 713, + 1943, + 713, + 1971, + 666, + 1971 + ], + "score": 0.3, + "latex": "5 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 240.0, + 836.0, + 240.0, + 836.0, + 300.0, + 316.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 237.0, + 1389.0, + 237.0, + 1389.0, + 301.0, + 869.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 291.0, + 831.0, + 291.0, + 831.0, + 354.0, + 321.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 285.0, + 1392.0, + 285.0, + 1392.0, + 359.0, + 859.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 354.0, + 472.0, + 354.0, + 472.0, + 406.0, + 323.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 358.0, + 523.0, + 358.0, + 523.0, + 404.0, + 474.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 355.0, + 579.0, + 355.0, + 579.0, + 405.0, + 527.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 367.0, + 620.0, + 367.0, + 620.0, + 395.0, + 583.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 355.0, + 680.0, + 355.0, + 680.0, + 405.0, + 628.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 359.0, + 724.0, + 359.0, + 724.0, + 403.0, + 684.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 357.0, + 780.0, + 357.0, + 780.0, + 403.0, + 735.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 354.0, + 830.0, + 354.0, + 830.0, + 401.0, + 784.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 350.0, + 922.0, + 350.0, + 922.0, + 404.0, + 871.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 358.0, + 969.0, + 358.0, + 969.0, + 398.0, + 924.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 348.0, + 1331.0, + 348.0, + 1331.0, + 409.0, + 978.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 357.0, + 1377.0, + 357.0, + 1377.0, + 399.0, + 1335.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 406.0, + 471.0, + 406.0, + 471.0, + 461.0, + 322.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 408.0, + 521.0, + 408.0, + 521.0, + 456.0, + 474.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 413.0, + 723.0, + 413.0, + 723.0, + 451.0, + 686.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 425.0, + 765.0, + 425.0, + 765.0, + 435.0, + 751.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 409.0, + 830.0, + 409.0, + 830.0, + 456.0, + 786.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 406.0, + 917.0, + 406.0, + 917.0, + 456.0, + 874.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 410.0, + 1071.0, + 410.0, + 1071.0, + 451.0, + 1028.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 406.0, + 1377.0, + 406.0, + 1377.0, + 457.0, + 1076.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 468.0, + 362.0, + 468.0, + 362.0, + 505.0, + 326.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 468.0, + 415.0, + 468.0, + 415.0, + 503.0, + 374.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 468.0, + 514.0, + 468.0, + 514.0, + 506.0, + 478.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 470.0, + 566.0, + 470.0, + 566.0, + 505.0, + 536.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 467.0, + 829.0, + 467.0, + 829.0, + 507.0, + 787.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 468.0, + 913.0, + 468.0, + 913.0, + 506.0, + 876.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 472.0, + 958.0, + 472.0, + 958.0, + 498.0, + 932.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 460.0, + 1383.0, + 460.0, + 1383.0, + 514.0, + 1020.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 522.0, + 362.0, + 522.0, + 362.0, + 559.0, + 328.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 527.0, + 469.0, + 527.0, + 469.0, + 558.0, + 421.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 521.0, + 577.0, + 521.0, + 577.0, + 560.0, + 529.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 531.0, + 615.0, + 531.0, + 615.0, + 554.0, + 580.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 531.0, + 667.0, + 531.0, + 667.0, + 554.0, + 640.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 521.0, + 827.0, + 521.0, + 827.0, + 562.0, + 787.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 505.0, + 1390.0, + 505.0, + 1390.0, + 575.0, + 860.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 580.0, + 602.0, + 580.0, + 602.0, + 626.0, + 552.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 578.0, + 1155.0, + 578.0, + 1155.0, + 627.0, + 1099.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 412.5, + 560.0, + 412.5, + 560.0, + 446.0, + 544.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.75, + 418.0, + 1009.75, + 418.0, + 1009.75, + 445.0, + 991.75, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.25, + 472.5, + 1014.25, + 472.5, + 1014.25, + 500.5, + 985.25, + 500.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 528.0, + 527.0, + 528.0, + 527.0, + 557.5, + 468.0, + 557.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 652.0, + 1404.0, + 652.0, + 1404.0, + 692.0, + 293.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 686.0, + 1404.0, + 686.0, + 1404.0, + 719.0, + 296.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 716.0, + 673.0, + 716.0, + 673.0, + 749.0, + 296.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1272.0, + 655.0, + 1272.0, + 655.0, + 1308.0, + 294.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1327.0, + 1403.0, + 1327.0, + 1403.0, + 1368.0, + 293.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1362.0, + 1403.0, + 1362.0, + 1403.0, + 1394.0, + 292.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 1404.0, + 1391.0, + 1404.0, + 1428.0, + 293.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 1405.0, + 1422.0, + 1405.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1452.0, + 1403.0, + 1452.0, + 1403.0, + 1488.0, + 293.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1407.0, + 1511.0, + 1407.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1545.0, + 1073.0, + 1545.0, + 1073.0, + 1578.0, + 296.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1624.0, + 296.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1619.0, + 1407.0, + 1619.0, + 1407.0, + 1657.0, + 292.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1687.0, + 292.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1679.0, + 597.0, + 1679.0, + 597.0, + 1717.0, + 293.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1717.0, + 632.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1709.0, + 332.0, + 1709.0, + 332.0, + 1748.0, + 293.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1709.0, + 805.0, + 1709.0, + 805.0, + 1748.0, + 368.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1709.0, + 1209.0, + 1709.0, + 1209.0, + 1748.0, + 841.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1709.0, + 1405.0, + 1709.0, + 1405.0, + 1748.0, + 1243.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1739.0, + 1404.0, + 1739.0, + 1404.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1798.0, + 448.0, + 1798.0, + 448.0, + 1843.0, + 292.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1798.0, + 1128.0, + 1798.0, + 1128.0, + 1843.0, + 484.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1798.0, + 1405.0, + 1798.0, + 1405.0, + 1843.0, + 1166.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1832.0, + 435.0, + 1832.0, + 435.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1406.0, + 1017.0, + 1406.0, + 1056.0, + 294.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1051.0, + 1406.0, + 1051.0, + 1406.0, + 1086.0, + 295.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1082.0, + 1403.0, + 1082.0, + 1403.0, + 1116.0, + 295.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1112.0, + 822.0, + 1112.0, + 822.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1112.0, + 908.0, + 1112.0, + 908.0, + 1146.0, + 899.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1112.0, + 1040.0, + 1112.0, + 1040.0, + 1146.0, + 984.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1112.0, + 1405.0, + 1112.0, + 1405.0, + 1146.0, + 1116.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1142.0, + 1405.0, + 1142.0, + 1405.0, + 1177.0, + 295.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1173.0, + 897.0, + 1173.0, + 897.0, + 1207.0, + 294.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1173.0, + 986.0, + 1173.0, + 986.0, + 1207.0, + 974.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1173.0, + 1121.0, + 1173.0, + 1121.0, + 1207.0, + 1061.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1173.0, + 1405.0, + 1173.0, + 1405.0, + 1207.0, + 1197.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1201.0, + 879.0, + 1201.0, + 879.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 855.0, + 294.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 849.0, + 1404.0, + 849.0, + 1404.0, + 887.0, + 294.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 918.0, + 292.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 911.0, + 1406.0, + 911.0, + 1406.0, + 949.0, + 291.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 946.0, + 776.0, + 946.0, + 776.0, + 977.0, + 295.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 946.0, + 1404.0, + 946.0, + 1404.0, + 977.0, + 811.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 974.0, + 1370.0, + 974.0, + 1370.0, + 1005.0, + 296.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1403.0, + 1909.0, + 1403.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 539.0, + 1942.0, + 539.0, + 1975.0, + 296.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1942.0, + 665.0, + 1942.0, + 665.0, + 1975.0, + 559.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1975.0, + 714.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1407.0, + 1969.0, + 1407.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1370, + 1403, + 1370, + 1403, + 1616, + 298, + 1616 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1043, + 1403, + 1043, + 1403, + 1259, + 298, + 1259 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1631, + 1402, + 1631, + 1402, + 1784, + 299, + 1784 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 814, + 1403, + 814, + 1403, + 1029, + 297, + 1029 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 303, + 1881, + 1402, + 1881, + 1402, + 2034, + 303, + 2034 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 296, + 589, + 1405, + 589, + 1405, + 743, + 296, + 743 + ], + "score": 0.965 + }, + { + "category_id": 3, + "poly": [ + 311, + 251, + 1392, + 251, + 1392, + 557, + 311, + 557 + ], + "score": 0.963 + }, + { + "category_id": 0, + "poly": [ + 300, + 1301, + 529, + 1301, + 529, + 1337, + 300, + 1337 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 300, + 1830, + 488, + 1830, + 488, + 1863, + 300, + 1863 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.777 + }, + { + "category_id": 13, + "poly": [ + 820, + 684, + 860, + 684, + 860, + 712, + 820, + 712 + ], + "score": 0.86, + "latex": "3 \\times" + }, + { + "category_id": 13, + "poly": [ + 884, + 653, + 924, + 653, + 924, + 681, + 884, + 681 + ], + "score": 0.85, + "latex": "3 \\times" + }, + { + "category_id": 13, + "poly": [ + 305, + 968, + 347, + 968, + 347, + 998, + 305, + 998 + ], + "score": 0.83, + "latex": "3 \\times" + }, + { + "category_id": 13, + "poly": [ + 1272, + 815, + 1326, + 815, + 1326, + 844, + 1272, + 844 + ], + "score": 0.59, + "latex": "4 . 9 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 591.0, + 1403.0, + 591.0, + 1403.0, + 624.0, + 295.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 621.0, + 1406.0, + 621.0, + 1406.0, + 655.0, + 293.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 883.0, + 649.0, + 883.0, + 689.0, + 294.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 689.0, + 925.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 683.0, + 819.0, + 683.0, + 819.0, + 716.0, + 296.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 683.0, + 1403.0, + 683.0, + 1403.0, + 716.0, + 861.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 709.0, + 480.0, + 709.0, + 480.0, + 747.0, + 295.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 315.0, + 959.0, + 315.0, + 959.0, + 330.0, + 941.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 433.0, + 1144.0, + 433.0, + 1144.0, + 443.0, + 1134.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 527.0, + 495.0, + 527.0, + 495.0, + 560.0, + 453.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 527.0, + 874.0, + 527.0, + 874.0, + 560.0, + 831.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 527.0, + 1250.0, + 527.0, + 1250.0, + 560.0, + 1208.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.75, + 465.0, + 708.75, + 465.0, + 708.75, + 477.0, + 691.75, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1300.0, + 534.0, + 1300.0, + 534.0, + 1344.0, + 292.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1829.0, + 491.0, + 1829.0, + 491.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1372.0, + 1405.0, + 1372.0, + 1405.0, + 1405.0, + 296.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 1404.0, + 1403.0, + 1404.0, + 1437.0, + 294.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1463.0, + 1403.0, + 1463.0, + 1403.0, + 1497.0, + 294.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1491.0, + 1405.0, + 1491.0, + 1405.0, + 1531.0, + 292.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1522.0, + 1405.0, + 1522.0, + 1405.0, + 1559.0, + 293.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1555.0, + 1404.0, + 1555.0, + 1404.0, + 1589.0, + 293.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1584.0, + 941.0, + 1584.0, + 941.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1042.0, + 1403.0, + 1042.0, + 1403.0, + 1079.0, + 293.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1076.0, + 1405.0, + 1076.0, + 1405.0, + 1110.0, + 294.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1104.0, + 1404.0, + 1104.0, + 1404.0, + 1142.0, + 293.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1137.0, + 1404.0, + 1137.0, + 1404.0, + 1168.0, + 296.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1165.0, + 1405.0, + 1165.0, + 1405.0, + 1200.0, + 293.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 1407.0, + 1195.0, + 1407.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1228.0, + 790.0, + 1228.0, + 790.0, + 1262.0, + 293.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1630.0, + 1404.0, + 1630.0, + 1404.0, + 1665.0, + 295.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1661.0, + 1402.0, + 1661.0, + 1402.0, + 1694.0, + 294.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1693.0, + 1403.0, + 1693.0, + 1403.0, + 1725.0, + 293.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1721.0, + 1404.0, + 1721.0, + 1404.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1754.0, + 679.0, + 1754.0, + 679.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 814.0, + 1271.0, + 814.0, + 1271.0, + 849.0, + 295.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 814.0, + 1403.0, + 814.0, + 1403.0, + 849.0, + 1327.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 846.0, + 1404.0, + 846.0, + 1404.0, + 881.0, + 296.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 876.0, + 1404.0, + 876.0, + 1404.0, + 910.0, + 295.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 903.0, + 1404.0, + 903.0, + 1404.0, + 944.0, + 294.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 934.0, + 1407.0, + 934.0, + 1407.0, + 973.0, + 292.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 965.0, + 304.0, + 965.0, + 304.0, + 1005.0, + 292.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 965.0, + 1407.0, + 965.0, + 1407.0, + 1005.0, + 348.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 998.0, + 442.0, + 998.0, + 442.0, + 1037.0, + 291.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1882.0, + 1403.0, + 1882.0, + 1403.0, + 1915.0, + 298.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 324.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1979.0, + 323.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 323.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2038.0, + 323.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 324, + 228, + 1402, + 228, + 1402, + 353, + 324, + 353 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 302, + 372, + 1404, + 372, + 1404, + 464, + 302, + 464 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 296, + 482, + 1400, + 482, + 1400, + 547, + 296, + 547 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 298, + 839, + 1403, + 839, + 1403, + 933, + 298, + 933 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 298, + 565, + 1400, + 565, + 1400, + 629, + 298, + 629 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 295, + 758, + 1401, + 758, + 1401, + 823, + 295, + 823 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 297, + 1175, + 1400, + 1175, + 1400, + 1241, + 297, + 1241 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 294, + 1747, + 1404, + 1747, + 1404, + 1840, + 294, + 1840 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 297, + 646, + 1404, + 646, + 1404, + 740, + 297, + 740 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 299, + 951, + 1397, + 951, + 1397, + 1017, + 299, + 1017 + ], + "score": 0.87 + }, + { + "category_id": 1, + "poly": [ + 297, + 1258, + 1399, + 1258, + 1399, + 1323, + 297, + 1323 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 299, + 1583, + 1399, + 1583, + 1399, + 1648, + 299, + 1648 + ], + "score": 0.854 + }, + { + "category_id": 1, + "poly": [ + 294, + 1971, + 1398, + 1971, + 1398, + 2034, + 294, + 2034 + ], + "score": 0.85 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.846 + }, + { + "category_id": 1, + "poly": [ + 298, + 1338, + 1399, + 1338, + 1399, + 1404, + 298, + 1404 + ], + "score": 0.844 + }, + { + "category_id": 1, + "poly": [ + 299, + 1501, + 1399, + 1501, + 1399, + 1567, + 299, + 1567 + ], + "score": 0.844 + }, + { + "category_id": 1, + "poly": [ + 295, + 1665, + 1402, + 1665, + 1402, + 1730, + 295, + 1730 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 298, + 1420, + 1399, + 1420, + 1399, + 1486, + 298, + 1486 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 296, + 1033, + 1404, + 1033, + 1404, + 1159, + 296, + 1159 + ], + "score": 0.797 + }, + { + "category_id": 1, + "poly": [ + 297, + 1859, + 1404, + 1859, + 1404, + 1953, + 297, + 1953 + ], + "score": 0.722 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 321.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1404.0, + 259.0, + 1404.0, + 298.0, + 321.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 327.0, + 321.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 319.0, + 496.0, + 319.0, + 496.0, + 359.0, + 323.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 370.0, + 1404.0, + 370.0, + 1404.0, + 410.0, + 296.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 402.0, + 1404.0, + 402.0, + 1404.0, + 438.0, + 322.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 428.0, + 400.0, + 428.0, + 400.0, + 469.0, + 320.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 524.0, + 294.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 514.0, + 1310.0, + 514.0, + 1310.0, + 551.0, + 324.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 839.0, + 1402.0, + 839.0, + 1402.0, + 876.0, + 294.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 910.0, + 318.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 899.0, + 399.0, + 899.0, + 399.0, + 936.0, + 320.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 562.0, + 1405.0, + 562.0, + 1405.0, + 605.0, + 293.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 594.0, + 1402.0, + 594.0, + 1402.0, + 630.0, + 323.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 760.0, + 1403.0, + 760.0, + 1403.0, + 796.0, + 295.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 789.0, + 1294.0, + 789.0, + 1294.0, + 824.0, + 322.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1212.0, + 296.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1208.0, + 871.0, + 1208.0, + 871.0, + 1242.0, + 323.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1748.0, + 1402.0, + 1748.0, + 1402.0, + 1781.0, + 294.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1779.0, + 1402.0, + 1779.0, + 1402.0, + 1809.0, + 325.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1809.0, + 514.0, + 1809.0, + 514.0, + 1841.0, + 323.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 643.0, + 1405.0, + 643.0, + 1405.0, + 686.0, + 292.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 676.0, + 1405.0, + 676.0, + 1405.0, + 716.0, + 319.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 709.0, + 491.0, + 709.0, + 491.0, + 740.0, + 324.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 946.0, + 1403.0, + 946.0, + 1403.0, + 994.0, + 292.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 984.0, + 1076.0, + 984.0, + 1076.0, + 1018.0, + 321.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1258.0, + 1405.0, + 1258.0, + 1405.0, + 1294.0, + 296.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1288.0, + 947.0, + 1288.0, + 947.0, + 1324.0, + 323.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1582.0, + 1404.0, + 1582.0, + 1404.0, + 1621.0, + 295.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1616.0, + 1159.0, + 1616.0, + 1159.0, + 1649.0, + 323.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1969.0, + 1403.0, + 1969.0, + 1403.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 815.0, + 2003.0, + 815.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1339.0, + 1403.0, + 1339.0, + 1403.0, + 1375.0, + 296.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1372.0, + 1386.0, + 1372.0, + 1386.0, + 1405.0, + 326.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1403.0, + 1500.0, + 1403.0, + 1538.0, + 294.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1535.0, + 1154.0, + 1535.0, + 1154.0, + 1568.0, + 325.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1404.0, + 1665.0, + 1404.0, + 1701.0, + 294.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1696.0, + 935.0, + 1696.0, + 935.0, + 1731.0, + 323.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1419.0, + 1403.0, + 1419.0, + 1403.0, + 1457.0, + 293.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1452.0, + 1155.0, + 1452.0, + 1155.0, + 1488.0, + 324.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1031.0, + 1405.0, + 1031.0, + 1405.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1065.0, + 1404.0, + 1065.0, + 1404.0, + 1098.0, + 324.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1097.0, + 1404.0, + 1097.0, + 1404.0, + 1130.0, + 323.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1127.0, + 870.0, + 1127.0, + 870.0, + 1160.0, + 323.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1895.0, + 294.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1889.0, + 1405.0, + 1889.0, + 1405.0, + 1927.0, + 321.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1921.0, + 485.0, + 1921.0, + 485.0, + 1956.0, + 324.0, + 1956.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 622, + 1400, + 622, + 1400, + 686, + 299, + 686 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 298, + 509, + 1400, + 509, + 1400, + 602, + 298, + 602 + ], + "score": 0.807 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1402, + 229, + 1402, + 294, + 293, + 294 + ], + "score": 0.763 + }, + { + "category_id": 1, + "poly": [ + 293, + 426, + 1401, + 426, + 1401, + 489, + 293, + 489 + ], + "score": 0.733 + }, + { + "category_id": 1, + "poly": [ + 299, + 311, + 1399, + 311, + 1399, + 407, + 299, + 407 + ], + "score": 0.668 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 857, + 74, + 857, + 105, + 298, + 105 + ], + "score": 0.467 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.371 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 622.0, + 1404.0, + 622.0, + 1404.0, + 658.0, + 295.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 653.0, + 1130.0, + 653.0, + 1130.0, + 687.0, + 322.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 507.0, + 1404.0, + 507.0, + 1404.0, + 547.0, + 294.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 540.0, + 1405.0, + 540.0, + 1405.0, + 576.0, + 321.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 571.0, + 601.0, + 571.0, + 601.0, + 605.0, + 324.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1405.0, + 225.0, + 1405.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 262.0, + 1274.0, + 262.0, + 1274.0, + 295.0, + 321.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 422.0, + 1404.0, + 422.0, + 1404.0, + 464.0, + 293.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 458.0, + 810.0, + 458.0, + 810.0, + 491.0, + 320.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 314.0, + 1401.0, + 314.0, + 1401.0, + 348.0, + 295.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 343.0, + 1405.0, + 343.0, + 1405.0, + 381.0, + 323.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 375.0, + 1331.0, + 375.0, + 1331.0, + 409.0, + 325.0, + 409.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1109, + 1405, + 1109, + 1405, + 1235, + 297, + 1235 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 538, + 1404, + 538, + 1404, + 665, + 296, + 665 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 676, + 1404, + 676, + 1404, + 772, + 297, + 772 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 517, + 1494, + 1183, + 1494, + 1183, + 1583, + 517, + 1583 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 661, + 1270, + 1038, + 1270, + 1038, + 1357, + 661, + 1357 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 292, + 1602, + 1402, + 1602, + 1402, + 1669, + 292, + 1669 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 570, + 1702, + 1126, + 1702, + 1126, + 1774, + 570, + 1774 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 300, + 895, + 1400, + 895, + 1400, + 961, + 300, + 961 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 296, + 351, + 1402, + 351, + 1402, + 416, + 296, + 416 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 499, + 811, + 1198, + 811, + 1198, + 870, + 499, + 870 + ], + "score": 0.94 + }, + { + "category_id": 8, + "poly": [ + 493, + 456, + 1205, + 456, + 1205, + 513, + 493, + 513 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 299, + 1379, + 822, + 1379, + 822, + 1414, + 299, + 1414 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 857, + 73, + 857, + 106, + 297, + 106 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 298, + 1426, + 898, + 1426, + 898, + 1461, + 298, + 1461 + ], + "score": 0.923 + }, + { + "category_id": 8, + "poly": [ + 321, + 1018, + 1322, + 1018, + 1322, + 1077, + 321, + 1077 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 298, + 297, + 1385, + 297, + 1385, + 329, + 298, + 329 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1522, + 1400, + 1522, + 1400, + 1554, + 1352, + 1554 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1351, + 461, + 1400, + 461, + 1400, + 493, + 1351, + 493 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1352, + 817, + 1400, + 817, + 1400, + 848, + 1352, + 848 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1714, + 1400, + 1714, + 1400, + 1746, + 1351, + 1746 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1025, + 1400, + 1025, + 1400, + 1057, + 1353, + 1057 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1297, + 1400, + 1297, + 1400, + 1329, + 1351, + 1329 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 866, + 2087, + 866, + 2113, + 836, + 2113 + ], + "score": 0.861 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 450, + 227, + 450, + 261, + 300, + 261 + ], + "score": 0.834 + }, + { + "category_id": 13, + "poly": [ + 537, + 1141, + 712, + 1141, + 712, + 1175, + 537, + 1175 + ], + "score": 0.94, + "latex": "p ( Z | Z _ { R } = z _ { R } )" + }, + { + "category_id": 13, + "poly": [ + 496, + 540, + 622, + 540, + 622, + 573, + 496, + 573 + ], + "score": 0.93, + "latex": "q ( X , Z , Y )" + }, + { + "category_id": 13, + "poly": [ + 822, + 895, + 897, + 895, + 897, + 931, + 822, + 931 + ], + "score": 0.93, + "latex": "q ^ { 0 } ( X )" + }, + { + "category_id": 13, + "poly": [ + 790, + 677, + 967, + 677, + 967, + 712, + 790, + 712 + ], + "score": 0.93, + "latex": "D _ { \\mathrm { K L } } ( q | | p ) \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 1241, + 601, + 1404, + 601, + 1404, + 635, + 1241, + 635 + ], + "score": 0.93, + "latex": "\\overset { \\cdot } { q } ( Y | Z ; \\theta _ { 2 } ) =" + }, + { + "category_id": 13, + "poly": [ + 700, + 601, + 763, + 601, + 763, + 635, + 700, + 635 + ], + "score": 0.92, + "latex": "q ( X )" + }, + { + "category_id": 13, + "poly": [ + 584, + 1634, + 628, + 1634, + 628, + 1664, + 584, + 1664 + ], + "score": 0.92, + "latex": "x ^ { ( i ) }" + }, + { + "category_id": 14, + "poly": [ + 516, + 1491, + 1186, + 1491, + 1186, + 1586, + 516, + 1586 + ], + "score": 0.92, + "latex": "\\begin{array} { c } { { p ( X | Z = f _ { \\theta _ { 1 } } ( x ) ; \\theta _ { 1 } ^ { \\prime } ) = p ( X | X _ { R } = g _ { \\theta _ { 1 } ^ { \\prime } } ( f _ { \\theta _ { 1 } } ( x ) ) ) } } \\\\ { { p ( Z | Y = f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) ; \\theta _ { 2 } ^ { \\prime } ) = p ( Z | Z _ { R } = g _ { \\theta _ { 2 } ^ { \\prime } } ( f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( x ) ) ) , } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 899, + 570, + 1050, + 570, + 1050, + 603, + 899, + 603 + ], + "score": 0.92, + "latex": "Z = f _ { \\boldsymbol { \\theta } _ { 1 } } ( { \\boldsymbol { X } } )" + }, + { + "category_id": 13, + "poly": [ + 450, + 898, + 513, + 898, + 513, + 930, + 450, + 930 + ], + "score": 0.91, + "latex": "q ( X )" + }, + { + "category_id": 13, + "poly": [ + 1106, + 571, + 1254, + 571, + 1254, + 603, + 1106, + 603 + ], + "score": 0.91, + "latex": "Y = f _ { \\theta _ { 2 } } ( Z )" + }, + { + "category_id": 14, + "poly": [ + 570, + 1702, + 1130, + 1702, + 1130, + 1777, + 570, + 1777 + ], + "score": 0.91, + "latex": "\\operatorname * { m i n } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\sum _ { i } L _ { L 2 } ( x ^ { ( i ) } , x _ { R } ^ { ( i ) } ) + L _ { L 2 } ( z ^ { ( i ) } , z _ { R } ^ { ( i ) } ) \\ : ." + }, + { + "category_id": 14, + "poly": [ + 491, + 457, + 1207, + 457, + 1207, + 515, + 491, + 515 + ], + "score": 0.91, + "latex": "\\operatorname* { m a x } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\mathbb { E } _ { q ( X , Z , Y ) } \\left[ \\log q ( X | Z ) \\right] + \\mathbb { E } _ { q ( X , Z , Y ) } \\left[ \\log q ( Z | Y ) \\right] ." + }, + { + "category_id": 13, + "poly": [ + 298, + 633, + 468, + 633, + 468, + 666, + 298, + 666 + ], + "score": 0.91, + "latex": "\\delta ( Y - f _ { \\theta _ { 2 } } ( Z ) )" + }, + { + "category_id": 13, + "poly": [ + 298, + 601, + 425, + 601, + 425, + 633, + 298, + 633 + ], + "score": 0.9, + "latex": "q ( X , Z , Y )" + }, + { + "category_id": 13, + "poly": [ + 1239, + 711, + 1285, + 711, + 1285, + 743, + 1239, + 743 + ], + "score": 0.9, + "latex": "p ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 662, + 1268, + 1042, + 1268, + 1042, + 1359, + 662, + 1359 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { p ( { \\cal X } | { \\cal X } _ { R } = x _ { R } ) = N ( x _ { R } , \\sigma _ { 0 } ^ { 2 } { \\bf I } ) } \\\\ { p ( { \\cal Z } | { \\cal Z } _ { R } = z _ { R } ) = N ( z _ { R } , \\sigma _ { 0 } ^ { 2 } { \\bf I } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1141, + 482, + 1141, + 482, + 1175, + 297, + 1175 + ], + "score": 0.9, + "latex": "p ( X | X _ { R } = x _ { R } )" + }, + { + "category_id": 13, + "poly": [ + 833, + 603, + 1184, + 603, + 1184, + 635, + 833, + 635 + ], + "score": 0.89, + "latex": "q ( Z | X ; \\theta _ { 1 } ) = \\delta ( Z - f _ { \\theta _ { 1 } } ( X ) )" + }, + { + "category_id": 14, + "poly": [ + 499, + 812, + 1199, + 812, + 1199, + 870, + 499, + 870 + ], + "score": 0.89, + "latex": "\\operatorname* { m a x } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\mathbb { E } _ { q ( X ) } \\left[ \\log p ( X | Z ; \\theta _ { 1 } ^ { \\prime } ) \\right] + \\mathbb { E } _ { q ( X ) } \\left[ \\log p ( Z | Y ; \\theta _ { 2 } ^ { \\prime } ) \\right] ." + }, + { + "category_id": 14, + "poly": [ + 317, + 1019, + 1336, + 1019, + 1336, + 1080, + 317, + 1080 + ], + "score": 0.89, + "latex": "\\operatorname* { m a x } _ { \\{ \\theta _ { 1 } , \\theta _ { 1 } ^ { \\prime } , \\theta _ { 2 } , \\theta _ { 2 } ^ { \\prime } \\} } \\mathbb { E } _ { q ^ { 0 } ( X ) } \\left[ \\log p ( X | Z = f _ { \\theta _ { 1 } } ( X ) ; \\theta _ { 1 } ^ { \\prime } ) \\right] + \\mathbb { E } _ { q ^ { 0 } ( X ) } \\left[ \\log p ( Z | Y = f _ { \\theta _ { 2 } } ( f _ { \\theta _ { 1 } } ( X ) ) ; \\theta _ { 2 } ^ { \\prime } ) \\right] ~ ." + }, + { + "category_id": 13, + "poly": [ + 952, + 1176, + 987, + 1176, + 987, + 1202, + 952, + 1202 + ], + "score": 0.88, + "latex": "z _ { R }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 1116, + 1034, + 1116, + 1034, + 1141, + 1000, + 1141 + ], + "score": 0.88, + "latex": "z _ { R }" + }, + { + "category_id": 13, + "poly": [ + 865, + 1176, + 903, + 1176, + 903, + 1203, + 865, + 1203 + ], + "score": 0.88, + "latex": "x _ { R }" + }, + { + "category_id": 13, + "poly": [ + 916, + 1116, + 953, + 1116, + 953, + 1141, + 916, + 1141 + ], + "score": 0.87, + "latex": "x _ { R }" + }, + { + "category_id": 13, + "poly": [ + 554, + 1379, + 814, + 1379, + 814, + 1415, + 554, + 1415 + ], + "score": 0.86, + "latex": "- \\log p ( \\cdot | \\cdot ) \\propto L _ { L 2 } ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1179, + 331, + 1179, + 331, + 1202, + 298, + 1202 + ], + "score": 0.85, + "latex": "z _ { R }" + }, + { + "category_id": 13, + "poly": [ + 1315, + 1145, + 1353, + 1145, + 1353, + 1172, + 1315, + 1172 + ], + "score": 0.85, + "latex": "x _ { R }" + }, + { + "category_id": 13, + "poly": [ + 621, + 1112, + 647, + 1112, + 647, + 1138, + 621, + 1138 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1108, + 542, + 1132, + 542, + 1132, + 568, + 1108, + 568 + ], + "score": 0.83, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 995, + 1147, + 1014, + 1147, + 1014, + 1169, + 995, + 1169 + ], + "score": 0.82, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 484, + 931, + 508, + 931, + 508, + 956, + 484, + 956 + ], + "score": 0.81, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1188, + 542, + 1214, + 542, + 1214, + 568, + 1188, + 568 + ], + "score": 0.8, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1261, + 686, + 1279, + 686, + 1279, + 710, + 1261, + 710 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 924, + 1147, + 943, + 1147, + 943, + 1169, + 924, + 1169 + ], + "score": 0.8, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1333, + 685, + 1350, + 685, + 1350, + 711, + 1333, + 711 + ], + "score": 0.79, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 570, + 1116, + 590, + 1116, + 590, + 1138, + 570, + 1138 + ], + "score": 0.79, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 549, + 633, + 566, + 633, + 566, + 659, + 549, + 659 + ], + "score": 0.77, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 758, + 1176, + 775, + 1176, + 775, + 1200, + 758, + 1200 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 689, + 1178, + 708, + 1178, + 708, + 1199, + 689, + 1199 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 371, + 930, + 398, + 930, + 398, + 956, + 371, + 956 + ], + "score": 0.61, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 410, + 930, + 433, + 930, + 433, + 956, + 410, + 956 + ], + "score": 0.53, + "latex": "Z" + }, + { + "category_id": 13, + "poly": [ + 370, + 929, + 434, + 929, + 434, + 958, + 370, + 958 + ], + "score": 0.34, + "latex": "X , Z" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 297.0, + 1388.0, + 297.0, + 1388.0, + 329.0, + 296.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 227.0, + 453.0, + 227.0, + 453.0, + 265.0, + 298.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1110.0, + 569.0, + 1110.0, + 569.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1110.0, + 620.0, + 1110.0, + 620.0, + 1146.0, + 591.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1110.0, + 915.0, + 1110.0, + 915.0, + 1146.0, + 648.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1110.0, + 999.0, + 1110.0, + 999.0, + 1146.0, + 954.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1110.0, + 1405.0, + 1110.0, + 1405.0, + 1146.0, + 1035.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1141.0, + 296.0, + 1141.0, + 296.0, + 1176.0, + 291.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1141.0, + 536.0, + 1141.0, + 536.0, + 1176.0, + 483.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1141.0, + 923.0, + 1141.0, + 923.0, + 1176.0, + 713.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1141.0, + 994.0, + 1141.0, + 994.0, + 1176.0, + 944.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1141.0, + 1314.0, + 1141.0, + 1314.0, + 1176.0, + 1015.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1141.0, + 1406.0, + 1141.0, + 1406.0, + 1176.0, + 1354.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1173.0, + 297.0, + 1173.0, + 297.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1173.0, + 688.0, + 1173.0, + 688.0, + 1205.0, + 332.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1173.0, + 757.0, + 1173.0, + 757.0, + 1205.0, + 709.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1173.0, + 864.0, + 1173.0, + 864.0, + 1205.0, + 776.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1173.0, + 951.0, + 1173.0, + 951.0, + 1205.0, + 904.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1173.0, + 1405.0, + 1173.0, + 1405.0, + 1205.0, + 988.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1201.0, + 900.0, + 1201.0, + 900.0, + 1236.0, + 294.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 538.0, + 495.0, + 538.0, + 495.0, + 574.0, + 293.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 538.0, + 1107.0, + 538.0, + 1107.0, + 574.0, + 623.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 538.0, + 1187.0, + 538.0, + 1187.0, + 574.0, + 1133.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 538.0, + 1405.0, + 538.0, + 1405.0, + 574.0, + 1215.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 570.0, + 898.0, + 570.0, + 898.0, + 606.0, + 294.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 570.0, + 1105.0, + 570.0, + 1105.0, + 606.0, + 1051.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 570.0, + 1404.0, + 570.0, + 1404.0, + 606.0, + 1255.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 600.0, + 297.0, + 600.0, + 297.0, + 637.0, + 294.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 600.0, + 699.0, + 600.0, + 699.0, + 637.0, + 426.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 600.0, + 832.0, + 600.0, + 832.0, + 637.0, + 764.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 600.0, + 1240.0, + 600.0, + 1240.0, + 637.0, + 1185.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 630.0, + 548.0, + 630.0, + 548.0, + 667.0, + 469.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 630.0, + 898.0, + 630.0, + 898.0, + 667.0, + 567.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 675.0, + 789.0, + 675.0, + 789.0, + 717.0, + 292.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 675.0, + 1260.0, + 675.0, + 1260.0, + 717.0, + 968.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 675.0, + 1332.0, + 675.0, + 1332.0, + 717.0, + 1280.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 675.0, + 1406.0, + 675.0, + 1406.0, + 717.0, + 1351.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 708.0, + 1238.0, + 708.0, + 1238.0, + 746.0, + 295.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 708.0, + 1407.0, + 708.0, + 1407.0, + 746.0, + 1286.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 741.0, + 562.0, + 741.0, + 562.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1602.0, + 1405.0, + 1602.0, + 1405.0, + 1638.0, + 295.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1632.0, + 583.0, + 1632.0, + 583.0, + 1672.0, + 292.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1632.0, + 746.0, + 1632.0, + 746.0, + 1672.0, + 629.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 892.0, + 449.0, + 892.0, + 449.0, + 932.0, + 295.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 892.0, + 821.0, + 892.0, + 821.0, + 932.0, + 514.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 892.0, + 1401.0, + 892.0, + 1401.0, + 932.0, + 898.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 927.0, + 369.0, + 927.0, + 369.0, + 963.0, + 295.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 927.0, + 483.0, + 927.0, + 483.0, + 963.0, + 435.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 927.0, + 669.0, + 927.0, + 669.0, + 963.0, + 509.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 350.0, + 1404.0, + 350.0, + 1404.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 384.0, + 1387.0, + 384.0, + 1387.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 553.0, + 1374.0, + 553.0, + 1420.0, + 294.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1374.0, + 824.0, + 1374.0, + 824.0, + 1420.0, + 815.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1420.0, + 897.0, + 1420.0, + 897.0, + 1466.0, + 296.0, + 1466.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 286, + 1405, + 286, + 1405, + 623, + 297, + 623 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 391, + 759, + 1308, + 759, + 1308, + 1200, + 391, + 1200 + ], + "score": 0.98, + "html": "
set No.1 (# training examples per class)1020501002005001k2k (all) 5k
ladder network model; Figure 325.8516.489.266.004.663.072.151.260.91
proposed-perturb (semantic); Figure 1(c)19.7612.338.776.064.592.931.871.310.93
set No.2 (# training examples per class)1020501002005001k2k
ladder network model; Figure 333.1417.4610.446.674.432.821.941.37
proposed-perturb (semantic); Figure 1(c)23.3615.359.435.754.432.991.871.39
set No.3 (# training examples per class)1020501002005001k2k
ladder network model; Figure 329.9916.999.737.344.393.002.121.47
proposed-perturb (semantic); Figure 1(c)23.2113.988.836.514.322.942.221.49
mean over 3 random trials1020501002005001k2k (all) 5k
ladder network model; Figure 329.6616.989.816.674.492.962.071.370.91
proposed-perturb (semantic); Figure 1(c)22.1113.899.016.114.452.951.991.400.93
" + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 856, + 76, + 856, + 104, + 298, + 104 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 6, + "poly": [ + 295, + 655, + 1399, + 655, + 1399, + 718, + 295, + 718 + ], + "score": 0.839 + }, + { + "category_id": 0, + "poly": [ + 342, + 232, + 1306, + 232, + 1306, + 261, + 342, + 261 + ], + "score": 0.802 + }, + { + "category_id": 0, + "poly": [ + 302, + 231, + 1329, + 231, + 1329, + 262, + 302, + 262 + ], + "score": 0.417 + }, + { + "category_id": 13, + "poly": [ + 1233, + 592, + 1323, + 592, + 1323, + 621, + 1233, + 621 + ], + "score": 0.89, + "latex": "2 9 . 6 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 637, + 592, + 726, + 592, + 726, + 621, + 637, + 621 + ], + "score": 0.88, + "latex": "2 2 . 1 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 848, + 657, + 874, + 657, + 874, + 684, + 848, + 684 + ], + "score": 0.8, + "latex": "\\%" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 655.0, + 847.0, + 655.0, + 847.0, + 690.0, + 295.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 655.0, + 1404.0, + 655.0, + 1404.0, + 690.0, + 875.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 685.0, + 1092.0, + 685.0, + 1092.0, + 721.0, + 295.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 231.0, + 1313.0, + 231.0, + 1313.0, + 266.0, + 335.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1334.0, + 229.0, + 1334.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 1405.0, + 287.0, + 1405.0, + 322.0, + 295.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1406.0, + 317.0, + 1406.0, + 351.0, + 295.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 350.0, + 1402.0, + 350.0, + 1402.0, + 381.0, + 296.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 378.0, + 1403.0, + 378.0, + 1403.0, + 411.0, + 292.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 410.0, + 1405.0, + 410.0, + 1405.0, + 445.0, + 297.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 440.0, + 1405.0, + 440.0, + 1405.0, + 474.0, + 294.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 471.0, + 1405.0, + 471.0, + 1405.0, + 505.0, + 295.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 499.0, + 1405.0, + 499.0, + 1405.0, + 535.0, + 292.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 531.0, + 1403.0, + 531.0, + 1403.0, + 566.0, + 295.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 560.0, + 1405.0, + 560.0, + 1405.0, + 597.0, + 294.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 591.0, + 636.0, + 591.0, + 636.0, + 626.0, + 295.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 591.0, + 1232.0, + 591.0, + 1232.0, + 626.0, + 727.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 591.0, + 1343.0, + 591.0, + 1343.0, + 626.0, + 1324.0, + 626.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 344, + 631, + 1352, + 631, + 1352, + 1717, + 344, + 1717 + ], + "score": 0.982, + "html": "
Set No.1(#train examples per class)
feed-forward model; Figure 2(a)10 22.6120 14.2050 11.25100 6.37200 4.34500 2.631k 1.832k 1.56(all) 5k 1.04
joint learning model with recon-one; Figure 2(b)18.6912.217.845.174.022.581.791.471.12
joint learning model with recon-one with denoising constraints20.3911.917.414.643.652.571.971.530.97
joint learning model with recon-all; Figure 2(b)18.8212.829.346.435.234.122.682.421.87
joint learning model with recon-all with denoisingconstraints17.9311.767.324.783.913.042.521.991.36
proposed-base; Figure 1(b)20.2310.186.473.893.041.891.330.910.80
proposed-base with denoising constraints19.8810.896.624.263.402.442.111.541.13
proposed-perturb (random); Figure 1(c)18.3810.586.643.783.141.901.210.890.65
proposed-perturb (semantic); Figure 1(c)19.339.725.983.472.841.841.160.840.62
Set No.2 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)28.8417.3610.146.204.783.021.611.41
joint learning model with recon-one; Figure 2(b)26.0914.407.985.184.172.291.941.52
joint learning model with recon-one with denoising constraints27.6913.116.955.073.542.371.831.28
joint learning model with recon-all; Figure 2(b)24.0114.138.986.845.443.512.982.18
joint learning model with recon-all with denoising constraints23.0513.297.795.123.923.012.271.84
proposed-base; Figure 1(b)22.9512.986.274.433.222.141.370.96
proposed-base with denoising constraints26.9612.216.454.623.132.531.881.49
proposed-perturb (random); Figure 1(c)22.1012.525.974.262.861.941.230.92
proposed-perturb (semantic); Figure 1(c)21.2211.525.753.912.611.731.140.89
Set No.3 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)22.2016.439.677.165.023.172.251.39
joint learning model with recon-one; Figure 2(b)20.2314.197.735.964.222.621.791.35
joint learning model with recon-one with denoising constraints19.3212.257.445.393.582.371.491.56
joint learning model with recon-al; Figure 2(b)17.5114.129.127.045.494.053.082.25
joint learning model with recon-all with denoising constraints17.0712.507.865.484.052.972.021.98
proposed-base; Figure 1(b)20.8611.796.254.632.961.911.160.96
proposed-base with denoising constraints19.8911.306.264.573.502.631.611.47
proposed-perturb (random); Figure 1(c)20.0211.946.124.323.131.811.281.08
proposed-perturb (semantic); Figure 1(c)19.7810.536.034.002.701.761.140.92
" + }, + { + "category_id": 1, + "poly": [ + 297, + 286, + 1404, + 286, + 1404, + 503, + 297, + 503 + ], + "score": 0.979 + }, + { + "category_id": 0, + "poly": [ + 300, + 230, + 681, + 230, + 681, + 262, + 300, + 262 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 6, + "poly": [ + 293, + 533, + 1400, + 533, + 1400, + 596, + 293, + 596 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 104, + 298, + 104 + ], + "score": 0.785 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 105, + 298, + 105 + ], + "score": 0.143 + }, + { + "category_id": 13, + "poly": [ + 870, + 535, + 897, + 535, + 897, + 562, + 870, + 562 + ], + "score": 0.8, + "latex": "\\%" + }, + { + "category_id": 13, + "poly": [ + 803, + 288, + 849, + 288, + 849, + 317, + 803, + 317 + ], + "score": 0.4, + "latex": "5 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 896, + 288, + 942, + 288, + 942, + 317, + 896, + 317 + ], + "score": 0.39, + "latex": "4 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 686.0, + 230.0, + 686.0, + 269.0, + 296.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 530.0, + 869.0, + 530.0, + 869.0, + 570.0, + 293.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 530.0, + 1405.0, + 530.0, + 1405.0, + 570.0, + 898.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 563.0, + 618.0, + 563.0, + 618.0, + 599.0, + 295.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 802.0, + 287.0, + 802.0, + 321.0, + 295.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 287.0, + 895.0, + 287.0, + 895.0, + 321.0, + 850.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 287.0, + 1404.0, + 287.0, + 1404.0, + 321.0, + 943.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 354.0, + 297.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 345.0, + 1404.0, + 345.0, + 1404.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 378.0, + 1405.0, + 378.0, + 1405.0, + 413.0, + 296.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 408.0, + 1404.0, + 408.0, + 1404.0, + 443.0, + 294.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 440.0, + 1406.0, + 440.0, + 1406.0, + 475.0, + 295.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 469.0, + 1297.0, + 469.0, + 1297.0, + 506.0, + 294.0, + 506.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 325, + 321, + 1373, + 321, + 1373, + 1411, + 325, + 1411 + ], + "score": 0.981, + "html": "
Set No.1(# train examples per class)1020501002005001k2k(all) 4k
feed-forward model; Figure 2(a)73.3069.2562.4255.6547.7134.3027.0421.0617.80
joint learning model with recon-one; Figure 2(b)75.1970.3862.2555.3046.8934.1226.6321.0517.68
joint learning model with recon-one with denoising constraints73.7268.2061.9955.2346.6436.3729.7825.5321.73
joint learning model with recon-all; Figure 2(b)74.7968.3362.9256.2451.3740.3030.9126.4922.71
joint learning model with recon-all with denoising constraints76.5669.6764.5357.8852.7442.2436.9030.9327.41
proposed-base; Figure 1(b)70.7966.5759.9152.9843.2932.2526.1920.9217.45
proposed-base with denoising constraints71.0367.4960.3753.5244.2833.4028.0025.0621.34
proposed-perturb (random); Figure 1(c)71.8967.1259.2252.7943.8731.8225.0420.9717.43
proposed-perturb (semantic); Figure 1(c)71.5966.9058.6452.3442.7430.9424.4520.1016.16
Set No.2 (# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)72.3969.4960.4554.8546.9133.3926.7321.00
joint learning model with recon-one; Figure 2(b)74.0669.1460.7154.5445.7033.5427.4320.90
joint learning model with recon-one with denoising constraints76.4069.3360.2855.3847.4036.2929.3124.60
joint learning model with recon-all; Figure 2(b)72.2867.6061.5356.6549.9942.0832.9926.33
joint learning model with recon-all with denoising constraints73.9069.2361.9057.9952.3545.1237.2330.14
proposed-base; Figure 1(b)72.4965.6257.8252.6643.2032.2425.6021.32
proposed-base with denoising constraints72.9966.7557.7853.8144.3333.5628.4025.03
proposed-perturb (random); Figure 1(c)71.8465.9858.0853.3743.4431.5625.6921.03
proposed-perturb (semantic); Figure 1(c)72.8566.6557.4452.2142.7431.1724.9920.54
Set No.3(# train examples per class)1020501002005001k2k
feed-forward model; Figure 2(a)75.7868.2461.0254.2946.2833.3826.1120.85
joint learning model with recon-one; Figure 2(b)77.7967.6261.3755.2245.9633.2126.2921.81
joint learning model with recon-one with denoising constraints76.6069.2761.1355.1047.5037.1229.6324.88
joint learning model with recon-all; Figure 2(b)72.9266.9763.3156.2350.1641.4133.7526.31
joint learning model with recon-all with denoising constraints76.8368.5365.5858.2952.4345.4239.0132.32
proposed-base; Figure 1(b)71.6066.3158.9952.3043.8831.1025.4820.95
proposed-base with denoising constraints72.3967.2060.6052.6444.6233.5228.0125.25
proposed-perturb (random); Figure 1(c)71.3467.1559.5552.8643.8132.0125.7820.42
proposed-perturb (semantic); Figure 1(c)70.0667.0758.8352.4143.4730.6125.0019.94
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 298, + 224, + 1400, + 224, + 1400, + 287, + 298, + 287 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 13, + "poly": [ + 870, + 226, + 897, + 226, + 897, + 253, + 870, + 253 + ], + "score": 0.78, + "latex": "\\%" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 869.0, + 221.0, + 869.0, + 261.0, + 294.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 221.0, + 1405.0, + 221.0, + 1405.0, + 261.0, + 898.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 253.0, + 649.0, + 253.0, + 649.0, + 289.0, + 296.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 314, + 400, + 1389, + 400, + 1389, + 1172, + 314, + 1172 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 297, + 1208, + 1406, + 1208, + 1406, + 1302, + 297, + 1302 + ], + "score": 0.93 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 294, + 286, + 1404, + 286, + 1404, + 351, + 294, + 351 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 688, + 229, + 688, + 262, + 300, + 262 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.845 + }, + { + "category_id": 15, + "poly": [ + 410.0, + 391.0, + 747.0, + 391.0, + 747.0, + 434.0, + 410.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 391.0, + 1299.0, + 391.0, + 1299.0, + 434.0, + 949.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 448.0, + 526.0, + 448.0, + 526.0, + 490.0, + 369.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 450.0, + 730.0, + 450.0, + 730.0, + 490.0, + 684.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 455.0, + 784.0, + 455.0, + 784.0, + 484.0, + 744.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 449.0, + 839.0, + 449.0, + 839.0, + 497.0, + 785.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 450.0, + 913.0, + 450.0, + 913.0, + 492.0, + 870.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 440.0, + 1379.0, + 440.0, + 1379.0, + 501.0, + 920.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 500.0, + 523.0, + 500.0, + 523.0, + 554.0, + 331.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 505.0, + 738.0, + 505.0, + 738.0, + 549.0, + 527.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 522.0, + 765.0, + 522.0, + 765.0, + 531.0, + 751.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 507.0, + 838.0, + 507.0, + 838.0, + 549.0, + 787.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 490.0, + 1343.0, + 490.0, + 1343.0, + 560.0, + 869.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 514.0, + 1370.0, + 514.0, + 1370.0, + 540.0, + 1344.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 551.0, + 778.0, + 551.0, + 778.0, + 618.0, + 368.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 563.0, + 833.0, + 563.0, + 833.0, + 603.0, + 788.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 558.0, + 916.0, + 558.0, + 916.0, + 605.0, + 870.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 548.0, + 1379.0, + 548.0, + 1379.0, + 616.0, + 919.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 588.0, + 492.0, + 588.0, + 492.0, + 673.0, + 355.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 609.0, + 586.0, + 609.0, + 586.0, + 661.0, + 523.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 606.0, + 743.0, + 606.0, + 743.0, + 663.0, + 674.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 618.0, + 836.0, + 618.0, + 836.0, + 657.0, + 793.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 611.0, + 916.0, + 611.0, + 916.0, + 658.0, + 870.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 619.0, + 962.0, + 619.0, + 962.0, + 656.0, + 925.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 614.0, + 1016.0, + 614.0, + 1016.0, + 656.0, + 974.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 619.0, + 1061.0, + 619.0, + 1061.0, + 651.0, + 1030.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 615.0, + 1121.0, + 615.0, + 1121.0, + 656.0, + 1080.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 614.0, + 1175.0, + 614.0, + 1175.0, + 655.0, + 1131.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 605.0, + 1333.0, + 605.0, + 1333.0, + 662.0, + 1182.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 620.0, + 1374.0, + 620.0, + 1374.0, + 650.0, + 1339.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 681.0, + 361.0, + 681.0, + 361.0, + 702.0, + 337.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 667.0, + 475.0, + 667.0, + 475.0, + 713.0, + 373.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 665.0, + 736.0, + 665.0, + 736.0, + 715.0, + 526.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 676.0, + 828.0, + 676.0, + 828.0, + 703.0, + 801.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 671.0, + 910.0, + 671.0, + 910.0, + 709.0, + 877.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 670.0, + 965.0, + 670.0, + 965.0, + 709.0, + 921.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 672.0, + 1016.0, + 672.0, + 1016.0, + 708.0, + 978.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 657.0, + 1390.0, + 657.0, + 1390.0, + 722.0, + 1022.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 714.0, + 845.0, + 714.0, + 845.0, + 777.0, + 318.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 708.0, + 1391.0, + 708.0, + 1391.0, + 781.0, + 857.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 791.0, + 746.0, + 791.0, + 746.0, + 834.0, + 411.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 792.0, + 1298.0, + 792.0, + 1298.0, + 832.0, + 951.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 854.0, + 565.0, + 854.0, + 565.0, + 875.0, + 543.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 854.0, + 905.0, + 854.0, + 905.0, + 886.0, + 879.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 849.0, + 1064.0, + 849.0, + 1064.0, + 878.0, + 1027.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 849.0, + 1120.0, + 849.0, + 1120.0, + 878.0, + 1080.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 855.0, + 1266.0, + 855.0, + 1266.0, + 869.0, + 1244.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 854.0, + 1373.0, + 854.0, + 1373.0, + 876.0, + 1341.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 897.0, + 1390.0, + 897.0, + 1390.0, + 1175.0, + 844.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 909.0, + 828.0, + 909.0, + 828.0, + 933.0, + 801.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 911.0, + 564.0, + 911.0, + 564.0, + 933.0, + 543.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 917.0, + 670.0, + 917.0, + 670.0, + 937.0, + 650.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 977.0, + 355.0, + 977.0, + 355.0, + 987.0, + 340.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 963.0, + 418.0, + 963.0, + 418.0, + 994.0, + 384.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 959.0, + 523.0, + 959.0, + 523.0, + 999.0, + 479.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 957.0, + 579.0, + 957.0, + 579.0, + 1004.0, + 531.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 963.0, + 624.0, + 963.0, + 624.0, + 998.0, + 583.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 971.0, + 676.0, + 971.0, + 676.0, + 994.0, + 646.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 963.0, + 728.0, + 963.0, + 728.0, + 998.0, + 691.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 971.0, + 900.0, + 971.0, + 900.0, + 988.0, + 887.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 966.0, + 1012.0, + 966.0, + 1012.0, + 997.0, + 978.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 958.0, + 1069.0, + 958.0, + 1069.0, + 1005.0, + 1022.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 958.0, + 1176.0, + 958.0, + 1176.0, + 1000.0, + 1073.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1012.0, + 368.0, + 1012.0, + 368.0, + 1055.0, + 331.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1012.0, + 415.0, + 1012.0, + 415.0, + 1053.0, + 376.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 991.0, + 628.0, + 991.0, + 628.0, + 1078.0, + 416.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1008.0, + 738.0, + 1008.0, + 738.0, + 1063.0, + 639.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 1018.0, + 827.0, + 1018.0, + 827.0, + 1044.0, + 803.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1013.0, + 916.0, + 1013.0, + 916.0, + 1055.0, + 874.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1008.0, + 1055.0, + 1008.0, + 1055.0, + 1021.0, + 1046.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1011.0, + 1262.0, + 1011.0, + 1262.0, + 1024.0, + 1239.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1075.0, + 363.0, + 1075.0, + 363.0, + 1099.0, + 334.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1070.0, + 413.0, + 1070.0, + 413.0, + 1104.0, + 384.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1062.0, + 471.0, + 1062.0, + 471.0, + 1112.0, + 426.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1060.0, + 517.0, + 1060.0, + 517.0, + 1102.0, + 486.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1063.0, + 572.0, + 1063.0, + 572.0, + 1108.0, + 533.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1073.0, + 904.0, + 1073.0, + 904.0, + 1099.0, + 886.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1040.0, + 1036.0, + 1040.0, + 1036.0, + 1115.0, + 956.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1125.0, + 364.0, + 1125.0, + 364.0, + 1156.0, + 331.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1124.0, + 416.0, + 1124.0, + 416.0, + 1159.0, + 378.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1132.0, + 564.0, + 1132.0, + 564.0, + 1153.0, + 543.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1123.0, + 915.0, + 1123.0, + 915.0, + 1164.0, + 875.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1115.0, + 1236.0, + 1115.0, + 1236.0, + 1169.0, + 962.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1125.0, + 1279.0, + 1125.0, + 1279.0, + 1158.0, + 1237.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1134.0, + 1320.0, + 1134.0, + 1320.0, + 1153.0, + 1290.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 439.0, + 639.0, + 439.0, + 639.0, + 495.0, + 518.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.25, + 456.5, + 673.25, + 456.5, + 673.25, + 477.0, + 635.25, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 614.0, + 500.0, + 614.0, + 500.0, + 653.5, + 495.0, + 653.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 621.5, + 676.0, + 621.5, + 676.0, + 655.0, + 645.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.25, + 614.5, + 376.25, + 614.5, + 376.25, + 651.5, + 320.25, + 651.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 623.5, + 776.0, + 623.5, + 776.0, + 646.5, + 743.0, + 646.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.75, + 616.5, + 639.75, + 616.5, + 639.75, + 652.5, + 567.75, + 652.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 673.5, + 523.0, + 673.5, + 523.0, + 701.5, + 480.0, + 701.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.25, + 908.0, + 1049.25, + 908.0, + 1049.25, + 938.0, + 1040.25, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 905.5, + 1112.0, + 905.5, + 1112.0, + 941.5, + 1086.0, + 941.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 905.5, + 915.0, + 905.5, + 915.0, + 938.0, + 873.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 915.0, + 1220.0, + 915.0, + 1220.0, + 932.0, + 1191.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 966.5, + 831.0, + 966.5, + 831.0, + 989.0, + 795.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1123.5, + 971.0, + 1123.5, + 971.0, + 1160.5, + 928.0, + 1160.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.25, + 1127.0, + 1376.25, + 1127.0, + 1376.25, + 1157.5, + 1331.25, + 1157.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1209.0, + 1404.0, + 1209.0, + 1404.0, + 1243.0, + 296.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1238.0, + 1404.0, + 1238.0, + 1404.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1271.0, + 458.0, + 1271.0, + 458.0, + 1303.0, + 292.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 692.0, + 230.0, + 692.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 286.0, + 1405.0, + 286.0, + 1405.0, + 322.0, + 296.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 319.0, + 1323.0, + 319.0, + 1323.0, + 353.0, + 293.0, + 353.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 297, + 74, + 857, + 74, + 857, + 105, + 297, + 105 + ], + "score": 0.908 + }, + { + "category_id": 3, + "poly": [ + 352, + 564, + 1356, + 564, + 1356, + 2018, + 352, + 2018 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 295, + 352, + 1402, + 352, + 1402, + 416, + 295, + 416 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.81 + }, + { + "category_id": 0, + "poly": [ + 300, + 230, + 689, + 230, + 689, + 262, + 300, + 262 + ], + "score": 0.755 + }, + { + "category_id": 1, + "poly": [ + 300, + 230, + 689, + 230, + 689, + 262, + 300, + 262 + ], + "score": 0.111 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 841.0, + 895.0, + 841.0, + 895.0, + 857.0, + 859.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 836.0, + 1253.0, + 836.0, + 1253.0, + 856.0, + 1230.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 852.0, + 540.0, + 852.0, + 540.0, + 878.0, + 512.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 851.0, + 1273.0, + 851.0, + 1273.0, + 878.0, + 1227.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1152.0, + 466.0, + 1152.0, + 466.0, + 1174.0, + 429.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1174.0, + 456.0, + 1174.0, + 456.0, + 1191.0, + 443.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1322.0, + 547.0, + 1322.0, + 547.0, + 1354.0, + 511.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1323.0, + 907.0, + 1323.0, + 907.0, + 1349.0, + 880.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1448.0, + 843.0, + 1448.0, + 843.0, + 1501.0, + 760.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1555.0, + 548.0, + 1555.0, + 548.0, + 1587.0, + 511.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1801.0, + 900.0, + 1801.0, + 900.0, + 1812.0, + 884.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 713.5, + 1173.0, + 713.5, + 1173.0, + 754.0, + 1080.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1590.5, + 452.0, + 1590.5, + 452.0, + 1649.5, + 445.0, + 1649.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1619.5, + 936.0, + 1619.5, + 936.0, + 1651.5, + 900.0, + 1651.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 232.0, + 691.0, + 232.0, + 691.0, + 268.0, + 298.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 1404.0, + 352.0, + 1404.0, + 388.0, + 295.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 385.0, + 877.0, + 385.0, + 877.0, + 417.0, + 297.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 232.0, + 691.0, + 232.0, + 691.0, + 268.0, + 298.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 354, + 201, + 1360, + 201, + 1360, + 1152, + 354, + 1152 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 297, + 1194, + 1404, + 1194, + 1404, + 1288, + 297, + 1288 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.842 + }, + { + "category_id": 4, + "poly": [ + 297, + 1194, + 1403, + 1194, + 1403, + 1288, + 297, + 1288 + ], + "score": 0.129 + }, + { + "category_id": 13, + "poly": [ + 734, + 1230, + 758, + 1230, + 758, + 1254, + 734, + 1254 + ], + "score": 0.77, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 382, + 1259, + 406, + 1259, + 406, + 1284, + 382, + 1284 + ], + "score": 0.76, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 250.0, + 795.0, + 250.0, + 795.0, + 258.0, + 784.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 367.0, + 1122.0, + 367.0, + 1122.0, + 375.0, + 1106.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 479.0, + 1243.0, + 479.0, + 1243.0, + 488.0, + 1233.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 502.0, + 501.0, + 502.0, + 501.0, + 516.0, + 485.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 507.0, + 545.0, + 507.0, + 545.0, + 527.0, + 494.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 531.0, + 883.0, + 531.0, + 883.0, + 543.0, + 858.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 601.0, + 486.0, + 601.0, + 486.0, + 610.0, + 472.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 594.0, + 913.0, + 594.0, + 913.0, + 610.0, + 881.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 716.0, + 549.0, + 716.0, + 549.0, + 729.0, + 526.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 768.0, + 1253.0, + 768.0, + 1253.0, + 776.0, + 1229.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 779.0, + 510.0, + 779.0, + 510.0, + 844.0, + 423.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 806.0, + 551.0, + 806.0, + 551.0, + 818.0, + 531.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 811.0, + 1186.0, + 811.0, + 1186.0, + 820.0, + 1173.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 944.0, + 560.0, + 944.0, + 560.0, + 997.0, + 488.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 974.0, + 1256.0, + 974.0, + 1256.0, + 986.0, + 1233.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1028.0, + 503.0, + 1028.0, + 503.0, + 1105.0, + 383.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1039.0, + 552.0, + 1039.0, + 552.0, + 1051.0, + 529.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 241.5, + 464.0, + 241.5, + 464.0, + 252.0, + 449.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 260.0, + 486.0, + 260.0, + 486.0, + 282.0, + 445.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 715.0, + 876.0, + 715.0, + 876.0, + 725.0, + 864.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.75, + 818.5, + 1204.75, + 818.5, + 1204.75, + 840.5, + 1162.75, + 840.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 982.5, + 539.0, + 982.5, + 539.0, + 999.0, + 497.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1194.0, + 1404.0, + 1194.0, + 1404.0, + 1231.0, + 295.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1226.0, + 733.0, + 1226.0, + 733.0, + 1260.0, + 295.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1226.0, + 1404.0, + 1226.0, + 1404.0, + 1260.0, + 759.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1256.0, + 381.0, + 1256.0, + 381.0, + 1289.0, + 295.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1256.0, + 1054.0, + 1256.0, + 1054.0, + 1289.0, + 407.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1194.0, + 1403.0, + 1194.0, + 1403.0, + 1231.0, + 295.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1226.0, + 733.0, + 1226.0, + 733.0, + 1260.0, + 295.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1226.0, + 1404.0, + 1226.0, + 1404.0, + 1260.0, + 759.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1257.0, + 381.0, + 1257.0, + 381.0, + 1288.0, + 295.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1257.0, + 1055.0, + 1257.0, + 1055.0, + 1288.0, + 407.0, + 1288.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh.md b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh.md new file mode 100644 index 0000000000000000000000000000000000000000..0cbf21ef7177b2d1ddbfea0212295c75f0b7361b --- /dev/null +++ b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh.md @@ -0,0 +1,346 @@ +# LEARNING WITH FEATURE-DEPENDENT LABEL NOISE:A PROGRESSIVE APPROACH + +Yikai $\mathbf { Z h a n g ^ { 1 * } }$ , Songzhu Zheng2∗, Pengxiang $\mathbf { W _ { u } } ^ { 1 * }$ , Mayank Goswami3, Chao Chen2 + +1Rutgers University, $\{ \mathtt { y z 4 2 2 } , \mathtt { p w 2 4 1 } \}$ @cs.rutgers.edu +2Stony Brook University, {zheng.songzhu,chao.chen.1}@stonybrook.edu +3City University of New York, mayank.isi@gmail.com + +# ABSTRACT + +Label noise is frequently observed in real-world large-scale datasets. The noise is introduced due to a variety of reasons; it is heterogeneous and feature-dependent. Most existing approaches to handling noisy labels fall into two categories: they either assume an ideal feature-independent noise, or remain heuristic without theoretical guarantees. In this paper, we propose to target a new family of featuredependent label noise, which is much more general than commonly used i.i.d. label noise and encompasses a broad spectrum of noise patterns. Focusing on this general noise family, we propose a progressive label correction algorithm that iteratively corrects labels and refines the model. We provide theoretical guarantees showing that for a wide variety of (unknown) noise patterns, a classifier trained with this strategy converges to be consistent with the Bayes classifier. In experiments, our method outperforms SOTA baselines and is robust to various noise types and levels. + +# 1 INTRODUCTION + +Addressing noise in training set labels is an important problem in supervised learning. Incorrect annotation of data is inevitable in large-scale data collection, due to intrinsic ambiguity of data/class and mistakes of human/automatic annotators (Yan et al., 2014; Andreas et al., 2017). Developing methods that are resilient to label noise is therefore crucial in real-life applications. + +Classical approaches take a rather simplistic i.i.d. assumption on the label noise, i.e., the label corruption is independent and identically distributed and thus is feature-independent. Methods based on this assumption either explicitly estimate the noise pattern (Reed et al., 2014; Patrini et al., 2017; Dan et al., 2019; Xu et al., 2019) or introduce extra regularizer/loss terms (Natarajan et al., 2013; Van Rooyen et al., 2015; Xiao et al., 2015; Zhang & Sabuncu, 2018; Ma et al., 2018; Arazo et al., 2019; Shen & Sanghavi, 2019). Some results prove that the commonly used losses are naturally robust against such i.i.d. label noise (Manwani & Sastry, 2013; Ghosh et al., 2015; Gao et al., 2016; Ghosh et al., 2017; Charoenphakdee et al., 2019; Hu et al., 2020). + +Although these methods come with theoretical guarantees, they usually do not perform as well as expected in practice due to the unrealistic i.i.d. assumption on noise. This is likely because label noise is heterogeneous and feature-dependent. A cat with an intrinsically ambiguous appearance is more likely to be mislabeled as a dog. An image with poor lighting or severe occlusion can be mislabeled, as important visual clues are imperceptible. Methods that can combat label noise of a much more general form are very much needed to address real-world challenges. + +To adapt to the heterogeneous label noise, state-of-the-arts (SOTAs) often resort to a data-recalibrating strategy. They progressively identify trustworthy data or correct data labels, and then train using these data (Tanaka et al., 2018; Wang et al., 2018; Lu et al., 2018; Li et al., 2019). The models gradually improve as more clean data are collected or more labels are corrected, eventually converging to models of high accuracy. These data-recalibrating methods best leverage the learning power of deep neural nets and achieve superior performance in practice. However, their underlying mechanism remains a mystery. No methods in this category can provide theoretical insights as to why the model can converge to an ideal one. Thus, these methods require careful hyperparameter tuning and are hard to generalize. + +In this paper, we propose a novel and principled method that specifically targets the heterogeneous, feature-dependent label noise. Unlike previous methods, we target a much more general family of noise, called Polynomial Margin Diminishing (PMD) label noise. In this noise family, we allow arbitrary noise level except for data far away from the true decision boundary. This is consistent with the real-world scenario; data near the decision boundary are harder to distinguish and more likely to be mislabeled. Meanwhile, a datum far away from the decision boundary is a typical example of its true class and should have a reasonably bounded noise level. + +Assuming this new PMD noise family, we propose a theoretically-guaranteed data-recalibrating algorithm that gradually corrects labels based on the noisy classifier’s confidence. We start from data points with high confidence, and correct the labels of these data using the predictions of the noisy classifier. Next, the model is improved using cleaned labels. We continue alternating the label correction and model improvement until it converges. See Figure 1 for an illustration. Our main theorem shows that with a theory-informed criterion for label correction at each iteration, the improvement of the label purity is guaranteed. Thus the model is guaranteed to improve with sufficient rate through iterations and eventually becomes consistent with the Bayes optimal classifier. + +Beside the theoretical strength, we also demonstrate the power of our method in practice. Our method outperforms others on CIFAR-10/100 with various synthetic noise patterns. We also evaluate our method against SOTAs on three real-world datasets with unknown noise patterns. + +To the best of our knowledge, our method is the first data-recalibrating method that is theoretically guaranteed to converge to an ideal model. The PMD noise family encompasses a broad spectrum of heterogeneous and feature-dependent noise, and better approximates the real-world scenario. It also provides a novel theoretical setting for the study of label noise. + +Related works. We review works that do not assume an i.i.d. label noise. Menon et al. (2018) generalized the work of (Ghosh et al., 2015) and provided an elegant theoretical framework, showing that loss functions fulfilling certain conditions naturally resist instance-dependent noise. The method can achieve even better theoretical properties (i.e., Bayes-consistency) with stronger assumption on the clean posterior probability $\eta$ . In practice, this method has not been extended to deep neural networks. Cheng et al. (2020) proposed an active learning method for instance-dependent label noise. + +![](images/e1c5bd95d246a39360bb9537b3313db61b0fbc917271c8e5928ac07345e50890.jpg) +Figure 1: Illustration of the algorithm using synthetic data. (a) Gaussian blob with clean label $( \eta ^ { * } ( \pmb { x } ) )$ . (b) Data with corrupted labels. (c) Final corrected data. Black dots are the data that have their clean labels. Red dots are the noisy data. Points that remain un-corrected are closer to the decision boundary. Our algorithm corrects most of the noise only using noisy classifier’s confidence. (d) Data after label correction. (e)-(h) We show the intermediate results at different iterations. Gray region is the area where the classifier has high confidence. Labels within this region are corrected. + +The algorithm iteratively queries clean labels from an oracle on carefully selected data. However, this approach is not applicable to settings where kosher annotations are unavailable. Another contemporary work (Chen et al., 2021) showed that the noise in real-world dataset is unlikely to be i.i.d., and proposed to fix the noisy labels by averaging the network predictions on each instance over the whole training process. While being effective, their method lacks theoretical guarantees. Chen et al. (2019) showed by regulating the topology of a classifier’s decision boundary, one can improve the model’s robustness against label noise. + +Data-recalibrating methods use noisy networks’ predictions to iteratively select/correct data and improve the models. Tanaka et al. (2018) introduced a joint training framework which simultaneously enforces the network to be consistent with its own predictions and corrects the noisy labels during training. Wang et al. (2018) identified noisy labels as outliers based on their label consistencies with surrounding data. Lu et al. (2018) used a curriculum learning strategy where the teacher net is trained on a small kosher dataset to determine if a datum is clean; then the learnt curriculum that gives the weight to each datum is fed into the student net for the training and inference. (Yu et al., 2019; Bo et al., 2018) trained two synchronized networks; the confidence and consistency of the two networks are utilized to identify clean data. Wu et al. (2020) selected the clean data by investigating the topological structures of the training data in the learned feature space. For completeness, we also refer to other methods of similar design (Li et al., 2017; Vahdat, 2017; Andreas et al., 2017; Ma et al., 2018; Thulasidasan et al., 2019; Arazo et al., 2019; Shu et al., 2019; Yi & Wu, 2019). + +As for theoretical guarantees, Ren et al. (2018) proposed an algorithm that iteratively re-weights each data point by solving an optimization problem. They proved the convergence of the training, but provided no guarantees that the model converges to an ideal one. Amid et al. (2019b) generalized the work of (Amid et al., 2019a) and proposed a tempered matching loss. They showed that when the final softmax layer is replaced by the bi-tempered loss, the resulting classifier will be Bayes consistent. Zheng et al. (2020) proved a one-shot guarantee for their data-recalibrating method; but the convergence of the model is not guaranteed. Our method is the first data-recalibrating method which is guaranteed to converge to a well-behaved classifier. + +# 2 METHOD + +We start by introducing the family of Poly-Margin Diminishing (PMD) label noise. In Section 2.2, we present our main algorithm. Finally, we prove the correctness of our algorithm in Section 3. + +Notations and preliminaries. Although the noise setting and algorithm naturally generalize to multiclass, for simplicity we focus on binary classification. Let the feature space be $\mathcal { X }$ . We assume the data $( { \pmb x } , y )$ is sampled from an underlying distribution $D$ on $\mathcal { X } \times \{ 0 , 1 \}$ . Define the posterior probability $\eta ( \pmb { x } ) = \mathbb { P } [ y = 1 \mid \pmb { x } ]$ . Let $\tau _ { 0 , 1 } ( \pmb { x } ) = \mathbb { P } [ \widetilde { y } = 1 \ | \ y = 0 , \pmb { x } ]$ and $\bar { \tau _ { 1 , 0 } } ( \pmb { x } ) = \mathbb { P } [ \widetilde { y } = 0 \ | \ y = 1 , \pmb { x } ]$ be the noise functions, where $\widetilde { y }$ e e denotes the corrupted label. For example, if a datum $_ { \textbf { \em x } }$ has true label $y = 0$ , it has $\tau _ { 0 , 1 } ( \pmb { x } )$ echance to be corrupted to 1. Similarly, it has $\tau _ { 1 , 0 } ( \pmb { x } )$ chance to be corrupted from 1 to 0. Let $\boldsymbol { \widetilde { \eta } } ( \boldsymbol { \dot { \mathbf { x } } } ) = \mathbb { P } [ \boldsymbol { \widetilde { y } } = 1 \mid \boldsymbol { \mathbf { x } } ]$ be the noisy posterior probability of $\widetilde y = 1$ given feature $_ { \textbf { \em x } }$ . Let $\eta ^ { * } ( \pmb { x } ) = \mathbb { I } _ { \{ \eta ( \pmb { x } ) \geq \frac { 1 } { 2 } \} }$ ebe the (clean) Bayes optimal classifier, where $\mathbb { I } _ { A }$ eequals 1 if $A$ is true, and 0 otherwise. Finally, let $f ( \pmb { x } ) : \mathcal { X } [ 0 , 1 ]$ be the classifier scoring function (the softmax output of a neural network in this paper). + +# 2.1 POLY-MARGIN DIMINISHING NOISE + +We first introduce the family of noise functions $\tau$ this paper will address. We introduce the concept of polynomial margin diminishing noise (PMD noise), which only upper bounds the noise $\tau$ in a certain level set of $\eta ( { \pmb x } )$ , thus allowing $\tau$ to be arbitrarily high outside the restricted domain. This formulation not only covers the feature-independent scenario but also generalizes scenarios proposed by (Du & Cai, 2015; Menon et al., 2018; Cheng et al., 2020). + +Definition 1 (PMD noise). $A$ pair of noise functions $\tau _ { 0 , 1 } ( \pmb { x } )$ and $\tau _ { 1 , 0 } ( \pmb { x } )$ are polynomial-margin diminishing (PMD), if there exist constants $t _ { 0 } \in ( 0 , \frac { 1 } { 2 } )$ , and $c _ { 1 } , c _ { 2 } > 0$ such that: + +$$ +\begin{array} { c l l } { \displaystyle \tau _ { 1 , 0 } ( { \pmb x } ) \leq c _ { 1 } [ 1 - \eta ( { \pmb x } ) ] ^ { 1 + c _ { 2 } } ; \forall \eta ( { \pmb x } ) \geq \displaystyle \frac { 1 } { 2 } + t _ { 0 } , a n d } \\ { \displaystyle \tau _ { 0 , 1 } ( { \pmb x } ) \leq c _ { 1 } \eta ( { \pmb x } ) ^ { 1 + c _ { 2 } } ; \forall \eta ( { \pmb x } ) \leq \displaystyle \frac { 1 } { 2 } - t _ { 0 } . } \end{array} +$$ + +We abuse notation by referring to $t _ { 0 }$ as the “margin” of $\tau$ . Note that the PMD condition only requires the upper bound on $\tau$ to be polynomial and monotonically decreasing in the region where the Bayes classifier is fairly confident. For the region $\begin{array} { r } \{ \pmb { x } : | \eta ( \pmb { x } ) - \frac { 1 } { 2 } \} < t _ { 0 } \} \end{array}$ , we allow both $\tau _ { 0 , 1 } ( \pmb { x } )$ and $\tau _ { 1 , 0 } ( \pmb { x } )$ to be arbitrary. Figure 2(d) illustrates the upper bound (orange curve) and a sample noise function (blue curve). We also show the corrupted data according to this noise function (black points are the clean data whereas red points are the data with corrupted labels). + +The PMD noise family is much more general than existing noise assumptions. For example, the boundary consistent noise (BCN) (Du & Cai, 2015; Menon et al., 2018) assumes a noise function that monotonically decreases as the data are moving away from the decision boundary. See Figure 2(c) for an illustration. This noise is much more restrictive compared to our PMD noise which (1) only requires a monotonic upper bound, and (2) allows arbitrary noise strength in a wide buffer near the decision boundary. Figure 2(b) shows a traditional feature-independent noise pattern (Reed et al., 2014; Patrini et al., 2017), which assumes $\tau _ { 0 , 1 } ( \pmb { x } )$ (resp. $\tau _ { 1 , 0 } ( \pmb { x } ) )$ ) to be a constant independent of $_ { \textbf { \em x } }$ . + +![](images/9dade37a8bec0e5a6be0825212c7aaf748d14562d66b3b8371b3b613d5ff59f3.jpg) +Figure 2: Illustration of different noise functions. (a) The original data: Gaussian blob with clean labels (by clean label, we refer to the prediction of the Bayes optimal classifier $\eta ^ { * } ( \pmb { x } )$ , not $y$ ). Confident region of $\eta$ (and thus $f$ ) in this case is the place where $\eta ( { \pmb x } )$ is close to 0 or 1. Blue and green dots correspond to different classes. (b) Uniform label noise: each point has an equal probability to be flipped. Red dots are data with corrupted labels; black dots correspond to data that are not corrupted. (c) BCN noise: the level of noise is decreasing as $\eta ^ { * } ( \pmb { x } )$ becomes confident. (d) PMD noise: noise level (blue) is only upper bounded by diminishing polynomial function when $\eta ( { \pmb x } )$ is higher or lower than certain threshold. The upper bound is shown in solid orange curve. The dashed orange curve means the noise level near the decision boundary is unbounded. + +# 2.2 THE PROGRESSIVE CORRECTION ALGORITHM + +Our algorithm iteratively trains a neural network and corrects labels. We start with a warm-up period, in which we train the neural network (NN) with the original noisy data. This allows us to attain a reasonable network before it starts fitting noise (Zhang et al., 2017). After the warm-up period, the classifier can be used for label correction. We only correct a label on which the classifier $f$ has a very high confidence. The intuition is that under the noise assumption, there exists a “pure region” in which the prediction of the noisy classifier $f$ is highly confident and is consistent with the clean Bayes optimal classifier $\eta ^ { * }$ . Thus the label correction gives clean labels within this pure region. In particular, we select a high threshold $\theta$ . If $f$ predicts a different label than $\tilde { y }$ and its confidence is above the threshold, $| f ( \bar { \pmb { x } } ) - 1 / 2 | > \theta$ , we flip the label $\tilde { y }$ to the prediction of $f$ . We repeatedly correct labels and improve the network until no label is corrected. Next, we slightly decrease the threshold $\theta$ , use the decreased threshold for label correction, and improve the model accordingly. We continue the process until convergence. For convenience in theoretical analysis, in the algorithm, we define a continuous increasing threshold $T$ and let $\theta = 1 / 2 - T$ . Our algorithm is summarized in Algorithm 1. We term our algorithm as PLC (Progressive Label Correction). In Section 3, we will show that this iterative algorithm will converge to be consistent with clean Bayes optimal classifier $\eta ^ { * } ( \pmb { x } )$ for most of the input instances. + +# Algorithm 1 Progressive Label Correction + +Input: Dataset $\tilde { S } = \{ ( \pmb { x } _ { 1 } , \widetilde { y } _ { 1 } ^ { 0 } ) , \cdot \cdot \cdot , ( \pmb { x } _ { n } , \widetilde { y } _ { n } ^ { 0 } ) \}$ , initial NN $f ( { \pmb x } )$ , step size $\beta$ , +einitial and end thresholds $( T _ { 0 } , T _ { e n d } )$ e, warm-up $m$ , total round $N$ +Output: $f _ { f i n a l } ( \cdot )$ +1: $T \gets T _ { 0 }$ +2: $\theta 1 / 2 - T _ { 0 }$ +3: for $t \gets 1 , \cdots , N$ do +4: Train $f ( { \pmb x } )$ on $\tilde { S }$ +5: for all $( { \pmb x } _ { i } , \widetilde { { y } } _ { i } ^ { t - 1 } ) \in \tilde { S }$ and $\begin{array} { r } { | f ( \pmb { x } _ { i } ) - \frac { 1 } { 2 } | \geq \theta } \end{array}$ do +6: y ti ← I{f (xi)≥ 1 } +7: end for +8: if $t \geq m$ then +9: $\theta 1 / 2 - T$ +10: $\mathbf { f } \forall i \in [ 1 , \cdot \cdot \cdot , n ] , \widetilde { y } _ { i } ^ { t } = \widetilde { y } _ { i } ^ { t - 1 } \mathbf { t h }$ en +11: $T \gets \operatorname* { m i n } ( T ( 1 + \beta ) , \bar { T _ { e n d } } )$ +12: end if +13: end if +14: $\tilde { S } \gets \{ ( { \pmb x } _ { 1 } , \widetilde { y } _ { 1 } ^ { t } ) , \cdot \cdot \cdot , ( { \pmb x } _ { n } , \widetilde { y } _ { n } ^ { t } ) \}$ +15: end for + +Generalizing to the multi-class scenario. In multi-class scenario, denote by $f _ { i } ( { \pmb x } )$ the classifier’s prediction probability of label $i$ . Let $h _ { x }$ be the classifier’s class prediction, i.e., $h _ { \pmb { x } } = \arg \operatorname* { m a x } _ { i } f _ { i } ( \pmb { x } )$ . We change the $\begin{array} { r } { | f ( \pmb { x } ) - \frac { 1 } { 2 } | } \end{array}$ term to the gap between the highest confidence $f _ { h _ { x } } ( \pmb { x } )$ and the confidence on $\widetilde { y }$ , $f _ { \widetilde { y } } ( { \pmb x } )$ . If the absolute difference between these two confidences is larger than certain threshold $\theta$ e e, then we correct $\widetilde { y }$ to $h _ { x }$ . In practice, we find using the difference of logarithms will be more robust. + +# 3 ANALYSIS + +Our analysis focuses on the asymptotic case and answers the following question: Given infinitely many data with corrupted labels, is it possible to learn a reasonably good classifier? We show that if the noise satisfies the arguably-general PMD condition, the answer is yes. Assuming mild conditions on the hypothesis class of the machine learning model and the distribution $D$ , we prove that Algorithm 1 obtains a nearly clean classifier. This reduces the challenge of noisy label learning from a realizable problem into a sample complexity problem. In this work we only focus on the asymptotic case, and leave the sample complexity for future work. + +# 3.1 ASSUMPTIONS + +Our first assumption restricts the model to be able to at least approximate the true Bayes classifier $\eta ( { \pmb x } )$ . This condition assumes that given a hypothesis class $\mathcal { H }$ with sufficient complexity, the approximation gap between a classifier $f ( { \pmb x } )$ in this class and $\eta ( { \pmb x } )$ is determined by the inconsistency between the noisy labels and the Bayes optimal classifier. + +Definition 2 (Level set $( \alpha , \epsilon )$ -consistency). Suppose data are sampled as $( { \pmb x } , \tilde { y } ) \sim D ( { \pmb x } , \tilde { \eta } ( { \pmb x } ) )$ and $f ( \pmb { x } ) = \arg \operatorname* { m i n } _ { h \in \mathcal { H } } E _ { ( \pmb { x } , \tilde { y } ) \sim D ( \pmb { x } , \tilde { \eta } ( \pmb { x } ) ) } L o s s ( h ( \pmb { x } ) , \tilde { y } )$ . Given $\varepsilon < \textstyle { \frac { 1 } { 2 } }$ , we call $\mathcal { H }$ is $( \alpha , \epsilon )$ -consistent $i f$ : + +$$ +| f ( \pmb { x } ) - \eta ( \pmb { x } ) | \leq \alpha \mathbb { E } _ { ( \pmb { z } , \bar { \pmb { y } } ) \sim D ( \pmb { z } , \bar { \eta } ( \pmb { z } ) ) } [ 1 _ { \{ \bar { \pmb { y } } _ { \pmb { z } } \neq \eta ^ { * } ( \pmb { z } ) \} } ( \pmb { z } ) | | \eta ( \pmb { z } ) - \frac { 1 } { 2 } | \geq | \eta ( \pmb { x } ) - \frac { 1 } { 2 } | ] + \epsilon . +$$ + +For two input instances $_ { z }$ and $_ { \textbf { \em x } }$ such that $\eta ( z ) ~ > ~ \eta ( \pmb { x } )$ (and hence the clean Bayes optimal classifier $\eta ^ { \ast } ( \pmb { x } )$ has higher confidence at $_ z$ than it does at $_ { \textbf { \em x } }$ ), the indicator function $\begin{array} { r } { \mathbb { 1 } _ { \{ \tilde { y } _ { z } \neq \eta ^ { * } ( z ) \} } \left( z : \left| \eta ( z ) - \frac { 1 } { 2 } \right| \ge \left| \eta ( \pmb { x } ) - \frac { 1 } { 2 } \right| \right) } \end{array}$ equals to 1 if the label of the more confident point $_ z$ is inconsistent with $\eta ^ { * } ( \pmb { x } )$ . This condition says that the approximation error of the classifier at $_ { \textbf { \em x } }$ should be controlled by the risk of $\eta ^ { * } ( \cdot )$ at points $_ z$ where $\eta ^ { * } ( \cdot )$ is more confident than it is at $_ { \textbf { \em x } }$ . + +We next define a regularity condition of data distribution which describes the continuity of the level set density function. + +Definition 3 (Level set bounded distribution). Define the margin $\begin{array} { r } { t ( { \pmb x } ) = | \eta ( { \pmb x } ) - \frac { 1 } { 2 } | } \end{array}$ and $G ( t )$ be the cdf of $t$ : $G ( t ) = \mathbb { P } _ { \mathbf { x } \sim D } ( | \eta ( \mathbf { x } ) - \frac { 1 } { 2 } | \leq t )$ . Let $g ( t ) = G ^ { \prime } ( t )$ be the density function of $t$ . We say the distribution $D$ is $( c _ { * } , c ^ { * } )$ -bounded if for all $0 \leq t \leq 1 / 2$ , $0 < c _ { * } \leq g ( t ) \leq c ^ { * }$ . If $D$ is $( c _ { * } , c ^ { * } )$ -bounded, we define the worst-case density-imbalance ratio of $D$ by $\begin{array} { r } { \ell _ { D } . = \frac { c ^ { * } } { c _ { * } } } \end{array}$ + +The above condition enforces the continuity of level set density function. This is crucial in the analysis since such a continuity allows one to borrow information from its neighborhood region so that a clean neighbor can help correct the corrupted label. To simplify the notation, we will omit $D$ in the subscript when we mention $\ell$ . From now on, we will assume: + +Assumption 1. There exist constants $\alpha , \epsilon , c _ { * } , c ^ { * }$ such that the hypothesis class $\mathcal { H }$ is $( \alpha , \epsilon )$ -consistent and the unknown distribution $D$ is $( c _ { * } , c ^ { * } )$ -bounded. + +# 3.2 MAIN RESULT AND PROOF SKETCH + +In this section we first state our main result, and then present the supporting claims. Complete proofs can be found in the appendix. Our main result below states that if our starting function is trained correctly, i.e., $f ( \pmb { x } ) = \arg \operatorname* { m i n } _ { h \in \mathcal { H } } E _ { ( \pmb { x } , \tilde { y } ) \sim D ( \pmb { x } , \tilde { \eta } ( \pmb { x } ) ) } L o s s ( h ( \pmb { x } ) , \tilde { y } )$ , then Algorithm 1 terminates with most of the final labels matching the Bayes optimal classifier labels. In practice, minimizing true risk is not achievable. Instead, the empirical risk is used to estimate true risk, approaching true risk asymptotically. For a scoring function $f$ , we will denote by $y _ { f ( \pmb { x } ) } : = \mathbb { I } ( f ( \pmb { x } ) \breve { \geq } 1 / 2 )$ the label predicted by $f$ . + +Theorem 1. Under Assumption $\begin{array} { r } { m a x ( t _ { 0 } , \frac { \alpha + \varepsilon } { 1 + 2 \alpha } ) } \end{array}$ . Then for the output of Algorithm $^ { l }$ , for any noise $^ { l }$ $\tau$ with which is PMD with margin $f$ as above and with the following initial- $t _ { 0 }$ , define $e _ { 0 } =$ izations: $( l ) T _ { 0 } < { \frac { 1 } { 2 } } - e _ { 0 }$ , (2) $\begin{array} { r } { m \geq \frac { \ell \alpha } { \varepsilon } \log \bigl ( \frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \bigr ) } \end{array}$ , (3) $\begin{array} { r } { N \ge m + \frac { 1 } { \beta } \log ( \frac { T _ { 0 } } { 3 \varepsilon } ) , } \end{array}$ , (4) $T _ { e n d } \leq 3 \epsilon$ and (5) $\begin{array} { r } { \frac { \varepsilon } { \alpha \ell } \le \beta \le \frac { 2 \varepsilon } { \alpha \ell } } \end{array}$ , we have: + +$$ +\mathbb { P } _ { \pmb { x } \sim D } \big [ y _ { f _ { f i n a l } ( \pmb { x } ) } = \eta ^ { * } ( \pmb { x } ) \big ] \geq 1 - 3 c ^ { * } \epsilon . +$$ + +In the remainder of this section we shall assume that the noise $\tau$ is PMD with margin $t _ { 0 }$ . To prove our result we first define a “pure” level set. + +Definition 4 (Pure $( e , f , \eta )$ -level set). $A$ set $L ( e , \eta ) : = \{ \pmb { x } | | \eta ( \pmb { x } ) - \frac { 1 } { 2 } | \geq e \}$ is pure for $f$ if $y _ { f ( \pmb { x } ) } =$ $\eta ^ { * } ( \pmb { x } )$ for all $\pmb { x } \in L ( e , \eta )$ . + +We now state a lemma that forms the foundation of our progressive correction algorithm. We show that given a tiny region where the model is reliable, we can move one step forward by trusting the model. Although the improvement is slight in a single round, it empowers a conservatively recursive step in the Algorithm 1. + +Lemma 1 (One round purity improvement). Suppose Assumption $^ { l }$ is satisfied, and assume an $f$ such that there exists a pure $( e , f , \eta )$ -level set with $\begin{array} { r } { 3 \dot { \epsilon } \le e < \frac { 1 } { 2 } } \end{array}$ . Let $\widetilde { \eta } _ { n e w } ( \pmb { x } ) = y _ { f ( \pmb { x } ) } i f \vert f ( \pmb { x } ) - 1 / 2 \vert \geq e$ and $\tilde { \eta } ( \pmb { x } ) i f | f ( \pmb { x } ) - 1 / 2 | < e$ , and assume $f _ { n e w } = \arg \operatorname* { m i n } _ { h \in \mathcal { H } } E _ { ( \pmb { x } , \tilde { y } ) \sim D ( \pmb { x } , \tilde { \eta } _ { n e w } ( \pmb { x } ) ) } L o s s ( h ( \pmb { x } ) , \tilde { y } )$ . Let $e _ { n e w } = \operatorname* { m i n } \{ e | e > 0$ , $L ( e , \eta )$ is pure for $f _ { n e w } \}$ . Then $\begin{array} { r } { \frac { 1 } { 2 } - e _ { n e w } \geq ( 1 + \frac { \varepsilon } { \alpha \ell } ) ( \frac { 1 } { 2 } - e ) . } \end{array}$ . + +The above lemma states that the cleansed region will be enlarged by at least a constant factor. In the following lemma, we justify the functionality of the first $m$ warm-up rounds. Since the initial neural network can behave badly, the region where we can trust the classifier can be very limited. Before starting the flipping procedure in a relatively larger level set, one first needs to expand the initial tiny region $\mathbf { \frac { 1 } { 2 } } - e _ { 0 }$ to a constant $T _ { 0 }$ . + +Lemma 2 is pure for arm-up r. Given uppose for a given function , after running Algorithm $f _ { 0 }$ thor $L ( e _ { 0 } , \eta )$ which, there $f _ { 0 }$ $T _ { 0 } < 1 / 2$ $\begin{array} { r } { m \geq \frac { \ell \alpha } { \varepsilon } \log \bigl ( \frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \bigr ) } \end{array}$ exists a level set $L ( \frac { 1 } { 2 } - T _ { 0 } , \eta )$ that is pure for $f _ { 0 }$ . + +Next we present our final lemma that combines the previous two lemmata. + +Lemma 3. Suppose Assumption $I$ is satisfied, and for a given function $f _ { 0 }$ there exists a level set $L ( e _ { 0 } , \eta )$ h is , (2) $f _ { 0 }$ ns A, (3) $I$ $f _ { 0 }$ an(4) ons:and $( l )$ $T _ { 0 } < \frac { 1 } { 2 } - e _ { 0 }$ $\begin{array} { r } { m \ge \frac { \ell \alpha } { \varepsilon } \log \bigl ( \frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \bigr ) } \end{array}$ $\begin{array} { r } { N \ge m + \frac { 1 } { \beta } \log ( \frac { 1 - 6 \varepsilon } { 2 T _ { 0 } } ) } \end{array}$ $T _ { e n d } \leq { \textstyle \frac { 1 } { 2 } } - 3 \epsilon$ $\begin{array} { r } { \frac { \varepsilon } { \alpha \ell } \le \beta \le \frac { 2 \varepsilon } { \alpha \ell } } \end{array}$ , then we have $\mathbb { P } _ { \mathbf { x } \sim D } [ y _ { f _ { f i n a l } ( \mathbf { x } ) } = \eta ^ { * } ( \pmb { x } ) ] \geq 1 - 3 c ^ { * } \epsilon$ . + +This lemma states that given an initial model that has a reasonably pure super level set, one can manage to progressively correct a large fraction of corrupted labels by running Algorithm 1 for a sufficient long time with carefully chosen parameters. The limit of Algorithm 1 will depend on the approximation ability of the neural network, which is characterized by parameter $\varepsilon$ in Definition 2. To prove Theorem 1 using Lemma 3, it suffices to get a model which has a reliable region. This is provably achievable by training with a family of good scoring functions on PMD noisy data. + +# 4 EXPERIMENTS + +We evaluate our method on both synthetic and real-world datasets. We first conduct synthetic experiments on two public datasets CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009). To synthesize the label noise, we first approximate the true posterior probability $\eta$ using the confidence prediction of a clean neural network (trained with the original clean labels). We call these original labels raw labels. Then we sample $y _ { \pmb { x } } \sim \eta ( \pmb { x } )$ for each instance $_ { \textbf { \em x } }$ . Instead of using raw labels, we use these sampled labels $y _ { x }$ as the clean labels, whose posterior probabilities are exactly $\eta ( { \pmb x } )$ ; and therefore the neural network is the Bayes optimal classifier $\eta ^ { * } : \mathcal { X } \{ 1 , \cdot \cdot , C \}$ , where $C$ is the number of classes. Note that in multi-class setting, $\eta ( { \pmb x } )$ has a vector output and $\dot { \eta _ { i } } ( \pmb { x } )$ is the $i$ -th element of this vector. + +Noise generation. We consider a generic family of noise. We consider not only feature-dependent noise, but also hybrid noise that consists of both feature-dependent noise and i.i.d. noise. + +For feature-dependent noise, we use three types of noise functions within the PMD noise family. To make the noise challenging enough, for input $_ { \textbf { \em x } }$ we always corrupt label from the most confident category $u _ { x }$ to the second confident category ${ s } _ { x }$ , according to $\eta ( { \pmb x } )$ . Because $s _ { x }$ is the class that confuses $\eta ^ { * } ( \pmb { x } )$ the most, this noise will hurt the network’s performance the most. Note that $y _ { x }$ is sampled from $\eta ( { \pmb x } )$ , which has quite an extreme confidence. Thus we generally assume $y _ { x }$ is $u _ { x }$ . For each datum $_ { \textbf { \em x } }$ , we only flip it to $s _ { x }$ or keep it as $u _ { x }$ . The three noise functions are as follows: + +$$ +\begin{array} { r l } & { \mathrm { T y p e - I : } \tau _ { u _ { \alpha } , s _ { \alpha } } = - \cfrac { 1 } { 2 } \left[ \eta _ { u _ { \alpha } } ( { \pmb x } ) - \eta _ { s _ { \alpha } } ( { \pmb x } ) \right] ^ { 2 } + \cfrac { 1 } { 2 } , \quad \mathrm { T y p e - I I : } \tau _ { u _ { \alpha } , s _ { \alpha } } = 1 - \left[ \eta _ { u _ { \alpha } } ( { \pmb x } ) - \eta _ { s _ { \alpha } } ( { \pmb x } ) \right] ^ { 3 } , } \\ & { \mathrm { T y p e - I I I : } \tau _ { u _ { \alpha } , s _ { \alpha } } = 1 - \cfrac { 1 } { 3 } \left[ \left[ \eta _ { u _ { \alpha } } ( { \pmb x } ) - \eta _ { s _ { \alpha } } ( { \pmb x } ) \right] ^ { 3 } + \left[ \eta _ { u _ { \alpha } } ( { \pmb x } ) - \eta _ { s _ { \alpha } } ( { \pmb x } ) \right] ^ { 2 } + \left[ \eta _ { u _ { \alpha } } ( { \pmb x } ) - \eta _ { s _ { \alpha } } ( { \pmb x } ) \right] \right] . } \end{array} +$$ + +Notice that the noise level is determined by the $\eta ( { \pmb x } )$ naturally and we cannot control it directly. To change the noise level, we multiply $\tau _ { u _ { x } , s _ { x } }$ by a certain constant factor such that the final proportion of noise matches our requirement. For PMD noise only, we test noise levels $3 5 \%$ and $70 \%$ , meaning that $3 5 \%$ and $70 \%$ of the data are corrupted due to the noise, respectively. + +For i.i.d. noise we follow the convention and adopt the commonly used uniform noise and asymmetric noise (Patrini et al., 2017). We artificially corrupt the labels by constructing the noise transition matrix $T$ , where $T _ { i j } = P ( \widetilde { y } = j | y = i ) = \tau _ { i j }$ defines the probability that a true label $y = i$ is flipped to $j$ e. Then for each sample with label $i$ , we replace its label with the one sampled from the probability distribution given by the $i$ -th row of matrix $T$ . We consider two kinds of i.i.d. noise in this work. (1) Uniform noise: the true label $i$ is corrupted uniformly to other classes, i.e., $T _ { i j } = \tau / ( C - 1 )$ for $i \neq j$ , and $T _ { i i } = 1 - \tau$ , where $\tau$ is the constant noise level; (2) Asymmetric noise: the true label $i$ is flipped to $j$ or stays unchanged with probabilities $T _ { i j } = \tau$ and $T _ { i i } = 1 - \tau$ , respectively. + +Baselines. We compare our method with several recently proposed approaches. (1) GCE (Zhang & Sabuncu, 2018); (2) Co-teaching $^ +$ (Yu et al., 2019); (3) SL (Wang et al., 2019); (4) LRT (Zheng et al., 2020). All these methods are generic and handle the label noise without assuming the noise structures. Finally, we also provide the results by standard method, which simply trains the deep network on noisy datasets in a standard manner. + +During training, we use a batch size of 128 and train the network for 180 epochs to ensure the convergence of all methods. We train the network with SGD optimizer, with initial learning rate 0.01. We randomly repeat the experiments 3 times, and report the mean and standard deviation values. Our code is available at https://github.com/pxiangwu/PLC. + +Results. Table 1 lists the performance of different methods under three types of feature-dependent noise at noise levels $3 5 \%$ and $70 \%$ . We observe that our method achieves the best performance across different noise settings. Moreover, notice that some of the baseline methods’ performances are inferior to the standard approach. Possible reasons are that these methods behave too conservatively in dealing with noise. Thus they only make use of a small subset of the original training set, which is not representative enough to grant the model good discriminative ability. + +In Table 2 we show the results on datasets corrupted with a combination of feature-dependent noise and i.i.d. noise, which ends up to real noise levels ranging from $50 \%$ to $70 \%$ (in terms of the proportion of corrupted labels). I.i.d. noise is overlayed on the feature-dependent noise. Our method outperforms baselines under these more complicated noise patterns. In contrast, when the noise level is high like the cases where we further apply additional $30 \%$ and $60 \%$ uniform noise, performances of a few baselines deteriorate and become worse than the standard approach. + +We carry out the ablation studies on hyper-parameters $\theta _ { 0 }$ (determining the initial confidence threshold for label correction, see Algorithm 1) and $\beta$ (the step size). In Tables 3 and 4, we show that our method is robust against the choice of $\theta _ { 0 }$ and $\beta$ up to a wide range. Notice that to compare against the threshold $\theta _ { 0 }$ , here we are calculating the absolute difference of $\log f _ { \widetilde { y } } ( { \pmb x } )$ and $\log f _ { h _ { x } } ( x )$ . As mentioned in Section 2.2, this operation gives a good performance in practice. + +Table 1: Test accuracy $( \% )$ on CIFAR-10 and CIFAR-100 under different feature-dependent noise types and levels. The average accuracy and standard deviation over 3 trials are reported. + +
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I(35%)78.11 ± 0.7479.97 ± 0.1580.65 ±0.3979.76 ± 0.7280.98 ±0.8082.80±0.27
Type-I(70%)41.98 ± 1.9640.69 ± 1.9936.52 ± 1.6236.29 ±0.6641.52 ± 4.5342.74± 2.14
Type-II(35%)76.65±0.5777.34 ± 0.4477.60± 0.8877.92 ±0.8980.74±0.2581.54± 0.47
Type-II(70%)45.57 ± 1.1245.44 ± 0.6440.30 ± 1.4641.11 ± 1.9244.67 ± 3.8946.04±2.20
Type-III (35%)76.89±0.7978.38 ±0.6779.18 ± 0.6178.81 ±0.2981.08 ± 0.3581.50±0.50
CIFAR-100Type-III(70%)43.32 ± 1.0041.90 ± 0.8637.10 ± 0.5938.49 ± 1.4644.47 ± 1.2345.05 ±1.13
Type-I(35%)57.68 ±0.2956.70 ±0.7158.37 ±0.1855.20±0.3356.74 ±0.3460.01±0.43
Type-I(70%)39.32 ± 0.4339.53 ± 0.2840.01 ± 0.7140.02 ±0.8545.29 ± 0.4345.92 ± 0.61
Type-II (35%)57.83 ± 0.2556.57 ± 0.5258.11 ± 1.0556.10±0.7357.25 ±0.6863.68±0.29
Type-II(70%)39.30± 0.3236.84± 0.3937.75 ± 0.4638.45 ± 0.4543.71 ± 0.5145.03 ±0.50
Type-III (35%) Type-II(70%)56.07 ± 0.79 40.01 ± 0.1855.77 ± 0.98 35.37 ± 2.6557.51 ± 1.16 40.53 ±0.6056.04 ± 0.74 39.94 ± 0.8456.57 ±0.30 44.41 ± 0.1963.68 ± 0.29 44.45 ± 0.62
+ +Table 2: Test accuracy $( \% )$ on CIFAR-10 and CIFAR-100 under different hybrid noise types and levels. The average accuracy and standard deviation over 3 trials are reported. + +
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I+30% Uniform75.26 ± 0.3278.72± 0.5378.08±0.6677.79± 0.4675.97± 0.2779.04±0.50
Type-I+ 60% Uniform64.25 ±0.7855.49 ± 2.1167.43 ± 1.4367.63 ± 1.3659.22 ± 0.7472.21 ± 2.92
Type-I+ 30% Asymmetric75.21 ± 0.6475.43 ± 2.9676.91 ± 0.5677.14± 0.7076.96 ± 0.4578.31 ± 0.41
Type-II +30% Uniform74.92 ± 0.6375.19 ± 0.5475.69 ± 0.2175.08± 0.4775.94 ± 0.5880.08±0.37
Type-II+ 60% Uniform64.02 ±0.6659.89 ±0.6366.39 ± 0.2966.76 ± 1.6058.99 ± 1.4371.21 ± 1.46
Type-II+ 30% Asymmetric74.28 ± 0.3973.37 ± 0.8375.30 ± 0.8175.43 ± 0.4277.03 ± 0.6277.63± 0.30
Type-III+30%Uniform74.00 ± 0.3877.31 ± 0.1177.00 ± 0.1276.22 ± 0.1275.66 ± 0.5780.06 ± 0.47
Type-III+ 60% Uniform63.96 ±0.6956.78± 1.5667.53 ± 0.5167.79 ± 0.5459.36 ± 0.9373.48 ±1.84
Type-III+ 30% Asymmetric75.31 ± 0.3474.62 ± 1.7175.70 ± 0.9176.09 ±0.1077.19 ± 0.7477.54±0.70
CIFAR-100Type-I+30%Uniform48.86±0.5652.33±0.6452.90±0.5351.34±0.6445.66 ± 1.6060.09±0.15
Type-I+ 60% Uniform35.97 ± 1.1227.17 ± 1.6638.62 ± 1.6537.57 ± 0.4323.37 ± 0.7251.68± 0.10
Type-I+ 30% Asymmetric45.85 ± 0.9351.21 ± 0.3152.69 ± 1.1450.18 ± 0.9752.04± 0.15
Type-II+30%Uniform49.32± 0.3651.99 ± 0.7553.61± 0.4643.86 ± 1.3156.40±0.34
Type-II + 60% Uniform35.16 ±0.0525.91 ± 0.6439.58 ±3.1350.58± 0.25 37.93± 0.2223.05 ±0.9960.01±0.63
Type-II + 30% Asymmetric46.50 ± 0.9551.07 ± 1.4451.98 ± 0.3749.46 ± 0.2352.11 ± 0.4649.35 ±1.53
Type-III +30% Uniform48.94± 0.6149.94± 0.4452.07±0.3550.18±0.5442.79 ± 1.7861.43±0.33 60.14±0.97
Type-III + 60% Uniform34.67± 0.1622.89 ± 0.7536.82 ±0.4937.65 ± 1.4222.81 ± 0.7250.73±2.16
Type-III + 30% Asymmetric45.70 ± 0.1249.38 ±0.8650.87 ± 1.1248.15±0.9050.31±0.3954.56±1.11
+ +Table 3: The effect of $\theta _ { 0 }$ on the performance. We use CIFAR-10 with $3 5 \%$ feature-dependent noise, and set $\beta = 0 . 1$ . + +
exp(00)0.20.30.40.5
Type-INoise83.3383.0482.6682.94
Type-II Noise81.8481.1881.0981.24
Type-III Noise81.7981.7581.9882.06
+ +Table 4: The effect of $\beta$ on the performance. We use CIFAR-10 with $3 5 \%$ feature-dependent noise, and set ex $) ( \theta _ { 0 } ) = 0 . 3$ . + +
β0.050.10.20.3
Type-I Noise83.5883.0483.2883.31
Type-II Noise80.9481.1880.9880.86
Type-III Noise81.9181.7582.1382.39
+ +Results on real-world noisy datasets. To test the effectiveness of the proposed method under realworld label noise, we conduct experiments on the Clothing1M dataset (Xiao et al., 2015). This dataset contains 1 million clothing images obtained from online shopping websites with 14 categories. The labels in this dataset are quite noisy with an unknown underlying structure. This dataset provides $5 0 k$ , + +$1 4 k$ and $1 0 k$ manually verified clean data for training, validation and testing, respectively. Following (Tanaka et al., 2018; Yi & Wu, 2019), in our experiment we discard the $5 0 k$ clean training data and evaluate the classification accuracy on the $1 0 k$ clean data. Also, following (Yi & Wu, 2019), we use a randomly sampled pseudo-balanced subset as the training set, which includes about $2 6 0 k$ images. We set the batch size 32, learning rate 0.001, and adopt SGD optimizer and use ResNet-50 with weights pre-trained on ImageNet, as in (Tanaka et al., 2018; Yi & Wu, 2019). + +We compare our method with the following baselines. (1) Standard; (2) Forward Correction (Patrini et al., 2017); (3) D2L (Ma et al., 2018); (4) JO (Tanaka et al., 2018); (5) PENCIL (Yi & Wu, 2019); (6) DY (Arazo et al., 2019); (7) GCE (Zhang & Sabuncu, 2018); (8) SL (Wang et al., 2019); (9) MLNT (Li et al., 2019); (10) LRT (Zheng et al., 2020). In Table 5 we observe that our method achieves the best performance, suggesting the applicability of our label correction strategy in real-world scenarios. + +Apart from Clothing1M, we also test our method on another smaller dataset, Food-101N (Lee et al., 2018). Food-101N is a dataset for food classification, and consists of 310k training images collected from the web. The estimated label purity is $80 \%$ . Following (Lee et al., 2018), the classification accuracy is evaluated on the Food-101 (Bossard et al., 2014) testing set, which contains 25k images with curated annotations. We use ResNet-50 pre-trained on ImageNet. We train the network for 30 epochs with SGD optimizer. The batch size is 32 and the initial learning rate is 0.005, which is divided by 10 every 10 epochs. We also adopt simple data augmentation procedures, including random horizontal flip, and resizing the image with a short edge of 256 and then randomly cropping a $2 2 4 \mathbf { x } 2 2 4$ patch from the resized image. We repeat the experiments with 3 random trials and report the mean value and standard deviation. The results are shown in Table 6. Our method much improves upon the previous approaches. + +Finally, we test our method on a recently proposed real-world dataset, ANIMAL-10N (Song et al., 2019). This dataset contains human-labeled online images for 10 animals with confusing appearance. The estimated label noise rate is $8 \%$ . There are 50,000 training and 5,000 testing images. Following (Song et al., 2019), we use VGG-19 with batch normalization. The SGD optimizer is employed. Also following (Song et al., 2019), we train the network for 100 epochs and use an initial learning rate of 0.1, which is divided by 5 at $50 \%$ and $7 5 \%$ of the total number of epochs. We repeat the experiments with 3 random trials and report the mean value and standard deviation. As is shown in Table 7, our method outperforms the existing baselines. + +Table 5: Test accuracy $( \% )$ on Clothing1M. + +
MethodStandardForwardD2LJOPENCILDYGCESLMLNTLRTPLC (ours)
Accuracy68.9469.8469.4772.2373.4971.0069.7571.0273.4771.7474.02
+ +Table 6: Test accuracy $( \% )$ on Food-101N. + +
MethodAccuracy
Standard81.67
CleanNet (Lee et al., 2018)83.95
PLC (ours)85.28 ± 0.04
+ +Table 7: Test accuracy $( \% )$ on ANIMAL-10N. + +
MethodAccuracy
Standard79.4 ± 0.14
SELFIE (Song et al., 2019)81.8 ± 0.09
PLC (ours)83.4±( 0.43
+ +# 5 CONCLUSION + +We propose a novel family of feature-dependent label noise that is much more general than the traditional i.i.d. noise pattern. Building upon this noise assumption, we propose the first datarecalibrating method that is theoretically guaranteed to converge to a well-behaved classifier. On the synthetic datasets, we show that our method outperforms various baselines under different featuredependent noise patterns subject to our assumption. Also, we test our method on different real-world noisy datasets and observe superior performances over existing approaches. The proposed noise family offers a new theoretical setting for the study of label noise. + +Acknowledgement. The authors acknowledge support from US National Science Foundation (NSF) awards CRII-1755791, CCF-1910873, CCF-1855760. This effort was partially supported by the Intelligence Advanced Research Projects Agency (IARPA) under the contract W911NF20C0038. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred. + +# REFERENCES + +Ehsan Amid, Manfred K Warmuth, and Sriram Srinivasan. Two-temperature logistic regression based on the tsallis divergence. In AISTATS, pp. 2388–2396, 2019a. + +Ehsan Amid, Manfred K. K Warmuth, Rohan Anil, and Tomer Koren. Robust bi-tempered logistic loss based on bregman divergences. In NeurIPS, pp. 15013–15022, 2019b. + +Veit Andreas, Alldrin Neil, Chechik Gal, Krasin Ivan, Gupta Abhinav, and J. Belongie Serge. Learning from noisy large-scale datasets with minimal supervision. In CVPR, pp. 6575–6583, 2017. + +Eric Arazo, Diego Ortego, Paul Albert, Noel E O’Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In ICML, 2019. + +Han Bo, Yao Quanming, Yu Xingrui, Niu Gang, Xu Miao, Hu Weihua, W. Tsang Ivor, and Sugiyama Masashi. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pp. 8536–8546, 2018. + +Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. In ECCV, pp. 446–461, 2014. + +Nontawat Charoenphakdee, Jongyeong Lee, and Masashi Sugiyama. On symmetric losses for learning from corrupted labels. In ICML, pp. 961–970, 2019. + +Chao Chen, Xiuyan Ni, Qinxun Bai, and Yusu Wang. A topological regularizer for classifiers via persistent homology. In AISTATS, pp. 2573–2582, 2019. + +Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond classconditional assumption: A primary attempt to combat instance-dependent label noise. In AAAI, 2021. + +Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance- and label-dependent label noise. In ICML, pp. 1789–1799, 2020. + +Hendrycks Dan, Lee Kimin, and Mazeika Mantas. Using pre-training can improve model robustness and uncertainty. In ICML, pp. 2712–2721, 2019. + +Jun Du and Zhihua Cai. Modelling class noise with symmetric and asymmetric distributions. In AAAI, pp. 2589–2595, 2015. + +Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy labels. arXiv preprint arXiv:1607.07526, 2016. + +Aritra Ghosh, Naresh Manwani, and P.S. Sastry. Making risk minimization tolerant to label noise. Neurocomput, 160:93–107, 2015. + +Aritra Ghosh, Himanshu Kumar, and PS Sastry. Robust loss functions under label noise for deep neural networks. In AAAI, 2017. + +Wei Hu, Zhiyuan Li, and Dingli Yu. Simple and effective regularization methods for training on noisily labeled data with generalization guarantee. In ICLR, 2020. + +Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. URL https://www.cs.toronto.edu/˜kriz/learning-features-2009-TR. pdf. + +Kuang-Huei Lee, Xiaodong He, Lei Zhang, and Linjun Yang. Cleannet: Transfer learning for scalable image classifier training with label noise. In CVPR, pp. 5447–5456, 2018. + +Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In CVPR, pp. 5051–5059, 2019. + +Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from noisy labels with distillation. In ICCV, pp. 1928–1936, 2017. + +Jiang Lu, Zhou Zhengyuan, Leung Thomas, Li Li-Jia, and Fei-Fei Li. Mentornet: Learning datadriven curriculum for very deep neural networks on corrupted labels. In ICML, pp. 2309–2318, 2018. + +Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi N. R. Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In ICML, pp. 3361–3370, 2018. + +Naresh Manwani and PS Sastry. Noise tolerance under risk minimization. IEEE transactions on cybernetics, 43(3):1146–1151, 2013. + +Aditya Krishna Menon, Brendan Van Rooyen, and Nagarajan Natarajan. Learning from binary labels with instance-dependent noise. Machine Learning, 107(8):1561–1595, 2018. + +Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. In NeurIPS, pp. 1196–1204, 2013. + +Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pp. 2233–2241, 2017. + +Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. Training deep neural networks on noisy labels with bootstrapping. In ICLR Workshop, 2014. + +Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, pp. 4331–4340, 2018. + +Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss minimization. In ICML, pp. 5739–5748, 2019. + +Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Metaweight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pp. 1917–1928, 2019. + +Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In ICML, pp. 5907–5915, 2019. + +Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, pp. 5552–5560, 2018. + +Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. In ICML, pp. 6234–6243, 2019. + +Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks. In NeurIPS, pp. 5596–5605, 2017. + +Brendan Van Rooyen, Aditya Menon, and Robert C Williamson. Learning with symmetric label noise: The importance of being unhinged. In NeurIPS, pp. 10–18, 2015. + +Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pp. 8688–8696, 2018. + +Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In CVPR, pp. 322–330, 2019. + +Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. In NeurIPS, 2020. + +Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, pp. 2691–2699, 2015. + +Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L dmi: A novel information-theoretic loss function for training deep nets robust to label noise. In NeurIPS, pp. 6222–6233, 2019. + +Yan Yan, Romer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from ´ multiple annotators with varying expertise. Machine learning, 95(3):291–327, 2014. + +Kun Yi and Jianxin Wu. Probabilistic end-to-end noise correction for learning with noisy labels. In CVPR, pp. 7017–7025, 2019. + +Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In ICML, pp. 7164–7173, 2019. + +Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2017. + +Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, pp. 8778–8788, 2018. + +Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded correction of noisy labels. In ICML, pp. 11447–11457, 2020. + +# A APPENDIX + +Lemma 1 (One round purity improvement). Suppose Assumption $^ { l }$ is satisfied, and assume an $f$ such that there exists a pure $( e , f , \eta )$ -level set with $\begin{array} { r } { 3 \dot { \epsilon } \le e < \frac { 1 } { 2 } } \end{array}$ . Let $\tilde { \eta } _ { n e w } ( { \pmb x } ) = y _ { f ( { \pmb x } ) } i f | f ( { \pmb x } ) - 1 / 2 | \ge e$ and $\tilde { \eta } ( \pmb { x } ) i f | f ( \pmb { x } ) - 1 / 2 | < e$ , and assume $f _ { n e w } = \arg \operatorname* { m i n } _ { h \in \mathcal { H } } E _ { ( \pmb { x } , \tilde { y } ) \sim D ( \pmb { x } , \tilde { \eta } _ { n e w } ( \pmb { x } ) ) } L o s s ( h ( \pmb { x } ) , \tilde { y } )$ . Let $e _ { n e w } = \operatorname* { m i n } \{ e | e > 0$ , $L ( e , \eta )$ is pure for $f _ { n e w } \}$ . Then $\begin{array} { r } { \frac { 1 } { 2 } - e _ { n e w } \geq ( 1 + \frac { \varepsilon } { \alpha \ell } ) ( \frac { 1 } { 2 } - e ) . } \end{array}$ . + +Proof: We analyze the case where $\begin{array} { r } { \eta ( \pmb { x } ) > \frac { 1 } { 2 } } \end{array}$ . The analysis on the other side can be derived similarly. Due to the fact that there exists a level set $L ( e , \eta )$ pure to $f$ , we have $\begin{array} { r } { e \leq | \eta ( \pmb { x } ) - \frac { 1 } { 2 } | , \forall \pmb { x } } \end{array}$ : + +$$ +\mathbb { E } _ { ( z , \tilde { y } ) \sim ( D , \tilde { \eta } _ { n e w } ( z ) ) } \left[ \mathbb { 1 } _ { \{ \tilde { y } _ { z } \neq \eta ^ { * } ( z ) \} } ( z ) \bigg | \bigg | \eta ( z ) - \frac { 1 } { 2 } \bigg | \ge \bigg | \eta ( x ) - \frac { 1 } { 2 } \bigg | \right] = 0 . +$$ + +Now consider $_ { \textbf { \em x } }$ where $e - \gamma \leq | \eta ( \pmb { x } ) - \frac { 1 } { 2 } |$ . Since the distribution $D$ is $( c _ { * } , c ^ { * } )$ -bounded, we have: + +$$ +\begin{array} { r l } & { \mathbb { E } _ { ( z \sim \hat { y } ) \sim ( D , \hat { y } \sim \alpha ( z ) ) } [ \hat { 1 } ( s _ { z \sim \hat { y } } , \hat { x } ) ( z ) ] \lVert ( z ( z ) - \frac { 1 } { 2 } \vert \geq | y | ( x ) - \frac { 1 } { 2 } \vert ] } \\ & { = \mathbb { E } _ { ( \hat { z } ) \sim } f ( \hat { x } ) \lVert ( z ) - \frac { 1 } { 2 } [ \hat { y } \langle x \rangle - \frac { 1 } { 2 } ] } \\ & { \frac { \mathbb { E } _ { ( \hat { z } ) \sim } f ( \hat { x } ) \cdot \mathcal { Y } ( z ) \cdot \mathcal { Y } ( z ) \cdot \mathcal { Y } ( z ) \cdot \frac { 1 } { 2 } } { \lVert \mathcal { Z } \rVert ( z ) ^ { 2 } - \frac { 1 } { 2 } } [ \frac { \mathbb { E } _ { \hat { z } } \cdot \mathcal { Y } ( z ) \cdot \mathcal { Y } ( z ) } { 1 - \frac { 1 } { 2 } } ] } \\ & { \leq \frac { \mathbb { E } _ { ( \hat { z } ) \sim } f ( \hat { x } ) \cdot \mathcal { Y } ( z ) \cdot \frac { 1 } { 2 } + \epsilon - \epsilon - \mathbb { E } _ { z } \cdot \mathcal { Z } ( \hat { z } ) \cdot \frac { \mathbb { E } _ { \hat { z } } } { 2 } + \epsilon } { \mathcal { Z } \cdot { \mathcal { Z } } ( z ) ^ { 2 } - \frac { \mathbb { E } _ { z } \cdot \mathcal { Y } ( z ) \cdot \mathcal { Y } ( z ) \cdot \mathcal { Y } ( z ) \cdot \mathcal { Z } ( z ) \cdot \frac { 1 } { 2 } - \frac { \epsilon } { 2 } } [ \mathcal { Z } \cdot \mathcal { Y } ( z ) \leq \frac { \frac { 1 } { 2 } } { 2 } - \epsilon + \frac { \mathbb { E } _ { z } } { 1 } ] } } \\ & \quad + \frac \mathbb { E } _ { ( \hat { z } ) \sim } f ( \hat { x } ) \cdot \mathcal { Z } ( z ) \cdot \frac { 1 } { 2 } + \epsilon \mathcal { Z } ( \hat { z } ) \cdot \frac { \mathbb { E } _ { z } \cdot \mathcal { Y } ( z ) \cdot \mathcal { Y } ( z ) \cdot \mathcal { Z } ( z ) \cdot \mathcal { Z } ( z ) } \epsilon - \frac \mathbb { E } _ { z } \cdot \{ { y } \cdot \mathcal { Z } \} \cdot \mathcal { Y } ( z ) \cdot \mathcal { Z } \cdot \mathcal { Y } ( z \end{array} +$$ + +If $\begin{array} { r } { \frac { \varepsilon } { \ell \alpha } ( \frac { 1 } { 2 } - e ) \le \gamma \le \frac { 2 \varepsilon } { \ell \alpha } ( \frac { 1 } { 2 } - e ) } \end{array}$ , the impurity in super level set $\begin{array} { r } { \frac { 1 } { 2 } + e - \gamma } \end{array}$ is at most $\frac { 2 \epsilon } { \alpha }$ . The level set $( \alpha , \varepsilon )$ -consistency condition implies $| f _ { n e w } ( \pmb { x } ) - \eta ( \pmb { x } ) | \le 3 \epsilon$ for $_ { \textbf { \em x } }$ s.t. $\begin{array} { r } { e - \gamma \le | \eta ( \pmb { x } ) - \frac { 1 } { 2 } | } \end{array}$ . If $e \geq 3 \epsilon$ , ${ \dot { f } } _ { n e w } ( { \pmb x } )$ will give the same label as $\eta ^ { * } ( \pmb { x } )$ and thus $( e - \gamma , \eta )$ -level set becomes pure for $f$ . Meanwhile, the choice of $\gamma$ ensures that $\begin{array} { r } { \frac { 1 } { 2 } - e _ { n e w } \geq ( 1 + \frac { \varepsilon } { \ell \alpha } ) ( \frac { 1 } { 2 } - e ) } \end{array}$ . □ + +Lemma 2 is pure for arm-up r. Given uppose for a given function, after running Algorithm $f _ { 0 }$ thor $L ( e _ { 0 } , \eta )$ which, there $f _ { 0 }$ $T _ { 0 } < 1 / 2$ $^ { l }$ $\begin{array} { r } { m \ge \frac { \ell \alpha } { \varepsilon } \log \bigl ( \frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \bigr ) } \end{array}$ exists a level set $L ( \frac { 1 } { 2 } - T _ { 0 } , \eta )$ that is pure for $f _ { 0 }$ . + +Proof: The proof follows from the fact that each round of label flipping improves the purity by a factor of $\textstyle { \left( { 1 + { \frac { \varepsilon } { \ell \alpha } } } \right) }$ . To obtain an at least $T _ { 0 }$ pure region, it suffices to repeat the flipping step for m ≥ \`αε log( T01−2e ) rounds. □ + +Lemma 3. Suppose Assumption $I$ is satisfied, and for a given function $f _ { 0 }$ there exists a level set $L ( e _ { 0 } , \eta )$ h is p, (2) $f _ { 0 }$ ns A, (3) $I$ $f _ { 0 }$ an(4) ons:and $( l )$ $T _ { 0 } < \frac { 1 } { 2 } - e _ { 0 }$ $\begin{array} { r } { m \ge \frac { \ell \alpha } { \varepsilon } \log \bigl ( \frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \bigr ) } \end{array}$ $\begin{array} { r } { N \ge m + \frac { 1 } { \beta } \log ( \frac { 1 - 6 \varepsilon } { 2 T _ { 0 } } ) , } \end{array}$ $T _ { e n d } \leq { \textstyle \frac { 1 } { 2 } } - 3 \epsilon$ $\begin{array} { r } { \frac { \varepsilon } { \alpha \ell } \le \beta \le \frac { 2 \varepsilon } { \alpha \ell } } \end{array}$ , then we have $\mathbb { P } _ { \pmb { x } \sim D } \big [ y _ { f _ { f i n a l } ( \pmb { x } ) } = \eta ^ { * } ( \pmb { x } ) \big ] \geq 1 - 3 c ^ { * } \epsilon$ . + +Proof: The proof can be done by combining Lemma 1 and Lemma 2. In the first $m$ iterations, by Lemma 2, we can guarantee a level set $\bigl ( \frac { 1 } { 2 } - T _ { 0 } , \eta \bigr )$ pure to $f$ . In the rest of the iterations we ensure the level set $\begin{array} { r } { | \eta ( \pmb { x } ) - \frac 1 2 | \geq \frac 1 2 - T } \end{array}$ is pure. We increase $T$ by a reasonable factor of $\beta$ to avoid incurring too many corrupted labels while ensuring enough progress in label purification, i.e., $\begin{array} { r } { \frac { \varepsilon } { \alpha \ell } \le \beta \le \frac { 2 \varepsilon } { \alpha \ell } } \end{array}$ such that in the level set $\begin{array} { r } { | \eta ( { \pmb x } ) - \frac 1 2 | \geq \frac 1 2 - T } \end{array}$ we have $| f ( { \pmb x } ) - \eta ( { \pmb x } ) | \leq 3 \varepsilon$ . This condition ensures the correctness of flipping when $T \leq { \frac { 1 } { 2 } } - 3 \varepsilon$ . The purity cannot be improved once $\begin{array} { r } { T \ge \frac { 1 } { 2 } - 3 \varepsilon = T _ { e n d } } \end{array}$ since there is no guarantee that $f ( { \pmb x } )$ has consistent label with $\eta ( { \pmb x } )$ when $| \eta ( \pmb { x } ) - \frac { 1 } { 2 } | < 3 \varepsilon$ and $| \eta ( { \pmb x } ) - f ( { \pmb x } ) | \leq \bar { 3 } \varepsilon$ . By $( c _ { * } , c ^ { * } )$ -bounded assumption on $D$ , its mass of impure $3 \varepsilon$ level set region is at most $3 c ^ { * } \epsilon$ . □ + +Theorem 1. Under assumption $I$ , for any noise $\tau$ which is PMD with margin $t _ { 0 }$ , define $e _ { 0 } =$ $\begin{array} { r } { m a x ( t _ { 0 } , \frac { \alpha + \varepsilon } { 1 + 2 \alpha } ) } \end{array}$ . Then for the output of Algorithm $^ { l }$ with $f$ as above and with the following initializations: $( I ) T _ { 0 } < { \textstyle \frac { 1 } { 2 } } - e _ { 0 } $ , (2) $\begin{array} { r } { m \geq \frac { \ell \alpha } { \varepsilon } \log \bigl ( \frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \bigr ) } \end{array}$ , (3) $\begin{array} { r } { N \ge m + \frac { 1 } { \beta } \log \bigl ( \frac { T _ { 0 } } { 3 \varepsilon } \bigr ) , } \end{array}$ , (4) $T _ { e n d } \leq 3 \epsilon$ and (5) $\begin{array} { r } { \frac { \varepsilon } { \alpha \ell } \le \beta \le \frac { 2 \varepsilon } { \alpha \ell } } \end{array}$ , we have: + +$$ +\mathbb { P } _ { \pmb { x } \sim D } \big [ y _ { f _ { f i n a l } ( \pmb { x } ) } = \eta ^ { * } ( \pmb { x } ) \big ] \geq 1 - 3 c ^ { * } \epsilon . +$$ + +Proof: The proof is based on Lemma 3 plus a verification of the existence of $f _ { 0 }$ for which there exists a pure $( e _ { 0 } , f , \eta )$ -level set. Let: + +$$ +\tau ( { \pmb x } ) = \left\{ \begin{array} { l l } { \tau _ { 1 0 } ( { \pmb x } ) } & { \eta ( { \pmb x } ) \geq 1 / 2 } \\ { \tau _ { 0 1 } ( { \pmb x } ) } & { \eta ( { \pmb x } ) < 1 / 2 } \end{array} \right. . +$$ + +In the level set $\begin{array} { r } { | \eta ( \pmb { x } ) - \frac { 1 } { 2 } | \geq e _ { 0 } , \mathbb { P } _ { z } [ \tilde { y } _ { z } \neq \eta ^ { * } ( z ) | | \eta ( z ) - \frac { 1 } { 2 } | \geq e _ { 0 } ] \leq \frac { 1 } { 2 } - e _ { 0 } + \tau ( z ) } \end{array}$ . By level set $( \alpha , \varepsilon )$ -consistency, it suffices to satisfy $\alpha ( \textstyle { \frac { 1 } { 2 } } - e _ { o } + \tau ) + \varepsilon \le e _ { 0 }$ to ensure that $f ( { \pmb x } )$ has the same prediction with $\eta ( { \pmb x } )$ when $| \eta ( \pmb { x } ) - \frac { 1 } { 2 } | \geq \overline { { e } } _ { 0 }$ . By polynomial level set diminishing noise, we have $\tau ( \pmb { x } ) \leq { \frac { 1 } { 2 } } - e _ { 0 }$ if $e _ { 0 } > t _ { 0 }$ , and thus by choosing $\begin{array} { r } { e _ { 0 } = m a x ( t _ { 0 } , \frac { \alpha + \varepsilon } { 1 + 2 \alpha } ) } \end{array}$ one can ensure that initial $f _ { 0 }$ has a pure $( e _ { 0 } , f _ { 0 } , \eta )$ -level set. The rest of the proof follows from Lemma 3. □ \ No newline at end of file diff --git a/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_content_list.json b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..756144cee2daa20618776476abf87068bd232d9c --- /dev/null +++ b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "text", + "text": "LEARNING WITH FEATURE-DEPENDENT LABEL NOISE:A PROGRESSIVE APPROACH", + "text_level": 1, + "bbox": [ + 174, + 101, + 826, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yikai $\\mathbf { Z h a n g ^ { 1 * } }$ , Songzhu Zheng2∗, Pengxiang $\\mathbf { W _ { u } } ^ { 1 * }$ , Mayank Goswami3, Chao Chen2 ", + "bbox": [ + 184, + 169, + 777, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Rutgers University, $\\{ \\mathtt { y z 4 2 2 } , \\mathtt { p w 2 4 1 } \\}$ @cs.rutgers.edu \n2Stony Brook University, {zheng.songzhu,chao.chen.1}@stonybrook.edu \n3City University of New York, mayank.isi@gmail.com ", + "bbox": [ + 187, + 185, + 766, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 263, + 544, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Label noise is frequently observed in real-world large-scale datasets. The noise is introduced due to a variety of reasons; it is heterogeneous and feature-dependent. Most existing approaches to handling noisy labels fall into two categories: they either assume an ideal feature-independent noise, or remain heuristic without theoretical guarantees. In this paper, we propose to target a new family of featuredependent label noise, which is much more general than commonly used i.i.d. label noise and encompasses a broad spectrum of noise patterns. Focusing on this general noise family, we propose a progressive label correction algorithm that iteratively corrects labels and refines the model. We provide theoretical guarantees showing that for a wide variety of (unknown) noise patterns, a classifier trained with this strategy converges to be consistent with the Bayes classifier. In experiments, our method outperforms SOTA baselines and is robust to various noise types and levels. ", + "bbox": [ + 233, + 296, + 766, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 494, + 334, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Addressing noise in training set labels is an important problem in supervised learning. Incorrect annotation of data is inevitable in large-scale data collection, due to intrinsic ambiguity of data/class and mistakes of human/automatic annotators (Yan et al., 2014; Andreas et al., 2017). Developing methods that are resilient to label noise is therefore crucial in real-life applications. ", + "bbox": [ + 176, + 527, + 823, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classical approaches take a rather simplistic i.i.d. assumption on the label noise, i.e., the label corruption is independent and identically distributed and thus is feature-independent. Methods based on this assumption either explicitly estimate the noise pattern (Reed et al., 2014; Patrini et al., 2017; Dan et al., 2019; Xu et al., 2019) or introduce extra regularizer/loss terms (Natarajan et al., 2013; Van Rooyen et al., 2015; Xiao et al., 2015; Zhang & Sabuncu, 2018; Ma et al., 2018; Arazo et al., 2019; Shen & Sanghavi, 2019). Some results prove that the commonly used losses are naturally robust against such i.i.d. label noise (Manwani & Sastry, 2013; Ghosh et al., 2015; Gao et al., 2016; Ghosh et al., 2017; Charoenphakdee et al., 2019; Hu et al., 2020). ", + "bbox": [ + 174, + 590, + 825, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although these methods come with theoretical guarantees, they usually do not perform as well as expected in practice due to the unrealistic i.i.d. assumption on noise. This is likely because label noise is heterogeneous and feature-dependent. A cat with an intrinsically ambiguous appearance is more likely to be mislabeled as a dog. An image with poor lighting or severe occlusion can be mislabeled, as important visual clues are imperceptible. Methods that can combat label noise of a much more general form are very much needed to address real-world challenges. ", + "bbox": [ + 174, + 709, + 825, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To adapt to the heterogeneous label noise, state-of-the-arts (SOTAs) often resort to a data-recalibrating strategy. They progressively identify trustworthy data or correct data labels, and then train using these data (Tanaka et al., 2018; Wang et al., 2018; Lu et al., 2018; Li et al., 2019). The models gradually improve as more clean data are collected or more labels are corrected, eventually converging to models of high accuracy. These data-recalibrating methods best leverage the learning power of deep neural nets and achieve superior performance in practice. However, their underlying mechanism remains a mystery. No methods in this category can provide theoretical insights as to why the model can converge to an ideal one. Thus, these methods require careful hyperparameter tuning and are hard to generalize. ", + "bbox": [ + 174, + 799, + 823, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel and principled method that specifically targets the heterogeneous, feature-dependent label noise. Unlike previous methods, we target a much more general family of noise, called Polynomial Margin Diminishing (PMD) label noise. In this noise family, we allow arbitrary noise level except for data far away from the true decision boundary. This is consistent with the real-world scenario; data near the decision boundary are harder to distinguish and more likely to be mislabeled. Meanwhile, a datum far away from the decision boundary is a typical example of its true class and should have a reasonably bounded noise level. ", + "bbox": [ + 174, + 138, + 825, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assuming this new PMD noise family, we propose a theoretically-guaranteed data-recalibrating algorithm that gradually corrects labels based on the noisy classifier’s confidence. We start from data points with high confidence, and correct the labels of these data using the predictions of the noisy classifier. Next, the model is improved using cleaned labels. We continue alternating the label correction and model improvement until it converges. See Figure 1 for an illustration. Our main theorem shows that with a theory-informed criterion for label correction at each iteration, the improvement of the label purity is guaranteed. Thus the model is guaranteed to improve with sufficient rate through iterations and eventually becomes consistent with the Bayes optimal classifier. ", + "bbox": [ + 174, + 243, + 825, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beside the theoretical strength, we also demonstrate the power of our method in practice. Our method outperforms others on CIFAR-10/100 with various synthetic noise patterns. We also evaluate our method against SOTAs on three real-world datasets with unknown noise patterns. ", + "bbox": [ + 176, + 362, + 825, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, our method is the first data-recalibrating method that is theoretically guaranteed to converge to an ideal model. The PMD noise family encompasses a broad spectrum of heterogeneous and feature-dependent noise, and better approximates the real-world scenario. It also provides a novel theoretical setting for the study of label noise. ", + "bbox": [ + 176, + 410, + 825, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Related works. We review works that do not assume an i.i.d. label noise. Menon et al. (2018) generalized the work of (Ghosh et al., 2015) and provided an elegant theoretical framework, showing that loss functions fulfilling certain conditions naturally resist instance-dependent noise. The method can achieve even better theoretical properties (i.e., Bayes-consistency) with stronger assumption on the clean posterior probability $\\eta$ . In practice, this method has not been extended to deep neural networks. Cheng et al. (2020) proposed an active learning method for instance-dependent label noise. ", + "bbox": [ + 174, + 473, + 825, + 558 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e1c5bd95d246a39360bb9537b3313db61b0fbc917271c8e5928ac07345e50890.jpg", + "image_caption": [ + "Figure 1: Illustration of the algorithm using synthetic data. (a) Gaussian blob with clean label $( \\eta ^ { * } ( \\pmb { x } ) )$ . (b) Data with corrupted labels. (c) Final corrected data. Black dots are the data that have their clean labels. Red dots are the noisy data. Points that remain un-corrected are closer to the decision boundary. Our algorithm corrects most of the noise only using noisy classifier’s confidence. (d) Data after label correction. (e)-(h) We show the intermediate results at different iterations. Gray region is the area where the classifier has high confidence. Labels within this region are corrected. " + ], + "image_footnote": [], + "bbox": [ + 197, + 584, + 800, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The algorithm iteratively queries clean labels from an oracle on carefully selected data. However, this approach is not applicable to settings where kosher annotations are unavailable. Another contemporary work (Chen et al., 2021) showed that the noise in real-world dataset is unlikely to be i.i.d., and proposed to fix the noisy labels by averaging the network predictions on each instance over the whole training process. While being effective, their method lacks theoretical guarantees. Chen et al. (2019) showed by regulating the topology of a classifier’s decision boundary, one can improve the model’s robustness against label noise. ", + "bbox": [ + 173, + 103, + 825, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-recalibrating methods use noisy networks’ predictions to iteratively select/correct data and improve the models. Tanaka et al. (2018) introduced a joint training framework which simultaneously enforces the network to be consistent with its own predictions and corrects the noisy labels during training. Wang et al. (2018) identified noisy labels as outliers based on their label consistencies with surrounding data. Lu et al. (2018) used a curriculum learning strategy where the teacher net is trained on a small kosher dataset to determine if a datum is clean; then the learnt curriculum that gives the weight to each datum is fed into the student net for the training and inference. (Yu et al., 2019; Bo et al., 2018) trained two synchronized networks; the confidence and consistency of the two networks are utilized to identify clean data. Wu et al. (2020) selected the clean data by investigating the topological structures of the training data in the learned feature space. For completeness, we also refer to other methods of similar design (Li et al., 2017; Vahdat, 2017; Andreas et al., 2017; Ma et al., 2018; Thulasidasan et al., 2019; Arazo et al., 2019; Shu et al., 2019; Yi & Wu, 2019). ", + "bbox": [ + 174, + 208, + 825, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for theoretical guarantees, Ren et al. (2018) proposed an algorithm that iteratively re-weights each data point by solving an optimization problem. They proved the convergence of the training, but provided no guarantees that the model converges to an ideal one. Amid et al. (2019b) generalized the work of (Amid et al., 2019a) and proposed a tempered matching loss. They showed that when the final softmax layer is replaced by the bi-tempered loss, the resulting classifier will be Bayes consistent. Zheng et al. (2020) proved a one-shot guarantee for their data-recalibrating method; but the convergence of the model is not guaranteed. Our method is the first data-recalibrating method which is guaranteed to converge to a well-behaved classifier. ", + "bbox": [ + 174, + 381, + 825, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 METHOD ", + "text_level": 1, + "bbox": [ + 174, + 512, + 281, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by introducing the family of Poly-Margin Diminishing (PMD) label noise. In Section 2.2, we present our main algorithm. Finally, we prove the correctness of our algorithm in Section 3. ", + "bbox": [ + 174, + 542, + 825, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notations and preliminaries. Although the noise setting and algorithm naturally generalize to multiclass, for simplicity we focus on binary classification. Let the feature space be $\\mathcal { X }$ . We assume the data $( { \\pmb x } , y )$ is sampled from an underlying distribution $D$ on $\\mathcal { X } \\times \\{ 0 , 1 \\}$ . Define the posterior probability $\\eta ( \\pmb { x } ) = \\mathbb { P } [ y = 1 \\mid \\pmb { x } ]$ . Let $\\tau _ { 0 , 1 } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 1 \\ | \\ y = 0 , \\pmb { x } ]$ and $\\bar { \\tau _ { 1 , 0 } } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 0 \\ | \\ y = 1 , \\pmb { x } ]$ be the noise functions, where $\\widetilde { y }$ e e denotes the corrupted label. For example, if a datum $_ { \\textbf { \\em x } }$ has true label $y = 0$ , it has $\\tau _ { 0 , 1 } ( \\pmb { x } )$ echance to be corrupted to 1. Similarly, it has $\\tau _ { 1 , 0 } ( \\pmb { x } )$ chance to be corrupted from 1 to 0. Let $\\boldsymbol { \\widetilde { \\eta } } ( \\boldsymbol { \\dot { \\mathbf { x } } } ) = \\mathbb { P } [ \\boldsymbol { \\widetilde { y } } = 1 \\mid \\boldsymbol { \\mathbf { x } } ]$ be the noisy posterior probability of $\\widetilde y = 1$ given feature $_ { \\textbf { \\em x } }$ . Let $\\eta ^ { * } ( \\pmb { x } ) = \\mathbb { I } _ { \\{ \\eta ( \\pmb { x } ) \\geq \\frac { 1 } { 2 } \\} }$ ebe the (clean) Bayes optimal classifier, where $\\mathbb { I } _ { A }$ eequals 1 if $A$ is true, and 0 otherwise. Finally, let $f ( \\pmb { x } ) : \\mathcal { X } [ 0 , 1 ]$ be the classifier scoring function (the softmax output of a neural network in this paper). ", + "bbox": [ + 173, + 578, + 825, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 POLY-MARGIN DIMINISHING NOISE ", + "text_level": 1, + "bbox": [ + 178, + 737, + 465, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the family of noise functions $\\tau$ this paper will address. We introduce the concept of polynomial margin diminishing noise (PMD noise), which only upper bounds the noise $\\tau$ in a certain level set of $\\eta ( { \\pmb x } )$ , thus allowing $\\tau$ to be arbitrarily high outside the restricted domain. This formulation not only covers the feature-independent scenario but also generalizes scenarios proposed by (Du & Cai, 2015; Menon et al., 2018; Cheng et al., 2020). ", + "bbox": [ + 174, + 762, + 825, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (PMD noise). $A$ pair of noise functions $\\tau _ { 0 , 1 } ( \\pmb { x } )$ and $\\tau _ { 1 , 0 } ( \\pmb { x } )$ are polynomial-margin diminishing (PMD), if there exist constants $t _ { 0 } \\in ( 0 , \\frac { 1 } { 2 } )$ , and $c _ { 1 } , c _ { 2 } > 0$ such that: ", + "bbox": [ + 173, + 834, + 820, + 866 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b81884617b954fafce98fcc351dfb3f087b1a26114f5ce5e15a5a23e3cf26faa.jpg", + "text": "$$\n\\begin{array} { c l l } { \\displaystyle \\tau _ { 1 , 0 } ( { \\pmb x } ) \\leq c _ { 1 } [ 1 - \\eta ( { \\pmb x } ) ] ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\geq \\displaystyle \\frac { 1 } { 2 } + t _ { 0 } , a n d } \\\\ { \\displaystyle \\tau _ { 0 , 1 } ( { \\pmb x } ) \\leq c _ { 1 } \\eta ( { \\pmb x } ) ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\leq \\displaystyle \\frac { 1 } { 2 } - t _ { 0 } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 328, + 868, + 673, + 930 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We abuse notation by referring to $t _ { 0 }$ as the “margin” of $\\tau$ . Note that the PMD condition only requires the upper bound on $\\tau$ to be polynomial and monotonically decreasing in the region where the Bayes classifier is fairly confident. For the region $\\begin{array} { r } \\{ \\pmb { x } : | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\} < t _ { 0 } \\} \\end{array}$ , we allow both $\\tau _ { 0 , 1 } ( \\pmb { x } )$ and $\\tau _ { 1 , 0 } ( \\pmb { x } )$ to be arbitrary. Figure 2(d) illustrates the upper bound (orange curve) and a sample noise function (blue curve). We also show the corrupted data according to this noise function (black points are the clean data whereas red points are the data with corrupted labels). ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The PMD noise family is much more general than existing noise assumptions. For example, the boundary consistent noise (BCN) (Du & Cai, 2015; Menon et al., 2018) assumes a noise function that monotonically decreases as the data are moving away from the decision boundary. See Figure 2(c) for an illustration. This noise is much more restrictive compared to our PMD noise which (1) only requires a monotonic upper bound, and (2) allows arbitrary noise strength in a wide buffer near the decision boundary. Figure 2(b) shows a traditional feature-independent noise pattern (Reed et al., 2014; Patrini et al., 2017), which assumes $\\tau _ { 0 , 1 } ( \\pmb { x } )$ (resp. $\\tau _ { 1 , 0 } ( \\pmb { x } ) )$ ) to be a constant independent of $_ { \\textbf { \\em x } }$ . ", + "bbox": [ + 173, + 194, + 825, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9dade37a8bec0e5a6be0825212c7aaf748d14562d66b3b8371b3b613d5ff59f3.jpg", + "image_caption": [ + "Figure 2: Illustration of different noise functions. (a) The original data: Gaussian blob with clean labels (by clean label, we refer to the prediction of the Bayes optimal classifier $\\eta ^ { * } ( \\pmb { x } )$ , not $y$ ). Confident region of $\\eta$ (and thus $f$ ) in this case is the place where $\\eta ( { \\pmb x } )$ is close to 0 or 1. Blue and green dots correspond to different classes. (b) Uniform label noise: each point has an equal probability to be flipped. Red dots are data with corrupted labels; black dots correspond to data that are not corrupted. (c) BCN noise: the level of noise is decreasing as $\\eta ^ { * } ( \\pmb { x } )$ becomes confident. (d) PMD noise: noise level (blue) is only upper bounded by diminishing polynomial function when $\\eta ( { \\pmb x } )$ is higher or lower than certain threshold. The upper bound is shown in solid orange curve. The dashed orange curve means the noise level near the decision boundary is unbounded. " + ], + "image_footnote": [], + "bbox": [ + 184, + 309, + 813, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 THE PROGRESSIVE CORRECTION ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 674, + 535, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our algorithm iteratively trains a neural network and corrects labels. We start with a warm-up period, in which we train the neural network (NN) with the original noisy data. This allows us to attain a reasonable network before it starts fitting noise (Zhang et al., 2017). After the warm-up period, the classifier can be used for label correction. We only correct a label on which the classifier $f$ has a very high confidence. The intuition is that under the noise assumption, there exists a “pure region” in which the prediction of the noisy classifier $f$ is highly confident and is consistent with the clean Bayes optimal classifier $\\eta ^ { * }$ . Thus the label correction gives clean labels within this pure region. In particular, we select a high threshold $\\theta$ . If $f$ predicts a different label than $\\tilde { y }$ and its confidence is above the threshold, $| f ( \\bar { \\pmb { x } } ) - 1 / 2 | > \\theta$ , we flip the label $\\tilde { y }$ to the prediction of $f$ . We repeatedly correct labels and improve the network until no label is corrected. Next, we slightly decrease the threshold $\\theta$ , use the decreased threshold for label correction, and improve the model accordingly. We continue the process until convergence. For convenience in theoretical analysis, in the algorithm, we define a continuous increasing threshold $T$ and let $\\theta = 1 / 2 - T$ . Our algorithm is summarized in Algorithm 1. We term our algorithm as PLC (Progressive Label Correction). In Section 3, we will show that this iterative algorithm will converge to be consistent with clean Bayes optimal classifier $\\eta ^ { * } ( \\pmb { x } )$ for most of the input instances. ", + "bbox": [ + 173, + 702, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Progressive Label Correction ", + "text_level": 1, + "bbox": [ + 174, + 118, + 539, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Dataset $\\tilde { S } = \\{ ( \\pmb { x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { 0 } ) , \\cdot \\cdot \\cdot , ( \\pmb { x } _ { n } , \\widetilde { y } _ { n } ^ { 0 } ) \\}$ , initial NN $f ( { \\pmb x } )$ , step size $\\beta$ , \neinitial and end thresholds $( T _ { 0 } , T _ { e n d } )$ e, warm-up $m$ , total round $N$ \nOutput: $f _ { f i n a l } ( \\cdot )$ \n1: $T \\gets T _ { 0 }$ \n2: $\\theta 1 / 2 - T _ { 0 }$ \n3: for $t \\gets 1 , \\cdots , N$ do \n4: Train $f ( { \\pmb x } )$ on $\\tilde { S }$ \n5: for all $( { \\pmb x } _ { i } , \\widetilde { { y } } _ { i } ^ { t - 1 } ) \\in \\tilde { S }$ and $\\begin{array} { r } { | f ( \\pmb { x } _ { i } ) - \\frac { 1 } { 2 } | \\geq \\theta } \\end{array}$ do \n6: y ti ← I{f (xi)≥ 1 } \n7: end for \n8: if $t \\geq m$ then \n9: $\\theta 1 / 2 - T$ \n10: $\\mathbf { f } \\forall i \\in [ 1 , \\cdot \\cdot \\cdot , n ] , \\widetilde { y } _ { i } ^ { t } = \\widetilde { y } _ { i } ^ { t - 1 } \\mathbf { t h }$ en \n11: $T \\gets \\operatorname* { m i n } ( T ( 1 + \\beta ) , \\bar { T _ { e n d } } )$ \n12: end if \n13: end if \n14: $\\tilde { S } \\gets \\{ ( { \\pmb x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { t } ) , \\cdot \\cdot \\cdot , ( { \\pmb x } _ { n } , \\widetilde { y } _ { n } ^ { t } ) \\}$ \n15: end for ", + "bbox": [ + 176, + 136, + 679, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalizing to the multi-class scenario. In multi-class scenario, denote by $f _ { i } ( { \\pmb x } )$ the classifier’s prediction probability of label $i$ . Let $h _ { x }$ be the classifier’s class prediction, i.e., $h _ { \\pmb { x } } = \\arg \\operatorname* { m a x } _ { i } f _ { i } ( \\pmb { x } )$ . We change the $\\begin{array} { r } { | f ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}$ term to the gap between the highest confidence $f _ { h _ { x } } ( \\pmb { x } )$ and the confidence on $\\widetilde { y }$ , $f _ { \\widetilde { y } } ( { \\pmb x } )$ . If the absolute difference between these two confidences is larger than certain threshold $\\theta$ e e, then we correct $\\widetilde { y }$ to $h _ { x }$ . In practice, we find using the difference of logarithms will be more robust. ", + "bbox": [ + 173, + 422, + 826, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 ANALYSIS ", + "text_level": 1, + "bbox": [ + 174, + 515, + 292, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our analysis focuses on the asymptotic case and answers the following question: Given infinitely many data with corrupted labels, is it possible to learn a reasonably good classifier? We show that if the noise satisfies the arguably-general PMD condition, the answer is yes. Assuming mild conditions on the hypothesis class of the machine learning model and the distribution $D$ , we prove that Algorithm 1 obtains a nearly clean classifier. This reduces the challenge of noisy label learning from a realizable problem into a sample complexity problem. In this work we only focus on the asymptotic case, and leave the sample complexity for future work. ", + "bbox": [ + 174, + 539, + 825, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 ASSUMPTIONS ", + "text_level": 1, + "bbox": [ + 174, + 654, + 316, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our first assumption restricts the model to be able to at least approximate the true Bayes classifier $\\eta ( { \\pmb x } )$ . This condition assumes that given a hypothesis class $\\mathcal { H }$ with sufficient complexity, the approximation gap between a classifier $f ( { \\pmb x } )$ in this class and $\\eta ( { \\pmb x } )$ is determined by the inconsistency between the noisy labels and the Bayes optimal classifier. ", + "bbox": [ + 173, + 672, + 826, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 2 (Level set $( \\alpha , \\epsilon )$ -consistency). Suppose data are sampled as $( { \\pmb x } , \\tilde { y } ) \\sim D ( { \\pmb x } , \\tilde { \\eta } ( { \\pmb x } ) )$ and $f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )$ . Given $\\varepsilon < \\textstyle { \\frac { 1 } { 2 } }$ , we call $\\mathcal { H }$ is $( \\alpha , \\epsilon )$ -consistent $i f$ : ", + "bbox": [ + 173, + 732, + 825, + 766 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0221483bf87e70974dc0f3cf4272dfa250b87d8eb4c293a7b95cd75797de6873.jpg", + "text": "$$\n| f ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\leq \\alpha \\mathbb { E } _ { ( \\pmb { z } , \\bar { \\pmb { y } } ) \\sim D ( \\pmb { z } , \\bar { \\eta } ( \\pmb { z } ) ) } [ 1 _ { \\{ \\bar { \\pmb { y } } _ { \\pmb { z } } \\neq \\eta ^ { * } ( \\pmb { z } ) \\} } ( \\pmb { z } ) | | \\eta ( \\pmb { z } ) - \\frac { 1 } { 2 } | \\geq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | ] + \\epsilon .\n$$", + "text_format": "latex", + "bbox": [ + 210, + 771, + 785, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For two input instances $_ { z }$ and $_ { \\textbf { \\em x } }$ such that $\\eta ( z ) ~ > ~ \\eta ( \\pmb { x } )$ (and hence the clean Bayes optimal classifier $\\eta ^ { \\ast } ( \\pmb { x } )$ has higher confidence at $_ z$ than it does at $_ { \\textbf { \\em x } }$ ), the indicator function $\\begin{array} { r } { \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } \\left( z : \\left| \\eta ( z ) - \\frac { 1 } { 2 } \\right| \\ge \\left| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\right| \\right) } \\end{array}$ equals to 1 if the label of the more confident point $_ z$ is inconsistent with $\\eta ^ { * } ( \\pmb { x } )$ . This condition says that the approximation error of the classifier at $_ { \\textbf { \\em x } }$ should be controlled by the risk of $\\eta ^ { * } ( \\cdot )$ at points $_ z$ where $\\eta ^ { * } ( \\cdot )$ is more confident than it is at $_ { \\textbf { \\em x } }$ . ", + "bbox": [ + 173, + 815, + 825, + 890 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We next define a regularity condition of data distribution which describes the continuity of the level set density function. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 3 (Level set bounded distribution). Define the margin $\\begin{array} { r } { t ( { \\pmb x } ) = | \\eta ( { \\pmb x } ) - \\frac { 1 } { 2 } | } \\end{array}$ and $G ( t )$ be the cdf of $t$ : $G ( t ) = \\mathbb { P } _ { \\mathbf { x } \\sim D } ( | \\eta ( \\mathbf { x } ) - \\frac { 1 } { 2 } | \\leq t )$ . Let $g ( t ) = G ^ { \\prime } ( t )$ be the density function of $t$ . We say the distribution $D$ is $( c _ { * } , c ^ { * } )$ -bounded if for all $0 \\leq t \\leq 1 / 2$ , $0 < c _ { * } \\leq g ( t ) \\leq c ^ { * }$ . If $D$ is $( c _ { * } , c ^ { * } )$ -bounded, we define the worst-case density-imbalance ratio of $D$ by $\\begin{array} { r } { \\ell _ { D } . = \\frac { c ^ { * } } { c _ { * } } } \\end{array}$ ", + "bbox": [ + 173, + 102, + 825, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above condition enforces the continuity of level set density function. This is crucial in the analysis since such a continuity allows one to borrow information from its neighborhood region so that a clean neighbor can help correct the corrupted label. To simplify the notation, we will omit $D$ in the subscript when we mention $\\ell$ . From now on, we will assume: ", + "bbox": [ + 173, + 176, + 825, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assumption 1. There exist constants $\\alpha , \\epsilon , c _ { * } , c ^ { * }$ such that the hypothesis class $\\mathcal { H }$ is $( \\alpha , \\epsilon )$ -consistent and the unknown distribution $D$ is $( c _ { * } , c ^ { * } )$ -bounded. ", + "bbox": [ + 171, + 239, + 823, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 MAIN RESULT AND PROOF SKETCH ", + "text_level": 1, + "bbox": [ + 176, + 284, + 462, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we first state our main result, and then present the supporting claims. Complete proofs can be found in the appendix. Our main result below states that if our starting function is trained correctly, i.e., $f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )$ , then Algorithm 1 terminates with most of the final labels matching the Bayes optimal classifier labels. In practice, minimizing true risk is not achievable. Instead, the empirical risk is used to estimate true risk, approaching true risk asymptotically. For a scoring function $f$ , we will denote by $y _ { f ( \\pmb { x } ) } : = \\mathbb { I } ( f ( \\pmb { x } ) \\breve { \\geq } 1 / 2 )$ the label predicted by $f$ . ", + "bbox": [ + 173, + 309, + 825, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1. Under Assumption $\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}$ . Then for the output of Algorithm $^ { l }$ , for any noise $^ { l }$ $\\tau$ with which is PMD with margin $f$ as above and with the following initial- $t _ { 0 }$ , define $e _ { 0 } =$ izations: $( l ) T _ { 0 } < { \\frac { 1 } { 2 } } - e _ { 0 }$ , (2) $\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}$ , (3) $\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { T _ { 0 } } { 3 \\varepsilon } ) , } \\end{array}$ , (4) $T _ { e n d } \\leq 3 \\epsilon$ and (5) $\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}$ , we have: ", + "bbox": [ + 173, + 416, + 825, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/bb59e04dd3b3b7fe323b6aa7c8572b23ae3d6a71998e9e2d6e0ac2bf7b7e18d3.jpg", + "text": "$$\n\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon .\n$$", + "text_format": "latex", + "bbox": [ + 367, + 479, + 629, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the remainder of this section we shall assume that the noise $\\tau$ is PMD with margin $t _ { 0 }$ . To prove our result we first define a “pure” level set. ", + "bbox": [ + 173, + 505, + 823, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 4 (Pure $( e , f , \\eta )$ -level set). $A$ set $L ( e , \\eta ) : = \\{ \\pmb { x } | | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e \\}$ is pure for $f$ if $y _ { f ( \\pmb { x } ) } =$ $\\eta ^ { * } ( \\pmb { x } )$ for all $\\pmb { x } \\in L ( e , \\eta )$ . ", + "bbox": [ + 174, + 535, + 821, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now state a lemma that forms the foundation of our progressive correction algorithm. We show that given a tiny region where the model is reliable, we can move one step forward by trusting the model. Although the improvement is slight in a single round, it empowers a conservatively recursive step in the Algorithm 1. ", + "bbox": [ + 173, + 577, + 825, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 1 (One round purity improvement). Suppose Assumption $^ { l }$ is satisfied, and assume an $f$ such that there exists a pure $( e , f , \\eta )$ -level set with $\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}$ . Let $\\widetilde { \\eta } _ { n e w } ( \\pmb { x } ) = y _ { f ( \\pmb { x } ) } i f \\vert f ( \\pmb { x } ) - 1 / 2 \\vert \\geq e$ and $\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e$ , and assume $f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )$ . Let $e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0$ , $L ( e , \\eta )$ is pure for $f _ { n e w } \\}$ . Then $\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}$ . ", + "bbox": [ + 176, + 636, + 825, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 181, + 686, + 733, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above lemma states that the cleansed region will be enlarged by at least a constant factor. In the following lemma, we justify the functionality of the first $m$ warm-up rounds. Since the initial neural network can behave badly, the region where we can trust the classifier can be very limited. Before starting the flipping procedure in a relatively larger level set, one first needs to expand the initial tiny region $\\mathbf { \\frac { 1 } { 2 } } - e _ { 0 }$ to a constant $T _ { 0 }$ . ", + "bbox": [ + 173, + 712, + 825, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 2 is pure for arm-up r. Given uppose for a given function , after running Algorithm $f _ { 0 }$ thor $L ( e _ { 0 } , \\eta )$ which, there $f _ { 0 }$ $T _ { 0 } < 1 / 2$ $\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}$ exists a level set $L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )$ that is pure for $f _ { 0 }$ . ", + "bbox": [ + 173, + 785, + 825, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next we present our final lemma that combines the previous two lemmata. ", + "bbox": [ + 174, + 843, + 660, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 3. Suppose Assumption $I$ is satisfied, and for a given function $f _ { 0 }$ there exists a level set $L ( e _ { 0 } , \\eta )$ h is , (2) $f _ { 0 }$ ns A, (3) $I$ $f _ { 0 }$ an(4) ons:and $( l )$ $T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }$ $\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}$ $\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) } \\end{array}$ $T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon$ $\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}$ , then we have $\\mathbb { P } _ { \\mathbf { x } \\sim D } [ y _ { f _ { f i n a l } ( \\mathbf { x } ) } = \\eta ^ { * } ( \\pmb { x } ) ] \\geq 1 - 3 c ^ { * } \\epsilon$ . ", + "bbox": [ + 174, + 862, + 825, + 928 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This lemma states that given an initial model that has a reasonably pure super level set, one can manage to progressively correct a large fraction of corrupted labels by running Algorithm 1 for a sufficient long time with carefully chosen parameters. The limit of Algorithm 1 will depend on the approximation ability of the neural network, which is characterized by parameter $\\varepsilon$ in Definition 2. To prove Theorem 1 using Lemma 3, it suffices to get a model which has a reliable region. This is provably achievable by training with a family of good scoring functions on PMD noisy data. ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 207, + 326, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our method on both synthetic and real-world datasets. We first conduct synthetic experiments on two public datasets CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009). To synthesize the label noise, we first approximate the true posterior probability $\\eta$ using the confidence prediction of a clean neural network (trained with the original clean labels). We call these original labels raw labels. Then we sample $y _ { \\pmb { x } } \\sim \\eta ( \\pmb { x } )$ for each instance $_ { \\textbf { \\em x } }$ . Instead of using raw labels, we use these sampled labels $y _ { x }$ as the clean labels, whose posterior probabilities are exactly $\\eta ( { \\pmb x } )$ ; and therefore the neural network is the Bayes optimal classifier $\\eta ^ { * } : \\mathcal { X } \\{ 1 , \\cdot \\cdot , C \\}$ , where $C$ is the number of classes. Note that in multi-class setting, $\\eta ( { \\pmb x } )$ has a vector output and $\\dot { \\eta _ { i } } ( \\pmb { x } )$ is the $i$ -th element of this vector. ", + "bbox": [ + 174, + 238, + 826, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise generation. We consider a generic family of noise. We consider not only feature-dependent noise, but also hybrid noise that consists of both feature-dependent noise and i.i.d. noise. ", + "bbox": [ + 174, + 356, + 821, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For feature-dependent noise, we use three types of noise functions within the PMD noise family. To make the noise challenging enough, for input $_ { \\textbf { \\em x } }$ we always corrupt label from the most confident category $u _ { x }$ to the second confident category ${ s } _ { x }$ , according to $\\eta ( { \\pmb x } )$ . Because $s _ { x }$ is the class that confuses $\\eta ^ { * } ( \\pmb { x } )$ the most, this noise will hurt the network’s performance the most. Note that $y _ { x }$ is sampled from $\\eta ( { \\pmb x } )$ , which has quite an extreme confidence. Thus we generally assume $y _ { x }$ is $u _ { x }$ . For each datum $_ { \\textbf { \\em x } }$ , we only flip it to $s _ { x }$ or keep it as $u _ { x }$ . The three noise functions are as follows: ", + "bbox": [ + 174, + 392, + 825, + 477 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/e090a158d4119449aa64536b038b227fae4ba4256d9894939ae1a4a1a42e836a.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathrm { T y p e - I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = - \\cfrac { 1 } { 2 } \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\cfrac { 1 } { 2 } , \\quad \\mathrm { T y p e - I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } , } \\\\ & { \\mathrm { T y p e - I I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\cfrac { 1 } { 3 } \\left[ \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 481, + 815, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notice that the noise level is determined by the $\\eta ( { \\pmb x } )$ naturally and we cannot control it directly. To change the noise level, we multiply $\\tau _ { u _ { x } , s _ { x } }$ by a certain constant factor such that the final proportion of noise matches our requirement. For PMD noise only, we test noise levels $3 5 \\%$ and $70 \\%$ , meaning that $3 5 \\%$ and $70 \\%$ of the data are corrupted due to the noise, respectively. ", + "bbox": [ + 174, + 546, + 825, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For i.i.d. noise we follow the convention and adopt the commonly used uniform noise and asymmetric noise (Patrini et al., 2017). We artificially corrupt the labels by constructing the noise transition matrix $T$ , where $T _ { i j } = P ( \\widetilde { y } = j | y = i ) = \\tau _ { i j }$ defines the probability that a true label $y = i$ is flipped to $j$ e. Then for each sample with label $i$ , we replace its label with the one sampled from the probability distribution given by the $i$ -th row of matrix $T$ . We consider two kinds of i.i.d. noise in this work. (1) Uniform noise: the true label $i$ is corrupted uniformly to other classes, i.e., $T _ { i j } = \\tau / ( C - 1 )$ for $i \\neq j$ , and $T _ { i i } = 1 - \\tau$ , where $\\tau$ is the constant noise level; (2) Asymmetric noise: the true label $i$ is flipped to $j$ or stays unchanged with probabilities $T _ { i j } = \\tau$ and $T _ { i i } = 1 - \\tau$ , respectively. ", + "bbox": [ + 173, + 609, + 825, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our method with several recently proposed approaches. (1) GCE (Zhang & Sabuncu, 2018); (2) Co-teaching $^ +$ (Yu et al., 2019); (3) SL (Wang et al., 2019); (4) LRT (Zheng et al., 2020). All these methods are generic and handle the label noise without assuming the noise structures. Finally, we also provide the results by standard method, which simply trains the deep network on noisy datasets in a standard manner. ", + "bbox": [ + 174, + 728, + 825, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, we use a batch size of 128 and train the network for 180 epochs to ensure the convergence of all methods. We train the network with SGD optimizer, with initial learning rate 0.01. We randomly repeat the experiments 3 times, and report the mean and standard deviation values. Our code is available at https://github.com/pxiangwu/PLC. ", + "bbox": [ + 174, + 804, + 825, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Table 1 lists the performance of different methods under three types of feature-dependent noise at noise levels $3 5 \\%$ and $70 \\%$ . We observe that our method achieves the best performance across different noise settings. Moreover, notice that some of the baseline methods’ performances are inferior to the standard approach. Possible reasons are that these methods behave too conservatively in dealing with noise. Thus they only make use of a small subset of the original training set, which is not representative enough to grant the model good discriminative ability. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 2 we show the results on datasets corrupted with a combination of feature-dependent noise and i.i.d. noise, which ends up to real noise levels ranging from $50 \\%$ to $70 \\%$ (in terms of the proportion of corrupted labels). I.i.d. noise is overlayed on the feature-dependent noise. Our method outperforms baselines under these more complicated noise patterns. In contrast, when the noise level is high like the cases where we further apply additional $30 \\%$ and $60 \\%$ uniform noise, performances of a few baselines deteriorate and become worse than the standard approach. ", + "bbox": [ + 174, + 138, + 825, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We carry out the ablation studies on hyper-parameters $\\theta _ { 0 }$ (determining the initial confidence threshold for label correction, see Algorithm 1) and $\\beta$ (the step size). In Tables 3 and 4, we show that our method is robust against the choice of $\\theta _ { 0 }$ and $\\beta$ up to a wide range. Notice that to compare against the threshold $\\theta _ { 0 }$ , here we are calculating the absolute difference of $\\log f _ { \\widetilde { y } } ( { \\pmb x } )$ and $\\log f _ { h _ { x } } ( x )$ . As mentioned in Section 2.2, this operation gives a good performance in practice. ", + "bbox": [ + 174, + 229, + 825, + 300 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b90b8a188e7c7188a0dfbb1868b8e64d34ee71f85aab2ed2920c335dfe4546e3.jpg", + "table_caption": [ + "Table 1: Test accuracy $( \\% )$ on CIFAR-10 and CIFAR-100 under different feature-dependent noise types and levels. The average accuracy and standard deviation over 3 trials are reported. " + ], + "table_footnote": [], + "table_body": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I(35%)78.11 ± 0.7479.97 ± 0.1580.65 ±0.3979.76 ± 0.7280.98 ±0.8082.80±0.27
Type-I(70%)41.98 ± 1.9640.69 ± 1.9936.52 ± 1.6236.29 ±0.6641.52 ± 4.5342.74± 2.14
Type-II(35%)76.65±0.5777.34 ± 0.4477.60± 0.8877.92 ±0.8980.74±0.2581.54± 0.47
Type-II(70%)45.57 ± 1.1245.44 ± 0.6440.30 ± 1.4641.11 ± 1.9244.67 ± 3.8946.04±2.20
Type-III (35%)76.89±0.7978.38 ±0.6779.18 ± 0.6178.81 ±0.2981.08 ± 0.3581.50±0.50
CIFAR-100Type-III(70%)43.32 ± 1.0041.90 ± 0.8637.10 ± 0.5938.49 ± 1.4644.47 ± 1.2345.05 ±1.13
Type-I(35%)57.68 ±0.2956.70 ±0.7158.37 ±0.1855.20±0.3356.74 ±0.3460.01±0.43
Type-I(70%)39.32 ± 0.4339.53 ± 0.2840.01 ± 0.7140.02 ±0.8545.29 ± 0.4345.92 ± 0.61
Type-II (35%)57.83 ± 0.2556.57 ± 0.5258.11 ± 1.0556.10±0.7357.25 ±0.6863.68±0.29
Type-II(70%)39.30± 0.3236.84± 0.3937.75 ± 0.4638.45 ± 0.4543.71 ± 0.5145.03 ±0.50
Type-III (35%) Type-II(70%)56.07 ± 0.79 40.01 ± 0.1855.77 ± 0.98 35.37 ± 2.6557.51 ± 1.16 40.53 ±0.6056.04 ± 0.74 39.94 ± 0.8456.57 ±0.30 44.41 ± 0.1963.68 ± 0.29 44.45 ± 0.62
", + "bbox": [ + 174, + 356, + 823, + 496 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/85024251e7bf09e737cf617285400769cb8888555e8e37371d55a7e43087cf09.jpg", + "table_caption": [ + "Table 2: Test accuracy $( \\% )$ on CIFAR-10 and CIFAR-100 under different hybrid noise types and levels. The average accuracy and standard deviation over 3 trials are reported. " + ], + "table_footnote": [], + "table_body": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I+30% Uniform75.26 ± 0.3278.72± 0.5378.08±0.6677.79± 0.4675.97± 0.2779.04±0.50
Type-I+ 60% Uniform64.25 ±0.7855.49 ± 2.1167.43 ± 1.4367.63 ± 1.3659.22 ± 0.7472.21 ± 2.92
Type-I+ 30% Asymmetric75.21 ± 0.6475.43 ± 2.9676.91 ± 0.5677.14± 0.7076.96 ± 0.4578.31 ± 0.41
Type-II +30% Uniform74.92 ± 0.6375.19 ± 0.5475.69 ± 0.2175.08± 0.4775.94 ± 0.5880.08±0.37
Type-II+ 60% Uniform64.02 ±0.6659.89 ±0.6366.39 ± 0.2966.76 ± 1.6058.99 ± 1.4371.21 ± 1.46
Type-II+ 30% Asymmetric74.28 ± 0.3973.37 ± 0.8375.30 ± 0.8175.43 ± 0.4277.03 ± 0.6277.63± 0.30
Type-III+30%Uniform74.00 ± 0.3877.31 ± 0.1177.00 ± 0.1276.22 ± 0.1275.66 ± 0.5780.06 ± 0.47
Type-III+ 60% Uniform63.96 ±0.6956.78± 1.5667.53 ± 0.5167.79 ± 0.5459.36 ± 0.9373.48 ±1.84
Type-III+ 30% Asymmetric75.31 ± 0.3474.62 ± 1.7175.70 ± 0.9176.09 ±0.1077.19 ± 0.7477.54±0.70
CIFAR-100Type-I+30%Uniform48.86±0.5652.33±0.6452.90±0.5351.34±0.6445.66 ± 1.6060.09±0.15
Type-I+ 60% Uniform35.97 ± 1.1227.17 ± 1.6638.62 ± 1.6537.57 ± 0.4323.37 ± 0.7251.68± 0.10
Type-I+ 30% Asymmetric45.85 ± 0.9351.21 ± 0.3152.69 ± 1.1450.18 ± 0.9752.04± 0.15
Type-II+30%Uniform49.32± 0.3651.99 ± 0.7553.61± 0.4643.86 ± 1.3156.40±0.34
Type-II + 60% Uniform35.16 ±0.0525.91 ± 0.6439.58 ±3.1350.58± 0.25 37.93± 0.2223.05 ±0.9960.01±0.63
Type-II + 30% Asymmetric46.50 ± 0.9551.07 ± 1.4451.98 ± 0.3749.46 ± 0.2352.11 ± 0.4649.35 ±1.53
Type-III +30% Uniform48.94± 0.6149.94± 0.4452.07±0.3550.18±0.5442.79 ± 1.7861.43±0.33 60.14±0.97
Type-III + 60% Uniform34.67± 0.1622.89 ± 0.7536.82 ±0.4937.65 ± 1.4222.81 ± 0.7250.73±2.16
Type-III + 30% Asymmetric45.70 ± 0.1249.38 ±0.8650.87 ± 1.1248.15±0.9050.31±0.3954.56±1.11
", + "bbox": [ + 176, + 558, + 823, + 742 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/048bb69b12be8ea61594802e4852ee4b92fc34386c563c9fb6079d903a8c1881.jpg", + "table_caption": [ + "Table 3: The effect of $\\theta _ { 0 }$ on the performance. We use CIFAR-10 with $3 5 \\%$ feature-dependent noise, and set $\\beta = 0 . 1$ . " + ], + "table_footnote": [], + "table_body": "
exp(00)0.20.30.40.5
Type-INoise83.3383.0482.6682.94
Type-II Noise81.8481.1881.0981.24
Type-III Noise81.7981.7581.9882.06
", + "bbox": [ + 178, + 800, + 485, + 853 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/efcc5cfbae92099fc153cb1ca450ce7f28558202ae4f98d60f9018201af565f5.jpg", + "table_caption": [ + "Table 4: The effect of $\\beta$ on the performance. We use CIFAR-10 with $3 5 \\%$ feature-dependent noise, and set ex $) ( \\theta _ { 0 } ) = 0 . 3$ . " + ], + "table_footnote": [], + "table_body": "
β0.050.10.20.3
Type-I Noise83.5883.0483.2883.31
Type-II Noise80.9481.1880.9880.86
Type-III Noise81.9181.7582.1382.39
", + "bbox": [ + 509, + 801, + 816, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on real-world noisy datasets. To test the effectiveness of the proposed method under realworld label noise, we conduct experiments on the Clothing1M dataset (Xiao et al., 2015). This dataset contains 1 million clothing images obtained from online shopping websites with 14 categories. The labels in this dataset are quite noisy with an unknown underlying structure. This dataset provides $5 0 k$ , ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$1 4 k$ and $1 0 k$ manually verified clean data for training, validation and testing, respectively. Following (Tanaka et al., 2018; Yi & Wu, 2019), in our experiment we discard the $5 0 k$ clean training data and evaluate the classification accuracy on the $1 0 k$ clean data. Also, following (Yi & Wu, 2019), we use a randomly sampled pseudo-balanced subset as the training set, which includes about $2 6 0 k$ images. We set the batch size 32, learning rate 0.001, and adopt SGD optimizer and use ResNet-50 with weights pre-trained on ImageNet, as in (Tanaka et al., 2018; Yi & Wu, 2019). ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare our method with the following baselines. (1) Standard; (2) Forward Correction (Patrini et al., 2017); (3) D2L (Ma et al., 2018); (4) JO (Tanaka et al., 2018); (5) PENCIL (Yi & Wu, 2019); (6) DY (Arazo et al., 2019); (7) GCE (Zhang & Sabuncu, 2018); (8) SL (Wang et al., 2019); (9) MLNT (Li et al., 2019); (10) LRT (Zheng et al., 2020). In Table 5 we observe that our method achieves the best performance, suggesting the applicability of our label correction strategy in real-world scenarios. ", + "bbox": [ + 174, + 194, + 825, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Apart from Clothing1M, we also test our method on another smaller dataset, Food-101N (Lee et al., 2018). Food-101N is a dataset for food classification, and consists of 310k training images collected from the web. The estimated label purity is $80 \\%$ . Following (Lee et al., 2018), the classification accuracy is evaluated on the Food-101 (Bossard et al., 2014) testing set, which contains 25k images with curated annotations. We use ResNet-50 pre-trained on ImageNet. We train the network for 30 epochs with SGD optimizer. The batch size is 32 and the initial learning rate is 0.005, which is divided by 10 every 10 epochs. We also adopt simple data augmentation procedures, including random horizontal flip, and resizing the image with a short edge of 256 and then randomly cropping a $2 2 4 \\mathbf { x } 2 2 4$ patch from the resized image. We repeat the experiments with 3 random trials and report the mean value and standard deviation. The results are shown in Table 6. Our method much improves upon the previous approaches. ", + "bbox": [ + 174, + 270, + 825, + 424 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we test our method on a recently proposed real-world dataset, ANIMAL-10N (Song et al., 2019). This dataset contains human-labeled online images for 10 animals with confusing appearance. The estimated label noise rate is $8 \\%$ . There are 50,000 training and 5,000 testing images. Following (Song et al., 2019), we use VGG-19 with batch normalization. The SGD optimizer is employed. Also following (Song et al., 2019), we train the network for 100 epochs and use an initial learning rate of 0.1, which is divided by 5 at $50 \\%$ and $7 5 \\%$ of the total number of epochs. We repeat the experiments with 3 random trials and report the mean value and standard deviation. As is shown in Table 7, our method outperforms the existing baselines. ", + "bbox": [ + 173, + 431, + 825, + 542 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/55cd00a15a47662b4ebbb40d2ac60b6e00cde8207e7807eacee88ab144bca0dc.jpg", + "table_caption": [ + "Table 5: Test accuracy $( \\% )$ on Clothing1M. " + ], + "table_footnote": [], + "table_body": "
MethodStandardForwardD2LJOPENCILDYGCESLMLNTLRTPLC (ours)
Accuracy68.9469.8469.4772.2373.4971.0069.7571.0273.4771.7474.02
", + "bbox": [ + 176, + 573, + 825, + 598 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c46d02ae8ff7d1c13f38ca2f049218de3d9fc3c1f93be2a428df3d5956f2157c.jpg", + "table_caption": [ + "Table 6: Test accuracy $( \\% )$ on Food-101N. " + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Standard81.67
CleanNet (Lee et al., 2018)83.95
PLC (ours)85.28 ± 0.04
", + "bbox": [ + 179, + 633, + 483, + 691 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/99edab4c835c0240b7ed3b062d65e038ecfb04fbb7c282bbf751e0051e904ed9.jpg", + "table_caption": [ + "Table 7: Test accuracy $( \\% )$ on ANIMAL-10N. " + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Standard79.4 ± 0.14
SELFIE (Song et al., 2019)81.8 ± 0.09
PLC (ours)83.4±( 0.43
", + "bbox": [ + 509, + 632, + 815, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 718, + 318, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose a novel family of feature-dependent label noise that is much more general than the traditional i.i.d. noise pattern. Building upon this noise assumption, we propose the first datarecalibrating method that is theoretically guaranteed to converge to a well-behaved classifier. On the synthetic datasets, we show that our method outperforms various baselines under different featuredependent noise patterns subject to our assumption. Also, we test our method on different real-world noisy datasets and observe superior performances over existing approaches. The proposed noise family offers a new theoretical setting for the study of label noise. ", + "bbox": [ + 173, + 750, + 825, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. The authors acknowledge support from US National Science Foundation (NSF) awards CRII-1755791, CCF-1910873, CCF-1855760. This effort was partially supported by the Intelligence Advanced Research Projects Agency (IARPA) under the contract W911NF20C0038. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 102, + 287, + 118 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ehsan Amid, Manfred K Warmuth, and Sriram Srinivasan. Two-temperature logistic regression based on the tsallis divergence. In AISTATS, pp. 2388–2396, 2019a. ", + "bbox": [ + 174, + 126, + 825, + 155 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ehsan Amid, Manfred K. K Warmuth, Rohan Anil, and Tomer Koren. Robust bi-tempered logistic loss based on bregman divergences. In NeurIPS, pp. 15013–15022, 2019b. ", + "bbox": [ + 173, + 164, + 823, + 194 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Veit Andreas, Alldrin Neil, Chechik Gal, Krasin Ivan, Gupta Abhinav, and J. Belongie Serge. Learning from noisy large-scale datasets with minimal supervision. In CVPR, pp. 6575–6583, 2017. ", + "bbox": [ + 173, + 202, + 823, + 232 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Eric Arazo, Diego Ortego, Paul Albert, Noel E O’Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In ICML, 2019. ", + "bbox": [ + 171, + 239, + 823, + 270 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Han Bo, Yao Quanming, Yu Xingrui, Niu Gang, Xu Miao, Hu Weihua, W. Tsang Ivor, and Sugiyama Masashi. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pp. 8536–8546, 2018. ", + "bbox": [ + 178, + 279, + 823, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. In ECCV, pp. 446–461, 2014. ", + "bbox": [ + 173, + 330, + 823, + 361 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nontawat Charoenphakdee, Jongyeong Lee, and Masashi Sugiyama. On symmetric losses for learning from corrupted labels. In ICML, pp. 961–970, 2019. ", + "bbox": [ + 174, + 369, + 821, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chao Chen, Xiuyan Ni, Qinxun Bai, and Yusu Wang. A topological regularizer for classifiers via persistent homology. In AISTATS, pp. 2573–2582, 2019. ", + "bbox": [ + 174, + 407, + 823, + 438 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond classconditional assumption: A primary attempt to combat instance-dependent label noise. In AAAI, 2021. ", + "bbox": [ + 174, + 445, + 823, + 489 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded instance- and label-dependent label noise. In ICML, pp. 1789–1799, 2020. ", + "bbox": [ + 173, + 497, + 823, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hendrycks Dan, Lee Kimin, and Mazeika Mantas. Using pre-training can improve model robustness and uncertainty. In ICML, pp. 2712–2721, 2019. ", + "bbox": [ + 173, + 536, + 821, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jun Du and Zhihua Cai. Modelling class noise with symmetric and asymmetric distributions. In AAAI, pp. 2589–2595, 2015. ", + "bbox": [ + 173, + 574, + 823, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy labels. arXiv preprint arXiv:1607.07526, 2016. ", + "bbox": [ + 173, + 613, + 825, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aritra Ghosh, Naresh Manwani, and P.S. Sastry. Making risk minimization tolerant to label noise. Neurocomput, 160:93–107, 2015. ", + "bbox": [ + 171, + 651, + 825, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aritra Ghosh, Himanshu Kumar, and PS Sastry. Robust loss functions under label noise for deep neural networks. In AAAI, 2017. ", + "bbox": [ + 171, + 689, + 823, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Wei Hu, Zhiyuan Li, and Dingli Yu. Simple and effective regularization methods for training on noisily labeled data with generalization guarantee. In ICLR, 2020. ", + "bbox": [ + 169, + 727, + 823, + 757 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. URL https://www.cs.toronto.edu/˜kriz/learning-features-2009-TR. pdf. ", + "bbox": [ + 176, + 766, + 826, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kuang-Huei Lee, Xiaodong He, Lei Zhang, and Linjun Yang. Cleannet: Transfer learning for scalable image classifier training with label noise. In CVPR, pp. 5447–5456, 2018. ", + "bbox": [ + 171, + 818, + 823, + 848 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In CVPR, pp. 5051–5059, 2019. ", + "bbox": [ + 173, + 856, + 821, + 886 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from noisy labels with distillation. In ICCV, pp. 1928–1936, 2017. ", + "bbox": [ + 174, + 895, + 820, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiang Lu, Zhou Zhengyuan, Leung Thomas, Li Li-Jia, and Fei-Fei Li. Mentornet: Learning datadriven curriculum for very deep neural networks on corrupted labels. In ICML, pp. 2309–2318, 2018. ", + "bbox": [ + 174, + 103, + 825, + 146 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi N. R. Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In ICML, pp. 3361–3370, 2018. ", + "bbox": [ + 176, + 155, + 823, + 198 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Naresh Manwani and PS Sastry. Noise tolerance under risk minimization. IEEE transactions on cybernetics, 43(3):1146–1151, 2013. ", + "bbox": [ + 173, + 207, + 821, + 236 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aditya Krishna Menon, Brendan Van Rooyen, and Nagarajan Natarajan. Learning from binary labels with instance-dependent noise. Machine Learning, 107(8):1561–1595, 2018. ", + "bbox": [ + 173, + 244, + 820, + 273 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. In NeurIPS, pp. 1196–1204, 2013. ", + "bbox": [ + 173, + 282, + 820, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pp. 2233–2241, 2017. ", + "bbox": [ + 174, + 320, + 825, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. Training deep neural networks on noisy labels with bootstrapping. In ICLR Workshop, 2014. ", + "bbox": [ + 176, + 372, + 825, + 415 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, pp. 4331–4340, 2018. ", + "bbox": [ + 174, + 424, + 823, + 454 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss minimization. In ICML, pp. 5739–5748, 2019. ", + "bbox": [ + 173, + 462, + 823, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Metaweight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pp. 1917–1928, 2019. ", + "bbox": [ + 174, + 501, + 823, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In ICML, pp. 5907–5915, 2019. ", + "bbox": [ + 171, + 553, + 825, + 582 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, pp. 5552–5560, 2018. ", + "bbox": [ + 171, + 590, + 823, + 621 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. In ICML, pp. 6234–6243, 2019. ", + "bbox": [ + 171, + 628, + 825, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks. In NeurIPS, pp. 5596–5605, 2017. ", + "bbox": [ + 173, + 666, + 825, + 695 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Brendan Van Rooyen, Aditya Menon, and Robert C Williamson. Learning with symmetric label noise: The importance of being unhinged. In NeurIPS, pp. 10–18, 2015. ", + "bbox": [ + 171, + 704, + 823, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pp. 8688–8696, 2018. ", + "bbox": [ + 171, + 742, + 825, + 772 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In CVPR, pp. 322–330, 2019. ", + "bbox": [ + 173, + 780, + 823, + 810 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. In NeurIPS, 2020. ", + "bbox": [ + 171, + 819, + 823, + 848 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, pp. 2691–2699, 2015. ", + "bbox": [ + 174, + 857, + 823, + 887 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L dmi: A novel information-theoretic loss function for training deep nets robust to label noise. In NeurIPS, pp. 6222–6233, 2019. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yan Yan, Romer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from ´ multiple annotators with varying expertise. Machine learning, 95(3):291–327, 2014. ", + "bbox": [ + 173, + 103, + 823, + 133 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kun Yi and Jianxin Wu. Probabilistic end-to-end noise correction for learning with noisy labels. In CVPR, pp. 7017–7025, 2019. ", + "bbox": [ + 176, + 140, + 823, + 169 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In ICML, pp. 7164–7173, 2019. ", + "bbox": [ + 178, + 176, + 823, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2017. ", + "bbox": [ + 176, + 213, + 823, + 242 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, pp. 8778–8788, 2018. ", + "bbox": [ + 176, + 250, + 821, + 280 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded correction of noisy labels. In ICML, pp. 11447–11457, 2020. ", + "bbox": [ + 176, + 286, + 826, + 315 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 339, + 299, + 356 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lemma 1 (One round purity improvement). Suppose Assumption $^ { l }$ is satisfied, and assume an $f$ such that there exists a pure $( e , f , \\eta )$ -level set with $\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}$ . Let $\\tilde { \\eta } _ { n e w } ( { \\pmb x } ) = y _ { f ( { \\pmb x } ) } i f | f ( { \\pmb x } ) - 1 / 2 | \\ge e$ and $\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e$ , and assume $f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )$ . Let $e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0$ , $L ( e , \\eta )$ is pure for $f _ { n e w } \\}$ . Then $\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}$ . ", + "bbox": [ + 173, + 371, + 826, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proof: We analyze the case where $\\begin{array} { r } { \\eta ( \\pmb { x } ) > \\frac { 1 } { 2 } } \\end{array}$ . The analysis on the other side can be derived similarly. Due to the fact that there exists a level set $L ( e , \\eta )$ pure to $f$ , we have $\\begin{array} { r } { e \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | , \\forall \\pmb { x } } \\end{array}$ : ", + "bbox": [ + 173, + 448, + 826, + 482 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/6608e21c0baa7bea63213fe867bd9128ba2b4062fd173a75a8ba24988cdcd8bc.jpg", + "text": "$$\n\\mathbb { E } _ { ( z , \\tilde { y } ) \\sim ( D , \\tilde { \\eta } _ { n e w } ( z ) ) } \\left[ \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } ( z ) \\bigg | \\bigg | \\eta ( z ) - \\frac { 1 } { 2 } \\bigg | \\ge \\bigg | \\eta ( x ) - \\frac { 1 } { 2 } \\bigg | \\right] = 0 .\n$$", + "text_format": "latex", + "bbox": [ + 267, + 484, + 727, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Now consider $_ { \\textbf { \\em x } }$ where $e - \\gamma \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } |$ . Since the distribution $D$ is $( c _ { * } , c ^ { * } )$ -bounded, we have: ", + "bbox": [ + 176, + 523, + 820, + 541 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/ded246c9b5f3fae10ae340fe6ab0cd4549373b09adcfa8afabf54518949d53ae.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbb { E } _ { ( z \\sim \\hat { y } ) \\sim ( D , \\hat { y } \\sim \\alpha ( z ) ) } [ \\hat { 1 } ( s _ { z \\sim \\hat { y } } , \\hat { x } ) ( z ) ] \\lVert ( z ( z ) - \\frac { 1 } { 2 } \\vert \\geq | y | ( x ) - \\frac { 1 } { 2 } \\vert ] } \\\\ & { = \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\lVert ( z ) - \\frac { 1 } { 2 } [ \\hat { y } \\langle x \\rangle - \\frac { 1 } { 2 } ] } \\\\ & { \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } } { \\lVert \\mathcal { Z } \\rVert ( z ) ^ { 2 } - \\frac { 1 } { 2 } } [ \\frac { \\mathbb { E } _ { \\hat { z } } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) } { 1 - \\frac { 1 } { 2 } } ] } \\\\ & { \\leq \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon - \\epsilon - \\mathbb { E } _ { z } \\cdot \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { \\hat { z } } } { 2 } + \\epsilon } { \\mathcal { Z } \\cdot { \\mathcal { Z } } ( z ) ^ { 2 } - \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } - \\frac { \\epsilon } { 2 } } [ \\mathcal { Z } \\cdot \\mathcal { Y } ( z ) \\leq \\frac { \\frac { 1 } { 2 } } { 2 } - \\epsilon + \\frac { \\mathbb { E } _ { z } } { 1 } ] } } \\\\ & \\quad + \\frac \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\mathcal { Z } ( z ) } \\epsilon - \\frac \\mathbb { E } _ { z } \\cdot \\{ { y } \\cdot \\mathcal { Z } \\} \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } \\cdot \\mathcal { Y } ( z \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 545, + 821, + 804 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "If $\\begin{array} { r } { \\frac { \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) \\le \\gamma \\le \\frac { 2 \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) } \\end{array}$ , the impurity in super level set $\\begin{array} { r } { \\frac { 1 } { 2 } + e - \\gamma } \\end{array}$ is at most $\\frac { 2 \\epsilon } { \\alpha }$ . The level set $( \\alpha , \\varepsilon )$ -consistency condition implies $| f _ { n e w } ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\le 3 \\epsilon$ for $_ { \\textbf { \\em x } }$ s.t. $\\begin{array} { r } { e - \\gamma \\le | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}$ . If $e \\geq 3 \\epsilon$ , ${ \\dot { f } } _ { n e w } ( { \\pmb x } )$ will give the same label as $\\eta ^ { * } ( \\pmb { x } )$ and thus $( e - \\gamma , \\eta )$ -level set becomes pure for $f$ . Meanwhile, the choice of $\\gamma$ ensures that $\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\ell \\alpha } ) ( \\frac { 1 } { 2 } - e ) } \\end{array}$ . □ ", + "bbox": [ + 173, + 811, + 826, + 875 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lemma 2 is pure for arm-up r. Given uppose for a given function, after running Algorithm $f _ { 0 }$ thor $L ( e _ { 0 } , \\eta )$ which, there $f _ { 0 }$ $T _ { 0 } < 1 / 2$ $^ { l }$ $\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}$ exists a level set $L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )$ that is pure for $f _ { 0 }$ . ", + "bbox": [ + 174, + 876, + 825, + 926 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proof: The proof follows from the fact that each round of label flipping improves the purity by a factor of $\\textstyle { \\left( { 1 + { \\frac { \\varepsilon } { \\ell \\alpha } } } \\right) }$ . To obtain an at least $T _ { 0 }$ pure region, it suffices to repeat the flipping step for m ≥ \\`αε log( T01−2e ) rounds. □ ", + "bbox": [ + 173, + 103, + 825, + 151 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lemma 3. Suppose Assumption $I$ is satisfied, and for a given function $f _ { 0 }$ there exists a level set $L ( e _ { 0 } , \\eta )$ h is p, (2) $f _ { 0 }$ ns A, (3) $I$ $f _ { 0 }$ an(4) ons:and $( l )$ $T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }$ $\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}$ $\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) , } \\end{array}$ $T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon$ $\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}$ , then we have $\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon$ . ", + "bbox": [ + 174, + 155, + 825, + 219 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof: The proof can be done by combining Lemma 1 and Lemma 2. In the first $m$ iterations, by Lemma 2, we can guarantee a level set $\\bigl ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta \\bigr )$ pure to $f$ . In the rest of the iterations we ensure the level set $\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}$ is pure. We increase $T$ by a reasonable factor of $\\beta$ to avoid incurring too many corrupted labels while ensuring enough progress in label purification, i.e., $\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}$ such that in the level set $\\begin{array} { r } { | \\eta ( { \\pmb x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}$ we have $| f ( { \\pmb x } ) - \\eta ( { \\pmb x } ) | \\leq 3 \\varepsilon$ . This condition ensures the correctness of flipping when $T \\leq { \\frac { 1 } { 2 } } - 3 \\varepsilon$ . The purity cannot be improved once $\\begin{array} { r } { T \\ge \\frac { 1 } { 2 } - 3 \\varepsilon = T _ { e n d } } \\end{array}$ since there is no guarantee that $f ( { \\pmb x } )$ has consistent label with $\\eta ( { \\pmb x } )$ when $| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | < 3 \\varepsilon$ and $| \\eta ( { \\pmb x } ) - f ( { \\pmb x } ) | \\leq \\bar { 3 } \\varepsilon$ . By $( c _ { * } , c ^ { * } )$ -bounded assumption on $D$ , its mass of impure $3 \\varepsilon$ level set region is at most $3 c ^ { * } \\epsilon$ . □ ", + "bbox": [ + 173, + 228, + 826, + 366 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Theorem 1. Under assumption $I$ , for any noise $\\tau$ which is PMD with margin $t _ { 0 }$ , define $e _ { 0 } =$ $\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}$ . Then for the output of Algorithm $^ { l }$ with $f$ as above and with the following initializations: $( I ) T _ { 0 } < { \\textstyle \\frac { 1 } { 2 } } - e _ { 0 } $ , (2) $\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}$ , (3) $\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log \\bigl ( \\frac { T _ { 0 } } { 3 \\varepsilon } \\bigr ) , } \\end{array}$ , (4) $T _ { e n d } \\leq 3 \\epsilon$ and (5) $\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}$ , we have: ", + "bbox": [ + 173, + 369, + 826, + 435 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/993df6e3b032810b65fd96058668ff86b288a8997eae0b4bd1ec86fd1e575228.jpg", + "text": "$$\n\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon .\n$$", + "text_format": "latex", + "bbox": [ + 367, + 433, + 629, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof: The proof is based on Lemma 3 plus a verification of the existence of $f _ { 0 }$ for which there exists a pure $( e _ { 0 } , f , \\eta )$ -level set. Let: ", + "bbox": [ + 173, + 459, + 823, + 488 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/fcca459cbff85ce5ba955e929ea9650afaabd969859502d5eaa924a5bc2a5788.jpg", + "text": "$$\n\\tau ( { \\pmb x } ) = \\left\\{ \\begin{array} { l l } { \\tau _ { 1 0 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) \\geq 1 / 2 } \\\\ { \\tau _ { 0 1 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) < 1 / 2 } \\end{array} \\right. .\n$$", + "text_format": "latex", + "bbox": [ + 382, + 494, + 612, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the level set $\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } , \\mathbb { P } _ { z } [ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) | | \\eta ( z ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } ] \\leq \\frac { 1 } { 2 } - e _ { 0 } + \\tau ( z ) } \\end{array}$ . By level set $( \\alpha , \\varepsilon )$ -consistency, it suffices to satisfy $\\alpha ( \\textstyle { \\frac { 1 } { 2 } } - e _ { o } + \\tau ) + \\varepsilon \\le e _ { 0 }$ to ensure that $f ( { \\pmb x } )$ has the same prediction with $\\eta ( { \\pmb x } )$ when $| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq \\overline { { e } } _ { 0 }$ . By polynomial level set diminishing noise, we have $\\tau ( \\pmb { x } ) \\leq { \\frac { 1 } { 2 } } - e _ { 0 }$ if $e _ { 0 } > t _ { 0 }$ , and thus by choosing $\\begin{array} { r } { e _ { 0 } = m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}$ one can ensure that initial $f _ { 0 }$ has a pure $( e _ { 0 } , f _ { 0 } , \\eta )$ -level set. The rest of the proof follows from Lemma 3. □ ", + "bbox": [ + 173, + 536, + 825, + 618 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_middle.json b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..671f76f38a86c7cb01ca63240b0efbd8645f534f --- /dev/null +++ b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_middle.json @@ -0,0 +1,46164 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 508, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 508, + 98 + ], + "score": 1.0, + "content": "LEARNING WITH FEATURE-DEPENDENT LABEL NOISE:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 102, + 308, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 308, + 118 + ], + "score": 1.0, + "content": "A PROGRESSIVE APPROACH", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 476, + 147 + ], + "lines": [ + { + "bbox": [ + 112, + 133, + 478, + 149 + ], + "spans": [ + { + "bbox": [ + 112, + 133, + 138, + 149 + ], + "score": 1.0, + "content": "Yikai", + "type": "text" + }, + { + "bbox": [ + 138, + 134, + 175, + 146 + ], + "score": 0.27, + "content": "\\mathbf { Z h a n g ^ { 1 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 133, + 307, + 149 + ], + "score": 1.0, + "content": ", Songzhu Zheng2∗, Pengxiang", + "type": "text" + }, + { + "bbox": [ + 307, + 134, + 332, + 146 + ], + "score": 0.83, + "content": "\\mathbf { W _ { u } } ^ { 1 * }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 133, + 478, + 149 + ], + "score": 1.0, + "content": ", Mayank Goswami3, Chao Chen2", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 115, + 147, + 469, + 180 + ], + "lines": [ + { + "bbox": [ + 112, + 145, + 365, + 160 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 196, + 160 + ], + "score": 1.0, + "content": "1Rutgers University,", + "type": "text" + }, + { + "bbox": [ + 197, + 147, + 274, + 158 + ], + "score": 0.67, + "content": "\\{ \\mathtt { y z 4 2 2 } , \\mathtt { p w 2 4 1 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 145, + 365, + 160 + ], + "score": 1.0, + "content": "@cs.rutgers.edu", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 155, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 470, + 172 + ], + "score": 1.0, + "content": "2Stony Brook University, {zheng.songzhu,chao.chen.1}@stonybrook.edu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 362, + 183 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 362, + 183 + ], + "score": 1.0, + "content": "3City University of New York, mayank.isi@gmail.com", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 209, + 333, + 221 + ], + "lines": [ + { + "bbox": [ + 276, + 209, + 335, + 222 + ], + "spans": [ + { + "bbox": [ + 276, + 209, + 335, + 222 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 235, + 469, + 367 + ], + "lines": [ + { + "bbox": [ + 142, + 236, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 142, + 236, + 469, + 248 + ], + "score": 1.0, + "content": "Label noise is frequently observed in real-world large-scale datasets. The noise is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "score": 1.0, + "content": "introduced due to a variety of reasons; it is heterogeneous and feature-dependent.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "score": 1.0, + "content": "Most existing approaches to handling noisy labels fall into two categories: they", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 469, + 281 + ], + "score": 1.0, + "content": "either assume an ideal feature-independent noise, or remain heuristic without", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 471, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 471, + 293 + ], + "score": 1.0, + "content": "theoretical guarantees. In this paper, we propose to target a new family of feature-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "score": 1.0, + "content": "dependent label noise, which is much more general than commonly used i.i.d. label", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "score": 1.0, + "content": "noise and encompasses a broad spectrum of noise patterns. Focusing on this general", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 313, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 470, + 325 + ], + "score": 1.0, + "content": "noise family, we propose a progressive label correction algorithm that iteratively", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "score": 1.0, + "content": "corrects labels and refines the model. We provide theoretical guarantees showing", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "that for a wide variety of (unknown) noise patterns, a classifier trained with this", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 471, + 359 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 471, + 359 + ], + "score": 1.0, + "content": "strategy converges to be consistent with the Bayes classifier. In experiments, our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "score": 1.0, + "content": "method outperforms SOTA baselines and is robust to various noise types and levels.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 392, + 205, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 208, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 208, + 408 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 418, + 504, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "Addressing noise in training set labels is an important problem in supervised learning. Incorrect", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "annotation of data is inevitable in large-scale data collection, due to intrinsic ambiguity of data/class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "and mistakes of human/automatic annotators (Yan et al., 2014; Andreas et al., 2017). Developing", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 451, + 439, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 439, + 464 + ], + "score": 1.0, + "content": "methods that are resilient to label noise is therefore crucial in real-life applications.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "Classical approaches take a rather simplistic i.i.d. assumption on the label noise, i.e., the label", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 480, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 491 + ], + "score": 1.0, + "content": "corruption is independent and identically distributed and thus is feature-independent. Methods based", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "on this assumption either explicitly estimate the noise pattern (Reed et al., 2014; Patrini et al., 2017;", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "Dan et al., 2019; Xu et al., 2019) or introduce extra regularizer/loss terms (Natarajan et al., 2013;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "score": 1.0, + "content": "Van Rooyen et al., 2015; Xiao et al., 2015; Zhang & Sabuncu, 2018; Ma et al., 2018; Arazo et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "2019; Shen & Sanghavi, 2019). Some results prove that the commonly used losses are naturally", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "score": 1.0, + "content": "robust against such i.i.d. label noise (Manwani & Sastry, 2013; Ghosh et al., 2015; Gao et al., 2016;", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 545, + 371, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 371, + 557 + ], + "score": 1.0, + "content": "Ghosh et al., 2017; Charoenphakdee et al., 2019; Hu et al., 2020).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 562, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Although these methods come with theoretical guarantees, they usually do not perform as well as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "expected in practice due to the unrealistic i.i.d. assumption on noise. This is likely because label", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "noise is heterogeneous and feature-dependent. A cat with an intrinsically ambiguous appearance", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "is more likely to be mislabeled as a dog. An image with poor lighting or severe occlusion can be", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "mislabeled, as important visual clues are imperceptible. Methods that can combat label noise of a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 616, + 431, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 431, + 630 + ], + "score": 1.0, + "content": "much more general form are very much needed to address real-world challenges.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 504, + 711 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 648 + ], + "score": 1.0, + "content": "To adapt to the heterogeneous label noise, state-of-the-arts (SOTAs) often resort to a data-recalibrating", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "strategy. They progressively identify trustworthy data or correct data labels, and then train using these", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "data (Tanaka et al., 2018; Wang et al., 2018; Lu et al., 2018; Li et al., 2019). The models gradually", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "improve as more clean data are collected or more labels are corrected, eventually converging to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "models of high accuracy. These data-recalibrating methods best leverage the learning power of deep", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "neural nets and achieve superior performance in practice. However, their underlying mechanism", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "remains a mystery. No methods in this category can provide theoretical insights as to why the model", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 722, + 196, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 198, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 198, + 733 + ], + "score": 1.0, + "content": "∗Equal contributions.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 508, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 508, + 98 + ], + "score": 1.0, + "content": "LEARNING WITH FEATURE-DEPENDENT LABEL NOISE:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 102, + 308, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 308, + 118 + ], + "score": 1.0, + "content": "A PROGRESSIVE APPROACH", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 476, + 147 + ], + "lines": [ + { + "bbox": [ + 112, + 133, + 478, + 149 + ], + "spans": [ + { + "bbox": [ + 112, + 133, + 138, + 149 + ], + "score": 1.0, + "content": "Yikai", + "type": "text" + }, + { + "bbox": [ + 138, + 134, + 175, + 146 + ], + "score": 0.27, + "content": "\\mathbf { Z h a n g ^ { 1 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 133, + 307, + 149 + ], + "score": 1.0, + "content": ", Songzhu Zheng2∗, Pengxiang", + "type": "text" + }, + { + "bbox": [ + 307, + 134, + 332, + 146 + ], + "score": 0.83, + "content": "\\mathbf { W _ { u } } ^ { 1 * }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 133, + 478, + 149 + ], + "score": 1.0, + "content": ", Mayank Goswami3, Chao Chen2", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 112, + 133, + 478, + 149 + ] + }, + { + "type": "index", + "bbox": [ + 115, + 147, + 469, + 180 + ], + "lines": [ + { + "bbox": [ + 112, + 145, + 365, + 160 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 196, + 160 + ], + "score": 1.0, + "content": "1Rutgers University,", + "type": "text" + }, + { + "bbox": [ + 197, + 147, + 274, + 158 + ], + "score": 0.67, + "content": "\\{ \\mathtt { y z 4 2 2 } , \\mathtt { p w 2 4 1 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 145, + 365, + 160 + ], + "score": 1.0, + "content": "@cs.rutgers.edu", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 155, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 111, + 155, + 470, + 172 + ], + "score": 1.0, + "content": "2Stony Brook University, {zheng.songzhu,chao.chen.1}@stonybrook.edu", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 168, + 362, + 183 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 362, + 183 + ], + "score": 1.0, + "content": "3City University of New York, mayank.isi@gmail.com", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 111, + 145, + 470, + 183 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 209, + 333, + 221 + ], + "lines": [ + { + "bbox": [ + 276, + 209, + 335, + 222 + ], + "spans": [ + { + "bbox": [ + 276, + 209, + 335, + 222 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 235, + 469, + 367 + ], + "lines": [ + { + "bbox": [ + 142, + 236, + 469, + 248 + ], + "spans": [ + { + "bbox": [ + 142, + 236, + 469, + 248 + ], + "score": 1.0, + "content": "Label noise is frequently observed in real-world large-scale datasets. The noise is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "score": 1.0, + "content": "introduced due to a variety of reasons; it is heterogeneous and feature-dependent.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "score": 1.0, + "content": "Most existing approaches to handling noisy labels fall into two categories: they", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 469, + 281 + ], + "score": 1.0, + "content": "either assume an ideal feature-independent noise, or remain heuristic without", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 471, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 471, + 293 + ], + "score": 1.0, + "content": "theoretical guarantees. In this paper, we propose to target a new family of feature-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 303 + ], + "score": 1.0, + "content": "dependent label noise, which is much more general than commonly used i.i.d. label", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "score": 1.0, + "content": "noise and encompasses a broad spectrum of noise patterns. Focusing on this general", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 313, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 470, + 325 + ], + "score": 1.0, + "content": "noise family, we propose a progressive label correction algorithm that iteratively", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "score": 1.0, + "content": "corrects labels and refines the model. We provide theoretical guarantees showing", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "that for a wide variety of (unknown) noise patterns, a classifier trained with this", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 471, + 359 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 471, + 359 + ], + "score": 1.0, + "content": "strategy converges to be consistent with the Bayes classifier. In experiments, our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "score": 1.0, + "content": "method outperforms SOTA baselines and is robust to various noise types and levels.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5, + "bbox_fs": [ + 141, + 236, + 471, + 369 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 392, + 205, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 208, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 208, + 408 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 418, + 504, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "Addressing noise in training set labels is an important problem in supervised learning. Incorrect", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "annotation of data is inevitable in large-scale data collection, due to intrinsic ambiguity of data/class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "and mistakes of human/automatic annotators (Yan et al., 2014; Andreas et al., 2017). Developing", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 451, + 439, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 439, + 464 + ], + "score": 1.0, + "content": "methods that are resilient to label noise is therefore crucial in real-life applications.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 419, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "Classical approaches take a rather simplistic i.i.d. assumption on the label noise, i.e., the label", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 480, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 491 + ], + "score": 1.0, + "content": "corruption is independent and identically distributed and thus is feature-independent. Methods based", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "on this assumption either explicitly estimate the noise pattern (Reed et al., 2014; Patrini et al., 2017;", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "Dan et al., 2019; Xu et al., 2019) or introduce extra regularizer/loss terms (Natarajan et al., 2013;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 524 + ], + "score": 1.0, + "content": "Van Rooyen et al., 2015; Xiao et al., 2015; Zhang & Sabuncu, 2018; Ma et al., 2018; Arazo et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "2019; Shen & Sanghavi, 2019). Some results prove that the commonly used losses are naturally", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 546 + ], + "score": 1.0, + "content": "robust against such i.i.d. label noise (Manwani & Sastry, 2013; Ghosh et al., 2015; Gao et al., 2016;", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 545, + 371, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 371, + 557 + ], + "score": 1.0, + "content": "Ghosh et al., 2017; Charoenphakdee et al., 2019; Hu et al., 2020).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 468, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 562, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Although these methods come with theoretical guarantees, they usually do not perform as well as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "expected in practice due to the unrealistic i.i.d. assumption on noise. This is likely because label", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "noise is heterogeneous and feature-dependent. A cat with an intrinsically ambiguous appearance", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "is more likely to be mislabeled as a dog. An image with poor lighting or severe occlusion can be", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "mislabeled, as important visual clues are imperceptible. Methods that can combat label noise of a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 616, + 431, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 431, + 630 + ], + "score": 1.0, + "content": "much more general form are very much needed to address real-world challenges.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 561, + 506, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 504, + 711 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 648 + ], + "score": 1.0, + "content": "To adapt to the heterogeneous label noise, state-of-the-arts (SOTAs) often resort to a data-recalibrating", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "strategy. They progressively identify trustworthy data or correct data labels, and then train using these", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "data (Tanaka et al., 2018; Wang et al., 2018; Lu et al., 2018; Li et al., 2019). The models gradually", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "improve as more clean data are collected or more labels are corrected, eventually converging to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "models of high accuracy. These data-recalibrating methods best leverage the learning power of deep", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "neural nets and achieve superior performance in practice. However, their underlying mechanism", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "remains a mystery. No methods in this category can provide theoretical insights as to why the model", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "can converge to an ideal one. Thus, these methods require careful hyperparameter tuning and are", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 183, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 183, + 107 + ], + "score": 1.0, + "content": "hard to generalize.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 632, + 506, + 712 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "can converge to an ideal one. Thus, these methods require careful hyperparameter tuning and are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 183, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 183, + 107 + ], + "score": 1.0, + "content": "hard to generalize.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "In this paper, we propose a novel and principled method that specifically targets the heterogeneous,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "feature-dependent label noise. Unlike previous methods, we target a much more general family of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "noise, called Polynomial Margin Diminishing (PMD) label noise. In this noise family, we allow", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "arbitrary noise level except for data far away from the true decision boundary. This is consistent with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "the real-world scenario; data near the decision boundary are harder to distinguish and more likely to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "be mislabeled. Meanwhile, a datum far away from the decision boundary is a typical example of its", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 350, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 350, + 188 + ], + "score": 1.0, + "content": "true class and should have a reasonably bounded noise level.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "Assuming this new PMD noise family, we propose a theoretically-guaranteed data-recalibrating", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "algorithm that gradually corrects labels based on the noisy classifier’s confidence. We start from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "data points with high confidence, and correct the labels of these data using the predictions of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "noisy classifier. Next, the model is improved using cleaned labels. We continue alternating the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "label correction and model improvement until it converges. See Figure 1 for an illustration. Our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "main theorem shows that with a theory-informed criterion for label correction at each iteration,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "the improvement of the label purity is guaranteed. Thus the model is guaranteed to improve with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 269, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 507, + 282 + ], + "score": 1.0, + "content": "sufficient rate through iterations and eventually becomes consistent with the Bayes optimal classifier.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 108, + 287, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "Beside the theoretical strength, we also demonstrate the power of our method in practice. Our method", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "outperforms others on CIFAR-10/100 with various synthetic noise patterns. We also evaluate our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 308, + 433, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 433, + 321 + ], + "score": 1.0, + "content": "method against SOTAs on three real-world datasets with unknown noise patterns.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 325, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 326, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "To the best of our knowledge, our method is the first data-recalibrating method that is theoretically", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "guaranteed to converge to an ideal model. The PMD noise family encompasses a broad spectrum of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "heterogeneous and feature-dependent noise, and better approximates the real-world scenario. It also", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 359, + 360, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 360, + 371 + ], + "score": 1.0, + "content": "provides a novel theoretical setting for the study of label noise.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 375, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "Related works. We review works that do not assume an i.i.d. label noise. Menon et al. (2018)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "score": 1.0, + "content": "generalized the work of (Ghosh et al., 2015) and provided an elegant theoretical framework, showing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "that loss functions fulfilling certain conditions naturally resist instance-dependent noise. The method", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "can achieve even better theoretical properties (i.e., Bayes-consistency) with stronger assumption", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 243, + 432 + ], + "score": 1.0, + "content": "on the clean posterior probability", + "type": "text" + }, + { + "bbox": [ + 244, + 421, + 250, + 431 + ], + "score": 0.68, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 419, + 506, + 432 + ], + "score": 1.0, + "content": ". In practice, this method has not been extended to deep neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "networks. Cheng et al. (2020) proposed an active learning method for instance-dependent label noise.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "image", + "bbox": [ + 121, + 463, + 490, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 463, + 490, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 463, + 490, + 654 + ], + "spans": [ + { + "bbox": [ + 121, + 463, + 490, + 654 + ], + "score": 0.978, + "type": "image", + "image_path": "e1c5bd95d246a39360bb9537b3313db61b0fbc917271c8e5928ac07345e50890.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 121, + 463, + 490, + 526.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 121, + 526.6666666666666, + 490, + 590.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 121, + 590.3333333333333, + 490, + 653.9999999999999 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 663, + 506, + 730 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 663, + 507, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 471, + 676 + ], + "score": 1.0, + "content": "Figure 1: Illustration of the algorithm using synthetic data. (a) Gaussian blob with clean label", + "type": "text" + }, + { + "bbox": [ + 471, + 664, + 502, + 675 + ], + "score": 0.89, + "content": "( \\eta ^ { * } ( \\pmb { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 663, + 507, + 676 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "(b) Data with corrupted labels. (c) Final corrected data. Black dots are the data that have their", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "score": 1.0, + "content": "clean labels. Red dots are the noisy data. Points that remain un-corrected are closer to the decision", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "boundary. Our algorithm corrects most of the noise only using noisy classifier’s confidence. (d) Data", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "score": 1.0, + "content": "after label correction. (e)-(h) We show the intermediate results at different iterations. Gray region is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 718, + 463, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 463, + 730 + ], + "score": 1.0, + "content": "the area where the classifier has high confidence. Labels within this region are corrected.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + } + ], + "index": 33.25 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 107 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "In this paper, we propose a novel and principled method that specifically targets the heterogeneous,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "feature-dependent label noise. Unlike previous methods, we target a much more general family of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "noise, called Polynomial Margin Diminishing (PMD) label noise. In this noise family, we allow", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "arbitrary noise level except for data far away from the true decision boundary. This is consistent with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "the real-world scenario; data near the decision boundary are harder to distinguish and more likely to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "be mislabeled. Meanwhile, a datum far away from the decision boundary is a typical example of its", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 350, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 350, + 188 + ], + "score": 1.0, + "content": "true class and should have a reasonably bounded noise level.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 110, + 506, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "Assuming this new PMD noise family, we propose a theoretically-guaranteed data-recalibrating", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "algorithm that gradually corrects labels based on the noisy classifier’s confidence. We start from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "data points with high confidence, and correct the labels of these data using the predictions of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "noisy classifier. Next, the model is improved using cleaned labels. We continue alternating the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "label correction and model improvement until it converges. See Figure 1 for an illustration. Our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "main theorem shows that with a theory-informed criterion for label correction at each iteration,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "the improvement of the label purity is guaranteed. Thus the model is guaranteed to improve with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 269, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 507, + 282 + ], + "score": 1.0, + "content": "sufficient rate through iterations and eventually becomes consistent with the Bayes optimal classifier.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 192, + 507, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 287, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "Beside the theoretical strength, we also demonstrate the power of our method in practice. Our method", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 309 + ], + "score": 1.0, + "content": "outperforms others on CIFAR-10/100 with various synthetic noise patterns. We also evaluate our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 308, + 433, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 433, + 321 + ], + "score": 1.0, + "content": "method against SOTAs on three real-world datasets with unknown noise patterns.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 287, + 505, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 325, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 326, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "To the best of our knowledge, our method is the first data-recalibrating method that is theoretically", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "guaranteed to converge to an ideal model. The PMD noise family encompasses a broad spectrum of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "heterogeneous and feature-dependent noise, and better approximates the real-world scenario. It also", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 359, + 360, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 360, + 371 + ], + "score": 1.0, + "content": "provides a novel theoretical setting for the study of label noise.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 106, + 326, + 506, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 375, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "Related works. We review works that do not assume an i.i.d. label noise. Menon et al. (2018)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "score": 1.0, + "content": "generalized the work of (Ghosh et al., 2015) and provided an elegant theoretical framework, showing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "that loss functions fulfilling certain conditions naturally resist instance-dependent noise. The method", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "can achieve even better theoretical properties (i.e., Bayes-consistency) with stronger assumption", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 243, + 432 + ], + "score": 1.0, + "content": "on the clean posterior probability", + "type": "text" + }, + { + "bbox": [ + 244, + 421, + 250, + 431 + ], + "score": 0.68, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 419, + 506, + 432 + ], + "score": 1.0, + "content": ". In practice, this method has not been extended to deep neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "networks. Cheng et al. (2020) proposed an active learning method for instance-dependent label noise.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 375, + 506, + 443 + ] + }, + { + "type": "image", + "bbox": [ + 121, + 463, + 490, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 463, + 490, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 121, + 463, + 490, + 654 + ], + "spans": [ + { + "bbox": [ + 121, + 463, + 490, + 654 + ], + "score": 0.978, + "type": "image", + "image_path": "e1c5bd95d246a39360bb9537b3313db61b0fbc917271c8e5928ac07345e50890.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 121, + 463, + 490, + 526.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 121, + 526.6666666666666, + 490, + 590.3333333333333 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 121, + 590.3333333333333, + 490, + 653.9999999999999 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 663, + 506, + 730 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 663, + 507, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 471, + 676 + ], + "score": 1.0, + "content": "Figure 1: Illustration of the algorithm using synthetic data. (a) Gaussian blob with clean label", + "type": "text" + }, + { + "bbox": [ + 471, + 664, + 502, + 675 + ], + "score": 0.89, + "content": "( \\eta ^ { * } ( \\pmb { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 663, + 507, + 676 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "(b) Data with corrupted labels. (c) Final corrected data. Black dots are the data that have their", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "score": 1.0, + "content": "clean labels. Red dots are the noisy data. Points that remain un-corrected are closer to the decision", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "boundary. Our algorithm corrects most of the noise only using noisy classifier’s confidence. (d) Data", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "score": 1.0, + "content": "after label correction. (e)-(h) We show the intermediate results at different iterations. Gray region is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 718, + 463, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 463, + 730 + ], + "score": 1.0, + "content": "the area where the classifier has high confidence. Labels within this region are corrected.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + } + ], + "index": 33.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "The algorithm iteratively queries clean labels from an oracle on carefully selected data. However,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "this approach is not applicable to settings where kosher annotations are unavailable. Another", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "contemporary work (Chen et al., 2021) showed that the noise in real-world dataset is unlikely to be", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "i.i.d., and proposed to fix the noisy labels by averaging the network predictions on each instance over", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "the whole training process. While being effective, their method lacks theoretical guarantees. Chen", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "et al. (2019) showed by regulating the topology of a classifier’s decision boundary, one can improve", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 278, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 278, + 161 + ], + "score": 1.0, + "content": "the model’s robustness against label noise.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "Data-recalibrating methods use noisy networks’ predictions to iteratively select/correct data and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "improve the models. Tanaka et al. (2018) introduced a joint training framework which simultaneously", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "enforces the network to be consistent with its own predictions and corrects the noisy labels during", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "training. Wang et al. (2018) identified noisy labels as outliers based on their label consistencies", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "with surrounding data. Lu et al. (2018) used a curriculum learning strategy where the teacher net", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "is trained on a small kosher dataset to determine if a datum is clean; then the learnt curriculum that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "gives the weight to each datum is fed into the student net for the training and inference. (Yu et al.,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "2019; Bo et al., 2018) trained two synchronized networks; the confidence and consistency of the two", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 251, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 506, + 267 + ], + "score": 1.0, + "content": "networks are utilized to identify clean data. Wu et al. (2020) selected the clean data by investigating", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "the topological structures of the training data in the learned feature space. For completeness, we also", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "refer to other methods of similar design (Li et al., 2017; Vahdat, 2017; Andreas et al., 2017; Ma et al.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 450, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 450, + 298 + ], + "score": 1.0, + "content": "2018; Thulasidasan et al., 2019; Arazo et al., 2019; Shu et al., 2019; Yi & Wu, 2019).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "As for theoretical guarantees, Ren et al. (2018) proposed an algorithm that iteratively re-weights each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "score": 1.0, + "content": "data point by solving an optimization problem. They proved the convergence of the training, but", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "provided no guarantees that the model converges to an ideal one. Amid et al. (2019b) generalized", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "the work of (Amid et al., 2019a) and proposed a tempered matching loss. They showed that when", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "the final softmax layer is replaced by the bi-tempered loss, the resulting classifier will be Bayes", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "consistent. Zheng et al. (2020) proved a one-shot guarantee for their data-recalibrating method; but", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "the convergence of the model is not guaranteed. Our method is the first data-recalibrating method", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 351, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 351, + 392 + ], + "score": 1.0, + "content": "which is guaranteed to converge to a well-behaved classifier.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 107, + 406, + 172, + 419 + ], + "lines": [ + { + "bbox": [ + 104, + 404, + 174, + 422 + ], + "spans": [ + { + "bbox": [ + 104, + 404, + 174, + 422 + ], + "score": 1.0, + "content": "2 METHOD", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "We start by introducing the family of Poly-Margin Diminishing (PMD) label noise. In Section 2.2,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 442, + 488, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 488, + 454 + ], + "score": 1.0, + "content": "we present our main algorithm. Finally, we prove the correctness of our algorithm in Section 3.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 458, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Notations and preliminaries. Although the noise setting and algorithm naturally generalize to multi-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 412, + 481 + ], + "score": 1.0, + "content": "class, for simplicity we focus on binary classification. Let the feature space be", + "type": "text" + }, + { + "bbox": [ + 412, + 470, + 422, + 479 + ], + "score": 0.8, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 469, + 505, + 481 + ], + "score": 1.0, + "content": ". We assume the data", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 479, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 130, + 492 + ], + "score": 0.92, + "content": "( { \\pmb x } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 479, + 304, + 494 + ], + "score": 1.0, + "content": "is sampled from an underlying distribution", + "type": "text" + }, + { + "bbox": [ + 304, + 481, + 314, + 490 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 479, + 327, + 494 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 327, + 480, + 373, + 492 + ], + "score": 0.93, + "content": "\\mathcal { X } \\times \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 479, + 506, + 494 + ], + "score": 1.0, + "content": ". Define the posterior probability", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 107, + 491, + 195, + 504 + ], + "score": 0.92, + "content": "\\eta ( \\pmb { x } ) = \\mathbb { P } [ y = 1 \\mid \\pmb { x } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 491, + 217, + 505 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 217, + 491, + 344, + 504 + ], + "score": 0.91, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 1 \\ | \\ y = 0 , \\pmb { x } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 491, + 363, + 505 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 364, + 491, + 492, + 504 + ], + "score": 0.91, + "content": "\\bar { \\tau _ { 1 , 0 } } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 0 \\ | \\ y = 1 , \\pmb { x } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 491, + 506, + 505 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 215, + 514 + ], + "score": 1.0, + "content": "the noise functions, where", + "type": "text" + }, + { + "bbox": [ + 216, + 503, + 223, + 514 + ], + "score": 0.84, + "content": "\\widetilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 502, + 439, + 514 + ], + "score": 1.0, + "content": "e e denotes the corrupted label. For example, if a datum", + "type": "text" + }, + { + "bbox": [ + 439, + 505, + 447, + 512 + ], + "score": 0.69, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 502, + 506, + 514 + ], + "score": 1.0, + "content": "has true label", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 132, + 525 + ], + "score": 0.89, + "content": "y = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 513, + 160, + 527 + ], + "score": 1.0, + "content": ", it has", + "type": "text" + }, + { + "bbox": [ + 161, + 513, + 191, + 525 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 513, + 377, + 527 + ], + "score": 1.0, + "content": "echance to be corrupted to 1. Similarly, it has", + "type": "text" + }, + { + "bbox": [ + 378, + 514, + 408, + 525 + ], + "score": 0.93, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "chance to be corrupted", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 175, + 537 + ], + "score": 1.0, + "content": "from 1 to 0. Let", + "type": "text" + }, + { + "bbox": [ + 175, + 524, + 261, + 536 + ], + "score": 0.92, + "content": "\\boldsymbol { \\widetilde { \\eta } } ( \\boldsymbol { \\dot { \\mathbf { x } } } ) = \\mathbb { P } [ \\boldsymbol { \\widetilde { y } } = 1 \\mid \\boldsymbol { \\mathbf { x } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 524, + 412, + 537 + ], + "score": 1.0, + "content": "be the noisy posterior probability of", + "type": "text" + }, + { + "bbox": [ + 412, + 524, + 438, + 536 + ], + "score": 0.93, + "content": "\\widetilde y = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 524, + 495, + 537 + ], + "score": 1.0, + "content": "given feature", + "type": "text" + }, + { + "bbox": [ + 495, + 526, + 503, + 534 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 524, + 506, + 537 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 533, + 507, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 122, + 551 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 536, + 200, + 549 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( \\pmb { x } ) = \\mathbb { I } _ { \\{ \\eta ( \\pmb { x } ) \\geq \\frac { 1 } { 2 } \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 533, + 384, + 551 + ], + "score": 1.0, + "content": "ebe the (clean) Bayes optimal classifier, where", + "type": "text" + }, + { + "bbox": [ + 384, + 536, + 396, + 547 + ], + "score": 0.87, + "content": "\\mathbb { I } _ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 533, + 441, + 551 + ], + "score": 1.0, + "content": "eequals 1 if", + "type": "text" + }, + { + "bbox": [ + 442, + 536, + 450, + 545 + ], + "score": 0.76, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 533, + 507, + 551 + ], + "score": 1.0, + "content": "is true, and 0", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 198, + 561 + ], + "score": 1.0, + "content": "otherwise. Finally, let", + "type": "text" + }, + { + "bbox": [ + 198, + 548, + 271, + 561 + ], + "score": 0.92, + "content": "f ( \\pmb { x } ) : \\mathcal { X } [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "be the classifier scoring function (the softmax output of a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 559, + 227, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 227, + 572 + ], + "score": 1.0, + "content": "neural network in this paper).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 109, + 584, + 285, + 595 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 286, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 286, + 596 + ], + "score": 1.0, + "content": "2.1 POLY-MARGIN DIMINISHING NOISE", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 297, + 617 + ], + "score": 1.0, + "content": "We first introduce the family of noise functions", + "type": "text" + }, + { + "bbox": [ + 298, + 607, + 304, + 614 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "this paper will address. We introduce the concept", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 615, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 478, + 627 + ], + "score": 1.0, + "content": "of polynomial margin diminishing noise (PMD noise), which only upper bounds the noise", + "type": "text" + }, + { + "bbox": [ + 479, + 617, + 486, + 625 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 615, + 506, + 627 + ], + "score": 1.0, + "content": "in a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 183, + 639 + ], + "score": 1.0, + "content": "certain level set of", + "type": "text" + }, + { + "bbox": [ + 184, + 627, + 204, + 639 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 626, + 265, + 639 + ], + "score": 1.0, + "content": ", thus allowing", + "type": "text" + }, + { + "bbox": [ + 266, + 628, + 273, + 636 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "to be arbitrarily high outside the restricted domain. This", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "formulation not only covers the feature-independent scenario but also generalizes scenarios proposed", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 648, + 352, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 352, + 661 + ], + "score": 1.0, + "content": "by (Du & Cai, 2015; Menon et al., 2018; Cheng et al., 2020).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 106, + 661, + 502, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 660, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 219, + 675 + ], + "score": 1.0, + "content": "Definition 1 (PMD noise).", + "type": "text" + }, + { + "bbox": [ + 219, + 662, + 228, + 672 + ], + "score": 0.31, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 660, + 325, + 675 + ], + "score": 1.0, + "content": "pair of noise functions", + "type": "text" + }, + { + "bbox": [ + 325, + 661, + 356, + 674 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 660, + 375, + 675 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 376, + 662, + 406, + 674 + ], + "score": 0.92, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 660, + 505, + 675 + ], + "score": 1.0, + "content": "are polynomial-margin", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 673, + 432, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 281, + 688 + ], + "score": 1.0, + "content": "diminishing (PMD), if there exist constants", + "type": "text" + }, + { + "bbox": [ + 281, + 673, + 325, + 687 + ], + "score": 0.93, + "content": "t _ { 0 } \\in ( 0 , \\frac { 1 } { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 673, + 347, + 688 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 347, + 675, + 388, + 686 + ], + "score": 0.9, + "content": "c _ { 1 } , c _ { 2 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 673, + 432, + 688 + ], + "score": 1.0, + "content": "such that:", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + }, + { + "type": "interline_equation", + "bbox": [ + 201, + 688, + 412, + 737 + ], + "lines": [ + { + "bbox": [ + 201, + 688, + 412, + 737 + ], + "spans": [ + { + "bbox": [ + 201, + 688, + 412, + 737 + ], + "score": 0.89, + "content": "\\begin{array} { c l l } { \\displaystyle \\tau _ { 1 , 0 } ( { \\pmb x } ) \\leq c _ { 1 } [ 1 - \\eta ( { \\pmb x } ) ] ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\geq \\displaystyle \\frac { 1 } { 2 } + t _ { 0 } , a n d } \\\\ { \\displaystyle \\tau _ { 0 , 1 } ( { \\pmb x } ) \\leq c _ { 1 } \\eta ( { \\pmb x } ) ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\leq \\displaystyle \\frac { 1 } { 2 } - t _ { 0 } . } \\end{array}", + "type": "interline_equation", + "image_path": "b81884617b954fafce98fcc351dfb3f087b1a26114f5ce5e15a5a23e3cf26faa.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 201, + 688, + 412, + 704.3333333333334 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 201, + 704.3333333333334, + 412, + 720.6666666666667 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 201, + 720.6666666666667, + 412, + 737.0000000000001 + ], + "spans": [], + "index": 50 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "The algorithm iteratively queries clean labels from an oracle on carefully selected data. However,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "this approach is not applicable to settings where kosher annotations are unavailable. Another", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "contemporary work (Chen et al., 2021) showed that the noise in real-world dataset is unlikely to be", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "i.i.d., and proposed to fix the noisy labels by averaging the network predictions on each instance over", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "the whole training process. While being effective, their method lacks theoretical guarantees. Chen", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "et al. (2019) showed by regulating the topology of a classifier’s decision boundary, one can improve", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 278, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 278, + 161 + ], + "score": 1.0, + "content": "the model’s robustness against label noise.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 82, + 506, + 161 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "Data-recalibrating methods use noisy networks’ predictions to iteratively select/correct data and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "improve the models. Tanaka et al. (2018) introduced a joint training framework which simultaneously", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "enforces the network to be consistent with its own predictions and corrects the noisy labels during", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "training. Wang et al. (2018) identified noisy labels as outliers based on their label consistencies", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "with surrounding data. Lu et al. (2018) used a curriculum learning strategy where the teacher net", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "is trained on a small kosher dataset to determine if a datum is clean; then the learnt curriculum that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "gives the weight to each datum is fed into the student net for the training and inference. (Yu et al.,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "2019; Bo et al., 2018) trained two synchronized networks; the confidence and consistency of the two", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 251, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 506, + 267 + ], + "score": 1.0, + "content": "networks are utilized to identify clean data. Wu et al. (2020) selected the clean data by investigating", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "the topological structures of the training data in the learned feature space. For completeness, we also", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "refer to other methods of similar design (Li et al., 2017; Vahdat, 2017; Andreas et al., 2017; Ma et al.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 450, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 450, + 298 + ], + "score": 1.0, + "content": "2018; Thulasidasan et al., 2019; Arazo et al., 2019; Shu et al., 2019; Yi & Wu, 2019).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 165, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "As for theoretical guarantees, Ren et al. (2018) proposed an algorithm that iteratively re-weights each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "score": 1.0, + "content": "data point by solving an optimization problem. They proved the convergence of the training, but", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "provided no guarantees that the model converges to an ideal one. Amid et al. (2019b) generalized", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "the work of (Amid et al., 2019a) and proposed a tempered matching loss. They showed that when", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "the final softmax layer is replaced by the bi-tempered loss, the resulting classifier will be Bayes", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "consistent. Zheng et al. (2020) proved a one-shot guarantee for their data-recalibrating method; but", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "the convergence of the model is not guaranteed. Our method is the first data-recalibrating method", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 351, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 351, + 392 + ], + "score": 1.0, + "content": "which is guaranteed to converge to a well-behaved classifier.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 302, + 506, + 392 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 406, + 172, + 419 + ], + "lines": [ + { + "bbox": [ + 104, + 404, + 174, + 422 + ], + "spans": [ + { + "bbox": [ + 104, + 404, + 174, + 422 + ], + "score": 1.0, + "content": "2 METHOD", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "We start by introducing the family of Poly-Margin Diminishing (PMD) label noise. In Section 2.2,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 442, + 488, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 488, + 454 + ], + "score": 1.0, + "content": "we present our main algorithm. Finally, we prove the correctness of our algorithm in Section 3.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 430, + 506, + 454 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 458, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Notations and preliminaries. Although the noise setting and algorithm naturally generalize to multi-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 412, + 481 + ], + "score": 1.0, + "content": "class, for simplicity we focus on binary classification. Let the feature space be", + "type": "text" + }, + { + "bbox": [ + 412, + 470, + 422, + 479 + ], + "score": 0.8, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 469, + 505, + 481 + ], + "score": 1.0, + "content": ". We assume the data", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 479, + 506, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 130, + 492 + ], + "score": 0.92, + "content": "( { \\pmb x } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 479, + 304, + 494 + ], + "score": 1.0, + "content": "is sampled from an underlying distribution", + "type": "text" + }, + { + "bbox": [ + 304, + 481, + 314, + 490 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 479, + 327, + 494 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 327, + 480, + 373, + 492 + ], + "score": 0.93, + "content": "\\mathcal { X } \\times \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 479, + 506, + 494 + ], + "score": 1.0, + "content": ". Define the posterior probability", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 107, + 491, + 195, + 504 + ], + "score": 0.92, + "content": "\\eta ( \\pmb { x } ) = \\mathbb { P } [ y = 1 \\mid \\pmb { x } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 491, + 217, + 505 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 217, + 491, + 344, + 504 + ], + "score": 0.91, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 1 \\ | \\ y = 0 , \\pmb { x } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 491, + 363, + 505 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 364, + 491, + 492, + 504 + ], + "score": 0.91, + "content": "\\bar { \\tau _ { 1 , 0 } } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 0 \\ | \\ y = 1 , \\pmb { x } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 491, + 506, + 505 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 215, + 514 + ], + "score": 1.0, + "content": "the noise functions, where", + "type": "text" + }, + { + "bbox": [ + 216, + 503, + 223, + 514 + ], + "score": 0.84, + "content": "\\widetilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 502, + 439, + 514 + ], + "score": 1.0, + "content": "e e denotes the corrupted label. For example, if a datum", + "type": "text" + }, + { + "bbox": [ + 439, + 505, + 447, + 512 + ], + "score": 0.69, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 502, + 506, + 514 + ], + "score": 1.0, + "content": "has true label", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 132, + 525 + ], + "score": 0.89, + "content": "y = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 513, + 160, + 527 + ], + "score": 1.0, + "content": ", it has", + "type": "text" + }, + { + "bbox": [ + 161, + 513, + 191, + 525 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 513, + 377, + 527 + ], + "score": 1.0, + "content": "echance to be corrupted to 1. Similarly, it has", + "type": "text" + }, + { + "bbox": [ + 378, + 514, + 408, + 525 + ], + "score": 0.93, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "chance to be corrupted", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 175, + 537 + ], + "score": 1.0, + "content": "from 1 to 0. Let", + "type": "text" + }, + { + "bbox": [ + 175, + 524, + 261, + 536 + ], + "score": 0.92, + "content": "\\boldsymbol { \\widetilde { \\eta } } ( \\boldsymbol { \\dot { \\mathbf { x } } } ) = \\mathbb { P } [ \\boldsymbol { \\widetilde { y } } = 1 \\mid \\boldsymbol { \\mathbf { x } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 524, + 412, + 537 + ], + "score": 1.0, + "content": "be the noisy posterior probability of", + "type": "text" + }, + { + "bbox": [ + 412, + 524, + 438, + 536 + ], + "score": 0.93, + "content": "\\widetilde y = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 524, + 495, + 537 + ], + "score": 1.0, + "content": "given feature", + "type": "text" + }, + { + "bbox": [ + 495, + 526, + 503, + 534 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 524, + 506, + 537 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 533, + 507, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 122, + 551 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 536, + 200, + 549 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( \\pmb { x } ) = \\mathbb { I } _ { \\{ \\eta ( \\pmb { x } ) \\geq \\frac { 1 } { 2 } \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 533, + 384, + 551 + ], + "score": 1.0, + "content": "ebe the (clean) Bayes optimal classifier, where", + "type": "text" + }, + { + "bbox": [ + 384, + 536, + 396, + 547 + ], + "score": 0.87, + "content": "\\mathbb { I } _ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 533, + 441, + 551 + ], + "score": 1.0, + "content": "eequals 1 if", + "type": "text" + }, + { + "bbox": [ + 442, + 536, + 450, + 545 + ], + "score": 0.76, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 533, + 507, + 551 + ], + "score": 1.0, + "content": "is true, and 0", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 198, + 561 + ], + "score": 1.0, + "content": "otherwise. Finally, let", + "type": "text" + }, + { + "bbox": [ + 198, + 548, + 271, + 561 + ], + "score": 0.92, + "content": "f ( \\pmb { x } ) : \\mathcal { X } [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "be the classifier scoring function (the softmax output of a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 559, + 227, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 227, + 572 + ], + "score": 1.0, + "content": "neural network in this paper).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 458, + 507, + 572 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 584, + 285, + 595 + ], + "lines": [ + { + "bbox": [ + 106, + 583, + 286, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 286, + 596 + ], + "score": 1.0, + "content": "2.1 POLY-MARGIN DIMINISHING NOISE", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 297, + 617 + ], + "score": 1.0, + "content": "We first introduce the family of noise functions", + "type": "text" + }, + { + "bbox": [ + 298, + 607, + 304, + 614 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "this paper will address. We introduce the concept", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 615, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 478, + 627 + ], + "score": 1.0, + "content": "of polynomial margin diminishing noise (PMD noise), which only upper bounds the noise", + "type": "text" + }, + { + "bbox": [ + 479, + 617, + 486, + 625 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 615, + 506, + 627 + ], + "score": 1.0, + "content": "in a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 183, + 639 + ], + "score": 1.0, + "content": "certain level set of", + "type": "text" + }, + { + "bbox": [ + 184, + 627, + 204, + 639 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 626, + 265, + 639 + ], + "score": 1.0, + "content": ", thus allowing", + "type": "text" + }, + { + "bbox": [ + 266, + 628, + 273, + 636 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "to be arbitrarily high outside the restricted domain. This", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "formulation not only covers the feature-independent scenario but also generalizes scenarios proposed", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 648, + 352, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 352, + 661 + ], + "score": 1.0, + "content": "by (Du & Cai, 2015; Menon et al., 2018; Cheng et al., 2020).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 603, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 661, + 502, + 686 + ], + "lines": [ + { + "bbox": [ + 105, + 660, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 219, + 675 + ], + "score": 1.0, + "content": "Definition 1 (PMD noise).", + "type": "text" + }, + { + "bbox": [ + 219, + 662, + 228, + 672 + ], + "score": 0.31, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 660, + 325, + 675 + ], + "score": 1.0, + "content": "pair of noise functions", + "type": "text" + }, + { + "bbox": [ + 325, + 661, + 356, + 674 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 660, + 375, + 675 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 376, + 662, + 406, + 674 + ], + "score": 0.92, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 660, + 505, + 675 + ], + "score": 1.0, + "content": "are polynomial-margin", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 673, + 432, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 281, + 688 + ], + "score": 1.0, + "content": "diminishing (PMD), if there exist constants", + "type": "text" + }, + { + "bbox": [ + 281, + 673, + 325, + 687 + ], + "score": 0.93, + "content": "t _ { 0 } \\in ( 0 , \\frac { 1 } { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 673, + 347, + 688 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 347, + 675, + 388, + 686 + ], + "score": 0.9, + "content": "c _ { 1 } , c _ { 2 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 673, + 432, + 688 + ], + "score": 1.0, + "content": "such that:", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 660, + 505, + 688 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 201, + 688, + 412, + 737 + ], + "lines": [ + { + "bbox": [ + 201, + 688, + 412, + 737 + ], + "spans": [ + { + "bbox": [ + 201, + 688, + 412, + 737 + ], + "score": 0.89, + "content": "\\begin{array} { c l l } { \\displaystyle \\tau _ { 1 , 0 } ( { \\pmb x } ) \\leq c _ { 1 } [ 1 - \\eta ( { \\pmb x } ) ] ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\geq \\displaystyle \\frac { 1 } { 2 } + t _ { 0 } , a n d } \\\\ { \\displaystyle \\tau _ { 0 , 1 } ( { \\pmb x } ) \\leq c _ { 1 } \\eta ( { \\pmb x } ) ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\leq \\displaystyle \\frac { 1 } { 2 } - t _ { 0 } . } \\end{array}", + "type": "interline_equation", + "image_path": "b81884617b954fafce98fcc351dfb3f087b1a26114f5ce5e15a5a23e3cf26faa.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 201, + 688, + 412, + 704.3333333333334 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 201, + 704.3333333333334, + 412, + 720.6666666666667 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 201, + 720.6666666666667, + 412, + 737.0000000000001 + ], + "spans": [], + "index": 50 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 239, + 95 + ], + "score": 1.0, + "content": "We abuse notation by referring to", + "type": "text" + }, + { + "bbox": [ + 239, + 84, + 249, + 93 + ], + "score": 0.86, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 81, + 324, + 95 + ], + "score": 1.0, + "content": "as the “margin” of", + "type": "text" + }, + { + "bbox": [ + 324, + 84, + 331, + 92 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ". Note that the PMD condition only requires", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 186, + 106 + ], + "score": 1.0, + "content": "the upper bound on", + "type": "text" + }, + { + "bbox": [ + 186, + 96, + 193, + 104 + ], + "score": 0.73, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "to be polynomial and monotonically decreasing in the region where the Bayes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 275, + 117 + ], + "score": 1.0, + "content": "classifier is fairly confident. For the region", + "type": "text" + }, + { + "bbox": [ + 275, + 104, + 365, + 117 + ], + "score": 0.92, + "content": "\\begin{array} { r } \\{ \\pmb { x } : | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\} < t _ { 0 } \\} \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 104, + 425, + 117 + ], + "score": 1.0, + "content": ", we allow both", + "type": "text" + }, + { + "bbox": [ + 426, + 105, + 456, + 117 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 104, + 474, + 117 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 474, + 104, + 505, + 117 + ], + "score": 0.92, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } )", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "to be arbitrary. Figure 2(d) illustrates the upper bound (orange curve) and a sample noise function", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "(blue curve). We also show the corrupted data according to this noise function (black points are the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 366, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 366, + 149 + ], + "score": 1.0, + "content": "clean data whereas red points are the data with corrupted labels).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "The PMD noise family is much more general than existing noise assumptions. For example, the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "boundary consistent noise (BCN) (Du & Cai, 2015; Menon et al., 2018) assumes a noise function", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "that monotonically decreases as the data are moving away from the decision boundary. See Figure", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "2(c) for an illustration. This noise is much more restrictive compared to our PMD noise which (1)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 506, + 210 + ], + "score": 1.0, + "content": "only requires a monotonic upper bound, and (2) allows arbitrary noise strength in a wide buffer near", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "the decision boundary. Figure 2(b) shows a traditional feature-independent noise pattern (Reed et al.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 275, + 233 + ], + "score": 1.0, + "content": "2014; Patrini et al., 2017), which assumes", + "type": "text" + }, + { + "bbox": [ + 276, + 220, + 306, + 232 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 219, + 332, + 233 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 332, + 220, + 365, + 232 + ], + "score": 0.91, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 219, + 495, + 233 + ], + "score": 1.0, + "content": ") to be a constant independent of", + "type": "text" + }, + { + "bbox": [ + 495, + 222, + 502, + 230 + ], + "score": 0.64, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 219, + 506, + 233 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 113, + 245, + 498, + 403 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 245, + 498, + 403 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 245, + 498, + 403 + ], + "spans": [ + { + "bbox": [ + 113, + 245, + 498, + 403 + ], + "score": 0.973, + "type": "image", + "image_path": "9dade37a8bec0e5a6be0825212c7aaf748d14562d66b3b8371b3b613d5ff59f3.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 113, + 245, + 498, + 297.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 113, + 297.6666666666667, + 498, + 350.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 113, + 350.33333333333337, + 498, + 403.00000000000006 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 412, + 506, + 511 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "Figure 2: Illustration of different noise functions. (a) The original data: Gaussian blob with clean", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 445, + 437 + ], + "score": 1.0, + "content": "labels (by clean label, we refer to the prediction of the Bayes optimal classifier", + "type": "text" + }, + { + "bbox": [ + 446, + 423, + 470, + 435 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 423, + 493, + 437 + ], + "score": 1.0, + "content": ", not", + "type": "text" + }, + { + "bbox": [ + 493, + 425, + 500, + 435 + ], + "score": 0.63, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 423, + 506, + 437 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 190, + 447 + ], + "score": 1.0, + "content": "Confident region of", + "type": "text" + }, + { + "bbox": [ + 190, + 436, + 198, + 446 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 434, + 240, + 447 + ], + "score": 1.0, + "content": "(and thus", + "type": "text" + }, + { + "bbox": [ + 241, + 435, + 249, + 446 + ], + "score": 0.8, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 434, + 381, + 447 + ], + "score": 1.0, + "content": ") in this case is the place where", + "type": "text" + }, + { + "bbox": [ + 382, + 434, + 403, + 446 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "is close to 0 or 1. Blue", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "and green dots correspond to different classes. (b) Uniform label noise: each point has an equal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "score": 1.0, + "content": "probability to be flipped. Red dots are data with corrupted labels; black dots correspond to data that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 383, + 480 + ], + "score": 1.0, + "content": "are not corrupted. (c) BCN noise: the level of noise is decreasing as", + "type": "text" + }, + { + "bbox": [ + 383, + 467, + 408, + 479 + ], + "score": 0.93, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "becomes confident. (d)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "PMD noise: noise level (blue) is only upper bounded by diminishing polynomial function when", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 127, + 501 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 489, + 506, + 501 + ], + "score": 1.0, + "content": "is higher or lower than certain threshold. The upper bound is shown in solid orange curve. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 500, + 447, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 447, + 513 + ], + "score": 1.0, + "content": "dashed orange curve means the noise level near the decision boundary is unbounded.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + } + ], + "index": 17.0 + }, + { + "type": "title", + "bbox": [ + 107, + 534, + 328, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 329, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 329, + 547 + ], + "score": 1.0, + "content": "2.2 THE PROGRESSIVE CORRECTION ALGORITHM", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 556, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "Our algorithm iteratively trains a neural network and corrects labels. We start with a warm-up period,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "score": 1.0, + "content": "in which we train the neural network (NN) with the original noisy data. This allows us to attain a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "reasonable network before it starts fitting noise (Zhang et al., 2017). After the warm-up period, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 473, + 602 + ], + "score": 1.0, + "content": "classifier can be used for label correction. We only correct a label on which the classifier", + "type": "text" + }, + { + "bbox": [ + 473, + 590, + 480, + 601 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "has a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 600, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 506, + 612 + ], + "score": 1.0, + "content": "very high confidence. The intuition is that under the noise assumption, there exists a “pure region”", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 290, + 624 + ], + "score": 1.0, + "content": "in which the prediction of the noisy classifier", + "type": "text" + }, + { + "bbox": [ + 291, + 612, + 298, + 623 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "is highly confident and is consistent with the clean", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 205, + 635 + ], + "score": 1.0, + "content": "Bayes optimal classifier", + "type": "text" + }, + { + "bbox": [ + 205, + 622, + 216, + 633 + ], + "score": 0.87, + "content": "\\eta ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 622, + 505, + 635 + ], + "score": 1.0, + "content": ". Thus the label correction gives clean labels within this pure region. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 259, + 646 + ], + "score": 1.0, + "content": "particular, we select a high threshold", + "type": "text" + }, + { + "bbox": [ + 259, + 633, + 266, + 643 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 633, + 280, + 646 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 280, + 633, + 288, + 644 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 633, + 411, + 646 + ], + "score": 1.0, + "content": "predicts a different label than", + "type": "text" + }, + { + "bbox": [ + 411, + 633, + 418, + 644 + ], + "score": 0.83, + "content": "\\tilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "and its confidence is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 192, + 657 + ], + "score": 1.0, + "content": "above the threshold,", + "type": "text" + }, + { + "bbox": [ + 192, + 644, + 267, + 656 + ], + "score": 0.94, + "content": "| f ( \\bar { \\pmb { x } } ) - 1 / 2 | > \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 643, + 340, + 657 + ], + "score": 1.0, + "content": ", we flip the label", + "type": "text" + }, + { + "bbox": [ + 341, + 644, + 348, + 655 + ], + "score": 0.83, + "content": "\\tilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 643, + 431, + 657 + ], + "score": 1.0, + "content": "to the prediction of", + "type": "text" + }, + { + "bbox": [ + 431, + 644, + 438, + 655 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 643, + 505, + 657 + ], + "score": 1.0, + "content": ". We repeatedly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "correct labels and improve the network until no label is corrected. Next, we slightly decrease the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 145, + 678 + ], + "score": 1.0, + "content": "threshold", + "type": "text" + }, + { + "bbox": [ + 146, + 666, + 151, + 676 + ], + "score": 0.76, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 665, + 505, + 678 + ], + "score": 1.0, + "content": ", use the decreased threshold for label correction, and improve the model accordingly. We", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "continue the process until convergence. For convenience in theoretical analysis, in the algorithm, we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 273, + 700 + ], + "score": 1.0, + "content": "define a continuous increasing threshold", + "type": "text" + }, + { + "bbox": [ + 273, + 689, + 281, + 698 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 688, + 313, + 700 + ], + "score": 1.0, + "content": "and let", + "type": "text" + }, + { + "bbox": [ + 313, + 687, + 367, + 700 + ], + "score": 0.94, + "content": "\\theta = 1 / 2 - T", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 688, + 505, + 700 + ], + "score": 1.0, + "content": ". Our algorithm is summarized in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Algorithm 1. We term our algorithm as PLC (Progressive Label Correction). In Section 3, we will", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "show that this iterative algorithm will converge to be consistent with clean Bayes optimal classifier", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 260, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 132, + 732 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 720, + 260, + 734 + ], + "score": 1.0, + "content": "for most of the input instances.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 33.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 239, + 95 + ], + "score": 1.0, + "content": "We abuse notation by referring to", + "type": "text" + }, + { + "bbox": [ + 239, + 84, + 249, + 93 + ], + "score": 0.86, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 81, + 324, + 95 + ], + "score": 1.0, + "content": "as the “margin” of", + "type": "text" + }, + { + "bbox": [ + 324, + 84, + 331, + 92 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ". Note that the PMD condition only requires", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 186, + 106 + ], + "score": 1.0, + "content": "the upper bound on", + "type": "text" + }, + { + "bbox": [ + 186, + 96, + 193, + 104 + ], + "score": 0.73, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "to be polynomial and monotonically decreasing in the region where the Bayes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 275, + 117 + ], + "score": 1.0, + "content": "classifier is fairly confident. For the region", + "type": "text" + }, + { + "bbox": [ + 275, + 104, + 365, + 117 + ], + "score": 0.92, + "content": "\\begin{array} { r } \\{ \\pmb { x } : | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\} < t _ { 0 } \\} \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 104, + 425, + 117 + ], + "score": 1.0, + "content": ", we allow both", + "type": "text" + }, + { + "bbox": [ + 426, + 105, + 456, + 117 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 104, + 474, + 117 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 474, + 104, + 505, + 117 + ], + "score": 0.92, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } )", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "to be arbitrary. Figure 2(d) illustrates the upper bound (orange curve) and a sample noise function", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "(blue curve). We also show the corrupted data according to this noise function (black points are the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 366, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 366, + 149 + ], + "score": 1.0, + "content": "clean data whereas red points are the data with corrupted labels).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 81, + 506, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "The PMD noise family is much more general than existing noise assumptions. For example, the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "boundary consistent noise (BCN) (Du & Cai, 2015; Menon et al., 2018) assumes a noise function", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "that monotonically decreases as the data are moving away from the decision boundary. See Figure", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "2(c) for an illustration. This noise is much more restrictive compared to our PMD noise which (1)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 506, + 210 + ], + "score": 1.0, + "content": "only requires a monotonic upper bound, and (2) allows arbitrary noise strength in a wide buffer near", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 222 + ], + "score": 1.0, + "content": "the decision boundary. Figure 2(b) shows a traditional feature-independent noise pattern (Reed et al.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 275, + 233 + ], + "score": 1.0, + "content": "2014; Patrini et al., 2017), which assumes", + "type": "text" + }, + { + "bbox": [ + 276, + 220, + 306, + 232 + ], + "score": 0.92, + "content": "\\tau _ { 0 , 1 } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 219, + 332, + 233 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 332, + 220, + 365, + 232 + ], + "score": 0.91, + "content": "\\tau _ { 1 , 0 } ( \\pmb { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 219, + 495, + 233 + ], + "score": 1.0, + "content": ") to be a constant independent of", + "type": "text" + }, + { + "bbox": [ + 495, + 222, + 502, + 230 + ], + "score": 0.64, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 219, + 506, + 233 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 154, + 506, + 233 + ] + }, + { + "type": "image", + "bbox": [ + 113, + 245, + 498, + 403 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 245, + 498, + 403 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 245, + 498, + 403 + ], + "spans": [ + { + "bbox": [ + 113, + 245, + 498, + 403 + ], + "score": 0.973, + "type": "image", + "image_path": "9dade37a8bec0e5a6be0825212c7aaf748d14562d66b3b8371b3b613d5ff59f3.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 113, + 245, + 498, + 297.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 113, + 297.6666666666667, + 498, + 350.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 113, + 350.33333333333337, + 498, + 403.00000000000006 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 412, + 506, + 511 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "Figure 2: Illustration of different noise functions. (a) The original data: Gaussian blob with clean", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 445, + 437 + ], + "score": 1.0, + "content": "labels (by clean label, we refer to the prediction of the Bayes optimal classifier", + "type": "text" + }, + { + "bbox": [ + 446, + 423, + 470, + 435 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 423, + 493, + 437 + ], + "score": 1.0, + "content": ", not", + "type": "text" + }, + { + "bbox": [ + 493, + 425, + 500, + 435 + ], + "score": 0.63, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 423, + 506, + 437 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 190, + 447 + ], + "score": 1.0, + "content": "Confident region of", + "type": "text" + }, + { + "bbox": [ + 190, + 436, + 198, + 446 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 434, + 240, + 447 + ], + "score": 1.0, + "content": "(and thus", + "type": "text" + }, + { + "bbox": [ + 241, + 435, + 249, + 446 + ], + "score": 0.8, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 434, + 381, + 447 + ], + "score": 1.0, + "content": ") in this case is the place where", + "type": "text" + }, + { + "bbox": [ + 382, + 434, + 403, + 446 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "is close to 0 or 1. Blue", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "and green dots correspond to different classes. (b) Uniform label noise: each point has an equal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "score": 1.0, + "content": "probability to be flipped. Red dots are data with corrupted labels; black dots correspond to data that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 383, + 480 + ], + "score": 1.0, + "content": "are not corrupted. (c) BCN noise: the level of noise is decreasing as", + "type": "text" + }, + { + "bbox": [ + 383, + 467, + 408, + 479 + ], + "score": 0.93, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "becomes confident. (d)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "PMD noise: noise level (blue) is only upper bounded by diminishing polynomial function when", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 127, + 501 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 489, + 506, + 501 + ], + "score": 1.0, + "content": "is higher or lower than certain threshold. The upper bound is shown in solid orange curve. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 500, + 447, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 447, + 513 + ], + "score": 1.0, + "content": "dashed orange curve means the noise level near the decision boundary is unbounded.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + } + ], + "index": 17.0 + }, + { + "type": "title", + "bbox": [ + 107, + 534, + 328, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 329, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 329, + 547 + ], + "score": 1.0, + "content": "2.2 THE PROGRESSIVE CORRECTION ALGORITHM", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 556, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "Our algorithm iteratively trains a neural network and corrects labels. We start with a warm-up period,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 579 + ], + "score": 1.0, + "content": "in which we train the neural network (NN) with the original noisy data. This allows us to attain a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "reasonable network before it starts fitting noise (Zhang et al., 2017). After the warm-up period, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 473, + 602 + ], + "score": 1.0, + "content": "classifier can be used for label correction. We only correct a label on which the classifier", + "type": "text" + }, + { + "bbox": [ + 473, + 590, + 480, + 601 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "has a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 600, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 506, + 612 + ], + "score": 1.0, + "content": "very high confidence. The intuition is that under the noise assumption, there exists a “pure region”", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 290, + 624 + ], + "score": 1.0, + "content": "in which the prediction of the noisy classifier", + "type": "text" + }, + { + "bbox": [ + 291, + 612, + 298, + 623 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "is highly confident and is consistent with the clean", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 205, + 635 + ], + "score": 1.0, + "content": "Bayes optimal classifier", + "type": "text" + }, + { + "bbox": [ + 205, + 622, + 216, + 633 + ], + "score": 0.87, + "content": "\\eta ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 622, + 505, + 635 + ], + "score": 1.0, + "content": ". Thus the label correction gives clean labels within this pure region. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 259, + 646 + ], + "score": 1.0, + "content": "particular, we select a high threshold", + "type": "text" + }, + { + "bbox": [ + 259, + 633, + 266, + 643 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 633, + 280, + 646 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 280, + 633, + 288, + 644 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 633, + 411, + 646 + ], + "score": 1.0, + "content": "predicts a different label than", + "type": "text" + }, + { + "bbox": [ + 411, + 633, + 418, + 644 + ], + "score": 0.83, + "content": "\\tilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "and its confidence is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 192, + 657 + ], + "score": 1.0, + "content": "above the threshold,", + "type": "text" + }, + { + "bbox": [ + 192, + 644, + 267, + 656 + ], + "score": 0.94, + "content": "| f ( \\bar { \\pmb { x } } ) - 1 / 2 | > \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 643, + 340, + 657 + ], + "score": 1.0, + "content": ", we flip the label", + "type": "text" + }, + { + "bbox": [ + 341, + 644, + 348, + 655 + ], + "score": 0.83, + "content": "\\tilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 643, + 431, + 657 + ], + "score": 1.0, + "content": "to the prediction of", + "type": "text" + }, + { + "bbox": [ + 431, + 644, + 438, + 655 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 643, + 505, + 657 + ], + "score": 1.0, + "content": ". We repeatedly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "correct labels and improve the network until no label is corrected. Next, we slightly decrease the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 145, + 678 + ], + "score": 1.0, + "content": "threshold", + "type": "text" + }, + { + "bbox": [ + 146, + 666, + 151, + 676 + ], + "score": 0.76, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 665, + 505, + 678 + ], + "score": 1.0, + "content": ", use the decreased threshold for label correction, and improve the model accordingly. We", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "continue the process until convergence. For convenience in theoretical analysis, in the algorithm, we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 273, + 700 + ], + "score": 1.0, + "content": "define a continuous increasing threshold", + "type": "text" + }, + { + "bbox": [ + 273, + 689, + 281, + 698 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 688, + 313, + 700 + ], + "score": 1.0, + "content": "and let", + "type": "text" + }, + { + "bbox": [ + 313, + 687, + 367, + 700 + ], + "score": 0.94, + "content": "\\theta = 1 / 2 - T", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 688, + 505, + 700 + ], + "score": 1.0, + "content": ". Our algorithm is summarized in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Algorithm 1. We term our algorithm as PLC (Progressive Label Correction). In Section 3, we will", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "show that this iterative algorithm will converge to be consistent with clean Bayes optimal classifier", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 260, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 132, + 732 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 720, + 260, + 734 + ], + "score": 1.0, + "content": "for most of the input instances.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 556, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 94, + 330, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 330, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 330, + 107 + ], + "score": 1.0, + "content": "Algorithm 1 Progressive Label Correction", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 108, + 416, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 107, + 413, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 107, + 171, + 123 + ], + "score": 1.0, + "content": "Input: Dataset", + "type": "text" + }, + { + "bbox": [ + 171, + 108, + 293, + 122 + ], + "score": 0.9, + "content": "\\tilde { S } = \\{ ( \\pmb { x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { 0 } ) , \\cdot \\cdot \\cdot , ( \\pmb { x } _ { n } , \\widetilde { y } _ { n } ^ { 0 } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 107, + 340, + 123 + ], + "score": 1.0, + "content": ", initial NN", + "type": "text" + }, + { + "bbox": [ + 340, + 110, + 361, + 122 + ], + "score": 0.88, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 107, + 401, + 123 + ], + "score": 1.0, + "content": ", step size", + "type": "text" + }, + { + "bbox": [ + 401, + 110, + 409, + 121 + ], + "score": 0.74, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 107, + 413, + 123 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 122, + 120, + 381, + 133 + ], + "spans": [ + { + "bbox": [ + 122, + 120, + 227, + 132 + ], + "score": 1.0, + "content": "einitial and end thresholds", + "type": "text" + }, + { + "bbox": [ + 227, + 122, + 268, + 133 + ], + "score": 0.9, + "content": "( T _ { 0 } , T _ { e n d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 120, + 311, + 132 + ], + "score": 1.0, + "content": "e, warm-up", + "type": "text" + }, + { + "bbox": [ + 311, + 123, + 321, + 131 + ], + "score": 0.76, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 120, + 371, + 132 + ], + "score": 1.0, + "content": ", total round", + "type": "text" + }, + { + "bbox": [ + 371, + 121, + 381, + 131 + ], + "score": 0.68, + "content": "N", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 130, + 182, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 145, + 146 + ], + "score": 1.0, + "content": "Output:", + "type": "text" + }, + { + "bbox": [ + 146, + 132, + 182, + 144 + ], + "score": 0.79, + "content": "f _ { f i n a l } ( \\cdot )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 141, + 158, + 155 + ], + "spans": [ + { + "bbox": [ + 109, + 141, + 123, + 155 + ], + "score": 1.0, + "content": "1:", + "type": "text" + }, + { + "bbox": [ + 123, + 143, + 158, + 153 + ], + "score": 0.84, + "content": "T \\gets T _ { 0 }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 151, + 183, + 166 + ], + "spans": [ + { + "bbox": [ + 110, + 151, + 123, + 166 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 124, + 154, + 183, + 165 + ], + "score": 0.87, + "content": "\\theta 1 / 2 - T _ { 0 }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 164, + 211, + 176 + ], + "spans": [ + { + "bbox": [ + 110, + 164, + 138, + 176 + ], + "score": 1.0, + "content": "3: for", + "type": "text" + }, + { + "bbox": [ + 138, + 165, + 197, + 176 + ], + "score": 0.83, + "content": "t \\gets 1 , \\cdots , N", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 164, + 211, + 176 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 176, + 201, + 189 + ], + "spans": [ + { + "bbox": [ + 111, + 177, + 122, + 188 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 131, + 176, + 157, + 188 + ], + "score": 1.0, + "content": "Train", + "type": "text" + }, + { + "bbox": [ + 157, + 177, + 178, + 189 + ], + "score": 0.89, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 176, + 192, + 188 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 192, + 176, + 201, + 187 + ], + "score": 0.8, + "content": "\\tilde { S }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 187, + 324, + 202 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 122, + 201 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 132, + 187, + 162, + 202 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 162, + 188, + 223, + 201 + ], + "score": 0.89, + "content": "( { \\pmb x } _ { i } , \\widetilde { { y } } _ { i } ^ { t - 1 } ) \\in \\tilde { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 187, + 242, + 202 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 188, + 310, + 202 + ], + "score": 0.88, + "content": "\\begin{array} { r } { | f ( \\pmb { x } _ { i } ) - \\frac { 1 } { 2 } | \\geq \\theta } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 187, + 324, + 202 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 202, + 213, + 216 + ], + "spans": [ + { + "bbox": [ + 110, + 203, + 122, + 213 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 142, + 202, + 213, + 216 + ], + "score": 1.0, + "content": "y ti ← I{f (xi)≥ 1 }", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 214, + 167, + 226 + ], + "spans": [ + { + "bbox": [ + 110, + 214, + 122, + 226 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 132, + 214, + 167, + 225 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 225, + 192, + 237 + ], + "spans": [ + { + "bbox": [ + 110, + 226, + 121, + 237 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 132, + 225, + 142, + 237 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 142, + 226, + 169, + 236 + ], + "score": 0.88, + "content": "t \\geq m", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 225, + 192, + 237 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 236, + 200, + 248 + ], + "spans": [ + { + "bbox": [ + 110, + 236, + 122, + 248 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 143, + 236, + 200, + 248 + ], + "score": 0.71, + "content": "\\theta 1 / 2 - T", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 244, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 107, + 249, + 122, + 260 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 147, + 247, + 270, + 260 + ], + "score": 0.54, + "content": "\\mathbf { f } \\forall i \\in [ 1 , \\cdot \\cdot \\cdot , n ] , \\widetilde { y } _ { i } ^ { t } = \\widetilde { y } _ { i } ^ { t - 1 } \\mathbf { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 244, + 287, + 265 + ], + "score": 1.0, + "content": "en", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 263, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 122, + 271 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 153, + 260, + 263, + 271 + ], + "score": 0.78, + "content": "T \\gets \\operatorname* { m i n } ( T ( 1 + \\beta ) , \\bar { T _ { e n d } } )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 269, + 171, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 122, + 282 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 141, + 270, + 171, + 281 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 281, + 162, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 122, + 293 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 132, + 281, + 162, + 292 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 291, + 258, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 122, + 303 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 133, + 291, + 258, + 305 + ], + "score": 0.6, + "content": "\\tilde { S } \\gets \\{ ( { \\pmb x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { t } ) , \\cdot \\cdot \\cdot , ( { \\pmb x } _ { n } , \\widetilde { y } _ { n } ^ { t } ) \\}", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 302, + 157, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 157, + 315 + ], + "score": 1.0, + "content": "15: end for", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 506, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 419, + 348 + ], + "score": 1.0, + "content": "Generalizing to the multi-class scenario. In multi-class scenario, denote by", + "type": "text" + }, + { + "bbox": [ + 420, + 335, + 443, + 347 + ], + "score": 0.9, + "content": "f _ { i } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "the classifier’s", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 225, + 360 + ], + "score": 1.0, + "content": "prediction probability of label", + "type": "text" + }, + { + "bbox": [ + 226, + 347, + 230, + 356 + ], + "score": 0.69, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 345, + 249, + 360 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 250, + 347, + 262, + 357 + ], + "score": 0.89, + "content": "h _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 345, + 415, + 360 + ], + "score": 1.0, + "content": "be the classifier’s class prediction, i.e.,", + "type": "text" + }, + { + "bbox": [ + 415, + 347, + 503, + 358 + ], + "score": 0.84, + "content": "h _ { \\pmb { x } } = \\arg \\operatorname* { m a x } _ { i } f _ { i } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 345, + 506, + 360 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 357, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 165, + 372 + ], + "score": 1.0, + "content": "We change the", + "type": "text" + }, + { + "bbox": [ + 166, + 357, + 212, + 371 + ], + "score": 0.94, + "content": "\\begin{array} { r } { | f ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 357, + 398, + 372 + ], + "score": 1.0, + "content": "term to the gap between the highest confidence", + "type": "text" + }, + { + "bbox": [ + 398, + 358, + 428, + 370 + ], + "score": 0.93, + "content": "f _ { h _ { x } } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 357, + 506, + 372 + ], + "score": 1.0, + "content": "and the confidence", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 369, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 118, + 384 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 370, + 126, + 382 + ], + "score": 0.56, + "content": "\\widetilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 369, + 129, + 384 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 129, + 370, + 154, + 383 + ], + "score": 0.82, + "content": "f _ { \\widetilde { y } } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 369, + 506, + 384 + ], + "score": 1.0, + "content": ". If the absolute difference between these two confidences is larger than certain threshold", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 381, + 507, + 394 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 113, + 391 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 381, + 178, + 394 + ], + "score": 1.0, + "content": "e e, then we correct", + "type": "text" + }, + { + "bbox": [ + 179, + 381, + 186, + 393 + ], + "score": 0.85, + "content": "\\widetilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 381, + 196, + 394 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 197, + 382, + 209, + 392 + ], + "score": 0.88, + "content": "h _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 381, + 507, + 394 + ], + "score": 1.0, + "content": ". In practice, we find using the difference of logarithms will be more robust.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 408, + 179, + 421 + ], + "lines": [ + { + "bbox": [ + 104, + 406, + 180, + 423 + ], + "spans": [ + { + "bbox": [ + 104, + 406, + 180, + 423 + ], + "score": 1.0, + "content": "3 ANALYSIS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "Our analysis focuses on the asymptotic case and answers the following question: Given infinitely", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "many data with corrupted labels, is it possible to learn a reasonably good classifier? We show", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 448, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 463 + ], + "score": 1.0, + "content": "that if the noise satisfies the arguably-general PMD condition, the answer is yes. Assuming mild", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 452, + 473 + ], + "score": 1.0, + "content": "conditions on the hypothesis class of the machine learning model and the distribution", + "type": "text" + }, + { + "bbox": [ + 452, + 461, + 461, + 470 + ], + "score": 0.79, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 460, + 505, + 473 + ], + "score": 1.0, + "content": ", we prove", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "score": 1.0, + "content": "that Algorithm 1 obtains a nearly clean classifier. This reduces the challenge of noisy label learning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "from a realizable problem into a sample complexity problem. In this work we only focus on the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 494, + 373, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 373, + 506 + ], + "score": 1.0, + "content": "asymptotic case, and leave the sample complexity for future work.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 107, + 518, + 194, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 517, + 195, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 195, + 531 + ], + "score": 1.0, + "content": "3.1 ASSUMPTIONS", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 533, + 506, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 482, + 547 + ], + "score": 1.0, + "content": "Our first assumption restricts the model to be able to at least approximate the true Bayes classifier", + "type": "text" + }, + { + "bbox": [ + 483, + 534, + 503, + 546 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 532, + 507, + 547 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 313, + 557 + ], + "score": 1.0, + "content": "This condition assumes that given a hypothesis class", + "type": "text" + }, + { + "bbox": [ + 314, + 545, + 324, + 555 + ], + "score": 0.8, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "with sufficient complexity, the approximation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 204, + 568 + ], + "score": 1.0, + "content": "gap between a classifier", + "type": "text" + }, + { + "bbox": [ + 204, + 556, + 225, + 568 + ], + "score": 0.92, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 556, + 292, + 568 + ], + "score": 1.0, + "content": "in this class and", + "type": "text" + }, + { + "bbox": [ + 293, + 555, + 313, + 568 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "is determined by the inconsistency between the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 567, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 287, + 578 + ], + "score": 1.0, + "content": "noisy labels and the Bayes optimal classifier.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 580, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 200, + 593 + ], + "score": 1.0, + "content": "Definition 2 (Level set", + "type": "text" + }, + { + "bbox": [ + 201, + 580, + 224, + 592 + ], + "score": 0.89, + "content": "( \\alpha , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 579, + 401, + 593 + ], + "score": 1.0, + "content": "-consistency). Suppose data are sampled as", + "type": "text" + }, + { + "bbox": [ + 402, + 580, + 486, + 592 + ], + "score": 0.92, + "content": "( { \\pmb x } , \\tilde { y } ) \\sim D ( { \\pmb x } , \\tilde { \\eta } ( { \\pmb x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 579, + 506, + 593 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 589, + 494, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 302, + 608 + ], + "score": 0.87, + "content": "f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 589, + 332, + 609 + ], + "score": 1.0, + "content": ". Given", + "type": "text" + }, + { + "bbox": [ + 333, + 591, + 358, + 604 + ], + "score": 0.92, + "content": "\\varepsilon < \\textstyle { \\frac { 1 } { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 589, + 393, + 609 + ], + "score": 1.0, + "content": ", we call", + "type": "text" + }, + { + "bbox": [ + 393, + 592, + 403, + 601 + ], + "score": 0.78, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 589, + 414, + 609 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 414, + 592, + 437, + 603 + ], + "score": 0.91, + "content": "( \\alpha , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 589, + 482, + 609 + ], + "score": 1.0, + "content": "-consistent", + "type": "text" + }, + { + "bbox": [ + 482, + 592, + 488, + 603 + ], + "score": 0.53, + "content": "i f", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 589, + 494, + 609 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 129, + 611, + 481, + 639 + ], + "lines": [ + { + "bbox": [ + 129, + 611, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 129, + 611, + 481, + 639 + ], + "score": 0.88, + "content": "| f ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\leq \\alpha \\mathbb { E } _ { ( \\pmb { z } , \\bar { \\pmb { y } } ) \\sim D ( \\pmb { z } , \\bar { \\eta } ( \\pmb { z } ) ) } [ 1 _ { \\{ \\bar { \\pmb { y } } _ { \\pmb { z } } \\neq \\eta ^ { * } ( \\pmb { z } ) \\} } ( \\pmb { z } ) | | \\eta ( \\pmb { z } ) - \\frac { 1 } { 2 } | \\geq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | ] + \\epsilon .", + "type": "interline_equation", + "image_path": "0221483bf87e70974dc0f3cf4272dfa250b87d8eb4c293a7b95cd75797de6873.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 129, + 611, + 481, + 639 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 213, + 659 + ], + "score": 1.0, + "content": "For two input instances", + "type": "text" + }, + { + "bbox": [ + 213, + 649, + 221, + 657 + ], + "score": 0.76, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 645, + 244, + 659 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 649, + 252, + 657 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 645, + 299, + 659 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 300, + 646, + 362, + 659 + ], + "score": 0.93, + "content": "\\eta ( z ) ~ > ~ \\eta ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 645, + 506, + 659 + ], + "score": 1.0, + "content": "(and hence the clean Bayes op-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 173, + 670 + ], + "score": 1.0, + "content": "timal classifier", + "type": "text" + }, + { + "bbox": [ + 173, + 658, + 199, + 669 + ], + "score": 0.9, + "content": "\\eta ^ { \\ast } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 657, + 312, + 670 + ], + "score": 1.0, + "content": "has higher confidence at", + "type": "text" + }, + { + "bbox": [ + 313, + 660, + 321, + 668 + ], + "score": 0.79, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 657, + 392, + 670 + ], + "score": 1.0, + "content": "than it does at", + "type": "text" + }, + { + "bbox": [ + 392, + 659, + 400, + 668 + ], + "score": 0.7, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "), the indicator function", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 108, + 668, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 108, + 669, + 283, + 683 + ], + "score": 0.78, + "content": "\\begin{array} { r } { \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } \\left( z : \\left| \\eta ( z ) - \\frac { 1 } { 2 } \\right| \\ge \\left| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\right| \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 668, + 506, + 683 + ], + "score": 1.0, + "content": "equals to 1 if the label of the more confident point", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 682, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 107, + 684, + 114, + 691 + ], + "score": 0.7, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 682, + 195, + 694 + ], + "score": 1.0, + "content": "is inconsistent with", + "type": "text" + }, + { + "bbox": [ + 195, + 682, + 221, + 693 + ], + "score": 0.89, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 682, + 496, + 694 + ], + "score": 1.0, + "content": ". This condition says that the approximation error of the classifier at", + "type": "text" + }, + { + "bbox": [ + 496, + 685, + 504, + 691 + ], + "score": 0.64, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 692, + 494, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 245, + 705 + ], + "score": 1.0, + "content": "should be controlled by the risk of", + "type": "text" + }, + { + "bbox": [ + 246, + 693, + 267, + 705 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 692, + 305, + 705 + ], + "score": 1.0, + "content": "at points", + "type": "text" + }, + { + "bbox": [ + 305, + 694, + 312, + 703 + ], + "score": 0.77, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 692, + 340, + 705 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 341, + 693, + 362, + 705 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 692, + 482, + 705 + ], + "score": 1.0, + "content": "is more confident than it is at", + "type": "text" + }, + { + "bbox": [ + 482, + 695, + 489, + 703 + ], + "score": 0.76, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 692, + 494, + 705 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "We next define a regularity condition of data distribution which describes the continuity of the level", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 189, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 189, + 732 + ], + "score": 1.0, + "content": "set density function.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 94, + 330, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 330, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 330, + 107 + ], + "score": 1.0, + "content": "Algorithm 1 Progressive Label Correction", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "index", + "bbox": [ + 108, + 108, + 416, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 107, + 413, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 107, + 171, + 123 + ], + "score": 1.0, + "content": "Input: Dataset", + "type": "text" + }, + { + "bbox": [ + 171, + 108, + 293, + 122 + ], + "score": 0.9, + "content": "\\tilde { S } = \\{ ( \\pmb { x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { 0 } ) , \\cdot \\cdot \\cdot , ( \\pmb { x } _ { n } , \\widetilde { y } _ { n } ^ { 0 } ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 107, + 340, + 123 + ], + "score": 1.0, + "content": ", initial NN", + "type": "text" + }, + { + "bbox": [ + 340, + 110, + 361, + 122 + ], + "score": 0.88, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 107, + 401, + 123 + ], + "score": 1.0, + "content": ", step size", + "type": "text" + }, + { + "bbox": [ + 401, + 110, + 409, + 121 + ], + "score": 0.74, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 107, + 413, + 123 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 122, + 120, + 381, + 133 + ], + "spans": [ + { + "bbox": [ + 122, + 120, + 227, + 132 + ], + "score": 1.0, + "content": "einitial and end thresholds", + "type": "text" + }, + { + "bbox": [ + 227, + 122, + 268, + 133 + ], + "score": 0.9, + "content": "( T _ { 0 } , T _ { e n d } )", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 120, + 311, + 132 + ], + "score": 1.0, + "content": "e, warm-up", + "type": "text" + }, + { + "bbox": [ + 311, + 123, + 321, + 131 + ], + "score": 0.76, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 120, + 371, + 132 + ], + "score": 1.0, + "content": ", total round", + "type": "text" + }, + { + "bbox": [ + 371, + 121, + 381, + 131 + ], + "score": 0.68, + "content": "N", + "type": "inline_equation" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 130, + 182, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 145, + 146 + ], + "score": 1.0, + "content": "Output:", + "type": "text" + }, + { + "bbox": [ + 146, + 132, + 182, + 144 + ], + "score": 0.79, + "content": "f _ { f i n a l } ( \\cdot )", + "type": "inline_equation" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 141, + 158, + 155 + ], + "spans": [ + { + "bbox": [ + 109, + 141, + 123, + 155 + ], + "score": 1.0, + "content": "1:", + "type": "text" + }, + { + "bbox": [ + 123, + 143, + 158, + 153 + ], + "score": 0.84, + "content": "T \\gets T _ { 0 }", + "type": "inline_equation" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 151, + 183, + 166 + ], + "spans": [ + { + "bbox": [ + 110, + 151, + 123, + 166 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 124, + 154, + 183, + 165 + ], + "score": 0.87, + "content": "\\theta 1 / 2 - T _ { 0 }", + "type": "inline_equation" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 164, + 211, + 176 + ], + "spans": [ + { + "bbox": [ + 110, + 164, + 138, + 176 + ], + "score": 1.0, + "content": "3: for", + "type": "text" + }, + { + "bbox": [ + 138, + 165, + 197, + 176 + ], + "score": 0.83, + "content": "t \\gets 1 , \\cdots , N", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 164, + 211, + 176 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 176, + 201, + 189 + ], + "spans": [ + { + "bbox": [ + 111, + 177, + 122, + 188 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 131, + 176, + 157, + 188 + ], + "score": 1.0, + "content": "Train", + "type": "text" + }, + { + "bbox": [ + 157, + 177, + 178, + 189 + ], + "score": 0.89, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 176, + 192, + 188 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 192, + 176, + 201, + 187 + ], + "score": 0.8, + "content": "\\tilde { S }", + "type": "inline_equation" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 187, + 324, + 202 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 122, + 201 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 132, + 187, + 162, + 202 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 162, + 188, + 223, + 201 + ], + "score": 0.89, + "content": "( { \\pmb x } _ { i } , \\widetilde { { y } } _ { i } ^ { t - 1 } ) \\in \\tilde { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 187, + 242, + 202 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 188, + 310, + 202 + ], + "score": 0.88, + "content": "\\begin{array} { r } { | f ( \\pmb { x } _ { i } ) - \\frac { 1 } { 2 } | \\geq \\theta } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 187, + 324, + 202 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 202, + 213, + 216 + ], + "spans": [ + { + "bbox": [ + 110, + 203, + 122, + 213 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 142, + 202, + 213, + 216 + ], + "score": 1.0, + "content": "y ti ← I{f (xi)≥ 1 }", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 214, + 167, + 226 + ], + "spans": [ + { + "bbox": [ + 110, + 214, + 122, + 226 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 132, + 214, + 167, + 225 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 225, + 192, + 237 + ], + "spans": [ + { + "bbox": [ + 110, + 226, + 121, + 237 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 132, + 225, + 142, + 237 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 142, + 226, + 169, + 236 + ], + "score": 0.88, + "content": "t \\geq m", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 225, + 192, + 237 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 236, + 200, + 248 + ], + "spans": [ + { + "bbox": [ + 110, + 236, + 122, + 248 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 143, + 236, + 200, + 248 + ], + "score": 0.71, + "content": "\\theta 1 / 2 - T", + "type": "inline_equation" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 107, + 244, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 107, + 249, + 122, + 260 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 147, + 247, + 270, + 260 + ], + "score": 0.54, + "content": "\\mathbf { f } \\forall i \\in [ 1 , \\cdot \\cdot \\cdot , n ] , \\widetilde { y } _ { i } ^ { t } = \\widetilde { y } _ { i } ^ { t - 1 } \\mathbf { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 244, + 287, + 265 + ], + "score": 1.0, + "content": "en", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 260, + 263, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 122, + 271 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 153, + 260, + 263, + 271 + ], + "score": 0.78, + "content": "T \\gets \\operatorname* { m i n } ( T ( 1 + \\beta ) , \\bar { T _ { e n d } } )", + "type": "inline_equation" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 269, + 171, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 122, + 282 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 141, + 270, + 171, + 281 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 281, + 162, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 122, + 293 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 132, + 281, + 162, + 292 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 291, + 258, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 122, + 303 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 133, + 291, + 258, + 305 + ], + "score": 0.6, + "content": "\\tilde { S } \\gets \\{ ( { \\pmb x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { t } ) , \\cdot \\cdot \\cdot , ( { \\pmb x } _ { n } , \\widetilde { y } _ { n } ^ { t } ) \\}", + "type": "inline_equation" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 302, + 157, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 157, + 315 + ], + "score": 1.0, + "content": "15: end for", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 107, + 413, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 506, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 419, + 348 + ], + "score": 1.0, + "content": "Generalizing to the multi-class scenario. In multi-class scenario, denote by", + "type": "text" + }, + { + "bbox": [ + 420, + 335, + 443, + 347 + ], + "score": 0.9, + "content": "f _ { i } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "the classifier’s", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 225, + 360 + ], + "score": 1.0, + "content": "prediction probability of label", + "type": "text" + }, + { + "bbox": [ + 226, + 347, + 230, + 356 + ], + "score": 0.69, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 345, + 249, + 360 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 250, + 347, + 262, + 357 + ], + "score": 0.89, + "content": "h _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 345, + 415, + 360 + ], + "score": 1.0, + "content": "be the classifier’s class prediction, i.e.,", + "type": "text" + }, + { + "bbox": [ + 415, + 347, + 503, + 358 + ], + "score": 0.84, + "content": "h _ { \\pmb { x } } = \\arg \\operatorname* { m a x } _ { i } f _ { i } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 345, + 506, + 360 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 357, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 165, + 372 + ], + "score": 1.0, + "content": "We change the", + "type": "text" + }, + { + "bbox": [ + 166, + 357, + 212, + 371 + ], + "score": 0.94, + "content": "\\begin{array} { r } { | f ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 357, + 398, + 372 + ], + "score": 1.0, + "content": "term to the gap between the highest confidence", + "type": "text" + }, + { + "bbox": [ + 398, + 358, + 428, + 370 + ], + "score": 0.93, + "content": "f _ { h _ { x } } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 357, + 506, + 372 + ], + "score": 1.0, + "content": "and the confidence", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 369, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 118, + 384 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 370, + 126, + 382 + ], + "score": 0.56, + "content": "\\widetilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 369, + 129, + 384 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 129, + 370, + 154, + 383 + ], + "score": 0.82, + "content": "f _ { \\widetilde { y } } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 369, + 506, + 384 + ], + "score": 1.0, + "content": ". If the absolute difference between these two confidences is larger than certain threshold", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 381, + 507, + 394 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 113, + 391 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 381, + 178, + 394 + ], + "score": 1.0, + "content": "e e, then we correct", + "type": "text" + }, + { + "bbox": [ + 179, + 381, + 186, + 393 + ], + "score": 0.85, + "content": "\\widetilde { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 381, + 196, + 394 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 197, + 382, + 209, + 392 + ], + "score": 0.88, + "content": "h _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 381, + 507, + 394 + ], + "score": 1.0, + "content": ". In practice, we find using the difference of logarithms will be more robust.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 335, + 507, + 394 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 408, + 179, + 421 + ], + "lines": [ + { + "bbox": [ + 104, + 406, + 180, + 423 + ], + "spans": [ + { + "bbox": [ + 104, + 406, + 180, + 423 + ], + "score": 1.0, + "content": "3 ANALYSIS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "Our analysis focuses on the asymptotic case and answers the following question: Given infinitely", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "many data with corrupted labels, is it possible to learn a reasonably good classifier? We show", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 448, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 463 + ], + "score": 1.0, + "content": "that if the noise satisfies the arguably-general PMD condition, the answer is yes. Assuming mild", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 452, + 473 + ], + "score": 1.0, + "content": "conditions on the hypothesis class of the machine learning model and the distribution", + "type": "text" + }, + { + "bbox": [ + 452, + 461, + 461, + 470 + ], + "score": 0.79, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 460, + 505, + 473 + ], + "score": 1.0, + "content": ", we prove", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "score": 1.0, + "content": "that Algorithm 1 obtains a nearly clean classifier. This reduces the challenge of noisy label learning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "from a realizable problem into a sample complexity problem. In this work we only focus on the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 494, + 373, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 373, + 506 + ], + "score": 1.0, + "content": "asymptotic case, and leave the sample complexity for future work.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 426, + 506, + 506 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 518, + 194, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 517, + 195, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 195, + 531 + ], + "score": 1.0, + "content": "3.1 ASSUMPTIONS", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 533, + 506, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 482, + 547 + ], + "score": 1.0, + "content": "Our first assumption restricts the model to be able to at least approximate the true Bayes classifier", + "type": "text" + }, + { + "bbox": [ + 483, + 534, + 503, + 546 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 532, + 507, + 547 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 313, + 557 + ], + "score": 1.0, + "content": "This condition assumes that given a hypothesis class", + "type": "text" + }, + { + "bbox": [ + 314, + 545, + 324, + 555 + ], + "score": 0.8, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "with sufficient complexity, the approximation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 204, + 568 + ], + "score": 1.0, + "content": "gap between a classifier", + "type": "text" + }, + { + "bbox": [ + 204, + 556, + 225, + 568 + ], + "score": 0.92, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 556, + 292, + 568 + ], + "score": 1.0, + "content": "in this class and", + "type": "text" + }, + { + "bbox": [ + 293, + 555, + 313, + 568 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "is determined by the inconsistency between the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 567, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 287, + 578 + ], + "score": 1.0, + "content": "noisy labels and the Bayes optimal classifier.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 532, + 507, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 580, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 200, + 593 + ], + "score": 1.0, + "content": "Definition 2 (Level set", + "type": "text" + }, + { + "bbox": [ + 201, + 580, + 224, + 592 + ], + "score": 0.89, + "content": "( \\alpha , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 579, + 401, + 593 + ], + "score": 1.0, + "content": "-consistency). Suppose data are sampled as", + "type": "text" + }, + { + "bbox": [ + 402, + 580, + 486, + 592 + ], + "score": 0.92, + "content": "( { \\pmb x } , \\tilde { y } ) \\sim D ( { \\pmb x } , \\tilde { \\eta } ( { \\pmb x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 579, + 506, + 593 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 589, + 494, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 302, + 608 + ], + "score": 0.87, + "content": "f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 589, + 332, + 609 + ], + "score": 1.0, + "content": ". Given", + "type": "text" + }, + { + "bbox": [ + 333, + 591, + 358, + 604 + ], + "score": 0.92, + "content": "\\varepsilon < \\textstyle { \\frac { 1 } { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 589, + 393, + 609 + ], + "score": 1.0, + "content": ", we call", + "type": "text" + }, + { + "bbox": [ + 393, + 592, + 403, + 601 + ], + "score": 0.78, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 589, + 414, + 609 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 414, + 592, + 437, + 603 + ], + "score": 0.91, + "content": "( \\alpha , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 589, + 482, + 609 + ], + "score": 1.0, + "content": "-consistent", + "type": "text" + }, + { + "bbox": [ + 482, + 592, + 488, + 603 + ], + "score": 0.53, + "content": "i f", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 589, + 494, + 609 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 579, + 506, + 609 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 129, + 611, + 481, + 639 + ], + "lines": [ + { + "bbox": [ + 129, + 611, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 129, + 611, + 481, + 639 + ], + "score": 0.88, + "content": "| f ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\leq \\alpha \\mathbb { E } _ { ( \\pmb { z } , \\bar { \\pmb { y } } ) \\sim D ( \\pmb { z } , \\bar { \\eta } ( \\pmb { z } ) ) } [ 1 _ { \\{ \\bar { \\pmb { y } } _ { \\pmb { z } } \\neq \\eta ^ { * } ( \\pmb { z } ) \\} } ( \\pmb { z } ) | | \\eta ( \\pmb { z } ) - \\frac { 1 } { 2 } | \\geq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | ] + \\epsilon .", + "type": "interline_equation", + "image_path": "0221483bf87e70974dc0f3cf4272dfa250b87d8eb4c293a7b95cd75797de6873.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 129, + 611, + 481, + 639 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 213, + 659 + ], + "score": 1.0, + "content": "For two input instances", + "type": "text" + }, + { + "bbox": [ + 213, + 649, + 221, + 657 + ], + "score": 0.76, + "content": "_ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 645, + 244, + 659 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 649, + 252, + 657 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 645, + 299, + 659 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 300, + 646, + 362, + 659 + ], + "score": 0.93, + "content": "\\eta ( z ) ~ > ~ \\eta ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 645, + 506, + 659 + ], + "score": 1.0, + "content": "(and hence the clean Bayes op-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 173, + 670 + ], + "score": 1.0, + "content": "timal classifier", + "type": "text" + }, + { + "bbox": [ + 173, + 658, + 199, + 669 + ], + "score": 0.9, + "content": "\\eta ^ { \\ast } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 657, + 312, + 670 + ], + "score": 1.0, + "content": "has higher confidence at", + "type": "text" + }, + { + "bbox": [ + 313, + 660, + 321, + 668 + ], + "score": 0.79, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 657, + 392, + 670 + ], + "score": 1.0, + "content": "than it does at", + "type": "text" + }, + { + "bbox": [ + 392, + 659, + 400, + 668 + ], + "score": 0.7, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "), the indicator function", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 108, + 668, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 108, + 669, + 283, + 683 + ], + "score": 0.78, + "content": "\\begin{array} { r } { \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } \\left( z : \\left| \\eta ( z ) - \\frac { 1 } { 2 } \\right| \\ge \\left| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\right| \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 668, + 506, + 683 + ], + "score": 1.0, + "content": "equals to 1 if the label of the more confident point", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 682, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 107, + 684, + 114, + 691 + ], + "score": 0.7, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 682, + 195, + 694 + ], + "score": 1.0, + "content": "is inconsistent with", + "type": "text" + }, + { + "bbox": [ + 195, + 682, + 221, + 693 + ], + "score": 0.89, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 682, + 496, + 694 + ], + "score": 1.0, + "content": ". This condition says that the approximation error of the classifier at", + "type": "text" + }, + { + "bbox": [ + 496, + 685, + 504, + 691 + ], + "score": 0.64, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 692, + 494, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 245, + 705 + ], + "score": 1.0, + "content": "should be controlled by the risk of", + "type": "text" + }, + { + "bbox": [ + 246, + 693, + 267, + 705 + ], + "score": 0.92, + "content": "\\eta ^ { * } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 692, + 305, + 705 + ], + "score": 1.0, + "content": "at points", + "type": "text" + }, + { + "bbox": [ + 305, + 694, + 312, + 703 + ], + "score": 0.77, + "content": "_ z", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 692, + 340, + 705 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 341, + 693, + 362, + 705 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 692, + 482, + 705 + ], + "score": 1.0, + "content": "is more confident than it is at", + "type": "text" + }, + { + "bbox": [ + 482, + 695, + 489, + 703 + ], + "score": 0.76, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 692, + 494, + 705 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 645, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "We next define a regularity condition of data distribution which describes the continuity of the level", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 189, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 189, + 732 + ], + "score": 1.0, + "content": "set density function.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 708, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 132 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 374, + 97 + ], + "score": 1.0, + "content": "Definition 3 (Level set bounded distribution). Define the margin", + "type": "text" + }, + { + "bbox": [ + 374, + 82, + 452, + 96 + ], + "score": 0.92, + "content": "\\begin{array} { r } { t ( { \\pmb x } ) = | \\eta ( { \\pmb x } ) - \\frac { 1 } { 2 } | } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 80, + 471, + 97 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 471, + 82, + 491, + 95 + ], + "score": 0.9, + "content": "G ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 80, + 506, + 97 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 148, + 109 + ], + "score": 1.0, + "content": "the cdf of", + "type": "text" + }, + { + "bbox": [ + 149, + 96, + 154, + 105 + ], + "score": 0.47, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 95, + 159, + 109 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 160, + 94, + 291, + 108 + ], + "score": 0.89, + "content": "G ( t ) = \\mathbb { P } _ { \\mathbf { x } \\sim D } ( | \\eta ( \\mathbf { x } ) - \\frac { 1 } { 2 } | \\leq t )", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 95, + 315, + 109 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 315, + 95, + 370, + 107 + ], + "score": 0.91, + "content": "g ( t ) = G ^ { \\prime } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 95, + 479, + 109 + ], + "score": 1.0, + "content": "be the density function of", + "type": "text" + }, + { + "bbox": [ + 479, + 97, + 484, + 105 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 95, + 505, + 109 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 189, + 119 + ], + "score": 1.0, + "content": "say the distribution", + "type": "text" + }, + { + "bbox": [ + 189, + 107, + 199, + 117 + ], + "score": 0.75, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 106, + 211, + 119 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 212, + 107, + 241, + 118 + ], + "score": 0.92, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 106, + 320, + 119 + ], + "score": 1.0, + "content": "-bounded if for all", + "type": "text" + }, + { + "bbox": [ + 320, + 107, + 376, + 118 + ], + "score": 0.79, + "content": "0 \\leq t \\leq 1 / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 106, + 380, + 119 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 381, + 106, + 468, + 118 + ], + "score": 0.74, + "content": "0 < c _ { * } \\leq g ( t ) \\leq c ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 106, + 484, + 119 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 484, + 107, + 494, + 117 + ], + "score": 0.78, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 118, + 443, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 136, + 131 + ], + "score": 0.91, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 118, + 385, + 132 + ], + "score": 1.0, + "content": "-bounded, we define the worst-case density-imbalance ratio of", + "type": "text" + }, + { + "bbox": [ + 386, + 119, + 395, + 129 + ], + "score": 0.62, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 118, + 408, + 132 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 408, + 118, + 443, + 133 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\ell _ { D } . = \\frac { c ^ { * } } { c _ { * } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "The above condition enforces the continuity of level set density function. This is crucial in the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "score": 1.0, + "content": "analysis since such a continuity allows one to borrow information from its neighborhood region so", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 484, + 174 + ], + "score": 1.0, + "content": "that a clean neighbor can help correct the corrupted label. To simplify the notation, we will omit", + "type": "text" + }, + { + "bbox": [ + 484, + 163, + 494, + 172 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 163, + 505, + 174 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 174, + 368, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 233, + 185 + ], + "score": 1.0, + "content": "the subscript when we mention", + "type": "text" + }, + { + "bbox": [ + 233, + 174, + 239, + 183 + ], + "score": 0.45, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 174, + 368, + 185 + ], + "score": 1.0, + "content": ". From now on, we will assume:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 190, + 504, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 255, + 203 + ], + "score": 1.0, + "content": "Assumption 1. There exist constants", + "type": "text" + }, + { + "bbox": [ + 255, + 191, + 298, + 201 + ], + "score": 0.89, + "content": "\\alpha , \\epsilon , c _ { * } , c ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 190, + 417, + 203 + ], + "score": 1.0, + "content": "such that the hypothesis class", + "type": "text" + }, + { + "bbox": [ + 417, + 191, + 427, + 200 + ], + "score": 0.83, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 190, + 438, + 203 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 438, + 190, + 461, + 202 + ], + "score": 0.93, + "content": "( \\alpha , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "-consistent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 200, + 318, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 226, + 213 + ], + "score": 1.0, + "content": "and the unknown distribution", + "type": "text" + }, + { + "bbox": [ + 226, + 201, + 235, + 211 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 200, + 246, + 213 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 246, + 201, + 276, + 213 + ], + "score": 0.95, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 200, + 318, + 213 + ], + "score": 1.0, + "content": "-bounded.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 225, + 283, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 285, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 285, + 238 + ], + "score": 1.0, + "content": "3.2 MAIN RESULT AND PROOF SKETCH", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "score": 1.0, + "content": "In this section we first state our main result, and then present the supporting claims. Complete proofs", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "can be found in the appendix. Our main result below states that if our starting function is trained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 267, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 165, + 286 + ], + "score": 1.0, + "content": "correctly, i.e.,", + "type": "text" + }, + { + "bbox": [ + 165, + 268, + 362, + 285 + ], + "score": 0.91, + "content": "f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 267, + 506, + 286 + ], + "score": 1.0, + "content": ", then Algorithm 1 terminates with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 283, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 297 + ], + "score": 1.0, + "content": "most of the final labels matching the Bayes optimal classifier labels. In practice, minimizing true", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "risk is not achievable. Instead, the empirical risk is used to estimate true risk, approaching true", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 304, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 281, + 320 + ], + "score": 1.0, + "content": "risk asymptotically. For a scoring function", + "type": "text" + }, + { + "bbox": [ + 282, + 306, + 289, + 317 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 304, + 367, + 320 + ], + "score": 1.0, + "content": ", we will denote by", + "type": "text" + }, + { + "bbox": [ + 367, + 306, + 466, + 319 + ], + "score": 0.93, + "content": "y _ { f ( \\pmb { x } ) } : = \\mathbb { I } ( f ( \\pmb { x } ) \\breve { \\geq } 1 / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 304, + 506, + 320 + ], + "score": 1.0, + "content": "the label", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 170, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 158, + 329 + ], + "score": 1.0, + "content": "predicted by", + "type": "text" + }, + { + "bbox": [ + 159, + 317, + 165, + 329 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 317, + 170, + 329 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 505, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 328, + 510, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 242, + 344 + ], + "score": 1.0, + "content": "Theorem 1. Under Assumption", + "type": "text" + }, + { + "bbox": [ + 106, + 342, + 168, + 356 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 336, + 308, + 360 + ], + "score": 1.0, + "content": ". Then for the output of Algorithm", + "type": "text" + }, + { + "bbox": [ + 242, + 332, + 248, + 341 + ], + "score": 0.55, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 328, + 312, + 344 + ], + "score": 1.0, + "content": ", for any noise", + "type": "text" + }, + { + "bbox": [ + 309, + 343, + 315, + 352 + ], + "score": 0.49, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 333, + 319, + 341 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 336, + 336, + 360 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 320, + 328, + 438, + 344 + ], + "score": 1.0, + "content": "which is PMD with margin", + "type": "text" + }, + { + "bbox": [ + 336, + 342, + 343, + 353 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 336, + 510, + 360 + ], + "score": 1.0, + "content": "as above and with the following initial-", + "type": "text" + }, + { + "bbox": [ + 438, + 332, + 447, + 342 + ], + "score": 0.81, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 328, + 480, + 344 + ], + "score": 1.0, + "content": ", define", + "type": "text" + }, + { + "bbox": [ + 481, + 332, + 505, + 342 + ], + "score": 0.85, + "content": "e _ { 0 } =", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 102, + 350, + 509, + 375 + ], + "spans": [ + { + "bbox": [ + 102, + 350, + 144, + 375 + ], + "score": 1.0, + "content": "izations:", + "type": "text" + }, + { + "bbox": [ + 144, + 355, + 210, + 369 + ], + "score": 0.81, + "content": "( l ) T _ { 0 } < { \\frac { 1 } { 2 } } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 350, + 228, + 375 + ], + "score": 1.0, + "content": ", (2)", + "type": "text" + }, + { + "bbox": [ + 228, + 354, + 309, + 370 + ], + "score": 0.83, + "content": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 350, + 327, + 375 + ], + "score": 1.0, + "content": ", (3)", + "type": "text" + }, + { + "bbox": [ + 327, + 354, + 412, + 370 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { T _ { 0 } } { 3 \\varepsilon } ) , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 350, + 430, + 375 + ], + "score": 1.0, + "content": ", (4)", + "type": "text" + }, + { + "bbox": [ + 430, + 356, + 473, + 368 + ], + "score": 0.88, + "content": "T _ { e n d } \\leq 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 350, + 509, + 375 + ], + "score": 1.0, + "content": "and (5)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 367, + 206, + 386 + ], + "spans": [ + { + "bbox": [ + 107, + 369, + 162, + 383 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 367, + 206, + 386 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 380, + 385, + 395 + ], + "lines": [ + { + "bbox": [ + 225, + 380, + 385, + 395 + ], + "spans": [ + { + "bbox": [ + 225, + 380, + 385, + 395 + ], + "score": 0.9, + "content": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon .", + "type": "interline_equation", + "image_path": "bb59e04dd3b3b7fe323b6aa7c8572b23ae3d6a71998e9e2d6e0ac2bf7b7e18d3.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 225, + 380, + 385, + 395 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 360, + 414 + ], + "score": 1.0, + "content": "In the remainder of this section we shall assume that the noise", + "type": "text" + }, + { + "bbox": [ + 360, + 403, + 367, + 411 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 398, + 453, + 414 + ], + "score": 1.0, + "content": "is PMD with margin", + "type": "text" + }, + { + "bbox": [ + 453, + 402, + 462, + 412 + ], + "score": 0.86, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 398, + 506, + 414 + ], + "score": 1.0, + "content": ". To prove", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 279, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 279, + 424 + ], + "score": 1.0, + "content": "our result we first define a “pure” level set.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 503, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 182, + 441 + ], + "score": 1.0, + "content": "Definition 4 (Pure", + "type": "text" + }, + { + "bbox": [ + 183, + 425, + 215, + 438 + ], + "score": 0.92, + "content": "( e , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 422, + 259, + 441 + ], + "score": 1.0, + "content": "-level set).", + "type": "text" + }, + { + "bbox": [ + 260, + 426, + 268, + 435 + ], + "score": 0.28, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 422, + 282, + 441 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 283, + 424, + 410, + 439 + ], + "score": 0.92, + "content": "L ( e , \\eta ) : = \\{ \\pmb { x } | | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 422, + 454, + 441 + ], + "score": 1.0, + "content": "is pure for", + "type": "text" + }, + { + "bbox": [ + 454, + 426, + 462, + 437 + ], + "score": 0.53, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 422, + 470, + 441 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 471, + 426, + 505, + 439 + ], + "score": 0.78, + "content": "y _ { f ( \\pmb { x } ) } =", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 437, + 213, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 132, + 450 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 437, + 160, + 451 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 160, + 438, + 209, + 450 + ], + "score": 0.93, + "content": "\\pmb { x } \\in L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 437, + 213, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 505, + 470 + ], + "score": 1.0, + "content": "We now state a lemma that forms the foundation of our progressive correction algorithm. We show", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "that given a tiny region where the model is reliable, we can move one step forward by trusting the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "model. Although the improvement is slight in a single round, it empowers a conservatively recursive", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 490, + 204, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 204, + 503 + ], + "score": 1.0, + "content": "step in the Algorithm 1.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 504, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 365, + 517 + ], + "score": 1.0, + "content": "Lemma 1 (One round purity improvement). Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 365, + 505, + 371, + 514 + ], + "score": 0.29, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 504, + 477, + 517 + ], + "score": 1.0, + "content": "is satisfied, and assume an", + "type": "text" + }, + { + "bbox": [ + 477, + 505, + 484, + 516 + ], + "score": 0.79, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 504, + 505, + 517 + ], + "score": 1.0, + "content": "such", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 198, + 529 + ], + "score": 1.0, + "content": "that there exists a pure", + "type": "text" + }, + { + "bbox": [ + 198, + 515, + 230, + 527 + ], + "score": 0.91, + "content": "( e , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 514, + 285, + 529 + ], + "score": 1.0, + "content": "-level set with", + "type": "text" + }, + { + "bbox": [ + 286, + 515, + 333, + 528 + ], + "score": 0.9, + "content": "\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 514, + 353, + 529 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 353, + 515, + 504, + 528 + ], + "score": 0.87, + "content": "\\widetilde { \\eta } _ { n e w } ( \\pmb { x } ) = y _ { f ( \\pmb { x } ) } i f \\vert f ( \\pmb { x } ) - 1 / 2 \\vert \\geq e", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 526, + 507, + 546 + ], + "spans": [ + { + "bbox": [ + 104, + 526, + 124, + 546 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 125, + 527, + 228, + 540 + ], + "score": 0.85, + "content": "\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 526, + 283, + 546 + ], + "score": 1.0, + "content": ", and assume", + "type": "text" + }, + { + "bbox": [ + 284, + 528, + 483, + 543 + ], + "score": 0.85, + "content": "f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 526, + 507, + 546 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 111, + 544, + 449, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 450, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 193, + 557 + ], + "score": 0.76, + "content": "e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 542, + 198, + 560 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 545, + 228, + 557 + ], + "score": 0.68, + "content": "L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 542, + 273, + 560 + ], + "score": 1.0, + "content": "is pure for", + "type": "text" + }, + { + "bbox": [ + 273, + 545, + 299, + 558 + ], + "score": 0.88, + "content": "f _ { n e w } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 542, + 327, + 560 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 327, + 545, + 447, + 558 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 542, + 450, + 560 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "The above lemma states that the cleansed region will be enlarged by at least a constant factor. In the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 331, + 588 + ], + "score": 1.0, + "content": "following lemma, we justify the functionality of the first", + "type": "text" + }, + { + "bbox": [ + 331, + 578, + 342, + 586 + ], + "score": 0.58, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "warm-up rounds. Since the initial neural", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "network can behave badly, the region where we can trust the classifier can be very limited. Before", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 596, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 505, + 612 + ], + "score": 1.0, + "content": "starting the flipping procedure in a relatively larger level set, one first needs to expand the initial tiny", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 608, + 233, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 135, + 623 + ], + "score": 1.0, + "content": "region", + "type": "text" + }, + { + "bbox": [ + 135, + 608, + 163, + 622 + ], + "score": 0.93, + "content": "\\mathbf { \\frac { 1 } { 2 } } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 609, + 217, + 623 + ], + "score": 1.0, + "content": "to a constant", + "type": "text" + }, + { + "bbox": [ + 217, + 609, + 228, + 620 + ], + "score": 0.88, + "content": "T _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 609, + 233, + 623 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 101, + 622, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 101, + 622, + 152, + 655 + ], + "score": 1.0, + "content": "Lemma 2 is pure for", + "type": "text" + }, + { + "bbox": [ + 163, + 622, + 195, + 655 + ], + "score": 1.0, + "content": "arm-up r. Given", + "type": "text" + }, + { + "bbox": [ + 237, + 622, + 344, + 655 + ], + "score": 1.0, + "content": "uppose for a given function , after running Algorithm", + "type": "text" + }, + { + "bbox": [ + 344, + 623, + 354, + 635 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 622, + 365, + 655 + ], + "score": 1.0, + "content": "thor", + "type": "text" + }, + { + "bbox": [ + 444, + 623, + 478, + 635 + ], + "score": 0.93, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 622, + 505, + 655 + ], + "score": 1.0, + "content": "which, there", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 152, + 634, + 446, + 649 + ], + "spans": [ + { + "bbox": [ + 152, + 635, + 162, + 647 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 635, + 236, + 647 + ], + "score": 0.92, + "content": "T _ { 0 } < 1 / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 634, + 446, + 649 + ], + "score": 0.93, + "content": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 304, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 173, + 663 + ], + "score": 1.0, + "content": "exists a level set", + "type": "text" + }, + { + "bbox": [ + 173, + 648, + 227, + 662 + ], + "score": 0.93, + "content": "L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 648, + 290, + 663 + ], + "score": 1.0, + "content": "that is pure for", + "type": "text" + }, + { + "bbox": [ + 290, + 649, + 300, + 661 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 648, + 304, + 663 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 404, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 405, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 405, + 682 + ], + "score": 1.0, + "content": "Next we present our final lemma that combines the previous two lemmata.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 683, + 505, + 735 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 507, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 242, + 696 + ], + "score": 1.0, + "content": "Lemma 3. Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 242, + 684, + 248, + 693 + ], + "score": 0.45, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 681, + 400, + 696 + ], + "score": 1.0, + "content": "is satisfied, and for a given function", + "type": "text" + }, + { + "bbox": [ + 401, + 684, + 411, + 695 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 681, + 507, + 696 + ], + "score": 1.0, + "content": "there exists a level set", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 694, + 505, + 726 + ], + "spans": [ + { + "bbox": [ + 107, + 694, + 141, + 706 + ], + "score": 0.91, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 694, + 182, + 726 + ], + "score": 1.0, + "content": "h is , (2)", + "type": "text" + }, + { + "bbox": [ + 215, + 695, + 225, + 705 + ], + "score": 0.84, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 694, + 286, + 726 + ], + "score": 1.0, + "content": "ns A, (3)", + "type": "text" + }, + { + "bbox": [ + 322, + 695, + 327, + 704 + ], + "score": 0.34, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 695, + 394, + 705 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 694, + 405, + 726 + ], + "score": 1.0, + "content": "an(4)", + "type": "text" + }, + { + "bbox": [ + 471, + 694, + 491, + 726 + ], + "score": 1.0, + "content": "ons:and", + "type": "text" + }, + { + "bbox": [ + 491, + 694, + 505, + 705 + ], + "score": 0.44, + "content": "( l )", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 705, + 471, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 706, + 163, + 719 + ], + "score": 0.91, + "content": "T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 705, + 267, + 720 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 705, + 385, + 720 + ], + "score": 0.88, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 705, + 471, + 720 + ], + "score": 0.92, + "content": "T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 714, + 384, + 740 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 162, + 734 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 714, + 220, + 740 + ], + "score": 1.0, + "content": ", then we have", + "type": "text" + }, + { + "bbox": [ + 221, + 721, + 376, + 734 + ], + "score": 0.93, + "content": "\\mathbb { P } _ { \\mathbf { x } \\sim D } [ y _ { f _ { f i n a l } ( \\mathbf { x } ) } = \\eta ^ { * } ( \\pmb { x } ) ] \\geq 1 - 3 c ^ { * } \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 714, + 384, + 740 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 132 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 374, + 97 + ], + "score": 1.0, + "content": "Definition 3 (Level set bounded distribution). Define the margin", + "type": "text" + }, + { + "bbox": [ + 374, + 82, + 452, + 96 + ], + "score": 0.92, + "content": "\\begin{array} { r } { t ( { \\pmb x } ) = | \\eta ( { \\pmb x } ) - \\frac { 1 } { 2 } | } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 80, + 471, + 97 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 471, + 82, + 491, + 95 + ], + "score": 0.9, + "content": "G ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 80, + 506, + 97 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 148, + 109 + ], + "score": 1.0, + "content": "the cdf of", + "type": "text" + }, + { + "bbox": [ + 149, + 96, + 154, + 105 + ], + "score": 0.47, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 95, + 159, + 109 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 160, + 94, + 291, + 108 + ], + "score": 0.89, + "content": "G ( t ) = \\mathbb { P } _ { \\mathbf { x } \\sim D } ( | \\eta ( \\mathbf { x } ) - \\frac { 1 } { 2 } | \\leq t )", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 95, + 315, + 109 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 315, + 95, + 370, + 107 + ], + "score": 0.91, + "content": "g ( t ) = G ^ { \\prime } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 95, + 479, + 109 + ], + "score": 1.0, + "content": "be the density function of", + "type": "text" + }, + { + "bbox": [ + 479, + 97, + 484, + 105 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 95, + 505, + 109 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 189, + 119 + ], + "score": 1.0, + "content": "say the distribution", + "type": "text" + }, + { + "bbox": [ + 189, + 107, + 199, + 117 + ], + "score": 0.75, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 106, + 211, + 119 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 212, + 107, + 241, + 118 + ], + "score": 0.92, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 106, + 320, + 119 + ], + "score": 1.0, + "content": "-bounded if for all", + "type": "text" + }, + { + "bbox": [ + 320, + 107, + 376, + 118 + ], + "score": 0.79, + "content": "0 \\leq t \\leq 1 / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 106, + 380, + 119 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 381, + 106, + 468, + 118 + ], + "score": 0.74, + "content": "0 < c _ { * } \\leq g ( t ) \\leq c ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 106, + 484, + 119 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 484, + 107, + 494, + 117 + ], + "score": 0.78, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 118, + 443, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 136, + 131 + ], + "score": 0.91, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 118, + 385, + 132 + ], + "score": 1.0, + "content": "-bounded, we define the worst-case density-imbalance ratio of", + "type": "text" + }, + { + "bbox": [ + 386, + 119, + 395, + 129 + ], + "score": 0.62, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 118, + 408, + 132 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 408, + 118, + 443, + 133 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\ell _ { D } . = \\frac { c ^ { * } } { c _ { * } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 104, + 80, + 506, + 133 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "The above condition enforces the continuity of level set density function. This is crucial in the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "score": 1.0, + "content": "analysis since such a continuity allows one to borrow information from its neighborhood region so", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 163, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 484, + 174 + ], + "score": 1.0, + "content": "that a clean neighbor can help correct the corrupted label. To simplify the notation, we will omit", + "type": "text" + }, + { + "bbox": [ + 484, + 163, + 494, + 172 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 163, + 505, + 174 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 174, + 368, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 233, + 185 + ], + "score": 1.0, + "content": "the subscript when we mention", + "type": "text" + }, + { + "bbox": [ + 233, + 174, + 239, + 183 + ], + "score": 0.45, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 174, + 368, + 185 + ], + "score": 1.0, + "content": ". From now on, we will assume:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 140, + 505, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 190, + 504, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 255, + 203 + ], + "score": 1.0, + "content": "Assumption 1. There exist constants", + "type": "text" + }, + { + "bbox": [ + 255, + 191, + 298, + 201 + ], + "score": 0.89, + "content": "\\alpha , \\epsilon , c _ { * } , c ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 190, + 417, + 203 + ], + "score": 1.0, + "content": "such that the hypothesis class", + "type": "text" + }, + { + "bbox": [ + 417, + 191, + 427, + 200 + ], + "score": 0.83, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 190, + 438, + 203 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 438, + 190, + 461, + 202 + ], + "score": 0.93, + "content": "( \\alpha , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "-consistent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 200, + 318, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 226, + 213 + ], + "score": 1.0, + "content": "and the unknown distribution", + "type": "text" + }, + { + "bbox": [ + 226, + 201, + 235, + 211 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 200, + 246, + 213 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 246, + 201, + 276, + 213 + ], + "score": 0.95, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 200, + 318, + 213 + ], + "score": 1.0, + "content": "-bounded.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 190, + 505, + 213 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 225, + 283, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 285, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 285, + 238 + ], + "score": 1.0, + "content": "3.2 MAIN RESULT AND PROOF SKETCH", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 259 + ], + "score": 1.0, + "content": "In this section we first state our main result, and then present the supporting claims. Complete proofs", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "can be found in the appendix. Our main result below states that if our starting function is trained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 267, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 165, + 286 + ], + "score": 1.0, + "content": "correctly, i.e.,", + "type": "text" + }, + { + "bbox": [ + 165, + 268, + 362, + 285 + ], + "score": 0.91, + "content": "f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 267, + 506, + 286 + ], + "score": 1.0, + "content": ", then Algorithm 1 terminates with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 283, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 297 + ], + "score": 1.0, + "content": "most of the final labels matching the Bayes optimal classifier labels. In practice, minimizing true", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "risk is not achievable. Instead, the empirical risk is used to estimate true risk, approaching true", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 304, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 281, + 320 + ], + "score": 1.0, + "content": "risk asymptotically. For a scoring function", + "type": "text" + }, + { + "bbox": [ + 282, + 306, + 289, + 317 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 304, + 367, + 320 + ], + "score": 1.0, + "content": ", we will denote by", + "type": "text" + }, + { + "bbox": [ + 367, + 306, + 466, + 319 + ], + "score": 0.93, + "content": "y _ { f ( \\pmb { x } ) } : = \\mathbb { I } ( f ( \\pmb { x } ) \\breve { \\geq } 1 / 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 304, + 506, + 320 + ], + "score": 1.0, + "content": "the label", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 170, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 158, + 329 + ], + "score": 1.0, + "content": "predicted by", + "type": "text" + }, + { + "bbox": [ + 159, + 317, + 165, + 329 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 317, + 170, + 329 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 245, + 506, + 329 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 505, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 328, + 510, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 242, + 344 + ], + "score": 1.0, + "content": "Theorem 1. Under Assumption", + "type": "text" + }, + { + "bbox": [ + 106, + 342, + 168, + 356 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 336, + 308, + 360 + ], + "score": 1.0, + "content": ". Then for the output of Algorithm", + "type": "text" + }, + { + "bbox": [ + 242, + 332, + 248, + 341 + ], + "score": 0.55, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 328, + 312, + 344 + ], + "score": 1.0, + "content": ", for any noise", + "type": "text" + }, + { + "bbox": [ + 309, + 343, + 315, + 352 + ], + "score": 0.49, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 333, + 319, + 341 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 336, + 336, + 360 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 320, + 328, + 438, + 344 + ], + "score": 1.0, + "content": "which is PMD with margin", + "type": "text" + }, + { + "bbox": [ + 336, + 342, + 343, + 353 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 336, + 510, + 360 + ], + "score": 1.0, + "content": "as above and with the following initial-", + "type": "text" + }, + { + "bbox": [ + 438, + 332, + 447, + 342 + ], + "score": 0.81, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 328, + 480, + 344 + ], + "score": 1.0, + "content": ", define", + "type": "text" + }, + { + "bbox": [ + 481, + 332, + 505, + 342 + ], + "score": 0.85, + "content": "e _ { 0 } =", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 102, + 350, + 509, + 375 + ], + "spans": [ + { + "bbox": [ + 102, + 350, + 144, + 375 + ], + "score": 1.0, + "content": "izations:", + "type": "text" + }, + { + "bbox": [ + 144, + 355, + 210, + 369 + ], + "score": 0.81, + "content": "( l ) T _ { 0 } < { \\frac { 1 } { 2 } } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 350, + 228, + 375 + ], + "score": 1.0, + "content": ", (2)", + "type": "text" + }, + { + "bbox": [ + 228, + 354, + 309, + 370 + ], + "score": 0.83, + "content": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 350, + 327, + 375 + ], + "score": 1.0, + "content": ", (3)", + "type": "text" + }, + { + "bbox": [ + 327, + 354, + 412, + 370 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { T _ { 0 } } { 3 \\varepsilon } ) , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 350, + 430, + 375 + ], + "score": 1.0, + "content": ", (4)", + "type": "text" + }, + { + "bbox": [ + 430, + 356, + 473, + 368 + ], + "score": 0.88, + "content": "T _ { e n d } \\leq 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 350, + 509, + 375 + ], + "score": 1.0, + "content": "and (5)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 367, + 206, + 386 + ], + "spans": [ + { + "bbox": [ + 107, + 369, + 162, + 383 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 367, + 206, + 386 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 102, + 328, + 510, + 386 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 380, + 385, + 395 + ], + "lines": [ + { + "bbox": [ + 225, + 380, + 385, + 395 + ], + "spans": [ + { + "bbox": [ + 225, + 380, + 385, + 395 + ], + "score": 0.9, + "content": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon .", + "type": "interline_equation", + "image_path": "bb59e04dd3b3b7fe323b6aa7c8572b23ae3d6a71998e9e2d6e0ac2bf7b7e18d3.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 225, + 380, + 385, + 395 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 504, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 360, + 414 + ], + "score": 1.0, + "content": "In the remainder of this section we shall assume that the noise", + "type": "text" + }, + { + "bbox": [ + 360, + 403, + 367, + 411 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 398, + 453, + 414 + ], + "score": 1.0, + "content": "is PMD with margin", + "type": "text" + }, + { + "bbox": [ + 453, + 402, + 462, + 412 + ], + "score": 0.86, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 398, + 506, + 414 + ], + "score": 1.0, + "content": ". To prove", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 279, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 279, + 424 + ], + "score": 1.0, + "content": "our result we first define a “pure” level set.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 398, + 506, + 424 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 503, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 182, + 441 + ], + "score": 1.0, + "content": "Definition 4 (Pure", + "type": "text" + }, + { + "bbox": [ + 183, + 425, + 215, + 438 + ], + "score": 0.92, + "content": "( e , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 422, + 259, + 441 + ], + "score": 1.0, + "content": "-level set).", + "type": "text" + }, + { + "bbox": [ + 260, + 426, + 268, + 435 + ], + "score": 0.28, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 422, + 282, + 441 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 283, + 424, + 410, + 439 + ], + "score": 0.92, + "content": "L ( e , \\eta ) : = \\{ \\pmb { x } | | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 422, + 454, + 441 + ], + "score": 1.0, + "content": "is pure for", + "type": "text" + }, + { + "bbox": [ + 454, + 426, + 462, + 437 + ], + "score": 0.53, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 422, + 470, + 441 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 471, + 426, + 505, + 439 + ], + "score": 0.78, + "content": "y _ { f ( \\pmb { x } ) } =", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 437, + 213, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 132, + 450 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 437, + 160, + 451 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 160, + 438, + 209, + 450 + ], + "score": 0.93, + "content": "\\pmb { x } \\in L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 437, + 213, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 422, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 458, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 505, + 470 + ], + "score": 1.0, + "content": "We now state a lemma that forms the foundation of our progressive correction algorithm. We show", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "that given a tiny region where the model is reliable, we can move one step forward by trusting the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "model. Although the improvement is slight in a single round, it empowers a conservatively recursive", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 490, + 204, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 204, + 503 + ], + "score": 1.0, + "content": "step in the Algorithm 1.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 458, + 506, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 504, + 505, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 365, + 517 + ], + "score": 1.0, + "content": "Lemma 1 (One round purity improvement). Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 365, + 505, + 371, + 514 + ], + "score": 0.29, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 504, + 477, + 517 + ], + "score": 1.0, + "content": "is satisfied, and assume an", + "type": "text" + }, + { + "bbox": [ + 477, + 505, + 484, + 516 + ], + "score": 0.79, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 504, + 505, + 517 + ], + "score": 1.0, + "content": "such", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 198, + 529 + ], + "score": 1.0, + "content": "that there exists a pure", + "type": "text" + }, + { + "bbox": [ + 198, + 515, + 230, + 527 + ], + "score": 0.91, + "content": "( e , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 514, + 285, + 529 + ], + "score": 1.0, + "content": "-level set with", + "type": "text" + }, + { + "bbox": [ + 286, + 515, + 333, + 528 + ], + "score": 0.9, + "content": "\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 514, + 353, + 529 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 353, + 515, + 504, + 528 + ], + "score": 0.87, + "content": "\\widetilde { \\eta } _ { n e w } ( \\pmb { x } ) = y _ { f ( \\pmb { x } ) } i f \\vert f ( \\pmb { x } ) - 1 / 2 \\vert \\geq e", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 526, + 507, + 546 + ], + "spans": [ + { + "bbox": [ + 104, + 526, + 124, + 546 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 125, + 527, + 228, + 540 + ], + "score": 0.85, + "content": "\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 526, + 283, + 546 + ], + "score": 1.0, + "content": ", and assume", + "type": "text" + }, + { + "bbox": [ + 284, + 528, + 483, + 543 + ], + "score": 0.85, + "content": "f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 526, + 507, + 546 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 542, + 450, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 193, + 557 + ], + "score": 0.76, + "content": "e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 542, + 198, + 560 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 545, + 228, + 557 + ], + "score": 0.68, + "content": "L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 542, + 273, + 560 + ], + "score": 1.0, + "content": "is pure for", + "type": "text" + }, + { + "bbox": [ + 273, + 545, + 299, + 558 + ], + "score": 0.88, + "content": "f _ { n e w } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 542, + 327, + 560 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 327, + 545, + 447, + 558 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 542, + 450, + 560 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 504, + 507, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 111, + 544, + 449, + 559 + ], + "lines": [], + "index": 33, + "bbox_fs": [ + 106, + 542, + 450, + 560 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "The above lemma states that the cleansed region will be enlarged by at least a constant factor. In the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 331, + 588 + ], + "score": 1.0, + "content": "following lemma, we justify the functionality of the first", + "type": "text" + }, + { + "bbox": [ + 331, + 578, + 342, + 586 + ], + "score": 0.58, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "warm-up rounds. Since the initial neural", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "network can behave badly, the region where we can trust the classifier can be very limited. Before", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 596, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 505, + 612 + ], + "score": 1.0, + "content": "starting the flipping procedure in a relatively larger level set, one first needs to expand the initial tiny", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 608, + 233, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 135, + 623 + ], + "score": 1.0, + "content": "region", + "type": "text" + }, + { + "bbox": [ + 135, + 608, + 163, + 622 + ], + "score": 0.93, + "content": "\\mathbf { \\frac { 1 } { 2 } } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 609, + 217, + 623 + ], + "score": 1.0, + "content": "to a constant", + "type": "text" + }, + { + "bbox": [ + 217, + 609, + 228, + 620 + ], + "score": 0.88, + "content": "T _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 609, + 233, + 623 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 564, + 505, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 101, + 622, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 101, + 622, + 152, + 655 + ], + "score": 1.0, + "content": "Lemma 2 is pure for", + "type": "text" + }, + { + "bbox": [ + 163, + 622, + 195, + 655 + ], + "score": 1.0, + "content": "arm-up r. Given", + "type": "text" + }, + { + "bbox": [ + 237, + 622, + 344, + 655 + ], + "score": 1.0, + "content": "uppose for a given function , after running Algorithm", + "type": "text" + }, + { + "bbox": [ + 344, + 623, + 354, + 635 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 622, + 365, + 655 + ], + "score": 1.0, + "content": "thor", + "type": "text" + }, + { + "bbox": [ + 444, + 623, + 478, + 635 + ], + "score": 0.93, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 622, + 505, + 655 + ], + "score": 1.0, + "content": "which, there", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 152, + 634, + 446, + 649 + ], + "spans": [ + { + "bbox": [ + 152, + 635, + 162, + 647 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 635, + 236, + 647 + ], + "score": 0.92, + "content": "T _ { 0 } < 1 / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 634, + 446, + 649 + ], + "score": 0.93, + "content": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 304, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 173, + 663 + ], + "score": 1.0, + "content": "exists a level set", + "type": "text" + }, + { + "bbox": [ + 173, + 648, + 227, + 662 + ], + "score": 0.93, + "content": "L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 648, + 290, + 663 + ], + "score": 1.0, + "content": "that is pure for", + "type": "text" + }, + { + "bbox": [ + 290, + 649, + 300, + 661 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 648, + 304, + 663 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 101, + 622, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 404, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 405, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 405, + 682 + ], + "score": 1.0, + "content": "Next we present our final lemma that combines the previous two lemmata.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 668, + 405, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 683, + 505, + 735 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 507, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 242, + 696 + ], + "score": 1.0, + "content": "Lemma 3. Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 242, + 684, + 248, + 693 + ], + "score": 0.45, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 681, + 400, + 696 + ], + "score": 1.0, + "content": "is satisfied, and for a given function", + "type": "text" + }, + { + "bbox": [ + 401, + 684, + 411, + 695 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 681, + 507, + 696 + ], + "score": 1.0, + "content": "there exists a level set", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 694, + 505, + 726 + ], + "spans": [ + { + "bbox": [ + 107, + 694, + 141, + 706 + ], + "score": 0.91, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 694, + 182, + 726 + ], + "score": 1.0, + "content": "h is , (2)", + "type": "text" + }, + { + "bbox": [ + 215, + 695, + 225, + 705 + ], + "score": 0.84, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 694, + 286, + 726 + ], + "score": 1.0, + "content": "ns A, (3)", + "type": "text" + }, + { + "bbox": [ + 322, + 695, + 327, + 704 + ], + "score": 0.34, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 695, + 394, + 705 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 694, + 405, + 726 + ], + "score": 1.0, + "content": "an(4)", + "type": "text" + }, + { + "bbox": [ + 471, + 694, + 491, + 726 + ], + "score": 1.0, + "content": "ons:and", + "type": "text" + }, + { + "bbox": [ + 491, + 694, + 505, + 705 + ], + "score": 0.44, + "content": "( l )", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 705, + 471, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 706, + 163, + 719 + ], + "score": 0.91, + "content": "T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 705, + 267, + 720 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 705, + 385, + 720 + ], + "score": 0.88, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 705, + 471, + 720 + ], + "score": 0.92, + "content": "T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 714, + 384, + 740 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 162, + 734 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 714, + 220, + 740 + ], + "score": 1.0, + "content": ", then we have", + "type": "text" + }, + { + "bbox": [ + 221, + 721, + 376, + 734 + ], + "score": 0.93, + "content": "\\mathbb { P } _ { \\mathbf { x } \\sim D } [ y _ { f _ { f i n a l } ( \\mathbf { x } ) } = \\eta ^ { * } ( \\pmb { x } ) ] \\geq 1 - 3 c ^ { * } \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 714, + 384, + 740 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 681, + 507, + 740 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "This lemma states that given an initial model that has a reasonably pure super level set, one can", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "manage to progressively correct a large fraction of corrupted labels by running Algorithm 1 for a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "sufficient long time with carefully chosen parameters. The limit of Algorithm 1 will depend on the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 435, + 128 + ], + "score": 1.0, + "content": "approximation ability of the neural network, which is characterized by parameter", + "type": "text" + }, + { + "bbox": [ + 436, + 117, + 442, + 126 + ], + "score": 0.72, + "content": "\\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 115, + 507, + 128 + ], + "score": 1.0, + "content": "in Definition 2.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "score": 1.0, + "content": "To prove Theorem 1 using Lemma 3, it suffices to get a model which has a reliable region. This is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 477, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 477, + 150 + ], + "score": 1.0, + "content": "provably achievable by training with a family of good scoring functions on PMD noisy data.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 164, + 200, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 201, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 201, + 179 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 189, + 506, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 189, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 506, + 201 + ], + "score": 1.0, + "content": "We evaluate our method on both synthetic and real-world datasets. We first conduct synthetic experi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "ments on two public datasets CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009). To synthesize the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 212, + 506, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 353, + 224 + ], + "score": 1.0, + "content": "label noise, we first approximate the true posterior probability", + "type": "text" + }, + { + "bbox": [ + 353, + 213, + 360, + 223 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 212, + 506, + 224 + ], + "score": 1.0, + "content": "using the confidence prediction of a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 506, + 235 + ], + "score": 1.0, + "content": "clean neural network (trained with the original clean labels). We call these original labels raw labels.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 174, + 246 + ], + "score": 1.0, + "content": "Then we sample", + "type": "text" + }, + { + "bbox": [ + 174, + 234, + 219, + 245 + ], + "score": 0.93, + "content": "y _ { \\pmb { x } } \\sim \\eta ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 233, + 291, + 246 + ], + "score": 1.0, + "content": "for each instance", + "type": "text" + }, + { + "bbox": [ + 291, + 235, + 299, + 244 + ], + "score": 0.65, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 233, + 506, + 246 + ], + "score": 1.0, + "content": ". Instead of using raw labels, we use these sampled", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 132, + 257 + ], + "score": 1.0, + "content": "labels", + "type": "text" + }, + { + "bbox": [ + 132, + 246, + 144, + 256 + ], + "score": 0.86, + "content": "y _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 243, + 383, + 257 + ], + "score": 1.0, + "content": "as the clean labels, whose posterior probabilities are exactly", + "type": "text" + }, + { + "bbox": [ + 384, + 244, + 405, + 256 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "; and therefore the neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 254, + 507, + 268 + ], + "spans": [ + { + "bbox": [ + 104, + 254, + 268, + 268 + ], + "score": 1.0, + "content": "network is the Bayes optimal classifier", + "type": "text" + }, + { + "bbox": [ + 268, + 255, + 361, + 267 + ], + "score": 0.93, + "content": "\\eta ^ { * } : \\mathcal { X } \\{ 1 , \\cdot \\cdot , C \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 254, + 392, + 268 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 393, + 256, + 402, + 265 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 254, + 507, + 268 + ], + "score": 1.0, + "content": "is the number of classes.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 266, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 233, + 279 + ], + "score": 1.0, + "content": "Note that in multi-class setting,", + "type": "text" + }, + { + "bbox": [ + 234, + 266, + 254, + 278 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 266, + 350, + 279 + ], + "score": 1.0, + "content": "has a vector output and", + "type": "text" + }, + { + "bbox": [ + 351, + 266, + 374, + 278 + ], + "score": 0.93, + "content": "\\dot { \\eta _ { i } } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 266, + 399, + 279 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 399, + 267, + 404, + 276 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 266, + 507, + 279 + ], + "score": 1.0, + "content": "-th element of this vector.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 503, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Noise generation. We consider a generic family of noise. We consider not only feature-dependent", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 461, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 461, + 307 + ], + "score": 1.0, + "content": "noise, but also hybrid noise that consists of both feature-dependent noise and i.i.d. noise.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "score": 1.0, + "content": "For feature-dependent noise, we use three types of noise functions within the PMD noise family. To", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 294, + 334 + ], + "score": 1.0, + "content": "make the noise challenging enough, for input", + "type": "text" + }, + { + "bbox": [ + 295, + 324, + 303, + 332 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "we always corrupt label from the most confident", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 144, + 346 + ], + "score": 1.0, + "content": "category", + "type": "text" + }, + { + "bbox": [ + 144, + 335, + 156, + 344 + ], + "score": 0.85, + "content": "u _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 333, + 294, + 346 + ], + "score": 1.0, + "content": "to the second confident category", + "type": "text" + }, + { + "bbox": [ + 294, + 335, + 305, + 344 + ], + "score": 0.85, + "content": "{ s } _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 333, + 363, + 346 + ], + "score": 1.0, + "content": ", according to", + "type": "text" + }, + { + "bbox": [ + 363, + 333, + 384, + 345 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 333, + 426, + 346 + ], + "score": 1.0, + "content": ". Because", + "type": "text" + }, + { + "bbox": [ + 426, + 334, + 438, + 344 + ], + "score": 0.87, + "content": "s _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "is the class that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 144, + 356 + ], + "score": 1.0, + "content": "confuses", + "type": "text" + }, + { + "bbox": [ + 145, + 344, + 170, + 356 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 344, + 482, + 356 + ], + "score": 1.0, + "content": "the most, this noise will hurt the network’s performance the most. Note that", + "type": "text" + }, + { + "bbox": [ + 483, + 345, + 494, + 355 + ], + "score": 0.87, + "content": "y _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 163, + 368 + ], + "score": 1.0, + "content": "sampled from", + "type": "text" + }, + { + "bbox": [ + 163, + 355, + 183, + 367 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 355, + 450, + 368 + ], + "score": 1.0, + "content": ", which has quite an extreme confidence. Thus we generally assume", + "type": "text" + }, + { + "bbox": [ + 450, + 357, + 462, + 367 + ], + "score": 0.87, + "content": "y _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 355, + 472, + 368 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 473, + 357, + 484, + 366 + ], + "score": 0.87, + "content": "u _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 355, + 506, + 368 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 366, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 155, + 378 + ], + "score": 1.0, + "content": "each datum", + "type": "text" + }, + { + "bbox": [ + 155, + 368, + 162, + 376 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 366, + 234, + 378 + ], + "score": 1.0, + "content": ", we only flip it to", + "type": "text" + }, + { + "bbox": [ + 235, + 367, + 246, + 377 + ], + "score": 0.86, + "content": "s _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 366, + 298, + 378 + ], + "score": 1.0, + "content": "or keep it as", + "type": "text" + }, + { + "bbox": [ + 298, + 367, + 310, + 377 + ], + "score": 0.86, + "content": "u _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 366, + 479, + 378 + ], + "score": 1.0, + "content": ". The three noise functions are as follows:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 381, + 499, + 430 + ], + "lines": [ + { + "bbox": [ + 112, + 381, + 499, + 430 + ], + "spans": [ + { + "bbox": [ + 112, + 381, + 499, + 430 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathrm { T y p e - I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = - \\cfrac { 1 } { 2 } \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\cfrac { 1 } { 2 } , \\quad \\mathrm { T y p e - I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } , } \\\\ & { \\mathrm { T y p e - I I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\cfrac { 1 } { 3 } \\left[ \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "e090a158d4119449aa64536b038b227fae4ba4256d9894939ae1a4a1a42e836a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 112, + 381, + 499, + 397.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 112, + 397.3333333333333, + 499, + 413.66666666666663 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 112, + 413.66666666666663, + 499, + 429.99999999999994 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 297, + 445 + ], + "score": 1.0, + "content": "Notice that the noise level is determined by the", + "type": "text" + }, + { + "bbox": [ + 297, + 433, + 318, + 446 + ], + "score": 0.93, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 434, + 505, + 445 + ], + "score": 1.0, + "content": "naturally and we cannot control it directly. To", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 444, + 507, + 459 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 249, + 459 + ], + "score": 1.0, + "content": "change the noise level, we multiply", + "type": "text" + }, + { + "bbox": [ + 250, + 446, + 275, + 457 + ], + "score": 0.9, + "content": "\\tau _ { u _ { x } , s _ { x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 444, + 507, + 459 + ], + "score": 1.0, + "content": "by a certain constant factor such that the final proportion", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 408, + 469 + ], + "score": 1.0, + "content": "of noise matches our requirement. For PMD noise only, we test noise levels", + "type": "text" + }, + { + "bbox": [ + 408, + 456, + 428, + 466 + ], + "score": 0.89, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 455, + 445, + 469 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 446, + 456, + 464, + 466 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 455, + 505, + 469 + ], + "score": 1.0, + "content": ", meaning", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 466, + 402, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 123, + 479 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 467, + 144, + 477 + ], + "score": 0.87, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 466, + 161, + 479 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 467, + 181, + 477 + ], + "score": 0.86, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 466, + 402, + 479 + ], + "score": 1.0, + "content": "of the data are corrupted due to the noise, respectively.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "For i.i.d. noise we follow the convention and adopt the commonly used uniform noise and asymmetric", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "noise (Patrini et al., 2017). We artificially corrupt the labels by constructing the noise transition", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 505, + 134, + 519 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 135, + 506, + 143, + 515 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 505, + 173, + 519 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 173, + 505, + 288, + 518 + ], + "score": 0.92, + "content": "T _ { i j } = P ( \\widetilde { y } = j | y = i ) = \\tau _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 505, + 442, + 519 + ], + "score": 1.0, + "content": "defines the probability that a true label", + "type": "text" + }, + { + "bbox": [ + 442, + 506, + 465, + 517 + ], + "score": 0.91, + "content": "y = i", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "is flipped", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 117, + 529 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 517, + 122, + 528 + ], + "score": 0.79, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 516, + 254, + 529 + ], + "score": 1.0, + "content": "e. Then for each sample with label", + "type": "text" + }, + { + "bbox": [ + 254, + 518, + 258, + 526 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 516, + 505, + 529 + ], + "score": 1.0, + "content": ", we replace its label with the one sampled from the probability", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 209, + 540 + ], + "score": 1.0, + "content": "distribution given by the", + "type": "text" + }, + { + "bbox": [ + 209, + 528, + 214, + 537 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 528, + 285, + 540 + ], + "score": 1.0, + "content": "-th row of matrix", + "type": "text" + }, + { + "bbox": [ + 285, + 528, + 294, + 537 + ], + "score": 0.7, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 528, + 506, + 540 + ], + "score": 1.0, + "content": ". We consider two kinds of i.i.d. noise in this work.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 238, + 551 + ], + "score": 1.0, + "content": "(1) Uniform noise: the true label", + "type": "text" + }, + { + "bbox": [ + 239, + 539, + 243, + 548 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 538, + 419, + 551 + ], + "score": 1.0, + "content": "is corrupted uniformly to other classes, i.e.,", + "type": "text" + }, + { + "bbox": [ + 420, + 538, + 489, + 550 + ], + "score": 0.92, + "content": "T _ { i j } = \\tau / ( C - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 107, + 550, + 129, + 561 + ], + "score": 0.89, + "content": "i \\neq j", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 549, + 150, + 561 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 150, + 549, + 199, + 560 + ], + "score": 0.92, + "content": "T _ { i i } = 1 - \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 549, + 228, + 561 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 229, + 551, + 236, + 559 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 549, + 490, + 561 + ], + "score": 1.0, + "content": "is the constant noise level; (2) Asymmetric noise: the true label", + "type": "text" + }, + { + "bbox": [ + 490, + 550, + 495, + 559 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 560, + 461, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 147, + 573 + ], + "score": 1.0, + "content": "flipped to", + "type": "text" + }, + { + "bbox": [ + 147, + 561, + 153, + 572 + ], + "score": 0.85, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 561, + 305, + 573 + ], + "score": 1.0, + "content": "or stays unchanged with probabilities", + "type": "text" + }, + { + "bbox": [ + 306, + 560, + 338, + 573 + ], + "score": 0.93, + "content": "T _ { i j } = \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 561, + 356, + 573 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 560, + 406, + 571 + ], + "score": 0.92, + "content": "T _ { i i } = 1 - \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 561, + 461, + 573 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "score": 1.0, + "content": "Baselines. We compare our method with several recently proposed approaches. (1) GCE (Zhang", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 586, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 249, + 602 + ], + "score": 1.0, + "content": "& Sabuncu, 2018); (2) Co-teaching", + "type": "text" + }, + { + "bbox": [ + 249, + 590, + 256, + 598 + ], + "score": 0.33, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 586, + 505, + 602 + ], + "score": 1.0, + "content": "(Yu et al., 2019); (3) SL (Wang et al., 2019); (4) LRT (Zheng", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "et al., 2020). All these methods are generic and handle the label noise without assuming the noise", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "structures. Finally, we also provide the results by standard method, which simply trains the deep", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 621, + 300, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 300, + 634 + ], + "score": 1.0, + "content": "network on noisy datasets in a standard manner.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "During training, we use a batch size of 128 and train the network for 180 epochs to ensure the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 649, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 660 + ], + "score": 1.0, + "content": "convergence of all methods. We train the network with SGD optimizer, with initial learning rate 0.01.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "We randomly repeat the experiments 3 times, and report the mean and standard deviation values. Our", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "score": 1.0, + "content": "code is available at https://github.com/pxiangwu/PLC.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "Results. Table 1 lists the performance of different methods under three types of feature-dependent", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 192, + 711 + ], + "score": 1.0, + "content": "noise at noise levels", + "type": "text" + }, + { + "bbox": [ + 192, + 699, + 213, + 709 + ], + "score": 0.89, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 699, + 232, + 711 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 232, + 699, + 252, + 709 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 699, + 506, + 711 + ], + "score": 1.0, + "content": ". We observe that our method achieves the best performance", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "across different noise settings. Moreover, notice that some of the baseline methods’ performances are", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "inferior to the standard approach. Possible reasons are that these methods behave too conservatively", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "This lemma states that given an initial model that has a reasonably pure super level set, one can", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "manage to progressively correct a large fraction of corrupted labels by running Algorithm 1 for a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "sufficient long time with carefully chosen parameters. The limit of Algorithm 1 will depend on the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 507, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 435, + 128 + ], + "score": 1.0, + "content": "approximation ability of the neural network, which is characterized by parameter", + "type": "text" + }, + { + "bbox": [ + 436, + 117, + 442, + 126 + ], + "score": 0.72, + "content": "\\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 115, + 507, + 128 + ], + "score": 1.0, + "content": "in Definition 2.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 139 + ], + "score": 1.0, + "content": "To prove Theorem 1 using Lemma 3, it suffices to get a model which has a reliable region. This is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 477, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 477, + 150 + ], + "score": 1.0, + "content": "provably achievable by training with a family of good scoring functions on PMD noisy data.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 81, + 507, + 150 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 164, + 200, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 201, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 201, + 179 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 189, + 506, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 189, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 506, + 201 + ], + "score": 1.0, + "content": "We evaluate our method on both synthetic and real-world datasets. We first conduct synthetic experi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "ments on two public datasets CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009). To synthesize the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 212, + 506, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 353, + 224 + ], + "score": 1.0, + "content": "label noise, we first approximate the true posterior probability", + "type": "text" + }, + { + "bbox": [ + 353, + 213, + 360, + 223 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 212, + 506, + 224 + ], + "score": 1.0, + "content": "using the confidence prediction of a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 506, + 235 + ], + "score": 1.0, + "content": "clean neural network (trained with the original clean labels). We call these original labels raw labels.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 174, + 246 + ], + "score": 1.0, + "content": "Then we sample", + "type": "text" + }, + { + "bbox": [ + 174, + 234, + 219, + 245 + ], + "score": 0.93, + "content": "y _ { \\pmb { x } } \\sim \\eta ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 233, + 291, + 246 + ], + "score": 1.0, + "content": "for each instance", + "type": "text" + }, + { + "bbox": [ + 291, + 235, + 299, + 244 + ], + "score": 0.65, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 233, + 506, + 246 + ], + "score": 1.0, + "content": ". Instead of using raw labels, we use these sampled", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 132, + 257 + ], + "score": 1.0, + "content": "labels", + "type": "text" + }, + { + "bbox": [ + 132, + 246, + 144, + 256 + ], + "score": 0.86, + "content": "y _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 243, + 383, + 257 + ], + "score": 1.0, + "content": "as the clean labels, whose posterior probabilities are exactly", + "type": "text" + }, + { + "bbox": [ + 384, + 244, + 405, + 256 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "; and therefore the neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 254, + 507, + 268 + ], + "spans": [ + { + "bbox": [ + 104, + 254, + 268, + 268 + ], + "score": 1.0, + "content": "network is the Bayes optimal classifier", + "type": "text" + }, + { + "bbox": [ + 268, + 255, + 361, + 267 + ], + "score": 0.93, + "content": "\\eta ^ { * } : \\mathcal { X } \\{ 1 , \\cdot \\cdot , C \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 254, + 392, + 268 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 393, + 256, + 402, + 265 + ], + "score": 0.84, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 254, + 507, + 268 + ], + "score": 1.0, + "content": "is the number of classes.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 266, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 233, + 279 + ], + "score": 1.0, + "content": "Note that in multi-class setting,", + "type": "text" + }, + { + "bbox": [ + 234, + 266, + 254, + 278 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 266, + 350, + 279 + ], + "score": 1.0, + "content": "has a vector output and", + "type": "text" + }, + { + "bbox": [ + 351, + 266, + 374, + 278 + ], + "score": 0.93, + "content": "\\dot { \\eta _ { i } } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 266, + 399, + 279 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 399, + 267, + 404, + 276 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 266, + 507, + 279 + ], + "score": 1.0, + "content": "-th element of this vector.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 189, + 507, + 279 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 503, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Noise generation. We consider a generic family of noise. We consider not only feature-dependent", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 461, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 461, + 307 + ], + "score": 1.0, + "content": "noise, but also hybrid noise that consists of both feature-dependent noise and i.i.d. noise.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 282, + 505, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 323 + ], + "score": 1.0, + "content": "For feature-dependent noise, we use three types of noise functions within the PMD noise family. To", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 294, + 334 + ], + "score": 1.0, + "content": "make the noise challenging enough, for input", + "type": "text" + }, + { + "bbox": [ + 295, + 324, + 303, + 332 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "we always corrupt label from the most confident", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 144, + 346 + ], + "score": 1.0, + "content": "category", + "type": "text" + }, + { + "bbox": [ + 144, + 335, + 156, + 344 + ], + "score": 0.85, + "content": "u _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 333, + 294, + 346 + ], + "score": 1.0, + "content": "to the second confident category", + "type": "text" + }, + { + "bbox": [ + 294, + 335, + 305, + 344 + ], + "score": 0.85, + "content": "{ s } _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 333, + 363, + 346 + ], + "score": 1.0, + "content": ", according to", + "type": "text" + }, + { + "bbox": [ + 363, + 333, + 384, + 345 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 333, + 426, + 346 + ], + "score": 1.0, + "content": ". Because", + "type": "text" + }, + { + "bbox": [ + 426, + 334, + 438, + 344 + ], + "score": 0.87, + "content": "s _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "is the class that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 144, + 356 + ], + "score": 1.0, + "content": "confuses", + "type": "text" + }, + { + "bbox": [ + 145, + 344, + 170, + 356 + ], + "score": 0.91, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 344, + 482, + 356 + ], + "score": 1.0, + "content": "the most, this noise will hurt the network’s performance the most. Note that", + "type": "text" + }, + { + "bbox": [ + 483, + 345, + 494, + 355 + ], + "score": 0.87, + "content": "y _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 163, + 368 + ], + "score": 1.0, + "content": "sampled from", + "type": "text" + }, + { + "bbox": [ + 163, + 355, + 183, + 367 + ], + "score": 0.91, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 355, + 450, + 368 + ], + "score": 1.0, + "content": ", which has quite an extreme confidence. Thus we generally assume", + "type": "text" + }, + { + "bbox": [ + 450, + 357, + 462, + 367 + ], + "score": 0.87, + "content": "y _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 355, + 472, + 368 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 473, + 357, + 484, + 366 + ], + "score": 0.87, + "content": "u _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 355, + 506, + 368 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 366, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 155, + 378 + ], + "score": 1.0, + "content": "each datum", + "type": "text" + }, + { + "bbox": [ + 155, + 368, + 162, + 376 + ], + "score": 0.75, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 366, + 234, + 378 + ], + "score": 1.0, + "content": ", we only flip it to", + "type": "text" + }, + { + "bbox": [ + 235, + 367, + 246, + 377 + ], + "score": 0.86, + "content": "s _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 366, + 298, + 378 + ], + "score": 1.0, + "content": "or keep it as", + "type": "text" + }, + { + "bbox": [ + 298, + 367, + 310, + 377 + ], + "score": 0.86, + "content": "u _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 366, + 479, + 378 + ], + "score": 1.0, + "content": ". The three noise functions are as follows:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 311, + 506, + 378 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 381, + 499, + 430 + ], + "lines": [ + { + "bbox": [ + 112, + 381, + 499, + 430 + ], + "spans": [ + { + "bbox": [ + 112, + 381, + 499, + 430 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathrm { T y p e - I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = - \\cfrac { 1 } { 2 } \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\cfrac { 1 } { 2 } , \\quad \\mathrm { T y p e - I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } , } \\\\ & { \\mathrm { T y p e - I I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\cfrac { 1 } { 3 } \\left[ \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "e090a158d4119449aa64536b038b227fae4ba4256d9894939ae1a4a1a42e836a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 112, + 381, + 499, + 397.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 112, + 397.3333333333333, + 499, + 413.66666666666663 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 112, + 413.66666666666663, + 499, + 429.99999999999994 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 297, + 445 + ], + "score": 1.0, + "content": "Notice that the noise level is determined by the", + "type": "text" + }, + { + "bbox": [ + 297, + 433, + 318, + 446 + ], + "score": 0.93, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 434, + 505, + 445 + ], + "score": 1.0, + "content": "naturally and we cannot control it directly. To", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 444, + 507, + 459 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 249, + 459 + ], + "score": 1.0, + "content": "change the noise level, we multiply", + "type": "text" + }, + { + "bbox": [ + 250, + 446, + 275, + 457 + ], + "score": 0.9, + "content": "\\tau _ { u _ { x } , s _ { x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 444, + 507, + 459 + ], + "score": 1.0, + "content": "by a certain constant factor such that the final proportion", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 408, + 469 + ], + "score": 1.0, + "content": "of noise matches our requirement. For PMD noise only, we test noise levels", + "type": "text" + }, + { + "bbox": [ + 408, + 456, + 428, + 466 + ], + "score": 0.89, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 455, + 445, + 469 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 446, + 456, + 464, + 466 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 455, + 505, + 469 + ], + "score": 1.0, + "content": ", meaning", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 466, + 402, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 123, + 479 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 467, + 144, + 477 + ], + "score": 0.87, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 466, + 161, + 479 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 162, + 467, + 181, + 477 + ], + "score": 0.86, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 466, + 402, + 479 + ], + "score": 1.0, + "content": "of the data are corrupted due to the noise, respectively.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 433, + 507, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "For i.i.d. noise we follow the convention and adopt the commonly used uniform noise and asymmetric", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "noise (Patrini et al., 2017). We artificially corrupt the labels by constructing the noise transition", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 505, + 134, + 519 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 135, + 506, + 143, + 515 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 505, + 173, + 519 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 173, + 505, + 288, + 518 + ], + "score": 0.92, + "content": "T _ { i j } = P ( \\widetilde { y } = j | y = i ) = \\tau _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 505, + 442, + 519 + ], + "score": 1.0, + "content": "defines the probability that a true label", + "type": "text" + }, + { + "bbox": [ + 442, + 506, + 465, + 517 + ], + "score": 0.91, + "content": "y = i", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "is flipped", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 117, + 529 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 517, + 122, + 528 + ], + "score": 0.79, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 516, + 254, + 529 + ], + "score": 1.0, + "content": "e. Then for each sample with label", + "type": "text" + }, + { + "bbox": [ + 254, + 518, + 258, + 526 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 516, + 505, + 529 + ], + "score": 1.0, + "content": ", we replace its label with the one sampled from the probability", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 209, + 540 + ], + "score": 1.0, + "content": "distribution given by the", + "type": "text" + }, + { + "bbox": [ + 209, + 528, + 214, + 537 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 528, + 285, + 540 + ], + "score": 1.0, + "content": "-th row of matrix", + "type": "text" + }, + { + "bbox": [ + 285, + 528, + 294, + 537 + ], + "score": 0.7, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 528, + 506, + 540 + ], + "score": 1.0, + "content": ". We consider two kinds of i.i.d. noise in this work.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 238, + 551 + ], + "score": 1.0, + "content": "(1) Uniform noise: the true label", + "type": "text" + }, + { + "bbox": [ + 239, + 539, + 243, + 548 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 538, + 419, + 551 + ], + "score": 1.0, + "content": "is corrupted uniformly to other classes, i.e.,", + "type": "text" + }, + { + "bbox": [ + 420, + 538, + 489, + 550 + ], + "score": 0.92, + "content": "T _ { i j } = \\tau / ( C - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 107, + 550, + 129, + 561 + ], + "score": 0.89, + "content": "i \\neq j", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 549, + 150, + 561 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 150, + 549, + 199, + 560 + ], + "score": 0.92, + "content": "T _ { i i } = 1 - \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 549, + 228, + 561 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 229, + 551, + 236, + 559 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 549, + 490, + 561 + ], + "score": 1.0, + "content": "is the constant noise level; (2) Asymmetric noise: the true label", + "type": "text" + }, + { + "bbox": [ + 490, + 550, + 495, + 559 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 560, + 461, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 147, + 573 + ], + "score": 1.0, + "content": "flipped to", + "type": "text" + }, + { + "bbox": [ + 147, + 561, + 153, + 572 + ], + "score": 0.85, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 561, + 305, + 573 + ], + "score": 1.0, + "content": "or stays unchanged with probabilities", + "type": "text" + }, + { + "bbox": [ + 306, + 560, + 338, + 573 + ], + "score": 0.93, + "content": "T _ { i j } = \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 561, + 356, + 573 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 560, + 406, + 571 + ], + "score": 0.92, + "content": "T _ { i i } = 1 - \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 561, + 461, + 573 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 483, + 506, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 591 + ], + "score": 1.0, + "content": "Baselines. We compare our method with several recently proposed approaches. (1) GCE (Zhang", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 586, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 249, + 602 + ], + "score": 1.0, + "content": "& Sabuncu, 2018); (2) Co-teaching", + "type": "text" + }, + { + "bbox": [ + 249, + 590, + 256, + 598 + ], + "score": 0.33, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 586, + 505, + 602 + ], + "score": 1.0, + "content": "(Yu et al., 2019); (3) SL (Wang et al., 2019); (4) LRT (Zheng", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "et al., 2020). All these methods are generic and handle the label noise without assuming the noise", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "structures. Finally, we also provide the results by standard method, which simply trains the deep", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 621, + 300, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 300, + 634 + ], + "score": 1.0, + "content": "network on noisy datasets in a standard manner.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 575, + 506, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "During training, we use a batch size of 128 and train the network for 180 epochs to ensure the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 649, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 660 + ], + "score": 1.0, + "content": "convergence of all methods. We train the network with SGD optimizer, with initial learning rate 0.01.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "We randomly repeat the experiments 3 times, and report the mean and standard deviation values. Our", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "score": 1.0, + "content": "code is available at https://github.com/pxiangwu/PLC.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 637, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "Results. Table 1 lists the performance of different methods under three types of feature-dependent", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 192, + 711 + ], + "score": 1.0, + "content": "noise at noise levels", + "type": "text" + }, + { + "bbox": [ + 192, + 699, + 213, + 709 + ], + "score": 0.89, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 699, + 232, + 711 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 232, + 699, + 252, + 709 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 699, + 506, + 711 + ], + "score": 1.0, + "content": ". We observe that our method achieves the best performance", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "across different noise settings. Moreover, notice that some of the baseline methods’ performances are", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "inferior to the standard approach. Possible reasons are that these methods behave too conservatively", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "in dealing with noise. Thus they only make use of a small subset of the original training set, which is", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 398, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 398, + 106 + ], + "score": 1.0, + "content": "not representative enough to grant the model good discriminative ability.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 688, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "in dealing with noise. Thus they only make use of a small subset of the original training set, which is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 398, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 398, + 106 + ], + "score": 1.0, + "content": "not representative enough to grant the model good discriminative ability.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "In Table 2 we show the results on datasets corrupted with a combination of feature-dependent noise", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 351, + 133 + ], + "score": 1.0, + "content": "and i.i.d. noise, which ends up to real noise levels ranging from", + "type": "text" + }, + { + "bbox": [ + 351, + 122, + 370, + 132 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 120, + 380, + 133 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 380, + 122, + 400, + 132 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "(in terms of the proportion", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "of corrupted labels). I.i.d. noise is overlayed on the feature-dependent noise. Our method outperforms", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "baselines under these more complicated noise patterns. In contrast, when the noise level is high like", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 288, + 167 + ], + "score": 1.0, + "content": "the cases where we further apply additional", + "type": "text" + }, + { + "bbox": [ + 288, + 154, + 308, + 165 + ], + "score": 0.87, + "content": "30 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 154, + 326, + 167 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 327, + 154, + 347, + 165 + ], + "score": 0.87, + "content": "60 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "uniform noise, performances of a few", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 379, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 379, + 178 + ], + "score": 1.0, + "content": "baselines deteriorate and become worse than the standard approach.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 318, + 194 + ], + "score": 1.0, + "content": "We carry out the ablation studies on hyper-parameters", + "type": "text" + }, + { + "bbox": [ + 318, + 183, + 328, + 193 + ], + "score": 0.88, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "(determining the initial confidence threshold", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 281, + 205 + ], + "score": 1.0, + "content": "for label correction, see Algorithm 1) and", + "type": "text" + }, + { + "bbox": [ + 281, + 194, + 289, + 204 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "(the step size). In Tables 3 and 4, we show that our", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 261, + 217 + ], + "score": 1.0, + "content": "method is robust against the choice of", + "type": "text" + }, + { + "bbox": [ + 262, + 205, + 272, + 215 + ], + "score": 0.88, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 204, + 290, + 217 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 205, + 298, + 216 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "up to a wide range. Notice that to compare against", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 162, + 228 + ], + "score": 1.0, + "content": "the threshold", + "type": "text" + }, + { + "bbox": [ + 162, + 215, + 173, + 226 + ], + "score": 0.88, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 214, + 382, + 228 + ], + "score": 1.0, + "content": ", here we are calculating the absolute difference of", + "type": "text" + }, + { + "bbox": [ + 382, + 215, + 421, + 227 + ], + "score": 0.93, + "content": "\\log f _ { \\widetilde { y } } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 214, + 441, + 228 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 441, + 215, + 486, + 227 + ], + "score": 0.93, + "content": "\\log f _ { h _ { x } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 214, + 506, + 228 + ], + "score": 1.0, + "content": ". As", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 227, + 420, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 420, + 239 + ], + "score": 1.0, + "content": "mentioned in Section 2.2, this operation gives a good performance in practice.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 107, + 282, + 504, + 393 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 255, + 504, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 254, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 200, + 269 + ], + "score": 1.0, + "content": "Table 1: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 200, + 257, + 216, + 267 + ], + "score": 0.78, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 254, + 505, + 269 + ], + "score": 1.0, + "content": "on CIFAR-10 and CIFAR-100 under different feature-dependent noise", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 267, + 459, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 459, + 280 + ], + "score": 1.0, + "content": "types and levels. The average accuracy and standard deviation over 3 trials are reported.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 282, + 504, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 282, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 107, + 282, + 504, + 393 + ], + "score": 0.981, + "html": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I(35%)78.11 ± 0.7479.97 ± 0.1580.65 ±0.3979.76 ± 0.7280.98 ±0.8082.80±0.27
Type-I(70%)41.98 ± 1.9640.69 ± 1.9936.52 ± 1.6236.29 ±0.6641.52 ± 4.5342.74± 2.14
Type-II(35%)76.65±0.5777.34 ± 0.4477.60± 0.8877.92 ±0.8980.74±0.2581.54± 0.47
Type-II(70%)45.57 ± 1.1245.44 ± 0.6440.30 ± 1.4641.11 ± 1.9244.67 ± 3.8946.04±2.20
Type-III (35%)76.89±0.7978.38 ±0.6779.18 ± 0.6178.81 ±0.2981.08 ± 0.3581.50±0.50
CIFAR-100Type-III(70%)43.32 ± 1.0041.90 ± 0.8637.10 ± 0.5938.49 ± 1.4644.47 ± 1.2345.05 ±1.13
Type-I(35%)57.68 ±0.2956.70 ±0.7158.37 ±0.1855.20±0.3356.74 ±0.3460.01±0.43
Type-I(70%)39.32 ± 0.4339.53 ± 0.2840.01 ± 0.7140.02 ±0.8545.29 ± 0.4345.92 ± 0.61
Type-II (35%)57.83 ± 0.2556.57 ± 0.5258.11 ± 1.0556.10±0.7357.25 ±0.6863.68±0.29
Type-II(70%)39.30± 0.3236.84± 0.3937.75 ± 0.4638.45 ± 0.4543.71 ± 0.5145.03 ±0.50
Type-III (35%) Type-II(70%)56.07 ± 0.79 40.01 ± 0.1855.77 ± 0.98 35.37 ± 2.6557.51 ± 1.16 40.53 ±0.6056.04 ± 0.74 39.94 ± 0.8456.57 ±0.30 44.41 ± 0.1963.68 ± 0.29 44.45 ± 0.62
", + "type": "table", + "image_path": "b90b8a188e7c7188a0dfbb1868b8e64d34ee71f85aab2ed2920c335dfe4546e3.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 107, + 282, + 504, + 319.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 107, + 319.0, + 504, + 356.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 107, + 356.0, + 504, + 393.0 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 14.75 + }, + { + "type": "table", + "bbox": [ + 108, + 442, + 504, + 588 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 406, + 503, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 203, + 419 + ], + "score": 1.0, + "content": "Table 2: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 203, + 407, + 217, + 417 + ], + "score": 0.59, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "on CIFAR-10 and CIFAR-100 under different hybrid noise types and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 417, + 419, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 419, + 430 + ], + "score": 1.0, + "content": "levels. The average accuracy and standard deviation over 3 trials are reported.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 442, + 504, + 588 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 442, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 108, + 442, + 504, + 588 + ], + "score": 0.983, + "html": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I+30% Uniform75.26 ± 0.3278.72± 0.5378.08±0.6677.79± 0.4675.97± 0.2779.04±0.50
Type-I+ 60% Uniform64.25 ±0.7855.49 ± 2.1167.43 ± 1.4367.63 ± 1.3659.22 ± 0.7472.21 ± 2.92
Type-I+ 30% Asymmetric75.21 ± 0.6475.43 ± 2.9676.91 ± 0.5677.14± 0.7076.96 ± 0.4578.31 ± 0.41
Type-II +30% Uniform74.92 ± 0.6375.19 ± 0.5475.69 ± 0.2175.08± 0.4775.94 ± 0.5880.08±0.37
Type-II+ 60% Uniform64.02 ±0.6659.89 ±0.6366.39 ± 0.2966.76 ± 1.6058.99 ± 1.4371.21 ± 1.46
Type-II+ 30% Asymmetric74.28 ± 0.3973.37 ± 0.8375.30 ± 0.8175.43 ± 0.4277.03 ± 0.6277.63± 0.30
Type-III+30%Uniform74.00 ± 0.3877.31 ± 0.1177.00 ± 0.1276.22 ± 0.1275.66 ± 0.5780.06 ± 0.47
Type-III+ 60% Uniform63.96 ±0.6956.78± 1.5667.53 ± 0.5167.79 ± 0.5459.36 ± 0.9373.48 ±1.84
Type-III+ 30% Asymmetric75.31 ± 0.3474.62 ± 1.7175.70 ± 0.9176.09 ±0.1077.19 ± 0.7477.54±0.70
CIFAR-100Type-I+30%Uniform48.86±0.5652.33±0.6452.90±0.5351.34±0.6445.66 ± 1.6060.09±0.15
Type-I+ 60% Uniform35.97 ± 1.1227.17 ± 1.6638.62 ± 1.6537.57 ± 0.4323.37 ± 0.7251.68± 0.10
Type-I+ 30% Asymmetric45.85 ± 0.9351.21 ± 0.3152.69 ± 1.1450.18 ± 0.9752.04± 0.15
Type-II+30%Uniform49.32± 0.3651.99 ± 0.7553.61± 0.4643.86 ± 1.3156.40±0.34
Type-II + 60% Uniform35.16 ±0.0525.91 ± 0.6439.58 ±3.1350.58± 0.25 37.93± 0.2223.05 ±0.9960.01±0.63
Type-II + 30% Asymmetric46.50 ± 0.9551.07 ± 1.4451.98 ± 0.3749.46 ± 0.2352.11 ± 0.4649.35 ±1.53
Type-III +30% Uniform48.94± 0.6149.94± 0.4452.07±0.3550.18±0.5442.79 ± 1.7861.43±0.33 60.14±0.97
Type-III + 60% Uniform34.67± 0.1622.89 ± 0.7536.82 ±0.4937.65 ± 1.4222.81 ± 0.7250.73±2.16
Type-III + 30% Asymmetric45.70 ± 0.1249.38 ±0.8650.87 ± 1.1248.15±0.9050.31±0.3954.56±1.11
", + "type": "table", + "image_path": "85024251e7bf09e737cf617285400769cb8888555e8e37371d55a7e43087cf09.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 108, + 442, + 504, + 490.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 108, + 490.6666666666667, + 504, + 539.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 108, + 539.3333333333334, + 504, + 588.0 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 19.75 + }, + { + "type": "table", + "bbox": [ + 109, + 634, + 297, + 676 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 601, + 301, + 634 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 107, + 600, + 302, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 600, + 204, + 614 + ], + "score": 1.0, + "content": "Table 3: The effect of", + "type": "text" + }, + { + "bbox": [ + 204, + 602, + 214, + 613 + ], + "score": 0.87, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 600, + 302, + 614 + ], + "score": 1.0, + "content": "on the performance.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 612, + 301, + 625 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 204, + 625 + ], + "score": 1.0, + "content": "We use CIFAR-10 with", + "type": "text" + }, + { + "bbox": [ + 205, + 613, + 224, + 623 + ], + "score": 0.83, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 612, + 301, + 625 + ], + "score": 1.0, + "content": "feature-dependent", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 108, + 624, + 201, + 635 + ], + "spans": [ + { + "bbox": [ + 108, + 624, + 164, + 635 + ], + "score": 1.0, + "content": "noise, and set", + "type": "text" + }, + { + "bbox": [ + 164, + 624, + 198, + 635 + ], + "score": 0.89, + "content": "\\beta = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 624, + 201, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "table_body", + "bbox": [ + 109, + 634, + 297, + 676 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 109, + 634, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 109, + 634, + 297, + 676 + ], + "score": 0.96, + "html": "
exp(00)0.20.30.40.5
Type-INoise83.3383.0482.6682.94
Type-II Noise81.8481.1881.0981.24
Type-III Noise81.7981.7581.9882.06
", + "type": "table", + "image_path": "048bb69b12be8ea61594802e4852ee4b92fc34386c563c9fb6079d903a8c1881.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 109, + 634, + 297, + 648.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 109, + 648.0, + 297, + 662.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 109, + 662.0, + 297, + 676.0 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "index": 27.5 + }, + { + "type": "table", + "bbox": [ + 312, + 635, + 500, + 676 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 309, + 602, + 503, + 634 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 309, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 309, + 600, + 408, + 614 + ], + "score": 1.0, + "content": "Table 4: The effect of", + "type": "text" + }, + { + "bbox": [ + 408, + 602, + 416, + 613 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "on the performance.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 310, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 310, + 611, + 407, + 624 + ], + "score": 1.0, + "content": "We use CIFAR-10 with", + "type": "text" + }, + { + "bbox": [ + 407, + 613, + 427, + 623 + ], + "score": 0.85, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 611, + 504, + 624 + ], + "score": 1.0, + "content": "feature-dependent", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 310, + 623, + 432, + 635 + ], + "spans": [ + { + "bbox": [ + 310, + 623, + 381, + 635 + ], + "score": 1.0, + "content": "noise, and set ex", + "type": "text" + }, + { + "bbox": [ + 381, + 623, + 428, + 635 + ], + "score": 0.78, + "content": ") ( \\theta _ { 0 } ) = 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 623, + 432, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "table_body", + "bbox": [ + 312, + 635, + 500, + 676 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 312, + 635, + 500, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 500, + 676 + ], + "score": 0.963, + "html": "
β0.050.10.20.3
Type-I Noise83.5883.0483.2883.31
Type-II Noise80.9481.1880.9880.86
Type-III Noise81.9181.7582.1382.39
", + "type": "table", + "image_path": "efcc5cfbae92099fc153cb1ca450ce7f28558202ae4f98d60f9018201af565f5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 312, + 635, + 500, + 648.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 312, + 648.6666666666666, + 500, + 662.3333333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 312, + 662.3333333333333, + 500, + 675.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "Results on real-world noisy datasets. To test the effectiveness of the proposed method under real-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "world label noise, we conduct experiments on the Clothing1M dataset (Xiao et al., 2015). This dataset", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "contains 1 million clothing images obtained from online shopping websites with 14 categories. The", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 487, + 733 + ], + "score": 1.0, + "content": "labels in this dataset are quite noisy with an unknown underlying structure. This dataset provides", + "type": "text" + }, + { + "bbox": [ + 487, + 721, + 503, + 731 + ], + "score": 0.77, + "content": "5 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 720, + 506, + 733 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "In Table 2 we show the results on datasets corrupted with a combination of feature-dependent noise", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 351, + 133 + ], + "score": 1.0, + "content": "and i.i.d. noise, which ends up to real noise levels ranging from", + "type": "text" + }, + { + "bbox": [ + 351, + 122, + 370, + 132 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 120, + 380, + 133 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 380, + 122, + 400, + 132 + ], + "score": 0.87, + "content": "70 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "(in terms of the proportion", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "of corrupted labels). I.i.d. noise is overlayed on the feature-dependent noise. Our method outperforms", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "baselines under these more complicated noise patterns. In contrast, when the noise level is high like", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 288, + 167 + ], + "score": 1.0, + "content": "the cases where we further apply additional", + "type": "text" + }, + { + "bbox": [ + 288, + 154, + 308, + 165 + ], + "score": 0.87, + "content": "30 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 154, + 326, + 167 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 327, + 154, + 347, + 165 + ], + "score": 0.87, + "content": "60 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "uniform noise, performances of a few", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 379, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 379, + 178 + ], + "score": 1.0, + "content": "baselines deteriorate and become worse than the standard approach.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 110, + 505, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 318, + 194 + ], + "score": 1.0, + "content": "We carry out the ablation studies on hyper-parameters", + "type": "text" + }, + { + "bbox": [ + 318, + 183, + 328, + 193 + ], + "score": 0.88, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "(determining the initial confidence threshold", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 281, + 205 + ], + "score": 1.0, + "content": "for label correction, see Algorithm 1) and", + "type": "text" + }, + { + "bbox": [ + 281, + 194, + 289, + 204 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "(the step size). In Tables 3 and 4, we show that our", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 261, + 217 + ], + "score": 1.0, + "content": "method is robust against the choice of", + "type": "text" + }, + { + "bbox": [ + 262, + 205, + 272, + 215 + ], + "score": 0.88, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 204, + 290, + 217 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 205, + 298, + 216 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "up to a wide range. Notice that to compare against", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 162, + 228 + ], + "score": 1.0, + "content": "the threshold", + "type": "text" + }, + { + "bbox": [ + 162, + 215, + 173, + 226 + ], + "score": 0.88, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 214, + 382, + 228 + ], + "score": 1.0, + "content": ", here we are calculating the absolute difference of", + "type": "text" + }, + { + "bbox": [ + 382, + 215, + 421, + 227 + ], + "score": 0.93, + "content": "\\log f _ { \\widetilde { y } } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 214, + 441, + 228 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 441, + 215, + 486, + 227 + ], + "score": 0.93, + "content": "\\log f _ { h _ { x } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 214, + 506, + 228 + ], + "score": 1.0, + "content": ". As", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 227, + 420, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 420, + 239 + ], + "score": 1.0, + "content": "mentioned in Section 2.2, this operation gives a good performance in practice.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 182, + 506, + 239 + ] + }, + { + "type": "table", + "bbox": [ + 107, + 282, + 504, + 393 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 255, + 504, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 254, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 200, + 269 + ], + "score": 1.0, + "content": "Table 1: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 200, + 257, + 216, + 267 + ], + "score": 0.78, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 254, + 505, + 269 + ], + "score": 1.0, + "content": "on CIFAR-10 and CIFAR-100 under different feature-dependent noise", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 267, + 459, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 459, + 280 + ], + "score": 1.0, + "content": "types and levels. The average accuracy and standard deviation over 3 trials are reported.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 282, + 504, + 393 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 282, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 107, + 282, + 504, + 393 + ], + "score": 0.981, + "html": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I(35%)78.11 ± 0.7479.97 ± 0.1580.65 ±0.3979.76 ± 0.7280.98 ±0.8082.80±0.27
Type-I(70%)41.98 ± 1.9640.69 ± 1.9936.52 ± 1.6236.29 ±0.6641.52 ± 4.5342.74± 2.14
Type-II(35%)76.65±0.5777.34 ± 0.4477.60± 0.8877.92 ±0.8980.74±0.2581.54± 0.47
Type-II(70%)45.57 ± 1.1245.44 ± 0.6440.30 ± 1.4641.11 ± 1.9244.67 ± 3.8946.04±2.20
Type-III (35%)76.89±0.7978.38 ±0.6779.18 ± 0.6178.81 ±0.2981.08 ± 0.3581.50±0.50
CIFAR-100Type-III(70%)43.32 ± 1.0041.90 ± 0.8637.10 ± 0.5938.49 ± 1.4644.47 ± 1.2345.05 ±1.13
Type-I(35%)57.68 ±0.2956.70 ±0.7158.37 ±0.1855.20±0.3356.74 ±0.3460.01±0.43
Type-I(70%)39.32 ± 0.4339.53 ± 0.2840.01 ± 0.7140.02 ±0.8545.29 ± 0.4345.92 ± 0.61
Type-II (35%)57.83 ± 0.2556.57 ± 0.5258.11 ± 1.0556.10±0.7357.25 ±0.6863.68±0.29
Type-II(70%)39.30± 0.3236.84± 0.3937.75 ± 0.4638.45 ± 0.4543.71 ± 0.5145.03 ±0.50
Type-III (35%) Type-II(70%)56.07 ± 0.79 40.01 ± 0.1855.77 ± 0.98 35.37 ± 2.6557.51 ± 1.16 40.53 ±0.6056.04 ± 0.74 39.94 ± 0.8456.57 ±0.30 44.41 ± 0.1963.68 ± 0.29 44.45 ± 0.62
", + "type": "table", + "image_path": "b90b8a188e7c7188a0dfbb1868b8e64d34ee71f85aab2ed2920c335dfe4546e3.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 107, + 282, + 504, + 319.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 107, + 319.0, + 504, + 356.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 107, + 356.0, + 504, + 393.0 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 14.75 + }, + { + "type": "table", + "bbox": [ + 108, + 442, + 504, + 588 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 406, + 503, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 203, + 419 + ], + "score": 1.0, + "content": "Table 2: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 203, + 407, + 217, + 417 + ], + "score": 0.59, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "on CIFAR-10 and CIFAR-100 under different hybrid noise types and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 417, + 419, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 419, + 430 + ], + "score": 1.0, + "content": "levels. The average accuracy and standard deviation over 3 trials are reported.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 442, + 504, + 588 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 442, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 108, + 442, + 504, + 588 + ], + "score": 0.983, + "html": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I+30% Uniform75.26 ± 0.3278.72± 0.5378.08±0.6677.79± 0.4675.97± 0.2779.04±0.50
Type-I+ 60% Uniform64.25 ±0.7855.49 ± 2.1167.43 ± 1.4367.63 ± 1.3659.22 ± 0.7472.21 ± 2.92
Type-I+ 30% Asymmetric75.21 ± 0.6475.43 ± 2.9676.91 ± 0.5677.14± 0.7076.96 ± 0.4578.31 ± 0.41
Type-II +30% Uniform74.92 ± 0.6375.19 ± 0.5475.69 ± 0.2175.08± 0.4775.94 ± 0.5880.08±0.37
Type-II+ 60% Uniform64.02 ±0.6659.89 ±0.6366.39 ± 0.2966.76 ± 1.6058.99 ± 1.4371.21 ± 1.46
Type-II+ 30% Asymmetric74.28 ± 0.3973.37 ± 0.8375.30 ± 0.8175.43 ± 0.4277.03 ± 0.6277.63± 0.30
Type-III+30%Uniform74.00 ± 0.3877.31 ± 0.1177.00 ± 0.1276.22 ± 0.1275.66 ± 0.5780.06 ± 0.47
Type-III+ 60% Uniform63.96 ±0.6956.78± 1.5667.53 ± 0.5167.79 ± 0.5459.36 ± 0.9373.48 ±1.84
Type-III+ 30% Asymmetric75.31 ± 0.3474.62 ± 1.7175.70 ± 0.9176.09 ±0.1077.19 ± 0.7477.54±0.70
CIFAR-100Type-I+30%Uniform48.86±0.5652.33±0.6452.90±0.5351.34±0.6445.66 ± 1.6060.09±0.15
Type-I+ 60% Uniform35.97 ± 1.1227.17 ± 1.6638.62 ± 1.6537.57 ± 0.4323.37 ± 0.7251.68± 0.10
Type-I+ 30% Asymmetric45.85 ± 0.9351.21 ± 0.3152.69 ± 1.1450.18 ± 0.9752.04± 0.15
Type-II+30%Uniform49.32± 0.3651.99 ± 0.7553.61± 0.4643.86 ± 1.3156.40±0.34
Type-II + 60% Uniform35.16 ±0.0525.91 ± 0.6439.58 ±3.1350.58± 0.25 37.93± 0.2223.05 ±0.9960.01±0.63
Type-II + 30% Asymmetric46.50 ± 0.9551.07 ± 1.4451.98 ± 0.3749.46 ± 0.2352.11 ± 0.4649.35 ±1.53
Type-III +30% Uniform48.94± 0.6149.94± 0.4452.07±0.3550.18±0.5442.79 ± 1.7861.43±0.33 60.14±0.97
Type-III + 60% Uniform34.67± 0.1622.89 ± 0.7536.82 ±0.4937.65 ± 1.4222.81 ± 0.7250.73±2.16
Type-III + 30% Asymmetric45.70 ± 0.1249.38 ±0.8650.87 ± 1.1248.15±0.9050.31±0.3954.56±1.11
", + "type": "table", + "image_path": "85024251e7bf09e737cf617285400769cb8888555e8e37371d55a7e43087cf09.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 108, + 442, + 504, + 490.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 108, + 490.6666666666667, + 504, + 539.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 108, + 539.3333333333334, + 504, + 588.0 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 19.75 + }, + { + "type": "table", + "bbox": [ + 109, + 634, + 297, + 676 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 601, + 301, + 634 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 107, + 600, + 302, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 600, + 204, + 614 + ], + "score": 1.0, + "content": "Table 3: The effect of", + "type": "text" + }, + { + "bbox": [ + 204, + 602, + 214, + 613 + ], + "score": 0.87, + "content": "\\theta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 600, + 302, + 614 + ], + "score": 1.0, + "content": "on the performance.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 612, + 301, + 625 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 204, + 625 + ], + "score": 1.0, + "content": "We use CIFAR-10 with", + "type": "text" + }, + { + "bbox": [ + 205, + 613, + 224, + 623 + ], + "score": 0.83, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 612, + 301, + 625 + ], + "score": 1.0, + "content": "feature-dependent", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 108, + 624, + 201, + 635 + ], + "spans": [ + { + "bbox": [ + 108, + 624, + 164, + 635 + ], + "score": 1.0, + "content": "noise, and set", + "type": "text" + }, + { + "bbox": [ + 164, + 624, + 198, + 635 + ], + "score": 0.89, + "content": "\\beta = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 624, + 201, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "table_body", + "bbox": [ + 109, + 634, + 297, + 676 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 109, + 634, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 109, + 634, + 297, + 676 + ], + "score": 0.96, + "html": "
exp(00)0.20.30.40.5
Type-INoise83.3383.0482.6682.94
Type-II Noise81.8481.1881.0981.24
Type-III Noise81.7981.7581.9882.06
", + "type": "table", + "image_path": "048bb69b12be8ea61594802e4852ee4b92fc34386c563c9fb6079d903a8c1881.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 109, + 634, + 297, + 648.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 109, + 648.0, + 297, + 662.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 109, + 662.0, + 297, + 676.0 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "index": 27.5 + }, + { + "type": "table", + "bbox": [ + 312, + 635, + 500, + 676 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 309, + 602, + 503, + 634 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 309, + 600, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 309, + 600, + 408, + 614 + ], + "score": 1.0, + "content": "Table 4: The effect of", + "type": "text" + }, + { + "bbox": [ + 408, + 602, + 416, + 613 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 600, + 505, + 614 + ], + "score": 1.0, + "content": "on the performance.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 310, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 310, + 611, + 407, + 624 + ], + "score": 1.0, + "content": "We use CIFAR-10 with", + "type": "text" + }, + { + "bbox": [ + 407, + 613, + 427, + 623 + ], + "score": 0.85, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 611, + 504, + 624 + ], + "score": 1.0, + "content": "feature-dependent", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 310, + 623, + 432, + 635 + ], + "spans": [ + { + "bbox": [ + 310, + 623, + 381, + 635 + ], + "score": 1.0, + "content": "noise, and set ex", + "type": "text" + }, + { + "bbox": [ + 381, + 623, + 428, + 635 + ], + "score": 0.78, + "content": ") ( \\theta _ { 0 } ) = 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 623, + 432, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "table_body", + "bbox": [ + 312, + 635, + 500, + 676 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 312, + 635, + 500, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 500, + 676 + ], + "score": 0.963, + "html": "
β0.050.10.20.3
Type-I Noise83.5883.0483.2883.31
Type-II Noise80.9481.1880.9880.86
Type-III Noise81.9181.7582.1382.39
", + "type": "table", + "image_path": "efcc5cfbae92099fc153cb1ca450ce7f28558202ae4f98d60f9018201af565f5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 312, + 635, + 500, + 648.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 312, + 648.6666666666666, + 500, + 662.3333333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 312, + 662.3333333333333, + 500, + 675.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "Results on real-world noisy datasets. To test the effectiveness of the proposed method under real-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "world label noise, we conduct experiments on the Clothing1M dataset (Xiao et al., 2015). This dataset", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "contains 1 million clothing images obtained from online shopping websites with 14 categories. The", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 487, + 733 + ], + "score": 1.0, + "content": "labels in this dataset are quite noisy with an unknown underlying structure. This dataset provides", + "type": "text" + }, + { + "bbox": [ + 487, + 721, + 503, + 731 + ], + "score": 0.77, + "content": "5 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 720, + 506, + 733 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 688, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 122, + 93 + ], + "score": 0.82, + "content": "1 4 k", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 81, + 140, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 140, + 83, + 155, + 93 + ], + "score": 0.82, + "content": "1 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "manually verified clean data for training, validation and testing, respectively. Following", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 394, + 106 + ], + "score": 1.0, + "content": "(Tanaka et al., 2018; Yi & Wu, 2019), in our experiment we discard the", + "type": "text" + }, + { + "bbox": [ + 394, + 94, + 411, + 104 + ], + "score": 0.79, + "content": "5 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "clean training data and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 272, + 118 + ], + "score": 1.0, + "content": "evaluate the classification accuracy on the", + "type": "text" + }, + { + "bbox": [ + 273, + 105, + 289, + 115 + ], + "score": 0.8, + "content": "1 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "clean data. Also, following (Yi & Wu, 2019), we use a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 434, + 128 + ], + "score": 1.0, + "content": "randomly sampled pseudo-balanced subset as the training set, which includes about", + "type": "text" + }, + { + "bbox": [ + 435, + 115, + 456, + 126 + ], + "score": 0.82, + "content": "2 6 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "images. We", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "set the batch size 32, learning rate 0.001, and adopt SGD optimizer and use ResNet-50 with weights", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 384, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 384, + 150 + ], + "score": 1.0, + "content": "pre-trained on ImageNet, as in (Tanaka et al., 2018; Yi & Wu, 2019).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "We compare our method with the following baselines. (1) Standard; (2) Forward Correction (Patrini", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "et al., 2017); (3) D2L (Ma et al., 2018); (4) JO (Tanaka et al., 2018); (5) PENCIL (Yi & Wu, 2019); (6)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "DY (Arazo et al., 2019); (7) GCE (Zhang & Sabuncu, 2018); (8) SL (Wang et al., 2019); (9) MLNT", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "score": 1.0, + "content": "(Li et al., 2019); (10) LRT (Zheng et al., 2020). In Table 5 we observe that our method achieves the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "score": 1.0, + "content": "best performance, suggesting the applicability of our label correction strategy in real-world scenarios.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 214, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 506, + 227 + ], + "score": 1.0, + "content": "Apart from Clothing1M, we also test our method on another smaller dataset, Food-101N (Lee et al.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "2018). Food-101N is a dataset for food classification, and consists of 310k training images collected", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 288, + 249 + ], + "score": 1.0, + "content": "from the web. The estimated label purity is", + "type": "text" + }, + { + "bbox": [ + 288, + 237, + 308, + 248 + ], + "score": 0.88, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 237, + 505, + 249 + ], + "score": 1.0, + "content": ". Following (Lee et al., 2018), the classification", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "accuracy is evaluated on the Food-101 (Bossard et al., 2014) testing set, which contains 25k images", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 506, + 271 + ], + "score": 1.0, + "content": "with curated annotations. We use ResNet-50 pre-trained on ImageNet. We train the network for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "score": 1.0, + "content": "30 epochs with SGD optimizer. The batch size is 32 and the initial learning rate is 0.005, which", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "is divided by 10 every 10 epochs. We also adopt simple data augmentation procedures, including", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 290, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 506, + 306 + ], + "score": 1.0, + "content": "random horizontal flip, and resizing the image with a short edge of 256 and then randomly cropping", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 113, + 316 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 113, + 303, + 150, + 313 + ], + "score": 0.78, + "content": "2 2 4 \\mathbf { x } 2 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "patch from the resized image. We repeat the experiments with 3 random trials and report", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "the mean value and standard deviation. The results are shown in Table 6. Our method much improves", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 230, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 230, + 337 + ], + "score": 1.0, + "content": "upon the previous approaches.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 507, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 507, + 354 + ], + "score": 1.0, + "content": "Finally, we test our method on a recently proposed real-world dataset, ANIMAL-10N (Song et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 507, + 365 + ], + "score": 1.0, + "content": "2019). This dataset contains human-labeled online images for 10 animals with confusing appearance.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 361, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 236, + 377 + ], + "score": 1.0, + "content": "The estimated label noise rate is", + "type": "text" + }, + { + "bbox": [ + 236, + 363, + 250, + 374 + ], + "score": 0.87, + "content": "8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 361, + 506, + 377 + ], + "score": 1.0, + "content": ". There are 50,000 training and 5,000 testing images. Following", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "(Song et al., 2019), we use VGG-19 with batch normalization. The SGD optimizer is employed. Also", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "following (Song et al., 2019), we train the network for 100 epochs and use an initial learning rate of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 219, + 408 + ], + "score": 1.0, + "content": "0.1, which is divided by 5 at", + "type": "text" + }, + { + "bbox": [ + 220, + 397, + 239, + 407 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 396, + 257, + 408 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 397, + 276, + 407 + ], + "score": 0.87, + "content": "7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "of the total number of epochs. We repeat the experiments", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "with 3 random trials and report the mean value and standard deviation. As is shown in Table 7, our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 419, + 280, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 280, + 430 + ], + "score": 1.0, + "content": "method outperforms the existing baselines.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "table", + "bbox": [ + 108, + 454, + 505, + 474 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 219, + 443, + 392, + 453 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 217, + 442, + 393, + 456 + ], + "spans": [ + { + "bbox": [ + 217, + 442, + 310, + 456 + ], + "score": 1.0, + "content": "Table 5: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 311, + 443, + 326, + 453 + ], + "score": 0.6, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 442, + 393, + 456 + ], + "score": 1.0, + "content": "on Clothing1M.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "table_body", + "bbox": [ + 108, + 454, + 505, + 474 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 454, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 108, + 454, + 505, + 474 + ], + "score": 0.939, + "html": "
MethodStandardForwardD2LJOPENCILDYGCESLMLNTLRTPLC (ours)
Accuracy68.9469.8469.4772.2373.4971.0069.7571.0273.4771.7474.02
", + "type": "table", + "image_path": "55cd00a15a47662b4ebbb40d2ac60b6e00cde8207e7807eacee88ab144bca0dc.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 108, + 454, + 505, + 474 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 30.5 + }, + { + "type": "table", + "bbox": [ + 110, + 502, + 296, + 548 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 117, + 491, + 290, + 502 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 118, + 490, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 118, + 490, + 210, + 503 + ], + "score": 1.0, + "content": "Table 6: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 210, + 491, + 226, + 502 + ], + "score": 0.6, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 490, + 290, + 503 + ], + "score": 1.0, + "content": "on Food-101N.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "table_body", + "bbox": [ + 110, + 502, + 296, + 548 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 502, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 110, + 502, + 296, + 548 + ], + "score": 0.971, + "html": "
MethodAccuracy
Standard81.67
CleanNet (Lee et al., 2018)83.95
PLC (ours)85.28 ± 0.04
", + "type": "table", + "image_path": "c46d02ae8ff7d1c13f38ca2f049218de3d9fc3c1f93be2a428df3d5956f2157c.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 110, + 502, + 296, + 517.3333333333334 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 110, + 517.3333333333334, + 296, + 532.6666666666667 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 110, + 532.6666666666667, + 296, + 548.0000000000001 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 34.0 + }, + { + "type": "table", + "bbox": [ + 312, + 501, + 499, + 549 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 315, + 491, + 500, + 501 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 312, + 489, + 500, + 503 + ], + "spans": [ + { + "bbox": [ + 312, + 489, + 406, + 503 + ], + "score": 1.0, + "content": "Table 7: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 406, + 491, + 421, + 501 + ], + "score": 0.62, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 489, + 500, + 503 + ], + "score": 1.0, + "content": "on ANIMAL-10N.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "table_body", + "bbox": [ + 312, + 501, + 499, + 549 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 312, + 501, + 499, + 549 + ], + "spans": [ + { + "bbox": [ + 312, + 501, + 499, + 549 + ], + "score": 0.972, + "html": "
MethodAccuracy
Standard79.4 ± 0.14
SELFIE (Song et al., 2019)81.8 ± 0.09
PLC (ours)83.4±( 0.43
", + "type": "table", + "image_path": "99edab4c835c0240b7ed3b062d65e038ecfb04fbb7c282bbf751e0051e904ed9.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 312, + 501, + 499, + 517.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 312, + 517.0, + 499, + 533.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 312, + 533.0, + 499, + 549.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 35.0 + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 195, + 582 + ], + "lines": [ + { + "bbox": [ + 104, + 567, + 198, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 198, + 586 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "We propose a novel family of feature-dependent label noise that is much more general than the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 605, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 507, + 618 + ], + "score": 1.0, + "content": "traditional i.i.d. noise pattern. Building upon this noise assumption, we propose the first data-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "recalibrating method that is theoretically guaranteed to converge to a well-behaved classifier. On the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 627, + 507, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 507, + 639 + ], + "score": 1.0, + "content": "synthetic datasets, we show that our method outperforms various baselines under different feature-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "dependent noise patterns subject to our assumption. Also, we test our method on different real-world", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "noisy datasets and observe superior performances over existing approaches. The proposed noise", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 659, + 372, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 372, + 673 + ], + "score": 1.0, + "content": "family offers a new theoretical setting for the study of label noise.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "Acknowledgement. The authors acknowledge support from US National Science Foundation (NSF)", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "awards CRII-1755791, CCF-1910873, CCF-1855760. This effort was partially supported by the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "Intelligence Advanced Research Projects Agency (IARPA) under the contract W911NF20C0038.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 708, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 507, + 724 + ], + "score": 1.0, + "content": "The content of this paper does not necessarily reflect the position or the policy of the Government,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 296, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 296, + 733 + ], + "score": 1.0, + "content": "and no official endorsement should be inferred.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 122, + 93 + ], + "score": 0.82, + "content": "1 4 k", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 81, + 140, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 140, + 83, + 155, + 93 + ], + "score": 0.82, + "content": "1 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "manually verified clean data for training, validation and testing, respectively. Following", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 394, + 106 + ], + "score": 1.0, + "content": "(Tanaka et al., 2018; Yi & Wu, 2019), in our experiment we discard the", + "type": "text" + }, + { + "bbox": [ + 394, + 94, + 411, + 104 + ], + "score": 0.79, + "content": "5 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "clean training data and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 272, + 118 + ], + "score": 1.0, + "content": "evaluate the classification accuracy on the", + "type": "text" + }, + { + "bbox": [ + 273, + 105, + 289, + 115 + ], + "score": 0.8, + "content": "1 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "clean data. Also, following (Yi & Wu, 2019), we use a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 434, + 128 + ], + "score": 1.0, + "content": "randomly sampled pseudo-balanced subset as the training set, which includes about", + "type": "text" + }, + { + "bbox": [ + 435, + 115, + 456, + 126 + ], + "score": 0.82, + "content": "2 6 0 k", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "images. We", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "set the batch size 32, learning rate 0.001, and adopt SGD optimizer and use ResNet-50 with weights", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 384, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 384, + 150 + ], + "score": 1.0, + "content": "pre-trained on ImageNet, as in (Tanaka et al., 2018; Yi & Wu, 2019).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 81, + 506, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "We compare our method with the following baselines. (1) Standard; (2) Forward Correction (Patrini", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "et al., 2017); (3) D2L (Ma et al., 2018); (4) JO (Tanaka et al., 2018); (5) PENCIL (Yi & Wu, 2019); (6)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "DY (Arazo et al., 2019); (7) GCE (Zhang & Sabuncu, 2018); (8) SL (Wang et al., 2019); (9) MLNT", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "score": 1.0, + "content": "(Li et al., 2019); (10) LRT (Zheng et al., 2020). In Table 5 we observe that our method achieves the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 212 + ], + "score": 1.0, + "content": "best performance, suggesting the applicability of our label correction strategy in real-world scenarios.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 154, + 506, + 212 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 214, + 505, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 215, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 506, + 227 + ], + "score": 1.0, + "content": "Apart from Clothing1M, we also test our method on another smaller dataset, Food-101N (Lee et al.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "2018). Food-101N is a dataset for food classification, and consists of 310k training images collected", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 288, + 249 + ], + "score": 1.0, + "content": "from the web. The estimated label purity is", + "type": "text" + }, + { + "bbox": [ + 288, + 237, + 308, + 248 + ], + "score": 0.88, + "content": "80 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 237, + 505, + 249 + ], + "score": 1.0, + "content": ". Following (Lee et al., 2018), the classification", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "accuracy is evaluated on the Food-101 (Bossard et al., 2014) testing set, which contains 25k images", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 506, + 271 + ], + "score": 1.0, + "content": "with curated annotations. We use ResNet-50 pre-trained on ImageNet. We train the network for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "score": 1.0, + "content": "30 epochs with SGD optimizer. The batch size is 32 and the initial learning rate is 0.005, which", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "is divided by 10 every 10 epochs. We also adopt simple data augmentation procedures, including", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 290, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 506, + 306 + ], + "score": 1.0, + "content": "random horizontal flip, and resizing the image with a short edge of 256 and then randomly cropping", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 113, + 316 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 113, + 303, + 150, + 313 + ], + "score": 0.78, + "content": "2 2 4 \\mathbf { x } 2 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "patch from the resized image. We repeat the experiments with 3 random trials and report", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "the mean value and standard deviation. The results are shown in Table 6. Our method much improves", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 230, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 230, + 337 + ], + "score": 1.0, + "content": "upon the previous approaches.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 215, + 506, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 507, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 507, + 354 + ], + "score": 1.0, + "content": "Finally, we test our method on a recently proposed real-world dataset, ANIMAL-10N (Song et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 507, + 365 + ], + "score": 1.0, + "content": "2019). This dataset contains human-labeled online images for 10 animals with confusing appearance.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 361, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 236, + 377 + ], + "score": 1.0, + "content": "The estimated label noise rate is", + "type": "text" + }, + { + "bbox": [ + 236, + 363, + 250, + 374 + ], + "score": 0.87, + "content": "8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 361, + 506, + 377 + ], + "score": 1.0, + "content": ". There are 50,000 training and 5,000 testing images. Following", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "(Song et al., 2019), we use VGG-19 with batch normalization. The SGD optimizer is employed. Also", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "following (Song et al., 2019), we train the network for 100 epochs and use an initial learning rate of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 219, + 408 + ], + "score": 1.0, + "content": "0.1, which is divided by 5 at", + "type": "text" + }, + { + "bbox": [ + 220, + 397, + 239, + 407 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 396, + 257, + 408 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 397, + 276, + 407 + ], + "score": 0.87, + "content": "7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "of the total number of epochs. We repeat the experiments", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "with 3 random trials and report the mean value and standard deviation. As is shown in Table 7, our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 419, + 280, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 280, + 430 + ], + "score": 1.0, + "content": "method outperforms the existing baselines.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 340, + 507, + 430 + ] + }, + { + "type": "table", + "bbox": [ + 108, + 454, + 505, + 474 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 219, + 443, + 392, + 453 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 217, + 442, + 393, + 456 + ], + "spans": [ + { + "bbox": [ + 217, + 442, + 310, + 456 + ], + "score": 1.0, + "content": "Table 5: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 311, + 443, + 326, + 453 + ], + "score": 0.6, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 442, + 393, + 456 + ], + "score": 1.0, + "content": "on Clothing1M.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "table_body", + "bbox": [ + 108, + 454, + 505, + 474 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 454, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 108, + 454, + 505, + 474 + ], + "score": 0.939, + "html": "
MethodStandardForwardD2LJOPENCILDYGCESLMLNTLRTPLC (ours)
Accuracy68.9469.8469.4772.2373.4971.0069.7571.0273.4771.7474.02
", + "type": "table", + "image_path": "55cd00a15a47662b4ebbb40d2ac60b6e00cde8207e7807eacee88ab144bca0dc.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 108, + 454, + 505, + 474 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 30.5 + }, + { + "type": "table", + "bbox": [ + 110, + 502, + 296, + 548 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 117, + 491, + 290, + 502 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 118, + 490, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 118, + 490, + 210, + 503 + ], + "score": 1.0, + "content": "Table 6: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 210, + 491, + 226, + 502 + ], + "score": 0.6, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 490, + 290, + 503 + ], + "score": 1.0, + "content": "on Food-101N.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "table_body", + "bbox": [ + 110, + 502, + 296, + 548 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 502, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 110, + 502, + 296, + 548 + ], + "score": 0.971, + "html": "
MethodAccuracy
Standard81.67
CleanNet (Lee et al., 2018)83.95
PLC (ours)85.28 ± 0.04
", + "type": "table", + "image_path": "c46d02ae8ff7d1c13f38ca2f049218de3d9fc3c1f93be2a428df3d5956f2157c.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 110, + 502, + 296, + 517.3333333333334 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 110, + 517.3333333333334, + 296, + 532.6666666666667 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 110, + 532.6666666666667, + 296, + 548.0000000000001 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 34.0 + }, + { + "type": "table", + "bbox": [ + 312, + 501, + 499, + 549 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 315, + 491, + 500, + 501 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 312, + 489, + 500, + 503 + ], + "spans": [ + { + "bbox": [ + 312, + 489, + 406, + 503 + ], + "score": 1.0, + "content": "Table 7: Test accuracy", + "type": "text" + }, + { + "bbox": [ + 406, + 491, + 421, + 501 + ], + "score": 0.62, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 489, + 500, + 503 + ], + "score": 1.0, + "content": "on ANIMAL-10N.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "table_body", + "bbox": [ + 312, + 501, + 499, + 549 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 312, + 501, + 499, + 549 + ], + "spans": [ + { + "bbox": [ + 312, + 501, + 499, + 549 + ], + "score": 0.972, + "html": "
MethodAccuracy
Standard79.4 ± 0.14
SELFIE (Song et al., 2019)81.8 ± 0.09
PLC (ours)83.4±( 0.43
", + "type": "table", + "image_path": "99edab4c835c0240b7ed3b062d65e038ecfb04fbb7c282bbf751e0051e904ed9.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 312, + 501, + 499, + 517.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 312, + 517.0, + 499, + 533.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 312, + 533.0, + 499, + 549.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 35.0 + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 195, + 582 + ], + "lines": [ + { + "bbox": [ + 104, + 567, + 198, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 198, + 586 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "We propose a novel family of feature-dependent label noise that is much more general than the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 605, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 507, + 618 + ], + "score": 1.0, + "content": "traditional i.i.d. noise pattern. Building upon this noise assumption, we propose the first data-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "recalibrating method that is theoretically guaranteed to converge to a well-behaved classifier. On the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 627, + 507, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 507, + 639 + ], + "score": 1.0, + "content": "synthetic datasets, we show that our method outperforms various baselines under different feature-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "dependent noise patterns subject to our assumption. Also, we test our method on different real-world", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "noisy datasets and observe superior performances over existing approaches. The proposed noise", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 659, + 372, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 372, + 673 + ], + "score": 1.0, + "content": "family offers a new theoretical setting for the study of label noise.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 594, + 507, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "Acknowledgement. The authors acknowledge support from US National Science Foundation (NSF)", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "awards CRII-1755791, CCF-1910873, CCF-1855760. This effort was partially supported by the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "Intelligence Advanced Research Projects Agency (IARPA) under the contract W911NF20C0038.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 708, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 507, + 724 + ], + "score": 1.0, + "content": "The content of this paper does not necessarily reflect the position or the policy of the Government,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 296, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 296, + 733 + ], + "score": 1.0, + "content": "and no official endorsement should be inferred.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 676, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "score": 1.0, + "content": "Ehsan Amid, Manfred K Warmuth, and Sriram Srinivasan. Two-temperature logistic regression based", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 365, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 365, + 124 + ], + "score": 1.0, + "content": "on the tsallis divergence. In AISTATS, pp. 2388–2396, 2019a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 130, + 504, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "Ehsan Amid, Manfred K. K Warmuth, Rohan Anil, and Tomer Koren. Robust bi-tempered logistic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 142, + 417, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 417, + 154 + ], + "score": 1.0, + "content": "loss based on bregman divergences. In NeurIPS, pp. 15013–15022, 2019b.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 106, + 159, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 506, + 175 + ], + "score": 1.0, + "content": "Veit Andreas, Alldrin Neil, Chechik Gal, Krasin Ivan, Gupta Abhinav, and J. Belongie Serge. Learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 172, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 480, + 184 + ], + "score": 1.0, + "content": "from noisy large-scale datasets with minimal supervision. In CVPR, pp. 6575–6583, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 190, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "Eric Arazo, Diego Ortego, Paul Albert, Noel E O’Connor, and Kevin McGuinness. Unsupervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 348, + 215 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 348, + 215 + ], + "score": 1.0, + "content": "label noise modeling and loss correction. In ICML, 2019.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 109, + 221, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 235 + ], + "score": 1.0, + "content": "Han Bo, Yao Quanming, Yu Xingrui, Niu Gang, Xu Miao, Hu Weihua, W. Tsang Ivor, and Sugiyama", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 116, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "Masashi. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 243, + 245, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 245, + 257 + ], + "score": 1.0, + "content": "NeurIPS, pp. 8536–8546, 2018.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 504, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 277 + ], + "score": 1.0, + "content": "Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative compo-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 274, + 347, + 286 + ], + "spans": [ + { + "bbox": [ + 116, + 274, + 347, + 286 + ], + "score": 1.0, + "content": "nents with random forests. In ECCV, pp. 446–461, 2014.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 503, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Nontawat Charoenphakdee, Jongyeong Lee, and Masashi Sugiyama. On symmetric losses for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 304, + 363, + 317 + ], + "spans": [ + { + "bbox": [ + 116, + 304, + 363, + 317 + ], + "score": 1.0, + "content": "learning from corrupted labels. In ICML, pp. 961–970, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "Chao Chen, Xiuyan Ni, Qinxun Bai, and Yusu Wang. A topological regularizer for classifiers via", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 334, + 344, + 347 + ], + "spans": [ + { + "bbox": [ + 115, + 334, + 344, + 347 + ], + "score": 1.0, + "content": "persistent homology. In AISTATS, pp. 2573–2582, 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 504, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond class-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "conditional assumption: A primary attempt to combat instance-dependent label noise. In AAAI,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 375, + 143, + 388 + ], + "spans": [ + { + "bbox": [ + 114, + 375, + 143, + 388 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "score": 1.0, + "content": "Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 406, + 416, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 416, + 419 + ], + "score": 1.0, + "content": "instance- and label-dependent label noise. In ICML, pp. 1789–1799, 2020.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 503, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 439 + ], + "score": 1.0, + "content": "Hendrycks Dan, Lee Kimin, and Mazeika Mantas. Using pre-training can improve model robustness", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 437, + 313, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 437, + 313, + 448 + ], + "score": 1.0, + "content": "and uncertainty. In ICML, pp. 2712–2721, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 504, + 479 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "Jun Du and Zhihua Cai. Modelling class noise with symmetric and asymmetric distributions. In", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 466, + 232, + 479 + ], + "spans": [ + { + "bbox": [ + 114, + 466, + 232, + 479 + ], + "score": 1.0, + "content": "AAAI, pp. 2589–2595, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 500 + ], + "score": 1.0, + "content": "Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 497, + 308, + 509 + ], + "spans": [ + { + "bbox": [ + 116, + 497, + 308, + 509 + ], + "score": 1.0, + "content": "labels. arXiv preprint arXiv:1607.07526, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 105, + 516, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "Aritra Ghosh, Naresh Manwani, and P.S. Sastry. Making risk minimization tolerant to label noise.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 527, + 252, + 539 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 252, + 539 + ], + "score": 1.0, + "content": "Neurocomput, 160:93–107, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 546, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "score": 1.0, + "content": "Aritra Ghosh, Himanshu Kumar, and PS Sastry. Robust loss functions under label noise for deep", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 558, + 249, + 569 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 249, + 569 + ], + "score": 1.0, + "content": "neural networks. In AAAI, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 104, + 576, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "Wei Hu, Zhiyuan Li, and Dingli Yu. Simple and effective regularization methods for training on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 588, + 382, + 600 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 382, + 600 + ], + "score": 1.0, + "content": "noisily labeled data with generalization guarantee. In ICLR, 2020.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 108, + 607, + 506, + 641 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 506, + 620 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 618, + 508, + 631 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 508, + 631 + ], + "score": 1.0, + "content": "URL https://www.cs.toronto.edu/˜kriz/learning-features-2009-TR.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 630, + 141, + 642 + ], + "spans": [ + { + "bbox": [ + 114, + 630, + 141, + 642 + ], + "score": 1.0, + "content": "pdf.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 648, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "Kuang-Huei Lee, Xiaodong He, Lei Zhang, and Linjun Yang. Cleannet: Transfer learning for scalable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 659, + 414, + 673 + ], + "spans": [ + { + "bbox": [ + 116, + 659, + 414, + 673 + ], + "score": 1.0, + "content": "image classifier training with label noise. In CVPR, pp. 5447–5456, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "score": 1.0, + "content": "Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 689, + 301, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 689, + 301, + 703 + ], + "score": 1.0, + "content": "labeled data. In CVPR, pp. 5051–5059, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 720, + 363, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 363, + 733 + ], + "score": 1.0, + "content": "noisy labels with distillation. In ICCV, pp. 1928–1936, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "score": 1.0, + "content": "Ehsan Amid, Manfred K Warmuth, and Sriram Srinivasan. Two-temperature logistic regression based", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 365, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 365, + 124 + ], + "score": 1.0, + "content": "on the tsallis divergence. In AISTATS, pp. 2388–2396, 2019a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 99, + 505, + 124 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 130, + 504, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "Ehsan Amid, Manfred K. K Warmuth, Rohan Anil, and Tomer Koren. Robust bi-tempered logistic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 142, + 417, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 417, + 154 + ], + "score": 1.0, + "content": "loss based on bregman divergences. In NeurIPS, pp. 15013–15022, 2019b.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 129, + 505, + 154 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 106, + 159, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 506, + 175 + ], + "score": 1.0, + "content": "Veit Andreas, Alldrin Neil, Chechik Gal, Krasin Ivan, Gupta Abhinav, and J. Belongie Serge. Learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 172, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 480, + 184 + ], + "score": 1.0, + "content": "from noisy large-scale datasets with minimal supervision. In CVPR, pp. 6575–6583, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 159, + 506, + 184 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 190, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "Eric Arazo, Diego Ortego, Paul Albert, Noel E O’Connor, and Kevin McGuinness. Unsupervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 348, + 215 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 348, + 215 + ], + "score": 1.0, + "content": "label noise modeling and loss correction. In ICML, 2019.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 191, + 505, + 215 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 221, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 235 + ], + "score": 1.0, + "content": "Han Bo, Yao Quanming, Yu Xingrui, Niu Gang, Xu Miao, Hu Weihua, W. Tsang Ivor, and Sugiyama", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 116, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "Masashi. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 243, + 245, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 245, + 257 + ], + "score": 1.0, + "content": "NeurIPS, pp. 8536–8546, 2018.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 220, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 504, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 277 + ], + "score": 1.0, + "content": "Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative compo-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 274, + 347, + 286 + ], + "spans": [ + { + "bbox": [ + 116, + 274, + 347, + 286 + ], + "score": 1.0, + "content": "nents with random forests. In ECCV, pp. 446–461, 2014.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 261, + 505, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 503, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Nontawat Charoenphakdee, Jongyeong Lee, and Masashi Sugiyama. On symmetric losses for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 304, + 363, + 317 + ], + "spans": [ + { + "bbox": [ + 116, + 304, + 363, + 317 + ], + "score": 1.0, + "content": "learning from corrupted labels. In ICML, pp. 961–970, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 293, + 505, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "Chao Chen, Xiuyan Ni, Qinxun Bai, and Yusu Wang. A topological regularizer for classifiers via", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 334, + 344, + 347 + ], + "spans": [ + { + "bbox": [ + 115, + 334, + 344, + 347 + ], + "score": 1.0, + "content": "persistent homology. In AISTATS, pp. 2573–2582, 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 323, + 505, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 504, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, and Pheng-Ann Heng. Beyond class-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "conditional assumption: A primary attempt to combat instance-dependent label noise. In AAAI,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 375, + 143, + 388 + ], + "spans": [ + { + "bbox": [ + 114, + 375, + 143, + 388 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 353, + 506, + 388 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "score": 1.0, + "content": "Jiacheng Cheng, Tongliang Liu, Kotagiri Ramamohanarao, and Dacheng Tao. Learning with bounded", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 406, + 416, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 416, + 419 + ], + "score": 1.0, + "content": "instance- and label-dependent label noise. In ICML, pp. 1789–1799, 2020.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 394, + 505, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 503, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 439 + ], + "score": 1.0, + "content": "Hendrycks Dan, Lee Kimin, and Mazeika Mantas. Using pre-training can improve model robustness", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 437, + 313, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 437, + 313, + 448 + ], + "score": 1.0, + "content": "and uncertainty. In ICML, pp. 2712–2721, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 424, + 505, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 504, + 479 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "Jun Du and Zhihua Cai. Modelling class noise with symmetric and asymmetric distributions. In", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 466, + 232, + 479 + ], + "spans": [ + { + "bbox": [ + 114, + 466, + 232, + 479 + ], + "score": 1.0, + "content": "AAAI, pp. 2589–2595, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 455, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 500 + ], + "score": 1.0, + "content": "Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 497, + 308, + 509 + ], + "spans": [ + { + "bbox": [ + 116, + 497, + 308, + 509 + ], + "score": 1.0, + "content": "labels. arXiv preprint arXiv:1607.07526, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 484, + 505, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 516, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "Aritra Ghosh, Naresh Manwani, and P.S. Sastry. Making risk minimization tolerant to label noise.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 527, + 252, + 539 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 252, + 539 + ], + "score": 1.0, + "content": "Neurocomput, 160:93–107, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 516, + 506, + 539 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 546, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 560 + ], + "score": 1.0, + "content": "Aritra Ghosh, Himanshu Kumar, and PS Sastry. Robust loss functions under label noise for deep", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 558, + 249, + 569 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 249, + 569 + ], + "score": 1.0, + "content": "neural networks. In AAAI, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 545, + 505, + 569 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 576, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "Wei Hu, Zhiyuan Li, and Dingli Yu. Simple and effective regularization methods for training on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 588, + 382, + 600 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 382, + 600 + ], + "score": 1.0, + "content": "noisily labeled data with generalization guarantee. In ICLR, 2020.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 576, + 505, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 607, + 506, + 641 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 506, + 620 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 618, + 508, + 631 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 508, + 631 + ], + "score": 1.0, + "content": "URL https://www.cs.toronto.edu/˜kriz/learning-features-2009-TR.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 630, + 141, + 642 + ], + "spans": [ + { + "bbox": [ + 114, + 630, + 141, + 642 + ], + "score": 1.0, + "content": "pdf.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 608, + 508, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 648, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "Kuang-Huei Lee, Xiaodong He, Lei Zhang, and Linjun Yang. Cleannet: Transfer learning for scalable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 659, + 414, + 673 + ], + "spans": [ + { + "bbox": [ + 116, + 659, + 414, + 673 + ], + "score": 1.0, + "content": "image classifier training with label noise. In CVPR, pp. 5447–5456, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 648, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 693 + ], + "score": 1.0, + "content": "Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 689, + 301, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 689, + 301, + 703 + ], + "score": 1.0, + "content": "labeled data. In CVPR, pp. 5051–5059, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 677, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 720, + 363, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 363, + 733 + ], + "score": 1.0, + "content": "noisy labels with distillation. In ICCV, pp. 1928–1936, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 107, + 709, + 504, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Jiang Lu, Zhou Zhengyuan, Leung Thomas, Li Li-Jia, and Fei-Fei Li. Mentornet: Learning data-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "score": 1.0, + "content": "driven curriculum for very deep neural networks on corrupted labels. In ICML, pp. 2309–2318,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 123, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 505, + 147 + ], + "score": 1.0, + "content": "N. R. Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 235, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 235, + 158 + ], + "score": 1.0, + "content": "ICML, pp. 3361–3370, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 503, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Naresh Manwani and PS Sastry. Noise tolerance under risk minimization. IEEE transactions on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 176, + 266, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 176, + 266, + 187 + ], + "score": 1.0, + "content": "cybernetics, 43(3):1146–1151, 2013.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 502, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 504, + 207 + ], + "score": 1.0, + "content": "Aditya Krishna Menon, Brendan Van Rooyen, and Nagarajan Natarajan. Learning from binary labels", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 426, + 218 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 426, + 218 + ], + "score": 1.0, + "content": "with instance-dependent noise. Machine Learning, 107(8):1561–1595, 2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 502, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 504, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 504, + 237 + ], + "score": 1.0, + "content": "Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 236, + 309, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 236, + 309, + 248 + ], + "score": 1.0, + "content": "noisy labels. In NeurIPS, pp. 1196–1204, 2013.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 254, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 253, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 268 + ], + "score": 1.0, + "content": "Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 266, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 507, + 279 + ], + "score": 1.0, + "content": "deep neural networks robust to label noise: A loss correction approach. In CVPR, pp. 2233–2241,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 276, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 114, + 276, + 142, + 289 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 108, + 295, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "score": 1.0, + "content": "Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "Rabinovich. Training deep neural networks on noisy labels with bootstrapping. In ICLR Workshop,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 317, + 143, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 143, + 330 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "score": 1.0, + "content": "Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 347, + 334, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 347, + 334, + 360 + ], + "score": 1.0, + "content": "robust deep learning. In ICML, pp. 4331–4340, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 380 + ], + "score": 1.0, + "content": "Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 379, + 305, + 390 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 305, + 390 + ], + "score": 1.0, + "content": "minimization. In ICML, pp. 5739–5748, 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 397, + 504, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "score": 1.0, + "content": "Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pp. 1917–1928,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 418, + 143, + 432 + ], + "spans": [ + { + "bbox": [ + 114, + 418, + 143, + 432 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 105, + 438, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 104, + 435, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 506, + 453 + ], + "score": 1.0, + "content": "Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 449, + 285, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 449, + 285, + 461 + ], + "score": 1.0, + "content": "learning. In ICML, pp. 5907–5915, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 105, + 468, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 479, + 370, + 492 + ], + "spans": [ + { + "bbox": [ + 116, + 479, + 370, + 492 + ], + "score": 1.0, + "content": "for learning with noisy labels. In CVPR, pp. 5552–5560, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 105, + 498, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 509, + 475, + 522 + ], + "spans": [ + { + "bbox": [ + 116, + 509, + 475, + 522 + ], + "score": 1.0, + "content": "Combating label noise in deep learning using abstention. In ICML, pp. 6234–6243, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 539, + 256, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 539, + 256, + 551 + ], + "score": 1.0, + "content": "In NeurIPS, pp. 5596–5605, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 105, + 558, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "Brendan Van Rooyen, Aditya Menon, and Robert C Williamson. Learning with symmetric label", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 570, + 408, + 582 + ], + "spans": [ + { + "bbox": [ + 115, + 570, + 408, + 582 + ], + "score": 1.0, + "content": "noise: The importance of being unhinged. In NeurIPS, pp. 10–18, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 105, + 588, + 505, + 612 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 600, + 427, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 600, + 427, + 612 + ], + "score": 1.0, + "content": "Iterative learning with open-set noisy labels. In CVPR, pp. 8688–8696, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 618, + 504, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 633 + ], + "score": 1.0, + "content": "Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 630, + 420, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 630, + 420, + 642 + ], + "score": 1.0, + "content": "entropy for robust learning with noisy labels. In CVPR, pp. 322–330, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 105, + 649, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 660, + 331, + 673 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 331, + 673 + ], + "score": 1.0, + "content": "filter for learning with label noise. In NeurIPS, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "score": 1.0, + "content": "Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 690, + 397, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 397, + 703 + ], + "score": 1.0, + "content": "labeled data for image classification. In CVPR, pp. 2691–2699, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L dmi: A novel information-theoretic loss", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 466, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 466, + 732 + ], + "score": 1.0, + "content": "function for training deep nets robust to label noise. In NeurIPS, pp. 6222–6233, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Jiang Lu, Zhou Zhengyuan, Leung Thomas, Li Li-Jia, and Fei-Fei Li. Mentornet: Learning data-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "score": 1.0, + "content": "driven curriculum for very deep neural networks on corrupted labels. In ICML, pp. 2309–2318,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 143, + 117 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 507, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 123, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 505, + 147 + ], + "score": 1.0, + "content": "N. R. Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 235, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 235, + 158 + ], + "score": 1.0, + "content": "ICML, pp. 3361–3370, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 123, + 505, + 158 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 503, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Naresh Manwani and PS Sastry. Noise tolerance under risk minimization. IEEE transactions on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 176, + 266, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 176, + 266, + 187 + ], + "score": 1.0, + "content": "cybernetics, 43(3):1146–1151, 2013.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 164, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 502, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 504, + 207 + ], + "score": 1.0, + "content": "Aditya Krishna Menon, Brendan Van Rooyen, and Nagarajan Natarajan. Learning from binary labels", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 426, + 218 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 426, + 218 + ], + "score": 1.0, + "content": "with instance-dependent noise. Machine Learning, 107(8):1561–1595, 2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 194, + 504, + 218 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 502, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 504, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 504, + 237 + ], + "score": 1.0, + "content": "Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 236, + 309, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 236, + 309, + 248 + ], + "score": 1.0, + "content": "noisy labels. In NeurIPS, pp. 1196–1204, 2013.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 225, + 504, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 254, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 253, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 268 + ], + "score": 1.0, + "content": "Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 266, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 507, + 279 + ], + "score": 1.0, + "content": "deep neural networks robust to label noise: A loss correction approach. In CVPR, pp. 2233–2241,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 276, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 114, + 276, + 142, + 289 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 253, + 507, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 295, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "score": 1.0, + "content": "Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "Rabinovich. Training deep neural networks on noisy labels with bootstrapping. In ICLR Workshop,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 317, + 143, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 143, + 330 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 296, + 506, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 350 + ], + "score": 1.0, + "content": "Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 347, + 334, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 347, + 334, + 360 + ], + "score": 1.0, + "content": "robust deep learning. In ICML, pp. 4331–4340, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 337, + 505, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 380 + ], + "score": 1.0, + "content": "Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 379, + 305, + 390 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 305, + 390 + ], + "score": 1.0, + "content": "minimization. In ICML, pp. 5739–5748, 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 366, + 506, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 397, + 504, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 411 + ], + "score": 1.0, + "content": "Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, pp. 1917–1928,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 418, + 143, + 432 + ], + "spans": [ + { + "bbox": [ + 114, + 418, + 143, + 432 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 396, + 506, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 438, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 104, + 435, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 506, + 453 + ], + "score": 1.0, + "content": "Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 449, + 285, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 449, + 285, + 461 + ], + "score": 1.0, + "content": "learning. In ICML, pp. 5907–5915, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 435, + 506, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 468, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 479, + 370, + 492 + ], + "spans": [ + { + "bbox": [ + 116, + 479, + 370, + 492 + ], + "score": 1.0, + "content": "for learning with noisy labels. In CVPR, pp. 5552–5560, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 469, + 505, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 498, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 509, + 475, + 522 + ], + "spans": [ + { + "bbox": [ + 116, + 509, + 475, + 522 + ], + "score": 1.0, + "content": "Combating label noise in deep learning using abstention. In ICML, pp. 6234–6243, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 498, + 506, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 539, + 256, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 539, + 256, + 551 + ], + "score": 1.0, + "content": "In NeurIPS, pp. 5596–5605, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 529, + 506, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 558, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "Brendan Van Rooyen, Aditya Menon, and Robert C Williamson. Learning with symmetric label", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 570, + 408, + 582 + ], + "spans": [ + { + "bbox": [ + 115, + 570, + 408, + 582 + ], + "score": 1.0, + "content": "noise: The importance of being unhinged. In NeurIPS, pp. 10–18, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 559, + 505, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 588, + 505, + 612 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 600, + 427, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 600, + 427, + 612 + ], + "score": 1.0, + "content": "Iterative learning with open-set noisy labels. In CVPR, pp. 8688–8696, 2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 588, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 618, + 504, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 633 + ], + "score": 1.0, + "content": "Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 630, + 420, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 630, + 420, + 642 + ], + "score": 1.0, + "content": "entropy for robust learning with noisy labels. In CVPR, pp. 322–330, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 617, + 505, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 649, + 504, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 660, + 331, + 673 + ], + "spans": [ + { + "bbox": [ + 116, + 660, + 331, + 673 + ], + "score": 1.0, + "content": "filter for learning with label noise. In NeurIPS, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 649, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 679, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 693 + ], + "score": 1.0, + "content": "Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 690, + 397, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 397, + 703 + ], + "score": 1.0, + "content": "labeled data for image classification. In CVPR, pp. 2691–2699, 2015.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 678, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L dmi: A novel information-theoretic loss", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 466, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 466, + 732 + ], + "score": 1.0, + "content": "function for training deep nets robust to label noise. In NeurIPS, pp. 6222–6233, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Yan Yan, Romer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from ´", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 456, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 456, + 106 + ], + "score": 1.0, + "content": "multiple annotators with varying expertise. Machine learning, 95(3):291–327, 2014.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 108, + 111, + 504, + 134 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Kun Yi and Jianxin Wu. Probabilistic end-to-end noise correction for learning with noisy labels. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 236, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 236, + 135 + ], + "score": 1.0, + "content": "CVPR, pp. 7017–7025, 2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 109, + 140, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 107, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 107, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 150, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 150, + 482, + 164 + ], + "score": 1.0, + "content": "disagreement help generalization against label corruption? In ICML, pp. 7164–7173, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 169, + 504, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 183 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 180, + 376, + 192 + ], + "spans": [ + { + "bbox": [ + 115, + 180, + 376, + 192 + ], + "score": 1.0, + "content": "deep learning requires rethinking generalization. In ICLR, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 108, + 198, + 503, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 208, + 329, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 329, + 221 + ], + "score": 1.0, + "content": "with noisy labels. In NeurIPS, pp. 8778–8788, 2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 108, + 227, + 506, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 504, + 240 + ], + "score": 1.0, + "content": "Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 239, + 450, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 239, + 450, + 250 + ], + "score": 1.0, + "content": "Chen. Error-bounded correction of noisy labels. In ICML, pp. 11447–11457, 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 269, + 183, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 185, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 185, + 284 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 506, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 365, + 306 + ], + "score": 1.0, + "content": "Lemma 1 (One round purity improvement). Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 365, + 295, + 371, + 304 + ], + "score": 0.53, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 294, + 477, + 306 + ], + "score": 1.0, + "content": "is satisfied, and assume an", + "type": "text" + }, + { + "bbox": [ + 477, + 295, + 484, + 306 + ], + "score": 0.81, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 294, + 505, + 306 + ], + "score": 1.0, + "content": "such", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 303, + 504, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 198, + 318 + ], + "score": 1.0, + "content": "that there exists a pure", + "type": "text" + }, + { + "bbox": [ + 198, + 305, + 230, + 317 + ], + "score": 0.92, + "content": "( e , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 303, + 286, + 318 + ], + "score": 1.0, + "content": "-level set with", + "type": "text" + }, + { + "bbox": [ + 286, + 304, + 333, + 318 + ], + "score": 0.91, + "content": "\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 303, + 352, + 318 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 353, + 305, + 504, + 317 + ], + "score": 0.85, + "content": "\\tilde { \\eta } _ { n e w } ( { \\pmb x } ) = y _ { f ( { \\pmb x } ) } i f | f ( { \\pmb x } ) - 1 / 2 | \\ge e", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 315, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 124, + 335 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 125, + 317, + 228, + 329 + ], + "score": 0.89, + "content": "\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 315, + 284, + 335 + ], + "score": 1.0, + "content": ", and assume", + "type": "text" + }, + { + "bbox": [ + 284, + 318, + 484, + 332 + ], + "score": 0.85, + "content": "f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 315, + 507, + 335 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 451, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 193, + 347 + ], + "score": 0.61, + "content": "e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 330, + 198, + 350 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 334, + 228, + 347 + ], + "score": 0.51, + "content": "L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 330, + 273, + 350 + ], + "score": 1.0, + "content": "is pure for", + "type": "text" + }, + { + "bbox": [ + 274, + 334, + 299, + 347 + ], + "score": 0.86, + "content": "f _ { n e w } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 330, + 327, + 350 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 327, + 334, + 446, + 348 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 330, + 451, + 350 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 506, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 355, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 244, + 370 + ], + "score": 1.0, + "content": "Proof: We analyze the case where", + "type": "text" + }, + { + "bbox": [ + 244, + 355, + 284, + 369 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\eta ( \\pmb { x } ) > \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 355, + 507, + 370 + ], + "score": 1.0, + "content": ". The analysis on the other side can be derived similarly.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 466, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 275, + 383 + ], + "score": 1.0, + "content": "Due to the fact that there exists a level set", + "type": "text" + }, + { + "bbox": [ + 275, + 369, + 304, + 381 + ], + "score": 0.93, + "content": "L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 367, + 336, + 383 + ], + "score": 1.0, + "content": "pure to", + "type": "text" + }, + { + "bbox": [ + 337, + 369, + 343, + 380 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 367, + 382, + 383 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 383, + 368, + 462, + 382 + ], + "score": 0.92, + "content": "\\begin{array} { r } { e \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | , \\forall \\pmb { x } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 367, + 466, + 383 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 384, + 445, + 413 + ], + "lines": [ + { + "bbox": [ + 164, + 384, + 445, + 413 + ], + "spans": [ + { + "bbox": [ + 164, + 384, + 445, + 413 + ], + "score": 0.93, + "content": "\\mathbb { E } _ { ( z , \\tilde { y } ) \\sim ( D , \\tilde { \\eta } _ { n e w } ( z ) ) } \\left[ \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } ( z ) \\bigg | \\bigg | \\eta ( z ) - \\frac { 1 } { 2 } \\bigg | \\ge \\bigg | \\eta ( x ) - \\frac { 1 } { 2 } \\bigg | \\right] = 0 .", + "type": "interline_equation", + "image_path": "6608e21c0baa7bea63213fe867bd9128ba2b4062fd173a75a8ba24988cdcd8bc.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 164, + 384, + 445, + 393.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 164, + 393.6666666666667, + 445, + 403.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 164, + 403.33333333333337, + 445, + 413.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 415, + 502, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 504, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 164, + 430 + ], + "score": 1.0, + "content": "Now consider", + "type": "text" + }, + { + "bbox": [ + 164, + 419, + 172, + 426 + ], + "score": 0.76, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 414, + 200, + 430 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 201, + 415, + 281, + 429 + ], + "score": 0.92, + "content": "e - \\gamma \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } |", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 414, + 373, + 430 + ], + "score": 1.0, + "content": ". Since the distribution", + "type": "text" + }, + { + "bbox": [ + 373, + 417, + 383, + 426 + ], + "score": 0.84, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 414, + 393, + 430 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 394, + 416, + 424, + 429 + ], + "score": 0.93, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 414, + 504, + 430 + ], + "score": 1.0, + "content": "-bounded, we have:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 432, + 503, + 637 + ], + "lines": [ + { + "bbox": [ + 111, + 432, + 503, + 637 + ], + "spans": [ + { + "bbox": [ + 111, + 432, + 503, + 637 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\mathbb { E } _ { ( z \\sim \\hat { y } ) \\sim ( D , \\hat { y } \\sim \\alpha ( z ) ) } [ \\hat { 1 } ( s _ { z \\sim \\hat { y } } , \\hat { x } ) ( z ) ] \\lVert ( z ( z ) - \\frac { 1 } { 2 } \\vert \\geq | y | ( x ) - \\frac { 1 } { 2 } \\vert ] } \\\\ & { = \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\lVert ( z ) - \\frac { 1 } { 2 } [ \\hat { y } \\langle x \\rangle - \\frac { 1 } { 2 } ] } \\\\ & { \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } } { \\lVert \\mathcal { Z } \\rVert ( z ) ^ { 2 } - \\frac { 1 } { 2 } } [ \\frac { \\mathbb { E } _ { \\hat { z } } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) } { 1 - \\frac { 1 } { 2 } } ] } \\\\ & { \\leq \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon - \\epsilon - \\mathbb { E } _ { z } \\cdot \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { \\hat { z } } } { 2 } + \\epsilon } { \\mathcal { Z } \\cdot { \\mathcal { Z } } ( z ) ^ { 2 } - \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } - \\frac { \\epsilon } { 2 } } [ \\mathcal { Z } \\cdot \\mathcal { Y } ( z ) \\leq \\frac { \\frac { 1 } { 2 } } { 2 } - \\epsilon + \\frac { \\mathbb { E } _ { z } } { 1 } ] } } \\\\ & \\quad + \\frac \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\mathcal { Z } ( z ) } \\epsilon - \\frac \\mathbb { E } _ { z } \\cdot \\{ { y } \\cdot \\mathcal { Z } \\} \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } \\cdot \\mathcal { Y } ( z \\end{array}", + "type": "interline_equation", + "image_path": "ded246c9b5f3fae10ae340fe6ab0cd4549373b09adcfa8afabf54518949d53ae.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 111, + 432, + 503, + 500.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 500.3333333333333, + 503, + 568.6666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 111, + 568.6666666666666, + 503, + 637.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 506, + 693 + ], + "lines": [ + { + "bbox": [ + 104, + 643, + 507, + 659 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 116, + 659 + ], + "score": 1.0, + "content": "If", + "type": "text" + }, + { + "bbox": [ + 117, + 644, + 235, + 659 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) \\le \\gamma \\le \\frac { 2 \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 643, + 363, + 659 + ], + "score": 1.0, + "content": ", the impurity in super level set", + "type": "text" + }, + { + "bbox": [ + 363, + 644, + 406, + 659 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } + e - \\gamma } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 643, + 449, + 659 + ], + "score": 1.0, + "content": "is at most", + "type": "text" + }, + { + "bbox": [ + 450, + 645, + 460, + 659 + ], + "score": 0.87, + "content": "\\frac { 2 \\epsilon } { \\alpha }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 643, + 507, + 659 + ], + "score": 1.0, + "content": ". The level", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 657, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 121, + 671 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 121, + 659, + 145, + 670 + ], + "score": 0.9, + "content": "( \\alpha , \\varepsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 657, + 271, + 671 + ], + "score": 1.0, + "content": "-consistency condition implies", + "type": "text" + }, + { + "bbox": [ + 271, + 658, + 367, + 670 + ], + "score": 0.91, + "content": "| f _ { n e w } ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\le 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 657, + 383, + 671 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 384, + 660, + 391, + 669 + ], + "score": 0.78, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 657, + 408, + 671 + ], + "score": 1.0, + "content": "s.t.", + "type": "text" + }, + { + "bbox": [ + 409, + 658, + 491, + 671 + ], + "score": 0.89, + "content": "\\begin{array} { r } { e - \\gamma \\le | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 657, + 507, + 671 + ], + "score": 1.0, + "content": ". If", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 668, + 507, + 682 + ], + "spans": [ + { + "bbox": [ + 107, + 670, + 135, + 680 + ], + "score": 0.82, + "content": "e \\geq 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 668, + 138, + 682 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 139, + 669, + 174, + 681 + ], + "score": 0.88, + "content": "{ \\dot { f } } _ { n e w } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 668, + 283, + 682 + ], + "score": 1.0, + "content": "will give the same label as", + "type": "text" + }, + { + "bbox": [ + 283, + 670, + 308, + 681 + ], + "score": 0.86, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 668, + 346, + 682 + ], + "score": 1.0, + "content": "and thus", + "type": "text" + }, + { + "bbox": [ + 347, + 670, + 386, + 681 + ], + "score": 0.89, + "content": "( e - \\gamma , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 668, + 496, + 682 + ], + "score": 1.0, + "content": "-level set becomes pure for", + "type": "text" + }, + { + "bbox": [ + 496, + 670, + 503, + 681 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 668, + 507, + 682 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 210, + 694 + ], + "score": 1.0, + "content": "Meanwhile, the choice of", + "type": "text" + }, + { + "bbox": [ + 210, + 682, + 218, + 692 + ], + "score": 0.81, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 679, + 269, + 694 + ], + "score": 1.0, + "content": "ensures that", + "type": "text" + }, + { + "bbox": [ + 269, + 680, + 388, + 694 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\ell \\alpha } ) ( \\frac { 1 } { 2 } - e ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 679, + 392, + 694 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 494, + 680, + 506, + 691 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 694, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 101, + 694, + 505, + 727 + ], + "spans": [ + { + "bbox": [ + 101, + 694, + 152, + 727 + ], + "score": 1.0, + "content": "Lemma 2 is pure for", + "type": "text" + }, + { + "bbox": [ + 163, + 694, + 195, + 727 + ], + "score": 1.0, + "content": "arm-up r. Given", + "type": "text" + }, + { + "bbox": [ + 237, + 694, + 342, + 727 + ], + "score": 1.0, + "content": "uppose for a given function, after running Algorithm", + "type": "text" + }, + { + "bbox": [ + 344, + 695, + 354, + 706 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 694, + 365, + 727 + ], + "score": 1.0, + "content": "thor", + "type": "text" + }, + { + "bbox": [ + 444, + 694, + 478, + 707 + ], + "score": 0.92, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 694, + 505, + 727 + ], + "score": 1.0, + "content": "which, there", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 152, + 705, + 446, + 721 + ], + "spans": [ + { + "bbox": [ + 152, + 707, + 162, + 718 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 707, + 236, + 719 + ], + "score": 0.91, + "content": "T _ { 0 } < 1 / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 708, + 348, + 717 + ], + "score": 0.44, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 705, + 446, + 721 + ], + "score": 0.92, + "content": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 719, + 303, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 173, + 733 + ], + "score": 1.0, + "content": "exists a level set", + "type": "text" + }, + { + "bbox": [ + 173, + 719, + 227, + 734 + ], + "score": 0.93, + "content": "L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 720, + 290, + 733 + ], + "score": 1.0, + "content": "that is pure for", + "type": "text" + }, + { + "bbox": [ + 291, + 721, + 300, + 732 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 720, + 303, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Yan Yan, Romer Rosales, Glenn Fung, Ramanathan Subramanian, and Jennifer Dy. Learning from ´", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 456, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 456, + 106 + ], + "score": 1.0, + "content": "multiple annotators with varying expertise. Machine learning, 95(3):291–327, 2014.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 81, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 111, + 504, + 134 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Kun Yi and Jianxin Wu. Probabilistic end-to-end noise correction for learning with noisy labels. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 236, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 236, + 135 + ], + "score": 1.0, + "content": "CVPR, pp. 7017–7025, 2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 111, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 140, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 107, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 107, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 150, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 150, + 482, + 164 + ], + "score": 1.0, + "content": "disagreement help generalization against label corruption? In ICML, pp. 7164–7173, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 107, + 139, + 505, + 164 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 169, + 504, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 183 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 180, + 376, + 192 + ], + "spans": [ + { + "bbox": [ + 115, + 180, + 376, + 192 + ], + "score": 1.0, + "content": "deep learning requires rethinking generalization. In ICLR, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 168, + 505, + 192 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 198, + 503, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 208, + 329, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 329, + 221 + ], + "score": 1.0, + "content": "with noisy labels. In NeurIPS, pp. 8778–8788, 2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 198, + 505, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 227, + 506, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 504, + 240 + ], + "score": 1.0, + "content": "Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 239, + 450, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 239, + 450, + 250 + ], + "score": 1.0, + "content": "Chen. Error-bounded correction of noisy labels. In ICML, pp. 11447–11457, 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 228, + 504, + 250 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 269, + 183, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 185, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 185, + 284 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 506, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 365, + 306 + ], + "score": 1.0, + "content": "Lemma 1 (One round purity improvement). Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 365, + 295, + 371, + 304 + ], + "score": 0.53, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 294, + 477, + 306 + ], + "score": 1.0, + "content": "is satisfied, and assume an", + "type": "text" + }, + { + "bbox": [ + 477, + 295, + 484, + 306 + ], + "score": 0.81, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 294, + 505, + 306 + ], + "score": 1.0, + "content": "such", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 303, + 504, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 198, + 318 + ], + "score": 1.0, + "content": "that there exists a pure", + "type": "text" + }, + { + "bbox": [ + 198, + 305, + 230, + 317 + ], + "score": 0.92, + "content": "( e , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 303, + 286, + 318 + ], + "score": 1.0, + "content": "-level set with", + "type": "text" + }, + { + "bbox": [ + 286, + 304, + 333, + 318 + ], + "score": 0.91, + "content": "\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 303, + 352, + 318 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 353, + 305, + 504, + 317 + ], + "score": 0.85, + "content": "\\tilde { \\eta } _ { n e w } ( { \\pmb x } ) = y _ { f ( { \\pmb x } ) } i f | f ( { \\pmb x } ) - 1 / 2 | \\ge e", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 315, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 315, + 124, + 335 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 125, + 317, + 228, + 329 + ], + "score": 0.89, + "content": "\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 315, + 284, + 335 + ], + "score": 1.0, + "content": ", and assume", + "type": "text" + }, + { + "bbox": [ + 284, + 318, + 484, + 332 + ], + "score": 0.85, + "content": "f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 315, + 507, + 335 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 451, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 193, + 347 + ], + "score": 0.61, + "content": "e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 330, + 198, + 350 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 334, + 228, + 347 + ], + "score": 0.51, + "content": "L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 330, + 273, + 350 + ], + "score": 1.0, + "content": "is pure for", + "type": "text" + }, + { + "bbox": [ + 274, + 334, + 299, + 347 + ], + "score": 0.86, + "content": "f _ { n e w } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 330, + 327, + 350 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + }, + { + "bbox": [ + 327, + 334, + 446, + 348 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 330, + 451, + 350 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 294, + 507, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 506, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 355, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 244, + 370 + ], + "score": 1.0, + "content": "Proof: We analyze the case where", + "type": "text" + }, + { + "bbox": [ + 244, + 355, + 284, + 369 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\eta ( \\pmb { x } ) > \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 355, + 507, + 370 + ], + "score": 1.0, + "content": ". The analysis on the other side can be derived similarly.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 466, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 275, + 383 + ], + "score": 1.0, + "content": "Due to the fact that there exists a level set", + "type": "text" + }, + { + "bbox": [ + 275, + 369, + 304, + 381 + ], + "score": 0.93, + "content": "L ( e , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 367, + 336, + 383 + ], + "score": 1.0, + "content": "pure to", + "type": "text" + }, + { + "bbox": [ + 337, + 369, + 343, + 380 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 367, + 382, + 383 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 383, + 368, + 462, + 382 + ], + "score": 0.92, + "content": "\\begin{array} { r } { e \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | , \\forall \\pmb { x } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 367, + 466, + 383 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 355, + 507, + 383 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 384, + 445, + 413 + ], + "lines": [ + { + "bbox": [ + 164, + 384, + 445, + 413 + ], + "spans": [ + { + "bbox": [ + 164, + 384, + 445, + 413 + ], + "score": 0.93, + "content": "\\mathbb { E } _ { ( z , \\tilde { y } ) \\sim ( D , \\tilde { \\eta } _ { n e w } ( z ) ) } \\left[ \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } ( z ) \\bigg | \\bigg | \\eta ( z ) - \\frac { 1 } { 2 } \\bigg | \\ge \\bigg | \\eta ( x ) - \\frac { 1 } { 2 } \\bigg | \\right] = 0 .", + "type": "interline_equation", + "image_path": "6608e21c0baa7bea63213fe867bd9128ba2b4062fd173a75a8ba24988cdcd8bc.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 164, + 384, + 445, + 393.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 164, + 393.6666666666667, + 445, + 403.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 164, + 403.33333333333337, + 445, + 413.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 415, + 502, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 504, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 164, + 430 + ], + "score": 1.0, + "content": "Now consider", + "type": "text" + }, + { + "bbox": [ + 164, + 419, + 172, + 426 + ], + "score": 0.76, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 414, + 200, + 430 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 201, + 415, + 281, + 429 + ], + "score": 0.92, + "content": "e - \\gamma \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } |", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 414, + 373, + 430 + ], + "score": 1.0, + "content": ". Since the distribution", + "type": "text" + }, + { + "bbox": [ + 373, + 417, + 383, + 426 + ], + "score": 0.84, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 414, + 393, + 430 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 394, + 416, + 424, + 429 + ], + "score": 0.93, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 414, + 504, + 430 + ], + "score": 1.0, + "content": "-bounded, we have:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 414, + 504, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 432, + 503, + 637 + ], + "lines": [ + { + "bbox": [ + 111, + 432, + 503, + 637 + ], + "spans": [ + { + "bbox": [ + 111, + 432, + 503, + 637 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\mathbb { E } _ { ( z \\sim \\hat { y } ) \\sim ( D , \\hat { y } \\sim \\alpha ( z ) ) } [ \\hat { 1 } ( s _ { z \\sim \\hat { y } } , \\hat { x } ) ( z ) ] \\lVert ( z ( z ) - \\frac { 1 } { 2 } \\vert \\geq | y | ( x ) - \\frac { 1 } { 2 } \\vert ] } \\\\ & { = \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\lVert ( z ) - \\frac { 1 } { 2 } [ \\hat { y } \\langle x \\rangle - \\frac { 1 } { 2 } ] } \\\\ & { \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } } { \\lVert \\mathcal { Z } \\rVert ( z ) ^ { 2 } - \\frac { 1 } { 2 } } [ \\frac { \\mathbb { E } _ { \\hat { z } } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) } { 1 - \\frac { 1 } { 2 } } ] } \\\\ & { \\leq \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon - \\epsilon - \\mathbb { E } _ { z } \\cdot \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { \\hat { z } } } { 2 } + \\epsilon } { \\mathcal { Z } \\cdot { \\mathcal { Z } } ( z ) ^ { 2 } - \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } - \\frac { \\epsilon } { 2 } } [ \\mathcal { Z } \\cdot \\mathcal { Y } ( z ) \\leq \\frac { \\frac { 1 } { 2 } } { 2 } - \\epsilon + \\frac { \\mathbb { E } _ { z } } { 1 } ] } } \\\\ & \\quad + \\frac \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\mathcal { Z } ( z ) } \\epsilon - \\frac \\mathbb { E } _ { z } \\cdot \\{ { y } \\cdot \\mathcal { Z } \\} \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } \\cdot \\mathcal { Y } ( z \\end{array}", + "type": "interline_equation", + "image_path": "ded246c9b5f3fae10ae340fe6ab0cd4549373b09adcfa8afabf54518949d53ae.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 111, + 432, + 503, + 500.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 500.3333333333333, + 503, + 568.6666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 111, + 568.6666666666666, + 503, + 637.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 506, + 693 + ], + "lines": [ + { + "bbox": [ + 104, + 643, + 507, + 659 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 116, + 659 + ], + "score": 1.0, + "content": "If", + "type": "text" + }, + { + "bbox": [ + 117, + 644, + 235, + 659 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) \\le \\gamma \\le \\frac { 2 \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 643, + 363, + 659 + ], + "score": 1.0, + "content": ", the impurity in super level set", + "type": "text" + }, + { + "bbox": [ + 363, + 644, + 406, + 659 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } + e - \\gamma } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 643, + 449, + 659 + ], + "score": 1.0, + "content": "is at most", + "type": "text" + }, + { + "bbox": [ + 450, + 645, + 460, + 659 + ], + "score": 0.87, + "content": "\\frac { 2 \\epsilon } { \\alpha }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 643, + 507, + 659 + ], + "score": 1.0, + "content": ". The level", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 657, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 121, + 671 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 121, + 659, + 145, + 670 + ], + "score": 0.9, + "content": "( \\alpha , \\varepsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 657, + 271, + 671 + ], + "score": 1.0, + "content": "-consistency condition implies", + "type": "text" + }, + { + "bbox": [ + 271, + 658, + 367, + 670 + ], + "score": 0.91, + "content": "| f _ { n e w } ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\le 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 657, + 383, + 671 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 384, + 660, + 391, + 669 + ], + "score": 0.78, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 657, + 408, + 671 + ], + "score": 1.0, + "content": "s.t.", + "type": "text" + }, + { + "bbox": [ + 409, + 658, + 491, + 671 + ], + "score": 0.89, + "content": "\\begin{array} { r } { e - \\gamma \\le | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 657, + 507, + 671 + ], + "score": 1.0, + "content": ". If", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 668, + 507, + 682 + ], + "spans": [ + { + "bbox": [ + 107, + 670, + 135, + 680 + ], + "score": 0.82, + "content": "e \\geq 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 668, + 138, + 682 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 139, + 669, + 174, + 681 + ], + "score": 0.88, + "content": "{ \\dot { f } } _ { n e w } ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 668, + 283, + 682 + ], + "score": 1.0, + "content": "will give the same label as", + "type": "text" + }, + { + "bbox": [ + 283, + 670, + 308, + 681 + ], + "score": 0.86, + "content": "\\eta ^ { * } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 668, + 346, + 682 + ], + "score": 1.0, + "content": "and thus", + "type": "text" + }, + { + "bbox": [ + 347, + 670, + 386, + 681 + ], + "score": 0.89, + "content": "( e - \\gamma , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 668, + 496, + 682 + ], + "score": 1.0, + "content": "-level set becomes pure for", + "type": "text" + }, + { + "bbox": [ + 496, + 670, + 503, + 681 + ], + "score": 0.83, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 668, + 507, + 682 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 679, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 210, + 694 + ], + "score": 1.0, + "content": "Meanwhile, the choice of", + "type": "text" + }, + { + "bbox": [ + 210, + 682, + 218, + 692 + ], + "score": 0.81, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 679, + 269, + 694 + ], + "score": 1.0, + "content": "ensures that", + "type": "text" + }, + { + "bbox": [ + 269, + 680, + 388, + 694 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\ell \\alpha } ) ( \\frac { 1 } { 2 } - e ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 679, + 392, + 694 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 494, + 680, + 506, + 691 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 643, + 507, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 694, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 101, + 694, + 505, + 727 + ], + "spans": [ + { + "bbox": [ + 101, + 694, + 152, + 727 + ], + "score": 1.0, + "content": "Lemma 2 is pure for", + "type": "text" + }, + { + "bbox": [ + 163, + 694, + 195, + 727 + ], + "score": 1.0, + "content": "arm-up r. Given", + "type": "text" + }, + { + "bbox": [ + 237, + 694, + 342, + 727 + ], + "score": 1.0, + "content": "uppose for a given function, after running Algorithm", + "type": "text" + }, + { + "bbox": [ + 344, + 695, + 354, + 706 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 694, + 365, + 727 + ], + "score": 1.0, + "content": "thor", + "type": "text" + }, + { + "bbox": [ + 444, + 694, + 478, + 707 + ], + "score": 0.92, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 694, + 505, + 727 + ], + "score": 1.0, + "content": "which, there", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 152, + 705, + 446, + 721 + ], + "spans": [ + { + "bbox": [ + 152, + 707, + 162, + 718 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 707, + 236, + 719 + ], + "score": 0.91, + "content": "T _ { 0 } < 1 / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 708, + 348, + 717 + ], + "score": 0.44, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 705, + 446, + 721 + ], + "score": 0.92, + "content": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 719, + 303, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 173, + 733 + ], + "score": 1.0, + "content": "exists a level set", + "type": "text" + }, + { + "bbox": [ + 173, + 719, + 227, + 734 + ], + "score": 0.93, + "content": "L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 720, + 290, + 733 + ], + "score": 1.0, + "content": "that is pure for", + "type": "text" + }, + { + "bbox": [ + 291, + 721, + 300, + 732 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 720, + 303, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 101, + 694, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 120 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Proof: The proof follows from the fact that each round of label flipping improves the purity by a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 90, + 508, + 110 + ], + "spans": [ + { + "bbox": [ + 104, + 90, + 145, + 110 + ], + "score": 1.0, + "content": "factor of", + "type": "text" + }, + { + "bbox": [ + 145, + 93, + 182, + 106 + ], + "score": 0.88, + "content": "\\textstyle { \\left( { 1 + { \\frac { \\varepsilon } { \\ell \\alpha } } } \\right) }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 90, + 274, + 110 + ], + "score": 1.0, + "content": ". To obtain an at least", + "type": "text" + }, + { + "bbox": [ + 275, + 94, + 286, + 105 + ], + "score": 0.88, + "content": "T _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 90, + 508, + 110 + ], + "score": 1.0, + "content": "pure region, it suffices to repeat the flipping step for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 186, + 122 + ], + "score": 1.0, + "content": "m ≥ `αε log( T01−2e )", + "type": "text" + }, + { + "bbox": [ + 187, + 106, + 221, + 118 + ], + "score": 1.0, + "content": "rounds.", + "type": "text" + }, + { + "bbox": [ + 493, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "□", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 242, + 135 + ], + "score": 1.0, + "content": "Lemma 3. Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 242, + 124, + 248, + 133 + ], + "score": 0.53, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 123, + 400, + 135 + ], + "score": 1.0, + "content": "is satisfied, and for a given function", + "type": "text" + }, + { + "bbox": [ + 401, + 124, + 411, + 135 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 123, + 506, + 135 + ], + "score": 1.0, + "content": "there exists a level set", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 133, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 107, + 134, + 141, + 146 + ], + "score": 0.91, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 133, + 183, + 168 + ], + "score": 1.0, + "content": "h is p, (2)", + "type": "text" + }, + { + "bbox": [ + 214, + 135, + 225, + 145 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 133, + 286, + 168 + ], + "score": 1.0, + "content": "ns A, (3)", + "type": "text" + }, + { + "bbox": [ + 321, + 135, + 327, + 144 + ], + "score": 0.27, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 135, + 394, + 146 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 133, + 405, + 168 + ], + "score": 1.0, + "content": "an(4)", + "type": "text" + }, + { + "bbox": [ + 471, + 133, + 491, + 168 + ], + "score": 1.0, + "content": "ons:and", + "type": "text" + }, + { + "bbox": [ + 491, + 134, + 505, + 145 + ], + "score": 0.34, + "content": "( l )", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 145, + 471, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 163, + 159 + ], + "score": 0.9, + "content": "T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 145, + 267, + 160 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 146, + 385, + 160 + ], + "score": 0.89, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 145, + 471, + 159 + ], + "score": 0.93, + "content": "T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 157, + 382, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 162, + 174 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 157, + 220, + 176 + ], + "score": 1.0, + "content": ", then we have", + "type": "text" + }, + { + "bbox": [ + 220, + 161, + 376, + 174 + ], + "score": 0.91, + "content": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 157, + 382, + 176 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 438, + 194 + ], + "score": 1.0, + "content": "Proof: The proof can be done by combining Lemma 1 and Lemma 2. In the first", + "type": "text" + }, + { + "bbox": [ + 438, + 183, + 448, + 192 + ], + "score": 0.74, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "iterations, by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 264, + 206 + ], + "score": 1.0, + "content": "Lemma 2, we can guarantee a level set", + "type": "text" + }, + { + "bbox": [ + 264, + 192, + 311, + 205 + ], + "score": 0.93, + "content": "\\bigl ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 192, + 343, + 206 + ], + "score": 1.0, + "content": "pure to", + "type": "text" + }, + { + "bbox": [ + 343, + 193, + 350, + 204 + ], + "score": 0.82, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 192, + 506, + 206 + ], + "score": 1.0, + "content": ". In the rest of the iterations we ensure", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 204, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 155, + 219 + ], + "score": 1.0, + "content": "the level set", + "type": "text" + }, + { + "bbox": [ + 155, + 204, + 235, + 218 + ], + "score": 0.95, + "content": "\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 204, + 317, + 219 + ], + "score": 1.0, + "content": "is pure. We increase", + "type": "text" + }, + { + "bbox": [ + 317, + 206, + 326, + 215 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 204, + 425, + 219 + ], + "score": 1.0, + "content": "by a reasonable factor of", + "type": "text" + }, + { + "bbox": [ + 425, + 205, + 433, + 217 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 204, + 506, + 219 + ], + "score": 1.0, + "content": "to avoid incurring", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 212, + 502, + 237 + ], + "spans": [ + { + "bbox": [ + 102, + 212, + 447, + 237 + ], + "score": 1.0, + "content": "too many corrupted labels while ensuring enough progress in label purification, i.e.,", + "type": "text" + }, + { + "bbox": [ + 447, + 217, + 502, + 231 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 201, + 244 + ], + "score": 1.0, + "content": "such that in the level set", + "type": "text" + }, + { + "bbox": [ + 201, + 230, + 281, + 244 + ], + "score": 0.92, + "content": "\\begin{array} { r } { | \\eta ( { \\pmb x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 230, + 316, + 244 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 317, + 231, + 396, + 243 + ], + "score": 0.94, + "content": "| f ( { \\pmb x } ) - \\eta ( { \\pmb x } ) | \\leq 3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 230, + 506, + 244 + ], + "score": 1.0, + "content": ". This condition ensures the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 221, + 258 + ], + "score": 1.0, + "content": "correctness of flipping when", + "type": "text" + }, + { + "bbox": [ + 221, + 244, + 271, + 257 + ], + "score": 0.92, + "content": "T \\leq { \\frac { 1 } { 2 } } - 3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 243, + 421, + 258 + ], + "score": 1.0, + "content": ". The purity cannot be improved once", + "type": "text" + }, + { + "bbox": [ + 421, + 243, + 504, + 256 + ], + "score": 0.92, + "content": "\\begin{array} { r } { T \\ge \\frac { 1 } { 2 } - 3 \\varepsilon = T _ { e n d } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 238, + 271 + ], + "score": 1.0, + "content": "since there is no guarantee that", + "type": "text" + }, + { + "bbox": [ + 239, + 257, + 259, + 268 + ], + "score": 0.91, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 255, + 366, + 271 + ], + "score": 1.0, + "content": "has consistent label with", + "type": "text" + }, + { + "bbox": [ + 367, + 257, + 387, + 268 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 255, + 415, + 271 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 415, + 257, + 486, + 269 + ], + "score": 0.91, + "content": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | < 3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 255, + 506, + 271 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 267, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 107, + 267, + 188, + 280 + ], + "score": 0.91, + "content": "| \\eta ( { \\pmb x } ) - f ( { \\pmb x } ) | \\leq \\bar { 3 } \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 267, + 206, + 281 + ], + "score": 1.0, + "content": ". By", + "type": "text" + }, + { + "bbox": [ + 207, + 268, + 237, + 280 + ], + "score": 0.92, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 267, + 334, + 281 + ], + "score": 1.0, + "content": "-bounded assumption on", + "type": "text" + }, + { + "bbox": [ + 335, + 268, + 344, + 277 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 267, + 422, + 281 + ], + "score": 1.0, + "content": ", its mass of impure", + "type": "text" + }, + { + "bbox": [ + 423, + 269, + 433, + 277 + ], + "score": 0.73, + "content": "3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 267, + 506, + 281 + ], + "score": 1.0, + "content": "level set region is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 279, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 138, + 290 + ], + "score": 1.0, + "content": "at most", + "type": "text" + }, + { + "bbox": [ + 138, + 280, + 157, + 289 + ], + "score": 0.87, + "content": "3 c ^ { * } \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 279, + 162, + 290 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 494, + 279, + 506, + 290 + ], + "score": 1.0, + "content": "□", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 506, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 241, + 307 + ], + "score": 1.0, + "content": "Theorem 1. Under assumption", + "type": "text" + }, + { + "bbox": [ + 241, + 294, + 247, + 303 + ], + "score": 0.4, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 291, + 311, + 307 + ], + "score": 1.0, + "content": ", for any noise", + "type": "text" + }, + { + "bbox": [ + 311, + 296, + 318, + 303 + ], + "score": 0.71, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 291, + 438, + 307 + ], + "score": 1.0, + "content": "which is PMD with margin", + "type": "text" + }, + { + "bbox": [ + 438, + 294, + 447, + 304 + ], + "score": 0.8, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 291, + 480, + 307 + ], + "score": 1.0, + "content": ", define", + "type": "text" + }, + { + "bbox": [ + 480, + 294, + 505, + 305 + ], + "score": 0.87, + "content": "e _ { 0 } =", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 299, + 510, + 322 + ], + "spans": [ + { + "bbox": [ + 107, + 304, + 168, + 318 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 299, + 308, + 322 + ], + "score": 1.0, + "content": ". Then for the output of Algorithm", + "type": "text" + }, + { + "bbox": [ + 309, + 306, + 315, + 315 + ], + "score": 0.54, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 299, + 336, + 322 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 336, + 305, + 343, + 316 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 299, + 510, + 322 + ], + "score": 1.0, + "content": "as above and with the following initial-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 102, + 311, + 509, + 340 + ], + "spans": [ + { + "bbox": [ + 102, + 311, + 145, + 340 + ], + "score": 1.0, + "content": "izations:", + "type": "text" + }, + { + "bbox": [ + 145, + 317, + 211, + 332 + ], + "score": 0.7, + "content": "( I ) T _ { 0 } < { \\textstyle \\frac { 1 } { 2 } } - e _ { 0 } ", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 311, + 227, + 340 + ], + "score": 1.0, + "content": ", (2)", + "type": "text" + }, + { + "bbox": [ + 228, + 317, + 309, + 333 + ], + "score": 0.68, + "content": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 311, + 326, + 340 + ], + "score": 1.0, + "content": ", (3)", + "type": "text" + }, + { + "bbox": [ + 327, + 317, + 412, + 333 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log \\bigl ( \\frac { T _ { 0 } } { 3 \\varepsilon } \\bigr ) , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 311, + 430, + 340 + ], + "score": 1.0, + "content": ", (4)", + "type": "text" + }, + { + "bbox": [ + 430, + 318, + 473, + 330 + ], + "score": 0.89, + "content": "T _ { e n d } \\leq 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 311, + 509, + 340 + ], + "score": 1.0, + "content": "and (5)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 331, + 205, + 346 + ], + "spans": [ + { + "bbox": [ + 107, + 332, + 162, + 346 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 331, + 205, + 346 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 343, + 385, + 358 + ], + "lines": [ + { + "bbox": [ + 225, + 343, + 385, + 358 + ], + "spans": [ + { + "bbox": [ + 225, + 343, + 385, + 358 + ], + "score": 0.91, + "content": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon .", + "type": "interline_equation", + "image_path": "993df6e3b032810b65fd96058668ff86b288a8997eae0b4bd1ec86fd1e575228.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 225, + 343, + 385, + 358 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 364, + 504, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 407, + 377 + ], + "score": 1.0, + "content": "Proof: The proof is based on Lemma 3 plus a verification of the existence of", + "type": "text" + }, + { + "bbox": [ + 407, + 365, + 417, + 376 + ], + "score": 0.89, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "for which there exists", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 375, + 230, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 134, + 388 + ], + "score": 1.0, + "content": "a pure", + "type": "text" + }, + { + "bbox": [ + 134, + 376, + 171, + 388 + ], + "score": 0.93, + "content": "( e _ { 0 } , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 375, + 230, + 388 + ], + "score": 1.0, + "content": "-level set. Let:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 392, + 375, + 420 + ], + "lines": [ + { + "bbox": [ + 234, + 392, + 375, + 420 + ], + "spans": [ + { + "bbox": [ + 234, + 392, + 375, + 420 + ], + "score": 0.94, + "content": "\\tau ( { \\pmb x } ) = \\left\\{ \\begin{array} { l l } { \\tau _ { 1 0 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) \\geq 1 / 2 } \\\\ { \\tau _ { 0 1 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) < 1 / 2 } \\end{array} \\right. .", + "type": "interline_equation", + "image_path": "fcca459cbff85ce5ba955e929ea9650afaabd969859502d5eaa924a5bc2a5788.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 234, + 392, + 375, + 420 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 168, + 440 + ], + "score": 1.0, + "content": "In the level set", + "type": "text" + }, + { + "bbox": [ + 168, + 425, + 450, + 439 + ], + "score": 0.85, + "content": "\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } , \\mathbb { P } _ { z } [ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) | | \\eta ( z ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } ] \\leq \\frac { 1 } { 2 } - e _ { 0 } + \\tau ( z ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 424, + 506, + 440 + ], + "score": 1.0, + "content": ". By level set", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 437, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 129, + 451 + ], + "score": 0.91, + "content": "( \\alpha , \\varepsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 437, + 266, + 452 + ], + "score": 1.0, + "content": "-consistency, it suffices to satisfy", + "type": "text" + }, + { + "bbox": [ + 266, + 439, + 367, + 452 + ], + "score": 0.92, + "content": "\\alpha ( \\textstyle { \\frac { 1 } { 2 } } - e _ { o } + \\tau ) + \\varepsilon \\le e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 437, + 427, + 452 + ], + "score": 1.0, + "content": "to ensure that", + "type": "text" + }, + { + "bbox": [ + 428, + 439, + 448, + 451 + ], + "score": 0.92, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 437, + 506, + 452 + ], + "score": 1.0, + "content": "has the same", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 451, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 171, + 466 + ], + "score": 1.0, + "content": "prediction with", + "type": "text" + }, + { + "bbox": [ + 171, + 452, + 192, + 464 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 452, + 218, + 466 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 219, + 451, + 286, + 465 + ], + "score": 0.93, + "content": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq \\overline { { e } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 452, + 506, + 466 + ], + "score": 1.0, + "content": ". By polynomial level set diminishing noise, we have", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 461, + 504, + 482 + ], + "spans": [ + { + "bbox": [ + 107, + 464, + 168, + 478 + ], + "score": 0.93, + "content": "\\tau ( \\pmb { x } ) \\leq { \\frac { 1 } { 2 } } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 461, + 178, + 482 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 178, + 466, + 210, + 477 + ], + "score": 0.9, + "content": "e _ { 0 } > t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 461, + 301, + 482 + ], + "score": 1.0, + "content": ", and thus by choosing", + "type": "text" + }, + { + "bbox": [ + 301, + 464, + 386, + 479 + ], + "score": 0.92, + "content": "\\begin{array} { r } { e _ { 0 } = m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 461, + 493, + 482 + ], + "score": 1.0, + "content": "one can ensure that initial", + "type": "text" + }, + { + "bbox": [ + 493, + 465, + 504, + 477 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 150, + 491 + ], + "score": 1.0, + "content": "has a pure", + "type": "text" + }, + { + "bbox": [ + 150, + 478, + 191, + 490 + ], + "score": 0.92, + "content": "( e _ { 0 } , f _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 477, + 413, + 491 + ], + "score": 1.0, + "content": "-level set. The rest of the proof follows from Lemma 3.", + "type": "text" + }, + { + "bbox": [ + 493, + 478, + 506, + 489 + ], + "score": 0.997, + "content": "□", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 120 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Proof: The proof follows from the fact that each round of label flipping improves the purity by a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 90, + 508, + 110 + ], + "spans": [ + { + "bbox": [ + 104, + 90, + 145, + 110 + ], + "score": 1.0, + "content": "factor of", + "type": "text" + }, + { + "bbox": [ + 145, + 93, + 182, + 106 + ], + "score": 0.88, + "content": "\\textstyle { \\left( { 1 + { \\frac { \\varepsilon } { \\ell \\alpha } } } \\right) }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 90, + 274, + 110 + ], + "score": 1.0, + "content": ". To obtain an at least", + "type": "text" + }, + { + "bbox": [ + 275, + 94, + 286, + 105 + ], + "score": 0.88, + "content": "T _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 90, + 508, + 110 + ], + "score": 1.0, + "content": "pure region, it suffices to repeat the flipping step for", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 186, + 122 + ], + "score": 1.0, + "content": "m ≥ `αε log( T01−2e )", + "type": "text" + }, + { + "bbox": [ + 187, + 106, + 221, + 118 + ], + "score": 1.0, + "content": "rounds.", + "type": "text" + }, + { + "bbox": [ + 493, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "□", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 81, + 508, + 122 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 242, + 135 + ], + "score": 1.0, + "content": "Lemma 3. Suppose Assumption", + "type": "text" + }, + { + "bbox": [ + 242, + 124, + 248, + 133 + ], + "score": 0.53, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 123, + 400, + 135 + ], + "score": 1.0, + "content": "is satisfied, and for a given function", + "type": "text" + }, + { + "bbox": [ + 401, + 124, + 411, + 135 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 123, + 506, + 135 + ], + "score": 1.0, + "content": "there exists a level set", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 133, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 107, + 134, + 141, + 146 + ], + "score": 0.91, + "content": "L ( e _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 133, + 183, + 168 + ], + "score": 1.0, + "content": "h is p, (2)", + "type": "text" + }, + { + "bbox": [ + 214, + 135, + 225, + 145 + ], + "score": 0.86, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 133, + 286, + 168 + ], + "score": 1.0, + "content": "ns A, (3)", + "type": "text" + }, + { + "bbox": [ + 321, + 135, + 327, + 144 + ], + "score": 0.27, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 135, + 394, + 146 + ], + "score": 0.87, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 133, + 405, + 168 + ], + "score": 1.0, + "content": "an(4)", + "type": "text" + }, + { + "bbox": [ + 471, + 133, + 491, + 168 + ], + "score": 1.0, + "content": "ons:and", + "type": "text" + }, + { + "bbox": [ + 491, + 134, + 505, + 145 + ], + "score": 0.34, + "content": "( l )", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 145, + 471, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 163, + 159 + ], + "score": 0.9, + "content": "T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 145, + 267, + 160 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 146, + 385, + 160 + ], + "score": 0.89, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 145, + 471, + 159 + ], + "score": 0.93, + "content": "T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 157, + 382, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 162, + 174 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 157, + 220, + 176 + ], + "score": 1.0, + "content": ", then we have", + "type": "text" + }, + { + "bbox": [ + 220, + 161, + 376, + 174 + ], + "score": 0.91, + "content": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 157, + 382, + 176 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 123, + 506, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 438, + 194 + ], + "score": 1.0, + "content": "Proof: The proof can be done by combining Lemma 1 and Lemma 2. In the first", + "type": "text" + }, + { + "bbox": [ + 438, + 183, + 448, + 192 + ], + "score": 0.74, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "iterations, by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 264, + 206 + ], + "score": 1.0, + "content": "Lemma 2, we can guarantee a level set", + "type": "text" + }, + { + "bbox": [ + 264, + 192, + 311, + 205 + ], + "score": 0.93, + "content": "\\bigl ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 192, + 343, + 206 + ], + "score": 1.0, + "content": "pure to", + "type": "text" + }, + { + "bbox": [ + 343, + 193, + 350, + 204 + ], + "score": 0.82, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 192, + 506, + 206 + ], + "score": 1.0, + "content": ". In the rest of the iterations we ensure", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 204, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 155, + 219 + ], + "score": 1.0, + "content": "the level set", + "type": "text" + }, + { + "bbox": [ + 155, + 204, + 235, + 218 + ], + "score": 0.95, + "content": "\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 204, + 317, + 219 + ], + "score": 1.0, + "content": "is pure. We increase", + "type": "text" + }, + { + "bbox": [ + 317, + 206, + 326, + 215 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 204, + 425, + 219 + ], + "score": 1.0, + "content": "by a reasonable factor of", + "type": "text" + }, + { + "bbox": [ + 425, + 205, + 433, + 217 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 204, + 506, + 219 + ], + "score": 1.0, + "content": "to avoid incurring", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 212, + 502, + 237 + ], + "spans": [ + { + "bbox": [ + 102, + 212, + 447, + 237 + ], + "score": 1.0, + "content": "too many corrupted labels while ensuring enough progress in label purification, i.e.,", + "type": "text" + }, + { + "bbox": [ + 447, + 217, + 502, + 231 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 201, + 244 + ], + "score": 1.0, + "content": "such that in the level set", + "type": "text" + }, + { + "bbox": [ + 201, + 230, + 281, + 244 + ], + "score": 0.92, + "content": "\\begin{array} { r } { | \\eta ( { \\pmb x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 230, + 316, + 244 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 317, + 231, + 396, + 243 + ], + "score": 0.94, + "content": "| f ( { \\pmb x } ) - \\eta ( { \\pmb x } ) | \\leq 3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 230, + 506, + 244 + ], + "score": 1.0, + "content": ". This condition ensures the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 221, + 258 + ], + "score": 1.0, + "content": "correctness of flipping when", + "type": "text" + }, + { + "bbox": [ + 221, + 244, + 271, + 257 + ], + "score": 0.92, + "content": "T \\leq { \\frac { 1 } { 2 } } - 3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 243, + 421, + 258 + ], + "score": 1.0, + "content": ". The purity cannot be improved once", + "type": "text" + }, + { + "bbox": [ + 421, + 243, + 504, + 256 + ], + "score": 0.92, + "content": "\\begin{array} { r } { T \\ge \\frac { 1 } { 2 } - 3 \\varepsilon = T _ { e n d } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 238, + 271 + ], + "score": 1.0, + "content": "since there is no guarantee that", + "type": "text" + }, + { + "bbox": [ + 239, + 257, + 259, + 268 + ], + "score": 0.91, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 255, + 366, + 271 + ], + "score": 1.0, + "content": "has consistent label with", + "type": "text" + }, + { + "bbox": [ + 367, + 257, + 387, + 268 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 255, + 415, + 271 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 415, + 257, + 486, + 269 + ], + "score": 0.91, + "content": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | < 3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 255, + 506, + 271 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 267, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 107, + 267, + 188, + 280 + ], + "score": 0.91, + "content": "| \\eta ( { \\pmb x } ) - f ( { \\pmb x } ) | \\leq \\bar { 3 } \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 267, + 206, + 281 + ], + "score": 1.0, + "content": ". By", + "type": "text" + }, + { + "bbox": [ + 207, + 268, + 237, + 280 + ], + "score": 0.92, + "content": "( c _ { * } , c ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 267, + 334, + 281 + ], + "score": 1.0, + "content": "-bounded assumption on", + "type": "text" + }, + { + "bbox": [ + 335, + 268, + 344, + 277 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 267, + 422, + 281 + ], + "score": 1.0, + "content": ", its mass of impure", + "type": "text" + }, + { + "bbox": [ + 423, + 269, + 433, + 277 + ], + "score": 0.73, + "content": "3 \\varepsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 267, + 506, + 281 + ], + "score": 1.0, + "content": "level set region is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 279, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 138, + 290 + ], + "score": 1.0, + "content": "at most", + "type": "text" + }, + { + "bbox": [ + 138, + 280, + 157, + 289 + ], + "score": 0.87, + "content": "3 c ^ { * } \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 279, + 162, + 290 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 494, + 279, + 506, + 290 + ], + "score": 1.0, + "content": "□", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11, + "bbox_fs": [ + 102, + 180, + 506, + 290 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 506, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 241, + 307 + ], + "score": 1.0, + "content": "Theorem 1. Under assumption", + "type": "text" + }, + { + "bbox": [ + 241, + 294, + 247, + 303 + ], + "score": 0.4, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 291, + 311, + 307 + ], + "score": 1.0, + "content": ", for any noise", + "type": "text" + }, + { + "bbox": [ + 311, + 296, + 318, + 303 + ], + "score": 0.71, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 291, + 438, + 307 + ], + "score": 1.0, + "content": "which is PMD with margin", + "type": "text" + }, + { + "bbox": [ + 438, + 294, + 447, + 304 + ], + "score": 0.8, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 291, + 480, + 307 + ], + "score": 1.0, + "content": ", define", + "type": "text" + }, + { + "bbox": [ + 480, + 294, + 505, + 305 + ], + "score": 0.87, + "content": "e _ { 0 } =", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 299, + 510, + 322 + ], + "spans": [ + { + "bbox": [ + 107, + 304, + 168, + 318 + ], + "score": 0.91, + "content": "\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 299, + 308, + 322 + ], + "score": 1.0, + "content": ". Then for the output of Algorithm", + "type": "text" + }, + { + "bbox": [ + 309, + 306, + 315, + 315 + ], + "score": 0.54, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 299, + 336, + 322 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 336, + 305, + 343, + 316 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 299, + 510, + 322 + ], + "score": 1.0, + "content": "as above and with the following initial-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 102, + 311, + 509, + 340 + ], + "spans": [ + { + "bbox": [ + 102, + 311, + 145, + 340 + ], + "score": 1.0, + "content": "izations:", + "type": "text" + }, + { + "bbox": [ + 145, + 317, + 211, + 332 + ], + "score": 0.7, + "content": "( I ) T _ { 0 } < { \\textstyle \\frac { 1 } { 2 } } - e _ { 0 } ", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 311, + 227, + 340 + ], + "score": 1.0, + "content": ", (2)", + "type": "text" + }, + { + "bbox": [ + 228, + 317, + 309, + 333 + ], + "score": 0.68, + "content": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 311, + 326, + 340 + ], + "score": 1.0, + "content": ", (3)", + "type": "text" + }, + { + "bbox": [ + 327, + 317, + 412, + 333 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log \\bigl ( \\frac { T _ { 0 } } { 3 \\varepsilon } \\bigr ) , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 311, + 430, + 340 + ], + "score": 1.0, + "content": ", (4)", + "type": "text" + }, + { + "bbox": [ + 430, + 318, + 473, + 330 + ], + "score": 0.89, + "content": "T _ { e n d } \\leq 3 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 311, + 509, + 340 + ], + "score": 1.0, + "content": "and (5)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 331, + 205, + 346 + ], + "spans": [ + { + "bbox": [ + 107, + 332, + 162, + 346 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 331, + 205, + 346 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 102, + 291, + 510, + 346 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 343, + 385, + 358 + ], + "lines": [ + { + "bbox": [ + 225, + 343, + 385, + 358 + ], + "spans": [ + { + "bbox": [ + 225, + 343, + 385, + 358 + ], + "score": 0.91, + "content": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon .", + "type": "interline_equation", + "image_path": "993df6e3b032810b65fd96058668ff86b288a8997eae0b4bd1ec86fd1e575228.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 225, + 343, + 385, + 358 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 364, + 504, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 407, + 377 + ], + "score": 1.0, + "content": "Proof: The proof is based on Lemma 3 plus a verification of the existence of", + "type": "text" + }, + { + "bbox": [ + 407, + 365, + 417, + 376 + ], + "score": 0.89, + "content": "f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "for which there exists", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 375, + 230, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 134, + 388 + ], + "score": 1.0, + "content": "a pure", + "type": "text" + }, + { + "bbox": [ + 134, + 376, + 171, + 388 + ], + "score": 0.93, + "content": "( e _ { 0 } , f , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 375, + 230, + 388 + ], + "score": 1.0, + "content": "-level set. Let:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 363, + 505, + 388 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 392, + 375, + 420 + ], + "lines": [ + { + "bbox": [ + 234, + 392, + 375, + 420 + ], + "spans": [ + { + "bbox": [ + 234, + 392, + 375, + 420 + ], + "score": 0.94, + "content": "\\tau ( { \\pmb x } ) = \\left\\{ \\begin{array} { l l } { \\tau _ { 1 0 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) \\geq 1 / 2 } \\\\ { \\tau _ { 0 1 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) < 1 / 2 } \\end{array} \\right. .", + "type": "interline_equation", + "image_path": "fcca459cbff85ce5ba955e929ea9650afaabd969859502d5eaa924a5bc2a5788.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 234, + 392, + 375, + 420 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 168, + 440 + ], + "score": 1.0, + "content": "In the level set", + "type": "text" + }, + { + "bbox": [ + 168, + 425, + 450, + 439 + ], + "score": 0.85, + "content": "\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } , \\mathbb { P } _ { z } [ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) | | \\eta ( z ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } ] \\leq \\frac { 1 } { 2 } - e _ { 0 } + \\tau ( z ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 424, + 506, + 440 + ], + "score": 1.0, + "content": ". By level set", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 437, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 129, + 451 + ], + "score": 0.91, + "content": "( \\alpha , \\varepsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 437, + 266, + 452 + ], + "score": 1.0, + "content": "-consistency, it suffices to satisfy", + "type": "text" + }, + { + "bbox": [ + 266, + 439, + 367, + 452 + ], + "score": 0.92, + "content": "\\alpha ( \\textstyle { \\frac { 1 } { 2 } } - e _ { o } + \\tau ) + \\varepsilon \\le e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 437, + 427, + 452 + ], + "score": 1.0, + "content": "to ensure that", + "type": "text" + }, + { + "bbox": [ + 428, + 439, + 448, + 451 + ], + "score": 0.92, + "content": "f ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 437, + 506, + 452 + ], + "score": 1.0, + "content": "has the same", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 451, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 171, + 466 + ], + "score": 1.0, + "content": "prediction with", + "type": "text" + }, + { + "bbox": [ + 171, + 452, + 192, + 464 + ], + "score": 0.92, + "content": "\\eta ( { \\pmb x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 452, + 218, + 466 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 219, + 451, + 286, + 465 + ], + "score": 0.93, + "content": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq \\overline { { e } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 452, + 506, + 466 + ], + "score": 1.0, + "content": ". By polynomial level set diminishing noise, we have", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 461, + 504, + 482 + ], + "spans": [ + { + "bbox": [ + 107, + 464, + 168, + 478 + ], + "score": 0.93, + "content": "\\tau ( \\pmb { x } ) \\leq { \\frac { 1 } { 2 } } - e _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 461, + 178, + 482 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 178, + 466, + 210, + 477 + ], + "score": 0.9, + "content": "e _ { 0 } > t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 461, + 301, + 482 + ], + "score": 1.0, + "content": ", and thus by choosing", + "type": "text" + }, + { + "bbox": [ + 301, + 464, + 386, + 479 + ], + "score": 0.92, + "content": "\\begin{array} { r } { e _ { 0 } = m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 461, + 493, + 482 + ], + "score": 1.0, + "content": "one can ensure that initial", + "type": "text" + }, + { + "bbox": [ + 493, + 465, + 504, + 477 + ], + "score": 0.85, + "content": "f _ { 0 }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 150, + 491 + ], + "score": 1.0, + "content": "has a pure", + "type": "text" + }, + { + "bbox": [ + 150, + 478, + 191, + 490 + ], + "score": 0.92, + "content": "( e _ { 0 } , f _ { 0 } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 477, + 413, + 491 + ], + "score": 1.0, + "content": "-level set. The rest of the proof follows from Lemma 3.", + "type": "text" + }, + { + "bbox": [ + 493, + 478, + 506, + 489 + ], + "score": 0.997, + "content": "□", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 424, + 506, + 491 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_model.json b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_model.json new file mode 100644 index 0000000000000000000000000000000000000000..757976dc9a56321baefbd566d4a3e18a5dc4a9dc --- /dev/null +++ b/parse/train/ZPa2SyGcbwh/ZPa2SyGcbwh_model.json @@ -0,0 +1,24564 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 655, + 1304, + 655, + 1304, + 1022, + 398, + 1022 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1301, + 1404, + 1301, + 1404, + 1546, + 298, + 1546 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1761, + 1402, + 1761, + 1402, + 1976, + 298, + 1976 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1562, + 1403, + 1562, + 1403, + 1745, + 299, + 1745 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1162, + 1402, + 1162, + 1402, + 1284, + 300, + 1284 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 299, + 223, + 1407, + 223, + 1407, + 323, + 299, + 323 + ], + "score": 0.956 + }, + { + "category_id": 2, + "poly": [ + 334, + 2006, + 545, + 2006, + 545, + 2033, + 334, + 2033 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 302, + 1090, + 572, + 1090, + 572, + 1125, + 302, + 1125 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 773, + 583, + 926, + 583, + 926, + 616, + 773, + 616 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 320, + 411, + 1305, + 411, + 1305, + 502, + 320, + 502 + ], + "score": 0.749 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 856, + 2088, + 856, + 2112, + 840, + 2112 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 315, + 373, + 1324, + 373, + 1324, + 409, + 315, + 409 + ], + "score": 0.717 + }, + { + "category_id": 13, + "poly": [ + 854, + 374, + 923, + 374, + 923, + 407, + 854, + 407 + ], + "score": 0.83, + "latex": "\\mathbf { W _ { u } } ^ { 1 * }" + }, + { + "category_id": 13, + "poly": [ + 548, + 409, + 762, + 409, + 762, + 441, + 548, + 441 + ], + "score": 0.67, + "latex": "\\{ \\mathtt { y z 4 2 2 } , \\mathtt { p w 2 4 1 } \\}" + }, + { + "category_id": 13, + "poly": [ + 385, + 374, + 487, + 374, + 487, + 408, + 385, + 408 + ], + "score": 0.27, + "latex": "\\mathbf { Z h a n g ^ { 1 * } }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 222.0, + 1412.0, + 222.0, + 1412.0, + 273.0, + 293.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 285.0, + 857.0, + 285.0, + 857.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2000.0, + 550.0, + 2000.0, + 550.0, + 2037.0, + 330.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1087.0, + 579.0, + 1087.0, + 579.0, + 1134.0, + 293.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 582.0, + 932.0, + 582.0, + 932.0, + 619.0, + 769.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 657.0, + 1304.0, + 657.0, + 1304.0, + 689.0, + 395.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 685.0, + 1308.0, + 685.0, + 1308.0, + 723.0, + 393.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 715.0, + 1306.0, + 715.0, + 1306.0, + 754.0, + 392.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 747.0, + 1305.0, + 747.0, + 1305.0, + 782.0, + 394.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 777.0, + 1309.0, + 777.0, + 1309.0, + 815.0, + 393.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 808.0, + 1308.0, + 808.0, + 1308.0, + 844.0, + 394.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 841.0, + 1306.0, + 841.0, + 1306.0, + 873.0, + 393.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 870.0, + 1306.0, + 870.0, + 1306.0, + 905.0, + 393.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 898.0, + 1305.0, + 898.0, + 1305.0, + 938.0, + 392.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 931.0, + 1305.0, + 931.0, + 1305.0, + 964.0, + 395.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 960.0, + 1309.0, + 960.0, + 1309.0, + 999.0, + 392.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 991.0, + 1310.0, + 991.0, + 1310.0, + 1026.0, + 393.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1301.0, + 1406.0, + 1301.0, + 1406.0, + 1335.0, + 296.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 1406.0, + 1334.0, + 1406.0, + 1365.0, + 293.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1362.0, + 1407.0, + 1362.0, + 1407.0, + 1396.0, + 293.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1394.0, + 1408.0, + 1394.0, + 1408.0, + 1427.0, + 294.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1408.0, + 1423.0, + 1408.0, + 1457.0, + 294.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1451.0, + 1406.0, + 1451.0, + 1406.0, + 1491.0, + 292.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1485.0, + 1408.0, + 1485.0, + 1408.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1516.0, + 1031.0, + 1516.0, + 1031.0, + 1549.0, + 296.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1800.0, + 293.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1827.0, + 294.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1404.0, + 1820.0, + 1404.0, + 1857.0, + 293.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1407.0, + 1849.0, + 1407.0, + 1891.0, + 292.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 1407.0, + 1880.0, + 1407.0, + 1921.0, + 292.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1914.0, + 1406.0, + 1914.0, + 1406.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1946.0, + 1404.0, + 1946.0, + 1404.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1597.0, + 294.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1592.0, + 1407.0, + 1592.0, + 1407.0, + 1628.0, + 293.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1622.0, + 1405.0, + 1622.0, + 1405.0, + 1659.0, + 292.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1654.0, + 1405.0, + 1654.0, + 1405.0, + 1689.0, + 294.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1717.0, + 297.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1198.0, + 1713.0, + 1198.0, + 1751.0, + 294.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1164.0, + 1404.0, + 1164.0, + 1404.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1195.0, + 1404.0, + 1195.0, + 1404.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1221.0, + 1406.0, + 1221.0, + 1406.0, + 1261.0, + 293.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1253.0, + 1222.0, + 1253.0, + 1222.0, + 1289.0, + 294.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 404.0, + 547.0, + 404.0, + 547.0, + 445.0, + 313.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 404.0, + 1014.0, + 404.0, + 1014.0, + 445.0, + 763.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 431.0, + 1308.0, + 431.0, + 1308.0, + 480.0, + 311.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 467.0, + 1007.0, + 467.0, + 1007.0, + 509.0, + 314.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 371.0, + 384.0, + 371.0, + 384.0, + 414.0, + 312.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 371.0, + 853.0, + 371.0, + 853.0, + 414.0, + 488.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 371.0, + 1330.0, + 371.0, + 1330.0, + 414.0, + 924.0, + 414.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 522, + 298, + 522 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1044, + 1405, + 1044, + 1405, + 1229, + 298, + 1229 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 537, + 1405, + 537, + 1405, + 782, + 298, + 782 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 338, + 1287, + 1362, + 1287, + 1362, + 1818, + 338, + 1818 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 905, + 1403, + 905, + 1403, + 1028, + 300, + 1028 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 798, + 1404, + 798, + 1404, + 889, + 300, + 889 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 295, + 1844, + 1407, + 1844, + 1407, + 2028, + 295, + 2028 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1401, + 229, + 1401, + 293, + 297, + 293 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.723 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.109 + }, + { + "category_id": 13, + "poly": [ + 1311, + 1845, + 1397, + 1845, + 1397, + 1877, + 1311, + 1877 + ], + "score": 0.89, + "latex": "( \\eta ^ { * } ( \\pmb { x } ) )" + }, + { + "category_id": 13, + "poly": [ + 678, + 1171, + 697, + 1171, + 697, + 1198, + 678, + 1198 + ], + "score": 0.68, + "latex": "\\eta" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1300.0, + 455.0, + 1300.0, + 455.0, + 1318.0, + 402.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1300.0, + 719.0, + 1300.0, + 719.0, + 1318.0, + 666.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1300.0, + 992.0, + 1300.0, + 992.0, + 1318.0, + 924.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1297.0, + 1241.0, + 1297.0, + 1241.0, + 1317.0, + 1191.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1315.0, + 455.0, + 1315.0, + 455.0, + 1338.0, + 400.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1315.0, + 720.0, + 1315.0, + 720.0, + 1338.0, + 665.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1323.0, + 914.0, + 1323.0, + 914.0, + 1332.0, + 904.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1314.0, + 995.0, + 1314.0, + 995.0, + 1334.0, + 921.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1315.0, + 1240.0, + 1315.0, + 1240.0, + 1334.0, + 1191.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1352.0, + 395.0, + 1352.0, + 395.0, + 1367.0, + 382.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1352.0, + 674.0, + 1352.0, + 674.0, + 1369.0, + 645.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1354.0, + 920.0, + 1354.0, + 920.0, + 1365.0, + 911.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1348.0, + 1206.0, + 1348.0, + 1206.0, + 1372.0, + 1172.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1378.0, + 362.0, + 1378.0, + 362.0, + 1406.0, + 344.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1381.0, + 431.0, + 1381.0, + 431.0, + 1391.0, + 421.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1381.0, + 623.0, + 1381.0, + 623.0, + 1404.0, + 604.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1381.0, + 886.0, + 1381.0, + 886.0, + 1404.0, + 869.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1374.0, + 1000.0, + 1374.0, + 1000.0, + 1383.0, + 989.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1377.0, + 1153.0, + 1377.0, + 1153.0, + 1406.0, + 1132.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1411.0, + 992.0, + 1411.0, + 992.0, + 1435.0, + 974.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1439.0, + 978.0, + 1439.0, + 978.0, + 1448.0, + 967.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1491.0, + 477.0, + 1491.0, + 477.0, + 1514.0, + 451.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1491.0, + 1005.0, + 1491.0, + 1005.0, + 1514.0, + 978.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1491.0, + 1268.0, + 1491.0, + 1268.0, + 1514.0, + 1241.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1519.0, + 541.0, + 1519.0, + 541.0, + 1553.0, + 370.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1518.0, + 830.0, + 1518.0, + 830.0, + 1555.0, + 609.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1519.0, + 1091.0, + 1519.0, + 1091.0, + 1553.0, + 875.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1519.0, + 1364.0, + 1519.0, + 1364.0, + 1551.0, + 1124.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1565.0, + 383.0, + 1565.0, + 383.0, + 1575.0, + 375.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1559.0, + 444.0, + 1559.0, + 444.0, + 1578.0, + 388.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1559.0, + 708.0, + 1559.0, + 708.0, + 1578.0, + 651.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1561.0, + 971.0, + 1561.0, + 971.0, + 1576.0, + 916.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1559.0, + 1236.0, + 1559.0, + 1236.0, + 1578.0, + 1179.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1577.0, + 383.0, + 1577.0, + 383.0, + 1587.0, + 372.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1572.0, + 444.0, + 1572.0, + 444.0, + 1591.0, + 388.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1578.0, + 646.0, + 1578.0, + 646.0, + 1587.0, + 637.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1574.0, + 708.0, + 1574.0, + 708.0, + 1589.0, + 649.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1578.0, + 910.0, + 1578.0, + 910.0, + 1587.0, + 899.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1572.0, + 972.0, + 1572.0, + 972.0, + 1591.0, + 914.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1572.0, + 1235.0, + 1572.0, + 1235.0, + 1591.0, + 1179.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1585.0, + 499.0, + 1585.0, + 499.0, + 1604.0, + 386.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1585.0, + 763.0, + 1585.0, + 763.0, + 1604.0, + 651.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1585.0, + 1024.0, + 1585.0, + 1024.0, + 1604.0, + 914.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1585.0, + 1290.0, + 1585.0, + 1290.0, + 1604.0, + 1178.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1619.0, + 744.0, + 1619.0, + 744.0, + 1653.0, + 717.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1626.0, + 1253.0, + 1626.0, + 1253.0, + 1631.0, + 1247.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1645.0, + 360.0, + 1645.0, + 360.0, + 1669.0, + 341.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1644.0, + 625.0, + 1644.0, + 625.0, + 1671.0, + 606.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1639.0, + 739.0, + 1639.0, + 739.0, + 1739.0, + 696.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1643.0, + 889.0, + 1643.0, + 889.0, + 1672.0, + 867.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1631.0, + 1003.0, + 1631.0, + 1003.0, + 1662.0, + 988.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1644.0, + 1152.0, + 1644.0, + 1152.0, + 1670.0, + 1130.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1671.0, + 995.0, + 1671.0, + 995.0, + 1708.0, + 971.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1677.0, + 1255.0, + 1677.0, + 1255.0, + 1699.0, + 1242.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1757.0, + 739.0, + 1757.0, + 739.0, + 1776.0, + 717.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1757.0, + 1002.0, + 1757.0, + 1002.0, + 1776.0, + 980.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1755.0, + 1267.0, + 1755.0, + 1267.0, + 1779.0, + 1242.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1784.0, + 532.0, + 1784.0, + 532.0, + 1817.0, + 377.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1782.0, + 793.0, + 1782.0, + 793.0, + 1818.0, + 643.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1783.0, + 1061.0, + 1783.0, + 1061.0, + 1820.0, + 901.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1782.0, + 1337.0, + 1782.0, + 1337.0, + 1818.0, + 1152.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.25, + 1357.0, + 411.25, + 1357.0, + 411.25, + 1368.5, + 393.25, + 1368.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.25, + 1371.5, + 429.25, + 1371.5, + 429.25, + 1384.0, + 415.25, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1490.5, + 739.0, + 1490.5, + 739.0, + 1512.5, + 717.0, + 1512.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1755.0, + 474.0, + 1755.0, + 474.0, + 1777.0, + 453.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1843.0, + 1310.0, + 1843.0, + 1310.0, + 1879.0, + 294.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1843.0, + 1409.0, + 1843.0, + 1409.0, + 1879.0, + 1398.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1876.0, + 1405.0, + 1876.0, + 1405.0, + 1908.0, + 294.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1903.0, + 1407.0, + 1903.0, + 1407.0, + 1940.0, + 293.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1936.0, + 1405.0, + 1936.0, + 1405.0, + 1972.0, + 294.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1964.0, + 1407.0, + 1964.0, + 1407.0, + 2002.0, + 293.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1996.0, + 1288.0, + 1996.0, + 1288.0, + 2030.0, + 294.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 1407.0, + 306.0, + 1407.0, + 345.0, + 293.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 338.0, + 1407.0, + 338.0, + 1407.0, + 376.0, + 293.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 368.0, + 1405.0, + 368.0, + 1405.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 402.0, + 1404.0, + 402.0, + 1404.0, + 432.0, + 296.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 426.0, + 1407.0, + 426.0, + 1407.0, + 467.0, + 292.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 497.0, + 293.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 492.0, + 974.0, + 492.0, + 974.0, + 523.0, + 296.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1044.0, + 1406.0, + 1044.0, + 1406.0, + 1078.0, + 295.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1073.0, + 1407.0, + 1073.0, + 1407.0, + 1112.0, + 293.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1105.0, + 1406.0, + 1105.0, + 1406.0, + 1139.0, + 293.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1135.0, + 1405.0, + 1135.0, + 1405.0, + 1171.0, + 295.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1166.0, + 677.0, + 1166.0, + 677.0, + 1201.0, + 295.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1166.0, + 1406.0, + 1166.0, + 1406.0, + 1201.0, + 698.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1196.0, + 1408.0, + 1196.0, + 1408.0, + 1232.0, + 292.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 1406.0, + 535.0, + 1406.0, + 574.0, + 292.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 568.0, + 1405.0, + 568.0, + 1405.0, + 602.0, + 295.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 598.0, + 1403.0, + 598.0, + 1403.0, + 632.0, + 296.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 663.0, + 293.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 658.0, + 1406.0, + 658.0, + 1406.0, + 695.0, + 293.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 689.0, + 1407.0, + 689.0, + 1407.0, + 723.0, + 295.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 722.0, + 1405.0, + 722.0, + 1405.0, + 755.0, + 295.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 749.0, + 1409.0, + 749.0, + 1409.0, + 785.0, + 295.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 906.0, + 1404.0, + 906.0, + 1404.0, + 939.0, + 295.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 937.0, + 1407.0, + 937.0, + 1407.0, + 970.0, + 295.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 968.0, + 1405.0, + 968.0, + 1405.0, + 1000.0, + 295.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 999.0, + 1001.0, + 999.0, + 1001.0, + 1032.0, + 296.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1403.0, + 798.0, + 1403.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 828.0, + 1403.0, + 828.0, + 1403.0, + 861.0, + 296.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 858.0, + 1203.0, + 858.0, + 1203.0, + 894.0, + 295.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1403.0, + 226.0, + 1403.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 258.0, + 511.0, + 258.0, + 511.0, + 298.0, + 292.0, + 298.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 459, + 1405, + 459, + 1405, + 827, + 298, + 827 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1274, + 1405, + 1274, + 1405, + 1588, + 297, + 1588 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 841, + 1404, + 841, + 1404, + 1087, + 298, + 1087 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1404, + 229, + 1404, + 445, + 297, + 445 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1679, + 1404, + 1679, + 1404, + 1833, + 298, + 1833 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 561, + 1914, + 1140, + 1914, + 1140, + 2044, + 561, + 2044 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 298, + 1196, + 1403, + 1196, + 1403, + 1259, + 298, + 1259 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 295, + 1838, + 1397, + 1838, + 1397, + 1908, + 295, + 1908 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 303, + 1623, + 792, + 1623, + 792, + 1654, + 303, + 1654 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 1128, + 480, + 1128, + 480, + 1165, + 299, + 1165 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1964, + 1400, + 1964, + 1400, + 1994, + 1366, + 1994 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.668 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.494 + }, + { + "category_id": 13, + "poly": [ + 911, + 1335, + 1038, + 1335, + 1038, + 1368, + 911, + 1368 + ], + "score": 0.93, + "latex": "\\mathcal { X } \\times \\{ 0 , 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 782, + 1871, + 904, + 1871, + 904, + 1909, + 782, + 1909 + ], + "score": 0.93, + "latex": "t _ { 0 } \\in ( 0 , \\frac { 1 } { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1147, + 1458, + 1218, + 1458, + 1218, + 1489, + 1147, + 1489 + ], + "score": 0.93, + "latex": "\\widetilde y = 1" + }, + { + "category_id": 13, + "poly": [ + 1050, + 1428, + 1135, + 1428, + 1135, + 1461, + 1050, + 1461 + ], + "score": 0.93, + "latex": "\\tau _ { 1 , 0 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1336, + 363, + 1336, + 363, + 1367, + 296, + 1367 + ], + "score": 0.92, + "latex": "( { \\pmb x } , y )" + }, + { + "category_id": 13, + "poly": [ + 904, + 1838, + 989, + 1838, + 989, + 1873, + 904, + 1873 + ], + "score": 0.92, + "latex": "\\tau _ { 0 , 1 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 448, + 1427, + 533, + 1427, + 533, + 1459, + 448, + 1459 + ], + "score": 0.92, + "latex": "\\tau _ { 0 , 1 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 488, + 1457, + 727, + 1457, + 727, + 1491, + 488, + 1491 + ], + "score": 0.92, + "latex": "\\boldsymbol { \\widetilde { \\eta } } ( \\boldsymbol { \\dot { \\mathbf { x } } } ) = \\mathbb { P } [ \\boldsymbol { \\widetilde { y } } = 1 \\mid \\boldsymbol { \\mathbf { x } } ]" + }, + { + "category_id": 13, + "poly": [ + 1045, + 1839, + 1130, + 1839, + 1130, + 1874, + 1045, + 1874 + ], + "score": 0.92, + "latex": "\\tau _ { 1 , 0 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 551, + 1524, + 753, + 1524, + 753, + 1559, + 551, + 1559 + ], + "score": 0.92, + "latex": "f ( \\pmb { x } ) : \\mathcal { X } [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 512, + 1742, + 569, + 1742, + 569, + 1775, + 512, + 1775 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 341, + 1489, + 558, + 1489, + 558, + 1527, + 341, + 1527 + ], + "score": 0.92, + "latex": "\\eta ^ { * } ( \\pmb { x } ) = \\mathbb { I } _ { \\{ \\eta ( \\pmb { x } ) \\geq \\frac { 1 } { 2 } \\} }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1366, + 542, + 1366, + 542, + 1400, + 298, + 1400 + ], + "score": 0.92, + "latex": "\\eta ( \\pmb { x } ) = \\mathbb { P } [ y = 1 \\mid \\pmb { x } ]" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1366, + 1367, + 1366, + 1367, + 1400, + 1012, + 1400 + ], + "score": 0.91, + "latex": "\\bar { \\tau _ { 1 , 0 } } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 0 \\ | \\ y = 1 , \\pmb { x } ]" + }, + { + "category_id": 13, + "poly": [ + 604, + 1366, + 958, + 1366, + 958, + 1400, + 604, + 1400 + ], + "score": 0.91, + "latex": "\\tau _ { 0 , 1 } ( \\pmb { x } ) = \\mathbb { P } [ \\widetilde { y } = 1 \\ | \\ y = 0 , \\pmb { x } ]" + }, + { + "category_id": 13, + "poly": [ + 965, + 1876, + 1080, + 1876, + 1080, + 1906, + 965, + 1906 + ], + "score": 0.9, + "latex": "c _ { 1 } , c _ { 2 } > 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 1429, + 368, + 1429, + 368, + 1459, + 297, + 1459 + ], + "score": 0.89, + "latex": "y = 0" + }, + { + "category_id": 14, + "poly": [ + 559, + 1913, + 1145, + 1913, + 1145, + 2049, + 559, + 2049 + ], + "score": 0.89, + "latex": "\\begin{array} { c l l } { \\displaystyle \\tau _ { 1 , 0 } ( { \\pmb x } ) \\leq c _ { 1 } [ 1 - \\eta ( { \\pmb x } ) ] ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\geq \\displaystyle \\frac { 1 } { 2 } + t _ { 0 } , a n d } \\\\ { \\displaystyle \\tau _ { 0 , 1 } ( { \\pmb x } ) \\leq c _ { 1 } \\eta ( { \\pmb x } ) ^ { 1 + c _ { 2 } } ; \\forall \\eta ( { \\pmb x } ) \\leq \\displaystyle \\frac { 1 } { 2 } - t _ { 0 } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1490, + 1102, + 1490, + 1102, + 1520, + 1069, + 1520 + ], + "score": 0.87, + "latex": "\\mathbb { I } _ { A }" + }, + { + "category_id": 13, + "poly": [ + 600, + 1398, + 620, + 1398, + 620, + 1428, + 600, + 1428 + ], + "score": 0.84, + "latex": "\\widetilde { y }" + }, + { + "category_id": 13, + "poly": [ + 846, + 1337, + 873, + 1337, + 873, + 1363, + 846, + 1363 + ], + "score": 0.81, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1146, + 1307, + 1173, + 1307, + 1173, + 1333, + 1146, + 1333 + ], + "score": 0.8, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 828, + 1687, + 847, + 1687, + 847, + 1708, + 828, + 1708 + ], + "score": 0.78, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 739, + 1747, + 759, + 1747, + 759, + 1769, + 739, + 1769 + ], + "score": 0.76, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1331, + 1716, + 1351, + 1716, + 1351, + 1738, + 1331, + 1738 + ], + "score": 0.76, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1228, + 1490, + 1252, + 1490, + 1252, + 1516, + 1228, + 1516 + ], + "score": 0.76, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 1377, + 1463, + 1398, + 1463, + 1398, + 1485, + 1377, + 1485 + ], + "score": 0.74, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1403, + 1243, + 1403, + 1243, + 1424, + 1221, + 1424 + ], + "score": 0.69, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 611, + 1841, + 634, + 1841, + 634, + 1868, + 611, + 1868 + ], + "score": 0.31, + "latex": "A" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1621.0, + 797.0, + 1621.0, + 797.0, + 1657.0, + 296.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1123.0, + 486.0, + 1123.0, + 486.0, + 1173.0, + 290.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 460.0, + 1405.0, + 460.0, + 1405.0, + 494.0, + 295.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 525.0, + 295.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 520.0, + 1405.0, + 520.0, + 1405.0, + 556.0, + 293.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 1403.0, + 552.0, + 1403.0, + 586.0, + 295.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 583.0, + 1407.0, + 583.0, + 1407.0, + 617.0, + 295.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 647.0, + 293.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 642.0, + 1408.0, + 642.0, + 1408.0, + 678.0, + 292.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 709.0, + 293.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 698.0, + 1406.0, + 698.0, + 1406.0, + 744.0, + 291.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 769.0, + 295.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 765.0, + 1408.0, + 765.0, + 1408.0, + 799.0, + 293.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 793.0, + 1250.0, + 793.0, + 1250.0, + 829.0, + 293.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1273.0, + 1407.0, + 1273.0, + 1407.0, + 1309.0, + 295.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1304.0, + 1145.0, + 1304.0, + 1145.0, + 1337.0, + 296.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1304.0, + 1403.0, + 1304.0, + 1403.0, + 1337.0, + 1174.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1333.0, + 295.0, + 1333.0, + 295.0, + 1373.0, + 292.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1333.0, + 845.0, + 1333.0, + 845.0, + 1373.0, + 364.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1333.0, + 910.0, + 1333.0, + 910.0, + 1373.0, + 874.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1333.0, + 1406.0, + 1333.0, + 1406.0, + 1373.0, + 1039.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1364.0, + 297.0, + 1364.0, + 297.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1364.0, + 603.0, + 1364.0, + 603.0, + 1404.0, + 543.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1364.0, + 1011.0, + 1364.0, + 1011.0, + 1404.0, + 959.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1364.0, + 1406.0, + 1364.0, + 1406.0, + 1404.0, + 1368.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1395.0, + 599.0, + 1395.0, + 599.0, + 1430.0, + 294.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1395.0, + 1220.0, + 1395.0, + 1220.0, + 1430.0, + 621.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1395.0, + 1406.0, + 1395.0, + 1406.0, + 1430.0, + 1244.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1425.0, + 296.0, + 1425.0, + 296.0, + 1465.0, + 292.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1425.0, + 447.0, + 1425.0, + 447.0, + 1465.0, + 369.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1425.0, + 1049.0, + 1425.0, + 1049.0, + 1465.0, + 534.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1425.0, + 1407.0, + 1425.0, + 1407.0, + 1465.0, + 1136.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1457.0, + 487.0, + 1457.0, + 487.0, + 1493.0, + 295.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1457.0, + 1146.0, + 1457.0, + 1146.0, + 1493.0, + 728.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1457.0, + 1376.0, + 1457.0, + 1376.0, + 1493.0, + 1219.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1457.0, + 1408.0, + 1457.0, + 1408.0, + 1493.0, + 1399.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1481.0, + 340.0, + 1481.0, + 340.0, + 1532.0, + 289.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1481.0, + 1068.0, + 1481.0, + 1068.0, + 1532.0, + 559.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1481.0, + 1227.0, + 1481.0, + 1227.0, + 1532.0, + 1103.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1481.0, + 1411.0, + 1481.0, + 1411.0, + 1532.0, + 1253.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1524.0, + 550.0, + 1524.0, + 550.0, + 1560.0, + 296.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1560.0, + 754.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1555.0, + 631.0, + 1555.0, + 631.0, + 1591.0, + 294.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 841.0, + 1405.0, + 841.0, + 1405.0, + 879.0, + 293.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 874.0, + 1406.0, + 874.0, + 1406.0, + 908.0, + 294.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 904.0, + 1405.0, + 904.0, + 1405.0, + 938.0, + 293.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 934.0, + 1405.0, + 934.0, + 1405.0, + 968.0, + 294.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 960.0, + 1406.0, + 960.0, + 1406.0, + 1000.0, + 292.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 994.0, + 1406.0, + 994.0, + 1406.0, + 1029.0, + 292.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1025.0, + 1405.0, + 1025.0, + 1405.0, + 1059.0, + 294.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1053.0, + 977.0, + 1053.0, + 977.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1407.0, + 228.0, + 1407.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1406.0, + 262.0, + 1406.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 325.0, + 292.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 357.0, + 295.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1404.0, + 351.0, + 1404.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 381.0, + 1404.0, + 381.0, + 1404.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 412.0, + 774.0, + 412.0, + 774.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1677.0, + 827.0, + 1677.0, + 827.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1677.0, + 1406.0, + 1677.0, + 1406.0, + 1716.0, + 848.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1711.0, + 1330.0, + 1711.0, + 1330.0, + 1744.0, + 296.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1711.0, + 1406.0, + 1711.0, + 1406.0, + 1744.0, + 1352.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1739.0, + 511.0, + 1739.0, + 511.0, + 1777.0, + 292.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1739.0, + 738.0, + 1739.0, + 738.0, + 1777.0, + 570.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1739.0, + 1406.0, + 1739.0, + 1406.0, + 1777.0, + 760.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1801.0, + 979.0, + 1801.0, + 979.0, + 1837.0, + 293.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1197.0, + 1407.0, + 1197.0, + 1407.0, + 1233.0, + 296.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1228.0, + 1356.0, + 1228.0, + 1356.0, + 1262.0, + 293.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1836.0, + 610.0, + 1836.0, + 610.0, + 1877.0, + 294.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1836.0, + 903.0, + 1836.0, + 903.0, + 1877.0, + 635.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1836.0, + 1044.0, + 1836.0, + 1044.0, + 1877.0, + 990.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1836.0, + 1403.0, + 1836.0, + 1403.0, + 1877.0, + 1131.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1870.0, + 781.0, + 1870.0, + 781.0, + 1912.0, + 295.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1870.0, + 964.0, + 1870.0, + 964.0, + 1912.0, + 905.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1870.0, + 1202.0, + 1870.0, + 1202.0, + 1912.0, + 1081.0, + 1912.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1547, + 1405, + 1547, + 1405, + 2034, + 297, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 414, + 298, + 414 + ], + "score": 0.978 + }, + { + "category_id": 4, + "poly": [ + 296, + 1145, + 1406, + 1145, + 1406, + 1422, + 296, + 1422 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 314, + 682, + 1386, + 682, + 1386, + 1121, + 314, + 1121 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 429, + 1405, + 429, + 1405, + 646, + 297, + 646 + ], + "score": 0.968 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 299, + 1486, + 913, + 1486, + 913, + 1517, + 299, + 1517 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.769 + }, + { + "category_id": 13, + "poly": [ + 872, + 1911, + 1022, + 1911, + 1022, + 1945, + 872, + 1945 + ], + "score": 0.94, + "latex": "\\theta = 1 / 2 - T" + }, + { + "category_id": 13, + "poly": [ + 536, + 1789, + 744, + 1789, + 744, + 1823, + 536, + 1823 + ], + "score": 0.94, + "latex": "| f ( \\bar { \\pmb { x } } ) - 1 / 2 | > \\theta" + }, + { + "category_id": 13, + "poly": [ + 1066, + 1299, + 1136, + 1299, + 1136, + 1333, + 1066, + 1333 + ], + "score": 0.93, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 767, + 613, + 852, + 613, + 852, + 646, + 767, + 646 + ], + "score": 0.92, + "latex": "\\tau _ { 0 , 1 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 1184, + 292, + 1269, + 292, + 1269, + 326, + 1184, + 326 + ], + "score": 0.92, + "latex": "\\tau _ { 0 , 1 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 766, + 289, + 1015, + 289, + 1015, + 327, + 766, + 327 + ], + "score": 0.92, + "latex": "\\begin{array} { r } \\{ \\pmb { x } : | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\} < t _ { 0 } \\} \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1319, + 291, + 1403, + 291, + 1403, + 325, + 1319, + 325 + ], + "score": 0.92, + "latex": "\\tau _ { 1 , 0 } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 2003, + 367, + 2003, + 367, + 2036, + 297, + 2036 + ], + "score": 0.92, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1208, + 1120, + 1208, + 1120, + 1241, + 1062, + 1241 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1177, + 1308, + 1177, + 1308, + 1210, + 1239, + 1210 + ], + "score": 0.91, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 924, + 613, + 1015, + 613, + 1015, + 646, + 924, + 646 + ], + "score": 0.91, + "latex": "\\tau _ { 1 , 0 } ( \\pmb { x } ) )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1360, + 355, + 1360, + 355, + 1393, + 297, + 1393 + ], + "score": 0.91, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 572, + 1730, + 601, + 1730, + 601, + 1761, + 572, + 1761 + ], + "score": 0.87, + "latex": "\\eta ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 665, + 234, + 692, + 234, + 692, + 261, + 665, + 261 + ], + "score": 0.86, + "latex": "t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1791, + 1219, + 1791, + 1219, + 1822, + 1199, + 1822 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 809, + 1700, + 829, + 1700, + 829, + 1731, + 809, + 1731 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 780, + 1761, + 800, + 1761, + 800, + 1791, + 780, + 1791 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1315, + 1639, + 1336, + 1639, + 1336, + 1670, + 1315, + 1670 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 948, + 1791, + 967, + 1791, + 967, + 1822, + 948, + 1822 + ], + "score": 0.83, + "latex": "\\tilde { y }" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1761, + 1163, + 1761, + 1163, + 1791, + 1143, + 1791 + ], + "score": 0.83, + "latex": "\\tilde { y }" + }, + { + "category_id": 13, + "poly": [ + 761, + 1914, + 783, + 1914, + 783, + 1939, + 761, + 1939 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 670, + 1209, + 692, + 1209, + 692, + 1240, + 670, + 1240 + ], + "score": 0.8, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 530, + 1212, + 551, + 1212, + 551, + 1240, + 530, + 1240 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 406, + 1852, + 422, + 1852, + 422, + 1878, + 406, + 1878 + ], + "score": 0.76, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 722, + 1761, + 739, + 1761, + 739, + 1787, + 722, + 1787 + ], + "score": 0.75, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 519, + 267, + 538, + 267, + 538, + 289, + 519, + 289 + ], + "score": 0.73, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 902, + 236, + 920, + 236, + 920, + 258, + 902, + 258 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1376, + 618, + 1396, + 618, + 1396, + 640, + 1376, + 640 + ], + "score": 0.64, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1371, + 1181, + 1391, + 1181, + 1391, + 1210, + 1371, + 1210 + ], + "score": 0.63, + "latex": "y" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1143.0, + 1404.0, + 1143.0, + 1404.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1175.0, + 1238.0, + 1175.0, + 1238.0, + 1214.0, + 291.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1175.0, + 1370.0, + 1175.0, + 1370.0, + 1214.0, + 1309.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1392.0, + 1175.0, + 1408.0, + 1175.0, + 1408.0, + 1214.0, + 1392.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 529.0, + 1206.0, + 529.0, + 1243.0, + 294.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1206.0, + 669.0, + 1206.0, + 669.0, + 1243.0, + 552.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1206.0, + 1061.0, + 1206.0, + 1061.0, + 1243.0, + 693.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1206.0, + 1407.0, + 1206.0, + 1407.0, + 1243.0, + 1121.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1266.0, + 1408.0, + 1266.0, + 1408.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1299.0, + 1065.0, + 1299.0, + 1065.0, + 1336.0, + 293.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1299.0, + 1407.0, + 1299.0, + 1407.0, + 1336.0, + 1137.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1329.0, + 1406.0, + 1329.0, + 1406.0, + 1366.0, + 293.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 296.0, + 1359.0, + 296.0, + 1393.0, + 293.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1359.0, + 1406.0, + 1359.0, + 1406.0, + 1393.0, + 356.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 1243.0, + 1390.0, + 1243.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 690.0, + 443.0, + 690.0, + 443.0, + 717.0, + 380.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 692.0, + 741.0, + 692.0, + 741.0, + 715.0, + 656.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 692.0, + 1017.0, + 692.0, + 1017.0, + 715.0, + 932.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 692.0, + 1292.0, + 692.0, + 1292.0, + 715.0, + 1205.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 707.0, + 440.0, + 707.0, + 440.0, + 735.0, + 382.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 710.0, + 740.0, + 710.0, + 740.0, + 734.0, + 657.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 709.0, + 1016.0, + 709.0, + 1016.0, + 736.0, + 932.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 711.0, + 1290.0, + 711.0, + 1290.0, + 734.0, + 1206.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 776.0, + 341.0, + 776.0, + 341.0, + 805.0, + 318.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 774.0, + 618.0, + 774.0, + 618.0, + 808.0, + 594.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 779.0, + 890.0, + 779.0, + 890.0, + 804.0, + 871.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 781.0, + 1163.0, + 781.0, + 1163.0, + 805.0, + 1144.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 794.0, + 1214.0, + 794.0, + 1214.0, + 810.0, + 1200.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 812.0, + 1222.0, + 812.0, + 1222.0, + 827.0, + 1203.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 894.0, + 464.0, + 894.0, + 464.0, + 922.0, + 432.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 894.0, + 739.0, + 894.0, + 739.0, + 922.0, + 707.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 894.0, + 1015.0, + 894.0, + 1015.0, + 922.0, + 983.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 897.0, + 1286.0, + 897.0, + 1286.0, + 920.0, + 1259.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 962.0, + 354.0, + 962.0, + 354.0, + 970.0, + 342.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 931.0, + 626.0, + 931.0, + 626.0, + 1084.0, + 587.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 928.0, + 910.0, + 928.0, + 910.0, + 1086.0, + 862.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 924.0, + 1175.0, + 924.0, + 1175.0, + 1085.0, + 1135.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 958.0, + 1342.0, + 958.0, + 1342.0, + 974.0, + 1220.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 975.0, + 356.0, + 975.0, + 356.0, + 991.0, + 341.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 976.0, + 907.0, + 976.0, + 907.0, + 987.0, + 895.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 987.0, + 356.0, + 987.0, + 356.0, + 1034.0, + 316.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 991.0, + 910.0, + 991.0, + 910.0, + 1006.0, + 893.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 988.0, + 1344.0, + 988.0, + 1344.0, + 1007.0, + 1271.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1015.0, + 353.0, + 1015.0, + 353.0, + 1023.0, + 342.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1001.0, + 635.0, + 1001.0, + 635.0, + 1016.0, + 617.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1006.0, + 911.0, + 1006.0, + 911.0, + 1025.0, + 892.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1000.0, + 1380.0, + 1000.0, + 1380.0, + 1026.0, + 1269.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1028.0, + 356.0, + 1028.0, + 356.0, + 1042.0, + 340.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1023.0, + 911.0, + 1023.0, + 911.0, + 1042.0, + 892.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1036.0, + 1354.0, + 1036.0, + 1354.0, + 1042.0, + 1348.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1043.0, + 360.0, + 1043.0, + 360.0, + 1062.0, + 338.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1043.0, + 910.0, + 1043.0, + 910.0, + 1057.0, + 893.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1045.0, + 1178.0, + 1045.0, + 1178.0, + 1058.0, + 1170.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1055.0, + 473.0, + 1055.0, + 473.0, + 1084.0, + 439.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1053.0, + 751.0, + 1053.0, + 751.0, + 1081.0, + 718.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1053.0, + 1026.0, + 1053.0, + 1026.0, + 1081.0, + 993.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1054.0, + 1299.0, + 1054.0, + 1299.0, + 1082.0, + 1266.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1086.0, + 531.0, + 1086.0, + 531.0, + 1121.0, + 344.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1086.0, + 819.0, + 1086.0, + 819.0, + 1124.0, + 607.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1084.0, + 1077.0, + 1084.0, + 1077.0, + 1126.0, + 901.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1086.0, + 1352.0, + 1086.0, + 1352.0, + 1124.0, + 1175.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1484.0, + 916.0, + 1484.0, + 916.0, + 1520.0, + 295.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1545.0, + 1407.0, + 1545.0, + 1407.0, + 1581.0, + 294.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1576.0, + 1408.0, + 1576.0, + 1408.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1606.0, + 1405.0, + 1606.0, + 1405.0, + 1643.0, + 292.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1634.0, + 1314.0, + 1634.0, + 1314.0, + 1674.0, + 292.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1634.0, + 1408.0, + 1634.0, + 1408.0, + 1674.0, + 1337.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1668.0, + 1408.0, + 1668.0, + 1408.0, + 1702.0, + 295.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1698.0, + 808.0, + 1698.0, + 808.0, + 1735.0, + 292.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1735.0, + 830.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1729.0, + 571.0, + 1729.0, + 571.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1764.0, + 602.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1759.0, + 721.0, + 1759.0, + 721.0, + 1795.0, + 291.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1759.0, + 779.0, + 1759.0, + 779.0, + 1795.0, + 740.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1759.0, + 1142.0, + 1759.0, + 1142.0, + 1795.0, + 801.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1759.0, + 1406.0, + 1759.0, + 1406.0, + 1795.0, + 1164.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 535.0, + 1788.0, + 535.0, + 1825.0, + 294.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1788.0, + 947.0, + 1788.0, + 947.0, + 1825.0, + 745.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1788.0, + 1198.0, + 1788.0, + 1198.0, + 1825.0, + 968.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1788.0, + 1405.0, + 1788.0, + 1405.0, + 1825.0, + 1220.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 405.0, + 1849.0, + 405.0, + 1886.0, + 292.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1886.0, + 423.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1879.0, + 1406.0, + 1879.0, + 1406.0, + 1919.0, + 291.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 760.0, + 1912.0, + 760.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1912.0, + 871.0, + 1912.0, + 871.0, + 1946.0, + 784.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 1023.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1403.0, + 1941.0, + 1403.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 2002.0, + 724.0, + 2002.0, + 724.0, + 2040.0, + 368.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 664.0, + 227.0, + 664.0, + 264.0, + 293.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 227.0, + 901.0, + 227.0, + 901.0, + 264.0, + 693.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 264.0, + 921.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 259.0, + 518.0, + 259.0, + 518.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 259.0, + 1407.0, + 259.0, + 1407.0, + 296.0, + 539.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 765.0, + 291.0, + 765.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 291.0, + 1183.0, + 291.0, + 1183.0, + 327.0, + 1016.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 291.0, + 1318.0, + 291.0, + 1318.0, + 327.0, + 1270.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 357.0, + 294.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 384.0, + 1019.0, + 384.0, + 1019.0, + 416.0, + 296.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 430.0, + 1405.0, + 430.0, + 1405.0, + 464.0, + 295.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 461.0, + 1405.0, + 461.0, + 1405.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 489.0, + 1405.0, + 489.0, + 1405.0, + 526.0, + 292.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 521.0, + 1406.0, + 521.0, + 1406.0, + 556.0, + 295.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 554.0, + 1406.0, + 554.0, + 1406.0, + 585.0, + 296.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 581.0, + 1407.0, + 581.0, + 1407.0, + 618.0, + 294.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 611.0, + 766.0, + 611.0, + 766.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 611.0, + 923.0, + 611.0, + 923.0, + 648.0, + 853.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 611.0, + 1375.0, + 611.0, + 1375.0, + 648.0, + 1016.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 611.0, + 1407.0, + 611.0, + 1407.0, + 648.0, + 1397.0, + 648.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1188, + 1405, + 1188, + 1405, + 1405, + 298, + 1405 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 1795, + 1405, + 1795, + 1405, + 1959, + 297, + 1959 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 932, + 1407, + 932, + 1407, + 1093, + 296, + 1093 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1483, + 1406, + 1483, + 1406, + 1606, + 297, + 1606 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 295, + 1612, + 1403, + 1612, + 1403, + 1688, + 295, + 1688 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1402, + 1971, + 1402, + 2035, + 296, + 2035 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 362, + 1698, + 1333, + 1698, + 1333, + 1775, + 362, + 1775 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 298, + 1135, + 498, + 1135, + 498, + 1170, + 298, + 1170 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 1440, + 541, + 1440, + 541, + 1471, + 299, + 1471 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1722, + 1400, + 1722, + 1400, + 1751, + 1367, + 1751 + ], + "score": 0.82 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 300, + 302, + 1156, + 302, + 1156, + 880, + 300, + 880 + ], + "score": 0.704 + }, + { + "category_id": 0, + "poly": [ + 298, + 263, + 918, + 263, + 918, + 294, + 298, + 294 + ], + "score": 0.258 + }, + { + "category_id": 13, + "poly": [ + 462, + 993, + 590, + 993, + 590, + 1033, + 462, + 1033 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { | f ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1107, + 996, + 1191, + 996, + 1191, + 1030, + 1107, + 1030 + ], + "score": 0.93, + "latex": "f _ { h _ { x } } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 834, + 1796, + 1006, + 1796, + 1006, + 1831, + 834, + 1831 + ], + "score": 0.93, + "latex": "\\eta ( z ) ~ > ~ \\eta ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 568, + 1545, + 627, + 1545, + 627, + 1578, + 568, + 1578 + ], + "score": 0.92, + "latex": "f ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 684, + 1925, + 743, + 1925, + 743, + 1959, + 684, + 1959 + ], + "score": 0.92, + "latex": "\\eta ^ { * } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 814, + 1544, + 871, + 1544, + 871, + 1578, + 814, + 1578 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 925, + 1642, + 995, + 1642, + 995, + 1680, + 925, + 1680 + ], + "score": 0.92, + "latex": "\\varepsilon < \\textstyle { \\frac { 1 } { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1613, + 1350, + 1613, + 1350, + 1646, + 1117, + 1646 + ], + "score": 0.92, + "latex": "( { \\pmb x } , \\tilde { y } ) \\sim D ( { \\pmb x } , \\tilde { \\eta } ( { \\pmb x } ) )" + }, + { + "category_id": 13, + "poly": [ + 1151, + 1645, + 1214, + 1645, + 1214, + 1677, + 1151, + 1677 + ], + "score": 0.91, + "latex": "( \\alpha , \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 948, + 1926, + 1006, + 1926, + 1006, + 1959, + 948, + 1959 + ], + "score": 0.91, + "latex": "\\eta ^ { * } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1342, + 1484, + 1398, + 1484, + 1398, + 1517, + 1342, + 1517 + ], + "score": 0.91, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1167, + 933, + 1232, + 933, + 1232, + 966, + 1167, + 966 + ], + "score": 0.9, + "latex": "f _ { i } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 632, + 339, + 747, + 339, + 747, + 370, + 632, + 370 + ], + "score": 0.9, + "latex": "( T _ { 0 } , T _ { e n d } )" + }, + { + "category_id": 13, + "poly": [ + 477, + 302, + 816, + 302, + 816, + 340, + 477, + 340 + ], + "score": 0.9, + "latex": "\\tilde { S } = \\{ ( \\pmb { x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { 0 } ) , \\cdot \\cdot \\cdot , ( \\pmb { x } _ { n } , \\widetilde { y } _ { n } ^ { 0 } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 483, + 1828, + 553, + 1828, + 553, + 1861, + 483, + 1861 + ], + "score": 0.9, + "latex": "\\eta ^ { \\ast } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 544, + 1897, + 614, + 1897, + 614, + 1927, + 544, + 1927 + ], + "score": 0.89, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 452, + 523, + 620, + 523, + 620, + 561, + 452, + 561 + ], + "score": 0.89, + "latex": "( { \\pmb x } _ { i } , \\widetilde { { y } } _ { i } ^ { t - 1 } ) \\in \\tilde { S }" + }, + { + "category_id": 13, + "poly": [ + 438, + 492, + 497, + 492, + 497, + 525, + 438, + 525 + ], + "score": 0.89, + "latex": "f ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 559, + 1612, + 623, + 1612, + 623, + 1646, + 559, + 1646 + ], + "score": 0.89, + "latex": "( \\alpha , \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 695, + 965, + 730, + 965, + 730, + 994, + 695, + 994 + ], + "score": 0.89, + "latex": "h _ { x }" + }, + { + "category_id": 14, + "poly": [ + 361, + 1698, + 1337, + 1698, + 1337, + 1776, + 361, + 1776 + ], + "score": 0.88, + "latex": "| f ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\leq \\alpha \\mathbb { E } _ { ( \\pmb { z } , \\bar { \\pmb { y } } ) \\sim D ( \\pmb { z } , \\bar { \\eta } ( \\pmb { z } ) ) } [ 1 _ { \\{ \\bar { \\pmb { y } } _ { \\pmb { z } } \\neq \\eta ^ { * } ( \\pmb { z } ) \\} } ( \\pmb { z } ) | | \\eta ( \\pmb { z } ) - \\frac { 1 } { 2 } | \\geq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | ] + \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 396, + 629, + 471, + 629, + 471, + 657, + 396, + 657 + ], + "score": 0.88, + "latex": "t \\geq m" + }, + { + "category_id": 13, + "poly": [ + 946, + 306, + 1004, + 306, + 1004, + 339, + 946, + 339 + ], + "score": 0.88, + "latex": "f ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 548, + 1062, + 582, + 1062, + 582, + 1091, + 548, + 1091 + ], + "score": 0.88, + "latex": "h _ { x }" + }, + { + "category_id": 13, + "poly": [ + 675, + 524, + 862, + 524, + 862, + 563, + 675, + 563 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { | f ( \\pmb { x } _ { i } ) - \\frac { 1 } { 2 } | \\geq \\theta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1644, + 840, + 1644, + 840, + 1690, + 297, + 1690 + ], + "score": 0.87, + "latex": "f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )" + }, + { + "category_id": 13, + "poly": [ + 345, + 428, + 509, + 428, + 509, + 460, + 345, + 460 + ], + "score": 0.87, + "latex": "\\theta 1 / 2 - T _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 498, + 1061, + 518, + 1061, + 518, + 1092, + 498, + 1092 + ], + "score": 0.85, + "latex": "\\widetilde { y }" + }, + { + "category_id": 13, + "poly": [ + 344, + 399, + 439, + 399, + 439, + 427, + 344, + 427 + ], + "score": 0.84, + "latex": "T \\gets T _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1155, + 966, + 1398, + 966, + 1398, + 996, + 1155, + 996 + ], + "score": 0.84, + "latex": "h _ { \\pmb { x } } = \\arg \\operatorname* { m a x } _ { i } f _ { i } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 386, + 461, + 548, + 461, + 548, + 490, + 386, + 490 + ], + "score": 0.83, + "latex": "t \\gets 1 , \\cdots , N" + }, + { + "category_id": 13, + "poly": [ + 360, + 1029, + 429, + 1029, + 429, + 1064, + 360, + 1064 + ], + "score": 0.82, + "latex": "f _ { \\widetilde { y } } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 873, + 1515, + 900, + 1515, + 900, + 1542, + 873, + 1542 + ], + "score": 0.8, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 536, + 489, + 559, + 489, + 559, + 521, + 536, + 521 + ], + "score": 0.8, + "latex": "\\tilde { S }" + }, + { + "category_id": 13, + "poly": [ + 1257, + 1282, + 1283, + 1282, + 1283, + 1308, + 1257, + 1308 + ], + "score": 0.79, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 406, + 367, + 507, + 367, + 507, + 400, + 406, + 400 + ], + "score": 0.79, + "latex": "f _ { f i n a l } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 870, + 1835, + 892, + 1835, + 892, + 1856, + 870, + 1856 + ], + "score": 0.79, + "latex": "_ z" + }, + { + "category_id": 13, + "poly": [ + 298, + 1062, + 314, + 1062, + 314, + 1088, + 298, + 1088 + ], + "score": 0.78, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 300, + 1859, + 788, + 1859, + 788, + 1898, + 300, + 1898 + ], + "score": 0.78, + "latex": "\\begin{array} { r } { \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } \\left( z : \\left| \\eta ( z ) - \\frac { 1 } { 2 } \\right| \\ge \\left| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } \\right| \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 427, + 723, + 731, + 723, + 731, + 755, + 427, + 755 + ], + "score": 0.78, + "latex": "T \\gets \\operatorname* { m i n } ( T ( 1 + \\beta ) , \\bar { T _ { e n d } } )" + }, + { + "category_id": 13, + "poly": [ + 1094, + 1646, + 1120, + 1646, + 1120, + 1671, + 1094, + 1671 + ], + "score": 0.78, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 849, + 1930, + 869, + 1930, + 869, + 1953, + 849, + 1953 + ], + "score": 0.77, + "latex": "_ z" + }, + { + "category_id": 13, + "poly": [ + 866, + 343, + 894, + 343, + 894, + 365, + 866, + 365 + ], + "score": 0.76, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 594, + 1804, + 616, + 1804, + 616, + 1825, + 594, + 1825 + ], + "score": 0.76, + "latex": "_ { z }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1932, + 1360, + 1932, + 1360, + 1953, + 1340, + 1953 + ], + "score": 0.76, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 308, + 1137, + 308, + 1137, + 338, + 1116, + 338 + ], + "score": 0.74, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 679, + 1804, + 702, + 1804, + 702, + 1826, + 679, + 1826 + ], + "score": 0.74, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 963, + 1398, + 963, + 1398, + 997, + 1333, + 997 + ], + "score": 0.74, + "latex": "f _ { i } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 399, + 657, + 557, + 657, + 557, + 690, + 399, + 690 + ], + "score": 0.71, + "latex": "\\theta 1 / 2 - T" + }, + { + "category_id": 13, + "poly": [ + 298, + 1901, + 317, + 1901, + 317, + 1922, + 298, + 1922 + ], + "score": 0.7, + "latex": "_ z" + }, + { + "category_id": 13, + "poly": [ + 1091, + 1833, + 1113, + 1833, + 1113, + 1857, + 1091, + 1857 + ], + "score": 0.7, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 628, + 966, + 641, + 966, + 641, + 991, + 628, + 991 + ], + "score": 0.69, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1032, + 338, + 1061, + 338, + 1061, + 364, + 1032, + 364 + ], + "score": 0.68, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1380, + 1904, + 1400, + 1904, + 1400, + 1922, + 1380, + 1922 + ], + "score": 0.64, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 371, + 809, + 718, + 809, + 718, + 848, + 371, + 848 + ], + "score": 0.6, + "latex": "\\tilde { S } \\gets \\{ ( { \\pmb x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { t } ) , \\cdot \\cdot \\cdot , ( { \\pmb x } _ { n } , \\widetilde { y } _ { n } ^ { t } ) \\}" + }, + { + "category_id": 14, + "poly": [ + 371, + 810, + 719, + 810, + 719, + 848, + 371, + 848 + ], + "score": 0.59, + "latex": "\\tilde { S } \\gets \\{ ( { \\pmb x } _ { 1 } , \\widetilde { y } _ { 1 } ^ { t } ) , \\cdot \\cdot \\cdot , ( { \\pmb x } _ { n } , \\widetilde { y } _ { n } ^ { t } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 331, + 1030, + 350, + 1030, + 350, + 1062, + 331, + 1062 + ], + "score": 0.56, + "latex": "\\widetilde { y }" + }, + { + "category_id": 13, + "poly": [ + 409, + 688, + 751, + 688, + 751, + 723, + 409, + 723 + ], + "score": 0.54, + "latex": "\\mathbf { f } \\forall i \\in [ 1 , \\cdot \\cdot \\cdot , n ] , \\widetilde { y } _ { i } ^ { t } = \\widetilde { y } _ { i } ^ { t - 1 } \\mathbf { t h }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1647, + 1358, + 1647, + 1358, + 1675, + 1340, + 1675 + ], + "score": 0.53, + "latex": "i f" + }, + { + "category_id": 13, + "poly": [ + 300, + 1861, + 434, + 1861, + 434, + 1898, + 300, + 1898 + ], + "score": 0.49, + "latex": "\\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} }" + }, + { + "category_id": 13, + "poly": [ + 399, + 561, + 589, + 561, + 589, + 602, + 399, + 602 + ], + "score": 0.44, + "latex": "\\widetilde { y } _ { i } ^ { t } \\gets \\mathbb { I } _ { \\{ f ( \\pmb { x } _ { i } ) \\geq \\frac { 1 } { 2 } \\} }" + }, + { + "category_id": 14, + "poly": [ + 399, + 561, + 591, + 561, + 591, + 602, + 399, + 602 + ], + "score": 0.3, + "latex": "\\widetilde { y } _ { i } ^ { t } \\gets \\mathbb { I } _ { \\{ f ( \\pmb { x } _ { i } ) \\geq \\frac { 1 } { 2 } \\} }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1130.0, + 502.0, + 1130.0, + 502.0, + 1177.0, + 291.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1437.0, + 543.0, + 1437.0, + 543.0, + 1476.0, + 294.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 260.0, + 919.0, + 260.0, + 919.0, + 299.0, + 296.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1186.0, + 1405.0, + 1186.0, + 1405.0, + 1226.0, + 295.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1221.0, + 1406.0, + 1221.0, + 1406.0, + 1256.0, + 295.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1247.0, + 1406.0, + 1247.0, + 1406.0, + 1287.0, + 292.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 1256.0, + 1280.0, + 1256.0, + 1316.0, + 293.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1280.0, + 1405.0, + 1280.0, + 1405.0, + 1316.0, + 1284.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1306.0, + 1405.0, + 1306.0, + 1405.0, + 1350.0, + 292.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1341.0, + 1403.0, + 1341.0, + 1403.0, + 1376.0, + 295.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1373.0, + 1038.0, + 1373.0, + 1038.0, + 1407.0, + 294.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1792.0, + 593.0, + 1792.0, + 593.0, + 1833.0, + 292.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1792.0, + 678.0, + 1792.0, + 678.0, + 1833.0, + 617.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1792.0, + 833.0, + 1792.0, + 833.0, + 1833.0, + 703.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1792.0, + 1408.0, + 1792.0, + 1408.0, + 1833.0, + 1007.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1825.0, + 482.0, + 1825.0, + 482.0, + 1862.0, + 294.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1825.0, + 869.0, + 1825.0, + 869.0, + 1862.0, + 554.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1825.0, + 1090.0, + 1825.0, + 1090.0, + 1862.0, + 893.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1825.0, + 1405.0, + 1825.0, + 1405.0, + 1862.0, + 1114.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1858.0, + 299.0, + 1858.0, + 299.0, + 1899.0, + 292.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1858.0, + 1408.0, + 1858.0, + 1408.0, + 1899.0, + 789.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1895.0, + 543.0, + 1895.0, + 543.0, + 1928.0, + 318.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1895.0, + 1379.0, + 1895.0, + 1379.0, + 1928.0, + 615.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1895.0, + 1405.0, + 1895.0, + 1405.0, + 1928.0, + 1401.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1924.0, + 683.0, + 1924.0, + 683.0, + 1961.0, + 294.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1924.0, + 848.0, + 1924.0, + 848.0, + 1961.0, + 744.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1924.0, + 947.0, + 1924.0, + 947.0, + 1961.0, + 870.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1924.0, + 1339.0, + 1924.0, + 1339.0, + 1961.0, + 1007.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1924.0, + 1374.0, + 1924.0, + 1374.0, + 1961.0, + 1361.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 931.0, + 1166.0, + 931.0, + 1166.0, + 969.0, + 295.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 931.0, + 1407.0, + 931.0, + 1407.0, + 969.0, + 1233.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 627.0, + 959.0, + 627.0, + 1002.0, + 294.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 959.0, + 694.0, + 959.0, + 694.0, + 1002.0, + 642.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 959.0, + 1154.0, + 959.0, + 1154.0, + 1002.0, + 731.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 959.0, + 1408.0, + 959.0, + 1408.0, + 1002.0, + 1399.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 993.0, + 461.0, + 993.0, + 461.0, + 1036.0, + 294.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 993.0, + 1106.0, + 993.0, + 1106.0, + 1036.0, + 591.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 993.0, + 1408.0, + 993.0, + 1408.0, + 1036.0, + 1192.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1025.0, + 330.0, + 1025.0, + 330.0, + 1068.0, + 291.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1025.0, + 359.0, + 1025.0, + 359.0, + 1068.0, + 351.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1025.0, + 1408.0, + 1025.0, + 1408.0, + 1068.0, + 430.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1059.0, + 297.0, + 1059.0, + 297.0, + 1097.0, + 294.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1059.0, + 497.0, + 1059.0, + 497.0, + 1097.0, + 315.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1059.0, + 547.0, + 1059.0, + 547.0, + 1097.0, + 519.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1059.0, + 1409.0, + 1059.0, + 1409.0, + 1097.0, + 583.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1478.0, + 1341.0, + 1478.0, + 1341.0, + 1521.0, + 292.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1478.0, + 1409.0, + 1478.0, + 1409.0, + 1521.0, + 1399.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1511.0, + 872.0, + 1511.0, + 872.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1511.0, + 1406.0, + 1511.0, + 1406.0, + 1549.0, + 901.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1545.0, + 567.0, + 1545.0, + 567.0, + 1579.0, + 292.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1545.0, + 813.0, + 1545.0, + 813.0, + 1579.0, + 628.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1545.0, + 1406.0, + 1545.0, + 1406.0, + 1579.0, + 872.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1576.0, + 799.0, + 1576.0, + 799.0, + 1608.0, + 295.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1609.0, + 558.0, + 1609.0, + 558.0, + 1649.0, + 292.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1609.0, + 1116.0, + 1609.0, + 1116.0, + 1649.0, + 624.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1609.0, + 1406.0, + 1609.0, + 1406.0, + 1649.0, + 1351.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1638.0, + 296.0, + 1638.0, + 296.0, + 1694.0, + 290.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1638.0, + 924.0, + 1638.0, + 924.0, + 1694.0, + 841.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1638.0, + 1093.0, + 1638.0, + 1093.0, + 1694.0, + 996.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1638.0, + 1150.0, + 1638.0, + 1150.0, + 1694.0, + 1121.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1638.0, + 1339.0, + 1638.0, + 1339.0, + 1694.0, + 1215.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1638.0, + 1374.0, + 1638.0, + 1374.0, + 1694.0, + 1359.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 527.0, + 2002.0, + 527.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 299.0, + 476.0, + 299.0, + 476.0, + 343.0, + 294.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 299.0, + 945.0, + 299.0, + 945.0, + 343.0, + 817.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 299.0, + 1115.0, + 299.0, + 1115.0, + 343.0, + 1005.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 299.0, + 1148.0, + 299.0, + 1148.0, + 343.0, + 1138.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 334.0, + 631.0, + 334.0, + 631.0, + 369.0, + 341.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 334.0, + 865.0, + 334.0, + 865.0, + 369.0, + 748.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 334.0, + 1031.0, + 334.0, + 1031.0, + 369.0, + 895.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 363.0, + 405.0, + 363.0, + 405.0, + 406.0, + 295.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 363.0, + 511.0, + 363.0, + 511.0, + 406.0, + 508.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 393.0, + 343.0, + 393.0, + 343.0, + 432.0, + 305.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 393.0, + 446.0, + 393.0, + 446.0, + 432.0, + 440.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 421.0, + 344.0, + 421.0, + 344.0, + 462.0, + 306.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 421.0, + 513.0, + 421.0, + 513.0, + 462.0, + 510.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 457.0, + 385.0, + 457.0, + 385.0, + 490.0, + 306.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 457.0, + 588.0, + 457.0, + 588.0, + 490.0, + 549.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 494.0, + 339.0, + 494.0, + 339.0, + 523.0, + 309.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 489.0, + 437.0, + 489.0, + 437.0, + 524.0, + 366.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 489.0, + 535.0, + 489.0, + 535.0, + 524.0, + 498.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 489.0, + 565.0, + 489.0, + 565.0, + 524.0, + 560.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 528.0, + 339.0, + 528.0, + 339.0, + 559.0, + 309.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 521.0, + 451.0, + 521.0, + 451.0, + 563.0, + 368.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 521.0, + 674.0, + 521.0, + 674.0, + 563.0, + 621.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 521.0, + 902.0, + 521.0, + 902.0, + 563.0, + 863.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 564.0, + 339.0, + 564.0, + 339.0, + 593.0, + 308.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 595.0, + 340.0, + 595.0, + 340.0, + 628.0, + 307.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 595.0, + 466.0, + 595.0, + 466.0, + 626.0, + 367.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 628.0, + 338.0, + 628.0, + 338.0, + 659.0, + 308.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 626.0, + 395.0, + 626.0, + 395.0, + 659.0, + 369.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 626.0, + 534.0, + 626.0, + 534.0, + 659.0, + 472.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 658.0, + 340.0, + 658.0, + 340.0, + 690.0, + 306.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 656.0, + 398.0, + 656.0, + 398.0, + 688.0, + 393.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 692.0, + 339.0, + 692.0, + 339.0, + 723.0, + 298.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 680.0, + 408.0, + 680.0, + 408.0, + 737.0, + 386.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 680.0, + 798.0, + 680.0, + 798.0, + 737.0, + 752.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 723.0, + 339.0, + 723.0, + 339.0, + 753.0, + 297.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 717.0, + 426.0, + 717.0, + 426.0, + 760.0, + 422.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 717.0, + 736.0, + 717.0, + 736.0, + 760.0, + 732.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 749.0, + 340.0, + 749.0, + 340.0, + 786.0, + 296.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 752.0, + 477.0, + 752.0, + 477.0, + 781.0, + 394.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 340.0, + 783.0, + 340.0, + 816.0, + 296.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 782.0, + 451.0, + 782.0, + 451.0, + 813.0, + 368.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 814.0, + 339.0, + 814.0, + 339.0, + 844.0, + 297.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 807.0, + 370.0, + 807.0, + 370.0, + 851.0, + 365.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 839.0, + 437.0, + 839.0, + 437.0, + 876.0, + 295.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 561.5, + 592.0, + 561.5, + 592.0, + 601.5, + 396.0, + 601.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 683, + 1405, + 683, + 1405, + 915, + 296, + 915 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 390, + 1405, + 390, + 1405, + 514, + 297, + 514 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1569, + 1404, + 1569, + 1404, + 1726, + 297, + 1726 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1730, + 1404, + 1730, + 1404, + 1840, + 297, + 1840 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 1270, + 1405, + 1270, + 1405, + 1395, + 296, + 1395 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1898, + 1404, + 1898, + 1404, + 2042, + 298, + 2042 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 1404, + 227, + 1404, + 367, + 297, + 367 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 919, + 1405, + 919, + 1405, + 1062, + 297, + 1062 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 294, + 528, + 1402, + 528, + 1402, + 592, + 294, + 592 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 295, + 1113, + 1402, + 1113, + 1402, + 1176, + 295, + 1176 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 298, + 1180, + 1399, + 1180, + 1399, + 1252, + 298, + 1252 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 298, + 1858, + 1123, + 1858, + 1123, + 1892, + 298, + 1892 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 301, + 627, + 788, + 627, + 788, + 660, + 301, + 660 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.906 + }, + { + "category_id": 8, + "poly": [ + 628, + 1059, + 1068, + 1059, + 1068, + 1098, + 628, + 1098 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.783 + }, + { + "category_id": 1, + "poly": [ + 300, + 1401, + 1403, + 1401, + 1403, + 1510, + 300, + 1510 + ], + "score": 0.71 + }, + { + "category_id": 1, + "poly": [ + 310, + 1513, + 1249, + 1513, + 1249, + 1553, + 310, + 1553 + ], + "score": 0.515 + }, + { + "category_id": 13, + "poly": [ + 686, + 561, + 769, + 561, + 769, + 593, + 686, + 593 + ], + "score": 0.95, + "latex": "( c _ { * } , c ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 1218, + 529, + 1282, + 529, + 1282, + 563, + 1218, + 563 + ], + "score": 0.93, + "latex": "( \\alpha , \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 1022, + 850, + 1295, + 850, + 1295, + 887, + 1022, + 887 + ], + "score": 0.93, + "latex": "y _ { f ( \\pmb { x } ) } : = \\mathbb { I } ( f ( \\pmb { x } ) \\breve { \\geq } 1 / 2 )" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1762, + 1241, + 1762, + 1241, + 1804, + 1016, + 1804 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 299, + 1027, + 450, + 1027, + 450, + 1066, + 299, + 1066 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 614, + 2003, + 1046, + 2003, + 1046, + 2039, + 614, + 2039 + ], + "score": 0.93, + "latex": "\\mathbb { P } _ { \\mathbf { x } \\sim D } [ y _ { f _ { f i n a l } ( \\mathbf { x } ) } = \\eta ^ { * } ( \\pmb { x } ) ] \\geq 1 - 3 c ^ { * } \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 376, + 1691, + 454, + 1691, + 454, + 1729, + 376, + 1729 + ], + "score": 0.93, + "latex": "\\mathbf { \\frac { 1 } { 2 } } - e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 483, + 1802, + 632, + 1802, + 632, + 1840, + 483, + 1840 + ], + "score": 0.93, + "latex": "L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 447, + 1218, + 581, + 1218, + 581, + 1251, + 447, + 1251 + ], + "score": 0.93, + "latex": "\\pmb { x } \\in L ( e , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 1234, + 1731, + 1328, + 1731, + 1328, + 1765, + 1234, + 1765 + ], + "score": 0.93, + "latex": "L ( e _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 910, + 986, + 1147, + 986, + 1147, + 1030, + 910, + 1030 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { T _ { 0 } } { 3 \\varepsilon } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1040, + 228, + 1256, + 228, + 1256, + 267, + 1040, + 267 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { t ( { \\pmb x } ) = | \\eta ( { \\pmb x } ) - \\frac { 1 } { 2 } | } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 589, + 299, + 672, + 299, + 672, + 330, + 589, + 330 + ], + "score": 0.92, + "latex": "( c _ { * } , c ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 299, + 2000, + 450, + 2000, + 450, + 2040, + 299, + 2040 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 544, + 1765, + 658, + 1765, + 658, + 1799, + 544, + 1799 + ], + "score": 0.92, + "latex": "T _ { 0 } < 1 / 2" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1960, + 1310, + 1960, + 1310, + 2000, + 1127, + 2000 + ], + "score": 0.92, + "latex": "T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 509, + 1182, + 599, + 1182, + 599, + 1217, + 509, + 1217 + ], + "score": 0.92, + "latex": "( e , f , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 787, + 1180, + 1139, + 1180, + 1139, + 1220, + 787, + 1220 + ], + "score": 0.92, + "latex": "L ( e , \\eta ) : = \\{ \\pmb { x } | | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e \\}" + }, + { + "category_id": 13, + "poly": [ + 551, + 1433, + 641, + 1433, + 641, + 1466, + 551, + 1466 + ], + "score": 0.91, + "latex": "( e , f , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 460, + 746, + 1007, + 746, + 1007, + 793, + 460, + 793 + ], + "score": 0.91, + "latex": "f ( \\pmb { x } ) = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1929, + 393, + 1929, + 393, + 1962, + 298, + 1962 + ], + "score": 0.91, + "latex": "L ( e _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 297, + 950, + 468, + 950, + 468, + 989, + 297, + 989 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 295, + 332, + 379, + 332, + 379, + 366, + 295, + 366 + ], + "score": 0.91, + "latex": "( c _ { * } , c ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 509, + 1961, + 742, + 1961, + 742, + 2001, + 509, + 2001 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1962, + 454, + 1962, + 454, + 1999, + 297, + 1999 + ], + "score": 0.91, + "latex": "T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1217, + 367, + 1217, + 367, + 1251, + 297, + 1251 + ], + "score": 0.91, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 876, + 265, + 1028, + 265, + 1028, + 298, + 876, + 298 + ], + "score": 0.91, + "latex": "g ( t ) = G ^ { \\prime } ( t )" + }, + { + "category_id": 14, + "poly": [ + 627, + 1056, + 1071, + 1056, + 1071, + 1098, + 627, + 1098 + ], + "score": 0.9, + "latex": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 795, + 1432, + 927, + 1432, + 927, + 1468, + 795, + 1468 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1310, + 229, + 1366, + 229, + 1366, + 264, + 1310, + 264 + ], + "score": 0.9, + "latex": "G ( t )" + }, + { + "category_id": 13, + "poly": [ + 910, + 1514, + 1242, + 1514, + 1242, + 1552, + 910, + 1552 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 711, + 531, + 828, + 531, + 828, + 561, + 711, + 561 + ], + "score": 0.89, + "latex": "\\alpha , \\epsilon , c _ { * } , c ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1135, + 330, + 1232, + 330, + 1232, + 371, + 1135, + 371 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\ell _ { D } . = \\frac { c ^ { * } } { c _ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 445, + 263, + 810, + 263, + 810, + 300, + 445, + 300 + ], + "score": 0.89, + "latex": "G ( t ) = \\mathbb { P } _ { \\mathbf { x } \\sim D } ( | \\eta ( \\mathbf { x } ) - \\frac { 1 } { 2 } | \\leq t )" + }, + { + "category_id": 13, + "poly": [ + 1196, + 989, + 1315, + 989, + 1315, + 1023, + 1196, + 1023 + ], + "score": 0.88, + "latex": "T _ { e n d } \\leq 3 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 761, + 1515, + 831, + 1515, + 831, + 1550, + 761, + 1550 + ], + "score": 0.88, + "latex": "f _ { n e w } \\}" + }, + { + "category_id": 13, + "poly": [ + 605, + 1694, + 636, + 1694, + 636, + 1724, + 605, + 1724 + ], + "score": 0.88, + "latex": "T _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 798, + 1961, + 1071, + 1961, + 1071, + 2002, + 798, + 2002 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1900, + 1143, + 1900, + 1143, + 1931, + 1114, + 1931 + ], + "score": 0.87, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1067, + 1931, + 1095, + 1931, + 1095, + 1961, + 1067, + 1961 + ], + "score": 0.87, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 983, + 1433, + 1402, + 1433, + 1402, + 1468, + 983, + 1468 + ], + "score": 0.87, + "latex": "\\widetilde { \\eta } _ { n e w } ( \\pmb { x } ) = y _ { f ( \\pmb { x } ) } i f \\vert f ( \\pmb { x } ) - 1 / 2 \\vert \\geq e" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1118, + 1285, + 1118, + 1285, + 1146, + 1260, + 1146 + ], + "score": 0.86, + "latex": "t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 424, + 1766, + 452, + 1766, + 452, + 1798, + 424, + 1798 + ], + "score": 0.85, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 348, + 1466, + 635, + 1466, + 635, + 1501, + 348, + 1501 + ], + "score": 0.85, + "latex": "\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e" + }, + { + "category_id": 13, + "poly": [ + 1337, + 924, + 1403, + 924, + 1403, + 952, + 1337, + 952 + ], + "score": 0.85, + "latex": "e _ { 0 } =" + }, + { + "category_id": 13, + "poly": [ + 957, + 1733, + 985, + 1733, + 985, + 1764, + 957, + 1764 + ], + "score": 0.85, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 789, + 1469, + 1344, + 1469, + 1344, + 1511, + 789, + 1511 + ], + "score": 0.85, + "latex": "f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )" + }, + { + "category_id": 13, + "poly": [ + 808, + 1805, + 836, + 1805, + 836, + 1837, + 808, + 1837 + ], + "score": 0.85, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 442, + 883, + 461, + 883, + 461, + 914, + 442, + 914 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 784, + 852, + 803, + 852, + 803, + 882, + 784, + 882 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 598, + 1931, + 625, + 1931, + 625, + 1961, + 598, + 1961 + ], + "score": 0.84, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 635, + 985, + 860, + 985, + 860, + 1028, + 635, + 1028 + ], + "score": 0.83, + "latex": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1161, + 531, + 1188, + 531, + 1188, + 558, + 1161, + 558 + ], + "score": 0.83, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 935, + 952, + 955, + 952, + 955, + 983, + 935, + 983 + ], + "score": 0.83, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 629, + 561, + 655, + 561, + 655, + 587, + 629, + 587 + ], + "score": 0.83, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1346, + 454, + 1373, + 454, + 1373, + 480, + 1346, + 480 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1218, + 923, + 1244, + 923, + 1244, + 951, + 1218, + 951 + ], + "score": 0.81, + "latex": "t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 402, + 987, + 586, + 987, + 586, + 1026, + 402, + 1026 + ], + "score": 0.81, + "latex": "( l ) T _ { 0 } < { \\frac { 1 } { 2 } } - e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1001, + 1121, + 1020, + 1121, + 1020, + 1142, + 1001, + 1142 + ], + "score": 0.79, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 890, + 298, + 1045, + 298, + 1045, + 330, + 890, + 330 + ], + "score": 0.79, + "latex": "0 \\leq t \\leq 1 / 2" + }, + { + "category_id": 13, + "poly": [ + 1326, + 1404, + 1345, + 1404, + 1345, + 1434, + 1326, + 1434 + ], + "score": 0.79, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1309, + 1185, + 1405, + 1185, + 1405, + 1220, + 1309, + 1220 + ], + "score": 0.78, + "latex": "y _ { f ( \\pmb { x } ) } =" + }, + { + "category_id": 13, + "poly": [ + 1347, + 298, + 1373, + 298, + 1373, + 325, + 1347, + 325 + ], + "score": 0.78, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 297, + 1514, + 537, + 1514, + 537, + 1549, + 297, + 1549 + ], + "score": 0.76, + "latex": "e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0" + }, + { + "category_id": 13, + "poly": [ + 868, + 927, + 888, + 927, + 888, + 948, + 868, + 948 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 527, + 299, + 554, + 299, + 554, + 325, + 527, + 325 + ], + "score": 0.75, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1059, + 297, + 1301, + 297, + 1301, + 330, + 1059, + 330 + ], + "score": 0.74, + "latex": "0 < c _ { * } \\leq g ( t ) \\leq c ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 552, + 1514, + 636, + 1514, + 636, + 1549, + 552, + 1549 + ], + "score": 0.68, + "latex": "L ( e , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 1333, + 270, + 1346, + 270, + 1346, + 294, + 1333, + 294 + ], + "score": 0.65, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1073, + 333, + 1099, + 333, + 1099, + 361, + 1073, + 361 + ], + "score": 0.62, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 922, + 1607, + 950, + 1607, + 950, + 1629, + 922, + 1629 + ], + "score": 0.58, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 674, + 923, + 690, + 923, + 690, + 948, + 674, + 948 + ], + "score": 0.55, + "latex": "^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1184, + 1284, + 1184, + 1284, + 1216, + 1263, + 1216 + ], + "score": 0.53, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 348, + 1466, + 404, + 1466, + 404, + 1500, + 348, + 1500 + ], + "score": 0.5, + "latex": "\\widetilde { \\eta } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 859, + 953, + 875, + 953, + 875, + 978, + 859, + 978 + ], + "score": 0.49, + "latex": "^ { l }" + }, + { + "category_id": 13, + "poly": [ + 414, + 269, + 428, + 269, + 428, + 294, + 414, + 294 + ], + "score": 0.47, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 674, + 1901, + 690, + 1901, + 690, + 1926, + 674, + 1926 + ], + "score": 0.45, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 649, + 484, + 664, + 484, + 664, + 510, + 649, + 510 + ], + "score": 0.45, + "latex": "\\ell" + }, + { + "category_id": 13, + "poly": [ + 1366, + 1930, + 1403, + 1930, + 1403, + 1961, + 1366, + 1961 + ], + "score": 0.44, + "latex": "( l )" + }, + { + "category_id": 13, + "poly": [ + 895, + 1932, + 910, + 1932, + 910, + 1957, + 895, + 1957 + ], + "score": 0.34, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1405, + 1031, + 1405, + 1031, + 1430, + 1016, + 1430 + ], + "score": 0.29, + "latex": "^ { l }" + }, + { + "category_id": 13, + "poly": [ + 723, + 1185, + 745, + 1185, + 745, + 1211, + 723, + 1211 + ], + "score": 0.28, + "latex": "A" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 625.0, + 792.0, + 625.0, + 792.0, + 662.0, + 293.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 682.0, + 1403.0, + 682.0, + 1403.0, + 720.0, + 294.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 716.0, + 1403.0, + 716.0, + 1403.0, + 748.0, + 296.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 743.0, + 459.0, + 743.0, + 459.0, + 796.0, + 291.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 743.0, + 1407.0, + 743.0, + 1407.0, + 796.0, + 1008.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 787.0, + 1405.0, + 787.0, + 1405.0, + 826.0, + 293.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 855.0, + 294.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 847.0, + 783.0, + 847.0, + 783.0, + 891.0, + 291.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 847.0, + 1021.0, + 847.0, + 1021.0, + 891.0, + 804.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 847.0, + 1408.0, + 847.0, + 1408.0, + 891.0, + 1296.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 881.0, + 441.0, + 881.0, + 441.0, + 916.0, + 292.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 881.0, + 474.0, + 881.0, + 474.0, + 916.0, + 462.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 391.0, + 1403.0, + 391.0, + 1403.0, + 424.0, + 295.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 422.0, + 1405.0, + 422.0, + 1405.0, + 456.0, + 294.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 453.0, + 1345.0, + 453.0, + 1345.0, + 486.0, + 295.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 453.0, + 1405.0, + 453.0, + 1405.0, + 486.0, + 1374.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 484.0, + 648.0, + 484.0, + 648.0, + 516.0, + 295.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 484.0, + 1023.0, + 484.0, + 1023.0, + 516.0, + 665.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1569.0, + 1404.0, + 1569.0, + 1404.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1602.0, + 921.0, + 1602.0, + 921.0, + 1635.0, + 296.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1602.0, + 1404.0, + 1602.0, + 1404.0, + 1635.0, + 951.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1631.0, + 1405.0, + 1631.0, + 1405.0, + 1665.0, + 295.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1657.0, + 1405.0, + 1657.0, + 1405.0, + 1702.0, + 291.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1692.0, + 375.0, + 1692.0, + 375.0, + 1733.0, + 292.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1692.0, + 604.0, + 1692.0, + 604.0, + 1733.0, + 455.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1692.0, + 648.0, + 1692.0, + 648.0, + 1733.0, + 637.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1730.0, + 423.0, + 1730.0, + 423.0, + 1822.0, + 283.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1730.0, + 543.0, + 1730.0, + 543.0, + 1822.0, + 453.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1730.0, + 956.0, + 1730.0, + 956.0, + 1822.0, + 659.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1730.0, + 1015.0, + 1730.0, + 1015.0, + 1822.0, + 986.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1822.0, + 1329.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 482.0, + 1802.0, + 482.0, + 1843.0, + 294.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1802.0, + 807.0, + 1802.0, + 807.0, + 1843.0, + 633.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1802.0, + 847.0, + 1802.0, + 847.0, + 1843.0, + 837.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.25, + 1759.0, + 1219.25, + 1759.0, + 1219.25, + 1788.0, + 1174.25, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1273.0, + 1405.0, + 1273.0, + 1405.0, + 1306.0, + 295.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1304.0, + 1403.0, + 1304.0, + 1403.0, + 1337.0, + 295.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1332.0, + 1406.0, + 1332.0, + 1406.0, + 1369.0, + 293.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1363.0, + 569.0, + 1363.0, + 569.0, + 1398.0, + 292.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1894.0, + 673.0, + 1894.0, + 673.0, + 1934.0, + 292.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1894.0, + 1113.0, + 1894.0, + 1113.0, + 1934.0, + 691.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1894.0, + 1409.0, + 1894.0, + 1409.0, + 1934.0, + 1144.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1929.0, + 296.0, + 1929.0, + 296.0, + 2018.0, + 278.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1929.0, + 508.0, + 1929.0, + 508.0, + 2018.0, + 455.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1929.0, + 797.0, + 1929.0, + 797.0, + 2018.0, + 743.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1929.0, + 1126.0, + 1929.0, + 1126.0, + 2018.0, + 1096.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1929.0, + 1365.0, + 1929.0, + 1365.0, + 2018.0, + 1311.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1929.0, + 1420.0, + 1929.0, + 1420.0, + 2018.0, + 1404.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1985.0, + 298.0, + 1985.0, + 298.0, + 2058.0, + 284.0, + 2058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1985.0, + 613.0, + 1985.0, + 613.0, + 2058.0, + 451.0, + 2058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1985.0, + 1068.0, + 1985.0, + 1068.0, + 2058.0, + 1047.0, + 2058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 223.0, + 1039.0, + 223.0, + 1039.0, + 270.0, + 291.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 223.0, + 1309.0, + 223.0, + 1309.0, + 270.0, + 1257.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 223.0, + 1406.0, + 223.0, + 1406.0, + 270.0, + 1367.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 265.0, + 413.0, + 265.0, + 413.0, + 303.0, + 294.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 265.0, + 444.0, + 265.0, + 444.0, + 303.0, + 429.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 265.0, + 875.0, + 265.0, + 875.0, + 303.0, + 811.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 265.0, + 1332.0, + 265.0, + 1332.0, + 303.0, + 1029.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 265.0, + 1405.0, + 265.0, + 1405.0, + 303.0, + 1347.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 297.0, + 526.0, + 297.0, + 526.0, + 332.0, + 294.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 297.0, + 588.0, + 297.0, + 588.0, + 332.0, + 555.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 297.0, + 889.0, + 297.0, + 889.0, + 332.0, + 673.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 297.0, + 1058.0, + 297.0, + 1058.0, + 332.0, + 1046.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 297.0, + 1346.0, + 297.0, + 1346.0, + 332.0, + 1302.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 297.0, + 1405.0, + 297.0, + 1405.0, + 332.0, + 1374.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 330.0, + 1072.0, + 330.0, + 1072.0, + 368.0, + 380.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 330.0, + 1134.0, + 330.0, + 1134.0, + 368.0, + 1100.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 327.5, + 1243.0, + 327.5, + 1243.0, + 374.0, + 1189.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 913.0, + 673.0, + 913.0, + 673.0, + 958.0, + 292.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 913.0, + 867.0, + 913.0, + 867.0, + 958.0, + 691.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 913.0, + 1217.0, + 913.0, + 1217.0, + 958.0, + 889.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 913.0, + 1336.0, + 913.0, + 1336.0, + 958.0, + 1245.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 935.0, + 296.0, + 935.0, + 296.0, + 1002.0, + 285.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 935.0, + 858.0, + 935.0, + 858.0, + 1002.0, + 469.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 935.0, + 934.0, + 935.0, + 934.0, + 1002.0, + 876.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 935.0, + 1417.0, + 935.0, + 1417.0, + 1002.0, + 956.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 974.0, + 401.0, + 974.0, + 401.0, + 1044.0, + 286.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 974.0, + 634.0, + 974.0, + 634.0, + 1044.0, + 587.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 974.0, + 909.0, + 974.0, + 909.0, + 1044.0, + 861.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 974.0, + 1195.0, + 974.0, + 1195.0, + 1044.0, + 1148.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 974.0, + 1415.0, + 974.0, + 1415.0, + 1044.0, + 1316.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 298.0, + 1022.0, + 298.0, + 1074.0, + 292.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1022.0, + 574.0, + 1022.0, + 574.0, + 1074.0, + 451.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 528.0, + 710.0, + 528.0, + 710.0, + 564.0, + 296.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 528.0, + 1160.0, + 528.0, + 1160.0, + 564.0, + 829.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 528.0, + 1217.0, + 528.0, + 1217.0, + 564.0, + 1189.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 528.0, + 1405.0, + 528.0, + 1405.0, + 564.0, + 1283.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 557.0, + 628.0, + 557.0, + 628.0, + 593.0, + 294.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 557.0, + 685.0, + 557.0, + 685.0, + 593.0, + 656.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 557.0, + 886.0, + 557.0, + 886.0, + 593.0, + 770.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1108.0, + 1000.0, + 1108.0, + 1000.0, + 1152.0, + 293.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1108.0, + 1259.0, + 1108.0, + 1259.0, + 1152.0, + 1021.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1108.0, + 1406.0, + 1108.0, + 1406.0, + 1152.0, + 1286.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1143.0, + 777.0, + 1143.0, + 777.0, + 1179.0, + 295.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1173.0, + 508.0, + 1173.0, + 508.0, + 1227.0, + 292.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1173.0, + 722.0, + 1173.0, + 722.0, + 1227.0, + 600.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1173.0, + 786.0, + 1173.0, + 786.0, + 1227.0, + 746.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1173.0, + 1262.0, + 1173.0, + 1262.0, + 1227.0, + 1140.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1173.0, + 1308.0, + 1173.0, + 1308.0, + 1227.0, + 1285.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1216.0, + 446.0, + 1216.0, + 446.0, + 1255.0, + 368.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1216.0, + 592.0, + 1216.0, + 592.0, + 1255.0, + 582.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1857.0, + 1126.0, + 1857.0, + 1126.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1401.0, + 1015.0, + 1401.0, + 1015.0, + 1437.0, + 295.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1401.0, + 1325.0, + 1401.0, + 1325.0, + 1437.0, + 1032.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1401.0, + 1405.0, + 1401.0, + 1405.0, + 1437.0, + 1346.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1429.0, + 550.0, + 1429.0, + 550.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1429.0, + 794.0, + 1429.0, + 794.0, + 1470.0, + 642.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1429.0, + 982.0, + 1429.0, + 982.0, + 1470.0, + 928.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1429.0, + 1407.0, + 1429.0, + 1407.0, + 1470.0, + 1403.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1462.0, + 347.0, + 1462.0, + 347.0, + 1517.0, + 291.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1462.0, + 788.0, + 1462.0, + 788.0, + 1517.0, + 636.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1462.0, + 1409.0, + 1462.0, + 1409.0, + 1517.0, + 1345.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1508.0, + 551.0, + 1508.0, + 551.0, + 1558.0, + 538.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1508.0, + 760.0, + 1508.0, + 760.0, + 1558.0, + 637.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1508.0, + 909.0, + 1508.0, + 909.0, + 1558.0, + 832.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1508.0, + 1251.0, + 1508.0, + 1251.0, + 1558.0, + 1243.0, + 1558.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1343, + 1405, + 1343, + 1405, + 1591, + 296, + 1591 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 525, + 1406, + 525, + 1406, + 774, + 298, + 774 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 864, + 1405, + 864, + 1405, + 1050, + 298, + 1050 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 415, + 298, + 415 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1603, + 1404, + 1603, + 1404, + 1759, + 298, + 1759 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1771, + 1403, + 1771, + 1403, + 1897, + 298, + 1897 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1204, + 1403, + 1204, + 1403, + 1329, + 298, + 1329 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 299, + 786, + 1398, + 786, + 1398, + 850, + 299, + 850 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 300, + 458, + 557, + 458, + 557, + 493, + 300, + 493 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.898 + }, + { + "category_id": 8, + "poly": [ + 313, + 1062, + 1390, + 1062, + 1390, + 1194, + 313, + 1194 + ], + "score": 0.832 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2111, + 841, + 2111 + ], + "score": 0.753 + }, + { + "category_id": 1, + "poly": [ + 313, + 1062, + 1390, + 1062, + 1390, + 1194, + 313, + 1194 + ], + "score": 0.179 + }, + { + "category_id": 13, + "poly": [ + 486, + 650, + 609, + 650, + 609, + 683, + 486, + 683 + ], + "score": 0.93, + "latex": "y _ { \\pmb { x } } \\sim \\eta ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 975, + 741, + 1040, + 741, + 1040, + 774, + 975, + 774 + ], + "score": 0.93, + "latex": "\\dot { \\eta _ { i } } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 850, + 1558, + 941, + 1558, + 941, + 1592, + 850, + 1592 + ], + "score": 0.93, + "latex": "T _ { i j } = \\tau" + }, + { + "category_id": 14, + "poly": [ + 312, + 1060, + 1387, + 1060, + 1387, + 1197, + 312, + 1197 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathrm { T y p e - I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = - \\cfrac { 1 } { 2 } \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\cfrac { 1 } { 2 } , \\quad \\mathrm { T y p e - I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } , } \\\\ & { \\mathrm { T y p e - I I I : } \\tau _ { u _ { \\alpha } , s _ { \\alpha } } = 1 - \\cfrac { 1 } { 3 } \\left[ \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 3 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] ^ { 2 } + \\left[ \\eta _ { u _ { \\alpha } } ( { \\pmb x } ) - \\eta _ { s _ { \\alpha } } ( { \\pmb x } ) \\right] \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 747, + 711, + 1004, + 711, + 1004, + 744, + 747, + 744 + ], + "score": 0.93, + "latex": "\\eta ^ { * } : \\mathcal { X } \\{ 1 , \\cdot \\cdot , C \\}" + }, + { + "category_id": 13, + "poly": [ + 827, + 1205, + 884, + 1205, + 884, + 1239, + 827, + 1239 + ], + "score": 0.93, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 991, + 1558, + 1128, + 1558, + 1128, + 1588, + 991, + 1588 + ], + "score": 0.92, + "latex": "T _ { i i } = 1 - \\tau" + }, + { + "category_id": 13, + "poly": [ + 1167, + 1496, + 1361, + 1496, + 1361, + 1530, + 1167, + 1530 + ], + "score": 0.92, + "latex": "T _ { i j } = \\tau / ( C - 1 )" + }, + { + "category_id": 13, + "poly": [ + 418, + 1527, + 553, + 1527, + 553, + 1558, + 418, + 1558 + ], + "score": 0.92, + "latex": "T _ { i i } = 1 - \\tau" + }, + { + "category_id": 13, + "poly": [ + 650, + 741, + 707, + 741, + 707, + 774, + 650, + 774 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1011, + 927, + 1068, + 927, + 1068, + 960, + 1011, + 960 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 482, + 1405, + 802, + 1405, + 802, + 1439, + 482, + 1439 + ], + "score": 0.92, + "latex": "T _ { i j } = P ( \\widetilde { y } = j | y = i ) = \\tau _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 403, + 957, + 473, + 957, + 473, + 989, + 403, + 989 + ], + "score": 0.91, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 455, + 988, + 511, + 988, + 511, + 1021, + 455, + 1021 + ], + "score": 0.91, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1067, + 680, + 1125, + 680, + 1125, + 713, + 1067, + 713 + ], + "score": 0.91, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1229, + 1407, + 1294, + 1407, + 1294, + 1437, + 1229, + 1437 + ], + "score": 0.91, + "latex": "y = i" + }, + { + "category_id": 13, + "poly": [ + 695, + 1240, + 766, + 1240, + 766, + 1270, + 695, + 1270 + ], + "score": 0.9, + "latex": "\\tau _ { u _ { x } , s _ { x } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1528, + 360, + 1528, + 360, + 1559, + 298, + 1559 + ], + "score": 0.89, + "latex": "i \\neq j" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1267, + 1189, + 1267, + 1189, + 1296, + 1135, + 1296 + ], + "score": 0.89, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 536, + 1943, + 592, + 1943, + 592, + 1972, + 536, + 1972 + ], + "score": 0.89, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1314, + 992, + 1347, + 992, + 1347, + 1018, + 1314, + 1018 + ], + "score": 0.87, + "latex": "u _ { x }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 993, + 1285, + 993, + 1285, + 1020, + 1252, + 1020 + ], + "score": 0.87, + "latex": "y _ { x }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1298, + 400, + 1298, + 400, + 1327, + 345, + 1327 + ], + "score": 0.87, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1267, + 1291, + 1267, + 1291, + 1296, + 1239, + 1296 + ], + "score": 0.87, + "latex": "70 \\%" + }, + { + "category_id": 13, + "poly": [ + 1342, + 960, + 1374, + 960, + 1374, + 988, + 1342, + 988 + ], + "score": 0.87, + "latex": "y _ { x }" + }, + { + "category_id": 13, + "poly": [ + 646, + 1943, + 701, + 1943, + 701, + 1972, + 646, + 1972 + ], + "score": 0.87, + "latex": "70 \\%" + }, + { + "category_id": 13, + "poly": [ + 1186, + 930, + 1218, + 930, + 1218, + 957, + 1186, + 957 + ], + "score": 0.87, + "latex": "s _ { x }" + }, + { + "category_id": 13, + "poly": [ + 830, + 1022, + 863, + 1022, + 863, + 1048, + 830, + 1048 + ], + "score": 0.86, + "latex": "u _ { x }" + }, + { + "category_id": 13, + "poly": [ + 653, + 1022, + 684, + 1022, + 684, + 1048, + 653, + 1048 + ], + "score": 0.86, + "latex": "s _ { x }" + }, + { + "category_id": 13, + "poly": [ + 368, + 685, + 400, + 685, + 400, + 712, + 368, + 712 + ], + "score": 0.86, + "latex": "y _ { x }" + }, + { + "category_id": 13, + "poly": [ + 450, + 1298, + 504, + 1298, + 504, + 1327, + 450, + 1327 + ], + "score": 0.86, + "latex": "70 \\%" + }, + { + "category_id": 13, + "poly": [ + 402, + 931, + 436, + 931, + 436, + 957, + 402, + 957 + ], + "score": 0.85, + "latex": "u _ { x }" + }, + { + "category_id": 13, + "poly": [ + 818, + 932, + 849, + 932, + 849, + 957, + 818, + 957 + ], + "score": 0.85, + "latex": "{ s } _ { x }" + }, + { + "category_id": 13, + "poly": [ + 410, + 1560, + 426, + 1560, + 426, + 1590, + 410, + 1590 + ], + "score": 0.85, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1092, + 713, + 1117, + 713, + 1117, + 738, + 1092, + 738 + ], + "score": 0.84, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 375, + 1407, + 398, + 1407, + 398, + 1433, + 375, + 1433 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 326, + 1438, + 341, + 1438, + 341, + 1467, + 326, + 1467 + ], + "score": 0.79, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 664, + 1499, + 676, + 1499, + 676, + 1524, + 664, + 1524 + ], + "score": 0.79, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 983, + 594, + 1002, + 594, + 1002, + 621, + 983, + 621 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1363, + 1530, + 1375, + 1530, + 1375, + 1553, + 1363, + 1553 + ], + "score": 0.78, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 637, + 1532, + 657, + 1532, + 657, + 1554, + 637, + 1554 + ], + "score": 0.78, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 707, + 1440, + 719, + 1440, + 719, + 1463, + 707, + 1463 + ], + "score": 0.76, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1110, + 743, + 1123, + 743, + 1123, + 768, + 1110, + 768 + ], + "score": 0.76, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 432, + 1024, + 452, + 1024, + 452, + 1045, + 432, + 1045 + ], + "score": 0.75, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 1212, + 327, + 1229, + 327, + 1229, + 350, + 1212, + 350 + ], + "score": 0.72, + "latex": "\\varepsilon" + }, + { + "category_id": 13, + "poly": [ + 820, + 901, + 842, + 901, + 842, + 923, + 820, + 923 + ], + "score": 0.72, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 582, + 1469, + 595, + 1469, + 595, + 1494, + 582, + 1494 + ], + "score": 0.71, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 794, + 1468, + 818, + 1468, + 818, + 1494, + 794, + 1494 + ], + "score": 0.7, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 810, + 655, + 831, + 655, + 831, + 678, + 810, + 678 + ], + "score": 0.65, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 694, + 1639, + 712, + 1639, + 712, + 1663, + 694, + 1663 + ], + "score": 0.33, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 455.0, + 561.0, + 455.0, + 561.0, + 499.0, + 293.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1378.0, + 295.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1409.0, + 294.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1403.0, + 374.0, + 1403.0, + 374.0, + 1444.0, + 291.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1403.0, + 481.0, + 1403.0, + 481.0, + 1444.0, + 399.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 1403.0, + 1228.0, + 1403.0, + 1228.0, + 1444.0, + 803.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1403.0, + 1407.0, + 1403.0, + 1407.0, + 1444.0, + 1295.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1434.0, + 325.0, + 1434.0, + 325.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1434.0, + 706.0, + 1434.0, + 706.0, + 1471.0, + 342.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1471.0, + 720.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 581.0, + 1467.0, + 581.0, + 1501.0, + 294.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1467.0, + 793.0, + 1467.0, + 793.0, + 1501.0, + 596.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1467.0, + 1408.0, + 1467.0, + 1408.0, + 1501.0, + 819.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1497.0, + 663.0, + 1497.0, + 663.0, + 1531.0, + 294.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1497.0, + 1166.0, + 1497.0, + 1166.0, + 1531.0, + 677.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1497.0, + 1407.0, + 1497.0, + 1407.0, + 1531.0, + 1362.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1526.0, + 297.0, + 1526.0, + 297.0, + 1560.0, + 293.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1526.0, + 417.0, + 1526.0, + 417.0, + 1560.0, + 361.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1526.0, + 636.0, + 1526.0, + 636.0, + 1560.0, + 554.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1526.0, + 1362.0, + 1526.0, + 1362.0, + 1560.0, + 658.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1526.0, + 1407.0, + 1526.0, + 1407.0, + 1560.0, + 1376.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1559.0, + 409.0, + 1559.0, + 409.0, + 1593.0, + 295.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1559.0, + 849.0, + 1559.0, + 849.0, + 1593.0, + 427.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1559.0, + 990.0, + 1559.0, + 990.0, + 1593.0, + 942.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1559.0, + 1281.0, + 1559.0, + 1281.0, + 1593.0, + 1129.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 527.0, + 1408.0, + 527.0, + 1408.0, + 561.0, + 295.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 560.0, + 1404.0, + 560.0, + 1404.0, + 590.0, + 295.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 589.0, + 982.0, + 589.0, + 982.0, + 623.0, + 295.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 589.0, + 1406.0, + 589.0, + 1406.0, + 623.0, + 1003.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 619.0, + 1408.0, + 619.0, + 1408.0, + 653.0, + 295.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 648.0, + 485.0, + 648.0, + 485.0, + 686.0, + 293.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 648.0, + 809.0, + 648.0, + 809.0, + 686.0, + 610.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 648.0, + 1407.0, + 648.0, + 1407.0, + 686.0, + 832.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 677.0, + 367.0, + 677.0, + 367.0, + 715.0, + 293.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 677.0, + 1066.0, + 677.0, + 1066.0, + 715.0, + 401.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 677.0, + 1406.0, + 677.0, + 1406.0, + 715.0, + 1126.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 708.0, + 746.0, + 708.0, + 746.0, + 746.0, + 291.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 708.0, + 1091.0, + 708.0, + 1091.0, + 746.0, + 1005.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 708.0, + 1411.0, + 708.0, + 1411.0, + 746.0, + 1118.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 739.0, + 649.0, + 739.0, + 649.0, + 777.0, + 293.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 739.0, + 974.0, + 739.0, + 974.0, + 777.0, + 708.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 739.0, + 1109.0, + 739.0, + 1109.0, + 777.0, + 1041.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 739.0, + 1411.0, + 739.0, + 1411.0, + 777.0, + 1124.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 898.0, + 295.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 894.0, + 819.0, + 894.0, + 819.0, + 929.0, + 293.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 894.0, + 1406.0, + 894.0, + 1406.0, + 929.0, + 843.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 926.0, + 401.0, + 926.0, + 401.0, + 962.0, + 295.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 926.0, + 817.0, + 926.0, + 817.0, + 962.0, + 437.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 926.0, + 1010.0, + 926.0, + 1010.0, + 962.0, + 850.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 926.0, + 1185.0, + 926.0, + 1185.0, + 962.0, + 1069.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 926.0, + 1407.0, + 926.0, + 1407.0, + 962.0, + 1219.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 956.0, + 402.0, + 956.0, + 402.0, + 991.0, + 295.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 956.0, + 1341.0, + 956.0, + 1341.0, + 991.0, + 474.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 956.0, + 1405.0, + 956.0, + 1405.0, + 991.0, + 1375.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 988.0, + 454.0, + 988.0, + 454.0, + 1023.0, + 295.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 988.0, + 1251.0, + 988.0, + 1251.0, + 1023.0, + 512.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 988.0, + 1313.0, + 988.0, + 1313.0, + 1023.0, + 1286.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 988.0, + 1407.0, + 988.0, + 1407.0, + 1023.0, + 1348.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1017.0, + 431.0, + 1017.0, + 431.0, + 1051.0, + 293.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1017.0, + 652.0, + 1017.0, + 652.0, + 1051.0, + 453.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1017.0, + 829.0, + 1017.0, + 829.0, + 1051.0, + 685.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1017.0, + 1333.0, + 1017.0, + 1333.0, + 1051.0, + 864.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 321.0, + 1211.0, + 321.0, + 1211.0, + 356.0, + 291.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 321.0, + 1410.0, + 321.0, + 1410.0, + 356.0, + 1230.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 353.0, + 1406.0, + 353.0, + 1406.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 383.0, + 1326.0, + 383.0, + 1326.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1598.0, + 1406.0, + 1598.0, + 1406.0, + 1644.0, + 292.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1629.0, + 693.0, + 1629.0, + 693.0, + 1674.0, + 291.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1674.0, + 713.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1661.0, + 1406.0, + 1661.0, + 1406.0, + 1702.0, + 292.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1692.0, + 1405.0, + 1692.0, + 1405.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 835.0, + 1726.0, + 835.0, + 1762.0, + 293.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1805.0, + 1408.0, + 1805.0, + 1408.0, + 1835.0, + 293.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1407.0, + 1835.0, + 1407.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1862.0, + 1043.0, + 1862.0, + 1043.0, + 1900.0, + 292.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 535.0, + 1943.0, + 535.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1943.0, + 645.0, + 1943.0, + 645.0, + 1976.0, + 593.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1976.0, + 702.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 826.0, + 1206.0, + 826.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1206.0, + 1403.0, + 1206.0, + 1403.0, + 1238.0, + 885.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1234.0, + 694.0, + 1234.0, + 694.0, + 1277.0, + 291.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1234.0, + 1409.0, + 1234.0, + 1409.0, + 1277.0, + 767.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1264.0, + 1134.0, + 1264.0, + 1134.0, + 1303.0, + 292.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1264.0, + 1238.0, + 1264.0, + 1238.0, + 1303.0, + 1190.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1303.0, + 1292.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1295.0, + 344.0, + 1295.0, + 344.0, + 1332.0, + 293.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1295.0, + 449.0, + 1295.0, + 449.0, + 1332.0, + 401.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1295.0, + 1118.0, + 1295.0, + 1118.0, + 1332.0, + 505.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 786.0, + 1403.0, + 786.0, + 1403.0, + 822.0, + 296.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 818.0, + 1283.0, + 818.0, + 1283.0, + 854.0, + 295.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1388.0, + 1058.0, + 1398.0, + 1058.0, + 1398.0, + 1134.0, + 1388.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1128.0, + 311.0, + 1128.0, + 311.0, + 1197.0, + 300.0, + 1197.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 300, + 1228, + 1401, + 1228, + 1401, + 1634, + 300, + 1634 + ], + "score": 0.983, + "html": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I+30% Uniform75.26 ± 0.3278.72± 0.5378.08±0.6677.79± 0.4675.97± 0.2779.04±0.50
Type-I+ 60% Uniform64.25 ±0.7855.49 ± 2.1167.43 ± 1.4367.63 ± 1.3659.22 ± 0.7472.21 ± 2.92
Type-I+ 30% Asymmetric75.21 ± 0.6475.43 ± 2.9676.91 ± 0.5677.14± 0.7076.96 ± 0.4578.31 ± 0.41
Type-II +30% Uniform74.92 ± 0.6375.19 ± 0.5475.69 ± 0.2175.08± 0.4775.94 ± 0.5880.08±0.37
Type-II+ 60% Uniform64.02 ±0.6659.89 ±0.6366.39 ± 0.2966.76 ± 1.6058.99 ± 1.4371.21 ± 1.46
Type-II+ 30% Asymmetric74.28 ± 0.3973.37 ± 0.8375.30 ± 0.8175.43 ± 0.4277.03 ± 0.6277.63± 0.30
Type-III+30%Uniform74.00 ± 0.3877.31 ± 0.1177.00 ± 0.1276.22 ± 0.1275.66 ± 0.5780.06 ± 0.47
Type-III+ 60% Uniform63.96 ±0.6956.78± 1.5667.53 ± 0.5167.79 ± 0.5459.36 ± 0.9373.48 ±1.84
Type-III+ 30% Asymmetric75.31 ± 0.3474.62 ± 1.7175.70 ± 0.9176.09 ±0.1077.19 ± 0.7477.54±0.70
CIFAR-100Type-I+30%Uniform48.86±0.5652.33±0.6452.90±0.5351.34±0.6445.66 ± 1.6060.09±0.15
Type-I+ 60% Uniform35.97 ± 1.1227.17 ± 1.6638.62 ± 1.6537.57 ± 0.4323.37 ± 0.7251.68± 0.10
Type-I+ 30% Asymmetric45.85 ± 0.9351.21 ± 0.3152.69 ± 1.1450.18 ± 0.9752.04± 0.15
Type-II+30%Uniform49.32± 0.3651.99 ± 0.7553.61± 0.4643.86 ± 1.3156.40±0.34
Type-II + 60% Uniform35.16 ±0.0525.91 ± 0.6439.58 ±3.1350.58± 0.25 37.93± 0.2223.05 ±0.9960.01±0.63
Type-II + 30% Asymmetric46.50 ± 0.9551.07 ± 1.4451.98 ± 0.3749.46 ± 0.2352.11 ± 0.4649.35 ±1.53
Type-III +30% Uniform48.94± 0.6149.94± 0.4452.07±0.3550.18±0.5442.79 ± 1.7861.43±0.33 60.14±0.97
Type-III + 60% Uniform34.67± 0.1622.89 ± 0.7536.82 ±0.4937.65 ± 1.4222.81 ± 0.7250.73±2.16
Type-III + 30% Asymmetric45.70 ± 0.1249.38 ±0.8650.87 ± 1.1248.15±0.9050.31±0.3954.56±1.11
" + }, + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 491, + 298, + 491 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 299, + 785, + 1400, + 785, + 1400, + 1092, + 299, + 1092 + ], + "score": 0.981, + "html": "
DatasetNoiseStandardCo-teaching+GCESLLRTPLC (ours)
CIFAR-10Type-I(35%)78.11 ± 0.7479.97 ± 0.1580.65 ±0.3979.76 ± 0.7280.98 ±0.8082.80±0.27
Type-I(70%)41.98 ± 1.9640.69 ± 1.9936.52 ± 1.6236.29 ±0.6641.52 ± 4.5342.74± 2.14
Type-II(35%)76.65±0.5777.34 ± 0.4477.60± 0.8877.92 ±0.8980.74±0.2581.54± 0.47
Type-II(70%)45.57 ± 1.1245.44 ± 0.6440.30 ± 1.4641.11 ± 1.9244.67 ± 3.8946.04±2.20
Type-III (35%)76.89±0.7978.38 ±0.6779.18 ± 0.6178.81 ±0.2981.08 ± 0.3581.50±0.50
CIFAR-100Type-III(70%)43.32 ± 1.0041.90 ± 0.8637.10 ± 0.5938.49 ± 1.4644.47 ± 1.2345.05 ±1.13
Type-I(35%)57.68 ±0.2956.70 ±0.7158.37 ±0.1855.20±0.3356.74 ±0.3460.01±0.43
Type-I(70%)39.32 ± 0.4339.53 ± 0.2840.01 ± 0.7140.02 ±0.8545.29 ± 0.4345.92 ± 0.61
Type-II (35%)57.83 ± 0.2556.57 ± 0.5258.11 ± 1.0556.10±0.7357.25 ±0.6863.68±0.29
Type-II(70%)39.30± 0.3236.84± 0.3937.75 ± 0.4638.45 ± 0.4543.71 ± 0.5145.03 ±0.50
Type-III (35%) Type-II(70%)56.07 ± 0.79 40.01 ± 0.1855.77 ± 0.98 35.37 ± 2.6557.51 ± 1.16 40.53 ±0.6056.04 ± 0.74 39.94 ± 0.8456.57 ±0.30 44.41 ± 0.1963.68 ± 0.29 44.45 ± 0.62
" + }, + { + "category_id": 1, + "poly": [ + 298, + 506, + 1403, + 506, + 1403, + 662, + 298, + 662 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1911, + 1405, + 1911, + 1405, + 2035, + 298, + 2035 + ], + "score": 0.964 + }, + { + "category_id": 5, + "poly": [ + 867, + 1764, + 1389, + 1764, + 1389, + 1880, + 867, + 1880 + ], + "score": 0.963, + "html": "
β0.050.10.20.3
Type-I Noise83.5883.0483.2883.31
Type-II Noise80.9481.1880.9880.86
Type-III Noise81.9181.7582.1382.39
" + }, + { + "category_id": 5, + "poly": [ + 305, + 1763, + 826, + 1763, + 826, + 1879, + 305, + 1879 + ], + "score": 0.96, + "html": "
exp(00)0.20.30.40.5
Type-INoise83.3383.0482.6682.94
Type-II Noise81.8481.1881.0981.24
Type-III Noise81.7981.7581.9882.06
" + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1401, + 229, + 1401, + 293, + 294, + 293 + ], + "score": 0.957 + }, + { + "category_id": 6, + "poly": [ + 300, + 1672, + 837, + 1672, + 837, + 1762, + 300, + 1762 + ], + "score": 0.941 + }, + { + "category_id": 6, + "poly": [ + 299, + 711, + 1401, + 711, + 1401, + 775, + 299, + 775 + ], + "score": 0.934 + }, + { + "category_id": 6, + "poly": [ + 299, + 1130, + 1399, + 1130, + 1399, + 1193, + 299, + 1193 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.902 + }, + { + "category_id": 6, + "poly": [ + 861, + 1673, + 1399, + 1673, + 1399, + 1763, + 861, + 1763 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.808 + }, + { + "category_id": 13, + "poly": [ + 1226, + 599, + 1350, + 599, + 1350, + 632, + 1226, + 632 + ], + "score": 0.93, + "latex": "\\log f _ { h _ { x } } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1063, + 599, + 1172, + 599, + 1172, + 633, + 1063, + 633 + ], + "score": 0.93, + "latex": "\\log f _ { \\widetilde { y } } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 458, + 1734, + 552, + 1734, + 552, + 1764, + 458, + 1764 + ], + "score": 0.89, + "latex": "\\beta = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 728, + 570, + 757, + 570, + 757, + 599, + 728, + 599 + ], + "score": 0.88, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 452, + 599, + 481, + 599, + 481, + 629, + 452, + 629 + ], + "score": 0.88, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 885, + 509, + 913, + 509, + 913, + 538, + 885, + 538 + ], + "score": 0.88, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 801, + 430, + 857, + 430, + 857, + 459, + 801, + 459 + ], + "score": 0.87, + "latex": "30 \\%" + }, + { + "category_id": 13, + "poly": [ + 909, + 430, + 965, + 430, + 965, + 459, + 909, + 459 + ], + "score": 0.87, + "latex": "60 \\%" + }, + { + "category_id": 13, + "poly": [ + 569, + 1674, + 597, + 1674, + 597, + 1703, + 569, + 1703 + ], + "score": 0.87, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1058, + 339, + 1112, + 339, + 1112, + 368, + 1058, + 368 + ], + "score": 0.87, + "latex": "70 \\%" + }, + { + "category_id": 13, + "poly": [ + 977, + 339, + 1030, + 339, + 1030, + 368, + 977, + 368 + ], + "score": 0.86, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1704, + 1187, + 1704, + 1187, + 1731, + 1133, + 1731 + ], + "score": 0.85, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1674, + 1158, + 1674, + 1158, + 1703, + 1136, + 1703 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 808, + 570, + 830, + 570, + 830, + 600, + 808, + 600 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 783, + 539, + 804, + 539, + 804, + 569, + 783, + 569 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 570, + 1705, + 624, + 1705, + 624, + 1732, + 570, + 1732 + ], + "score": 0.83, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1061, + 1733, + 1191, + 1733, + 1191, + 1766, + 1061, + 1766 + ], + "score": 0.78, + "latex": ") ( \\theta _ { 0 } ) = 0 . 3" + }, + { + "category_id": 13, + "poly": [ + 557, + 714, + 601, + 714, + 601, + 743, + 557, + 743 + ], + "score": 0.78, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1355, + 2003, + 1398, + 2003, + 1398, + 2032, + 1355, + 2032 + ], + "score": 0.77, + "latex": "5 0 k" + }, + { + "category_id": 13, + "poly": [ + 485, + 1505, + 537, + 1505, + 537, + 1524, + 485, + 1524 + ], + "score": 0.67, + "latex": "4 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 485, + 1315, + 537, + 1315, + 537, + 1333, + 485, + 1333 + ], + "score": 0.59, + "latex": "\\mp 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 565, + 1132, + 603, + 1132, + 603, + 1160, + 565, + 1160 + ], + "score": 0.59, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 491, + 1567, + 542, + 1567, + 542, + 1586, + 491, + 1586 + ], + "score": 0.56, + "latex": "\\overline { { + 3 0 \\% } }" + }, + { + "category_id": 13, + "poly": [ + 478, + 1485, + 531, + 1485, + 531, + 1503, + 478, + 1503 + ], + "score": 0.47, + "latex": "+ 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 471, + 1336, + 537, + 1336, + 537, + 1354, + 471, + 1354 + ], + "score": 0.47, + "latex": "\\Pi + 6 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 484, + 1526, + 536, + 1526, + 536, + 1545, + 484, + 1545 + ], + "score": 0.44, + "latex": "+ 6 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 397, + 1763, + 438, + 1763, + 438, + 1794, + 397, + 1794 + ], + "score": 0.43, + "latex": "\\overline { { ( \\theta _ { 0 } ) } }" + }, + { + "category_id": 13, + "poly": [ + 475, + 1547, + 537, + 1547, + 537, + 1565, + 475, + 1565 + ], + "score": 0.42, + "latex": "\\mathrm { I } + 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 950, + 1765, + 969, + 1765, + 969, + 1793, + 950, + 1793 + ], + "score": 0.41, + "latex": "\\overline { { \\beta } }" + }, + { + "category_id": 13, + "poly": [ + 491, + 1418, + 544, + 1418, + 544, + 1437, + 491, + 1437 + ], + "score": 0.39, + "latex": "+ 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 490, + 1397, + 543, + 1397, + 543, + 1416, + 490, + 1416 + ], + "score": 0.37, + "latex": "+ 6 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 479, + 1464, + 531, + 1464, + 531, + 1483, + 479, + 1483 + ], + "score": 0.35, + "latex": "+ 6 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 471, + 1356, + 537, + 1356, + 537, + 1374, + 471, + 1374 + ], + "score": 0.34, + "latex": "\\Pi + 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 486, + 1608, + 544, + 1608, + 544, + 1628, + 486, + 1628 + ], + "score": 0.33, + "latex": "1 + 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 491, + 1376, + 543, + 1376, + 543, + 1395, + 491, + 1395 + ], + "score": 0.32, + "latex": "\\mp 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 468, + 1588, + 543, + 1588, + 543, + 1607, + 468, + 1607 + ], + "score": 0.28, + "latex": ". \\mathrm { I I I } + 6 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 503, + 859, + 566, + 859, + 566, + 880, + 503, + 880 + ], + "score": 0.26, + "latex": "( 3 5 \\% )" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1669.0, + 568.0, + 1669.0, + 568.0, + 1706.0, + 298.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1669.0, + 841.0, + 1669.0, + 841.0, + 1706.0, + 598.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1700.0, + 569.0, + 1700.0, + 569.0, + 1737.0, + 299.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1700.0, + 837.0, + 1700.0, + 837.0, + 1737.0, + 625.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1734.0, + 457.0, + 1734.0, + 457.0, + 1764.0, + 300.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1734.0, + 561.0, + 1734.0, + 561.0, + 1764.0, + 553.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 707.0, + 556.0, + 707.0, + 556.0, + 749.0, + 293.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 707.0, + 1405.0, + 707.0, + 1405.0, + 749.0, + 602.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 743.0, + 1276.0, + 743.0, + 1276.0, + 778.0, + 294.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1130.0, + 564.0, + 1130.0, + 564.0, + 1165.0, + 295.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1130.0, + 1404.0, + 1130.0, + 1404.0, + 1165.0, + 604.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1164.0, + 1159.0, + 1164.0, + 1195.0, + 294.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1668.0, + 1135.0, + 1668.0, + 1135.0, + 1706.0, + 861.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1668.0, + 1404.0, + 1668.0, + 1404.0, + 1706.0, + 1159.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1699.0, + 1132.0, + 1699.0, + 1132.0, + 1736.0, + 862.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1699.0, + 1401.0, + 1699.0, + 1401.0, + 1736.0, + 1188.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1732.0, + 1060.0, + 1732.0, + 1060.0, + 1766.0, + 863.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1732.0, + 1202.0, + 1732.0, + 1202.0, + 1766.0, + 1192.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 1405.0, + 306.0, + 1405.0, + 343.0, + 293.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 336.0, + 976.0, + 336.0, + 976.0, + 372.0, + 294.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 336.0, + 1057.0, + 336.0, + 1057.0, + 372.0, + 1031.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 336.0, + 1404.0, + 336.0, + 1404.0, + 372.0, + 1113.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 368.0, + 1405.0, + 368.0, + 1405.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 434.0, + 293.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 429.0, + 800.0, + 429.0, + 800.0, + 464.0, + 294.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 429.0, + 908.0, + 429.0, + 908.0, + 464.0, + 858.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 429.0, + 1405.0, + 429.0, + 1405.0, + 464.0, + 966.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 460.0, + 1054.0, + 460.0, + 1054.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 507.0, + 884.0, + 507.0, + 884.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 507.0, + 1404.0, + 507.0, + 1404.0, + 540.0, + 914.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 537.0, + 782.0, + 537.0, + 782.0, + 571.0, + 296.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 537.0, + 1405.0, + 537.0, + 1405.0, + 571.0, + 805.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 569.0, + 727.0, + 569.0, + 727.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 569.0, + 807.0, + 569.0, + 807.0, + 603.0, + 758.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 569.0, + 1405.0, + 569.0, + 1405.0, + 603.0, + 831.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 597.0, + 451.0, + 597.0, + 451.0, + 635.0, + 294.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 597.0, + 1062.0, + 597.0, + 1062.0, + 635.0, + 482.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 597.0, + 1225.0, + 597.0, + 1225.0, + 635.0, + 1173.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 597.0, + 1407.0, + 597.0, + 1407.0, + 635.0, + 1351.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 631.0, + 1167.0, + 631.0, + 1167.0, + 664.0, + 296.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 1408.0, + 1912.0, + 1408.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1976.0, + 292.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1354.0, + 2002.0, + 1354.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 2002.0, + 1408.0, + 2002.0, + 1408.0, + 2038.0, + 1399.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1107.0, + 261.0, + 1107.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 950, + 1405, + 950, + 1405, + 1195, + 297, + 1195 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1650, + 1404, + 1650, + 1404, + 1866, + 297, + 1866 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 597, + 1405, + 597, + 1405, + 935, + 298, + 935 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 414, + 298, + 414 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1405, + 1881, + 1405, + 2034, + 299, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 428, + 1404, + 428, + 1404, + 583, + 298, + 583 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 869, + 1394, + 1387, + 1394, + 1387, + 1525, + 869, + 1525 + ], + "score": 0.972, + "html": "
MethodAccuracy
Standard79.4 ± 0.14
SELFIE (Song et al., 2019)81.8 ± 0.09
PLC (ours)83.4±( 0.43
" + }, + { + "category_id": 5, + "poly": [ + 307, + 1395, + 824, + 1395, + 824, + 1524, + 307, + 1524 + ], + "score": 0.971, + "html": "
MethodAccuracy
Standard81.67
CleanNet (Lee et al., 2018)83.95
PLC (ours)85.28 ± 0.04
" + }, + { + "category_id": 5, + "poly": [ + 301, + 1262, + 1405, + 1262, + 1405, + 1317, + 301, + 1317 + ], + "score": 0.939, + "html": "
MethodStandardForwardD2LJOPENCILDYGCESLMLNTLRTPLC (ours)
Accuracy68.9469.8469.4772.2373.4971.0069.7571.0273.4771.7474.02
" + }, + { + "category_id": 0, + "poly": [ + 299, + 1583, + 544, + 1583, + 544, + 1619, + 299, + 1619 + ], + "score": 0.912 + }, + { + "category_id": 6, + "poly": [ + 875, + 1364, + 1389, + 1364, + 1389, + 1394, + 875, + 1394 + ], + "score": 0.908 + }, + { + "category_id": 6, + "poly": [ + 609, + 1232, + 1090, + 1232, + 1090, + 1261, + 609, + 1261 + ], + "score": 0.902 + }, + { + "category_id": 6, + "poly": [ + 326, + 1365, + 808, + 1365, + 808, + 1395, + 326, + 1395 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.792 + }, + { + "category_id": 13, + "poly": [ + 801, + 660, + 856, + 660, + 856, + 689, + 801, + 689 + ], + "score": 0.88, + "latex": "80 \\%" + }, + { + "category_id": 13, + "poly": [ + 715, + 1103, + 769, + 1103, + 769, + 1132, + 715, + 1132 + ], + "score": 0.87, + "latex": "7 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 612, + 1103, + 666, + 1103, + 666, + 1132, + 612, + 1132 + ], + "score": 0.87, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 657, + 1011, + 697, + 1011, + 697, + 1039, + 657, + 1039 + ], + "score": 0.87, + "latex": "8 \\%" + }, + { + "category_id": 13, + "poly": [ + 390, + 231, + 433, + 231, + 433, + 259, + 390, + 259 + ], + "score": 0.82, + "latex": "1 0 k" + }, + { + "category_id": 13, + "poly": [ + 296, + 231, + 340, + 231, + 340, + 260, + 296, + 260 + ], + "score": 0.82, + "latex": "1 4 k" + }, + { + "category_id": 13, + "poly": [ + 1209, + 322, + 1267, + 322, + 1267, + 350, + 1209, + 350 + ], + "score": 0.82, + "latex": "2 6 0 k" + }, + { + "category_id": 13, + "poly": [ + 759, + 292, + 803, + 292, + 803, + 320, + 759, + 320 + ], + "score": 0.8, + "latex": "1 0 k" + }, + { + "category_id": 13, + "poly": [ + 1096, + 262, + 1142, + 262, + 1142, + 290, + 1096, + 290 + ], + "score": 0.79, + "latex": "5 0 k" + }, + { + "category_id": 13, + "poly": [ + 316, + 843, + 417, + 843, + 417, + 872, + 316, + 872 + ], + "score": 0.78, + "latex": "2 2 4 \\mathbf { x } 2 2 4" + }, + { + "category_id": 13, + "poly": [ + 1130, + 1365, + 1171, + 1365, + 1171, + 1393, + 1130, + 1393 + ], + "score": 0.62, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 864, + 1233, + 907, + 1233, + 907, + 1261, + 864, + 1261 + ], + "score": 0.6, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 586, + 1366, + 629, + 1366, + 629, + 1395, + 586, + 1395 + ], + "score": 0.6, + "latex": "( \\% )" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1577.0, + 550.0, + 1577.0, + 550.0, + 1628.0, + 290.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1361.0, + 1129.0, + 1361.0, + 1129.0, + 1398.0, + 868.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1361.0, + 1391.0, + 1361.0, + 1391.0, + 1398.0, + 1172.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1229.0, + 863.0, + 1229.0, + 863.0, + 1268.0, + 604.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1229.0, + 1094.0, + 1229.0, + 1094.0, + 1268.0, + 908.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1363.0, + 585.0, + 1363.0, + 585.0, + 1398.0, + 329.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1363.0, + 807.0, + 1363.0, + 807.0, + 1398.0, + 630.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 947.0, + 1410.0, + 947.0, + 1410.0, + 986.0, + 292.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 978.0, + 1410.0, + 978.0, + 1410.0, + 1016.0, + 294.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1004.0, + 656.0, + 1004.0, + 656.0, + 1049.0, + 291.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1004.0, + 1406.0, + 1004.0, + 1406.0, + 1049.0, + 698.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1041.0, + 1405.0, + 1041.0, + 1405.0, + 1075.0, + 295.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1070.0, + 1407.0, + 1070.0, + 1407.0, + 1107.0, + 292.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1102.0, + 611.0, + 1102.0, + 611.0, + 1136.0, + 294.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1102.0, + 714.0, + 1102.0, + 714.0, + 1136.0, + 667.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1102.0, + 1405.0, + 1102.0, + 1405.0, + 1136.0, + 770.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1132.0, + 1405.0, + 1132.0, + 1405.0, + 1166.0, + 295.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1166.0, + 778.0, + 1166.0, + 778.0, + 1196.0, + 296.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1686.0, + 295.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1683.0, + 1409.0, + 1683.0, + 1409.0, + 1717.0, + 295.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1711.0, + 1407.0, + 1711.0, + 1407.0, + 1749.0, + 292.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1743.0, + 1409.0, + 1743.0, + 1409.0, + 1775.0, + 292.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1774.0, + 1406.0, + 1774.0, + 1406.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1839.0, + 295.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1832.0, + 1035.0, + 1832.0, + 1035.0, + 1872.0, + 292.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 598.0, + 1407.0, + 598.0, + 1407.0, + 633.0, + 295.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 627.0, + 1406.0, + 627.0, + 1406.0, + 664.0, + 293.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 659.0, + 800.0, + 659.0, + 800.0, + 694.0, + 295.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 659.0, + 1405.0, + 659.0, + 1405.0, + 694.0, + 857.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 690.0, + 1403.0, + 690.0, + 1403.0, + 725.0, + 295.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 720.0, + 1406.0, + 720.0, + 1406.0, + 755.0, + 295.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 749.0, + 1406.0, + 749.0, + 1406.0, + 786.0, + 292.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1405.0, + 782.0, + 1405.0, + 817.0, + 293.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 807.0, + 1406.0, + 807.0, + 1406.0, + 851.0, + 291.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 841.0, + 315.0, + 841.0, + 315.0, + 878.0, + 292.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 841.0, + 1405.0, + 841.0, + 1405.0, + 878.0, + 418.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 873.0, + 1405.0, + 873.0, + 1405.0, + 908.0, + 295.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 905.0, + 639.0, + 905.0, + 639.0, + 938.0, + 294.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 225.0, + 295.0, + 225.0, + 295.0, + 269.0, + 291.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 225.0, + 389.0, + 225.0, + 389.0, + 269.0, + 341.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 225.0, + 1405.0, + 225.0, + 1405.0, + 269.0, + 434.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 259.0, + 1095.0, + 259.0, + 1095.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 296.0, + 1143.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 758.0, + 288.0, + 758.0, + 329.0, + 292.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 288.0, + 1408.0, + 288.0, + 1408.0, + 329.0, + 804.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 322.0, + 1208.0, + 322.0, + 1208.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 356.0, + 1268.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1403.0, + 351.0, + 1403.0, + 386.0, + 294.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1068.0, + 381.0, + 1068.0, + 418.0, + 292.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1878.0, + 1406.0, + 1878.0, + 1406.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1408.0, + 1940.0, + 1408.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1967.0, + 1409.0, + 1967.0, + 1409.0, + 2012.0, + 293.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 824.0, + 2004.0, + 824.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 428.0, + 1405.0, + 428.0, + 1405.0, + 466.0, + 294.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1406.0, + 459.0, + 1406.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 488.0, + 1405.0, + 488.0, + 1405.0, + 529.0, + 292.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 1405.0, + 519.0, + 1405.0, + 560.0, + 293.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1408.0, + 552.0, + 1408.0, + 589.0, + 294.0, + 589.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.82 + }, + { + "category_id": 0, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 262, + 299, + 262 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 293, + 1518, + 1402, + 1518, + 1402, + 1584, + 293, + 1584 + ], + "score": 0.564 + }, + { + "category_id": 1, + "poly": [ + 292, + 1434, + 1405, + 1434, + 1405, + 1500, + 292, + 1500 + ], + "score": 0.562 + }, + { + "category_id": 1, + "poly": [ + 298, + 279, + 1403, + 279, + 1403, + 344, + 298, + 344 + ], + "score": 0.542 + }, + { + "category_id": 1, + "poly": [ + 297, + 446, + 1402, + 446, + 1402, + 513, + 297, + 513 + ], + "score": 0.511 + }, + { + "category_id": 1, + "poly": [ + 300, + 1688, + 1406, + 1688, + 1406, + 1783, + 300, + 1783 + ], + "score": 0.51 + }, + { + "category_id": 1, + "poly": [ + 291, + 1602, + 1402, + 1602, + 1402, + 1669, + 291, + 1669 + ], + "score": 0.507 + }, + { + "category_id": 1, + "poly": [ + 296, + 362, + 1401, + 362, + 1401, + 428, + 296, + 428 + ], + "score": 0.504 + }, + { + "category_id": 1, + "poly": [ + 295, + 1350, + 1403, + 1350, + 1403, + 1416, + 295, + 1416 + ], + "score": 0.503 + }, + { + "category_id": 1, + "poly": [ + 298, + 982, + 1401, + 982, + 1401, + 1078, + 298, + 1078 + ], + "score": 0.495 + }, + { + "category_id": 1, + "poly": [ + 298, + 898, + 1400, + 898, + 1400, + 965, + 298, + 965 + ], + "score": 0.49 + }, + { + "category_id": 1, + "poly": [ + 297, + 1265, + 1401, + 1265, + 1401, + 1332, + 297, + 1332 + ], + "score": 0.489 + }, + { + "category_id": 1, + "poly": [ + 298, + 814, + 1399, + 814, + 1399, + 880, + 298, + 880 + ], + "score": 0.485 + }, + { + "category_id": 1, + "poly": [ + 292, + 530, + 1401, + 530, + 1401, + 597, + 292, + 597 + ], + "score": 0.484 + }, + { + "category_id": 1, + "poly": [ + 303, + 615, + 1401, + 615, + 1401, + 711, + 303, + 711 + ], + "score": 0.472 + }, + { + "category_id": 1, + "poly": [ + 296, + 1182, + 1399, + 1182, + 1399, + 1248, + 296, + 1248 + ], + "score": 0.471 + }, + { + "category_id": 1, + "poly": [ + 297, + 729, + 1400, + 729, + 1400, + 796, + 297, + 796 + ], + "score": 0.471 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1397, + 1971, + 1397, + 2036, + 299, + 2036 + ], + "score": 0.466 + }, + { + "category_id": 1, + "poly": [ + 296, + 1097, + 1400, + 1097, + 1400, + 1163, + 296, + 1163 + ], + "score": 0.464 + }, + { + "category_id": 1, + "poly": [ + 297, + 1886, + 1399, + 1886, + 1399, + 1952, + 297, + 1952 + ], + "score": 0.442 + }, + { + "category_id": 1, + "poly": [ + 292, + 1801, + 1402, + 1801, + 1402, + 1868, + 292, + 1868 + ], + "score": 0.415 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1515.0, + 1405.0, + 1515.0, + 1405.0, + 1558.0, + 293.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1551.0, + 693.0, + 1551.0, + 693.0, + 1583.0, + 322.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1435.0, + 1408.0, + 1435.0, + 1408.0, + 1471.0, + 296.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1466.0, + 702.0, + 1466.0, + 702.0, + 1498.0, + 320.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 275.0, + 1405.0, + 275.0, + 1405.0, + 318.0, + 293.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 309.0, + 1015.0, + 309.0, + 1015.0, + 346.0, + 321.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 442.0, + 1406.0, + 442.0, + 1406.0, + 487.0, + 295.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 478.0, + 1336.0, + 478.0, + 1336.0, + 513.0, + 322.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1689.0, + 1408.0, + 1689.0, + 1408.0, + 1723.0, + 296.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1718.0, + 1413.0, + 1718.0, + 1413.0, + 1754.0, + 322.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1751.0, + 392.0, + 1751.0, + 392.0, + 1785.0, + 317.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1641.0, + 295.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1636.0, + 1063.0, + 1636.0, + 1063.0, + 1669.0, + 322.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 361.0, + 1403.0, + 361.0, + 1403.0, + 400.0, + 294.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 395.0, + 1160.0, + 395.0, + 1160.0, + 429.0, + 321.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1345.0, + 1405.0, + 1345.0, + 1405.0, + 1391.0, + 294.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1383.0, + 857.0, + 1383.0, + 857.0, + 1415.0, + 323.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 982.0, + 1406.0, + 982.0, + 1406.0, + 1020.0, + 294.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1014.0, + 1408.0, + 1014.0, + 1408.0, + 1052.0, + 321.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1042.0, + 398.0, + 1042.0, + 398.0, + 1079.0, + 319.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 898.0, + 1405.0, + 898.0, + 1405.0, + 938.0, + 294.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 930.0, + 958.0, + 930.0, + 958.0, + 965.0, + 322.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1266.0, + 1405.0, + 1266.0, + 1405.0, + 1302.0, + 295.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1297.0, + 646.0, + 1297.0, + 646.0, + 1332.0, + 317.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 816.0, + 1404.0, + 816.0, + 1404.0, + 852.0, + 294.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 846.0, + 1011.0, + 846.0, + 1011.0, + 881.0, + 323.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 531.0, + 1403.0, + 531.0, + 1403.0, + 567.0, + 295.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 562.0, + 967.0, + 562.0, + 967.0, + 599.0, + 323.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 1405.0, + 613.0, + 1405.0, + 654.0, + 294.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 646.0, + 1404.0, + 646.0, + 1404.0, + 683.0, + 323.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 676.0, + 682.0, + 676.0, + 682.0, + 714.0, + 322.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1180.0, + 1404.0, + 1180.0, + 1404.0, + 1220.0, + 294.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1215.0, + 870.0, + 1215.0, + 870.0, + 1247.0, + 324.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 725.0, + 1404.0, + 725.0, + 1404.0, + 770.0, + 292.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 763.0, + 964.0, + 763.0, + 964.0, + 795.0, + 323.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1971.0, + 1402.0, + 1971.0, + 1402.0, + 2007.0, + 298.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2001.0, + 1011.0, + 2001.0, + 1011.0, + 2037.0, + 323.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1095.0, + 1405.0, + 1095.0, + 1405.0, + 1137.0, + 292.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1129.0, + 1158.0, + 1129.0, + 1158.0, + 1164.0, + 322.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1927.0, + 292.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1916.0, + 837.0, + 1916.0, + 837.0, + 1953.0, + 323.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1838.0, + 296.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1833.0, + 1152.0, + 1833.0, + 1152.0, + 1870.0, + 324.0, + 1870.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 862, + 2088, + 862, + 2113, + 835, + 2113 + ], + "score": 0.808 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1403, + 228, + 1403, + 323, + 299, + 323 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 300, + 821, + 1404, + 821, + 1404, + 915, + 300, + 915 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 298, + 707, + 1404, + 707, + 1404, + 802, + 298, + 802 + ], + "score": 0.595 + }, + { + "category_id": 1, + "poly": [ + 299, + 1103, + 1402, + 1103, + 1402, + 1197, + 299, + 1197 + ], + "score": 0.555 + }, + { + "category_id": 1, + "poly": [ + 296, + 624, + 1397, + 624, + 1397, + 690, + 296, + 690 + ], + "score": 0.552 + }, + { + "category_id": 1, + "poly": [ + 294, + 1552, + 1402, + 1552, + 1402, + 1618, + 294, + 1618 + ], + "score": 0.551 + }, + { + "category_id": 1, + "poly": [ + 297, + 1019, + 1401, + 1019, + 1401, + 1086, + 297, + 1086 + ], + "score": 0.55 + }, + { + "category_id": 1, + "poly": [ + 298, + 936, + 1400, + 936, + 1400, + 1002, + 298, + 1002 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 292, + 1217, + 1404, + 1217, + 1404, + 1282, + 292, + 1282 + ], + "score": 0.547 + }, + { + "category_id": 1, + "poly": [ + 294, + 1301, + 1402, + 1301, + 1402, + 1367, + 294, + 1367 + ], + "score": 0.537 + }, + { + "category_id": 1, + "poly": [ + 295, + 1468, + 1403, + 1468, + 1403, + 1533, + 295, + 1533 + ], + "score": 0.529 + }, + { + "category_id": 1, + "poly": [ + 294, + 1635, + 1404, + 1635, + 1404, + 1701, + 294, + 1701 + ], + "score": 0.529 + }, + { + "category_id": 1, + "poly": [ + 294, + 1384, + 1403, + 1384, + 1403, + 1451, + 294, + 1451 + ], + "score": 0.512 + }, + { + "category_id": 1, + "poly": [ + 295, + 1719, + 1401, + 1719, + 1401, + 1785, + 295, + 1785 + ], + "score": 0.499 + }, + { + "category_id": 1, + "poly": [ + 296, + 456, + 1399, + 456, + 1399, + 522, + 296, + 522 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 295, + 541, + 1397, + 541, + 1397, + 605, + 295, + 605 + ], + "score": 0.496 + }, + { + "category_id": 1, + "poly": [ + 300, + 343, + 1400, + 343, + 1400, + 437, + 300, + 437 + ], + "score": 0.478 + }, + { + "category_id": 1, + "poly": [ + 301, + 1971, + 1400, + 1971, + 1400, + 2035, + 301, + 2035 + ], + "score": 0.457 + }, + { + "category_id": 1, + "poly": [ + 294, + 1804, + 1402, + 1804, + 1402, + 1868, + 294, + 1868 + ], + "score": 0.451 + }, + { + "category_id": 1, + "poly": [ + 298, + 1887, + 1402, + 1887, + 1402, + 1953, + 298, + 1953 + ], + "score": 0.437 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1408.0, + 228.0, + 1408.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1409.0, + 260.0, + 1409.0, + 297.0, + 322.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 285.0, + 399.0, + 285.0, + 399.0, + 326.0, + 320.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 823.0, + 1403.0, + 823.0, + 1403.0, + 857.0, + 296.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 850.0, + 1408.0, + 850.0, + 1408.0, + 892.0, + 320.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 881.0, + 399.0, + 881.0, + 399.0, + 918.0, + 320.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 705.0, + 1405.0, + 705.0, + 1405.0, + 746.0, + 295.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 739.0, + 1410.0, + 739.0, + 1410.0, + 776.0, + 321.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 767.0, + 397.0, + 767.0, + 397.0, + 803.0, + 318.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1101.0, + 1407.0, + 1101.0, + 1407.0, + 1142.0, + 293.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1134.0, + 1408.0, + 1134.0, + 1408.0, + 1172.0, + 320.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1162.0, + 398.0, + 1162.0, + 398.0, + 1200.0, + 319.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 625.0, + 1402.0, + 625.0, + 1402.0, + 661.0, + 295.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 656.0, + 861.0, + 656.0, + 861.0, + 689.0, + 321.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1553.0, + 1404.0, + 1553.0, + 1404.0, + 1589.0, + 296.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1584.0, + 1134.0, + 1584.0, + 1134.0, + 1619.0, + 321.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1017.0, + 1406.0, + 1017.0, + 1406.0, + 1058.0, + 296.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1053.0, + 848.0, + 1053.0, + 848.0, + 1084.0, + 323.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 938.0, + 1404.0, + 938.0, + 1404.0, + 974.0, + 294.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 966.0, + 928.0, + 966.0, + 928.0, + 1001.0, + 322.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1211.0, + 1407.0, + 1211.0, + 1407.0, + 1260.0, + 291.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1248.0, + 792.0, + 1248.0, + 792.0, + 1283.0, + 321.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1304.0, + 1403.0, + 1304.0, + 1403.0, + 1337.0, + 297.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1333.0, + 1029.0, + 1333.0, + 1029.0, + 1369.0, + 323.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1470.0, + 1406.0, + 1470.0, + 1406.0, + 1506.0, + 297.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1499.0, + 713.0, + 1499.0, + 713.0, + 1531.0, + 322.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1634.0, + 1407.0, + 1634.0, + 1407.0, + 1673.0, + 296.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1668.0, + 1188.0, + 1668.0, + 1188.0, + 1702.0, + 322.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1384.0, + 1406.0, + 1384.0, + 1406.0, + 1420.0, + 296.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1416.0, + 1322.0, + 1416.0, + 1322.0, + 1451.0, + 323.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1716.0, + 1405.0, + 1716.0, + 1405.0, + 1759.0, + 295.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1752.0, + 1168.0, + 1752.0, + 1168.0, + 1784.0, + 322.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 457.0, + 1405.0, + 457.0, + 1405.0, + 493.0, + 295.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 490.0, + 741.0, + 490.0, + 741.0, + 521.0, + 324.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 541.0, + 1402.0, + 541.0, + 1402.0, + 577.0, + 296.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 572.0, + 1185.0, + 572.0, + 1185.0, + 608.0, + 323.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 343.0, + 1404.0, + 343.0, + 1404.0, + 378.0, + 294.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 375.0, + 1404.0, + 375.0, + 1404.0, + 409.0, + 322.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 403.0, + 654.0, + 403.0, + 654.0, + 441.0, + 321.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 1295.0, + 2002.0, + 1295.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1804.0, + 1404.0, + 1804.0, + 1404.0, + 1840.0, + 296.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1835.0, + 921.0, + 1835.0, + 921.0, + 1871.0, + 323.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1884.0, + 1404.0, + 1884.0, + 1404.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1917.0, + 1103.0, + 1917.0, + 1103.0, + 1954.0, + 323.0, + 1954.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1788, + 1406, + 1788, + 1406, + 1927, + 297, + 1927 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 296, + 817, + 1406, + 817, + 1406, + 969, + 296, + 969 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 311, + 1202, + 1390, + 1202, + 1390, + 1767, + 311, + 1767 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1930, + 1403, + 1930, + 1403, + 2039, + 298, + 2039 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 297, + 987, + 1408, + 987, + 1408, + 1062, + 297, + 1062 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 457, + 1070, + 1236, + 1070, + 1236, + 1147, + 457, + 1147 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 302, + 1155, + 1396, + 1155, + 1396, + 1194, + 302, + 1194 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 749, + 510, + 749, + 510, + 785, + 300, + 785 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 300, + 631, + 1406, + 631, + 1406, + 697, + 300, + 697 + ], + "score": 0.634 + }, + { + "category_id": 1, + "poly": [ + 300, + 551, + 1399, + 551, + 1399, + 617, + 300, + 617 + ], + "score": 0.483 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1400, + 228, + 1400, + 295, + 297, + 295 + ], + "score": 0.466 + }, + { + "category_id": 1, + "poly": [ + 301, + 471, + 1400, + 471, + 1400, + 535, + 301, + 535 + ], + "score": 0.42 + }, + { + "category_id": 1, + "poly": [ + 302, + 310, + 1400, + 310, + 1400, + 374, + 302, + 374 + ], + "score": 0.384 + }, + { + "category_id": 1, + "poly": [ + 303, + 389, + 1401, + 389, + 1401, + 455, + 303, + 455 + ], + "score": 0.374 + }, + { + "category_id": 13, + "poly": [ + 679, + 988, + 791, + 988, + 791, + 1027, + 679, + 1027 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\eta ( \\pmb { x } ) > \\frac { 1 } { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1158, + 1178, + 1158, + 1178, + 1192, + 1095, + 1192 + ], + "score": 0.93, + "latex": "( c _ { * } , c ^ { * } )" + }, + { + "category_id": 14, + "poly": [ + 458, + 1069, + 1238, + 1069, + 1238, + 1148, + 458, + 1148 + ], + "score": 0.93, + "latex": "\\mathbb { E } _ { ( z , \\tilde { y } ) \\sim ( D , \\tilde { \\eta } _ { n e w } ( z ) ) } \\left[ \\mathbb { 1 } _ { \\{ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) \\} } ( z ) \\bigg | \\bigg | \\eta ( z ) - \\frac { 1 } { 2 } \\bigg | \\ge \\bigg | \\eta ( x ) - \\frac { 1 } { 2 } \\bigg | \\right] = 0 ." + }, + { + "category_id": 13, + "poly": [ + 483, + 1999, + 632, + 1999, + 632, + 2039, + 483, + 2039 + ], + "score": 0.93, + "latex": "L ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 765, + 1027, + 847, + 1027, + 847, + 1060, + 765, + 1060 + ], + "score": 0.93, + "latex": "L ( e , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1961, + 1241, + 1961, + 1241, + 2003, + 1016, + 2003 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 559, + 1155, + 781, + 1155, + 781, + 1194, + 559, + 1194 + ], + "score": 0.92, + "latex": "e - \\gamma \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } |" + }, + { + "category_id": 13, + "poly": [ + 749, + 1891, + 1078, + 1891, + 1078, + 1928, + 749, + 1928 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\ell \\alpha } ) ( \\frac { 1 } { 2 } - e ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 325, + 1791, + 653, + 1791, + 653, + 1832, + 325, + 1832 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) \\le \\gamma \\le \\frac { 2 \\varepsilon } { \\ell \\alpha } ( \\frac { 1 } { 2 } - e ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1234, + 1930, + 1328, + 1930, + 1328, + 1964, + 1234, + 1964 + ], + "score": 0.92, + "latex": "L ( e _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 551, + 849, + 641, + 849, + 641, + 881, + 551, + 881 + ], + "score": 0.92, + "latex": "( e , f , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1791, + 1129, + 1791, + 1129, + 1831, + 1011, + 1831 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { 1 } { 2 } + e - \\gamma } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1064, + 1024, + 1286, + 1024, + 1286, + 1063, + 1064, + 1063 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { e \\leq | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | , \\forall \\pmb { x } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 544, + 1964, + 658, + 1964, + 658, + 1998, + 544, + 1998 + ], + "score": 0.91, + "latex": "T _ { 0 } < 1 / 2" + }, + { + "category_id": 13, + "poly": [ + 796, + 847, + 927, + 847, + 927, + 884, + 796, + 884 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { 3 \\dot { \\epsilon } \\le e < \\frac { 1 } { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 755, + 1829, + 1020, + 1829, + 1020, + 1863, + 755, + 1863 + ], + "score": 0.91, + "latex": "| f _ { n e w } ( \\pmb { x } ) - \\eta ( \\pmb { x } ) | \\le 3 \\epsilon" + }, + { + "category_id": 14, + "poly": [ + 311, + 1201, + 1398, + 1201, + 1398, + 1772, + 311, + 1772 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { \\mathbb { E } _ { ( z \\sim \\hat { y } ) \\sim ( D , \\hat { y } \\sim \\alpha ( z ) ) } [ \\hat { 1 } ( s _ { z \\sim \\hat { y } } , \\hat { x } ) ( z ) ] \\lVert ( z ( z ) - \\frac { 1 } { 2 } \\vert \\geq | y | ( x ) - \\frac { 1 } { 2 } \\vert ] } \\\\ & { = \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\lVert ( z ) - \\frac { 1 } { 2 } [ \\hat { y } \\langle x \\rangle - \\frac { 1 } { 2 } ] } \\\\ & { \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } } { \\lVert \\mathcal { Z } \\rVert ( z ) ^ { 2 } - \\frac { 1 } { 2 } } [ \\frac { \\mathbb { E } _ { \\hat { z } } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) } { 1 - \\frac { 1 } { 2 } } ] } \\\\ & { \\leq \\frac { \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Y } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon - \\epsilon - \\mathbb { E } _ { z } \\cdot \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { \\hat { z } } } { 2 } + \\epsilon } { \\mathcal { Z } \\cdot { \\mathcal { Z } } ( z ) ^ { 2 } - \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } - \\frac { \\epsilon } { 2 } } [ \\mathcal { Z } \\cdot \\mathcal { Y } ( z ) \\leq \\frac { \\frac { 1 } { 2 } } { 2 } - \\epsilon + \\frac { \\mathbb { E } _ { z } } { 1 } ] } } \\\\ & \\quad + \\frac \\mathbb { E } _ { ( \\hat { z } ) \\sim } f ( \\hat { x } ) \\cdot \\mathcal { Z } ( z ) \\cdot \\frac { 1 } { 2 } + \\epsilon \\mathcal { Z } ( \\hat { z } ) \\cdot \\frac { \\mathbb { E } _ { z } \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } ( z ) \\cdot \\mathcal { Z } ( z ) } \\epsilon - \\frac \\mathbb { E } _ { z } \\cdot \\{ { y } \\cdot \\mathcal { Z } \\} \\cdot \\mathcal { Y } ( z ) \\cdot \\mathcal { Z } \\cdot \\mathcal { Y } ( z \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 338, + 1833, + 403, + 1833, + 403, + 1862, + 338, + 1862 + ], + "score": 0.9, + "latex": "( \\alpha , \\varepsilon )" + }, + { + "category_id": 13, + "poly": [ + 910, + 928, + 1241, + 928, + 1241, + 967, + 910, + 967 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\frac { 1 } { 2 } - e _ { n e w } \\geq ( 1 + \\frac { \\varepsilon } { \\alpha \\ell } ) ( \\frac { 1 } { 2 } - e ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 348, + 881, + 636, + 881, + 636, + 916, + 348, + 916 + ], + "score": 0.89, + "latex": "\\tilde { \\eta } ( \\pmb { x } ) i f | f ( \\pmb { x } ) - 1 / 2 | < e" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1830, + 1364, + 1830, + 1364, + 1865, + 1137, + 1865 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { e - \\gamma \\le | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 964, + 1862, + 1074, + 1862, + 1074, + 1892, + 964, + 1892 + ], + "score": 0.89, + "latex": "( e - \\gamma , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 387, + 1861, + 484, + 1861, + 484, + 1894, + 387, + 1894 + ], + "score": 0.88, + "latex": "{ \\dot { f } } _ { n e w } ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1792, + 1278, + 1792, + 1278, + 1831, + 1250, + 1831 + ], + "score": 0.87, + "latex": "\\frac { 2 \\epsilon } { \\alpha }" + }, + { + "category_id": 13, + "poly": [ + 762, + 930, + 832, + 930, + 832, + 964, + 762, + 964 + ], + "score": 0.86, + "latex": "f _ { n e w } \\}" + }, + { + "category_id": 13, + "poly": [ + 957, + 1932, + 985, + 1932, + 985, + 1963, + 957, + 1963 + ], + "score": 0.86, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 424, + 1965, + 452, + 1965, + 452, + 1997, + 424, + 1997 + ], + "score": 0.86, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1863, + 857, + 1863, + 857, + 1893, + 788, + 1893 + ], + "score": 0.86, + "latex": "\\eta ^ { * } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 809, + 2003, + 836, + 2003, + 836, + 2035, + 809, + 2035 + ], + "score": 0.86, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 981, + 848, + 1402, + 848, + 1402, + 883, + 981, + 883 + ], + "score": 0.85, + "latex": "\\tilde { \\eta } _ { n e w } ( { \\pmb x } ) = y _ { f ( { \\pmb x } ) } i f | f ( { \\pmb x } ) - 1 / 2 | \\ge e" + }, + { + "category_id": 13, + "poly": [ + 790, + 885, + 1346, + 885, + 1346, + 924, + 790, + 924 + ], + "score": 0.85, + "latex": "f _ { n e w } = \\arg \\operatorname* { m i n } _ { h \\in \\mathcal { H } } E _ { ( \\pmb { x } , \\tilde { y } ) \\sim D ( \\pmb { x } , \\tilde { \\eta } _ { n e w } ( \\pmb { x } ) ) } L o s s ( h ( \\pmb { x } ) , \\tilde { y } )" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1159, + 1064, + 1159, + 1064, + 1186, + 1038, + 1186 + ], + "score": 0.84, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 937, + 1027, + 955, + 1027, + 955, + 1058, + 937, + 1058 + ], + "score": 0.83, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1379, + 1862, + 1398, + 1862, + 1398, + 1892, + 1379, + 1892 + ], + "score": 0.83, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 298, + 1863, + 375, + 1863, + 375, + 1891, + 298, + 1891 + ], + "score": 0.82, + "latex": "e \\geq 3 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 585, + 1896, + 606, + 1896, + 606, + 1924, + 585, + 1924 + ], + "score": 0.81, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 1327, + 820, + 1345, + 820, + 1345, + 850, + 1327, + 850 + ], + "score": 0.81, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1067, + 1836, + 1088, + 1836, + 1088, + 1859, + 1067, + 1859 + ], + "score": 0.78, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 458, + 1164, + 480, + 1164, + 480, + 1186, + 458, + 1186 + ], + "score": 0.76, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 928, + 537, + 928, + 537, + 964, + 297, + 964 + ], + "score": 0.61, + "latex": "e _ { n e w } = \\operatorname* { m i n } \\{ e | e > 0" + }, + { + "category_id": 13, + "poly": [ + 1016, + 821, + 1031, + 821, + 1031, + 846, + 1016, + 846 + ], + "score": 0.53, + "latex": "^ { l }" + }, + { + "category_id": 13, + "poly": [ + 552, + 929, + 636, + 929, + 636, + 964, + 552, + 964 + ], + "score": 0.51, + "latex": "L ( e , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 952, + 1967, + 968, + 1967, + 968, + 1993, + 952, + 1993 + ], + "score": 0.44, + "latex": "^ { l }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 747.0, + 514.0, + 747.0, + 514.0, + 791.0, + 295.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1788.0, + 324.0, + 1788.0, + 324.0, + 1833.0, + 290.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1788.0, + 1010.0, + 1788.0, + 1010.0, + 1833.0, + 654.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1788.0, + 1249.0, + 1788.0, + 1249.0, + 1833.0, + 1130.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1788.0, + 1409.0, + 1788.0, + 1409.0, + 1833.0, + 1279.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1825.0, + 337.0, + 1825.0, + 337.0, + 1866.0, + 291.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1825.0, + 754.0, + 1825.0, + 754.0, + 1866.0, + 404.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1825.0, + 1066.0, + 1825.0, + 1066.0, + 1866.0, + 1021.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1825.0, + 1136.0, + 1825.0, + 1136.0, + 1866.0, + 1089.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1825.0, + 1409.0, + 1825.0, + 1409.0, + 1866.0, + 1365.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1858.0, + 297.0, + 1858.0, + 297.0, + 1896.0, + 292.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1858.0, + 386.0, + 1858.0, + 386.0, + 1896.0, + 376.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1858.0, + 787.0, + 1858.0, + 787.0, + 1896.0, + 485.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1858.0, + 963.0, + 1858.0, + 963.0, + 1896.0, + 858.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1858.0, + 1378.0, + 1858.0, + 1378.0, + 1896.0, + 1075.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1858.0, + 1411.0, + 1858.0, + 1411.0, + 1896.0, + 1399.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1887.0, + 584.0, + 1887.0, + 584.0, + 1928.0, + 292.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1887.0, + 748.0, + 1887.0, + 748.0, + 1928.0, + 607.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1887.0, + 1090.0, + 1887.0, + 1090.0, + 1928.0, + 1079.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1891.0, + 1406.0, + 1891.0, + 1406.0, + 1921.0, + 1373.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 818.0, + 1015.0, + 818.0, + 1015.0, + 851.0, + 295.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 818.0, + 1326.0, + 818.0, + 1326.0, + 851.0, + 1032.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 818.0, + 1403.0, + 818.0, + 1403.0, + 851.0, + 1346.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 842.0, + 550.0, + 842.0, + 550.0, + 886.0, + 293.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 842.0, + 795.0, + 842.0, + 795.0, + 886.0, + 642.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 842.0, + 980.0, + 842.0, + 980.0, + 886.0, + 928.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 842.0, + 1407.0, + 842.0, + 1407.0, + 886.0, + 1403.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 875.0, + 347.0, + 875.0, + 347.0, + 932.0, + 290.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 875.0, + 789.0, + 875.0, + 789.0, + 932.0, + 637.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 875.0, + 1411.0, + 875.0, + 1411.0, + 932.0, + 1347.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 917.0, + 296.0, + 917.0, + 296.0, + 974.0, + 289.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 917.0, + 551.0, + 917.0, + 551.0, + 974.0, + 538.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 917.0, + 761.0, + 917.0, + 761.0, + 974.0, + 637.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 917.0, + 909.0, + 917.0, + 909.0, + 974.0, + 833.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 917.0, + 1254.0, + 917.0, + 1254.0, + 974.0, + 1242.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 1928.0, + 423.0, + 1928.0, + 423.0, + 2020.0, + 282.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1928.0, + 543.0, + 1928.0, + 543.0, + 2020.0, + 453.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1928.0, + 951.0, + 1928.0, + 951.0, + 2020.0, + 659.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1928.0, + 1015.0, + 1928.0, + 1015.0, + 2020.0, + 986.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1928.0, + 1405.0, + 1928.0, + 1405.0, + 2020.0, + 1329.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2001.0, + 482.0, + 2001.0, + 482.0, + 2038.0, + 296.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 2001.0, + 808.0, + 2001.0, + 808.0, + 2038.0, + 633.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2001.0, + 844.0, + 2001.0, + 844.0, + 2038.0, + 837.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1958.0, + 1220.0, + 1958.0, + 1220.0, + 1985.0, + 1175.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 988.0, + 678.0, + 988.0, + 678.0, + 1030.0, + 294.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 988.0, + 1411.0, + 988.0, + 1411.0, + 1030.0, + 792.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1022.0, + 764.0, + 1022.0, + 764.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1022.0, + 936.0, + 1022.0, + 936.0, + 1064.0, + 848.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1022.0, + 1063.0, + 1022.0, + 1063.0, + 1064.0, + 956.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1022.0, + 1297.0, + 1022.0, + 1297.0, + 1064.0, + 1287.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1151.0, + 457.0, + 1151.0, + 457.0, + 1196.0, + 296.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1151.0, + 558.0, + 1151.0, + 558.0, + 1196.0, + 481.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1151.0, + 1037.0, + 1151.0, + 1037.0, + 1196.0, + 782.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1151.0, + 1094.0, + 1151.0, + 1094.0, + 1196.0, + 1065.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1151.0, + 1400.0, + 1151.0, + 1400.0, + 1196.0, + 1179.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 634.0, + 1402.0, + 634.0, + 1402.0, + 667.0, + 296.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 664.0, + 1252.0, + 664.0, + 1252.0, + 696.0, + 324.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 1403.0, + 552.0, + 1403.0, + 588.0, + 295.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 580.0, + 916.0, + 580.0, + 916.0, + 616.0, + 322.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1268.0, + 260.0, + 1268.0, + 295.0, + 322.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 467.0, + 1405.0, + 467.0, + 1405.0, + 511.0, + 295.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 502.0, + 1045.0, + 502.0, + 1045.0, + 536.0, + 321.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 309.0, + 1403.0, + 309.0, + 1403.0, + 345.0, + 296.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 341.0, + 656.0, + 341.0, + 656.0, + 375.0, + 322.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 388.0, + 1405.0, + 388.0, + 1405.0, + 427.0, + 298.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 419.0, + 1340.0, + 419.0, + 1340.0, + 456.0, + 324.0, + 456.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 503, + 1406, + 503, + 1406, + 807, + 296, + 807 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 1181, + 1405, + 1181, + 1405, + 1363, + 296, + 1363 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 342, + 1404, + 342, + 1404, + 486, + 298, + 486 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1405, + 228, + 1405, + 335, + 297, + 335 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 814, + 1406, + 814, + 1406, + 960, + 297, + 960 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 653, + 1092, + 1037, + 1092, + 1037, + 1168, + 653, + 1168 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 295, + 1013, + 1400, + 1013, + 1400, + 1077, + 295, + 1077 + ], + "score": 0.949 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.912 + }, + { + "category_id": 8, + "poly": [ + 627, + 956, + 1070, + 956, + 1070, + 995, + 627, + 995 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2112, + 835, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 13, + "poly": [ + 432, + 568, + 655, + 568, + 655, + 608, + 432, + 608 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 881, + 642, + 1101, + 642, + 1101, + 677, + 881, + 677 + ], + "score": 0.94, + "latex": "| f ( { \\pmb x } ) - \\eta ( { \\pmb x } ) | \\leq 3 \\varepsilon" + }, + { + "category_id": 14, + "poly": [ + 651, + 1091, + 1043, + 1091, + 1043, + 1169, + 651, + 1169 + ], + "score": 0.94, + "latex": "\\tau ( { \\pmb x } ) = \\left\\{ \\begin{array} { l l } { \\tau _ { 1 0 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) \\geq 1 / 2 } \\\\ { \\tau _ { 0 1 } ( { \\pmb x } ) } & { \\eta ( { \\pmb x } ) < 1 / 2 } \\end{array} \\right. ." + }, + { + "category_id": 13, + "poly": [ + 1243, + 604, + 1396, + 604, + 1396, + 643, + 1243, + 643 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 609, + 1255, + 797, + 1255, + 797, + 1293, + 609, + 1293 + ], + "score": 0.93, + "latex": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq \\overline { { e } } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1290, + 468, + 1290, + 468, + 1329, + 298, + 1329 + ], + "score": 0.93, + "latex": "\\tau ( \\pmb { x } ) \\leq { \\frac { 1 } { 2 } } - e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1045, + 477, + 1045, + 477, + 1078, + 374, + 1078 + ], + "score": 0.93, + "latex": "( e _ { 0 } , f , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 735, + 534, + 865, + 534, + 865, + 572, + 735, + 572 + ], + "score": 0.93, + "latex": "\\bigl ( \\frac { 1 } { 2 } - T _ { 0 } , \\eta \\bigr )" + }, + { + "category_id": 13, + "poly": [ + 1128, + 405, + 1310, + 405, + 1310, + 444, + 1128, + 444 + ], + "score": 0.93, + "latex": "T _ { e n d } \\leq { \\textstyle \\frac { 1 } { 2 } } - 3 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 299, + 923, + 450, + 923, + 450, + 963, + 299, + 963 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 909, + 882, + 1147, + 882, + 1147, + 927, + 909, + 927 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log \\bigl ( \\frac { T _ { 0 } } { 3 \\varepsilon } \\bigr ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 419, + 1329, + 531, + 1329, + 531, + 1362, + 419, + 1362 + ], + "score": 0.92, + "latex": "( e _ { 0 } , f _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 575, + 745, + 659, + 745, + 659, + 778, + 575, + 778 + ], + "score": 0.92, + "latex": "( c _ { * } , c ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 561, + 640, + 783, + 640, + 783, + 678, + 561, + 678 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { | \\eta ( { \\pmb x } ) - \\frac 1 2 | \\geq \\frac 1 2 - T } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 616, + 678, + 755, + 678, + 755, + 714, + 616, + 714 + ], + "score": 0.92, + "latex": "T \\leq { \\frac { 1 } { 2 } } - 3 \\varepsilon" + }, + { + "category_id": 13, + "poly": [ + 838, + 1291, + 1074, + 1291, + 1074, + 1331, + 838, + 1331 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { e _ { 0 } = m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 741, + 1221, + 1022, + 1221, + 1022, + 1257, + 741, + 1257 + ], + "score": 0.92, + "latex": "\\alpha ( \\textstyle { \\frac { 1 } { 2 } } - e _ { o } + \\tau ) + \\varepsilon \\le e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1020, + 714, + 1077, + 714, + 1077, + 747, + 1020, + 747 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1189, + 1222, + 1247, + 1222, + 1247, + 1255, + 1189, + 1255 + ], + "score": 0.92, + "latex": "f ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1172, + 675, + 1401, + 675, + 1401, + 713, + 1172, + 713 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { T \\ge \\frac { 1 } { 2 } - 3 \\varepsilon = T _ { e n d } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 477, + 1257, + 534, + 1257, + 534, + 1290, + 477, + 1290 + ], + "score": 0.92, + "latex": "\\eta ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 1154, + 714, + 1351, + 714, + 1351, + 749, + 1154, + 749 + ], + "score": 0.91, + "latex": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | < 3 \\varepsilon" + }, + { + "category_id": 13, + "poly": [ + 664, + 715, + 722, + 715, + 722, + 747, + 664, + 747 + ], + "score": 0.91, + "latex": "f ( { \\pmb x } )" + }, + { + "category_id": 13, + "poly": [ + 299, + 445, + 450, + 445, + 450, + 484, + 299, + 484 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\frac { \\varepsilon } { \\alpha \\ell } \\le \\beta \\le \\frac { 2 \\varepsilon } { \\alpha \\ell } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 294, + 521, + 294, + 521, + 337, + 298, + 337 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { \\bar { T } _ { 0 } ^ { \\cdots } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 625, + 953, + 1072, + 953, + 1072, + 995, + 625, + 995 + ], + "score": 0.91, + "latex": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 298, + 744, + 523, + 744, + 523, + 778, + 298, + 778 + ], + "score": 0.91, + "latex": "| \\eta ( { \\pmb x } ) - f ( { \\pmb x } ) | \\leq \\bar { 3 } \\varepsilon" + }, + { + "category_id": 13, + "poly": [ + 613, + 448, + 1045, + 448, + 1045, + 484, + 613, + 484 + ], + "score": 0.91, + "latex": "\\mathbb { P } _ { \\pmb { x } \\sim D } \\big [ y _ { f _ { f i n a l } ( \\pmb { x } ) } = \\eta ^ { * } ( \\pmb { x } ) \\big ] \\geq 1 - 3 c ^ { * } \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 295, + 1220, + 361, + 1220, + 361, + 1255, + 295, + 1255 + ], + "score": 0.91, + "latex": "( \\alpha , \\varepsilon )" + }, + { + "category_id": 13, + "poly": [ + 298, + 373, + 393, + 373, + 393, + 406, + 298, + 406 + ], + "score": 0.91, + "latex": "L ( e _ { 0 } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 510, + 405, + 742, + 405, + 742, + 446, + 510, + 446 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { m \\ge \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 847, + 468, + 847, + 468, + 885, + 298, + 885 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { m a x ( t _ { 0 } , \\frac { \\alpha + \\varepsilon } { 1 + 2 \\alpha } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 497, + 1295, + 584, + 1295, + 584, + 1325, + 497, + 1325 + ], + "score": 0.9, + "latex": "e _ { 0 } > t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 407, + 454, + 407, + 454, + 444, + 297, + 444 + ], + "score": 0.9, + "latex": "T _ { 0 } < \\frac { 1 } { 2 } - e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 885, + 1315, + 885, + 1315, + 919, + 1196, + 919 + ], + "score": 0.89, + "latex": "T _ { e n d } \\leq 3 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1016, + 1161, + 1016, + 1161, + 1047, + 1133, + 1047 + ], + "score": 0.89, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 798, + 406, + 1072, + 406, + 1072, + 447, + 798, + 447 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { N \\ge m + \\frac { 1 } { \\beta } \\log ( \\frac { 1 - 6 \\varepsilon } { 2 T _ { 0 } } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 404, + 261, + 506, + 261, + 506, + 295, + 404, + 295 + ], + "score": 0.88, + "latex": "\\textstyle { \\left( { 1 + { \\frac { \\varepsilon } { \\ell \\alpha } } } \\right) }" + }, + { + "category_id": 13, + "poly": [ + 764, + 262, + 796, + 262, + 796, + 293, + 764, + 293 + ], + "score": 0.88, + "latex": "T _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1067, + 375, + 1095, + 375, + 1095, + 406, + 1067, + 406 + ], + "score": 0.87, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1114, + 345, + 1143, + 345, + 1143, + 375, + 1114, + 375 + ], + "score": 0.87, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 385, + 778, + 438, + 778, + 438, + 805, + 385, + 805 + ], + "score": 0.87, + "latex": "3 c ^ { * } \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1336, + 819, + 1403, + 819, + 1403, + 848, + 1336, + 848 + ], + "score": 0.87, + "latex": "e _ { 0 } =" + }, + { + "category_id": 13, + "poly": [ + 597, + 375, + 625, + 375, + 625, + 405, + 597, + 405 + ], + "score": 0.86, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1182, + 572, + 1203, + 572, + 1203, + 604, + 1182, + 604 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1372, + 1293, + 1401, + 1293, + 1401, + 1326, + 1372, + 1326 + ], + "score": 0.85, + "latex": "f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 469, + 1183, + 1252, + 1183, + 1252, + 1222, + 469, + 1222 + ], + "score": 0.85, + "latex": "\\begin{array} { r } { | \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } , \\mathbb { P } _ { z } [ \\tilde { y } _ { z } \\neq \\eta ^ { * } ( z ) | | \\eta ( z ) - \\frac { 1 } { 2 } | \\geq e _ { 0 } ] \\leq \\frac { 1 } { 2 } - e _ { 0 } + \\tau ( z ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 935, + 849, + 955, + 849, + 955, + 879, + 935, + 879 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 954, + 537, + 974, + 537, + 974, + 568, + 954, + 568 + ], + "score": 0.82, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 470, + 1183, + 657, + 1183, + 657, + 1222, + 470, + 1222 + ], + "score": 0.82, + "latex": "| \\eta ( \\pmb { x } ) - \\frac { 1 } { 2 } | \\geq e _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 931, + 746, + 957, + 746, + 957, + 772, + 931, + 772 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 883, + 573, + 907, + 573, + 907, + 599, + 883, + 599 + ], + "score": 0.8, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1218, + 819, + 1243, + 819, + 1243, + 847, + 1218, + 847 + ], + "score": 0.8, + "latex": "t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1219, + 511, + 1247, + 511, + 1247, + 534, + 1219, + 534 + ], + "score": 0.74, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 1175, + 749, + 1205, + 749, + 1205, + 772, + 1175, + 772 + ], + "score": 0.73, + "latex": "3 \\varepsilon" + }, + { + "category_id": 13, + "poly": [ + 866, + 823, + 886, + 823, + 886, + 844, + 866, + 844 + ], + "score": 0.71, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 404, + 883, + 587, + 883, + 587, + 923, + 404, + 923 + ], + "score": 0.7, + "latex": "( I ) T _ { 0 } < { \\textstyle \\frac { 1 } { 2 } } - e _ { 0 } " + }, + { + "category_id": 13, + "poly": [ + 634, + 882, + 859, + 882, + 859, + 925, + 634, + 925 + ], + "score": 0.68, + "latex": "\\begin{array} { r } { m \\geq \\frac { \\ell \\alpha } { \\varepsilon } \\log \\bigl ( \\frac { 2 T _ { 0 } } { 1 - 2 e _ { 0 } } \\bigr ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 859, + 850, + 875, + 850, + 875, + 875, + 859, + 875 + ], + "score": 0.54, + "latex": "^ { l }" + }, + { + "category_id": 13, + "poly": [ + 674, + 346, + 690, + 346, + 690, + 371, + 674, + 371 + ], + "score": 0.53, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 671, + 819, + 687, + 819, + 687, + 844, + 671, + 844 + ], + "score": 0.4, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 1366, + 374, + 1403, + 374, + 1403, + 405, + 1366, + 405 + ], + "score": 0.34, + "latex": "( l )" + }, + { + "category_id": 13, + "poly": [ + 894, + 376, + 910, + 376, + 910, + 401, + 894, + 401 + ], + "score": 0.27, + "latex": "I" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 502.0, + 1218.0, + 502.0, + 1218.0, + 540.0, + 293.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 502.0, + 1404.0, + 502.0, + 1404.0, + 540.0, + 1248.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 534.0, + 734.0, + 534.0, + 734.0, + 573.0, + 293.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 534.0, + 953.0, + 534.0, + 953.0, + 573.0, + 866.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 534.0, + 1406.0, + 534.0, + 1406.0, + 573.0, + 975.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 568.0, + 431.0, + 568.0, + 431.0, + 611.0, + 293.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 568.0, + 882.0, + 568.0, + 882.0, + 611.0, + 656.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 568.0, + 1181.0, + 568.0, + 1181.0, + 611.0, + 908.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 568.0, + 1407.0, + 568.0, + 1407.0, + 611.0, + 1204.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 589.0, + 1242.0, + 589.0, + 1242.0, + 660.0, + 284.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 589.0, + 1419.0, + 589.0, + 1419.0, + 660.0, + 1397.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 641.0, + 560.0, + 641.0, + 560.0, + 680.0, + 294.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 641.0, + 880.0, + 641.0, + 880.0, + 680.0, + 784.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 641.0, + 1407.0, + 641.0, + 1407.0, + 680.0, + 1102.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 675.0, + 615.0, + 675.0, + 615.0, + 718.0, + 293.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 675.0, + 1171.0, + 675.0, + 1171.0, + 718.0, + 756.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 675.0, + 1407.0, + 675.0, + 1407.0, + 718.0, + 1402.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 711.0, + 663.0, + 711.0, + 663.0, + 753.0, + 293.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 711.0, + 1019.0, + 711.0, + 1019.0, + 753.0, + 723.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 711.0, + 1153.0, + 711.0, + 1153.0, + 753.0, + 1078.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 711.0, + 1407.0, + 711.0, + 1407.0, + 753.0, + 1352.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 742.0, + 574.0, + 742.0, + 574.0, + 781.0, + 524.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 742.0, + 930.0, + 742.0, + 930.0, + 781.0, + 660.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 742.0, + 1174.0, + 742.0, + 1174.0, + 781.0, + 958.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 742.0, + 1407.0, + 742.0, + 1407.0, + 781.0, + 1206.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 775.0, + 384.0, + 775.0, + 384.0, + 808.0, + 294.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 775.0, + 451.0, + 775.0, + 451.0, + 808.0, + 439.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 775.0, + 1406.0, + 775.0, + 1406.0, + 807.0, + 1373.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1179.0, + 468.0, + 1179.0, + 468.0, + 1223.0, + 293.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1179.0, + 1407.0, + 1179.0, + 1407.0, + 1223.0, + 1253.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1215.0, + 294.0, + 1215.0, + 294.0, + 1258.0, + 291.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1215.0, + 740.0, + 1215.0, + 740.0, + 1258.0, + 362.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1215.0, + 1188.0, + 1215.0, + 1188.0, + 1258.0, + 1023.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1215.0, + 1406.0, + 1215.0, + 1406.0, + 1258.0, + 1248.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1256.0, + 476.0, + 1256.0, + 476.0, + 1295.0, + 294.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1256.0, + 608.0, + 1256.0, + 608.0, + 1295.0, + 535.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1256.0, + 1406.0, + 1256.0, + 1406.0, + 1295.0, + 798.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1282.0, + 297.0, + 1282.0, + 297.0, + 1341.0, + 290.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1282.0, + 496.0, + 1282.0, + 496.0, + 1341.0, + 469.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1282.0, + 837.0, + 1282.0, + 837.0, + 1341.0, + 585.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1282.0, + 1371.0, + 1282.0, + 1371.0, + 1341.0, + 1075.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1282.0, + 1412.0, + 1282.0, + 1412.0, + 1341.0, + 1402.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1325.0, + 418.0, + 1325.0, + 418.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1325.0, + 1149.0, + 1325.0, + 1149.0, + 1364.0, + 532.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1329.0, + 1406.0, + 1329.0, + 1406.0, + 1359.0, + 1372.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 342.0, + 673.0, + 342.0, + 673.0, + 377.0, + 294.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 342.0, + 1113.0, + 342.0, + 1113.0, + 377.0, + 691.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 342.0, + 1408.0, + 342.0, + 1408.0, + 377.0, + 1144.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 371.0, + 296.0, + 371.0, + 296.0, + 467.0, + 279.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 371.0, + 509.0, + 371.0, + 509.0, + 467.0, + 455.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 371.0, + 797.0, + 371.0, + 797.0, + 467.0, + 743.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 371.0, + 1127.0, + 371.0, + 1127.0, + 467.0, + 1096.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 371.0, + 1365.0, + 371.0, + 1365.0, + 467.0, + 1311.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 371.0, + 1412.0, + 371.0, + 1412.0, + 467.0, + 1404.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 438.0, + 298.0, + 438.0, + 298.0, + 491.0, + 289.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 438.0, + 612.0, + 438.0, + 612.0, + 491.0, + 451.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 438.0, + 1063.0, + 438.0, + 1063.0, + 491.0, + 1046.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1407.0, + 226.0, + 1407.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 250.0, + 403.0, + 250.0, + 403.0, + 307.0, + 289.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 250.0, + 763.0, + 250.0, + 763.0, + 307.0, + 507.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 250.0, + 1412.0, + 250.0, + 1412.0, + 307.0, + 797.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 297.0, + 614.0, + 297.0, + 614.0, + 329.0, + 522.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 297.0, + 1405.0, + 297.0, + 1405.0, + 329.0, + 1372.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 287.5, + 518.0, + 287.5, + 518.0, + 341.5, + 289.0, + 341.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 810.0, + 670.0, + 810.0, + 670.0, + 854.0, + 292.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 810.0, + 865.0, + 810.0, + 865.0, + 854.0, + 688.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 810.0, + 1217.0, + 810.0, + 1217.0, + 854.0, + 887.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 810.0, + 1335.0, + 810.0, + 1335.0, + 854.0, + 1244.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 833.0, + 297.0, + 833.0, + 297.0, + 897.0, + 287.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 833.0, + 858.0, + 833.0, + 858.0, + 897.0, + 469.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 833.0, + 934.0, + 833.0, + 934.0, + 897.0, + 876.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 833.0, + 1417.0, + 833.0, + 1417.0, + 897.0, + 956.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 866.0, + 403.0, + 866.0, + 403.0, + 945.0, + 285.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 866.0, + 633.0, + 866.0, + 633.0, + 945.0, + 588.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 866.0, + 908.0, + 866.0, + 908.0, + 945.0, + 860.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 866.0, + 1195.0, + 866.0, + 1195.0, + 945.0, + 1148.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 866.0, + 1416.0, + 866.0, + 1416.0, + 945.0, + 1316.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 922.0, + 298.0, + 922.0, + 298.0, + 963.0, + 293.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 922.0, + 572.0, + 922.0, + 572.0, + 963.0, + 451.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1010.0, + 1132.0, + 1010.0, + 1132.0, + 1049.0, + 294.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1010.0, + 1404.0, + 1010.0, + 1404.0, + 1049.0, + 1162.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1044.0, + 373.0, + 1044.0, + 373.0, + 1078.0, + 293.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1044.0, + 641.0, + 1044.0, + 641.0, + 1078.0, + 478.0, + 1078.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/qda7-sVg84/qda7-sVg84.md b/parse/train/qda7-sVg84/qda7-sVg84.md new file mode 100644 index 0000000000000000000000000000000000000000..136d66aa37ff44f63fe55134010b8ae2264f7a8e --- /dev/null +++ b/parse/train/qda7-sVg84/qda7-sVg84.md @@ -0,0 +1,775 @@ +# CONTRASTIVE BEHAVIORAL SIMILARITY EMBEDDINGS FOR GENERALIZATION IN REINFORCEMENT LEARNING + +Rishabh Agarwal∗ Marlos C. Machado‡ Pablo Samuel Castro Marc G. Bellemare Google Research, Brain Team {rishabhagarwal, marlosm, psc, bellemare}@google.com + +# ABSTRACT + +Reinforcement learning methods trained on few environments rarely learn policies that generalize to unseen environments. To improve generalization, we incorporate the inherent sequential structure in reinforcement learning into the representation learning process. This approach is orthogonal to recent approaches, which rarely exploit this structure explicitly. Specifically, we introduce a theoretically motivated policy similarity metric (PSM) for measuring behavioral similarity between states. PSM assigns high similarity to states for which the optimal policies in those states as well as in future states are similar. We also present a contrastive representation learning procedure to embed any state similarity metric, which we instantiate with PSM to obtain policy similarity embeddings $( \mathrm { P S E s ^ { 1 } } )$ ). We demonstrate that PSEs improve generalization on diverse benchmarks, including LQR with spurious correlations, a jumping task from pixels, and Distracting DM Control Suite. + +# 1 INTRODUCTION + +Current reinforcement learning (RL) approaches often learn policies that do not generalize to environments different than those the agent was trained on, even when these environments are semantically equivalent (Tachet des Combes et al., 2018; Song et al., 2019; Cobbe et al., 2019). For example, consider a jumping task where an agent, learning from pixels, needs to jump over an obstacle (Figure 1). Deep RL agents trained on a few of these tasks with different obstacle positions struggle to solve test tasks where obstacles are at previously unseen locations. + +![](images/abed9d708ee10498bdd0d0bf48ab3b19c58d372b5155e6350f85107e57f370fa.jpg) +Figure 1: Jumping task: The agent (white block), learning from pixels, needs to jump over an obstacle (grey square). The challenge is to generalize to unseen obstacle positions and floor heights in test tasks using a small number of training tasks. We show the agent’s trajectories using faded blocks. + +Recent solutions to circumvent poor generalization in RL are adapted from supervised learning, and, as such, largely ignore the sequential aspect of RL. Most of these solutions revolve around enhancing the learning process, including data augmentation (e.g., Kostrikov et al., 2020; Lee et al., 2020a), regularization (Cobbe et al., 2019; Farebrother et al., 2018), noise injection (Igl et al., 2019), and diverse training conditions (Tobin et al., 2017); they rarely exploit properties of the sequential decision making problem such as similarity in actions across temporal observations. + +Instead, we tackle generalization by incorporating properties of the RL problem into the representation learning process. Our approach exploits the fact that an agent, when operating in environments with similar underlying mechanics, exhibits at least short sequences of behaviors that are similar across these environments. Concretely, the agent is optimized to learn an embedding in which states are close when the agent’s optimal policies in these states and future states are similar. This notion of proximity is general and it is applicable to observations from different environments. + +Specifically, inspired by bisimulation metrics (Castro, 2020; Ferns et al., 2004), we propose a novel policy similarity metric (PSM). PSM (Section 3) defines a notion of similarity between states originated from different environments by the proximity of the long-term optimal behavior from these states. PSM is reward-agnostic, making it more robust for generalization compared to approaches that rely on reward information. We prove that PSM yields an upper bound on suboptimality of policies transferred from one environment to another (Theorem 1), which is not attainable with bisimulation. + +We employ PSM for representation learning and introduce policy similarity embeddings (PSEs) for deep RL. To do so, we present a general contrastive procedure (Section 4) to learn an embedding based on any state similarity metric. PSEs are the instantiation of this procedure with PSM. PSEs are appealing for generalization as they encode task-relevant invariances by putting behaviorally equivalent states together. This is unlike prior approaches, which rely on capturing such invariances without being explicitly trained to do so, for example, through value function similarities across states (e.g., Castro & Precup, 2010), or being robust to fixed transformations of the observation space (e.g., Kostrikov et al., 2020; Laskin et al., 2020a). + +PSEs lead to better generalization while being orthogonal to how most of the field has been tackling generalization. We illustrate the efficacy and broad applicability of our approach on three existing benchmarks specifically designed to test generalization: (i) jumping task from pixels (Tachet des Combes et al., 2018) (Section 5), (ii) LQR with spurious correlations (Song et al., 2019) (Section 6.1), and (iii) Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b) (Section 6.2). Our approach improves generalization compared to a wide variety of approaches including standard regularization (Farebrother et al., 2018; Cobbe et al., 2019), bisimulation (Castro & Precup, 2010; Castro, 2020; Zhang et al., 2021), out-of-distribution generalization (Arjovsky et al., 2019) and state-of-the-art data augmentation (Kostrikov et al., 2020; Lee et al., 2020a). + +# 2 PRELIMINARIES + +We describe an environment as a Markov decision process (MDP) (Puterman, 1994) $\begin{array} { r l } { \mathcal { M } } & { { } = } \end{array}$ $( \mathcal { X } , \mathcal { A } , R , P , \gamma )$ , with a state space $\mathcal { X }$ , an action space $\mathcal { A }$ , a reward function $R$ , transition dynamics $P$ , and a discount factor $\gamma \in [ 0 , 1 )$ . A policy $\pi ( \cdot \mid x )$ maps states $x \in \mathcal { X }$ to distributions over actions. Whenever convenient, we abuse notation and write $\pi ( x )$ to describe the probability distribution $\pi ( \cdot | x )$ , treating $\pi ( x )$ as a vector. In RL, the goal is to find an optimal policy $\pi ^ { * }$ that maximizes the cumulative expected return $\mathbb E _ { a _ { t } \sim \pi ( \cdot \textit { \textbf { \ i } } x _ { t } ) } [ \sum _ { t } \bar { \gamma } ^ { t } R ( x _ { t } , a _ { t } ) ]$ starting from an initial state $x _ { 0 }$ . + +We are interested in learning a policy that generalizes across related environments. We formalize this by considering a collection $\rho$ of MDPs, sharing an action space $\mathcal { A }$ but with disjoint state spaces. We use $\mathcal { X }$ and $\mathcal { V }$ to denote the state spaces of specific environments, and write $R _ { \mathcal { X } }$ , $P _ { \mathcal { X } }$ for the reward and transition functions of the MDP whose state space is $\mathcal { X }$ , and $\pi _ { \mathcal { X } } ^ { * }$ for its optimal policy, which we assume unique without loss of generality. For a given policy $\pi$ , we further specialize these into $R _ { \mathcal { X } } ^ { \pi }$ and $P _ { \mathcal { X } } ^ { \pi }$ , the reward and state-to-state transition dynamics arising from following $\pi$ in that MDP. + +We write $s$ for the union of the state spaces of the MDPs in $\rho$ . Concretely, different MDPs correspond to specific scenarios in a problem class (Figure 1), and $s$ is the space of all possible configurations. Used without subscripts, $R , P$ , and $\pi$ refer to the reward and transition function of this “union MDP”, and a policy defined over $s$ ; this notation simplifies the exposition. We measure distances between states across environments using pseudometrics2 on $s$ ; the set of all such pseudometrics is $\mathbb { M }$ , and $\mathbb { M } _ { p }$ is the set of metrics on probability distributions over $s$ . + +In our setting, the learner has access to a collection of training MDPs $\{ \mathcal { M } _ { i } \} _ { i = 1 } ^ { N }$ , drawn from $\rho$ . After interacting with these environments, the learner must produce a policy $\pi$ over the entire state space $s$ , which is then evaluated on unseen MDPs from $\rho$ . Similar in spirit to the setting of transfer learning (Taylor & Stone, 2009), here we evaluate the policy’s zero-shot performance on $\rho$ . + +Our policy similarity metric (Section 3) builds on the concept of $\pi$ -bisimulation (Castro, 2020). Under the $\pi$ -bisimulation metric, the distance between two states, $x$ and $y$ , is defined in terms of the difference between the expected rewards obtained when following policy $\pi$ . The $\pi$ -bisimulation metric $d _ { \pi }$ satisfies a recursive equation based on the 1-Wasserstein metric $\mathcal { W } _ { 1 } : \mathbb { M } \mathbb { M } _ { p }$ , where $\mathcal { W } _ { 1 } ( d ) ( A , B )$ is the minimal cost of transporting probability mass from $A$ to $B$ (two probability distributions on $s$ ) under the base metric $d$ (Villani, 2008). The recursion is + +$$ +\begin{array} { r } { d _ { \pi } ( x , y ) = | R ^ { \pi } ( x ) - R ^ { \pi } ( y ) | + \gamma \mathcal { W } _ { 1 } ( d _ { \pi } ) \big ( P ^ { \pi } ( \cdot | x ) , P ^ { \pi } ( \cdot | y ) \big ) , \qquad x , y \in \mathcal { S } . } \end{array} +$$ + +To achieve good generalization properties, we learn an embedding function $z _ { \theta } : \mathcal { S } \mathbb { R } ^ { k }$ that reflects the information encoded in the policy similarity metric; this yields a policy similarity embedding (Section 4). We use contrastive methods (Hadsell et al., 2006; Oord et al., 2018) whose track record for representation learning is well-established. We adapt SimCLR (Chen et al., 2020), a popular contrastive method for learning embeddings of image inputs. Given two inputs $x$ and $y$ , their embedding similarity is $s _ { \theta } ( x , y ) = s i m ( z _ { \theta } ( x ) , z _ { \theta } ( y ) )$ , where $\begin{array} { r } { s i m ( u , v ) = \frac { u ^ { T } v } { \| u \| \| v \| } } \end{array}$ denotes the cosine similarity function. SimCLR aims to maximize similarity between augmented versions of an image (e.g., crops, colour changes) while minimizing its similarity to other images. The loss used by SimCLR for two versions $x , y$ of an image, and a set $\mathcal { X } ^ { \prime }$ containing other images is: + +$$ +\ell _ { \theta } ( x , y ; \mathcal { X } ^ { \prime } ) = - \log \frac { \exp ( \lambda s _ { \theta } ( x , y ) ) } { \exp ( \lambda s _ { \theta } ( x , y ) ) + \sum _ { x ^ { \prime } \in \mathcal { X } ^ { \prime } \backslash \left\{ x \right\} } \exp ( \lambda s _ { \theta } ( x ^ { \prime } , y ) ) } +$$ + +where $\lambda$ is an inverse temperature hyperparameter. The overall SimCLR loss is then the expected value of $\ell _ { \theta } ( x , y ; S )$ , when $x , y$ , and $\mathcal { X } ^ { \prime }$ are drawn from some augmented training distribution. + +# 3 POLICY SIMILARITY METRIC + +A useful tool in learning a policy that generalizes is to understand which states result in similar behavior, and which do not. To be maximally effective, this similarity should go beyond the immediately chosen action and consider long-term behavior. In this regards, the $\pi$ -bisimulation metrics are interesting as they are based on the full sequence of future rewards received from different states. However, considering rewards can be both too restrictive (when the policies are the same, but the obtained rewards are not; see Figure D.1) or too permissive (when the policies are different, but the obtained rewards are not; see Figure 3a). In fact, $\pi$ -bisimulation metrics actually lead to poor generalization in our experiments (Sections 5.1 and 5.2). + +To address this issue, we instead consider the similarity between policies themselves. We replace the absolute reward difference by a probability pseudometric between policies, denoted DIST. Additionally, since we would like to perform well in unseen environments, we are interested in similarity in optimal behavior. We thus use $\pi ^ { * }$ as the grounding policy. This yields the policy similarity metric (PSM), for which states are close when the optimal policies in these states and future states are similar. For a given DIST, the PSM $d ^ { * } : \mathcal { S } \times \mathcal { S } \to \mathbb { R }$ satisfies the recursive equation + +$$ +\begin{array} { r l r } & { } & { d ^ { * } ( x , y ) = \underbrace { \mathrm { D I S T } \big ( \pi ^ { * } ( x ) , \pi ^ { * } ( y ) \big ) } _ { \mathrm { ( A ) } } + \underbrace { \gamma \mathcal { W } _ { 1 } ( d ^ { * } ) \big ( P ^ { \pi ^ { * } } ( \cdot | x ) , P ^ { \pi ^ { * } } ( \cdot | y ) \big ) } _ { \mathrm { ( B ) } } . } \end{array} +$$ + +The DIST term captures the difference in local optimal behavior (A) while $\mathcal { W } _ { 1 }$ captures long-term optimal behavior difference (B); the exact weights assigned to the two are given by the discount factor. Furthermore, when DIST is bounded, $d ^ { * }$ is guaranteed to be finite. While there are technically multiple PSMs (one for each DIST), we omit this distinction whenever clear from context. A proof of the uniqueness of $d ^ { * }$ is given in Proposition B.1. + +Our main use of PSM will be to compare states across environments. In this context, we identify the terms in Equation 3 with specific environments for clarity and write (despite its technical inaccuracy) + +$$ +\begin{array} { r } { d ^ { * } ( x , y ) = \mathrm { D I S T } \big ( \pi _ { \mathcal { X } } ^ { * } ( x ) , \pi _ { \mathcal { Y } } ^ { * } ( y ) \big ) + \gamma \mathcal { W } _ { 1 } \big ( d ^ { * } ) \big ( P _ { \mathcal { X } } ^ { \pi ^ { * } } ( \cdot | x ) , P _ { \mathcal { Y } } ^ { \pi ^ { * } } ( \cdot | y ) \big ) . } \end{array} +$$ + +PSM is applicable to both discrete and continuous action spaces. In our experiments, DIST is the total variation distance $( T V )$ when $\mathcal { A }$ is discrete, and we use the $\ell _ { 1 }$ distance between the mean actions of the two policies when $\mathcal { A }$ is continuous. PSM can be computed iteratively using dynamic programming (Ferns et al., 2011) (more details in Section D.1). Furthermore, when $\pi ^ { * }$ is unavailable on training environments, we replace it by an approximation $\hat { \pi } ^ { * }$ to obtain an approximate PSM, which is close to the exact PSM depending on the suboptimality of $\hat { \pi } ^ { * }$ (Proposition D.3). + +Despite resembling $\pi$ -bisimulation metrics in form, the PSM possesses different characteristics which are better suited to the problem of generalizing a learned policy. To illustrate this point, consider the following simple nearest-neighbour scheme: Given a state $y \in \mathcal { V }$ , denote its closest match in $\mathcal { X }$ by $\begin{array} { r } { \tilde { x } _ { y } : = \arg \operatorname* { m i n } _ { x \in \mathcal { X } } d ^ { * } ( x , y ) } \end{array}$ . Suppose that we use this scheme to transfer $\pi _ { \mathcal { X } } ^ { * }$ to $\mathcal { M } _ { \mathcal { Y } }$ , in the sense that we behave according to the policy $\tilde { \pi } ( y ) = \pi _ { \chi } ^ { * } ( \tilde { x } _ { y } )$ . We can then bound the difference between $\tilde { \pi }$ and $\pi _ { \mathcal { Y } } ^ { * }$ , something that is not possible if $d ^ { * }$ is replaced by a $\pi$ -bisimulation metric. + +Theorem 1. [Bound on policy transfer] For any $y \in \mathcal { V }$ , let $Y _ { y } ^ { t } \sim P ^ { \tilde { \pi } } ( \cdot | Y _ { y } ^ { t - 1 } )$ define the sequence of random states encountered starting in $Y _ { y } ^ { 0 } = y$ and following policy $\tilde { \pi }$ . We have: + +$$ +\mathbb { E } _ { Y _ { y } ^ { t } } \left[ \sum _ { t \ge 0 } \gamma ^ { t } T V \left( \tilde { \pi } ( Y _ { y } ^ { t } ) , \pi ^ { * } ( Y _ { y } ^ { t } ) \right) \right] \le \frac { 1 + \gamma } { 1 - \gamma } d ^ { * } ( \tilde { x } _ { y } , y ) . +$$ + +The proof is in the Appendix (Section B). Theorem 1 is non-vacuous whenever $d ^ { * } ( \tilde { x } _ { y } , y ) < 1 / ( 1 + \gamma )$ . In particular, $d ^ { * } ( \tilde { x } _ { y } , y ) = 0$ implies that the transferred policy is optimal. Although this scheme is not practical (computing $d ^ { * }$ requires knowledge of $\pi _ { \mathcal { D } } ^ { * }$ ), it shows that meaningful policy generalization can be obtained if we can find a mapping that generalizes across $s$ . Put another way, PSM gives us a principled way of lifting generalization across inputs (a supervised learning problem) to generalization across environments. We now describe how to employ PSM for learning representations that put together states in which the agent’s long-term optimal behavior is similar. + +# 4 LEARNING CONTRASTIVE METRIC EMBEDDINGS + +To generalize a learned policy to new environments, we build on the success of contrastive representations (Section 2). Given a state similarity metric $d$ , we develop a general procedure (Algorithm 1) to learn contrastive metric embeddings (CMEs) for $d$ . We utilize the metric $d$ for defining the set of positive and negative pairs, as well as assigning importance weights to these pairs in the contrastive loss (Equation 4). + +# Algorithm 1 Contrastive Metric Embeddings (CMEs) + +1: Given: State embedding $z _ { \theta } ( \cdot )$ , Metric $d ( \cdot , \cdot )$ Training environments $\{ \mathcal { M } _ { i } \} _ { i = 1 } ^ { N }$ · · · . Hyperparameters: Temperature $1 / \lambda$ , Scale $\beta$ , Total training steps $K$ +2: for each step $k = 1 , \ldots , K$ do +3: Sample a pair of training MDPs $\mathcal { M } _ { \mathcal { X } } , \mathcal { M } _ { \mathcal { Y } }$ +4: Update $\theta$ to minimize ${ \mathcal { L } } _ { \mathrm { C M E } }$ where $\mathcal { L } _ { \mathrm { C M E } } = \mathbb { E } _ { \mathcal { M } _ { \mathcal { X } } , \mathcal { M } _ { \mathcal { Y } } \sim \rho } \left[ L _ { \theta } ( \mathcal { M } _ { \mathcal { X } } , \mathcal { M } _ { \mathcal { Y } } ) \right]$ +5: end for + +We first apply a transformation to convert $d$ to a similarity measure $\Gamma$ , bounded in [0, 1] for “soft” similarities. In this work, we transform $d$ using the Gaussian kernel with a positive scale parameter $\beta$ , that is, $\Gamma ( x , y ) = \exp ( - d ( x , y ) / \beta )$ . $\beta$ controls the sensitivity of the similarity measure to $d$ . + +Second, we select the positive and negative pairs given a set of states ${ \mathcal { X } } ^ { \prime } \subseteq { \mathcal { X } }$ and $\mathcal { V }$ from MDPs $\mathcal { M } _ { \mathcal { X } } , \mathcal { M } _ { \mathcal { Y } }$ respectively. For each anchor state $y \in \mathcal { V }$ , we use its nearest neighbor in $\mathcal { X } ^ { \prime }$ based on the similarity measure $\Gamma$ to define the positive pairs $\{ ( \tilde { x } _ { y } , y ) \}$ , where $\tilde { x } _ { y } = \underset { x \in \mathcal { X } ^ { \prime } } { \operatorname { a r g m a x } } \Gamma ( x , y )$ . The remaining states in $\mathcal { X } ^ { \prime }$ , paired with $y$ , are used as negative pairs. This choice of pairs is motivated by Theorem 1, which shows that if we transfer the optimal policy in $\mathcal { M } _ { \mathcal { X } }$ to the nearest neighbors defined using PSM, its performance in $\mathcal { M } _ { \mathcal { Y } }$ has suboptimality bounded by PSM. + +Next, we define a soft version of the SimCLR contrastive loss (Equation 2) for learning the function $z _ { \theta }$ , which maps states (usually high-dimensional) to embeddings. Given a positive state pair $( \tilde { x } _ { y } , y )$ , the set $\mathcal { X } ^ { \prime }$ , and the similarity measure $\Gamma$ , the loss (pseudocode provided in Section J.2) is given by + +$$ +\ell _ { \boldsymbol { \theta } } ( \tilde { x } _ { y } , y ; \mathcal { X } ^ { \prime } ) = - \log \frac { \Gamma ( \tilde { x } _ { y } , y ) \exp ( \lambda s _ { \boldsymbol { \theta } } ( \tilde { x } _ { y } , y ) ) } { \Gamma ( \tilde { x } _ { y } , y ) \exp ( \lambda s _ { \boldsymbol { \theta } } ( \tilde { x } _ { y } , y ) ) + \sum _ { \boldsymbol { x } ^ { \prime } \in \mathcal { X } ^ { \prime } \backslash \{ \tilde { x } _ { y } \} } ( 1 - \Gamma ( \boldsymbol { x } ^ { \prime } , y ) ) \exp ( \lambda s _ { \boldsymbol { \theta } } ( \boldsymbol { x } ^ { \prime } , y ) ) } +$$ + +where we use the same notation as in Equation 2. Following SimCLR, we use a non-linear projection of the representation as $z _ { \theta }$ (Figure A.1). The agent’s policy is an affine function of the representation. + +The total contrastive loss for $\mathcal { M } _ { \mathcal { X } }$ and $\mathcal { M } _ { \mathcal { Y } }$ utilizes the optimal trajectories $\tau _ { \mathcal { X } } ^ { * } = \{ x _ { t } \} _ { t = 1 } ^ { N }$ and $\tau _ { y } ^ { * } = \{ y _ { t } \} _ { t = 1 } ^ { N }$ , where $x _ { t + 1 } \sim P _ { \mathcal { X } } ^ { \pi ^ { * } } ( \cdot | x _ { t } )$ and $y _ { t + 1 } \sim P _ { y } ^ { \pi ^ { * } } ( \cdot | y _ { t } )$ . We set $\mathcal { X } ^ { \prime } = \tau _ { \mathcal { X } } ^ { * }$ and define + +$$ +L _ { \theta } ( \mathcal { M } _ { \mathcal { X } } , \mathcal { M } _ { \mathcal { Y } } ) = \mathbb { E } _ { y \sim \tau _ { \mathcal { Y } } ^ { * } } \left[ \ell _ { \theta } ( \tilde { x } _ { y } , y ; \tau _ { \mathcal { X } } ^ { * } ) \right] \qquad \mathrm { w h e r e } \ \tilde { x } _ { y } = \underset { x \in \tau _ { \mathcal { X } } ^ { * } } { \mathrm { a r g m a x } } \ \Gamma ( x , y ) . +$$ + +We refer to CMEs learned with policy similarity metric as policy similarity embeddings $( P S E s )$ . PSEs can trivially be combined with data augmentation by using augmented states when computing state embeddings. We simultaneously learn PSEs with the RL agent by adding ${ \mathcal { L } } _ { \mathrm { C M E } }$ (Algorithm 1) as an auxiliary objective during training. Next, we illustrate the benefits from this auxiliary objective. + +# 5 JUMPING TASK FROM PIXELS: A CASE STUDY + +Task Description. The jumping task (Tachet des Combes et al., 2018) (Figure 1) captures, using well-defined factors of variations, whether agents can learn the correct invariances required for generalization, directly from image inputs. The task consists of an agent trying to jump over an obstacle. The agent has access to two actions: right and jump. The agent needs to time the jump precisely, at a specific distance from the obstacle, otherwise it will eventually hit the obstacle. Different tasks consist in shifting the floor height and/or the obstacle position. To generalize, the agent needs to be invariant to the floor height while jump based on the obstacle position. The obstacle can be in 26 different locations while the floor has 11 different heights, totaling 286 tasks. + +![](images/fe5d30be2f8515ffad0c843ba9528ed9477ed5a3bbb49088ef023877b2fee29f.jpg) +Figure 2: Jumping Task: Visualization of average performance of PSEs with data augmentation across different configurations. We plot the median performance across 100 runs. Each tile in the grid represents a different task (obstacle position/floor height combination). For each grid configuration, the height varies along the $_ y$ -axis (11 heights) while the obstacle position varies along the $x -$ -axis (26 locations). The red letter $\top$ indicates the training tasks. Random grid depicts only one instance, each run consisted of a different test/train split. Beige tiles are tasks PSEs solved while black tiles are tasks PSEs did not solve when used with data augmentation. These results were chosen across all the 100 runs to demonstrate what the average reported performance looks like. + +Problem Setup. We split the problem into 18 seen (training) and 268 unseen (test) tasks to stress test generalization using a few changes in the underlying factors of variations seen during training. The small number of positive examples3 results in a highly unbalanced classification problem with low amounts of data, making it challenging without additional inductive biases. Thus, we evaluate generalization in regimes with and without data augmentation. The different grids configurations (Figure 2) capture different types of generalization: the “wide” grid tests generalization via “interpolation”, the “narrow” grid tests out-of-distribution generalization via “extrapolation”, and the random grid instances evaluate generalization similar to supervised learning where train and test samples are drawn i.i.d. from the same distribution. + +We used RandConv (Lee et al., 2020a), a state-of-the-art data augmentation for generalization. For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the “wide” grid (Figure 2a) and pick the parameters with the best validation performance. We use these fixed parameters for all grid configurations to show the robustness of PSEs to hyperparameter tuning. We first compute the optimal trajectories in the training tasks. Using these trajectories, we compute PSM using dynamic programming (Section D.1). We train the agent by imitation learning, combined with an auxiliary loss for PSEs (Section 4). More details are in Section G. + +# 5.1 EVALUATING GENERALIZATION ON JUMPING TASK + +We show the efficacy of PSEs compared to common generalization approaches such as regularization (e.g., Cobbe et al., 2019; Farebrother et al., 2018), and data augmentation (e.g., Lee et al., 2020a; Laskin et al., 2020a), which are quite effective on pixel-based RL tasks. We also contrast PSEs with bisimulation transfer (Castro & Precup, 2010), a tabular state-based transfer approach based on bisimulation metrics which does not do any learning and bisimulation preserving representations (Zhang et al., 2021), showing the advantages of PSM over a prevalent state similarity metric. + +![](images/f6cb8503f0d51f326842d71126a63d55028d7bdc17f24dcab3810ef2b38ded2e.jpg) +Figure 3: Embedding visualization. (a) Optimal trajectories on original jumping task (visualized as coloured blocks) with different obstacle positions. We visualize the hidden representations using UMAP, where the color of points indicate the tasks of the corresponding observations. Points with the same number label correspond to same distance of the agent from the obstacle, the underlying optimal invariant feature across tasks. + +We first investigated how well PSEs generalize over existing methods without incorporating additional domain knowledge during training. Table 1 summarizes, in the setting without data augmentation, the performance of these methods in different train/test splits $\cdot \cdot f .$ Figure 2 for a detailed description). PSEs, with only 18 examples, already leads to better performance than standard regularization. + +PSEs also outperform bisimulation transfer in the “wide” and random grids. Although bisimulation transfer is impractical4 when evaluating zero-shot generalization, we still performed this comparison, unfair to PSEs, to highlight their efficacy. PSEs perform better because, in contrast to bisimulation, PSM is reward agnostic ( $( c . f .$ Proposition D.1) – the expected return of the jump action is quite different depending on the obstacle position $( c . f .$ Figure G.2 for a visual juxtaposition of PSM and bisimulation). Overall, these results are promising because they place PSEs as an effective generalization method that does not rely on data augmentation. + +Nevertheless, PSEs are complementary to data augmentation, which consistently improves generalization in deep RL. We compared RandConv combined with PSEs to simply using RandConv. Domain-specific augmentation also succeeds in the jumping task. Thus, it is not surprising that RandConv is so effective compared to techniques without augmentation. Table 1 ( $2 ^ { n d }$ row) shows that PSEs substantially improve the performance of RandConv across all grid configurations. Moreover, Table 1 ( $2 ^ { n d }$ row) illustrates that when combined with RandConv, bisimulation preserving representations (Zhang et al., 2021) diminish generalization by $3 0 - 5 0 \%$ relative to PSEs. + +Notably, Table 1 ( $\cdot ^ { 1 ^ { s t } }$ row) indicates that learning-based methods are ineffective on the “narrow” grid without data augmentation. That said, PSEs do work quite well when combined with RandConv. However, even with data augmentation, generalization in “narrow” grid happens only around the vicinity of training tasks, exhibiting the challenge this grid poses for learning-based methods. We believe this is due to the poor extrapolation ability of neural networks (e.g., Haley & Soloway, 1992; Xu et al., 2020), which is more perceptible without prior inductive bias from data augmentation. + +5.2 UNDERSTANDING GAINS FROM PSES: ABLATIONS AND VISUALIZATIONS + +PSEs are contrastive metric embeddings (CMEs) learned with PSM. We investigate the gains from CMEs (Section 4) and PSM (Section 3) by ablating them. CMEs can be learned with any state similarity metric – we use $\pi ^ { * }$ - bisimulation (Castro, 2020) as an alternative. Similarly, PSM can be used with any metric embedding – we use $\ell _ { 2 }$ -embeddings (Section E) as an alternative, which Zhang et al. + +Table 2: Ablating PSEs. Percentage $( \% )$ of test tasks solved when we ablate the similarity metric and learning procedure for metric embeddings in the data augmentation setting on wide grid. PSEs, which combine CMEs with PSM, considerably outperform other embeddings. We report the average performance across 100 runs with standard deviation between parentheses. All ablations except PSEs deteriorate performance compared to just using data augmentation (RandConv), as reported in Table 1. + +
Metric /Embeddingl2-embeddingsCMEs
π*-bisimulation41.4 (17.6)23.1 (7.6)
PSM17.5 (8.4)87.0 (10.1)
+ +(2021) employed with $\pi ^ { * }$ -bisimulation for learning representations in a single task RL setting. For a fair comparison, we tune hyperparameters for 128 trials for each ablation entry in Table 2. + +Table 2 show that PSEs $( = \mathrm { P S M } + \mathrm { C M E s } )$ generalize significantly better than $\pi ^ { * }$ -bisimulation with CMEs or $\ell _ { 2 }$ -embeddings, both of which significantly degrade performance ( $- 6 0 \%$ and $- 4 5 \%$ , respectively). This is expected, as $\pi ^ { * }$ -bisimulation imposes the incorrect invariances for the jumping task $\cdot { c . f . }$ Figures G.2a and G.2d). Additionally, looking at the rows of Table 2, CMEs are superior than $\ell _ { 2 }$ -embeddings for PSM while inferior for $\pi ^ { * }$ -bisimulation. This result is in line with the hypothesis that CMEs better enforce the invariances encoded by a similarity metric compared to $\ell _ { 2 }$ -embeddings $\left( c . f . \right.$ Figures 3b and 3c). + +Visualizing learned representations. We visualize the metric embeddings in the ablation above by projecting them to two dimensions with UMAP (McInnes et al., 2018), a popular visualization technique for high dimensional data which better preserves the data’s global structure compared to other methods such as t-SNE (Coenen & Pearce, 2019). + +Figure 3 shows that PSEs partition the states into two sets: (1) states where a single suboptimal action leads to failure (all states before jump) and (2) states where actions do not affect the final outcome (states after jump). Additionally, PSEs align the labeled states in the first set, which have a PSM distance of zero. These aligned states have the same distance from the obstacle, the invariant feature that generalizes across tasks. On the other hand, $\ell _ { 2 }$ -embeddings (Zhang et al., 2021) with PSM do not align states with zero PSM except the state with the jump action – poor generalization, as observed empirically, is likely due to states with the same optimal behavior ending up with distant embeddings. CMEs with $\pi ^ { * }$ -bisimulation align states with $\pi ^ { * }$ -bisimulation distance of zero – such states are equidistant from the start state and have different optimal behavior for any pair of tasks with different obstacle positions (Figure G.2c). + +5.3 JUMPING TASK WITH COLORS: WHERE TASK-DEPENDENT INVARIANCE MATTERS + +PSEs capture invariances that are usually orthogonal to task-agnostic invariances from data augmentation. This difference is important because, for certain tasks, data augmentation can erroneously alias states with different optimal behavior. Domain knowledge is often required to select appropriate augmentations, otherwise augmentations can hurt generalization. In contrast, PSEs do not require any domain knowledge but instead exploit the inherent structure of the RL tasks. + +To demonstrate the difference between PSEs and data augmentation, we simply include colored obstacles in the jumping task (see Figure G.5). In this modified task, the optimal behavior of the agent depends on the obstacle color: the agent needs to jump over the red obstacle but strike the green obstacle to get a high return. The red obstacle task has the same difficulty as the original jumping task while the green obstacle task is easier. We jointly train the agent with 18 training tasks each, for both obstacle colors, and evaluate generalization on unseen red tasks. + +![](images/dd27daffdbeb11b7f10511fabc09304681a9967704b663bae3974721a6096f03.jpg) +Figure 4: Percentage $( \% )$ of red obstacle test tasks solved when trained, jointly with red and green obstacles, on the “wide” grid. We report the mean across 100 runs. Error bars show $9 9 \%$ confidence interval for the mean. + +Figure 4 shows the large performance gap between PSEs and data augmentation with RandConv. All methods solve the green obstacle tasks (Table G.1). As opposed to the original jumping task ( $\cdot f .$ Table 1), data augmentation inhibits generalization since RandConv forces the agent to ignore color, conflating the red and green tasks (Figure G.6). PSEs still outperform regularization and data augmentation. Furthermore, data augmentation performs better when combined with PSEs. Thus, PSEs are effective even when data augmentation hurts performance. + +# 5.4 EFFECT OF POLICY SUBOPTIMALITY ON PSES + +To understand the sensitivity of learning effective PSEs to the quality of the policies, we compute PSEs using $\epsilon$ -suboptimal policies on the jumping task, which take the optimal action with probability $1 - \epsilon$ and the subopotimal action with probability $\epsilon$ . We evaluate the generalization performance of PSEs for increasingly suboptimal policies, ranging from the optimal policy $( \epsilon = 0$ ) to the uniform random policy $\epsilon = 0 . 5$ ). To isolate the effect of suboptimality on PSEs, the agent still imitates the optimal actions during training for all $\epsilon$ . + +Figure 5 shows that PSEs show near-optimal generalization with $\epsilon \leq 0 . 4$ while degrade generalization with an uniform random policy. This result is well-aligned with Proposition D.3, which shows that for policies with decreasing suboptimality, PSM approximation becomes more accurate, resulting in improved PSEs. Overall, this study confirms that the utility of PSEs for generalization is robust to suboptimality. One reason for this robustness is that PSEs are likely to align states with similar long-term greedy optimal actions, resulting in good performance even with suboptimal policies that preserve these greedy actions. + +# 6 ADDITIONAL EMPIRICAL EVALUATION + +In this section, we exhibit that PSM ignores spurious information + +![](images/cc90cd32460e2a251f32af744d09a75a8fd7b373561f554a97a64367ed0acb68.jpg) +Figure 5: $\%$ of test tasks solved using PSEs computed using $\epsilon \cdot$ suboptimal policies on the “wide” grid. We report the mean across 100 runs. Error bars show one standard deviation. + +for generalization using a LQR task (Song et al., 2019) with non-image inputs. Then, we demonstrate the scalability of PSEs without explicit access to optimal policies in an RL setting, with continuous actions, using Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b). + +# 6.1 LQR WITH SPURIOUS CORRELATIONS + +We show how representations learned using PSM, when faced with semantically equivalent environments, can learn the main factors of variation and ignore spurious correlations that hinder generalization. We use LQR with distractors (Song et al., 2019; Sonar et al., 2020) to assess generalization in a featurebased RL setting with linear function approximation. The distractors are input features that are spuriously correlated with optimal actions and can be used for predicting these actions during training, but hurt generalization. The agent learns a linear policy using 2 environments with fixed distractors. This policy is evaluated on environments with unseen distractors. + +We aggregate state pairs in training environments with near-zero PSM. We contrast this approach with (i) IPO (Sonar et al., 2020), a policy optimization method based on IRM (Arjovsky et al., 2019) for outof-distribution generalization, (ii) overparametrization, which leads to better generalization via implicit regularization (Song et al., 2019), and (iii) weight sparsity using $\ell _ { 1 }$ -regularization since the policy weights which generalize in this task are sparse. + +![](images/6faf7cc10acce48febb5bb654d3e6e540adb0aa2e7913c37cb528653ffb5a9f1.jpg) +Figure 6: LQR generalization: Absolute test error in LQR cost relative to the oracle (which has access to true state), of various methods trained with $n _ { d }$ distractors on 2 training environments. Lower error is better. PSM achieves near-optimal performance. We report mean across 100 different seeds. Error bars (non-existent for most methods) show $9 9 \%$ confidence interval for mean. + +All methods optimally solve the training environments; however, the baselines perform abysmally in terms of generalization compared to state aggregation with PSM (Figure 6), indicating their reliance on distractors. PSM obtains near-optimal generalization which we corroborate through this conjecture (Section H.1): Assuming zero state aggregation error with PSM, the policy learned using gradient descent is independent of the distractors. Refer to Section H for a detailed discussion. + +# 6.2 DISTRACTING DM CONTROL SUITE + +Finally, we demonstrate scalability of PSEs on the Distracting DM Control Suite (DCS) (Stone et al., 2021), which tests whether agents can ignore high-dimensional visual distractors irrelevant to the RL task. Since we do not have access to optimal training policies, we use learned policies as proxy for $\pi ^ { * }$ for computing PSM as well as collecting data for optimizing PSEs. Even with this approximation, PSEs outperform stateof-the-art data augmentation. DCS extends DM Control (Tassa et al., 2020) with visual distractions. We use the dynamic background distractions (Stone et al., 2021; Zhang et al., 2018b) where a video is played in the background from a specific frame. The video and the frame are randomly sampled every new episode. We use 2 videos during training (Figure 7) and evaluate generalization on 30 unseen videos (Figure I.1). + +![](images/ff872ad14af14eeee7e2d8c38cd29313df3392bf8308eb9fbdc234e87dd58a41.jpg) +Figure 7: Distracting Control Suite: Snapshots of training environments. + +Table 3: Generalization performance with unseen distractions in the Distracting Suite at 500K steps. We report the average scores across 5 seeds $\pm$ standard error. All methods are added to SAC (Haarnoja et al., 2018). Pretrained initialization uses $\mathrm { D r Q }$ trained for 500K steps. Figures I.2 and I.3 show generalization curves. + +
InitializationMethodBiC-catchC-swingupC-runF-spinR-easyW-walk
RandomDrQ747±28582±42220±12646±54931±14549±83
DrQ + PSEs821±17749±19308±12779±49955±10789±28
PretrainedDrQ748±30689±22219±10764±48943±10709±29
DrQ +PSEs805±25753±13282±8803±19962±11829±21
+ +All agents are built on top of SAC (Haarnoja et al., 2018) combined with $\mathrm { D r Q }$ (Kostrikov et al., 2020), an augmentation method with state-of-the-art performance on DM control. Notably, DrQ outperforms CURL (Laskin et al., 2020b), a strong contrastive method for RL. Without data augmentation on DM control, SAC performs poorly, even during training (Kostrikov et al., 2020). We augment $\mathrm { D r Q }$ with an auxiliary loss for learning PSEs and compare it with DrQ (Table 3). Orthogonal to DrQ, PSEs align representations of different states across environments based on PSM $( c . f .$ Figure A.1). All agents are trained for 500K environment steps with random crop augmentation. For computing PSM, we use policies learned by $\mathrm { D r Q }$ pretrained on training environments for 500K steps. + +First, we investigate how much better PSEs generalize relative to DrQ, assuming the agent is provided with PSM beforehand. The agent’s policy is randomly initialized so that additional gains over DrQ can be attributed to the auxiliary information from PSM. The substantial gains in Table 3 indicate that PSEs are more effective than $_ \mathrm { D r Q }$ for encoding invariance to distractors. + +Since PSEs utilize PSM approximated using pretrained policies, we also compare to a $_ \mathrm { D r Q }$ agent where we initialize it using these pretrained policies. This comparison provides the same auxiliary information to DrQ as available to PSEs, thus, the generalization difference stems from how they utilize this information. Table 3 demonstrates that PSEs outperform DrQ with pretrained initialization, indicating that the additional pretraining steps are more judiciously utilized for computing PSM as opposed to just longer training with DrQ. More details, including learning curves, are in Section I. + +# 7 RELATED WORK + +PSM (Section 3) is inspired by bisimulation metrics (Section C). However, different than traditional bisimulation (e.g., Larsen & Skou, 1991; Givan et al., 2003; Ferns et al., 2011), PSM is more tractable as it defined with respect to a single policy similar to the recently proposed $\pi ^ { * }$ -bisimulation (Castro, 2020; Zhang et al., 2021). However, in contrast to PSM, bisimulation metrics rely on reward information and may not provide a meaningful notion of behavioral similarity in certain environments (Section 5). For example, states similar under PSM would have similar optimal policies, yet can have arbitrarily large $\pi ^ { * }$ -bisimulation distance between them (Proposition D.1). + +PSEs (Section 4) use contrastive learning to encode behavior similarity (Section 3) across MDPs. Previously, contrastive learning has been applied for imposing state self-consistency (Laskin et al., 2020b), capturing predictive information (Oord et al., 2018; Mazoure et al., 2020; Lee et al., 2020b) or encoding transition dynamics (van der Pol et al., 2020; Stooke et al., 2020; Schwarzer et al., 2020) within an MDP. These methods can be integrated with PSEs to encode additional invariances. Interestingly, in a similar spirit to PSEs, Pacchiano et al. (2020); Moskovitz et al. (2021) explore comparing behavioral similarity between policies to guide policy optimization within an MDP. + +PSEs are complementary to data augmentation methods (Kostrikov et al., 2020; Lee et al., 2020a; Raileanu et al., 2020; Ye et al., 2020), which have recently been shown to significantly improve agents’ generalization capabilities. In fact, we combine PSEs to state-of-the-art augmentation methods including random convolutions (Lee et al., 2020a; Laskin et al., 2020a) in the jumping task and DrQ (Kostrikov et al., 2020) on Distracting Control Suite, leading to performance improvement. + +# 8 CONCLUSION + +This paper advances generalization in RL by two contributions: (1) the policy similarity metric (PSM) which provides a new notion of state similarity based on behavior proximity, and (2) contrastive metric embeddings, which harness the benefits of contrastive learning for representations based on a similarity metric. PSEs combine these two ideas to improve generalization. Overall, this paper shows the benefits of exploiting the inherent structure in RL for learning effective representations. + +# REFERENCES + +Rishabh Agarwal, Chen Liang, Dale Schuurmans, and Mohammad Norouzi. Learning to generalize from sparse and underspecified rewards. In ICML, 2019. + +Martín Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. CoRR, abs/1907.02893, 2019. + +Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In Advances in Neural Information Processing Systems, pp. 7413–7424, 2019. + +Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta-regularization. In Advances in Neural Information Processing Systems, pp. 998–1008, 2018. + +Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic Markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. + +Pablo Samuel Castro and Doina Precup. Using bisimulation for policy transfer in MDPs. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2010. + +Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. CoRR, abs/2002.05709, 2020. + +Karl Cobbe, Oleg Klimov, Christopher Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2019. + +Andy Coenen and Adam Pearce. Understanding umap, 2019. URL https://pair-code.github.io/ understanding-umap/. + +Jesse Farebrother, Marlos C. Machado, and Michael Bowling. Generalization and regularization in DQN. In NeurIPS Deep Reinforcement Learning Workshop, 2018. + +Norm Ferns, Prakash Panangaden, and Doina Precup. Metrics for finite Markov decision processes. In Proceedings of the Conference in Uncertainty in Artificial Intelligence (UAI), 2004. + +Norm Ferns, Pablo Samuel Castro, Doina Precup, and Prakash Panangaden. Methods for computing state similarity in Markov decision processes. In Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence, UAI ’06. AUAI Press, 2006. + +Norm Ferns, Prakash Panangaden, and Doina Precup. Bisimulation metrics for continuous markov decision processes. SIAM Journal on Computing, 40(6):1662–1714, 2011. + +Norman Ferns and Doina Precup. Bisimulation metrics are optimal value functions. In The 30th Conference on Uncertainty in Artificial Intelligence, pp. 10, 2014. + +Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International Conference on Machine Learning (ICML), 2017. + +Robert Givan, Thomas Dean, and Matthew Greig. Equivalence notions and model minimization in markov decision processes. Artificial Intelligence, 147(1-2):163–223, 2003. + +Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and D Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pp. 1487–1495, 2017. + +Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit regularization in matrix factorization. In Advances in Neural Information Processing Systems, pp. 6151–6159, 2017. + +Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. arXiv preprint arXiv:1801.01290, 2018. + +Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2, pp. 1735–1742. IEEE, 2006. + +Pamela J Haley and DONALD Soloway. Extrapolation limitations of multilayer feedforward neural networks. In [Proceedings 1992] IJCNN International Joint Conference on Neural Networks, volume 4, pp. 25–30. IEEE, 1992. + +Maximilian Igl, Kamil Ciosek, Yingzhen Li, Sebastian Tschiatschek, Cheng Zhang, Sam Devlin, and Katja Hofmann. Generalization in reinforcement learning with selective noise injection and information bottleneck. In Advances in Neural Information Processing Systems, pp. 13978–13990, 2019. + +Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Domain extrapolation via regret minimization. arXiv preprint arXiv:2006.03908, 2020. + +Arthur Juliani, Ahmed Khalifa, Vincent-Pierre Berges, Jonathan Harper, Ervin Teng, Hunter Henry, Adam Crespi, Julian Togelius, and Danny Lange. Obstacle Tower: A generalization challenge in vision, control, and planning. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2019. + +Niels Justesen, Ruben Rodriguez Torrado, Philip Bontrager, Ahmed Khalifa, Julian Togelius, and Sebastian Risi. Illuminating generalization in deep reinforcement learning through procedural level generation. arXiv preprint arXiv:1806.10729, 2018. + +Taylor W. Killian, George Dimitri Konidaris, and Finale Doshi-Velez. Robust and efficient transfer learning with hidden parameter Markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pp. 4949–4950, 2017. + +Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. CoRR, abs/2004.13649, 2020. + +Kim G Larsen and Arne Skou. Bisimulation through probabilistic testing. Information and computation, 94(1): 1–28, 1991. + +Michael Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. CoRR, abs/2004.14990, 2020a. + +Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. Proceedings of the 37th International Conference on Machine Learning, Vienna, Austria, PMLR 119, 2020b. arXiv:2003.06417. + +Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for generalization in deep reinforcement learning. In The International Conference on Learning Representations (ICLR), 2020a. + +Kuang-Huei Lee, Ian Fischer, Anthony Liu, Yijie Guo, Honglak Lee, John Canny, and Sergio Guadarrama. Predictive information accelerates learning in rl. arXiv preprint arXiv:2007.12401, 2020b. + +Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. + +Bogdan Mazoure, Remi Tachet des Combes, Thang Long DOAN, Philip Bachman, and R Devon Hjelm. Deep reinforcement and infomax learning. Advances in Neural Information Processing Systems, 33, 2020. + +L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiv e-prints, February 2018. + +Ted Moskovitz, Michael Arbel, Ferenc Huszar, and Arthur Gretton. Efficient wasserstein natural gradients for reinforcement learning. In International Conference on Learning Representations, 2021. + +Junhyuk Oh, Satinder P. Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2017. + +Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. + +Aldo Pacchiano, Jack Parker-Holder, Yunhao Tang, Krzysztof Choromanski, Anna Choromanska, and Michael Jordan. Learning to score behaviors for guided policy optimization. In International Conference on Machine Learning, 2020. + +Charles Packer, Katelyn Gao, Jernej Kos, Philipp Krähenbühl, Vladlen Koltun, and Dawn Song. Assessing generalization in deep reinforcement learning. arXiv preprint arXiv:1810.12282, 2018. + +Christian F. Perez, Felipe Petroski Such, and Theofanis Karaletsos. Generalized hidden parameter mdps transferable model-based rl in a handful of trials. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. + +Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. +Martin L Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley & Sons, Inc., 1994. +Roberta Raileanu, Max Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in deep reinforcement learning. arXiv preprint arXiv:2006.12862, 2020. +Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards generalization and simplicity in continuous control. In Advances in Neural Information Processing Systems (NeurIPS), 2017. +Benjamin Recht. A tour of reinforcement learning: The view from continuous control. Annual Review of Control, Robotics, and Autonomous Systems, 2019. +Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Dataefficient reinforcement learning with momentum predictive representations. arXiv preprint arXiv:2007.05929, 2020. +Anoopkumar Sonar, Vincent Pacelli, and Anirudha Majumdar. Invariant policy optimization: Towards stronger generalization in reinforcement learning. arXiv preprint arXiv:2006.01096, 2020. +Xingyou Song, Yiding Jiang, Yilun Du, and Behnam Neyshabur. Observational overfitting in reinforcement learning. In The International Conference on Learning Representations (ICLR), 2019. +Austin Stone, Oscar Ramirez, Kurt Konolige, and Rico Jonschkowski. The distracting control suite – a challenging benchmark for reinforcement learning from pixels. arXiv preprint arXiv:2101.02722, 2021. +Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from reinforcement learning. arXiv preprint arXiv:2009.08319, 2020. +Remi Tachet des Combes, Philip Bachman, and Harm van Seijen. Learning invariances for policy generalization. In Workshop track at the International Conference on Learning Representations (ICLR), 2018. +Yujin Tang, Duong Nguyen, and David Ha. Neuroevolution of self-interpretable agents. arXiv preprint arXiv:2003.08165, 2020. +Yuval Tassa, Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom Erez, Timothy Lillicrap, and Nicolas Heess. dm_control: Software and tasks for continuous control. arXiv preprint arXiv:2006.12983, 2020. +Matthew E. Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research, 10:1633–1685, 2009. +Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 23–30. IEEE, 2017. +Elise van der Pol, Thomas Kipf, Frans A Oliehoek, and Max Welling. Plannable approximations to mdp homomorphisms: Equivariance under actions. In Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems, pp. 1431–1439, 2020. +Cédric Villani. Optimal transport: old and new. Springer, 2008. +Sam Witty, Jun Ki Lee, Emma Tosch, Akanksha Atrey, Michael Littman, and David Jensen. Measuring and characterizing generalization in deep reinforcement learning. In NeurIPS Critiquing and Correcting Trends in Machine Learning Workshop, 2018. +Keyulu Xu, Jingling Li, Mozhi Zhang, Simon S Du, Ken-ichi Kawarabayashi, and Stefanie Jegelka. How neural networks extrapolate: From feedforward to graph neural networks. arXiv preprint arXiv:2009.11848, 2020. +Chang Ye, Ahmed Khalifa, Philip Bontrager, and Julian Togelius. Rotation, translation, and cropping for zero-shot generalization. CoRR, abs/2001.09908, 2020. +Amy Zhang, Nicolas Ballas, and Joelle Pineau. A Dissection of Overfitting and Generalization in Continuous Reinforcement Learning. CoRR, abs/1806.07937, 2018a. +Amy Zhang, Yuxin Wu, and Joelle Pineau. Natural environment benchmarks for reinforcement learning. arXiv preprint arXiv:1811.06032, 2018b. + +Amy Zhang, Clare Lyle, Shagun Sodhani, Angelos Filos, Marta Kwiatkowska, Joelle Pineau, Yarin Gal, and Doina Precup. Invariant causal prediction for block mdps. In Proceedings of the International Conference on Machine Learning (ICML), 2020. + +Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. The International Conference on Learning Representations (ICLR), 2021. + +Chiyuan Zhang, Oriol Vinyals, Rémi Munos, and Samy Bengio. A study on overfitting in deep reinforcement learning. CoRR, abs/1804.06893, 2018c. + +# Appendix + +# A LEARNING CONTRASTIVE METRIC EMBEDDINGS + +![](images/d19f18e3e5447a5a82c4130e02a0d8375292bd0ba4e9e7a07050c01ca385d18a.jpg) +Figure A.1: Architecture for learning CMEs. Given an input pair $( x , y )$ , we first apply the (optional) data augmentation operator $\Psi$ to produce the input augmentations $\bar { \Psi } _ { x } : = \Psi ( x ) , \Psi _ { y } : = \bar { \Psi } ( y )$ . When not using data augmentation, $\Psi$ is equal to the identity operator, that is, ∀ $x \Psi ( x ) = x$ . The agent’s policy network then outputs the representations for these augmentations by applying the encoder $f _ { \theta }$ , that is, ${ \bar { f } } _ { x } = f _ { \theta } ( \Psi _ { x } )$ , $f _ { y } =$ $f _ { \theta } ( \hat { \Psi } _ { y } )$ . These representations are projected using a non-linear projector $h _ { \theta }$ to obtain the embedding $z _ { \theta }$ , that is, $z _ { \theta } ( x ) = h _ { \theta } ( f _ { x } )$ , $z _ { \theta } ( y ) = h _ { \theta } ( f _ { y } )$ . These metric embeddings are trained using the contrastive loss defined in Equation (4). The policy $\pi \theta$ is an affine function of the representation, that is, $\pi _ { \boldsymbol { \theta } } ( \cdot | \boldsymbol { y } ) = \boldsymbol { W } ^ { T } f _ { \boldsymbol { y } } + \boldsymbol { b }$ , where $W , b$ are learned weights and biases. The entire network is trained end-to-end jointly using the reinforcement learning (or imitation learning) loss in conjunction with the auxiliary contrastive loss. + +# B PROOFS + +We begin by defining some notation which will be used throughout these results: + +• We denote $\begin{array} { r } { { ! } \mathbb { E } _ { t \ge 0 } [ \gamma ^ { t } T V ( \tilde { \pi } ( Y _ { y } ^ { t } ) , \pi ^ { * } ( Y _ { y } ^ { t } ) ) ] = \mathbb { E } _ { Y _ { y } ^ { t } } \left[ \sum _ { t \ge 0 } \gamma ^ { t } T V ( \tilde { \pi } ( Y _ { y } ^ { t } ) , \pi ^ { * } ( Y _ { y } ^ { t } ) ) \right] } \end{array}$ • For any $y \in Y$ , let $Y _ { y } ^ { t } \sim P ^ { \tilde { \pi } } ( \cdot | Y _ { y } ^ { t - 1 } )$ , where $Y _ { y } ^ { 0 } = y$ . +• $T V ^ { n } ( Y _ { y } ^ { k } ) = \mathbb { E } _ { 0 \leq t < n } \gamma ^ { t } T V ( \tilde { \pi } ( Y _ { y } ^ { k + t } ) , \pi ^ { * } ( Y _ { y } ^ { k + t } ) ) .$ + +We now proceed with some technical lemmas necessary for the main result. + +Lemma 1. Given any two pseudometrics5 $d , d ^ { \prime } \in \mathbb { M }$ and probability distributions $P _ { X } , P _ { y }$ where $\mathcal { X } , \mathcal { Y } \subset \mathcal { S }$ , we have: + +$$ +\mathcal { W } ^ { 1 } ( d ) ( P _ { \mathcal { X } } , P _ { \mathcal { Y } } ) \leq \| d - d ^ { \prime } \| + \mathcal { W } ^ { 1 } ( d ^ { \prime } ) ( P _ { \mathcal { X } } , P _ { \mathcal { Y } } ) +$$ + +Proof. Note that the dual of the linear program for computing $\mathcal { W } ^ { 1 } ( d ) ( P _ { \mathcal { X } } , P _ { \mathcal { Y } } )$ is given by + +$$ +\begin{array} { l } { \displaystyle \mathop { \operatorname* { m i n } } _ { \textsc { i r } } \sum _ { x \in \mathcal { X } , \ y \in \mathcal { V } } \Gamma ( x , y ) d ( x , y ) } \\ { \mathrm { s u b j e c t ~ t o ~ } \displaystyle \sum _ { x } \Gamma ( x , y ) = P y ( y ) \forall y , \quad \displaystyle \sum _ { y } \Gamma ( x , y ) = P _ { \mathcal { X } } ( x ) \forall x , \quad \Gamma ( x , y ) \geq 0 \forall x , y } \end{array} +$$ + +5Pseudometrics are generalization of metrics where the distance between two distinct states can be zero. + +Using the dual formulation subject to the constraints above, $\mathscr { W } ^ { 1 } ( d )$ can be written as + +$$ +\begin{array} { l } { \displaystyle \mathcal { W } ^ { 1 } ( d ) ( P _ { \mathcal { X } } , P _ { \mathcal { Y } } ) \leq \| d - d ^ { \prime } \| = \mathcal { W } ^ { 1 } ( d - d ^ { \prime } + d ^ { \prime } ) ( P _ { \mathcal { X } } , P _ { \mathcal { Y } } ) \leq \| d - d ^ { \prime } \| } \\ { \displaystyle \quad \quad \quad = \operatorname* { m i n } _ { \displaystyle \Gamma } \sum _ { x \in \mathcal { X } , y \in \mathcal { Y } } \Gamma ( x , y ) \left( d ( x , y ) - d ^ { \prime } ( x , y ) + d ^ { \prime } ( x , y ) \right) } \\ { \displaystyle \quad \quad \leq \operatorname* { m i n } _ { \displaystyle \Gamma } \sum _ { x \in \mathcal { X } , y \in \mathcal { Y } } \Gamma ( x , y ) \left( \| d - d ^ { \prime } \| + d ^ { \prime } ( x , y ) \right) } \\ { \displaystyle \quad \quad = \| d - d ^ { \prime } \| + \operatorname* { m i n } _ { \displaystyle \Gamma } \sum _ { x \in \mathcal { X } , y \in \mathcal { Y } } \Gamma ( x , y ) d ^ { \prime } ( x , y ) } \\ { \displaystyle \quad \quad = \| d - d ^ { \prime } \| + \mathcal { W } ^ { 1 } ( d ^ { \prime } ) ( P _ { \mathcal { X } } , P _ { \mathcal { Y } } ) } \end{array} +$$ + +Lemma 2. Given any $y _ { 0 } \in Y$ , we have: + +$$ +\sum _ { y _ { 1 } \in Y } \left( P ^ { \tilde { \pi } } ( y _ { 1 } | y _ { 0 } ) - P ^ { \pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \right) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \leq \frac { 2 } { 1 - \gamma } T V ( \tilde { \pi } ( y _ { 0 } ) , \pi ^ { * } ( y _ { 0 } ) ) +$$ + +Proof. + +$$ +\begin{array} { r l } { \displaystyle \sum _ { i \in \mathcal { K } } ( P ^ { \kappa } ( g _ { 1 | | | B \rangle } ) - P ^ { \kappa } ( g _ { 1 | | B \rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) \leq | \displaystyle \sum _ { i \in \mathcal { K } } ( P ^ { \kappa } ( g _ { 1 | | B \rangle } ) - P ^ { \kappa ^ { \star } } ( g _ { 1 | | B \rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) | } & { } \\ & { \leq \displaystyle \sum _ { i \in \mathcal { K } } | \displaystyle \sum _ { i \in \mathcal { K } } P ( g _ { 1 } | \mathfrak { g } _ { 1 } , \alpha | \hat { \sigma } ( a ) | \mathfrak { g } _ { 1 } ) - \pi ^ { \star } ( a | \mathfrak { g } _ { 1 } ) ) | T V ^ { n } ( Y _ { n } ^ { 0 } ) } \\ & { \leq \frac { 1 } { 1 - \gamma } \displaystyle \sum _ { i \in \mathcal { K } } \displaystyle \sum _ { i \in \mathcal { K } } \displaystyle \sum _ { i \in \mathcal { K } } \sum _ { i \in \mathcal { K } } P ^ { \kappa } ( g _ { 1 } | \mathfrak { g } _ { 1 } , \alpha ) | \overline { { \tau } } ( \mathfrak { g } ( | \mathfrak { g } _ { 1 } \rangle ) - \pi ^ { \star } ( \mathfrak { g } _ { 1 } ) | ) } \\ & { = \displaystyle \frac { 1 } { 1 - \gamma } \displaystyle \sum _ { i \in \mathcal { K } } | \overline { { \tau } } ( \mathfrak { g } _ { 1 } | \mathfrak { g } _ { 1 } ) - \pi ^ { \star } ( \mathfrak { g } _ { 1 } | \mathfrak { g } _ { 1 } ) | \displaystyle \sum _ { i \in \mathcal { K } } P ( g _ { 1 } | \mathfrak { g } _ { 1 } , \alpha ) } \\ & { = \displaystyle \frac { 1 } { 1 - \gamma } \displaystyle \sum _ { i \in \mathcal { K } } | \overline { { \tau } } ( \mathfrak { g } _ { 1 } | \mathfrak { g } _ { 1 } ) - \pi ^ { \star } ( \mathfrak { g } _ { 1 } | \mathfrak { g } _ { 1 } ) | } \\ & { = \displaystyle \frac { 1 } { 1 - \gamma } \displaystyle \sum _ { i \in \mathcal { K } } | \overline { { \tau } } ( \mathfrak { g } _ { 1 } | \mathfrak { g } _ { 1 } ) - \pi ^ { \star } ( \mathfrak { g } _ { 1 } | \mathfrak { g } _ { 1 } ) | } \\ & = \ \end{array} +$$ + +Lemma 3. Given any $y _ { 0 } \in Y$ , if $\begin{array} { r } { \mathbf { \dot { \mathbf { \theta } } T } V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \le \frac { 1 + \gamma } { 1 - \gamma } d ^ { * } ( \tilde { x } _ { y _ { 1 } } , y _ { 1 } ) } \end{array}$ , we have: + +$$ +\sum _ { y _ { 1 } \in Y } { P ^ { \pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \leq { \frac { 1 + \gamma } { 1 - \gamma } } W _ { 1 } ( d ^ { * } ) \left( P ^ { \pi ^ { * } } ( \cdot | { \tilde { x } _ { y _ { 0 } } } ) , P ^ { \pi ^ { * } } ( \cdot | y _ { 0 } ) \right) +$$ + +Proof. Note that we have the following equality, where 0 is a vector of zeros: + +$$ +\sum _ { y _ { 1 } \in Y } P ^ { \pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) = \sum _ { y _ { 1 } \in Y } P ^ { \pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) - \sum _ { x \in X } P ^ { \pi ^ { * } } ( x | \tilde { x } _ { y _ { 0 } } ) \mathbf { 0 } +$$ + +which is the same form as the primal LP for $W _ { 1 } ( d ^ { * } ) ( P ^ { \pi ^ { * } } ( \cdot | y _ { 0 } ) , P ^ { \pi ^ { * } } ( \cdot | \tilde { x } _ { y _ { 0 } } ) )$ . By assumption, we have that + +$$ +T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \leq \frac { 1 + \gamma } { 1 - \gamma } d ^ { \ast } ( \tilde { x } _ { y _ { 1 } } , y _ { 1 } ) +$$ + +This implies that $\textstyle { \frac { 1 - \gamma } { 1 + \gamma } } T V ^ { n } ( Y _ { . } ^ { 0 } )$ is a feasible solution to $W _ { 1 } ( d ^ { * } ) ( P ^ { \pi ^ { * } } ( \cdot | y _ { 0 } ) , P ^ { \pi ^ { * } } ( \cdot | \tilde { x } _ { y _ { 0 } } ) )$ + +$$ +\sum _ { y _ { 1 } \in Y } { P ^ { \pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \frac { 1 - \gamma } { 1 + \gamma } T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \leq W _ { 1 } ( d ^ { * } ) \left( P ^ { \pi ^ { * } } ( \cdot | \tilde { x } _ { y _ { 0 } } ) , P ^ { \pi ^ { * } } ( \cdot | y _ { 0 } ) \right) +$$ + +and the result follows. + +Proposition B.1. The operator $\mathcal { F }$ given by: + +$$ +\mathcal F ( d ) ( x , y ) = \mathrm { D I S T } \big ( \pi ^ { * } ( x ) , \pi ^ { * } ( y ) \big ) + \gamma \mathcal W _ { 1 } ( d ) \big ( P _ { \mathcal X } ^ { \pi ^ { * } } ( \cdot | x ) , P _ { \mathcal V } ^ { \pi ^ { * } } ( \cdot | y ) \big ) +$$ + +is a contraction mapping and has a unique fixed point for a bounded dist. + +Proof. We first prove that $\mathcal { F }$ is contraction mapping. Then, a simple application of the Banach Fixed Point Theorem asserts that $\mathcal { F }$ has a unique fixed point. Note that for all pseudometrics $d , d ^ { \prime } \in \mathbb { M }$ , and for all states $x \in \mathcal { X }$ , $y \in \mathcal { V }$ , + +$$ +\begin{array} { r l } & { \mathcal { F } ( d ) ( x , y ) - \mathcal { F } ( d ^ { \prime } ) ( x , y ) } \\ & { \quad = \gamma \left( \mathcal { W } _ { 1 } ( d ) ( P ^ { \pi _ { \mathcal { X } } ^ { * } } ( \cdot | x ) , P _ { y } ^ { \pi ^ { * } } ( \cdot | y ) ) - \mathcal { W } _ { 1 } ( d ^ { \prime } ) ( P ^ { \pi _ { \mathcal { X } } ^ { * } } ( \cdot | x ) , P _ { y } ^ { \pi ^ { * } } ( \cdot | y ) ) \right) } \\ & { \quad \overset { L e m m a ~ \ : 1 } { \leq } \gamma \left( \| d - d ^ { \prime } \| + \mathcal { W } _ { 1 } ( d ^ { \prime } ) ( P ^ { \pi _ { \mathcal { X } } ^ { * } } ( \cdot | x ) , P _ { y } ^ { \pi ^ { * } } ( \cdot | y ) ) - \mathcal { W } _ { 1 } ( d ^ { \prime } ) ( P ^ { \pi _ { \mathcal { X } } ^ { * } } ( \cdot | x ) , P _ { y } ^ { \pi ^ { * } } ( \cdot | y ) ) \right) } \\ & { \quad = \gamma \| d - d ^ { \prime } \| } \end{array} +$$ + +Thus, $\| \mathcal F ( d ) - \mathcal F ( d ^ { \prime } ) \| \le \gamma \| d - d ^ { \prime } \|$ , so that $\mathcal { F }$ is a contracting mapping for $\gamma < 1$ and has an unique fixed point $d ^ { * }$ . + +Theorem 1. [Bound on policy transfer] For any $y \in \mathcal { V }$ , let $Y _ { y } ^ { t } \sim P ^ { \tilde { \pi } } ( \cdot | Y _ { y } ^ { t - 1 } )$ define the sequence of random states encountered starting in $Y _ { y } ^ { 0 } = y$ and following policy $\tilde { \pi }$ . We have: + +$$ +\mathbb { E } _ { Y _ { y } ^ { t } } \left[ \sum _ { t \ge 0 } \gamma ^ { t } T V \left( \tilde { \pi } ( Y _ { y } ^ { t } ) , \pi ^ { * } ( Y _ { y } ^ { t } ) \right) \right] \le \frac { 1 + \gamma } { 1 - \gamma } d ^ { * } ( \tilde { x } _ { y } , y ) . +$$ + +Proof. We will prove this by induction. Assuming that the bound holds for $T V ^ { n }$ , we prove the bound holds for $T V ^ { n + \bar { 1 } }$ . The base case for $n = 1$ follows from $T V ( \tilde { \pi } ( y ) , \pi ^ { * } ( y ) ) = T V ( \pi ^ { * } ( \tilde { x } _ { y } ) , \pi ^ { * } ( y ) ) \leq$ d∗(˜xy, y). Note that T V n ≤ 1−γn+11−γ since the $T V$ distance per time-step can be at most 1. + +Let $P _ { t } ^ { \pi } ( y ^ { \prime } | y )$ denote the probability of ending in state $y ^ { \prime } \in Y$ after $t$ steps when following policy $\pi$ and starting from state $y$ . We then have: + +$$ +\begin{array} { r l } & { \gamma ^ { \varepsilon } \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) - \sum _ { s \in \mathcal { N } } ^ { \varepsilon } \mathcal { K } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) - \gamma \varepsilon \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) } \\ & { - \sum _ { s \in \mathcal { N } } ^ { \varepsilon } \Bigg [ \varepsilon \gamma \varepsilon ( \mathbf { x } ^ { \varepsilon } ) \Delta \mathbf { x } ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) + \varepsilon \gamma ( \mathbf { x } ^ { \varepsilon } ) \Delta \mathbf { x } ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) - \sum _ { s \in \mathcal { N } } ^ { \varepsilon } \varepsilon ( \Phi _ { 0 } ^ { \varepsilon } ) \Delta \mathbf { x } ^ { \varepsilon } \Bigg ( \mathbf { x } ^ { \varepsilon } ) \Bigg ] } \\ & { - \sum _ { s \in \mathcal { N } } ^ { \varepsilon } \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) \Delta \mathbf { x } ^ { \varepsilon } \Bigg ( \mathbf { x } ^ { \varepsilon } ) } \\ & { + \sum _ { s \in \mathcal { N } } ^ { \varepsilon } \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) + \sum _ { s \in \mathcal { N } } ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) ( \mathbf { x } ^ { \varepsilon } ) ( \mathbf { x } ^ { \varepsilon } ) ( \mathbf { x } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) } \\ & { \quad + \sum _ { s \in \mathcal { N } } ^ { \varepsilon } \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) + \gamma \gamma ^ { \varepsilon } ( \Phi _ { 0 } ^ { \varepsilon } ) \gamma ^ { \varepsilon } ( \mathbf { x } ^ { \varepsilon } ) } \\ & \quad - \sum _ { s \in \mathcal { N } } ^ \end{array} +$$ + +Thus, by induction, it follows that for all $n$ : + +$$ +T V ^ { n } ( Y _ { y } ^ { 0 } ) \leq \frac { 1 + \gamma } { 1 - \gamma } d ^ { * } ( \tilde { x } _ { y } , y ) , +$$ + +which completes the proof. + +# C BISIMULATION METRICS + +Notation. We use the notation as defined in Section 2. + +Bisimulation metrics (Givan et al., 2003; Ferns et al., 2011) define a pseudometric $d _ { \sim } : \mathcal { S } \times \mathcal { S } \to \mathbb { R }$ where $d _ { \sim } ( x , y )$ is defined in terms of distance between immediate rewards and next state distributions. Define $\mathcal { F } _ { \sim } ^ { e } : \mathbb { M } \mathbb { M }$ by + +$$ +\mathcal { F } _ { \sim } ^ { e } ( d ) ( x , y ) = \operatorname* { m a x } _ { a \in A } \vert R ( x , a ) - R ( y , a ) \vert + \gamma \mathcal { W } _ { 1 } ( d ) \left( P ^ { a } ( \cdot \vert x ) , P ^ { a } ( \cdot \vert y ) \right) +$$ + +then, $\mathcal { F } _ { \sim } ^ { e }$ has a unique fixed point $d _ { \sim }$ which is a bisimulation metric. $\mathcal { F } _ { \sim } ^ { e }$ uses the 1-Wasserstein metric $\mathcal { W } _ { 1 } : \mathbb { M } \mathbb { M } _ { p }$ . The 1-Wasserstein distance $\mathscr { W } _ { 1 } ( d )$ under the pseudometric $d$ can be computed using the dual linear program: + +$$ +\operatorname* { m a x } _ { \mathbf { u } , \mathbf { v } } \sum _ { x \in \mathcal { X } } P ( x ) u _ { x } - \sum _ { y \in \mathcal { Y } } P ( y ) v _ { y } \quad \mathrm { s u b j e c t ~ t o } \forall x \in \mathcal { X } , y \in \mathcal { Y } \quad u _ { x } - v _ { y } \leq d ( x , y ) +$$ + +Since we are only interested in computing the coupling between the states in $\mathcal { X }$ and $\mathcal { V }$ , the above formulation assumes that $P _ { \mathcal { X } } ( y ) = 0$ for all $y \in \mathcal { V }$ and $P _ { \mathcal { Y } } ( x ) = 0$ for all $x \in \mathcal { X }$ . The computation of $d _ { \sim }$ is expensive and requires a tabular representation of the states, rendering it impractical for large state spaces. On-policy bisimulation (Castro, 2020) (e.g., $\pi ^ { * }$ -bisimulation) is tied to specific behavior policies and is much easier to approximate than bisimulation. + +# D POLICY SIMILARITY METRIC + +# D.1 COMPUTING PSM + +In general, PSM for a given DIST across MDPs $\mathcal { M } _ { \mathcal { X } }$ and $\mathcal { M } _ { \mathcal { Y } }$ is given by + +$$ +\begin{array} { r } { d ^ { * } ( x , y ) = \mathrm { D I S T } \big ( \pi _ { \mathcal { X } } ^ { * } ( x ) , \pi _ { \mathcal { Y } } ^ { * } ( y ) \big ) + \gamma \mathcal { W } _ { 1 } \big ( d ^ { * } ) \big ( P _ { \mathcal { X } } ^ { \pi ^ { * } } ( \cdot | x ) , P _ { \mathcal { Y } } ^ { \pi ^ { * } } ( \cdot | y ) \big ) . } \end{array} +$$ + +Since our main focus is showing the utility of PSM for generalization, we simply use environments where PSM can be computed using dynamic programming. Using a similar observation to Castro (2020), we assert that the recursion for $d ^ { * }$ takes the following form in deterministic environments: + +$$ +\begin{array} { r } { d ^ { * } ( x , y ) = \mathrm { D I S T } \big ( \pi _ { \mathcal { X } } ^ { * } ( x ) , \pi _ { \mathcal { Y } } ^ { * } ( y ) \big ) + \gamma d ^ { * } ( x ^ { \prime } , y ^ { \prime } ) . } \end{array} +$$ + +where $x ^ { \prime } = P _ { \mathcal { X } } ^ { \pi ^ { * } } ( x )$ , $y ^ { \prime } = P _ { y } ^ { \pi ^ { * } } ( y )$ are the next states from taking actions $\pi _ { \mathcal { X } } ^ { \ast } ( x ) , \pi _ { \mathcal { X } } ^ { \ast } ( y )$ from states $x , y$ respectively. Furthermore, we assume that DIST between terminal states from $\mathcal { M } _ { \mathcal { X } }$ and $\mathcal { M } _ { \mathcal { Y } }$ is zero. Note that the form of Equation D.2 closely resembles the update rule in Q-learning, and as such, can be efficiently computed with samples using approximate dynamic programming. Given access to optimal trajectories $\tau _ { \mathcal { X } } ^ { * } = \{ x _ { t } \} _ { t = 1 } ^ { N }$ and $\tau _ { y } ^ { * } = \{ y _ { t } \} _ { t = 1 } ^ { N }$ , where $x _ { t + 1 } = P _ { \mathcal { X } } ^ { \pi ^ { * } } ( x _ { t } )$ and $y _ { t + 1 } = P _ { y } ^ { \pi ^ { * } } ( y _ { t } )$ , Equation D.2 can be solved using exact dynamic programming; we provide pseudocode in Section J.1. + +There are other ways to approximate the Wasserstein distance in bisimulation metrics (e.g., Ferns et al., 2006; 2011; Castro, 2020; Zhang et al., 2021). That said, approximating bisimulation (or PSM) for stochastic environments remains an exciting research direction (Castro, 2020). Investigating other distance metrics for long-term behavior difference in PSM is also interesting for future work. + +# D.2 PSM CONNECTIONS TO DATA AUGMENTATION AND BISIMULATION + +Connection to bisimulation. Although bisimulation metrics have appealing properties such as bounding value function differences (e.g., (Ferns & Precup, 2014)), they rely on reward information and may not provide a meaningful notion of behavioral similarity in certain environments. Proposition D.1 implies that states similar under PSM would have similar optimal policies yet can have arbitrarily large bisimulation distance between them. + +![](images/3d8fd583f31c9cbc260975c387fd75a3ea68a6f1e11f6b2895873d41fe0189c2.jpg) +Figure D.1: Cyan edges represent actions with a positive reward, which are also the optimal actions. Zero rewards everywhere else. $x _ { 0 } , y _ { 0 }$ are the start states while $x _ { 2 } , y _ { 2 }$ are the terminal states. + +Proposition D.1. There exists environments $M _ { \mathcal { X } }$ and $M _ { \mathcal { Y } }$ such that $\forall ( x , y ) \in ~ { \mathcal { L } }$ where ${ \mathcal { L } } =$ $\{ ( x , y ) ~ | x \in \mathcal { X }$ , $y \in \mathcal { V }$ , $d ^ { * } ( x , y ) = 0 \}$ , $\begin{array} { r } { d _ { \sim } ^ { \ast } ( x , y ) = \frac { \left| R _ { \mathrm { m a x } } - R _ { \mathrm { m i n } } \right| } { 1 - \gamma } - \epsilon } \end{array}$ for any given $\epsilon > 0$ . + +For example, consider the two semantically equivalent environments in Figure D.1 with $\pi _ { \mathcal { X } } ^ { * } ( x _ { 0 } ) =$ $\pi _ { y } ^ { * } ( y _ { 0 } ) = a _ { 0 }$ and $\pi _ { \mathcal { X } } ^ { \ast } ( x _ { 1 } ) = \pi _ { \mathcal { Y } } ^ { \ast } ( y _ { 1 } ) = a _ { 1 }$ but different rewards $r _ { x } , r _ { y }$ respectively. Whenever $r _ { y } > ( 1 + 1 / \gamma ) r _ { x }$ , bisimulation metrics incorrectly imply that $x _ { 0 }$ is more behaviorally similar to $y _ { 1 }$ than $y _ { 0 }$ . + +For the MDPs shown in Figure D.1, to determine which $y$ state is behaviorally equivalent to $x _ { 0 }$ , we look at the distances computed by bisimulation metric $d _ { \sim }$ and $\pi ^ { * }$ -bisimulation metric $d _ { \sim } ^ { \ast }$ : + +$$ +\begin{array} { r l } & { d _ { \sim } ( x _ { 0 } , y _ { 0 } ) = d _ { \sim } ^ { \ast } ( x _ { 0 } , y _ { 0 } ) = ( 1 + \gamma ) | r _ { y } - r _ { x } | } \\ & { d _ { \sim } ( x _ { 0 } , y _ { 1 } ) = \operatorname* { m a x } \left( ( 1 + \gamma ) r _ { x } , r _ { y } \right) , \ d _ { \sim } ^ { \ast } ( x _ { 0 } , y } \end{array} +$$ + +Thus, $r _ { y } > \left( 1 + 1 / \gamma \right) r _ { x }$ implies that $d _ { \sim } ( x _ { 0 } , y _ { 1 } ) < d _ { \sim } ( x _ { 0 } , y _ { 0 } )$ as well as $d _ { \sim } ^ { \ast } ( x _ { 0 } , y _ { 1 } ) < d _ { \sim } ^ { \ast } ( x _ { 0 } , y _ { 0 } )$ + +Connection to data augmentation. Data augmentation often assumes access to optimality invariant transformations, e.g., random crops or flips in image-based benchmarks (Laskin et al., 2020a; Kostrikov et al., 2020). However, for certain RL tasks, such augmentations can erroneously alias states with different optimal behavior and hurt generalization. For example, if the image observation is flipped in a goal reaching task with left and right actions, the optimal actions would also be flipped to take left actions instead of right and vice versa. Proposition D.2 states that PSMs can precisely quantify the invariance of such augmentations. + +Proposition D.2. For an MDP $M _ { \mathcal { X } }$ and its transformed version $M _ { \psi ( \chi ) }$ for the data augmentation $\psi$ , $d ^ { * } ( x , \psi ( x ) )$ indicates the optimality invariance of $\psi$ for any $x \in \mathcal { X }$ . + +# D.3 PSM WITH APPROXIMATELY-OPTIMAL POLICIES + +Generalized Policy Similarity Metric for arbitrary policies. For a given DIST, we define a generalized PSM $\dot { d } : ( \boldsymbol { S } \times \boldsymbol { \Pi } ) \times ( \boldsymbol { S } \times \boldsymbol { \Pi } ) \mathbb { R }$ where $\Pi$ is the set of all policies over $s$ . $d$ satisfies the recursive equation: + +$$ +d \bigl ( ( x , \pi _ { 1 } ) , ( y , \pi _ { 2 } ) \bigr ) = \mathrm { D I S T } \bigl ( \pi _ { 1 } ( x ) , \pi _ { 2 } ( y ) \bigr ) + \gamma { \mathcal W } _ { 1 } ( d ) \bigl ( P ^ { \pi _ { 1 } } ( \cdot | x ) , P ^ { \pi _ { 2 } } ( \cdot | y ) \bigr ) . +$$ + +Since DIST is assumed to be a pseudometric and $\mathcal { W } _ { 1 }$ is a probability metric, it implies that $d$ is a pseudometric as (1) $d$ is non-negative, that is, $d \big ( ( x , \pi _ { 1 } \bar { ) } , ( y , \pi _ { 2 } ) \big ) ^ { \cdot } \geq 0$ , (2) $d$ is symmetric, that is, $d \big ( ( x , \pi _ { 1 } ) , ( y , \pi _ { 2 } ) \big ) \ = \ d \big ( ( x , \pi _ { 1 } ) , ( y , \pi _ { 2 } ) \big )$ , and $d$ satisfies the triangle inequality, that is, $d \bigl ( ( x , \pi _ { 1 } ) , ( y , \pi _ { 2 } ) \bigr ) < d \bigl ( ( x , \pi _ { 1 } ) , ( z , \pi _ { 3 } ) \bigr ) + d \bigl ( ( z , \pi _ { 3 } ) , ( y , \pi _ { 2 } ) \bigr )$ . + +Using this notion of generalized PSM, we show that the approximation error in PSM from using a suboptimal policy is bounded by the policy’s suboptimality. Thus, for policies with decreasing suboptimality, the PSM approximation becomes more accurate, resulting in improved PSEs. + +Proposition D.3. [Approximation error in PSM] Let $\hat { d } : \mathcal { S } \times \mathcal { S } \mathbb { R }$ be the approximate PSM computed using a suboptimal policy $\hat { \pi }$ defined over $s$ , that is, $\hat { d } ( x , y ) ~ = ~ \mathrm { D I S T } \big ( \hat { \pi } ( x ) , \hat { \pi } ( y ) \big ) ~ + ~$ $\gamma \mathcal { W } _ { 1 } ( \hat { d } ) \big ( P ^ { \hat { \pi } } ( \cdot \vert x ) , P ^ { \hat { \pi } } ( \cdot \vert y ) \big )$ . We have: + +$$ +| d ^ { * } ( x , y ) - { \hat { d } } ( x , y ) | < { { \cal A } } \bigl ( ( x , \pi ^ { * } ) , ( x , { \hat { \pi } } ) \bigr ) _ { \bot } + { { \cal A } } \bigl ( ( y , { \hat { \pi } } ) , ( y , \pi ^ { * } ) \bigr ) _ { \bot } . +$$ + +Proof. The PSM $d ^ { * }$ and approximate PSM $\hat { d }$ are instantiations of the generalized PSM (Equation D.3) with both input policies as $\pi ^ { * }$ and $\hat { \pi }$ respectively. + +$$ +\begin{array} { r l } & { d ^ { * } ( x , y ) = d ( ( x , \pi ^ { * } ) , ( y , \pi ^ { * } ) ) < d \big ( ( x , \pi ^ { * } ) , ( x , \hat { \pi } ) \big ) + d \big ( ( x , \hat { \pi } ) , ( y , \pi ^ { * } ) \big ) } \\ & { \qquad < d \big ( ( x , \pi ^ { * } ) , ( x , \hat { \pi } ) \big ) + d \big ( ( y , \hat { \pi } ) , ( y , \pi ^ { * } ) \big ) + d \big ( ( x , \hat { \pi } ) , ( y , \hat { \pi } ) \big ) } \\ & { d ^ { * } ( x , y ) - \hat { d } ( x , y ) < d \big ( ( x , \pi ^ { * } ) , ( x , \hat { \pi } ) \big ) + d \big ( ( y , \hat { \pi } ) , ( y , \pi ^ { * } ) \big ) \quad : \ \hat { d } ( x , y ) = d \big ( ( x , \hat { \pi } ) , ( y , \hat { \pi } ) \big ) } \\ & { { \mathrm { S i m i l a r l y } } , \ \hat { d } ( x , y ) - d ^ { * } ( x , y ) < d \big ( ( x , \pi ^ { * } ) , ( x , \hat { \pi } ) \big ) + d \big ( ( y , \hat { \pi } ) , ( y , \pi ^ { * } ) \big ) } \end{array} +$$ + +# E L2 METRIC EMBEDDINGS + +Another common choice (Zhang et al., 2021) for learning metric embeddings is to use the squared loss (i.e., l2-loss) for matching the euclidean distance between the representations of a pair of states to the metric distance between those states. Concretely, for a given $d ^ { * }$ and representation $f _ { \theta }$ , the loss $\mathcal { L } ( \theta ) = \mathbb { E } _ { s _ { i } , s _ { j } } [ ( \| f _ { \theta } ( s _ { i } ) - f _ { \theta } ( s _ { j } ) \| _ { 2 } - d ^ { * } ( s _ { i } , s _ { j } ) ) ^ { \bar { 2 } } ]$ is minimized. However, it might be too restrictive to match the exact metric distances, which we demonstrate empirically by comparing $l 2$ metric embeddings with CMEs (Section 5.2). + +# F EXTENDED RELATED WORK + +Generalization across different tasks used to be described as transfer learning. In the past, most transfer learning approaches relied on fixed representations and tackled different problem formulations (e.g., assuming shared state space). Taylor & Stone (2009) present a comprehensive survey of the techniques at the time, before representation learning became so prevalent in RL. Recently, the problem of performing well in a different, but related task, started to be seen as a problem of generalization; with the community highlighting that deep RL agents tend to overfit to the environments they are trained on (Cobbe et al., 2019; Witty et al., 2018; Farebrother et al., 2018; Juliani et al., 2019; Kostrikov et al., 2020; Song et al., 2019; Justesen et al., 2018; Packer et al., 2018). + +Prior generalization approaches are typically adapted from supervised learning, including data augmentation, regularization (Cobbe et al., 2019; Farebrother et al., 2018), stochasticity (Zhang et al., 2018c), noise injection (Igl et al., 2019; Zhang et al., 2018a), more diverse training conditions (Rajeswaran et al., 2017; Witty et al., 2018) and self-attention architectures (Tang et al., 2020). In contrast, PSEs exploits behavior similarity (Section 3), a property related to the sequential aspect of RL. Furthermore, for certain RL tasks, it can be unclear what an optimality invariant data augmentation would look like (Section 5.3). PSM can quantify the invariance of such augmentations (Proposition D.2). + +Meta-learning is also related to generalization. Meta-learning methods try to find a parametrization that requires a small number of gradient steps to achieve good performance on a new task (Finn et al., 2017). In this context, various meta-learning approaches capable of zero-shot generalization have been proposed (Li et al., 2018; Agarwal et al., 2019; Balaji et al., 2018). These approaches typically consist in minimizing the loss in the environments the agent is while adding an auxiliary loss for ensuring improvement in the other (validation) environments available to the agent. Nevertheless, Tachet des Combes et al. (2018) has shown that meta-learning approaches fail in the jumping task which we also observed empirically. Others have also reported similar findings (e.g., Farebrother et al., 2018). + +There are several other approaches for tackling zero-shot generalization in RL, but they often rely on domain-specific information. Some examples include knowledge about equivalences between entities in the environment (Oh et al., 2017) and about what is under the agent’s control (Ye et al., 2020). Causality-based methods are a different way of tackling generalization, but current solutions do not scale to high-dimensional observation spaces (e.g., Killian et al., 2017; Perez et al., 2020; Zhang et al., 2020). + +# G JUMPING TASK WITH PIXELS + +![](images/ea6505432a67100e306e9753d67f0ba9bec335abc00370ec65ac6866a23b14e2.jpg) +Figure G.1: Optimal trajectories on the jumping tasks for two different environments. Note that the optimal trajectory is a sequence of right actions, followed by a single jump at a certain distance from the obstacle, followed by right actions. + +Detailed Task Description. The jumping task consists of an agent trying to jump over an obstacle on a floor. The environment is deterministic with the agent observing a reward of $+ 1$ at each time step. If the agent successfully reaches the rightmost side of the screen, it receives a bonus reward of $+ 1 0 0$ ; if the agent touches the obstacle, the episode terminates. The observation space is the pixel representation of the environment, as depicted in Figure 1. The agent has access to two actions: right and jump. The jump action moves the agent vertically and horizontally to the right. + +Architecture. The neural network used for Jumping Task experiment is adapted from the Nature DQN architecture. Specifically, the network consists of 3 convolutional layers of sizes 32, 64, 64 with filter sizes $8 \times 8$ , $4 \times 4$ and $3 \times 3$ and strides 4, 2, and 1, respectively. The output of the convnet is fed into a single fully connected layer of size 256 followed by ’ReLU’ non-linearity. Finally, this FC layer output is fed into a linear layer which computes the policy which outputs the probability of the jump and right actions. + +Contrastive Embedding. For all our experiments, we use a single ReLU layer with $k = 6 4$ units for the non-linear projection to obtain the embedding $z _ { \theta }$ (Figure A.1). We compute the embedding using the penultimate layer in the jumping task network. Hyperparameters are reported in Table G.2. + +Total Loss. For jumpy world, the total loss is given by $\mathcal { L } _ { \mathrm { I L } } + \alpha \mathcal { L } _ { \mathrm { C M E } }$ where ${ \mathcal { L } } _ { \mathrm { I L } }$ is the imitation learning loss, $\mathcal { L } _ { \mathrm { C M E } }$ is the auxiliary loss for learning PSEs with the coefficient $\alpha$ . + +![](images/b0d8815c5d7275437c5956e7dcd2666091c0d6503fae943844e0b9a282c3036a.jpg) +Figure G.3: Test performance curves in the setting without data augmentation on the “wide”, “narrow”, and random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25 and 75 percentiles. + +![](images/7ae61140cd1c0eae07e313b43e44f832d172fa581accd9edd3d1ff30eb85643b.jpg) +Figure G.2: PSM vs. Bisimulation. Visualizing PSM and bisimulation metrics with discount factor $\gamma = 0 . 9 9$ . $x _ { i }$ and $y _ { i }$ correspond to the states visited by optimal policies in the two environments with obstacle at positions 25 and 45, respectively. For each grid, the $( \dot { i } , j ) ^ { t h }$ location shows the distance assigned by the metric to the states $x _ { i }$ and $y _ { j }$ . Lower distances (lighter shades) imply higher similarity between two states. The bottom right figure shows which states are equivalent to each other. The ranges are different for each metric as bisimulation metrics utilize reward differences while PSM uses the total variation $( T V )$ distance between policies. Note that the large bisimulation distances are due to the fact that the reward at the terminal state is set to 100. + +![](images/42a79abb3d752df4e248de208365ef50ff1cf065c5fbce26bd1ba3fcb36ec3f1.jpg) +Figure G.4: Test performance curves in the setting with data augmentation on the “wide”, “narrow”, and random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25 and 75 percentiles. + +![](images/23cbe4e7878d1818c802facc5015190dbca31d0a9ec6d66d453c3d286303c25d.jpg) +Figure G.5: Jumping task with colored obstacles. The agent needs to jump over the red obstacle but strike the green obstacle. + +Table G.1: Percentage $( \% )$ of test tasks solved when trained on the “wide” grid with both red and green obstacles. The numbers we report are averaged across 100 runs. Standard error is reported between parentheses. + +
MethodRed (%)Green (%)
RandConv6.2 (0.4)99.6 (0.2)
Dropout and l2 reg.19.5(0.2)100.0 (0.0)
RandConv +PSEs29.8 (1.3)99.6 (0.2)
PSEs37.9 (1.9)100.0 (0.0)
+ +![](images/b97fc7dbb58cf30e9e1722e2dc460df60c4711f902a11b8b59a27bf8a70f5846.jpg) +Figure G.6: Randconv enforces color invariance. The first column shows the original observation where the top row corresponds to the observation from task with green obstacle jumping task while the bottom row corresponds to the red obstacle jumping task. Columns 2-6 in each row show 5 augmentations by applying RandConv. The augmentations show that RandConv tries to encode invariance with respect to the obstacle color. + +# G.2 HYPERPARAMETERS + +For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the “wide” grid and pick the parameters with the best validation performance. The validation set (Figure G.7) was selected by using the environments nearby to the training environments whose floor height differ by 1 or whose obstacle position differ by 1. + +Table G.2: Common hyperparameters across all methods for all jumping task experiments. + +
HyperparameterValue
Learning rate decay Training epochs0.999
Optimizer2000 Adam
Batch size (Imitation)256
Num training tasks I-scale Parameter (β)18
Embedding size (k)0.01
Batch Size (LCME)64
57
|T57
+ +![](images/c50466fd3830e13b4efa16578edae345ae90199f2967bef1d51b49352ce01705.jpg) +Figure G.7: Unseen 56 validation tasks, labeled with V, used for hyperparameter selection. + +Table G.3: Optimal hyperparameters for reporting results in Table 1. These hyperparameters are selected using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. All grid configurations in Table 1 use these hyperparameters. + +
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-33.2×10-37×10-32.6×10-3
l2-reg. coefficient4.3×10-41×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)11.05×10-1
Auxiliary loss coefficient (α)11×10115.0
+ +Table G.4: Optimal hyperparameters for reporting results in Figure 5.3. These hyperparameters are selected using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. + +
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-36×10-35×10-32.6×10-3
l2-reg. coefficient4.3×10-47×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)15×10-115×10-1
Auxiliary loss coefficient (α)15.015.0
+ +Table G.5: Optimal hyperparameters for reporting ablation results in Table 2. These hyperparameters are selected using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. + +
HyperparameterPSMπ*-bisimulation
CMEsl2-embeddingsCMEsl2-embeddings
Learning Rate4×10-35×10-44.7×10-41×10-4
Contrastive Temperature (1/入)1.015×10-11
Auxiliary loss coefficient (α)5.01×10-11×10-11×10-6
+ +Please note that Table G.3 and Table G.4 correspond to two different tasks: one uses the standard jumping task with white obstacles, while the other uses colored obstacles where the optimal policies depend on color. For fair comparison, we tune hyperparameters for all the methods using Bayesian optimization (Golovin et al., 2017). We use the best parameters among these tuned hyperparameters and the ones found in Table G.3, leading to different parameters for both PSEs as well as RandConv. Evaluating PSEs with the jumping task hyperparameters from Table G.3 instead of the ones in Table G.4 leads to a small drop $( - 4 \% )$ on the jumping task with colors (Section 5.3). Nevertheless, PSEs still outperform other methods in Section 5.3. + +# H LQR: ADDITIONAL DETAILS + +Optimal control with linear dynamics and quadratic cost, commonly known as LQR, has been increasingly used as a simplified surrogate for deep RL problems (Recht, 2019). Following Song et al. (2019); Sonar et al. (2020), we analyze the following LQR problem for assessing generalization: + +$$ +\begin{array} { r l } & { \mathrm { m i n i m i z e } \quad E _ { s _ { 0 } \sim \mathcal { D } } \left[ \frac { 1 } { 2 } \sum _ { t = 0 } ^ { \infty } s _ { t } ^ { T } Q s _ { t } + a _ { t } ^ { T } R a _ { t } \right] , } \\ & { \mathrm { s u b j e c t \ t o } \quad s _ { t + 1 } = A s _ { t } + B a _ { t } , o _ { t } = \left[ \begin{array} { l } { 0 . 1 W _ { c } } \\ { W _ { d } } \end{array} \right] s _ { t } , a _ { t } = K o _ { t } , } \end{array} +$$ + +where $\mathcal { D }$ is the initial state distribution, $s _ { t } \in \mathbb { R } ^ { n _ { s } }$ is the (hidden) true state at time t, $a _ { t } \in \mathbb { R } ^ { n _ { a } }$ is the control action, and $K$ is the linear policy matrix. The agent receives the input observation $o _ { t }$ , which is a linear transformation of the state $s _ { t }$ . $W _ { c }$ and $W _ { d }$ are semi-orthogonal matrices to prevent information loss for predicting optimal actions. An environment corresponds to a particular choice of $W _ { d }$ ; all other system parameters $( A , B , Q , R , W _ { c } )$ are fixed matrices which are shared across environments and unknown to the agent. The agent learns the policy matrix $K$ using $N$ training environments based on Equation H.1. At test time, the learned policy is evaluated on environments with unseen $W _ { d }$ . + +Table H.1: LQR generalization performance: Absolute error in LQR cost, w.r.t. the oracle solver (which has access to true state), of various methods trained with $n _ { d }$ distractors on $N = 2$ environments. The reported mean and standard deviations are across 100 different seeds. Lower error is better. + +
MethodNumber of Distractors (nd)
500100010000
Overparametrization (Song et al., 2019)25.8 (1.5)24.9 (1.1)24.9 (0.4)
IPO (Sonar et al., 202O) (IRM + Policy opt.)32.6 (5.0)27.3 (2.8)24.8 (0.4)
Weight Sparsity (ll-reg.)28.2 (0.0)28.2 (0.0)28.2 (0.0)
PSM (State aggregation)0.03 (0.0)0.03 (0.0)0.02 (0.0)
+ +The generalization challenge in this setup is to ignore the distractors: $W _ { c } s _ { t } \in \mathbb { R } ^ { n _ { s } }$ represents the state features invariant across environments while $W _ { d } s _ { t } \in \mathbb { R } ^ { n _ { d } }$ is a high-dimensional distractor of size $n _ { s } , n _ { d }$ , respectively, such that $n _ { s } < < n _ { d }$ . Furthermore, the policy matrix which generalizes across all environments is $K _ { \star } = \bigg [ 1 0 W _ { c } P _ { \star } ^ { ~ T } \bigg ] ^ { T }$ , where $P _ { \star }$ corresponds to the optimal LQR solution with access to state $s _ { t }$ . However, for a single environment with distractor $W _ { d }$ , multiple solutions exist, for instance, $K _ { \star } { } ^ { \prime } = \left[ \begin{array} { c c } { 1 0 \alpha W _ { c } P _ { \star } ^ { \mathrm { ~ \top ~ } } } \\ { ( 1 - \alpha ) W _ { d } P _ { \star } ^ { \mathrm { ~ \top ~ } } } \end{array} \right] ^ { T } \quad \forall \alpha \in [ 0 , 1 ]$ . Note that the distractors are an order of magnitude larger than the invariant features in $o _ { t }$ and dependence on them is likely to cause the agent to act erratically on inputs with unseen distractors, resulting in poor generalization. + +We use overparametrized policies with two linear layers, i.e., $K = K _ { 1 } K _ { 2 }$ , where $K _ { 1 } ( o )$ is the learned representation for observation $o$ . We learn $K$ using gradient descent using the combined cost on 2 training environments with varying number of distractors. We aggregate observation pairs with near-zero PSM by matching their representations using a squared loss. We use the open-source code released by Sonar et al. (2020) for our experiments. + +Table H.2: An overview of hyper-parameters for LQR. + +
ParameterSeting
AOrthogonal matrix, scaled 0.8
BI20×20
nx20
na20
QI20×20
RI20×20
KiAiOrthogonal Initialization, scaled 0.001
WdRandom semi-orthogonal matrix
+ +The reliance on distractors for IPO also highlights a limitation of IRM: if a model can achieve a solution with zero training error, then any such solution is acceptable by IRM regardless of its generalization ability – a common scenario with overparametrized deep neural nets (Jin et al., 2020). + +# H.1 NEAR-OPTIMALITY OF PSM AGGREGATION + +Conjecture 1. Assuming zero state aggregation error with policy similarity metric (PSM), the policy matrix $K$ learned using gradient descent is independent of the distractors. + +Proof. For LQR domains $x , y$ , an observation pair $( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )$ has zero PSM iff the underlying state $s _ { t }$ is same for both the observations in the pair. This is true, as (a) both domains has the same transition + +dynamics, as specified by Equation H.1, and (b) the optimal policy is deterministic and is completely determined by the current state $s _ { t }$ at any time $t$ . + +Assume $o _ { t } { } ^ { x } = \left[ \begin{array} { c } { 0 . 1 W _ { c } } \\ { W _ { d ^ { x } } } \end{array} \right] s _ { t }$ and $o _ { t } { } ^ { y } = \left[ \stackrel { 0 . 1 } { W } _ { d ^ { y } } \right] s _ { t }$ for distractor semi-orthogonal matrices $W _ { d ^ { x } }$ and $W _ { d } y$ , respectively. Furthermore, the representation is given by $K _ { 1 } ( o _ { t } ^ { x } )$ and $K _ { 1 } ( o _ { t } ^ { y } )$ respectively. Assume that $K _ { 1 } = [ K _ { s } \quad K _ { d } ]$ where $K _ { s } \in \mathbb { R } ^ { h \times n _ { s } }$ and $K _ { d } \in \mathbb { R } ^ { h \times n _ { d } }$ and $K _ { 1 } \in \mathbb { R } ^ { h \times ( n _ { s } + n _ { d } ) }$ . + +Zero state-aggregation error with squared loss implies that for pair $( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )$ corresponding to $s _ { t }$ , + +$$ +K _ { 1 } ( o _ { t } { ^ x } - o _ { t } { ^ y } ) = K _ { 1 } [ W _ { d { ^ x } } - W _ { d { ^ y } } ] s _ { t } = 0 \implies K _ { d } ( W _ { d { ^ x } } - W _ { d { ^ y } } ) s _ { t } = 0 +$$ + +As Equation H.2 holds for all states visited by the optimal policy in an infinite horizon LQR, it follows that $K _ { d } ( W _ { d ^ { x } } - W _ { d ^ { y } } ) = 0$ . + +Furthermore, it is well-known that gradient descent tends to find low-rank solutions due to implicit regularization (Arora et al., 2019; Gunasekar et al., 2017), e.g.,with small enough step sizes and initialization close enough to the origin, gradient descent on matrix factorization converges to the minimum nuclear norm solution for 2 layer linear networks (Gunasekar et al., 2017). Based on this, we conjecture that $K _ { d } = 0$ which we found to be true in practice. □ + +# I DISTRACTING CONTROL SUITE + +We use the same setup as Kostrikov et al. (2020); Stone et al. (2021) for implementation details and training protocol. For completeness, we describe the details below. + +Dynamic Background Distractions. In Distracting Control Suite (Stone et al., 2021), random backgrounds are projected from scenes of the DAVIS 2017 dataset (Pont-Tuset et al., 2017) onto the skybox of the scene. To make these backgrounds visible for all tasks and views, the floor grid is semi-transparent with transparency coefficient 0.3. We take the first 2 videos in the DAVIS 2017 training set and randomly sample a scene and a frame from those at the start of every episode. In the dynamic setting, the video plays forwards or backwards until the last or first frame is reached at which point the video is played backwards. This way, the background motion is always smooth and without “cuts”. + +Soft Actor-Critic. Soft Actor-Critic (SAC) (Haarnoja et al., 2018) learns a state-action value function $Q _ { \theta }$ , a stochastic policy $\pi _ { \theta }$ and a temperature $\alpha$ to find an optimal policy for an MDP $( S , \mathcal { A } , p , r , \gamma )$ by optimizing a $\gamma$ -discounted maximum-entropy objective. $\theta$ is used generically to denote the parameters updated through training in each part of the model. The actor policy $\pi _ { \boldsymbol { \theta } } ( a _ { t } | \boldsymbol { s } _ { t } )$ is a parametric tanhGaussian that given $s _ { t }$ samples $a _ { t } \dot { = } \operatorname { t a n h } ( \mu _ { \theta } ( s _ { t } ) + \sigma _ { \theta } ( s _ { t } ) \epsilon )$ , where $\epsilon \sim \mathcal { N } ( 0 , 1 )$ and $\mu _ { \theta }$ and $\sigma _ { \theta }$ are parametric mean and standard deviation. + +The policy evaluation step learns the critic $Q _ { \theta } ( s _ { t } , a _ { t } )$ network by optimizing a single-step of the soft Bellman residual + +$$ +\begin{array} { r l } & { J _ { Q } ( \mathcal { D } ) = E _ { ( s _ { t } , a _ { t } , s _ { t } ^ { \prime } ) \sim \mathcal { D } } [ ( Q _ { \theta } ( s _ { t } , a _ { t } ) - y _ { t } ) ^ { 2 } ] } \\ & { \qquad a _ { t } ^ { \prime } { \sim } \pi ( \cdot | s _ { t } ^ { \prime } ) } \\ & { y _ { t } = r ( s _ { t } , a _ { t } ) + \gamma [ Q _ { \theta ^ { \prime } } ( s _ { t } ^ { \prime } , a _ { t } ^ { \prime } ) - \alpha \log \pi _ { \theta } ( a _ { t } ^ { \prime } | s _ { t } ^ { \prime } ) ] , } \end{array} +$$ + +where $\mathcal { D }$ is a replay buffer of transitions, $\theta ^ { \prime }$ is an exponential moving average of the weights. SAC uses clipped double-Q learning, which we omit for simplicity but employ in practice. + +The policy improvement step then fits the actor policy $\pi _ { \boldsymbol { \theta } } ( a _ { t } | \boldsymbol { s } _ { t } )$ network by optimizing the objective + +$$ +J _ { \pi } ( \mathcal { D } ) = E _ { s _ { t } \sim \mathcal { D } } [ D _ { \mathrm { K L } } ( \pi _ { \theta } ( \cdot | s _ { t } ) | | \exp \{ \frac { 1 } { \alpha } Q _ { \theta } ( s _ { t } , \cdot ) \} ) ] . +$$ + +Finally, the temperature $\alpha$ is learned with the loss + +$$ +J _ { \alpha } ( \mathcal { D } ) = E _ { \underset { a _ { t } \sim \pi _ { \theta } ( \cdot | s _ { t } ) } { s _ { t } \sim \mathcal { D } } } [ - \alpha \log \pi _ { \theta } ( a _ { t } | s _ { t } ) - \alpha \bar { \mathcal { H } } ] , +$$ + +where $\bar { \mathcal { H } } \in \mathbb { R }$ is the target entropy hyper-parameter that the policy tries to match, which in practice is usually set to $\bar { \mathcal { H } } = - | \bar { \mathcal { A } } |$ . + +![](images/7c1ed64c41245f4fe76e367be7d81f27ce231d876cb7515942b5a20a75e1bb2a.jpg) +Figure I.1: DCS Test Environments: Snapshots of test environments used for evaluating generalization on Distracting Control Suite. Random backgrounds are projected from scenes of the first 30 videos of DAVIS 2017 validation dataset. + +# I.1 ACTOR AND CRITIC NETWORKS + +Following Kostrikov et al. (2020), we use clipped double Q-learning for the critic, where each $Q$ -function is parametrized as a 3-layer MLP with ReLU activations after each layer except of the last. The actor is also a 3-layer MLP with ReLUs that outputs mean and covariance for the diagonal Gaussian that represents the policy. The hidden dimension is set to 1024 for both the critic and actor. + +# I.2 ENCODER NETWORK + +We employ the encoder architecture from Kostrikov et al. (2020). This encoder consists of four convolution layers with $3 \times 3$ kernels and 32 channels. The ReLU activation is applied after each convolutional layer. We use stride to 1 everywhere, except of the first convolutional layer, which has stride 2. The output of the convnet is feed into a single fully-connected layer normalized by LayerNorm. Finally, we apply tanh nonlinearity to the 50 dimensional output of the fullyconnected layer. We initialize the weight matrix of fully-connected and convolutional layers with the orthogonal initialization and set the bias to be zero. The actor and critic networks both have separate encoders, although we share the weights of the conv layers between them. Furthermore, only the critic optimizer is allowed to update these weights (i.e.,we stop the gradients from the actor before they propagate to the shared convolutional layers). + +# I.3 CONTRASTIVE METRIC EMBEDDING LOSS + +For all our experiments, we use a single ReLU layer with $k = 2 5 6$ units for the non-linear projection to obtain the embedding $z _ { \theta }$ (Figure A.1). We compute the embedding using the penultimate layer in the actor network. For picking the hyperparameters, we used 3 temperatures [0.1, 0.01, 1.0] and 3 auxiliary ${ \mathcal { L } } _ { \mathrm { C M E } }$ loss coefficients [1, 3, 10] using “Ball In Cup Catch” as the validation environment. All other hyperparameters are the same as prior work (see Table I.2). + +We approximate optimal policies with the policies obtained after training a DrQ agent for 500K environment steps. Since a given action sequence from this approximate policy has the same performance across different training environments, we compute the PSM across training environments, via dynamic programming (see Section J.1 for pseudo-code), using such action sequences. + +Total Loss. The total loss is given by $\mathcal { L } _ { \mathrm { R L } } + \alpha \mathcal { L } _ { \mathrm { C M E } }$ where $\mathcal { L } _ { \mathrm { R L } }$ is the reinforcement learning loss which combines $J _ { \pi } ( \mathcal { D } )$ , $J _ { \pi } ( \mathcal { D } )$ , and $J _ { \alpha } ( { \mathcal { D } } ) _ { , }$ ), while ${ \mathcal { L } } _ { \mathrm { C M E } }$ is the auxiliary loss for learning PSEs with the coefficient $\alpha$ . + +# I.4 TRAINING AND EVALUATION SETUP + +For evaluation, we use the first 30 videos from the DAVIS 2017 validation dataset (see Figure I.1). Each checkpoint is evaluated by computing the average episode return over 100 episodes from the unseen environments. All experiments are performed with five random seeds per task used to compute means and standard deviations/errors of their evaluations. We use $K = 2 , M = 2$ as prescribed by Kostrikov et al. (2020) for $\mathrm { D r Q }$ . Following Kostrikov et al. (2020) and Stone et al. (2021), we use a different action repeat hyper-parameter for each task, which we summarize in Table I.3. We construct an observational input as a 3-stack of consecutive frames (Kostrikov et al., 2020), where each frame is an RGB rendering of size $8 4 \times 8 4$ from the 0th camera. We then divide each pixel by 255 to scale it down to [0, 1] range. For data augmentation, we maintain temporal consistency by using the same crop augmentation across consecutive frames. + +Table I.1: Optimal hyperparameters for PSE auxiliary loss for reporting results in Table 3. + +
HyperparameterSetting
Contrastive temperature (1/入)0.1
Auxiliary loss coefficient (α)1.0
I-scale parameter (β)0.1
Batch Size(LcME)128
|1000 // Action Repeat
+ +Table I.2: Hyper-parameters taken from Kostrikov et al. (2020) in the Distracting Control Suite experiments. + +
ParameterSetting
Replay buffer capacity100,000
Seed steps1,000
Batch size (DrQ)512
Discount γ0.99
OptimizerAdam
Learning rate10-3
Critic target update frequency2
Critic Q-function soft-update rate T0.01
Actor update frequency2
Actor log stddev bounds
Init temperature[-10,2] 0.1
+ +Table I.3: The action repeat hyper-parameter used for each task in the Distracting Control Suite benchmark. + +
Task nameAction repeat
Cartpole Swingup8
Reacher Easy4
Cheetah Run4
Finger Spin2
Ball In Cup Catch4
Walker Walk2
+ +![](images/9058a9690116b883329a6466ccc12d2c432e08d5b417ce9c0fc8928c6ebed906.jpg) +Figure I.2: Random Initialization. Generalization performance on unseen environments over the course of training. The agent is initialized randomly. $_ \mathrm { D r Q }$ augmented with PSEs outperforms DrQ. We plot the average episode return across 5 seeds and the shaded region shows the standard deviation. Each checkpoint is evaluated using 100 episodes with unseen backgrounds. + +![](images/48d74111701471e8fc0c0104c5c1128c88f60454f9197c2358f703f65f4db49e.jpg) +Figure I.3: Pretrained Initialization. Generalization performance on unseen environments over the course of training. The agent is initialized using a pretrained $\mathrm { D r Q }$ agent. DrQ augmented with PSEs outperforms $\mathrm { D r Q }$ on most of the environments. We plot the average return across 5 seeds and the shaded region shows the standard deviation. Each checkpoint is evaluated using 100 episodes with unseen backgrounds. + +# J PSEUDO CODE + +# J.1 DYNAMIC PROGRAMMING FOR COMPUTING PSM + +1 def metric_fixed_point(cost_matrix, gamma ${ } = 0$ .99, ep $s { = } 1 { \mathrm { e - } } 7$ ): +2 """DP for calculating PSM in environments with deterministic dynamics. +3 +4 Args: +5 cost_matrix: DIST matrix where entries at index (i, j) is DIST(x_i, +y_j) +6 gamma: Metric discount factor. +7 eps: Threshold for stopping the fixed point iteration. +8 """ +9 d $=$ np.zeros_like(cost_matrix) +10 def operator(d_cur): +11 d_new $\ c = ~ 1$ $\star$ cost_matrix +12 discounted_d_cur $=$ gamma $\star$ d_cur +13 d_new[:-1, :-1] $+ =$ discounted_d_cur[1:, 1:] +14 d_new[:-1, -1] $+ =$ discounted_d_cur[1:, -1] +15 d_new[-1, :-1] $+ =$ discounted_d_cur[-1, 1:] +16 return d_new +17 +18 while True: +19 d_new $=$ operator(d) +20 if np.sum(np.abs(d - d_new)) $<$ eps: +21 break +22 else: +23 d $=$ d_new[:] +24 return d + +# J.2 CONTRASTIVE LOSS + +1 def contrastive_loss(similarity_matrix, +2 metric_values, +3 temperature, +4 bet $\gimel = 1 \digamma . 0$ ): +5 """Contrative Loss with embedding similarity .""" +6 metric_shape $=$ tf.shape(metric_values) +7 ## z_\theta(X): embedding_1 $=$ nn_model.representation(X) +8 ## z_\theta(Y): embedding_ $\_ =$ nn_model.representation(Y) +9 ## similarity_matrix $=$ cosine_similarity(embedding_1, embedding_2 +10 ## metric_values $=$ PSM(X, Y) +11 similarity_matrix / $=$ temperature +12 neg_logits1 $=$ similarity_matrix +13 +14 col_indices $=$ tf.cast(tf.argmin(metric_values, axi $\ S = 1$ ), dtype $=$ tf.int32) +15 pos_indices1 $=$ tf.stack( +16 (tf.range(metric_shape[0], dtype $=$ tf.int32), col_indices), axi $\mathsf { S } = 1$ ) +17 pos_logits1 $=$ tf.gather_nd(similarity_matrix, pos_indices1) +18 +19 metric_values / $=$ beta +20 similarity_measure $=$ tf.exp(-metric_values) +21 pos_weights1 $=$ -tf.gather_nd(metric_values, pos_indices1) +22 pos_logits1 $+ =$ pos_weights1 +23 negative_weights $=$ tf.math.log((1.0 - similarity_measure) $^ +$ 1e-8) +24 neg_logits1 $+ =$ tf.tensor_scatter_nd_update( +25 negative_weights, pos_indices1, pos_weights1) +26 +27 neg_logits1 $=$ tf.math.reduce_logsumexp(neg_logits1, axis $^ { = 1 }$ ) +28 return tf.reduce_mean(neg_logits1 - pos_logits1) # Equation 4 \ No newline at end of file diff --git a/parse/train/qda7-sVg84/qda7-sVg84_content_list.json b/parse/train/qda7-sVg84/qda7-sVg84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6451b5f25c9a7303844fa2941b728245f8209d6 --- /dev/null +++ b/parse/train/qda7-sVg84/qda7-sVg84_content_list.json @@ -0,0 +1,3575 @@ +[ + { + "type": "text", + "text": "CONTRASTIVE BEHAVIORAL SIMILARITY EMBEDDINGS FOR GENERALIZATION IN REINFORCEMENT LEARNING ", + "text_level": 1, + "bbox": [ + 174, + 99, + 671, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rishabh Agarwal∗ Marlos C. Machado‡ Pablo Samuel Castro Marc G. Bellemare Google Research, Brain Team {rishabhagarwal, marlosm, psc, bellemare}@google.com ", + "bbox": [ + 183, + 193, + 790, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 256, + 544, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reinforcement learning methods trained on few environments rarely learn policies that generalize to unseen environments. To improve generalization, we incorporate the inherent sequential structure in reinforcement learning into the representation learning process. This approach is orthogonal to recent approaches, which rarely exploit this structure explicitly. Specifically, we introduce a theoretically motivated policy similarity metric (PSM) for measuring behavioral similarity between states. PSM assigns high similarity to states for which the optimal policies in those states as well as in future states are similar. We also present a contrastive representation learning procedure to embed any state similarity metric, which we instantiate with PSM to obtain policy similarity embeddings $( \\mathrm { P S E s ^ { 1 } } )$ ). We demonstrate that PSEs improve generalization on diverse benchmarks, including LQR with spurious correlations, a jumping task from pixels, and Distracting DM Control Suite. ", + "bbox": [ + 233, + 280, + 766, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 465, + 336, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current reinforcement learning (RL) approaches often learn policies that do not generalize to environments different than those the agent was trained on, even when these environments are semantically equivalent (Tachet des Combes et al., 2018; Song et al., 2019; Cobbe et al., 2019). For example, consider a jumping task where an agent, learning from pixels, needs to jump over an obstacle (Figure 1). Deep RL agents trained on a few of these tasks with different obstacle positions struggle to solve test tasks where obstacles are at previously unseen locations. ", + "bbox": [ + 174, + 488, + 517, + 640 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abed9d708ee10498bdd0d0bf48ab3b19c58d372b5155e6350f85107e57f370fa.jpg", + "image_caption": [ + "Figure 1: Jumping task: The agent (white block), learning from pixels, needs to jump over an obstacle (grey square). The challenge is to generalize to unseen obstacle positions and floor heights in test tasks using a small number of training tasks. We show the agent’s trajectories using faded blocks. " + ], + "image_footnote": [], + "bbox": [ + 549, + 482, + 800, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent solutions to circumvent poor generalization in RL are adapted from supervised learning, and, as such, largely ignore the sequential aspect of RL. Most of these solutions revolve around enhancing the learning process, including data augmentation (e.g., Kostrikov et al., 2020; Lee et al., 2020a), regularization (Cobbe et al., 2019; Farebrother et al., 2018), noise injection (Igl et al., 2019), and diverse training conditions (Tobin et al., 2017); they rarely exploit properties of the sequential decision making problem such as similarity in actions across temporal observations. ", + "bbox": [ + 174, + 648, + 825, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instead, we tackle generalization by incorporating properties of the RL problem into the representation learning process. Our approach exploits the fact that an agent, when operating in environments with similar underlying mechanics, exhibits at least short sequences of behaviors that are similar across these environments. Concretely, the agent is optimized to learn an embedding in which states are close when the agent’s optimal policies in these states and future states are similar. This notion of proximity is general and it is applicable to observations from different environments. ", + "bbox": [ + 174, + 739, + 825, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, inspired by bisimulation metrics (Castro, 2020; Ferns et al., 2004), we propose a novel policy similarity metric (PSM). PSM (Section 3) defines a notion of similarity between states originated from different environments by the proximity of the long-term optimal behavior from these states. PSM is reward-agnostic, making it more robust for generalization compared to approaches that rely on reward information. We prove that PSM yields an upper bound on suboptimality of policies transferred from one environment to another (Theorem 1), which is not attainable with bisimulation. ", + "bbox": [ + 174, + 829, + 825, + 886 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We employ PSM for representation learning and introduce policy similarity embeddings (PSEs) for deep RL. To do so, we present a general contrastive procedure (Section 4) to learn an embedding based on any state similarity metric. PSEs are the instantiation of this procedure with PSM. PSEs are appealing for generalization as they encode task-relevant invariances by putting behaviorally equivalent states together. This is unlike prior approaches, which rely on capturing such invariances without being explicitly trained to do so, for example, through value function similarities across states (e.g., Castro & Precup, 2010), or being robust to fixed transformations of the observation space (e.g., Kostrikov et al., 2020; Laskin et al., 2020a). ", + "bbox": [ + 174, + 138, + 825, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PSEs lead to better generalization while being orthogonal to how most of the field has been tackling generalization. We illustrate the efficacy and broad applicability of our approach on three existing benchmarks specifically designed to test generalization: (i) jumping task from pixels (Tachet des Combes et al., 2018) (Section 5), (ii) LQR with spurious correlations (Song et al., 2019) (Section 6.1), and (iii) Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b) (Section 6.2). Our approach improves generalization compared to a wide variety of approaches including standard regularization (Farebrother et al., 2018; Cobbe et al., 2019), bisimulation (Castro & Precup, 2010; Castro, 2020; Zhang et al., 2021), out-of-distribution generalization (Arjovsky et al., 2019) and state-of-the-art data augmentation (Kostrikov et al., 2020; Lee et al., 2020a). ", + "bbox": [ + 174, + 257, + 825, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 174, + 392, + 339, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We describe an environment as a Markov decision process (MDP) (Puterman, 1994) $\\begin{array} { r l } { \\mathcal { M } } & { { } = } \\end{array}$ $( \\mathcal { X } , \\mathcal { A } , R , P , \\gamma )$ , with a state space $\\mathcal { X }$ , an action space $\\mathcal { A }$ , a reward function $R$ , transition dynamics $P$ , and a discount factor $\\gamma \\in [ 0 , 1 )$ . A policy $\\pi ( \\cdot \\mid x )$ maps states $x \\in \\mathcal { X }$ to distributions over actions. Whenever convenient, we abuse notation and write $\\pi ( x )$ to describe the probability distribution $\\pi ( \\cdot | x )$ , treating $\\pi ( x )$ as a vector. In RL, the goal is to find an optimal policy $\\pi ^ { * }$ that maximizes the cumulative expected return $\\mathbb E _ { a _ { t } \\sim \\pi ( \\cdot \\textit { \\textbf { \\ i } } x _ { t } ) } [ \\sum _ { t } \\bar { \\gamma } ^ { t } R ( x _ { t } , a _ { t } ) ]$ starting from an initial state $x _ { 0 }$ . ", + "bbox": [ + 173, + 416, + 825, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are interested in learning a policy that generalizes across related environments. We formalize this by considering a collection $\\rho$ of MDPs, sharing an action space $\\mathcal { A }$ but with disjoint state spaces. We use $\\mathcal { X }$ and $\\mathcal { V }$ to denote the state spaces of specific environments, and write $R _ { \\mathcal { X } }$ , $P _ { \\mathcal { X } }$ for the reward and transition functions of the MDP whose state space is $\\mathcal { X }$ , and $\\pi _ { \\mathcal { X } } ^ { * }$ for its optimal policy, which we assume unique without loss of generality. For a given policy $\\pi$ , we further specialize these into $R _ { \\mathcal { X } } ^ { \\pi }$ and $P _ { \\mathcal { X } } ^ { \\pi }$ , the reward and state-to-state transition dynamics arising from following $\\pi$ in that MDP. ", + "bbox": [ + 174, + 507, + 825, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We write $s$ for the union of the state spaces of the MDPs in $\\rho$ . Concretely, different MDPs correspond to specific scenarios in a problem class (Figure 1), and $s$ is the space of all possible configurations. Used without subscripts, $R , P$ , and $\\pi$ refer to the reward and transition function of this “union MDP”, and a policy defined over $s$ ; this notation simplifies the exposition. We measure distances between states across environments using pseudometrics2 on $s$ ; the set of all such pseudometrics is $\\mathbb { M }$ , and $\\mathbb { M } _ { p }$ is the set of metrics on probability distributions over $s$ . ", + "bbox": [ + 174, + 598, + 825, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our setting, the learner has access to a collection of training MDPs $\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }$ , drawn from $\\rho$ . After interacting with these environments, the learner must produce a policy $\\pi$ over the entire state space $s$ , which is then evaluated on unseen MDPs from $\\rho$ . Similar in spirit to the setting of transfer learning (Taylor & Stone, 2009), here we evaluate the policy’s zero-shot performance on $\\rho$ . ", + "bbox": [ + 174, + 689, + 825, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our policy similarity metric (Section 3) builds on the concept of $\\pi$ -bisimulation (Castro, 2020). Under the $\\pi$ -bisimulation metric, the distance between two states, $x$ and $y$ , is defined in terms of the difference between the expected rewards obtained when following policy $\\pi$ . The $\\pi$ -bisimulation metric $d _ { \\pi }$ satisfies a recursive equation based on the 1-Wasserstein metric $\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }$ , where $\\mathcal { W } _ { 1 } ( d ) ( A , B )$ is the minimal cost of transporting probability mass from $A$ to $B$ (two probability distributions on $s$ ) under the base metric $d$ (Villani, 2008). The recursion is ", + "bbox": [ + 174, + 752, + 825, + 837 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/620aa4a43bc93eafeecfb704f3618718837319bec60a3d823ad690a9332d2a62.jpg", + "text": "$$\n\\begin{array} { r } { d _ { \\pi } ( x , y ) = | R ^ { \\pi } ( x ) - R ^ { \\pi } ( y ) | + \\gamma \\mathcal { W } _ { 1 } ( d _ { \\pi } ) \\big ( P ^ { \\pi } ( \\cdot | x ) , P ^ { \\pi } ( \\cdot | y ) \\big ) , \\qquad x , y \\in \\mathcal { S } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 238, + 843, + 759, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve good generalization properties, we learn an embedding function $z _ { \\theta } : \\mathcal { S } \\mathbb { R } ^ { k }$ that reflects the information encoded in the policy similarity metric; this yields a policy similarity embedding (Section 4). We use contrastive methods (Hadsell et al., 2006; Oord et al., 2018) whose track record for representation learning is well-established. We adapt SimCLR (Chen et al., 2020), a popular contrastive method for learning embeddings of image inputs. Given two inputs $x$ and $y$ , their embedding similarity is $s _ { \\theta } ( x , y ) = s i m ( z _ { \\theta } ( x ) , z _ { \\theta } ( y ) )$ , where $\\begin{array} { r } { s i m ( u , v ) = \\frac { u ^ { T } v } { \\| u \\| \\| v \\| } } \\end{array}$ denotes the cosine similarity function. SimCLR aims to maximize similarity between augmented versions of an image (e.g., crops, colour changes) while minimizing its similarity to other images. The loss used by SimCLR for two versions $x , y$ of an image, and a set $\\mathcal { X } ^ { \\prime }$ containing other images is: ", + "bbox": [ + 176, + 871, + 823, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 102, + 826, + 208 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/e728c707965bd848cfff2b88e2d41e2810b434960e667d469684e6e39308eb85.jpg", + "text": "$$\n\\ell _ { \\theta } ( x , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) } { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) + \\sum _ { x ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\left\\{ x \\right\\} } \\exp ( \\lambda s _ { \\theta } ( x ^ { \\prime } , y ) ) }\n$$", + "text_format": "latex", + "bbox": [ + 271, + 212, + 725, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda$ is an inverse temperature hyperparameter. The overall SimCLR loss is then the expected value of $\\ell _ { \\theta } ( x , y ; S )$ , when $x , y$ , and $\\mathcal { X } ^ { \\prime }$ are drawn from some augmented training distribution. ", + "bbox": [ + 174, + 256, + 823, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 POLICY SIMILARITY METRIC ", + "text_level": 1, + "bbox": [ + 176, + 292, + 450, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A useful tool in learning a policy that generalizes is to understand which states result in similar behavior, and which do not. To be maximally effective, this similarity should go beyond the immediately chosen action and consider long-term behavior. In this regards, the $\\pi$ -bisimulation metrics are interesting as they are based on the full sequence of future rewards received from different states. However, considering rewards can be both too restrictive (when the policies are the same, but the obtained rewards are not; see Figure D.1) or too permissive (when the policies are different, but the obtained rewards are not; see Figure 3a). In fact, $\\pi$ -bisimulation metrics actually lead to poor generalization in our experiments (Sections 5.1 and 5.2). ", + "bbox": [ + 173, + 318, + 826, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we instead consider the similarity between policies themselves. We replace the absolute reward difference by a probability pseudometric between policies, denoted DIST. Additionally, since we would like to perform well in unseen environments, we are interested in similarity in optimal behavior. We thus use $\\pi ^ { * }$ as the grounding policy. This yields the policy similarity metric (PSM), for which states are close when the optimal policies in these states and future states are similar. For a given DIST, the PSM $d ^ { * } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }$ satisfies the recursive equation ", + "bbox": [ + 173, + 435, + 826, + 520 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/9ac02354c480f08d23c9d5364cc9a542168e6e21775e16792bd3b5b35a59511a.jpg", + "text": "$$\n\\begin{array} { r l r } & { } & { d ^ { * } ( x , y ) = \\underbrace { \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) } _ { \\mathrm { ( A ) } } + \\underbrace { \\gamma \\mathcal { W } _ { 1 } ( d ^ { * } ) \\big ( P ^ { \\pi ^ { * } } ( \\cdot | x ) , P ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) } _ { \\mathrm { ( B ) } } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 272, + 525, + 725, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The DIST term captures the difference in local optimal behavior (A) while $\\mathcal { W } _ { 1 }$ captures long-term optimal behavior difference (B); the exact weights assigned to the two are given by the discount factor. Furthermore, when DIST is bounded, $d ^ { * }$ is guaranteed to be finite. While there are technically multiple PSMs (one for each DIST), we omit this distinction whenever clear from context. A proof of the uniqueness of $d ^ { * }$ is given in Proposition B.1. ", + "bbox": [ + 173, + 574, + 825, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main use of PSM will be to compare states across environments. In this context, we identify the terms in Equation 3 with specific environments for clarity and write (despite its technical inaccuracy) ", + "bbox": [ + 173, + 651, + 825, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/ef10847f7eac97b0d10c96ae00eaa9d1f6547c2a192ff9eb67cf97571f068e91.jpg", + "text": "$$\n\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 269, + 685, + 727, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PSM is applicable to both discrete and continuous action spaces. In our experiments, DIST is the total variation distance $( T V )$ when $\\mathcal { A }$ is discrete, and we use the $\\ell _ { 1 }$ distance between the mean actions of the two policies when $\\mathcal { A }$ is continuous. PSM can be computed iteratively using dynamic programming (Ferns et al., 2011) (more details in Section D.1). Furthermore, when $\\pi ^ { * }$ is unavailable on training environments, we replace it by an approximation $\\hat { \\pi } ^ { * }$ to obtain an approximate PSM, which is close to the exact PSM depending on the suboptimality of $\\hat { \\pi } ^ { * }$ (Proposition D.3). ", + "bbox": [ + 173, + 712, + 825, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite resembling $\\pi$ -bisimulation metrics in form, the PSM possesses different characteristics which are better suited to the problem of generalizing a learned policy. To illustrate this point, consider the following simple nearest-neighbour scheme: Given a state $y \\in \\mathcal { V }$ , denote its closest match in $\\mathcal { X }$ by $\\begin{array} { r } { \\tilde { x } _ { y } : = \\arg \\operatorname* { m i n } _ { x \\in \\mathcal { X } } d ^ { * } ( x , y ) } \\end{array}$ . Suppose that we use this scheme to transfer $\\pi _ { \\mathcal { X } } ^ { * }$ to $\\mathcal { M } _ { \\mathcal { Y } }$ , in the sense that we behave according to the policy $\\tilde { \\pi } ( y ) = \\pi _ { \\chi } ^ { * } ( \\tilde { x } _ { y } )$ . We can then bound the difference between $\\tilde { \\pi }$ and $\\pi _ { \\mathcal { Y } } ^ { * }$ , something that is not possible if $d ^ { * }$ is replaced by a $\\pi$ -bisimulation metric. ", + "bbox": [ + 173, + 801, + 825, + 887 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. [Bound on policy transfer] For any $y \\in \\mathcal { V }$ , let $Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )$ define the sequence of random states encountered starting in $Y _ { y } ^ { 0 } = y$ and following policy $\\tilde { \\pi }$ . We have: ", + "bbox": [ + 173, + 891, + 820, + 926 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/cd132237f035791e31f3d0bfc25305f208b4467de5d698d767c39bc00821433b.jpg", + "text": "$$\n\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) .\n$$", + "text_format": "latex", + "bbox": [ + 312, + 99, + 683, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof is in the Appendix (Section B). Theorem 1 is non-vacuous whenever $d ^ { * } ( \\tilde { x } _ { y } , y ) < 1 / ( 1 + \\gamma )$ . In particular, $d ^ { * } ( \\tilde { x } _ { y } , y ) = 0$ implies that the transferred policy is optimal. Although this scheme is not practical (computing $d ^ { * }$ requires knowledge of $\\pi _ { \\mathcal { D } } ^ { * }$ ), it shows that meaningful policy generalization can be obtained if we can find a mapping that generalizes across $s$ . Put another way, PSM gives us a principled way of lifting generalization across inputs (a supervised learning problem) to generalization across environments. We now describe how to employ PSM for learning representations that put together states in which the agent’s long-term optimal behavior is similar. ", + "bbox": [ + 173, + 155, + 826, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 LEARNING CONTRASTIVE METRIC EMBEDDINGS ", + "text_level": 1, + "bbox": [ + 174, + 262, + 616, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generalize a learned policy to new environments, we build on the success of contrastive representations (Section 2). Given a state similarity metric $d$ , we develop a general procedure (Algorithm 1) to learn contrastive metric embeddings (CMEs) for $d$ . We utilize the metric $d$ for defining the set of positive and negative pairs, as well as assigning importance weights to these pairs in the contrastive loss (Equation 4). ", + "bbox": [ + 173, + 286, + 452, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Contrastive Metric Embeddings (CMEs) ", + "text_level": 1, + "bbox": [ + 464, + 290, + 821, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Given: State embedding $z _ { \\theta } ( \\cdot )$ , Metric $d ( \\cdot , \\cdot )$ Training environments $\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }$ · · · . Hyperparameters: Temperature $1 / \\lambda$ , Scale $\\beta$ , Total training steps $K$ \n2: for each step $k = 1 , \\ldots , K$ do \n3: Sample a pair of training MDPs $\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }$ \n4: Update $\\theta$ to minimize ${ \\mathcal { L } } _ { \\mathrm { C M E } }$ where $\\mathcal { L } _ { \\mathrm { C M E } } = \\mathbb { E } _ { \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } \\sim \\rho } \\left[ L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) \\right]$ \n5: end for ", + "bbox": [ + 470, + 310, + 825, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first apply a transformation to convert $d$ to a similarity measure $\\Gamma$ , bounded in [0, 1] for “soft” similarities. In this work, we transform $d$ using the Gaussian kernel with a positive scale parameter $\\beta$ , that is, $\\Gamma ( x , y ) = \\exp ( - d ( x , y ) / \\beta )$ . $\\beta$ controls the sensitivity of the similarity measure to $d$ . ", + "bbox": [ + 173, + 431, + 826, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we select the positive and negative pairs given a set of states ${ \\mathcal { X } } ^ { \\prime } \\subseteq { \\mathcal { X } }$ and $\\mathcal { V }$ from MDPs $\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }$ respectively. For each anchor state $y \\in \\mathcal { V }$ , we use its nearest neighbor in $\\mathcal { X } ^ { \\prime }$ based on the similarity measure $\\Gamma$ to define the positive pairs $\\{ ( \\tilde { x } _ { y } , y ) \\}$ , where $\\tilde { x } _ { y } = \\underset { x \\in \\mathcal { X } ^ { \\prime } } { \\operatorname { a r g m a x } } \\Gamma ( x , y )$ . The remaining states in $\\mathcal { X } ^ { \\prime }$ , paired with $y$ , are used as negative pairs. This choice of pairs is motivated by Theorem 1, which shows that if we transfer the optimal policy in $\\mathcal { M } _ { \\mathcal { X } }$ to the nearest neighbors defined using PSM, its performance in $\\mathcal { M } _ { \\mathcal { Y } }$ has suboptimality bounded by PSM. ", + "bbox": [ + 173, + 481, + 826, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we define a soft version of the SimCLR contrastive loss (Equation 2) for learning the function $z _ { \\theta }$ , which maps states (usually high-dimensional) to embeddings. Given a positive state pair $( \\tilde { x } _ { y } , y )$ , the set $\\mathcal { X } ^ { \\prime }$ , and the similarity measure $\\Gamma$ , the loss (pseudocode provided in Section J.2) is given by ", + "bbox": [ + 174, + 580, + 825, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/50d884eead6fd8b8db5a7642e5b04245f374580b54d8c1e73dc93eeb51e99deb.jpg", + "text": "$$\n\\ell _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) } { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) + \\sum _ { \\boldsymbol { x } ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\{ \\tilde { x } _ { y } \\} } ( 1 - \\Gamma ( \\boldsymbol { x } ^ { \\prime } , y ) ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\boldsymbol { x } ^ { \\prime } , y ) ) }\n$$", + "text_format": "latex", + "bbox": [ + 171, + 628, + 803, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use the same notation as in Equation 2. Following SimCLR, we use a non-linear projection of the representation as $z _ { \\theta }$ (Figure A.1). The agent’s policy is an affine function of the representation. ", + "bbox": [ + 173, + 671, + 825, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total contrastive loss for $\\mathcal { M } _ { \\mathcal { X } }$ and $\\mathcal { M } _ { \\mathcal { Y } }$ utilizes the optimal trajectories $\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }$ and $\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }$ , where $x _ { t + 1 } \\sim P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x _ { t } )$ and $y _ { t + 1 } \\sim P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y _ { t } )$ . We set $\\mathcal { X } ^ { \\prime } = \\tau _ { \\mathcal { X } } ^ { * }$ and define ", + "bbox": [ + 173, + 705, + 825, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/9d03779755fcd19362a56a25c5ae58750a0c2ce96aa98c5ed4bdc23a3e1cce64.jpg", + "text": "$$\nL _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) = \\mathbb { E } _ { y \\sim \\tau _ { \\mathcal { Y } } ^ { * } } \\left[ \\ell _ { \\theta } ( \\tilde { x } _ { y } , y ; \\tau _ { \\mathcal { X } } ^ { * } ) \\right] \\qquad \\mathrm { w h e r e } \\ \\tilde { x } _ { y } = \\underset { x \\in \\tau _ { \\mathcal { X } } ^ { * } } { \\mathrm { a r g m a x } } \\ \\Gamma ( x , y ) .\n$$", + "text_format": "latex", + "bbox": [ + 248, + 743, + 748, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We refer to CMEs learned with policy similarity metric as policy similarity embeddings $( P S E s )$ . PSEs can trivially be combined with data augmentation by using augmented states when computing state embeddings. We simultaneously learn PSEs with the RL agent by adding ${ \\mathcal { L } } _ { \\mathrm { C M E } }$ (Algorithm 1) as an auxiliary objective during training. Next, we illustrate the benefits from this auxiliary objective. ", + "bbox": [ + 173, + 773, + 825, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 JUMPING TASK FROM PIXELS: A CASE STUDY ", + "text_level": 1, + "bbox": [ + 174, + 843, + 593, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task Description. The jumping task (Tachet des Combes et al., 2018) (Figure 1) captures, using well-defined factors of variations, whether agents can learn the correct invariances required for generalization, directly from image inputs. The task consists of an agent trying to jump over an obstacle. The agent has access to two actions: right and jump. The agent needs to time the jump precisely, at a specific distance from the obstacle, otherwise it will eventually hit the obstacle. Different tasks consist in shifting the floor height and/or the obstacle position. To generalize, the agent needs to be invariant to the floor height while jump based on the obstacle position. The obstacle can be in 26 different locations while the floor has 11 different heights, totaling 286 tasks. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe5d30be2f8515ffad0c843ba9528ed9477ed5a3bbb49088ef023877b2fee29f.jpg", + "image_caption": [ + "Figure 2: Jumping Task: Visualization of average performance of PSEs with data augmentation across different configurations. We plot the median performance across 100 runs. Each tile in the grid represents a different task (obstacle position/floor height combination). For each grid configuration, the height varies along the $_ y$ -axis (11 heights) while the obstacle position varies along the $x -$ -axis (26 locations). The red letter $\\top$ indicates the training tasks. Random grid depicts only one instance, each run consisted of a different test/train split. Beige tiles are tasks PSEs solved while black tiles are tasks PSEs did not solve when used with data augmentation. These results were chosen across all the 100 runs to demonstrate what the average reported performance looks like. " + ], + "image_footnote": [], + "bbox": [ + 181, + 147, + 808, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 486, + 825, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Setup. We split the problem into 18 seen (training) and 268 unseen (test) tasks to stress test generalization using a few changes in the underlying factors of variations seen during training. The small number of positive examples3 results in a highly unbalanced classification problem with low amounts of data, making it challenging without additional inductive biases. Thus, we evaluate generalization in regimes with and without data augmentation. The different grids configurations (Figure 2) capture different types of generalization: the “wide” grid tests generalization via “interpolation”, the “narrow” grid tests out-of-distribution generalization via “extrapolation”, and the random grid instances evaluate generalization similar to supervised learning where train and test samples are drawn i.i.d. from the same distribution. ", + "bbox": [ + 173, + 549, + 825, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used RandConv (Lee et al., 2020a), a state-of-the-art data augmentation for generalization. For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the “wide” grid (Figure 2a) and pick the parameters with the best validation performance. We use these fixed parameters for all grid configurations to show the robustness of PSEs to hyperparameter tuning. We first compute the optimal trajectories in the training tasks. Using these trajectories, we compute PSM using dynamic programming (Section D.1). We train the agent by imitation learning, combined with an auxiliary loss for PSEs (Section 4). More details are in Section G. ", + "bbox": [ + 173, + 681, + 825, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 EVALUATING GENERALIZATION ON JUMPING TASK ", + "text_level": 1, + "bbox": [ + 173, + 784, + 570, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show the efficacy of PSEs compared to common generalization approaches such as regularization (e.g., Cobbe et al., 2019; Farebrother et al., 2018), and data augmentation (e.g., Lee et al., 2020a; Laskin et al., 2020a), which are quite effective on pixel-based RL tasks. We also contrast PSEs with bisimulation transfer (Castro & Precup, 2010), a tabular state-based transfer approach based on bisimulation metrics which does not do any learning and bisimulation preserving representations (Zhang et al., 2021), showing the advantages of PSM over a prevalent state similarity metric. ", + "bbox": [ + 174, + 804, + 825, + 887 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6cb8503f0d51f326842d71126a63d55028d7bdc17f24dcab3810ef2b38ded2e.jpg", + "image_caption": [ + "Figure 3: Embedding visualization. (a) Optimal trajectories on original jumping task (visualized as coloured blocks) with different obstacle positions. We visualize the hidden representations using UMAP, where the color of points indicate the tasks of the corresponding observations. Points with the same number label correspond to same distance of the agent from the obstacle, the underlying optimal invariant feature across tasks. " + ], + "image_footnote": [], + "bbox": [ + 176, + 103, + 820, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first investigated how well PSEs generalize over existing methods without incorporating additional domain knowledge during training. Table 1 summarizes, in the setting without data augmentation, the performance of these methods in different train/test splits $\\cdot \\cdot f .$ Figure 2 for a detailed description). PSEs, with only 18 examples, already leads to better performance than standard regularization. ", + "bbox": [ + 174, + 338, + 826, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PSEs also outperform bisimulation transfer in the “wide” and random grids. Although bisimulation transfer is impractical4 when evaluating zero-shot generalization, we still performed this comparison, unfair to PSEs, to highlight their efficacy. PSEs perform better because, in contrast to bisimulation, PSM is reward agnostic ( $( c . f .$ Proposition D.1) – the expected return of the jump action is quite different depending on the obstacle position $( c . f .$ Figure G.2 for a visual juxtaposition of PSM and bisimulation). Overall, these results are promising because they place PSEs as an effective generalization method that does not rely on data augmentation. ", + "bbox": [ + 173, + 400, + 825, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nevertheless, PSEs are complementary to data augmentation, which consistently improves generalization in deep RL. We compared RandConv combined with PSEs to simply using RandConv. Domain-specific augmentation also succeeds in the jumping task. Thus, it is not surprising that RandConv is so effective compared to techniques without augmentation. Table 1 ( $2 ^ { n d }$ row) shows that PSEs substantially improve the performance of RandConv across all grid configurations. Moreover, Table 1 ( $2 ^ { n d }$ row) illustrates that when combined with RandConv, bisimulation preserving representations (Zhang et al., 2021) diminish generalization by $3 0 - 5 0 \\%$ relative to PSEs. ", + "bbox": [ + 173, + 505, + 825, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, Table 1 ( $\\cdot ^ { 1 ^ { s t } }$ row) indicates that learning-based methods are ineffective on the “narrow” grid without data augmentation. That said, PSEs do work quite well when combined with RandConv. However, even with data augmentation, generalization in “narrow” grid happens only around the vicinity of training tasks, exhibiting the challenge this grid poses for learning-based methods. We believe this is due to the poor extrapolation ability of neural networks (e.g., Haley & Soloway, 1992; Xu et al., 2020), which is more perceptible without prior inductive bias from data augmentation. ", + "bbox": [ + 174, + 609, + 825, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2 UNDERSTANDING GAINS FROM PSES: ABLATIONS AND VISUALIZATIONS ", + "bbox": [ + 176, + 699, + 722, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PSEs are contrastive metric embeddings (CMEs) learned with PSM. We investigate the gains from CMEs (Section 4) and PSM (Section 3) by ablating them. CMEs can be learned with any state similarity metric – we use $\\pi ^ { * }$ - bisimulation (Castro, 2020) as an alternative. Similarly, PSM can be used with any metric embedding – we use $\\ell _ { 2 }$ -embeddings (Section E) as an alternative, which Zhang et al. ", + "bbox": [ + 174, + 720, + 415, + 887 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Ablating PSEs. Percentage $( \\% )$ of test tasks solved when we ablate the similarity metric and learning procedure for metric embeddings in the data augmentation setting on wide grid. PSEs, which combine CMEs with PSM, considerably outperform other embeddings. We report the average performance across 100 runs with standard deviation between parentheses. All ablations except PSEs deteriorate performance compared to just using data augmentation (RandConv), as reported in Table 1. ", + "bbox": [ + 426, + 719, + 823, + 821 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e8f288563fbbcb7b1a69e06fdaeb32e47b99dd2f34b1560b3ad93961cec8a3f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metric /Embeddingl2-embeddingsCMEs
π*-bisimulation41.4 (17.6)23.1 (7.6)
PSM17.5 (8.4)87.0 (10.1)
", + "bbox": [ + 457, + 828, + 795, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(2021) employed with $\\pi ^ { * }$ -bisimulation for learning representations in a single task RL setting. For a fair comparison, we tune hyperparameters for 128 trials for each ablation entry in Table 2. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 show that PSEs $( = \\mathrm { P S M } + \\mathrm { C M E s } )$ generalize significantly better than $\\pi ^ { * }$ -bisimulation with CMEs or $\\ell _ { 2 }$ -embeddings, both of which significantly degrade performance ( $- 6 0 \\%$ and $- 4 5 \\%$ , respectively). This is expected, as $\\pi ^ { * }$ -bisimulation imposes the incorrect invariances for the jumping task $\\cdot { c . f . }$ Figures G.2a and G.2d). Additionally, looking at the rows of Table 2, CMEs are superior than $\\ell _ { 2 }$ -embeddings for PSM while inferior for $\\pi ^ { * }$ -bisimulation. This result is in line with the hypothesis that CMEs better enforce the invariances encoded by a similarity metric compared to $\\ell _ { 2 }$ -embeddings $\\left( c . f . \\right.$ Figures 3b and 3c). ", + "bbox": [ + 174, + 138, + 825, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualizing learned representations. We visualize the metric embeddings in the ablation above by projecting them to two dimensions with UMAP (McInnes et al., 2018), a popular visualization technique for high dimensional data which better preserves the data’s global structure compared to other methods such as t-SNE (Coenen & Pearce, 2019). ", + "bbox": [ + 174, + 243, + 825, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 3 shows that PSEs partition the states into two sets: (1) states where a single suboptimal action leads to failure (all states before jump) and (2) states where actions do not affect the final outcome (states after jump). Additionally, PSEs align the labeled states in the first set, which have a PSM distance of zero. These aligned states have the same distance from the obstacle, the invariant feature that generalizes across tasks. On the other hand, $\\ell _ { 2 }$ -embeddings (Zhang et al., 2021) with PSM do not align states with zero PSM except the state with the jump action – poor generalization, as observed empirically, is likely due to states with the same optimal behavior ending up with distant embeddings. CMEs with $\\pi ^ { * }$ -bisimulation align states with $\\pi ^ { * }$ -bisimulation distance of zero – such states are equidistant from the start state and have different optimal behavior for any pair of tasks with different obstacle positions (Figure G.2c). ", + "bbox": [ + 173, + 306, + 825, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 JUMPING TASK WITH COLORS: WHERE TASK-DEPENDENT INVARIANCE MATTERS ", + "bbox": [ + 171, + 448, + 784, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PSEs capture invariances that are usually orthogonal to task-agnostic invariances from data augmentation. This difference is important because, for certain tasks, data augmentation can erroneously alias states with different optimal behavior. Domain knowledge is often required to select appropriate augmentations, otherwise augmentations can hurt generalization. In contrast, PSEs do not require any domain knowledge but instead exploit the inherent structure of the RL tasks. ", + "bbox": [ + 174, + 468, + 529, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the difference between PSEs and data augmentation, we simply include colored obstacles in the jumping task (see Figure G.5). In this modified task, the optimal behavior of the agent depends on the obstacle color: the agent needs to jump over the red obstacle but strike the green obstacle to get a high return. The red obstacle task has the same difficulty as the original jumping task while the green obstacle task is easier. We jointly train the agent with 18 training tasks each, for both obstacle colors, and evaluate generalization on unseen red tasks. ", + "bbox": [ + 174, + 601, + 529, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd27daffdbeb11b7f10511fabc09304681a9967704b663bae3974721a6096f03.jpg", + "image_caption": [ + "Figure 4: Percentage $( \\% )$ of red obstacle test tasks solved when trained, jointly with red and green obstacles, on the “wide” grid. We report the mean across 100 runs. Error bars show $9 9 \\%$ confidence interval for the mean. " + ], + "image_footnote": [], + "bbox": [ + 555, + 469, + 812, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 698, + 825, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows the large performance gap between PSEs and data augmentation with RandConv. All methods solve the green obstacle tasks (Table G.1). As opposed to the original jumping task ( $\\cdot f .$ Table 1), data augmentation inhibits generalization since RandConv forces the agent to ignore color, conflating the red and green tasks (Figure G.6). PSEs still outperform regularization and data augmentation. Furthermore, data augmentation performs better when combined with PSEs. Thus, PSEs are effective even when data augmentation hurts performance. ", + "bbox": [ + 174, + 733, + 826, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4 EFFECT OF POLICY SUBOPTIMALITY ON PSES ", + "text_level": 1, + "bbox": [ + 174, + 819, + 537, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To understand the sensitivity of learning effective PSEs to the quality of the policies, we compute PSEs using $\\epsilon$ -suboptimal policies on the jumping task, which take the optimal action with probability $1 - \\epsilon$ and the subopotimal action with probability $\\epsilon$ . We evaluate the generalization performance of PSEs for increasingly suboptimal policies, ranging from the optimal policy $( \\epsilon = 0$ ) to the uniform random policy $\\epsilon = 0 . 5$ ). To isolate the effect of suboptimality on PSEs, the agent still imitates the optimal actions during training for all $\\epsilon$ . ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 shows that PSEs show near-optimal generalization with $\\epsilon \\leq 0 . 4$ while degrade generalization with an uniform random policy. This result is well-aligned with Proposition D.3, which shows that for policies with decreasing suboptimality, PSM approximation becomes more accurate, resulting in improved PSEs. Overall, this study confirms that the utility of PSEs for generalization is robust to suboptimality. One reason for this robustness is that PSEs are likely to align states with similar long-term greedy optimal actions, resulting in good performance even with suboptimal policies that preserve these greedy actions. ", + "bbox": [ + 173, + 103, + 602, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 ADDITIONAL EMPIRICAL EVALUATION ", + "text_level": 1, + "bbox": [ + 174, + 250, + 531, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we exhibit that PSM ignores spurious information ", + "bbox": [ + 178, + 273, + 602, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc90cd32460e2a251f32af744d09a75a8fd7b373561f554a97a64367ed0acb68.jpg", + "image_caption": [ + "Figure 5: $\\%$ of test tasks solved using PSEs computed using $\\epsilon \\cdot$ suboptimal policies on the “wide” grid. We report the mean across 100 runs. Error bars show one standard deviation. " + ], + "image_footnote": [], + "bbox": [ + 617, + 102, + 812, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for generalization using a LQR task (Song et al., 2019) with non-image inputs. Then, we demonstrate the scalability of PSEs without explicit access to optimal policies in an RL setting, with continuous actions, using Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b). ", + "bbox": [ + 173, + 287, + 825, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 LQR WITH SPURIOUS CORRELATIONS ", + "text_level": 1, + "bbox": [ + 174, + 337, + 480, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show how representations learned using PSM, when faced with semantically equivalent environments, can learn the main factors of variation and ignore spurious correlations that hinder generalization. We use LQR with distractors (Song et al., 2019; Sonar et al., 2020) to assess generalization in a featurebased RL setting with linear function approximation. The distractors are input features that are spuriously correlated with optimal actions and can be used for predicting these actions during training, but hurt generalization. The agent learns a linear policy using 2 environments with fixed distractors. This policy is evaluated on environments with unseen distractors. ", + "bbox": [ + 174, + 354, + 524, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We aggregate state pairs in training environments with near-zero PSM. We contrast this approach with (i) IPO (Sonar et al., 2020), a policy optimization method based on IRM (Arjovsky et al., 2019) for outof-distribution generalization, (ii) overparametrization, which leads to better generalization via implicit regularization (Song et al., 2019), and (iii) weight sparsity using $\\ell _ { 1 }$ -regularization since the policy weights which generalize in this task are sparse. ", + "bbox": [ + 173, + 542, + 522, + 611 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6faf7cc10acce48febb5bb654d3e6e540adb0aa2e7913c37cb528653ffb5a9f1.jpg", + "image_caption": [ + "Figure 6: LQR generalization: Absolute test error in LQR cost relative to the oracle (which has access to true state), of various methods trained with $n _ { d }$ distractors on 2 training environments. Lower error is better. PSM achieves near-optimal performance. We report mean across 100 different seeds. Error bars (non-existent for most methods) show $9 9 \\%$ confidence interval for mean. " + ], + "image_footnote": [], + "bbox": [ + 545, + 351, + 802, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 612, + 821, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All methods optimally solve the training environments; however, the baselines perform abysmally in terms of generalization compared to state aggregation with PSM (Figure 6), indicating their reliance on distractors. PSM obtains near-optimal generalization which we corroborate through this conjecture (Section H.1): Assuming zero state aggregation error with PSM, the policy learned using gradient descent is independent of the distractors. Refer to Section H for a detailed discussion. ", + "bbox": [ + 174, + 646, + 825, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2 DISTRACTING DM CONTROL SUITE ", + "text_level": 1, + "bbox": [ + 176, + 723, + 464, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we demonstrate scalability of PSEs on the Distracting DM Control Suite (DCS) (Stone et al., 2021), which tests whether agents can ignore high-dimensional visual distractors irrelevant to the RL task. Since we do not have access to optimal training policies, we use learned policies as proxy for $\\pi ^ { * }$ for computing PSM as well as collecting data for optimizing PSEs. Even with this approximation, PSEs outperform stateof-the-art data augmentation. DCS extends DM Control (Tassa et al., 2020) with visual distractions. We use the dynamic background distractions (Stone et al., 2021; Zhang et al., 2018b) where a video is played in the background from a specific frame. The video and the frame are randomly sampled every new episode. We use 2 videos during training (Figure 7) and evaluate generalization on 30 unseen videos (Figure I.1). ", + "bbox": [ + 174, + 743, + 581, + 897 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff872ad14af14eeee7e2d8c38cd29313df3392bf8308eb9fbdc234e87dd58a41.jpg", + "image_caption": [ + "Figure 7: Distracting Control Suite: Snapshots of training environments. " + ], + "image_footnote": [], + "bbox": [ + 594, + 743, + 823, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 897, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d3a20085648824a4fd02c4eee0f37a927f6224b136e454d7b0d14b5a1bf5010.jpg", + "table_caption": [ + "Table 3: Generalization performance with unseen distractions in the Distracting Suite at 500K steps. We report the average scores across 5 seeds $\\pm$ standard error. All methods are added to SAC (Haarnoja et al., 2018). Pretrained initialization uses $\\mathrm { D r Q }$ trained for 500K steps. Figures I.2 and I.3 show generalization curves. " + ], + "table_footnote": [], + "table_body": "
InitializationMethodBiC-catchC-swingupC-runF-spinR-easyW-walk
RandomDrQ747±28582±42220±12646±54931±14549±83
DrQ + PSEs821±17749±19308±12779±49955±10789±28
PretrainedDrQ748±30689±22219±10764±48943±10709±29
DrQ +PSEs805±25753±13282±8803±19962±11829±21
", + "bbox": [ + 179, + 147, + 820, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "All agents are built on top of SAC (Haarnoja et al., 2018) combined with $\\mathrm { D r Q }$ (Kostrikov et al., 2020), an augmentation method with state-of-the-art performance on DM control. Notably, DrQ outperforms CURL (Laskin et al., 2020b), a strong contrastive method for RL. Without data augmentation on DM control, SAC performs poorly, even during training (Kostrikov et al., 2020). We augment $\\mathrm { D r Q }$ with an auxiliary loss for learning PSEs and compare it with DrQ (Table 3). Orthogonal to DrQ, PSEs align representations of different states across environments based on PSM $( c . f .$ Figure A.1). All agents are trained for 500K environment steps with random crop augmentation. For computing PSM, we use policies learned by $\\mathrm { D r Q }$ pretrained on training environments for 500K steps. ", + "bbox": [ + 173, + 248, + 825, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, we investigate how much better PSEs generalize relative to DrQ, assuming the agent is provided with PSM beforehand. The agent’s policy is randomly initialized so that additional gains over DrQ can be attributed to the auxiliary information from PSM. The substantial gains in Table 3 indicate that PSEs are more effective than $_ \\mathrm { D r Q }$ for encoding invariance to distractors. ", + "bbox": [ + 174, + 367, + 825, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since PSEs utilize PSM approximated using pretrained policies, we also compare to a $_ \\mathrm { D r Q }$ agent where we initialize it using these pretrained policies. This comparison provides the same auxiliary information to DrQ as available to PSEs, thus, the generalization difference stems from how they utilize this information. Table 3 demonstrates that PSEs outperform DrQ with pretrained initialization, indicating that the additional pretraining steps are more judiciously utilized for computing PSM as opposed to just longer training with DrQ. More details, including learning curves, are in Section I. ", + "bbox": [ + 174, + 429, + 825, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 522, + 346, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PSM (Section 3) is inspired by bisimulation metrics (Section C). However, different than traditional bisimulation (e.g., Larsen & Skou, 1991; Givan et al., 2003; Ferns et al., 2011), PSM is more tractable as it defined with respect to a single policy similar to the recently proposed $\\pi ^ { * }$ -bisimulation (Castro, 2020; Zhang et al., 2021). However, in contrast to PSM, bisimulation metrics rely on reward information and may not provide a meaningful notion of behavioral similarity in certain environments (Section 5). For example, states similar under PSM would have similar optimal policies, yet can have arbitrarily large $\\pi ^ { * }$ -bisimulation distance between them (Proposition D.1). ", + "bbox": [ + 174, + 545, + 826, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PSEs (Section 4) use contrastive learning to encode behavior similarity (Section 3) across MDPs. Previously, contrastive learning has been applied for imposing state self-consistency (Laskin et al., 2020b), capturing predictive information (Oord et al., 2018; Mazoure et al., 2020; Lee et al., 2020b) or encoding transition dynamics (van der Pol et al., 2020; Stooke et al., 2020; Schwarzer et al., 2020) within an MDP. These methods can be integrated with PSEs to encode additional invariances. Interestingly, in a similar spirit to PSEs, Pacchiano et al. (2020); Moskovitz et al. (2021) explore comparing behavioral similarity between policies to guide policy optimization within an MDP. ", + "bbox": [ + 173, + 648, + 826, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PSEs are complementary to data augmentation methods (Kostrikov et al., 2020; Lee et al., 2020a; Raileanu et al., 2020; Ye et al., 2020), which have recently been shown to significantly improve agents’ generalization capabilities. In fact, we combine PSEs to state-of-the-art augmentation methods including random convolutions (Lee et al., 2020a; Laskin et al., 2020a) in the jumping task and DrQ (Kostrikov et al., 2020) on Distracting Control Suite, leading to performance improvement. ", + "bbox": [ + 174, + 753, + 825, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 832, + 320, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper advances generalization in RL by two contributions: (1) the policy similarity metric (PSM) which provides a new notion of state similarity based on behavior proximity, and (2) contrastive metric embeddings, which harness the benefits of contrastive learning for representations based on a similarity metric. PSEs combine these two ideas to improve generalization. Overall, this paper shows the benefits of exploiting the inherent structure in RL for learning effective representations. ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 103, + 285, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rishabh Agarwal, Chen Liang, Dale Schuurmans, and Mohammad Norouzi. Learning to generalize from sparse and underspecified rewards. In ICML, 2019. ", + "bbox": [ + 174, + 125, + 823, + 151 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Martín Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. CoRR, abs/1907.02893, 2019. ", + "bbox": [ + 173, + 160, + 821, + 186 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In Advances in Neural Information Processing Systems, pp. 7413–7424, 2019. ", + "bbox": [ + 176, + 195, + 821, + 223 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta-regularization. In Advances in Neural Information Processing Systems, pp. 998–1008, 2018. ", + "bbox": [ + 176, + 231, + 823, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic Markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. ", + "bbox": [ + 176, + 266, + 821, + 294 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pablo Samuel Castro and Doina Precup. Using bisimulation for policy transfer in MDPs. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2010. ", + "bbox": [ + 176, + 301, + 823, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. CoRR, abs/2002.05709, 2020. ", + "bbox": [ + 176, + 337, + 823, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Karl Cobbe, Oleg Klimov, Christopher Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2019. ", + "bbox": [ + 174, + 372, + 823, + 400 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andy Coenen and Adam Pearce. Understanding umap, 2019. URL https://pair-code.github.io/ understanding-umap/. ", + "bbox": [ + 174, + 407, + 823, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jesse Farebrother, Marlos C. Machado, and Michael Bowling. Generalization and regularization in DQN. In NeurIPS Deep Reinforcement Learning Workshop, 2018. ", + "bbox": [ + 174, + 443, + 823, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Norm Ferns, Prakash Panangaden, and Doina Precup. Metrics for finite Markov decision processes. In Proceedings of the Conference in Uncertainty in Artificial Intelligence (UAI), 2004. ", + "bbox": [ + 173, + 479, + 823, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Norm Ferns, Pablo Samuel Castro, Doina Precup, and Prakash Panangaden. Methods for computing state similarity in Markov decision processes. In Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence, UAI ’06. AUAI Press, 2006. ", + "bbox": [ + 176, + 515, + 823, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Norm Ferns, Prakash Panangaden, and Doina Precup. Bisimulation metrics for continuous markov decision processes. SIAM Journal on Computing, 40(6):1662–1714, 2011. ", + "bbox": [ + 176, + 563, + 823, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Norman Ferns and Doina Precup. Bisimulation metrics are optimal value functions. In The 30th Conference on Uncertainty in Artificial Intelligence, pp. 10, 2014. ", + "bbox": [ + 173, + 598, + 821, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International Conference on Machine Learning (ICML), 2017. ", + "bbox": [ + 174, + 633, + 823, + 661 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Robert Givan, Thomas Dean, and Matthew Greig. Equivalence notions and model minimization in markov decision processes. Artificial Intelligence, 147(1-2):163–223, 2003. ", + "bbox": [ + 173, + 669, + 823, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and D Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pp. 1487–1495, 2017. ", + "bbox": [ + 176, + 704, + 821, + 744 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit regularization in matrix factorization. In Advances in Neural Information Processing Systems, pp. 6151–6159, 2017. ", + "bbox": [ + 176, + 752, + 825, + 791 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. arXiv preprint arXiv:1801.01290, 2018. ", + "bbox": [ + 173, + 800, + 823, + 828 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2, pp. 1735–1742. IEEE, 2006. ", + "bbox": [ + 173, + 837, + 823, + 876 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pamela J Haley and DONALD Soloway. Extrapolation limitations of multilayer feedforward neural networks. In [Proceedings 1992] IJCNN International Joint Conference on Neural Networks, volume 4, pp. 25–30. IEEE, 1992. ", + "bbox": [ + 174, + 885, + 826, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Maximilian Igl, Kamil Ciosek, Yingzhen Li, Sebastian Tschiatschek, Cheng Zhang, Sam Devlin, and Katja Hofmann. Generalization in reinforcement learning with selective noise injection and information bottleneck. In Advances in Neural Information Processing Systems, pp. 13978–13990, 2019. ", + "bbox": [ + 176, + 104, + 823, + 143 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Domain extrapolation via regret minimization. arXiv preprint arXiv:2006.03908, 2020. ", + "bbox": [ + 171, + 151, + 823, + 179 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Arthur Juliani, Ahmed Khalifa, Vincent-Pierre Berges, Jonathan Harper, Ervin Teng, Hunter Henry, Adam Crespi, Julian Togelius, and Danny Lange. Obstacle Tower: A generalization challenge in vision, control, and planning. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2019. ", + "bbox": [ + 176, + 188, + 823, + 228 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Niels Justesen, Ruben Rodriguez Torrado, Philip Bontrager, Ahmed Khalifa, Julian Togelius, and Sebastian Risi. Illuminating generalization in deep reinforcement learning through procedural level generation. arXiv preprint arXiv:1806.10729, 2018. ", + "bbox": [ + 173, + 237, + 823, + 276 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Taylor W. Killian, George Dimitri Konidaris, and Finale Doshi-Velez. Robust and efficient transfer learning with hidden parameter Markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pp. 4949–4950, 2017. ", + "bbox": [ + 173, + 285, + 825, + 324 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. CoRR, abs/2004.13649, 2020. ", + "bbox": [ + 173, + 333, + 821, + 361 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kim G Larsen and Arne Skou. Bisimulation through probabilistic testing. Information and computation, 94(1): 1–28, 1991. ", + "bbox": [ + 174, + 368, + 823, + 396 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Michael Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. CoRR, abs/2004.14990, 2020a. ", + "bbox": [ + 173, + 405, + 823, + 431 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. Proceedings of the 37th International Conference on Machine Learning, Vienna, Austria, PMLR 119, 2020b. arXiv:2003.06417. ", + "bbox": [ + 174, + 440, + 823, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for generalization in deep reinforcement learning. In The International Conference on Learning Representations (ICLR), 2020a. ", + "bbox": [ + 174, + 488, + 823, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kuang-Huei Lee, Ian Fischer, Anthony Liu, Yijie Guo, Honglak Lee, John Canny, and Sergio Guadarrama. Predictive information accelerates learning in rl. arXiv preprint arXiv:2007.12401, 2020b. ", + "bbox": [ + 174, + 536, + 823, + 564 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. ", + "bbox": [ + 174, + 573, + 825, + 601 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Bogdan Mazoure, Remi Tachet des Combes, Thang Long DOAN, Philip Bachman, and R Devon Hjelm. Deep reinforcement and infomax learning. Advances in Neural Information Processing Systems, 33, 2020. ", + "bbox": [ + 173, + 608, + 823, + 636 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. ArXiv e-prints, February 2018. ", + "bbox": [ + 174, + 643, + 823, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ted Moskovitz, Michael Arbel, Ferenc Huszar, and Arthur Gretton. Efficient wasserstein natural gradients for reinforcement learning. In International Conference on Learning Representations, 2021. ", + "bbox": [ + 171, + 680, + 823, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Junhyuk Oh, Satinder P. Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2017. ", + "bbox": [ + 173, + 715, + 825, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. ", + "bbox": [ + 169, + 765, + 825, + 791 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aldo Pacchiano, Jack Parker-Holder, Yunhao Tang, Krzysztof Choromanski, Anna Choromanska, and Michael Jordan. Learning to score behaviors for guided policy optimization. In International Conference on Machine Learning, 2020. ", + "bbox": [ + 173, + 800, + 823, + 839 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Charles Packer, Katelyn Gao, Jernej Kos, Philipp Krähenbühl, Vladlen Koltun, and Dawn Song. Assessing generalization in deep reinforcement learning. arXiv preprint arXiv:1810.12282, 2018. ", + "bbox": [ + 171, + 848, + 821, + 876 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Christian F. Perez, Felipe Petroski Such, and Theofanis Karaletsos. Generalized hidden parameter mdps transferable model-based rl in a handful of trials. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. ", + "bbox": [ + 173, + 885, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. \nMartin L Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley & Sons, Inc., 1994. \nRoberta Raileanu, Max Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in deep reinforcement learning. arXiv preprint arXiv:2006.12862, 2020. \nAravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards generalization and simplicity in continuous control. In Advances in Neural Information Processing Systems (NeurIPS), 2017. \nBenjamin Recht. A tour of reinforcement learning: The view from continuous control. Annual Review of Control, Robotics, and Autonomous Systems, 2019. \nMax Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Dataefficient reinforcement learning with momentum predictive representations. arXiv preprint arXiv:2007.05929, 2020. \nAnoopkumar Sonar, Vincent Pacelli, and Anirudha Majumdar. Invariant policy optimization: Towards stronger generalization in reinforcement learning. arXiv preprint arXiv:2006.01096, 2020. \nXingyou Song, Yiding Jiang, Yilun Du, and Behnam Neyshabur. Observational overfitting in reinforcement learning. In The International Conference on Learning Representations (ICLR), 2019. \nAustin Stone, Oscar Ramirez, Kurt Konolige, and Rico Jonschkowski. The distracting control suite – a challenging benchmark for reinforcement learning from pixels. arXiv preprint arXiv:2101.02722, 2021. \nAdam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from reinforcement learning. arXiv preprint arXiv:2009.08319, 2020. \nRemi Tachet des Combes, Philip Bachman, and Harm van Seijen. Learning invariances for policy generalization. In Workshop track at the International Conference on Learning Representations (ICLR), 2018. \nYujin Tang, Duong Nguyen, and David Ha. Neuroevolution of self-interpretable agents. arXiv preprint arXiv:2003.08165, 2020. \nYuval Tassa, Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom Erez, Timothy Lillicrap, and Nicolas Heess. dm_control: Software and tasks for continuous control. arXiv preprint arXiv:2006.12983, 2020. \nMatthew E. Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research, 10:1633–1685, 2009. \nJosh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 23–30. IEEE, 2017. \nElise van der Pol, Thomas Kipf, Frans A Oliehoek, and Max Welling. Plannable approximations to mdp homomorphisms: Equivariance under actions. In Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems, pp. 1431–1439, 2020. \nCédric Villani. Optimal transport: old and new. Springer, 2008. \nSam Witty, Jun Ki Lee, Emma Tosch, Akanksha Atrey, Michael Littman, and David Jensen. Measuring and characterizing generalization in deep reinforcement learning. In NeurIPS Critiquing and Correcting Trends in Machine Learning Workshop, 2018. \nKeyulu Xu, Jingling Li, Mozhi Zhang, Simon S Du, Ken-ichi Kawarabayashi, and Stefanie Jegelka. How neural networks extrapolate: From feedforward to graph neural networks. arXiv preprint arXiv:2009.11848, 2020. \nChang Ye, Ahmed Khalifa, Philip Bontrager, and Julian Togelius. Rotation, translation, and cropping for zero-shot generalization. CoRR, abs/2001.09908, 2020. \nAmy Zhang, Nicolas Ballas, and Joelle Pineau. A Dissection of Overfitting and Generalization in Continuous Reinforcement Learning. CoRR, abs/1806.07937, 2018a. \nAmy Zhang, Yuxin Wu, and Joelle Pineau. Natural environment benchmarks for reinforcement learning. arXiv preprint arXiv:1811.06032, 2018b. ", + "bbox": [ + 171, + 65, + 828, + 926 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Amy Zhang, Clare Lyle, Shagun Sodhani, Angelos Filos, Marta Kwiatkowska, Joelle Pineau, Yarin Gal, and Doina Precup. Invariant causal prediction for block mdps. In Proceedings of the International Conference on Machine Learning (ICML), 2020. ", + "bbox": [ + 174, + 104, + 823, + 142 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. The International Conference on Learning Representations (ICLR), 2021. ", + "bbox": [ + 173, + 151, + 823, + 190 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chiyuan Zhang, Oriol Vinyals, Rémi Munos, and Samy Bengio. A study on overfitting in deep reinforcement learning. CoRR, abs/1804.06893, 2018c. ", + "bbox": [ + 173, + 199, + 825, + 226 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Appendix ", + "text_level": 1, + "bbox": [ + 174, + 99, + 315, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A LEARNING CONTRASTIVE METRIC EMBEDDINGS ", + "text_level": 1, + "bbox": [ + 173, + 148, + 620, + 166 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d19f18e3e5447a5a82c4130e02a0d8375292bd0ba4e9e7a07050c01ca385d18a.jpg", + "image_caption": [ + "Figure A.1: Architecture for learning CMEs. Given an input pair $( x , y )$ , we first apply the (optional) data augmentation operator $\\Psi$ to produce the input augmentations $\\bar { \\Psi } _ { x } : = \\Psi ( x ) , \\Psi _ { y } : = \\bar { \\Psi } ( y )$ . When not using data augmentation, $\\Psi$ is equal to the identity operator, that is, ∀ $x \\Psi ( x ) = x$ . The agent’s policy network then outputs the representations for these augmentations by applying the encoder $f _ { \\theta }$ , that is, ${ \\bar { f } } _ { x } = f _ { \\theta } ( \\Psi _ { x } )$ , $f _ { y } =$ $f _ { \\theta } ( \\hat { \\Psi } _ { y } )$ . These representations are projected using a non-linear projector $h _ { \\theta }$ to obtain the embedding $z _ { \\theta }$ , that is, $z _ { \\theta } ( x ) = h _ { \\theta } ( f _ { x } )$ , $z _ { \\theta } ( y ) = h _ { \\theta } ( f _ { y } )$ . These metric embeddings are trained using the contrastive loss defined in Equation (4). The policy $\\pi \\theta$ is an affine function of the representation, that is, $\\pi _ { \\boldsymbol { \\theta } } ( \\cdot | \\boldsymbol { y } ) = \\boldsymbol { W } ^ { T } f _ { \\boldsymbol { y } } + \\boldsymbol { b }$ , where $W , b$ are learned weights and biases. The entire network is trained end-to-end jointly using the reinforcement learning (or imitation learning) loss in conjunction with the auxiliary contrastive loss. " + ], + "image_footnote": [], + "bbox": [ + 176, + 196, + 812, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B PROOFS ", + "text_level": 1, + "bbox": [ + 174, + 556, + 276, + 573 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We begin by defining some notation which will be used throughout these results: ", + "bbox": [ + 173, + 589, + 700, + 606 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "• We denote $\\begin{array} { r } { { ! } \\mathbb { E } _ { t \\ge 0 } [ \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) ] = \\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) \\right] } \\end{array}$ • For any $y \\in Y$ , let $Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )$ , where $Y _ { y } ^ { 0 } = y$ . \n• $T V ^ { n } ( Y _ { y } ^ { k } ) = \\mathbb { E } _ { 0 \\leq t < n } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { k + t } ) , \\pi ^ { * } ( Y _ { y } ^ { k + t } ) ) .$ ", + "bbox": [ + 217, + 616, + 761, + 693 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We now proceed with some technical lemmas necessary for the main result. ", + "bbox": [ + 173, + 702, + 669, + 717 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Lemma 1. Given any two pseudometrics5 $d , d ^ { \\prime } \\in \\mathbb { M }$ and probability distributions $P _ { X } , P _ { y }$ where $\\mathcal { X } , \\mathcal { Y } \\subset \\mathcal { S }$ , we have: ", + "bbox": [ + 173, + 722, + 826, + 751 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/178d0eb40c25e708a5e9dde06ff4afb48b7840e525fb666121160255643bd9c7.jpg", + "text": "$$\n\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )\n$$", + "text_format": "latex", + "bbox": [ + 334, + 758, + 661, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof. Note that the dual of the linear program for computing $\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )$ is given by ", + "bbox": [ + 171, + 799, + 769, + 815 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/268d7bdcbcf866dab677354dd6a9c0f34f95aed562188eff455f54455025a4a9.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\mathop { \\operatorname* { m i n } } _ { \\textsc { i r } } \\sum _ { x \\in \\mathcal { X } , \\ y \\in \\mathcal { V } } \\Gamma ( x , y ) d ( x , y ) } \\\\ { \\mathrm { s u b j e c t ~ t o ~ } \\displaystyle \\sum _ { x } \\Gamma ( x , y ) = P y ( y ) \\forall y , \\quad \\displaystyle \\sum _ { y } \\Gamma ( x , y ) = P _ { \\mathcal { X } } ( x ) \\forall x , \\quad \\Gamma ( x , y ) \\geq 0 \\forall x , y } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 214, + 824, + 782, + 896 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5Pseudometrics are generalization of metrics where the distance between two distinct states can be zero. ", + "bbox": [ + 187, + 910, + 805, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Using the dual formulation subject to the constraints above, $\\mathscr { W } ^ { 1 } ( d )$ can be written as ", + "bbox": [ + 174, + 102, + 728, + 119 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/3e0255eb3d0bd9d3ccace9e0a20f8b685338115f46f5e790bb6653e56f406bba.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| = \\mathcal { W } ^ { 1 } ( d - d ^ { \\prime } + d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| } \\\\ { \\displaystyle \\quad \\quad \\quad = \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( d ( x , y ) - d ^ { \\prime } ( x , y ) + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad \\leq \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( \\| d - d ^ { \\prime } \\| + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) d ^ { \\prime } ( x , y ) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 217, + 122, + 782, + 271 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Lemma 2. Given any $y _ { 0 } \\in Y$ , we have: ", + "bbox": [ + 173, + 294, + 439, + 309 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/68534635c2d9afb6aac4e8a0c836e0007df72fd4318165876c6f465a2993771a.jpg", + "text": "$$\n\\sum _ { y _ { 1 } \\in Y } \\left( P ^ { \\tilde { \\pi } } ( y _ { 1 } | y _ { 0 } ) - P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\right) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 2 } { 1 - \\gamma } T V ( \\tilde { \\pi } ( y _ { 0 } ) , \\pi ^ { * } ( y _ { 0 } ) )\n$$", + "text_format": "latex", + "bbox": [ + 254, + 313, + 741, + 351 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Proof. ", + "bbox": [ + 173, + 363, + 217, + 378 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/01dff727690cafb5108175805d5f99671b922013be8e9969d5cb591cbf014423.jpg", + "text": "$$\n\\begin{array} { r l } { \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | | B \\rangle } ) - P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) \\leq | \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) - P ^ { \\kappa ^ { \\star } } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) | } & { } \\\\ & { \\leq \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha | \\hat { \\sigma } ( a ) | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( a | \\mathfrak { g } _ { 1 } ) ) | T V ^ { n } ( Y _ { n } ^ { 0 } ) } \\\\ & { \\leq \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\sum _ { i \\in \\mathcal { K } } P ^ { \\kappa } ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) | \\overline { { \\tau } } ( \\mathfrak { g } ( | \\mathfrak { g } _ { 1 } \\rangle ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } ) | ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & = \\ \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 381, + 839, + 630 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Lemma 3. Given any $y _ { 0 } \\in Y$ , if $\\begin{array} { r } { \\mathbf { \\dot { \\mathbf { \\theta } } T } V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } ) } \\end{array}$ , we have: ", + "bbox": [ + 173, + 647, + 655, + 667 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/520af0569c33b4cfea5e3279b677c5d78e474c4192797189faf1aec9512d738d.jpg", + "text": "$$\n\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq { \\frac { 1 + \\gamma } { 1 - \\gamma } } W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | { \\tilde { x } _ { y _ { 0 } } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 264, + 671, + 733, + 710 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Proof. Note that we have the following equality, where 0 is a vector of zeros: ", + "bbox": [ + 173, + 723, + 684, + 738 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/d8c5f18a057edae5a97b8e3461c34ea7880dc52b2cb709d7f988f6651fad4b25.jpg", + "text": "$$\n\\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) = \\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) - \\sum _ { x \\in X } P ^ { \\pi ^ { * } } ( x | \\tilde { x } _ { y _ { 0 } } ) \\mathbf { 0 }\n$$", + "text_format": "latex", + "bbox": [ + 235, + 741, + 761, + 776 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "which is the same form as the primal LP for $W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )$ . By assumption, we have that ", + "bbox": [ + 173, + 780, + 823, + 808 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/7d973b469ae0e5fef9c6eb192c403af32c2c01f146bf623d455ccb1d8ccc4515.jpg", + "text": "$$\nT V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { \\ast } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } )\n$$", + "text_format": "latex", + "bbox": [ + 392, + 804, + 604, + 837 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This implies that $\\textstyle { \\frac { 1 - \\gamma } { 1 + \\gamma } } T V ^ { n } ( Y _ { . } ^ { 0 } )$ is a feasible solution to $W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )$ ", + "bbox": [ + 173, + 845, + 758, + 864 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/d7571a1589112c9b127e22f040e1034d276c161a5718ca8181542e6a29f1989c.jpg", + "text": "$$\n\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\frac { 1 - \\gamma } { 1 + \\gamma } T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 264, + 867, + 733, + 906 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "and the result follows. ", + "bbox": [ + 173, + 909, + 320, + 924 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Proposition B.1. The operator $\\mathcal { F }$ given by: ", + "bbox": [ + 174, + 103, + 464, + 118 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/50037dbaadaa611b3c2b7307b310ef9effa7d80a0cfe71252bc574a5d4e1c2c3.jpg", + "text": "$$\n\\mathcal F ( d ) ( x , y ) = \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) + \\gamma \\mathcal W _ { 1 } ( d ) \\big ( P _ { \\mathcal X } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal V } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big )\n$$", + "text_format": "latex", + "bbox": [ + 277, + 123, + 720, + 142 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "is a contraction mapping and has a unique fixed point for a bounded dist. ", + "bbox": [ + 174, + 147, + 658, + 162 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proof. We first prove that $\\mathcal { F }$ is contraction mapping. Then, a simple application of the Banach Fixed Point Theorem asserts that $\\mathcal { F }$ has a unique fixed point. Note that for all pseudometrics $d , d ^ { \\prime } \\in \\mathbb { M }$ , and for all states $x \\in \\mathcal { X }$ , $y \\in \\mathcal { V }$ , ", + "bbox": [ + 174, + 176, + 825, + 220 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/e26031ddb7f6793c820c994471ce5075a5e1d92786e164ddfeb0013f881bdb40.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { F } ( d ) ( x , y ) - \\mathcal { F } ( d ^ { \\prime } ) ( x , y ) } \\\\ & { \\quad = \\gamma \\left( \\mathcal { W } _ { 1 } ( d ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad \\overset { L e m m a ~ \\ : 1 } { \\leq } \\gamma \\left( \\| d - d ^ { \\prime } \\| + \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad = \\gamma \\| d - d ^ { \\prime } \\| } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 209, + 223, + 787, + 321 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Thus, $\\| \\mathcal F ( d ) - \\mathcal F ( d ^ { \\prime } ) \\| \\le \\gamma \\| d - d ^ { \\prime } \\|$ , so that $\\mathcal { F }$ is a contracting mapping for $\\gamma < 1$ and has an unique fixed point $d ^ { * }$ . ", + "bbox": [ + 171, + 342, + 825, + 371 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Theorem 1. [Bound on policy transfer] For any $y \\in \\mathcal { V }$ , let $Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )$ define the sequence of random states encountered starting in $Y _ { y } ^ { 0 } = y$ and following policy $\\tilde { \\pi }$ . We have: ", + "bbox": [ + 173, + 377, + 820, + 411 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/78300f5f5ad236d336ba4cea178e615dab71051f4cf522680b2ec7241d02a039.jpg", + "text": "$$\n\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) .\n$$", + "text_format": "latex", + "bbox": [ + 312, + 415, + 684, + 465 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proof. We will prove this by induction. Assuming that the bound holds for $T V ^ { n }$ , we prove the bound holds for $T V ^ { n + \\bar { 1 } }$ . The base case for $n = 1$ follows from $T V ( \\tilde { \\pi } ( y ) , \\pi ^ { * } ( y ) ) = T V ( \\pi ^ { * } ( \\tilde { x } _ { y } ) , \\pi ^ { * } ( y ) ) \\leq$ d∗(˜xy, y). Note that T V n ≤ 1−γn+11−γ since the $T V$ distance per time-step can be at most 1. ", + "bbox": [ + 173, + 477, + 826, + 529 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Let $P _ { t } ^ { \\pi } ( y ^ { \\prime } | y )$ denote the probability of ending in state $y ^ { \\prime } \\in Y$ after $t$ steps when following policy $\\pi$ and starting from state $y$ . We then have: ", + "bbox": [ + 174, + 534, + 823, + 561 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/9806b9b18cd6be1835fe38e323ac6ae7a8a021231bd7e90489f9f032b6ffd66c.jpg", + "text": "$$\n\\begin{array} { r l } & { \\gamma ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\mathcal { K } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\gamma \\varepsilon \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\Bigg [ \\varepsilon \\gamma \\varepsilon ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\varepsilon \\gamma ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\varepsilon ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) \\Bigg ] } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { \\quad + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\gamma \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & \\quad - \\sum _ { s \\in \\mathcal { N } } ^ \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 564, + 839, + 915 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Thus, by induction, it follows that for all $n$ : ", + "bbox": [ + 174, + 103, + 459, + 118 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f0e3466874ac9aa6d3f381f52d738f7138816ba8defceed58d69a2f3532d74d6.jpg", + "text": "$$\nT V ^ { n } ( Y _ { y } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) ,\n$$", + "text_format": "latex", + "bbox": [ + 398, + 121, + 596, + 154 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "which completes the proof. ", + "bbox": [ + 174, + 156, + 352, + 171 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "C BISIMULATION METRICS ", + "text_level": 1, + "bbox": [ + 176, + 190, + 415, + 207 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Notation. We use the notation as defined in Section 2. ", + "bbox": [ + 173, + 220, + 529, + 236 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Bisimulation metrics (Givan et al., 2003; Ferns et al., 2011) define a pseudometric $d _ { \\sim } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }$ where $d _ { \\sim } ( x , y )$ is defined in terms of distance between immediate rewards and next state distributions. Define $\\mathcal { F } _ { \\sim } ^ { e } : \\mathbb { M } \\mathbb { M }$ by ", + "bbox": [ + 173, + 242, + 826, + 285 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/4961b6782cf1de03ed5b5de85a3e92827a028dd378284d7a9a4535a4ce354b1b.jpg", + "text": "$$\n\\mathcal { F } _ { \\sim } ^ { e } ( d ) ( x , y ) = \\operatorname* { m a x } _ { a \\in A } \\vert R ( x , a ) - R ( y , a ) \\vert + \\gamma \\mathcal { W } _ { 1 } ( d ) \\left( P ^ { a } ( \\cdot \\vert x ) , P ^ { a } ( \\cdot \\vert y ) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 258, + 287, + 738, + 311 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "then, $\\mathcal { F } _ { \\sim } ^ { e }$ has a unique fixed point $d _ { \\sim }$ which is a bisimulation metric. $\\mathcal { F } _ { \\sim } ^ { e }$ uses the 1-Wasserstein metric $\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }$ . The 1-Wasserstein distance $\\mathscr { W } _ { 1 } ( d )$ under the pseudometric $d$ can be computed using the dual linear program: ", + "bbox": [ + 173, + 315, + 826, + 358 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f8613eabbb2d2596c7a8e041d947c55bd72e986330d6682b84bffc223ed82188.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\mathbf { u } , \\mathbf { v } } \\sum _ { x \\in \\mathcal { X } } P ( x ) u _ { x } - \\sum _ { y \\in \\mathcal { Y } } P ( y ) v _ { y } \\quad \\mathrm { s u b j e c t ~ t o } \\forall x \\in \\mathcal { X } , y \\in \\mathcal { Y } \\quad u _ { x } - v _ { y } \\leq d ( x , y )\n$$", + "text_format": "latex", + "bbox": [ + 225, + 361, + 771, + 396 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Since we are only interested in computing the coupling between the states in $\\mathcal { X }$ and $\\mathcal { V }$ , the above formulation assumes that $P _ { \\mathcal { X } } ( y ) = 0$ for all $y \\in \\mathcal { V }$ and $P _ { \\mathcal { Y } } ( x ) = 0$ for all $x \\in \\mathcal { X }$ . The computation of $d _ { \\sim }$ is expensive and requires a tabular representation of the states, rendering it impractical for large state spaces. On-policy bisimulation (Castro, 2020) (e.g., $\\pi ^ { * }$ -bisimulation) is tied to specific behavior policies and is much easier to approximate than bisimulation. ", + "bbox": [ + 173, + 398, + 825, + 469 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D POLICY SIMILARITY METRIC ", + "text_level": 1, + "bbox": [ + 176, + 488, + 455, + 505 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D.1 COMPUTING PSM ", + "text_level": 1, + "bbox": [ + 174, + 518, + 346, + 534 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "In general, PSM for a given DIST across MDPs $\\mathcal { M } _ { \\mathcal { X } }$ and $\\mathcal { M } _ { \\mathcal { Y } }$ is given by ", + "bbox": [ + 174, + 545, + 658, + 560 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/397980339d5fd9d1b862a2716debbd7eb5a154289e77e3908ae1376c757b1ffb.jpg", + "text": "$$\n\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 269, + 564, + 727, + 584 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Since our main focus is showing the utility of PSM for generalization, we simply use environments where PSM can be computed using dynamic programming. Using a similar observation to Castro (2020), we assert that the recursion for $d ^ { * }$ takes the following form in deterministic environments: ", + "bbox": [ + 173, + 593, + 825, + 636 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/191926b32509785686fc5757bafe5efd5ea9a5819cee8360dc34c4ac99d9d785.jpg", + "text": "$$\n\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma d ^ { * } ( x ^ { \\prime } , y ^ { \\prime } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 339, + 638, + 658, + 659 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where $x ^ { \\prime } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x )$ , $y ^ { \\prime } = P _ { y } ^ { \\pi ^ { * } } ( y )$ are the next states from taking actions $\\pi _ { \\mathcal { X } } ^ { \\ast } ( x ) , \\pi _ { \\mathcal { X } } ^ { \\ast } ( y )$ from states $x , y$ respectively. Furthermore, we assume that DIST between terminal states from $\\mathcal { M } _ { \\mathcal { X } }$ and $\\mathcal { M } _ { \\mathcal { Y } }$ is zero. Note that the form of Equation D.2 closely resembles the update rule in Q-learning, and as such, can be efficiently computed with samples using approximate dynamic programming. Given access to optimal trajectories $\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }$ and $\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }$ , where $x _ { t + 1 } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x _ { t } )$ and $y _ { t + 1 } = P _ { y } ^ { \\pi ^ { * } } ( y _ { t } )$ , Equation D.2 can be solved using exact dynamic programming; we provide pseudocode in Section J.1. ", + "bbox": [ + 173, + 662, + 826, + 750 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "There are other ways to approximate the Wasserstein distance in bisimulation metrics (e.g., Ferns et al., 2006; 2011; Castro, 2020; Zhang et al., 2021). That said, approximating bisimulation (or PSM) for stochastic environments remains an exciting research direction (Castro, 2020). Investigating other distance metrics for long-term behavior difference in PSM is also interesting for future work. ", + "bbox": [ + 174, + 755, + 825, + 811 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D.2 PSM CONNECTIONS TO DATA AUGMENTATION AND BISIMULATION ", + "text_level": 1, + "bbox": [ + 174, + 827, + 687, + 843 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Connection to bisimulation. Although bisimulation metrics have appealing properties such as bounding value function differences (e.g., (Ferns & Precup, 2014)), they rely on reward information and may not provide a meaningful notion of behavioral similarity in certain environments. Proposition D.1 implies that states similar under PSM would have similar optimal policies yet can have arbitrarily large bisimulation distance between them. ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/3d8fd583f31c9cbc260975c387fd75a3ea68a6f1e11f6b2895873d41fe0189c2.jpg", + "image_caption": [ + "Figure D.1: Cyan edges represent actions with a positive reward, which are also the optimal actions. Zero rewards everywhere else. $x _ { 0 } , y _ { 0 }$ are the start states while $x _ { 2 } , y _ { 2 }$ are the terminal states. " + ], + "image_footnote": [], + "bbox": [ + 228, + 98, + 761, + 229 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Proposition D.1. There exists environments $M _ { \\mathcal { X } }$ and $M _ { \\mathcal { Y } }$ such that $\\forall ( x , y ) \\in ~ { \\mathcal { L } }$ where ${ \\mathcal { L } } =$ $\\{ ( x , y ) ~ | x \\in \\mathcal { X }$ , $y \\in \\mathcal { V }$ , $d ^ { * } ( x , y ) = 0 \\}$ , $\\begin{array} { r } { d _ { \\sim } ^ { \\ast } ( x , y ) = \\frac { \\left| R _ { \\mathrm { m a x } } - R _ { \\mathrm { m i n } } \\right| } { 1 - \\gamma } - \\epsilon } \\end{array}$ for any given $\\epsilon > 0$ . ", + "bbox": [ + 174, + 272, + 821, + 308 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "For example, consider the two semantically equivalent environments in Figure D.1 with $\\pi _ { \\mathcal { X } } ^ { * } ( x _ { 0 } ) =$ $\\pi _ { y } ^ { * } ( y _ { 0 } ) = a _ { 0 }$ and $\\pi _ { \\mathcal { X } } ^ { \\ast } ( x _ { 1 } ) = \\pi _ { \\mathcal { Y } } ^ { \\ast } ( y _ { 1 } ) = a _ { 1 }$ but different rewards $r _ { x } , r _ { y }$ respectively. Whenever $r _ { y } > ( 1 + 1 / \\gamma ) r _ { x }$ , bisimulation metrics incorrectly imply that $x _ { 0 }$ is more behaviorally similar to $y _ { 1 }$ than $y _ { 0 }$ . ", + "bbox": [ + 173, + 315, + 825, + 373 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "For the MDPs shown in Figure D.1, to determine which $y$ state is behaviorally equivalent to $x _ { 0 }$ , we look at the distances computed by bisimulation metric $d _ { \\sim }$ and $\\pi ^ { * }$ -bisimulation metric $d _ { \\sim } ^ { \\ast }$ : ", + "bbox": [ + 171, + 378, + 825, + 409 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/e4979757fb4a3e4cbd7241f251f17527e62bb56b8e6a2956053483f92eb86dca.jpg", + "text": "$$\n\\begin{array} { r l } & { d _ { \\sim } ( x _ { 0 } , y _ { 0 } ) = d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } ) = ( 1 + \\gamma ) | r _ { y } - r _ { x } | } \\\\ & { d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) = \\operatorname* { m a x } \\left( ( 1 + \\gamma ) r _ { x } , r _ { y } \\right) , \\ d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 274, + 409, + 578, + 445 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Thus, $r _ { y } > \\left( 1 + 1 / \\gamma \\right) r _ { x }$ implies that $d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ( x _ { 0 } , y _ { 0 } )$ as well as $d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } )$ ", + "bbox": [ + 171, + 445, + 821, + 462 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Connection to data augmentation. Data augmentation often assumes access to optimality invariant transformations, e.g., random crops or flips in image-based benchmarks (Laskin et al., 2020a; Kostrikov et al., 2020). However, for certain RL tasks, such augmentations can erroneously alias states with different optimal behavior and hurt generalization. For example, if the image observation is flipped in a goal reaching task with left and right actions, the optimal actions would also be flipped to take left actions instead of right and vice versa. Proposition D.2 states that PSMs can precisely quantify the invariance of such augmentations. ", + "bbox": [ + 173, + 467, + 826, + 565 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Proposition D.2. For an MDP $M _ { \\mathcal { X } }$ and its transformed version $M _ { \\psi ( \\chi ) }$ for the data augmentation $\\psi$ , $d ^ { * } ( x , \\psi ( x ) )$ indicates the optimality invariance of $\\psi$ for any $x \\in \\mathcal { X }$ . ", + "bbox": [ + 173, + 566, + 823, + 597 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "D.3 PSM WITH APPROXIMATELY-OPTIMAL POLICIES ", + "text_level": 1, + "bbox": [ + 173, + 612, + 560, + 627 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Generalized Policy Similarity Metric for arbitrary policies. For a given DIST, we define a generalized PSM $\\dot { d } : ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\times ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\mathbb { R }$ where $\\Pi$ is the set of all policies over $s$ . $d$ satisfies the recursive equation: ", + "bbox": [ + 173, + 638, + 825, + 681 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/dc3026d09841fbcd424cddad26526fa80b33737ec272f5efff6f2f6bbca80fdd.jpg", + "text": "$$\nd \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) = \\mathrm { D I S T } \\bigl ( \\pi _ { 1 } ( x ) , \\pi _ { 2 } ( y ) \\bigr ) + \\gamma { \\mathcal W } _ { 1 } ( d ) \\bigl ( P ^ { \\pi _ { 1 } } ( \\cdot | x ) , P ^ { \\pi _ { 2 } } ( \\cdot | y ) \\bigr ) .\n$$", + "text_format": "latex", + "bbox": [ + 243, + 681, + 751, + 700 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Since DIST is assumed to be a pseudometric and $\\mathcal { W } _ { 1 }$ is a probability metric, it implies that $d$ is a pseudometric as (1) $d$ is non-negative, that is, $d \\big ( ( x , \\pi _ { 1 } \\bar { ) } , ( y , \\pi _ { 2 } ) \\big ) ^ { \\cdot } \\geq 0$ , (2) $d$ is symmetric, that is, $d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big ) \\ = \\ d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big )$ , and $d$ satisfies the triangle inequality, that is, $d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) < d \\bigl ( ( x , \\pi _ { 1 } ) , ( z , \\pi _ { 3 } ) \\bigr ) + d \\bigl ( ( z , \\pi _ { 3 } ) , ( y , \\pi _ { 2 } ) \\bigr )$ . ", + "bbox": [ + 173, + 700, + 826, + 765 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Using this notion of generalized PSM, we show that the approximation error in PSM from using a suboptimal policy is bounded by the policy’s suboptimality. Thus, for policies with decreasing suboptimality, the PSM approximation becomes more accurate, resulting in improved PSEs. ", + "bbox": [ + 173, + 775, + 826, + 818 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Proposition D.3. [Approximation error in PSM] Let $\\hat { d } : \\mathcal { S } \\times \\mathcal { S } \\mathbb { R }$ be the approximate PSM computed using a suboptimal policy $\\hat { \\pi }$ defined over $s$ , that is, $\\hat { d } ( x , y ) ~ = ~ \\mathrm { D I S T } \\big ( \\hat { \\pi } ( x ) , \\hat { \\pi } ( y ) \\big ) ~ + ~$ $\\gamma \\mathcal { W } _ { 1 } ( \\hat { d } ) \\big ( P ^ { \\hat { \\pi } } ( \\cdot \\vert x ) , P ^ { \\hat { \\pi } } ( \\cdot \\vert y ) \\big )$ . We have: ", + "bbox": [ + 176, + 820, + 825, + 869 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/10f9d8ec51df5d0cf299b606569303f4deef5f621c8aef98d9b38f7d68392ca9.jpg", + "text": "$$\n| d ^ { * } ( x , y ) - { \\hat { d } } ( x , y ) | < { { \\cal A } } \\bigl ( ( x , \\pi ^ { * } ) , ( x , { \\hat { \\pi } } ) \\bigr ) _ { \\bot } + { { \\cal A } } \\bigl ( ( y , { \\hat { \\pi } } ) , ( y , \\pi ^ { * } ) \\bigr ) _ { \\bot } .\n$$", + "text_format": "latex", + "bbox": [ + 245, + 873, + 750, + 893 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Proof. The PSM $d ^ { * }$ and approximate PSM $\\hat { d }$ are instantiations of the generalized PSM (Equation D.3) with both input policies as $\\pi ^ { * }$ and $\\hat { \\pi }$ respectively. ", + "bbox": [ + 171, + 102, + 825, + 131 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/e95da61ecc378d62225eef9e66050b53f1ac386dd3f324fc9c32a1f64535e908.jpg", + "text": "$$\n\\begin{array} { r l } & { d ^ { * } ( x , y ) = d ( ( x , \\pi ^ { * } ) , ( y , \\pi ^ { * } ) ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\\\ & { \\qquad < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { d ^ { * } ( x , y ) - \\hat { d } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) \\quad : \\ \\hat { d } ( x , y ) = d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { { \\mathrm { S i m i l a r l y } } , \\ \\hat { d } ( x , y ) - d ^ { * } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 178, + 147, + 815, + 224 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E L2 METRIC EMBEDDINGS ", + "text_level": 1, + "bbox": [ + 174, + 301, + 424, + 316 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Another common choice (Zhang et al., 2021) for learning metric embeddings is to use the squared loss (i.e., l2-loss) for matching the euclidean distance between the representations of a pair of states to the metric distance between those states. Concretely, for a given $d ^ { * }$ and representation $f _ { \\theta }$ , the loss $\\mathcal { L } ( \\theta ) = \\mathbb { E } _ { s _ { i } , s _ { j } } [ ( \\| f _ { \\theta } ( s _ { i } ) - f _ { \\theta } ( s _ { j } ) \\| _ { 2 } - d ^ { * } ( s _ { i } , s _ { j } ) ) ^ { \\bar { 2 } } ]$ is minimized. However, it might be too restrictive to match the exact metric distances, which we demonstrate empirically by comparing $l 2$ metric embeddings with CMEs (Section 5.2). ", + "bbox": [ + 173, + 337, + 825, + 420 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "F EXTENDED RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 450, + 444, + 467 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Generalization across different tasks used to be described as transfer learning. In the past, most transfer learning approaches relied on fixed representations and tackled different problem formulations (e.g., assuming shared state space). Taylor & Stone (2009) present a comprehensive survey of the techniques at the time, before representation learning became so prevalent in RL. Recently, the problem of performing well in a different, but related task, started to be seen as a problem of generalization; with the community highlighting that deep RL agents tend to overfit to the environments they are trained on (Cobbe et al., 2019; Witty et al., 2018; Farebrother et al., 2018; Juliani et al., 2019; Kostrikov et al., 2020; Song et al., 2019; Justesen et al., 2018; Packer et al., 2018). ", + "bbox": [ + 173, + 486, + 825, + 598 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Prior generalization approaches are typically adapted from supervised learning, including data augmentation, regularization (Cobbe et al., 2019; Farebrother et al., 2018), stochasticity (Zhang et al., 2018c), noise injection (Igl et al., 2019; Zhang et al., 2018a), more diverse training conditions (Rajeswaran et al., 2017; Witty et al., 2018) and self-attention architectures (Tang et al., 2020). In contrast, PSEs exploits behavior similarity (Section 3), a property related to the sequential aspect of RL. Furthermore, for certain RL tasks, it can be unclear what an optimality invariant data augmentation would look like (Section 5.3). PSM can quantify the invariance of such augmentations (Proposition D.2). ", + "bbox": [ + 173, + 604, + 826, + 702 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Meta-learning is also related to generalization. Meta-learning methods try to find a parametrization that requires a small number of gradient steps to achieve good performance on a new task (Finn et al., 2017). In this context, various meta-learning approaches capable of zero-shot generalization have been proposed (Li et al., 2018; Agarwal et al., 2019; Balaji et al., 2018). These approaches typically consist in minimizing the loss in the environments the agent is while adding an auxiliary loss for ensuring improvement in the other (validation) environments available to the agent. Nevertheless, Tachet des Combes et al. (2018) has shown that meta-learning approaches fail in the jumping task which we also observed empirically. Others have also reported similar findings (e.g., Farebrother et al., 2018). ", + "bbox": [ + 173, + 709, + 825, + 835 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "There are several other approaches for tackling zero-shot generalization in RL, but they often rely on domain-specific information. Some examples include knowledge about equivalences between entities in the environment (Oh et al., 2017) and about what is under the agent’s control (Ye et al., 2020). Causality-based methods are a different way of tackling generalization, but current solutions do not scale to high-dimensional observation spaces (e.g., Killian et al., 2017; Perez et al., 2020; Zhang et al., 2020). ", + "bbox": [ + 173, + 842, + 825, + 925 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "G JUMPING TASK WITH PIXELS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 450, + 118 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/ea6505432a67100e306e9753d67f0ba9bec335abc00370ec65ac6866a23b14e2.jpg", + "image_caption": [ + "Figure G.1: Optimal trajectories on the jumping tasks for two different environments. Note that the optimal trajectory is a sequence of right actions, followed by a single jump at a certain distance from the obstacle, followed by right actions. " + ], + "image_footnote": [], + "bbox": [ + 186, + 176, + 812, + 359 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Detailed Task Description. The jumping task consists of an agent trying to jump over an obstacle on a floor. The environment is deterministic with the agent observing a reward of $+ 1$ at each time step. If the agent successfully reaches the rightmost side of the screen, it receives a bonus reward of $+ 1 0 0$ ; if the agent touches the obstacle, the episode terminates. The observation space is the pixel representation of the environment, as depicted in Figure 1. The agent has access to two actions: right and jump. The jump action moves the agent vertically and horizontally to the right. ", + "bbox": [ + 173, + 464, + 825, + 549 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Architecture. The neural network used for Jumping Task experiment is adapted from the Nature DQN architecture. Specifically, the network consists of 3 convolutional layers of sizes 32, 64, 64 with filter sizes $8 \\times 8$ , $4 \\times 4$ and $3 \\times 3$ and strides 4, 2, and 1, respectively. The output of the convnet is fed into a single fully connected layer of size 256 followed by ’ReLU’ non-linearity. Finally, this FC layer output is fed into a linear layer which computes the policy which outputs the probability of the jump and right actions. ", + "bbox": [ + 173, + 555, + 825, + 638 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Contrastive Embedding. For all our experiments, we use a single ReLU layer with $k = 6 4$ units for the non-linear projection to obtain the embedding $z _ { \\theta }$ (Figure A.1). We compute the embedding using the penultimate layer in the jumping task network. Hyperparameters are reported in Table G.2. ", + "bbox": [ + 173, + 646, + 825, + 689 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Total Loss. For jumpy world, the total loss is given by $\\mathcal { L } _ { \\mathrm { I L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }$ where ${ \\mathcal { L } } _ { \\mathrm { I L } }$ is the imitation learning loss, $\\mathcal { L } _ { \\mathrm { C M E } }$ is the auxiliary loss for learning PSEs with the coefficient $\\alpha$ . ", + "bbox": [ + 174, + 694, + 823, + 724 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/b0d8815c5d7275437c5956e7dcd2666091c0d6503fae943844e0b9a282c3036a.jpg", + "image_caption": [ + "Figure G.3: Test performance curves in the setting without data augmentation on the “wide”, “narrow”, and random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25 and 75 percentiles. " + ], + "image_footnote": [], + "bbox": [ + 176, + 750, + 820, + 869 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/7ae61140cd1c0eae07e313b43e44f832d172fa581accd9edd3d1ff30eb85643b.jpg", + "image_caption": [ + "Figure G.2: PSM vs. Bisimulation. Visualizing PSM and bisimulation metrics with discount factor $\\gamma = 0 . 9 9$ . $x _ { i }$ and $y _ { i }$ correspond to the states visited by optimal policies in the two environments with obstacle at positions 25 and 45, respectively. For each grid, the $( \\dot { i } , j ) ^ { t h }$ location shows the distance assigned by the metric to the states $x _ { i }$ and $y _ { j }$ . Lower distances (lighter shades) imply higher similarity between two states. The bottom right figure shows which states are equivalent to each other. The ranges are different for each metric as bisimulation metrics utilize reward differences while PSM uses the total variation $( T V )$ distance between policies. Note that the large bisimulation distances are due to the fact that the reward at the terminal state is set to 100. " + ], + "image_footnote": [], + "bbox": [ + 205, + 87, + 792, + 536 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/42a79abb3d752df4e248de208365ef50ff1cf065c5fbce26bd1ba3fcb36ec3f1.jpg", + "image_caption": [ + "Figure G.4: Test performance curves in the setting with data augmentation on the “wide”, “narrow”, and random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25 and 75 percentiles. " + ], + "image_footnote": [], + "bbox": [ + 176, + 767, + 820, + 883 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/23cbe4e7878d1818c802facc5015190dbca31d0a9ec6d66d453c3d286303c25d.jpg", + "image_caption": [ + "Figure G.5: Jumping task with colored obstacles. The agent needs to jump over the red obstacle but strike the green obstacle. " + ], + "image_footnote": [], + "bbox": [ + 186, + 148, + 419, + 237 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/22af8666fadce0f1cd2db61684e23b73b29f2c8cbdb18535ec2f9cbc313a93d6.jpg", + "table_caption": [ + "Table G.1: Percentage $( \\% )$ of test tasks solved when trained on the “wide” grid with both red and green obstacles. The numbers we report are averaged across 100 runs. Standard error is reported between parentheses. " + ], + "table_footnote": [], + "table_body": "
MethodRed (%)Green (%)
RandConv6.2 (0.4)99.6 (0.2)
Dropout and l2 reg.19.5(0.2)100.0 (0.0)
RandConv +PSEs29.8 (1.3)99.6 (0.2)
PSEs37.9 (1.9)100.0 (0.0)
", + "bbox": [ + 477, + 205, + 803, + 291 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/b97fc7dbb58cf30e9e1722e2dc460df60c4711f902a11b8b59a27bf8a70f5846.jpg", + "image_caption": [ + "Figure G.6: Randconv enforces color invariance. The first column shows the original observation where the top row corresponds to the observation from task with green obstacle jumping task while the bottom row corresponds to the red obstacle jumping task. Columns 2-6 in each row show 5 augmentations by applying RandConv. The augmentations show that RandConv tries to encode invariance with respect to the obstacle color. " + ], + "image_footnote": [], + "bbox": [ + 194, + 330, + 816, + 517 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "G.2 HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 174, + 623, + 361, + 637 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the “wide” grid and pick the parameters with the best validation performance. The validation set (Figure G.7) was selected by using the environments nearby to the training environments whose floor height differ by 1 or whose obstacle position differ by 1. ", + "bbox": [ + 173, + 651, + 825, + 708 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/84cb1c01e9bc78c25c4259b474b636fbd84f41c7ab251a87ba9b20cfba802612.jpg", + "table_caption": [ + "Table G.2: Common hyperparameters across all methods for all jumping task experiments. " + ], + "table_footnote": [], + "table_body": "
HyperparameterValue
Learning rate decay Training epochs0.999
Optimizer2000 Adam
Batch size (Imitation)256
Num training tasks I-scale Parameter (β)18
Embedding size (k)0.01
Batch Size (LCME)64
57
|T57
", + "bbox": [ + 191, + 765, + 418, + 921 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/c50466fd3830e13b4efa16578edae345ae90199f2967bef1d51b49352ce01705.jpg", + "image_caption": [ + "Figure G.7: Unseen 56 validation tasks, labeled with V, used for hyperparameter selection. " + ], + "image_footnote": [], + "bbox": [ + 472, + 750, + 816, + 863 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/857d0992e2052ecea10fe6587da7814aadcd89ca25d72bfa8443ab2ae5e7762f.jpg", + "table_caption": [ + "Table G.3: Optimal hyperparameters for reporting results in Table 1. These hyperparameters are selected using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. All grid configurations in Table 1 use these hyperparameters. " + ], + "table_footnote": [], + "table_body": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-33.2×10-37×10-32.6×10-3
l2-reg. coefficient4.3×10-41×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)11.05×10-1
Auxiliary loss coefficient (α)11×10115.0
", + "bbox": [ + 173, + 150, + 831, + 253 + ], + "page_idx": 22 + }, + { + "type": "table", + "img_path": "images/cea865ec2db0c3f5d013c54414495264361f7e8172a2396a85bd7388ab250bdc.jpg", + "table_caption": [ + "Table G.4: Optimal hyperparameters for reporting results in Figure 5.3. These hyperparameters are selected using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. " + ], + "table_footnote": [], + "table_body": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-36×10-35×10-32.6×10-3
l2-reg. coefficient4.3×10-47×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)15×10-115×10-1
Auxiliary loss coefficient (α)15.015.0
", + "bbox": [ + 179, + 318, + 815, + 421 + ], + "page_idx": 22 + }, + { + "type": "table", + "img_path": "images/06ec11b3bfb0af64be94de637f254be00cc9b2c1aefaac1af5da15fd92dff8a0.jpg", + "table_caption": [ + "Table G.5: Optimal hyperparameters for reporting ablation results in Table 2. These hyperparameters are selected using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. " + ], + "table_footnote": [], + "table_body": "
HyperparameterPSMπ*-bisimulation
CMEsl2-embeddingsCMEsl2-embeddings
Learning Rate4×10-35×10-44.7×10-41×10-4
Contrastive Temperature (1/入)1.015×10-11
Auxiliary loss coefficient (α)5.01×10-11×10-11×10-6
", + "bbox": [ + 192, + 483, + 800, + 579 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Please note that Table G.3 and Table G.4 correspond to two different tasks: one uses the standard jumping task with white obstacles, while the other uses colored obstacles where the optimal policies depend on color. For fair comparison, we tune hyperparameters for all the methods using Bayesian optimization (Golovin et al., 2017). We use the best parameters among these tuned hyperparameters and the ones found in Table G.3, leading to different parameters for both PSEs as well as RandConv. Evaluating PSEs with the jumping task hyperparameters from Table G.3 instead of the ones in Table G.4 leads to a small drop $( - 4 \\% )$ on the jumping task with colors (Section 5.3). Nevertheless, PSEs still outperform other methods in Section 5.3. ", + "bbox": [ + 173, + 598, + 826, + 712 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "H LQR: ADDITIONAL DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 731, + 450, + 747 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Optimal control with linear dynamics and quadratic cost, commonly known as LQR, has been increasingly used as a simplified surrogate for deep RL problems (Recht, 2019). Following Song et al. (2019); Sonar et al. (2020), we analyze the following LQR problem for assessing generalization: ", + "bbox": [ + 174, + 762, + 825, + 805 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/aa837e6fd580fa352b89e2b6bf422ef74b965af98b4dcb5124af5c96245c3d33.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathrm { m i n i m i z e } \\quad E _ { s _ { 0 } \\sim \\mathcal { D } } \\left[ \\frac { 1 } { 2 } \\sum _ { t = 0 } ^ { \\infty } s _ { t } ^ { T } Q s _ { t } + a _ { t } ^ { T } R a _ { t } \\right] , } \\\\ & { \\mathrm { s u b j e c t \\ t o } \\quad s _ { t + 1 } = A s _ { t } + B a _ { t } , o _ { t } = \\left[ \\begin{array} { l } { 0 . 1 W _ { c } } \\\\ { W _ { d } } \\end{array} \\right] s _ { t } , a _ { t } = K o _ { t } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 292, + 810, + 700, + 862 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "where $\\mathcal { D }$ is the initial state distribution, $s _ { t } \\in \\mathbb { R } ^ { n _ { s } }$ is the (hidden) true state at time t, $a _ { t } \\in \\mathbb { R } ^ { n _ { a } }$ is the control action, and $K$ is the linear policy matrix. The agent receives the input observation $o _ { t }$ , which is a linear transformation of the state $s _ { t }$ . $W _ { c }$ and $W _ { d }$ are semi-orthogonal matrices to prevent information loss for predicting optimal actions. An environment corresponds to a particular choice of $W _ { d }$ ; all other system parameters $( A , B , Q , R , W _ { c } )$ are fixed matrices which are shared across environments and unknown to the agent. The agent learns the policy matrix $K$ using $N$ training environments based on Equation H.1. At test time, the learned policy is evaluated on environments with unseen $W _ { d }$ . ", + "bbox": [ + 173, + 867, + 826, + 924 + ], + "page_idx": 22 + }, + { + "type": "table", + "img_path": "images/caafa97b2aa45cc81bd6aca58750d161d74f297e6d88f6d11513e838fea19323.jpg", + "table_caption": [ + "Table H.1: LQR generalization performance: Absolute error in LQR cost, w.r.t. the oracle solver (which has access to true state), of various methods trained with $n _ { d }$ distractors on $N = 2$ environments. The reported mean and standard deviations are across 100 different seeds. Lower error is better. " + ], + "table_footnote": [], + "table_body": "
MethodNumber of Distractors (nd)
500100010000
Overparametrization (Song et al., 2019)25.8 (1.5)24.9 (1.1)24.9 (0.4)
IPO (Sonar et al., 202O) (IRM + Policy opt.)32.6 (5.0)27.3 (2.8)24.8 (0.4)
Weight Sparsity (ll-reg.)28.2 (0.0)28.2 (0.0)28.2 (0.0)
PSM (State aggregation)0.03 (0.0)0.03 (0.0)0.02 (0.0)
", + "bbox": [ + 223, + 150, + 769, + 258 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 276, + 825, + 333 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "The generalization challenge in this setup is to ignore the distractors: $W _ { c } s _ { t } \\in \\mathbb { R } ^ { n _ { s } }$ represents the state features invariant across environments while $W _ { d } s _ { t } \\in \\mathbb { R } ^ { n _ { d } }$ is a high-dimensional distractor of size $n _ { s } , n _ { d }$ , respectively, such that $n _ { s } < < n _ { d }$ . Furthermore, the policy matrix which generalizes across all environments is $K _ { \\star } = \\bigg [ 1 0 W _ { c } P _ { \\star } ^ { ~ T } \\bigg ] ^ { T }$ , where $P _ { \\star }$ corresponds to the optimal LQR solution with access to state $s _ { t }$ . However, for a single environment with distractor $W _ { d }$ , multiple solutions exist, for instance, $K _ { \\star } { } ^ { \\prime } = \\left[ \\begin{array} { c c } { 1 0 \\alpha W _ { c } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\\\ { ( 1 - \\alpha ) W _ { d } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\end{array} \\right] ^ { T } \\quad \\forall \\alpha \\in [ 0 , 1 ]$ . Note that the distractors are an order of magnitude larger than the invariant features in $o _ { t }$ and dependence on them is likely to cause the agent to act erratically on inputs with unseen distractors, resulting in poor generalization. ", + "bbox": [ + 173, + 338, + 825, + 491 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "We use overparametrized policies with two linear layers, i.e., $K = K _ { 1 } K _ { 2 }$ , where $K _ { 1 } ( o )$ is the learned representation for observation $o$ . We learn $K$ using gradient descent using the combined cost on 2 training environments with varying number of distractors. We aggregate observation pairs with near-zero PSM by matching their representations using a squared loss. We use the open-source code released by Sonar et al. (2020) for our experiments. ", + "bbox": [ + 174, + 497, + 825, + 568 + ], + "page_idx": 23 + }, + { + "type": "table", + "img_path": "images/3940a54e5cdfb949bb831f0567a12d25d62c13a9fb0f20c25a21209eed78195a.jpg", + "table_caption": [ + "Table H.2: An overview of hyper-parameters for LQR. " + ], + "table_footnote": [], + "table_body": "
ParameterSeting
AOrthogonal matrix, scaled 0.8
BI20×20
nx20
na20
QI20×20
RI20×20
KiAiOrthogonal Initialization, scaled 0.001
WdRandom semi-orthogonal matrix
", + "bbox": [ + 320, + 606, + 679, + 747 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "The reliance on distractors for IPO also highlights a limitation of IRM: if a model can achieve a solution with zero training error, then any such solution is acceptable by IRM regardless of its generalization ability – a common scenario with overparametrized deep neural nets (Jin et al., 2020). ", + "bbox": [ + 174, + 762, + 825, + 805 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "H.1 NEAR-OPTIMALITY OF PSM AGGREGATION ", + "text_level": 1, + "bbox": [ + 174, + 821, + 524, + 837 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "Conjecture 1. Assuming zero state aggregation error with policy similarity metric (PSM), the policy matrix $K$ learned using gradient descent is independent of the distractors. ", + "bbox": [ + 174, + 849, + 821, + 877 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "Proof. For LQR domains $x , y$ , an observation pair $( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )$ has zero PSM iff the underlying state $s _ { t }$ is same for both the observations in the pair. This is true, as (a) both domains has the same transition ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "dynamics, as specified by Equation H.1, and (b) the optimal policy is deterministic and is completely determined by the current state $s _ { t }$ at any time $t$ . ", + "bbox": [ + 171, + 103, + 825, + 133 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Assume $o _ { t } { } ^ { x } = \\left[ \\begin{array} { c } { 0 . 1 W _ { c } } \\\\ { W _ { d ^ { x } } } \\end{array} \\right] s _ { t }$ and $o _ { t } { } ^ { y } = \\left[ \\stackrel { 0 . 1 } { W } _ { d ^ { y } } \\right] s _ { t }$ for distractor semi-orthogonal matrices $W _ { d ^ { x } }$ and $W _ { d } y$ , respectively. Furthermore, the representation is given by $K _ { 1 } ( o _ { t } ^ { x } )$ and $K _ { 1 } ( o _ { t } ^ { y } )$ respectively. Assume that $K _ { 1 } = [ K _ { s } \\quad K _ { d } ]$ where $K _ { s } \\in \\mathbb { R } ^ { h \\times n _ { s } }$ and $K _ { d } \\in \\mathbb { R } ^ { h \\times n _ { d } }$ and $K _ { 1 } \\in \\mathbb { R } ^ { h \\times ( n _ { s } + n _ { d } ) }$ . ", + "bbox": [ + 173, + 137, + 826, + 200 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Zero state-aggregation error with squared loss implies that for pair $( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )$ corresponding to $s _ { t }$ , ", + "bbox": [ + 173, + 205, + 812, + 222 + ], + "page_idx": 24 + }, + { + "type": "equation", + "img_path": "images/88ac353b7950e601d0879780a90531fb54fce1de1a5d6ff0b4a4c0b3673b63c3.jpg", + "text": "$$\nK _ { 1 } ( o _ { t } { ^ x } - o _ { t } { ^ y } ) = K _ { 1 } [ W _ { d { ^ x } } - W _ { d { ^ y } } ] s _ { t } = 0 \\implies K _ { d } ( W _ { d { ^ x } } - W _ { d { ^ y } } ) s _ { t } = 0\n$$", + "text_format": "latex", + "bbox": [ + 253, + 226, + 745, + 261 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "As Equation H.2 holds for all states visited by the optimal policy in an infinite horizon LQR, it follows that $K _ { d } ( W _ { d ^ { x } } - W _ { d ^ { y } } ) = 0$ . ", + "bbox": [ + 176, + 265, + 823, + 295 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Furthermore, it is well-known that gradient descent tends to find low-rank solutions due to implicit regularization (Arora et al., 2019; Gunasekar et al., 2017), e.g.,with small enough step sizes and initialization close enough to the origin, gradient descent on matrix factorization converges to the minimum nuclear norm solution for 2 layer linear networks (Gunasekar et al., 2017). Based on this, we conjecture that $K _ { d } = 0$ which we found to be true in practice. □ ", + "bbox": [ + 173, + 300, + 825, + 371 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "I DISTRACTING CONTROL SUITE ", + "text_level": 1, + "bbox": [ + 176, + 390, + 464, + 406 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We use the same setup as Kostrikov et al. (2020); Stone et al. (2021) for implementation details and training protocol. For completeness, we describe the details below. ", + "bbox": [ + 174, + 421, + 823, + 450 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Dynamic Background Distractions. In Distracting Control Suite (Stone et al., 2021), random backgrounds are projected from scenes of the DAVIS 2017 dataset (Pont-Tuset et al., 2017) onto the skybox of the scene. To make these backgrounds visible for all tasks and views, the floor grid is semi-transparent with transparency coefficient 0.3. We take the first 2 videos in the DAVIS 2017 training set and randomly sample a scene and a frame from those at the start of every episode. In the dynamic setting, the video plays forwards or backwards until the last or first frame is reached at which point the video is played backwards. This way, the background motion is always smooth and without “cuts”. ", + "bbox": [ + 173, + 455, + 825, + 568 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Soft Actor-Critic. Soft Actor-Critic (SAC) (Haarnoja et al., 2018) learns a state-action value function $Q _ { \\theta }$ , a stochastic policy $\\pi _ { \\theta }$ and a temperature $\\alpha$ to find an optimal policy for an MDP $( S , \\mathcal { A } , p , r , \\gamma )$ by optimizing a $\\gamma$ -discounted maximum-entropy objective. $\\theta$ is used generically to denote the parameters updated through training in each part of the model. The actor policy $\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )$ is a parametric tanhGaussian that given $s _ { t }$ samples $a _ { t } \\dot { = } \\operatorname { t a n h } ( \\mu _ { \\theta } ( s _ { t } ) + \\sigma _ { \\theta } ( s _ { t } ) \\epsilon )$ , where $\\epsilon \\sim \\mathcal { N } ( 0 , 1 )$ and $\\mu _ { \\theta }$ and $\\sigma _ { \\theta }$ are parametric mean and standard deviation. ", + "bbox": [ + 173, + 574, + 826, + 660 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "The policy evaluation step learns the critic $Q _ { \\theta } ( s _ { t } , a _ { t } )$ network by optimizing a single-step of the soft Bellman residual ", + "bbox": [ + 173, + 665, + 823, + 695 + ], + "page_idx": 24 + }, + { + "type": "equation", + "img_path": "images/d967d8f701d309f8a2b9d1f8aed651c62ceb99e0c454aba8f437771b1a610922.jpg", + "text": "$$\n\\begin{array} { r l } & { J _ { Q } ( \\mathcal { D } ) = E _ { ( s _ { t } , a _ { t } , s _ { t } ^ { \\prime } ) \\sim \\mathcal { D } } [ ( Q _ { \\theta } ( s _ { t } , a _ { t } ) - y _ { t } ) ^ { 2 } ] } \\\\ & { \\qquad a _ { t } ^ { \\prime } { \\sim } \\pi ( \\cdot | s _ { t } ^ { \\prime } ) } \\\\ & { y _ { t } = r ( s _ { t } , a _ { t } ) + \\gamma [ Q _ { \\theta ^ { \\prime } } ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) - \\alpha \\log \\pi _ { \\theta } ( a _ { t } ^ { \\prime } | s _ { t } ^ { \\prime } ) ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 318, + 698, + 678, + 750 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "where $\\mathcal { D }$ is a replay buffer of transitions, $\\theta ^ { \\prime }$ is an exponential moving average of the weights. SAC uses clipped double-Q learning, which we omit for simplicity but employ in practice. ", + "bbox": [ + 171, + 752, + 823, + 782 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "The policy improvement step then fits the actor policy $\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )$ network by optimizing the objective ", + "bbox": [ + 174, + 787, + 821, + 804 + ], + "page_idx": 24 + }, + { + "type": "equation", + "img_path": "images/b8de65fbed10cf47b2b18c028ad713f54b791fcc6be2ce542047cddf3eb2f735.jpg", + "text": "$$\nJ _ { \\pi } ( \\mathcal { D } ) = E _ { s _ { t } \\sim \\mathcal { D } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\theta } ( \\cdot | s _ { t } ) | | \\exp \\{ \\frac { 1 } { \\alpha } Q _ { \\theta } ( s _ { t } , \\cdot ) \\} ) ] .\n$$", + "text_format": "latex", + "bbox": [ + 323, + 809, + 673, + 838 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Finally, the temperature $\\alpha$ is learned with the loss ", + "bbox": [ + 174, + 842, + 500, + 858 + ], + "page_idx": 24 + }, + { + "type": "equation", + "img_path": "images/2cb81c1d6d5abb9132324b381e379d4fad3b4465a9273c43452d653b03c371b4.jpg", + "text": "$$\nJ _ { \\alpha } ( \\mathcal { D } ) = E _ { \\underset { a _ { t } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t } ) } { s _ { t } \\sim \\mathcal { D } } } [ - \\alpha \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) - \\alpha \\bar { \\mathcal { H } } ] ,\n$$", + "text_format": "latex", + "bbox": [ + 339, + 862, + 656, + 890 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "where $\\bar { \\mathcal { H } } \\in \\mathbb { R }$ is the target entropy hyper-parameter that the policy tries to match, which in practice is usually set to $\\bar { \\mathcal { H } } = - | \\bar { \\mathcal { A } } |$ . ", + "bbox": [ + 174, + 895, + 825, + 925 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/7c1ed64c41245f4fe76e367be7d81f27ce231d876cb7515942b5a20a75e1bb2a.jpg", + "image_caption": [ + "Figure I.1: DCS Test Environments: Snapshots of test environments used for evaluating generalization on Distracting Control Suite. Random backgrounds are projected from scenes of the first 30 videos of DAVIS 2017 validation dataset. " + ], + "image_footnote": [], + "bbox": [ + 196, + 107, + 813, + 888 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "I.1 ACTOR AND CRITIC NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 434, + 117 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "Following Kostrikov et al. (2020), we use clipped double Q-learning for the critic, where each $Q$ -function is parametrized as a 3-layer MLP with ReLU activations after each layer except of the last. The actor is also a 3-layer MLP with ReLUs that outputs mean and covariance for the diagonal Gaussian that represents the policy. The hidden dimension is set to 1024 for both the critic and actor. ", + "bbox": [ + 174, + 131, + 825, + 186 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "I.2 ENCODER NETWORK ", + "text_level": 1, + "bbox": [ + 174, + 207, + 357, + 220 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "We employ the encoder architecture from Kostrikov et al. (2020). This encoder consists of four convolution layers with $3 \\times 3$ kernels and 32 channels. The ReLU activation is applied after each convolutional layer. We use stride to 1 everywhere, except of the first convolutional layer, which has stride 2. The output of the convnet is feed into a single fully-connected layer normalized by LayerNorm. Finally, we apply tanh nonlinearity to the 50 dimensional output of the fullyconnected layer. We initialize the weight matrix of fully-connected and convolutional layers with the orthogonal initialization and set the bias to be zero. The actor and critic networks both have separate encoders, although we share the weights of the conv layers between them. Furthermore, only the critic optimizer is allowed to update these weights (i.e.,we stop the gradients from the actor before they propagate to the shared convolutional layers). ", + "bbox": [ + 173, + 234, + 825, + 373 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "I.3 CONTRASTIVE METRIC EMBEDDING LOSS ", + "text_level": 1, + "bbox": [ + 174, + 386, + 508, + 401 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "For all our experiments, we use a single ReLU layer with $k = 2 5 6$ units for the non-linear projection to obtain the embedding $z _ { \\theta }$ (Figure A.1). We compute the embedding using the penultimate layer in the actor network. For picking the hyperparameters, we used 3 temperatures [0.1, 0.01, 1.0] and 3 auxiliary ${ \\mathcal { L } } _ { \\mathrm { C M E } }$ loss coefficients [1, 3, 10] using “Ball In Cup Catch” as the validation environment. All other hyperparameters are the same as prior work (see Table I.2). ", + "bbox": [ + 174, + 414, + 825, + 484 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "We approximate optimal policies with the policies obtained after training a DrQ agent for 500K environment steps. Since a given action sequence from this approximate policy has the same performance across different training environments, we compute the PSM across training environments, via dynamic programming (see Section J.1 for pseudo-code), using such action sequences. ", + "bbox": [ + 173, + 491, + 825, + 546 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "Total Loss. The total loss is given by $\\mathcal { L } _ { \\mathrm { R L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }$ where $\\mathcal { L } _ { \\mathrm { R L } }$ is the reinforcement learning loss which combines $J _ { \\pi } ( \\mathcal { D } )$ , $J _ { \\pi } ( \\mathcal { D } )$ , and $J _ { \\alpha } ( { \\mathcal { D } } ) _ { , }$ ), while ${ \\mathcal { L } } _ { \\mathrm { C M E } }$ is the auxiliary loss for learning PSEs with the coefficient $\\alpha$ . ", + "bbox": [ + 176, + 553, + 823, + 595 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "I.4 TRAINING AND EVALUATION SETUP ", + "text_level": 1, + "bbox": [ + 176, + 616, + 462, + 631 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "For evaluation, we use the first 30 videos from the DAVIS 2017 validation dataset (see Figure I.1). Each checkpoint is evaluated by computing the average episode return over 100 episodes from the unseen environments. All experiments are performed with five random seeds per task used to compute means and standard deviations/errors of their evaluations. We use $K = 2 , M = 2$ as prescribed by Kostrikov et al. (2020) for $\\mathrm { D r Q }$ . Following Kostrikov et al. (2020) and Stone et al. (2021), we use a different action repeat hyper-parameter for each task, which we summarize in Table I.3. We construct an observational input as a 3-stack of consecutive frames (Kostrikov et al., 2020), where each frame is an RGB rendering of size $8 4 \\times 8 4$ from the 0th camera. We then divide each pixel by 255 to scale it down to [0, 1] range. For data augmentation, we maintain temporal consistency by using the same crop augmentation across consecutive frames. ", + "bbox": [ + 173, + 643, + 825, + 782 + ], + "page_idx": 26 + }, + { + "type": "table", + "img_path": "images/1d4f1d425398184d92075449048aba2f6b98a942f7b3046738afc0165b55abf2.jpg", + "table_caption": [ + "Table I.1: Optimal hyperparameters for PSE auxiliary loss for reporting results in Table 3. " + ], + "table_footnote": [], + "table_body": "
HyperparameterSetting
Contrastive temperature (1/入)0.1
Auxiliary loss coefficient (α)1.0
I-scale parameter (β)0.1
Batch Size(LcME)128
|1000 // Action Repeat
", + "bbox": [ + 313, + 824, + 678, + 925 + ], + "page_idx": 26 + }, + { + "type": "table", + "img_path": "images/bb9329f1383c2eca2834c2f994d252c85f7b9e348c37d854ea8cbdce690b2b7e.jpg", + "table_caption": [ + "Table I.2: Hyper-parameters taken from Kostrikov et al. (2020) in the Distracting Control Suite experiments. " + ], + "table_footnote": [], + "table_body": "
ParameterSetting
Replay buffer capacity100,000
Seed steps1,000
Batch size (DrQ)512
Discount γ0.99
OptimizerAdam
Learning rate10-3
Critic target update frequency2
Critic Q-function soft-update rate T0.01
Actor update frequency2
Actor log stddev bounds
Init temperature[-10,2] 0.1
", + "bbox": [ + 331, + 126, + 663, + 309 + ], + "page_idx": 27 + }, + { + "type": "table", + "img_path": "images/da8d4384f50c36616da3269b10b8f38eb7bac21c398f4c93c4a028bf3daf07c3.jpg", + "table_caption": [ + "Table I.3: The action repeat hyper-parameter used for each task in the Distracting Control Suite benchmark. " + ], + "table_footnote": [], + "table_body": "
Task nameAction repeat
Cartpole Swingup8
Reacher Easy4
Cheetah Run4
Finger Spin2
Ball In Cup Catch4
Walker Walk2
", + "bbox": [ + 383, + 378, + 614, + 492 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/9058a9690116b883329a6466ccc12d2c432e08d5b417ce9c0fc8928c6ebed906.jpg", + "image_caption": [ + "Figure I.2: Random Initialization. Generalization performance on unseen environments over the course of training. The agent is initialized randomly. $_ \\mathrm { D r Q }$ augmented with PSEs outperforms DrQ. We plot the average episode return across 5 seeds and the shaded region shows the standard deviation. Each checkpoint is evaluated using 100 episodes with unseen backgrounds. " + ], + "image_footnote": [], + "bbox": [ + 173, + 574, + 825, + 857 + ], + "page_idx": 27 + }, + { + "type": "image", + "img_path": "images/48d74111701471e8fc0c0104c5c1128c88f60454f9197c2358f703f65f4db49e.jpg", + "image_caption": [ + "Figure I.3: Pretrained Initialization. Generalization performance on unseen environments over the course of training. The agent is initialized using a pretrained $\\mathrm { D r Q }$ agent. DrQ augmented with PSEs outperforms $\\mathrm { D r Q }$ on most of the environments. We plot the average return across 5 seeds and the shaded region shows the standard deviation. Each checkpoint is evaluated using 100 episodes with unseen backgrounds. " + ], + "image_footnote": [], + "bbox": [ + 173, + 104, + 825, + 386 + ], + "page_idx": 28 + }, + { + "type": "text", + "text": "J PSEUDO CODE ", + "text_level": 1, + "bbox": [ + 174, + 477, + 325, + 492 + ], + "page_idx": 28 + }, + { + "type": "text", + "text": "J.1 DYNAMIC PROGRAMMING FOR COMPUTING PSM ", + "text_level": 1, + "bbox": [ + 173, + 508, + 558, + 522 + ], + "page_idx": 28 + }, + { + "type": "text", + "text": "1 def metric_fixed_point(cost_matrix, gamma ${ } = 0$ .99, ep $s { = } 1 { \\mathrm { e - } } 7$ ): \n2 \"\"\"DP for calculating PSM in environments with deterministic dynamics. \n3 \n4 Args: \n5 cost_matrix: DIST matrix where entries at index (i, j) is DIST(x_i, \ny_j) \n6 gamma: Metric discount factor. \n7 eps: Threshold for stopping the fixed point iteration. \n8 \"\"\" \n9 d $=$ np.zeros_like(cost_matrix) \n10 def operator(d_cur): \n11 d_new $\\ c = ~ 1$ $\\star$ cost_matrix \n12 discounted_d_cur $=$ gamma $\\star$ d_cur \n13 d_new[:-1, :-1] $+ =$ discounted_d_cur[1:, 1:] \n14 d_new[:-1, -1] $+ =$ discounted_d_cur[1:, -1] \n15 d_new[-1, :-1] $+ =$ discounted_d_cur[-1, 1:] \n16 return d_new \n17 \n18 while True: \n19 d_new $=$ operator(d) \n20 if np.sum(np.abs(d - d_new)) $<$ eps: \n21 break \n22 else: \n23 d $=$ d_new[:] \n24 return d ", + "bbox": [ + 155, + 531, + 803, + 848 + ], + "page_idx": 28 + }, + { + "type": "text", + "text": "J.2 CONTRASTIVE LOSS ", + "text_level": 1, + "bbox": [ + 173, + 861, + 356, + 875 + ], + "page_idx": 28 + }, + { + "type": "text", + "text": "1 def contrastive_loss(similarity_matrix, \n2 metric_values, \n3 temperature, \n4 bet $\\gimel = 1 \\digamma . 0$ ): \n5 \"\"\"Contrative Loss with embedding similarity .\"\"\" \n6 metric_shape $=$ tf.shape(metric_values) \n7 ## z_\\theta(X): embedding_1 $=$ nn_model.representation(X) \n8 ## z_\\theta(Y): embedding_ $\\_ =$ nn_model.representation(Y) \n9 ## similarity_matrix $=$ cosine_similarity(embedding_1, embedding_2 \n10 ## metric_values $=$ PSM(X, Y) \n11 similarity_matrix / $=$ temperature \n12 neg_logits1 $=$ similarity_matrix \n13 \n14 col_indices $=$ tf.cast(tf.argmin(metric_values, axi $\\ S = 1$ ), dtype $=$ tf.int32) \n15 pos_indices1 $=$ tf.stack( \n16 (tf.range(metric_shape[0], dtype $=$ tf.int32), col_indices), axi $\\mathsf { S } = 1$ ) \n17 pos_logits1 $=$ tf.gather_nd(similarity_matrix, pos_indices1) \n18 \n19 metric_values / $=$ beta \n20 similarity_measure $=$ tf.exp(-metric_values) \n21 pos_weights1 $=$ -tf.gather_nd(metric_values, pos_indices1) \n22 pos_logits1 $+ =$ pos_weights1 \n23 negative_weights $=$ tf.math.log((1.0 - similarity_measure) $^ +$ 1e-8) \n24 neg_logits1 $+ =$ tf.tensor_scatter_nd_update( \n25 negative_weights, pos_indices1, pos_weights1) \n26 \n27 neg_logits1 $=$ tf.math.reduce_logsumexp(neg_logits1, axis $^ { = 1 }$ ) \n28 return tf.reduce_mean(neg_logits1 - pos_logits1) # Equation 4 ", + "bbox": [ + 160, + 886, + 516, + 924 + ], + "page_idx": 28 + }, + { + "type": "text", + "text": "", + "bbox": [ + 153, + 103, + 818, + 422 + ], + "page_idx": 29 + } +] \ No newline at end of file diff --git a/parse/train/qda7-sVg84/qda7-sVg84_middle.json b/parse/train/qda7-sVg84/qda7-sVg84_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..f295cd88be90b21edb7f21000d51f6dd337e7139 --- /dev/null +++ b/parse/train/qda7-sVg84/qda7-sVg84_middle.json @@ -0,0 +1,86575 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 411, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 413, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 413, + 96 + ], + "score": 1.0, + "content": "CONTRASTIVE BEHAVIORAL SIMILARITY", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 401, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 401, + 117 + ], + "score": 1.0, + "content": "EMBEDDINGS FOR GENERALIZATION IN", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 119, + 321, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 321, + 136 + ], + "score": 1.0, + "content": "REINFORCEMENT LEARNING", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 153, + 484, + 188 + ], + "lines": [ + { + "bbox": [ + 111, + 153, + 486, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 153, + 486, + 168 + ], + "score": 1.0, + "content": "Rishabh Agarwal∗ Marlos C. Machado‡ Pablo Samuel Castro Marc G. Bellemare", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 165, + 235, + 178 + ], + "spans": [ + { + "bbox": [ + 112, + 165, + 235, + 178 + ], + "score": 1.0, + "content": "Google Research, Brain Team", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 177, + 427, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 177, + 427, + 190 + ], + "score": 1.0, + "content": "{rishabhagarwal, marlosm, psc, bellemare}@google.com", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 203, + 333, + 215 + ], + "lines": [ + { + "bbox": [ + 277, + 203, + 335, + 216 + ], + "spans": [ + { + "bbox": [ + 277, + 203, + 335, + 216 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 222, + 469, + 354 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "Reinforcement learning methods trained on few environments rarely learn policies", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "that generalize to unseen environments. To improve generalization, we incorporate", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "the inherent sequential structure in reinforcement learning into the representation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "learning process. This approach is orthogonal to recent approaches, which rarely", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "exploit this structure explicitly. Specifically, we introduce a theoretically motivated", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 471, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 471, + 290 + ], + "score": 1.0, + "content": "policy similarity metric (PSM) for measuring behavioral similarity between states.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "score": 1.0, + "content": "PSM assigns high similarity to states for which the optimal policies in those states", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "score": 1.0, + "content": "as well as in future states are similar. We also present a contrastive representation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "learning procedure to embed any state similarity metric, which we instantiate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 345, + 334 + ], + "score": 1.0, + "content": "with PSM to obtain policy similarity embeddings", + "type": "text" + }, + { + "bbox": [ + 345, + 321, + 376, + 333 + ], + "score": 0.28, + "content": "( \\mathrm { P S E s ^ { 1 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "). We demonstrate that", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "score": 1.0, + "content": "PSEs improve generalization on diverse benchmarks, including LQR with spurious", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 343, + 447, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 447, + 356 + ], + "score": 1.0, + "content": "correlations, a jumping task from pixels, and Distracting DM Control Suite.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 369, + 206, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 208, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 208, + 384 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 317, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 318, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 318, + 399 + ], + "score": 1.0, + "content": "Current reinforcement learning (RL) approaches of-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 398, + 318, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 318, + 410 + ], + "score": 1.0, + "content": "ten learn policies that do not generalize to environ-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 409, + 318, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 318, + 421 + ], + "score": 1.0, + "content": "ments different than those the agent was trained on,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 420, + 318, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 318, + 431 + ], + "score": 1.0, + "content": "even when these environments are semantically equiv-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 430, + 318, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 318, + 443 + ], + "score": 1.0, + "content": "alent (Tachet des Combes et al., 2018; Song et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 442, + 317, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 317, + 453 + ], + "score": 1.0, + "content": "2019; Cobbe et al., 2019). For example, consider a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 453, + 318, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 318, + 465 + ], + "score": 1.0, + "content": "jumping task where an agent, learning from pixels,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 464, + 317, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 317, + 475 + ], + "score": 1.0, + "content": "needs to jump over an obstacle (Figure 1). Deep RL", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 475, + 317, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 317, + 486 + ], + "score": 1.0, + "content": "agents trained on a few of these tasks with different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 487, + 317, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 317, + 497 + ], + "score": 1.0, + "content": "obstacle positions struggle to solve test tasks where", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 497, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 286, + 509 + ], + "score": 1.0, + "content": "obstacles are at previously unseen locations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 336, + 382, + 490, + 442 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 336, + 382, + 490, + 442 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 336, + 382, + 490, + 442 + ], + "spans": [ + { + "bbox": [ + 336, + 382, + 490, + 442 + ], + "score": 0.968, + "type": "image", + "image_path": "abed9d708ee10498bdd0d0bf48ab3b19c58d372b5155e6350f85107e57f370fa.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 336, + 382, + 490, + 394.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 336, + 394.0, + 490, + 406.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 336, + 406.0, + 490, + 418.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 336, + 418.0, + 490, + 430.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 336, + 430.0, + 490, + 442.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 446, + 505, + 506 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 324, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 324, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "Figure 1: Jumping task: The agent (white block),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 323, + 456, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 323, + 456, + 506, + 468 + ], + "score": 1.0, + "content": "learning from pixels, needs to jump over an obsta-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 324, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 324, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "cle (grey square). The challenge is to generalize to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 324, + 477, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 324, + 477, + 505, + 487 + ], + "score": 1.0, + "content": "unseen obstacle positions and floor heights in test", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 324, + 487, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 324, + 487, + 504, + 497 + ], + "score": 1.0, + "content": "tasks using a small number of training tasks. We", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 324, + 496, + 500, + 507 + ], + "spans": [ + { + "bbox": [ + 324, + 496, + 500, + 507 + ], + "score": 1.0, + "content": "show the agent’s trajectories using faded blocks.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + } + ], + "index": 35.75 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 507, + 527 + ], + "score": 1.0, + "content": "Recent solutions to circumvent poor generalization in RL are adapted from supervised learning,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 525, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 536 + ], + "score": 1.0, + "content": "and, as such, largely ignore the sequential aspect of RL. Most of these solutions revolve around", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 535, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 506, + 548 + ], + "score": 1.0, + "content": "enhancing the learning process, including data augmentation (e.g., Kostrikov et al., 2020; Lee et al.,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "2020a), regularization (Cobbe et al., 2019; Farebrother et al., 2018), noise injection (Igl et al., 2019),", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 557, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 569 + ], + "score": 1.0, + "content": "and diverse training conditions (Tobin et al., 2017); they rarely exploit properties of the sequential", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 568, + 444, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 444, + 581 + ], + "score": 1.0, + "content": "decision making problem such as similarity in actions across temporal observations.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "Instead, we tackle generalization by incorporating properties of the RL problem into the representation", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "score": 1.0, + "content": "learning process. Our approach exploits the fact that an agent, when operating in environments with", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "similar underlying mechanics, exhibits at least short sequences of behaviors that are similar across", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "these environments. Concretely, the agent is optimized to learn an embedding in which states are", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "close when the agent’s optimal policies in these states and future states are similar. This notion of", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 641, + 447, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 447, + 653 + ], + "score": 1.0, + "content": "proximity is general and it is applicable to observations from different environments.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5 + }, + { + "type": "text", + "bbox": [ + 107, + 657, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 507, + 671 + ], + "score": 1.0, + "content": "Specifically, inspired by bisimulation metrics (Castro, 2020; Ferns et al., 2004), we propose a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "novel policy similarity metric (PSM). PSM (Section 3) defines a notion of similarity between states", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "originated from different environments by the proximity of the long-term optimal behavior from these", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "score": 1.0, + "content": "states. PSM is reward-agnostic, making it more robust for generalization compared to approaches that", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 710, + 334, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 334, + 722 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 334, + 722 + ], + "score": 1.0, + "content": "∗Also at Mila, Université de Montréal. ‡Now at DeepMind.", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 720, + 196, + 732 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 196, + 732 + ], + "score": 1.0, + "content": "1Pronounce ‘pisces’.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 411, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 413, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 413, + 96 + ], + "score": 1.0, + "content": "CONTRASTIVE BEHAVIORAL SIMILARITY", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 401, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 401, + 117 + ], + "score": 1.0, + "content": "EMBEDDINGS FOR GENERALIZATION IN", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 119, + 321, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 321, + 136 + ], + "score": 1.0, + "content": "REINFORCEMENT LEARNING", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 153, + 484, + 188 + ], + "lines": [ + { + "bbox": [ + 111, + 153, + 486, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 153, + 486, + 168 + ], + "score": 1.0, + "content": "Rishabh Agarwal∗ Marlos C. Machado‡ Pablo Samuel Castro Marc G. Bellemare", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 165, + 235, + 178 + ], + "spans": [ + { + "bbox": [ + 112, + 165, + 235, + 178 + ], + "score": 1.0, + "content": "Google Research, Brain Team", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 177, + 427, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 177, + 427, + 190 + ], + "score": 1.0, + "content": "{rishabhagarwal, marlosm, psc, bellemare}@google.com", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 111, + 153, + 486, + 190 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 203, + 333, + 215 + ], + "lines": [ + { + "bbox": [ + 277, + 203, + 335, + 216 + ], + "spans": [ + { + "bbox": [ + 277, + 203, + 335, + 216 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 222, + 469, + 354 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "Reinforcement learning methods trained on few environments rarely learn policies", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "that generalize to unseen environments. To improve generalization, we incorporate", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "the inherent sequential structure in reinforcement learning into the representation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "learning process. This approach is orthogonal to recent approaches, which rarely", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 279 + ], + "score": 1.0, + "content": "exploit this structure explicitly. Specifically, we introduce a theoretically motivated", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 471, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 471, + 290 + ], + "score": 1.0, + "content": "policy similarity metric (PSM) for measuring behavioral similarity between states.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "score": 1.0, + "content": "PSM assigns high similarity to states for which the optimal policies in those states", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "score": 1.0, + "content": "as well as in future states are similar. We also present a contrastive representation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "learning procedure to embed any state similarity metric, which we instantiate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 345, + 334 + ], + "score": 1.0, + "content": "with PSM to obtain policy similarity embeddings", + "type": "text" + }, + { + "bbox": [ + 345, + 321, + 376, + 333 + ], + "score": 0.28, + "content": "( \\mathrm { P S E s ^ { 1 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "). We demonstrate that", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "score": 1.0, + "content": "PSEs improve generalization on diverse benchmarks, including LQR with spurious", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 343, + 447, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 447, + 356 + ], + "score": 1.0, + "content": "correlations, a jumping task from pixels, and Distracting DM Control Suite.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5, + "bbox_fs": [ + 141, + 222, + 471, + 356 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 369, + 206, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 208, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 208, + 384 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 317, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 318, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 318, + 399 + ], + "score": 1.0, + "content": "Current reinforcement learning (RL) approaches of-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 398, + 318, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 318, + 410 + ], + "score": 1.0, + "content": "ten learn policies that do not generalize to environ-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 409, + 318, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 318, + 421 + ], + "score": 1.0, + "content": "ments different than those the agent was trained on,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 420, + 318, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 318, + 431 + ], + "score": 1.0, + "content": "even when these environments are semantically equiv-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 430, + 318, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 318, + 443 + ], + "score": 1.0, + "content": "alent (Tachet des Combes et al., 2018; Song et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 442, + 317, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 317, + 453 + ], + "score": 1.0, + "content": "2019; Cobbe et al., 2019). For example, consider a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 453, + 318, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 318, + 465 + ], + "score": 1.0, + "content": "jumping task where an agent, learning from pixels,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 464, + 317, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 317, + 475 + ], + "score": 1.0, + "content": "needs to jump over an obstacle (Figure 1). Deep RL", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 475, + 317, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 317, + 486 + ], + "score": 1.0, + "content": "agents trained on a few of these tasks with different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 487, + 317, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 317, + 497 + ], + "score": 1.0, + "content": "obstacle positions struggle to solve test tasks where", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 497, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 286, + 509 + ], + "score": 1.0, + "content": "obstacles are at previously unseen locations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 387, + 318, + 509 + ] + }, + { + "type": "image", + "bbox": [ + 336, + 382, + 490, + 442 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 336, + 382, + 490, + 442 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 336, + 382, + 490, + 442 + ], + "spans": [ + { + "bbox": [ + 336, + 382, + 490, + 442 + ], + "score": 0.968, + "type": "image", + "image_path": "abed9d708ee10498bdd0d0bf48ab3b19c58d372b5155e6350f85107e57f370fa.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 336, + 382, + 490, + 394.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 336, + 394.0, + 490, + 406.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 336, + 406.0, + 490, + 418.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 336, + 418.0, + 490, + 430.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 336, + 430.0, + 490, + 442.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 446, + 505, + 506 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 324, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 324, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "Figure 1: Jumping task: The agent (white block),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 323, + 456, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 323, + 456, + 506, + 468 + ], + "score": 1.0, + "content": "learning from pixels, needs to jump over an obsta-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 324, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 324, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "cle (grey square). The challenge is to generalize to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 324, + 477, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 324, + 477, + 505, + 487 + ], + "score": 1.0, + "content": "unseen obstacle positions and floor heights in test", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 324, + 487, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 324, + 487, + 504, + 497 + ], + "score": 1.0, + "content": "tasks using a small number of training tasks. We", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 324, + 496, + 500, + 507 + ], + "spans": [ + { + "bbox": [ + 324, + 496, + 500, + 507 + ], + "score": 1.0, + "content": "show the agent’s trajectories using faded blocks.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + } + ], + "index": 35.75 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 507, + 527 + ], + "score": 1.0, + "content": "Recent solutions to circumvent poor generalization in RL are adapted from supervised learning,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 525, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 536 + ], + "score": 1.0, + "content": "and, as such, largely ignore the sequential aspect of RL. Most of these solutions revolve around", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 535, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 506, + 548 + ], + "score": 1.0, + "content": "enhancing the learning process, including data augmentation (e.g., Kostrikov et al., 2020; Lee et al.,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "2020a), regularization (Cobbe et al., 2019; Farebrother et al., 2018), noise injection (Igl et al., 2019),", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 557, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 569 + ], + "score": 1.0, + "content": "and diverse training conditions (Tobin et al., 2017); they rarely exploit properties of the sequential", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 568, + 444, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 444, + 581 + ], + "score": 1.0, + "content": "decision making problem such as similarity in actions across temporal observations.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 511, + 507, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 598 + ], + "score": 1.0, + "content": "Instead, we tackle generalization by incorporating properties of the RL problem into the representation", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 504, + 609 + ], + "score": 1.0, + "content": "learning process. Our approach exploits the fact that an agent, when operating in environments with", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "similar underlying mechanics, exhibits at least short sequences of behaviors that are similar across", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "these environments. Concretely, the agent is optimized to learn an embedding in which states are", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "close when the agent’s optimal policies in these states and future states are similar. This notion of", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 641, + 447, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 447, + 653 + ], + "score": 1.0, + "content": "proximity is general and it is applicable to observations from different environments.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 585, + 506, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 657, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 507, + 671 + ], + "score": 1.0, + "content": "Specifically, inspired by bisimulation metrics (Castro, 2020; Ferns et al., 2004), we propose a", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "novel policy similarity metric (PSM). PSM (Section 3) defines a notion of similarity between states", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "originated from different environments by the proximity of the long-term optimal behavior from these", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "score": 1.0, + "content": "states. PSM is reward-agnostic, making it more robust for generalization compared to approaches that", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "rely on reward information. We prove that PSM yields an upper bound on suboptimality of policies", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 105 + ], + "score": 1.0, + "content": "transferred from one environment to another (Theorem 1), which is not attainable with bisimulation.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 55.5, + "bbox_fs": [ + 105, + 657, + 507, + 702 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "rely on reward information. We prove that PSM yields an upper bound on suboptimality of policies", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 105 + ], + "score": 1.0, + "content": "transferred from one environment to another (Theorem 1), which is not attainable with bisimulation.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "We employ PSM for representation learning and introduce policy similarity embeddings (PSEs) for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "deep RL. To do so, we present a general contrastive procedure (Section 4) to learn an embedding", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "based on any state similarity metric. PSEs are the instantiation of this procedure with PSM. PSEs", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 504, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 504, + 156 + ], + "score": 1.0, + "content": "are appealing for generalization as they encode task-relevant invariances by putting behaviorally", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "equivalent states together. This is unlike prior approaches, which rely on capturing such invariances", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "without being explicitly trained to do so, for example, through value function similarities across", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 188 + ], + "score": 1.0, + "content": "states (e.g., Castro & Precup, 2010), or being robust to fixed transformations of the observation", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 331, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 331, + 200 + ], + "score": 1.0, + "content": "space (e.g., Kostrikov et al., 2020; Laskin et al., 2020a).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 505, + 217 + ], + "score": 1.0, + "content": "PSEs lead to better generalization while being orthogonal to how most of the field has been tackling", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "generalization. We illustrate the efficacy and broad applicability of our approach on three existing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 226, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 506, + 239 + ], + "score": 1.0, + "content": "benchmarks specifically designed to test generalization: (i) jumping task from pixels (Tachet des", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "Combes et al., 2018) (Section 5), (ii) LQR with spurious correlations (Song et al., 2019) (Section 6.1),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "and (iii) Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b) (Section 6.2). Our", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "approach improves generalization compared to a wide variety of approaches including standard", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "score": 1.0, + "content": "regularization (Farebrother et al., 2018; Cobbe et al., 2019), bisimulation (Castro & Precup, 2010;", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "Castro, 2020; Zhang et al., 2021), out-of-distribution generalization (Arjovsky et al., 2019) and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 412, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 412, + 303 + ], + "score": 1.0, + "content": "state-of-the-art data augmentation (Kostrikov et al., 2020; Lee et al., 2020a).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 311, + 208, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 209, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 209, + 326 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 476, + 343 + ], + "score": 1.0, + "content": "We describe an environment as a Markov decision process (MDP) (Puterman, 1994)", + "type": "text" + }, + { + "bbox": [ + 476, + 331, + 505, + 342 + ], + "score": 0.83, + "content": "\\begin{array} { r l } { \\mathcal { M } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 168, + 354 + ], + "score": 0.91, + "content": "( \\mathcal { X } , \\mathcal { A } , R , P , \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 340, + 245, + 354 + ], + "score": 1.0, + "content": ", with a state space", + "type": "text" + }, + { + "bbox": [ + 246, + 342, + 255, + 352 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 340, + 324, + 354 + ], + "score": 1.0, + "content": ", an action space", + "type": "text" + }, + { + "bbox": [ + 324, + 342, + 333, + 352 + ], + "score": 0.79, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 340, + 410, + 354 + ], + "score": 1.0, + "content": ", a reward function", + "type": "text" + }, + { + "bbox": [ + 410, + 342, + 419, + 352 + ], + "score": 0.8, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 340, + 506, + 354 + ], + "score": 1.0, + "content": ", transition dynamics", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 107, + 354, + 115, + 362 + ], + "score": 0.79, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 352, + 203, + 365 + ], + "score": 1.0, + "content": ", and a discount factor", + "type": "text" + }, + { + "bbox": [ + 203, + 353, + 243, + 365 + ], + "score": 0.92, + "content": "\\gamma \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 352, + 283, + 365 + ], + "score": 1.0, + "content": ". A policy", + "type": "text" + }, + { + "bbox": [ + 284, + 353, + 316, + 365 + ], + "score": 0.93, + "content": "\\pi ( \\cdot \\mid x )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 352, + 365, + 365 + ], + "score": 1.0, + "content": "maps states", + "type": "text" + }, + { + "bbox": [ + 365, + 353, + 393, + 363 + ], + "score": 0.91, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "to distributions over actions.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 321, + 376 + ], + "score": 1.0, + "content": "Whenever convenient, we abuse notation and write", + "type": "text" + }, + { + "bbox": [ + 321, + 364, + 342, + 375 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "to describe the probability distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 136, + 387 + ], + "score": 0.92, + "content": "\\pi ( \\cdot | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 374, + 172, + 387 + ], + "score": 1.0, + "content": ", treating", + "type": "text" + }, + { + "bbox": [ + 173, + 375, + 193, + 386 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 374, + 414, + 387 + ], + "score": 1.0, + "content": "as a vector. In RL, the goal is to find an optimal policy", + "type": "text" + }, + { + "bbox": [ + 414, + 375, + 426, + 385 + ], + "score": 0.85, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "that maximizes the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 385, + 467, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 217, + 399 + ], + "score": 1.0, + "content": "cumulative expected return", + "type": "text" + }, + { + "bbox": [ + 217, + 385, + 337, + 398 + ], + "score": 0.94, + "content": "\\mathbb E _ { a _ { t } \\sim \\pi ( \\cdot \\textit { \\textbf { \\ i } } x _ { t } ) } [ \\sum _ { t } \\bar { \\gamma } ^ { t } R ( x _ { t } , a _ { t } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 385, + 451, + 399 + ], + "score": 1.0, + "content": "starting from an initial state", + "type": "text" + }, + { + "bbox": [ + 451, + 387, + 462, + 397 + ], + "score": 0.85, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 385, + 467, + 399 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "We are interested in learning a policy that generalizes across related environments. We formalize this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 216, + 425 + ], + "score": 1.0, + "content": "by considering a collection", + "type": "text" + }, + { + "bbox": [ + 216, + 415, + 223, + 425 + ], + "score": 0.82, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 413, + 360, + 425 + ], + "score": 1.0, + "content": "of MDPs, sharing an action space", + "type": "text" + }, + { + "bbox": [ + 360, + 414, + 369, + 423 + ], + "score": 0.73, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "but with disjoint state spaces. We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 122, + 437 + ], + "score": 1.0, + "content": "use", + "type": "text" + }, + { + "bbox": [ + 123, + 425, + 132, + 434 + ], + "score": 0.82, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 424, + 150, + 437 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 151, + 425, + 160, + 435 + ], + "score": 0.84, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 424, + 410, + 437 + ], + "score": 1.0, + "content": "to denote the state spaces of specific environments, and write", + "type": "text" + }, + { + "bbox": [ + 410, + 425, + 425, + 435 + ], + "score": 0.86, + "content": "R _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 424, + 429, + 437 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 429, + 425, + 444, + 435 + ], + "score": 0.84, + "content": "P _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "for the reward", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 332, + 447 + ], + "score": 1.0, + "content": "and transition functions of the MDP whose state space is", + "type": "text" + }, + { + "bbox": [ + 332, + 436, + 342, + 445 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 436, + 362, + 447 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 362, + 436, + 376, + 447 + ], + "score": 0.9, + "content": "\\pi _ { \\mathcal { X } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 436, + 505, + 447 + ], + "score": 1.0, + "content": "for its optimal policy, which we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 444, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 349, + 462 + ], + "score": 1.0, + "content": "assume unique without loss of generality. For a given policy", + "type": "text" + }, + { + "bbox": [ + 350, + 448, + 357, + 456 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 444, + 488, + 462 + ], + "score": 1.0, + "content": ", we further specialize these into", + "type": "text" + }, + { + "bbox": [ + 488, + 446, + 504, + 458 + ], + "score": 0.89, + "content": "R _ { \\mathcal { X } } ^ { \\pi }", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 456, + 491, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 123, + 469 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 458, + 138, + 469 + ], + "score": 0.89, + "content": "P _ { \\mathcal { X } } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 456, + 429, + 469 + ], + "score": 1.0, + "content": ", the reward and state-to-state transition dynamics arising from following", + "type": "text" + }, + { + "bbox": [ + 430, + 459, + 438, + 467 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 456, + 491, + 469 + ], + "score": 1.0, + "content": "in that MDP.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 143, + 487 + ], + "score": 1.0, + "content": "We write", + "type": "text" + }, + { + "bbox": [ + 144, + 475, + 151, + 484 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 473, + 338, + 487 + ], + "score": 1.0, + "content": "for the union of the state spaces of the MDPs in", + "type": "text" + }, + { + "bbox": [ + 339, + 476, + 345, + 486 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 473, + 505, + 487 + ], + "score": 1.0, + "content": ". Concretely, different MDPs correspond", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 486, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 327, + 497 + ], + "score": 1.0, + "content": "to specific scenarios in a problem class (Figure 1), and", + "type": "text" + }, + { + "bbox": [ + 327, + 486, + 336, + 495 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 486, + 506, + 497 + ], + "score": 1.0, + "content": "is the space of all possible configurations.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 205, + 508 + ], + "score": 1.0, + "content": "Used without subscripts,", + "type": "text" + }, + { + "bbox": [ + 205, + 496, + 227, + 507 + ], + "score": 0.58, + "content": "R , P", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 495, + 246, + 508 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 246, + 498, + 254, + 506 + ], + "score": 0.78, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "refer to the reward and transition function of this “union MDP”,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 507, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 210, + 519 + ], + "score": 1.0, + "content": "and a policy defined over", + "type": "text" + }, + { + "bbox": [ + 210, + 507, + 218, + 517 + ], + "score": 0.79, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 507, + 506, + 519 + ], + "score": 1.0, + "content": "; this notation simplifies the exposition. We measure distances between", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 518, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 318, + 530 + ], + "score": 1.0, + "content": "states across environments using pseudometrics2 on", + "type": "text" + }, + { + "bbox": [ + 318, + 518, + 326, + 528 + ], + "score": 0.79, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 518, + 473, + 530 + ], + "score": 1.0, + "content": "; the set of all such pseudometrics is", + "type": "text" + }, + { + "bbox": [ + 473, + 518, + 484, + 528 + ], + "score": 0.44, + "content": "\\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 518, + 506, + 530 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 528, + 345, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 122, + 541 + ], + "score": 0.89, + "content": "\\mathbb { M } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 528, + 333, + 542 + ], + "score": 1.0, + "content": "is the set of metrics on probability distributions over", + "type": "text" + }, + { + "bbox": [ + 334, + 529, + 342, + 539 + ], + "score": 0.78, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 528, + 345, + 542 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 104, + 545, + 507, + 562 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 380, + 562 + ], + "score": 1.0, + "content": "In our setting, the learner has access to a collection of training MDPs", + "type": "text" + }, + { + "bbox": [ + 380, + 546, + 419, + 559 + ], + "score": 0.93, + "content": "\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 545, + 471, + 562 + ], + "score": 1.0, + "content": ", drawn from", + "type": "text" + }, + { + "bbox": [ + 471, + 549, + 478, + 558 + ], + "score": 0.72, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 545, + 507, + 562 + ], + "score": 1.0, + "content": ". After", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 559, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 383, + 570 + ], + "score": 1.0, + "content": "interacting with these environments, the learner must produce a policy", + "type": "text" + }, + { + "bbox": [ + 383, + 561, + 390, + 568 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 559, + 494, + 570 + ], + "score": 1.0, + "content": "over the entire state space", + "type": "text" + }, + { + "bbox": [ + 495, + 559, + 502, + 568 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 559, + 506, + 570 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 567, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 295, + 583 + ], + "score": 1.0, + "content": "which is then evaluated on unseen MDPs from", + "type": "text" + }, + { + "bbox": [ + 295, + 571, + 302, + 581 + ], + "score": 0.77, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 567, + 505, + 583 + ], + "score": 1.0, + "content": ". Similar in spirit to the setting of transfer learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 437, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 426, + 594 + ], + "score": 1.0, + "content": "(Taylor & Stone, 2009), here we evaluate the policy’s zero-shot performance on", + "type": "text" + }, + { + "bbox": [ + 426, + 582, + 433, + 592 + ], + "score": 0.8, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 579, + 437, + 594 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 507, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 375, + 609 + ], + "score": 1.0, + "content": "Our policy similarity metric (Section 3) builds on the concept of", + "type": "text" + }, + { + "bbox": [ + 376, + 599, + 384, + 607 + ], + "score": 0.81, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 596, + 507, + 609 + ], + "score": 1.0, + "content": "-bisimulation (Castro, 2020).", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 150, + 620 + ], + "score": 1.0, + "content": "Under the", + "type": "text" + }, + { + "bbox": [ + 150, + 609, + 157, + 618 + ], + "score": 0.79, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 607, + 378, + 620 + ], + "score": 1.0, + "content": "-bisimulation metric, the distance between two states,", + "type": "text" + }, + { + "bbox": [ + 379, + 610, + 386, + 618 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 607, + 404, + 620 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 610, + 411, + 619 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 607, + 506, + 620 + ], + "score": 1.0, + "content": ", is defined in terms of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 619, + 504, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 414, + 631 + ], + "score": 1.0, + "content": "the difference between the expected rewards obtained when following policy", + "type": "text" + }, + { + "bbox": [ + 414, + 621, + 421, + 629 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 619, + 443, + 631 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 443, + 620, + 451, + 629 + ], + "score": 0.8, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 619, + 504, + 631 + ], + "score": 1.0, + "content": "-bisimulation", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 135, + 642 + ], + "score": 1.0, + "content": "metric", + "type": "text" + }, + { + "bbox": [ + 135, + 630, + 147, + 641 + ], + "score": 0.88, + "content": "d _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 629, + 410, + 642 + ], + "score": 1.0, + "content": "satisfies a recursive equation based on the 1-Wasserstein metric", + "type": "text" + }, + { + "bbox": [ + 410, + 630, + 474, + 642 + ], + "score": 0.92, + "content": "\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 629, + 505, + 642 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 107, + 641, + 162, + 653 + ], + "score": 0.92, + "content": "\\mathcal { W } _ { 1 } ( d ) ( A , B )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 639, + 404, + 654 + ], + "score": 1.0, + "content": "is the minimal cost of transporting probability mass from", + "type": "text" + }, + { + "bbox": [ + 404, + 641, + 413, + 651 + ], + "score": 0.77, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 639, + 425, + 654 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 426, + 642, + 435, + 651 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 639, + 505, + 654 + ], + "score": 1.0, + "content": "(two probability", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 651, + 411, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 171, + 664 + ], + "score": 1.0, + "content": "distributions on", + "type": "text" + }, + { + "bbox": [ + 171, + 653, + 179, + 662 + ], + "score": 0.78, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 651, + 271, + 664 + ], + "score": 1.0, + "content": ") under the base metric", + "type": "text" + }, + { + "bbox": [ + 272, + 652, + 278, + 662 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 651, + 411, + 664 + ], + "score": 1.0, + "content": "(Villani, 2008). The recursion is", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + }, + { + "type": "interline_equation", + "bbox": [ + 146, + 668, + 465, + 683 + ], + "lines": [ + { + "bbox": [ + 146, + 668, + 465, + 683 + ], + "spans": [ + { + "bbox": [ + 146, + 668, + 465, + 683 + ], + "score": 0.89, + "content": "\\begin{array} { r } { d _ { \\pi } ( x , y ) = | R ^ { \\pi } ( x ) - R ^ { \\pi } ( y ) | + \\gamma \\mathcal { W } _ { 1 } ( d _ { \\pi } ) \\big ( P ^ { \\pi } ( \\cdot | x ) , P ^ { \\pi } ( \\cdot | y ) \\big ) , \\qquad x , y \\in \\mathcal { S } . } \\end{array}", + "type": "interline_equation", + "image_path": "620aa4a43bc93eafeecfb704f3618718837319bec60a3d823ad690a9332d2a62.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 146, + 668, + 465, + 683 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 690, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 424, + 703 + ], + "score": 1.0, + "content": "To achieve good generalization properties, we learn an embedding function", + "type": "text" + }, + { + "bbox": [ + 424, + 689, + 485, + 701 + ], + "score": 0.92, + "content": "z _ { \\theta } : \\mathcal { S } \\mathbb { R } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 689, + 506, + 703 + ], + "score": 1.0, + "content": "that", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 700, + 504, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 714 + ], + "score": 1.0, + "content": "reflects the information encoded in the policy similarity metric; this yields a policy similarity", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 114, + 721, + 494, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 495, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 495, + 733 + ], + "score": 1.0, + "content": "2Pseudometrics are generalization of metrics where the distance between two distinct states can be zero.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 506, + 105 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "We employ PSM for representation learning and introduce policy similarity embeddings (PSEs) for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "deep RL. To do so, we present a general contrastive procedure (Section 4) to learn an embedding", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "based on any state similarity metric. PSEs are the instantiation of this procedure with PSM. PSEs", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 504, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 504, + 156 + ], + "score": 1.0, + "content": "are appealing for generalization as they encode task-relevant invariances by putting behaviorally", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "equivalent states together. This is unlike prior approaches, which rely on capturing such invariances", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "without being explicitly trained to do so, for example, through value function similarities across", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 188 + ], + "score": 1.0, + "content": "states (e.g., Castro & Precup, 2010), or being robust to fixed transformations of the observation", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 331, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 331, + 200 + ], + "score": 1.0, + "content": "space (e.g., Kostrikov et al., 2020; Laskin et al., 2020a).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 110, + 506, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 505, + 217 + ], + "score": 1.0, + "content": "PSEs lead to better generalization while being orthogonal to how most of the field has been tackling", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "generalization. We illustrate the efficacy and broad applicability of our approach on three existing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 226, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 506, + 239 + ], + "score": 1.0, + "content": "benchmarks specifically designed to test generalization: (i) jumping task from pixels (Tachet des", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "Combes et al., 2018) (Section 5), (ii) LQR with spurious correlations (Song et al., 2019) (Section 6.1),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "and (iii) Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b) (Section 6.2). Our", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "approach improves generalization compared to a wide variety of approaches including standard", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 282 + ], + "score": 1.0, + "content": "regularization (Farebrother et al., 2018; Cobbe et al., 2019), bisimulation (Castro & Precup, 2010;", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "Castro, 2020; Zhang et al., 2021), out-of-distribution generalization (Arjovsky et al., 2019) and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 412, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 412, + 303 + ], + "score": 1.0, + "content": "state-of-the-art data augmentation (Kostrikov et al., 2020; Lee et al., 2020a).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 202, + 506, + 303 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 311, + 208, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 209, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 209, + 326 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 476, + 343 + ], + "score": 1.0, + "content": "We describe an environment as a Markov decision process (MDP) (Puterman, 1994)", + "type": "text" + }, + { + "bbox": [ + 476, + 331, + 505, + 342 + ], + "score": 0.83, + "content": "\\begin{array} { r l } { \\mathcal { M } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 168, + 354 + ], + "score": 0.91, + "content": "( \\mathcal { X } , \\mathcal { A } , R , P , \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 340, + 245, + 354 + ], + "score": 1.0, + "content": ", with a state space", + "type": "text" + }, + { + "bbox": [ + 246, + 342, + 255, + 352 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 340, + 324, + 354 + ], + "score": 1.0, + "content": ", an action space", + "type": "text" + }, + { + "bbox": [ + 324, + 342, + 333, + 352 + ], + "score": 0.79, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 340, + 410, + 354 + ], + "score": 1.0, + "content": ", a reward function", + "type": "text" + }, + { + "bbox": [ + 410, + 342, + 419, + 352 + ], + "score": 0.8, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 340, + 506, + 354 + ], + "score": 1.0, + "content": ", transition dynamics", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 107, + 354, + 115, + 362 + ], + "score": 0.79, + "content": "P", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 352, + 203, + 365 + ], + "score": 1.0, + "content": ", and a discount factor", + "type": "text" + }, + { + "bbox": [ + 203, + 353, + 243, + 365 + ], + "score": 0.92, + "content": "\\gamma \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 352, + 283, + 365 + ], + "score": 1.0, + "content": ". A policy", + "type": "text" + }, + { + "bbox": [ + 284, + 353, + 316, + 365 + ], + "score": 0.93, + "content": "\\pi ( \\cdot \\mid x )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 352, + 365, + 365 + ], + "score": 1.0, + "content": "maps states", + "type": "text" + }, + { + "bbox": [ + 365, + 353, + 393, + 363 + ], + "score": 0.91, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "to distributions over actions.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 321, + 376 + ], + "score": 1.0, + "content": "Whenever convenient, we abuse notation and write", + "type": "text" + }, + { + "bbox": [ + 321, + 364, + 342, + 375 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "to describe the probability distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 136, + 387 + ], + "score": 0.92, + "content": "\\pi ( \\cdot | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 374, + 172, + 387 + ], + "score": 1.0, + "content": ", treating", + "type": "text" + }, + { + "bbox": [ + 173, + 375, + 193, + 386 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 374, + 414, + 387 + ], + "score": 1.0, + "content": "as a vector. In RL, the goal is to find an optimal policy", + "type": "text" + }, + { + "bbox": [ + 414, + 375, + 426, + 385 + ], + "score": 0.85, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "that maximizes the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 385, + 467, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 217, + 399 + ], + "score": 1.0, + "content": "cumulative expected return", + "type": "text" + }, + { + "bbox": [ + 217, + 385, + 337, + 398 + ], + "score": 0.94, + "content": "\\mathbb E _ { a _ { t } \\sim \\pi ( \\cdot \\textit { \\textbf { \\ i } } x _ { t } ) } [ \\sum _ { t } \\bar { \\gamma } ^ { t } R ( x _ { t } , a _ { t } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 385, + 451, + 399 + ], + "score": 1.0, + "content": "starting from an initial state", + "type": "text" + }, + { + "bbox": [ + 451, + 387, + 462, + 397 + ], + "score": 0.85, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 385, + 467, + 399 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 330, + 506, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "We are interested in learning a policy that generalizes across related environments. We formalize this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 216, + 425 + ], + "score": 1.0, + "content": "by considering a collection", + "type": "text" + }, + { + "bbox": [ + 216, + 415, + 223, + 425 + ], + "score": 0.82, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 413, + 360, + 425 + ], + "score": 1.0, + "content": "of MDPs, sharing an action space", + "type": "text" + }, + { + "bbox": [ + 360, + 414, + 369, + 423 + ], + "score": 0.73, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "but with disjoint state spaces. We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 122, + 437 + ], + "score": 1.0, + "content": "use", + "type": "text" + }, + { + "bbox": [ + 123, + 425, + 132, + 434 + ], + "score": 0.82, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 424, + 150, + 437 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 151, + 425, + 160, + 435 + ], + "score": 0.84, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 424, + 410, + 437 + ], + "score": 1.0, + "content": "to denote the state spaces of specific environments, and write", + "type": "text" + }, + { + "bbox": [ + 410, + 425, + 425, + 435 + ], + "score": 0.86, + "content": "R _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 424, + 429, + 437 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 429, + 425, + 444, + 435 + ], + "score": 0.84, + "content": "P _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "for the reward", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 436, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 332, + 447 + ], + "score": 1.0, + "content": "and transition functions of the MDP whose state space is", + "type": "text" + }, + { + "bbox": [ + 332, + 436, + 342, + 445 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 436, + 362, + 447 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 362, + 436, + 376, + 447 + ], + "score": 0.9, + "content": "\\pi _ { \\mathcal { X } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 436, + 505, + 447 + ], + "score": 1.0, + "content": "for its optimal policy, which we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 444, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 444, + 349, + 462 + ], + "score": 1.0, + "content": "assume unique without loss of generality. For a given policy", + "type": "text" + }, + { + "bbox": [ + 350, + 448, + 357, + 456 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 444, + 488, + 462 + ], + "score": 1.0, + "content": ", we further specialize these into", + "type": "text" + }, + { + "bbox": [ + 488, + 446, + 504, + 458 + ], + "score": 0.89, + "content": "R _ { \\mathcal { X } } ^ { \\pi }", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 456, + 491, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 123, + 469 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 458, + 138, + 469 + ], + "score": 0.89, + "content": "P _ { \\mathcal { X } } ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 456, + 429, + 469 + ], + "score": 1.0, + "content": ", the reward and state-to-state transition dynamics arising from following", + "type": "text" + }, + { + "bbox": [ + 430, + 459, + 438, + 467 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 456, + 491, + 469 + ], + "score": 1.0, + "content": "in that MDP.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 401, + 505, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 143, + 487 + ], + "score": 1.0, + "content": "We write", + "type": "text" + }, + { + "bbox": [ + 144, + 475, + 151, + 484 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 473, + 338, + 487 + ], + "score": 1.0, + "content": "for the union of the state spaces of the MDPs in", + "type": "text" + }, + { + "bbox": [ + 339, + 476, + 345, + 486 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 473, + 505, + 487 + ], + "score": 1.0, + "content": ". Concretely, different MDPs correspond", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 486, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 327, + 497 + ], + "score": 1.0, + "content": "to specific scenarios in a problem class (Figure 1), and", + "type": "text" + }, + { + "bbox": [ + 327, + 486, + 336, + 495 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 486, + 506, + 497 + ], + "score": 1.0, + "content": "is the space of all possible configurations.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 205, + 508 + ], + "score": 1.0, + "content": "Used without subscripts,", + "type": "text" + }, + { + "bbox": [ + 205, + 496, + 227, + 507 + ], + "score": 0.58, + "content": "R , P", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 495, + 246, + 508 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 246, + 498, + 254, + 506 + ], + "score": 0.78, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "refer to the reward and transition function of this “union MDP”,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 507, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 210, + 519 + ], + "score": 1.0, + "content": "and a policy defined over", + "type": "text" + }, + { + "bbox": [ + 210, + 507, + 218, + 517 + ], + "score": 0.79, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 507, + 506, + 519 + ], + "score": 1.0, + "content": "; this notation simplifies the exposition. We measure distances between", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 518, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 318, + 530 + ], + "score": 1.0, + "content": "states across environments using pseudometrics2 on", + "type": "text" + }, + { + "bbox": [ + 318, + 518, + 326, + 528 + ], + "score": 0.79, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 518, + 473, + 530 + ], + "score": 1.0, + "content": "; the set of all such pseudometrics is", + "type": "text" + }, + { + "bbox": [ + 473, + 518, + 484, + 528 + ], + "score": 0.44, + "content": "\\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 518, + 506, + 530 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 528, + 345, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 122, + 541 + ], + "score": 0.89, + "content": "\\mathbb { M } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 528, + 333, + 542 + ], + "score": 1.0, + "content": "is the set of metrics on probability distributions over", + "type": "text" + }, + { + "bbox": [ + 334, + 529, + 342, + 539 + ], + "score": 0.78, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 528, + 345, + 542 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 473, + 506, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 104, + 545, + 507, + 562 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 380, + 562 + ], + "score": 1.0, + "content": "In our setting, the learner has access to a collection of training MDPs", + "type": "text" + }, + { + "bbox": [ + 380, + 546, + 419, + 559 + ], + "score": 0.93, + "content": "\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 545, + 471, + 562 + ], + "score": 1.0, + "content": ", drawn from", + "type": "text" + }, + { + "bbox": [ + 471, + 549, + 478, + 558 + ], + "score": 0.72, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 545, + 507, + 562 + ], + "score": 1.0, + "content": ". After", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 559, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 383, + 570 + ], + "score": 1.0, + "content": "interacting with these environments, the learner must produce a policy", + "type": "text" + }, + { + "bbox": [ + 383, + 561, + 390, + 568 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 559, + 494, + 570 + ], + "score": 1.0, + "content": "over the entire state space", + "type": "text" + }, + { + "bbox": [ + 495, + 559, + 502, + 568 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 559, + 506, + 570 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 567, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 295, + 583 + ], + "score": 1.0, + "content": "which is then evaluated on unseen MDPs from", + "type": "text" + }, + { + "bbox": [ + 295, + 571, + 302, + 581 + ], + "score": 0.77, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 567, + 505, + 583 + ], + "score": 1.0, + "content": ". Similar in spirit to the setting of transfer learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 437, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 426, + 594 + ], + "score": 1.0, + "content": "(Taylor & Stone, 2009), here we evaluate the policy’s zero-shot performance on", + "type": "text" + }, + { + "bbox": [ + 426, + 582, + 433, + 592 + ], + "score": 0.8, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 579, + 437, + 594 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 545, + 507, + 594 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 507, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 375, + 609 + ], + "score": 1.0, + "content": "Our policy similarity metric (Section 3) builds on the concept of", + "type": "text" + }, + { + "bbox": [ + 376, + 599, + 384, + 607 + ], + "score": 0.81, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 596, + 507, + 609 + ], + "score": 1.0, + "content": "-bisimulation (Castro, 2020).", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 150, + 620 + ], + "score": 1.0, + "content": "Under the", + "type": "text" + }, + { + "bbox": [ + 150, + 609, + 157, + 618 + ], + "score": 0.79, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 607, + 378, + 620 + ], + "score": 1.0, + "content": "-bisimulation metric, the distance between two states,", + "type": "text" + }, + { + "bbox": [ + 379, + 610, + 386, + 618 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 607, + 404, + 620 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 610, + 411, + 619 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 607, + 506, + 620 + ], + "score": 1.0, + "content": ", is defined in terms of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 619, + 504, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 414, + 631 + ], + "score": 1.0, + "content": "the difference between the expected rewards obtained when following policy", + "type": "text" + }, + { + "bbox": [ + 414, + 621, + 421, + 629 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 619, + 443, + 631 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 443, + 620, + 451, + 629 + ], + "score": 0.8, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 619, + 504, + 631 + ], + "score": 1.0, + "content": "-bisimulation", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 135, + 642 + ], + "score": 1.0, + "content": "metric", + "type": "text" + }, + { + "bbox": [ + 135, + 630, + 147, + 641 + ], + "score": 0.88, + "content": "d _ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 629, + 410, + 642 + ], + "score": 1.0, + "content": "satisfies a recursive equation based on the 1-Wasserstein metric", + "type": "text" + }, + { + "bbox": [ + 410, + 630, + 474, + 642 + ], + "score": 0.92, + "content": "\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 629, + 505, + 642 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 107, + 641, + 162, + 653 + ], + "score": 0.92, + "content": "\\mathcal { W } _ { 1 } ( d ) ( A , B )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 639, + 404, + 654 + ], + "score": 1.0, + "content": "is the minimal cost of transporting probability mass from", + "type": "text" + }, + { + "bbox": [ + 404, + 641, + 413, + 651 + ], + "score": 0.77, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 639, + 425, + 654 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 426, + 642, + 435, + 651 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 639, + 505, + 654 + ], + "score": 1.0, + "content": "(two probability", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 651, + 411, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 171, + 664 + ], + "score": 1.0, + "content": "distributions on", + "type": "text" + }, + { + "bbox": [ + 171, + 653, + 179, + 662 + ], + "score": 0.78, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 651, + 271, + 664 + ], + "score": 1.0, + "content": ") under the base metric", + "type": "text" + }, + { + "bbox": [ + 272, + 652, + 278, + 662 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 651, + 411, + 664 + ], + "score": 1.0, + "content": "(Villani, 2008). The recursion is", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 596, + 507, + 664 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 146, + 668, + 465, + 683 + ], + "lines": [ + { + "bbox": [ + 146, + 668, + 465, + 683 + ], + "spans": [ + { + "bbox": [ + 146, + 668, + 465, + 683 + ], + "score": 0.89, + "content": "\\begin{array} { r } { d _ { \\pi } ( x , y ) = | R ^ { \\pi } ( x ) - R ^ { \\pi } ( y ) | + \\gamma \\mathcal { W } _ { 1 } ( d _ { \\pi } ) \\big ( P ^ { \\pi } ( \\cdot | x ) , P ^ { \\pi } ( \\cdot | y ) \\big ) , \\qquad x , y \\in \\mathcal { S } . } \\end{array}", + "type": "interline_equation", + "image_path": "620aa4a43bc93eafeecfb704f3618718837319bec60a3d823ad690a9332d2a62.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 146, + 668, + 465, + 683 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 690, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 424, + 703 + ], + "score": 1.0, + "content": "To achieve good generalization properties, we learn an embedding function", + "type": "text" + }, + { + "bbox": [ + 424, + 689, + 485, + 701 + ], + "score": 0.92, + "content": "z _ { \\theta } : \\mathcal { S } \\mathbb { R } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 689, + 506, + 703 + ], + "score": 1.0, + "content": "that", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 700, + 504, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 714 + ], + "score": 1.0, + "content": "reflects the information encoded in the policy similarity metric; this yields a policy similarity", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "embedding (Section 4). We use contrastive methods (Hadsell et al., 2006; Oord et al., 2018) whose", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 507, + 106 + ], + "score": 1.0, + "content": "track record for representation learning is well-established. We adapt SimCLR (Chen et al., 2020),", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 470, + 118 + ], + "score": 1.0, + "content": "a popular contrastive method for learning embeddings of image inputs. Given two inputs", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 471, + 106, + 477, + 114 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 478, + 104, + 496, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 496, + 107, + 502, + 116 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 503, + 104, + 507, + 118 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 223, + 135 + ], + "score": 1.0, + "content": "their embedding similarity is", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 223, + 117, + 343, + 130 + ], + "score": 0.93, + "content": "s _ { \\theta } ( x , y ) = s i m ( z _ { \\theta } ( x ) , z _ { \\theta } ( y ) )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 344, + 114, + 374, + 135 + ], + "score": 1.0, + "content": ", where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 375, + 115, + 456, + 133 + ], + "score": 0.92, + "content": "\\begin{array} { r } { s i m ( u , v ) = \\frac { u ^ { T } v } { \\| u \\| \\| v \\| } } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 456, + 114, + 506, + 135 + ], + "score": 1.0, + "content": "denotes the", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "score": 1.0, + "content": "cosine similarity function. SimCLR aims to maximize similarity between augmented versions of an", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "image (e.g., crops, colour changes) while minimizing its similarity to other images. The loss used by", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 443, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 211, + 165 + ], + "score": 1.0, + "content": "SimCLR for two versions", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 212, + 155, + 228, + 164 + ], + "score": 0.87, + "content": "x , y", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 228, + 152, + 318, + 165 + ], + "score": 1.0, + "content": "of an image, and a set", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 319, + 153, + 331, + 163 + ], + "score": 0.88, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 331, + 152, + 443, + 165 + ], + "score": 1.0, + "content": "containing other images is:", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 49.5, + "bbox_fs": [ + 106, + 689, + 506, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "embedding (Section 4). We use contrastive methods (Hadsell et al., 2006; Oord et al., 2018) whose", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 507, + 106 + ], + "score": 1.0, + "content": "track record for representation learning is well-established. We adapt SimCLR (Chen et al., 2020),", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 470, + 118 + ], + "score": 1.0, + "content": "a popular contrastive method for learning embeddings of image inputs. Given two inputs", + "type": "text" + }, + { + "bbox": [ + 471, + 106, + 477, + 114 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 104, + 496, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 496, + 107, + 502, + 116 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 104, + 507, + 118 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 223, + 135 + ], + "score": 1.0, + "content": "their embedding similarity is", + "type": "text" + }, + { + "bbox": [ + 223, + 117, + 343, + 130 + ], + "score": 0.93, + "content": "s _ { \\theta } ( x , y ) = s i m ( z _ { \\theta } ( x ) , z _ { \\theta } ( y ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 114, + 374, + 135 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 375, + 115, + 456, + 133 + ], + "score": 0.92, + "content": "\\begin{array} { r } { s i m ( u , v ) = \\frac { u ^ { T } v } { \\| u \\| \\| v \\| } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 114, + 506, + 135 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "score": 1.0, + "content": "cosine similarity function. SimCLR aims to maximize similarity between augmented versions of an", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 154 + ], + "score": 1.0, + "content": "image (e.g., crops, colour changes) while minimizing its similarity to other images. The loss used by", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 443, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 211, + 165 + ], + "score": 1.0, + "content": "SimCLR for two versions", + "type": "text" + }, + { + "bbox": [ + 212, + 155, + 228, + 164 + ], + "score": 0.87, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 152, + 318, + 165 + ], + "score": 1.0, + "content": "of an image, and a set", + "type": "text" + }, + { + "bbox": [ + 319, + 153, + 331, + 163 + ], + "score": 0.88, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 152, + 443, + 165 + ], + "score": 1.0, + "content": "containing other images is:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 168, + 444, + 198 + ], + "lines": [ + { + "bbox": [ + 166, + 168, + 444, + 198 + ], + "spans": [ + { + "bbox": [ + 166, + 168, + 444, + 198 + ], + "score": 0.93, + "content": "\\ell _ { \\theta } ( x , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) } { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) + \\sum _ { x ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\left\\{ x \\right\\} } \\exp ( \\lambda s _ { \\theta } ( x ^ { \\prime } , y ) ) }", + "type": "interline_equation", + "image_path": "e728c707965bd848cfff2b88e2d41e2810b434960e667d469684e6e39308eb85.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 166, + 168, + 444, + 178.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 166, + 178.0, + 444, + 188.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 166, + 188.0, + 444, + 198.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 203, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 133, + 216 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 204, + 141, + 213 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "is an inverse temperature hyperparameter. The overall SimCLR loss is then the expected", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 141, + 227 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 141, + 214, + 185, + 226 + ], + "score": 0.93, + "content": "\\ell _ { \\theta } ( x , y ; S )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 214, + 212, + 227 + ], + "score": 1.0, + "content": ", when", + "type": "text" + }, + { + "bbox": [ + 213, + 216, + 230, + 226 + ], + "score": 0.59, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 214, + 250, + 227 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 251, + 214, + 263, + 224 + ], + "score": 0.87, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 214, + 482, + 227 + ], + "score": 1.0, + "content": "are drawn from some augmented training distribution.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 232, + 276, + 246 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 277, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 277, + 248 + ], + "score": 1.0, + "content": "3 POLICY SIMILARITY METRIC", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 506, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "A useful tool in learning a policy that generalizes is to understand which states result in similar", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "behavior, and which do not. To be maximally effective, this similarity should go beyond the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 442, + 287 + ], + "score": 1.0, + "content": "immediately chosen action and consider long-term behavior. In this regards, the", + "type": "text" + }, + { + "bbox": [ + 442, + 276, + 450, + 285 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "-bisimulation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 286, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 297 + ], + "score": 1.0, + "content": "metrics are interesting as they are based on the full sequence of future rewards received from different", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 297, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 308 + ], + "score": 1.0, + "content": "states. However, considering rewards can be both too restrictive (when the policies are the same, but", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "the obtained rewards are not; see Figure D.1) or too permissive (when the policies are different, but", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 323, + 331 + ], + "score": 1.0, + "content": "the obtained rewards are not; see Figure 3a). In fact,", + "type": "text" + }, + { + "bbox": [ + 324, + 320, + 330, + 329 + ], + "score": 0.79, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "-bisimulation metrics actually lead to poor", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 330, + 334, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 334, + 341 + ], + "score": 1.0, + "content": "generalization in our experiments (Sections 5.1 and 5.2).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 506, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "To address this issue, we instead consider the similarity between policies themselves. We replace", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "the absolute reward difference by a probability pseudometric between policies, denoted DIST. Addi-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "tionally, since we would like to perform well in unseen environments, we are interested in similarity", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 249, + 392 + ], + "score": 1.0, + "content": "in optimal behavior. We thus use", + "type": "text" + }, + { + "bbox": [ + 250, + 379, + 262, + 389 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "as the grounding policy. This yields the policy similarity", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "metric (PSM), for which states are close when the optimal policies in these states and future states", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 401, + 457, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 264, + 414 + ], + "score": 1.0, + "content": "are similar. For a given DIST, the PSM", + "type": "text" + }, + { + "bbox": [ + 265, + 401, + 332, + 411 + ], + "score": 0.9, + "content": "d ^ { * } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 401, + 457, + 414 + ], + "score": 1.0, + "content": "satisfies the recursive equation", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 416, + 444, + 451 + ], + "lines": [ + { + "bbox": [ + 167, + 416, + 444, + 451 + ], + "spans": [ + { + "bbox": [ + 167, + 416, + 444, + 451 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } & { } & { d ^ { * } ( x , y ) = \\underbrace { \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) } _ { \\mathrm { ( A ) } } + \\underbrace { \\gamma \\mathcal { W } _ { 1 } ( d ^ { * } ) \\big ( P ^ { \\pi ^ { * } } ( \\cdot | x ) , P ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) } _ { \\mathrm { ( B ) } } . } \\end{array}", + "type": "interline_equation", + "image_path": "9ac02354c480f08d23c9d5364cc9a542168e6e21775e16792bd3b5b35a59511a.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 167, + 416, + 444, + 427.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 167, + 427.6666666666667, + 444, + 439.33333333333337 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 167, + 439.33333333333337, + 444, + 451.00000000000006 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 504, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 409, + 468 + ], + "score": 1.0, + "content": "The DIST term captures the difference in local optimal behavior (A) while", + "type": "text" + }, + { + "bbox": [ + 409, + 456, + 425, + 467 + ], + "score": 0.89, + "content": "\\mathcal { W } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 455, + 504, + 468 + ], + "score": 1.0, + "content": "captures long-term", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "optimal behavior difference (B); the exact weights assigned to the two are given by the discount", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 283, + 490 + ], + "score": 1.0, + "content": "factor. Furthermore, when DIST is bounded,", + "type": "text" + }, + { + "bbox": [ + 283, + 478, + 294, + 488 + ], + "score": 0.86, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "is guaranteed to be finite. While there are technically", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 488, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 506, + 500 + ], + "score": 1.0, + "content": "multiple PSMs (one for each DIST), we omit this distinction whenever clear from context. A proof of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 500, + 301, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 178, + 511 + ], + "score": 1.0, + "content": "the uniqueness of", + "type": "text" + }, + { + "bbox": [ + 179, + 500, + 190, + 509 + ], + "score": 0.87, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 500, + 301, + 511 + ], + "score": 1.0, + "content": "is given in Proposition B.1.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "Our main use of PSM will be to compare states across environments. In this context, we identify the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "terms in Equation 3 with specific environments for clarity and write (despite its technical inaccuracy)", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 543, + 445, + 560 + ], + "lines": [ + { + "bbox": [ + 165, + 543, + 445, + 560 + ], + "spans": [ + { + "bbox": [ + 165, + 543, + 445, + 560 + ], + "score": 0.9, + "content": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ef10847f7eac97b0d10c96ae00eaa9d1f6547c2a192ff9eb67cf97571f068e91.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 165, + 543, + 445, + 560 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "score": 1.0, + "content": "PSM is applicable to both discrete and continuous action spaces. In our experiments, DIST is the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 204, + 588 + ], + "score": 1.0, + "content": "total variation distance", + "type": "text" + }, + { + "bbox": [ + 204, + 576, + 227, + 587 + ], + "score": 0.8, + "content": "( T V )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 574, + 254, + 588 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 254, + 576, + 264, + 586 + ], + "score": 0.81, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 574, + 378, + 588 + ], + "score": 1.0, + "content": "is discrete, and we use the", + "type": "text" + }, + { + "bbox": [ + 379, + 576, + 389, + 586 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 574, + 505, + 588 + ], + "score": 1.0, + "content": "distance between the mean", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 587, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 238, + 599 + ], + "score": 1.0, + "content": "actions of the two policies when", + "type": "text" + }, + { + "bbox": [ + 239, + 587, + 248, + 596 + ], + "score": 0.79, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 587, + 504, + 599 + ], + "score": 1.0, + "content": "is continuous. PSM can be computed iteratively using dynamic", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 435, + 609 + ], + "score": 1.0, + "content": "programming (Ferns et al., 2011) (more details in Section D.1). Furthermore, when", + "type": "text" + }, + { + "bbox": [ + 436, + 598, + 447, + 608 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "is unavailable", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 342, + 621 + ], + "score": 1.0, + "content": "on training environments, we replace it by an approximation", + "type": "text" + }, + { + "bbox": [ + 343, + 609, + 354, + 619 + ], + "score": 0.88, + "content": "\\hat { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "to obtain an approximate PSM, which", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 436, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 348, + 632 + ], + "score": 1.0, + "content": "is close to the exact PSM depending on the suboptimality of", + "type": "text" + }, + { + "bbox": [ + 349, + 620, + 360, + 630 + ], + "score": 0.86, + "content": "\\hat { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 619, + 436, + 632 + ], + "score": 1.0, + "content": "(Proposition D.3).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 184, + 648 + ], + "score": 1.0, + "content": "Despite resembling", + "type": "text" + }, + { + "bbox": [ + 184, + 638, + 191, + 646 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "-bisimulation metrics in form, the PSM possesses different characteristics which", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 647, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 505, + 660 + ], + "score": 1.0, + "content": "are better suited to the problem of generalizing a learned policy. To illustrate this point, consider the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 656, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 343, + 672 + ], + "score": 1.0, + "content": "following simple nearest-neighbour scheme: Given a state", + "type": "text" + }, + { + "bbox": [ + 343, + 659, + 369, + 669 + ], + "score": 0.92, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 656, + 482, + 672 + ], + "score": 1.0, + "content": ", denote its closest match in", + "type": "text" + }, + { + "bbox": [ + 482, + 659, + 492, + 668 + ], + "score": 0.8, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 656, + 506, + 672 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 668, + 507, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 219, + 681 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\tilde { x } _ { y } : = \\arg \\operatorname* { m i n } _ { x \\in \\mathcal { X } } d ^ { * } ( x , y ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 668, + 405, + 685 + ], + "score": 1.0, + "content": ". Suppose that we use this scheme to transfer", + "type": "text" + }, + { + "bbox": [ + 405, + 669, + 419, + 681 + ], + "score": 0.9, + "content": "\\pi _ { \\mathcal { X } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 668, + 431, + 685 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 431, + 669, + 451, + 681 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 668, + 507, + 685 + ], + "score": 1.0, + "content": ", in the sense", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 679, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 259, + 694 + ], + "score": 1.0, + "content": "that we behave according to the policy", + "type": "text" + }, + { + "bbox": [ + 259, + 680, + 324, + 692 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ( y ) = \\pi _ { \\chi } ^ { * } ( \\tilde { x } _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 679, + 496, + 694 + ], + "score": 1.0, + "content": ". We can then bound the difference between", + "type": "text" + }, + { + "bbox": [ + 496, + 681, + 504, + 690 + ], + "score": 0.79, + "content": "\\tilde { \\pi }", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 690, + 440, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 123, + 705 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 691, + 136, + 704 + ], + "score": 0.9, + "content": "\\pi _ { \\mathcal { Y } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 690, + 270, + 705 + ], + "score": 1.0, + "content": ", something that is not possible if", + "type": "text" + }, + { + "bbox": [ + 270, + 692, + 280, + 700 + ], + "score": 0.87, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 690, + 347, + 705 + ], + "score": 1.0, + "content": "is replaced by a", + "type": "text" + }, + { + "bbox": [ + 347, + 693, + 354, + 701 + ], + "score": 0.8, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 690, + 440, + 705 + ], + "score": 1.0, + "content": "-bisimulation metric.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 106, + 706, + 502, + 734 + ], + "lines": [ + { + "bbox": [ + 104, + 704, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 704, + 303, + 725 + ], + "score": 1.0, + "content": "Theorem 1. [Bound on policy transfer] For any", + "type": "text" + }, + { + "bbox": [ + 304, + 708, + 329, + 719 + ], + "score": 0.87, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 704, + 346, + 725 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 346, + 707, + 423, + 721 + ], + "score": 0.92, + "content": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 704, + 506, + 725 + ], + "score": 1.0, + "content": "define the sequence", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 716, + 440, + 737 + ], + "spans": [ + { + "bbox": [ + 104, + 716, + 270, + 737 + ], + "score": 1.0, + "content": "of random states encountered starting in", + "type": "text" + }, + { + "bbox": [ + 271, + 720, + 302, + 734 + ], + "score": 0.93, + "content": "Y _ { y } ^ { 0 } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 716, + 388, + 737 + ], + "score": 1.0, + "content": "and following policy", + "type": "text" + }, + { + "bbox": [ + 388, + 722, + 395, + 730 + ], + "score": 0.76, + "content": "\\tilde { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 716, + 440, + 737 + ], + "score": 1.0, + "content": ". We have:", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 165 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 104, + 82, + 507, + 165 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 166, + 168, + 444, + 198 + ], + "lines": [ + { + "bbox": [ + 166, + 168, + 444, + 198 + ], + "spans": [ + { + "bbox": [ + 166, + 168, + 444, + 198 + ], + "score": 0.93, + "content": "\\ell _ { \\theta } ( x , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) } { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) + \\sum _ { x ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\left\\{ x \\right\\} } \\exp ( \\lambda s _ { \\theta } ( x ^ { \\prime } , y ) ) }", + "type": "interline_equation", + "image_path": "e728c707965bd848cfff2b88e2d41e2810b434960e667d469684e6e39308eb85.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 166, + 168, + 444, + 178.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 166, + 178.0, + 444, + 188.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 166, + 188.0, + 444, + 198.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 203, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 133, + 216 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 204, + 141, + 213 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "is an inverse temperature hyperparameter. The overall SimCLR loss is then the expected", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 141, + 227 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 141, + 214, + 185, + 226 + ], + "score": 0.93, + "content": "\\ell _ { \\theta } ( x , y ; S )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 214, + 212, + 227 + ], + "score": 1.0, + "content": ", when", + "type": "text" + }, + { + "bbox": [ + 213, + 216, + 230, + 226 + ], + "score": 0.59, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 214, + 250, + 227 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 251, + 214, + 263, + 224 + ], + "score": 0.87, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 214, + 482, + 227 + ], + "score": 1.0, + "content": "are drawn from some augmented training distribution.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 203, + 505, + 227 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 232, + 276, + 246 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 277, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 277, + 248 + ], + "score": 1.0, + "content": "3 POLICY SIMILARITY METRIC", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 506, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 264 + ], + "score": 1.0, + "content": "A useful tool in learning a policy that generalizes is to understand which states result in similar", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "behavior, and which do not. To be maximally effective, this similarity should go beyond the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 442, + 287 + ], + "score": 1.0, + "content": "immediately chosen action and consider long-term behavior. In this regards, the", + "type": "text" + }, + { + "bbox": [ + 442, + 276, + 450, + 285 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "-bisimulation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 286, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 297 + ], + "score": 1.0, + "content": "metrics are interesting as they are based on the full sequence of future rewards received from different", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 297, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 308 + ], + "score": 1.0, + "content": "states. However, considering rewards can be both too restrictive (when the policies are the same, but", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "the obtained rewards are not; see Figure D.1) or too permissive (when the policies are different, but", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 323, + 331 + ], + "score": 1.0, + "content": "the obtained rewards are not; see Figure 3a). In fact,", + "type": "text" + }, + { + "bbox": [ + 324, + 320, + 330, + 329 + ], + "score": 0.79, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "-bisimulation metrics actually lead to poor", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 330, + 334, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 334, + 341 + ], + "score": 1.0, + "content": "generalization in our experiments (Sections 5.1 and 5.2).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 252, + 506, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 506, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "To address this issue, we instead consider the similarity between policies themselves. We replace", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "the absolute reward difference by a probability pseudometric between policies, denoted DIST. Addi-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "tionally, since we would like to perform well in unseen environments, we are interested in similarity", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 249, + 392 + ], + "score": 1.0, + "content": "in optimal behavior. We thus use", + "type": "text" + }, + { + "bbox": [ + 250, + 379, + 262, + 389 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "as the grounding policy. This yields the policy similarity", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 402 + ], + "score": 1.0, + "content": "metric (PSM), for which states are close when the optimal policies in these states and future states", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 401, + 457, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 264, + 414 + ], + "score": 1.0, + "content": "are similar. For a given DIST, the PSM", + "type": "text" + }, + { + "bbox": [ + 265, + 401, + 332, + 411 + ], + "score": 0.9, + "content": "d ^ { * } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 401, + 457, + 414 + ], + "score": 1.0, + "content": "satisfies the recursive equation", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 345, + 506, + 414 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 416, + 444, + 451 + ], + "lines": [ + { + "bbox": [ + 167, + 416, + 444, + 451 + ], + "spans": [ + { + "bbox": [ + 167, + 416, + 444, + 451 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } & { } & { d ^ { * } ( x , y ) = \\underbrace { \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) } _ { \\mathrm { ( A ) } } + \\underbrace { \\gamma \\mathcal { W } _ { 1 } ( d ^ { * } ) \\big ( P ^ { \\pi ^ { * } } ( \\cdot | x ) , P ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) } _ { \\mathrm { ( B ) } } . } \\end{array}", + "type": "interline_equation", + "image_path": "9ac02354c480f08d23c9d5364cc9a542168e6e21775e16792bd3b5b35a59511a.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 167, + 416, + 444, + 427.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 167, + 427.6666666666667, + 444, + 439.33333333333337 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 167, + 439.33333333333337, + 444, + 451.00000000000006 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 504, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 409, + 468 + ], + "score": 1.0, + "content": "The DIST term captures the difference in local optimal behavior (A) while", + "type": "text" + }, + { + "bbox": [ + 409, + 456, + 425, + 467 + ], + "score": 0.89, + "content": "\\mathcal { W } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 455, + 504, + 468 + ], + "score": 1.0, + "content": "captures long-term", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "optimal behavior difference (B); the exact weights assigned to the two are given by the discount", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 283, + 490 + ], + "score": 1.0, + "content": "factor. Furthermore, when DIST is bounded,", + "type": "text" + }, + { + "bbox": [ + 283, + 478, + 294, + 488 + ], + "score": 0.86, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "is guaranteed to be finite. While there are technically", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 488, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 506, + 500 + ], + "score": 1.0, + "content": "multiple PSMs (one for each DIST), we omit this distinction whenever clear from context. A proof of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 500, + 301, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 178, + 511 + ], + "score": 1.0, + "content": "the uniqueness of", + "type": "text" + }, + { + "bbox": [ + 179, + 500, + 190, + 509 + ], + "score": 0.87, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 500, + 301, + 511 + ], + "score": 1.0, + "content": "is given in Proposition B.1.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 455, + 506, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "Our main use of PSM will be to compare states across environments. In this context, we identify the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "terms in Equation 3 with specific environments for clarity and write (despite its technical inaccuracy)", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 516, + 506, + 540 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 543, + 445, + 560 + ], + "lines": [ + { + "bbox": [ + 165, + 543, + 445, + 560 + ], + "spans": [ + { + "bbox": [ + 165, + 543, + 445, + 560 + ], + "score": 0.9, + "content": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ef10847f7eac97b0d10c96ae00eaa9d1f6547c2a192ff9eb67cf97571f068e91.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 165, + 543, + 445, + 560 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "score": 1.0, + "content": "PSM is applicable to both discrete and continuous action spaces. In our experiments, DIST is the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 204, + 588 + ], + "score": 1.0, + "content": "total variation distance", + "type": "text" + }, + { + "bbox": [ + 204, + 576, + 227, + 587 + ], + "score": 0.8, + "content": "( T V )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 574, + 254, + 588 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 254, + 576, + 264, + 586 + ], + "score": 0.81, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 574, + 378, + 588 + ], + "score": 1.0, + "content": "is discrete, and we use the", + "type": "text" + }, + { + "bbox": [ + 379, + 576, + 389, + 586 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 574, + 505, + 588 + ], + "score": 1.0, + "content": "distance between the mean", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 587, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 238, + 599 + ], + "score": 1.0, + "content": "actions of the two policies when", + "type": "text" + }, + { + "bbox": [ + 239, + 587, + 248, + 596 + ], + "score": 0.79, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 587, + 504, + 599 + ], + "score": 1.0, + "content": "is continuous. PSM can be computed iteratively using dynamic", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 435, + 609 + ], + "score": 1.0, + "content": "programming (Ferns et al., 2011) (more details in Section D.1). Furthermore, when", + "type": "text" + }, + { + "bbox": [ + 436, + 598, + 447, + 608 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "is unavailable", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 342, + 621 + ], + "score": 1.0, + "content": "on training environments, we replace it by an approximation", + "type": "text" + }, + { + "bbox": [ + 343, + 609, + 354, + 619 + ], + "score": 0.88, + "content": "\\hat { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "to obtain an approximate PSM, which", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 436, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 348, + 632 + ], + "score": 1.0, + "content": "is close to the exact PSM depending on the suboptimality of", + "type": "text" + }, + { + "bbox": [ + 349, + 620, + 360, + 630 + ], + "score": 0.86, + "content": "\\hat { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 619, + 436, + 632 + ], + "score": 1.0, + "content": "(Proposition D.3).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 564, + 505, + 632 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 184, + 648 + ], + "score": 1.0, + "content": "Despite resembling", + "type": "text" + }, + { + "bbox": [ + 184, + 638, + 191, + 646 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "-bisimulation metrics in form, the PSM possesses different characteristics which", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 647, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 505, + 660 + ], + "score": 1.0, + "content": "are better suited to the problem of generalizing a learned policy. To illustrate this point, consider the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 656, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 343, + 672 + ], + "score": 1.0, + "content": "following simple nearest-neighbour scheme: Given a state", + "type": "text" + }, + { + "bbox": [ + 343, + 659, + 369, + 669 + ], + "score": 0.92, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 656, + 482, + 672 + ], + "score": 1.0, + "content": ", denote its closest match in", + "type": "text" + }, + { + "bbox": [ + 482, + 659, + 492, + 668 + ], + "score": 0.8, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 656, + 506, + 672 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 668, + 507, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 219, + 681 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\tilde { x } _ { y } : = \\arg \\operatorname* { m i n } _ { x \\in \\mathcal { X } } d ^ { * } ( x , y ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 668, + 405, + 685 + ], + "score": 1.0, + "content": ". Suppose that we use this scheme to transfer", + "type": "text" + }, + { + "bbox": [ + 405, + 669, + 419, + 681 + ], + "score": 0.9, + "content": "\\pi _ { \\mathcal { X } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 668, + 431, + 685 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 431, + 669, + 451, + 681 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 668, + 507, + 685 + ], + "score": 1.0, + "content": ", in the sense", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 679, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 259, + 694 + ], + "score": 1.0, + "content": "that we behave according to the policy", + "type": "text" + }, + { + "bbox": [ + 259, + 680, + 324, + 692 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ( y ) = \\pi _ { \\chi } ^ { * } ( \\tilde { x } _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 679, + 496, + 694 + ], + "score": 1.0, + "content": ". We can then bound the difference between", + "type": "text" + }, + { + "bbox": [ + 496, + 681, + 504, + 690 + ], + "score": 0.79, + "content": "\\tilde { \\pi }", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 690, + 440, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 123, + 705 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 691, + 136, + 704 + ], + "score": 0.9, + "content": "\\pi _ { \\mathcal { Y } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 690, + 270, + 705 + ], + "score": 1.0, + "content": ", something that is not possible if", + "type": "text" + }, + { + "bbox": [ + 270, + 692, + 280, + 700 + ], + "score": 0.87, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 690, + 347, + 705 + ], + "score": 1.0, + "content": "is replaced by a", + "type": "text" + }, + { + "bbox": [ + 347, + 693, + 354, + 701 + ], + "score": 0.8, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 690, + 440, + 705 + ], + "score": 1.0, + "content": "-bisimulation metric.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 104, + 636, + 507, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 706, + 502, + 734 + ], + "lines": [ + { + "bbox": [ + 104, + 704, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 704, + 303, + 725 + ], + "score": 1.0, + "content": "Theorem 1. [Bound on policy transfer] For any", + "type": "text" + }, + { + "bbox": [ + 304, + 708, + 329, + 719 + ], + "score": 0.87, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 704, + 346, + 725 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 346, + 707, + 423, + 721 + ], + "score": 0.92, + "content": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 704, + 506, + 725 + ], + "score": 1.0, + "content": "define the sequence", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 716, + 440, + 737 + ], + "spans": [ + { + "bbox": [ + 104, + 716, + 270, + 737 + ], + "score": 1.0, + "content": "of random states encountered starting in", + "type": "text" + }, + { + "bbox": [ + 271, + 720, + 302, + 734 + ], + "score": 0.93, + "content": "Y _ { y } ^ { 0 } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 716, + 388, + 737 + ], + "score": 1.0, + "content": "and following policy", + "type": "text" + }, + { + "bbox": [ + 388, + 722, + 395, + 730 + ], + "score": 0.76, + "content": "\\tilde { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 716, + 440, + 737 + ], + "score": 1.0, + "content": ". We have:", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5, + "bbox_fs": [ + 104, + 704, + 506, + 737 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 191, + 79, + 418, + 119 + ], + "lines": [ + { + "bbox": [ + 191, + 79, + 418, + 119 + ], + "spans": [ + { + "bbox": [ + 191, + 79, + 418, + 119 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) .", + "type": "interline_equation", + "image_path": "cd132237f035791e31f3d0bfc25305f208b4467de5d698d767c39bc00821433b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 191, + 79, + 418, + 92.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 191, + 92.33333333333333, + 418, + 105.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 191, + 105.66666666666666, + 418, + 118.99999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 506, + 201 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 413, + 137 + ], + "score": 1.0, + "content": "The proof is in the Appendix (Section B). Theorem 1 is non-vacuous whenever", + "type": "text" + }, + { + "bbox": [ + 414, + 123, + 503, + 136 + ], + "score": 0.93, + "content": "d ^ { * } ( \\tilde { x } _ { y } , y ) < 1 / ( 1 + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 123, + 506, + 137 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 159, + 147 + ], + "score": 1.0, + "content": "In particular,", + "type": "text" + }, + { + "bbox": [ + 159, + 135, + 216, + 146 + ], + "score": 0.9, + "content": "d ^ { * } ( \\tilde { x } _ { y } , y ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "implies that the transferred policy is optimal. Although this scheme is not", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 192, + 158 + ], + "score": 1.0, + "content": "practical (computing", + "type": "text" + }, + { + "bbox": [ + 192, + 146, + 203, + 155 + ], + "score": 0.84, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 145, + 296, + 158 + ], + "score": 1.0, + "content": "requires knowledge of", + "type": "text" + }, + { + "bbox": [ + 297, + 146, + 311, + 158 + ], + "score": 0.87, + "content": "\\pi _ { \\mathcal { D } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "), it shows that meaningful policy generalization", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 361, + 170 + ], + "score": 1.0, + "content": "can be obtained if we can find a mapping that generalizes across", + "type": "text" + }, + { + "bbox": [ + 362, + 157, + 370, + 166 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 156, + 506, + 170 + ], + "score": 1.0, + "content": ". Put another way, PSM gives us a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 167, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 506, + 180 + ], + "score": 1.0, + "content": "principled way of lifting generalization across inputs (a supervised learning problem) to generalization", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "across environments. We now describe how to employ PSM for learning representations that put", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 189, + 402, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 402, + 203 + ], + "score": 1.0, + "content": "together states in which the agent’s long-term optimal behavior is similar.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 107, + 208, + 377, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 206, + 378, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 378, + 223 + ], + "score": 1.0, + "content": "4 LEARNING CONTRASTIVE METRIC EMBEDDINGS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 277, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 227, + 279, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 279, + 239 + ], + "score": 1.0, + "content": "To generalize a learned policy to new envi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 279, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 279, + 250 + ], + "score": 1.0, + "content": "ronments, we build on the success of con-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 277, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 277, + 261 + ], + "score": 1.0, + "content": "trastive representations (Section 2). Given", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 278, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 208, + 272 + ], + "score": 1.0, + "content": "a state similarity metric", + "type": "text" + }, + { + "bbox": [ + 208, + 261, + 215, + 270 + ], + "score": 0.69, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 260, + 278, + 272 + ], + "score": 1.0, + "content": ", we develop a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 272, + 277, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 277, + 282 + ], + "score": 1.0, + "content": "general procedure (Algorithm 1) to learn", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 282, + 277, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 277, + 294 + ], + "score": 1.0, + "content": "contrastive metric embeddings (CMEs) for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 292, + 277, + 305 + ], + "spans": [ + { + "bbox": [ + 107, + 293, + 113, + 303 + ], + "score": 0.54, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 292, + 204, + 305 + ], + "score": 1.0, + "content": ". We utilize the metric", + "type": "text" + }, + { + "bbox": [ + 204, + 294, + 211, + 303 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 292, + 277, + 305 + ], + "score": 1.0, + "content": "for defining the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 304, + 277, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 277, + 316 + ], + "score": 1.0, + "content": "set of positive and negative pairs, as well", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 277, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 277, + 327 + ], + "score": 1.0, + "content": "as assigning importance weights to these", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 272, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 272, + 338 + ], + "score": 1.0, + "content": "pairs in the contrastive loss (Equation 4).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 284, + 230, + 503, + 242 + ], + "lines": [ + { + "bbox": [ + 283, + 228, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 283, + 228, + 503, + 245 + ], + "score": 1.0, + "content": "Algorithm 1 Contrastive Metric Embeddings (CMEs)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 288, + 246, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 290, + 246, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 290, + 246, + 401, + 258 + ], + "score": 1.0, + "content": "1: Given: State embedding", + "type": "text" + }, + { + "bbox": [ + 401, + 246, + 421, + 258 + ], + "score": 0.88, + "content": "z _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 246, + 454, + 258 + ], + "score": 1.0, + "content": ", Metric", + "type": "text" + }, + { + "bbox": [ + 455, + 246, + 478, + 258 + ], + "score": 0.9, + "content": "d ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 246, + 506, + 258 + ], + "score": 1.0, + "content": "Train-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 251, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 299, + 251, + 377, + 274 + ], + "score": 1.0, + "content": "ing environments", + "type": "text" + }, + { + "bbox": [ + 378, + 257, + 416, + 269 + ], + "score": 0.92, + "content": "\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 251, + 506, + 274 + ], + "score": 1.0, + "content": "· · · . Hyperparameters:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 301, + 268, + 500, + 280 + ], + "spans": [ + { + "bbox": [ + 301, + 268, + 354, + 280 + ], + "score": 1.0, + "content": "Temperature", + "type": "text" + }, + { + "bbox": [ + 354, + 268, + 371, + 280 + ], + "score": 0.88, + "content": "1 / \\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 268, + 399, + 280 + ], + "score": 1.0, + "content": ", Scale", + "type": "text" + }, + { + "bbox": [ + 399, + 270, + 406, + 280 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 268, + 488, + 280 + ], + "score": 1.0, + "content": ", Total training steps", + "type": "text" + }, + { + "bbox": [ + 489, + 268, + 500, + 278 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 290, + 280, + 427, + 291 + ], + "spans": [ + { + "bbox": [ + 290, + 280, + 298, + 289 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 300, + 280, + 356, + 291 + ], + "score": 1.0, + "content": "for each step", + "type": "text" + }, + { + "bbox": [ + 356, + 280, + 412, + 291 + ], + "score": 0.87, + "content": "k = 1 , \\ldots , K", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 280, + 427, + 291 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 289, + 290, + 484, + 302 + ], + "spans": [ + { + "bbox": [ + 289, + 290, + 299, + 301 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 311, + 290, + 441, + 302 + ], + "score": 1.0, + "content": "Sample a pair of training MDPs", + "type": "text" + }, + { + "bbox": [ + 441, + 290, + 484, + 302 + ], + "score": 0.8, + "content": "\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 289, + 301, + 458, + 313 + ], + "spans": [ + { + "bbox": [ + 289, + 302, + 299, + 313 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 311, + 301, + 343, + 313 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 343, + 302, + 349, + 311 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 301, + 400, + 313 + ], + "score": 1.0, + "content": "to minimize", + "type": "text" + }, + { + "bbox": [ + 401, + 302, + 429, + 312 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 301, + 458, + 313 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 311, + 312, + 475, + 325 + ], + "spans": [ + { + "bbox": [ + 311, + 312, + 475, + 325 + ], + "score": 0.81, + "content": "\\mathcal { L } _ { \\mathrm { C M E } } = \\mathbb { E } _ { \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } \\sim \\rho } \\left[ L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) \\right]", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 289, + 322, + 336, + 335 + ], + "spans": [ + { + "bbox": [ + 289, + 322, + 336, + 335 + ], + "score": 1.0, + "content": "5: end for", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 506, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 507, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 278, + 355 + ], + "score": 1.0, + "content": "We first apply a transformation to convert", + "type": "text" + }, + { + "bbox": [ + 278, + 343, + 285, + 353 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 342, + 381, + 355 + ], + "score": 1.0, + "content": "to a similarity measure", + "type": "text" + }, + { + "bbox": [ + 381, + 343, + 389, + 353 + ], + "score": 0.41, + "content": "\\Gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 342, + 507, + 355 + ], + "score": 1.0, + "content": ", bounded in [0, 1] for “soft”", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 261, + 366 + ], + "score": 1.0, + "content": "similarities. In this work, we transform", + "type": "text" + }, + { + "bbox": [ + 262, + 354, + 268, + 363 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 353, + 495, + 366 + ], + "score": 1.0, + "content": "using the Gaussian kernel with a positive scale parameter", + "type": "text" + }, + { + "bbox": [ + 495, + 354, + 502, + 365 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 353, + 506, + 366 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 363, + 478, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 136, + 378 + ], + "score": 1.0, + "content": "that is,", + "type": "text" + }, + { + "bbox": [ + 136, + 365, + 249, + 377 + ], + "score": 0.9, + "content": "\\Gamma ( x , y ) = \\exp ( - d ( x , y ) / \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 363, + 253, + 378 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 254, + 365, + 262, + 376 + ], + "score": 0.8, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 363, + 467, + 378 + ], + "score": 1.0, + "content": "controls the sensitivity of the similarity measure to", + "type": "text" + }, + { + "bbox": [ + 468, + 365, + 474, + 375 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 363, + 478, + 378 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 506, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 390, + 394 + ], + "score": 1.0, + "content": "Second, we select the positive and negative pairs given a set of states", + "type": "text" + }, + { + "bbox": [ + 390, + 381, + 425, + 393 + ], + "score": 0.91, + "content": "{ \\mathcal { X } } ^ { \\prime } \\subseteq { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 381, + 444, + 394 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 444, + 382, + 453, + 392 + ], + "score": 0.8, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "from MDPs", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 107, + 393, + 150, + 405 + ], + "score": 0.89, + "content": "\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 392, + 296, + 405 + ], + "score": 1.0, + "content": "respectively. For each anchor state", + "type": "text" + }, + { + "bbox": [ + 296, + 393, + 323, + 404 + ], + "score": 0.91, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 392, + 452, + 405 + ], + "score": 1.0, + "content": ", we use its nearest neighbor in", + "type": "text" + }, + { + "bbox": [ + 452, + 393, + 465, + 402 + ], + "score": 0.86, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "based on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 403, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 200, + 418 + ], + "score": 1.0, + "content": "the similarity measure", + "type": "text" + }, + { + "bbox": [ + 200, + 404, + 208, + 414 + ], + "score": 0.6, + "content": "\\Gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 403, + 320, + 418 + ], + "score": 1.0, + "content": "to define the positive pairs", + "type": "text" + }, + { + "bbox": [ + 320, + 403, + 359, + 416 + ], + "score": 0.93, + "content": "\\{ ( \\tilde { x } _ { y } , y ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 403, + 390, + 418 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 391, + 404, + 481, + 422 + ], + "score": 0.74, + "content": "\\tilde { x } _ { y } = \\underset { x \\in \\mathcal { X } ^ { \\prime } } { \\operatorname { a r g m a x } } \\Gamma ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 403, + 506, + 418 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 421, + 186, + 434 + ], + "score": 1.0, + "content": "remaining states in", + "type": "text" + }, + { + "bbox": [ + 186, + 422, + 198, + 432 + ], + "score": 0.86, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 421, + 250, + 434 + ], + "score": 1.0, + "content": ", paired with", + "type": "text" + }, + { + "bbox": [ + 251, + 424, + 257, + 433 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 421, + 505, + 434 + ], + "score": 1.0, + "content": ", are used as negative pairs. This choice of pairs is motivated", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 384, + 446 + ], + "score": 1.0, + "content": "by Theorem 1, which shows that if we transfer the optimal policy in", + "type": "text" + }, + { + "bbox": [ + 384, + 433, + 405, + 444 + ], + "score": 0.91, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "to the nearest neighbors", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 444, + 430, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 262, + 456 + ], + "score": 1.0, + "content": "defined using PSM, its performance in", + "type": "text" + }, + { + "bbox": [ + 263, + 444, + 282, + 456 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 444, + 430, + 456 + ], + "score": 1.0, + "content": "has suboptimality bounded by PSM.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "Next, we define a soft version of the SimCLR contrastive loss (Equation 2) for learning the function", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 471, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 107, + 474, + 117, + 483 + ], + "score": 0.83, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 471, + 474, + 485 + ], + "score": 1.0, + "content": ", which maps states (usually high-dimensional) to embeddings. Given a positive state pair", + "type": "text" + }, + { + "bbox": [ + 474, + 472, + 502, + 484 + ], + "score": 0.92, + "content": "( \\tilde { x } _ { y } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 471, + 506, + 485 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 482, + 499, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 135, + 496 + ], + "score": 1.0, + "content": "the set", + "type": "text" + }, + { + "bbox": [ + 135, + 483, + 147, + 493 + ], + "score": 0.88, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 482, + 259, + 496 + ], + "score": 1.0, + "content": ", and the similarity measure", + "type": "text" + }, + { + "bbox": [ + 259, + 483, + 266, + 493 + ], + "score": 0.64, + "content": "\\Gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 482, + 499, + 496 + ], + "score": 1.0, + "content": ", the loss (pseudocode provided in Section J.2) is given by", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 105, + 498, + 492, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 492, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 492, + 528 + ], + "score": 0.94, + "content": "\\ell _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) } { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) + \\sum _ { \\boldsymbol { x } ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\{ \\tilde { x } _ { y } \\} } ( 1 - \\Gamma ( \\boldsymbol { x } ^ { \\prime } , y ) ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\boldsymbol { x } ^ { \\prime } , y ) ) }", + "type": "interline_equation", + "image_path": "50d884eead6fd8b8db5a7642e5b04245f374580b54d8c1e73dc93eeb51e99deb.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 105, + 498, + 492, + 508.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 105, + 508.0, + 492, + 518.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 105, + 518.0, + 492, + 528.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "where we use the same notation as in Equation 2. Following SimCLR, we use a non-linear projection", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 200, + 556 + ], + "score": 1.0, + "content": "of the representation as", + "type": "text" + }, + { + "bbox": [ + 200, + 545, + 210, + 554 + ], + "score": 0.84, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 542, + 506, + 556 + ], + "score": 1.0, + "content": "(Figure A.1). The agent’s policy is an affine function of the representation.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 586 + ], + "lines": [ + { + "bbox": [ + 103, + 555, + 508, + 577 + ], + "spans": [ + { + "bbox": [ + 103, + 555, + 229, + 577 + ], + "score": 1.0, + "content": "The total contrastive loss for", + "type": "text" + }, + { + "bbox": [ + 229, + 560, + 250, + 571 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 555, + 270, + 577 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 560, + 290, + 572 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 555, + 423, + 577 + ], + "score": 1.0, + "content": "utilizes the optimal trajectories", + "type": "text" + }, + { + "bbox": [ + 423, + 559, + 485, + 572 + ], + "score": 0.94, + "content": "\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 555, + 508, + 577 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 570, + 483, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 163, + 586 + ], + "score": 0.93, + "content": "\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 570, + 194, + 587 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 194, + 572, + 272, + 585 + ], + "score": 0.92, + "content": "x _ { t + 1 } \\sim P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 570, + 289, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 571, + 365, + 586 + ], + "score": 0.93, + "content": "y _ { t + 1 } \\sim P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 570, + 398, + 587 + ], + "score": 1.0, + "content": ". We set", + "type": "text" + }, + { + "bbox": [ + 398, + 572, + 435, + 585 + ], + "score": 0.92, + "content": "\\mathcal { X } ^ { \\prime } = \\tau _ { \\mathcal { X } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 570, + 483, + 587 + ], + "score": 1.0, + "content": "and define", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 589, + 458, + 611 + ], + "lines": [ + { + "bbox": [ + 152, + 589, + 458, + 611 + ], + "spans": [ + { + "bbox": [ + 152, + 589, + 458, + 611 + ], + "score": 0.88, + "content": "L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) = \\mathbb { E } _ { y \\sim \\tau _ { \\mathcal { Y } } ^ { * } } \\left[ \\ell _ { \\theta } ( \\tilde { x } _ { y } , y ; \\tau _ { \\mathcal { X } } ^ { * } ) \\right] \\qquad \\mathrm { w h e r e } \\ \\tilde { x } _ { y } = \\underset { x \\in \\tau _ { \\mathcal { X } } ^ { * } } { \\mathrm { a r g m a x } } \\ \\Gamma ( x , y ) .", + "type": "interline_equation", + "image_path": "9d03779755fcd19362a56a25c5ae58750a0c2ce96aa98c5ed4bdc23a3e1cce64.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 152, + 589, + 458, + 611 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 450, + 627 + ], + "score": 1.0, + "content": "We refer to CMEs learned with policy similarity metric as policy similarity embeddings", + "type": "text" + }, + { + "bbox": [ + 450, + 614, + 477, + 625 + ], + "score": 0.67, + "content": "( P S E s )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 613, + 505, + 627 + ], + "score": 1.0, + "content": ". PSEs", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 625, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 638 + ], + "score": 1.0, + "content": "can trivially be combined with data augmentation by using augmented states when computing state", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 395, + 649 + ], + "score": 1.0, + "content": "embeddings. We simultaneously learn PSEs with the RL agent by adding", + "type": "text" + }, + { + "bbox": [ + 396, + 636, + 424, + 647 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "(Algorithm 1) as an", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 646, + 490, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 490, + 660 + ], + "score": 1.0, + "content": "auxiliary objective during training. Next, we illustrate the benefits from this auxiliary objective.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51.5 + }, + { + "type": "title", + "bbox": [ + 107, + 668, + 363, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 363, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 363, + 682 + ], + "score": 1.0, + "content": "5 JUMPING TASK FROM PIXELS: A CASE STUDY", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "score": 1.0, + "content": "Task Description. The jumping task (Tachet des Combes et al., 2018) (Figure 1) captures, using", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "well-defined factors of variations, whether agents can learn the correct invariances required for", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "generalization, directly from image inputs. The task consists of an agent trying to jump over an", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "obstacle. The agent has access to two actions: right and jump. The agent needs to time the jump", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 56.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 191, + 79, + 418, + 119 + ], + "lines": [ + { + "bbox": [ + 191, + 79, + 418, + 119 + ], + "spans": [ + { + "bbox": [ + 191, + 79, + 418, + 119 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) .", + "type": "interline_equation", + "image_path": "cd132237f035791e31f3d0bfc25305f208b4467de5d698d767c39bc00821433b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 191, + 79, + 418, + 92.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 191, + 92.33333333333333, + 418, + 105.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 191, + 105.66666666666666, + 418, + 118.99999999999999 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 506, + 201 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 413, + 137 + ], + "score": 1.0, + "content": "The proof is in the Appendix (Section B). Theorem 1 is non-vacuous whenever", + "type": "text" + }, + { + "bbox": [ + 414, + 123, + 503, + 136 + ], + "score": 0.93, + "content": "d ^ { * } ( \\tilde { x } _ { y } , y ) < 1 / ( 1 + \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 123, + 506, + 137 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 159, + 147 + ], + "score": 1.0, + "content": "In particular,", + "type": "text" + }, + { + "bbox": [ + 159, + 135, + 216, + 146 + ], + "score": 0.9, + "content": "d ^ { * } ( \\tilde { x } _ { y } , y ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "implies that the transferred policy is optimal. Although this scheme is not", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 192, + 158 + ], + "score": 1.0, + "content": "practical (computing", + "type": "text" + }, + { + "bbox": [ + 192, + 146, + 203, + 155 + ], + "score": 0.84, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 145, + 296, + 158 + ], + "score": 1.0, + "content": "requires knowledge of", + "type": "text" + }, + { + "bbox": [ + 297, + 146, + 311, + 158 + ], + "score": 0.87, + "content": "\\pi _ { \\mathcal { D } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "), it shows that meaningful policy generalization", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 361, + 170 + ], + "score": 1.0, + "content": "can be obtained if we can find a mapping that generalizes across", + "type": "text" + }, + { + "bbox": [ + 362, + 157, + 370, + 166 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 156, + 506, + 170 + ], + "score": 1.0, + "content": ". Put another way, PSM gives us a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 167, + 506, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 506, + 180 + ], + "score": 1.0, + "content": "principled way of lifting generalization across inputs (a supervised learning problem) to generalization", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "across environments. We now describe how to employ PSM for learning representations that put", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 189, + 402, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 402, + 203 + ], + "score": 1.0, + "content": "together states in which the agent’s long-term optimal behavior is similar.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 123, + 506, + 203 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 208, + 377, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 206, + 378, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 378, + 223 + ], + "score": 1.0, + "content": "4 LEARNING CONTRASTIVE METRIC EMBEDDINGS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 277, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 227, + 279, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 279, + 239 + ], + "score": 1.0, + "content": "To generalize a learned policy to new envi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 279, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 279, + 250 + ], + "score": 1.0, + "content": "ronments, we build on the success of con-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 277, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 277, + 261 + ], + "score": 1.0, + "content": "trastive representations (Section 2). Given", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 278, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 208, + 272 + ], + "score": 1.0, + "content": "a state similarity metric", + "type": "text" + }, + { + "bbox": [ + 208, + 261, + 215, + 270 + ], + "score": 0.69, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 260, + 278, + 272 + ], + "score": 1.0, + "content": ", we develop a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 272, + 277, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 277, + 282 + ], + "score": 1.0, + "content": "general procedure (Algorithm 1) to learn", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 282, + 277, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 277, + 294 + ], + "score": 1.0, + "content": "contrastive metric embeddings (CMEs) for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 292, + 277, + 305 + ], + "spans": [ + { + "bbox": [ + 107, + 293, + 113, + 303 + ], + "score": 0.54, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 292, + 204, + 305 + ], + "score": 1.0, + "content": ". We utilize the metric", + "type": "text" + }, + { + "bbox": [ + 204, + 294, + 211, + 303 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 292, + 277, + 305 + ], + "score": 1.0, + "content": "for defining the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 304, + 277, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 277, + 316 + ], + "score": 1.0, + "content": "set of positive and negative pairs, as well", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 277, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 277, + 327 + ], + "score": 1.0, + "content": "as assigning importance weights to these", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 272, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 272, + 338 + ], + "score": 1.0, + "content": "pairs in the contrastive loss (Equation 4).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 227, + 279, + 338 + ] + }, + { + "type": "title", + "bbox": [ + 284, + 230, + 503, + 242 + ], + "lines": [ + { + "bbox": [ + 283, + 228, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 283, + 228, + 503, + 245 + ], + "score": 1.0, + "content": "Algorithm 1 Contrastive Metric Embeddings (CMEs)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "list", + "bbox": [ + 288, + 246, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 290, + 246, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 290, + 246, + 401, + 258 + ], + "score": 1.0, + "content": "1: Given: State embedding", + "type": "text" + }, + { + "bbox": [ + 401, + 246, + 421, + 258 + ], + "score": 0.88, + "content": "z _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 246, + 454, + 258 + ], + "score": 1.0, + "content": ", Metric", + "type": "text" + }, + { + "bbox": [ + 455, + 246, + 478, + 258 + ], + "score": 0.9, + "content": "d ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 246, + 506, + 258 + ], + "score": 1.0, + "content": "Train-", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 299, + 251, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 299, + 251, + 377, + 274 + ], + "score": 1.0, + "content": "ing environments", + "type": "text" + }, + { + "bbox": [ + 378, + 257, + 416, + 269 + ], + "score": 0.92, + "content": "\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 251, + 506, + 274 + ], + "score": 1.0, + "content": "· · · . Hyperparameters:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 301, + 268, + 500, + 280 + ], + "spans": [ + { + "bbox": [ + 301, + 268, + 354, + 280 + ], + "score": 1.0, + "content": "Temperature", + "type": "text" + }, + { + "bbox": [ + 354, + 268, + 371, + 280 + ], + "score": 0.88, + "content": "1 / \\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 268, + 399, + 280 + ], + "score": 1.0, + "content": ", Scale", + "type": "text" + }, + { + "bbox": [ + 399, + 270, + 406, + 280 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 268, + 488, + 280 + ], + "score": 1.0, + "content": ", Total training steps", + "type": "text" + }, + { + "bbox": [ + 489, + 268, + 500, + 278 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 290, + 280, + 427, + 291 + ], + "spans": [ + { + "bbox": [ + 290, + 280, + 298, + 289 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 300, + 280, + 356, + 291 + ], + "score": 1.0, + "content": "for each step", + "type": "text" + }, + { + "bbox": [ + 356, + 280, + 412, + 291 + ], + "score": 0.87, + "content": "k = 1 , \\ldots , K", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 280, + 427, + 291 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 289, + 290, + 484, + 302 + ], + "spans": [ + { + "bbox": [ + 289, + 290, + 299, + 301 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 311, + 290, + 441, + 302 + ], + "score": 1.0, + "content": "Sample a pair of training MDPs", + "type": "text" + }, + { + "bbox": [ + 441, + 290, + 484, + 302 + ], + "score": 0.8, + "content": "\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + } + ], + "index": 26, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 289, + 301, + 458, + 313 + ], + "spans": [ + { + "bbox": [ + 289, + 302, + 299, + 313 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 311, + 301, + 343, + 313 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 343, + 302, + 349, + 311 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 301, + 400, + 313 + ], + "score": 1.0, + "content": "to minimize", + "type": "text" + }, + { + "bbox": [ + 401, + 302, + 429, + 312 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 301, + 458, + 313 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 311, + 312, + 475, + 325 + ], + "spans": [ + { + "bbox": [ + 311, + 312, + 475, + 325 + ], + "score": 0.81, + "content": "\\mathcal { L } _ { \\mathrm { C M E } } = \\mathbb { E } _ { \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } \\sim \\rho } \\left[ L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) \\right]", + "type": "inline_equation" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 289, + 322, + 336, + 335 + ], + "spans": [ + { + "bbox": [ + 289, + 322, + 336, + 335 + ], + "score": 1.0, + "content": "5: end for", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 25.5, + "bbox_fs": [ + 289, + 246, + 506, + 335 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 506, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 507, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 278, + 355 + ], + "score": 1.0, + "content": "We first apply a transformation to convert", + "type": "text" + }, + { + "bbox": [ + 278, + 343, + 285, + 353 + ], + "score": 0.77, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 342, + 381, + 355 + ], + "score": 1.0, + "content": "to a similarity measure", + "type": "text" + }, + { + "bbox": [ + 381, + 343, + 389, + 353 + ], + "score": 0.41, + "content": "\\Gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 342, + 507, + 355 + ], + "score": 1.0, + "content": ", bounded in [0, 1] for “soft”", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 261, + 366 + ], + "score": 1.0, + "content": "similarities. In this work, we transform", + "type": "text" + }, + { + "bbox": [ + 262, + 354, + 268, + 363 + ], + "score": 0.8, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 353, + 495, + 366 + ], + "score": 1.0, + "content": "using the Gaussian kernel with a positive scale parameter", + "type": "text" + }, + { + "bbox": [ + 495, + 354, + 502, + 365 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 353, + 506, + 366 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 363, + 478, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 136, + 378 + ], + "score": 1.0, + "content": "that is,", + "type": "text" + }, + { + "bbox": [ + 136, + 365, + 249, + 377 + ], + "score": 0.9, + "content": "\\Gamma ( x , y ) = \\exp ( - d ( x , y ) / \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 363, + 253, + 378 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 254, + 365, + 262, + 376 + ], + "score": 0.8, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 363, + 467, + 378 + ], + "score": 1.0, + "content": "controls the sensitivity of the similarity measure to", + "type": "text" + }, + { + "bbox": [ + 468, + 365, + 474, + 375 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 363, + 478, + 378 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 342, + 507, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 506, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 390, + 394 + ], + "score": 1.0, + "content": "Second, we select the positive and negative pairs given a set of states", + "type": "text" + }, + { + "bbox": [ + 390, + 381, + 425, + 393 + ], + "score": 0.91, + "content": "{ \\mathcal { X } } ^ { \\prime } \\subseteq { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 381, + 444, + 394 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 444, + 382, + 453, + 392 + ], + "score": 0.8, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "from MDPs", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 107, + 393, + 150, + 405 + ], + "score": 0.89, + "content": "\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 392, + 296, + 405 + ], + "score": 1.0, + "content": "respectively. For each anchor state", + "type": "text" + }, + { + "bbox": [ + 296, + 393, + 323, + 404 + ], + "score": 0.91, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 392, + 452, + 405 + ], + "score": 1.0, + "content": ", we use its nearest neighbor in", + "type": "text" + }, + { + "bbox": [ + 452, + 393, + 465, + 402 + ], + "score": 0.86, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "based on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 403, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 200, + 418 + ], + "score": 1.0, + "content": "the similarity measure", + "type": "text" + }, + { + "bbox": [ + 200, + 404, + 208, + 414 + ], + "score": 0.6, + "content": "\\Gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 403, + 320, + 418 + ], + "score": 1.0, + "content": "to define the positive pairs", + "type": "text" + }, + { + "bbox": [ + 320, + 403, + 359, + 416 + ], + "score": 0.93, + "content": "\\{ ( \\tilde { x } _ { y } , y ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 403, + 390, + 418 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 391, + 404, + 481, + 422 + ], + "score": 0.74, + "content": "\\tilde { x } _ { y } = \\underset { x \\in \\mathcal { X } ^ { \\prime } } { \\operatorname { a r g m a x } } \\Gamma ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 403, + 506, + 418 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 421, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 421, + 186, + 434 + ], + "score": 1.0, + "content": "remaining states in", + "type": "text" + }, + { + "bbox": [ + 186, + 422, + 198, + 432 + ], + "score": 0.86, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 421, + 250, + 434 + ], + "score": 1.0, + "content": ", paired with", + "type": "text" + }, + { + "bbox": [ + 251, + 424, + 257, + 433 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 421, + 505, + 434 + ], + "score": 1.0, + "content": ", are used as negative pairs. This choice of pairs is motivated", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 384, + 446 + ], + "score": 1.0, + "content": "by Theorem 1, which shows that if we transfer the optimal policy in", + "type": "text" + }, + { + "bbox": [ + 384, + 433, + 405, + 444 + ], + "score": 0.91, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "to the nearest neighbors", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 444, + 430, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 262, + 456 + ], + "score": 1.0, + "content": "defined using PSM, its performance in", + "type": "text" + }, + { + "bbox": [ + 263, + 444, + 282, + 456 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 444, + 430, + 456 + ], + "score": 1.0, + "content": "has suboptimality bounded by PSM.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 381, + 506, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "Next, we define a soft version of the SimCLR contrastive loss (Equation 2) for learning the function", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 471, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 107, + 474, + 117, + 483 + ], + "score": 0.83, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 471, + 474, + 485 + ], + "score": 1.0, + "content": ", which maps states (usually high-dimensional) to embeddings. Given a positive state pair", + "type": "text" + }, + { + "bbox": [ + 474, + 472, + 502, + 484 + ], + "score": 0.92, + "content": "( \\tilde { x } _ { y } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 471, + 506, + 485 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 482, + 499, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 135, + 496 + ], + "score": 1.0, + "content": "the set", + "type": "text" + }, + { + "bbox": [ + 135, + 483, + 147, + 493 + ], + "score": 0.88, + "content": "\\mathcal { X } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 482, + 259, + 496 + ], + "score": 1.0, + "content": ", and the similarity measure", + "type": "text" + }, + { + "bbox": [ + 259, + 483, + 266, + 493 + ], + "score": 0.64, + "content": "\\Gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 482, + 499, + 496 + ], + "score": 1.0, + "content": ", the loss (pseudocode provided in Section J.2) is given by", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 461, + 506, + 496 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 105, + 498, + 492, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 492, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 492, + 528 + ], + "score": 0.94, + "content": "\\ell _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) } { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) + \\sum _ { \\boldsymbol { x } ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\{ \\tilde { x } _ { y } \\} } ( 1 - \\Gamma ( \\boldsymbol { x } ^ { \\prime } , y ) ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\boldsymbol { x } ^ { \\prime } , y ) ) }", + "type": "interline_equation", + "image_path": "50d884eead6fd8b8db5a7642e5b04245f374580b54d8c1e73dc93eeb51e99deb.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 105, + 498, + 492, + 508.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 105, + 508.0, + 492, + 518.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 105, + 518.0, + 492, + 528.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "where we use the same notation as in Equation 2. Following SimCLR, we use a non-linear projection", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 200, + 556 + ], + "score": 1.0, + "content": "of the representation as", + "type": "text" + }, + { + "bbox": [ + 200, + 545, + 210, + 554 + ], + "score": 0.84, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 542, + 506, + 556 + ], + "score": 1.0, + "content": "(Figure A.1). The agent’s policy is an affine function of the representation.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 532, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 505, + 586 + ], + "lines": [ + { + "bbox": [ + 103, + 555, + 508, + 577 + ], + "spans": [ + { + "bbox": [ + 103, + 555, + 229, + 577 + ], + "score": 1.0, + "content": "The total contrastive loss for", + "type": "text" + }, + { + "bbox": [ + 229, + 560, + 250, + 571 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 555, + 270, + 577 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 560, + 290, + 572 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 555, + 423, + 577 + ], + "score": 1.0, + "content": "utilizes the optimal trajectories", + "type": "text" + }, + { + "bbox": [ + 423, + 559, + 485, + 572 + ], + "score": 0.94, + "content": "\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 555, + 508, + 577 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 570, + 483, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 163, + 586 + ], + "score": 0.93, + "content": "\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 570, + 194, + 587 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 194, + 572, + 272, + 585 + ], + "score": 0.92, + "content": "x _ { t + 1 } \\sim P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 570, + 289, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 571, + 365, + 586 + ], + "score": 0.93, + "content": "y _ { t + 1 } \\sim P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 570, + 398, + 587 + ], + "score": 1.0, + "content": ". We set", + "type": "text" + }, + { + "bbox": [ + 398, + 572, + 435, + 585 + ], + "score": 0.92, + "content": "\\mathcal { X } ^ { \\prime } = \\tau _ { \\mathcal { X } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 570, + 483, + 587 + ], + "score": 1.0, + "content": "and define", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 103, + 555, + 508, + 587 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 152, + 589, + 458, + 611 + ], + "lines": [ + { + "bbox": [ + 152, + 589, + 458, + 611 + ], + "spans": [ + { + "bbox": [ + 152, + 589, + 458, + 611 + ], + "score": 0.88, + "content": "L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) = \\mathbb { E } _ { y \\sim \\tau _ { \\mathcal { Y } } ^ { * } } \\left[ \\ell _ { \\theta } ( \\tilde { x } _ { y } , y ; \\tau _ { \\mathcal { X } } ^ { * } ) \\right] \\qquad \\mathrm { w h e r e } \\ \\tilde { x } _ { y } = \\underset { x \\in \\tau _ { \\mathcal { X } } ^ { * } } { \\mathrm { a r g m a x } } \\ \\Gamma ( x , y ) .", + "type": "interline_equation", + "image_path": "9d03779755fcd19362a56a25c5ae58750a0c2ce96aa98c5ed4bdc23a3e1cce64.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 152, + 589, + 458, + 611 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 450, + 627 + ], + "score": 1.0, + "content": "We refer to CMEs learned with policy similarity metric as policy similarity embeddings", + "type": "text" + }, + { + "bbox": [ + 450, + 614, + 477, + 625 + ], + "score": 0.67, + "content": "( P S E s )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 613, + 505, + 627 + ], + "score": 1.0, + "content": ". PSEs", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 625, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 638 + ], + "score": 1.0, + "content": "can trivially be combined with data augmentation by using augmented states when computing state", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 395, + 649 + ], + "score": 1.0, + "content": "embeddings. We simultaneously learn PSEs with the RL agent by adding", + "type": "text" + }, + { + "bbox": [ + 396, + 636, + 424, + 647 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "(Algorithm 1) as an", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 646, + 490, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 490, + 660 + ], + "score": 1.0, + "content": "auxiliary objective during training. Next, we illustrate the benefits from this auxiliary objective.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 613, + 505, + 660 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 668, + 363, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 363, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 363, + 682 + ], + "score": 1.0, + "content": "5 JUMPING TASK FROM PIXELS: A CASE STUDY", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "score": 1.0, + "content": "Task Description. The jumping task (Tachet des Combes et al., 2018) (Figure 1) captures, using", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "well-defined factors of variations, whether agents can learn the correct invariances required for", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "generalization, directly from image inputs. The task consists of an agent trying to jump over an", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "obstacle. The agent has access to two actions: right and jump. The agent needs to time the jump", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "precisely, at a specific distance from the obstacle, otherwise it will eventually hit the obstacle.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "Different tasks consist in shifting the floor height and/or the obstacle position. To generalize, the", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "agent needs to be invariant to the floor height while jump based on the obstacle position. The obstacle", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 418, + 467, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 467, + 430 + ], + "score": 1.0, + "content": "can be in 26 different locations while the floor has 11 different heights, totaling 286 tasks.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 56.5, + "bbox_fs": [ + 105, + 686, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 117, + 495, + 306 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 117, + 495, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 117, + 495, + 306 + ], + "spans": [ + { + "bbox": [ + 111, + 117, + 495, + 306 + ], + "score": 0.877, + "type": "image", + "image_path": "fe5d30be2f8515ffad0c843ba9528ed9477ed5a3bbb49088ef023877b2fee29f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 117, + 495, + 180.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 180.0, + 495, + 243.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 243.0, + 495, + 306.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 311, + 505, + 381 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 311, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 506, + 322 + ], + "score": 1.0, + "content": "Figure 2: Jumping Task: Visualization of average performance of PSEs with data augmentation across different", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 321, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 331 + ], + "score": 1.0, + "content": "configurations. We plot the median performance across 100 runs. Each tile in the grid represents a different task", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 466, + 342 + ], + "score": 1.0, + "content": "(obstacle position/floor height combination). For each grid configuration, the height varies along the", + "type": "text" + }, + { + "bbox": [ + 467, + 332, + 473, + 341 + ], + "score": 0.79, + "content": "_ y", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "-axis (11", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 291, + 353 + ], + "score": 1.0, + "content": "heights) while the obstacle position varies along the", + "type": "text" + }, + { + "bbox": [ + 291, + 343, + 298, + 350 + ], + "score": 0.75, + "content": "x -", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 339, + 419, + 353 + ], + "score": 1.0, + "content": "-axis (26 locations). The red letter", + "type": "text" + }, + { + "bbox": [ + 420, + 341, + 428, + 350 + ], + "score": 0.67, + "content": "\\top", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "indicates the training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "score": 1.0, + "content": "tasks. Random grid depicts only one instance, each run consisted of a different test/train split. Beige tiles are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "tasks PSEs solved while black tiles are tasks PSEs did not solve when used with data augmentation. These", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 370, + 498, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 498, + 382 + ], + "score": 1.0, + "content": "results were chosen across all the 100 runs to demonstrate what the average reported performance looks like.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "precisely, at a specific distance from the obstacle, otherwise it will eventually hit the obstacle.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "Different tasks consist in shifting the floor height and/or the obstacle position. To generalize, the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "agent needs to be invariant to the floor height while jump based on the obstacle position. The obstacle", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 418, + 467, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 467, + 430 + ], + "score": 1.0, + "content": "can be in 26 different locations while the floor has 11 different heights, totaling 286 tasks.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "Problem Setup. We split the problem into 18 seen (training) and 268 unseen (test) tasks to stress test", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "generalization using a few changes in the underlying factors of variations seen during training. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "small number of positive examples3 results in a highly unbalanced classification problem with low", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "amounts of data, making it challenging without additional inductive biases. Thus, we evaluate gener-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "alization in regimes with and without data augmentation. The different grids configurations (Figure 2)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "capture different types of generalization: the “wide” grid tests generalization via “interpolation”,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "the “narrow” grid tests out-of-distribution generalization via “extrapolation”, and the random grid", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "instances evaluate generalization similar to supervised learning where train and test samples are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 523, + 264, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 264, + 535 + ], + "score": 1.0, + "content": "drawn i.i.d. from the same distribution.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 540, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "We used RandConv (Lee et al., 2020a), a state-of-the-art data augmentation for generalization. For", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "score": 1.0, + "content": "“wide” grid (Figure 2a) and pick the parameters with the best validation performance. We use these", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 571, + 507, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 507, + 587 + ], + "score": 1.0, + "content": "fixed parameters for all grid configurations to show the robustness of PSEs to hyperparameter tuning.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "We first compute the optimal trajectories in the training tasks. Using these trajectories, we compute", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "PSM using dynamic programming (Section D.1). We train the agent by imitation learning, combined", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 606, + 403, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 403, + 617 + ], + "score": 1.0, + "content": "with an auxiliary loss for PSEs (Section 4). More details are in Section G.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 106, + 621, + 349, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 351, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 351, + 633 + ], + "score": 1.0, + "content": "5.1 EVALUATING GENERALIZATION ON JUMPING TASK", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 507, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 507, + 649 + ], + "score": 1.0, + "content": "We show the efficacy of PSEs compared to common generalization approaches such as regulariza-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "tion (e.g., Cobbe et al., 2019; Farebrother et al., 2018), and data augmentation (e.g., Lee et al., 2020a;", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Laskin et al., 2020a), which are quite effective on pixel-based RL tasks. We also contrast PSEs with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "bisimulation transfer (Castro & Precup, 2010), a tabular state-based transfer approach based on bisim-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "ulation metrics which does not do any learning and bisimulation preserving representations (Zhang", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 691, + 447, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 447, + 704 + ], + "score": 1.0, + "content": "et al., 2021), showing the advantages of PSM over a prevalent state similarity metric.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 712, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 505, + 724 + ], + "score": 1.0, + "content": "3We have 18 different trajectories with several examples for the action right, but only one instance of jump", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 370, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 370, + 733 + ], + "score": 1.0, + "content": "action per trajectory, leading to just 18 total instances of the action jump.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 117, + 495, + 306 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 117, + 495, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 117, + 495, + 306 + ], + "spans": [ + { + "bbox": [ + 111, + 117, + 495, + 306 + ], + "score": 0.877, + "type": "image", + "image_path": "fe5d30be2f8515ffad0c843ba9528ed9477ed5a3bbb49088ef023877b2fee29f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 117, + 495, + 180.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 180.0, + 495, + 243.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 243.0, + 495, + 306.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 311, + 505, + 381 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 311, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 506, + 322 + ], + "score": 1.0, + "content": "Figure 2: Jumping Task: Visualization of average performance of PSEs with data augmentation across different", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 321, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 331 + ], + "score": 1.0, + "content": "configurations. We plot the median performance across 100 runs. Each tile in the grid represents a different task", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 466, + 342 + ], + "score": 1.0, + "content": "(obstacle position/floor height combination). For each grid configuration, the height varies along the", + "type": "text" + }, + { + "bbox": [ + 467, + 332, + 473, + 341 + ], + "score": 0.79, + "content": "_ y", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "-axis (11", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 291, + 353 + ], + "score": 1.0, + "content": "heights) while the obstacle position varies along the", + "type": "text" + }, + { + "bbox": [ + 291, + 343, + 298, + 350 + ], + "score": 0.75, + "content": "x -", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 339, + 419, + 353 + ], + "score": 1.0, + "content": "-axis (26 locations). The red letter", + "type": "text" + }, + { + "bbox": [ + 420, + 341, + 428, + 350 + ], + "score": 0.67, + "content": "\\top", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "indicates the training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "score": 1.0, + "content": "tasks. Random grid depicts only one instance, each run consisted of a different test/train split. Beige tiles are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "tasks PSEs solved while black tiles are tasks PSEs did not solve when used with data augmentation. These", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 370, + 498, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 498, + 382 + ], + "score": 1.0, + "content": "results were chosen across all the 100 runs to demonstrate what the average reported performance looks like.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 429 + ], + "lines": [], + "index": 11.5, + "bbox_fs": [ + 105, + 385, + 506, + 430 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "Problem Setup. We split the problem into 18 seen (training) and 268 unseen (test) tasks to stress test", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "generalization using a few changes in the underlying factors of variations seen during training. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "small number of positive examples3 results in a highly unbalanced classification problem with low", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "amounts of data, making it challenging without additional inductive biases. Thus, we evaluate gener-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "alization in regimes with and without data augmentation. The different grids configurations (Figure 2)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "capture different types of generalization: the “wide” grid tests generalization via “interpolation”,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 513 + ], + "score": 1.0, + "content": "the “narrow” grid tests out-of-distribution generalization via “extrapolation”, and the random grid", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "instances evaluate generalization similar to supervised learning where train and test samples are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 523, + 264, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 264, + 535 + ], + "score": 1.0, + "content": "drawn i.i.d. from the same distribution.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 435, + 506, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 540, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "We used RandConv (Lee et al., 2020a), a state-of-the-art data augmentation for generalization. For", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 104, + 560, + 506, + 575 + ], + "score": 1.0, + "content": "“wide” grid (Figure 2a) and pick the parameters with the best validation performance. We use these", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 571, + 507, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 507, + 587 + ], + "score": 1.0, + "content": "fixed parameters for all grid configurations to show the robustness of PSEs to hyperparameter tuning.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "We first compute the optimal trajectories in the training tasks. Using these trajectories, we compute", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "PSM using dynamic programming (Section D.1). We train the agent by imitation learning, combined", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 606, + 403, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 403, + 617 + ], + "score": 1.0, + "content": "with an auxiliary loss for PSEs (Section 4). More details are in Section G.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 539, + 507, + 617 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 621, + 349, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 351, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 351, + 633 + ], + "score": 1.0, + "content": "5.1 EVALUATING GENERALIZATION ON JUMPING TASK", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 507, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 507, + 649 + ], + "score": 1.0, + "content": "We show the efficacy of PSEs compared to common generalization approaches such as regulariza-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "tion (e.g., Cobbe et al., 2019; Farebrother et al., 2018), and data augmentation (e.g., Lee et al., 2020a;", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Laskin et al., 2020a), which are quite effective on pixel-based RL tasks. We also contrast PSEs with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "bisimulation transfer (Castro & Precup, 2010), a tabular state-based transfer approach based on bisim-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "ulation metrics which does not do any learning and bisimulation preserving representations (Zhang", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 691, + 447, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 447, + 704 + ], + "score": 1.0, + "content": "et al., 2021), showing the advantages of PSM over a prevalent state similarity metric.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 636, + 507, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 502, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 502, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 502, + 214 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 502, + 214 + ], + "score": 0.964, + "type": "image", + "image_path": "f6cb8503f0d51f326842d71126a63d55028d7bdc17f24dcab3810ef2b38ded2e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 502, + 126.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 126.0, + 502, + 170.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 170.0, + 502, + 214.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 223, + 505, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "Figure 3: Embedding visualization. (a) Optimal trajectories on original jumping task (visualized as coloured", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "blocks) with different obstacle positions. We visualize the hidden representations using UMAP, where the color", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "of points indicate the tasks of the corresponding observations. Points with the same number label correspond to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 254, + 461, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 461, + 264 + ], + "score": 1.0, + "content": "same distance of the agent from the obstacle, the underlying optimal invariant feature across tasks.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 506, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "We first investigated how well PSEs generalize over existing methods without incorporating additional", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "domain knowledge during training. Table 1 summarizes, in the setting without data augmentation,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 290, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 350, + 302 + ], + "score": 1.0, + "content": "the performance of these methods in different train/test splits", + "type": "text" + }, + { + "bbox": [ + 351, + 290, + 363, + 302 + ], + "score": 0.65, + "content": "\\cdot \\cdot f .", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 290, + 506, + 302 + ], + "score": 1.0, + "content": "Figure 2 for a detailed description).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 300, + 487, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 487, + 313 + ], + "score": 1.0, + "content": "PSEs, with only 18 examples, already leads to better performance than standard regularization.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "PSEs also outperform bisimulation transfer in the “wide” and random grids. Although bisimulation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "score": 1.0, + "content": "transfer is impractical4 when evaluating zero-shot generalization, we still performed this comparison,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "unfair to PSEs, to highlight their efficacy. PSEs perform better because, in contrast to bisimulation,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 212, + 363 + ], + "score": 1.0, + "content": "PSM is reward agnostic (", + "type": "text" + }, + { + "bbox": [ + 212, + 351, + 225, + 362 + ], + "score": 0.28, + "content": "( c . f .", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "Proposition D.1) – the expected return of the jump action is quite", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 297, + 375 + ], + "score": 1.0, + "content": "different depending on the obstacle position", + "type": "text" + }, + { + "bbox": [ + 297, + 362, + 311, + 373 + ], + "score": 0.75, + "content": "( c . f .", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "Figure G.2 for a visual juxtaposition of PSM", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "and bisimulation). Overall, these results are promising because they place PSEs as an effective", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 384, + 360, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 360, + 396 + ], + "score": 1.0, + "content": "generalization method that does not rely on data augmentation.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "Nevertheless, PSEs are complementary to data augmentation, which consistently improves gener-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "alization in deep RL. We compared RandConv combined with PSEs to simply using RandConv.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 423, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 506, + 435 + ], + "score": 1.0, + "content": "Domain-specific augmentation also succeeds in the jumping task. Thus, it is not surprising that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 439, + 447 + ], + "score": 1.0, + "content": "RandConv is so effective compared to techniques without augmentation. Table 1 (", + "type": "text" + }, + { + "bbox": [ + 439, + 433, + 455, + 444 + ], + "score": 0.81, + "content": "2 ^ { n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 432, + 506, + 447 + ], + "score": 1.0, + "content": "row) shows", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "that PSEs substantially improve the performance of RandConv across all grid configurations. More-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 453, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 167, + 469 + ], + "score": 1.0, + "content": "over, Table 1 (", + "type": "text" + }, + { + "bbox": [ + 167, + 455, + 183, + 466 + ], + "score": 0.82, + "content": "2 ^ { n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 453, + 506, + 469 + ], + "score": 1.0, + "content": "row) illustrates that when combined with RandConv, bisimulation preserving", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 466, + 470, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 359, + 478 + ], + "score": 1.0, + "content": "representations (Zhang et al., 2021) diminish generalization by", + "type": "text" + }, + { + "bbox": [ + 360, + 466, + 401, + 477 + ], + "score": 0.9, + "content": "3 0 - 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 467, + 470, + 478 + ], + "score": 1.0, + "content": "relative to PSEs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 177, + 496 + ], + "score": 1.0, + "content": "Notably, Table 1 (", + "type": "text" + }, + { + "bbox": [ + 177, + 483, + 190, + 494 + ], + "score": 0.8, + "content": "\\cdot ^ { 1 ^ { s t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 482, + 505, + 496 + ], + "score": 1.0, + "content": "row) indicates that learning-based methods are ineffective on the “narrow” grid", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 494, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 506, + 506 + ], + "score": 1.0, + "content": "without data augmentation. That said, PSEs do work quite well when combined with RandConv.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "score": 1.0, + "content": "However, even with data augmentation, generalization in “narrow” grid happens only around the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "vicinity of training tasks, exhibiting the challenge this grid poses for learning-based methods. We", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "believe this is due to the poor extrapolation ability of neural networks (e.g., Haley & Soloway, 1992;", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 537, + 492, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 492, + 550 + ], + "score": 1.0, + "content": "Xu et al., 2020), which is more perceptible without prior inductive bias from data augmentation.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 554, + 442, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 444, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 444, + 567 + ], + "score": 1.0, + "content": "5.2 UNDERSTANDING GAINS FROM PSES: ABLATIONS AND VISUALIZATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 254, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 254, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 254, + 582 + ], + "score": 1.0, + "content": "PSEs are contrastive metric embed-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 582, + 255, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 255, + 593 + ], + "score": 1.0, + "content": "dings (CMEs) learned with PSM.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "score": 1.0, + "content": "We investigate the gains from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 603, + 255, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 255, + 615 + ], + "score": 1.0, + "content": "CMEs (Section 4) and PSM (Sec-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 614, + 254, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 254, + 627 + ], + "score": 1.0, + "content": "tion 3) by ablating them. CMEs", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 626, + 254, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 254, + 637 + ], + "score": 1.0, + "content": "can be learned with any state", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 637, + 255, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 238, + 648 + ], + "score": 1.0, + "content": "similarity metric – we use", + "type": "text" + }, + { + "bbox": [ + 239, + 637, + 250, + 647 + ], + "score": 0.8, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 637, + 255, + 648 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 648, + 255, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 255, + 659 + ], + "score": 1.0, + "content": "bisimulation (Castro, 2020) as an al-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 659, + 253, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 253, + 670 + ], + "score": 1.0, + "content": "ternative. Similarly, PSM can be", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 669, + 255, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 255, + 682 + ], + "score": 1.0, + "content": "used with any metric embedding –", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 680, + 254, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 138, + 692 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 139, + 681, + 149, + 691 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 680, + 254, + 692 + ], + "score": 1.0, + "content": "-embeddings (Section E)", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 691, + 255, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 255, + 703 + ], + "score": 1.0, + "content": "as an alternative, which Zhang et al.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 261, + 570, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 261, + 571, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 261, + 571, + 396, + 581 + ], + "score": 1.0, + "content": "Table 2: Ablating PSEs. Percentage", + "type": "text" + }, + { + "bbox": [ + 396, + 571, + 410, + 581 + ], + "score": 0.73, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 571, + 505, + 581 + ], + "score": 1.0, + "content": "of test tasks solved when", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 260, + 581, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 260, + 581, + 505, + 592 + ], + "score": 1.0, + "content": "we ablate the similarity metric and learning procedure for metric", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 261, + 591, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 261, + 591, + 505, + 601 + ], + "score": 1.0, + "content": "embeddings in the data augmentation setting on wide grid. PSEs,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 260, + 600, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 260, + 600, + 506, + 612 + ], + "score": 1.0, + "content": "which combine CMEs with PSM, considerably outperform other", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 260, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 260, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "embeddings. We report the average performance across 100 runs with", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 261, + 621, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 261, + 621, + 505, + 631 + ], + "score": 1.0, + "content": "standard deviation between parentheses. All ablations except PSEs", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 260, + 631, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 260, + 631, + 505, + 641 + ], + "score": 1.0, + "content": "deteriorate performance compared to just using data augmentation", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 260, + 640, + 391, + 652 + ], + "spans": [ + { + "bbox": [ + 260, + 640, + 391, + 652 + ], + "score": 1.0, + "content": "(RandConv), as reported in Table 1.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5 + }, + { + "type": "table", + "bbox": [ + 280, + 656, + 487, + 701 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 280, + 656, + 487, + 701 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 280, + 656, + 487, + 701 + ], + "spans": [ + { + "bbox": [ + 280, + 656, + 487, + 701 + ], + "score": 0.968, + "html": "
Metric /Embeddingl2-embeddingsCMEs
π*-bisimulation41.4 (17.6)23.1 (7.6)
PSM17.5 (8.4)87.0 (10.1)
", + "type": "table", + "image_path": "6e8f288563fbbcb7b1a69e06fdaeb32e47b99dd2f34b1560b3ad93961cec8a3f.jpg" + } + ] + } + ], + "index": 53, + "virtual_lines": [ + { + "bbox": [ + 280, + 656, + 487, + 671.0 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 280, + 671.0, + 487, + 686.0 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 280, + 686.0, + 487, + 701.0 + ], + "spans": [], + "index": 54 + } + ] + } + ], + "index": 53 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 712, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "4Bisimulation transfer assumes oracle access to dynamics and rewards on unseen environments as well as", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 369, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 369, + 732 + ], + "score": 1.0, + "content": "tabular state space to compute the exact bisimulation metric (Section C).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 82, + 502, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 82, + 502, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 82, + 502, + 214 + ], + "spans": [ + { + "bbox": [ + 108, + 82, + 502, + 214 + ], + "score": 0.964, + "type": "image", + "image_path": "f6cb8503f0d51f326842d71126a63d55028d7bdc17f24dcab3810ef2b38ded2e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 82, + 502, + 126.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 126.0, + 502, + 170.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 170.0, + 502, + 214.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 223, + 505, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "Figure 3: Embedding visualization. (a) Optimal trajectories on original jumping task (visualized as coloured", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "blocks) with different obstacle positions. We visualize the hidden representations using UMAP, where the color", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "of points indicate the tasks of the corresponding observations. Points with the same number label correspond to", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 254, + 461, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 461, + 264 + ], + "score": 1.0, + "content": "same distance of the agent from the obstacle, the underlying optimal invariant feature across tasks.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 506, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "We first investigated how well PSEs generalize over existing methods without incorporating additional", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "domain knowledge during training. Table 1 summarizes, in the setting without data augmentation,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 290, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 350, + 302 + ], + "score": 1.0, + "content": "the performance of these methods in different train/test splits", + "type": "text" + }, + { + "bbox": [ + 351, + 290, + 363, + 302 + ], + "score": 0.65, + "content": "\\cdot \\cdot f .", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 290, + 506, + 302 + ], + "score": 1.0, + "content": "Figure 2 for a detailed description).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 300, + 487, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 487, + 313 + ], + "score": 1.0, + "content": "PSEs, with only 18 examples, already leads to better performance than standard regularization.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 268, + 506, + 313 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "PSEs also outperform bisimulation transfer in the “wide” and random grids. Although bisimulation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 507, + 342 + ], + "score": 1.0, + "content": "transfer is impractical4 when evaluating zero-shot generalization, we still performed this comparison,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "unfair to PSEs, to highlight their efficacy. PSEs perform better because, in contrast to bisimulation,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 212, + 363 + ], + "score": 1.0, + "content": "PSM is reward agnostic (", + "type": "text" + }, + { + "bbox": [ + 212, + 351, + 225, + 362 + ], + "score": 0.28, + "content": "( c . f .", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "Proposition D.1) – the expected return of the jump action is quite", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 297, + 375 + ], + "score": 1.0, + "content": "different depending on the obstacle position", + "type": "text" + }, + { + "bbox": [ + 297, + 362, + 311, + 373 + ], + "score": 0.75, + "content": "( c . f .", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "Figure G.2 for a visual juxtaposition of PSM", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "and bisimulation). Overall, these results are promising because they place PSEs as an effective", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 384, + 360, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 360, + 396 + ], + "score": 1.0, + "content": "generalization method that does not rely on data augmentation.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 317, + 507, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "Nevertheless, PSEs are complementary to data augmentation, which consistently improves gener-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "alization in deep RL. We compared RandConv combined with PSEs to simply using RandConv.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 423, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 506, + 435 + ], + "score": 1.0, + "content": "Domain-specific augmentation also succeeds in the jumping task. Thus, it is not surprising that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 439, + 447 + ], + "score": 1.0, + "content": "RandConv is so effective compared to techniques without augmentation. Table 1 (", + "type": "text" + }, + { + "bbox": [ + 439, + 433, + 455, + 444 + ], + "score": 0.81, + "content": "2 ^ { n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 432, + 506, + 447 + ], + "score": 1.0, + "content": "row) shows", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "that PSEs substantially improve the performance of RandConv across all grid configurations. More-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 453, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 167, + 469 + ], + "score": 1.0, + "content": "over, Table 1 (", + "type": "text" + }, + { + "bbox": [ + 167, + 455, + 183, + 466 + ], + "score": 0.82, + "content": "2 ^ { n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 453, + 506, + 469 + ], + "score": 1.0, + "content": "row) illustrates that when combined with RandConv, bisimulation preserving", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 466, + 470, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 359, + 478 + ], + "score": 1.0, + "content": "representations (Zhang et al., 2021) diminish generalization by", + "type": "text" + }, + { + "bbox": [ + 360, + 466, + 401, + 477 + ], + "score": 0.9, + "content": "3 0 - 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 467, + 470, + 478 + ], + "score": 1.0, + "content": "relative to PSEs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 400, + 506, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 177, + 496 + ], + "score": 1.0, + "content": "Notably, Table 1 (", + "type": "text" + }, + { + "bbox": [ + 177, + 483, + 190, + 494 + ], + "score": 0.8, + "content": "\\cdot ^ { 1 ^ { s t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 482, + 505, + 496 + ], + "score": 1.0, + "content": "row) indicates that learning-based methods are ineffective on the “narrow” grid", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 494, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 506, + 506 + ], + "score": 1.0, + "content": "without data augmentation. That said, PSEs do work quite well when combined with RandConv.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "score": 1.0, + "content": "However, even with data augmentation, generalization in “narrow” grid happens only around the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "vicinity of training tasks, exhibiting the challenge this grid poses for learning-based methods. We", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "believe this is due to the poor extrapolation ability of neural networks (e.g., Haley & Soloway, 1992;", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 537, + 492, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 492, + 550 + ], + "score": 1.0, + "content": "Xu et al., 2020), which is more perceptible without prior inductive bias from data augmentation.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 482, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 554, + 442, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 444, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 444, + 567 + ], + "score": 1.0, + "content": "5.2 UNDERSTANDING GAINS FROM PSES: ABLATIONS AND VISUALIZATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 554, + 444, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 254, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 254, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 254, + 582 + ], + "score": 1.0, + "content": "PSEs are contrastive metric embed-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 582, + 255, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 255, + 593 + ], + "score": 1.0, + "content": "dings (CMEs) learned with PSM.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "score": 1.0, + "content": "We investigate the gains from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 603, + 255, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 255, + 615 + ], + "score": 1.0, + "content": "CMEs (Section 4) and PSM (Sec-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 614, + 254, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 254, + 627 + ], + "score": 1.0, + "content": "tion 3) by ablating them. CMEs", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 626, + 254, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 254, + 637 + ], + "score": 1.0, + "content": "can be learned with any state", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 637, + 255, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 238, + 648 + ], + "score": 1.0, + "content": "similarity metric – we use", + "type": "text" + }, + { + "bbox": [ + 239, + 637, + 250, + 647 + ], + "score": 0.8, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 637, + 255, + 648 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 648, + 255, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 255, + 659 + ], + "score": 1.0, + "content": "bisimulation (Castro, 2020) as an al-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 659, + 253, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 253, + 670 + ], + "score": 1.0, + "content": "ternative. Similarly, PSM can be", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 669, + 255, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 255, + 682 + ], + "score": 1.0, + "content": "used with any metric embedding –", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 680, + 254, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 138, + 692 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 139, + 681, + 149, + 691 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 680, + 254, + 692 + ], + "score": 1.0, + "content": "-embeddings (Section E)", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 691, + 255, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 255, + 703 + ], + "score": 1.0, + "content": "as an alternative, which Zhang et al.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 571, + 255, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 261, + 570, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 261, + 571, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 261, + 571, + 396, + 581 + ], + "score": 1.0, + "content": "Table 2: Ablating PSEs. Percentage", + "type": "text" + }, + { + "bbox": [ + 396, + 571, + 410, + 581 + ], + "score": 0.73, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 571, + 505, + 581 + ], + "score": 1.0, + "content": "of test tasks solved when", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 260, + 581, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 260, + 581, + 505, + 592 + ], + "score": 1.0, + "content": "we ablate the similarity metric and learning procedure for metric", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 261, + 591, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 261, + 591, + 505, + 601 + ], + "score": 1.0, + "content": "embeddings in the data augmentation setting on wide grid. PSEs,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 260, + 600, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 260, + 600, + 506, + 612 + ], + "score": 1.0, + "content": "which combine CMEs with PSM, considerably outperform other", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 260, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 260, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "embeddings. We report the average performance across 100 runs with", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 261, + 621, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 261, + 621, + 505, + 631 + ], + "score": 1.0, + "content": "standard deviation between parentheses. All ablations except PSEs", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 260, + 631, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 260, + 631, + 505, + 641 + ], + "score": 1.0, + "content": "deteriorate performance compared to just using data augmentation", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 260, + 640, + 391, + 652 + ], + "spans": [ + { + "bbox": [ + 260, + 640, + 391, + 652 + ], + "score": 1.0, + "content": "(RandConv), as reported in Table 1.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5, + "bbox_fs": [ + 260, + 571, + 506, + 652 + ] + }, + { + "type": "table", + "bbox": [ + 280, + 656, + 487, + 701 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 280, + 656, + 487, + 701 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 280, + 656, + 487, + 701 + ], + "spans": [ + { + "bbox": [ + 280, + 656, + 487, + 701 + ], + "score": 0.968, + "html": "
Metric /Embeddingl2-embeddingsCMEs
π*-bisimulation41.4 (17.6)23.1 (7.6)
PSM17.5 (8.4)87.0 (10.1)
", + "type": "table", + "image_path": "6e8f288563fbbcb7b1a69e06fdaeb32e47b99dd2f34b1560b3ad93961cec8a3f.jpg" + } + ] + } + ], + "index": 53, + "virtual_lines": [ + { + "bbox": [ + 280, + 656, + 487, + 671.0 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 280, + 671.0, + 487, + 686.0 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 280, + 686.0, + 487, + 701.0 + ], + "spans": [], + "index": 54 + } + ] + } + ], + "index": 53 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 196, + 95 + ], + "score": 1.0, + "content": "(2021) employed with", + "type": "text" + }, + { + "bbox": [ + 196, + 83, + 208, + 93 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "-bisimulation for learning representations in a single task RL setting. For a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 468, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 468, + 106 + ], + "score": 1.0, + "content": "fair comparison, we tune hyperparameters for 128 trials for each ablation entry in Table 2.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 211, + 123 + ], + "score": 1.0, + "content": "Table 2 show that PSEs", + "type": "text" + }, + { + "bbox": [ + 212, + 111, + 287, + 122 + ], + "score": 0.83, + "content": "( = \\mathrm { P S M } + \\mathrm { C M E s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 110, + 437, + 123 + ], + "score": 1.0, + "content": "generalize significantly better than", + "type": "text" + }, + { + "bbox": [ + 438, + 111, + 449, + 120 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "-bisimulation", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 165, + 135 + ], + "score": 1.0, + "content": "with CMEs or", + "type": "text" + }, + { + "bbox": [ + 165, + 122, + 175, + 132 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 121, + 430, + 135 + ], + "score": 1.0, + "content": "-embeddings, both of which significantly degrade performance (", + "type": "text" + }, + { + "bbox": [ + 430, + 121, + 457, + 132 + ], + "score": 0.87, + "content": "- 6 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 121, + 475, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 121, + 502, + 132 + ], + "score": 0.89, + "content": "- 4 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 121, + 506, + 135 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 242, + 146 + ], + "score": 1.0, + "content": "respectively). This is expected, as", + "type": "text" + }, + { + "bbox": [ + 242, + 133, + 254, + 142 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "-bisimulation imposes the incorrect invariances for the jumping", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 128, + 156 + ], + "score": 1.0, + "content": "task", + "type": "text" + }, + { + "bbox": [ + 128, + 144, + 142, + 155 + ], + "score": 0.72, + "content": "\\cdot { c . f . }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "Figures G.2a and G.2d). Additionally, looking at the rows of Table 2, CMEs are superior", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 127, + 165 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 155, + 137, + 165 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 154, + 306, + 165 + ], + "score": 1.0, + "content": "-embeddings for PSM while inferior for", + "type": "text" + }, + { + "bbox": [ + 307, + 155, + 318, + 164 + ], + "score": 0.88, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 154, + 504, + 165 + ], + "score": 1.0, + "content": "-bisimulation. This result is in line with the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "hypothesis that CMEs better enforce the invariances encoded by a similarity metric compared to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 176, + 267, + 189 + ], + "spans": [ + { + "bbox": [ + 107, + 177, + 116, + 187 + ], + "score": 0.86, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 176, + 172, + 189 + ], + "score": 1.0, + "content": "-embeddings", + "type": "text" + }, + { + "bbox": [ + 172, + 177, + 185, + 188 + ], + "score": 0.51, + "content": "\\left( c . f . \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 176, + 267, + 189 + ], + "score": 1.0, + "content": "Figures 3b and 3c).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "Visualizing learned representations. We visualize the metric embeddings in the ablation above", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "score": 1.0, + "content": "by projecting them to two dimensions with UMAP (McInnes et al., 2018), a popular visualization", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "technique for high dimensional data which better preserves the data’s global structure compared to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 330, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 330, + 237 + ], + "score": 1.0, + "content": "other methods such as t-SNE (Coenen & Pearce, 2019).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "Figure 3 shows that PSEs partition the states into two sets: (1) states where a single suboptimal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "score": 1.0, + "content": "action leads to failure (all states before jump) and (2) states where actions do not affect the final", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 266, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 506, + 277 + ], + "score": 1.0, + "content": "outcome (states after jump). Additionally, PSEs align the labeled states in the first set, which have a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "PSM distance of zero. These aligned states have the same distance from the obstacle, the invariant", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 336, + 299 + ], + "score": 1.0, + "content": "feature that generalizes across tasks. On the other hand,", + "type": "text" + }, + { + "bbox": [ + 337, + 287, + 347, + 298 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "-embeddings (Zhang et al., 2021) with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "PSM do not align states with zero PSM except the state with the jump action – poor generalization,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "as observed empirically, is likely due to states with the same optimal behavior ending up with distant", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 209, + 331 + ], + "score": 1.0, + "content": "embeddings. CMEs with", + "type": "text" + }, + { + "bbox": [ + 209, + 320, + 221, + 330 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 320, + 343, + 331 + ], + "score": 1.0, + "content": "-bisimulation align states with", + "type": "text" + }, + { + "bbox": [ + 344, + 320, + 355, + 330 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 320, + 505, + 331 + ], + "score": 1.0, + "content": "-bisimulation distance of zero – such", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "states are equidistant from the start state and have different optimal behavior for any pair of tasks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 342, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 295, + 354 + ], + "score": 1.0, + "content": "with different obstacle positions (Figure G.2c).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 105, + 355, + 480, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 480, + 366 + ], + "score": 1.0, + "content": "5.3 JUMPING TASK WITH COLORS: WHERE TASK-DEPENDENT INVARIANCE MATTERS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 324, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 325, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 325, + 384 + ], + "score": 1.0, + "content": "PSEs capture invariances that are usually orthogonal to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 383, + 325, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 325, + 394 + ], + "score": 1.0, + "content": "task-agnostic invariances from data augmentation. This", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 393, + 325, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 325, + 406 + ], + "score": 1.0, + "content": "difference is important because, for certain tasks, data", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 406, + 325, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 325, + 416 + ], + "score": 1.0, + "content": "augmentation can erroneously alias states with different", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 416, + 325, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 325, + 428 + ], + "score": 1.0, + "content": "optimal behavior. Domain knowledge is often required", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 427, + 326, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 326, + 438 + ], + "score": 1.0, + "content": "to select appropriate augmentations, otherwise augmen-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 437, + 325, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 325, + 449 + ], + "score": 1.0, + "content": "tations can hurt generalization. In contrast, PSEs do", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 448, + 325, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 325, + 460 + ], + "score": 1.0, + "content": "not require any domain knowledge but instead exploit", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 460, + 259, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 259, + 470 + ], + "score": 1.0, + "content": "the inherent structure of the RL tasks.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 324, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 325, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 325, + 487 + ], + "score": 1.0, + "content": "To demonstrate the difference between PSEs and data", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 488, + 324, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 324, + 498 + ], + "score": 1.0, + "content": "augmentation, we simply include colored obstacles in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 497, + 325, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 325, + 510 + ], + "score": 1.0, + "content": "the jumping task (see Figure G.5). In this modified", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 509, + 324, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 324, + 521 + ], + "score": 1.0, + "content": "task, the optimal behavior of the agent depends on", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 520, + 325, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 325, + 532 + ], + "score": 1.0, + "content": "the obstacle color: the agent needs to jump over the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 531, + 325, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 325, + 543 + ], + "score": 1.0, + "content": "red obstacle but strike the green obstacle to get a high", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 541, + 325, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 325, + 554 + ], + "score": 1.0, + "content": "return. The red obstacle task has the same difficulty as", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 340, + 372, + 497, + 492 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 340, + 372, + 497, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 340, + 372, + 497, + 492 + ], + "spans": [ + { + "bbox": [ + 340, + 372, + 497, + 492 + ], + "score": 0.967, + "type": "image", + "image_path": "dd27daffdbeb11b7f10511fabc09304681a9967704b663bae3974721a6096f03.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 340, + 372, + 497, + 384.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 340, + 384.0, + 497, + 396.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 340, + 396.0, + 497, + 408.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 340, + 408.0, + 497, + 420.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 340, + 420.0, + 497, + 432.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 340, + 432.0, + 497, + 444.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 340, + 444.0, + 497, + 456.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 340, + 456.0, + 497, + 468.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 340, + 468.0, + 497, + 480.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 340, + 480.0, + 497, + 492.0 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 332, + 496, + 505, + 546 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 331, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 331, + 495, + 414, + 507 + ], + "score": 1.0, + "content": "Figure 4: Percentage", + "type": "text" + }, + { + "bbox": [ + 415, + 496, + 429, + 506 + ], + "score": 0.73, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "of red obstacle test", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 331, + 506, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 331, + 506, + 505, + 516 + ], + "score": 1.0, + "content": "tasks solved when trained, jointly with red and", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 331, + 516, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 331, + 516, + 505, + 526 + ], + "score": 1.0, + "content": "green obstacles, on the “wide” grid. We report", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 331, + 525, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 331, + 525, + 487, + 536 + ], + "score": 1.0, + "content": "the mean across 100 runs. Error bars show", + "type": "text" + }, + { + "bbox": [ + 487, + 526, + 505, + 535 + ], + "score": 0.84, + "content": "9 9 \\%", + "type": "inline_equation" + } + ], + "index": 53 + }, + { + "bbox": [ + 331, + 535, + 453, + 546 + ], + "spans": [ + { + "bbox": [ + 331, + 535, + 453, + 546 + ], + "score": 1.0, + "content": "confidence interval for the mean.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52 + } + ], + "index": 48.25 + }, + { + "type": "text", + "bbox": [ + 108, + 553, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "the original jumping task while the green obstacle task is easier. We jointly train the agent with 18", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 564, + 477, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 477, + 576 + ], + "score": 1.0, + "content": "training tasks each, for both obstacle colors, and evaluate generalization on unseen red tasks.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 55.5 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 506, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "score": 1.0, + "content": "Figure 4 shows the large performance gap between PSEs and data augmentation with RandConv. All", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 592, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 477, + 604 + ], + "score": 1.0, + "content": "methods solve the green obstacle tasks (Table G.1). As opposed to the original jumping task (", + "type": "text" + }, + { + "bbox": [ + 477, + 592, + 490, + 604 + ], + "score": 0.68, + "content": "\\cdot f .", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 592, + 505, + 604 + ], + "score": 1.0, + "content": "Ta-", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "score": 1.0, + "content": "ble 1), data augmentation inhibits generalization since RandConv forces the agent to ignore color,", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 628 + ], + "score": 1.0, + "content": "conflating the red and green tasks (Figure G.6). PSEs still outperform regularization and data aug-", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "mentation. Furthermore, data augmentation performs better when combined with PSEs. Thus, PSEs", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 635, + 357, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 357, + 649 + ], + "score": 1.0, + "content": "are effective even when data augmentation hurts performance.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59.5 + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 329, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 330, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 330, + 661 + ], + "score": 1.0, + "content": "5.4 EFFECT OF POLICY SUBOPTIMALITY ON PSES", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "To understand the sensitivity of learning effective PSEs to the quality of the policies, we compute", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 153, + 690 + ], + "score": 1.0, + "content": "PSEs using", + "type": "text" + }, + { + "bbox": [ + 154, + 679, + 159, + 687 + ], + "score": 0.61, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "-suboptimal policies on the jumping task, which take the optimal action with probability", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 129, + 698 + ], + "score": 0.87, + "content": "1 - \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 688, + 307, + 700 + ], + "score": 1.0, + "content": "and the subopotimal action with probability", + "type": "text" + }, + { + "bbox": [ + 307, + 692, + 312, + 697 + ], + "score": 0.6, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 688, + 505, + 700 + ], + "score": 1.0, + "content": ". We evaluate the generalization performance of", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 415, + 712 + ], + "score": 1.0, + "content": "PSEs for increasingly suboptimal policies, ranging from the optimal policy", + "type": "text" + }, + { + "bbox": [ + 416, + 699, + 440, + 709 + ], + "score": 0.84, + "content": "( \\epsilon = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 698, + 505, + 712 + ], + "score": 1.0, + "content": ") to the uniform", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 170, + 722 + ], + "score": 1.0, + "content": "random policy", + "type": "text" + }, + { + "bbox": [ + 170, + 710, + 203, + 721 + ], + "score": 0.84, + "content": "\\epsilon = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "). To isolate the effect of suboptimality on PSEs, the agent still imitates the", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 720, + 267, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 259, + 733 + ], + "score": 1.0, + "content": "optimal actions during training for all", + "type": "text" + }, + { + "bbox": [ + 259, + 723, + 264, + 730 + ], + "score": 0.6, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 720, + 267, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 66.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 196, + 95 + ], + "score": 1.0, + "content": "(2021) employed with", + "type": "text" + }, + { + "bbox": [ + 196, + 83, + 208, + 93 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "-bisimulation for learning representations in a single task RL setting. For a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 468, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 468, + 106 + ], + "score": 1.0, + "content": "fair comparison, we tune hyperparameters for 128 trials for each ablation entry in Table 2.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 211, + 123 + ], + "score": 1.0, + "content": "Table 2 show that PSEs", + "type": "text" + }, + { + "bbox": [ + 212, + 111, + 287, + 122 + ], + "score": 0.83, + "content": "( = \\mathrm { P S M } + \\mathrm { C M E s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 110, + 437, + 123 + ], + "score": 1.0, + "content": "generalize significantly better than", + "type": "text" + }, + { + "bbox": [ + 438, + 111, + 449, + 120 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "-bisimulation", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 165, + 135 + ], + "score": 1.0, + "content": "with CMEs or", + "type": "text" + }, + { + "bbox": [ + 165, + 122, + 175, + 132 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 121, + 430, + 135 + ], + "score": 1.0, + "content": "-embeddings, both of which significantly degrade performance (", + "type": "text" + }, + { + "bbox": [ + 430, + 121, + 457, + 132 + ], + "score": 0.87, + "content": "- 6 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 121, + 475, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 121, + 502, + 132 + ], + "score": 0.89, + "content": "- 4 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 121, + 506, + 135 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 242, + 146 + ], + "score": 1.0, + "content": "respectively). This is expected, as", + "type": "text" + }, + { + "bbox": [ + 242, + 133, + 254, + 142 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "-bisimulation imposes the incorrect invariances for the jumping", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 128, + 156 + ], + "score": 1.0, + "content": "task", + "type": "text" + }, + { + "bbox": [ + 128, + 144, + 142, + 155 + ], + "score": 0.72, + "content": "\\cdot { c . f . }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "Figures G.2a and G.2d). Additionally, looking at the rows of Table 2, CMEs are superior", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 127, + 165 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 155, + 137, + 165 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 154, + 306, + 165 + ], + "score": 1.0, + "content": "-embeddings for PSM while inferior for", + "type": "text" + }, + { + "bbox": [ + 307, + 155, + 318, + 164 + ], + "score": 0.88, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 154, + 504, + 165 + ], + "score": 1.0, + "content": "-bisimulation. This result is in line with the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "hypothesis that CMEs better enforce the invariances encoded by a similarity metric compared to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 176, + 267, + 189 + ], + "spans": [ + { + "bbox": [ + 107, + 177, + 116, + 187 + ], + "score": 0.86, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 176, + 172, + 189 + ], + "score": 1.0, + "content": "-embeddings", + "type": "text" + }, + { + "bbox": [ + 172, + 177, + 185, + 188 + ], + "score": 0.51, + "content": "\\left( c . f . \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 176, + 267, + 189 + ], + "score": 1.0, + "content": "Figures 3b and 3c).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 110, + 506, + 189 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "Visualizing learned representations. We visualize the metric embeddings in the ablation above", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 216 + ], + "score": 1.0, + "content": "by projecting them to two dimensions with UMAP (McInnes et al., 2018), a popular visualization", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "technique for high dimensional data which better preserves the data’s global structure compared to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 330, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 330, + 237 + ], + "score": 1.0, + "content": "other methods such as t-SNE (Coenen & Pearce, 2019).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 193, + 506, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "Figure 3 shows that PSEs partition the states into two sets: (1) states where a single suboptimal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "score": 1.0, + "content": "action leads to failure (all states before jump) and (2) states where actions do not affect the final", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 266, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 506, + 277 + ], + "score": 1.0, + "content": "outcome (states after jump). Additionally, PSEs align the labeled states in the first set, which have a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "PSM distance of zero. These aligned states have the same distance from the obstacle, the invariant", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 336, + 299 + ], + "score": 1.0, + "content": "feature that generalizes across tasks. On the other hand,", + "type": "text" + }, + { + "bbox": [ + 337, + 287, + 347, + 298 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "-embeddings (Zhang et al., 2021) with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "PSM do not align states with zero PSM except the state with the jump action – poor generalization,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "as observed empirically, is likely due to states with the same optimal behavior ending up with distant", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 209, + 331 + ], + "score": 1.0, + "content": "embeddings. CMEs with", + "type": "text" + }, + { + "bbox": [ + 209, + 320, + 221, + 330 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 320, + 343, + 331 + ], + "score": 1.0, + "content": "-bisimulation align states with", + "type": "text" + }, + { + "bbox": [ + 344, + 320, + 355, + 330 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 320, + 505, + 331 + ], + "score": 1.0, + "content": "-bisimulation distance of zero – such", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "states are equidistant from the start state and have different optimal behavior for any pair of tasks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 342, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 295, + 354 + ], + "score": 1.0, + "content": "with different obstacle positions (Figure G.2c).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 243, + 506, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 355, + 480, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 480, + 366 + ], + "score": 1.0, + "content": "5.3 JUMPING TASK WITH COLORS: WHERE TASK-DEPENDENT INVARIANCE MATTERS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 354, + 480, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 324, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 325, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 325, + 384 + ], + "score": 1.0, + "content": "PSEs capture invariances that are usually orthogonal to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 383, + 325, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 325, + 394 + ], + "score": 1.0, + "content": "task-agnostic invariances from data augmentation. This", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 393, + 325, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 325, + 406 + ], + "score": 1.0, + "content": "difference is important because, for certain tasks, data", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 406, + 325, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 325, + 416 + ], + "score": 1.0, + "content": "augmentation can erroneously alias states with different", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 416, + 325, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 325, + 428 + ], + "score": 1.0, + "content": "optimal behavior. Domain knowledge is often required", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 427, + 326, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 326, + 438 + ], + "score": 1.0, + "content": "to select appropriate augmentations, otherwise augmen-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 437, + 325, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 325, + 449 + ], + "score": 1.0, + "content": "tations can hurt generalization. In contrast, PSEs do", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 448, + 325, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 325, + 460 + ], + "score": 1.0, + "content": "not require any domain knowledge but instead exploit", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 460, + 259, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 259, + 470 + ], + "score": 1.0, + "content": "the inherent structure of the RL tasks.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 370, + 326, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 324, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 325, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 325, + 487 + ], + "score": 1.0, + "content": "To demonstrate the difference between PSEs and data", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 488, + 324, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 324, + 498 + ], + "score": 1.0, + "content": "augmentation, we simply include colored obstacles in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 497, + 325, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 325, + 510 + ], + "score": 1.0, + "content": "the jumping task (see Figure G.5). In this modified", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 509, + 324, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 324, + 521 + ], + "score": 1.0, + "content": "task, the optimal behavior of the agent depends on", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 520, + 325, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 325, + 532 + ], + "score": 1.0, + "content": "the obstacle color: the agent needs to jump over the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 531, + 325, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 325, + 543 + ], + "score": 1.0, + "content": "red obstacle but strike the green obstacle to get a high", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 541, + 325, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 325, + 554 + ], + "score": 1.0, + "content": "return. The red obstacle task has the same difficulty as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "the original jumping task while the green obstacle task is easier. We jointly train the agent with 18", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 564, + 477, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 477, + 576 + ], + "score": 1.0, + "content": "training tasks each, for both obstacle colors, and evaluate generalization on unseen red tasks.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 477, + 325, + 554 + ] + }, + { + "type": "image", + "bbox": [ + 340, + 372, + 497, + 492 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 340, + 372, + 497, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 340, + 372, + 497, + 492 + ], + "spans": [ + { + "bbox": [ + 340, + 372, + 497, + 492 + ], + "score": 0.967, + "type": "image", + "image_path": "dd27daffdbeb11b7f10511fabc09304681a9967704b663bae3974721a6096f03.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 340, + 372, + 497, + 384.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 340, + 384.0, + 497, + 396.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 340, + 396.0, + 497, + 408.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 340, + 408.0, + 497, + 420.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 340, + 420.0, + 497, + 432.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 340, + 432.0, + 497, + 444.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 340, + 444.0, + 497, + 456.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 340, + 456.0, + 497, + 468.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 340, + 468.0, + 497, + 480.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 340, + 480.0, + 497, + 492.0 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 332, + 496, + 505, + 546 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 331, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 331, + 495, + 414, + 507 + ], + "score": 1.0, + "content": "Figure 4: Percentage", + "type": "text" + }, + { + "bbox": [ + 415, + 496, + 429, + 506 + ], + "score": 0.73, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "of red obstacle test", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 331, + 506, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 331, + 506, + 505, + 516 + ], + "score": 1.0, + "content": "tasks solved when trained, jointly with red and", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 331, + 516, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 331, + 516, + 505, + 526 + ], + "score": 1.0, + "content": "green obstacles, on the “wide” grid. We report", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 331, + 525, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 331, + 525, + 487, + 536 + ], + "score": 1.0, + "content": "the mean across 100 runs. Error bars show", + "type": "text" + }, + { + "bbox": [ + 487, + 526, + 505, + 535 + ], + "score": 0.84, + "content": "9 9 \\%", + "type": "inline_equation" + } + ], + "index": 53 + }, + { + "bbox": [ + 331, + 535, + 453, + 546 + ], + "spans": [ + { + "bbox": [ + 331, + 535, + 453, + 546 + ], + "score": 1.0, + "content": "confidence interval for the mean.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52 + } + ], + "index": 48.25 + }, + { + "type": "text", + "bbox": [ + 108, + 553, + 505, + 575 + ], + "lines": [], + "index": 55.5, + "bbox_fs": [ + 105, + 553, + 505, + 576 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 506, + 647 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 593 + ], + "score": 1.0, + "content": "Figure 4 shows the large performance gap between PSEs and data augmentation with RandConv. All", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 592, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 477, + 604 + ], + "score": 1.0, + "content": "methods solve the green obstacle tasks (Table G.1). As opposed to the original jumping task (", + "type": "text" + }, + { + "bbox": [ + 477, + 592, + 490, + 604 + ], + "score": 0.68, + "content": "\\cdot f .", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 592, + 505, + 604 + ], + "score": 1.0, + "content": "Ta-", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "score": 1.0, + "content": "ble 1), data augmentation inhibits generalization since RandConv forces the agent to ignore color,", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 628 + ], + "score": 1.0, + "content": "conflating the red and green tasks (Figure G.6). PSEs still outperform regularization and data aug-", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "mentation. Furthermore, data augmentation performs better when combined with PSEs. Thus, PSEs", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 635, + 357, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 357, + 649 + ], + "score": 1.0, + "content": "are effective even when data augmentation hurts performance.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59.5, + "bbox_fs": [ + 105, + 581, + 506, + 649 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 649, + 329, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 330, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 330, + 661 + ], + "score": 1.0, + "content": "5.4 EFFECT OF POLICY SUBOPTIMALITY ON PSES", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "To understand the sensitivity of learning effective PSEs to the quality of the policies, we compute", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 153, + 690 + ], + "score": 1.0, + "content": "PSEs using", + "type": "text" + }, + { + "bbox": [ + 154, + 679, + 159, + 687 + ], + "score": 0.61, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "-suboptimal policies on the jumping task, which take the optimal action with probability", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 129, + 698 + ], + "score": 0.87, + "content": "1 - \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 688, + 307, + 700 + ], + "score": 1.0, + "content": "and the subopotimal action with probability", + "type": "text" + }, + { + "bbox": [ + 307, + 692, + 312, + 697 + ], + "score": 0.6, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 688, + 505, + 700 + ], + "score": 1.0, + "content": ". We evaluate the generalization performance of", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 415, + 712 + ], + "score": 1.0, + "content": "PSEs for increasingly suboptimal policies, ranging from the optimal policy", + "type": "text" + }, + { + "bbox": [ + 416, + 699, + 440, + 709 + ], + "score": 0.84, + "content": "( \\epsilon = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 698, + 505, + 712 + ], + "score": 1.0, + "content": ") to the uniform", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 170, + 722 + ], + "score": 1.0, + "content": "random policy", + "type": "text" + }, + { + "bbox": [ + 170, + 710, + 203, + 721 + ], + "score": 0.84, + "content": "\\epsilon = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "). To isolate the effect of suboptimality on PSEs, the agent still imitates the", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 720, + 267, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 259, + 733 + ], + "score": 1.0, + "content": "optimal actions during training for all", + "type": "text" + }, + { + "bbox": [ + 259, + 723, + 264, + 730 + ], + "score": 0.6, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 720, + 267, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 66.5, + "bbox_fs": [ + 105, + 664, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 369, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 368, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 368, + 94 + ], + "score": 1.0, + "content": "Figure 5 shows that PSEs show near-optimal generalization with", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 94, + 370, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 94, + 138, + 105 + ], + "score": 0.89, + "content": "\\epsilon \\leq 0 . 4", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 94, + 370, + 106 + ], + "score": 1.0, + "content": "while degrade generalization with an uniform random pol-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 369, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 369, + 117 + ], + "score": 1.0, + "content": "icy. This result is well-aligned with Proposition D.3, which shows", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 370, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 370, + 128 + ], + "score": 1.0, + "content": "that for policies with decreasing suboptimality, PSM approxima-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 370, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 370, + 139 + ], + "score": 1.0, + "content": "tion becomes more accurate, resulting in improved PSEs. Overall,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 369, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 369, + 149 + ], + "score": 1.0, + "content": "this study confirms that the utility of PSEs for generalization is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 369, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 369, + 160 + ], + "score": 1.0, + "content": "robust to suboptimality. One reason for this robustness is that", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 370, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 370, + 172 + ], + "score": 1.0, + "content": "PSEs are likely to align states with similar long-term greedy op-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 369, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 369, + 183 + ], + "score": 1.0, + "content": "timal actions, resulting in good performance even with suboptimal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 280, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 280, + 194 + ], + "score": 1.0, + "content": "policies that preserve these greedy actions.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 107, + 198, + 325, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 326, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 326, + 212 + ], + "score": 1.0, + "content": "6 ADDITIONAL EMPIRICAL EVALUATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 109, + 217, + 369, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 368, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 368, + 230 + ], + "score": 1.0, + "content": "In this section, we exhibit that PSM ignores spurious information", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 81, + 497, + 163 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 378, + 81, + 497, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 378, + 81, + 497, + 163 + ], + "spans": [ + { + "bbox": [ + 378, + 81, + 497, + 163 + ], + "score": 0.962, + "type": "image", + "image_path": "cc90cd32460e2a251f32af744d09a75a8fd7b373561f554a97a64367ed0acb68.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 378, + 81, + 497, + 122.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 378, + 122.0, + 497, + 163.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 375, + 165, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 375, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 375, + 164, + 416, + 176 + ], + "score": 1.0, + "content": "Figure 5:", + "type": "text" + }, + { + "bbox": [ + 417, + 165, + 426, + 174 + ], + "score": 0.74, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "of test tasks solved", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 175, + 503, + 186 + ], + "spans": [ + { + "bbox": [ + 375, + 175, + 497, + 186 + ], + "score": 1.0, + "content": "using PSEs computed using", + "type": "text" + }, + { + "bbox": [ + 497, + 177, + 503, + 185 + ], + "score": 0.51, + "content": "\\epsilon \\cdot", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 185, + 507, + 195 + ], + "spans": [ + { + "bbox": [ + 375, + 185, + 507, + 195 + ], + "score": 1.0, + "content": "suboptimal policies on the “wide”", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 195, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 375, + 195, + 506, + 206 + ], + "score": 1.0, + "content": "grid. We report the mean across 100", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 375, + 205, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 375, + 205, + 505, + 216 + ], + "score": 1.0, + "content": "runs. Error bars show one standard", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 214, + 414, + 226 + ], + "spans": [ + { + "bbox": [ + 375, + 214, + 414, + 226 + ], + "score": 1.0, + "content": "deviation.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 228, + 505, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 241 + ], + "score": 1.0, + "content": "for generalization using a LQR task (Song et al., 2019) with non-image inputs. Then, we demonstrate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 239, + 504, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 504, + 250 + ], + "score": 1.0, + "content": "the scalability of PSEs without explicit access to optimal policies in an RL setting, with continuous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 250, + 448, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 448, + 262 + ], + "score": 1.0, + "content": "actions, using Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 267, + 294, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 295, + 279 + ], + "score": 1.0, + "content": "6.1 LQR WITH SPURIOUS CORRELATIONS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 321, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 322, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 322, + 293 + ], + "score": 1.0, + "content": "We show how representations learned using PSM,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 291, + 322, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 322, + 304 + ], + "score": 1.0, + "content": "when faced with semantically equivalent environ-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 303, + 321, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 321, + 314 + ], + "score": 1.0, + "content": "ments, can learn the main factors of variation and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 314, + 322, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 322, + 325 + ], + "score": 1.0, + "content": "ignore spurious correlations that hinder generaliza-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 325, + 322, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 322, + 336 + ], + "score": 1.0, + "content": "tion. We use LQR with distractors (Song et al., 2019;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 336, + 322, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 322, + 347 + ], + "score": 1.0, + "content": "Sonar et al., 2020) to assess generalization in a feature-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 346, + 322, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 322, + 358 + ], + "score": 1.0, + "content": "based RL setting with linear function approximation.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 357, + 321, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 321, + 370 + ], + "score": 1.0, + "content": "The distractors are input features that are spuriously", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 369, + 322, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 322, + 381 + ], + "score": 1.0, + "content": "correlated with optimal actions and can be used for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 380, + 322, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 322, + 392 + ], + "score": 1.0, + "content": "predicting these actions during training, but hurt gen-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 391, + 321, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 321, + 403 + ], + "score": 1.0, + "content": "eralization. The agent learns a linear policy using 2", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 401, + 321, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 321, + 414 + ], + "score": 1.0, + "content": "environments with fixed distractors. This policy is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 414, + 312, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 312, + 424 + ], + "score": 1.0, + "content": "evaluated on environments with unseen distractors.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 320, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 321, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 321, + 442 + ], + "score": 1.0, + "content": "We aggregate state pairs in training environments", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 440, + 321, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 321, + 452 + ], + "score": 1.0, + "content": "with near-zero PSM. We contrast this approach with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 451, + 321, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 321, + 463 + ], + "score": 1.0, + "content": "(i) IPO (Sonar et al., 2020), a policy optimization", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 462, + 322, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 322, + 474 + ], + "score": 1.0, + "content": "method based on IRM (Arjovsky et al., 2019) for out-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 473, + 322, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 322, + 486 + ], + "score": 1.0, + "content": "of-distribution generalization, (ii) overparametrization,", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 334, + 278, + 491, + 392 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 278, + 491, + 392 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 278, + 491, + 392 + ], + "spans": [ + { + "bbox": [ + 334, + 278, + 491, + 392 + ], + "score": 0.965, + "type": "image", + "image_path": "6faf7cc10acce48febb5bb654d3e6e540adb0aa2e7913c37cb528653ffb5a9f1.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 334, + 278, + 491, + 290.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 334, + 290.6666666666667, + 491, + 303.33333333333337 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 334, + 303.33333333333337, + 491, + 316.00000000000006 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 334, + 316.00000000000006, + 491, + 328.66666666666674 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 334, + 328.66666666666674, + 491, + 341.3333333333334 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 334, + 341.3333333333334, + 491, + 354.0000000000001 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 334, + 354.0000000000001, + 491, + 366.6666666666668 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 334, + 366.6666666666668, + 491, + 379.3333333333335 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 334, + 379.3333333333335, + 491, + 392.00000000000017 + ], + "spans": [], + "index": 50 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 328, + 396, + 506, + 477 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 328, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 328, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "Figure 6: LQR generalization: Absolute test er-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 327, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 327, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "ror in LQR cost relative to the oracle (which has", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 327, + 417, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 327, + 417, + 505, + 426 + ], + "score": 1.0, + "content": "access to true state), of various methods trained", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 327, + 426, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 327, + 426, + 347, + 437 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 347, + 428, + 358, + 436 + ], + "score": 0.85, + "content": "n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 426, + 506, + 437 + ], + "score": 1.0, + "content": "distractors on 2 training environments.", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 327, + 436, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 327, + 436, + 506, + 447 + ], + "score": 1.0, + "content": "Lower error is better. PSM achieves near-optimal", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 327, + 447, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 327, + 447, + 506, + 456 + ], + "score": 1.0, + "content": "performance. We report mean across 100 differ-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 327, + 457, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 327, + 457, + 506, + 466 + ], + "score": 1.0, + "content": "ent seeds. Error bars (non-existent for most meth-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 328, + 466, + 492, + 477 + ], + "spans": [ + { + "bbox": [ + 328, + 466, + 366, + 477 + ], + "score": 1.0, + "content": "ods) show", + "type": "text" + }, + { + "bbox": [ + 367, + 466, + 385, + 475 + ], + "score": 0.86, + "content": "9 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 466, + 492, + 477 + ], + "score": 1.0, + "content": "confidence interval for mean.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 54.5 + } + ], + "index": 50.25 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 503, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "which leads to better generalization via implicit regularization (Song et al., 2019), and (iii) weight", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 494, + 492, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 164, + 508 + ], + "score": 1.0, + "content": "sparsity using", + "type": "text" + }, + { + "bbox": [ + 164, + 496, + 174, + 506 + ], + "score": 0.85, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 494, + 492, + 508 + ], + "score": 1.0, + "content": "-regularization since the policy weights which generalize in this task are sparse.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 59.5 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "All methods optimally solve the training environments; however, the baselines perform abysmally", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "in terms of generalization compared to state aggregation with PSM (Figure 6), indicating their", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "reliance on distractors. PSM obtains near-optimal generalization which we corroborate through this", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "conjecture (Section H.1): Assuming zero state aggregation error with PSM, the policy learned using", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 556, + 487, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 487, + 569 + ], + "score": 1.0, + "content": "gradient descent is independent of the distractors. Refer to Section H for a detailed discussion.", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 63 + }, + { + "type": "title", + "bbox": [ + 108, + 573, + 284, + 584 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 286, + 586 + ], + "score": 1.0, + "content": "6.2 DISTRACTING DM CONTROL SUITE", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 66 + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 356, + 711 + ], + "lines": [ + { + "bbox": [ + 106, + 589, + 358, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 358, + 601 + ], + "score": 1.0, + "content": "Finally, we demonstrate scalability of PSEs on the Distract-", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 600, + 356, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 356, + 612 + ], + "score": 1.0, + "content": "ing DM Control Suite (DCS) (Stone et al., 2021), which tests", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 612, + 356, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 356, + 622 + ], + "score": 1.0, + "content": "whether agents can ignore high-dimensional visual distractors", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 106, + 622, + 357, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 357, + 633 + ], + "score": 1.0, + "content": "irrelevant to the RL task. Since we do not have access to opti-", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 633, + 355, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 343, + 645 + ], + "score": 1.0, + "content": "mal training policies, we use learned policies as proxy for", + "type": "text" + }, + { + "bbox": [ + 343, + 633, + 355, + 643 + ], + "score": 0.84, + "content": "\\pi ^ { * }", + "type": "inline_equation" + } + ], + "index": 71 + }, + { + "bbox": [ + 105, + 642, + 357, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 357, + 657 + ], + "score": 1.0, + "content": "for computing PSM as well as collecting data for optimizing", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 654, + 358, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 358, + 667 + ], + "score": 1.0, + "content": "PSEs. Even with this approximation, PSEs outperform state-", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 106, + 666, + 357, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 357, + 677 + ], + "score": 1.0, + "content": "of-the-art data augmentation. DCS extends DM Control (Tassa", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 677, + 358, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 358, + 688 + ], + "score": 1.0, + "content": "et al., 2020) with visual distractions. We use the dynamic back-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 106, + 687, + 357, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 357, + 700 + ], + "score": 1.0, + "content": "ground distractions (Stone et al., 2021; Zhang et al., 2018b)", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 106, + 699, + 356, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 356, + 711 + ], + "score": 1.0, + "content": "where a video is played in the background from a specific", + "type": "text" + } + ], + "index": 77 + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 364, + 589, + 504, + 681 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 364, + 589, + 504, + 681 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 364, + 589, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 364, + 589, + 504, + 681 + ], + "score": 0.968, + "type": "image", + "image_path": "ff872ad14af14eeee7e2d8c38cd29313df3392bf8308eb9fbdc234e87dd58a41.jpg" + } + ] + } + ], + "index": 78.5, + "virtual_lines": [ + { + "bbox": [ + 364, + 589, + 504, + 635.0 + ], + "spans": [], + "index": 78 + }, + { + "bbox": [ + 364, + 635.0, + 504, + 681.0 + ], + "spans": [], + "index": 79 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 364, + 689, + 505, + 709 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 363, + 688, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 363, + 688, + 506, + 699 + ], + "score": 1.0, + "content": "Figure 7: Distracting Control Suite:", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 363, + 699, + 495, + 709 + ], + "spans": [ + { + "bbox": [ + 363, + 699, + 495, + 709 + ], + "score": 1.0, + "content": "Snapshots of training environments.", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 80.5 + } + ], + "index": 79.5 + }, + { + "type": "text", + "bbox": [ + 106, + 711, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "frame. The video and the frame are randomly sampled every new episode. We use 2 videos during", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 106, + 720, + 429, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 429, + 733 + ], + "score": 1.0, + "content": "training (Figure 7) and evaluate generalization on 30 unseen videos (Figure I.1).", + "type": "text" + } + ], + "index": 83 + } + ], + "index": 82.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 369, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 368, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 368, + 94 + ], + "score": 1.0, + "content": "Figure 5 shows that PSEs show near-optimal generalization with", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 94, + 370, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 94, + 138, + 105 + ], + "score": 0.89, + "content": "\\epsilon \\leq 0 . 4", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 94, + 370, + 106 + ], + "score": 1.0, + "content": "while degrade generalization with an uniform random pol-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 369, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 369, + 117 + ], + "score": 1.0, + "content": "icy. This result is well-aligned with Proposition D.3, which shows", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 370, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 370, + 128 + ], + "score": 1.0, + "content": "that for policies with decreasing suboptimality, PSM approxima-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 370, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 370, + 139 + ], + "score": 1.0, + "content": "tion becomes more accurate, resulting in improved PSEs. Overall,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 369, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 369, + 149 + ], + "score": 1.0, + "content": "this study confirms that the utility of PSEs for generalization is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 369, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 369, + 160 + ], + "score": 1.0, + "content": "robust to suboptimality. One reason for this robustness is that", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 370, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 370, + 172 + ], + "score": 1.0, + "content": "PSEs are likely to align states with similar long-term greedy op-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 369, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 369, + 183 + ], + "score": 1.0, + "content": "timal actions, resulting in good performance even with suboptimal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 280, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 280, + 194 + ], + "score": 1.0, + "content": "policies that preserve these greedy actions.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 82, + 370, + 194 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 198, + 325, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 326, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 326, + 212 + ], + "score": 1.0, + "content": "6 ADDITIONAL EMPIRICAL EVALUATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 109, + 217, + 369, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 368, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 368, + 230 + ], + "score": 1.0, + "content": "In this section, we exhibit that PSM ignores spurious information", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 216, + 368, + 230 + ] + }, + { + "type": "image", + "bbox": [ + 378, + 81, + 497, + 163 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 378, + 81, + 497, + 163 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 378, + 81, + 497, + 163 + ], + "spans": [ + { + "bbox": [ + 378, + 81, + 497, + 163 + ], + "score": 0.962, + "type": "image", + "image_path": "cc90cd32460e2a251f32af744d09a75a8fd7b373561f554a97a64367ed0acb68.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 378, + 81, + 497, + 122.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 378, + 122.0, + 497, + 163.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 375, + 165, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 375, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 375, + 164, + 416, + 176 + ], + "score": 1.0, + "content": "Figure 5:", + "type": "text" + }, + { + "bbox": [ + 417, + 165, + 426, + 174 + ], + "score": 0.74, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "of test tasks solved", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 175, + 503, + 186 + ], + "spans": [ + { + "bbox": [ + 375, + 175, + 497, + 186 + ], + "score": 1.0, + "content": "using PSEs computed using", + "type": "text" + }, + { + "bbox": [ + 497, + 177, + 503, + 185 + ], + "score": 0.51, + "content": "\\epsilon \\cdot", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 185, + 507, + 195 + ], + "spans": [ + { + "bbox": [ + 375, + 185, + 507, + 195 + ], + "score": 1.0, + "content": "suboptimal policies on the “wide”", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 195, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 375, + 195, + 506, + 206 + ], + "score": 1.0, + "content": "grid. We report the mean across 100", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 375, + 205, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 375, + 205, + 505, + 216 + ], + "score": 1.0, + "content": "runs. Error bars show one standard", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 214, + 414, + 226 + ], + "spans": [ + { + "bbox": [ + 375, + 214, + 414, + 226 + ], + "score": 1.0, + "content": "deviation.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 228, + 505, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 241 + ], + "score": 1.0, + "content": "for generalization using a LQR task (Song et al., 2019) with non-image inputs. Then, we demonstrate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 239, + 504, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 504, + 250 + ], + "score": 1.0, + "content": "the scalability of PSEs without explicit access to optimal policies in an RL setting, with continuous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 250, + 448, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 448, + 262 + ], + "score": 1.0, + "content": "actions, using Distracting DM Control Suite (Stone et al., 2021; Zhang et al., 2018b).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 227, + 505, + 262 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 267, + 294, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 295, + 279 + ], + "score": 1.0, + "content": "6.1 LQR WITH SPURIOUS CORRELATIONS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 321, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 322, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 322, + 293 + ], + "score": 1.0, + "content": "We show how representations learned using PSM,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 291, + 322, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 322, + 304 + ], + "score": 1.0, + "content": "when faced with semantically equivalent environ-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 303, + 321, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 321, + 314 + ], + "score": 1.0, + "content": "ments, can learn the main factors of variation and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 314, + 322, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 322, + 325 + ], + "score": 1.0, + "content": "ignore spurious correlations that hinder generaliza-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 325, + 322, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 322, + 336 + ], + "score": 1.0, + "content": "tion. We use LQR with distractors (Song et al., 2019;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 336, + 322, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 322, + 347 + ], + "score": 1.0, + "content": "Sonar et al., 2020) to assess generalization in a feature-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 346, + 322, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 322, + 358 + ], + "score": 1.0, + "content": "based RL setting with linear function approximation.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 357, + 321, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 321, + 370 + ], + "score": 1.0, + "content": "The distractors are input features that are spuriously", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 369, + 322, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 322, + 381 + ], + "score": 1.0, + "content": "correlated with optimal actions and can be used for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 380, + 322, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 322, + 392 + ], + "score": 1.0, + "content": "predicting these actions during training, but hurt gen-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 391, + 321, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 321, + 403 + ], + "score": 1.0, + "content": "eralization. The agent learns a linear policy using 2", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 401, + 321, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 321, + 414 + ], + "score": 1.0, + "content": "environments with fixed distractors. This policy is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 414, + 312, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 312, + 424 + ], + "score": 1.0, + "content": "evaluated on environments with unseen distractors.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 280, + 322, + 424 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 320, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 321, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 321, + 442 + ], + "score": 1.0, + "content": "We aggregate state pairs in training environments", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 440, + 321, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 321, + 452 + ], + "score": 1.0, + "content": "with near-zero PSM. We contrast this approach with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 451, + 321, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 321, + 463 + ], + "score": 1.0, + "content": "(i) IPO (Sonar et al., 2020), a policy optimization", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 462, + 322, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 322, + 474 + ], + "score": 1.0, + "content": "method based on IRM (Arjovsky et al., 2019) for out-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 473, + 322, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 322, + 486 + ], + "score": 1.0, + "content": "of-distribution generalization, (ii) overparametrization,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "which leads to better generalization via implicit regularization (Song et al., 2019), and (iii) weight", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 494, + 492, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 164, + 508 + ], + "score": 1.0, + "content": "sparsity using", + "type": "text" + }, + { + "bbox": [ + 164, + 496, + 174, + 506 + ], + "score": 0.85, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 494, + 492, + 508 + ], + "score": 1.0, + "content": "-regularization since the policy weights which generalize in this task are sparse.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 429, + 322, + 486 + ] + }, + { + "type": "image", + "bbox": [ + 334, + 278, + 491, + 392 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 278, + 491, + 392 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 278, + 491, + 392 + ], + "spans": [ + { + "bbox": [ + 334, + 278, + 491, + 392 + ], + "score": 0.965, + "type": "image", + "image_path": "6faf7cc10acce48febb5bb654d3e6e540adb0aa2e7913c37cb528653ffb5a9f1.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 334, + 278, + 491, + 290.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 334, + 290.6666666666667, + 491, + 303.33333333333337 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 334, + 303.33333333333337, + 491, + 316.00000000000006 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 334, + 316.00000000000006, + 491, + 328.66666666666674 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 334, + 328.66666666666674, + 491, + 341.3333333333334 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 334, + 341.3333333333334, + 491, + 354.0000000000001 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 334, + 354.0000000000001, + 491, + 366.6666666666668 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 334, + 366.6666666666668, + 491, + 379.3333333333335 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 334, + 379.3333333333335, + 491, + 392.00000000000017 + ], + "spans": [], + "index": 50 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 328, + 396, + 506, + 477 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 328, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 328, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "Figure 6: LQR generalization: Absolute test er-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 327, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 327, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "ror in LQR cost relative to the oracle (which has", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 327, + 417, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 327, + 417, + 505, + 426 + ], + "score": 1.0, + "content": "access to true state), of various methods trained", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 327, + 426, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 327, + 426, + 347, + 437 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 347, + 428, + 358, + 436 + ], + "score": 0.85, + "content": "n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 426, + 506, + 437 + ], + "score": 1.0, + "content": "distractors on 2 training environments.", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 327, + 436, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 327, + 436, + 506, + 447 + ], + "score": 1.0, + "content": "Lower error is better. PSM achieves near-optimal", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 327, + 447, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 327, + 447, + 506, + 456 + ], + "score": 1.0, + "content": "performance. We report mean across 100 differ-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 327, + 457, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 327, + 457, + 506, + 466 + ], + "score": 1.0, + "content": "ent seeds. Error bars (non-existent for most meth-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 328, + 466, + 492, + 477 + ], + "spans": [ + { + "bbox": [ + 328, + 466, + 366, + 477 + ], + "score": 1.0, + "content": "ods) show", + "type": "text" + }, + { + "bbox": [ + 367, + 466, + 385, + 475 + ], + "score": 0.86, + "content": "9 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 466, + 492, + 477 + ], + "score": 1.0, + "content": "confidence interval for mean.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 54.5 + } + ], + "index": 50.25 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 503, + 507 + ], + "lines": [], + "index": 59.5, + "bbox_fs": [ + 105, + 483, + 505, + 508 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "All methods optimally solve the training environments; however, the baselines perform abysmally", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "in terms of generalization compared to state aggregation with PSM (Figure 6), indicating their", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "reliance on distractors. PSM obtains near-optimal generalization which we corroborate through this", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "conjecture (Section H.1): Assuming zero state aggregation error with PSM, the policy learned using", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 556, + 487, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 487, + 569 + ], + "score": 1.0, + "content": "gradient descent is independent of the distractors. Refer to Section H for a detailed discussion.", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 63, + "bbox_fs": [ + 105, + 511, + 506, + 569 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 573, + 284, + 584 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 286, + 586 + ], + "score": 1.0, + "content": "6.2 DISTRACTING DM CONTROL SUITE", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 66 + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 356, + 711 + ], + "lines": [ + { + "bbox": [ + 106, + 589, + 358, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 358, + 601 + ], + "score": 1.0, + "content": "Finally, we demonstrate scalability of PSEs on the Distract-", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 106, + 600, + 356, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 356, + 612 + ], + "score": 1.0, + "content": "ing DM Control Suite (DCS) (Stone et al., 2021), which tests", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 612, + 356, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 356, + 622 + ], + "score": 1.0, + "content": "whether agents can ignore high-dimensional visual distractors", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 106, + 622, + 357, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 357, + 633 + ], + "score": 1.0, + "content": "irrelevant to the RL task. Since we do not have access to opti-", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 106, + 633, + 355, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 343, + 645 + ], + "score": 1.0, + "content": "mal training policies, we use learned policies as proxy for", + "type": "text" + }, + { + "bbox": [ + 343, + 633, + 355, + 643 + ], + "score": 0.84, + "content": "\\pi ^ { * }", + "type": "inline_equation" + } + ], + "index": 71 + }, + { + "bbox": [ + 105, + 642, + 357, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 357, + 657 + ], + "score": 1.0, + "content": "for computing PSM as well as collecting data for optimizing", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 105, + 654, + 358, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 358, + 667 + ], + "score": 1.0, + "content": "PSEs. Even with this approximation, PSEs outperform state-", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 106, + 666, + 357, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 357, + 677 + ], + "score": 1.0, + "content": "of-the-art data augmentation. DCS extends DM Control (Tassa", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 677, + 358, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 358, + 688 + ], + "score": 1.0, + "content": "et al., 2020) with visual distractions. We use the dynamic back-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 106, + 687, + 357, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 357, + 700 + ], + "score": 1.0, + "content": "ground distractions (Stone et al., 2021; Zhang et al., 2018b)", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 106, + 699, + 356, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 356, + 711 + ], + "score": 1.0, + "content": "where a video is played in the background from a specific", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "frame. The video and the frame are randomly sampled every new episode. We use 2 videos during", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 106, + 720, + 429, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 429, + 733 + ], + "score": 1.0, + "content": "training (Figure 7) and evaluate generalization on 30 unseen videos (Figure I.1).", + "type": "text" + } + ], + "index": 83 + } + ], + "index": 72, + "bbox_fs": [ + 105, + 589, + 358, + 711 + ] + }, + { + "type": "image", + "bbox": [ + 364, + 589, + 504, + 681 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 364, + 589, + 504, + 681 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 364, + 589, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 364, + 589, + 504, + 681 + ], + "score": 0.968, + "type": "image", + "image_path": "ff872ad14af14eeee7e2d8c38cd29313df3392bf8308eb9fbdc234e87dd58a41.jpg" + } + ] + } + ], + "index": 78.5, + "virtual_lines": [ + { + "bbox": [ + 364, + 589, + 504, + 635.0 + ], + "spans": [], + "index": 78 + }, + { + "bbox": [ + 364, + 635.0, + 504, + 681.0 + ], + "spans": [], + "index": 79 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 364, + 689, + 505, + 709 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 363, + 688, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 363, + 688, + 506, + 699 + ], + "score": 1.0, + "content": "Figure 7: Distracting Control Suite:", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 363, + 699, + 495, + 709 + ], + "spans": [ + { + "bbox": [ + 363, + 699, + 495, + 709 + ], + "score": 1.0, + "content": "Snapshots of training environments.", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 80.5 + } + ], + "index": 79.5 + }, + { + "type": "text", + "bbox": [ + 106, + 711, + 504, + 732 + ], + "lines": [], + "index": 82.5, + "bbox_fs": [ + 105, + 708, + 505, + 733 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 117, + 502, + 189 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 506, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 91 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 505, + 91 + ], + "score": 1.0, + "content": "Table 3: Generalization performance with unseen distractions in the Distracting Suite at 500K steps. We", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 248, + 101 + ], + "score": 1.0, + "content": "report the average scores across 5 seeds", + "type": "text" + }, + { + "bbox": [ + 248, + 91, + 258, + 100 + ], + "score": 0.66, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 90, + 505, + 101 + ], + "score": 1.0, + "content": "standard error. All methods are added to SAC (Haarnoja et al., 2018).", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 99, + 484, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 210, + 111 + ], + "score": 1.0, + "content": "Pretrained initialization uses", + "type": "text" + }, + { + "bbox": [ + 211, + 100, + 228, + 111 + ], + "score": 0.41, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 99, + 484, + 111 + ], + "score": 1.0, + "content": "trained for 500K steps. Figures I.2 and I.3 show generalization curves.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 110, + 117, + 502, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 117, + 502, + 189 + ], + "spans": [ + { + "bbox": [ + 110, + 117, + 502, + 189 + ], + "score": 0.978, + "html": "
InitializationMethodBiC-catchC-swingupC-runF-spinR-easyW-walk
RandomDrQ747±28582±42220±12646±54931±14549±83
DrQ + PSEs821±17749±19308±12779±49955±10789±28
PretrainedDrQ748±30689±22219±10764±48943±10709±29
DrQ +PSEs805±25753±13282±8803±19962±11829±21
", + "type": "table", + "image_path": "5d3a20085648824a4fd02c4eee0f37a927f6224b136e454d7b0d14b5a1bf5010.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 110, + 117, + 502, + 141.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 141.0, + 502, + 165.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 165.0, + 502, + 189.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 197, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 390, + 210 + ], + "score": 1.0, + "content": "All agents are built on top of SAC (Haarnoja et al., 2018) combined with", + "type": "text" + }, + { + "bbox": [ + 391, + 198, + 410, + 209 + ], + "score": 0.28, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "(Kostrikov et al., 2020),", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "score": 1.0, + "content": "an augmentation method with state-of-the-art performance on DM control. Notably, DrQ outperforms", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "CURL (Laskin et al., 2020b), a strong contrastive method for RL. Without data augmentation on DM", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 464, + 243 + ], + "score": 1.0, + "content": "control, SAC performs poorly, even during training (Kostrikov et al., 2020). We augment", + "type": "text" + }, + { + "bbox": [ + 464, + 231, + 484, + 241 + ], + "score": 0.32, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "an auxiliary loss for learning PSEs and compare it with DrQ (Table 3). Orthogonal to DrQ, PSEs", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 253, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 417, + 264 + ], + "score": 1.0, + "content": "align representations of different states across environments based on PSM", + "type": "text" + }, + { + "bbox": [ + 418, + 253, + 433, + 264 + ], + "score": 0.56, + "content": "( c . f .", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 253, + 505, + 264 + ], + "score": 1.0, + "content": "Figure A.1). All", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "agents are trained for 500K environment steps with random crop augmentation. For computing PSM,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 274, + 443, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 213, + 288 + ], + "score": 1.0, + "content": "we use policies learned by", + "type": "text" + }, + { + "bbox": [ + 214, + 275, + 233, + 286 + ], + "score": 0.43, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 274, + 443, + 288 + ], + "score": 1.0, + "content": "pretrained on training environments for 500K steps.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 304 + ], + "score": 1.0, + "content": "First, we investigate how much better PSEs generalize relative to DrQ, assuming the agent is provided", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "with PSM beforehand. The agent’s policy is randomly initialized so that additional gains over DrQ", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 314, + 504, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 504, + 326 + ], + "score": 1.0, + "content": "can be attributed to the auxiliary information from PSM. The substantial gains in Table 3 indicate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 324, + 415, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 241, + 336 + ], + "score": 1.0, + "content": "that PSEs are more effective than", + "type": "text" + }, + { + "bbox": [ + 241, + 324, + 261, + 335 + ], + "score": 0.3, + "content": "_ \\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 324, + 415, + 336 + ], + "score": 1.0, + "content": "for encoding invariance to distractors.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 459, + 354 + ], + "score": 1.0, + "content": "Since PSEs utilize PSM approximated using pretrained policies, we also compare to a", + "type": "text" + }, + { + "bbox": [ + 460, + 341, + 480, + 352 + ], + "score": 0.27, + "content": "_ \\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "agent", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "where we initialize it using these pretrained policies. This comparison provides the same auxiliary", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "information to DrQ as available to PSEs, thus, the generalization difference stems from how they", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 373, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 507, + 387 + ], + "score": 1.0, + "content": "utilize this information. Table 3 demonstrates that PSEs outperform DrQ with pretrained initialization,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "indicating that the additional pretraining steps are more judiciously utilized for computing PSM as", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 396, + 501, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 501, + 409 + ], + "score": 1.0, + "content": "opposed to just longer training with DrQ. More details, including learning curves, are in Section I.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 212, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 213, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 213, + 428 + ], + "score": 1.0, + "content": "7 RELATED WORK", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 506, + 509 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "PSM (Section 3) is inspired by bisimulation metrics (Section C). However, different than traditional", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "bisimulation (e.g., Larsen & Skou, 1991; Givan et al., 2003; Ferns et al., 2011), PSM is more tractable", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 453, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 414, + 466 + ], + "score": 1.0, + "content": "as it defined with respect to a single policy similar to the recently proposed", + "type": "text" + }, + { + "bbox": [ + 415, + 454, + 426, + 464 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 453, + 506, + 466 + ], + "score": 1.0, + "content": "-bisimulation (Cas-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 465, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 476 + ], + "score": 1.0, + "content": "tro, 2020; Zhang et al., 2021). However, in contrast to PSM, bisimulation metrics rely on reward", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 476, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 487 + ], + "score": 1.0, + "content": "information and may not provide a meaningful notion of behavioral similarity in certain environ-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "ments (Section 5). For example, states similar under PSM would have similar optimal policies, yet", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 498, + 442, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 208, + 509 + ], + "score": 1.0, + "content": "can have arbitrarily large", + "type": "text" + }, + { + "bbox": [ + 208, + 498, + 220, + 507 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 498, + 442, + 509 + ], + "score": 1.0, + "content": "-bisimulation distance between them (Proposition D.1).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 506, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "score": 1.0, + "content": "PSEs (Section 4) use contrastive learning to encode behavior similarity (Section 3) across MDPs.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "Previously, contrastive learning has been applied for imposing state self-consistency (Laskin et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "2020b), capturing predictive information (Oord et al., 2018; Mazoure et al., 2020; Lee et al., 2020b)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "score": 1.0, + "content": "or encoding transition dynamics (van der Pol et al., 2020; Stooke et al., 2020; Schwarzer et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 557, + 507, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 507, + 570 + ], + "score": 1.0, + "content": "2020) within an MDP. These methods can be integrated with PSEs to encode additional invariances.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "Interestingly, in a similar spirit to PSEs, Pacchiano et al. (2020); Moskovitz et al. (2021) explore", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 580, + 485, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 485, + 592 + ], + "score": 1.0, + "content": "comparing behavioral similarity between policies to guide policy optimization within an MDP.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "PSEs are complementary to data augmentation methods (Kostrikov et al., 2020; Lee et al., 2020a;", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "score": 1.0, + "content": "Raileanu et al., 2020; Ye et al., 2020), which have recently been shown to significantly improve", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "agents’ generalization capabilities. In fact, we combine PSEs to state-of-the-art augmentation methods", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "including random convolutions (Lee et al., 2020a; Laskin et al., 2020a) in the jumping task and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 640, + 492, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 492, + 654 + ], + "score": 1.0, + "content": "DrQ (Kostrikov et al., 2020) on Distracting Control Suite, leading to performance improvement.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "title", + "bbox": [ + 107, + 659, + 196, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 198, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 198, + 675 + ], + "score": 1.0, + "content": "8 CONCLUSION", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "This paper advances generalization in RL by two contributions: (1) the policy similarity metric (PSM)", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "which provides a new notion of state similarity based on behavior proximity, and (2) contrastive", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "metric embeddings, which harness the benefits of contrastive learning for representations based on a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "similarity metric. PSEs combine these two ideas to improve generalization. Overall, this paper shows", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "score": 1.0, + "content": "the benefits of exploiting the inherent structure in RL for learning effective representations.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 117, + 502, + 189 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 506, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 91 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 505, + 91 + ], + "score": 1.0, + "content": "Table 3: Generalization performance with unseen distractions in the Distracting Suite at 500K steps. We", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 248, + 101 + ], + "score": 1.0, + "content": "report the average scores across 5 seeds", + "type": "text" + }, + { + "bbox": [ + 248, + 91, + 258, + 100 + ], + "score": 0.66, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 90, + 505, + 101 + ], + "score": 1.0, + "content": "standard error. All methods are added to SAC (Haarnoja et al., 2018).", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 99, + 484, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 210, + 111 + ], + "score": 1.0, + "content": "Pretrained initialization uses", + "type": "text" + }, + { + "bbox": [ + 211, + 100, + 228, + 111 + ], + "score": 0.41, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 99, + 484, + 111 + ], + "score": 1.0, + "content": "trained for 500K steps. Figures I.2 and I.3 show generalization curves.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 110, + 117, + 502, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 117, + 502, + 189 + ], + "spans": [ + { + "bbox": [ + 110, + 117, + 502, + 189 + ], + "score": 0.978, + "html": "
InitializationMethodBiC-catchC-swingupC-runF-spinR-easyW-walk
RandomDrQ747±28582±42220±12646±54931±14549±83
DrQ + PSEs821±17749±19308±12779±49955±10789±28
PretrainedDrQ748±30689±22219±10764±48943±10709±29
DrQ +PSEs805±25753±13282±8803±19962±11829±21
", + "type": "table", + "image_path": "5d3a20085648824a4fd02c4eee0f37a927f6224b136e454d7b0d14b5a1bf5010.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 110, + 117, + 502, + 141.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 141.0, + 502, + 165.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 165.0, + 502, + 189.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 197, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 390, + 210 + ], + "score": 1.0, + "content": "All agents are built on top of SAC (Haarnoja et al., 2018) combined with", + "type": "text" + }, + { + "bbox": [ + 391, + 198, + 410, + 209 + ], + "score": 0.28, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "(Kostrikov et al., 2020),", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "score": 1.0, + "content": "an augmentation method with state-of-the-art performance on DM control. Notably, DrQ outperforms", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "CURL (Laskin et al., 2020b), a strong contrastive method for RL. Without data augmentation on DM", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 464, + 243 + ], + "score": 1.0, + "content": "control, SAC performs poorly, even during training (Kostrikov et al., 2020). We augment", + "type": "text" + }, + { + "bbox": [ + 464, + 231, + 484, + 241 + ], + "score": 0.32, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "an auxiliary loss for learning PSEs and compare it with DrQ (Table 3). Orthogonal to DrQ, PSEs", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 253, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 417, + 264 + ], + "score": 1.0, + "content": "align representations of different states across environments based on PSM", + "type": "text" + }, + { + "bbox": [ + 418, + 253, + 433, + 264 + ], + "score": 0.56, + "content": "( c . f .", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 253, + 505, + 264 + ], + "score": 1.0, + "content": "Figure A.1). All", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "agents are trained for 500K environment steps with random crop augmentation. For computing PSM,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 274, + 443, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 213, + 288 + ], + "score": 1.0, + "content": "we use policies learned by", + "type": "text" + }, + { + "bbox": [ + 214, + 275, + 233, + 286 + ], + "score": 0.43, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 274, + 443, + 288 + ], + "score": 1.0, + "content": "pretrained on training environments for 500K steps.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 197, + 506, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 304 + ], + "score": 1.0, + "content": "First, we investigate how much better PSEs generalize relative to DrQ, assuming the agent is provided", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "with PSM beforehand. The agent’s policy is randomly initialized so that additional gains over DrQ", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 314, + 504, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 504, + 326 + ], + "score": 1.0, + "content": "can be attributed to the auxiliary information from PSM. The substantial gains in Table 3 indicate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 324, + 415, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 241, + 336 + ], + "score": 1.0, + "content": "that PSEs are more effective than", + "type": "text" + }, + { + "bbox": [ + 241, + 324, + 261, + 335 + ], + "score": 0.3, + "content": "_ \\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 324, + 415, + 336 + ], + "score": 1.0, + "content": "for encoding invariance to distractors.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 290, + 506, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 340, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 459, + 354 + ], + "score": 1.0, + "content": "Since PSEs utilize PSM approximated using pretrained policies, we also compare to a", + "type": "text" + }, + { + "bbox": [ + 460, + 341, + 480, + 352 + ], + "score": 0.27, + "content": "_ \\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "agent", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "where we initialize it using these pretrained policies. This comparison provides the same auxiliary", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "information to DrQ as available to PSEs, thus, the generalization difference stems from how they", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 373, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 507, + 387 + ], + "score": 1.0, + "content": "utilize this information. Table 3 demonstrates that PSEs outperform DrQ with pretrained initialization,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "indicating that the additional pretraining steps are more judiciously utilized for computing PSM as", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 396, + 501, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 501, + 409 + ], + "score": 1.0, + "content": "opposed to just longer training with DrQ. More details, including learning curves, are in Section I.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 340, + 507, + 409 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 212, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 213, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 213, + 428 + ], + "score": 1.0, + "content": "7 RELATED WORK", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 506, + 509 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "PSM (Section 3) is inspired by bisimulation metrics (Section C). However, different than traditional", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "bisimulation (e.g., Larsen & Skou, 1991; Givan et al., 2003; Ferns et al., 2011), PSM is more tractable", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 453, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 414, + 466 + ], + "score": 1.0, + "content": "as it defined with respect to a single policy similar to the recently proposed", + "type": "text" + }, + { + "bbox": [ + 415, + 454, + 426, + 464 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 453, + 506, + 466 + ], + "score": 1.0, + "content": "-bisimulation (Cas-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 465, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 505, + 476 + ], + "score": 1.0, + "content": "tro, 2020; Zhang et al., 2021). However, in contrast to PSM, bisimulation metrics rely on reward", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 476, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 487 + ], + "score": 1.0, + "content": "information and may not provide a meaningful notion of behavioral similarity in certain environ-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "ments (Section 5). For example, states similar under PSM would have similar optimal policies, yet", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 498, + 442, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 208, + 509 + ], + "score": 1.0, + "content": "can have arbitrarily large", + "type": "text" + }, + { + "bbox": [ + 208, + 498, + 220, + 507 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 498, + 442, + 509 + ], + "score": 1.0, + "content": "-bisimulation distance between them (Proposition D.1).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 431, + 506, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 514, + 506, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "score": 1.0, + "content": "PSEs (Section 4) use contrastive learning to encode behavior similarity (Section 3) across MDPs.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "Previously, contrastive learning has been applied for imposing state self-consistency (Laskin et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "2020b), capturing predictive information (Oord et al., 2018; Mazoure et al., 2020; Lee et al., 2020b)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 506, + 559 + ], + "score": 1.0, + "content": "or encoding transition dynamics (van der Pol et al., 2020; Stooke et al., 2020; Schwarzer et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 557, + 507, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 507, + 570 + ], + "score": 1.0, + "content": "2020) within an MDP. These methods can be integrated with PSEs to encode additional invariances.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "Interestingly, in a similar spirit to PSEs, Pacchiano et al. (2020); Moskovitz et al. (2021) explore", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 580, + 485, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 485, + 592 + ], + "score": 1.0, + "content": "comparing behavioral similarity between policies to guide policy optimization within an MDP.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 513, + 507, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "PSEs are complementary to data augmentation methods (Kostrikov et al., 2020; Lee et al., 2020a;", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "score": 1.0, + "content": "Raileanu et al., 2020; Ye et al., 2020), which have recently been shown to significantly improve", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "agents’ generalization capabilities. In fact, we combine PSEs to state-of-the-art augmentation methods", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "including random convolutions (Lee et al., 2020a; Laskin et al., 2020a) in the jumping task and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 640, + 492, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 492, + 654 + ], + "score": 1.0, + "content": "DrQ (Kostrikov et al., 2020) on Distracting Control Suite, leading to performance improvement.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 596, + 506, + 654 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 659, + 196, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 198, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 198, + 675 + ], + "score": 1.0, + "content": "8 CONCLUSION", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "This paper advances generalization in RL by two contributions: (1) the policy similarity metric (PSM)", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "which provides a new notion of state similarity based on behavior proximity, and (2) contrastive", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "metric embeddings, which harness the benefits of contrastive learning for representations based on a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "similarity metric. PSEs combine these two ideas to improve generalization. Overall, this paper shows", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "score": 1.0, + "content": "the benefits of exploiting the inherent structure in RL for learning effective representations.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 99, + 504, + 120 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 505, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 505, + 111 + ], + "score": 1.0, + "content": "Rishabh Agarwal, Chen Liang, Dale Schuurmans, and Mohammad Norouzi. Learning to generalize from sparse", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 109, + 278, + 120 + ], + "spans": [ + { + "bbox": [ + 116, + 109, + 278, + 120 + ], + "score": 1.0, + "content": "and underspecified rewards. In ICML, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 503, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "Martín Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. CoRR,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 137, + 201, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 201, + 148 + ], + "score": 1.0, + "content": "abs/1907.02893, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 108, + 155, + 503, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 389, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 389, + 177 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 7413–7424, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 471, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 471, + 205 + ], + "score": 1.0, + "content": "meta-regularization. In Advances in Neural Information Processing Systems, pp. 998–1008, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 108, + 211, + 503, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "score": 1.0, + "content": "Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic Markov decision", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 443, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 443, + 234 + ], + "score": 1.0, + "content": "processes. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 108, + 239, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "Pablo Samuel Castro and Doina Precup. Using bisimulation for policy transfer in MDPs. In Proceedings of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 250, + 323, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 323, + 261 + ], + "score": 1.0, + "content": "AAAI Conference on Artificial Intelligence (AAAI), 2010.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 108, + 267, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 278, + 351, + 289 + ], + "spans": [ + { + "bbox": [ + 116, + 278, + 351, + 289 + ], + "score": 1.0, + "content": "learning of visual representations. CoRR, abs/2002.05709, 2020.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 504, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "Karl Cobbe, Oleg Klimov, Christopher Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 306, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 306, + 506, + 318 + ], + "score": 1.0, + "content": "reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 504, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "Andy Coenen and Adam Pearce. Understanding umap, 2019. URL https://pair-code.github.io/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 333, + 223, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 223, + 346 + ], + "score": 1.0, + "content": "understanding-umap/.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "Jesse Farebrother, Marlos C. Machado, and Michael Bowling. Generalization and regularization in DQN. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 362, + 322, + 374 + ], + "spans": [ + { + "bbox": [ + 114, + 362, + 322, + 374 + ], + "score": 1.0, + "content": "NeurIPS Deep Reinforcement Learning Workshop, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 392 + ], + "score": 1.0, + "content": "Norm Ferns, Prakash Panangaden, and Doina Precup. Metrics for finite Markov decision processes. In", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 390, + 418, + 402 + ], + "spans": [ + { + "bbox": [ + 116, + 390, + 418, + 402 + ], + "score": 1.0, + "content": "Proceedings of the Conference in Uncertainty in Artificial Intelligence (UAI), 2004.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 108, + 408, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "Norm Ferns, Pablo Samuel Castro, Doina Precup, and Prakash Panangaden. Methods for computing state", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 418, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 418, + 506, + 430 + ], + "score": 1.0, + "content": "similarity in Markov decision processes. In Proceedings of the 22nd Conference on Uncertainty in Artificial", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 267, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 267, + 440 + ], + "score": 1.0, + "content": "Intelligence, UAI ’06. AUAI Press, 2006.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 446, + 504, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "Norm Ferns, Prakash Panangaden, and Doina Precup. Bisimulation metrics for continuous markov decision", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 456, + 354, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 456, + 354, + 468 + ], + "score": 1.0, + "content": "processes. SIAM Journal on Computing, 40(6):1662–1714, 2011.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 503, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "score": 1.0, + "content": "Norman Ferns and Doina Precup. Bisimulation metrics are optimal value functions. In The 30th Conference on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 117, + 484, + 303, + 496 + ], + "spans": [ + { + "bbox": [ + 117, + 484, + 303, + 496 + ], + "score": 1.0, + "content": "Uncertainty in Artificial Intelligence, pp. 10, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 504, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 516 + ], + "score": 1.0, + "content": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 513, + 461, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 513, + 461, + 524 + ], + "score": 1.0, + "content": "networks. In Proceedings of the International Conference on Machine Learning (ICML), 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "score": 1.0, + "content": "Robert Givan, Thomas Dean, and Matthew Greig. Equivalence notions and model minimization in markov", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 541, + 361, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 361, + 551 + ], + "score": 1.0, + "content": "decision processes. Artificial Intelligence, 147(1-2):163–223, 2003.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 108, + 558, + 503, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and D Sculley. Google vizier:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 579, + 338, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 579, + 338, + 591 + ], + "score": 1.0, + "content": "knowledge discovery and data mining, pp. 1487–1495, 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 608 + ], + "score": 1.0, + "content": "Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 608, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 505, + 618 + ], + "score": 1.0, + "content": "regularization in matrix factorization. In Advances in Neural Information Processing Systems, pp. 6151–6159,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 615, + 140, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 615, + 140, + 628 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 504, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 645, + 480, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 645, + 480, + 657 + ], + "score": 1.0, + "content": "entropy deep reinforcement learning with a stochastic actor. arXiv preprint arXiv:1801.01290, 2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 663, + 504, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 673, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 115, + 673, + 506, + 685 + ], + "score": 1.0, + "content": "2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 684, + 222, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 684, + 222, + 694 + ], + "score": 1.0, + "content": "pp. 1735–1742. IEEE, 2006.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 701, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 702, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 712 + ], + "score": 1.0, + "content": "Pamela J Haley and DONALD Soloway. Extrapolation limitations of multilayer feedforward neural networks. In", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 712, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 506, + 722 + ], + "score": 1.0, + "content": "[Proceedings 1992] IJCNN International Joint Conference on Neural Networks, volume 4, pp. 25–30. IEEE,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 139, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 139, + 732 + ], + "score": 1.0, + "content": "1992.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 99, + 504, + 120 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 505, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 505, + 111 + ], + "score": 1.0, + "content": "Rishabh Agarwal, Chen Liang, Dale Schuurmans, and Mohammad Norouzi. Learning to generalize from sparse", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 109, + 278, + 120 + ], + "spans": [ + { + "bbox": [ + 116, + 109, + 278, + 120 + ], + "score": 1.0, + "content": "and underspecified rewards. In ICML, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 98, + 505, + 120 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 503, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "Martín Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. CoRR,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 137, + 201, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 201, + 148 + ], + "score": 1.0, + "content": "abs/1907.02893, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 127, + 505, + 148 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 155, + 503, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 168 + ], + "score": 1.0, + "content": "Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 389, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 389, + 177 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 7413–7424, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 154, + 505, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 471, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 471, + 205 + ], + "score": 1.0, + "content": "meta-regularization. In Advances in Neural Information Processing Systems, pp. 998–1008, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 182, + 505, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 211, + 503, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 505, + 224 + ], + "score": 1.0, + "content": "Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic Markov decision", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 443, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 443, + 234 + ], + "score": 1.0, + "content": "processes. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 211, + 505, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 239, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "Pablo Samuel Castro and Doina Precup. Using bisimulation for policy transfer in MDPs. In Proceedings of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 250, + 323, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 323, + 261 + ], + "score": 1.0, + "content": "AAAI Conference on Artificial Intelligence (AAAI), 2010.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 239, + 505, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 267, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 278, + 351, + 289 + ], + "spans": [ + { + "bbox": [ + 116, + 278, + 351, + 289 + ], + "score": 1.0, + "content": "learning of visual representations. CoRR, abs/2002.05709, 2020.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 267, + 505, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 504, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "Karl Cobbe, Oleg Klimov, Christopher Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 306, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 306, + 506, + 318 + ], + "score": 1.0, + "content": "reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 295, + 506, + 318 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 504, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "Andy Coenen and Adam Pearce. Understanding umap, 2019. URL https://pair-code.github.io/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 333, + 223, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 223, + 346 + ], + "score": 1.0, + "content": "understanding-umap/.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 323, + 505, + 346 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "Jesse Farebrother, Marlos C. Machado, and Michael Bowling. Generalization and regularization in DQN. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 362, + 322, + 374 + ], + "spans": [ + { + "bbox": [ + 114, + 362, + 322, + 374 + ], + "score": 1.0, + "content": "NeurIPS Deep Reinforcement Learning Workshop, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 351, + 505, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 392 + ], + "score": 1.0, + "content": "Norm Ferns, Prakash Panangaden, and Doina Precup. Metrics for finite Markov decision processes. In", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 390, + 418, + 402 + ], + "spans": [ + { + "bbox": [ + 116, + 390, + 418, + 402 + ], + "score": 1.0, + "content": "Proceedings of the Conference in Uncertainty in Artificial Intelligence (UAI), 2004.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 106, + 381, + 505, + 402 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 408, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "Norm Ferns, Pablo Samuel Castro, Doina Precup, and Prakash Panangaden. Methods for computing state", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 418, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 418, + 506, + 430 + ], + "score": 1.0, + "content": "similarity in Markov decision processes. In Proceedings of the 22nd Conference on Uncertainty in Artificial", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 267, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 267, + 440 + ], + "score": 1.0, + "content": "Intelligence, UAI ’06. AUAI Press, 2006.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 408, + 506, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 446, + 504, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "Norm Ferns, Prakash Panangaden, and Doina Precup. Bisimulation metrics for continuous markov decision", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 456, + 354, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 456, + 354, + 468 + ], + "score": 1.0, + "content": "processes. SIAM Journal on Computing, 40(6):1662–1714, 2011.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 446, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 503, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "score": 1.0, + "content": "Norman Ferns and Doina Precup. Bisimulation metrics are optimal value functions. In The 30th Conference on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 117, + 484, + 303, + 496 + ], + "spans": [ + { + "bbox": [ + 117, + 484, + 303, + 496 + ], + "score": 1.0, + "content": "Uncertainty in Artificial Intelligence, pp. 10, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 473, + 505, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 504, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 516 + ], + "score": 1.0, + "content": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 513, + 461, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 513, + 461, + 524 + ], + "score": 1.0, + "content": "networks. In Proceedings of the International Conference on Machine Learning (ICML), 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 501, + 505, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "score": 1.0, + "content": "Robert Givan, Thomas Dean, and Matthew Greig. Equivalence notions and model minimization in markov", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 541, + 361, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 361, + 551 + ], + "score": 1.0, + "content": "decision processes. Artificial Intelligence, 147(1-2):163–223, 2003.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 531, + 505, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 558, + 503, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and D Sculley. Google vizier:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 579, + 338, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 579, + 338, + 591 + ], + "score": 1.0, + "content": "knowledge discovery and data mining, pp. 1487–1495, 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 559, + 505, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 608 + ], + "score": 1.0, + "content": "Suriya Gunasekar, Blake E Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 608, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 505, + 618 + ], + "score": 1.0, + "content": "regularization in matrix factorization. In Advances in Neural Information Processing Systems, pp. 6151–6159,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 615, + 140, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 615, + 140, + 628 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 597, + 505, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 504, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 645, + 480, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 645, + 480, + 657 + ], + "score": 1.0, + "content": "entropy deep reinforcement learning with a stochastic actor. arXiv preprint arXiv:1801.01290, 2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 634, + 505, + 657 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 663, + 504, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 673, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 115, + 673, + 506, + 685 + ], + "score": 1.0, + "content": "2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 684, + 222, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 684, + 222, + 694 + ], + "score": 1.0, + "content": "pp. 1735–1742. IEEE, 2006.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 662, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 701, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 702, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 712 + ], + "score": 1.0, + "content": "Pamela J Haley and DONALD Soloway. Extrapolation limitations of multilayer feedforward neural networks. In", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 712, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 506, + 722 + ], + "score": 1.0, + "content": "[Proceedings 1992] IJCNN International Joint Conference on Neural Networks, volume 4, pp. 25–30. IEEE,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 139, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 139, + 732 + ], + "score": 1.0, + "content": "1992.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 702, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 83, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Maximilian Igl, Kamil Ciosek, Yingzhen Li, Sebastian Tschiatschek, Cheng Zhang, Sam Devlin, and Katja", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 104 + ], + "score": 1.0, + "content": "Hofmann. Generalization in reinforcement learning with selective noise injection and information bottleneck.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 103, + 408, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 408, + 115 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 13978–13990, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 120, + 504, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Domain extrapolation via regret minimization. arXiv", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 131, + 240, + 142 + ], + "spans": [ + { + "bbox": [ + 114, + 131, + 240, + 142 + ], + "score": 1.0, + "content": "preprint arXiv:2006.03908, 2020.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 108, + 149, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 161 + ], + "score": 1.0, + "content": "Arthur Juliani, Ahmed Khalifa, Vincent-Pierre Berges, Jonathan Harper, Ervin Teng, Hunter Henry, Adam", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 116, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "Crespi, Julian Togelius, and Danny Lange. Obstacle Tower: A generalization challenge in vision, control,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "and planning. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 188, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "Niels Justesen, Ruben Rodriguez Torrado, Philip Bontrager, Ahmed Khalifa, Julian Togelius, and Sebastian", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 199, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 117, + 199, + 505, + 209 + ], + "score": 1.0, + "content": "Risi. Illuminating generalization in deep reinforcement learning through procedural level generation. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 209, + 241, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 241, + 219 + ], + "score": 1.0, + "content": "preprint arXiv:1806.10729, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "Taylor W. Killian, George Dimitri Konidaris, and Finale Doshi-Velez. Robust and efficient transfer learning", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 237, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 505, + 248 + ], + "score": 1.0, + "content": "with hidden parameter Markov decision processes. In Proceedings of the AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 246, + 272, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 246, + 272, + 258 + ], + "score": 1.0, + "content": "Intelligence (AAAI), pp. 4949–4950, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 278 + ], + "score": 1.0, + "content": "Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 275, + 357, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 357, + 286 + ], + "score": 1.0, + "content": "reinforcement learning from pixels. CoRR, abs/2004.13649, 2020.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 292, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "score": 1.0, + "content": "Kim G Larsen and Arne Skou. Bisimulation through probabilistic testing. Information and computation, 94(1):", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 303, + 162, + 314 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 162, + 314 + ], + "score": 1.0, + "content": "1–28, 1991.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "Michael Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 331, + 342, + 342 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 342, + 342 + ], + "score": 1.0, + "content": "learning with augmented data. CoRR, abs/2004.14990, 2020a.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "reinforcement learning. Proceedings of the 37th International Conference on Machine Learning, Vienna,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 370, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 288, + 381 + ], + "score": 1.0, + "content": "Austria, PMLR 119, 2020b. arXiv:2003.06417.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "score": 1.0, + "content": "Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for general-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 398, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 398, + 506, + 410 + ], + "score": 1.0, + "content": "ization in deep reinforcement learning. In The International Conference on Learning Representations (ICLR),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 406, + 144, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 144, + 419 + ], + "score": 1.0, + "content": "2020a.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "score": 1.0, + "content": "Kuang-Huei Lee, Ian Fischer, Anthony Liu, Yijie Guo, Honglak Lee, John Canny, and Sergio Guadarrama.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 436, + 443, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 443, + 448 + ], + "score": 1.0, + "content": "Predictive information accelerates learning in rl. arXiv preprint arXiv:2007.12401, 2020b.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 466 + ], + "score": 1.0, + "content": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 464, + 488, + 477 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 488, + 477 + ], + "score": 1.0, + "content": "domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 504, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "score": 1.0, + "content": "Bogdan Mazoure, Remi Tachet des Combes, Thang Long DOAN, Philip Bachman, and R Devon Hjelm. Deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 493, + 479, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 493, + 479, + 504 + ], + "score": 1.0, + "content": "reinforcement and infomax learning. Advances in Neural Information Processing Systems, 33, 2020.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 521, + 272, + 532 + ], + "spans": [ + { + "bbox": [ + 116, + 521, + 272, + 532 + ], + "score": 1.0, + "content": "Reduction. ArXiv e-prints, February 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 105, + 539, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "Ted Moskovitz, Michael Arbel, Ferenc Huszar, and Arthur Gretton. Efficient wasserstein natural gradients for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 549, + 438, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 549, + 438, + 561 + ], + "score": 1.0, + "content": "reinforcement learning. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 567, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 578 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder P. Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "deep reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML),", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 104, + 606, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 507, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 507, + 619 + ], + "score": 1.0, + "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 616, + 262, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 616, + 262, + 627 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1807.03748, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 504, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "score": 1.0, + "content": "Aldo Pacchiano, Jack Parker-Holder, Yunhao Tang, Krzysztof Choromanski, Anna Choromanska, and Michael", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "Jordan. Learning to score behaviors for guided policy optimization. In International Conference on Machine", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 655, + 176, + 666 + ], + "spans": [ + { + "bbox": [ + 116, + 655, + 176, + 666 + ], + "score": 1.0, + "content": "Learning, 2020.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 105, + 672, + 503, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 686 + ], + "score": 1.0, + "content": "Charles Packer, Katelyn Gao, Jernej Kos, Philipp Krähenbühl, Vladlen Koltun, and Dawn Song. Assessing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 683, + 432, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 683, + 432, + 694 + ], + "score": 1.0, + "content": "generalization in deep reinforcement learning. arXiv preprint arXiv:1810.12282, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 106, + 701, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "Christian F. Perez, Felipe Petroski Such, and Theofanis Karaletsos. Generalized hidden parameter mdps", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "score": 1.0, + "content": "transferable model-based rl in a handful of trials. In Proceedings of the AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 213, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 213, + 732 + ], + "score": 1.0, + "content": "Intelligence (AAAI), 2020.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 83, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Maximilian Igl, Kamil Ciosek, Yingzhen Li, Sebastian Tschiatschek, Cheng Zhang, Sam Devlin, and Katja", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 104 + ], + "score": 1.0, + "content": "Hofmann. Generalization in reinforcement learning with selective noise injection and information bottleneck.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 103, + 408, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 408, + 115 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 13978–13990, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 120, + 504, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Domain extrapolation via regret minimization. arXiv", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 131, + 240, + 142 + ], + "spans": [ + { + "bbox": [ + 114, + 131, + 240, + 142 + ], + "score": 1.0, + "content": "preprint arXiv:2006.03908, 2020.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 120, + 505, + 142 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 149, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 161 + ], + "score": 1.0, + "content": "Arthur Juliani, Ahmed Khalifa, Vincent-Pierre Berges, Jonathan Harper, Ervin Teng, Hunter Henry, Adam", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 116, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "Crespi, Julian Togelius, and Danny Lange. Obstacle Tower: A generalization challenge in vision, control,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "and planning. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 150, + 506, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 188, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "Niels Justesen, Ruben Rodriguez Torrado, Philip Bontrager, Ahmed Khalifa, Julian Togelius, and Sebastian", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 199, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 117, + 199, + 505, + 209 + ], + "score": 1.0, + "content": "Risi. Illuminating generalization in deep reinforcement learning through procedural level generation. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 209, + 241, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 241, + 219 + ], + "score": 1.0, + "content": "preprint arXiv:1806.10729, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 188, + 505, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "Taylor W. Killian, George Dimitri Konidaris, and Finale Doshi-Velez. Robust and efficient transfer learning", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 237, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 505, + 248 + ], + "score": 1.0, + "content": "with hidden parameter Markov decision processes. In Proceedings of the AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 246, + 272, + 258 + ], + "spans": [ + { + "bbox": [ + 116, + 246, + 272, + 258 + ], + "score": 1.0, + "content": "Intelligence (AAAI), pp. 4949–4950, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 225, + 505, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 278 + ], + "score": 1.0, + "content": "Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 275, + 357, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 357, + 286 + ], + "score": 1.0, + "content": "reinforcement learning from pixels. CoRR, abs/2004.13649, 2020.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 262, + 505, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 292, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "score": 1.0, + "content": "Kim G Larsen and Arne Skou. Bisimulation through probabilistic testing. Information and computation, 94(1):", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 303, + 162, + 314 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 162, + 314 + ], + "score": 1.0, + "content": "1–28, 1991.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 291, + 506, + 314 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "Michael Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 331, + 342, + 342 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 342, + 342 + ], + "score": 1.0, + "content": "learning with augmented data. CoRR, abs/2004.14990, 2020a.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 321, + 505, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "reinforcement learning. Proceedings of the 37th International Conference on Machine Learning, Vienna,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 370, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 288, + 381 + ], + "score": 1.0, + "content": "Austria, PMLR 119, 2020b. arXiv:2003.06417.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 349, + 506, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "score": 1.0, + "content": "Kimin Lee, Kibok Lee, Jinwoo Shin, and Honglak Lee. Network randomization: A simple technique for general-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 398, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 398, + 506, + 410 + ], + "score": 1.0, + "content": "ization in deep reinforcement learning. In The International Conference on Learning Representations (ICLR),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 406, + 144, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 144, + 419 + ], + "score": 1.0, + "content": "2020a.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 388, + 506, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 504, + 438 + ], + "score": 1.0, + "content": "Kuang-Huei Lee, Ian Fischer, Anthony Liu, Yijie Guo, Honglak Lee, John Canny, and Sergio Guadarrama.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 436, + 443, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 443, + 448 + ], + "score": 1.0, + "content": "Predictive information accelerates learning in rl. arXiv preprint arXiv:2007.12401, 2020b.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 426, + 504, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 466 + ], + "score": 1.0, + "content": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 464, + 488, + 477 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 488, + 477 + ], + "score": 1.0, + "content": "domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 455, + 505, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 504, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "score": 1.0, + "content": "Bogdan Mazoure, Remi Tachet des Combes, Thang Long DOAN, Philip Bachman, and R Devon Hjelm. Deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 493, + 479, + 504 + ], + "spans": [ + { + "bbox": [ + 116, + 493, + 479, + 504 + ], + "score": 1.0, + "content": "reinforcement and infomax learning. Advances in Neural Information Processing Systems, 33, 2020.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 481, + 505, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 521, + 272, + 532 + ], + "spans": [ + { + "bbox": [ + 116, + 521, + 272, + 532 + ], + "score": 1.0, + "content": "Reduction. ArXiv e-prints, February 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 510, + 505, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 539, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "Ted Moskovitz, Michael Arbel, Ferenc Huszar, and Arthur Gretton. Efficient wasserstein natural gradients for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 549, + 438, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 549, + 438, + 561 + ], + "score": 1.0, + "content": "reinforcement learning. In International Conference on Learning Representations, 2021.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 539, + 505, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 567, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 578 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder P. Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "deep reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML),", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 568, + 506, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 606, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 507, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 507, + 619 + ], + "score": 1.0, + "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 616, + 262, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 616, + 262, + 627 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1807.03748, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 604, + 507, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 504, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "score": 1.0, + "content": "Aldo Pacchiano, Jack Parker-Holder, Yunhao Tang, Krzysztof Choromanski, Anna Choromanska, and Michael", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "Jordan. Learning to score behaviors for guided policy optimization. In International Conference on Machine", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 655, + 176, + 666 + ], + "spans": [ + { + "bbox": [ + 116, + 655, + 176, + 666 + ], + "score": 1.0, + "content": "Learning, 2020.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 634, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 672, + 503, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 686 + ], + "score": 1.0, + "content": "Charles Packer, Katelyn Gao, Jernej Kos, Philipp Krähenbühl, Vladlen Koltun, and Dawn Song. Assessing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 683, + 432, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 683, + 432, + 694 + ], + "score": 1.0, + "content": "generalization in deep reinforcement learning. arXiv preprint arXiv:1810.12282, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 670, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 701, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "Christian F. Perez, Felipe Petroski Such, and Theofanis Karaletsos. Generalized hidden parameter mdps", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "score": 1.0, + "content": "transferable model-based rl in a handful of trials. In Proceedings of the AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 213, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 213, + 732 + ], + "score": 1.0, + "content": "Intelligence (AAAI), 2020.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 106, + 700, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 52, + 507, + 734 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "score": 1.0, + "content": "Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 466, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 466, + 105 + ], + "score": 1.0, + "content": "The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 104, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "Martin L Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley &", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 180, + 133 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 180, + 133 + ], + "score": 1.0, + "content": "Sons, Inc., 1994.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 137, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 506, + 153 + ], + "score": 1.0, + "content": "Roberta Raileanu, Max Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 149, + 445, + 162 + ], + "spans": [ + { + "bbox": [ + 113, + 149, + 445, + 162 + ], + "score": 1.0, + "content": "for generalization in deep reinforcement learning. arXiv preprint arXiv:2006.12862, 2020.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards generalization and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 178, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 178, + 501, + 190 + ], + "score": 1.0, + "content": "simplicity in continuous control. In Advances in Neural Information Processing Systems (NeurIPS), 2017.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "Benjamin Recht. A tour of reinforcement learning: The view from continuous control. Annual Review of Control,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 207, + 270, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 270, + 217 + ], + "score": 1.0, + "content": "Robotics, and Autonomous Systems, 2019.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 235 + ], + "score": 1.0, + "content": "Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "efficient reinforcement learning with momentum predictive representations. arXiv preprint arXiv:2007.05929,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 243, + 141, + 255 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 141, + 255 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 262, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 506, + 274 + ], + "score": 1.0, + "content": "Anoopkumar Sonar, Vincent Pacelli, and Anirudha Majumdar. Invariant policy optimization: Towards stronger", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 113, + 271, + 413, + 284 + ], + "spans": [ + { + "bbox": [ + 113, + 271, + 413, + 284 + ], + "score": 1.0, + "content": "generalization in reinforcement learning. arXiv preprint arXiv:2006.01096, 2020.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 302 + ], + "score": 1.0, + "content": "Xingyou Song, Yiding Jiang, Yilun Du, and Behnam Neyshabur. Observational overfitting in reinforcement", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 299, + 428, + 311 + ], + "spans": [ + { + "bbox": [ + 114, + 299, + 428, + 311 + ], + "score": 1.0, + "content": "learning. In The International Conference on Learning Representations (ICLR), 2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 316, + 507, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 507, + 331 + ], + "score": 1.0, + "content": "Austin Stone, Oscar Ramirez, Kurt Konolige, and Rico Jonschkowski. The distracting control suite – a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 328, + 492, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 328, + 492, + 340 + ], + "score": 1.0, + "content": "challenging benchmark for reinforcement learning from pixels. arXiv preprint arXiv:2101.02722, 2021.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 113, + 355, + 351, + 368 + ], + "spans": [ + { + "bbox": [ + 113, + 355, + 351, + 368 + ], + "score": 1.0, + "content": "reinforcement learning. arXiv preprint arXiv:2009.08319, 2020.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 372, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 104, + 372, + 507, + 387 + ], + "score": 1.0, + "content": "Remi Tachet des Combes, Philip Bachman, and Harm van Seijen. Learning invariances for policy generalization.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 114, + 384, + 458, + 396 + ], + "spans": [ + { + "bbox": [ + 114, + 384, + 458, + 396 + ], + "score": 1.0, + "content": "In Workshop track at the International Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "score": 1.0, + "content": "Yujin Tang, Duong Nguyen, and David Ha. Neuroevolution of self-interpretable agents. arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 411, + 210, + 423 + ], + "spans": [ + { + "bbox": [ + 114, + 411, + 210, + 423 + ], + "score": 1.0, + "content": "arXiv:2003.08165, 2020.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "Yuval Tassa, Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 438, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 114, + 438, + 506, + 453 + ], + "score": 1.0, + "content": "Erez, Timothy Lillicrap, and Nicolas Heess. dm_control: Software and tasks for continuous control. arXiv", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 113, + 449, + 242, + 462 + ], + "spans": [ + { + "bbox": [ + 113, + 449, + 242, + 462 + ], + "score": 1.0, + "content": "preprint arXiv:2006.12983, 2020.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "score": 1.0, + "content": "Matthew E. Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. Journal of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 478, + 301, + 490 + ], + "spans": [ + { + "bbox": [ + 114, + 478, + 301, + 490 + ], + "score": 1.0, + "content": "Machine Learning Research, 10:1633–1685, 2009.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomiza-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 113, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 113, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "tion for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ International", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 113, + 515, + 402, + 528 + ], + "spans": [ + { + "bbox": [ + 113, + 515, + 402, + 528 + ], + "score": 1.0, + "content": "Conference on Intelligent Robots and Systems (IROS), pp. 23–30. IEEE, 2017.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "Elise van der Pol, Thomas Kipf, Frans A Oliehoek, and Max Welling. Plannable approximations to mdp", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "homomorphisms: Equivariance under actions. In Proceedings of the 19th International Conference on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 113, + 553, + 363, + 566 + ], + "spans": [ + { + "bbox": [ + 113, + 553, + 363, + 566 + ], + "score": 1.0, + "content": "Autonomous Agents and MultiAgent Systems, pp. 1431–1439, 2020.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 571, + 339, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 339, + 583 + ], + "score": 1.0, + "content": "Cédric Villani. Optimal transport: old and new. Springer, 2008.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Sam Witty, Jun Ki Lee, Emma Tosch, Akanksha Atrey, Michael Littman, and David Jensen. Measuring and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 600, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 600, + 506, + 612 + ], + "score": 1.0, + "content": "characterizing generalization in deep reinforcement learning. In NeurIPS Critiquing and Correcting Trends in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 610, + 247, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 610, + 247, + 622 + ], + "score": 1.0, + "content": "Machine Learning Workshop, 2018.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Keyulu Xu, Jingling Li, Mozhi Zhang, Simon S Du, Ken-ichi Kawarabayashi, and Stefanie Jegelka. How neural", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 113, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 113, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "networks extrapolate: From feedforward to graph neural networks. arXiv preprint arXiv:2009.11848, 2020.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "Chang Ye, Ahmed Khalifa, Philip Bontrager, and Julian Togelius. Rotation, translation, and cropping for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 666, + 318, + 677 + ], + "spans": [ + { + "bbox": [ + 115, + 666, + 318, + 677 + ], + "score": 1.0, + "content": "zero-shot generalization. CoRR, abs/2001.09908, 2020.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 684, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 506, + 695 + ], + "score": 1.0, + "content": "Amy Zhang, Nicolas Ballas, and Joelle Pineau. A Dissection of Overfitting and Generalization in Continuous", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 693, + 325, + 707 + ], + "spans": [ + { + "bbox": [ + 114, + 693, + 325, + 707 + ], + "score": 1.0, + "content": "Reinforcement Learning. CoRR, abs/1806.07937, 2018a.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 711, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 725 + ], + "score": 1.0, + "content": "Amy Zhang, Yuxin Wu, and Joelle Pineau. Natural environment benchmarks for reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 113, + 721, + 246, + 734 + ], + "spans": [ + { + "bbox": [ + 113, + 721, + 246, + 734 + ], + "score": 1.0, + "content": "preprint arXiv:1811.06032, 2018b.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 23.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 52, + 507, + 734 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 95 + ], + "score": 1.0, + "content": "Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool.", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 93, + 466, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 466, + 105 + ], + "score": 1.0, + "content": "The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017.", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 104, + 110, + 506, + 124 + ], + "score": 1.0, + "content": "Martin L Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley &", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 122, + 180, + 133 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 180, + 133 + ], + "score": 1.0, + "content": "Sons, Inc., 1994.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 137, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 506, + 153 + ], + "score": 1.0, + "content": "Roberta Raileanu, Max Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 149, + 445, + 162 + ], + "spans": [ + { + "bbox": [ + 113, + 149, + 445, + 162 + ], + "score": 1.0, + "content": "for generalization in deep reinforcement learning. arXiv preprint arXiv:2006.12862, 2020.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "score": 1.0, + "content": "Aravind Rajeswaran, Kendall Lowrey, Emanuel Todorov, and Sham M. Kakade. Towards generalization and", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 178, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 178, + 501, + 190 + ], + "score": 1.0, + "content": "simplicity in continuous control. In Advances in Neural Information Processing Systems (NeurIPS), 2017.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "Benjamin Recht. A tour of reinforcement learning: The view from continuous control. Annual Review of Control,", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 207, + 270, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 270, + 217 + ], + "score": 1.0, + "content": "Robotics, and Autonomous Systems, 2019.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 223, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 235 + ], + "score": 1.0, + "content": "Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "efficient reinforcement learning with momentum predictive representations. arXiv preprint arXiv:2007.05929,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 243, + 141, + 255 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 141, + 255 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 262, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 506, + 274 + ], + "score": 1.0, + "content": "Anoopkumar Sonar, Vincent Pacelli, and Anirudha Majumdar. Invariant policy optimization: Towards stronger", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 271, + 413, + 284 + ], + "spans": [ + { + "bbox": [ + 113, + 271, + 413, + 284 + ], + "score": 1.0, + "content": "generalization in reinforcement learning. arXiv preprint arXiv:2006.01096, 2020.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 288, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 302 + ], + "score": 1.0, + "content": "Xingyou Song, Yiding Jiang, Yilun Du, and Behnam Neyshabur. Observational overfitting in reinforcement", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 299, + 428, + 311 + ], + "spans": [ + { + "bbox": [ + 114, + 299, + 428, + 311 + ], + "score": 1.0, + "content": "learning. In The International Conference on Learning Representations (ICLR), 2019.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 316, + 507, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 507, + 331 + ], + "score": 1.0, + "content": "Austin Stone, Oscar Ramirez, Kurt Konolige, and Rico Jonschkowski. The distracting control suite – a", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 328, + 492, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 328, + 492, + 340 + ], + "score": 1.0, + "content": "challenging benchmark for reinforcement learning from pixels. arXiv preprint arXiv:2101.02722, 2021.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "Adam Stooke, Kimin Lee, Pieter Abbeel, and Michael Laskin. Decoupling representation learning from", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 355, + 351, + 368 + ], + "spans": [ + { + "bbox": [ + 113, + 355, + 351, + 368 + ], + "score": 1.0, + "content": "reinforcement learning. arXiv preprint arXiv:2009.08319, 2020.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 372, + 507, + 387 + ], + "spans": [ + { + "bbox": [ + 104, + 372, + 507, + 387 + ], + "score": 1.0, + "content": "Remi Tachet des Combes, Philip Bachman, and Harm van Seijen. Learning invariances for policy generalization.", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 384, + 458, + 396 + ], + "spans": [ + { + "bbox": [ + 114, + 384, + 458, + 396 + ], + "score": 1.0, + "content": "In Workshop track at the International Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "score": 1.0, + "content": "Yujin Tang, Duong Nguyen, and David Ha. Neuroevolution of self-interpretable agents. arXiv preprint", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 411, + 210, + 423 + ], + "spans": [ + { + "bbox": [ + 114, + 411, + 210, + 423 + ], + "score": 1.0, + "content": "arXiv:2003.08165, 2020.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "Yuval Tassa, Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 438, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 114, + 438, + 506, + 453 + ], + "score": 1.0, + "content": "Erez, Timothy Lillicrap, and Nicolas Heess. dm_control: Software and tasks for continuous control. arXiv", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 113, + 449, + 242, + 462 + ], + "spans": [ + { + "bbox": [ + 113, + 449, + 242, + 462 + ], + "score": 1.0, + "content": "preprint arXiv:2006.12983, 2020.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 506, + 479 + ], + "score": 1.0, + "content": "Matthew E. Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. Journal of", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 478, + 301, + 490 + ], + "spans": [ + { + "bbox": [ + 114, + 478, + 301, + 490 + ], + "score": 1.0, + "content": "Machine Learning Research, 10:1633–1685, 2009.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomiza-", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 113, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "tion for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ International", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 113, + 515, + 402, + 528 + ], + "spans": [ + { + "bbox": [ + 113, + 515, + 402, + 528 + ], + "score": 1.0, + "content": "Conference on Intelligent Robots and Systems (IROS), pp. 23–30. IEEE, 2017.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "Elise van der Pol, Thomas Kipf, Frans A Oliehoek, and Max Welling. Plannable approximations to mdp", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "homomorphisms: Equivariance under actions. In Proceedings of the 19th International Conference on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 113, + 553, + 363, + 566 + ], + "spans": [ + { + "bbox": [ + 113, + 553, + 363, + 566 + ], + "score": 1.0, + "content": "Autonomous Agents and MultiAgent Systems, pp. 1431–1439, 2020.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 571, + 339, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 339, + 583 + ], + "score": 1.0, + "content": "Cédric Villani. Optimal transport: old and new. Springer, 2008.", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Sam Witty, Jun Ki Lee, Emma Tosch, Akanksha Atrey, Michael Littman, and David Jensen. Measuring and", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 600, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 600, + 506, + 612 + ], + "score": 1.0, + "content": "characterizing generalization in deep reinforcement learning. In NeurIPS Critiquing and Correcting Trends in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 610, + 247, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 610, + 247, + 622 + ], + "score": 1.0, + "content": "Machine Learning Workshop, 2018.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Keyulu Xu, Jingling Li, Mozhi Zhang, Simon S Du, Ken-ichi Kawarabayashi, and Stefanie Jegelka. How neural", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 113, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "networks extrapolate: From feedforward to graph neural networks. arXiv preprint arXiv:2009.11848, 2020.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "Chang Ye, Ahmed Khalifa, Philip Bontrager, and Julian Togelius. Rotation, translation, and cropping for", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 666, + 318, + 677 + ], + "spans": [ + { + "bbox": [ + 115, + 666, + 318, + 677 + ], + "score": 1.0, + "content": "zero-shot generalization. CoRR, abs/2001.09908, 2020.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 684, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 506, + 695 + ], + "score": 1.0, + "content": "Amy Zhang, Nicolas Ballas, and Joelle Pineau. A Dissection of Overfitting and Generalization in Continuous", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 693, + 325, + 707 + ], + "spans": [ + { + "bbox": [ + 114, + 693, + 325, + 707 + ], + "score": 1.0, + "content": "Reinforcement Learning. CoRR, abs/1806.07937, 2018a.", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 711, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 725 + ], + "score": 1.0, + "content": "Amy Zhang, Yuxin Wu, and Joelle Pineau. Natural environment benchmarks for reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 721, + 246, + 734 + ], + "spans": [ + { + "bbox": [ + 113, + 721, + 246, + 734 + ], + "score": 1.0, + "content": "preprint arXiv:1811.06032, 2018b.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 83, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 504, + 113 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 94 + ], + "score": 1.0, + "content": "Amy Zhang, Clare Lyle, Shagun Sodhani, Angelos Filos, Marta Kwiatkowska, Joelle Pineau, Yarin Gal, and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "Doina Precup. Invariant causal prediction for block mdps. In Proceedings of the International Conference on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 239, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 239, + 114 + ], + "score": 1.0, + "content": "Machine Learning (ICML), 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 504, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant repre-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 130, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 114, + 130, + 506, + 144 + ], + "score": 1.0, + "content": "sentations for reinforcement learning without reconstruction. The International Conference on Learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 141, + 228, + 153 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 228, + 153 + ], + "score": 1.0, + "content": "Representations (ICLR), 2021.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 106, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Oriol Vinyals, Rémi Munos, and Samy Bengio. A study on overfitting in deep reinforcement", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 169, + 265, + 180 + ], + "spans": [ + { + "bbox": [ + 116, + 169, + 265, + 180 + ], + "score": 1.0, + "content": "learning. CoRR, abs/1804.06893, 2018c.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 504, + 113 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 94 + ], + "score": 1.0, + "content": "Amy Zhang, Clare Lyle, Shagun Sodhani, Angelos Filos, Marta Kwiatkowska, Joelle Pineau, Yarin Gal, and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "Doina Precup. Invariant causal prediction for block mdps. In Proceedings of the International Conference on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 239, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 239, + 114 + ], + "score": 1.0, + "content": "Machine Learning (ICML), 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 83, + 506, + 114 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 504, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant repre-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 130, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 114, + 130, + 506, + 144 + ], + "score": 1.0, + "content": "sentations for reinforcement learning without reconstruction. The International Conference on Learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 141, + 228, + 153 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 228, + 153 + ], + "score": 1.0, + "content": "Representations (ICLR), 2021.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 120, + 506, + 153 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 106, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Oriol Vinyals, Rémi Munos, and Samy Bengio. A study on overfitting in deep reinforcement", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 169, + 265, + 180 + ], + "spans": [ + { + "bbox": [ + 116, + 169, + 265, + 180 + ], + "score": 1.0, + "content": "learning. CoRR, abs/1804.06893, 2018c.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 158, + 505, + 180 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 193, + 101 + ], + "lines": [ + { + "bbox": [ + 103, + 77, + 196, + 103 + ], + "spans": [ + { + "bbox": [ + 103, + 77, + 196, + 103 + ], + "score": 1.0, + "content": "Appendix", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 106, + 118, + 380, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 118, + 380, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 380, + 134 + ], + "score": 1.0, + "content": "A LEARNING CONTRASTIVE METRIC EMBEDDINGS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 156, + 497, + 324 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 156, + 497, + 324 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 156, + 497, + 324 + ], + "spans": [ + { + "bbox": [ + 108, + 156, + 497, + 324 + ], + "score": 0.97, + "type": "image", + "image_path": "d19f18e3e5447a5a82c4130e02a0d8375292bd0ba4e9e7a07050c01ca385d18a.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 156, + 497, + 212.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 212.0, + 497, + 268.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 268.0, + 497, + 324.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 506, + 426 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 356, + 347 + ], + "score": 1.0, + "content": "Figure A.1: Architecture for learning CMEs. Given an input pair", + "type": "text" + }, + { + "bbox": [ + 356, + 335, + 378, + 345 + ], + "score": 0.89, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 334, + 506, + 347 + ], + "score": 1.0, + "content": ", we first apply the (optional) data", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 342, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 342, + 192, + 358 + ], + "score": 1.0, + "content": "augmentation operator", + "type": "text" + }, + { + "bbox": [ + 192, + 345, + 201, + 353 + ], + "score": 0.82, + "content": "\\Psi", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 342, + 336, + 358 + ], + "score": 1.0, + "content": "to produce the input augmentations", + "type": "text" + }, + { + "bbox": [ + 337, + 344, + 438, + 355 + ], + "score": 0.9, + "content": "\\bar { \\Psi } _ { x } : = \\Psi ( x ) , \\Psi _ { y } : = \\bar { \\Psi } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 342, + 506, + 358 + ], + "score": 1.0, + "content": ". When not using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 355, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 178, + 366 + ], + "score": 1.0, + "content": "data augmentation,", + "type": "text" + }, + { + "bbox": [ + 178, + 355, + 186, + 363 + ], + "score": 0.82, + "content": "\\Psi", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 355, + 338, + 366 + ], + "score": 1.0, + "content": "is equal to the identity operator, that is, ∀", + "type": "text" + }, + { + "bbox": [ + 338, + 355, + 382, + 365 + ], + "score": 0.88, + "content": "x \\Psi ( x ) = x", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 355, + 505, + 366 + ], + "score": 1.0, + "content": ". The agent’s policy network then", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 388, + 377 + ], + "score": 1.0, + "content": "outputs the representations for these augmentations by applying the encoder", + "type": "text" + }, + { + "bbox": [ + 388, + 365, + 397, + 375 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 364, + 428, + 377 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + }, + { + "bbox": [ + 428, + 364, + 478, + 375 + ], + "score": 0.81, + "content": "{ \\bar { f } } _ { x } = f _ { \\theta } ( \\Psi _ { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 364, + 483, + 377 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 484, + 365, + 505, + 375 + ], + "score": 0.76, + "content": "f _ { y } =", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 374, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 107, + 374, + 135, + 385 + ], + "score": 0.89, + "content": "f _ { \\theta } ( \\hat { \\Psi } _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 375, + 367, + 386 + ], + "score": 1.0, + "content": ". These representations are projected using a non-linear projector", + "type": "text" + }, + { + "bbox": [ + 367, + 375, + 378, + 384 + ], + "score": 0.87, + "content": "h _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 375, + 467, + 386 + ], + "score": 1.0, + "content": "to obtain the embedding", + "type": "text" + }, + { + "bbox": [ + 467, + 376, + 477, + 384 + ], + "score": 0.83, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 375, + 506, + 386 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 384, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 107, + 384, + 167, + 395 + ], + "score": 0.84, + "content": "z _ { \\theta } ( x ) = h _ { \\theta } ( f _ { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 384, + 172, + 396 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 172, + 384, + 231, + 396 + ], + "score": 0.91, + "content": "z _ { \\theta } ( y ) = h _ { \\theta } ( f _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 384, + 506, + 396 + ], + "score": 1.0, + "content": ". These metric embeddings are trained using the contrastive loss defined in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 393, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 104, + 393, + 200, + 408 + ], + "score": 1.0, + "content": "Equation (4). The policy", + "type": "text" + }, + { + "bbox": [ + 200, + 397, + 211, + 406 + ], + "score": 0.85, + "content": "\\pi \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 393, + 395, + 408 + ], + "score": 1.0, + "content": "is an affine function of the representation, that is,", + "type": "text" + }, + { + "bbox": [ + 396, + 394, + 477, + 406 + ], + "score": 0.93, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\cdot | \\boldsymbol { y } ) = \\boldsymbol { W } ^ { T } f _ { \\boldsymbol { y } } + \\boldsymbol { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 393, + 506, + 408 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 107, + 406, + 124, + 416 + ], + "score": 0.83, + "content": "W , b", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "are learned weights and biases. The entire network is trained end-to-end jointly using the reinforcement", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 415, + 415, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 415, + 427 + ], + "score": 1.0, + "content": "learning (or imitation learning) loss in conjunction with the auxiliary contrastive loss.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + } + ], + "index": 6.0 + }, + { + "type": "title", + "bbox": [ + 107, + 441, + 169, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 171, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 171, + 456 + ], + "score": 1.0, + "content": "B PROOFS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 467, + 429, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 467, + 430, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 430, + 481 + ], + "score": 1.0, + "content": "We begin by defining some notation which will be used throughout these results:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 133, + 488, + 466, + 549 + ], + "lines": [ + { + "bbox": [ + 131, + 488, + 466, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 184, + 510 + ], + "score": 1.0, + "content": "• We denote", + "type": "text" + }, + { + "bbox": [ + 184, + 488, + 466, + 510 + ], + "score": 0.83, + "content": "\\begin{array} { r } { { ! } \\mathbb { E } _ { t \\ge 0 } [ \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) ] = \\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) \\right] } \\end{array}", + "type": "inline_equation", + "image_path": "9bba4b6ae03b6c67df9f5d6f5b309d2258c2dfc07746a4b9ad9adb1bf8de16d5.jpg" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 361, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 175, + 532 + ], + "score": 1.0, + "content": "• For any", + "type": "text" + }, + { + "bbox": [ + 175, + 515, + 201, + 527 + ], + "score": 0.9, + "content": "y \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 512, + 218, + 532 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 218, + 514, + 292, + 529 + ], + "score": 0.93, + "content": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 512, + 322, + 532 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 323, + 514, + 355, + 529 + ], + "score": 0.93, + "content": "Y _ { y } ^ { 0 } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 512, + 361, + 532 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 531, + 345, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 138, + 551 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 139, + 533, + 345, + 549 + ], + "score": 0.74, + "content": "T V ^ { n } ( Y _ { y } ^ { k } ) = \\mathbb { E } _ { 0 \\leq t < n } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { k + t } ) , \\pi ^ { * } ( Y _ { y } ^ { k + t } ) ) .", + "type": "inline_equation", + "image_path": "e5d7ac590f7cc953d552608c2475ce80df62af79cb577ab4284fa91cdaaf3a5d.jpg" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 556, + 410, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 410, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 410, + 570 + ], + "score": 1.0, + "content": "We now proceed with some technical lemmas necessary for the main result.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 506, + 595 + ], + "lines": [ + { + "bbox": [ + 104, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 280, + 586 + ], + "score": 1.0, + "content": "Lemma 1. Given any two pseudometrics5", + "type": "text" + }, + { + "bbox": [ + 280, + 572, + 322, + 584 + ], + "score": 0.9, + "content": "d , d ^ { \\prime } \\in \\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 570, + 442, + 586 + ], + "score": 1.0, + "content": "and probability distributions", + "type": "text" + }, + { + "bbox": [ + 442, + 573, + 477, + 585 + ], + "score": 0.73, + "content": "P _ { X } , P _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 583, + 190, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 148, + 595 + ], + "score": 0.91, + "content": "\\mathcal { X } , \\mathcal { Y } \\subset \\mathcal { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 583, + 190, + 595 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 601, + 405, + 617 + ], + "lines": [ + { + "bbox": [ + 205, + 601, + 405, + 617 + ], + "spans": [ + { + "bbox": [ + 205, + 601, + 405, + 617 + ], + "score": 0.86, + "content": "\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )", + "type": "interline_equation", + "image_path": "178d0eb40c25e708a5e9dde06ff4afb48b7840e525fb666121160255643bd9c7.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 205, + 601, + 405, + 617 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 633, + 471, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 471, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 356, + 648 + ], + "score": 1.0, + "content": "Proof. Note that the dual of the linear program for computing", + "type": "text" + }, + { + "bbox": [ + 356, + 633, + 423, + 646 + ], + "score": 0.93, + "content": "\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 631, + 471, + 648 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 131, + 653, + 479, + 710 + ], + "lines": [ + { + "bbox": [ + 131, + 653, + 479, + 710 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 479, + 710 + ], + "score": 0.92, + "content": "\\begin{array} { l } { \\displaystyle \\mathop { \\operatorname* { m i n } } _ { \\textsc { i r } } \\sum _ { x \\in \\mathcal { X } , \\ y \\in \\mathcal { V } } \\Gamma ( x , y ) d ( x , y ) } \\\\ { \\mathrm { s u b j e c t ~ t o ~ } \\displaystyle \\sum _ { x } \\Gamma ( x , y ) = P y ( y ) \\forall y , \\quad \\displaystyle \\sum _ { y } \\Gamma ( x , y ) = P _ { \\mathcal { X } } ( x ) \\forall x , \\quad \\Gamma ( x , y ) \\geq 0 \\forall x , y } \\end{array}", + "type": "interline_equation", + "image_path": "268d7bdcbcf866dab677354dd6a9c0f34f95aed562188eff455f54455025a4a9.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 131, + 653, + 479, + 672.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 131, + 672.0, + 479, + 691.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 131, + 691.0, + 479, + 710.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 115, + 721, + 493, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 718, + 495, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 718, + 495, + 734 + ], + "score": 1.0, + "content": "5Pseudometrics are generalization of metrics where the distance between two distinct states can be zero.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 193, + 101 + ], + "lines": [ + { + "bbox": [ + 103, + 77, + 196, + 103 + ], + "spans": [ + { + "bbox": [ + 103, + 77, + 196, + 103 + ], + "score": 1.0, + "content": "Appendix", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 106, + 118, + 380, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 118, + 380, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 380, + 134 + ], + "score": 1.0, + "content": "A LEARNING CONTRASTIVE METRIC EMBEDDINGS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 156, + 497, + 324 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 156, + 497, + 324 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 156, + 497, + 324 + ], + "spans": [ + { + "bbox": [ + 108, + 156, + 497, + 324 + ], + "score": 0.97, + "type": "image", + "image_path": "d19f18e3e5447a5a82c4130e02a0d8375292bd0ba4e9e7a07050c01ca385d18a.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 156, + 497, + 212.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 212.0, + 497, + 268.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 268.0, + 497, + 324.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 506, + 426 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 356, + 347 + ], + "score": 1.0, + "content": "Figure A.1: Architecture for learning CMEs. Given an input pair", + "type": "text" + }, + { + "bbox": [ + 356, + 335, + 378, + 345 + ], + "score": 0.89, + "content": "( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 334, + 506, + 347 + ], + "score": 1.0, + "content": ", we first apply the (optional) data", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 342, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 342, + 192, + 358 + ], + "score": 1.0, + "content": "augmentation operator", + "type": "text" + }, + { + "bbox": [ + 192, + 345, + 201, + 353 + ], + "score": 0.82, + "content": "\\Psi", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 342, + 336, + 358 + ], + "score": 1.0, + "content": "to produce the input augmentations", + "type": "text" + }, + { + "bbox": [ + 337, + 344, + 438, + 355 + ], + "score": 0.9, + "content": "\\bar { \\Psi } _ { x } : = \\Psi ( x ) , \\Psi _ { y } : = \\bar { \\Psi } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 342, + 506, + 358 + ], + "score": 1.0, + "content": ". When not using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 355, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 178, + 366 + ], + "score": 1.0, + "content": "data augmentation,", + "type": "text" + }, + { + "bbox": [ + 178, + 355, + 186, + 363 + ], + "score": 0.82, + "content": "\\Psi", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 355, + 338, + 366 + ], + "score": 1.0, + "content": "is equal to the identity operator, that is, ∀", + "type": "text" + }, + { + "bbox": [ + 338, + 355, + 382, + 365 + ], + "score": 0.88, + "content": "x \\Psi ( x ) = x", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 355, + 505, + 366 + ], + "score": 1.0, + "content": ". The agent’s policy network then", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 388, + 377 + ], + "score": 1.0, + "content": "outputs the representations for these augmentations by applying the encoder", + "type": "text" + }, + { + "bbox": [ + 388, + 365, + 397, + 375 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 364, + 428, + 377 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + }, + { + "bbox": [ + 428, + 364, + 478, + 375 + ], + "score": 0.81, + "content": "{ \\bar { f } } _ { x } = f _ { \\theta } ( \\Psi _ { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 364, + 483, + 377 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 484, + 365, + 505, + 375 + ], + "score": 0.76, + "content": "f _ { y } =", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 374, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 107, + 374, + 135, + 385 + ], + "score": 0.89, + "content": "f _ { \\theta } ( \\hat { \\Psi } _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 375, + 367, + 386 + ], + "score": 1.0, + "content": ". These representations are projected using a non-linear projector", + "type": "text" + }, + { + "bbox": [ + 367, + 375, + 378, + 384 + ], + "score": 0.87, + "content": "h _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 375, + 467, + 386 + ], + "score": 1.0, + "content": "to obtain the embedding", + "type": "text" + }, + { + "bbox": [ + 467, + 376, + 477, + 384 + ], + "score": 0.83, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 375, + 506, + 386 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 384, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 107, + 384, + 167, + 395 + ], + "score": 0.84, + "content": "z _ { \\theta } ( x ) = h _ { \\theta } ( f _ { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 384, + 172, + 396 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 172, + 384, + 231, + 396 + ], + "score": 0.91, + "content": "z _ { \\theta } ( y ) = h _ { \\theta } ( f _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 384, + 506, + 396 + ], + "score": 1.0, + "content": ". These metric embeddings are trained using the contrastive loss defined in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 393, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 104, + 393, + 200, + 408 + ], + "score": 1.0, + "content": "Equation (4). The policy", + "type": "text" + }, + { + "bbox": [ + 200, + 397, + 211, + 406 + ], + "score": 0.85, + "content": "\\pi \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 393, + 395, + 408 + ], + "score": 1.0, + "content": "is an affine function of the representation, that is,", + "type": "text" + }, + { + "bbox": [ + 396, + 394, + 477, + 406 + ], + "score": 0.93, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( \\cdot | \\boldsymbol { y } ) = \\boldsymbol { W } ^ { T } f _ { \\boldsymbol { y } } + \\boldsymbol { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 393, + 506, + 408 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 107, + 406, + 124, + 416 + ], + "score": 0.83, + "content": "W , b", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "are learned weights and biases. The entire network is trained end-to-end jointly using the reinforcement", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 415, + 415, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 415, + 427 + ], + "score": 1.0, + "content": "learning (or imitation learning) loss in conjunction with the auxiliary contrastive loss.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + } + ], + "index": 6.0 + }, + { + "type": "title", + "bbox": [ + 107, + 441, + 169, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 171, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 171, + 456 + ], + "score": 1.0, + "content": "B PROOFS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 467, + 429, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 467, + 430, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 430, + 481 + ], + "score": 1.0, + "content": "We begin by defining some notation which will be used throughout these results:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 467, + 430, + 481 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 488, + 466, + 549 + ], + "lines": [ + { + "bbox": [ + 131, + 488, + 466, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 184, + 510 + ], + "score": 1.0, + "content": "• We denote", + "type": "text" + }, + { + "bbox": [ + 184, + 488, + 466, + 510 + ], + "score": 0.83, + "content": "\\begin{array} { r } { { ! } \\mathbb { E } _ { t \\ge 0 } [ \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) ] = \\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) \\right] } \\end{array}", + "type": "inline_equation", + "image_path": "9bba4b6ae03b6c67df9f5d6f5b309d2258c2dfc07746a4b9ad9adb1bf8de16d5.jpg" + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 361, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 175, + 532 + ], + "score": 1.0, + "content": "• For any", + "type": "text" + }, + { + "bbox": [ + 175, + 515, + 201, + 527 + ], + "score": 0.9, + "content": "y \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 512, + 218, + 532 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 218, + 514, + 292, + 529 + ], + "score": 0.93, + "content": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 512, + 322, + 532 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 323, + 514, + 355, + 529 + ], + "score": 0.93, + "content": "Y _ { y } ^ { 0 } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 512, + 361, + 532 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 531, + 345, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 138, + 551 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 139, + 533, + 345, + 549 + ], + "score": 0.74, + "content": "T V ^ { n } ( Y _ { y } ^ { k } ) = \\mathbb { E } _ { 0 \\leq t < n } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { k + t } ) , \\pi ^ { * } ( Y _ { y } ^ { k + t } ) ) .", + "type": "inline_equation", + "image_path": "e5d7ac590f7cc953d552608c2475ce80df62af79cb577ab4284fa91cdaaf3a5d.jpg" + } + ], + "index": 18, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 17, + "bbox_fs": [ + 131, + 488, + 466, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 556, + 410, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 410, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 410, + 570 + ], + "score": 1.0, + "content": "We now proceed with some technical lemmas necessary for the main result.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 555, + 410, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 506, + 595 + ], + "lines": [ + { + "bbox": [ + 104, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 280, + 586 + ], + "score": 1.0, + "content": "Lemma 1. Given any two pseudometrics5", + "type": "text" + }, + { + "bbox": [ + 280, + 572, + 322, + 584 + ], + "score": 0.9, + "content": "d , d ^ { \\prime } \\in \\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 570, + 442, + 586 + ], + "score": 1.0, + "content": "and probability distributions", + "type": "text" + }, + { + "bbox": [ + 442, + 573, + 477, + 585 + ], + "score": 0.73, + "content": "P _ { X } , P _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 583, + 190, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 148, + 595 + ], + "score": 0.91, + "content": "\\mathcal { X } , \\mathcal { Y } \\subset \\mathcal { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 583, + 190, + 595 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 570, + 506, + 595 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 601, + 405, + 617 + ], + "lines": [ + { + "bbox": [ + 205, + 601, + 405, + 617 + ], + "spans": [ + { + "bbox": [ + 205, + 601, + 405, + 617 + ], + "score": 0.86, + "content": "\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )", + "type": "interline_equation", + "image_path": "178d0eb40c25e708a5e9dde06ff4afb48b7840e525fb666121160255643bd9c7.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 205, + 601, + 405, + 617 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 633, + 471, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 471, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 356, + 648 + ], + "score": 1.0, + "content": "Proof. Note that the dual of the linear program for computing", + "type": "text" + }, + { + "bbox": [ + 356, + 633, + 423, + 646 + ], + "score": 0.93, + "content": "\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 631, + 471, + 648 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 631, + 471, + 648 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 131, + 653, + 479, + 710 + ], + "lines": [ + { + "bbox": [ + 131, + 653, + 479, + 710 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 479, + 710 + ], + "score": 0.92, + "content": "\\begin{array} { l } { \\displaystyle \\mathop { \\operatorname* { m i n } } _ { \\textsc { i r } } \\sum _ { x \\in \\mathcal { X } , \\ y \\in \\mathcal { V } } \\Gamma ( x , y ) d ( x , y ) } \\\\ { \\mathrm { s u b j e c t ~ t o ~ } \\displaystyle \\sum _ { x } \\Gamma ( x , y ) = P y ( y ) \\forall y , \\quad \\displaystyle \\sum _ { y } \\Gamma ( x , y ) = P _ { \\mathcal { X } } ( x ) \\forall x , \\quad \\Gamma ( x , y ) \\geq 0 \\forall x , y } \\end{array}", + "type": "interline_equation", + "image_path": "268d7bdcbcf866dab677354dd6a9c0f34f95aed562188eff455f54455025a4a9.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 131, + 653, + 479, + 672.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 131, + 672.0, + 479, + 691.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 131, + 691.0, + 479, + 710.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 115, + 721, + 493, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 718, + 495, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 718, + 495, + 734 + ], + "score": 1.0, + "content": "5Pseudometrics are generalization of metrics where the distance between two distinct states can be zero.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 118, + 718, + 495, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 446, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 447, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 346, + 96 + ], + "score": 1.0, + "content": "Using the dual formulation subject to the constraints above,", + "type": "text" + }, + { + "bbox": [ + 347, + 82, + 376, + 95 + ], + "score": 0.93, + "content": "\\mathscr { W } ^ { 1 } ( d )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 81, + 447, + 96 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 97, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 97, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 97, + 479, + 215 + ], + "score": 0.95, + "content": "\\begin{array} { l } { \\displaystyle \\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| = \\mathcal { W } ^ { 1 } ( d - d ^ { \\prime } + d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| } \\\\ { \\displaystyle \\quad \\quad \\quad = \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( d ( x , y ) - d ^ { \\prime } ( x , y ) + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad \\leq \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( \\| d - d ^ { \\prime } \\| + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) d ^ { \\prime } ( x , y ) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) } \\end{array}", + "type": "interline_equation", + "image_path": "3e0255eb3d0bd9d3ccace9e0a20f8b685338115f46f5e790bb6653e56f406bba.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 133, + 97, + 479, + 136.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 133, + 136.33333333333334, + 479, + 175.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 133, + 175.66666666666669, + 479, + 215.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 269, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 270, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 197, + 246 + ], + "score": 1.0, + "content": "Lemma 2. Given any", + "type": "text" + }, + { + "bbox": [ + 198, + 234, + 227, + 245 + ], + "score": 0.91, + "content": "y _ { 0 } \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 232, + 270, + 246 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 248, + 454, + 278 + ], + "lines": [ + { + "bbox": [ + 156, + 248, + 454, + 278 + ], + "spans": [ + { + "bbox": [ + 156, + 248, + 454, + 278 + ], + "score": 0.91, + "content": "\\sum _ { y _ { 1 } \\in Y } \\left( P ^ { \\tilde { \\pi } } ( y _ { 1 } | y _ { 0 } ) - P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\right) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 2 } { 1 - \\gamma } T V ( \\tilde { \\pi } ( y _ { 0 } ) , \\pi ^ { * } ( y _ { 0 } ) )", + "type": "interline_equation", + "image_path": "68534635c2d9afb6aac4e8a0c836e0007df72fd4318165876c6f465a2993771a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 156, + 248, + 454, + 278 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 133, + 300 + ], + "lines": [ + { + "bbox": [ + 104, + 286, + 135, + 303 + ], + "spans": [ + { + "bbox": [ + 104, + 286, + 135, + 303 + ], + "score": 1.0, + "content": "Proof.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 302, + 514, + 499 + ], + "lines": [ + { + "bbox": [ + 111, + 302, + 514, + 499 + ], + "spans": [ + { + "bbox": [ + 111, + 302, + 514, + 499 + ], + "score": 0.88, + "content": "\\begin{array} { r l } { \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | | B \\rangle } ) - P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) \\leq | \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) - P ^ { \\kappa ^ { \\star } } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) | } & { } \\\\ & { \\leq \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha | \\hat { \\sigma } ( a ) | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( a | \\mathfrak { g } _ { 1 } ) ) | T V ^ { n } ( Y _ { n } ^ { 0 } ) } \\\\ & { \\leq \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\sum _ { i \\in \\mathcal { K } } P ^ { \\kappa } ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) | \\overline { { \\tau } } ( \\mathfrak { g } ( | \\mathfrak { g } _ { 1 } \\rangle ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } ) | ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & = \\ \\end{array}", + "type": "interline_equation", + "image_path": "01dff727690cafb5108175805d5f99671b922013be8e9969d5cb591cbf014423.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 111, + 302, + 514, + 367.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 367.6666666666667, + 514, + 433.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 433.33333333333337, + 514, + 499.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 401, + 529 + ], + "lines": [ + { + "bbox": [ + 103, + 511, + 404, + 532 + ], + "spans": [ + { + "bbox": [ + 103, + 511, + 197, + 532 + ], + "score": 1.0, + "content": "Lemma 3. Given any", + "type": "text" + }, + { + "bbox": [ + 198, + 516, + 228, + 527 + ], + "score": 0.86, + "content": "y _ { 0 } \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 511, + 238, + 532 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 239, + 514, + 359, + 530 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\mathbf { \\dot { \\mathbf { \\theta } } T } V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 511, + 404, + 532 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 532, + 449, + 563 + ], + "lines": [ + { + "bbox": [ + 162, + 532, + 449, + 563 + ], + "spans": [ + { + "bbox": [ + 162, + 532, + 449, + 563 + ], + "score": 0.9, + "content": "\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq { \\frac { 1 + \\gamma } { 1 - \\gamma } } W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | { \\tilde { x } _ { y _ { 0 } } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)", + "type": "interline_equation", + "image_path": "520af0569c33b4cfea5e3279b677c5d78e474c4192797189faf1aec9512d738d.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 162, + 532, + 449, + 542.3333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 162, + 542.3333333333334, + 449, + 552.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 162, + 552.6666666666667, + 449, + 563.0000000000001 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 573, + 419, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 419, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 419, + 586 + ], + "score": 1.0, + "content": "Proof. Note that we have the following equality, where 0 is a vector of zeros:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 587, + 466, + 615 + ], + "lines": [ + { + "bbox": [ + 144, + 587, + 466, + 615 + ], + "spans": [ + { + "bbox": [ + 144, + 587, + 466, + 615 + ], + "score": 0.91, + "content": "\\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) = \\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) - \\sum _ { x \\in X } P ^ { \\pi ^ { * } } ( x | \\tilde { x } _ { y _ { 0 } } ) \\mathbf { 0 }", + "type": "interline_equation", + "image_path": "d8c5f18a057edae5a97b8e3461c34ea7880dc52b2cb709d7f988f6651fad4b25.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 144, + 587, + 466, + 615 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 618, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 103, + 615, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 103, + 615, + 290, + 634 + ], + "score": 1.0, + "content": "which is the same form as the primal LP for", + "type": "text" + }, + { + "bbox": [ + 290, + 618, + 419, + 632 + ], + "score": 0.92, + "content": "W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 615, + 506, + 634 + ], + "score": 1.0, + "content": ". By assumption, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 629, + 146, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 146, + 642 + ], + "score": 1.0, + "content": "have that", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 637, + 370, + 663 + ], + "lines": [ + { + "bbox": [ + 240, + 637, + 370, + 663 + ], + "spans": [ + { + "bbox": [ + 240, + 637, + 370, + 663 + ], + "score": 0.93, + "content": "T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { \\ast } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } )", + "type": "interline_equation", + "image_path": "7d973b469ae0e5fef9c6eb192c403af32c2c01f146bf623d455ccb1d8ccc4515.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 240, + 637, + 370, + 663 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 464, + 685 + ], + "lines": [ + { + "bbox": [ + 104, + 667, + 461, + 688 + ], + "spans": [ + { + "bbox": [ + 104, + 667, + 176, + 688 + ], + "score": 1.0, + "content": "This implies that", + "type": "text" + }, + { + "bbox": [ + 176, + 670, + 235, + 686 + ], + "score": 0.93, + "content": "\\textstyle { \\frac { 1 - \\gamma } { 1 + \\gamma } } T V ^ { n } ( Y _ { . } ^ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 667, + 331, + 688 + ], + "score": 1.0, + "content": "is a feasible solution to", + "type": "text" + }, + { + "bbox": [ + 331, + 670, + 461, + 684 + ], + "score": 0.89, + "content": "W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 687, + 449, + 718 + ], + "lines": [ + { + "bbox": [ + 162, + 687, + 449, + 718 + ], + "spans": [ + { + "bbox": [ + 162, + 687, + 449, + 718 + ], + "score": 0.92, + "content": "\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\frac { 1 - \\gamma } { 1 + \\gamma } T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)", + "type": "interline_equation", + "image_path": "d7571a1589112c9b127e22f040e1034d276c161a5718ca8181542e6a29f1989c.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 162, + 687, + 449, + 697.3333333333334 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 162, + 697.3333333333334, + 449, + 707.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 162, + 707.6666666666667, + 449, + 718.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 196, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 196, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 196, + 732 + ], + "score": 1.0, + "content": "and the result follows.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 721, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 496, + 722, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 496, + 722, + 505, + 732 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 498, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 496, + 500, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 496, + 500, + 505, + 509 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 216, + 505, + 227 + ], + "lines": [ + { + "bbox": [ + 496, + 218, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 496, + 218, + 505, + 228 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 446, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 447, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 346, + 96 + ], + "score": 1.0, + "content": "Using the dual formulation subject to the constraints above,", + "type": "text" + }, + { + "bbox": [ + 347, + 82, + 376, + 95 + ], + "score": 0.93, + "content": "\\mathscr { W } ^ { 1 } ( d )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 81, + 447, + 96 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 447, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 97, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 97, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 97, + 479, + 215 + ], + "score": 0.95, + "content": "\\begin{array} { l } { \\displaystyle \\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| = \\mathcal { W } ^ { 1 } ( d - d ^ { \\prime } + d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| } \\\\ { \\displaystyle \\quad \\quad \\quad = \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( d ( x , y ) - d ^ { \\prime } ( x , y ) + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad \\leq \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( \\| d - d ^ { \\prime } \\| + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) d ^ { \\prime } ( x , y ) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) } \\end{array}", + "type": "interline_equation", + "image_path": "3e0255eb3d0bd9d3ccace9e0a20f8b685338115f46f5e790bb6653e56f406bba.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 133, + 97, + 479, + 136.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 133, + 136.33333333333334, + 479, + 175.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 133, + 175.66666666666669, + 479, + 215.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 269, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 270, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 197, + 246 + ], + "score": 1.0, + "content": "Lemma 2. Given any", + "type": "text" + }, + { + "bbox": [ + 198, + 234, + 227, + 245 + ], + "score": 0.91, + "content": "y _ { 0 } \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 232, + 270, + 246 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 232, + 270, + 246 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 248, + 454, + 278 + ], + "lines": [ + { + "bbox": [ + 156, + 248, + 454, + 278 + ], + "spans": [ + { + "bbox": [ + 156, + 248, + 454, + 278 + ], + "score": 0.91, + "content": "\\sum _ { y _ { 1 } \\in Y } \\left( P ^ { \\tilde { \\pi } } ( y _ { 1 } | y _ { 0 } ) - P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\right) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 2 } { 1 - \\gamma } T V ( \\tilde { \\pi } ( y _ { 0 } ) , \\pi ^ { * } ( y _ { 0 } ) )", + "type": "interline_equation", + "image_path": "68534635c2d9afb6aac4e8a0c836e0007df72fd4318165876c6f465a2993771a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 156, + 248, + 454, + 278 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 133, + 300 + ], + "lines": [ + { + "bbox": [ + 104, + 286, + 135, + 303 + ], + "spans": [ + { + "bbox": [ + 104, + 286, + 135, + 303 + ], + "score": 1.0, + "content": "Proof.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 104, + 286, + 135, + 303 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 302, + 514, + 499 + ], + "lines": [ + { + "bbox": [ + 111, + 302, + 514, + 499 + ], + "spans": [ + { + "bbox": [ + 111, + 302, + 514, + 499 + ], + "score": 0.88, + "content": "\\begin{array} { r l } { \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | | B \\rangle } ) - P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) \\leq | \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) - P ^ { \\kappa ^ { \\star } } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) | } & { } \\\\ & { \\leq \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha | \\hat { \\sigma } ( a ) | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( a | \\mathfrak { g } _ { 1 } ) ) | T V ^ { n } ( Y _ { n } ^ { 0 } ) } \\\\ & { \\leq \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\sum _ { i \\in \\mathcal { K } } P ^ { \\kappa } ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) | \\overline { { \\tau } } ( \\mathfrak { g } ( | \\mathfrak { g } _ { 1 } \\rangle ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } ) | ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & = \\ \\end{array}", + "type": "interline_equation", + "image_path": "01dff727690cafb5108175805d5f99671b922013be8e9969d5cb591cbf014423.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 111, + 302, + 514, + 367.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 367.6666666666667, + 514, + 433.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 433.33333333333337, + 514, + 499.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 401, + 529 + ], + "lines": [ + { + "bbox": [ + 103, + 511, + 404, + 532 + ], + "spans": [ + { + "bbox": [ + 103, + 511, + 197, + 532 + ], + "score": 1.0, + "content": "Lemma 3. Given any", + "type": "text" + }, + { + "bbox": [ + 198, + 516, + 228, + 527 + ], + "score": 0.86, + "content": "y _ { 0 } \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 511, + 238, + 532 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 239, + 514, + 359, + 530 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\mathbf { \\dot { \\mathbf { \\theta } } T } V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 511, + 404, + 532 + ], + "score": 1.0, + "content": ", we have:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 103, + 511, + 404, + 532 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 532, + 449, + 563 + ], + "lines": [ + { + "bbox": [ + 162, + 532, + 449, + 563 + ], + "spans": [ + { + "bbox": [ + 162, + 532, + 449, + 563 + ], + "score": 0.9, + "content": "\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq { \\frac { 1 + \\gamma } { 1 - \\gamma } } W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | { \\tilde { x } _ { y _ { 0 } } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)", + "type": "interline_equation", + "image_path": "520af0569c33b4cfea5e3279b677c5d78e474c4192797189faf1aec9512d738d.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 162, + 532, + 449, + 542.3333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 162, + 542.3333333333334, + 449, + 552.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 162, + 552.6666666666667, + 449, + 563.0000000000001 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 573, + 419, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 419, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 419, + 586 + ], + "score": 1.0, + "content": "Proof. Note that we have the following equality, where 0 is a vector of zeros:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 572, + 419, + 586 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 587, + 466, + 615 + ], + "lines": [ + { + "bbox": [ + 144, + 587, + 466, + 615 + ], + "spans": [ + { + "bbox": [ + 144, + 587, + 466, + 615 + ], + "score": 0.91, + "content": "\\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) = \\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) - \\sum _ { x \\in X } P ^ { \\pi ^ { * } } ( x | \\tilde { x } _ { y _ { 0 } } ) \\mathbf { 0 }", + "type": "interline_equation", + "image_path": "d8c5f18a057edae5a97b8e3461c34ea7880dc52b2cb709d7f988f6651fad4b25.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 144, + 587, + 466, + 615 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 618, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 103, + 615, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 103, + 615, + 290, + 634 + ], + "score": 1.0, + "content": "which is the same form as the primal LP for", + "type": "text" + }, + { + "bbox": [ + 290, + 618, + 419, + 632 + ], + "score": 0.92, + "content": "W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 615, + 506, + 634 + ], + "score": 1.0, + "content": ". By assumption, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 629, + 146, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 146, + 642 + ], + "score": 1.0, + "content": "have that", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 103, + 615, + 506, + 642 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 637, + 370, + 663 + ], + "lines": [ + { + "bbox": [ + 240, + 637, + 370, + 663 + ], + "spans": [ + { + "bbox": [ + 240, + 637, + 370, + 663 + ], + "score": 0.93, + "content": "T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { \\ast } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } )", + "type": "interline_equation", + "image_path": "7d973b469ae0e5fef9c6eb192c403af32c2c01f146bf623d455ccb1d8ccc4515.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 240, + 637, + 370, + 663 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 464, + 685 + ], + "lines": [ + { + "bbox": [ + 104, + 667, + 461, + 688 + ], + "spans": [ + { + "bbox": [ + 104, + 667, + 176, + 688 + ], + "score": 1.0, + "content": "This implies that", + "type": "text" + }, + { + "bbox": [ + 176, + 670, + 235, + 686 + ], + "score": 0.93, + "content": "\\textstyle { \\frac { 1 - \\gamma } { 1 + \\gamma } } T V ^ { n } ( Y _ { . } ^ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 667, + 331, + 688 + ], + "score": 1.0, + "content": "is a feasible solution to", + "type": "text" + }, + { + "bbox": [ + 331, + 670, + 461, + 684 + ], + "score": 0.89, + "content": "W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 667, + 461, + 688 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 687, + 449, + 718 + ], + "lines": [ + { + "bbox": [ + 162, + 687, + 449, + 718 + ], + "spans": [ + { + "bbox": [ + 162, + 687, + 449, + 718 + ], + "score": 0.92, + "content": "\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\frac { 1 - \\gamma } { 1 + \\gamma } T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)", + "type": "interline_equation", + "image_path": "d7571a1589112c9b127e22f040e1034d276c161a5718ca8181542e6a29f1989c.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 162, + 687, + 449, + 697.3333333333334 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 162, + 697.3333333333334, + 449, + 707.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 162, + 707.6666666666667, + 449, + 718.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 196, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 196, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 196, + 732 + ], + "score": 1.0, + "content": "and the result follows.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 720, + 196, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 284, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 284, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 234, + 97 + ], + "score": 1.0, + "content": "Proposition B.1. The operator", + "type": "text" + }, + { + "bbox": [ + 235, + 83, + 244, + 93 + ], + "score": 0.82, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 80, + 284, + 97 + ], + "score": 1.0, + "content": "given by:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 170, + 98, + 441, + 113 + ], + "lines": [ + { + "bbox": [ + 170, + 98, + 441, + 113 + ], + "spans": [ + { + "bbox": [ + 170, + 98, + 441, + 113 + ], + "score": 0.88, + "content": "\\mathcal F ( d ) ( x , y ) = \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) + \\gamma \\mathcal W _ { 1 } ( d ) \\big ( P _ { \\mathcal X } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal V } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big )", + "type": "interline_equation", + "image_path": "50037dbaadaa611b3c2b7307b310ef9effa7d80a0cfe71252bc574a5d4e1c2c3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 170, + 98, + 441, + 113 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 117, + 403, + 129 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 404, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 404, + 131 + ], + "score": 1.0, + "content": "is a contraction mapping and has a unique fixed point for a bounded dist.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 211, + 153 + ], + "score": 1.0, + "content": "Proof. We first prove that", + "type": "text" + }, + { + "bbox": [ + 211, + 142, + 220, + 151 + ], + "score": 0.86, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "is contraction mapping. Then, a simple application of the Banach Fixed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 152, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 218, + 164 + ], + "score": 1.0, + "content": "Point Theorem asserts that", + "type": "text" + }, + { + "bbox": [ + 218, + 153, + 227, + 162 + ], + "score": 0.86, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 152, + 461, + 164 + ], + "score": 1.0, + "content": "has a unique fixed point. Note that for all pseudometrics", + "type": "text" + }, + { + "bbox": [ + 461, + 152, + 502, + 163 + ], + "score": 0.92, + "content": "d , d ^ { \\prime } \\in \\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 152, + 506, + 164 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 163, + 236, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 174, + 174 + ], + "score": 1.0, + "content": "and for all states", + "type": "text" + }, + { + "bbox": [ + 175, + 163, + 202, + 173 + ], + "score": 0.84, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 163, + 206, + 174 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 206, + 163, + 232, + 174 + ], + "score": 0.82, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 163, + 236, + 174 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 177, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 128, + 177, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 128, + 177, + 482, + 255 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { F } ( d ) ( x , y ) - \\mathcal { F } ( d ^ { \\prime } ) ( x , y ) } \\\\ & { \\quad = \\gamma \\left( \\mathcal { W } _ { 1 } ( d ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad \\overset { L e m m a ~ \\ : 1 } { \\leq } \\gamma \\left( \\| d - d ^ { \\prime } \\| + \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad = \\gamma \\| d - d ^ { \\prime } \\| } \\end{array}", + "type": "interline_equation", + "image_path": "e26031ddb7f6793c820c994471ce5075a5e1d92786e164ddfeb0013f881bdb40.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 128, + 177, + 482, + 203.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 128, + 203.0, + 482, + 229.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 128, + 229.0, + 482, + 255.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 271, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 131, + 285 + ], + "score": 1.0, + "content": "Thus,", + "type": "text" + }, + { + "bbox": [ + 131, + 272, + 252, + 284 + ], + "score": 0.92, + "content": "\\| \\mathcal F ( d ) - \\mathcal F ( d ^ { \\prime } ) \\| \\le \\gamma \\| d - d ^ { \\prime } \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 270, + 283, + 285 + ], + "score": 1.0, + "content": ", so that", + "type": "text" + }, + { + "bbox": [ + 284, + 272, + 293, + 282 + ], + "score": 0.85, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 270, + 406, + 285 + ], + "score": 1.0, + "content": "is a contracting mapping for", + "type": "text" + }, + { + "bbox": [ + 406, + 273, + 431, + 284 + ], + "score": 0.91, + "content": "\\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 270, + 506, + 285 + ], + "score": 1.0, + "content": "and has an unique", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 281, + 167, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 152, + 295 + ], + "score": 1.0, + "content": "fixed point", + "type": "text" + }, + { + "bbox": [ + 152, + 284, + 162, + 293 + ], + "score": 0.85, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 281, + 167, + 295 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 502, + 326 + ], + "lines": [ + { + "bbox": [ + 103, + 296, + 507, + 319 + ], + "spans": [ + { + "bbox": [ + 103, + 296, + 303, + 319 + ], + "score": 1.0, + "content": "Theorem 1. [Bound on policy transfer] For any", + "type": "text" + }, + { + "bbox": [ + 303, + 301, + 329, + 312 + ], + "score": 0.9, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 296, + 345, + 319 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 346, + 300, + 423, + 314 + ], + "score": 0.93, + "content": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 296, + 507, + 319 + ], + "score": 1.0, + "content": "define the sequence", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 311, + 439, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 270, + 329 + ], + "score": 1.0, + "content": "of random states encountered starting in", + "type": "text" + }, + { + "bbox": [ + 271, + 313, + 303, + 327 + ], + "score": 0.92, + "content": "Y _ { y } ^ { 0 } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 311, + 388, + 329 + ], + "score": 1.0, + "content": "and following policy", + "type": "text" + }, + { + "bbox": [ + 388, + 315, + 395, + 324 + ], + "score": 0.64, + "content": "\\tilde { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 311, + 439, + 329 + ], + "score": 1.0, + "content": ". We have:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 329, + 419, + 369 + ], + "lines": [ + { + "bbox": [ + 191, + 329, + 419, + 369 + ], + "spans": [ + { + "bbox": [ + 191, + 329, + 419, + 369 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) .", + "type": "interline_equation", + "image_path": "78300f5f5ad236d336ba4cea178e615dab71051f4cf522680b2ec7241d02a039.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 329, + 419, + 349.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 191, + 349.0, + 419, + 369.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 378, + 506, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 400, + 390 + ], + "score": 1.0, + "content": "Proof. We will prove this by induction. Assuming that the bound holds for", + "type": "text" + }, + { + "bbox": [ + 401, + 380, + 422, + 389 + ], + "score": 0.87, + "content": "T V ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 378, + 505, + 390 + ], + "score": 1.0, + "content": ", we prove the bound", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 387, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 104, + 387, + 144, + 404 + ], + "score": 1.0, + "content": "holds for", + "type": "text" + }, + { + "bbox": [ + 144, + 390, + 175, + 400 + ], + "score": 0.9, + "content": "T V ^ { n + \\bar { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 387, + 250, + 404 + ], + "score": 1.0, + "content": ". The base case for", + "type": "text" + }, + { + "bbox": [ + 250, + 390, + 275, + 400 + ], + "score": 0.89, + "content": "n = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 387, + 329, + 404 + ], + "score": 1.0, + "content": "follows from", + "type": "text" + }, + { + "bbox": [ + 329, + 390, + 506, + 402 + ], + "score": 0.91, + "content": "T V ( \\tilde { \\pi } ( y ) , \\pi ^ { * } ( y ) ) = T V ( \\pi ^ { * } ( \\tilde { x } _ { y } ) , \\pi ^ { * } ( y ) ) \\leq", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 103, + 399, + 470, + 420 + ], + "spans": [ + { + "bbox": [ + 103, + 399, + 259, + 420 + ], + "score": 1.0, + "content": "d∗(˜xy, y). Note that T V n ≤ 1−γn+11−γ", + "type": "text" + }, + { + "bbox": [ + 255, + 404, + 293, + 417 + ], + "score": 1.0, + "content": "since the", + "type": "text" + }, + { + "bbox": [ + 294, + 405, + 310, + 415 + ], + "score": 0.79, + "content": "T V", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 404, + 470, + 417 + ], + "score": 1.0, + "content": "distance per time-step can be at most 1.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 504, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 504, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 122, + 437 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 424, + 160, + 436 + ], + "score": 0.93, + "content": "P _ { t } ^ { \\pi } ( y ^ { \\prime } | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 422, + 324, + 437 + ], + "score": 1.0, + "content": "denote the probability of ending in state", + "type": "text" + }, + { + "bbox": [ + 324, + 424, + 353, + 435 + ], + "score": 0.92, + "content": "y ^ { \\prime } \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 422, + 375, + 437 + ], + "score": 1.0, + "content": "after", + "type": "text" + }, + { + "bbox": [ + 375, + 425, + 380, + 434 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 422, + 496, + 437 + ], + "score": 1.0, + "content": "steps when following policy", + "type": "text" + }, + { + "bbox": [ + 497, + 426, + 504, + 434 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 434, + 268, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 198, + 446 + ], + "score": 1.0, + "content": "and starting from state", + "type": "text" + }, + { + "bbox": [ + 199, + 436, + 205, + 446 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 434, + 268, + 446 + ], + "score": 1.0, + "content": ". We then have:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 447, + 514, + 725 + ], + "lines": [ + { + "bbox": [ + 112, + 447, + 514, + 725 + ], + "spans": [ + { + "bbox": [ + 112, + 447, + 514, + 725 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\gamma ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\mathcal { K } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\gamma \\varepsilon \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\Bigg [ \\varepsilon \\gamma \\varepsilon ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\varepsilon \\gamma ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\varepsilon ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) \\Bigg ] } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { \\quad + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\gamma \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & \\quad - \\sum _ { s \\in \\mathcal { N } } ^ \\end{array}", + "type": "interline_equation", + "image_path": "9806b9b18cd6be1835fe38e323ac6ae7a8a021231bd7e90489f9f032b6ffd66c.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 112, + 447, + 514, + 539.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 112, + 539.6666666666666, + 514, + 632.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 112, + 632.3333333333333, + 514, + 724.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 284, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 284, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 234, + 97 + ], + "score": 1.0, + "content": "Proposition B.1. The operator", + "type": "text" + }, + { + "bbox": [ + 235, + 83, + 244, + 93 + ], + "score": 0.82, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 80, + 284, + 97 + ], + "score": 1.0, + "content": "given by:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 80, + 284, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 170, + 98, + 441, + 113 + ], + "lines": [ + { + "bbox": [ + 170, + 98, + 441, + 113 + ], + "spans": [ + { + "bbox": [ + 170, + 98, + 441, + 113 + ], + "score": 0.88, + "content": "\\mathcal F ( d ) ( x , y ) = \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) + \\gamma \\mathcal W _ { 1 } ( d ) \\big ( P _ { \\mathcal X } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal V } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big )", + "type": "interline_equation", + "image_path": "50037dbaadaa611b3c2b7307b310ef9effa7d80a0cfe71252bc574a5d4e1c2c3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 170, + 98, + 441, + 113 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 117, + 403, + 129 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 404, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 404, + 131 + ], + "score": 1.0, + "content": "is a contraction mapping and has a unique fixed point for a bounded dist.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 117, + 404, + 131 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 211, + 153 + ], + "score": 1.0, + "content": "Proof. We first prove that", + "type": "text" + }, + { + "bbox": [ + 211, + 142, + 220, + 151 + ], + "score": 0.86, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "is contraction mapping. Then, a simple application of the Banach Fixed", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 152, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 218, + 164 + ], + "score": 1.0, + "content": "Point Theorem asserts that", + "type": "text" + }, + { + "bbox": [ + 218, + 153, + 227, + 162 + ], + "score": 0.86, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 152, + 461, + 164 + ], + "score": 1.0, + "content": "has a unique fixed point. Note that for all pseudometrics", + "type": "text" + }, + { + "bbox": [ + 461, + 152, + 502, + 163 + ], + "score": 0.92, + "content": "d , d ^ { \\prime } \\in \\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 152, + 506, + 164 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 163, + 236, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 174, + 174 + ], + "score": 1.0, + "content": "and for all states", + "type": "text" + }, + { + "bbox": [ + 175, + 163, + 202, + 173 + ], + "score": 0.84, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 163, + 206, + 174 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 206, + 163, + 232, + 174 + ], + "score": 0.82, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 163, + 236, + 174 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 141, + 506, + 174 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 177, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 128, + 177, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 128, + 177, + 482, + 255 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { F } ( d ) ( x , y ) - \\mathcal { F } ( d ^ { \\prime } ) ( x , y ) } \\\\ & { \\quad = \\gamma \\left( \\mathcal { W } _ { 1 } ( d ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad \\overset { L e m m a ~ \\ : 1 } { \\leq } \\gamma \\left( \\| d - d ^ { \\prime } \\| + \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad = \\gamma \\| d - d ^ { \\prime } \\| } \\end{array}", + "type": "interline_equation", + "image_path": "e26031ddb7f6793c820c994471ce5075a5e1d92786e164ddfeb0013f881bdb40.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 128, + 177, + 482, + 203.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 128, + 203.0, + 482, + 229.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 128, + 229.0, + 482, + 255.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 271, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 131, + 285 + ], + "score": 1.0, + "content": "Thus,", + "type": "text" + }, + { + "bbox": [ + 131, + 272, + 252, + 284 + ], + "score": 0.92, + "content": "\\| \\mathcal F ( d ) - \\mathcal F ( d ^ { \\prime } ) \\| \\le \\gamma \\| d - d ^ { \\prime } \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 270, + 283, + 285 + ], + "score": 1.0, + "content": ", so that", + "type": "text" + }, + { + "bbox": [ + 284, + 272, + 293, + 282 + ], + "score": 0.85, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 270, + 406, + 285 + ], + "score": 1.0, + "content": "is a contracting mapping for", + "type": "text" + }, + { + "bbox": [ + 406, + 273, + 431, + 284 + ], + "score": 0.91, + "content": "\\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 270, + 506, + 285 + ], + "score": 1.0, + "content": "and has an unique", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 281, + 167, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 152, + 295 + ], + "score": 1.0, + "content": "fixed point", + "type": "text" + }, + { + "bbox": [ + 152, + 284, + 162, + 293 + ], + "score": 0.85, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 281, + 167, + 295 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 270, + 506, + 295 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 502, + 326 + ], + "lines": [ + { + "bbox": [ + 103, + 296, + 507, + 319 + ], + "spans": [ + { + "bbox": [ + 103, + 296, + 303, + 319 + ], + "score": 1.0, + "content": "Theorem 1. [Bound on policy transfer] For any", + "type": "text" + }, + { + "bbox": [ + 303, + 301, + 329, + 312 + ], + "score": 0.9, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 296, + 345, + 319 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 346, + 300, + 423, + 314 + ], + "score": 0.93, + "content": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 296, + 507, + 319 + ], + "score": 1.0, + "content": "define the sequence", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 311, + 439, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 270, + 329 + ], + "score": 1.0, + "content": "of random states encountered starting in", + "type": "text" + }, + { + "bbox": [ + 271, + 313, + 303, + 327 + ], + "score": 0.92, + "content": "Y _ { y } ^ { 0 } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 311, + 388, + 329 + ], + "score": 1.0, + "content": "and following policy", + "type": "text" + }, + { + "bbox": [ + 388, + 315, + 395, + 324 + ], + "score": 0.64, + "content": "\\tilde { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 311, + 439, + 329 + ], + "score": 1.0, + "content": ". We have:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 103, + 296, + 507, + 329 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 329, + 419, + 369 + ], + "lines": [ + { + "bbox": [ + 191, + 329, + 419, + 369 + ], + "spans": [ + { + "bbox": [ + 191, + 329, + 419, + 369 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) .", + "type": "interline_equation", + "image_path": "78300f5f5ad236d336ba4cea178e615dab71051f4cf522680b2ec7241d02a039.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 329, + 419, + 349.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 191, + 349.0, + 419, + 369.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 378, + 506, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 400, + 390 + ], + "score": 1.0, + "content": "Proof. We will prove this by induction. Assuming that the bound holds for", + "type": "text" + }, + { + "bbox": [ + 401, + 380, + 422, + 389 + ], + "score": 0.87, + "content": "T V ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 378, + 505, + 390 + ], + "score": 1.0, + "content": ", we prove the bound", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 387, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 104, + 387, + 144, + 404 + ], + "score": 1.0, + "content": "holds for", + "type": "text" + }, + { + "bbox": [ + 144, + 390, + 175, + 400 + ], + "score": 0.9, + "content": "T V ^ { n + \\bar { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 387, + 250, + 404 + ], + "score": 1.0, + "content": ". The base case for", + "type": "text" + }, + { + "bbox": [ + 250, + 390, + 275, + 400 + ], + "score": 0.89, + "content": "n = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 387, + 329, + 404 + ], + "score": 1.0, + "content": "follows from", + "type": "text" + }, + { + "bbox": [ + 329, + 390, + 506, + 402 + ], + "score": 0.91, + "content": "T V ( \\tilde { \\pi } ( y ) , \\pi ^ { * } ( y ) ) = T V ( \\pi ^ { * } ( \\tilde { x } _ { y } ) , \\pi ^ { * } ( y ) ) \\leq", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 103, + 399, + 470, + 420 + ], + "spans": [ + { + "bbox": [ + 103, + 399, + 259, + 420 + ], + "score": 1.0, + "content": "d∗(˜xy, y). Note that T V n ≤ 1−γn+11−γ", + "type": "text" + }, + { + "bbox": [ + 255, + 404, + 293, + 417 + ], + "score": 1.0, + "content": "since the", + "type": "text" + }, + { + "bbox": [ + 294, + 405, + 310, + 415 + ], + "score": 0.79, + "content": "T V", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 404, + 470, + 417 + ], + "score": 1.0, + "content": "distance per time-step can be at most 1.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 103, + 378, + 506, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 504, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 504, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 122, + 437 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 424, + 160, + 436 + ], + "score": 0.93, + "content": "P _ { t } ^ { \\pi } ( y ^ { \\prime } | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 422, + 324, + 437 + ], + "score": 1.0, + "content": "denote the probability of ending in state", + "type": "text" + }, + { + "bbox": [ + 324, + 424, + 353, + 435 + ], + "score": 0.92, + "content": "y ^ { \\prime } \\in Y", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 422, + 375, + 437 + ], + "score": 1.0, + "content": "after", + "type": "text" + }, + { + "bbox": [ + 375, + 425, + 380, + 434 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 422, + 496, + 437 + ], + "score": 1.0, + "content": "steps when following policy", + "type": "text" + }, + { + "bbox": [ + 497, + 426, + 504, + 434 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 434, + 268, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 198, + 446 + ], + "score": 1.0, + "content": "and starting from state", + "type": "text" + }, + { + "bbox": [ + 199, + 436, + 205, + 446 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 434, + 268, + 446 + ], + "score": 1.0, + "content": ". We then have:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 422, + 504, + 446 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 447, + 514, + 725 + ], + "lines": [ + { + "bbox": [ + 112, + 447, + 514, + 725 + ], + "spans": [ + { + "bbox": [ + 112, + 447, + 514, + 725 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\gamma ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\mathcal { K } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\gamma \\varepsilon \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\Bigg [ \\varepsilon \\gamma \\varepsilon ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\varepsilon \\gamma ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\varepsilon ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) \\Bigg ] } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { \\quad + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\gamma \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & \\quad - \\sum _ { s \\in \\mathcal { N } } ^ \\end{array}", + "type": "interline_equation", + "image_path": "9806b9b18cd6be1835fe38e323ac6ae7a8a021231bd7e90489f9f032b6ffd66c.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 112, + 447, + 514, + 539.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 112, + 539.6666666666666, + 514, + 632.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 112, + 632.3333333333333, + 514, + 724.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 281, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 281, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 270, + 95 + ], + "score": 1.0, + "content": "Thus, by induction, it follows that for all", + "type": "text" + }, + { + "bbox": [ + 271, + 85, + 277, + 92 + ], + "score": 0.73, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 82, + 281, + 95 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 96, + 365, + 122 + ], + "lines": [ + { + "bbox": [ + 244, + 96, + 365, + 122 + ], + "spans": [ + { + "bbox": [ + 244, + 96, + 365, + 122 + ], + "score": 0.94, + "content": "T V ^ { n } ( Y _ { y } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) ,", + "type": "interline_equation", + "image_path": "f0e3466874ac9aa6d3f381f52d738f7138816ba8defceed58d69a2f3532d74d6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 244, + 96, + 365, + 122 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 216, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 217, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 217, + 137 + ], + "score": 1.0, + "content": "which completes the proof.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 151, + 254, + 164 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 255, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 255, + 166 + ], + "score": 1.0, + "content": "C BISIMULATION METRICS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 175, + 324, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 325, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 325, + 188 + ], + "score": 1.0, + "content": "Notation. We use the notation as defined in Section 2.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 506, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 434, + 204 + ], + "score": 1.0, + "content": "Bisimulation metrics (Givan et al., 2003; Ferns et al., 2011) define a pseudometric", + "type": "text" + }, + { + "bbox": [ + 434, + 193, + 504, + 203 + ], + "score": 0.89, + "content": "d _ { \\sim } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 203, + 507, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 132, + 216 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 204, + 168, + 215 + ], + "score": 0.92, + "content": "d _ { \\sim } ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 203, + 507, + 216 + ], + "score": 1.0, + "content": "is defined in terms of distance between immediate rewards and next state distributions.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 213, + 207, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 136, + 227 + ], + "score": 1.0, + "content": "Define", + "type": "text" + }, + { + "bbox": [ + 136, + 215, + 194, + 226 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { \\sim } ^ { e } : \\mathbb { M } \\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 213, + 207, + 227 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 228, + 452, + 247 + ], + "lines": [ + { + "bbox": [ + 158, + 228, + 452, + 247 + ], + "spans": [ + { + "bbox": [ + 158, + 228, + 452, + 247 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { \\sim } ^ { e } ( d ) ( x , y ) = \\operatorname* { m a x } _ { a \\in A } \\vert R ( x , a ) - R ( y , a ) \\vert + \\gamma \\mathcal { W } _ { 1 } ( d ) \\left( P ^ { a } ( \\cdot \\vert x ) , P ^ { a } ( \\cdot \\vert y ) \\right)", + "type": "interline_equation", + "image_path": "4961b6782cf1de03ed5b5de85a3e92827a028dd378284d7a9a4535a4ce354b1b.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 158, + 228, + 452, + 247 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 506, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 129, + 262 + ], + "score": 1.0, + "content": "then,", + "type": "text" + }, + { + "bbox": [ + 129, + 251, + 144, + 262 + ], + "score": 0.88, + "content": "\\mathcal { F } _ { \\sim } ^ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 250, + 247, + 262 + ], + "score": 1.0, + "content": "has a unique fixed point", + "type": "text" + }, + { + "bbox": [ + 248, + 250, + 261, + 261 + ], + "score": 0.87, + "content": "d _ { \\sim }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 250, + 393, + 262 + ], + "score": 1.0, + "content": "which is a bisimulation metric.", + "type": "text" + }, + { + "bbox": [ + 393, + 251, + 408, + 262 + ], + "score": 0.89, + "content": "\\mathcal { F } _ { \\sim } ^ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "uses the 1-Wasserstein", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 260, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 133, + 273 + ], + "score": 1.0, + "content": "metric", + "type": "text" + }, + { + "bbox": [ + 134, + 261, + 196, + 273 + ], + "score": 0.92, + "content": "\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 260, + 308, + 273 + ], + "score": 1.0, + "content": ". The 1-Wasserstein distance", + "type": "text" + }, + { + "bbox": [ + 308, + 261, + 336, + 273 + ], + "score": 0.93, + "content": "\\mathscr { W } _ { 1 } ( d )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 260, + 430, + 273 + ], + "score": 1.0, + "content": "under the pseudometric", + "type": "text" + }, + { + "bbox": [ + 430, + 262, + 437, + 271 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 260, + 506, + 273 + ], + "score": 1.0, + "content": "can be computed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 271, + 229, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 229, + 285 + ], + "score": 1.0, + "content": "using the dual linear program:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 286, + 472, + 314 + ], + "lines": [ + { + "bbox": [ + 138, + 286, + 472, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 472, + 314 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\mathbf { u } , \\mathbf { v } } \\sum _ { x \\in \\mathcal { X } } P ( x ) u _ { x } - \\sum _ { y \\in \\mathcal { Y } } P ( y ) v _ { y } \\quad \\mathrm { s u b j e c t ~ t o } \\forall x \\in \\mathcal { X } , y \\in \\mathcal { Y } \\quad u _ { x } - v _ { y } \\leq d ( x , y )", + "type": "interline_equation", + "image_path": "f8613eabbb2d2596c7a8e041d947c55bd72e986330d6682b84bffc223ed82188.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 138, + 286, + 472, + 295.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 138, + 295.3333333333333, + 472, + 304.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 138, + 304.66666666666663, + 472, + 313.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 422, + 329 + ], + "score": 1.0, + "content": "Since we are only interested in computing the coupling between the states in", + "type": "text" + }, + { + "bbox": [ + 422, + 317, + 432, + 326 + ], + "score": 0.83, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 317, + 450, + 329 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 451, + 317, + 459, + 327 + ], + "score": 0.82, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 317, + 505, + 329 + ], + "score": 1.0, + "content": ", the above", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 327, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 210, + 341 + ], + "score": 1.0, + "content": "formulation assumes that", + "type": "text" + }, + { + "bbox": [ + 210, + 327, + 256, + 340 + ], + "score": 0.93, + "content": "P _ { \\mathcal { X } } ( y ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 327, + 284, + 341 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 284, + 328, + 311, + 339 + ], + "score": 0.91, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 327, + 328, + 341 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 329, + 327, + 374, + 340 + ], + "score": 0.94, + "content": "P _ { \\mathcal { Y } } ( x ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 327, + 402, + 341 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 402, + 328, + 430, + 338 + ], + "score": 0.9, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 327, + 505, + 341 + ], + "score": 1.0, + "content": ". The computation", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 337, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 104, + 337, + 117, + 353 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 339, + 130, + 349 + ], + "score": 0.88, + "content": "d _ { \\sim }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 337, + 506, + 353 + ], + "score": 1.0, + "content": "is expensive and requires a tabular representation of the states, rendering it impractical for large", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 331, + 362 + ], + "score": 1.0, + "content": "state spaces. On-policy bisimulation (Castro, 2020) (e.g.,", + "type": "text" + }, + { + "bbox": [ + 332, + 350, + 343, + 360 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 349, + 506, + 362 + ], + "score": 1.0, + "content": "-bisimulation) is tied to specific behavior", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 361, + 353, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 353, + 373 + ], + "score": 1.0, + "content": "policies and is much easier to approximate than bisimulation.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 387, + 279, + 400 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 280, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 280, + 402 + ], + "score": 1.0, + "content": "D POLICY SIMILARITY METRIC", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 411, + 212, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 213, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 213, + 425 + ], + "score": 1.0, + "content": "D.1 COMPUTING PSM", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 403, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 403, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 298, + 447 + ], + "score": 1.0, + "content": "In general, PSM for a given DIST across MDPs", + "type": "text" + }, + { + "bbox": [ + 299, + 433, + 318, + 443 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 430, + 337, + 447 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 337, + 433, + 356, + 444 + ], + "score": 0.91, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 430, + 403, + 447 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 447, + 445, + 463 + ], + "lines": [ + { + "bbox": [ + 165, + 447, + 445, + 463 + ], + "spans": [ + { + "bbox": [ + 165, + 447, + 445, + 463 + ], + "score": 0.89, + "content": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}", + "type": "interline_equation", + "image_path": "397980339d5fd9d1b862a2716debbd7eb5a154289e77e3908ae1376c757b1ffb.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 165, + 447, + 445, + 463 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "score": 1.0, + "content": "Since our main focus is showing the utility of PSM for generalization, we simply use environments", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "where PSM can be computed using dynamic programming. Using a similar observation to Castro", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 493, + 500, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 263, + 505 + ], + "score": 1.0, + "content": "(2020), we assert that the recursion for", + "type": "text" + }, + { + "bbox": [ + 263, + 493, + 273, + 503 + ], + "score": 0.87, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 493, + 500, + 505 + ], + "score": 1.0, + "content": "takes the following form in deterministic environments:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 506, + 403, + 522 + ], + "lines": [ + { + "bbox": [ + 208, + 506, + 403, + 522 + ], + "spans": [ + { + "bbox": [ + 208, + 506, + 403, + 522 + ], + "score": 0.92, + "content": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma d ^ { * } ( x ^ { \\prime } , y ^ { \\prime } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "191926b32509785686fc5757bafe5efd5ea9a5819cee8360dc34c4ac99d9d785.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 208, + 506, + 403, + 522 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 506, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 133, + 540 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 524, + 187, + 538 + ], + "score": 0.83, + "content": "x ^ { \\prime } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 524, + 190, + 540 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 191, + 524, + 243, + 539 + ], + "score": 0.87, + "content": "y ^ { \\prime } = P _ { y } ^ { \\pi ^ { * } } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 524, + 398, + 540 + ], + "score": 1.0, + "content": "are the next states from taking actions", + "type": "text" + }, + { + "bbox": [ + 399, + 526, + 457, + 538 + ], + "score": 0.29, + "content": "\\pi _ { \\mathcal { X } } ^ { \\ast } ( x ) , \\pi _ { \\mathcal { X } } ^ { \\ast } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 524, + 506, + 540 + ], + "score": 1.0, + "content": "from states", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 536, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 126, + 549 + ], + "score": 0.66, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 536, + 437, + 550 + ], + "score": 1.0, + "content": "respectively. Furthermore, we assume that DIST between terminal states from", + "type": "text" + }, + { + "bbox": [ + 437, + 538, + 457, + 548 + ], + "score": 0.89, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 536, + 475, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 537, + 495, + 549 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 536, + 506, + 550 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "zero. Note that the form of Equation D.2 closely resembles the update rule in Q-learning, and as such,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "can be efficiently computed with samples using approximate dynamic programming. Given access to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 567, + 508, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 185, + 587 + ], + "score": 1.0, + "content": "optimal trajectories", + "type": "text" + }, + { + "bbox": [ + 186, + 570, + 244, + 583 + ], + "score": 0.94, + "content": "\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 567, + 262, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 570, + 319, + 584 + ], + "score": 0.94, + "content": "\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 567, + 349, + 587 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 350, + 570, + 418, + 583 + ], + "score": 0.93, + "content": "x _ { t + 1 } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 567, + 436, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 436, + 569, + 502, + 584 + ], + "score": 0.93, + "content": "y _ { t + 1 } = P _ { y } ^ { \\pi ^ { * } } ( y _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 567, + 508, + 587 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 581, + 507, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 507, + 595 + ], + "score": 1.0, + "content": "Equation D.2 can be solved using exact dynamic programming; we provide pseudocode in Section J.1.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 598, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "There are other ways to approximate the Wasserstein distance in bisimulation metrics (e.g., Ferns", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "et al., 2006; 2011; Castro, 2020; Zhang et al., 2021). That said, approximating bisimulation (or PSM)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "score": 1.0, + "content": "for stochastic environments remains an exciting research direction (Castro, 2020). Investigating other", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 632, + 479, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 479, + 644 + ], + "score": 1.0, + "content": "distance metrics for long-term behavior difference in PSM is also interesting for future work.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 107, + 655, + 421, + 668 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 423, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 423, + 669 + ], + "score": 1.0, + "content": "D.2 PSM CONNECTIONS TO DATA AUGMENTATION AND BISIMULATION", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Connection to bisimulation. Although bisimulation metrics have appealing properties such as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "bounding value function differences (e.g., (Ferns & Precup, 2014)), they rely on reward information", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "and may not provide a meaningful notion of behavioral similarity in certain environments. Propo-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "sition D.1 implies that states similar under PSM would have similar optimal policies yet can have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 721, + 318, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 318, + 732 + ], + "score": 1.0, + "content": "arbitrarily large bisimulation distance between them.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 124, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 495, + 126, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 495, + 126, + 505, + 136 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 281, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 281, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 270, + 95 + ], + "score": 1.0, + "content": "Thus, by induction, it follows that for all", + "type": "text" + }, + { + "bbox": [ + 271, + 85, + 277, + 92 + ], + "score": 0.73, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 82, + 281, + 95 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 82, + 281, + 95 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 96, + 365, + 122 + ], + "lines": [ + { + "bbox": [ + 244, + 96, + 365, + 122 + ], + "spans": [ + { + "bbox": [ + 244, + 96, + 365, + 122 + ], + "score": 0.94, + "content": "T V ^ { n } ( Y _ { y } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) ,", + "type": "interline_equation", + "image_path": "f0e3466874ac9aa6d3f381f52d738f7138816ba8defceed58d69a2f3532d74d6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 244, + 96, + 365, + 122 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 216, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 217, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 217, + 137 + ], + "score": 1.0, + "content": "which completes the proof.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 123, + 217, + 137 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 151, + 254, + 164 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 255, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 255, + 166 + ], + "score": 1.0, + "content": "C BISIMULATION METRICS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 175, + 324, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 325, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 325, + 188 + ], + "score": 1.0, + "content": "Notation. We use the notation as defined in Section 2.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 174, + 325, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 506, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 434, + 204 + ], + "score": 1.0, + "content": "Bisimulation metrics (Givan et al., 2003; Ferns et al., 2011) define a pseudometric", + "type": "text" + }, + { + "bbox": [ + 434, + 193, + 504, + 203 + ], + "score": 0.89, + "content": "d _ { \\sim } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 203, + 507, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 132, + 216 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 204, + 168, + 215 + ], + "score": 0.92, + "content": "d _ { \\sim } ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 203, + 507, + 216 + ], + "score": 1.0, + "content": "is defined in terms of distance between immediate rewards and next state distributions.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 213, + 207, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 136, + 227 + ], + "score": 1.0, + "content": "Define", + "type": "text" + }, + { + "bbox": [ + 136, + 215, + 194, + 226 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { \\sim } ^ { e } : \\mathbb { M } \\mathbb { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 213, + 207, + 227 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 192, + 507, + 227 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 228, + 452, + 247 + ], + "lines": [ + { + "bbox": [ + 158, + 228, + 452, + 247 + ], + "spans": [ + { + "bbox": [ + 158, + 228, + 452, + 247 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { \\sim } ^ { e } ( d ) ( x , y ) = \\operatorname* { m a x } _ { a \\in A } \\vert R ( x , a ) - R ( y , a ) \\vert + \\gamma \\mathcal { W } _ { 1 } ( d ) \\left( P ^ { a } ( \\cdot \\vert x ) , P ^ { a } ( \\cdot \\vert y ) \\right)", + "type": "interline_equation", + "image_path": "4961b6782cf1de03ed5b5de85a3e92827a028dd378284d7a9a4535a4ce354b1b.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 158, + 228, + 452, + 247 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 506, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 129, + 262 + ], + "score": 1.0, + "content": "then,", + "type": "text" + }, + { + "bbox": [ + 129, + 251, + 144, + 262 + ], + "score": 0.88, + "content": "\\mathcal { F } _ { \\sim } ^ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 250, + 247, + 262 + ], + "score": 1.0, + "content": "has a unique fixed point", + "type": "text" + }, + { + "bbox": [ + 248, + 250, + 261, + 261 + ], + "score": 0.87, + "content": "d _ { \\sim }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 250, + 393, + 262 + ], + "score": 1.0, + "content": "which is a bisimulation metric.", + "type": "text" + }, + { + "bbox": [ + 393, + 251, + 408, + 262 + ], + "score": 0.89, + "content": "\\mathcal { F } _ { \\sim } ^ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "uses the 1-Wasserstein", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 260, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 133, + 273 + ], + "score": 1.0, + "content": "metric", + "type": "text" + }, + { + "bbox": [ + 134, + 261, + 196, + 273 + ], + "score": 0.92, + "content": "\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 260, + 308, + 273 + ], + "score": 1.0, + "content": ". The 1-Wasserstein distance", + "type": "text" + }, + { + "bbox": [ + 308, + 261, + 336, + 273 + ], + "score": 0.93, + "content": "\\mathscr { W } _ { 1 } ( d )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 260, + 430, + 273 + ], + "score": 1.0, + "content": "under the pseudometric", + "type": "text" + }, + { + "bbox": [ + 430, + 262, + 437, + 271 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 260, + 506, + 273 + ], + "score": 1.0, + "content": "can be computed", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 271, + 229, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 229, + 285 + ], + "score": 1.0, + "content": "using the dual linear program:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 250, + 506, + 285 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 286, + 472, + 314 + ], + "lines": [ + { + "bbox": [ + 138, + 286, + 472, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 472, + 314 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\mathbf { u } , \\mathbf { v } } \\sum _ { x \\in \\mathcal { X } } P ( x ) u _ { x } - \\sum _ { y \\in \\mathcal { Y } } P ( y ) v _ { y } \\quad \\mathrm { s u b j e c t ~ t o } \\forall x \\in \\mathcal { X } , y \\in \\mathcal { Y } \\quad u _ { x } - v _ { y } \\leq d ( x , y )", + "type": "interline_equation", + "image_path": "f8613eabbb2d2596c7a8e041d947c55bd72e986330d6682b84bffc223ed82188.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 138, + 286, + 472, + 295.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 138, + 295.3333333333333, + 472, + 304.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 138, + 304.66666666666663, + 472, + 313.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 422, + 329 + ], + "score": 1.0, + "content": "Since we are only interested in computing the coupling between the states in", + "type": "text" + }, + { + "bbox": [ + 422, + 317, + 432, + 326 + ], + "score": 0.83, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 317, + 450, + 329 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 451, + 317, + 459, + 327 + ], + "score": 0.82, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 317, + 505, + 329 + ], + "score": 1.0, + "content": ", the above", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 327, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 210, + 341 + ], + "score": 1.0, + "content": "formulation assumes that", + "type": "text" + }, + { + "bbox": [ + 210, + 327, + 256, + 340 + ], + "score": 0.93, + "content": "P _ { \\mathcal { X } } ( y ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 327, + 284, + 341 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 284, + 328, + 311, + 339 + ], + "score": 0.91, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 327, + 328, + 341 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 329, + 327, + 374, + 340 + ], + "score": 0.94, + "content": "P _ { \\mathcal { Y } } ( x ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 327, + 402, + 341 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 402, + 328, + 430, + 338 + ], + "score": 0.9, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 327, + 505, + 341 + ], + "score": 1.0, + "content": ". The computation", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 337, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 104, + 337, + 117, + 353 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 339, + 130, + 349 + ], + "score": 0.88, + "content": "d _ { \\sim }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 337, + 506, + 353 + ], + "score": 1.0, + "content": "is expensive and requires a tabular representation of the states, rendering it impractical for large", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 349, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 331, + 362 + ], + "score": 1.0, + "content": "state spaces. On-policy bisimulation (Castro, 2020) (e.g.,", + "type": "text" + }, + { + "bbox": [ + 332, + 350, + 343, + 360 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 349, + 506, + 362 + ], + "score": 1.0, + "content": "-bisimulation) is tied to specific behavior", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 361, + 353, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 353, + 373 + ], + "score": 1.0, + "content": "policies and is much easier to approximate than bisimulation.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 317, + 506, + 373 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 387, + 279, + 400 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 280, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 280, + 402 + ], + "score": 1.0, + "content": "D POLICY SIMILARITY METRIC", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 411, + 212, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 213, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 213, + 425 + ], + "score": 1.0, + "content": "D.1 COMPUTING PSM", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 403, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 403, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 298, + 447 + ], + "score": 1.0, + "content": "In general, PSM for a given DIST across MDPs", + "type": "text" + }, + { + "bbox": [ + 299, + 433, + 318, + 443 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 430, + 337, + 447 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 337, + 433, + 356, + 444 + ], + "score": 0.91, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 430, + 403, + 447 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 430, + 403, + 447 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 447, + 445, + 463 + ], + "lines": [ + { + "bbox": [ + 165, + 447, + 445, + 463 + ], + "spans": [ + { + "bbox": [ + 165, + 447, + 445, + 463 + ], + "score": 0.89, + "content": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}", + "type": "interline_equation", + "image_path": "397980339d5fd9d1b862a2716debbd7eb5a154289e77e3908ae1376c757b1ffb.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 165, + 447, + 445, + 463 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "score": 1.0, + "content": "Since our main focus is showing the utility of PSM for generalization, we simply use environments", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "where PSM can be computed using dynamic programming. Using a similar observation to Castro", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 493, + 500, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 263, + 505 + ], + "score": 1.0, + "content": "(2020), we assert that the recursion for", + "type": "text" + }, + { + "bbox": [ + 263, + 493, + 273, + 503 + ], + "score": 0.87, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 493, + 500, + 505 + ], + "score": 1.0, + "content": "takes the following form in deterministic environments:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 469, + 506, + 505 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 506, + 403, + 522 + ], + "lines": [ + { + "bbox": [ + 208, + 506, + 403, + 522 + ], + "spans": [ + { + "bbox": [ + 208, + 506, + 403, + 522 + ], + "score": 0.92, + "content": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma d ^ { * } ( x ^ { \\prime } , y ^ { \\prime } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "191926b32509785686fc5757bafe5efd5ea9a5819cee8360dc34c4ac99d9d785.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 208, + 506, + 403, + 522 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 506, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 133, + 540 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 524, + 187, + 538 + ], + "score": 0.83, + "content": "x ^ { \\prime } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 524, + 190, + 540 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 191, + 524, + 243, + 539 + ], + "score": 0.87, + "content": "y ^ { \\prime } = P _ { y } ^ { \\pi ^ { * } } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 524, + 398, + 540 + ], + "score": 1.0, + "content": "are the next states from taking actions", + "type": "text" + }, + { + "bbox": [ + 399, + 526, + 457, + 538 + ], + "score": 0.29, + "content": "\\pi _ { \\mathcal { X } } ^ { \\ast } ( x ) , \\pi _ { \\mathcal { X } } ^ { \\ast } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 524, + 506, + 540 + ], + "score": 1.0, + "content": "from states", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 536, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 126, + 549 + ], + "score": 0.66, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 536, + 437, + 550 + ], + "score": 1.0, + "content": "respectively. Furthermore, we assume that DIST between terminal states from", + "type": "text" + }, + { + "bbox": [ + 437, + 538, + 457, + 548 + ], + "score": 0.89, + "content": "\\mathcal { M } _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 536, + 475, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 537, + 495, + 549 + ], + "score": 0.9, + "content": "\\mathcal { M } _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 536, + 506, + 550 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "zero. Note that the form of Equation D.2 closely resembles the update rule in Q-learning, and as such,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "can be efficiently computed with samples using approximate dynamic programming. Given access to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 567, + 508, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 185, + 587 + ], + "score": 1.0, + "content": "optimal trajectories", + "type": "text" + }, + { + "bbox": [ + 186, + 570, + 244, + 583 + ], + "score": 0.94, + "content": "\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 567, + 262, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 570, + 319, + 584 + ], + "score": 0.94, + "content": "\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 567, + 349, + 587 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 350, + 570, + 418, + 583 + ], + "score": 0.93, + "content": "x _ { t + 1 } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 567, + 436, + 587 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 436, + 569, + 502, + 584 + ], + "score": 0.93, + "content": "y _ { t + 1 } = P _ { y } ^ { \\pi ^ { * } } ( y _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 567, + 508, + 587 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 581, + 507, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 507, + 595 + ], + "score": 1.0, + "content": "Equation D.2 can be solved using exact dynamic programming; we provide pseudocode in Section J.1.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 524, + 508, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 598, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "There are other ways to approximate the Wasserstein distance in bisimulation metrics (e.g., Ferns", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "et al., 2006; 2011; Castro, 2020; Zhang et al., 2021). That said, approximating bisimulation (or PSM)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "score": 1.0, + "content": "for stochastic environments remains an exciting research direction (Castro, 2020). Investigating other", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 632, + 479, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 479, + 644 + ], + "score": 1.0, + "content": "distance metrics for long-term behavior difference in PSM is also interesting for future work.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 598, + 506, + 644 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 655, + 421, + 668 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 423, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 423, + 669 + ], + "score": 1.0, + "content": "D.2 PSM CONNECTIONS TO DATA AUGMENTATION AND BISIMULATION", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Connection to bisimulation. Although bisimulation metrics have appealing properties such as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "bounding value function differences (e.g., (Ferns & Precup, 2014)), they rely on reward information", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "and may not provide a meaningful notion of behavioral similarity in certain environments. Propo-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "sition D.1 implies that states similar under PSM would have similar optimal policies yet can have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 721, + 318, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 318, + 732 + ], + "score": 1.0, + "content": "arbitrarily large bisimulation distance between them.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 677, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 78, + 466, + 182 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 140, + 78, + 466, + 182 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 78, + 466, + 182 + ], + "spans": [ + { + "bbox": [ + 140, + 78, + 466, + 182 + ], + "score": 0.968, + "type": "image", + "image_path": "3d8fd583f31c9cbc260975c387fd75a3ea68a6f1e11f6b2895873d41fe0189c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 140, + 78, + 466, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 140, + 112.66666666666666, + 466, + 147.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 140, + 147.33333333333331, + 466, + 181.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 191, + 504, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "Figure D.1: Cyan edges represent actions with a positive reward, which are also the optimal actions. Zero", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 202, + 420, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 199, + 213 + ], + "score": 1.0, + "content": "rewards everywhere else.", + "type": "text" + }, + { + "bbox": [ + 199, + 204, + 223, + 213 + ], + "score": 0.88, + "content": "x _ { 0 } , y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 202, + 312, + 213 + ], + "score": 1.0, + "content": "are the start states while", + "type": "text" + }, + { + "bbox": [ + 312, + 204, + 335, + 213 + ], + "score": 0.88, + "content": "x _ { 2 } , y _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 202, + 420, + 213 + ], + "score": 1.0, + "content": "are the terminal states.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 503, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 294, + 229 + ], + "score": 1.0, + "content": "Proposition D.1. There exists environments", + "type": "text" + }, + { + "bbox": [ + 294, + 217, + 312, + 227 + ], + "score": 0.88, + "content": "M _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 215, + 334, + 229 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 334, + 217, + 352, + 228 + ], + "score": 0.88, + "content": "M _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 215, + 396, + 229 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 397, + 216, + 451, + 228 + ], + "score": 0.91, + "content": "\\forall ( x , y ) \\in ~ { \\mathcal { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 215, + 482, + 229 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 482, + 216, + 505, + 228 + ], + "score": 0.85, + "content": "{ \\mathcal { L } } =", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 223, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 107, + 229, + 168, + 242 + ], + "score": 0.29, + "content": "\\{ ( x , y ) ~ | x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 223, + 172, + 247 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 173, + 230, + 200, + 241 + ], + "score": 0.54, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 223, + 204, + 247 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 205, + 229, + 262, + 242 + ], + "score": 0.72, + "content": "d ^ { * } ( x , y ) = 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 223, + 266, + 247 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 266, + 228, + 382, + 244 + ], + "score": 0.9, + "content": "\\begin{array} { r } { d _ { \\sim } ^ { \\ast } ( x , y ) = \\frac { \\left| R _ { \\mathrm { m a x } } - R _ { \\mathrm { m i n } } \\right| } { 1 - \\gamma } - \\epsilon } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 223, + 440, + 247 + ], + "score": 1.0, + "content": "for any given", + "type": "text" + }, + { + "bbox": [ + 441, + 230, + 464, + 240 + ], + "score": 0.87, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 223, + 470, + 247 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 104, + 249, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 249, + 461, + 265 + ], + "score": 1.0, + "content": "For example, consider the two semantically equivalent environments in Figure D.1 with", + "type": "text" + }, + { + "bbox": [ + 461, + 251, + 506, + 264 + ], + "score": 0.91, + "content": "\\pi _ { \\mathcal { X } } ^ { * } ( x _ { 0 } ) =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 163, + 275 + ], + "score": 0.89, + "content": "\\pi _ { y } ^ { * } ( y _ { 0 } ) = a _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 261, + 183, + 276 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 183, + 262, + 287, + 275 + ], + "score": 0.92, + "content": "\\pi _ { \\mathcal { X } } ^ { \\ast } ( x _ { 1 } ) = \\pi _ { \\mathcal { Y } } ^ { \\ast } ( y _ { 1 } ) = a _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 261, + 378, + 276 + ], + "score": 1.0, + "content": "but different rewards", + "type": "text" + }, + { + "bbox": [ + 378, + 264, + 403, + 275 + ], + "score": 0.89, + "content": "r _ { x } , r _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "respectively. Whenever", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 272, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 183, + 285 + ], + "score": 0.86, + "content": "r _ { y } > ( 1 + 1 / \\gamma ) r _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 272, + 357, + 288 + ], + "score": 1.0, + "content": ", bisimulation metrics incorrectly imply that", + "type": "text" + }, + { + "bbox": [ + 358, + 275, + 369, + 284 + ], + "score": 0.86, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 272, + 494, + 288 + ], + "score": 1.0, + "content": "is more behaviorally similar to", + "type": "text" + }, + { + "bbox": [ + 494, + 275, + 504, + 285 + ], + "score": 0.84, + "content": "y _ { 1 }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 283, + 142, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 126, + 298 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 126, + 287, + 136, + 296 + ], + "score": 0.86, + "content": "y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 283, + 142, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 105, + 300, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 331, + 314 + ], + "score": 1.0, + "content": "For the MDPs shown in Figure D.1, to determine which", + "type": "text" + }, + { + "bbox": [ + 332, + 303, + 339, + 312 + ], + "score": 0.82, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 299, + 475, + 314 + ], + "score": 1.0, + "content": "state is behaviorally equivalent to", + "type": "text" + }, + { + "bbox": [ + 476, + 303, + 487, + 312 + ], + "score": 0.86, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 299, + 505, + 314 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 312, + 467, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 324, + 325 + ], + "score": 1.0, + "content": "look at the distances computed by bisimulation metric", + "type": "text" + }, + { + "bbox": [ + 325, + 313, + 338, + 323 + ], + "score": 0.88, + "content": "d _ { \\sim }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 312, + 356, + 325 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 313, + 367, + 322 + ], + "score": 0.88, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 312, + 450, + 325 + ], + "score": 1.0, + "content": "-bisimulation metric", + "type": "text" + }, + { + "bbox": [ + 450, + 313, + 462, + 324 + ], + "score": 0.88, + "content": "d _ { \\sim } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 312, + 467, + 325 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 324, + 354, + 353 + ], + "lines": [ + { + "bbox": [ + 168, + 324, + 354, + 353 + ], + "spans": [ + { + "bbox": [ + 168, + 324, + 354, + 353 + ], + "score": 0.39, + "content": "\\begin{array} { r l } & { d _ { \\sim } ( x _ { 0 } , y _ { 0 } ) = d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } ) = ( 1 + \\gamma ) | r _ { y } - r _ { x } | } \\\\ & { d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) = \\operatorname* { m a x } \\left( ( 1 + \\gamma ) r _ { x } , r _ { y } \\right) , \\ d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y } \\end{array}", + "type": "interline_equation", + "image_path": "e4979757fb4a3e4cbd7241f251f17527e62bb56b8e6a2956053483f92eb86dca.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 168, + 324, + 354, + 338.5 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 168, + 338.5, + 354, + 353.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 353, + 503, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 503, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 131, + 368 + ], + "score": 1.0, + "content": "Thus,", + "type": "text" + }, + { + "bbox": [ + 131, + 353, + 208, + 366 + ], + "score": 0.9, + "content": "r _ { y } > \\left( 1 + 1 / \\gamma \\right) r _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 350, + 258, + 368 + ], + "score": 1.0, + "content": "implies that", + "type": "text" + }, + { + "bbox": [ + 258, + 353, + 359, + 365 + ], + "score": 0.89, + "content": "d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ( x _ { 0 } , y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 350, + 401, + 368 + ], + "score": 1.0, + "content": "as well as", + "type": "text" + }, + { + "bbox": [ + 401, + 353, + 503, + 365 + ], + "score": 0.91, + "content": "d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } )", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 370, + 506, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "score": 1.0, + "content": "Connection to data augmentation. Data augmentation often assumes access to optimality invariant", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "transformations, e.g., random crops or flips in image-based benchmarks (Laskin et al., 2020a;", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "Kostrikov et al., 2020). However, for certain RL tasks, such augmentations can erroneously alias", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "states with different optimal behavior and hurt generalization. For example, if the image observation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "is flipped in a goal reaching task with left and right actions, the optimal actions would also be flipped", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "to take left actions instead of right and vice versa. Proposition D.2 states that PSMs can precisely", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 436, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 295, + 448 + ], + "score": 1.0, + "content": "quantify the invariance of such augmentations.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 233, + 463 + ], + "score": 1.0, + "content": "Proposition D.2. For an MDP", + "type": "text" + }, + { + "bbox": [ + 234, + 449, + 252, + 460 + ], + "score": 0.87, + "content": "M _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 447, + 366, + 463 + ], + "score": 1.0, + "content": "and its transformed version", + "type": "text" + }, + { + "bbox": [ + 366, + 450, + 396, + 462 + ], + "score": 0.92, + "content": "M _ { \\psi ( \\chi ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 447, + 506, + 463 + ], + "score": 1.0, + "content": "for the data augmentation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 460, + 392, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 114, + 473 + ], + "score": 0.47, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 460, + 118, + 474 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 118, + 461, + 167, + 473 + ], + "score": 0.92, + "content": "d ^ { * } ( x , \\psi ( x ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 460, + 320, + 474 + ], + "score": 1.0, + "content": "indicates the optimality invariance of", + "type": "text" + }, + { + "bbox": [ + 320, + 462, + 327, + 473 + ], + "score": 0.75, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 460, + 360, + 474 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 360, + 462, + 387, + 471 + ], + "score": 0.89, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 460, + 392, + 474 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 106, + 485, + 343, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 343, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 343, + 498 + ], + "score": 1.0, + "content": "D.3 PSM WITH APPROXIMATELY-OPTIMAL POLICIES", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "Generalized Policy Similarity Metric for arbitrary policies. For a given DIST, we define a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 516, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 178, + 530 + ], + "score": 1.0, + "content": "generalized PSM", + "type": "text" + }, + { + "bbox": [ + 179, + 517, + 297, + 529 + ], + "score": 0.91, + "content": "\\dot { d } : ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\times ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 516, + 325, + 530 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 325, + 518, + 334, + 527 + ], + "score": 0.67, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 516, + 450, + 530 + ], + "score": 1.0, + "content": "is the set of all policies over", + "type": "text" + }, + { + "bbox": [ + 451, + 518, + 458, + 527 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 516, + 462, + 530 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 463, + 518, + 469, + 527 + ], + "score": 0.6, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 516, + 506, + 530 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 529, + 200, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 200, + 541 + ], + "score": 1.0, + "content": "the recursive equation:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 540, + 460, + 555 + ], + "lines": [ + { + "bbox": [ + 149, + 540, + 460, + 555 + ], + "spans": [ + { + "bbox": [ + 149, + 540, + 460, + 555 + ], + "score": 0.75, + "content": "d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) = \\mathrm { D I S T } \\bigl ( \\pi _ { 1 } ( x ) , \\pi _ { 2 } ( y ) \\bigr ) + \\gamma { \\mathcal W } _ { 1 } ( d ) \\bigl ( P ^ { \\pi _ { 1 } } ( \\cdot | x ) , P ^ { \\pi _ { 2 } } ( \\cdot | y ) \\bigr ) .", + "type": "interline_equation", + "image_path": "dc3026d09841fbcd424cddad26526fa80b33737ec272f5efff6f2f6bbca80fdd.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 149, + 540, + 460, + 555 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 506, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 317, + 568 + ], + "score": 1.0, + "content": "Since DIST is assumed to be a pseudometric and", + "type": "text" + }, + { + "bbox": [ + 317, + 556, + 333, + 566 + ], + "score": 0.87, + "content": "\\mathcal { W } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 555, + 497, + 568 + ], + "score": 1.0, + "content": "is a probability metric, it implies that", + "type": "text" + }, + { + "bbox": [ + 497, + 556, + 504, + 565 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 566, + 507, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 210, + 579 + ], + "score": 1.0, + "content": "is a pseudometric as (1)", + "type": "text" + }, + { + "bbox": [ + 211, + 567, + 217, + 577 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 567, + 321, + 579 + ], + "score": 1.0, + "content": "is non-negative, that is,", + "type": "text" + }, + { + "bbox": [ + 321, + 566, + 419, + 579 + ], + "score": 0.92, + "content": "d \\big ( ( x , \\pi _ { 1 } \\bar { ) } , ( y , \\pi _ { 2 } ) \\big ) ^ { \\cdot } \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 567, + 439, + 579 + ], + "score": 1.0, + "content": ", (2)", + "type": "text" + }, + { + "bbox": [ + 439, + 567, + 446, + 577 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 567, + 507, + 579 + ], + "score": 1.0, + "content": "is symmetric,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 578, + 507, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 138, + 593 + ], + "score": 1.0, + "content": "that is,", + "type": "text" + }, + { + "bbox": [ + 139, + 578, + 307, + 592 + ], + "score": 0.9, + "content": "d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big ) \\ = \\ d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 578, + 331, + 593 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 332, + 580, + 339, + 590 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 578, + 507, + 593 + ], + "score": 1.0, + "content": "satisfies the triangle inequality, that is,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 591, + 360, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 356, + 606 + ], + "score": 0.84, + "content": "d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) < d \\bigl ( ( x , \\pi _ { 1 } ) , ( z , \\pi _ { 3 } ) \\bigr ) + d \\bigl ( ( z , \\pi _ { 3 } ) , ( y , \\pi _ { 2 } ) \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 591, + 360, + 606 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 506, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "score": 1.0, + "content": "Using this notion of generalized PSM, we show that the approximation error in PSM from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 624, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 639 + ], + "score": 1.0, + "content": "using a suboptimal policy is bounded by the policy’s suboptimality. Thus, for policies with decreasing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 636, + 476, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 476, + 650 + ], + "score": 1.0, + "content": "suboptimality, the PSM approximation becomes more accurate, resulting in improved PSEs.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 650, + 505, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 328, + 664 + ], + "score": 1.0, + "content": "Proposition D.3. [Approximation error in PSM] Let", + "type": "text" + }, + { + "bbox": [ + 328, + 649, + 398, + 662 + ], + "score": 0.92, + "content": "\\hat { d } : \\mathcal { S } \\times \\mathcal { S } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 649, + 506, + 664 + ], + "score": 1.0, + "content": "be the approximate PSM", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 662, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 262, + 677 + ], + "score": 1.0, + "content": "computed using a suboptimal policy", + "type": "text" + }, + { + "bbox": [ + 262, + 664, + 270, + 674 + ], + "score": 0.53, + "content": "\\hat { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 663, + 327, + 677 + ], + "score": 1.0, + "content": "defined over", + "type": "text" + }, + { + "bbox": [ + 327, + 664, + 335, + 674 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 663, + 374, + 677 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + }, + { + "bbox": [ + 374, + 662, + 506, + 677 + ], + "score": 0.91, + "content": "\\hat { d } ( x , y ) ~ = ~ \\mathrm { D I S T } \\big ( \\hat { \\pi } ( x ) , \\hat { \\pi } ( y ) \\big ) ~ + ~", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 676, + 271, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 226, + 691 + ], + "score": 0.89, + "content": "\\gamma \\mathcal { W } _ { 1 } ( \\hat { d } ) \\big ( P ^ { \\hat { \\pi } } ( \\cdot \\vert x ) , P ^ { \\hat { \\pi } } ( \\cdot \\vert y ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 676, + 271, + 691 + ], + "score": 1.0, + "content": ". We have:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 692, + 459, + 708 + ], + "lines": [ + { + "bbox": [ + 150, + 692, + 459, + 708 + ], + "spans": [ + { + "bbox": [ + 150, + 692, + 459, + 708 + ], + "score": 0.79, + "content": "| d ^ { * } ( x , y ) - { \\hat { d } } ( x , y ) | < { { \\cal A } } \\bigl ( ( x , \\pi ^ { * } ) , ( x , { \\hat { \\pi } } ) \\bigr ) _ { \\bot } + { { \\cal A } } \\bigl ( ( y , { \\hat { \\pi } } ) , ( y , \\pi ^ { * } ) \\bigr ) _ { \\bot } .", + "type": "interline_equation", + "image_path": "10f9d8ec51df5d0cf299b606569303f4deef5f621c8aef98d9b38f7d68392ca9.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 150, + 692, + 459, + 708 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 78, + 466, + 182 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 140, + 78, + 466, + 182 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 78, + 466, + 182 + ], + "spans": [ + { + "bbox": [ + 140, + 78, + 466, + 182 + ], + "score": 0.968, + "type": "image", + "image_path": "3d8fd583f31c9cbc260975c387fd75a3ea68a6f1e11f6b2895873d41fe0189c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 140, + 78, + 466, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 140, + 112.66666666666666, + 466, + 147.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 140, + 147.33333333333331, + 466, + 181.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 191, + 504, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "Figure D.1: Cyan edges represent actions with a positive reward, which are also the optimal actions. Zero", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 202, + 420, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 199, + 213 + ], + "score": 1.0, + "content": "rewards everywhere else.", + "type": "text" + }, + { + "bbox": [ + 199, + 204, + 223, + 213 + ], + "score": 0.88, + "content": "x _ { 0 } , y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 202, + 312, + 213 + ], + "score": 1.0, + "content": "are the start states while", + "type": "text" + }, + { + "bbox": [ + 312, + 204, + 335, + 213 + ], + "score": 0.88, + "content": "x _ { 2 } , y _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 202, + 420, + 213 + ], + "score": 1.0, + "content": "are the terminal states.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 503, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 294, + 229 + ], + "score": 1.0, + "content": "Proposition D.1. There exists environments", + "type": "text" + }, + { + "bbox": [ + 294, + 217, + 312, + 227 + ], + "score": 0.88, + "content": "M _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 215, + 334, + 229 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 334, + 217, + 352, + 228 + ], + "score": 0.88, + "content": "M _ { \\mathcal { Y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 215, + 396, + 229 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 397, + 216, + 451, + 228 + ], + "score": 0.91, + "content": "\\forall ( x , y ) \\in ~ { \\mathcal { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 215, + 482, + 229 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 482, + 216, + 505, + 228 + ], + "score": 0.85, + "content": "{ \\mathcal { L } } =", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 223, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 107, + 229, + 168, + 242 + ], + "score": 0.29, + "content": "\\{ ( x , y ) ~ | x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 223, + 172, + 247 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 173, + 230, + 200, + 241 + ], + "score": 0.54, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 223, + 204, + 247 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 205, + 229, + 262, + 242 + ], + "score": 0.72, + "content": "d ^ { * } ( x , y ) = 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 223, + 266, + 247 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 266, + 228, + 382, + 244 + ], + "score": 0.9, + "content": "\\begin{array} { r } { d _ { \\sim } ^ { \\ast } ( x , y ) = \\frac { \\left| R _ { \\mathrm { m a x } } - R _ { \\mathrm { m i n } } \\right| } { 1 - \\gamma } - \\epsilon } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 223, + 440, + 247 + ], + "score": 1.0, + "content": "for any given", + "type": "text" + }, + { + "bbox": [ + 441, + 230, + 464, + 240 + ], + "score": 0.87, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 223, + 470, + 247 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 215, + 505, + 247 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 104, + 249, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 249, + 461, + 265 + ], + "score": 1.0, + "content": "For example, consider the two semantically equivalent environments in Figure D.1 with", + "type": "text" + }, + { + "bbox": [ + 461, + 251, + 506, + 264 + ], + "score": 0.91, + "content": "\\pi _ { \\mathcal { X } } ^ { * } ( x _ { 0 } ) =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 163, + 275 + ], + "score": 0.89, + "content": "\\pi _ { y } ^ { * } ( y _ { 0 } ) = a _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 261, + 183, + 276 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 183, + 262, + 287, + 275 + ], + "score": 0.92, + "content": "\\pi _ { \\mathcal { X } } ^ { \\ast } ( x _ { 1 } ) = \\pi _ { \\mathcal { Y } } ^ { \\ast } ( y _ { 1 } ) = a _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 261, + 378, + 276 + ], + "score": 1.0, + "content": "but different rewards", + "type": "text" + }, + { + "bbox": [ + 378, + 264, + 403, + 275 + ], + "score": 0.89, + "content": "r _ { x } , r _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "respectively. Whenever", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 272, + 504, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 183, + 285 + ], + "score": 0.86, + "content": "r _ { y } > ( 1 + 1 / \\gamma ) r _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 272, + 357, + 288 + ], + "score": 1.0, + "content": ", bisimulation metrics incorrectly imply that", + "type": "text" + }, + { + "bbox": [ + 358, + 275, + 369, + 284 + ], + "score": 0.86, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 272, + 494, + 288 + ], + "score": 1.0, + "content": "is more behaviorally similar to", + "type": "text" + }, + { + "bbox": [ + 494, + 275, + 504, + 285 + ], + "score": 0.84, + "content": "y _ { 1 }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 283, + 142, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 126, + 298 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 126, + 287, + 136, + 296 + ], + "score": 0.86, + "content": "y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 283, + 142, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 249, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 300, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 331, + 314 + ], + "score": 1.0, + "content": "For the MDPs shown in Figure D.1, to determine which", + "type": "text" + }, + { + "bbox": [ + 332, + 303, + 339, + 312 + ], + "score": 0.82, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 299, + 475, + 314 + ], + "score": 1.0, + "content": "state is behaviorally equivalent to", + "type": "text" + }, + { + "bbox": [ + 476, + 303, + 487, + 312 + ], + "score": 0.86, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 299, + 505, + 314 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 312, + 467, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 324, + 325 + ], + "score": 1.0, + "content": "look at the distances computed by bisimulation metric", + "type": "text" + }, + { + "bbox": [ + 325, + 313, + 338, + 323 + ], + "score": 0.88, + "content": "d _ { \\sim }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 312, + 356, + 325 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 313, + 367, + 322 + ], + "score": 0.88, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 312, + 450, + 325 + ], + "score": 1.0, + "content": "-bisimulation metric", + "type": "text" + }, + { + "bbox": [ + 450, + 313, + 462, + 324 + ], + "score": 0.88, + "content": "d _ { \\sim } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 312, + 467, + 325 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 299, + 505, + 325 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 324, + 354, + 353 + ], + "lines": [ + { + "bbox": [ + 168, + 324, + 354, + 353 + ], + "spans": [ + { + "bbox": [ + 168, + 324, + 354, + 353 + ], + "score": 0.39, + "content": "\\begin{array} { r l } & { d _ { \\sim } ( x _ { 0 } , y _ { 0 } ) = d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } ) = ( 1 + \\gamma ) | r _ { y } - r _ { x } | } \\\\ & { d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) = \\operatorname* { m a x } \\left( ( 1 + \\gamma ) r _ { x } , r _ { y } \\right) , \\ d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y } \\end{array}", + "type": "interline_equation", + "image_path": "e4979757fb4a3e4cbd7241f251f17527e62bb56b8e6a2956053483f92eb86dca.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 168, + 324, + 354, + 338.5 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 168, + 338.5, + 354, + 353.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 353, + 503, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 503, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 131, + 368 + ], + "score": 1.0, + "content": "Thus,", + "type": "text" + }, + { + "bbox": [ + 131, + 353, + 208, + 366 + ], + "score": 0.9, + "content": "r _ { y } > \\left( 1 + 1 / \\gamma \\right) r _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 350, + 258, + 368 + ], + "score": 1.0, + "content": "implies that", + "type": "text" + }, + { + "bbox": [ + 258, + 353, + 359, + 365 + ], + "score": 0.89, + "content": "d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ( x _ { 0 } , y _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 350, + 401, + 368 + ], + "score": 1.0, + "content": "as well as", + "type": "text" + }, + { + "bbox": [ + 401, + 353, + 503, + 365 + ], + "score": 0.91, + "content": "d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } )", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 350, + 503, + 368 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 370, + 506, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 382 + ], + "score": 1.0, + "content": "Connection to data augmentation. Data augmentation often assumes access to optimality invariant", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "transformations, e.g., random crops or flips in image-based benchmarks (Laskin et al., 2020a;", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "Kostrikov et al., 2020). However, for certain RL tasks, such augmentations can erroneously alias", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "states with different optimal behavior and hurt generalization. For example, if the image observation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "is flipped in a goal reaching task with left and right actions, the optimal actions would also be flipped", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "to take left actions instead of right and vice versa. Proposition D.2 states that PSMs can precisely", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 436, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 295, + 448 + ], + "score": 1.0, + "content": "quantify the invariance of such augmentations.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 370, + 506, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 233, + 463 + ], + "score": 1.0, + "content": "Proposition D.2. For an MDP", + "type": "text" + }, + { + "bbox": [ + 234, + 449, + 252, + 460 + ], + "score": 0.87, + "content": "M _ { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 447, + 366, + 463 + ], + "score": 1.0, + "content": "and its transformed version", + "type": "text" + }, + { + "bbox": [ + 366, + 450, + 396, + 462 + ], + "score": 0.92, + "content": "M _ { \\psi ( \\chi ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 447, + 506, + 463 + ], + "score": 1.0, + "content": "for the data augmentation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 460, + 392, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 114, + 473 + ], + "score": 0.47, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 460, + 118, + 474 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 118, + 461, + 167, + 473 + ], + "score": 0.92, + "content": "d ^ { * } ( x , \\psi ( x ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 460, + 320, + 474 + ], + "score": 1.0, + "content": "indicates the optimality invariance of", + "type": "text" + }, + { + "bbox": [ + 320, + 462, + 327, + 473 + ], + "score": 0.75, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 460, + 360, + 474 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 360, + 462, + 387, + 471 + ], + "score": 0.89, + "content": "x \\in \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 460, + 392, + 474 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 447, + 506, + 474 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 485, + 343, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 343, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 343, + 498 + ], + "score": 1.0, + "content": "D.3 PSM WITH APPROXIMATELY-OPTIMAL POLICIES", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "Generalized Policy Similarity Metric for arbitrary policies. For a given DIST, we define a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 516, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 178, + 530 + ], + "score": 1.0, + "content": "generalized PSM", + "type": "text" + }, + { + "bbox": [ + 179, + 517, + 297, + 529 + ], + "score": 0.91, + "content": "\\dot { d } : ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\times ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 516, + 325, + 530 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 325, + 518, + 334, + 527 + ], + "score": 0.67, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 516, + 450, + 530 + ], + "score": 1.0, + "content": "is the set of all policies over", + "type": "text" + }, + { + "bbox": [ + 451, + 518, + 458, + 527 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 516, + 462, + 530 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 463, + 518, + 469, + 527 + ], + "score": 0.6, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 516, + 506, + 530 + ], + "score": 1.0, + "content": "satisfies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 529, + 200, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 200, + 541 + ], + "score": 1.0, + "content": "the recursive equation:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 505, + 506, + 541 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 540, + 460, + 555 + ], + "lines": [ + { + "bbox": [ + 149, + 540, + 460, + 555 + ], + "spans": [ + { + "bbox": [ + 149, + 540, + 460, + 555 + ], + "score": 0.75, + "content": "d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) = \\mathrm { D I S T } \\bigl ( \\pi _ { 1 } ( x ) , \\pi _ { 2 } ( y ) \\bigr ) + \\gamma { \\mathcal W } _ { 1 } ( d ) \\bigl ( P ^ { \\pi _ { 1 } } ( \\cdot | x ) , P ^ { \\pi _ { 2 } } ( \\cdot | y ) \\bigr ) .", + "type": "interline_equation", + "image_path": "dc3026d09841fbcd424cddad26526fa80b33737ec272f5efff6f2f6bbca80fdd.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 149, + 540, + 460, + 555 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 506, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 317, + 568 + ], + "score": 1.0, + "content": "Since DIST is assumed to be a pseudometric and", + "type": "text" + }, + { + "bbox": [ + 317, + 556, + 333, + 566 + ], + "score": 0.87, + "content": "\\mathcal { W } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 555, + 497, + 568 + ], + "score": 1.0, + "content": "is a probability metric, it implies that", + "type": "text" + }, + { + "bbox": [ + 497, + 556, + 504, + 565 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 566, + 507, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 210, + 579 + ], + "score": 1.0, + "content": "is a pseudometric as (1)", + "type": "text" + }, + { + "bbox": [ + 211, + 567, + 217, + 577 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 567, + 321, + 579 + ], + "score": 1.0, + "content": "is non-negative, that is,", + "type": "text" + }, + { + "bbox": [ + 321, + 566, + 419, + 579 + ], + "score": 0.92, + "content": "d \\big ( ( x , \\pi _ { 1 } \\bar { ) } , ( y , \\pi _ { 2 } ) \\big ) ^ { \\cdot } \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 567, + 439, + 579 + ], + "score": 1.0, + "content": ", (2)", + "type": "text" + }, + { + "bbox": [ + 439, + 567, + 446, + 577 + ], + "score": 0.76, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 567, + 507, + 579 + ], + "score": 1.0, + "content": "is symmetric,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 578, + 507, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 138, + 593 + ], + "score": 1.0, + "content": "that is,", + "type": "text" + }, + { + "bbox": [ + 139, + 578, + 307, + 592 + ], + "score": 0.9, + "content": "d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big ) \\ = \\ d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 578, + 331, + 593 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 332, + 580, + 339, + 590 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 578, + 507, + 593 + ], + "score": 1.0, + "content": "satisfies the triangle inequality, that is,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 591, + 360, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 356, + 606 + ], + "score": 0.84, + "content": "d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) < d \\bigl ( ( x , \\pi _ { 1 } ) , ( z , \\pi _ { 3 } ) \\bigr ) + d \\bigl ( ( z , \\pi _ { 3 } ) , ( y , \\pi _ { 2 } ) \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 591, + 360, + 606 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 555, + 507, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 506, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 506, + 627 + ], + "score": 1.0, + "content": "Using this notion of generalized PSM, we show that the approximation error in PSM from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 624, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 639 + ], + "score": 1.0, + "content": "using a suboptimal policy is bounded by the policy’s suboptimality. Thus, for policies with decreasing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 636, + 476, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 476, + 650 + ], + "score": 1.0, + "content": "suboptimality, the PSM approximation becomes more accurate, resulting in improved PSEs.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 614, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 650, + 505, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 328, + 664 + ], + "score": 1.0, + "content": "Proposition D.3. [Approximation error in PSM] Let", + "type": "text" + }, + { + "bbox": [ + 328, + 649, + 398, + 662 + ], + "score": 0.92, + "content": "\\hat { d } : \\mathcal { S } \\times \\mathcal { S } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 649, + 506, + 664 + ], + "score": 1.0, + "content": "be the approximate PSM", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 662, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 262, + 677 + ], + "score": 1.0, + "content": "computed using a suboptimal policy", + "type": "text" + }, + { + "bbox": [ + 262, + 664, + 270, + 674 + ], + "score": 0.53, + "content": "\\hat { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 663, + 327, + 677 + ], + "score": 1.0, + "content": "defined over", + "type": "text" + }, + { + "bbox": [ + 327, + 664, + 335, + 674 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 663, + 374, + 677 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + }, + { + "bbox": [ + 374, + 662, + 506, + 677 + ], + "score": 0.91, + "content": "\\hat { d } ( x , y ) ~ = ~ \\mathrm { D I S T } \\big ( \\hat { \\pi } ( x ) , \\hat { \\pi } ( y ) \\big ) ~ + ~", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 676, + 271, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 226, + 691 + ], + "score": 0.89, + "content": "\\gamma \\mathcal { W } _ { 1 } ( \\hat { d } ) \\big ( P ^ { \\hat { \\pi } } ( \\cdot \\vert x ) , P ^ { \\hat { \\pi } } ( \\cdot \\vert y ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 676, + 271, + 691 + ], + "score": 1.0, + "content": ". We have:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 649, + 506, + 691 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 692, + 459, + 708 + ], + "lines": [ + { + "bbox": [ + 150, + 692, + 459, + 708 + ], + "spans": [ + { + "bbox": [ + 150, + 692, + 459, + 708 + ], + "score": 0.79, + "content": "| d ^ { * } ( x , y ) - { \\hat { d } } ( x , y ) | < { { \\cal A } } \\bigl ( ( x , \\pi ^ { * } ) , ( x , { \\hat { \\pi } } ) \\bigr ) _ { \\bot } + { { \\cal A } } \\bigl ( ( y , { \\hat { \\pi } } ) , ( y , \\pi ^ { * } ) \\bigr ) _ { \\bot } .", + "type": "interline_equation", + "image_path": "10f9d8ec51df5d0cf299b606569303f4deef5f621c8aef98d9b38f7d68392ca9.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 150, + 692, + 459, + 708 + ], + "spans": [], + "index": 40 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 175, + 96 + ], + "score": 1.0, + "content": "Proof. The PSM", + "type": "text" + }, + { + "bbox": [ + 175, + 83, + 186, + 92 + ], + "score": 0.84, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 80, + 276, + 96 + ], + "score": 1.0, + "content": "and approximate PSM", + "type": "text" + }, + { + "bbox": [ + 276, + 81, + 283, + 93 + ], + "score": 0.82, + "content": "\\hat { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "are instantiations of the generalized PSM (Equation D.3)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 304, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 213, + 106 + ], + "score": 1.0, + "content": "with both input policies as", + "type": "text" + }, + { + "bbox": [ + 214, + 93, + 226, + 102 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 92, + 243, + 106 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 93, + 251, + 102 + ], + "score": 0.8, + "content": "\\hat { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 92, + 304, + 106 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 109, + 117, + 499, + 178 + ], + "lines": [ + { + "bbox": [ + 109, + 117, + 499, + 178 + ], + "spans": [ + { + "bbox": [ + 109, + 117, + 499, + 178 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { d ^ { * } ( x , y ) = d ( ( x , \\pi ^ { * } ) , ( y , \\pi ^ { * } ) ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\\\ & { \\qquad < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { d ^ { * } ( x , y ) - \\hat { d } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) \\quad : \\ \\hat { d } ( x , y ) = d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { { \\mathrm { S i m i l a r l y } } , \\ \\hat { d } ( x , y ) - d ^ { * } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\end{array}", + "type": "interline_equation", + "image_path": "e95da61ecc378d62225eef9e66050b53f1ac386dd3f324fc9c32a1f64535e908.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 109, + 117, + 499, + 137.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 137.33333333333334, + 499, + 157.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 157.66666666666669, + 499, + 178.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 239, + 260, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 261, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 261, + 253 + ], + "score": 1.0, + "content": "E L2 METRIC EMBEDDINGS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "Another common choice (Zhang et al., 2021) for learning metric embeddings is to use the squared", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "loss (i.e., l2-loss) for matching the euclidean distance between the representations of a pair of states", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 384, + 302 + ], + "score": 1.0, + "content": "to the metric distance between those states. Concretely, for a given", + "type": "text" + }, + { + "bbox": [ + 385, + 290, + 396, + 299 + ], + "score": 0.86, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 289, + 475, + 302 + ], + "score": 1.0, + "content": "and representation", + "type": "text" + }, + { + "bbox": [ + 475, + 290, + 486, + 301 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 289, + 505, + 302 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 299, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 125, + 314 + ], + "score": 1.0, + "content": "loss", + "type": "text" + }, + { + "bbox": [ + 126, + 299, + 335, + 313 + ], + "score": 0.9, + "content": "\\mathcal { L } ( \\theta ) = \\mathbb { E } _ { s _ { i } , s _ { j } } [ ( \\| f _ { \\theta } ( s _ { i } ) - f _ { \\theta } ( s _ { j } ) \\| _ { 2 } - d ^ { * } ( s _ { i } , s _ { j } ) ) ^ { \\bar { 2 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 299, + 506, + 314 + ], + "score": 1.0, + "content": "is minimized. However, it might be too", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 311, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 494, + 324 + ], + "score": 1.0, + "content": "restrictive to match the exact metric distances, which we demonstrate empirically by comparing", + "type": "text" + }, + { + "bbox": [ + 495, + 311, + 504, + 321 + ], + "score": 0.36, + "content": "l 2", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 321, + 290, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 290, + 334 + ], + "score": 1.0, + "content": "metric embeddings with CMEs (Section 5.2).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 357, + 272, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 273, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 273, + 371 + ], + "score": 1.0, + "content": "F EXTENDED RELATED WORK", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "Generalization across different tasks used to be described as transfer learning. In the past, most", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "transfer learning approaches relied on fixed representations and tackled different problem formulations", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "(e.g., assuming shared state space). Taylor & Stone (2009) present a comprehensive survey of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "score": 1.0, + "content": "the techniques at the time, before representation learning became so prevalent in RL. Recently,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "the problem of performing well in a different, but related task, started to be seen as a problem", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "of generalization; with the community highlighting that deep RL agents tend to overfit to the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 452, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 463 + ], + "score": 1.0, + "content": "environments they are trained on (Cobbe et al., 2019; Witty et al., 2018; Farebrother et al., 2018;", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Juliani et al., 2019; Kostrikov et al., 2020; Song et al., 2019; Justesen et al., 2018; Packer et al., 2018).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 506, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 493 + ], + "score": 1.0, + "content": "Prior generalization approaches are typically adapted from supervised learning, including data aug-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "score": 1.0, + "content": "mentation, regularization (Cobbe et al., 2019; Farebrother et al., 2018), stochasticity (Zhang et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "2018c), noise injection (Igl et al., 2019; Zhang et al., 2018a), more diverse training conditions (Ra-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "jeswaran et al., 2017; Witty et al., 2018) and self-attention architectures (Tang et al., 2020). In contrast,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "score": 1.0, + "content": "PSEs exploits behavior similarity (Section 3), a property related to the sequential aspect of RL. Fur-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "score": 1.0, + "content": "thermore, for certain RL tasks, it can be unclear what an optimality invariant data augmentation would", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 545, + 501, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 501, + 558 + ], + "score": 1.0, + "content": "look like (Section 5.3). PSM can quantify the invariance of such augmentations (Proposition D.2).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 562, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "Meta-learning is also related to generalization. Meta-learning methods try to find a parametrization", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "score": 1.0, + "content": "that requires a small number of gradient steps to achieve good performance on a new task (Finn et al.,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "2017). In this context, various meta-learning approaches capable of zero-shot generalization have", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "score": 1.0, + "content": "been proposed (Li et al., 2018; Agarwal et al., 2019; Balaji et al., 2018). These approaches typically", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "consist in minimizing the loss in the environments the agent is while adding an auxiliary loss for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "ensuring improvement in the other (validation) environments available to the agent. Nevertheless,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "Tachet des Combes et al. (2018) has shown that meta-learning approaches fail in the jumping task", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "score": 1.0, + "content": "which we also observed empirically. Others have also reported similar findings (e.g., Farebrother", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 649, + 159, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 159, + 662 + ], + "score": 1.0, + "content": "et al., 2018).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "There are several other approaches for tackling zero-shot generalization in RL, but they often rely on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "domain-specific information. Some examples include knowledge about equivalences between entities", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "in the environment (Oh et al., 2017) and about what is under the agent’s control (Ye et al., 2020).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "Causality-based methods are a different way of tackling generalization, but current solutions do not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "scale to high-dimensional observation spaces (e.g., Killian et al., 2017; Perez et al., 2020; Zhang", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 722, + 158, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 722, + 158, + 734 + ], + "score": 1.0, + "content": "et al., 2020).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 495, + 192, + 505, + 202 + ], + "lines": [ + { + "bbox": [ + 496, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 496, + 192, + 505, + 204 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 175, + 96 + ], + "score": 1.0, + "content": "Proof. The PSM", + "type": "text" + }, + { + "bbox": [ + 175, + 83, + 186, + 92 + ], + "score": 0.84, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 80, + 276, + 96 + ], + "score": 1.0, + "content": "and approximate PSM", + "type": "text" + }, + { + "bbox": [ + 276, + 81, + 283, + 93 + ], + "score": 0.82, + "content": "\\hat { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "are instantiations of the generalized PSM (Equation D.3)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 304, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 213, + 106 + ], + "score": 1.0, + "content": "with both input policies as", + "type": "text" + }, + { + "bbox": [ + 214, + 93, + 226, + 102 + ], + "score": 0.87, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 92, + 243, + 106 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 93, + 251, + 102 + ], + "score": 0.8, + "content": "\\hat { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 92, + 304, + 106 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 80, + 506, + 106 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 109, + 117, + 499, + 178 + ], + "lines": [ + { + "bbox": [ + 109, + 117, + 499, + 178 + ], + "spans": [ + { + "bbox": [ + 109, + 117, + 499, + 178 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { d ^ { * } ( x , y ) = d ( ( x , \\pi ^ { * } ) , ( y , \\pi ^ { * } ) ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\\\ & { \\qquad < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { d ^ { * } ( x , y ) - \\hat { d } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) \\quad : \\ \\hat { d } ( x , y ) = d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { { \\mathrm { S i m i l a r l y } } , \\ \\hat { d } ( x , y ) - d ^ { * } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\end{array}", + "type": "interline_equation", + "image_path": "e95da61ecc378d62225eef9e66050b53f1ac386dd3f324fc9c32a1f64535e908.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 109, + 117, + 499, + 137.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 137.33333333333334, + 499, + 157.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 157.66666666666669, + 499, + 178.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 239, + 260, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 261, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 261, + 253 + ], + "score": 1.0, + "content": "E L2 METRIC EMBEDDINGS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "Another common choice (Zhang et al., 2021) for learning metric embeddings is to use the squared", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "loss (i.e., l2-loss) for matching the euclidean distance between the representations of a pair of states", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 384, + 302 + ], + "score": 1.0, + "content": "to the metric distance between those states. Concretely, for a given", + "type": "text" + }, + { + "bbox": [ + 385, + 290, + 396, + 299 + ], + "score": 0.86, + "content": "d ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 289, + 475, + 302 + ], + "score": 1.0, + "content": "and representation", + "type": "text" + }, + { + "bbox": [ + 475, + 290, + 486, + 301 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 289, + 505, + 302 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 299, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 125, + 314 + ], + "score": 1.0, + "content": "loss", + "type": "text" + }, + { + "bbox": [ + 126, + 299, + 335, + 313 + ], + "score": 0.9, + "content": "\\mathcal { L } ( \\theta ) = \\mathbb { E } _ { s _ { i } , s _ { j } } [ ( \\| f _ { \\theta } ( s _ { i } ) - f _ { \\theta } ( s _ { j } ) \\| _ { 2 } - d ^ { * } ( s _ { i } , s _ { j } ) ) ^ { \\bar { 2 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 299, + 506, + 314 + ], + "score": 1.0, + "content": "is minimized. However, it might be too", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 311, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 494, + 324 + ], + "score": 1.0, + "content": "restrictive to match the exact metric distances, which we demonstrate empirically by comparing", + "type": "text" + }, + { + "bbox": [ + 495, + 311, + 504, + 321 + ], + "score": 0.36, + "content": "l 2", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 321, + 290, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 290, + 334 + ], + "score": 1.0, + "content": "metric embeddings with CMEs (Section 5.2).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 266, + 506, + 334 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 357, + 272, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 357, + 273, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 273, + 371 + ], + "score": 1.0, + "content": "F EXTENDED RELATED WORK", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "Generalization across different tasks used to be described as transfer learning. In the past, most", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "transfer learning approaches relied on fixed representations and tackled different problem formulations", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "(e.g., assuming shared state space). Taylor & Stone (2009) present a comprehensive survey of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "score": 1.0, + "content": "the techniques at the time, before representation learning became so prevalent in RL. Recently,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "the problem of performing well in a different, but related task, started to be seen as a problem", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "of generalization; with the community highlighting that deep RL agents tend to overfit to the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 452, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 463 + ], + "score": 1.0, + "content": "environments they are trained on (Cobbe et al., 2019; Witty et al., 2018; Farebrother et al., 2018;", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Juliani et al., 2019; Kostrikov et al., 2020; Song et al., 2019; Justesen et al., 2018; Packer et al., 2018).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 385, + 506, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 506, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 493 + ], + "score": 1.0, + "content": "Prior generalization approaches are typically adapted from supervised learning, including data aug-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "score": 1.0, + "content": "mentation, regularization (Cobbe et al., 2019; Farebrother et al., 2018), stochasticity (Zhang et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 501, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 506, + 514 + ], + "score": 1.0, + "content": "2018c), noise injection (Igl et al., 2019; Zhang et al., 2018a), more diverse training conditions (Ra-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "jeswaran et al., 2017; Witty et al., 2018) and self-attention architectures (Tang et al., 2020). In contrast,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "score": 1.0, + "content": "PSEs exploits behavior similarity (Section 3), a property related to the sequential aspect of RL. Fur-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 546 + ], + "score": 1.0, + "content": "thermore, for certain RL tasks, it can be unclear what an optimality invariant data augmentation would", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 545, + 501, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 501, + 558 + ], + "score": 1.0, + "content": "look like (Section 5.3). PSM can quantify the invariance of such augmentations (Proposition D.2).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 478, + 506, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 562, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "Meta-learning is also related to generalization. Meta-learning methods try to find a parametrization", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 506, + 585 + ], + "score": 1.0, + "content": "that requires a small number of gradient steps to achieve good performance on a new task (Finn et al.,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "2017). In this context, various meta-learning approaches capable of zero-shot generalization have", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "score": 1.0, + "content": "been proposed (Li et al., 2018; Agarwal et al., 2019; Balaji et al., 2018). These approaches typically", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "consist in minimizing the loss in the environments the agent is while adding an auxiliary loss for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "ensuring improvement in the other (validation) environments available to the agent. Nevertheless,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "Tachet des Combes et al. (2018) has shown that meta-learning approaches fail in the jumping task", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "score": 1.0, + "content": "which we also observed empirically. Others have also reported similar findings (e.g., Farebrother", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 649, + 159, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 159, + 662 + ], + "score": 1.0, + "content": "et al., 2018).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 562, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "There are several other approaches for tackling zero-shot generalization in RL, but they often rely on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "domain-specific information. Some examples include knowledge about equivalences between entities", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "in the environment (Oh et al., 2017) and about what is under the agent’s control (Ye et al., 2020).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "Causality-based methods are a different way of tackling generalization, but current solutions do not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 709, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 104, + 709, + 506, + 725 + ], + "score": 1.0, + "content": "scale to high-dimensional observation spaces (e.g., Killian et al., 2017; Perez et al., 2020; Zhang", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 722, + 158, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 722, + 158, + 734 + ], + "score": 1.0, + "content": "et al., 2020).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 276, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 278, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 278, + 96 + ], + "score": 1.0, + "content": "G JUMPING TASK WITH PIXELS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 114, + 140, + 497, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 140, + 497, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 140, + 497, + 285 + ], + "spans": [ + { + "bbox": [ + 114, + 140, + 497, + 285 + ], + "score": 0.974, + "type": "image", + "image_path": "ea6505432a67100e306e9753d67f0ba9bec335abc00370ec65ac6866a23b14e2.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 114, + 140, + 497, + 188.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 188.33333333333334, + 497, + 236.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 114, + 236.66666666666669, + 497, + 285.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 297, + 506, + 328 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "Figure G.1: Optimal trajectories on the jumping tasks for two different environments. Note that the optimal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 307, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 506, + 319 + ], + "score": 1.0, + "content": "trajectory is a sequence of right actions, followed by a single jump at a certain distance from the obstacle,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 316, + 202, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 202, + 329 + ], + "score": 1.0, + "content": "followed by right actions.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 368, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "Detailed Task Description. The jumping task consists of an agent trying to jump over an obstacle", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 438, + 392 + ], + "score": 1.0, + "content": "on a floor. The environment is deterministic with the agent observing a reward of", + "type": "text" + }, + { + "bbox": [ + 438, + 380, + 452, + 390 + ], + "score": 0.83, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "at each time", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "step. If the agent successfully reaches the rightmost side of the screen, it receives a bonus reward of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 128, + 412 + ], + "score": 0.86, + "content": "+ 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "; if the agent touches the obstacle, the episode terminates. The observation space is the pixel", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "representation of the environment, as depicted in Figure 1. The agent has access to two actions: right", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 423, + 439, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 439, + 436 + ], + "score": 1.0, + "content": "and jump. The jump action moves the agent vertically and horizontally to the right.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "Architecture. The neural network used for Jumping Task experiment is adapted from the Nature", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 451, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 504, + 462 + ], + "score": 1.0, + "content": "DQN architecture. Specifically, the network consists of 3 convolutional layers of sizes 32, 64, 64", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 168, + 474 + ], + "score": 1.0, + "content": "with filter sizes", + "type": "text" + }, + { + "bbox": [ + 168, + 462, + 191, + 473 + ], + "score": 0.84, + "content": "8 \\times 8", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 462, + 194, + 474 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 195, + 462, + 218, + 473 + ], + "score": 0.86, + "content": "4 \\times 4", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 462, + 235, + 474 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 462, + 259, + 473 + ], + "score": 0.9, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "and strides 4, 2, and 1, respectively. The output of the convnet", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "is fed into a single fully connected layer of size 256 followed by ’ReLU’ non-linearity. Finally, this", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "FC layer output is fed into a linear layer which computes the policy which outputs the probability of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 495, + 216, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 216, + 508 + ], + "score": 1.0, + "content": "the jump and right actions.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 438, + 525 + ], + "score": 1.0, + "content": "Contrastive Embedding. For all our experiments, we use a single ReLU layer with", + "type": "text" + }, + { + "bbox": [ + 439, + 512, + 469, + 522 + ], + "score": 0.9, + "content": "k = 6 4", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 511, + 506, + 525 + ], + "score": 1.0, + "content": "units for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 303, + 537 + ], + "score": 1.0, + "content": "the non-linear projection to obtain the embedding", + "type": "text" + }, + { + "bbox": [ + 303, + 524, + 313, + 534 + ], + "score": 0.83, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 522, + 505, + 537 + ], + "score": 1.0, + "content": "(Figure A.1). We compute the embedding using", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 534, + 486, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 486, + 546 + ], + "score": 1.0, + "content": "the penultimate layer in the jumping task network. Hyperparameters are reported in Table G.2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 504, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 331, + 563 + ], + "score": 1.0, + "content": "Total Loss. For jumpy world, the total loss is given by", + "type": "text" + }, + { + "bbox": [ + 331, + 551, + 394, + 562 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { I L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 550, + 423, + 563 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 423, + 551, + 440, + 562 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { I L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "is the imitation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 561, + 436, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 162, + 574 + ], + "score": 1.0, + "content": "learning loss,", + "type": "text" + }, + { + "bbox": [ + 162, + 562, + 191, + 573 + ], + "score": 0.85, + "content": "\\mathcal { L } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 561, + 424, + 574 + ], + "score": 1.0, + "content": "is the auxiliary loss for learning PSEs with the coefficient", + "type": "text" + }, + { + "bbox": [ + 424, + 564, + 431, + 572 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 561, + 436, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "image", + "bbox": [ + 108, + 594, + 502, + 689 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 594, + 502, + 689 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 594, + 502, + 689 + ], + "spans": [ + { + "bbox": [ + 108, + 594, + 502, + 689 + ], + "score": 0.967, + "type": "image", + "image_path": "b0d8815c5d7275437c5956e7dcd2666091c0d6503fae943844e0b9a282c3036a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 108, + 594, + 502, + 625.6666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 108, + 625.6666666666666, + 502, + 657.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 108, + 657.3333333333333, + 502, + 688.9999999999999 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 698, + 504, + 729 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 709 + ], + "score": 1.0, + "content": "Figure G.3: Test performance curves in the setting without data augmentation on the “wide”, “narrow”, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 708, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 720 + ], + "score": 1.0, + "content": "random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 718, + 176, + 729 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 176, + 729 + ], + "score": 1.0, + "content": "and 75 percentiles.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "index": 26.5 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 276, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 278, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 278, + 96 + ], + "score": 1.0, + "content": "G JUMPING TASK WITH PIXELS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 114, + 140, + 497, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 140, + 497, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 140, + 497, + 285 + ], + "spans": [ + { + "bbox": [ + 114, + 140, + 497, + 285 + ], + "score": 0.974, + "type": "image", + "image_path": "ea6505432a67100e306e9753d67f0ba9bec335abc00370ec65ac6866a23b14e2.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 114, + 140, + 497, + 188.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 114, + 188.33333333333334, + 497, + 236.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 114, + 236.66666666666669, + 497, + 285.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 297, + 506, + 328 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 308 + ], + "score": 1.0, + "content": "Figure G.1: Optimal trajectories on the jumping tasks for two different environments. Note that the optimal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 307, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 506, + 319 + ], + "score": 1.0, + "content": "trajectory is a sequence of right actions, followed by a single jump at a certain distance from the obstacle,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 316, + 202, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 202, + 329 + ], + "score": 1.0, + "content": "followed by right actions.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 368, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "Detailed Task Description. The jumping task consists of an agent trying to jump over an obstacle", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 438, + 392 + ], + "score": 1.0, + "content": "on a floor. The environment is deterministic with the agent observing a reward of", + "type": "text" + }, + { + "bbox": [ + 438, + 380, + 452, + 390 + ], + "score": 0.83, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "at each time", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "step. If the agent successfully reaches the rightmost side of the screen, it receives a bonus reward of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 128, + 412 + ], + "score": 0.86, + "content": "+ 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "; if the agent touches the obstacle, the episode terminates. The observation space is the pixel", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "representation of the environment, as depicted in Figure 1. The agent has access to two actions: right", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 423, + 439, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 439, + 436 + ], + "score": 1.0, + "content": "and jump. The jump action moves the agent vertically and horizontally to the right.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 368, + 506, + 436 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "Architecture. The neural network used for Jumping Task experiment is adapted from the Nature", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 451, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 504, + 462 + ], + "score": 1.0, + "content": "DQN architecture. Specifically, the network consists of 3 convolutional layers of sizes 32, 64, 64", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 168, + 474 + ], + "score": 1.0, + "content": "with filter sizes", + "type": "text" + }, + { + "bbox": [ + 168, + 462, + 191, + 473 + ], + "score": 0.84, + "content": "8 \\times 8", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 462, + 194, + 474 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 195, + 462, + 218, + 473 + ], + "score": 0.86, + "content": "4 \\times 4", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 462, + 235, + 474 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 462, + 259, + 473 + ], + "score": 0.9, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "and strides 4, 2, and 1, respectively. The output of the convnet", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "is fed into a single fully connected layer of size 256 followed by ’ReLU’ non-linearity. Finally, this", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "FC layer output is fed into a linear layer which computes the policy which outputs the probability of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 495, + 216, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 216, + 508 + ], + "score": 1.0, + "content": "the jump and right actions.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 440, + 506, + 508 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 438, + 525 + ], + "score": 1.0, + "content": "Contrastive Embedding. For all our experiments, we use a single ReLU layer with", + "type": "text" + }, + { + "bbox": [ + 439, + 512, + 469, + 522 + ], + "score": 0.9, + "content": "k = 6 4", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 511, + 506, + 525 + ], + "score": 1.0, + "content": "units for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 303, + 537 + ], + "score": 1.0, + "content": "the non-linear projection to obtain the embedding", + "type": "text" + }, + { + "bbox": [ + 303, + 524, + 313, + 534 + ], + "score": 0.83, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 522, + 505, + 537 + ], + "score": 1.0, + "content": "(Figure A.1). We compute the embedding using", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 534, + 486, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 486, + 546 + ], + "score": 1.0, + "content": "the penultimate layer in the jumping task network. Hyperparameters are reported in Table G.2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 511, + 506, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 504, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 331, + 563 + ], + "score": 1.0, + "content": "Total Loss. For jumpy world, the total loss is given by", + "type": "text" + }, + { + "bbox": [ + 331, + 551, + 394, + 562 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { I L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 550, + 423, + 563 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 423, + 551, + 440, + 562 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { I L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "is the imitation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 561, + 436, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 162, + 574 + ], + "score": 1.0, + "content": "learning loss,", + "type": "text" + }, + { + "bbox": [ + 162, + 562, + 191, + 573 + ], + "score": 0.85, + "content": "\\mathcal { L } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 561, + 424, + 574 + ], + "score": 1.0, + "content": "is the auxiliary loss for learning PSEs with the coefficient", + "type": "text" + }, + { + "bbox": [ + 424, + 564, + 431, + 572 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 561, + 436, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 550, + 505, + 574 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 594, + 502, + 689 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 594, + 502, + 689 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 594, + 502, + 689 + ], + "spans": [ + { + "bbox": [ + 108, + 594, + 502, + 689 + ], + "score": 0.967, + "type": "image", + "image_path": "b0d8815c5d7275437c5956e7dcd2666091c0d6503fae943844e0b9a282c3036a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 108, + 594, + 502, + 625.6666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 108, + 625.6666666666666, + 502, + 657.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 108, + 657.3333333333333, + 502, + 688.9999999999999 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 698, + 504, + 729 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 709 + ], + "score": 1.0, + "content": "Figure G.3: Test performance curves in the setting without data augmentation on the “wide”, “narrow”, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 708, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 720 + ], + "score": 1.0, + "content": "random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 718, + 176, + 729 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 176, + 729 + ], + "score": 1.0, + "content": "and 75 percentiles.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "index": 26.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 69, + 485, + 425 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 69, + 485, + 425 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 69, + 485, + 425 + ], + "spans": [ + { + "bbox": [ + 126, + 69, + 485, + 425 + ], + "score": 0.978, + "type": "image", + "image_path": "7ae61140cd1c0eae07e313b43e44f832d172fa581accd9edd3d1ff30eb85643b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 126, + 69, + 485, + 187.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 126, + 187.66666666666669, + 485, + 306.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 126, + 306.33333333333337, + 485, + 425.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 434, + 506, + 505 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 434, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 468, + 446 + ], + "score": 1.0, + "content": "Figure G.2: PSM vs. Bisimulation. Visualizing PSM and bisimulation metrics with discount factor", + "type": "text" + }, + { + "bbox": [ + 468, + 435, + 503, + 445 + ], + "score": 0.9, + "content": "\\gamma = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 434, + 506, + 446 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 107, + 447, + 116, + 455 + ], + "score": 0.83, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 444, + 132, + 457 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 132, + 446, + 141, + 455 + ], + "score": 0.82, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "correspond to the states visited by optimal policies in the two environments with obstacle at positions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 452, + 507, + 468 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 265, + 468 + ], + "score": 1.0, + "content": "25 and 45, respectively. For each grid, the", + "type": "text" + }, + { + "bbox": [ + 265, + 454, + 292, + 466 + ], + "score": 0.92, + "content": "( \\dot { i } , j ) ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 452, + 507, + 468 + ], + "score": 1.0, + "content": "location shows the distance assigned by the metric to the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 128, + 477 + ], + "score": 1.0, + "content": "states", + "type": "text" + }, + { + "bbox": [ + 128, + 466, + 138, + 474 + ], + "score": 0.85, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 464, + 154, + 477 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 466, + 164, + 475 + ], + "score": 0.86, + "content": "y _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 464, + 505, + 477 + ], + "score": 1.0, + "content": ". Lower distances (lighter shades) imply higher similarity between two states. The bottom right", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "figure shows which states are equivalent to each other. The ranges are different for each metric as bisimulation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 485, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 352, + 496 + ], + "score": 1.0, + "content": "metrics utilize reward differences while PSM uses the total variation", + "type": "text" + }, + { + "bbox": [ + 352, + 485, + 373, + 495 + ], + "score": 0.64, + "content": "( T V )", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 485, + 505, + 496 + ], + "score": 1.0, + "content": "distance between policies. Note that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 494, + 465, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 465, + 506 + ], + "score": 1.0, + "content": "the large bisimulation distances are due to the fact that the reward at the terminal state is set to 100.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 108, + 608, + 502, + 700 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 608, + 502, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 608, + 502, + 700 + ], + "spans": [ + { + "bbox": [ + 108, + 608, + 502, + 700 + ], + "score": 0.957, + "type": "image", + "image_path": "42a79abb3d752df4e248de208365ef50ff1cf065c5fbce26bd1ba3fcb36ec3f1.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 608, + 502, + 638.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 638.6666666666666, + 502, + 669.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 669.3333333333333, + 502, + 699.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 709, + 506, + 740 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 720 + ], + "score": 1.0, + "content": "Figure G.4: Test performance curves in the setting with data augmentation on the “wide”, “narrow”, and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 720, + 505, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 730 + ], + "score": 1.0, + "content": "random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 730, + 176, + 741 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 176, + 741 + ], + "score": 1.0, + "content": "and 75 percentiles.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 761 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "21", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 69, + 485, + 425 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 69, + 485, + 425 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 69, + 485, + 425 + ], + "spans": [ + { + "bbox": [ + 126, + 69, + 485, + 425 + ], + "score": 0.978, + "type": "image", + "image_path": "7ae61140cd1c0eae07e313b43e44f832d172fa581accd9edd3d1ff30eb85643b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 126, + 69, + 485, + 187.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 126, + 187.66666666666669, + 485, + 306.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 126, + 306.33333333333337, + 485, + 425.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 434, + 506, + 505 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 434, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 468, + 446 + ], + "score": 1.0, + "content": "Figure G.2: PSM vs. Bisimulation. Visualizing PSM and bisimulation metrics with discount factor", + "type": "text" + }, + { + "bbox": [ + 468, + 435, + 503, + 445 + ], + "score": 0.9, + "content": "\\gamma = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 434, + 506, + 446 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 107, + 447, + 116, + 455 + ], + "score": 0.83, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 444, + 132, + 457 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 132, + 446, + 141, + 455 + ], + "score": 0.82, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "correspond to the states visited by optimal policies in the two environments with obstacle at positions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 452, + 507, + 468 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 265, + 468 + ], + "score": 1.0, + "content": "25 and 45, respectively. For each grid, the", + "type": "text" + }, + { + "bbox": [ + 265, + 454, + 292, + 466 + ], + "score": 0.92, + "content": "( \\dot { i } , j ) ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 452, + 507, + 468 + ], + "score": 1.0, + "content": "location shows the distance assigned by the metric to the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 128, + 477 + ], + "score": 1.0, + "content": "states", + "type": "text" + }, + { + "bbox": [ + 128, + 466, + 138, + 474 + ], + "score": 0.85, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 464, + 154, + 477 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 154, + 466, + 164, + 475 + ], + "score": 0.86, + "content": "y _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 464, + 505, + 477 + ], + "score": 1.0, + "content": ". Lower distances (lighter shades) imply higher similarity between two states. The bottom right", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "figure shows which states are equivalent to each other. The ranges are different for each metric as bisimulation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 485, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 352, + 496 + ], + "score": 1.0, + "content": "metrics utilize reward differences while PSM uses the total variation", + "type": "text" + }, + { + "bbox": [ + 352, + 485, + 373, + 495 + ], + "score": 0.64, + "content": "( T V )", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 485, + 505, + 496 + ], + "score": 1.0, + "content": "distance between policies. Note that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 494, + 465, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 465, + 506 + ], + "score": 1.0, + "content": "the large bisimulation distances are due to the fact that the reward at the terminal state is set to 100.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 108, + 608, + 502, + 700 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 608, + 502, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 608, + 502, + 700 + ], + "spans": [ + { + "bbox": [ + 108, + 608, + 502, + 700 + ], + "score": 0.957, + "type": "image", + "image_path": "42a79abb3d752df4e248de208365ef50ff1cf065c5fbce26bd1ba3fcb36ec3f1.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 608, + 502, + 638.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 638.6666666666666, + 502, + 669.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 669.3333333333333, + 502, + 699.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 709, + 506, + 740 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 720 + ], + "score": 1.0, + "content": "Figure G.4: Test performance curves in the setting with data augmentation on the “wide”, “narrow”, and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 720, + 505, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 730 + ], + "score": 1.0, + "content": "random grids described in Figure 2. We plot the median performance across 100 runs. Shaded regions show 25", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 730, + 176, + 741 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 176, + 741 + ], + "score": 1.0, + "content": "and 75 percentiles.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 118, + 257, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 118, + 257, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 118, + 257, + 188 + ], + "spans": [ + { + "bbox": [ + 114, + 118, + 257, + 188 + ], + "score": 0.964, + "type": "image", + "image_path": "23cbe4e7878d1818c802facc5015190dbca31d0a9ec6d66d453c3d286303c25d.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 114, + 118, + 257, + 153.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 153.0, + 257, + 188.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 268, + 227 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 197, + 268, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 268, + 208 + ], + "score": 1.0, + "content": "Figure G.5: Jumping task with colored ob-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 207, + 268, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 268, + 218 + ], + "score": 1.0, + "content": "stacles. The agent needs to jump over the red", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 217, + 243, + 228 + ], + "spans": [ + { + "bbox": [ + 107, + 217, + 243, + 228 + ], + "score": 1.0, + "content": "obstacle but strike the green obstacle.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + } + ], + "index": 4.75 + }, + { + "type": "table", + "bbox": [ + 292, + 163, + 492, + 231 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 281, + 113, + 502, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 281, + 113, + 502, + 124 + ], + "spans": [ + { + "bbox": [ + 281, + 113, + 365, + 124 + ], + "score": 1.0, + "content": "Table G.1: Percentage", + "type": "text" + }, + { + "bbox": [ + 366, + 114, + 380, + 124 + ], + "score": 0.72, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 113, + 502, + 124 + ], + "score": 1.0, + "content": "of test tasks solved when trained", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 281, + 123, + 503, + 134 + ], + "spans": [ + { + "bbox": [ + 281, + 123, + 503, + 134 + ], + "score": 1.0, + "content": "on the “wide” grid with both red and green obstacles. The", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 281, + 133, + 502, + 144 + ], + "spans": [ + { + "bbox": [ + 281, + 133, + 502, + 144 + ], + "score": 1.0, + "content": "numbers we report are averaged across 100 runs. Standard", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 282, + 143, + 420, + 154 + ], + "spans": [ + { + "bbox": [ + 282, + 143, + 420, + 154 + ], + "score": 1.0, + "content": "error is reported between parentheses.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 292, + 163, + 492, + 231 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 292, + 163, + 492, + 231 + ], + "spans": [ + { + "bbox": [ + 292, + 163, + 492, + 231 + ], + "score": 0.971, + "html": "
MethodRed (%)Green (%)
RandConv6.2 (0.4)99.6 (0.2)
Dropout and l2 reg.19.5(0.2)100.0 (0.0)
RandConv +PSEs29.8 (1.3)99.6 (0.2)
PSEs37.9 (1.9)100.0 (0.0)
", + "type": "table", + "image_path": "22af8666fadce0f1cd2db61684e23b73b29f2c8cbdb18535ec2f9cbc313a93d6.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 292, + 163, + 492, + 174.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 292, + 174.33333333333334, + 492, + 185.66666666666669 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 292, + 185.66666666666669, + 492, + 197.00000000000003 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 292, + 197.00000000000003, + 492, + 208.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 292, + 208.33333333333337, + 492, + 219.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 292, + 219.6666666666667, + 492, + 231.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "image", + "bbox": [ + 119, + 262, + 500, + 410 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 262, + 500, + 410 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 262, + 500, + 410 + ], + "spans": [ + { + "bbox": [ + 119, + 262, + 500, + 410 + ], + "score": 0.974, + "type": "image", + "image_path": "b97fc7dbb58cf30e9e1722e2dc460df60c4711f902a11b8b59a27bf8a70f5846.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 119, + 262, + 500, + 311.3333333333333 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 119, + 311.3333333333333, + 500, + 360.66666666666663 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 119, + 360.66666666666663, + 500, + 409.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 428, + 505, + 469 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "Figure G.6: Randconv enforces color invariance. The first column shows the original observation where", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "the top row corresponds to the observation from task with green obstacle jumping task while the bottom row", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "score": 1.0, + "content": "corresponds to the red obstacle jumping task. Columns 2-6 in each row show 5 augmentations by applying", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "RandConv. The augmentations show that RandConv tries to encode invariance with respect to the obstacle color.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 17.75 + }, + { + "type": "title", + "bbox": [ + 107, + 494, + 221, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 221, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 221, + 506 + ], + "score": 1.0, + "content": "G.2 HYPERPARAMETERS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "in the “wide” grid and pick the parameters with the best validation performance. The validation", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "set (Figure G.7) was selected by using the environments nearby to the training environments whose", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 549, + 355, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 355, + 563 + ], + "score": 1.0, + "content": "floor height differ by 1 or whose obstacle position differ by 1.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "table", + "bbox": [ + 117, + 606, + 256, + 730 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 577, + 267, + 598 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 577, + 267, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 267, + 588 + ], + "score": 1.0, + "content": "Table G.2: Common hyperparameters across", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 587, + 267, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 267, + 599 + ], + "score": 1.0, + "content": "all methods for all jumping task experiments.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "table_body", + "bbox": [ + 117, + 606, + 256, + 730 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 606, + 256, + 730 + ], + "spans": [ + { + "bbox": [ + 117, + 606, + 256, + 730 + ], + "score": 0.972, + "html": "
HyperparameterValue
Learning rate decay Training epochs0.999
Optimizer2000 Adam
Batch size (Imitation)256
Num training tasks I-scale Parameter (β)18
Embedding size (k)0.01
Batch Size (LCME)64
57
|T57
", + "type": "table", + "image_path": "84cb1c01e9bc78c25c4259b474b636fbd84f41c7ab251a87ba9b20cfba802612.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 117, + 606, + 256, + 668.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 117, + 668.0, + 256, + 730.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 30.5 + }, + { + "type": "image", + "bbox": [ + 289, + 594, + 500, + 684 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 289, + 594, + 500, + 684 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 289, + 594, + 500, + 684 + ], + "spans": [ + { + "bbox": [ + 289, + 594, + 500, + 684 + ], + "score": 0.967, + "type": "image", + "image_path": "c50466fd3830e13b4efa16578edae345ae90199f2967bef1d51b49352ce01705.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 289, + 594, + 500, + 605.25 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 289, + 605.25, + 500, + 616.5 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 289, + 616.5, + 500, + 627.75 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 289, + 627.75, + 500, + 639.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 289, + 639.0, + 500, + 650.25 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 289, + 650.25, + 500, + 661.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 289, + 661.5, + 500, + 672.75 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 289, + 672.75, + 500, + 684.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 284, + 696, + 505, + 717 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 284, + 696, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 284, + 696, + 505, + 708 + ], + "score": 1.0, + "content": "Figure G.7: Unseen 56 validation tasks, labeled with V, used", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 284, + 707, + 392, + 718 + ], + "spans": [ + { + "bbox": [ + 284, + 707, + 392, + 718 + ], + "score": 1.0, + "content": "for hyperparameter selection.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "index": 36.5 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 83, + 263, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 264, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 264, + 95 + ], + "score": 1.0, + "content": "G.1 JUMPING TASK FROM COLORS", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 118, + 257, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 118, + 257, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 118, + 257, + 188 + ], + "spans": [ + { + "bbox": [ + 114, + 118, + 257, + 188 + ], + "score": 0.964, + "type": "image", + "image_path": "23cbe4e7878d1818c802facc5015190dbca31d0a9ec6d66d453c3d286303c25d.jpg" + } + ] + } + ], + "index": 2.5, + "virtual_lines": [ + { + "bbox": [ + 114, + 118, + 257, + 153.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 114, + 153.0, + 257, + 188.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 268, + 227 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 197, + 268, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 268, + 208 + ], + "score": 1.0, + "content": "Figure G.5: Jumping task with colored ob-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 207, + 268, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 268, + 218 + ], + "score": 1.0, + "content": "stacles. The agent needs to jump over the red", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 217, + 243, + 228 + ], + "spans": [ + { + "bbox": [ + 107, + 217, + 243, + 228 + ], + "score": 1.0, + "content": "obstacle but strike the green obstacle.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + } + ], + "index": 4.75 + }, + { + "type": "table", + "bbox": [ + 292, + 163, + 492, + 231 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 281, + 113, + 502, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 281, + 113, + 502, + 124 + ], + "spans": [ + { + "bbox": [ + 281, + 113, + 365, + 124 + ], + "score": 1.0, + "content": "Table G.1: Percentage", + "type": "text" + }, + { + "bbox": [ + 366, + 114, + 380, + 124 + ], + "score": 0.72, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 113, + 502, + 124 + ], + "score": 1.0, + "content": "of test tasks solved when trained", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 281, + 123, + 503, + 134 + ], + "spans": [ + { + "bbox": [ + 281, + 123, + 503, + 134 + ], + "score": 1.0, + "content": "on the “wide” grid with both red and green obstacles. The", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 281, + 133, + 502, + 144 + ], + "spans": [ + { + "bbox": [ + 281, + 133, + 502, + 144 + ], + "score": 1.0, + "content": "numbers we report are averaged across 100 runs. Standard", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 282, + 143, + 420, + 154 + ], + "spans": [ + { + "bbox": [ + 282, + 143, + 420, + 154 + ], + "score": 1.0, + "content": "error is reported between parentheses.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 292, + 163, + 492, + 231 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 292, + 163, + 492, + 231 + ], + "spans": [ + { + "bbox": [ + 292, + 163, + 492, + 231 + ], + "score": 0.971, + "html": "
MethodRed (%)Green (%)
RandConv6.2 (0.4)99.6 (0.2)
Dropout and l2 reg.19.5(0.2)100.0 (0.0)
RandConv +PSEs29.8 (1.3)99.6 (0.2)
PSEs37.9 (1.9)100.0 (0.0)
", + "type": "table", + "image_path": "22af8666fadce0f1cd2db61684e23b73b29f2c8cbdb18535ec2f9cbc313a93d6.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 292, + 163, + 492, + 174.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 292, + 174.33333333333334, + 492, + 185.66666666666669 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 292, + 185.66666666666669, + 492, + 197.00000000000003 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 292, + 197.00000000000003, + 492, + 208.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 292, + 208.33333333333337, + 492, + 219.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 292, + 219.6666666666667, + 492, + 231.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "image", + "bbox": [ + 119, + 262, + 500, + 410 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 262, + 500, + 410 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 262, + 500, + 410 + ], + "spans": [ + { + "bbox": [ + 119, + 262, + 500, + 410 + ], + "score": 0.974, + "type": "image", + "image_path": "b97fc7dbb58cf30e9e1722e2dc460df60c4711f902a11b8b59a27bf8a70f5846.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 119, + 262, + 500, + 311.3333333333333 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 119, + 311.3333333333333, + 500, + 360.66666666666663 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 119, + 360.66666666666663, + 500, + 409.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 428, + 505, + 469 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "Figure G.6: Randconv enforces color invariance. The first column shows the original observation where", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "the top row corresponds to the observation from task with green obstacle jumping task while the bottom row", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "score": 1.0, + "content": "corresponds to the red obstacle jumping task. Columns 2-6 in each row show 5 augmentations by applying", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "RandConv. The augmentations show that RandConv tries to encode invariance with respect to the obstacle color.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 17.75 + }, + { + "type": "title", + "bbox": [ + 107, + 494, + 221, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 221, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 221, + 506 + ], + "score": 1.0, + "content": "G.2 HYPERPARAMETERS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "For hyperparameter selection, we evaluate all agents on a validation set containing 54 unseen tasks", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "in the “wide” grid and pick the parameters with the best validation performance. The validation", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "set (Figure G.7) was selected by using the environments nearby to the training environments whose", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 549, + 355, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 355, + 563 + ], + "score": 1.0, + "content": "floor height differ by 1 or whose obstacle position differ by 1.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 516, + 506, + 563 + ] + }, + { + "type": "table", + "bbox": [ + 117, + 606, + 256, + 730 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 577, + 267, + 598 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 577, + 267, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 267, + 588 + ], + "score": 1.0, + "content": "Table G.2: Common hyperparameters across", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 587, + 267, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 267, + 599 + ], + "score": 1.0, + "content": "all methods for all jumping task experiments.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "table_body", + "bbox": [ + 117, + 606, + 256, + 730 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 606, + 256, + 730 + ], + "spans": [ + { + "bbox": [ + 117, + 606, + 256, + 730 + ], + "score": 0.972, + "html": "
HyperparameterValue
Learning rate decay Training epochs0.999
Optimizer2000 Adam
Batch size (Imitation)256
Num training tasks I-scale Parameter (β)18
Embedding size (k)0.01
Batch Size (LCME)64
57
|T57
", + "type": "table", + "image_path": "84cb1c01e9bc78c25c4259b474b636fbd84f41c7ab251a87ba9b20cfba802612.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 117, + 606, + 256, + 668.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 117, + 668.0, + 256, + 730.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 30.5 + }, + { + "type": "image", + "bbox": [ + 289, + 594, + 500, + 684 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 289, + 594, + 500, + 684 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 289, + 594, + 500, + 684 + ], + "spans": [ + { + "bbox": [ + 289, + 594, + 500, + 684 + ], + "score": 0.967, + "type": "image", + "image_path": "c50466fd3830e13b4efa16578edae345ae90199f2967bef1d51b49352ce01705.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 289, + 594, + 500, + 605.25 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 289, + 605.25, + 500, + 616.5 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 289, + 616.5, + 500, + 627.75 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 289, + 627.75, + 500, + 639.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 289, + 639.0, + 500, + 650.25 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 289, + 650.25, + 500, + 661.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 289, + 661.5, + 500, + 672.75 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 289, + 672.75, + 500, + 684.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 284, + 696, + 505, + 717 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 284, + 696, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 284, + 696, + 505, + 708 + ], + "score": 1.0, + "content": "Figure G.7: Unseen 56 validation tasks, labeled with V, used", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 284, + 707, + 392, + 718 + ], + "spans": [ + { + "bbox": [ + 284, + 707, + 392, + 718 + ], + "score": 1.0, + "content": "for hyperparameter selection.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "index": 36.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 119, + 509, + 201 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 506, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "score": 1.0, + "content": "Table G.3: Optimal hyperparameters for reporting results in Table 1. These hyperparameters are selected using", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 90, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 90, + 505, + 102 + ], + "score": 1.0, + "content": "the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. All grid", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 100, + 297, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 297, + 111 + ], + "score": 1.0, + "content": "configurations in Table 1 use these hyperparameters.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 106, + 119, + 509, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 119, + 509, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 509, + 201 + ], + "score": 0.979, + "html": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-33.2×10-37×10-32.6×10-3
l2-reg. coefficient4.3×10-41×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)11.05×10-1
Auxiliary loss coefficient (α)11×10115.0
", + "type": "table", + "image_path": "857d0992e2052ecea10fe6587da7814aadcd89ca25d72bfa8443ab2ae5e7762f.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 119, + 509, + 146.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 146.33333333333334, + 509, + 173.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 173.66666666666669, + 509, + 201.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 110, + 252, + 499, + 334 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 222, + 504, + 244 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "Table G.4: Optimal hyperparameters for reporting results in Figure 5.3. These hyperparameters are selected", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 232, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 476, + 245 + ], + "score": 1.0, + "content": "using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 252, + 499, + 334 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 252, + 499, + 334 + ], + "spans": [ + { + "bbox": [ + 110, + 252, + 499, + 334 + ], + "score": 0.978, + "html": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-36×10-35×10-32.6×10-3
l2-reg. coefficient4.3×10-47×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)15×10-115×10-1
Auxiliary loss coefficient (α)15.015.0
", + "type": "table", + "image_path": "cea865ec2db0c3f5d013c54414495264361f7e8172a2396a85bd7388ab250bdc.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 252, + 499, + 279.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 279.3333333333333, + 499, + 306.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 306.66666666666663, + 499, + 333.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 118, + 383, + 490, + 459 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 354, + 504, + 375 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 354, + 504, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 504, + 365 + ], + "score": 1.0, + "content": "Table G.5: Optimal hyperparameters for reporting ablation results in Table 2. These hyperparameters are selected", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 364, + 476, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 476, + 376 + ], + "score": 1.0, + "content": "using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "table_body", + "bbox": [ + 118, + 383, + 490, + 459 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 118, + 383, + 490, + 459 + ], + "spans": [ + { + "bbox": [ + 118, + 383, + 490, + 459 + ], + "score": 0.982, + "html": "
HyperparameterPSMπ*-bisimulation
CMEsl2-embeddingsCMEsl2-embeddings
Learning Rate4×10-35×10-44.7×10-41×10-4
Contrastive Temperature (1/入)1.015×10-11
Auxiliary loss coefficient (α)5.01×10-11×10-11×10-6
", + "type": "table", + "image_path": "06ec11b3bfb0af64be94de637f254be00cc9b2c1aefaac1af5da15fd92dff8a0.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 118, + 383, + 490, + 408.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 118, + 408.3333333333333, + 490, + 433.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 118, + 433.66666666666663, + 490, + 458.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 506, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 486 + ], + "score": 1.0, + "content": "Please note that Table G.3 and Table G.4 correspond to two different tasks: one uses the standard", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "jumping task with white obstacles, while the other uses colored obstacles where the optimal policies", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "depend on color. For fair comparison, we tune hyperparameters for all the methods using Bayesian", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "optimization (Golovin et al., 2017). We use the best parameters among these tuned hyperparameters", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 519, + 507, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 507, + 531 + ], + "score": 1.0, + "content": "and the ones found in Table G.3, leading to different parameters for both PSEs as well as RandConv.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Evaluating PSEs with the jumping task hyperparameters from Table G.3 instead of the ones in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 540, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 235, + 553 + ], + "score": 1.0, + "content": "Table G.4 leads to a small drop", + "type": "text" + }, + { + "bbox": [ + 235, + 541, + 259, + 552 + ], + "score": 0.87, + "content": "( - 4 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 540, + 507, + 553 + ], + "score": 1.0, + "content": "on the jumping task with colors (Section 5.3). Nevertheless,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 551, + 315, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 315, + 564 + ], + "score": 1.0, + "content": "PSEs still outperform other methods in Section 5.3.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 108, + 579, + 276, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 278, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 278, + 595 + ], + "score": 1.0, + "content": "H LQR: ADDITIONAL DETAILS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "Optimal control with linear dynamics and quadratic cost, commonly known as LQR, has been", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "increasingly used as a simplified surrogate for deep RL problems (Recht, 2019). Following Song et al.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 626, + 493, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 493, + 640 + ], + "score": 1.0, + "content": "(2019); Sonar et al. (2020), we analyze the following LQR problem for assessing generalization:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 642, + 429, + 683 + ], + "lines": [ + { + "bbox": [ + 179, + 642, + 429, + 683 + ], + "spans": [ + { + "bbox": [ + 179, + 642, + 429, + 683 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathrm { m i n i m i z e } \\quad E _ { s _ { 0 } \\sim \\mathcal { D } } \\left[ \\frac { 1 } { 2 } \\sum _ { t = 0 } ^ { \\infty } s _ { t } ^ { T } Q s _ { t } + a _ { t } ^ { T } R a _ { t } \\right] , } \\\\ & { \\mathrm { s u b j e c t \\ t o } \\quad s _ { t + 1 } = A s _ { t } + B a _ { t } , o _ { t } = \\left[ \\begin{array} { l } { 0 . 1 W _ { c } } \\\\ { W _ { d } } \\end{array} \\right] s _ { t } , a _ { t } = K o _ { t } , } \\end{array}", + "type": "interline_equation", + "image_path": "aa837e6fd580fa352b89e2b6bf422ef74b965af98b4dcb5124af5c96245c3d33.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 179, + 642, + 429, + 655.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 179, + 655.6666666666666, + 429, + 669.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 179, + 669.3333333333333, + 429, + 682.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 687, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 133, + 700 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 688, + 143, + 698 + ], + "score": 0.8, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 687, + 270, + 700 + ], + "score": 1.0, + "content": "is the initial state distribution,", + "type": "text" + }, + { + "bbox": [ + 270, + 688, + 309, + 699 + ], + "score": 0.91, + "content": "s _ { t } \\in \\mathbb { R } ^ { n _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 687, + 453, + 700 + ], + "score": 1.0, + "content": "is the (hidden) true state at time t,", + "type": "text" + }, + { + "bbox": [ + 453, + 688, + 493, + 699 + ], + "score": 0.91, + "content": "a _ { t } \\in \\mathbb { R } ^ { n _ { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 201, + 712 + ], + "score": 1.0, + "content": "the control action, and", + "type": "text" + }, + { + "bbox": [ + 201, + 699, + 212, + 709 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 698, + 493, + 712 + ], + "score": 1.0, + "content": "is the linear policy matrix. The agent receives the input observation", + "type": "text" + }, + { + "bbox": [ + 493, + 701, + 502, + 710 + ], + "score": 0.78, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 698, + 507, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 281, + 722 + ], + "score": 1.0, + "content": "which is a linear transformation of the state", + "type": "text" + }, + { + "bbox": [ + 281, + 711, + 290, + 721 + ], + "score": 0.73, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 709, + 295, + 722 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 295, + 710, + 310, + 721 + ], + "score": 0.81, + "content": "W _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 709, + 328, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 710, + 343, + 721 + ], + "score": 0.89, + "content": "W _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "are semi-orthogonal matrices to prevent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "information loss for predicting optimal actions. An environment corresponds to a particular choice", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "23", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 119, + 509, + 201 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 506, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 504, + 92 + ], + "score": 1.0, + "content": "Table G.3: Optimal hyperparameters for reporting results in Table 1. These hyperparameters are selected using", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 90, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 90, + 505, + 102 + ], + "score": 1.0, + "content": "the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks. All grid", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 100, + 297, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 297, + 111 + ], + "score": 1.0, + "content": "configurations in Table 1 use these hyperparameters.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 106, + 119, + 509, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 119, + 509, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 509, + 201 + ], + "score": 0.979, + "html": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-33.2×10-37×10-32.6×10-3
l2-reg. coefficient4.3×10-41×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)11.05×10-1
Auxiliary loss coefficient (α)11×10115.0
", + "type": "table", + "image_path": "857d0992e2052ecea10fe6587da7814aadcd89ca25d72bfa8443ab2ae5e7762f.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 119, + 509, + 146.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 146.33333333333334, + 509, + 173.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 173.66666666666669, + 509, + 201.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 110, + 252, + 499, + 334 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 222, + 504, + 244 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "Table G.4: Optimal hyperparameters for reporting results in Figure 5.3. These hyperparameters are selected", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 232, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 476, + 245 + ], + "score": 1.0, + "content": "using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 252, + 499, + 334 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 252, + 499, + 334 + ], + "spans": [ + { + "bbox": [ + 110, + 252, + 499, + 334 + ], + "score": 0.978, + "html": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-36×10-35×10-32.6×10-3
l2-reg. coefficient4.3×10-47×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)15×10-115×10-1
Auxiliary loss coefficient (α)15.015.0
", + "type": "table", + "image_path": "cea865ec2db0c3f5d013c54414495264361f7e8172a2396a85bd7388ab250bdc.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 252, + 499, + 279.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 279.3333333333333, + 499, + 306.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 306.66666666666663, + 499, + 333.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 118, + 383, + 490, + 459 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 354, + 504, + 375 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 354, + 504, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 504, + 365 + ], + "score": 1.0, + "content": "Table G.5: Optimal hyperparameters for reporting ablation results in Table 2. These hyperparameters are selected", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 364, + 476, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 476, + 376 + ], + "score": 1.0, + "content": "using the “wide” grid by maximizing final performance on a validation set containing 56 unseen tasks.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "table_body", + "bbox": [ + 118, + 383, + 490, + 459 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 118, + 383, + 490, + 459 + ], + "spans": [ + { + "bbox": [ + 118, + 383, + 490, + 459 + ], + "score": 0.982, + "html": "
HyperparameterPSMπ*-bisimulation
CMEsl2-embeddingsCMEsl2-embeddings
Learning Rate4×10-35×10-44.7×10-41×10-4
Contrastive Temperature (1/入)1.015×10-11
Auxiliary loss coefficient (α)5.01×10-11×10-11×10-6
", + "type": "table", + "image_path": "06ec11b3bfb0af64be94de637f254be00cc9b2c1aefaac1af5da15fd92dff8a0.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 118, + 383, + 490, + 408.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 118, + 408.3333333333333, + 490, + 433.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 118, + 433.66666666666663, + 490, + 458.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 506, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 486 + ], + "score": 1.0, + "content": "Please note that Table G.3 and Table G.4 correspond to two different tasks: one uses the standard", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "jumping task with white obstacles, while the other uses colored obstacles where the optimal policies", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "depend on color. For fair comparison, we tune hyperparameters for all the methods using Bayesian", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "optimization (Golovin et al., 2017). We use the best parameters among these tuned hyperparameters", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 519, + 507, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 507, + 531 + ], + "score": 1.0, + "content": "and the ones found in Table G.3, leading to different parameters for both PSEs as well as RandConv.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Evaluating PSEs with the jumping task hyperparameters from Table G.3 instead of the ones in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 540, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 235, + 553 + ], + "score": 1.0, + "content": "Table G.4 leads to a small drop", + "type": "text" + }, + { + "bbox": [ + 235, + 541, + 259, + 552 + ], + "score": 0.87, + "content": "( - 4 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 540, + 507, + 553 + ], + "score": 1.0, + "content": "on the jumping task with colors (Section 5.3). Nevertheless,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 551, + 315, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 315, + 564 + ], + "score": 1.0, + "content": "PSEs still outperform other methods in Section 5.3.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 475, + 507, + 564 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 579, + 276, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 278, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 278, + 595 + ], + "score": 1.0, + "content": "H LQR: ADDITIONAL DETAILS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "Optimal control with linear dynamics and quadratic cost, commonly known as LQR, has been", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "increasingly used as a simplified surrogate for deep RL problems (Recht, 2019). Following Song et al.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 626, + 493, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 493, + 640 + ], + "score": 1.0, + "content": "(2019); Sonar et al. (2020), we analyze the following LQR problem for assessing generalization:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 604, + 506, + 640 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 642, + 429, + 683 + ], + "lines": [ + { + "bbox": [ + 179, + 642, + 429, + 683 + ], + "spans": [ + { + "bbox": [ + 179, + 642, + 429, + 683 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathrm { m i n i m i z e } \\quad E _ { s _ { 0 } \\sim \\mathcal { D } } \\left[ \\frac { 1 } { 2 } \\sum _ { t = 0 } ^ { \\infty } s _ { t } ^ { T } Q s _ { t } + a _ { t } ^ { T } R a _ { t } \\right] , } \\\\ & { \\mathrm { s u b j e c t \\ t o } \\quad s _ { t + 1 } = A s _ { t } + B a _ { t } , o _ { t } = \\left[ \\begin{array} { l } { 0 . 1 W _ { c } } \\\\ { W _ { d } } \\end{array} \\right] s _ { t } , a _ { t } = K o _ { t } , } \\end{array}", + "type": "interline_equation", + "image_path": "aa837e6fd580fa352b89e2b6bf422ef74b965af98b4dcb5124af5c96245c3d33.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 179, + 642, + 429, + 655.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 179, + 655.6666666666666, + 429, + 669.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 179, + 669.3333333333333, + 429, + 682.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 687, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 133, + 700 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 688, + 143, + 698 + ], + "score": 0.8, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 687, + 270, + 700 + ], + "score": 1.0, + "content": "is the initial state distribution,", + "type": "text" + }, + { + "bbox": [ + 270, + 688, + 309, + 699 + ], + "score": 0.91, + "content": "s _ { t } \\in \\mathbb { R } ^ { n _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 687, + 453, + 700 + ], + "score": 1.0, + "content": "is the (hidden) true state at time t,", + "type": "text" + }, + { + "bbox": [ + 453, + 688, + 493, + 699 + ], + "score": 0.91, + "content": "a _ { t } \\in \\mathbb { R } ^ { n _ { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 201, + 712 + ], + "score": 1.0, + "content": "the control action, and", + "type": "text" + }, + { + "bbox": [ + 201, + 699, + 212, + 709 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 698, + 493, + 712 + ], + "score": 1.0, + "content": "is the linear policy matrix. The agent receives the input observation", + "type": "text" + }, + { + "bbox": [ + 493, + 701, + 502, + 710 + ], + "score": 0.78, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 698, + 507, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 281, + 722 + ], + "score": 1.0, + "content": "which is a linear transformation of the state", + "type": "text" + }, + { + "bbox": [ + 281, + 711, + 290, + 721 + ], + "score": 0.73, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 709, + 295, + 722 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 295, + 710, + 310, + 721 + ], + "score": 0.81, + "content": "W _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 709, + 328, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 710, + 343, + 721 + ], + "score": 0.89, + "content": "W _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "are semi-orthogonal matrices to prevent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "information loss for predicting optimal actions. An environment corresponds to a particular choice", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 118, + 232 + ], + "score": 1.0, + "content": "of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 118, + 220, + 133, + 231 + ], + "score": 0.88, + "content": "W _ { d }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 134, + 219, + 255, + 232 + ], + "score": 1.0, + "content": "; all other system parameters", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 255, + 219, + 326, + 231 + ], + "score": 0.91, + "content": "( A , B , Q , R , W _ { c } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 326, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "are fixed matrices which are shared across", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 229, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 421, + 244 + ], + "score": 1.0, + "content": "environments and unknown to the agent. The agent learns the policy matrix", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 422, + 231, + 433, + 241 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 433, + 229, + 459, + 244 + ], + "score": 1.0, + "content": "using", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 459, + 231, + 469, + 241 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 470, + 229, + 505, + 244 + ], + "score": 1.0, + "content": "training", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "environments based on Equation H.1. At test time, the learned policy is evaluated on environments", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 252, + 176, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 156, + 264 + ], + "score": 1.0, + "content": "with unseen", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 156, + 253, + 172, + 263 + ], + "score": 0.89, + "content": "W _ { d }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 172, + 252, + 176, + 264 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 687, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 119, + 471, + 205 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 91 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 91 + ], + "score": 1.0, + "content": "Table H.1: LQR generalization performance: Absolute error in LQR cost, w.r.t. the oracle solver (which has", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 506, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 293, + 101 + ], + "score": 1.0, + "content": "access to true state), of various methods trained with", + "type": "text" + }, + { + "bbox": [ + 294, + 91, + 305, + 100 + ], + "score": 0.85, + "content": "n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 90, + 356, + 101 + ], + "score": 1.0, + "content": "distractors on", + "type": "text" + }, + { + "bbox": [ + 356, + 91, + 383, + 99 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 90, + 506, + 101 + ], + "score": 1.0, + "content": "environments. The reported mean", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 100, + 381, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 381, + 111 + ], + "score": 1.0, + "content": "and standard deviations are across 100 different seeds. Lower error is better.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 137, + 119, + 471, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 119, + 471, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 119, + 471, + 205 + ], + "score": 0.979, + "html": "
MethodNumber of Distractors (nd)
500100010000
Overparametrization (Song et al., 2019)25.8 (1.5)24.9 (1.1)24.9 (0.4)
IPO (Sonar et al., 202O) (IRM + Policy opt.)32.6 (5.0)27.3 (2.8)24.8 (0.4)
Weight Sparsity (ll-reg.)28.2 (0.0)28.2 (0.0)28.2 (0.0)
PSM (State aggregation)0.03 (0.0)0.03 (0.0)0.02 (0.0)
", + "type": "table", + "image_path": "caafa97b2aa45cc81bd6aca58750d161d74f297e6d88f6d11513e838fea19323.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 137, + 119, + 471, + 147.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 137, + 147.66666666666666, + 471, + 176.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 137, + 176.33333333333331, + 471, + 204.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 118, + 232 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 220, + 133, + 231 + ], + "score": 0.88, + "content": "W _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 219, + 255, + 232 + ], + "score": 1.0, + "content": "; all other system parameters", + "type": "text" + }, + { + "bbox": [ + 255, + 219, + 326, + 231 + ], + "score": 0.91, + "content": "( A , B , Q , R , W _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "are fixed matrices which are shared across", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 229, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 421, + 244 + ], + "score": 1.0, + "content": "environments and unknown to the agent. The agent learns the policy matrix", + "type": "text" + }, + { + "bbox": [ + 422, + 231, + 433, + 241 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 229, + 459, + 244 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 459, + 231, + 469, + 241 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 229, + 505, + 244 + ], + "score": 1.0, + "content": "training", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "environments based on Equation H.1. At test time, the learned policy is evaluated on environments", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 252, + 176, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 156, + 264 + ], + "score": 1.0, + "content": "with unseen", + "type": "text" + }, + { + "bbox": [ + 156, + 253, + 172, + 263 + ], + "score": 0.89, + "content": "W _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 252, + 176, + 264 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 392, + 281 + ], + "score": 1.0, + "content": "The generalization challenge in this setup is to ignore the distractors:", + "type": "text" + }, + { + "bbox": [ + 392, + 270, + 444, + 280 + ], + "score": 0.93, + "content": "W _ { c } s _ { t } \\in \\mathbb { R } ^ { n _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 310, + 292 + ], + "score": 1.0, + "content": "state features invariant across environments while", + "type": "text" + }, + { + "bbox": [ + 310, + 280, + 362, + 291 + ], + "score": 0.92, + "content": "W _ { d } s _ { t } \\in \\mathbb { R } ^ { n _ { d } }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "is a high-dimensional distractor of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 125, + 304 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 125, + 293, + 152, + 303 + ], + "score": 0.89, + "content": "n _ { s } , n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 291, + 249, + 304 + ], + "score": 1.0, + "content": ", respectively, such that", + "type": "text" + }, + { + "bbox": [ + 249, + 292, + 294, + 302 + ], + "score": 0.89, + "content": "n _ { s } < < n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 291, + 505, + 304 + ], + "score": 1.0, + "content": ". Furthermore, the policy matrix which generalizes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 301, + 504, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 210, + 324 + ], + "score": 1.0, + "content": "across all environments is", + "type": "text" + }, + { + "bbox": [ + 211, + 303, + 297, + 330 + ], + "score": 0.9, + "content": "K _ { \\star } = \\bigg [ 1 0 W _ { c } P _ { \\star } ^ { ~ T } \\bigg ] ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 301, + 328, + 324 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 328, + 312, + 340, + 322 + ], + "score": 0.86, + "content": "P _ { \\star }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 301, + 504, + 324 + ], + "score": 1.0, + "content": "corresponds to the optimal LQR solution", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 188, + 341 + ], + "score": 1.0, + "content": "with access to state", + "type": "text" + }, + { + "bbox": [ + 188, + 331, + 197, + 339 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 328, + 409, + 341 + ], + "score": 1.0, + "content": ". However, for a single environment with distractor", + "type": "text" + }, + { + "bbox": [ + 410, + 329, + 425, + 339 + ], + "score": 0.88, + "content": "W _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 328, + 505, + 341 + ], + "score": 1.0, + "content": ", multiple solutions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 339, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 183, + 362 + ], + "score": 1.0, + "content": "exist, for instance,", + "type": "text" + }, + { + "bbox": [ + 183, + 339, + 351, + 368 + ], + "score": 0.91, + "content": "K _ { \\star } { } ^ { \\prime } = \\left[ \\begin{array} { c c } { 1 0 \\alpha W _ { c } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\\\ { ( 1 - \\alpha ) W _ { d } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\end{array} \\right] ^ { T } \\quad \\forall \\alpha \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 349, + 505, + 361 + ], + "score": 1.0, + "content": ". Note that the distractors are an order", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 306, + 379 + ], + "score": 1.0, + "content": "of magnitude larger than the invariant features in", + "type": "text" + }, + { + "bbox": [ + 306, + 369, + 316, + 378 + ], + "score": 0.83, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "and dependence on them is likely to cause the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 377, + 464, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 464, + 390 + ], + "score": 1.0, + "content": "agent to act erratically on inputs with unseen distractors, resulting in poor generalization.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 344, + 406 + ], + "score": 1.0, + "content": "We use overparametrized policies with two linear layers, i.e.,", + "type": "text" + }, + { + "bbox": [ + 344, + 394, + 393, + 405 + ], + "score": 0.91, + "content": "K = K _ { 1 } K _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 394, + 423, + 406 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 423, + 394, + 450, + 406 + ], + "score": 0.92, + "content": "K _ { 1 } ( o )", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "is the learned", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 232, + 418 + ], + "score": 1.0, + "content": "representation for observation", + "type": "text" + }, + { + "bbox": [ + 232, + 408, + 238, + 415 + ], + "score": 0.63, + "content": "o", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 405, + 283, + 418 + ], + "score": 1.0, + "content": ". We learn", + "type": "text" + }, + { + "bbox": [ + 283, + 406, + 294, + 416 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "using gradient descent using the combined cost on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "2 training environments with varying number of distractors. We aggregate observation pairs with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "near-zero PSM by matching their representations using a squared loss. We use the open-source code", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 437, + 314, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 314, + 451 + ], + "score": 1.0, + "content": "released by Sonar et al. (2020) for our experiments.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 196, + 480, + 416, + 592 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 205, + 460, + 404, + 472 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 204, + 459, + 406, + 473 + ], + "spans": [ + { + "bbox": [ + 204, + 459, + 406, + 473 + ], + "score": 1.0, + "content": "Table H.2: An overview of hyper-parameters for LQR.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "table_body", + "bbox": [ + 196, + 480, + 416, + 592 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 196, + 480, + 416, + 592 + ], + "spans": [ + { + "bbox": [ + 196, + 480, + 416, + 592 + ], + "score": 0.979, + "html": "
ParameterSeting
AOrthogonal matrix, scaled 0.8
BI20×20
nx20
na20
QI20×20
RI20×20
KiAiOrthogonal Initialization, scaled 0.001
WdRandom semi-orthogonal matrix
", + "type": "table", + "image_path": "3940a54e5cdfb949bb831f0567a12d25d62c13a9fb0f20c25a21209eed78195a.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 480, + 416, + 494.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 196, + 494.0, + 416, + 508.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 196, + 508.0, + 416, + 522.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 196, + 522.0, + 416, + 536.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 196, + 536.0, + 416, + 550.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 196, + 550.0, + 416, + 564.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 196, + 564.0, + 416, + 578.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 196, + 578.0, + 416, + 592.0 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 25.25 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "The reliance on distractors for IPO also highlights a limitation of IRM: if a model can achieve", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "a solution with zero training error, then any such solution is acceptable by IRM regardless of its", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 626, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 638 + ], + "score": 1.0, + "content": "generalization ability – a common scenario with overparametrized deep neural nets (Jin et al., 2020).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 107, + 651, + 321, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 321, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 321, + 664 + ], + "score": 1.0, + "content": "H.1 NEAR-OPTIMALITY OF PSM AGGREGATION", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 673, + 503, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 672, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 687 + ], + "score": 1.0, + "content": "Conjecture 1. Assuming zero state aggregation error with policy similarity metric (PSM), the policy", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 683, + 405, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 135, + 697 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 135, + 684, + 146, + 694 + ], + "score": 0.62, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 683, + 405, + 697 + ], + "score": 1.0, + "content": "learned using gradient descent is independent of the distractors.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 209, + 723 + ], + "score": 1.0, + "content": "Proof. For LQR domains", + "type": "text" + }, + { + "bbox": [ + 209, + 712, + 226, + 721 + ], + "score": 0.83, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 708, + 307, + 723 + ], + "score": 1.0, + "content": ", an observation pair", + "type": "text" + }, + { + "bbox": [ + 307, + 710, + 345, + 721 + ], + "score": 0.88, + "content": "( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 708, + 494, + 723 + ], + "score": 1.0, + "content": "has zero PSM iff the underlying state", + "type": "text" + }, + { + "bbox": [ + 494, + 712, + 504, + 721 + ], + "score": 0.81, + "content": "s _ { t }", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "is same for both the observations in the pair. This is true, as (a) both domains has the same transition", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 119, + 471, + 205 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 91 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 91 + ], + "score": 1.0, + "content": "Table H.1: LQR generalization performance: Absolute error in LQR cost, w.r.t. the oracle solver (which has", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 506, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 293, + 101 + ], + "score": 1.0, + "content": "access to true state), of various methods trained with", + "type": "text" + }, + { + "bbox": [ + 294, + 91, + 305, + 100 + ], + "score": 0.85, + "content": "n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 90, + 356, + 101 + ], + "score": 1.0, + "content": "distractors on", + "type": "text" + }, + { + "bbox": [ + 356, + 91, + 383, + 99 + ], + "score": 0.9, + "content": "N = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 90, + 506, + 101 + ], + "score": 1.0, + "content": "environments. The reported mean", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 100, + 381, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 381, + 111 + ], + "score": 1.0, + "content": "and standard deviations are across 100 different seeds. Lower error is better.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 137, + 119, + 471, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 119, + 471, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 119, + 471, + 205 + ], + "score": 0.979, + "html": "
MethodNumber of Distractors (nd)
500100010000
Overparametrization (Song et al., 2019)25.8 (1.5)24.9 (1.1)24.9 (0.4)
IPO (Sonar et al., 202O) (IRM + Policy opt.)32.6 (5.0)27.3 (2.8)24.8 (0.4)
Weight Sparsity (ll-reg.)28.2 (0.0)28.2 (0.0)28.2 (0.0)
PSM (State aggregation)0.03 (0.0)0.03 (0.0)0.02 (0.0)
", + "type": "table", + "image_path": "caafa97b2aa45cc81bd6aca58750d161d74f297e6d88f6d11513e838fea19323.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 137, + 119, + 471, + 147.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 137, + 147.66666666666666, + 471, + 176.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 137, + 176.33333333333331, + 471, + 204.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 505, + 264 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 219, + 506, + 264 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 392, + 281 + ], + "score": 1.0, + "content": "The generalization challenge in this setup is to ignore the distractors:", + "type": "text" + }, + { + "bbox": [ + 392, + 270, + 444, + 280 + ], + "score": 0.93, + "content": "W _ { c } s _ { t } \\in \\mathbb { R } ^ { n _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "represents the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 310, + 292 + ], + "score": 1.0, + "content": "state features invariant across environments while", + "type": "text" + }, + { + "bbox": [ + 310, + 280, + 362, + 291 + ], + "score": 0.92, + "content": "W _ { d } s _ { t } \\in \\mathbb { R } ^ { n _ { d } }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "is a high-dimensional distractor of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 125, + 304 + ], + "score": 1.0, + "content": "size", + "type": "text" + }, + { + "bbox": [ + 125, + 293, + 152, + 303 + ], + "score": 0.89, + "content": "n _ { s } , n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 291, + 249, + 304 + ], + "score": 1.0, + "content": ", respectively, such that", + "type": "text" + }, + { + "bbox": [ + 249, + 292, + 294, + 302 + ], + "score": 0.89, + "content": "n _ { s } < < n _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 291, + 505, + 304 + ], + "score": 1.0, + "content": ". Furthermore, the policy matrix which generalizes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 301, + 504, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 210, + 324 + ], + "score": 1.0, + "content": "across all environments is", + "type": "text" + }, + { + "bbox": [ + 211, + 303, + 297, + 330 + ], + "score": 0.9, + "content": "K _ { \\star } = \\bigg [ 1 0 W _ { c } P _ { \\star } ^ { ~ T } \\bigg ] ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 301, + 328, + 324 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 328, + 312, + 340, + 322 + ], + "score": 0.86, + "content": "P _ { \\star }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 301, + 504, + 324 + ], + "score": 1.0, + "content": "corresponds to the optimal LQR solution", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 188, + 341 + ], + "score": 1.0, + "content": "with access to state", + "type": "text" + }, + { + "bbox": [ + 188, + 331, + 197, + 339 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 328, + 409, + 341 + ], + "score": 1.0, + "content": ". However, for a single environment with distractor", + "type": "text" + }, + { + "bbox": [ + 410, + 329, + 425, + 339 + ], + "score": 0.88, + "content": "W _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 328, + 505, + 341 + ], + "score": 1.0, + "content": ", multiple solutions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 339, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 183, + 362 + ], + "score": 1.0, + "content": "exist, for instance,", + "type": "text" + }, + { + "bbox": [ + 183, + 339, + 351, + 368 + ], + "score": 0.91, + "content": "K _ { \\star } { } ^ { \\prime } = \\left[ \\begin{array} { c c } { 1 0 \\alpha W _ { c } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\\\ { ( 1 - \\alpha ) W _ { d } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\end{array} \\right] ^ { T } \\quad \\forall \\alpha \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 349, + 505, + 361 + ], + "score": 1.0, + "content": ". Note that the distractors are an order", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 366, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 306, + 379 + ], + "score": 1.0, + "content": "of magnitude larger than the invariant features in", + "type": "text" + }, + { + "bbox": [ + 306, + 369, + 316, + 378 + ], + "score": 0.83, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 366, + 505, + 379 + ], + "score": 1.0, + "content": "and dependence on them is likely to cause the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 377, + 464, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 464, + 390 + ], + "score": 1.0, + "content": "agent to act erratically on inputs with unseen distractors, resulting in poor generalization.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 269, + 505, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 344, + 406 + ], + "score": 1.0, + "content": "We use overparametrized policies with two linear layers, i.e.,", + "type": "text" + }, + { + "bbox": [ + 344, + 394, + 393, + 405 + ], + "score": 0.91, + "content": "K = K _ { 1 } K _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 394, + 423, + 406 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 423, + 394, + 450, + 406 + ], + "score": 0.92, + "content": "K _ { 1 } ( o )", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "is the learned", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 232, + 418 + ], + "score": 1.0, + "content": "representation for observation", + "type": "text" + }, + { + "bbox": [ + 232, + 408, + 238, + 415 + ], + "score": 0.63, + "content": "o", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 405, + 283, + 418 + ], + "score": 1.0, + "content": ". We learn", + "type": "text" + }, + { + "bbox": [ + 283, + 406, + 294, + 416 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "using gradient descent using the combined cost on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "2 training environments with varying number of distractors. We aggregate observation pairs with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "near-zero PSM by matching their representations using a squared loss. We use the open-source code", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 437, + 314, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 314, + 451 + ], + "score": 1.0, + "content": "released by Sonar et al. (2020) for our experiments.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 394, + 505, + 451 + ] + }, + { + "type": "table", + "bbox": [ + 196, + 480, + 416, + 592 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 205, + 460, + 404, + 472 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 204, + 459, + 406, + 473 + ], + "spans": [ + { + "bbox": [ + 204, + 459, + 406, + 473 + ], + "score": 1.0, + "content": "Table H.2: An overview of hyper-parameters for LQR.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "table_body", + "bbox": [ + 196, + 480, + 416, + 592 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 196, + 480, + 416, + 592 + ], + "spans": [ + { + "bbox": [ + 196, + 480, + 416, + 592 + ], + "score": 0.979, + "html": "
ParameterSeting
AOrthogonal matrix, scaled 0.8
BI20×20
nx20
na20
QI20×20
RI20×20
KiAiOrthogonal Initialization, scaled 0.001
WdRandom semi-orthogonal matrix
", + "type": "table", + "image_path": "3940a54e5cdfb949bb831f0567a12d25d62c13a9fb0f20c25a21209eed78195a.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 480, + 416, + 494.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 196, + 494.0, + 416, + 508.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 196, + 508.0, + 416, + 522.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 196, + 522.0, + 416, + 536.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 196, + 536.0, + 416, + 550.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 196, + 550.0, + 416, + 564.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 196, + 564.0, + 416, + 578.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 196, + 578.0, + 416, + 592.0 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 25.25 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "The reliance on distractors for IPO also highlights a limitation of IRM: if a model can achieve", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "a solution with zero training error, then any such solution is acceptable by IRM regardless of its", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 626, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 638 + ], + "score": 1.0, + "content": "generalization ability – a common scenario with overparametrized deep neural nets (Jin et al., 2020).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 604, + 506, + 638 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 651, + 321, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 321, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 321, + 664 + ], + "score": 1.0, + "content": "H.1 NEAR-OPTIMALITY OF PSM AGGREGATION", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 673, + 503, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 672, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 687 + ], + "score": 1.0, + "content": "Conjecture 1. Assuming zero state aggregation error with policy similarity metric (PSM), the policy", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 683, + 405, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 135, + 697 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 135, + 684, + 146, + 694 + ], + "score": 0.62, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 683, + 405, + 697 + ], + "score": 1.0, + "content": "learned using gradient descent is independent of the distractors.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 672, + 505, + 697 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 209, + 723 + ], + "score": 1.0, + "content": "Proof. For LQR domains", + "type": "text" + }, + { + "bbox": [ + 209, + 712, + 226, + 721 + ], + "score": 0.83, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 708, + 307, + 723 + ], + "score": 1.0, + "content": ", an observation pair", + "type": "text" + }, + { + "bbox": [ + 307, + 710, + 345, + 721 + ], + "score": 0.88, + "content": "( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 708, + 494, + 723 + ], + "score": 1.0, + "content": "has zero PSM iff the underlying state", + "type": "text" + }, + { + "bbox": [ + 494, + 712, + 504, + 721 + ], + "score": 0.81, + "content": "s _ { t }", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "is same for both the observations in the pair. This is true, as (a) both domains has the same transition", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "dynamics, as specified by Equation H.1, and (b) the optimal policy is deterministic and is completely", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 299, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 232, + 106 + ], + "score": 1.0, + "content": "determined by the current state", + "type": "text" + }, + { + "bbox": [ + 232, + 96, + 242, + 105 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 93, + 289, + 106 + ], + "score": 1.0, + "content": "at any time", + "type": "text" + }, + { + "bbox": [ + 290, + 95, + 294, + 104 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 93, + 299, + 106 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 109, + 506, + 159 + ], + "lines": [ + { + "bbox": [ + 104, + 104, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 140, + 131 + ], + "score": 1.0, + "content": "Assume", + "type": "text" + }, + { + "bbox": [ + 140, + 109, + 216, + 136 + ], + "score": 0.91, + "content": "o _ { t } { } ^ { x } = \\left[ \\begin{array} { c } { 0 . 1 W _ { c } } \\\\ { W _ { d ^ { x } } } \\end{array} \\right] s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 104, + 234, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 234, + 109, + 311, + 136 + ], + "score": 0.92, + "content": "o _ { t } { } ^ { y } = \\left[ \\stackrel { 0 . 1 } { W } _ { d ^ { y } } \\right] s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 104, + 466, + 138 + ], + "score": 1.0, + "content": "for distractor semi-orthogonal matrices", + "type": "text" + }, + { + "bbox": [ + 466, + 118, + 487, + 128 + ], + "score": 0.91, + "content": "W _ { d ^ { x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 104, + 506, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 133, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 107, + 136, + 127, + 146 + ], + "score": 0.89, + "content": "W _ { d } y", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 133, + 363, + 148 + ], + "score": 1.0, + "content": ", respectively. Furthermore, the representation is given by", + "type": "text" + }, + { + "bbox": [ + 363, + 135, + 398, + 147 + ], + "score": 0.92, + "content": "K _ { 1 } ( o _ { t } ^ { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 133, + 417, + 148 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 417, + 135, + 452, + 146 + ], + "score": 0.91, + "content": "K _ { 1 } ( o _ { t } ^ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 133, + 506, + 148 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 145, + 478, + 159 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 158, + 159 + ], + "score": 1.0, + "content": "Assume that", + "type": "text" + }, + { + "bbox": [ + 159, + 147, + 227, + 159 + ], + "score": 0.91, + "content": "K _ { 1 } = [ K _ { s } \\quad K _ { d } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 145, + 255, + 159 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 255, + 146, + 308, + 158 + ], + "score": 0.93, + "content": "K _ { s } \\in \\mathbb { R } ^ { h \\times n _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 145, + 327, + 159 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 327, + 147, + 380, + 158 + ], + "score": 0.93, + "content": "K _ { d } \\in \\mathbb { R } ^ { h \\times n _ { d } }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 145, + 399, + 159 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 399, + 146, + 474, + 158 + ], + "score": 0.89, + "content": "K _ { 1 } \\in \\mathbb { R } ^ { h \\times ( n _ { s } + n _ { d } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 145, + 478, + 159 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 497, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 498, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 374, + 179 + ], + "score": 1.0, + "content": "Zero state-aggregation error with squared loss implies that for pair", + "type": "text" + }, + { + "bbox": [ + 374, + 165, + 413, + 176 + ], + "score": 0.89, + "content": "( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 162, + 484, + 179 + ], + "score": 1.0, + "content": "corresponding to", + "type": "text" + }, + { + "bbox": [ + 484, + 167, + 493, + 175 + ], + "score": 0.82, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 162, + 498, + 179 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 179, + 456, + 207 + ], + "lines": [ + { + "bbox": [ + 155, + 179, + 456, + 207 + ], + "spans": [ + { + "bbox": [ + 155, + 179, + 456, + 207 + ], + "score": 0.94, + "content": "K _ { 1 } ( o _ { t } { ^ x } - o _ { t } { ^ y } ) = K _ { 1 } [ W _ { d { ^ x } } - W _ { d { ^ y } } ] s _ { t } = 0 \\implies K _ { d } ( W _ { d { ^ x } } - W _ { d { ^ y } } ) s _ { t } = 0", + "type": "interline_equation", + "image_path": "88ac353b7950e601d0879780a90531fb54fce1de1a5d6ff0b4a4c0b3673b63c3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 155, + 179, + 456, + 188.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 188.33333333333334, + 456, + 197.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 197.66666666666669, + 456, + 207.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 210, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "As Equation H.2 holds for all states visited by the optimal policy in an infinite horizon LQR, it", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 249, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 156, + 234 + ], + "score": 1.0, + "content": "follows that", + "type": "text" + }, + { + "bbox": [ + 156, + 221, + 245, + 234 + ], + "score": 0.92, + "content": "K _ { d } ( W _ { d ^ { x } } - W _ { d ^ { y } } ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 221, + 249, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "Furthermore, it is well-known that gradient descent tends to find low-rank solutions due to implicit", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "regularization (Arora et al., 2019; Gunasekar et al., 2017), e.g.,with small enough step sizes and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "initialization close enough to the origin, gradient descent on matrix factorization converges to the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "minimum nuclear norm solution for 2 layer linear networks (Gunasekar et al., 2017). Based on this,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 182, + 295 + ], + "score": 1.0, + "content": "we conjecture that", + "type": "text" + }, + { + "bbox": [ + 182, + 283, + 214, + 293 + ], + "score": 0.92, + "content": "K _ { d } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 282, + 368, + 295 + ], + "score": 1.0, + "content": "which we found to be true in practice.", + "type": "text" + }, + { + "bbox": [ + 494, + 283, + 506, + 294 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 309, + 284, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 285, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 285, + 324 + ], + "score": 1.0, + "content": "I DISTRACTING CONTROL SUITE", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 504, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "We use the same setup as Kostrikov et al. (2020); Stone et al. (2021) for implementation details and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 345, + 375, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 375, + 357 + ], + "score": 1.0, + "content": "training protocol. For completeness, we describe the details below.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "Dynamic Background Distractions. In Distracting Control Suite (Stone et al., 2021), random", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "backgrounds are projected from scenes of the DAVIS 2017 dataset (Pont-Tuset et al., 2017) onto", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "the skybox of the scene. To make these backgrounds visible for all tasks and views, the floor grid", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 395, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 407 + ], + "score": 1.0, + "content": "is semi-transparent with transparency coefficient 0.3. We take the first 2 videos in the DAVIS 2017", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "training set and randomly sample a scene and a frame from those at the start of every episode. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "score": 1.0, + "content": "the dynamic setting, the video plays forwards or backwards until the last or first frame is reached at", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "which point the video is played backwards. This way, the background motion is always smooth and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 169, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 169, + 450 + ], + "score": 1.0, + "content": "without “cuts”.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 506, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "Soft Actor-Critic. Soft Actor-Critic (SAC) (Haarnoja et al., 2018) learns a state-action value function", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 466, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 107, + 467, + 120, + 478 + ], + "score": 0.88, + "content": "Q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 466, + 197, + 480 + ], + "score": 1.0, + "content": ", a stochastic policy", + "type": "text" + }, + { + "bbox": [ + 198, + 469, + 209, + 478 + ], + "score": 0.85, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 466, + 281, + 480 + ], + "score": 1.0, + "content": "and a temperature", + "type": "text" + }, + { + "bbox": [ + 281, + 469, + 289, + 477 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 466, + 436, + 480 + ], + "score": 1.0, + "content": "to find an optimal policy for an MDP", + "type": "text" + }, + { + "bbox": [ + 437, + 467, + 492, + 479 + ], + "score": 0.91, + "content": "( S , \\mathcal { A } , p , r , \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 466, + 506, + 480 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 157, + 491 + ], + "score": 1.0, + "content": "optimizing a", + "type": "text" + }, + { + "bbox": [ + 158, + 480, + 164, + 489 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 478, + 326, + 491 + ], + "score": 1.0, + "content": "-discounted maximum-entropy objective.", + "type": "text" + }, + { + "bbox": [ + 326, + 478, + 333, + 488 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "is used generically to denote the parameters", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 379, + 501 + ], + "score": 1.0, + "content": "updated through training in each part of the model. The actor policy", + "type": "text" + }, + { + "bbox": [ + 380, + 489, + 419, + 500 + ], + "score": 0.91, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "is a parametric tanh-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 185, + 513 + ], + "score": 1.0, + "content": "Gaussian that given", + "type": "text" + }, + { + "bbox": [ + 186, + 501, + 195, + 511 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 500, + 230, + 513 + ], + "score": 1.0, + "content": "samples", + "type": "text" + }, + { + "bbox": [ + 231, + 500, + 350, + 512 + ], + "score": 0.92, + "content": "a _ { t } \\dot { = } \\operatorname { t a n h } ( \\mu _ { \\theta } ( s _ { t } ) + \\sigma _ { \\theta } ( s _ { t } ) \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 500, + 379, + 513 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 380, + 500, + 430, + 512 + ], + "score": 0.92, + "content": "\\epsilon \\sim \\mathcal { N } ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 500, + 448, + 513 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 448, + 501, + 460, + 511 + ], + "score": 0.84, + "content": "\\mu _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 500, + 478, + 513 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 478, + 502, + 489, + 511 + ], + "score": 0.85, + "content": "\\sigma _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 271, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 271, + 523 + ], + "score": 1.0, + "content": "parametric mean and standard deviation.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 275, + 541 + ], + "score": 1.0, + "content": "The policy evaluation step learns the critic", + "type": "text" + }, + { + "bbox": [ + 275, + 528, + 317, + 540 + ], + "score": 0.93, + "content": "Q _ { \\theta } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "network by optimizing a single-step of the soft", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 538, + 177, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 177, + 550 + ], + "score": 1.0, + "content": "Bellman residual", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 553, + 415, + 594 + ], + "lines": [ + { + "bbox": [ + 195, + 553, + 415, + 594 + ], + "spans": [ + { + "bbox": [ + 195, + 553, + 415, + 594 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { J _ { Q } ( \\mathcal { D } ) = E _ { ( s _ { t } , a _ { t } , s _ { t } ^ { \\prime } ) \\sim \\mathcal { D } } [ ( Q _ { \\theta } ( s _ { t } , a _ { t } ) - y _ { t } ) ^ { 2 } ] } \\\\ & { \\qquad a _ { t } ^ { \\prime } { \\sim } \\pi ( \\cdot | s _ { t } ^ { \\prime } ) } \\\\ & { y _ { t } = r ( s _ { t } , a _ { t } ) + \\gamma [ Q _ { \\theta ^ { \\prime } } ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) - \\alpha \\log \\pi _ { \\theta } ( a _ { t } ^ { \\prime } | s _ { t } ^ { \\prime } ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "d967d8f701d309f8a2b9d1f8aed651c62ceb99e0c454aba8f437771b1a610922.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 195, + 553, + 415, + 566.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 195, + 566.6666666666666, + 415, + 580.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 195, + 580.3333333333333, + 415, + 593.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 596, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 133, + 610 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 597, + 142, + 607 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 597, + 272, + 610 + ], + "score": 1.0, + "content": "is a replay buffer of transitions,", + "type": "text" + }, + { + "bbox": [ + 273, + 597, + 281, + 607 + ], + "score": 0.85, + "content": "\\theta ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "is an exponential moving average of the weights. SAC", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 607, + 448, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 448, + 621 + ], + "score": 1.0, + "content": "uses clipped double-Q learning, which we omit for simplicity but employ in practice.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 503, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 320, + 639 + ], + "score": 1.0, + "content": "The policy improvement step then fits the actor policy", + "type": "text" + }, + { + "bbox": [ + 321, + 625, + 359, + 637 + ], + "score": 0.93, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 623, + 505, + 639 + ], + "score": 1.0, + "content": "network by optimizing the objective", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 641, + 412, + 664 + ], + "lines": [ + { + "bbox": [ + 198, + 641, + 412, + 664 + ], + "spans": [ + { + "bbox": [ + 198, + 641, + 412, + 664 + ], + "score": 0.92, + "content": "J _ { \\pi } ( \\mathcal { D } ) = E _ { s _ { t } \\sim \\mathcal { D } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\theta } ( \\cdot | s _ { t } ) | | \\exp \\{ \\frac { 1 } { \\alpha } Q _ { \\theta } ( s _ { t } , \\cdot ) \\} ) ] .", + "type": "interline_equation", + "image_path": "b8de65fbed10cf47b2b18c028ad713f54b791fcc6be2ce542047cddf3eb2f735.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 198, + 641, + 412, + 664 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 306, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 307, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 204, + 681 + ], + "score": 1.0, + "content": "Finally, the temperature", + "type": "text" + }, + { + "bbox": [ + 204, + 670, + 211, + 678 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 667, + 307, + 681 + ], + "score": 1.0, + "content": "is learned with the loss", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 683, + 402, + 705 + ], + "lines": [ + { + "bbox": [ + 208, + 683, + 402, + 705 + ], + "spans": [ + { + "bbox": [ + 208, + 683, + 402, + 705 + ], + "score": 0.93, + "content": "J _ { \\alpha } ( \\mathcal { D } ) = E _ { \\underset { a _ { t } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t } ) } { s _ { t } \\sim \\mathcal { D } } } [ - \\alpha \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) - \\alpha \\bar { \\mathcal { H } } ] ,", + "type": "interline_equation", + "image_path": "2cb81c1d6d5abb9132324b381e379d4fad3b4465a9273c43452d653b03c371b4.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 208, + 683, + 402, + 705 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 132, + 722 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 709, + 162, + 720 + ], + "score": 0.91, + "content": "\\bar { \\mathcal { H } } \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "is the target entropy hyper-parameter that the policy tries to match, which in practice is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 718, + 210, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 162, + 734 + ], + "score": 1.0, + "content": "usually set to", + "type": "text" + }, + { + "bbox": [ + 162, + 720, + 206, + 732 + ], + "score": 0.93, + "content": "\\bar { \\mathcal { H } } = - | \\bar { \\mathcal { A } } |", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 718, + 210, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 16, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "dynamics, as specified by Equation H.1, and (b) the optimal policy is deterministic and is completely", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 299, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 232, + 106 + ], + "score": 1.0, + "content": "determined by the current state", + "type": "text" + }, + { + "bbox": [ + 232, + 96, + 242, + 105 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 93, + 289, + 106 + ], + "score": 1.0, + "content": "at any time", + "type": "text" + }, + { + "bbox": [ + 290, + 95, + 294, + 104 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 93, + 299, + 106 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 109, + 506, + 159 + ], + "lines": [ + { + "bbox": [ + 104, + 104, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 140, + 131 + ], + "score": 1.0, + "content": "Assume", + "type": "text" + }, + { + "bbox": [ + 140, + 109, + 216, + 136 + ], + "score": 0.91, + "content": "o _ { t } { } ^ { x } = \\left[ \\begin{array} { c } { 0 . 1 W _ { c } } \\\\ { W _ { d ^ { x } } } \\end{array} \\right] s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 104, + 234, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 234, + 109, + 311, + 136 + ], + "score": 0.92, + "content": "o _ { t } { } ^ { y } = \\left[ \\stackrel { 0 . 1 } { W } _ { d ^ { y } } \\right] s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 104, + 466, + 138 + ], + "score": 1.0, + "content": "for distractor semi-orthogonal matrices", + "type": "text" + }, + { + "bbox": [ + 466, + 118, + 487, + 128 + ], + "score": 0.91, + "content": "W _ { d ^ { x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 104, + 506, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 133, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 107, + 136, + 127, + 146 + ], + "score": 0.89, + "content": "W _ { d } y", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 133, + 363, + 148 + ], + "score": 1.0, + "content": ", respectively. Furthermore, the representation is given by", + "type": "text" + }, + { + "bbox": [ + 363, + 135, + 398, + 147 + ], + "score": 0.92, + "content": "K _ { 1 } ( o _ { t } ^ { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 133, + 417, + 148 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 417, + 135, + 452, + 146 + ], + "score": 0.91, + "content": "K _ { 1 } ( o _ { t } ^ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 133, + 506, + 148 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 145, + 478, + 159 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 158, + 159 + ], + "score": 1.0, + "content": "Assume that", + "type": "text" + }, + { + "bbox": [ + 159, + 147, + 227, + 159 + ], + "score": 0.91, + "content": "K _ { 1 } = [ K _ { s } \\quad K _ { d } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 145, + 255, + 159 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 255, + 146, + 308, + 158 + ], + "score": 0.93, + "content": "K _ { s } \\in \\mathbb { R } ^ { h \\times n _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 145, + 327, + 159 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 327, + 147, + 380, + 158 + ], + "score": 0.93, + "content": "K _ { d } \\in \\mathbb { R } ^ { h \\times n _ { d } }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 145, + 399, + 159 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 399, + 146, + 474, + 158 + ], + "score": 0.89, + "content": "K _ { 1 } \\in \\mathbb { R } ^ { h \\times ( n _ { s } + n _ { d } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 145, + 478, + 159 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 104, + 506, + 159 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 497, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 498, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 374, + 179 + ], + "score": 1.0, + "content": "Zero state-aggregation error with squared loss implies that for pair", + "type": "text" + }, + { + "bbox": [ + 374, + 165, + 413, + 176 + ], + "score": 0.89, + "content": "( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 162, + 484, + 179 + ], + "score": 1.0, + "content": "corresponding to", + "type": "text" + }, + { + "bbox": [ + 484, + 167, + 493, + 175 + ], + "score": 0.82, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 162, + 498, + 179 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 162, + 498, + 179 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 179, + 456, + 207 + ], + "lines": [ + { + "bbox": [ + 155, + 179, + 456, + 207 + ], + "spans": [ + { + "bbox": [ + 155, + 179, + 456, + 207 + ], + "score": 0.94, + "content": "K _ { 1 } ( o _ { t } { ^ x } - o _ { t } { ^ y } ) = K _ { 1 } [ W _ { d { ^ x } } - W _ { d { ^ y } } ] s _ { t } = 0 \\implies K _ { d } ( W _ { d { ^ x } } - W _ { d { ^ y } } ) s _ { t } = 0", + "type": "interline_equation", + "image_path": "88ac353b7950e601d0879780a90531fb54fce1de1a5d6ff0b4a4c0b3673b63c3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 155, + 179, + 456, + 188.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 188.33333333333334, + 456, + 197.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 197.66666666666669, + 456, + 207.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 210, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "As Equation H.2 holds for all states visited by the optimal policy in an infinite horizon LQR, it", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 249, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 156, + 234 + ], + "score": 1.0, + "content": "follows that", + "type": "text" + }, + { + "bbox": [ + 156, + 221, + 245, + 234 + ], + "score": 0.92, + "content": "K _ { d } ( W _ { d ^ { x } } - W _ { d ^ { y } } ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 221, + 249, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 210, + 505, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "Furthermore, it is well-known that gradient descent tends to find low-rank solutions due to implicit", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "regularization (Arora et al., 2019; Gunasekar et al., 2017), e.g.,with small enough step sizes and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "initialization close enough to the origin, gradient descent on matrix factorization converges to the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "minimum nuclear norm solution for 2 layer linear networks (Gunasekar et al., 2017). Based on this,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 182, + 295 + ], + "score": 1.0, + "content": "we conjecture that", + "type": "text" + }, + { + "bbox": [ + 182, + 283, + 214, + 293 + ], + "score": 0.92, + "content": "K _ { d } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 282, + 368, + 295 + ], + "score": 1.0, + "content": "which we found to be true in practice.", + "type": "text" + }, + { + "bbox": [ + 494, + 283, + 506, + 294 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 238, + 506, + 295 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 309, + 284, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 285, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 285, + 324 + ], + "score": 1.0, + "content": "I DISTRACTING CONTROL SUITE", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 504, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "We use the same setup as Kostrikov et al. (2020); Stone et al. (2021) for implementation details and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 345, + 375, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 375, + 357 + ], + "score": 1.0, + "content": "training protocol. For completeness, we describe the details below.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 334, + 505, + 357 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "Dynamic Background Distractions. In Distracting Control Suite (Stone et al., 2021), random", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "backgrounds are projected from scenes of the DAVIS 2017 dataset (Pont-Tuset et al., 2017) onto", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "the skybox of the scene. To make these backgrounds visible for all tasks and views, the floor grid", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 395, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 407 + ], + "score": 1.0, + "content": "is semi-transparent with transparency coefficient 0.3. We take the first 2 videos in the DAVIS 2017", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "training set and randomly sample a scene and a frame from those at the start of every episode. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "score": 1.0, + "content": "the dynamic setting, the video plays forwards or backwards until the last or first frame is reached at", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "which point the video is played backwards. This way, the background motion is always smooth and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 439, + 169, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 169, + 450 + ], + "score": 1.0, + "content": "without “cuts”.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 362, + 506, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 506, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "Soft Actor-Critic. Soft Actor-Critic (SAC) (Haarnoja et al., 2018) learns a state-action value function", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 466, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 107, + 467, + 120, + 478 + ], + "score": 0.88, + "content": "Q _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 466, + 197, + 480 + ], + "score": 1.0, + "content": ", a stochastic policy", + "type": "text" + }, + { + "bbox": [ + 198, + 469, + 209, + 478 + ], + "score": 0.85, + "content": "\\pi _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 466, + 281, + 480 + ], + "score": 1.0, + "content": "and a temperature", + "type": "text" + }, + { + "bbox": [ + 281, + 469, + 289, + 477 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 466, + 436, + 480 + ], + "score": 1.0, + "content": "to find an optimal policy for an MDP", + "type": "text" + }, + { + "bbox": [ + 437, + 467, + 492, + 479 + ], + "score": 0.91, + "content": "( S , \\mathcal { A } , p , r , \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 466, + 506, + 480 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 157, + 491 + ], + "score": 1.0, + "content": "optimizing a", + "type": "text" + }, + { + "bbox": [ + 158, + 480, + 164, + 489 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 478, + 326, + 491 + ], + "score": 1.0, + "content": "-discounted maximum-entropy objective.", + "type": "text" + }, + { + "bbox": [ + 326, + 478, + 333, + 488 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "is used generically to denote the parameters", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 379, + 501 + ], + "score": 1.0, + "content": "updated through training in each part of the model. The actor policy", + "type": "text" + }, + { + "bbox": [ + 380, + 489, + 419, + 500 + ], + "score": 0.91, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "is a parametric tanh-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 500, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 185, + 513 + ], + "score": 1.0, + "content": "Gaussian that given", + "type": "text" + }, + { + "bbox": [ + 186, + 501, + 195, + 511 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 500, + 230, + 513 + ], + "score": 1.0, + "content": "samples", + "type": "text" + }, + { + "bbox": [ + 231, + 500, + 350, + 512 + ], + "score": 0.92, + "content": "a _ { t } \\dot { = } \\operatorname { t a n h } ( \\mu _ { \\theta } ( s _ { t } ) + \\sigma _ { \\theta } ( s _ { t } ) \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 500, + 379, + 513 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 380, + 500, + 430, + 512 + ], + "score": 0.92, + "content": "\\epsilon \\sim \\mathcal { N } ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 500, + 448, + 513 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 448, + 501, + 460, + 511 + ], + "score": 0.84, + "content": "\\mu _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 500, + 478, + 513 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 478, + 502, + 489, + 511 + ], + "score": 0.85, + "content": "\\sigma _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 500, + 506, + 513 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 271, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 271, + 523 + ], + "score": 1.0, + "content": "parametric mean and standard deviation.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 455, + 506, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 275, + 541 + ], + "score": 1.0, + "content": "The policy evaluation step learns the critic", + "type": "text" + }, + { + "bbox": [ + 275, + 528, + 317, + 540 + ], + "score": 0.93, + "content": "Q _ { \\theta } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "network by optimizing a single-step of the soft", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 538, + 177, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 177, + 550 + ], + "score": 1.0, + "content": "Bellman residual", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 527, + 506, + 550 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 553, + 415, + 594 + ], + "lines": [ + { + "bbox": [ + 195, + 553, + 415, + 594 + ], + "spans": [ + { + "bbox": [ + 195, + 553, + 415, + 594 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { J _ { Q } ( \\mathcal { D } ) = E _ { ( s _ { t } , a _ { t } , s _ { t } ^ { \\prime } ) \\sim \\mathcal { D } } [ ( Q _ { \\theta } ( s _ { t } , a _ { t } ) - y _ { t } ) ^ { 2 } ] } \\\\ & { \\qquad a _ { t } ^ { \\prime } { \\sim } \\pi ( \\cdot | s _ { t } ^ { \\prime } ) } \\\\ & { y _ { t } = r ( s _ { t } , a _ { t } ) + \\gamma [ Q _ { \\theta ^ { \\prime } } ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) - \\alpha \\log \\pi _ { \\theta } ( a _ { t } ^ { \\prime } | s _ { t } ^ { \\prime } ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "d967d8f701d309f8a2b9d1f8aed651c62ceb99e0c454aba8f437771b1a610922.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 195, + 553, + 415, + 566.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 195, + 566.6666666666666, + 415, + 580.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 195, + 580.3333333333333, + 415, + 593.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 596, + 504, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 133, + 610 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 597, + 142, + 607 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 597, + 272, + 610 + ], + "score": 1.0, + "content": "is a replay buffer of transitions,", + "type": "text" + }, + { + "bbox": [ + 273, + 597, + 281, + 607 + ], + "score": 0.85, + "content": "\\theta ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "is an exponential moving average of the weights. SAC", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 607, + 448, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 448, + 621 + ], + "score": 1.0, + "content": "uses clipped double-Q learning, which we omit for simplicity but employ in practice.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 597, + 505, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 503, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 320, + 639 + ], + "score": 1.0, + "content": "The policy improvement step then fits the actor policy", + "type": "text" + }, + { + "bbox": [ + 321, + 625, + 359, + 637 + ], + "score": 0.93, + "content": "\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 623, + 505, + 639 + ], + "score": 1.0, + "content": "network by optimizing the objective", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 623, + 505, + 639 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 641, + 412, + 664 + ], + "lines": [ + { + "bbox": [ + 198, + 641, + 412, + 664 + ], + "spans": [ + { + "bbox": [ + 198, + 641, + 412, + 664 + ], + "score": 0.92, + "content": "J _ { \\pi } ( \\mathcal { D } ) = E _ { s _ { t } \\sim \\mathcal { D } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\theta } ( \\cdot | s _ { t } ) | | \\exp \\{ \\frac { 1 } { \\alpha } Q _ { \\theta } ( s _ { t } , \\cdot ) \\} ) ] .", + "type": "interline_equation", + "image_path": "b8de65fbed10cf47b2b18c028ad713f54b791fcc6be2ce542047cddf3eb2f735.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 198, + 641, + 412, + 664 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 306, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 307, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 204, + 681 + ], + "score": 1.0, + "content": "Finally, the temperature", + "type": "text" + }, + { + "bbox": [ + 204, + 670, + 211, + 678 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 667, + 307, + 681 + ], + "score": 1.0, + "content": "is learned with the loss", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 667, + 307, + 681 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 683, + 402, + 705 + ], + "lines": [ + { + "bbox": [ + 208, + 683, + 402, + 705 + ], + "spans": [ + { + "bbox": [ + 208, + 683, + 402, + 705 + ], + "score": 0.93, + "content": "J _ { \\alpha } ( \\mathcal { D } ) = E _ { \\underset { a _ { t } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t } ) } { s _ { t } \\sim \\mathcal { D } } } [ - \\alpha \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) - \\alpha \\bar { \\mathcal { H } } ] ,", + "type": "interline_equation", + "image_path": "2cb81c1d6d5abb9132324b381e379d4fad3b4465a9273c43452d653b03c371b4.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 208, + 683, + 402, + 705 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 132, + 722 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 709, + 162, + 720 + ], + "score": 0.91, + "content": "\\bar { \\mathcal { H } } \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "is the target entropy hyper-parameter that the policy tries to match, which in practice is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 718, + 210, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 162, + 734 + ], + "score": 1.0, + "content": "usually set to", + "type": "text" + }, + { + "bbox": [ + 162, + 720, + 206, + 732 + ], + "score": 0.93, + "content": "\\bar { \\mathcal { H } } = - | \\bar { \\mathcal { A } } |", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 718, + 210, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 708, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 85, + 498, + 704 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 85, + 498, + 704 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 85, + 498, + 704 + ], + "spans": [ + { + "bbox": [ + 120, + 85, + 498, + 704 + ], + "score": 0.964, + "type": "image", + "image_path": "7c1ed64c41245f4fe76e367be7d81f27ce231d876cb7515942b5a20a75e1bb2a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 85, + 498, + 291.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 291.33333333333337, + 498, + 497.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 497.66666666666674, + 498, + 704.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 709, + 506, + 739 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 720 + ], + "score": 1.0, + "content": "Figure I.1: DCS Test Environments: Snapshots of test environments used for evaluating generalization on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 718, + 505, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 505, + 730 + ], + "score": 1.0, + "content": "Distracting Control Suite. Random backgrounds are projected from scenes of the first 30 videos of DAVIS 2017", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 730, + 174, + 740 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 174, + 740 + ], + "score": 1.0, + "content": "validation dataset.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 85, + 498, + 704 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 85, + 498, + 704 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 85, + 498, + 704 + ], + "spans": [ + { + "bbox": [ + 120, + 85, + 498, + 704 + ], + "score": 0.964, + "type": "image", + "image_path": "7c1ed64c41245f4fe76e367be7d81f27ce231d876cb7515942b5a20a75e1bb2a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 85, + 498, + 291.33333333333337 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 291.33333333333337, + 498, + 497.66666666666674 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 497.66666666666674, + 498, + 704.0000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 709, + 506, + 739 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 720 + ], + "score": 1.0, + "content": "Figure I.1: DCS Test Environments: Snapshots of test environments used for evaluating generalization on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 718, + 505, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 505, + 730 + ], + "score": 1.0, + "content": "Distracting Control Suite. Random backgrounds are projected from scenes of the first 30 videos of DAVIS 2017", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 730, + 174, + 740 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 174, + 740 + ], + "score": 1.0, + "content": "validation dataset.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 266, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 267, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 267, + 95 + ], + "score": 1.0, + "content": "I.1 ACTOR AND CRITIC NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 104, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 103, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 116 + ], + "score": 1.0, + "content": "Following Kostrikov et al. (2020), we use clipped double Q-learning for the critic, where each", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 107, + 115, + 115, + 127 + ], + "score": 0.83, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "-function is parametrized as a 3-layer MLP with ReLU activations after each layer except of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "last. The actor is also a 3-layer MLP with ReLUs that outputs mean and covariance for the diagonal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "score": 1.0, + "content": "Gaussian that represents the policy. The hidden dimension is set to 1024 for both the critic and actor.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 164, + 219, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 221, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 221, + 177 + ], + "score": 1.0, + "content": "I.2 ENCODER NETWORK", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 186, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 197 + ], + "score": 1.0, + "content": "We employ the encoder architecture from Kostrikov et al. (2020). This encoder consists of four", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 204, + 209 + ], + "score": 1.0, + "content": "convolution layers with", + "type": "text" + }, + { + "bbox": [ + 205, + 198, + 228, + 208 + ], + "score": 0.88, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "kernels and 32 channels. The ReLU activation is applied after each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "convolutional layer. We use stride to 1 everywhere, except of the first convolutional layer, which", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 219, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 231 + ], + "score": 1.0, + "content": "has stride 2. The output of the convnet is feed into a single fully-connected layer normalized", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "by LayerNorm. Finally, we apply tanh nonlinearity to the 50 dimensional output of the fully-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "connected layer. We initialize the weight matrix of fully-connected and convolutional layers with the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "orthogonal initialization and set the bias to be zero. The actor and critic networks both have separate", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "encoders, although we share the weights of the conv layers between them. Furthermore, only the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "critic optimizer is allowed to update these weights (i.e.,we stop the gradients from the actor before", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 285, + 310, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 310, + 297 + ], + "score": 1.0, + "content": "they propagate to the shared convolutional layers).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 306, + 311, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 312, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 312, + 320 + ], + "score": 1.0, + "content": "I.3 CONTRASTIVE METRIC EMBEDDING LOSS", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 329, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 333, + 340 + ], + "score": 1.0, + "content": "For all our experiments, we use a single ReLU layer with", + "type": "text" + }, + { + "bbox": [ + 334, + 329, + 369, + 339 + ], + "score": 0.88, + "content": "k = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 329, + 505, + 340 + ], + "score": 1.0, + "content": "units for the non-linear projection", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 204, + 352 + ], + "score": 1.0, + "content": "to obtain the embedding", + "type": "text" + }, + { + "bbox": [ + 205, + 340, + 216, + 351 + ], + "score": 0.85, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "(Figure A.1). We compute the embedding using the penultimate layer in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "the actor network. For picking the hyperparameters, we used 3 temperatures [0.1, 0.01, 1.0] and 3", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 361, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 144, + 374 + ], + "score": 1.0, + "content": "auxiliary", + "type": "text" + }, + { + "bbox": [ + 144, + 361, + 173, + 373 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 361, + 507, + 374 + ], + "score": 1.0, + "content": "loss coefficients [1, 3, 10] using “Ball In Cup Catch” as the validation environment.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 383, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 383, + 385 + ], + "score": 1.0, + "content": "All other hyperparameters are the same as prior work (see Table I.2).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 506, + 402 + ], + "score": 1.0, + "content": "We approximate optimal policies with the policies obtained after training a DrQ agent for 500K", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 401, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 412 + ], + "score": 1.0, + "content": "environment steps. Since a given action sequence from this approximate policy has the same perfor-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "mance across different training environments, we compute the PSM across training environments, via", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 423, + 454, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 454, + 434 + ], + "score": 1.0, + "content": "dynamic programming (see Section J.1 for pseudo-code), using such action sequences.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 108, + 438, + 504, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 256, + 452 + ], + "score": 1.0, + "content": "Total Loss. The total loss is given by", + "type": "text" + }, + { + "bbox": [ + 257, + 439, + 322, + 450 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { \\mathrm { R L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 437, + 349, + 452 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 350, + 439, + 370, + 450 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { R L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 437, + 505, + 452 + ], + "score": 1.0, + "content": "is the reinforcement learning loss", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 449, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 175, + 463 + ], + "score": 1.0, + "content": "which combines", + "type": "text" + }, + { + "bbox": [ + 176, + 450, + 203, + 462 + ], + "score": 0.9, + "content": "J _ { \\pi } ( \\mathcal { D } )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 449, + 207, + 463 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 208, + 450, + 235, + 462 + ], + "score": 0.9, + "content": "J _ { \\pi } ( \\mathcal { D } )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 449, + 257, + 463 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 257, + 450, + 286, + 462 + ], + "score": 0.9, + "content": "J _ { \\alpha } ( { \\mathcal { D } } ) _ { , }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 449, + 318, + 463 + ], + "score": 1.0, + "content": "), while", + "type": "text" + }, + { + "bbox": [ + 318, + 450, + 347, + 461 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 449, + 505, + 463 + ], + "score": 1.0, + "content": "is the auxiliary loss for learning PSEs", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 460, + 198, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 185, + 473 + ], + "score": 1.0, + "content": "with the coefficient", + "type": "text" + }, + { + "bbox": [ + 186, + 463, + 193, + 471 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 460, + 198, + 473 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 488, + 283, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 285, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 285, + 501 + ], + "score": 1.0, + "content": "I.4 TRAINING AND EVALUATION SETUP", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 506, + 522 + ], + "score": 1.0, + "content": "For evaluation, we use the first 30 videos from the DAVIS 2017 validation dataset (see Figure I.1).", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "Each checkpoint is evaluated by computing the average episode return over 100 episodes from the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "unseen environments. All experiments are performed with five random seeds per task used to compute", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 373, + 556 + ], + "score": 1.0, + "content": "means and standard deviations/errors of their evaluations. We use", + "type": "text" + }, + { + "bbox": [ + 374, + 543, + 436, + 554 + ], + "score": 0.52, + "content": "K = 2 , M = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "as prescribed by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 213, + 567 + ], + "score": 1.0, + "content": "Kostrikov et al. (2020) for", + "type": "text" + }, + { + "bbox": [ + 213, + 554, + 232, + 565 + ], + "score": 0.33, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 553, + 506, + 567 + ], + "score": 1.0, + "content": ". Following Kostrikov et al. (2020) and Stone et al. (2021), we use a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "score": 1.0, + "content": "different action repeat hyper-parameter for each task, which we summarize in Table I.3. We construct", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "an observational input as a 3-stack of consecutive frames (Kostrikov et al., 2020), where each frame", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 587, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 218, + 599 + ], + "score": 1.0, + "content": "is an RGB rendering of size", + "type": "text" + }, + { + "bbox": [ + 219, + 587, + 252, + 597 + ], + "score": 0.9, + "content": "8 4 \\times 8 4", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 587, + 504, + 599 + ], + "score": 1.0, + "content": "from the 0th camera. We then divide each pixel by 255 to scale", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "it down to [0, 1] range. For data augmentation, we maintain temporal consistency by using the same", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 609, + 292, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 292, + 621 + ], + "score": 1.0, + "content": "crop augmentation across consecutive frames.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5 + }, + { + "type": "table", + "bbox": [ + 192, + 653, + 415, + 733 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 142, + 633, + 466, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 633, + 469, + 646 + ], + "spans": [ + { + "bbox": [ + 142, + 633, + 469, + 646 + ], + "score": 1.0, + "content": "Table I.1: Optimal hyperparameters for PSE auxiliary loss for reporting results in Table 3.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "table_body", + "bbox": [ + 192, + 653, + 415, + 733 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 653, + 415, + 733 + ], + "spans": [ + { + "bbox": [ + 192, + 653, + 415, + 733 + ], + "score": 0.978, + "html": "
HyperparameterSetting
Contrastive temperature (1/入)0.1
Auxiliary loss coefficient (α)1.0
I-scale parameter (β)0.1
Batch Size(LcME)128
|1000 // Action Repeat
", + "type": "table", + "image_path": "1d4f1d425398184d92075449048aba2f6b98a942f7b3046738afc0165b55abf2.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 192, + 653, + 415, + 666.3333333333334 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 192, + 666.3333333333334, + 415, + 679.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 192, + 679.6666666666667, + 415, + 693.0000000000001 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 192, + 693.0000000000001, + 415, + 706.3333333333335 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 192, + 706.3333333333335, + 415, + 719.6666666666669 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 192, + 719.6666666666669, + 415, + 733.0000000000002 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "index": 41.75 + } + ], + "page_idx": 26, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 16, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 266, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 267, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 267, + 95 + ], + "score": 1.0, + "content": "I.1 ACTOR AND CRITIC NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 104, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 103, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 116 + ], + "score": 1.0, + "content": "Following Kostrikov et al. (2020), we use clipped double Q-learning for the critic, where each", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 107, + 115, + 115, + 127 + ], + "score": 0.83, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "-function is parametrized as a 3-layer MLP with ReLU activations after each layer except of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "last. The actor is also a 3-layer MLP with ReLUs that outputs mean and covariance for the diagonal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 148 + ], + "score": 1.0, + "content": "Gaussian that represents the policy. The hidden dimension is set to 1024 for both the critic and actor.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 103, + 506, + 148 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 164, + 219, + 175 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 221, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 221, + 177 + ], + "score": 1.0, + "content": "I.2 ENCODER NETWORK", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 186, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 197 + ], + "score": 1.0, + "content": "We employ the encoder architecture from Kostrikov et al. (2020). This encoder consists of four", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 204, + 209 + ], + "score": 1.0, + "content": "convolution layers with", + "type": "text" + }, + { + "bbox": [ + 205, + 198, + 228, + 208 + ], + "score": 0.88, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "kernels and 32 channels. The ReLU activation is applied after each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "convolutional layer. We use stride to 1 everywhere, except of the first convolutional layer, which", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 219, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 231 + ], + "score": 1.0, + "content": "has stride 2. The output of the convnet is feed into a single fully-connected layer normalized", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "by LayerNorm. Finally, we apply tanh nonlinearity to the 50 dimensional output of the fully-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "connected layer. We initialize the weight matrix of fully-connected and convolutional layers with the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "orthogonal initialization and set the bias to be zero. The actor and critic networks both have separate", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "encoders, although we share the weights of the conv layers between them. Furthermore, only the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "critic optimizer is allowed to update these weights (i.e.,we stop the gradients from the actor before", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 285, + 310, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 310, + 297 + ], + "score": 1.0, + "content": "they propagate to the shared convolutional layers).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 186, + 506, + 297 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 306, + 311, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 312, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 312, + 320 + ], + "score": 1.0, + "content": "I.3 CONTRASTIVE METRIC EMBEDDING LOSS", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 329, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 333, + 340 + ], + "score": 1.0, + "content": "For all our experiments, we use a single ReLU layer with", + "type": "text" + }, + { + "bbox": [ + 334, + 329, + 369, + 339 + ], + "score": 0.88, + "content": "k = 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 329, + 505, + 340 + ], + "score": 1.0, + "content": "units for the non-linear projection", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 204, + 352 + ], + "score": 1.0, + "content": "to obtain the embedding", + "type": "text" + }, + { + "bbox": [ + 205, + 340, + 216, + 351 + ], + "score": 0.85, + "content": "z _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "(Figure A.1). We compute the embedding using the penultimate layer in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "the actor network. For picking the hyperparameters, we used 3 temperatures [0.1, 0.01, 1.0] and 3", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 361, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 144, + 374 + ], + "score": 1.0, + "content": "auxiliary", + "type": "text" + }, + { + "bbox": [ + 144, + 361, + 173, + 373 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 361, + 507, + 374 + ], + "score": 1.0, + "content": "loss coefficients [1, 3, 10] using “Ball In Cup Catch” as the validation environment.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 383, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 383, + 385 + ], + "score": 1.0, + "content": "All other hyperparameters are the same as prior work (see Table I.2).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 329, + 507, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 506, + 402 + ], + "score": 1.0, + "content": "We approximate optimal policies with the policies obtained after training a DrQ agent for 500K", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 401, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 412 + ], + "score": 1.0, + "content": "environment steps. Since a given action sequence from this approximate policy has the same perfor-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "mance across different training environments, we compute the PSM across training environments, via", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 423, + 454, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 454, + 434 + ], + "score": 1.0, + "content": "dynamic programming (see Section J.1 for pseudo-code), using such action sequences.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 389, + 506, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 438, + 504, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 256, + 452 + ], + "score": 1.0, + "content": "Total Loss. The total loss is given by", + "type": "text" + }, + { + "bbox": [ + 257, + 439, + 322, + 450 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { \\mathrm { R L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 437, + 349, + 452 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 350, + 439, + 370, + 450 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { R L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 437, + 505, + 452 + ], + "score": 1.0, + "content": "is the reinforcement learning loss", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 449, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 175, + 463 + ], + "score": 1.0, + "content": "which combines", + "type": "text" + }, + { + "bbox": [ + 176, + 450, + 203, + 462 + ], + "score": 0.9, + "content": "J _ { \\pi } ( \\mathcal { D } )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 449, + 207, + 463 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 208, + 450, + 235, + 462 + ], + "score": 0.9, + "content": "J _ { \\pi } ( \\mathcal { D } )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 449, + 257, + 463 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 257, + 450, + 286, + 462 + ], + "score": 0.9, + "content": "J _ { \\alpha } ( { \\mathcal { D } } ) _ { , }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 449, + 318, + 463 + ], + "score": 1.0, + "content": "), while", + "type": "text" + }, + { + "bbox": [ + 318, + 450, + 347, + 461 + ], + "score": 0.89, + "content": "{ \\mathcal { L } } _ { \\mathrm { C M E } }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 449, + 505, + 463 + ], + "score": 1.0, + "content": "is the auxiliary loss for learning PSEs", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 460, + 198, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 185, + 473 + ], + "score": 1.0, + "content": "with the coefficient", + "type": "text" + }, + { + "bbox": [ + 186, + 463, + 193, + 471 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 460, + 198, + 473 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 437, + 505, + 473 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 488, + 283, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 285, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 285, + 501 + ], + "score": 1.0, + "content": "I.4 TRAINING AND EVALUATION SETUP", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 506, + 522 + ], + "score": 1.0, + "content": "For evaluation, we use the first 30 videos from the DAVIS 2017 validation dataset (see Figure I.1).", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "Each checkpoint is evaluated by computing the average episode return over 100 episodes from the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "unseen environments. All experiments are performed with five random seeds per task used to compute", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 373, + 556 + ], + "score": 1.0, + "content": "means and standard deviations/errors of their evaluations. We use", + "type": "text" + }, + { + "bbox": [ + 374, + 543, + 436, + 554 + ], + "score": 0.52, + "content": "K = 2 , M = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "as prescribed by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 213, + 567 + ], + "score": 1.0, + "content": "Kostrikov et al. (2020) for", + "type": "text" + }, + { + "bbox": [ + 213, + 554, + 232, + 565 + ], + "score": 0.33, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 553, + 506, + 567 + ], + "score": 1.0, + "content": ". Following Kostrikov et al. (2020) and Stone et al. (2021), we use a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 578 + ], + "score": 1.0, + "content": "different action repeat hyper-parameter for each task, which we summarize in Table I.3. We construct", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "an observational input as a 3-stack of consecutive frames (Kostrikov et al., 2020), where each frame", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 587, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 218, + 599 + ], + "score": 1.0, + "content": "is an RGB rendering of size", + "type": "text" + }, + { + "bbox": [ + 219, + 587, + 252, + 597 + ], + "score": 0.9, + "content": "8 4 \\times 8 4", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 587, + 504, + 599 + ], + "score": 1.0, + "content": "from the 0th camera. We then divide each pixel by 255 to scale", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "it down to [0, 1] range. For data augmentation, we maintain temporal consistency by using the same", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 609, + 292, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 292, + 621 + ], + "score": 1.0, + "content": "crop augmentation across consecutive frames.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 510, + 506, + 621 + ] + }, + { + "type": "table", + "bbox": [ + 192, + 653, + 415, + 733 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 142, + 633, + 466, + 645 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 633, + 469, + 646 + ], + "spans": [ + { + "bbox": [ + 142, + 633, + 469, + 646 + ], + "score": 1.0, + "content": "Table I.1: Optimal hyperparameters for PSE auxiliary loss for reporting results in Table 3.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "table_body", + "bbox": [ + 192, + 653, + 415, + 733 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 653, + 415, + 733 + ], + "spans": [ + { + "bbox": [ + 192, + 653, + 415, + 733 + ], + "score": 0.978, + "html": "
HyperparameterSetting
Contrastive temperature (1/入)0.1
Auxiliary loss coefficient (α)1.0
I-scale parameter (β)0.1
Batch Size(LcME)128
|1000 // Action Repeat
", + "type": "table", + "image_path": "1d4f1d425398184d92075449048aba2f6b98a942f7b3046738afc0165b55abf2.jpg" + } + ] + } + ], + "index": 43.5, + "virtual_lines": [ + { + "bbox": [ + 192, + 653, + 415, + 666.3333333333334 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 192, + 666.3333333333334, + 415, + 679.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 192, + 679.6666666666667, + 415, + 693.0000000000001 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 192, + 693.0000000000001, + 415, + 706.3333333333335 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 192, + 706.3333333333335, + 415, + 719.6666666666669 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 192, + 719.6666666666669, + 415, + 733.0000000000002 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "index": 41.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 203, + 100, + 406, + 245 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 501, + 91 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 502, + 92 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 502, + 92 + ], + "score": 1.0, + "content": "Table I.2: Hyper-parameters taken from Kostrikov et al. (2020) in the Distracting Control Suite experiments.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 203, + 100, + 406, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 203, + 100, + 406, + 245 + ], + "spans": [ + { + "bbox": [ + 203, + 100, + 406, + 245 + ], + "score": 0.979, + "html": "
ParameterSetting
Replay buffer capacity100,000
Seed steps1,000
Batch size (DrQ)512
Discount γ0.99
OptimizerAdam
Learning rate10-3
Critic target update frequency2
Critic Q-function soft-update rate T0.01
Actor update frequency2
Actor log stddev bounds
Init temperature[-10,2] 0.1
", + "type": "table", + "image_path": "bb9329f1383c2eca2834c2f994d252c85f7b9e348c37d854ea8cbdce690b2b7e.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 203, + 100, + 406, + 113.18181818181819 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 203, + 113.18181818181819, + 406, + 126.36363636363637 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 203, + 126.36363636363637, + 406, + 139.54545454545456 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 203, + 139.54545454545456, + 406, + 152.72727272727275 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 203, + 152.72727272727275, + 406, + 165.90909090909093 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 203, + 165.90909090909093, + 406, + 179.09090909090912 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 203, + 179.09090909090912, + 406, + 192.2727272727273 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 203, + 192.2727272727273, + 406, + 205.4545454545455 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 203, + 205.4545454545455, + 406, + 218.63636363636368 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 203, + 218.63636363636368, + 406, + 231.81818181818187 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 203, + 231.81818181818187, + 406, + 245.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 235, + 300, + 376, + 390 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 280, + 500, + 291 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 280, + 501, + 292 + ], + "spans": [ + { + "bbox": [ + 109, + 280, + 501, + 292 + ], + "score": 1.0, + "content": "Table I.3: The action repeat hyper-parameter used for each task in the Distracting Control Suite benchmark.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 235, + 300, + 376, + 390 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 235, + 300, + 376, + 390 + ], + "spans": [ + { + "bbox": [ + 235, + 300, + 376, + 390 + ], + "score": 0.971, + "html": "
Task nameAction repeat
Cartpole Swingup8
Reacher Easy4
Cheetah Run4
Finger Spin2
Ball In Cup Catch4
Walker Walk2
", + "type": "table", + "image_path": "da8d4384f50c36616da3269b10b8f38eb7bac21c398f4c93c4a028bf3daf07c3.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 300, + 376, + 345.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 235, + 345.0, + 376, + 390.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "image", + "bbox": [ + 106, + 455, + 505, + 679 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 455, + 505, + 679 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 679 + ], + "score": 0.973, + "type": "image", + "image_path": "9058a9690116b883329a6466ccc12d2c432e08d5b417ce9c0fc8928c6ebed906.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 106, + 455, + 505, + 529.6666666666666 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 529.6666666666666, + 505, + 604.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 604.3333333333333, + 505, + 678.9999999999999 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 688, + 505, + 729 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "Figure I.2: Random Initialization. Generalization performance on unseen environments over the course of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 263, + 712 + ], + "score": 1.0, + "content": "training. The agent is initialized randomly.", + "type": "text" + }, + { + "bbox": [ + 264, + 699, + 282, + 709 + ], + "score": 0.26, + "content": "_ \\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "augmented with PSEs outperforms DrQ. We plot the average", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "score": 1.0, + "content": "episode return across 5 seeds and the shaded region shows the standard deviation. Each checkpoint is evaluated", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 718, + 271, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 271, + 730 + ], + "score": 1.0, + "content": "using 100 episodes with unseen backgrounds.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 17.75 + } + ], + "page_idx": 27, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "28", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 417, + 245, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 246, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 246, + 429 + ], + "score": 1.0, + "content": "I.5 GENERALIZATION CURVES", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 203, + 100, + 406, + 245 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 501, + 91 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 80, + 502, + 92 + ], + "spans": [ + { + "bbox": [ + 108, + 80, + 502, + 92 + ], + "score": 1.0, + "content": "Table I.2: Hyper-parameters taken from Kostrikov et al. (2020) in the Distracting Control Suite experiments.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 203, + 100, + 406, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 203, + 100, + 406, + 245 + ], + "spans": [ + { + "bbox": [ + 203, + 100, + 406, + 245 + ], + "score": 0.979, + "html": "
ParameterSetting
Replay buffer capacity100,000
Seed steps1,000
Batch size (DrQ)512
Discount γ0.99
OptimizerAdam
Learning rate10-3
Critic target update frequency2
Critic Q-function soft-update rate T0.01
Actor update frequency2
Actor log stddev bounds
Init temperature[-10,2] 0.1
", + "type": "table", + "image_path": "bb9329f1383c2eca2834c2f994d252c85f7b9e348c37d854ea8cbdce690b2b7e.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 203, + 100, + 406, + 113.18181818181819 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 203, + 113.18181818181819, + 406, + 126.36363636363637 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 203, + 126.36363636363637, + 406, + 139.54545454545456 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 203, + 139.54545454545456, + 406, + 152.72727272727275 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 203, + 152.72727272727275, + 406, + 165.90909090909093 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 203, + 165.90909090909093, + 406, + 179.09090909090912 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 203, + 179.09090909090912, + 406, + 192.2727272727273 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 203, + 192.2727272727273, + 406, + 205.4545454545455 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 203, + 205.4545454545455, + 406, + 218.63636363636368 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 203, + 218.63636363636368, + 406, + 231.81818181818187 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 203, + 231.81818181818187, + 406, + 245.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 235, + 300, + 376, + 390 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 280, + 500, + 291 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 280, + 501, + 292 + ], + "spans": [ + { + "bbox": [ + 109, + 280, + 501, + 292 + ], + "score": 1.0, + "content": "Table I.3: The action repeat hyper-parameter used for each task in the Distracting Control Suite benchmark.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 235, + 300, + 376, + 390 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 235, + 300, + 376, + 390 + ], + "spans": [ + { + "bbox": [ + 235, + 300, + 376, + 390 + ], + "score": 0.971, + "html": "
Task nameAction repeat
Cartpole Swingup8
Reacher Easy4
Cheetah Run4
Finger Spin2
Ball In Cup Catch4
Walker Walk2
", + "type": "table", + "image_path": "da8d4384f50c36616da3269b10b8f38eb7bac21c398f4c93c4a028bf3daf07c3.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 300, + 376, + 345.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 235, + 345.0, + 376, + 390.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "image", + "bbox": [ + 106, + 455, + 505, + 679 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 455, + 505, + 679 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 679 + ], + "score": 0.973, + "type": "image", + "image_path": "9058a9690116b883329a6466ccc12d2c432e08d5b417ce9c0fc8928c6ebed906.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 106, + 455, + 505, + 529.6666666666666 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 529.6666666666666, + 505, + 604.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 604.3333333333333, + 505, + 678.9999999999999 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 688, + 505, + 729 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "Figure I.2: Random Initialization. Generalization performance on unseen environments over the course of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 697, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 263, + 712 + ], + "score": 1.0, + "content": "training. The agent is initialized randomly.", + "type": "text" + }, + { + "bbox": [ + 264, + 699, + 282, + 709 + ], + "score": 0.26, + "content": "_ \\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 697, + 505, + 712 + ], + "score": 1.0, + "content": "augmented with PSEs outperforms DrQ. We plot the average", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 506, + 720 + ], + "score": 1.0, + "content": "episode return across 5 seeds and the shaded region shows the standard deviation. Each checkpoint is evaluated", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 718, + 271, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 271, + 730 + ], + "score": 1.0, + "content": "using 100 episodes with unseen backgrounds.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 17.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 83, + 505, + 306 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 83, + 505, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 306 + ], + "score": 0.975, + "type": "image", + "image_path": "48d74111701471e8fc0c0104c5c1128c88f60454f9197c2358f703f65f4db49e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 83, + 505, + 157.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 157.33333333333331, + 505, + 231.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 231.66666666666663, + 505, + 305.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 316, + 505, + 357 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "Figure I.3: Pretrained Initialization. Generalization performance on unseen environments over the course of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 288, + 338 + ], + "score": 1.0, + "content": "training. The agent is initialized using a pretrained", + "type": "text" + }, + { + "bbox": [ + 289, + 326, + 306, + 336 + ], + "score": 0.32, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 325, + 475, + 338 + ], + "score": 1.0, + "content": "agent. DrQ augmented with PSEs outperforms", + "type": "text" + }, + { + "bbox": [ + 475, + 326, + 493, + 336 + ], + "score": 0.26, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "most of the environments. We plot the average return across 5 seeds and the shaded region shows the standard", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 347, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 416, + 357 + ], + "score": 1.0, + "content": "deviation. Each checkpoint is evaluated using 100 episodes with unseen backgrounds.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "title", + "bbox": [ + 107, + 378, + 199, + 390 + ], + "lines": [ + { + "bbox": [ + 104, + 376, + 200, + 393 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 200, + 393 + ], + "score": 1.0, + "content": "J PSEUDO CODE", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 106, + 403, + 342, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 342, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 342, + 415 + ], + "score": 1.0, + "content": "J.1 DYNAMIC PROGRAMMING FOR COMPUTING PSM", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 95, + 421, + 492, + 672 + ], + "lines": [ + { + "bbox": [ + 98, + 421, + 420, + 433 + ], + "spans": [ + { + "bbox": [ + 98, + 421, + 327, + 433 + ], + "score": 1.0, + "content": "1 def metric_fixed_point(cost_matrix, gamma", + "type": "text" + }, + { + "bbox": [ + 327, + 423, + 339, + 431 + ], + "score": 0.66, + "content": "{ } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 421, + 377, + 433 + ], + "score": 1.0, + "content": ".99, ep", + "type": "text" + }, + { + "bbox": [ + 377, + 423, + 409, + 432 + ], + "score": 0.48, + "content": "s { = } 1 { \\mathrm { e - } } 7", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 421, + 420, + 433 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 98, + 429, + 494, + 443 + ], + "spans": [ + { + "bbox": [ + 98, + 434, + 104, + 441 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 116, + 429, + 494, + 443 + ], + "score": 1.0, + "content": "\"\"\"DP for calculating PSM in environments with deterministic dynamics.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 444, + 104, + 451 + ], + "spans": [ + { + "bbox": [ + 99, + 444, + 104, + 451 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 98, + 451, + 147, + 463 + ], + "spans": [ + { + "bbox": [ + 98, + 455, + 104, + 461 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 116, + 451, + 147, + 463 + ], + "score": 1.0, + "content": "Args:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 98, + 461, + 489, + 473 + ], + "spans": [ + { + "bbox": [ + 98, + 464, + 104, + 471 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 127, + 461, + 489, + 473 + ], + "score": 1.0, + "content": "cost_matrix: DIST matrix where entries at index (i, j) is DIST(x_i,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 471, + 150, + 484 + ], + "spans": [ + { + "bbox": [ + 125, + 471, + 150, + 484 + ], + "score": 1.0, + "content": "y_j)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 97, + 481, + 291, + 493 + ], + "spans": [ + { + "bbox": [ + 97, + 484, + 104, + 491 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 126, + 481, + 291, + 493 + ], + "score": 1.0, + "content": "gamma: Metric discount factor.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 99, + 491, + 420, + 503 + ], + "spans": [ + { + "bbox": [ + 99, + 494, + 104, + 501 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 126, + 491, + 420, + 503 + ], + "score": 1.0, + "content": "eps: Threshold for stopping the fixed point iteration.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 97, + 501, + 136, + 512 + ], + "spans": [ + { + "bbox": [ + 97, + 504, + 104, + 512 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 117, + 501, + 136, + 510 + ], + "score": 1.0, + "content": "\"\"\"", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 97, + 510, + 280, + 523 + ], + "spans": [ + { + "bbox": [ + 97, + 514, + 104, + 521 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 115, + 510, + 127, + 523 + ], + "score": 1.0, + "content": "d", + "type": "text" + }, + { + "bbox": [ + 127, + 513, + 136, + 520 + ], + "score": 0.47, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 510, + 280, + 523 + ], + "score": 1.0, + "content": "np.zeros_like(cost_matrix)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 95, + 521, + 227, + 532 + ], + "spans": [ + { + "bbox": [ + 95, + 523, + 105, + 532 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 116, + 521, + 227, + 532 + ], + "score": 1.0, + "content": "def operator(d_cur):", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 95, + 532, + 254, + 542 + ], + "spans": [ + { + "bbox": [ + 95, + 534, + 104, + 541 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 127, + 532, + 159, + 542 + ], + "score": 1.0, + "content": "d_new", + "type": "text" + }, + { + "bbox": [ + 159, + 532, + 184, + 541 + ], + "score": 0.37, + "content": "\\ c = ~ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 533, + 189, + 541 + ], + "score": 0.25, + "content": "\\star", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 532, + 254, + 542 + ], + "score": 1.0, + "content": "cost_matrix", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 95, + 541, + 303, + 552 + ], + "spans": [ + { + "bbox": [ + 95, + 543, + 104, + 552 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 127, + 541, + 218, + 552 + ], + "score": 1.0, + "content": "discounted_d_cur", + "type": "text" + }, + { + "bbox": [ + 218, + 543, + 228, + 550 + ], + "score": 0.81, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 541, + 262, + 552 + ], + "score": 1.0, + "content": "gamma", + "type": "text" + }, + { + "bbox": [ + 262, + 543, + 270, + 551 + ], + "score": 0.46, + "content": "\\star", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 541, + 303, + 552 + ], + "score": 1.0, + "content": "d_cur", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 95, + 551, + 360, + 563 + ], + "spans": [ + { + "bbox": [ + 95, + 552, + 105, + 563 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 127, + 551, + 213, + 562 + ], + "score": 1.0, + "content": "d_new[:-1, :-1]", + "type": "text" + }, + { + "bbox": [ + 214, + 552, + 228, + 561 + ], + "score": 0.84, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 551, + 360, + 562 + ], + "score": 1.0, + "content": "discounted_d_cur[1:, 1:]", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 95, + 561, + 357, + 573 + ], + "spans": [ + { + "bbox": [ + 95, + 563, + 105, + 572 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 127, + 561, + 208, + 573 + ], + "score": 1.0, + "content": "d_new[:-1, -1]", + "type": "text" + }, + { + "bbox": [ + 209, + 563, + 222, + 570 + ], + "score": 0.85, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 561, + 357, + 573 + ], + "score": 1.0, + "content": "discounted_d_cur[1:, -1]", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 95, + 571, + 356, + 582 + ], + "spans": [ + { + "bbox": [ + 95, + 572, + 104, + 582 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 127, + 571, + 208, + 582 + ], + "score": 1.0, + "content": "d_new[-1, :-1]", + "type": "text" + }, + { + "bbox": [ + 208, + 573, + 222, + 581 + ], + "score": 0.82, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 571, + 356, + 582 + ], + "score": 1.0, + "content": "discounted_d_cur[-1, 1:]", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 95, + 581, + 196, + 592 + ], + "spans": [ + { + "bbox": [ + 95, + 583, + 105, + 592 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 127, + 581, + 196, + 592 + ], + "score": 1.0, + "content": "return d_new", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 95, + 592, + 104, + 602 + ], + "spans": [ + { + "bbox": [ + 95, + 592, + 104, + 602 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 95, + 600, + 179, + 612 + ], + "spans": [ + { + "bbox": [ + 95, + 602, + 105, + 612 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 116, + 600, + 179, + 612 + ], + "score": 1.0, + "content": "while True:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 95, + 611, + 231, + 622 + ], + "spans": [ + { + "bbox": [ + 95, + 612, + 105, + 622 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 127, + 611, + 159, + 622 + ], + "score": 1.0, + "content": "d_new", + "type": "text" + }, + { + "bbox": [ + 159, + 613, + 169, + 621 + ], + "score": 0.78, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 611, + 231, + 622 + ], + "score": 1.0, + "content": "operator(d)", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 95, + 620, + 318, + 633 + ], + "spans": [ + { + "bbox": [ + 95, + 622, + 105, + 632 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 127, + 620, + 282, + 633 + ], + "score": 1.0, + "content": "if np.sum(np.abs(d - d_new))", + "type": "text" + }, + { + "bbox": [ + 283, + 622, + 293, + 631 + ], + "score": 0.37, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 620, + 318, + 633 + ], + "score": 1.0, + "content": "eps:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 93, + 630, + 169, + 643 + ], + "spans": [ + { + "bbox": [ + 93, + 632, + 105, + 642 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 137, + 630, + 169, + 643 + ], + "score": 1.0, + "content": "break", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 94, + 641, + 157, + 653 + ], + "spans": [ + { + "bbox": [ + 94, + 642, + 105, + 653 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 127, + 641, + 157, + 651 + ], + "score": 1.0, + "content": "else:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 93, + 651, + 204, + 662 + ], + "spans": [ + { + "bbox": [ + 93, + 651, + 105, + 662 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 137, + 651, + 149, + 662 + ], + "score": 1.0, + "content": "d", + "type": "text" + }, + { + "bbox": [ + 149, + 652, + 158, + 660 + ], + "score": 0.42, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 651, + 204, + 662 + ], + "score": 1.0, + "content": "d_new[:]", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 94, + 660, + 165, + 672 + ], + "spans": [ + { + "bbox": [ + 94, + 662, + 105, + 671 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 117, + 660, + 165, + 672 + ], + "score": 1.0, + "content": "return d", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 106, + 682, + 218, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 219, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 219, + 694 + ], + "score": 1.0, + "content": "J.2 CONTRASTIVE LOSS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 98, + 702, + 316, + 732 + ], + "lines": [ + { + "bbox": [ + 97, + 700, + 318, + 713 + ], + "spans": [ + { + "bbox": [ + 97, + 700, + 318, + 713 + ], + "score": 1.0, + "content": "1 def contrastive_loss(similarity_matrix,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 97, + 711, + 296, + 722 + ], + "spans": [ + { + "bbox": [ + 97, + 714, + 104, + 721 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 219, + 711, + 296, + 722 + ], + "score": 1.0, + "content": "metric_values,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 97, + 722, + 285, + 732 + ], + "spans": [ + { + "bbox": [ + 97, + 723, + 104, + 731 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 219, + 722, + 285, + 732 + ], + "score": 1.0, + "content": "temperature,", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + } + ], + "page_idx": 28, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "29", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 83, + 505, + 306 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 83, + 505, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 306 + ], + "score": 0.975, + "type": "image", + "image_path": "48d74111701471e8fc0c0104c5c1128c88f60454f9197c2358f703f65f4db49e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 83, + 505, + 157.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 157.33333333333331, + 505, + 231.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 231.66666666666663, + 505, + 305.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 316, + 505, + 357 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "Figure I.3: Pretrained Initialization. Generalization performance on unseen environments over the course of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 288, + 338 + ], + "score": 1.0, + "content": "training. The agent is initialized using a pretrained", + "type": "text" + }, + { + "bbox": [ + 289, + 326, + 306, + 336 + ], + "score": 0.32, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 325, + 475, + 338 + ], + "score": 1.0, + "content": "agent. DrQ augmented with PSEs outperforms", + "type": "text" + }, + { + "bbox": [ + 475, + 326, + 493, + 336 + ], + "score": 0.26, + "content": "\\mathrm { D r Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "most of the environments. We plot the average return across 5 seeds and the shaded region shows the standard", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 347, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 416, + 357 + ], + "score": 1.0, + "content": "deviation. Each checkpoint is evaluated using 100 episodes with unseen backgrounds.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "title", + "bbox": [ + 107, + 378, + 199, + 390 + ], + "lines": [ + { + "bbox": [ + 104, + 376, + 200, + 393 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 200, + 393 + ], + "score": 1.0, + "content": "J PSEUDO CODE", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 106, + 403, + 342, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 342, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 342, + 415 + ], + "score": 1.0, + "content": "J.1 DYNAMIC PROGRAMMING FOR COMPUTING PSM", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "index", + "bbox": [ + 95, + 421, + 492, + 672 + ], + "lines": [ + { + "bbox": [ + 98, + 421, + 420, + 433 + ], + "spans": [ + { + "bbox": [ + 98, + 421, + 327, + 433 + ], + "score": 1.0, + "content": "1 def metric_fixed_point(cost_matrix, gamma", + "type": "text" + }, + { + "bbox": [ + 327, + 423, + 339, + 431 + ], + "score": 0.66, + "content": "{ } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 421, + 377, + 433 + ], + "score": 1.0, + "content": ".99, ep", + "type": "text" + }, + { + "bbox": [ + 377, + 423, + 409, + 432 + ], + "score": 0.48, + "content": "s { = } 1 { \\mathrm { e - } } 7", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 421, + 420, + 433 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 429, + 494, + 443 + ], + "spans": [ + { + "bbox": [ + 98, + 434, + 104, + 441 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 116, + 429, + 494, + 443 + ], + "score": 1.0, + "content": "\"\"\"DP for calculating PSM in environments with deterministic dynamics.", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 444, + 104, + 451 + ], + "spans": [ + { + "bbox": [ + 99, + 444, + 104, + 451 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 451, + 147, + 463 + ], + "spans": [ + { + "bbox": [ + 98, + 455, + 104, + 461 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 116, + 451, + 147, + 463 + ], + "score": 1.0, + "content": "Args:", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 98, + 461, + 489, + 473 + ], + "spans": [ + { + "bbox": [ + 98, + 464, + 104, + 471 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 127, + 461, + 489, + 473 + ], + "score": 1.0, + "content": "cost_matrix: DIST matrix where entries at index (i, j) is DIST(x_i,", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 471, + 150, + 484 + ], + "spans": [ + { + "bbox": [ + 125, + 471, + 150, + 484 + ], + "score": 1.0, + "content": "y_j)", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 481, + 291, + 493 + ], + "spans": [ + { + "bbox": [ + 97, + 484, + 104, + 491 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 126, + 481, + 291, + 493 + ], + "score": 1.0, + "content": "gamma: Metric discount factor.", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 491, + 420, + 503 + ], + "spans": [ + { + "bbox": [ + 99, + 494, + 104, + 501 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 126, + 491, + 420, + 503 + ], + "score": 1.0, + "content": "eps: Threshold for stopping the fixed point iteration.", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 501, + 136, + 512 + ], + "spans": [ + { + "bbox": [ + 97, + 504, + 104, + 512 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 117, + 501, + 136, + 510 + ], + "score": 1.0, + "content": "\"\"\"", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 510, + 280, + 523 + ], + "spans": [ + { + "bbox": [ + 97, + 514, + 104, + 521 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 115, + 510, + 127, + 523 + ], + "score": 1.0, + "content": "d", + "type": "text" + }, + { + "bbox": [ + 127, + 513, + 136, + 520 + ], + "score": 0.47, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 510, + 280, + 523 + ], + "score": 1.0, + "content": "np.zeros_like(cost_matrix)", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 521, + 227, + 532 + ], + "spans": [ + { + "bbox": [ + 95, + 523, + 105, + 532 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 116, + 521, + 227, + 532 + ], + "score": 1.0, + "content": "def operator(d_cur):", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 532, + 254, + 542 + ], + "spans": [ + { + "bbox": [ + 95, + 534, + 104, + 541 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 127, + 532, + 159, + 542 + ], + "score": 1.0, + "content": "d_new", + "type": "text" + }, + { + "bbox": [ + 159, + 532, + 184, + 541 + ], + "score": 0.37, + "content": "\\ c = ~ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 533, + 189, + 541 + ], + "score": 0.25, + "content": "\\star", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 532, + 254, + 542 + ], + "score": 1.0, + "content": "cost_matrix", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 541, + 303, + 552 + ], + "spans": [ + { + "bbox": [ + 95, + 543, + 104, + 552 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 127, + 541, + 218, + 552 + ], + "score": 1.0, + "content": "discounted_d_cur", + "type": "text" + }, + { + "bbox": [ + 218, + 543, + 228, + 550 + ], + "score": 0.81, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 541, + 262, + 552 + ], + "score": 1.0, + "content": "gamma", + "type": "text" + }, + { + "bbox": [ + 262, + 543, + 270, + 551 + ], + "score": 0.46, + "content": "\\star", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 541, + 303, + 552 + ], + "score": 1.0, + "content": "d_cur", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 551, + 360, + 563 + ], + "spans": [ + { + "bbox": [ + 95, + 552, + 105, + 563 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 127, + 551, + 213, + 562 + ], + "score": 1.0, + "content": "d_new[:-1, :-1]", + "type": "text" + }, + { + "bbox": [ + 214, + 552, + 228, + 561 + ], + "score": 0.84, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 551, + 360, + 562 + ], + "score": 1.0, + "content": "discounted_d_cur[1:, 1:]", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 561, + 357, + 573 + ], + "spans": [ + { + "bbox": [ + 95, + 563, + 105, + 572 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 127, + 561, + 208, + 573 + ], + "score": 1.0, + "content": "d_new[:-1, -1]", + "type": "text" + }, + { + "bbox": [ + 209, + 563, + 222, + 570 + ], + "score": 0.85, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 561, + 357, + 573 + ], + "score": 1.0, + "content": "discounted_d_cur[1:, -1]", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 571, + 356, + 582 + ], + "spans": [ + { + "bbox": [ + 95, + 572, + 104, + 582 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 127, + 571, + 208, + 582 + ], + "score": 1.0, + "content": "d_new[-1, :-1]", + "type": "text" + }, + { + "bbox": [ + 208, + 573, + 222, + 581 + ], + "score": 0.82, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 571, + 356, + 582 + ], + "score": 1.0, + "content": "discounted_d_cur[-1, 1:]", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 581, + 196, + 592 + ], + "spans": [ + { + "bbox": [ + 95, + 583, + 105, + 592 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 127, + 581, + 196, + 592 + ], + "score": 1.0, + "content": "return d_new", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 592, + 104, + 602 + ], + "spans": [ + { + "bbox": [ + 95, + 592, + 104, + 602 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 600, + 179, + 612 + ], + "spans": [ + { + "bbox": [ + 95, + 602, + 105, + 612 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 116, + 600, + 179, + 612 + ], + "score": 1.0, + "content": "while True:", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 611, + 231, + 622 + ], + "spans": [ + { + "bbox": [ + 95, + 612, + 105, + 622 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 127, + 611, + 159, + 622 + ], + "score": 1.0, + "content": "d_new", + "type": "text" + }, + { + "bbox": [ + 159, + 613, + 169, + 621 + ], + "score": 0.78, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 611, + 231, + 622 + ], + "score": 1.0, + "content": "operator(d)", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 620, + 318, + 633 + ], + "spans": [ + { + "bbox": [ + 95, + 622, + 105, + 632 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 127, + 620, + 282, + 633 + ], + "score": 1.0, + "content": "if np.sum(np.abs(d - d_new))", + "type": "text" + }, + { + "bbox": [ + 283, + 622, + 293, + 631 + ], + "score": 0.37, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 620, + 318, + 633 + ], + "score": 1.0, + "content": "eps:", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 630, + 169, + 643 + ], + "spans": [ + { + "bbox": [ + 93, + 632, + 105, + 642 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 137, + 630, + 169, + 643 + ], + "score": 1.0, + "content": "break", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 641, + 157, + 653 + ], + "spans": [ + { + "bbox": [ + 94, + 642, + 105, + 653 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 127, + 641, + 157, + 651 + ], + "score": 1.0, + "content": "else:", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 651, + 204, + 662 + ], + "spans": [ + { + "bbox": [ + 93, + 651, + 105, + 662 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 137, + 651, + 149, + 662 + ], + "score": 1.0, + "content": "d", + "type": "text" + }, + { + "bbox": [ + 149, + 652, + 158, + 660 + ], + "score": 0.42, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 651, + 204, + 662 + ], + "score": 1.0, + "content": "d_new[:]", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 660, + 165, + 672 + ], + "spans": [ + { + "bbox": [ + 94, + 662, + 105, + 671 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 117, + 660, + 165, + 672 + ], + "score": 1.0, + "content": "return d", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + } + ], + "index": 21, + "bbox_fs": [ + 93, + 421, + 494, + 672 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 682, + 218, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 219, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 219, + 694 + ], + "score": 1.0, + "content": "J.2 CONTRASTIVE LOSS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "index", + "bbox": [ + 98, + 702, + 316, + 732 + ], + "lines": [ + { + "bbox": [ + 97, + 700, + 318, + 713 + ], + "spans": [ + { + "bbox": [ + 97, + 700, + 318, + 713 + ], + "score": 1.0, + "content": "1 def contrastive_loss(similarity_matrix,", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 711, + 296, + 722 + ], + "spans": [ + { + "bbox": [ + 97, + 714, + 104, + 721 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 219, + 711, + 296, + 722 + ], + "score": 1.0, + "content": "metric_values,", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 722, + 285, + 732 + ], + "spans": [ + { + "bbox": [ + 97, + 723, + 104, + 731 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 219, + 722, + 285, + 732 + ], + "score": 1.0, + "content": "temperature,", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 83, + 275, + 94 + ], + "spans": [ + { + "bbox": [ + 99, + 86, + 104, + 92 + ], + "score": 1.0, + "content": "4", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 218, + 83, + 239, + 94 + ], + "score": 1.0, + "content": "bet", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 239, + 84, + 264, + 93 + ], + "score": 0.5, + "content": "\\gimel = 1 \\digamma . 0", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 264, + 83, + 275, + 94 + ], + "score": 1.0, + "content": "):", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 99, + 91, + 384, + 105 + ], + "spans": [ + { + "bbox": [ + 99, + 96, + 104, + 103 + ], + "score": 1.0, + "content": "5", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 116, + 91, + 384, + 105 + ], + "score": 1.0, + "content": "\"\"\"Contrative Loss with embedding similarity .\"\"\"", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 103, + 322, + 113 + ], + "spans": [ + { + "bbox": [ + 97, + 105, + 104, + 113 + ], + "score": 1.0, + "content": "6", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 117, + 103, + 186, + 113 + ], + "score": 1.0, + "content": "metric_shape", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 186, + 104, + 196, + 112 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 196, + 103, + 322, + 113 + ], + "score": 1.0, + "content": "tf.shape(metric_values)", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 113, + 419, + 124 + ], + "spans": [ + { + "bbox": [ + 97, + 115, + 104, + 123 + ], + "score": 1.0, + "content": "7", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 117, + 113, + 266, + 124 + ], + "score": 1.0, + "content": "## z_\\theta(X): embedding_1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 266, + 114, + 276, + 122 + ], + "score": 0.51, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 277, + 113, + 419, + 124 + ], + "score": 1.0, + "content": "nn_model.representation(X)", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 122, + 420, + 135 + ], + "spans": [ + { + "bbox": [ + 97, + 125, + 104, + 133 + ], + "score": 1.0, + "content": "8", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 116, + 122, + 256, + 135 + ], + "score": 1.0, + "content": "## z_\\theta(Y): embedding_", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 256, + 124, + 276, + 133 + ], + "score": 0.48, + "content": "\\_ =", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 277, + 122, + 420, + 135 + ], + "score": 1.0, + "content": "nn_model.representation(Y)", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 97, + 132, + 469, + 145 + ], + "spans": [ + { + "bbox": [ + 97, + 135, + 104, + 143 + ], + "score": 1.0, + "content": "9", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 132, + 228, + 145 + ], + "score": 1.0, + "content": "## similarity_matrix", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 229, + 134, + 239, + 142 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 239, + 132, + 469, + 145 + ], + "score": 1.0, + "content": "cosine_similarity(embedding_1, embedding_2", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 141, + 270, + 155 + ], + "spans": [ + { + "bbox": [ + 96, + 145, + 104, + 154 + ], + "score": 1.0, + "content": "10", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 141, + 207, + 155 + ], + "score": 1.0, + "content": "## metric_values", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 208, + 145, + 217, + 152 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 217, + 141, + 270, + 155 + ], + "score": 1.0, + "content": "PSM(X, Y)", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 96, + 154, + 291, + 163 + ], + "spans": [ + { + "bbox": [ + 96, + 155, + 104, + 163 + ], + "score": 1.0, + "content": "11", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 117, + 155, + 219, + 163 + ], + "score": 1.0, + "content": "similarity_matrix /", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 220, + 154, + 227, + 162 + ], + "score": 0.44, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 228, + 155, + 291, + 163 + ], + "score": 1.0, + "content": "temperature", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 164, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 95, + 164, + 105, + 174 + ], + "score": 1.0, + "content": "12", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 116, + 164, + 180, + 173 + ], + "score": 1.0, + "content": "neg_logits1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 180, + 164, + 190, + 172 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 190, + 164, + 286, + 173 + ], + "score": 1.0, + "content": "similarity_matrix", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 174, + 105, + 184 + ], + "spans": [ + { + "bbox": [ + 95, + 174, + 105, + 184 + ], + "score": 1.0, + "content": "13", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 183, + 500, + 194 + ], + "spans": [ + { + "bbox": [ + 95, + 185, + 105, + 194 + ], + "score": 1.0, + "content": "14", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 117, + 183, + 181, + 194 + ], + "score": 1.0, + "content": "col_indices", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 181, + 184, + 190, + 192 + ], + "score": 0.83, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 191, + 183, + 390, + 194 + ], + "score": 1.0, + "content": "tf.cast(tf.argmin(metric_values, axi", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 390, + 184, + 403, + 192 + ], + "score": 0.71, + "content": "\\ S = 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 404, + 183, + 446, + 194 + ], + "score": 1.0, + "content": "), dtype", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 446, + 185, + 452, + 192 + ], + "score": 0.62, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 452, + 183, + 500, + 194 + ], + "score": 1.0, + "content": "tf.int32)", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 193, + 249, + 203 + ], + "spans": [ + { + "bbox": [ + 95, + 194, + 105, + 203 + ], + "score": 1.0, + "content": "15", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 193, + 185, + 203 + ], + "score": 1.0, + "content": "pos_indices1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 186, + 194, + 196, + 202 + ], + "score": 0.79, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 196, + 193, + 249, + 203 + ], + "score": 1.0, + "content": "tf.stack(", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 203, + 489, + 214 + ], + "spans": [ + { + "bbox": [ + 95, + 204, + 105, + 214 + ], + "score": 1.0, + "content": "16", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 140, + 203, + 311, + 214 + ], + "score": 1.0, + "content": "(tf.range(metric_shape[0], dtype", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 312, + 205, + 317, + 212 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 318, + 203, + 468, + 214 + ], + "score": 1.0, + "content": "tf.int32), col_indices), axi", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 468, + 204, + 484, + 212 + ], + "score": 0.78, + "content": "\\mathsf { S } = 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 485, + 203, + 489, + 214 + ], + "score": 1.0, + "content": ")", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 212, + 436, + 225 + ], + "spans": [ + { + "bbox": [ + 95, + 214, + 104, + 224 + ], + "score": 1.0, + "content": "17", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 212, + 180, + 225 + ], + "score": 1.0, + "content": "pos_logits1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 180, + 214, + 190, + 222 + ], + "score": 0.81, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 191, + 212, + 436, + 225 + ], + "score": 1.0, + "content": "tf.gather_nd(similarity_matrix, pos_indices1)", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 225, + 105, + 234 + ], + "spans": [ + { + "bbox": [ + 95, + 225, + 105, + 234 + ], + "score": 1.0, + "content": "18", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 95, + 233, + 233, + 244 + ], + "spans": [ + { + "bbox": [ + 95, + 234, + 104, + 244 + ], + "score": 1.0, + "content": "19", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 116, + 233, + 198, + 244 + ], + "score": 1.0, + "content": "metric_values /", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 198, + 234, + 206, + 242 + ], + "score": 0.52, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 206, + 233, + 233, + 244 + ], + "score": 1.0, + "content": "beta", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 243, + 350, + 254 + ], + "spans": [ + { + "bbox": [ + 94, + 244, + 105, + 253 + ], + "score": 1.0, + "content": "20", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 117, + 243, + 218, + 254 + ], + "score": 1.0, + "content": "similarity_measure", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 218, + 244, + 228, + 252 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 228, + 243, + 350, + 254 + ], + "score": 1.0, + "content": "tf.exp(-metric_values)", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 253, + 425, + 264 + ], + "spans": [ + { + "bbox": [ + 94, + 254, + 104, + 263 + ], + "score": 1.0, + "content": "21", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 253, + 185, + 264 + ], + "score": 1.0, + "content": "pos_weights1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 186, + 254, + 196, + 262 + ], + "score": 0.81, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 196, + 253, + 425, + 264 + ], + "score": 1.0, + "content": "-tf.gather_nd(metric_values, pos_indices1)", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 263, + 265, + 273 + ], + "spans": [ + { + "bbox": [ + 94, + 264, + 104, + 273 + ], + "score": 1.0, + "content": "22", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 263, + 181, + 273 + ], + "score": 1.0, + "content": "pos_logits1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 181, + 264, + 195, + 272 + ], + "score": 0.84, + "content": "+ =", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 196, + 263, + 265, + 273 + ], + "score": 1.0, + "content": "pos_weights1", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 272, + 468, + 285 + ], + "spans": [ + { + "bbox": [ + 94, + 274, + 104, + 284 + ], + "score": 1.0, + "content": "23", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 116, + 272, + 207, + 285 + ], + "score": 1.0, + "content": "negative_weights", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 208, + 274, + 217, + 282 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 217, + 272, + 429, + 285 + ], + "score": 1.0, + "content": "tf.math.log((1.0 - similarity_measure)", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 429, + 274, + 437, + 282 + ], + "score": 0.33, + "content": "^ +", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 437, + 272, + 468, + 285 + ], + "score": 1.0, + "content": "1e-8)", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 283, + 351, + 294 + ], + "spans": [ + { + "bbox": [ + 94, + 284, + 105, + 293 + ], + "score": 1.0, + "content": "24", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 115, + 283, + 181, + 294 + ], + "score": 1.0, + "content": "neg_logits1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 182, + 284, + 196, + 292 + ], + "score": 0.81, + "content": "+ =", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 196, + 283, + 351, + 294 + ], + "score": 1.0, + "content": "tf.tensor_scatter_nd_update(", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 293, + 381, + 304 + ], + "spans": [ + { + "bbox": [ + 94, + 293, + 105, + 304 + ], + "score": 1.0, + "content": "25", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 138, + 293, + 381, + 304 + ], + "score": 1.0, + "content": "negative_weights, pos_indices1, pos_weights1)", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 304, + 105, + 313 + ], + "spans": [ + { + "bbox": [ + 94, + 304, + 105, + 313 + ], + "score": 1.0, + "content": "26", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 312, + 435, + 324 + ], + "spans": [ + { + "bbox": [ + 94, + 314, + 104, + 324 + ], + "score": 1.0, + "content": "27", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 116, + 312, + 180, + 324 + ], + "score": 1.0, + "content": "neg_logits1", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 181, + 314, + 190, + 322 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 191, + 312, + 419, + 324 + ], + "score": 1.0, + "content": "tf.math.reduce_logsumexp(neg_logits1, axis", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 419, + 314, + 430, + 322 + ], + "score": 0.68, + "content": "^ { = 1 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 431, + 312, + 435, + 324 + ], + "score": 1.0, + "content": ")", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 94, + 322, + 447, + 334 + ], + "spans": [ + { + "bbox": [ + 94, + 324, + 105, + 333 + ], + "score": 1.0, + "content": "28", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 117, + 322, + 447, + 334 + ], + "score": 1.0, + "content": "return tf.reduce_mean(neg_logits1 - pos_logits1) # Equation 4", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + } + ], + "index": 36, + "bbox_fs": [ + 97, + 700, + 318, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 94, + 82, + 501, + 335 + ], + "lines": [ + { + "bbox": [ + 99, + 83, + 275, + 94 + ], + "spans": [ + { + "bbox": [ + 99, + 86, + 104, + 92 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 218, + 83, + 239, + 94 + ], + "score": 1.0, + "content": "bet", + "type": "text" + }, + { + "bbox": [ + 239, + 84, + 264, + 93 + ], + "score": 0.5, + "content": "\\gimel = 1 \\digamma . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 83, + 275, + 94 + ], + "score": 1.0, + "content": "):", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 91, + 384, + 105 + ], + "spans": [ + { + "bbox": [ + 99, + 96, + 104, + 103 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 116, + 91, + 384, + 105 + ], + "score": 1.0, + "content": "\"\"\"Contrative Loss with embedding similarity .\"\"\"", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 103, + 322, + 113 + ], + "spans": [ + { + "bbox": [ + 97, + 105, + 104, + 113 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 117, + 103, + 186, + 113 + ], + "score": 1.0, + "content": "metric_shape", + "type": "text" + }, + { + "bbox": [ + 186, + 104, + 196, + 112 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 103, + 322, + 113 + ], + "score": 1.0, + "content": "tf.shape(metric_values)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 113, + 419, + 124 + ], + "spans": [ + { + "bbox": [ + 97, + 115, + 104, + 123 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 117, + 113, + 266, + 124 + ], + "score": 1.0, + "content": "## z_\\theta(X): embedding_1", + "type": "text" + }, + { + "bbox": [ + 266, + 114, + 276, + 122 + ], + "score": 0.51, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 113, + 419, + 124 + ], + "score": 1.0, + "content": "nn_model.representation(X)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 122, + 420, + 135 + ], + "spans": [ + { + "bbox": [ + 97, + 125, + 104, + 133 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 116, + 122, + 256, + 135 + ], + "score": 1.0, + "content": "## z_\\theta(Y): embedding_", + "type": "text" + }, + { + "bbox": [ + 256, + 124, + 276, + 133 + ], + "score": 0.48, + "content": "\\_ =", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 122, + 420, + 135 + ], + "score": 1.0, + "content": "nn_model.representation(Y)", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 132, + 469, + 145 + ], + "spans": [ + { + "bbox": [ + 97, + 135, + 104, + 143 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 115, + 132, + 228, + 145 + ], + "score": 1.0, + "content": "## similarity_matrix", + "type": "text" + }, + { + "bbox": [ + 229, + 134, + 239, + 142 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 132, + 469, + 145 + ], + "score": 1.0, + "content": "cosine_similarity(embedding_1, embedding_2", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 141, + 270, + 155 + ], + "spans": [ + { + "bbox": [ + 96, + 145, + 104, + 154 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 115, + 141, + 207, + 155 + ], + "score": 1.0, + "content": "## metric_values", + "type": "text" + }, + { + "bbox": [ + 208, + 145, + 217, + 152 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 141, + 270, + 155 + ], + "score": 1.0, + "content": "PSM(X, Y)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 154, + 291, + 163 + ], + "spans": [ + { + "bbox": [ + 96, + 155, + 104, + 163 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 117, + 155, + 219, + 163 + ], + "score": 1.0, + "content": "similarity_matrix /", + "type": "text" + }, + { + "bbox": [ + 220, + 154, + 227, + 162 + ], + "score": 0.44, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 155, + 291, + 163 + ], + "score": 1.0, + "content": "temperature", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 95, + 164, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 95, + 164, + 105, + 174 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 116, + 164, + 180, + 173 + ], + "score": 1.0, + "content": "neg_logits1", + "type": "text" + }, + { + "bbox": [ + 180, + 164, + 190, + 172 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 164, + 286, + 173 + ], + "score": 1.0, + "content": "similarity_matrix", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 95, + 174, + 105, + 184 + ], + "spans": [ + { + "bbox": [ + 95, + 174, + 105, + 184 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 95, + 183, + 500, + 194 + ], + "spans": [ + { + "bbox": [ + 95, + 185, + 105, + 194 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 117, + 183, + 181, + 194 + ], + "score": 1.0, + "content": "col_indices", + "type": "text" + }, + { + "bbox": [ + 181, + 184, + 190, + 192 + ], + "score": 0.83, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 183, + 390, + 194 + ], + "score": 1.0, + "content": "tf.cast(tf.argmin(metric_values, axi", + "type": "text" + }, + { + "bbox": [ + 390, + 184, + 403, + 192 + ], + "score": 0.71, + "content": "\\ S = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 183, + 446, + 194 + ], + "score": 1.0, + "content": "), dtype", + "type": "text" + }, + { + "bbox": [ + 446, + 185, + 452, + 192 + ], + "score": 0.62, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 183, + 500, + 194 + ], + "score": 1.0, + "content": "tf.int32)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 95, + 193, + 249, + 203 + ], + "spans": [ + { + "bbox": [ + 95, + 194, + 105, + 203 + ], + "score": 1.0, + "content": "15", + "type": "text" + }, + { + "bbox": [ + 115, + 193, + 185, + 203 + ], + "score": 1.0, + "content": "pos_indices1", + "type": "text" + }, + { + "bbox": [ + 186, + 194, + 196, + 202 + ], + "score": 0.79, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 193, + 249, + 203 + ], + "score": 1.0, + "content": "tf.stack(", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 95, + 203, + 489, + 214 + ], + "spans": [ + { + "bbox": [ + 95, + 204, + 105, + 214 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 140, + 203, + 311, + 214 + ], + "score": 1.0, + "content": "(tf.range(metric_shape[0], dtype", + "type": "text" + }, + { + "bbox": [ + 312, + 205, + 317, + 212 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 203, + 468, + 214 + ], + "score": 1.0, + "content": "tf.int32), col_indices), axi", + "type": "text" + }, + { + "bbox": [ + 468, + 204, + 484, + 212 + ], + "score": 0.78, + "content": "\\mathsf { S } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 203, + 489, + 214 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 95, + 212, + 436, + 225 + ], + "spans": [ + { + "bbox": [ + 95, + 214, + 104, + 224 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 115, + 212, + 180, + 225 + ], + "score": 1.0, + "content": "pos_logits1", + "type": "text" + }, + { + "bbox": [ + 180, + 214, + 190, + 222 + ], + "score": 0.81, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 212, + 436, + 225 + ], + "score": 1.0, + "content": "tf.gather_nd(similarity_matrix, pos_indices1)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 95, + 225, + 105, + 234 + ], + "spans": [ + { + "bbox": [ + 95, + 225, + 105, + 234 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 95, + 233, + 233, + 244 + ], + "spans": [ + { + "bbox": [ + 95, + 234, + 104, + 244 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 116, + 233, + 198, + 244 + ], + "score": 1.0, + "content": "metric_values /", + "type": "text" + }, + { + "bbox": [ + 198, + 234, + 206, + 242 + ], + "score": 0.52, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 233, + 233, + 244 + ], + "score": 1.0, + "content": "beta", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 94, + 243, + 350, + 254 + ], + "spans": [ + { + "bbox": [ + 94, + 244, + 105, + 253 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 117, + 243, + 218, + 254 + ], + "score": 1.0, + "content": "similarity_measure", + "type": "text" + }, + { + "bbox": [ + 218, + 244, + 228, + 252 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 243, + 350, + 254 + ], + "score": 1.0, + "content": "tf.exp(-metric_values)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 94, + 253, + 425, + 264 + ], + "spans": [ + { + "bbox": [ + 94, + 254, + 104, + 263 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 115, + 253, + 185, + 264 + ], + "score": 1.0, + "content": "pos_weights1", + "type": "text" + }, + { + "bbox": [ + 186, + 254, + 196, + 262 + ], + "score": 0.81, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 253, + 425, + 264 + ], + "score": 1.0, + "content": "-tf.gather_nd(metric_values, pos_indices1)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 94, + 263, + 265, + 273 + ], + "spans": [ + { + "bbox": [ + 94, + 264, + 104, + 273 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 115, + 263, + 181, + 273 + ], + "score": 1.0, + "content": "pos_logits1", + "type": "text" + }, + { + "bbox": [ + 181, + 264, + 195, + 272 + ], + "score": 0.84, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 263, + 265, + 273 + ], + "score": 1.0, + "content": "pos_weights1", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 94, + 272, + 468, + 285 + ], + "spans": [ + { + "bbox": [ + 94, + 274, + 104, + 284 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 116, + 272, + 207, + 285 + ], + "score": 1.0, + "content": "negative_weights", + "type": "text" + }, + { + "bbox": [ + 208, + 274, + 217, + 282 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 272, + 429, + 285 + ], + "score": 1.0, + "content": "tf.math.log((1.0 - similarity_measure)", + "type": "text" + }, + { + "bbox": [ + 429, + 274, + 437, + 282 + ], + "score": 0.33, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 272, + 468, + 285 + ], + "score": 1.0, + "content": "1e-8)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 94, + 283, + 351, + 294 + ], + "spans": [ + { + "bbox": [ + 94, + 284, + 105, + 293 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 115, + 283, + 181, + 294 + ], + "score": 1.0, + "content": "neg_logits1", + "type": "text" + }, + { + "bbox": [ + 182, + 284, + 196, + 292 + ], + "score": 0.81, + "content": "+ =", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 283, + 351, + 294 + ], + "score": 1.0, + "content": "tf.tensor_scatter_nd_update(", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 94, + 293, + 381, + 304 + ], + "spans": [ + { + "bbox": [ + 94, + 293, + 105, + 304 + ], + "score": 1.0, + "content": "25", + "type": "text" + }, + { + "bbox": [ + 138, + 293, + 381, + 304 + ], + "score": 1.0, + "content": "negative_weights, pos_indices1, pos_weights1)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 94, + 304, + 105, + 313 + ], + "spans": [ + { + "bbox": [ + 94, + 304, + 105, + 313 + ], + "score": 1.0, + "content": "26", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 94, + 312, + 435, + 324 + ], + "spans": [ + { + "bbox": [ + 94, + 314, + 104, + 324 + ], + "score": 1.0, + "content": "27", + "type": "text" + }, + { + "bbox": [ + 116, + 312, + 180, + 324 + ], + "score": 1.0, + "content": "neg_logits1", + "type": "text" + }, + { + "bbox": [ + 181, + 314, + 190, + 322 + ], + "score": 0.76, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 312, + 419, + 324 + ], + "score": 1.0, + "content": "tf.math.reduce_logsumexp(neg_logits1, axis", + "type": "text" + }, + { + "bbox": [ + 419, + 314, + 430, + 322 + ], + "score": 0.68, + "content": "^ { = 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 312, + 435, + 324 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 94, + 322, + 447, + 334 + ], + "spans": [ + { + "bbox": [ + 94, + 324, + 105, + 333 + ], + "score": 1.0, + "content": "28", + "type": "text" + }, + { + "bbox": [ + 117, + 322, + 447, + 334 + ], + "score": 1.0, + "content": "return tf.reduce_mean(neg_logits1 - pos_logits1) # Equation 4", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 12 + } + ], + "page_idx": 29, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "30", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 347, + 113, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 113, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 113, + 355 + ], + "score": 1.0, + "content": "a", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "index", + "bbox": [ + 94, + 82, + 501, + 335 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 94, + 83, + 500, + 334 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/qda7-sVg84/qda7-sVg84_model.json b/parse/train/qda7-sVg84/qda7-sVg84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a1527b9db6d278265b72a4cc5178713813109a4 --- /dev/null +++ b/parse/train/qda7-sVg84/qda7-sVg84_model.json @@ -0,0 +1,47960 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 619, + 1303, + 619, + 1303, + 986, + 398, + 986 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1077, + 882, + 1077, + 882, + 1411, + 299, + 1411 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1628, + 1404, + 1628, + 1404, + 1811, + 298, + 1811 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1827, + 1403, + 1827, + 1403, + 1950, + 299, + 1950 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 901, + 1241, + 1405, + 1241, + 1405, + 1407, + 901, + 1407 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 934, + 1063, + 1363, + 1063, + 1363, + 1228, + 934, + 1228 + ], + "score": 0.968 + }, + { + "category_id": 0, + "poly": [ + 298, + 220, + 1144, + 220, + 1144, + 379, + 298, + 379 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 298, + 1428, + 1405, + 1428, + 1405, + 1613, + 298, + 1613 + ], + "score": 0.962 + }, + { + "category_id": 2, + "poly": [ + 329, + 1974, + 929, + 1974, + 929, + 2033, + 329, + 2033 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 313, + 427, + 1346, + 427, + 1346, + 524, + 313, + 524 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 301, + 1027, + 574, + 1027, + 574, + 1061, + 301, + 1061 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 813, + 76, + 813, + 104, + 299, + 104 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 773, + 565, + 927, + 565, + 927, + 599, + 773, + 599 + ], + "score": 0.842 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.74 + }, + { + "category_id": 13, + "poly": [ + 961, + 892, + 1045, + 892, + 1045, + 925, + 961, + 925 + ], + "score": 0.28, + "latex": "( \\mathrm { P S E s ^ { 1 } } )" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1241.0, + 1406.0, + 1241.0, + 1406.0, + 1273.0, + 901.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1267.0, + 1407.0, + 1267.0, + 1407.0, + 1301.0, + 899.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1295.0, + 1404.0, + 1295.0, + 1404.0, + 1328.0, + 900.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1326.0, + 1404.0, + 1326.0, + 1404.0, + 1354.0, + 900.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1353.0, + 1402.0, + 1353.0, + 1402.0, + 1382.0, + 902.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1379.0, + 1390.0, + 1379.0, + 1390.0, + 1411.0, + 902.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1063.0, + 1354.0, + 1063.0, + 1354.0, + 1092.0, + 938.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 220.0, + 1149.0, + 220.0, + 1149.0, + 269.0, + 297.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 275.0, + 1116.0, + 275.0, + 1116.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 332.0, + 894.0, + 332.0, + 894.0, + 380.0, + 295.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1970.0, + 930.0, + 1970.0, + 930.0, + 2006.0, + 329.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 2001.0, + 545.0, + 2001.0, + 545.0, + 2036.0, + 332.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1022.0, + 578.0, + 1022.0, + 578.0, + 1068.0, + 293.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 564.0, + 932.0, + 564.0, + 932.0, + 602.0, + 770.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 617.0, + 1305.0, + 617.0, + 1305.0, + 656.0, + 393.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 650.0, + 1304.0, + 650.0, + 1304.0, + 686.0, + 394.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 681.0, + 1304.0, + 681.0, + 1304.0, + 715.0, + 394.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 712.0, + 1307.0, + 712.0, + 1307.0, + 747.0, + 393.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 742.0, + 1307.0, + 742.0, + 1307.0, + 777.0, + 394.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 773.0, + 1309.0, + 773.0, + 1309.0, + 808.0, + 392.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 802.0, + 1306.0, + 802.0, + 1306.0, + 837.0, + 393.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 834.0, + 1304.0, + 834.0, + 1304.0, + 866.0, + 394.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 865.0, + 1304.0, + 865.0, + 1304.0, + 897.0, + 394.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 893.0, + 960.0, + 893.0, + 960.0, + 929.0, + 393.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 893.0, + 1306.0, + 893.0, + 1306.0, + 929.0, + 1046.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 923.0, + 1305.0, + 923.0, + 1305.0, + 961.0, + 393.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 955.0, + 1242.0, + 955.0, + 1242.0, + 990.0, + 394.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1077.0, + 885.0, + 1077.0, + 885.0, + 1110.0, + 297.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1107.0, + 886.0, + 1107.0, + 886.0, + 1140.0, + 296.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 885.0, + 1138.0, + 885.0, + 1171.0, + 294.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1169.0, + 884.0, + 1169.0, + 884.0, + 1199.0, + 296.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1197.0, + 884.0, + 1197.0, + 884.0, + 1232.0, + 294.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1228.0, + 883.0, + 1228.0, + 883.0, + 1261.0, + 295.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1260.0, + 885.0, + 1260.0, + 885.0, + 1292.0, + 293.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1291.0, + 882.0, + 1291.0, + 882.0, + 1321.0, + 294.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1321.0, + 883.0, + 1321.0, + 883.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1353.0, + 881.0, + 1353.0, + 881.0, + 1383.0, + 297.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1382.0, + 795.0, + 1382.0, + 795.0, + 1414.0, + 297.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1626.0, + 1405.0, + 1626.0, + 1405.0, + 1663.0, + 293.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1660.0, + 1402.0, + 1660.0, + 1402.0, + 1692.0, + 296.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1725.0, + 293.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1720.0, + 1405.0, + 1720.0, + 1405.0, + 1755.0, + 294.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 1405.0, + 1749.0, + 1405.0, + 1783.0, + 294.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1781.0, + 1243.0, + 1781.0, + 1243.0, + 1814.0, + 293.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1825.0, + 1409.0, + 1825.0, + 1409.0, + 1864.0, + 294.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1858.0, + 1405.0, + 1858.0, + 1405.0, + 1891.0, + 295.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1889.0, + 1405.0, + 1889.0, + 1405.0, + 1925.0, + 294.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1919.0, + 1404.0, + 1919.0, + 1404.0, + 1951.0, + 295.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1422.0, + 1410.0, + 1422.0, + 1410.0, + 1466.0, + 292.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1459.0, + 1406.0, + 1459.0, + 1406.0, + 1491.0, + 293.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1487.0, + 1406.0, + 1487.0, + 1406.0, + 1523.0, + 295.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1519.0, + 1407.0, + 1519.0, + 1407.0, + 1555.0, + 295.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1549.0, + 1406.0, + 1549.0, + 1406.0, + 1583.0, + 293.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1578.0, + 1236.0, + 1578.0, + 1236.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 425.0, + 1352.0, + 425.0, + 1352.0, + 467.0, + 309.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 461.0, + 654.0, + 461.0, + 654.0, + 496.0, + 313.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 492.0, + 1187.0, + 492.0, + 1187.0, + 530.0, + 314.0, + 530.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 567, + 1405, + 567, + 1405, + 844, + 298, + 844 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1118, + 1403, + 1118, + 1403, + 1303, + 298, + 1303 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 918, + 1405, + 918, + 1405, + 1107, + 297, + 1107 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1404, + 307, + 1404, + 554, + 298, + 554 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1658, + 1404, + 1658, + 1404, + 1843, + 298, + 1843 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1519, + 1405, + 1519, + 1405, + 1645, + 298, + 1645 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1317, + 1405, + 1317, + 1405, + 1503, + 298, + 1503 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1402, + 229, + 1402, + 293, + 296, + 293 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 301, + 1917, + 1402, + 1917, + 1402, + 1981, + 301, + 1981 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 299, + 865, + 579, + 865, + 579, + 900, + 299, + 900 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 319, + 2004, + 1373, + 2004, + 1373, + 2034, + 319, + 2034 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.898 + }, + { + "category_id": 8, + "poly": [ + 405, + 1858, + 1289, + 1858, + 1289, + 1899, + 405, + 1899 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1864, + 1399, + 1864, + 1399, + 1893, + 1366, + 1893 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.755 + }, + { + "category_id": 13, + "poly": [ + 605, + 1071, + 938, + 1071, + 938, + 1108, + 605, + 1108 + ], + "score": 0.94, + "latex": "\\mathbb E _ { a _ { t } \\sim \\pi ( \\cdot \\textit { \\textbf { \\ i } } x _ { t } ) } [ \\sum _ { t } \\bar { \\gamma } ^ { t } R ( x _ { t } , a _ { t } ) ]" + }, + { + "category_id": 13, + "poly": [ + 1057, + 1518, + 1166, + 1518, + 1166, + 1555, + 1057, + 1555 + ], + "score": 0.93, + "latex": "\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 789, + 981, + 880, + 981, + 880, + 1014, + 789, + 1014 + ], + "score": 0.93, + "latex": "\\pi ( \\cdot \\mid x )" + }, + { + "category_id": 13, + "poly": [ + 299, + 1782, + 452, + 1782, + 452, + 1815, + 299, + 1815 + ], + "score": 0.92, + "latex": "\\mathcal { W } _ { 1 } ( d ) ( A , B )" + }, + { + "category_id": 13, + "poly": [ + 1180, + 1916, + 1348, + 1916, + 1348, + 1949, + 1180, + 1949 + ], + "score": 0.92, + "latex": "z _ { \\theta } : \\mathcal { S } \\mathbb { R } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 893, + 1012, + 950, + 1012, + 950, + 1044, + 893, + 1044 + ], + "score": 0.92, + "latex": "\\pi ( x )" + }, + { + "category_id": 13, + "poly": [ + 481, + 1042, + 538, + 1042, + 538, + 1074, + 481, + 1074 + ], + "score": 0.92, + "latex": "\\pi ( x )" + }, + { + "category_id": 13, + "poly": [ + 565, + 981, + 675, + 981, + 675, + 1014, + 565, + 1014 + ], + "score": 0.92, + "latex": "\\gamma \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1140, + 1751, + 1317, + 1751, + 1317, + 1784, + 1140, + 1784 + ], + "score": 0.92, + "latex": "\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1041, + 378, + 1041, + 378, + 1075, + 297, + 1075 + ], + "score": 0.92, + "latex": "\\pi ( \\cdot | x )" + }, + { + "category_id": 13, + "poly": [ + 296, + 950, + 468, + 950, + 468, + 984, + 296, + 984 + ], + "score": 0.91, + "latex": "( \\mathcal { X } , \\mathcal { A } , R , P , \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1015, + 982, + 1092, + 982, + 1092, + 1009, + 1015, + 1009 + ], + "score": 0.91, + "latex": "x \\in \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1008, + 1212, + 1047, + 1212, + 1047, + 1243, + 1008, + 1243 + ], + "score": 0.9, + "latex": "\\pi _ { \\mathcal { X } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1273, + 384, + 1273, + 384, + 1305, + 345, + 1305 + ], + "score": 0.89, + "latex": "P _ { \\mathcal { X } } ^ { \\pi }" + }, + { + "category_id": 14, + "poly": [ + 408, + 1857, + 1292, + 1857, + 1292, + 1899, + 408, + 1899 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { d _ { \\pi } ( x , y ) = | R ^ { \\pi } ( x ) - R ^ { \\pi } ( y ) | + \\gamma \\mathcal { W } _ { 1 } ( d _ { \\pi } ) \\big ( P ^ { \\pi } ( \\cdot | x ) , P ^ { \\pi } ( \\cdot | y ) \\big ) , \\qquad x , y \\in \\mathcal { S } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1471, + 339, + 1471, + 339, + 1505, + 297, + 1505 + ], + "score": 0.89, + "latex": "\\mathbb { M } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 1357, + 1241, + 1400, + 1241, + 1400, + 1274, + 1357, + 1274 + ], + "score": 0.89, + "latex": "R _ { \\mathcal { X } } ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 377, + 1751, + 409, + 1751, + 409, + 1781, + 377, + 1781 + ], + "score": 0.88, + "latex": "d _ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1140, + 1181, + 1183, + 1181, + 1183, + 1210, + 1140, + 1210 + ], + "score": 0.86, + "latex": "R _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 1152, + 1043, + 1185, + 1043, + 1185, + 1070, + 1152, + 1070 + ], + "score": 0.85, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1077, + 1286, + 1077, + 1286, + 1104, + 1255, + 1104 + ], + "score": 0.85, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 420, + 1181, + 445, + 1181, + 445, + 1210, + 420, + 1210 + ], + "score": 0.84, + "latex": "\\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 1194, + 1181, + 1235, + 1181, + 1235, + 1210, + 1194, + 1210 + ], + "score": 0.84, + "latex": "P _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 1324, + 920, + 1403, + 920, + 1403, + 951, + 1324, + 951 + ], + "score": 0.83, + "latex": "\\begin{array} { r l } { \\mathcal { M } } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 342, + 1181, + 369, + 1181, + 369, + 1207, + 342, + 1207 + ], + "score": 0.82, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1155, + 620, + 1155, + 620, + 1181, + 601, + 1181 + ], + "score": 0.82, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1045, + 1664, + 1067, + 1664, + 1067, + 1687, + 1045, + 1687 + ], + "score": 0.81, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 924, + 1212, + 951, + 1212, + 951, + 1238, + 924, + 1238 + ], + "score": 0.81, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 942, + 1323, + 960, + 1323, + 960, + 1350, + 942, + 1350 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1184, + 1784, + 1210, + 1784, + 1210, + 1809, + 1184, + 1809 + ], + "score": 0.81, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 400, + 1320, + 422, + 1320, + 422, + 1346, + 400, + 1346 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1375, + 1553, + 1397, + 1553, + 1397, + 1580, + 1375, + 1580 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 684, + 952, + 711, + 952, + 711, + 978, + 684, + 978 + ], + "score": 0.81, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 911, + 1350, + 934, + 1350, + 934, + 1376, + 911, + 1376 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1233, + 1724, + 1253, + 1724, + 1253, + 1748, + 1233, + 1748 + ], + "score": 0.8, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1125, + 1695, + 1144, + 1695, + 1144, + 1721, + 1125, + 1721 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1141, + 952, + 1166, + 952, + 1166, + 979, + 1141, + 979 + ], + "score": 0.8, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1186, + 1619, + 1203, + 1619, + 1203, + 1645, + 1186, + 1645 + ], + "score": 0.8, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 585, + 1411, + 607, + 1411, + 607, + 1438, + 585, + 1438 + ], + "score": 0.79, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 902, + 952, + 926, + 952, + 926, + 978, + 902, + 978 + ], + "score": 0.79, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 298, + 984, + 321, + 984, + 321, + 1008, + 298, + 1008 + ], + "score": 0.79, + "latex": "P" + }, + { + "category_id": 13, + "poly": [ + 418, + 1694, + 438, + 1694, + 438, + 1717, + 418, + 1717 + ], + "score": 0.79, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 885, + 1440, + 908, + 1440, + 908, + 1468, + 885, + 1468 + ], + "score": 0.79, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 686, + 1384, + 707, + 1384, + 707, + 1408, + 686, + 1408 + ], + "score": 0.78, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 477, + 1814, + 499, + 1814, + 499, + 1839, + 477, + 1839 + ], + "score": 0.78, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 928, + 1472, + 950, + 1472, + 950, + 1498, + 928, + 1498 + ], + "score": 0.78, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 822, + 1587, + 840, + 1587, + 840, + 1614, + 822, + 1614 + ], + "score": 0.77, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1195, + 1276, + 1217, + 1276, + 1217, + 1299, + 1195, + 1299 + ], + "score": 0.77, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 756, + 1813, + 774, + 1813, + 774, + 1839, + 756, + 1839 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1124, + 1783, + 1148, + 1783, + 1148, + 1809, + 1124, + 1809 + ], + "score": 0.77, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1696, + 1073, + 1696, + 1073, + 1718, + 1053, + 1718 + ], + "score": 0.76, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1151, + 1726, + 1171, + 1726, + 1171, + 1748, + 1151, + 1748 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 973, + 1247, + 992, + 1247, + 992, + 1268, + 973, + 1268 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1066, + 1559, + 1086, + 1559, + 1086, + 1579, + 1066, + 1579 + ], + "score": 0.74, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1001, + 1151, + 1026, + 1151, + 1026, + 1177, + 1001, + 1177 + ], + "score": 0.73, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 1311, + 1526, + 1329, + 1526, + 1329, + 1552, + 1311, + 1552 + ], + "score": 0.72, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 571, + 1380, + 631, + 1380, + 631, + 1409, + 571, + 1409 + ], + "score": 0.58, + "latex": "R , P" + }, + { + "category_id": 13, + "poly": [ + 1316, + 1440, + 1346, + 1440, + 1346, + 1467, + 1316, + 1467 + ], + "score": 0.44, + "latex": "\\mathbb { M }" + }, + { + "category_id": 13, + "poly": [ + 606, + 1381, + 631, + 1381, + 631, + 1408, + 606, + 1408 + ], + "score": 0.33, + "latex": "P" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 862.0, + 582.0, + 862.0, + 582.0, + 906.0, + 292.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1999.0, + 1376.0, + 1999.0, + 1376.0, + 2038.0, + 330.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 563.0, + 1405.0, + 563.0, + 1405.0, + 605.0, + 293.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 596.0, + 1406.0, + 596.0, + 1406.0, + 635.0, + 292.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 628.0, + 1406.0, + 628.0, + 1406.0, + 665.0, + 295.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 656.0, + 1407.0, + 656.0, + 1407.0, + 695.0, + 295.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 690.0, + 1406.0, + 690.0, + 1406.0, + 723.0, + 296.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 720.0, + 1405.0, + 720.0, + 1405.0, + 753.0, + 293.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 749.0, + 1407.0, + 749.0, + 1407.0, + 785.0, + 292.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 779.0, + 1406.0, + 779.0, + 1406.0, + 814.0, + 295.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 812.0, + 1147.0, + 812.0, + 1147.0, + 844.0, + 296.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1116.0, + 1404.0, + 1116.0, + 1404.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1149.0, + 600.0, + 1149.0, + 600.0, + 1181.0, + 296.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1149.0, + 1000.0, + 1149.0, + 1000.0, + 1181.0, + 621.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1149.0, + 1403.0, + 1149.0, + 1403.0, + 1181.0, + 1027.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 341.0, + 1179.0, + 341.0, + 1214.0, + 294.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1179.0, + 419.0, + 1179.0, + 419.0, + 1214.0, + 370.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1179.0, + 1139.0, + 1179.0, + 1139.0, + 1214.0, + 446.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1179.0, + 1193.0, + 1179.0, + 1193.0, + 1214.0, + 1184.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1214.0, + 1236.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1212.0, + 923.0, + 1212.0, + 923.0, + 1244.0, + 296.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1212.0, + 1007.0, + 1212.0, + 1007.0, + 1244.0, + 952.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1212.0, + 1404.0, + 1212.0, + 1404.0, + 1244.0, + 1048.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1236.0, + 972.0, + 1236.0, + 972.0, + 1284.0, + 291.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1236.0, + 1356.0, + 1236.0, + 1356.0, + 1284.0, + 993.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1236.0, + 1408.0, + 1236.0, + 1408.0, + 1284.0, + 1401.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1269.0, + 344.0, + 1269.0, + 344.0, + 1305.0, + 296.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1269.0, + 1194.0, + 1269.0, + 1194.0, + 1305.0, + 385.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1269.0, + 1365.0, + 1269.0, + 1365.0, + 1305.0, + 1218.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 1323.0, + 917.0, + 1323.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 947.0, + 295.0, + 947.0, + 295.0, + 985.0, + 292.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 947.0, + 683.0, + 947.0, + 683.0, + 985.0, + 469.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 947.0, + 901.0, + 947.0, + 901.0, + 985.0, + 712.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 947.0, + 1140.0, + 947.0, + 1140.0, + 985.0, + 927.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 947.0, + 1406.0, + 947.0, + 1406.0, + 985.0, + 1167.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 979.0, + 564.0, + 979.0, + 564.0, + 1015.0, + 322.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 979.0, + 788.0, + 979.0, + 788.0, + 1015.0, + 676.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 979.0, + 1014.0, + 979.0, + 1014.0, + 1015.0, + 881.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 979.0, + 1408.0, + 979.0, + 1408.0, + 1015.0, + 1093.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1009.0, + 892.0, + 1009.0, + 892.0, + 1046.0, + 292.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1046.0, + 951.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1041.0, + 480.0, + 1041.0, + 480.0, + 1077.0, + 379.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1041.0, + 1151.0, + 1041.0, + 1151.0, + 1077.0, + 539.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1041.0, + 1407.0, + 1041.0, + 1407.0, + 1077.0, + 1186.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1071.0, + 604.0, + 1071.0, + 604.0, + 1111.0, + 294.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1071.0, + 1254.0, + 1071.0, + 1254.0, + 1111.0, + 939.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1071.0, + 1299.0, + 1071.0, + 1299.0, + 1111.0, + 1287.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 306.0, + 1406.0, + 306.0, + 1406.0, + 343.0, + 295.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 335.0, + 1406.0, + 335.0, + 1406.0, + 375.0, + 293.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 1404.0, + 367.0, + 1404.0, + 403.0, + 293.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1402.0, + 400.0, + 1402.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 429.0, + 1405.0, + 429.0, + 1405.0, + 465.0, + 293.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1406.0, + 461.0, + 1406.0, + 494.0, + 294.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 490.0, + 1406.0, + 490.0, + 1406.0, + 524.0, + 292.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 521.0, + 922.0, + 521.0, + 922.0, + 556.0, + 293.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1658.0, + 1044.0, + 1658.0, + 1044.0, + 1693.0, + 294.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1658.0, + 1409.0, + 1658.0, + 1409.0, + 1693.0, + 1068.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1688.0, + 417.0, + 1688.0, + 417.0, + 1724.0, + 294.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1688.0, + 1052.0, + 1688.0, + 1052.0, + 1724.0, + 439.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1688.0, + 1124.0, + 1688.0, + 1124.0, + 1724.0, + 1074.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1688.0, + 1406.0, + 1688.0, + 1406.0, + 1724.0, + 1145.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1721.0, + 1150.0, + 1721.0, + 1150.0, + 1753.0, + 296.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1721.0, + 1232.0, + 1721.0, + 1232.0, + 1753.0, + 1172.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1721.0, + 1402.0, + 1721.0, + 1402.0, + 1753.0, + 1254.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 376.0, + 1749.0, + 376.0, + 1785.0, + 294.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1749.0, + 1139.0, + 1749.0, + 1139.0, + 1785.0, + 410.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1749.0, + 1404.0, + 1749.0, + 1404.0, + 1785.0, + 1318.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1777.0, + 298.0, + 1777.0, + 298.0, + 1819.0, + 293.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1777.0, + 1123.0, + 1777.0, + 1123.0, + 1819.0, + 453.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1777.0, + 1183.0, + 1777.0, + 1183.0, + 1819.0, + 1149.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1777.0, + 1405.0, + 1777.0, + 1405.0, + 1819.0, + 1211.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1809.0, + 476.0, + 1809.0, + 476.0, + 1846.0, + 293.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1809.0, + 755.0, + 1809.0, + 755.0, + 1846.0, + 500.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1809.0, + 1144.0, + 1809.0, + 1144.0, + 1846.0, + 775.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1515.0, + 1056.0, + 1515.0, + 1056.0, + 1563.0, + 289.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1515.0, + 1310.0, + 1515.0, + 1310.0, + 1563.0, + 1167.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 1515.0, + 1410.0, + 1515.0, + 1410.0, + 1563.0, + 1330.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 1065.0, + 1553.0, + 1065.0, + 1586.0, + 295.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1553.0, + 1374.0, + 1553.0, + 1374.0, + 1586.0, + 1087.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1553.0, + 1407.0, + 1553.0, + 1407.0, + 1586.0, + 1398.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1577.0, + 821.0, + 1577.0, + 821.0, + 1620.0, + 292.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1577.0, + 1405.0, + 1577.0, + 1405.0, + 1620.0, + 841.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1610.0, + 1185.0, + 1610.0, + 1185.0, + 1650.0, + 292.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1610.0, + 1216.0, + 1610.0, + 1216.0, + 1650.0, + 1204.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1316.0, + 399.0, + 1316.0, + 399.0, + 1353.0, + 293.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1316.0, + 941.0, + 1316.0, + 941.0, + 1353.0, + 423.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1316.0, + 1405.0, + 1316.0, + 1405.0, + 1353.0, + 961.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1350.0, + 910.0, + 1350.0, + 910.0, + 1382.0, + 296.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1382.0, + 935.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1377.0, + 570.0, + 1377.0, + 570.0, + 1413.0, + 293.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1377.0, + 685.0, + 1377.0, + 685.0, + 1413.0, + 632.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1413.0, + 708.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1409.0, + 584.0, + 1409.0, + 584.0, + 1444.0, + 295.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1409.0, + 1407.0, + 1409.0, + 1407.0, + 1444.0, + 608.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1439.0, + 884.0, + 1439.0, + 884.0, + 1474.0, + 292.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1439.0, + 1315.0, + 1439.0, + 1315.0, + 1474.0, + 909.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1439.0, + 1406.0, + 1439.0, + 1406.0, + 1474.0, + 1347.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 296.0, + 1467.0, + 296.0, + 1507.0, + 293.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1467.0, + 927.0, + 1467.0, + 927.0, + 1507.0, + 340.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1467.0, + 961.0, + 1467.0, + 961.0, + 1507.0, + 951.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 262.0, + 1407.0, + 262.0, + 1407.0, + 294.0, + 296.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1914.0, + 1179.0, + 1914.0, + 1179.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 1914.0, + 1407.0, + 1914.0, + 1407.0, + 1953.0, + 1349.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1945.0, + 1402.0, + 1945.0, + 1402.0, + 1985.0, + 295.0, + 1985.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 700, + 1406, + 700, + 1406, + 949, + 296, + 949 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 960, + 1406, + 960, + 1406, + 1147, + 297, + 1147 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1568, + 1404, + 1568, + 1404, + 1754, + 297, + 1754 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 1407, + 227, + 1407, + 459, + 296, + 459 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1766, + 1404, + 1766, + 1404, + 1954, + 297, + 1954 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1266, + 1405, + 1266, + 1405, + 1421, + 297, + 1421 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 564, + 1402, + 564, + 1402, + 628, + 298, + 628 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 464, + 1161, + 1232, + 1161, + 1232, + 1252, + 464, + 1252 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 462, + 472, + 1235, + 472, + 1235, + 552, + 462, + 552 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 295, + 1436, + 1403, + 1436, + 1403, + 1499, + 295, + 1499 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 295, + 1963, + 1397, + 1963, + 1397, + 2039, + 295, + 2039 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 459, + 1513, + 1237, + 1513, + 1237, + 1559, + 459, + 1559 + ], + "score": 0.928 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 301, + 647, + 769, + 647, + 769, + 684, + 301, + 684 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1366, + 493, + 1400, + 493, + 1400, + 523, + 1366, + 523 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1169, + 1400, + 1169, + 1400, + 1199, + 1366, + 1199 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.622 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.593 + }, + { + "category_id": 14, + "poly": [ + 462, + 469, + 1235, + 469, + 1235, + 552, + 462, + 552 + ], + "score": 0.93, + "latex": "\\ell _ { \\theta } ( x , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) } { \\exp ( \\lambda s _ { \\theta } ( x , y ) ) + \\sum _ { x ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\left\\{ x \\right\\} } \\exp ( \\lambda s _ { \\theta } ( x ^ { \\prime } , y ) ) }" + }, + { + "category_id": 13, + "poly": [ + 393, + 595, + 514, + 595, + 514, + 629, + 393, + 629 + ], + "score": 0.93, + "latex": "\\ell _ { \\theta } ( x , y ; S )" + }, + { + "category_id": 13, + "poly": [ + 722, + 1890, + 900, + 1890, + 900, + 1924, + 722, + 1924 + ], + "score": 0.93, + "latex": "\\tilde { \\pi } ( y ) = \\pi _ { \\chi } ^ { * } ( \\tilde { x } _ { y } )" + }, + { + "category_id": 13, + "poly": [ + 622, + 326, + 955, + 326, + 955, + 362, + 622, + 362 + ], + "score": 0.93, + "latex": "s _ { \\theta } ( x , y ) = s i m ( z _ { \\theta } ( x ) , z _ { \\theta } ( y ) )" + }, + { + "category_id": 14, + "poly": [ + 464, + 1158, + 1234, + 1158, + 1234, + 1254, + 464, + 1254 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } & { } & { d ^ { * } ( x , y ) = \\underbrace { \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) } _ { \\mathrm { ( A ) } } + \\underbrace { \\gamma \\mathcal { W } _ { 1 } ( d ^ { * } ) \\big ( P ^ { \\pi ^ { * } } ( \\cdot | x ) , P ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) } _ { \\mathrm { ( B ) } } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 753, + 2002, + 841, + 2002, + 841, + 2040, + 753, + 2040 + ], + "score": 0.93, + "latex": "Y _ { y } ^ { 0 } = y" + }, + { + "category_id": 13, + "poly": [ + 1042, + 322, + 1268, + 322, + 1268, + 370, + 1042, + 370 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { s i m ( u , v ) = \\frac { u ^ { T } v } { \\| u \\| \\| v \\| } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 963, + 1964, + 1175, + 1964, + 1175, + 2003, + 963, + 2003 + ], + "score": 0.92, + "latex": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 955, + 1831, + 1027, + 1831, + 1027, + 1861, + 955, + 1861 + ], + "score": 0.92, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1860, + 609, + 1860, + 609, + 1894, + 297, + 1894 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\tilde { x } _ { y } : = \\arg \\operatorname* { m i n } _ { x \\in \\mathcal { X } } d ^ { * } ( x , y ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 737, + 1115, + 924, + 1115, + 924, + 1144, + 737, + 1144 + ], + "score": 0.9, + "latex": "d ^ { * } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1860, + 1253, + 1860, + 1253, + 1893, + 1199, + 1893 + ], + "score": 0.9, + "latex": "\\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 1126, + 1861, + 1165, + 1861, + 1165, + 1893, + 1126, + 1893 + ], + "score": 0.9, + "latex": "\\pi _ { \\mathcal { X } } ^ { * }" + }, + { + "category_id": 14, + "poly": [ + 461, + 1511, + 1237, + 1511, + 1237, + 1558, + 461, + 1558 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 344, + 1922, + 380, + 1922, + 380, + 1958, + 344, + 1958 + ], + "score": 0.9, + "latex": "\\pi _ { \\mathcal { Y } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1269, + 1181, + 1269, + 1181, + 1298, + 1138, + 1298 + ], + "score": 0.89, + "latex": "\\mathcal { W } _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 887, + 426, + 921, + 426, + 921, + 453, + 887, + 453 + ], + "score": 0.88, + "latex": "\\mathcal { X } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 953, + 1693, + 985, + 1693, + 985, + 1720, + 953, + 1720 + ], + "score": 0.88, + "latex": "\\hat { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 752, + 1924, + 780, + 1924, + 780, + 1947, + 752, + 1947 + ], + "score": 0.87, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1600, + 1082, + 1600, + 1082, + 1630, + 1053, + 1630 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 845, + 1967, + 916, + 1967, + 916, + 1999, + 845, + 1999 + ], + "score": 0.87, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 698, + 596, + 732, + 596, + 732, + 623, + 698, + 623 + ], + "score": 0.87, + "latex": "\\mathcal { X } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 589, + 431, + 635, + 431, + 635, + 457, + 589, + 457 + ], + "score": 0.87, + "latex": "x , y" + }, + { + "category_id": 13, + "poly": [ + 498, + 1389, + 528, + 1389, + 528, + 1416, + 498, + 1416 + ], + "score": 0.87, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 695, + 1055, + 728, + 1055, + 728, + 1082, + 695, + 1082 + ], + "score": 0.86, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1662, + 1244, + 1662, + 1244, + 1689, + 1212, + 1689 + ], + "score": 0.86, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 970, + 1723, + 1001, + 1723, + 1001, + 1750, + 970, + 1750 + ], + "score": 0.86, + "latex": "\\hat { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1330, + 817, + 1330, + 817, + 1356, + 788, + 1356 + ], + "score": 0.86, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 708, + 1601, + 734, + 1601, + 734, + 1628, + 708, + 1628 + ], + "score": 0.81, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 373, + 567, + 392, + 567, + 392, + 593, + 373, + 593 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 568, + 1600, + 631, + 1600, + 631, + 1631, + 568, + 1631 + ], + "score": 0.8, + "latex": "( T V )" + }, + { + "category_id": 13, + "poly": [ + 966, + 1926, + 984, + 1926, + 984, + 1948, + 966, + 1948 + ], + "score": 0.8, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1831, + 1367, + 1831, + 1367, + 1857, + 1340, + 1857 + ], + "score": 0.8, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1380, + 1892, + 1401, + 1892, + 1401, + 1918, + 1380, + 1918 + ], + "score": 0.79, + "latex": "\\tilde { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 664, + 1631, + 690, + 1631, + 690, + 1658, + 664, + 1658 + ], + "score": 0.79, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 900, + 890, + 919, + 890, + 919, + 914, + 900, + 914 + ], + "score": 0.79, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1230, + 767, + 1250, + 767, + 1250, + 792, + 1230, + 792 + ], + "score": 0.77, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1309, + 297, + 1327, + 297, + 1327, + 319, + 1309, + 319 + ], + "score": 0.76, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1080, + 2006, + 1099, + 2006, + 1099, + 2030, + 1080, + 2030 + ], + "score": 0.76, + "latex": "\\tilde { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1379, + 299, + 1397, + 299, + 1397, + 324, + 1379, + 324 + ], + "score": 0.76, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 513, + 1773, + 533, + 1773, + 533, + 1797, + 513, + 1797 + ], + "score": 0.74, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 592, + 600, + 640, + 600, + 640, + 628, + 592, + 628 + ], + "score": 0.59, + "latex": "x , y" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 645.0, + 772.0, + 645.0, + 772.0, + 689.0, + 292.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1406.0, + 702.0, + 1406.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 730.0, + 1404.0, + 730.0, + 1404.0, + 766.0, + 294.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 764.0, + 1229.0, + 764.0, + 1229.0, + 798.0, + 295.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 764.0, + 1406.0, + 764.0, + 1406.0, + 798.0, + 1251.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 795.0, + 1404.0, + 795.0, + 1404.0, + 826.0, + 296.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 825.0, + 1406.0, + 825.0, + 1406.0, + 857.0, + 293.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 852.0, + 1406.0, + 852.0, + 1406.0, + 889.0, + 294.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 884.0, + 899.0, + 884.0, + 899.0, + 920.0, + 294.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 884.0, + 1406.0, + 884.0, + 1406.0, + 920.0, + 920.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 929.0, + 917.0, + 929.0, + 949.0, + 295.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 1404.0, + 959.0, + 1404.0, + 997.0, + 294.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 1408.0, + 992.0, + 1408.0, + 1028.0, + 295.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1021.0, + 1406.0, + 1021.0, + 1406.0, + 1060.0, + 294.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1054.0, + 694.0, + 1054.0, + 694.0, + 1089.0, + 294.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1054.0, + 1404.0, + 1054.0, + 1404.0, + 1089.0, + 729.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1084.0, + 1406.0, + 1084.0, + 1406.0, + 1119.0, + 292.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1114.0, + 736.0, + 1114.0, + 736.0, + 1150.0, + 294.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1114.0, + 1272.0, + 1114.0, + 1272.0, + 1150.0, + 925.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1568.0, + 1404.0, + 1568.0, + 1404.0, + 1602.0, + 295.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1597.0, + 567.0, + 1597.0, + 567.0, + 1634.0, + 292.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1597.0, + 707.0, + 1597.0, + 707.0, + 1634.0, + 632.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1597.0, + 1052.0, + 1597.0, + 1052.0, + 1634.0, + 735.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1597.0, + 1405.0, + 1597.0, + 1405.0, + 1634.0, + 1083.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 663.0, + 1632.0, + 663.0, + 1664.0, + 295.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1632.0, + 1402.0, + 1632.0, + 1402.0, + 1664.0, + 691.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1661.0, + 1211.0, + 1661.0, + 1211.0, + 1694.0, + 292.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1661.0, + 1404.0, + 1661.0, + 1404.0, + 1694.0, + 1245.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1692.0, + 952.0, + 1692.0, + 952.0, + 1725.0, + 292.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1692.0, + 1405.0, + 1692.0, + 1405.0, + 1725.0, + 986.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1721.0, + 969.0, + 1721.0, + 969.0, + 1757.0, + 292.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1721.0, + 1213.0, + 1721.0, + 1213.0, + 1757.0, + 1002.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 259.0, + 1409.0, + 259.0, + 1409.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 289.0, + 1308.0, + 289.0, + 1308.0, + 330.0, + 291.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 289.0, + 1378.0, + 289.0, + 1378.0, + 330.0, + 1328.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 289.0, + 1410.0, + 289.0, + 1410.0, + 330.0, + 1398.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 318.0, + 621.0, + 318.0, + 621.0, + 377.0, + 293.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 318.0, + 1041.0, + 318.0, + 1041.0, + 377.0, + 956.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 318.0, + 1407.0, + 318.0, + 1407.0, + 377.0, + 1269.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 400.0, + 295.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 394.0, + 1404.0, + 394.0, + 1404.0, + 430.0, + 294.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 423.0, + 588.0, + 423.0, + 588.0, + 461.0, + 294.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 423.0, + 886.0, + 423.0, + 886.0, + 461.0, + 636.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 423.0, + 1231.0, + 423.0, + 1231.0, + 461.0, + 922.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1769.0, + 512.0, + 1769.0, + 512.0, + 1801.0, + 296.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1769.0, + 1404.0, + 1769.0, + 1404.0, + 1801.0, + 534.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1798.0, + 1405.0, + 1798.0, + 1405.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1824.0, + 954.0, + 1824.0, + 954.0, + 1868.0, + 291.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1824.0, + 1339.0, + 1824.0, + 1339.0, + 1868.0, + 1028.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1824.0, + 1406.0, + 1824.0, + 1406.0, + 1868.0, + 1368.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1856.0, + 296.0, + 1856.0, + 296.0, + 1904.0, + 288.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1856.0, + 1125.0, + 1856.0, + 1125.0, + 1904.0, + 610.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1856.0, + 1198.0, + 1856.0, + 1198.0, + 1904.0, + 1166.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1856.0, + 1410.0, + 1856.0, + 1410.0, + 1904.0, + 1254.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1888.0, + 721.0, + 1888.0, + 721.0, + 1928.0, + 292.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1888.0, + 1379.0, + 1888.0, + 1379.0, + 1928.0, + 901.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1888.0, + 1407.0, + 1888.0, + 1407.0, + 1928.0, + 1402.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1919.0, + 343.0, + 1919.0, + 343.0, + 1959.0, + 294.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1919.0, + 751.0, + 1919.0, + 751.0, + 1959.0, + 381.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1919.0, + 965.0, + 1919.0, + 965.0, + 1959.0, + 781.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 1919.0, + 1223.0, + 1919.0, + 1223.0, + 1959.0, + 985.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1266.0, + 1137.0, + 1266.0, + 1137.0, + 1300.0, + 296.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1266.0, + 1402.0, + 1266.0, + 1402.0, + 1300.0, + 1182.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1298.0, + 1405.0, + 1298.0, + 1405.0, + 1332.0, + 296.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1329.0, + 787.0, + 1329.0, + 787.0, + 1362.0, + 296.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1362.0, + 818.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1358.0, + 1406.0, + 1358.0, + 1406.0, + 1391.0, + 296.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1389.0, + 497.0, + 1389.0, + 497.0, + 1422.0, + 296.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1389.0, + 838.0, + 1389.0, + 838.0, + 1422.0, + 529.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 564.0, + 372.0, + 564.0, + 372.0, + 600.0, + 294.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 564.0, + 1404.0, + 564.0, + 1404.0, + 600.0, + 393.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 595.0, + 392.0, + 595.0, + 392.0, + 631.0, + 296.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 595.0, + 591.0, + 595.0, + 591.0, + 631.0, + 515.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 595.0, + 697.0, + 595.0, + 697.0, + 631.0, + 641.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 595.0, + 1340.0, + 595.0, + 1340.0, + 631.0, + 733.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1436.0, + 1404.0, + 1436.0, + 1404.0, + 1471.0, + 297.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1465.0, + 1406.0, + 1465.0, + 1406.0, + 1502.0, + 294.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1956.0, + 844.0, + 1956.0, + 844.0, + 2016.0, + 289.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1956.0, + 962.0, + 1956.0, + 962.0, + 2016.0, + 917.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1956.0, + 1408.0, + 1956.0, + 1408.0, + 2016.0, + 1176.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1991.0, + 752.0, + 1991.0, + 752.0, + 2048.0, + 291.0, + 2048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1991.0, + 1079.0, + 1991.0, + 1079.0, + 2048.0, + 842.0, + 2048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1991.0, + 1224.0, + 1991.0, + 1224.0, + 2048.0, + 1100.0, + 2048.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 343, + 1407, + 343, + 1407, + 559, + 296, + 559 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 631, + 770, + 631, + 770, + 939, + 297, + 939 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 1705, + 1405, + 1705, + 1405, + 1830, + 296, + 1830 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1911, + 1404, + 1911, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1280, + 1405, + 1280, + 1405, + 1375, + 298, + 1375 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 952, + 1408, + 952, + 1408, + 1048, + 297, + 1048 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 532, + 220, + 1162, + 220, + 1162, + 330, + 532, + 330 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 1478, + 1404, + 1478, + 1404, + 1543, + 297, + 1543 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 303, + 1386, + 1364, + 1386, + 1364, + 1468, + 303, + 1468 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 297, + 1554, + 1404, + 1554, + 1404, + 1628, + 297, + 1628 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 801, + 684, + 1405, + 684, + 1405, + 930, + 801, + 930 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 423, + 1636, + 1273, + 1636, + 1273, + 1696, + 423, + 1696 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 298, + 1856, + 1009, + 1856, + 1009, + 1894, + 298, + 1894 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 299, + 578, + 1048, + 578, + 1048, + 613, + 299, + 613 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 105, + 298, + 105 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 295, + 1060, + 1406, + 1060, + 1406, + 1268, + 295, + 1268 + ], + "score": 0.871 + }, + { + "category_id": 9, + "poly": [ + 1375, + 1407, + 1417, + 1407, + 1417, + 1440, + 1375, + 1440 + ], + "score": 0.826 + }, + { + "category_id": 0, + "poly": [ + 791, + 640, + 1399, + 640, + 1399, + 674, + 791, + 674 + ], + "score": 0.815 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.778 + }, + { + "category_id": 13, + "poly": [ + 1177, + 1553, + 1349, + 1553, + 1349, + 1591, + 1177, + 1591 + ], + "score": 0.94, + "latex": "\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }" + }, + { + "category_id": 14, + "poly": [ + 293, + 1385, + 1368, + 1385, + 1368, + 1468, + 293, + 1468 + ], + "score": 0.94, + "latex": "\\ell _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ; \\mathcal { X } ^ { \\prime } ) = - \\log \\frac { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) } { \\Gamma ( \\tilde { x } _ { y } , y ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\tilde { x } _ { y } , y ) ) + \\sum _ { \\boldsymbol { x } ^ { \\prime } \\in \\mathcal { X } ^ { \\prime } \\backslash \\{ \\tilde { x } _ { y } \\} } ( 1 - \\Gamma ( \\boldsymbol { x } ^ { \\prime } , y ) ) \\exp ( \\lambda s _ { \\boldsymbol { \\theta } } ( \\boldsymbol { x } ^ { \\prime } , y ) ) }" + }, + { + "category_id": 14, + "poly": [ + 532, + 220, + 1163, + 220, + 1163, + 331, + 532, + 331 + ], + "score": 0.94, + "latex": "\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) ." + }, + { + "category_id": 13, + "poly": [ + 891, + 1122, + 998, + 1122, + 998, + 1157, + 891, + 1157 + ], + "score": 0.93, + "latex": "\\{ ( \\tilde { x } _ { y } , y ) \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1588, + 455, + 1588, + 455, + 1628, + 297, + 1628 + ], + "score": 0.93, + "latex": "\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 806, + 1588, + 1015, + 1588, + 1015, + 1629, + 806, + 1629 + ], + "score": 0.93, + "latex": "y _ { t + 1 } \\sim P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1150, + 344, + 1398, + 344, + 1398, + 378, + 1150, + 378 + ], + "score": 0.93, + "latex": "d ^ { * } ( \\tilde { x } _ { y } , y ) < 1 / ( 1 + \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 541, + 1589, + 756, + 1589, + 756, + 1627, + 541, + 1627 + ], + "score": 0.92, + "latex": "x _ { t + 1 } \\sim P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1591, + 1211, + 1591, + 1211, + 1626, + 1107, + 1626 + ], + "score": 0.92, + "latex": "\\mathcal { X } ^ { \\prime } = \\tau _ { \\mathcal { X } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1050, + 716, + 1158, + 716, + 1158, + 749, + 1050, + 749 + ], + "score": 0.92, + "latex": "\\{ \\mathcal { M } _ { i } \\} _ { i = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1318, + 1312, + 1396, + 1312, + 1396, + 1346, + 1318, + 1346 + ], + "score": 0.92, + "latex": "( \\tilde { x } _ { y } , y )" + }, + { + "category_id": 13, + "poly": [ + 1085, + 1061, + 1183, + 1061, + 1183, + 1092, + 1085, + 1092 + ], + "score": 0.91, + "latex": "{ \\mathcal { X } } ^ { \\prime } \\subseteq { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 824, + 1093, + 899, + 1093, + 899, + 1123, + 824, + 1123 + ], + "score": 0.91, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1204, + 1125, + 1204, + 1125, + 1234, + 1069, + 1234 + ], + "score": 0.91, + "latex": "\\mathcal { M } _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 731, + 1235, + 784, + 1235, + 784, + 1268, + 731, + 1268 + ], + "score": 0.9, + "latex": "\\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 751, + 1557, + 806, + 1557, + 806, + 1589, + 751, + 1589 + ], + "score": 0.9, + "latex": "\\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 444, + 375, + 601, + 375, + 601, + 408, + 444, + 408 + ], + "score": 0.9, + "latex": "d ^ { * } ( \\tilde { x } _ { y } , y ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1264, + 685, + 1330, + 685, + 1330, + 719, + 1264, + 719 + ], + "score": 0.9, + "latex": "d ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 379, + 1014, + 693, + 1014, + 693, + 1049, + 379, + 1049 + ], + "score": 0.9, + "latex": "\\Gamma ( x , y ) = \\exp ( - d ( x , y ) / \\beta )" + }, + { + "category_id": 13, + "poly": [ + 638, + 1557, + 695, + 1557, + 695, + 1587, + 638, + 1587 + ], + "score": 0.9, + "latex": "\\mathcal { M } _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1768, + 1180, + 1768, + 1180, + 1799, + 1100, + 1799 + ], + "score": 0.9, + "latex": "{ \\mathcal { L } } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 1114, + 839, + 1193, + 839, + 1193, + 868, + 1114, + 868 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1092, + 417, + 1092, + 417, + 1125, + 298, + 1125 + ], + "score": 0.89, + "latex": "\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 685, + 1172, + 685, + 1172, + 717, + 1116, + 717 + ], + "score": 0.88, + "latex": "z _ { \\theta } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 985, + 746, + 1032, + 746, + 1032, + 779, + 985, + 779 + ], + "score": 0.88, + "latex": "1 / \\lambda" + }, + { + "category_id": 14, + "poly": [ + 423, + 1637, + 1273, + 1637, + 1273, + 1698, + 423, + 1698 + ], + "score": 0.88, + "latex": "L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) = \\mathbb { E } _ { y \\sim \\tau _ { \\mathcal { Y } } ^ { * } } \\left[ \\ell _ { \\theta } ( \\tilde { x } _ { y } , y ; \\tau _ { \\mathcal { X } } ^ { * } ) \\right] \\qquad \\mathrm { w h e r e } \\ \\tilde { x } _ { y } = \\underset { x \\in \\tau _ { \\mathcal { X } } ^ { * } } { \\mathrm { a r g m a x } } \\ \\Gamma ( x , y ) ." + }, + { + "category_id": 13, + "poly": [ + 376, + 1343, + 410, + 1343, + 410, + 1370, + 376, + 1370 + ], + "score": 0.88, + "latex": "\\mathcal { X } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 825, + 406, + 864, + 406, + 864, + 441, + 825, + 441 + ], + "score": 0.87, + "latex": "\\pi _ { \\mathcal { D } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 991, + 779, + 1147, + 779, + 1147, + 809, + 991, + 809 + ], + "score": 0.87, + "latex": "k = 1 , \\ldots , K" + }, + { + "category_id": 13, + "poly": [ + 518, + 1174, + 552, + 1174, + 552, + 1201, + 518, + 1201 + ], + "score": 0.86, + "latex": "\\mathcal { X } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1258, + 1092, + 1292, + 1092, + 1292, + 1119, + 1258, + 1119 + ], + "score": 0.86, + "latex": "\\mathcal { X } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 536, + 408, + 565, + 408, + 565, + 432, + 536, + 432 + ], + "score": 0.84, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 557, + 1514, + 585, + 1514, + 585, + 1540, + 557, + 1540 + ], + "score": 0.84, + "latex": "z _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1110, + 750, + 1129, + 750, + 1129, + 778, + 1110, + 778 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 298, + 1317, + 325, + 1317, + 325, + 1342, + 298, + 1342 + ], + "score": 0.83, + "latex": "z _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1377, + 985, + 1397, + 985, + 1397, + 1015, + 1377, + 1015 + ], + "score": 0.82, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 866, + 869, + 1321, + 869, + 1321, + 903, + 866, + 903 + ], + "score": 0.81, + "latex": "\\mathcal { L } _ { \\mathrm { C M E } } = \\mathbb { E } _ { \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } \\sim \\rho } \\left[ L _ { \\theta } ( \\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } } ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 1227, + 808, + 1347, + 808, + 1347, + 840, + 1227, + 840 + ], + "score": 0.8, + "latex": "\\mathcal { M } _ { \\mathcal { X } } , \\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1062, + 1260, + 1062, + 1260, + 1091, + 1235, + 1091 + ], + "score": 0.8, + "latex": "\\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 698, + 1180, + 716, + 1180, + 716, + 1205, + 698, + 1205 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 706, + 1015, + 728, + 1015, + 728, + 1047, + 706, + 1047 + ], + "score": 0.8, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 728, + 985, + 747, + 985, + 747, + 1011, + 728, + 1011 + ], + "score": 0.8, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 954, + 840, + 971, + 840, + 971, + 866, + 954, + 866 + ], + "score": 0.79, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1359, + 747, + 1389, + 747, + 1389, + 774, + 1359, + 774 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1300, + 1016, + 1317, + 1016, + 1317, + 1042, + 1300, + 1042 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 774, + 955, + 792, + 955, + 792, + 981, + 774, + 981 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 569, + 817, + 588, + 817, + 588, + 844, + 569, + 844 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1006, + 437, + 1028, + 437, + 1028, + 463, + 1006, + 463 + ], + "score": 0.75, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1087, + 1123, + 1337, + 1123, + 1337, + 1174, + 1087, + 1174 + ], + "score": 0.74, + "latex": "\\tilde { x } _ { y } = \\underset { x \\in \\mathcal { X } ^ { \\prime } } { \\operatorname { a r g m a x } } \\Gamma ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 580, + 725, + 598, + 725, + 598, + 752, + 580, + 752 + ], + "score": 0.69, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1254, + 1122, + 1338, + 1122, + 1338, + 1155, + 1254, + 1155 + ], + "score": 0.68, + "latex": "\\Gamma ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1708, + 1327, + 1708, + 1327, + 1738, + 1251, + 1738 + ], + "score": 0.67, + "latex": "( P S E s )" + }, + { + "category_id": 13, + "poly": [ + 721, + 1344, + 741, + 1344, + 741, + 1370, + 721, + 1370 + ], + "score": 0.64, + "latex": "\\Gamma" + }, + { + "category_id": 13, + "poly": [ + 557, + 1124, + 579, + 1124, + 579, + 1150, + 557, + 1150 + ], + "score": 0.6, + "latex": "\\Gamma" + }, + { + "category_id": 13, + "poly": [ + 298, + 816, + 316, + 816, + 316, + 843, + 298, + 843 + ], + "score": 0.54, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1061, + 955, + 1082, + 955, + 1082, + 981, + 1061, + 981 + ], + "score": 0.41, + "latex": "\\Gamma" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1856.0, + 1011.0, + 1856.0, + 1011.0, + 1897.0, + 294.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 574.0, + 1050.0, + 574.0, + 1050.0, + 620.0, + 293.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 634.0, + 1399.0, + 634.0, + 1399.0, + 683.0, + 787.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 837.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 342.0, + 1149.0, + 342.0, + 1149.0, + 381.0, + 294.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 342.0, + 1408.0, + 342.0, + 1408.0, + 381.0, + 1399.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 373.0, + 443.0, + 373.0, + 443.0, + 411.0, + 294.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 373.0, + 1408.0, + 373.0, + 1408.0, + 411.0, + 602.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 404.0, + 535.0, + 404.0, + 535.0, + 441.0, + 293.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 404.0, + 824.0, + 404.0, + 824.0, + 441.0, + 566.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 404.0, + 1407.0, + 404.0, + 1407.0, + 441.0, + 865.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 434.0, + 1005.0, + 434.0, + 1005.0, + 473.0, + 293.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 434.0, + 1408.0, + 434.0, + 1408.0, + 473.0, + 1029.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 465.0, + 1407.0, + 465.0, + 1407.0, + 502.0, + 293.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 497.0, + 1405.0, + 497.0, + 1405.0, + 531.0, + 294.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 526.0, + 1119.0, + 526.0, + 1119.0, + 565.0, + 295.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 632.0, + 776.0, + 632.0, + 776.0, + 666.0, + 297.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 662.0, + 776.0, + 662.0, + 776.0, + 696.0, + 294.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 693.0, + 772.0, + 693.0, + 772.0, + 726.0, + 294.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 724.0, + 579.0, + 724.0, + 579.0, + 758.0, + 295.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 724.0, + 773.0, + 724.0, + 773.0, + 758.0, + 599.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 756.0, + 772.0, + 756.0, + 772.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 784.0, + 772.0, + 784.0, + 772.0, + 817.0, + 295.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 813.0, + 297.0, + 813.0, + 297.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 813.0, + 568.0, + 813.0, + 568.0, + 849.0, + 317.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 813.0, + 772.0, + 813.0, + 772.0, + 849.0, + 589.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 846.0, + 772.0, + 846.0, + 772.0, + 879.0, + 295.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 772.0, + 877.0, + 772.0, + 910.0, + 294.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 907.0, + 758.0, + 907.0, + 758.0, + 941.0, + 295.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1705.0, + 1250.0, + 1705.0, + 1250.0, + 1742.0, + 294.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1705.0, + 1403.0, + 1705.0, + 1403.0, + 1742.0, + 1328.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1737.0, + 1405.0, + 1737.0, + 1405.0, + 1773.0, + 294.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1767.0, + 1099.0, + 1767.0, + 1099.0, + 1803.0, + 295.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1767.0, + 1405.0, + 1767.0, + 1405.0, + 1803.0, + 1181.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1797.0, + 1362.0, + 1797.0, + 1362.0, + 1834.0, + 293.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1908.0, + 1405.0, + 1908.0, + 1405.0, + 1949.0, + 292.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2041.0, + 293.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1281.0, + 1403.0, + 1281.0, + 1403.0, + 1315.0, + 296.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 297.0, + 1309.0, + 297.0, + 1349.0, + 293.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1309.0, + 1317.0, + 1309.0, + 1317.0, + 1349.0, + 326.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1309.0, + 1408.0, + 1309.0, + 1408.0, + 1349.0, + 1397.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1340.0, + 375.0, + 1340.0, + 375.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1340.0, + 720.0, + 1340.0, + 720.0, + 1380.0, + 411.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1340.0, + 1388.0, + 1340.0, + 1388.0, + 1380.0, + 742.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 950.0, + 773.0, + 950.0, + 773.0, + 988.0, + 292.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 950.0, + 1060.0, + 950.0, + 1060.0, + 988.0, + 793.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 950.0, + 1410.0, + 950.0, + 1410.0, + 988.0, + 1083.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 982.0, + 727.0, + 982.0, + 727.0, + 1018.0, + 294.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 982.0, + 1376.0, + 982.0, + 1376.0, + 1018.0, + 748.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 982.0, + 1408.0, + 982.0, + 1408.0, + 1018.0, + 1398.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1011.0, + 378.0, + 1011.0, + 378.0, + 1050.0, + 292.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1011.0, + 705.0, + 1011.0, + 705.0, + 1050.0, + 694.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1011.0, + 1299.0, + 1011.0, + 1299.0, + 1050.0, + 729.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1011.0, + 1329.0, + 1011.0, + 1329.0, + 1050.0, + 1318.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1478.0, + 1404.0, + 1478.0, + 1404.0, + 1514.0, + 295.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1508.0, + 556.0, + 1508.0, + 556.0, + 1545.0, + 294.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1508.0, + 1407.0, + 1508.0, + 1407.0, + 1545.0, + 586.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1543.0, + 637.0, + 1543.0, + 637.0, + 1603.0, + 288.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1543.0, + 750.0, + 1543.0, + 750.0, + 1603.0, + 696.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1543.0, + 1176.0, + 1543.0, + 1176.0, + 1603.0, + 807.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1543.0, + 1413.0, + 1543.0, + 1413.0, + 1603.0, + 1350.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1584.0, + 296.0, + 1584.0, + 296.0, + 1632.0, + 292.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1584.0, + 540.0, + 1584.0, + 540.0, + 1632.0, + 456.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1584.0, + 805.0, + 1584.0, + 805.0, + 1632.0, + 757.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1584.0, + 1106.0, + 1584.0, + 1106.0, + 1632.0, + 1016.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1584.0, + 1342.0, + 1584.0, + 1342.0, + 1632.0, + 1212.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 684.0, + 1115.0, + 684.0, + 1115.0, + 719.0, + 806.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 684.0, + 1263.0, + 684.0, + 1263.0, + 719.0, + 1173.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 684.0, + 1408.0, + 684.0, + 1408.0, + 719.0, + 1331.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 699.0, + 1049.0, + 699.0, + 1049.0, + 762.0, + 831.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 699.0, + 1406.0, + 699.0, + 1406.0, + 762.0, + 1159.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 746.0, + 984.0, + 746.0, + 984.0, + 780.0, + 837.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 746.0, + 1109.0, + 746.0, + 1109.0, + 780.0, + 1033.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 746.0, + 1358.0, + 746.0, + 1358.0, + 780.0, + 1130.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 779.0, + 830.0, + 779.0, + 830.0, + 805.0, + 807.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 778.0, + 990.0, + 778.0, + 990.0, + 809.0, + 835.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 778.0, + 1188.0, + 778.0, + 1188.0, + 809.0, + 1148.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 808.0, + 833.0, + 808.0, + 833.0, + 837.0, + 804.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 807.0, + 1226.0, + 807.0, + 1226.0, + 841.0, + 864.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 840.0, + 833.0, + 840.0, + 833.0, + 870.0, + 804.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 837.0, + 953.0, + 837.0, + 953.0, + 872.0, + 865.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 837.0, + 1113.0, + 837.0, + 1113.0, + 872.0, + 972.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 837.0, + 1274.0, + 837.0, + 1274.0, + 872.0, + 1194.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 863.0, + 1325.0, + 863.0, + 1325.0, + 909.0, + 1322.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 896.0, + 934.0, + 896.0, + 934.0, + 932.0, + 803.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1059.0, + 1084.0, + 1059.0, + 1084.0, + 1095.0, + 295.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1059.0, + 1234.0, + 1059.0, + 1234.0, + 1095.0, + 1184.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1059.0, + 1404.0, + 1059.0, + 1404.0, + 1095.0, + 1261.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1089.0, + 297.0, + 1089.0, + 297.0, + 1127.0, + 294.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1089.0, + 823.0, + 1089.0, + 823.0, + 1127.0, + 418.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1089.0, + 1257.0, + 1089.0, + 1257.0, + 1127.0, + 900.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1089.0, + 1406.0, + 1089.0, + 1406.0, + 1127.0, + 1293.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1121.0, + 556.0, + 1121.0, + 556.0, + 1162.0, + 293.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1121.0, + 890.0, + 1121.0, + 890.0, + 1162.0, + 580.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1121.0, + 1086.0, + 1121.0, + 1086.0, + 1162.0, + 999.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1121.0, + 1407.0, + 1121.0, + 1407.0, + 1162.0, + 1339.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1172.0, + 517.0, + 1172.0, + 517.0, + 1207.0, + 290.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1172.0, + 697.0, + 1172.0, + 697.0, + 1207.0, + 553.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1172.0, + 1405.0, + 1172.0, + 1405.0, + 1207.0, + 717.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1202.0, + 1068.0, + 1202.0, + 1068.0, + 1240.0, + 294.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1202.0, + 1406.0, + 1202.0, + 1406.0, + 1240.0, + 1126.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1235.0, + 730.0, + 1235.0, + 730.0, + 1269.0, + 295.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 1235.0, + 1196.0, + 1235.0, + 1196.0, + 1269.0, + 785.0, + 1269.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1209, + 1405, + 1209, + 1405, + 1484, + 297, + 1484 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1500, + 1405, + 1500, + 1405, + 1715, + 297, + 1715 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1770, + 1405, + 1770, + 1405, + 1955, + 298, + 1955 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1070, + 1405, + 1070, + 1405, + 1194, + 299, + 1194 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 296, + 864, + 1404, + 864, + 1404, + 1059, + 296, + 1059 + ], + "score": 0.965 + }, + { + "category_id": 6, + "poly": [ + 296, + 224, + 1405, + 224, + 1405, + 309, + 296, + 309 + ], + "score": 0.933 + }, + { + "category_id": 2, + "poly": [ + 297, + 1978, + 1404, + 1978, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 297, + 1726, + 971, + 1726, + 971, + 1756, + 297, + 1756 + ], + "score": 0.878 + }, + { + "category_id": 3, + "poly": [ + 309, + 327, + 1377, + 327, + 1377, + 851, + 309, + 851 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.761 + }, + { + "category_id": 5, + "poly": [ + 309, + 327, + 1377, + 327, + 1377, + 851, + 309, + 851 + ], + "score": 0.096, + "html": "
Data AugmentationMethodGrid Configuration (%)
“Wide”“Narrow”Random
XDropout and l2 reg.17.8 (2.2)10.2 (4.6)9.3 (5.4)
Bisimulation Transfer417.9 (0.0)17.9 (0.0)30.9 (4.2)
PSEs33.6 (10.0)9.3 (5.3)37.7 (10.4)
[ Obstacle PositionRandConv RandConv + Bisimulation RandConv+PSEs50.7 (24.2) 41.4 (17.6)33.7 (11.8)71.3 (15.6)
17.4 (6.7)33.4 (15.6)
87.0 (10.1)52.4 (5.8)83.4 (10.1)
Floor
" + }, + { + "category_id": 13, + "poly": [ + 1298, + 924, + 1315, + 924, + 1315, + 949, + 1298, + 949 + ], + "score": 0.79, + "latex": "_ y" + }, + { + "category_id": 13, + "poly": [ + 811, + 953, + 829, + 953, + 829, + 973, + 811, + 973 + ], + "score": 0.75, + "latex": "x -" + }, + { + "category_id": 13, + "poly": [ + 1167, + 948, + 1190, + 948, + 1190, + 973, + 1167, + 973 + ], + "score": 0.67, + "latex": "\\top" + }, + { + "category_id": 13, + "poly": [ + 509, + 226, + 547, + 226, + 547, + 252, + 509, + 252 + ], + "score": 0.67, + "latex": "( \\% )" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 864.0, + 1406.0, + 864.0, + 1406.0, + 897.0, + 295.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 893.0, + 1405.0, + 893.0, + 1405.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 918.0, + 1297.0, + 918.0, + 1297.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 918.0, + 1405.0, + 918.0, + 1405.0, + 951.0, + 1316.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 944.0, + 810.0, + 944.0, + 810.0, + 983.0, + 293.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 944.0, + 1166.0, + 944.0, + 1166.0, + 983.0, + 830.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 944.0, + 1406.0, + 944.0, + 1406.0, + 983.0, + 1191.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 972.0, + 1406.0, + 972.0, + 1406.0, + 1008.0, + 294.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1000.0, + 1405.0, + 1000.0, + 1405.0, + 1035.0, + 294.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1029.0, + 1385.0, + 1029.0, + 1385.0, + 1062.0, + 294.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 508.0, + 221.0, + 508.0, + 256.0, + 291.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 221.0, + 1405.0, + 221.0, + 1405.0, + 256.0, + 548.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 249.0, + 1406.0, + 249.0, + 1406.0, + 284.0, + 291.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 278.0, + 1126.0, + 278.0, + 1126.0, + 312.0, + 293.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2012.0, + 329.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1028.0, + 2004.0, + 1028.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1722.0, + 975.0, + 1722.0, + 975.0, + 1761.0, + 292.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 330.0, + 546.0, + 330.0, + 546.0, + 396.0, + 382.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 348.0, + 766.0, + 348.0, + 766.0, + 384.0, + 668.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 332.0, + 1228.0, + 332.0, + 1228.0, + 373.0, + 961.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 375.0, + 982.0, + 375.0, + 982.0, + 415.0, + 884.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 376.0, + 1156.0, + 376.0, + 1156.0, + 412.0, + 1038.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 377.0, + 1292.0, + 377.0, + 1292.0, + 413.0, + 1186.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 414.0, + 831.0, + 414.0, + 831.0, + 464.0, + 603.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 416.0, + 1006.0, + 416.0, + 1006.0, + 459.0, + 886.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 420.0, + 1156.0, + 420.0, + 1156.0, + 456.0, + 1039.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 420.0, + 1293.0, + 420.0, + 1293.0, + 456.0, + 1186.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 439.0, + 489.0, + 439.0, + 489.0, + 494.0, + 439.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 448.0, + 849.0, + 448.0, + 849.0, + 486.0, + 584.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 449.0, + 1007.0, + 449.0, + 1007.0, + 489.0, + 884.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 446.0, + 1157.0, + 446.0, + 1157.0, + 489.0, + 1035.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 447.0, + 1309.0, + 447.0, + 1309.0, + 491.0, + 1186.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 478.0, + 755.0, + 478.0, + 755.0, + 520.0, + 679.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 479.0, + 1019.0, + 479.0, + 1019.0, + 520.0, + 884.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 479.0, + 1145.0, + 479.0, + 1145.0, + 520.0, + 1034.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 478.0, + 1320.0, + 478.0, + 1320.0, + 521.0, + 1185.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 520.0, + 781.0, + 520.0, + 781.0, + 560.0, + 651.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 522.0, + 1019.0, + 522.0, + 1019.0, + 562.0, + 884.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 522.0, + 1169.0, + 522.0, + 1169.0, + 562.0, + 1035.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 522.0, + 1321.0, + 522.0, + 1321.0, + 561.0, + 1185.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 554.0, + 859.0, + 554.0, + 859.0, + 589.0, + 566.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 548.0, + 1020.0, + 548.0, + 1020.0, + 593.0, + 883.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 550.0, + 1158.0, + 550.0, + 1158.0, + 593.0, + 1035.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 550.0, + 1320.0, + 550.0, + 1320.0, + 592.0, + 1187.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 585.0, + 822.0, + 585.0, + 822.0, + 617.0, + 613.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 579.0, + 1021.0, + 579.0, + 1021.0, + 622.0, + 883.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 579.0, + 1158.0, + 579.0, + 1158.0, + 622.0, + 1035.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 579.0, + 1321.0, + 579.0, + 1321.0, + 622.0, + 1184.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 633.0, + 596.0, + 633.0, + 596.0, + 661.0, + 444.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 663.0, + 568.0, + 663.0, + 568.0, + 673.0, + 558.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 701.0, + 360.0, + 701.0, + 360.0, + 732.0, + 305.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 721.0, + 368.0, + 721.0, + 368.0, + 758.0, + 300.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 818.0, + 586.0, + 818.0, + 586.0, + 857.0, + 416.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 817.0, + 976.0, + 817.0, + 976.0, + 856.0, + 783.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 818.0, + 1316.0, + 818.0, + 1316.0, + 855.0, + 1137.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1211.0, + 1406.0, + 1211.0, + 1406.0, + 1244.0, + 296.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1406.0, + 1241.0, + 1406.0, + 1276.0, + 294.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1271.0, + 1406.0, + 1271.0, + 1406.0, + 1306.0, + 294.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1301.0, + 1408.0, + 1301.0, + 1408.0, + 1337.0, + 295.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1332.0, + 1407.0, + 1332.0, + 1407.0, + 1369.0, + 295.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1361.0, + 1406.0, + 1361.0, + 1406.0, + 1397.0, + 292.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1393.0, + 1405.0, + 1393.0, + 1405.0, + 1426.0, + 296.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1460.0, + 292.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 735.0, + 1453.0, + 735.0, + 1487.0, + 295.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1557.0, + 1406.0, + 1557.0, + 1406.0, + 1599.0, + 290.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1587.0, + 1409.0, + 1587.0, + 1409.0, + 1631.0, + 292.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1658.0, + 294.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1651.0, + 1406.0, + 1651.0, + 1406.0, + 1690.0, + 292.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1685.0, + 1122.0, + 1685.0, + 1122.0, + 1716.0, + 296.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1767.0, + 1410.0, + 1767.0, + 1410.0, + 1804.0, + 293.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1799.0, + 1408.0, + 1799.0, + 1408.0, + 1835.0, + 293.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1829.0, + 1405.0, + 1829.0, + 1405.0, + 1865.0, + 293.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1860.0, + 1408.0, + 1860.0, + 1408.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1889.0, + 1406.0, + 1889.0, + 1406.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1921.0, + 1244.0, + 1921.0, + 1244.0, + 1957.0, + 295.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1070.0, + 1408.0, + 1070.0, + 1408.0, + 1105.0, + 293.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1103.0, + 1405.0, + 1103.0, + 1405.0, + 1136.0, + 295.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1134.0, + 1405.0, + 1134.0, + 1405.0, + 1167.0, + 295.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1162.0, + 1298.0, + 1162.0, + 1298.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1113, + 1405, + 1113, + 1405, + 1328, + 297, + 1328 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 882, + 1404, + 882, + 1404, + 1098, + 297, + 1098 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1343, + 1405, + 1343, + 1405, + 1527, + 298, + 1527 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 778, + 1823, + 1354, + 1823, + 1354, + 1948, + 778, + 1948 + ], + "score": 0.968, + "html": "
Metric /Embeddingl2-embeddingsCMEs
π*-bisimulation41.4 (17.6)23.1 (7.6)
PSM17.5 (8.4)87.0 (10.1)
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1587, + 706, + 1587, + 706, + 1954, + 298, + 1954 + ], + "score": 0.965 + }, + { + "category_id": 3, + "poly": [ + 300, + 228, + 1397, + 228, + 1397, + 596, + 300, + 596 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 298, + 745, + 1406, + 745, + 1406, + 868, + 298, + 868 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 297, + 622, + 1404, + 622, + 1404, + 735, + 297, + 735 + ], + "score": 0.938 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 727, + 1586, + 1402, + 1586, + 1402, + 1809, + 727, + 1809 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 297, + 1978, + 1409, + 1978, + 1409, + 2034, + 297, + 2034 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 300, + 1541, + 1230, + 1541, + 1230, + 1572, + 300, + 1572 + ], + "score": 0.768 + }, + { + "category_id": 0, + "poly": [ + 300, + 1541, + 1230, + 1541, + 1230, + 1572, + 300, + 1572 + ], + "score": 0.153 + }, + { + "category_id": 7, + "poly": [ + 297, + 1978, + 1409, + 1978, + 1409, + 2034, + 297, + 2034 + ], + "score": 0.099 + }, + { + "category_id": 13, + "poly": [ + 1000, + 1296, + 1116, + 1296, + 1116, + 1327, + 1000, + 1327 + ], + "score": 0.9, + "latex": "3 0 - 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 387, + 1892, + 415, + 1892, + 415, + 1922, + 387, + 1922 + ], + "score": 0.88, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1199, + 555, + 1229, + 555, + 1229, + 580, + 1199, + 580 + ], + "score": 0.85, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 465, + 1264, + 509, + 1264, + 509, + 1295, + 465, + 1295 + ], + "score": 0.82, + "latex": "2 ^ { n d }" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1204, + 1264, + 1204, + 1264, + 1234, + 1221, + 1234 + ], + "score": 0.81, + "latex": "2 ^ { n d }" + }, + { + "category_id": 13, + "poly": [ + 493, + 1342, + 530, + 1342, + 530, + 1373, + 493, + 1373 + ], + "score": 0.8, + "latex": "\\cdot ^ { 1 ^ { s t } }" + }, + { + "category_id": 13, + "poly": [ + 664, + 1772, + 696, + 1772, + 696, + 1799, + 664, + 1799 + ], + "score": 0.8, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 827, + 1007, + 865, + 1007, + 865, + 1038, + 827, + 1038 + ], + "score": 0.75, + "latex": "( c . f ." + }, + { + "category_id": 13, + "poly": [ + 1101, + 1588, + 1141, + 1588, + 1141, + 1614, + 1101, + 1614 + ], + "score": 0.73, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 975, + 808, + 1011, + 808, + 1011, + 839, + 975, + 839 + ], + "score": 0.65, + "latex": "\\cdot \\cdot f ." + }, + { + "category_id": 13, + "poly": [ + 1303, + 556, + 1323, + 556, + 1323, + 580, + 1303, + 580 + ], + "score": 0.61, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 919, + 539, + 1040, + 539, + 1040, + 568, + 919, + 568 + ], + "score": 0.58, + "latex": "\\mathrm { P S M ~ + ~ } \\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 590, + 977, + 627, + 977, + 627, + 1007, + 590, + 1007 + ], + "score": 0.28, + "latex": "( c . f ." + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 268.0, + 1282.0, + 268.0, + 1282.0, + 286.0, + 1261.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 272.0, + 1240.0, + 272.0, + 1240.0, + 282.0, + 1230.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 278.0, + 1295.0, + 278.0, + 1295.0, + 288.0, + 1282.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 283.0, + 327.0, + 283.0, + 327.0, + 293.0, + 315.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 283.0, + 347.0, + 283.0, + 347.0, + 293.0, + 335.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 284.0, + 1250.0, + 284.0, + 1250.0, + 302.0, + 1228.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 349.0, + 406.0, + 349.0, + 406.0, + 373.0, + 331.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 415.0, + 440.0, + 415.0, + 440.0, + 448.0, + 342.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 381.0, + 933.0, + 381.0, + 933.0, + 452.0, + 903.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 485.0, + 514.0, + 485.0, + 514.0, + 528.0, + 330.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 490.0, + 1063.0, + 490.0, + 1063.0, + 509.0, + 1042.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 536.0, + 918.0, + 536.0, + 918.0, + 570.0, + 876.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 536.0, + 1137.0, + 536.0, + 1137.0, + 570.0, + 1041.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 555.0, + 563.0, + 555.0, + 563.0, + 590.0, + 312.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 550.0, + 779.0, + 550.0, + 779.0, + 585.0, + 676.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 563.0, + 946.0, + 563.0, + 946.0, + 604.0, + 875.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 550.0, + 1198.0, + 550.0, + 1198.0, + 583.0, + 1161.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 550.0, + 1302.0, + 550.0, + 1302.0, + 583.0, + 1230.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 550.0, + 1395.0, + 550.0, + 1395.0, + 583.0, + 1324.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 284.5, + 1307.0, + 284.5, + 1307.0, + 309.0, + 1259.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.75, + 376.5, + 708.75, + 376.5, + 708.75, + 388.5, + 662.75, + 388.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 621.0, + 1405.0, + 621.0, + 1405.0, + 655.0, + 295.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 648.0, + 1405.0, + 648.0, + 1405.0, + 683.0, + 294.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 677.0, + 1405.0, + 677.0, + 1405.0, + 710.0, + 296.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 706.0, + 1283.0, + 706.0, + 1283.0, + 736.0, + 296.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2011.0, + 330.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 1026.0, + 2005.0, + 1026.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1539.0, + 1234.0, + 1539.0, + 1234.0, + 1576.0, + 293.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2011.0, + 330.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 1026.0, + 2005.0, + 1026.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1113.0, + 1408.0, + 1113.0, + 1408.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1143.0, + 1408.0, + 1143.0, + 1408.0, + 1180.0, + 294.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1175.0, + 1406.0, + 1175.0, + 1406.0, + 1209.0, + 295.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1200.0, + 1220.0, + 1200.0, + 1220.0, + 1243.0, + 292.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1200.0, + 1407.0, + 1200.0, + 1407.0, + 1243.0, + 1265.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1235.0, + 1407.0, + 1235.0, + 1407.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1261.0, + 464.0, + 1261.0, + 464.0, + 1304.0, + 292.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1261.0, + 1406.0, + 1261.0, + 1406.0, + 1304.0, + 510.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 999.0, + 1298.0, + 999.0, + 1329.0, + 294.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1298.0, + 1308.0, + 1298.0, + 1308.0, + 1329.0, + 1117.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 881.0, + 1404.0, + 881.0, + 1404.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 911.0, + 1409.0, + 911.0, + 1409.0, + 952.0, + 292.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 943.0, + 1407.0, + 943.0, + 1407.0, + 980.0, + 294.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 974.0, + 589.0, + 974.0, + 589.0, + 1010.0, + 294.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 974.0, + 1406.0, + 974.0, + 1406.0, + 1010.0, + 628.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1004.0, + 826.0, + 1004.0, + 826.0, + 1042.0, + 294.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1004.0, + 1406.0, + 1004.0, + 1406.0, + 1042.0, + 866.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 1405.0, + 1037.0, + 1405.0, + 1071.0, + 295.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1068.0, + 1000.0, + 1068.0, + 1000.0, + 1101.0, + 293.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1340.0, + 492.0, + 1340.0, + 492.0, + 1378.0, + 293.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1340.0, + 1405.0, + 1340.0, + 1405.0, + 1378.0, + 531.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1374.0, + 1407.0, + 1374.0, + 1407.0, + 1406.0, + 296.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1406.0, + 1403.0, + 1406.0, + 1403.0, + 1438.0, + 296.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1469.0, + 295.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1464.0, + 1407.0, + 1464.0, + 1407.0, + 1500.0, + 295.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1493.0, + 1367.0, + 1493.0, + 1367.0, + 1530.0, + 292.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1587.0, + 708.0, + 1587.0, + 708.0, + 1618.0, + 297.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1617.0, + 710.0, + 1617.0, + 710.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1644.0, + 708.0, + 1644.0, + 708.0, + 1684.0, + 293.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1677.0, + 709.0, + 1677.0, + 709.0, + 1711.0, + 295.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 706.0, + 1708.0, + 706.0, + 1742.0, + 295.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1740.0, + 706.0, + 1740.0, + 706.0, + 1771.0, + 296.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1770.0, + 663.0, + 1770.0, + 663.0, + 1802.0, + 294.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1770.0, + 711.0, + 1770.0, + 711.0, + 1802.0, + 697.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1801.0, + 709.0, + 1801.0, + 709.0, + 1832.0, + 296.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1831.0, + 705.0, + 1831.0, + 705.0, + 1862.0, + 296.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1861.0, + 709.0, + 1861.0, + 709.0, + 1895.0, + 295.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1891.0, + 386.0, + 1891.0, + 386.0, + 1923.0, + 294.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1891.0, + 706.0, + 1891.0, + 706.0, + 1923.0, + 416.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1920.0, + 709.0, + 1920.0, + 709.0, + 1955.0, + 293.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 746.0, + 1404.0, + 746.0, + 1404.0, + 779.0, + 295.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 776.0, + 1408.0, + 776.0, + 1408.0, + 809.0, + 296.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 806.0, + 974.0, + 806.0, + 974.0, + 841.0, + 293.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 806.0, + 1408.0, + 806.0, + 1408.0, + 841.0, + 1012.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 834.0, + 1354.0, + 834.0, + 1354.0, + 872.0, + 292.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1587.0, + 1100.0, + 1587.0, + 1100.0, + 1616.0, + 725.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1587.0, + 1404.0, + 1587.0, + 1404.0, + 1616.0, + 1142.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1615.0, + 1405.0, + 1615.0, + 1405.0, + 1646.0, + 724.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1644.0, + 1405.0, + 1644.0, + 1405.0, + 1672.0, + 726.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1700.0, + 724.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1729.0, + 724.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1727.0, + 1403.0, + 1727.0, + 1403.0, + 1755.0, + 725.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1753.0, + 1403.0, + 1753.0, + 1403.0, + 1783.0, + 724.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1780.0, + 1088.0, + 1780.0, + 1088.0, + 1812.0, + 724.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1539.0, + 1234.0, + 1539.0, + 1234.0, + 1576.0, + 293.0, + 1576.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1033, + 902, + 1033, + 902, + 1308, + 298, + 1308 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1404, + 307, + 1404, + 522, + 298, + 522 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1404, + 1850, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 676, + 1405, + 676, + 1405, + 981, + 297, + 981 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1614, + 1406, + 1614, + 1406, + 1799, + 298, + 1799 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1324, + 901, + 1324, + 901, + 1536, + 298, + 1536 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 537, + 1403, + 537, + 1403, + 661, + 298, + 661 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 946, + 1036, + 1382, + 1036, + 1382, + 1369, + 946, + 1369 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 923, + 1378, + 1403, + 1378, + 1403, + 1518, + 923, + 1518 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1400, + 229, + 1400, + 293, + 294, + 293 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 1538, + 1404, + 1538, + 1404, + 1599, + 300, + 1599 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 299, + 1804, + 916, + 1804, + 916, + 1835, + 299, + 1835 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.753 + }, + { + "category_id": 1, + "poly": [ + 294, + 987, + 1336, + 987, + 1336, + 1019, + 294, + 1019 + ], + "score": 0.384 + }, + { + "category_id": 0, + "poly": [ + 294, + 987, + 1336, + 987, + 1336, + 1019, + 294, + 1019 + ], + "score": 0.13 + }, + { + "category_id": 13, + "poly": [ + 1322, + 338, + 1397, + 338, + 1397, + 369, + 1322, + 369 + ], + "score": 0.89, + "latex": "- 4 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 460, + 339, + 487, + 339, + 487, + 369, + 460, + 369 + ], + "score": 0.88, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 853, + 431, + 886, + 431, + 886, + 458, + 853, + 458 + ], + "score": 0.88, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 937, + 798, + 965, + 798, + 965, + 829, + 937, + 829 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 583, + 891, + 615, + 891, + 615, + 917, + 583, + 917 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 956, + 891, + 988, + 891, + 988, + 917, + 956, + 917 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 338, + 1272, + 338, + 1272, + 369, + 1196, + 369 + ], + "score": 0.87, + "latex": "- 6 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1217, + 310, + 1249, + 310, + 1249, + 336, + 1217, + 336 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1913, + 359, + 1913, + 359, + 1941, + 296, + 1941 + ], + "score": 0.87, + "latex": "1 - \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 547, + 233, + 579, + 233, + 579, + 259, + 547, + 259 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 354, + 431, + 383, + 431, + 383, + 460, + 354, + 460 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 674, + 370, + 706, + 370, + 706, + 397, + 674, + 397 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 298, + 492, + 324, + 492, + 324, + 521, + 298, + 521 + ], + "score": 0.86, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1156, + 1944, + 1224, + 1944, + 1224, + 1972, + 1156, + 1972 + ], + "score": 0.84, + "latex": "( \\epsilon = 0" + }, + { + "category_id": 13, + "poly": [ + 474, + 1974, + 565, + 1974, + 565, + 2003, + 474, + 2003 + ], + "score": 0.84, + "latex": "\\epsilon = 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 1354, + 1462, + 1403, + 1462, + 1403, + 1488, + 1354, + 1488 + ], + "score": 0.84, + "latex": "9 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 589, + 309, + 799, + 309, + 799, + 339, + 589, + 339 + ], + "score": 0.83, + "latex": "( = \\mathrm { P S M } + \\mathrm { C M E s } )" + }, + { + "category_id": 13, + "poly": [ + 1153, + 1379, + 1193, + 1379, + 1193, + 1407, + 1153, + 1407 + ], + "score": 0.73, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 358, + 401, + 395, + 401, + 395, + 431, + 358, + 431 + ], + "score": 0.72, + "latex": "\\cdot { c . f . }" + }, + { + "category_id": 13, + "poly": [ + 1326, + 1647, + 1362, + 1647, + 1362, + 1678, + 1326, + 1678 + ], + "score": 0.68, + "latex": "\\cdot f ." + }, + { + "category_id": 13, + "poly": [ + 428, + 1887, + 442, + 1887, + 442, + 1910, + 428, + 1910 + ], + "score": 0.61, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 855, + 1924, + 868, + 1924, + 868, + 1938, + 855, + 1938 + ], + "score": 0.6, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 721, + 2010, + 734, + 2010, + 734, + 2030, + 721, + 2030 + ], + "score": 0.6, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 479, + 492, + 516, + 492, + 516, + 523, + 479, + 523 + ], + "score": 0.51, + "latex": "\\left( c . f . \\right." + }, + { + "category_id": 13, + "poly": [ + 1084, + 1338, + 1117, + 1338, + 1117, + 1369, + 1084, + 1369 + ], + "score": 0.35, + "latex": "\\%" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1033.0, + 1079.0, + 1033.0, + 1079.0, + 1071.0, + 956.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1081.0, + 1077.0, + 1081.0, + 1077.0, + 1116.0, + 946.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1121.0, + 1079.0, + 1121.0, + 1079.0, + 1159.0, + 968.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1145.0, + 1077.0, + 1145.0, + 1077.0, + 1187.0, + 942.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1177.0, + 1077.0, + 1177.0, + 1077.0, + 1212.0, + 945.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1203.0, + 1078.0, + 1203.0, + 1078.0, + 1239.0, + 981.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1243.0, + 1080.0, + 1243.0, + 1080.0, + 1284.0, + 1005.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1301.0, + 1098.0, + 1301.0, + 1098.0, + 1327.0, + 1075.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1301.0, + 1174.0, + 1301.0, + 1174.0, + 1329.0, + 1130.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1300.0, + 1364.0, + 1300.0, + 1364.0, + 1330.0, + 1191.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1336.0, + 1377.0, + 1336.0, + 1377.0, + 1373.0, + 1118.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1377.0, + 1152.0, + 1377.0, + 1152.0, + 1409.0, + 921.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1377.0, + 1406.0, + 1377.0, + 1406.0, + 1409.0, + 1194.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1406.0, + 1405.0, + 1406.0, + 1405.0, + 1435.0, + 922.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1463.0, + 922.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1461.0, + 1353.0, + 1461.0, + 1353.0, + 1490.0, + 922.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1487.0, + 1260.0, + 1487.0, + 1260.0, + 1519.0, + 921.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1803.0, + 917.0, + 1803.0, + 917.0, + 1838.0, + 295.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 985.0, + 1336.0, + 985.0, + 1336.0, + 1019.0, + 296.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 905.0, + 1030.0, + 905.0, + 1069.0, + 294.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1065.0, + 904.0, + 1065.0, + 904.0, + 1096.0, + 295.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1094.0, + 904.0, + 1094.0, + 904.0, + 1128.0, + 296.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1128.0, + 903.0, + 1128.0, + 903.0, + 1156.0, + 295.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1156.0, + 905.0, + 1156.0, + 905.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1187.0, + 907.0, + 1187.0, + 907.0, + 1217.0, + 295.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 905.0, + 1216.0, + 905.0, + 1249.0, + 294.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1247.0, + 904.0, + 1247.0, + 904.0, + 1280.0, + 294.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1278.0, + 721.0, + 1278.0, + 721.0, + 1308.0, + 296.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 307.0, + 588.0, + 307.0, + 588.0, + 342.0, + 294.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 307.0, + 1216.0, + 307.0, + 1216.0, + 342.0, + 800.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 307.0, + 1404.0, + 307.0, + 1404.0, + 342.0, + 1250.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 337.0, + 459.0, + 337.0, + 459.0, + 375.0, + 293.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 337.0, + 1195.0, + 337.0, + 1195.0, + 375.0, + 488.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 337.0, + 1321.0, + 337.0, + 1321.0, + 375.0, + 1273.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 337.0, + 1408.0, + 337.0, + 1408.0, + 375.0, + 1398.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 368.0, + 673.0, + 368.0, + 673.0, + 406.0, + 293.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 368.0, + 1404.0, + 368.0, + 1404.0, + 406.0, + 707.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 357.0, + 400.0, + 357.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 400.0, + 1406.0, + 400.0, + 1406.0, + 434.0, + 396.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 430.0, + 353.0, + 430.0, + 353.0, + 461.0, + 296.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 430.0, + 852.0, + 430.0, + 852.0, + 461.0, + 384.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 430.0, + 1402.0, + 430.0, + 1402.0, + 461.0, + 887.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 491.0, + 478.0, + 491.0, + 478.0, + 525.0, + 325.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 491.0, + 742.0, + 491.0, + 742.0, + 525.0, + 517.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1847.0, + 1405.0, + 1847.0, + 1405.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 427.0, + 1881.0, + 427.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 443.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1913.0, + 854.0, + 1913.0, + 854.0, + 1945.0, + 360.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1945.0, + 869.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1155.0, + 1941.0, + 1155.0, + 1980.0, + 292.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1980.0, + 1225.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 473.0, + 1972.0, + 473.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 566.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 720.0, + 2002.0, + 720.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 2002.0, + 744.0, + 2002.0, + 744.0, + 2037.0, + 735.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 675.0, + 1405.0, + 675.0, + 1405.0, + 711.0, + 295.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 706.0, + 1406.0, + 706.0, + 1406.0, + 741.0, + 295.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 766.0, + 1407.0, + 766.0, + 1407.0, + 802.0, + 295.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 797.0, + 936.0, + 797.0, + 936.0, + 832.0, + 294.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 797.0, + 1406.0, + 797.0, + 1406.0, + 832.0, + 966.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 826.0, + 1408.0, + 826.0, + 1408.0, + 864.0, + 294.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 859.0, + 1407.0, + 859.0, + 1407.0, + 892.0, + 294.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 891.0, + 582.0, + 891.0, + 582.0, + 922.0, + 296.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 891.0, + 955.0, + 891.0, + 955.0, + 922.0, + 616.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 891.0, + 1403.0, + 891.0, + 1403.0, + 922.0, + 989.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 920.0, + 1405.0, + 920.0, + 1405.0, + 955.0, + 294.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 950.0, + 822.0, + 950.0, + 822.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1614.0, + 1406.0, + 1614.0, + 1406.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1647.0, + 1325.0, + 1647.0, + 1325.0, + 1679.0, + 295.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1647.0, + 1404.0, + 1647.0, + 1404.0, + 1679.0, + 1363.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1673.0, + 1408.0, + 1673.0, + 1408.0, + 1712.0, + 293.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1703.0, + 1408.0, + 1703.0, + 1408.0, + 1745.0, + 292.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1736.0, + 1406.0, + 1736.0, + 1406.0, + 1770.0, + 293.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1766.0, + 992.0, + 1766.0, + 992.0, + 1803.0, + 293.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1325.0, + 903.0, + 1325.0, + 903.0, + 1355.0, + 296.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1356.0, + 902.0, + 1356.0, + 902.0, + 1384.0, + 295.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1383.0, + 905.0, + 1383.0, + 905.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1415.0, + 902.0, + 1415.0, + 902.0, + 1448.0, + 295.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1445.0, + 904.0, + 1445.0, + 904.0, + 1480.0, + 294.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1476.0, + 904.0, + 1476.0, + 904.0, + 1510.0, + 295.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1505.0, + 904.0, + 1505.0, + 904.0, + 1540.0, + 294.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 537.0, + 1404.0, + 537.0, + 1404.0, + 570.0, + 296.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 568.0, + 1407.0, + 568.0, + 1407.0, + 602.0, + 293.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 598.0, + 1405.0, + 598.0, + 1405.0, + 633.0, + 293.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 628.0, + 917.0, + 628.0, + 917.0, + 661.0, + 294.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 546.0, + 229.0, + 546.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 580.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1300.0, + 260.0, + 1300.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1537.0, + 1404.0, + 1537.0, + 1404.0, + 1572.0, + 296.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1567.0, + 1327.0, + 1567.0, + 1327.0, + 1601.0, + 294.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 985.0, + 1336.0, + 985.0, + 1336.0, + 1019.0, + 296.0, + 1019.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 782, + 893, + 782, + 893, + 1177, + 298, + 1177 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1423, + 1404, + 1423, + 1404, + 1577, + 298, + 1577 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1025, + 229, + 1025, + 536, + 297, + 536 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 912, + 1102, + 1406, + 1102, + 1406, + 1325, + 912, + 1325 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 1012, + 1637, + 1402, + 1637, + 1402, + 1894, + 1012, + 1894 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 930, + 773, + 1364, + 773, + 1364, + 1091, + 930, + 1091 + ], + "score": 0.965 + }, + { + "category_id": 3, + "poly": [ + 1052, + 227, + 1382, + 227, + 1382, + 453, + 1052, + 453 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 297, + 1195, + 891, + 1195, + 891, + 1347, + 297, + 1347 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 1044, + 460, + 1405, + 460, + 1405, + 627, + 1044, + 627 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 299, + 1637, + 991, + 1637, + 991, + 1976, + 299, + 1976 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 298, + 1348, + 1399, + 1348, + 1399, + 1409, + 298, + 1409 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 297, + 1975, + 1400, + 1975, + 1400, + 2034, + 297, + 2034 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 299, + 551, + 904, + 551, + 904, + 585, + 299, + 585 + ], + "score": 0.913 + }, + { + "category_id": 4, + "poly": [ + 1012, + 1914, + 1403, + 1914, + 1403, + 1970, + 1012, + 1970 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 302, + 1592, + 790, + 1592, + 790, + 1623, + 302, + 1623 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 742, + 817, + 742, + 817, + 772, + 299, + 772 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 297, + 636, + 1404, + 636, + 1404, + 727, + 297, + 727 + ], + "score": 0.744 + }, + { + "category_id": 1, + "poly": [ + 303, + 605, + 1027, + 605, + 1027, + 635, + 303, + 635 + ], + "score": 0.41 + }, + { + "category_id": 13, + "poly": [ + 298, + 262, + 385, + 262, + 385, + 292, + 298, + 292 + ], + "score": 0.89, + "latex": "\\epsilon \\leq 0 . 4" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1296, + 1070, + 1296, + 1070, + 1322, + 1020, + 1322 + ], + "score": 0.86, + "latex": "9 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 457, + 1378, + 484, + 1378, + 484, + 1408, + 457, + 1408 + ], + "score": 0.85, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 965, + 1190, + 997, + 1190, + 997, + 1213, + 965, + 1213 + ], + "score": 0.85, + "latex": "n _ { d }" + }, + { + "category_id": 13, + "poly": [ + 955, + 1761, + 988, + 1761, + 988, + 1788, + 955, + 1788 + ], + "score": 0.84, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1159, + 460, + 1185, + 460, + 1185, + 486, + 1159, + 486 + ], + "score": 0.74, + "latex": "\\%" + }, + { + "category_id": 13, + "poly": [ + 1382, + 492, + 1398, + 492, + 1398, + 514, + 1382, + 514 + ], + "score": 0.51, + "latex": "\\epsilon \\cdot" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1102.0, + 1408.0, + 1102.0, + 1408.0, + 1134.0, + 912.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1130.0, + 1405.0, + 1130.0, + 1405.0, + 1160.0, + 909.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1159.0, + 1404.0, + 1159.0, + 1404.0, + 1186.0, + 911.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1185.0, + 964.0, + 1185.0, + 964.0, + 1216.0, + 911.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1185.0, + 1407.0, + 1185.0, + 1407.0, + 1216.0, + 998.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1212.0, + 1406.0, + 1212.0, + 1406.0, + 1244.0, + 910.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1242.0, + 1407.0, + 1242.0, + 1407.0, + 1269.0, + 911.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1270.0, + 1407.0, + 1270.0, + 1407.0, + 1296.0, + 911.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1295.0, + 1019.0, + 1295.0, + 1019.0, + 1326.0, + 912.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1295.0, + 1369.0, + 1295.0, + 1369.0, + 1326.0, + 1071.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 767.0, + 1045.0, + 767.0, + 1045.0, + 810.0, + 994.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 774.0, + 970.0, + 774.0, + 970.0, + 1018.0, + 929.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 826.0, + 990.0, + 826.0, + 990.0, + 958.0, + 958.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 837.0, + 1037.0, + 837.0, + 1037.0, + 873.0, + 996.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 845.0, + 1256.0, + 845.0, + 1256.0, + 881.0, + 1074.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 893.0, + 1102.0, + 893.0, + 1102.0, + 906.0, + 1089.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 884.0, + 1241.0, + 884.0, + 1241.0, + 915.0, + 1131.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 904.0, + 1038.0, + 904.0, + 1038.0, + 940.0, + 996.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 922.0, + 1254.0, + 922.0, + 1254.0, + 954.0, + 1131.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 971.0, + 1044.0, + 971.0, + 1044.0, + 1009.0, + 971.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 973.0, + 1100.0, + 973.0, + 1100.0, + 982.0, + 1090.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 959.0, + 1267.0, + 959.0, + 1267.0, + 991.0, + 1130.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1006.0, + 1363.0, + 1006.0, + 1363.0, + 1044.0, + 1047.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1036.0, + 1194.0, + 1036.0, + 1194.0, + 1067.0, + 981.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1061.0, + 1265.0, + 1061.0, + 1265.0, + 1097.0, + 1153.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 229.0, + 1151.0, + 229.0, + 1151.0, + 262.0, + 1094.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 230.0, + 1086.0, + 230.0, + 1086.0, + 403.0, + 1053.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 244.0, + 1152.0, + 244.0, + 1152.0, + 391.0, + 1104.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 263.0, + 1106.0, + 263.0, + 1106.0, + 355.0, + 1080.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 358.0, + 1369.0, + 358.0, + 1369.0, + 377.0, + 1353.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 396.0, + 1384.0, + 396.0, + 1384.0, + 429.0, + 1144.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 422.0, + 1378.0, + 422.0, + 1378.0, + 458.0, + 1158.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 458.0, + 1158.0, + 458.0, + 1158.0, + 490.0, + 1043.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 458.0, + 1406.0, + 458.0, + 1406.0, + 490.0, + 1186.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 487.0, + 1381.0, + 487.0, + 1381.0, + 518.0, + 1042.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 487.0, + 1409.0, + 487.0, + 1409.0, + 518.0, + 1399.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 515.0, + 1409.0, + 515.0, + 1409.0, + 543.0, + 1042.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 543.0, + 1407.0, + 543.0, + 1407.0, + 573.0, + 1042.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 571.0, + 1405.0, + 571.0, + 1405.0, + 600.0, + 1042.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 596.0, + 1152.0, + 596.0, + 1152.0, + 629.0, + 1042.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 547.0, + 906.0, + 547.0, + 906.0, + 591.0, + 293.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1912.0, + 1407.0, + 1912.0, + 1407.0, + 1944.0, + 1010.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1942.0, + 1376.0, + 1942.0, + 1376.0, + 1971.0, + 1011.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1589.0, + 796.0, + 1589.0, + 796.0, + 1628.0, + 295.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 739.0, + 822.0, + 739.0, + 822.0, + 776.0, + 294.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 780.0, + 895.0, + 780.0, + 895.0, + 815.0, + 296.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 811.0, + 896.0, + 811.0, + 896.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 842.0, + 893.0, + 842.0, + 893.0, + 874.0, + 295.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 874.0, + 896.0, + 874.0, + 896.0, + 905.0, + 295.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 903.0, + 895.0, + 903.0, + 895.0, + 935.0, + 294.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 934.0, + 897.0, + 934.0, + 897.0, + 966.0, + 295.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 963.0, + 897.0, + 963.0, + 897.0, + 997.0, + 294.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 993.0, + 893.0, + 993.0, + 893.0, + 1028.0, + 294.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1027.0, + 895.0, + 1027.0, + 895.0, + 1059.0, + 295.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1056.0, + 896.0, + 1056.0, + 896.0, + 1091.0, + 295.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1088.0, + 893.0, + 1088.0, + 893.0, + 1120.0, + 295.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1116.0, + 893.0, + 1116.0, + 893.0, + 1150.0, + 295.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1150.0, + 867.0, + 1150.0, + 867.0, + 1179.0, + 297.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1421.0, + 1404.0, + 1421.0, + 1404.0, + 1460.0, + 293.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1454.0, + 1406.0, + 1454.0, + 1406.0, + 1491.0, + 293.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1484.0, + 1405.0, + 1484.0, + 1405.0, + 1521.0, + 293.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1515.0, + 1404.0, + 1515.0, + 1404.0, + 1553.0, + 293.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1545.0, + 1355.0, + 1545.0, + 1355.0, + 1581.0, + 293.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1024.0, + 230.0, + 1024.0, + 263.0, + 295.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 262.0, + 1028.0, + 262.0, + 1028.0, + 295.0, + 386.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 1025.0, + 292.0, + 1025.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 320.0, + 1028.0, + 320.0, + 1028.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1028.0, + 351.0, + 1028.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 383.0, + 1025.0, + 383.0, + 1025.0, + 416.0, + 295.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 413.0, + 1025.0, + 413.0, + 1025.0, + 447.0, + 295.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 440.0, + 1028.0, + 440.0, + 1028.0, + 480.0, + 293.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 474.0, + 1025.0, + 474.0, + 1025.0, + 509.0, + 295.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 779.0, + 505.0, + 779.0, + 539.0, + 294.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1194.0, + 894.0, + 1194.0, + 894.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 892.0, + 1224.0, + 892.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1253.0, + 893.0, + 1253.0, + 893.0, + 1287.0, + 292.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 896.0, + 1285.0, + 896.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1314.0, + 897.0, + 1314.0, + 897.0, + 1351.0, + 295.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1637.0, + 995.0, + 1637.0, + 995.0, + 1670.0, + 297.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1667.0, + 991.0, + 1667.0, + 991.0, + 1702.0, + 297.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1700.0, + 991.0, + 1700.0, + 991.0, + 1730.0, + 297.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1728.0, + 993.0, + 1728.0, + 993.0, + 1760.0, + 296.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1759.0, + 954.0, + 1759.0, + 954.0, + 1793.0, + 296.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1786.0, + 993.0, + 1786.0, + 993.0, + 1826.0, + 293.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 995.0, + 1818.0, + 995.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1850.0, + 992.0, + 1850.0, + 992.0, + 1883.0, + 296.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 995.0, + 1881.0, + 995.0, + 1913.0, + 294.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1911.0, + 992.0, + 1911.0, + 992.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1942.0, + 991.0, + 1942.0, + 991.0, + 1975.0, + 297.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1343.0, + 1404.0, + 1343.0, + 1404.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1373.0, + 456.0, + 1373.0, + 456.0, + 1412.0, + 294.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1373.0, + 1369.0, + 1373.0, + 1369.0, + 1412.0, + 485.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1967.0, + 1405.0, + 1967.0, + 1405.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 1192.0, + 2002.0, + 1192.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 633.0, + 1405.0, + 633.0, + 1405.0, + 670.0, + 295.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 664.0, + 1402.0, + 664.0, + 1402.0, + 697.0, + 296.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 695.0, + 1246.0, + 695.0, + 1246.0, + 729.0, + 293.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 601.0, + 1024.0, + 601.0, + 1024.0, + 641.0, + 296.0, + 641.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1429, + 1406, + 1429, + 1406, + 1645, + 297, + 1645 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1200, + 1406, + 1200, + 1406, + 1415, + 298, + 1415 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 947, + 1405, + 947, + 1405, + 1134, + 298, + 1134 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 549, + 1405, + 549, + 1405, + 796, + 297, + 796 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 306, + 326, + 1395, + 326, + 1395, + 526, + 306, + 526 + ], + "score": 0.978, + "html": "
InitializationMethodBiC-catchC-swingupC-runF-spinR-easyW-walk
RandomDrQ747±28582±42220±12646±54931±14549±83
DrQ + PSEs821±17749±19308±12779±49955±10789±28
PretrainedDrQ748±30689±22219±10764±48943±10709±29
DrQ +PSEs805±25753±13282±8803±19962±11829±21
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1659, + 1404, + 1659, + 1404, + 1814, + 298, + 1814 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 809, + 1404, + 809, + 1404, + 933, + 298, + 933 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1403, + 1880, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.977 + }, + { + "category_id": 6, + "poly": [ + 297, + 224, + 1407, + 224, + 1407, + 310, + 297, + 310 + ], + "score": 0.959 + }, + { + "category_id": 0, + "poly": [ + 300, + 1150, + 589, + 1150, + 589, + 1186, + 300, + 1186 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 299, + 1831, + 545, + 1831, + 545, + 1866, + 299, + 1866 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.797 + }, + { + "category_id": 13, + "poly": [ + 1153, + 1263, + 1185, + 1263, + 1185, + 1290, + 1153, + 1290 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 580, + 1385, + 612, + 1385, + 612, + 1411, + 580, + 1411 + ], + "score": 0.86, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 691, + 254, + 717, + 254, + 717, + 278, + 691, + 278 + ], + "score": 0.66, + "latex": "\\pm" + }, + { + "category_id": 13, + "poly": [ + 1162, + 704, + 1203, + 704, + 1203, + 734, + 1162, + 734 + ], + "score": 0.56, + "latex": "( c . f ." + }, + { + "category_id": 13, + "poly": [ + 595, + 764, + 649, + 764, + 649, + 795, + 595, + 795 + ], + "score": 0.43, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 587, + 280, + 636, + 280, + 636, + 309, + 587, + 309 + ], + "score": 0.41, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 1291, + 642, + 1346, + 642, + 1346, + 672, + 1291, + 672 + ], + "score": 0.32, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 672, + 902, + 726, + 902, + 726, + 933, + 672, + 933 + ], + "score": 0.3, + "latex": "_ \\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 1087, + 551, + 1141, + 551, + 1141, + 582, + 1087, + 582 + ], + "score": 0.28, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 1278, + 949, + 1334, + 949, + 1334, + 980, + 1278, + 980 + ], + "score": 0.27, + "latex": "_ \\mathrm { D r Q }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 222.0, + 1405.0, + 222.0, + 1405.0, + 255.0, + 292.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 250.0, + 690.0, + 250.0, + 690.0, + 282.0, + 294.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 250.0, + 1405.0, + 250.0, + 1405.0, + 282.0, + 718.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 277.0, + 586.0, + 277.0, + 586.0, + 311.0, + 292.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 277.0, + 1346.0, + 277.0, + 1346.0, + 311.0, + 637.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1145.0, + 593.0, + 1145.0, + 593.0, + 1191.0, + 292.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1824.0, + 550.0, + 1824.0, + 550.0, + 1875.0, + 292.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1427.0, + 1409.0, + 1427.0, + 1409.0, + 1464.0, + 294.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1460.0, + 1407.0, + 1460.0, + 1407.0, + 1495.0, + 295.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1489.0, + 1406.0, + 1489.0, + 1406.0, + 1525.0, + 294.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1522.0, + 1407.0, + 1522.0, + 1407.0, + 1553.0, + 295.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1549.0, + 1409.0, + 1549.0, + 1409.0, + 1585.0, + 294.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1581.0, + 1404.0, + 1581.0, + 1404.0, + 1616.0, + 295.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1612.0, + 1349.0, + 1612.0, + 1349.0, + 1646.0, + 294.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1199.0, + 1404.0, + 1199.0, + 1404.0, + 1234.0, + 295.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1230.0, + 1404.0, + 1230.0, + 1404.0, + 1264.0, + 295.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1261.0, + 1152.0, + 1261.0, + 1152.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1261.0, + 1406.0, + 1261.0, + 1406.0, + 1295.0, + 1186.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1293.0, + 1404.0, + 1293.0, + 1404.0, + 1323.0, + 296.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1323.0, + 1407.0, + 1323.0, + 1407.0, + 1354.0, + 296.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1351.0, + 1406.0, + 1351.0, + 1406.0, + 1386.0, + 293.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1385.0, + 579.0, + 1385.0, + 579.0, + 1416.0, + 296.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1385.0, + 1230.0, + 1385.0, + 1230.0, + 1416.0, + 613.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 946.0, + 1277.0, + 946.0, + 1277.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 946.0, + 1405.0, + 946.0, + 1405.0, + 984.0, + 1335.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 979.0, + 1403.0, + 979.0, + 1403.0, + 1014.0, + 295.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1007.0, + 1405.0, + 1007.0, + 1405.0, + 1046.0, + 293.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 1409.0, + 1037.0, + 1409.0, + 1076.0, + 295.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1070.0, + 1406.0, + 1070.0, + 1406.0, + 1106.0, + 295.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1101.0, + 1393.0, + 1101.0, + 1393.0, + 1138.0, + 294.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 548.0, + 1086.0, + 548.0, + 1086.0, + 586.0, + 294.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 548.0, + 1406.0, + 548.0, + 1406.0, + 586.0, + 1142.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 1407.0, + 580.0, + 1407.0, + 617.0, + 294.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 609.0, + 1406.0, + 609.0, + 1406.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 1290.0, + 642.0, + 1290.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 676.0, + 1347.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 707.0, + 292.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 1161.0, + 703.0, + 1161.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 703.0, + 1405.0, + 703.0, + 1405.0, + 736.0, + 1204.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 733.0, + 1408.0, + 733.0, + 1408.0, + 769.0, + 292.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 594.0, + 763.0, + 594.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 763.0, + 1231.0, + 763.0, + 1231.0, + 801.0, + 650.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1657.0, + 1408.0, + 1657.0, + 1408.0, + 1693.0, + 295.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1686.0, + 1406.0, + 1686.0, + 1406.0, + 1726.0, + 293.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1719.0, + 1405.0, + 1719.0, + 1405.0, + 1754.0, + 293.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1751.0, + 1405.0, + 1751.0, + 1405.0, + 1784.0, + 296.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1779.0, + 1367.0, + 1779.0, + 1367.0, + 1818.0, + 293.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 808.0, + 1406.0, + 808.0, + 1406.0, + 845.0, + 292.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 838.0, + 1406.0, + 838.0, + 1406.0, + 876.0, + 292.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 873.0, + 1402.0, + 873.0, + 1402.0, + 906.0, + 294.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 901.0, + 671.0, + 901.0, + 671.0, + 936.0, + 293.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 901.0, + 1154.0, + 901.0, + 1154.0, + 936.0, + 727.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1915.0, + 297.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1911.0, + 1403.0, + 1911.0, + 1403.0, + 1944.0, + 297.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1407.0, + 1971.0, + 1407.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1311.0, + 2001.0, + 1311.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 299, + 228, + 488, + 228, + 488, + 261, + 299, + 261 + ], + "score": 0.85 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 297, + 1317, + 1399, + 1317, + 1399, + 1378, + 297, + 1378 + ], + "score": 0.538 + }, + { + "category_id": 1, + "poly": [ + 299, + 899, + 1401, + 899, + 1401, + 961, + 299, + 961 + ], + "score": 0.508 + }, + { + "category_id": 1, + "poly": [ + 298, + 977, + 1401, + 977, + 1401, + 1038, + 298, + 1038 + ], + "score": 0.506 + }, + { + "category_id": 1, + "poly": [ + 299, + 1396, + 1400, + 1396, + 1400, + 1456, + 299, + 1456 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 300, + 743, + 1401, + 743, + 1401, + 803, + 300, + 803 + ], + "score": 0.495 + }, + { + "category_id": 1, + "poly": [ + 300, + 1239, + 1401, + 1239, + 1401, + 1300, + 300, + 1300 + ], + "score": 0.49 + }, + { + "category_id": 1, + "poly": [ + 299, + 821, + 1402, + 821, + 1402, + 883, + 299, + 883 + ], + "score": 0.487 + }, + { + "category_id": 1, + "poly": [ + 297, + 353, + 1399, + 353, + 1399, + 412, + 297, + 412 + ], + "score": 0.487 + }, + { + "category_id": 1, + "poly": [ + 300, + 1134, + 1401, + 1134, + 1401, + 1221, + 300, + 1221 + ], + "score": 0.475 + }, + { + "category_id": 1, + "poly": [ + 299, + 277, + 1400, + 277, + 1400, + 334, + 299, + 334 + ], + "score": 0.469 + }, + { + "category_id": 1, + "poly": [ + 300, + 1552, + 1399, + 1552, + 1399, + 1641, + 300, + 1641 + ], + "score": 0.468 + }, + { + "category_id": 1, + "poly": [ + 297, + 1842, + 1401, + 1842, + 1401, + 1929, + 297, + 1929 + ], + "score": 0.468 + }, + { + "category_id": 1, + "poly": [ + 300, + 510, + 1400, + 510, + 1400, + 570, + 300, + 570 + ], + "score": 0.463 + }, + { + "category_id": 1, + "poly": [ + 295, + 1056, + 1400, + 1056, + 1400, + 1117, + 295, + 1117 + ], + "score": 0.461 + }, + { + "category_id": 1, + "poly": [ + 301, + 431, + 1399, + 431, + 1399, + 492, + 301, + 492 + ], + "score": 0.45 + }, + { + "category_id": 1, + "poly": [ + 300, + 666, + 1401, + 666, + 1401, + 726, + 300, + 726 + ], + "score": 0.446 + }, + { + "category_id": 1, + "poly": [ + 296, + 1474, + 1400, + 1474, + 1400, + 1534, + 296, + 1534 + ], + "score": 0.443 + }, + { + "category_id": 1, + "poly": [ + 300, + 1657, + 1403, + 1657, + 1403, + 1744, + 300, + 1744 + ], + "score": 0.441 + }, + { + "category_id": 1, + "poly": [ + 301, + 588, + 1399, + 588, + 1399, + 648, + 301, + 648 + ], + "score": 0.437 + }, + { + "category_id": 1, + "poly": [ + 296, + 1763, + 1400, + 1763, + 1400, + 1824, + 296, + 1824 + ], + "score": 0.414 + }, + { + "category_id": 1, + "poly": [ + 298, + 1948, + 1406, + 1948, + 1406, + 2033, + 298, + 2033 + ], + "score": 0.348 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 228.0, + 490.0, + 228.0, + 490.0, + 265.0, + 297.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1315.0, + 1404.0, + 1315.0, + 1404.0, + 1354.0, + 293.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1347.0, + 843.0, + 1347.0, + 843.0, + 1379.0, + 325.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 898.0, + 1405.0, + 898.0, + 1405.0, + 934.0, + 295.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 926.0, + 622.0, + 926.0, + 622.0, + 962.0, + 322.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 1405.0, + 977.0, + 1405.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1006.0, + 897.0, + 1006.0, + 897.0, + 1040.0, + 319.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1393.0, + 1405.0, + 1393.0, + 1405.0, + 1434.0, + 294.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1425.0, + 1282.0, + 1425.0, + 1282.0, + 1457.0, + 320.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 743.0, + 1405.0, + 743.0, + 1405.0, + 778.0, + 296.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 774.0, + 976.0, + 774.0, + 976.0, + 803.0, + 323.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 1405.0, + 1239.0, + 1405.0, + 1275.0, + 295.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1268.0, + 986.0, + 1268.0, + 986.0, + 1302.0, + 321.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 820.0, + 1406.0, + 820.0, + 1406.0, + 857.0, + 294.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 852.0, + 1407.0, + 852.0, + 1407.0, + 884.0, + 323.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 355.0, + 1403.0, + 355.0, + 1403.0, + 386.0, + 296.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 381.0, + 560.0, + 381.0, + 560.0, + 412.0, + 322.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1134.0, + 1405.0, + 1134.0, + 1405.0, + 1168.0, + 294.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1163.0, + 1406.0, + 1163.0, + 1406.0, + 1196.0, + 323.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1189.0, + 743.0, + 1189.0, + 743.0, + 1223.0, + 321.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 274.0, + 1405.0, + 274.0, + 1405.0, + 310.0, + 294.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 304.0, + 774.0, + 304.0, + 774.0, + 334.0, + 323.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1553.0, + 1403.0, + 1553.0, + 1403.0, + 1587.0, + 292.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1580.0, + 1405.0, + 1580.0, + 1405.0, + 1615.0, + 320.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1609.0, + 941.0, + 1609.0, + 941.0, + 1643.0, + 322.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1841.0, + 1405.0, + 1841.0, + 1405.0, + 1876.0, + 293.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1871.0, + 1408.0, + 1871.0, + 1408.0, + 1904.0, + 322.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1901.0, + 617.0, + 1901.0, + 617.0, + 1930.0, + 321.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 546.0, + 296.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 539.0, + 1309.0, + 539.0, + 1309.0, + 571.0, + 321.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1090.0, + 296.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1085.0, + 1163.0, + 1085.0, + 1163.0, + 1117.0, + 323.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 430.0, + 1405.0, + 430.0, + 1405.0, + 467.0, + 293.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 461.0, + 1082.0, + 461.0, + 1082.0, + 493.0, + 320.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 1403.0, + 666.0, + 1403.0, + 701.0, + 294.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 696.0, + 898.0, + 696.0, + 898.0, + 727.0, + 321.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1476.0, + 1404.0, + 1476.0, + 1404.0, + 1508.0, + 296.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1504.0, + 1005.0, + 1504.0, + 1005.0, + 1533.0, + 322.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1660.0, + 1404.0, + 1660.0, + 1404.0, + 1689.0, + 296.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1689.0, + 1405.0, + 1689.0, + 1405.0, + 1718.0, + 321.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1711.0, + 391.0, + 1711.0, + 391.0, + 1747.0, + 320.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 588.0, + 1404.0, + 588.0, + 1404.0, + 623.0, + 295.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 617.0, + 1231.0, + 617.0, + 1231.0, + 650.0, + 321.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1762.0, + 1405.0, + 1762.0, + 1405.0, + 1799.0, + 295.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1794.0, + 1336.0, + 1794.0, + 1336.0, + 1825.0, + 321.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1950.0, + 1404.0, + 1950.0, + 1404.0, + 1979.0, + 296.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1979.0, + 1407.0, + 1979.0, + 1407.0, + 2008.0, + 324.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2000.0, + 388.0, + 2000.0, + 388.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 299, + 1419, + 1400, + 1419, + 1400, + 1480, + 299, + 1480 + ], + "score": 0.652 + }, + { + "category_id": 1, + "poly": [ + 294, + 1498, + 1401, + 1498, + 1401, + 1560, + 294, + 1560 + ], + "score": 0.642 + }, + { + "category_id": 1, + "poly": [ + 298, + 1262, + 1404, + 1262, + 1404, + 1323, + 298, + 1323 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 297, + 1341, + 1401, + 1341, + 1401, + 1401, + 297, + 1401 + ], + "score": 0.612 + }, + { + "category_id": 1, + "poly": [ + 298, + 813, + 1402, + 813, + 1402, + 873, + 298, + 873 + ], + "score": 0.601 + }, + { + "category_id": 1, + "poly": [ + 291, + 1684, + 1404, + 1684, + 1404, + 1744, + 291, + 1744 + ], + "score": 0.597 + }, + { + "category_id": 1, + "poly": [ + 297, + 1577, + 1403, + 1577, + 1403, + 1663, + 297, + 1663 + ], + "score": 0.578 + }, + { + "category_id": 1, + "poly": [ + 299, + 1183, + 1400, + 1183, + 1400, + 1244, + 299, + 1244 + ], + "score": 0.574 + }, + { + "category_id": 1, + "poly": [ + 299, + 970, + 1401, + 970, + 1401, + 1058, + 299, + 1058 + ], + "score": 0.572 + }, + { + "category_id": 1, + "poly": [ + 298, + 1077, + 1401, + 1077, + 1401, + 1163, + 298, + 1163 + ], + "score": 0.569 + }, + { + "category_id": 1, + "poly": [ + 295, + 892, + 1402, + 892, + 1402, + 951, + 295, + 951 + ], + "score": 0.567 + }, + { + "category_id": 1, + "poly": [ + 296, + 735, + 1399, + 735, + 1399, + 795, + 296, + 795 + ], + "score": 0.525 + }, + { + "category_id": 1, + "poly": [ + 300, + 231, + 1401, + 231, + 1401, + 318, + 300, + 318 + ], + "score": 0.522 + }, + { + "category_id": 1, + "poly": [ + 294, + 336, + 1401, + 336, + 1401, + 396, + 294, + 396 + ], + "score": 0.52 + }, + { + "category_id": 1, + "poly": [ + 296, + 1763, + 1402, + 1763, + 1402, + 1849, + 296, + 1849 + ], + "score": 0.517 + }, + { + "category_id": 1, + "poly": [ + 296, + 1948, + 1400, + 1948, + 1400, + 2034, + 296, + 2034 + ], + "score": 0.512 + }, + { + "category_id": 1, + "poly": [ + 301, + 415, + 1401, + 415, + 1401, + 504, + 301, + 504 + ], + "score": 0.505 + }, + { + "category_id": 1, + "poly": [ + 296, + 629, + 1403, + 629, + 1403, + 716, + 296, + 716 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 293, + 1868, + 1398, + 1868, + 1398, + 1929, + 293, + 1929 + ], + "score": 0.485 + }, + { + "category_id": 1, + "poly": [ + 297, + 523, + 1402, + 523, + 1402, + 609, + 297, + 609 + ], + "score": 0.485 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 1404.0, + 1419.0, + 1404.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1448.0, + 756.0, + 1448.0, + 756.0, + 1480.0, + 323.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1499.0, + 1404.0, + 1499.0, + 1404.0, + 1532.0, + 294.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1526.0, + 1217.0, + 1526.0, + 1217.0, + 1561.0, + 321.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1265.0, + 1404.0, + 1265.0, + 1404.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1291.0, + 1357.0, + 1291.0, + 1357.0, + 1325.0, + 322.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1337.0, + 1405.0, + 1337.0, + 1405.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1371.0, + 1332.0, + 1371.0, + 1332.0, + 1402.0, + 324.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 810.0, + 1406.0, + 810.0, + 1406.0, + 849.0, + 293.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 842.0, + 450.0, + 842.0, + 450.0, + 874.0, + 322.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1679.0, + 1409.0, + 1679.0, + 1409.0, + 1722.0, + 293.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1713.0, + 730.0, + 1713.0, + 730.0, + 1744.0, + 323.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1579.0, + 1405.0, + 1579.0, + 1405.0, + 1608.0, + 296.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1606.0, + 1408.0, + 1606.0, + 1408.0, + 1640.0, + 320.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1632.0, + 391.0, + 1632.0, + 391.0, + 1664.0, + 321.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1186.0, + 1402.0, + 1186.0, + 1402.0, + 1217.0, + 297.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1213.0, + 1231.0, + 1213.0, + 1231.0, + 1245.0, + 323.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 971.0, + 1405.0, + 971.0, + 1405.0, + 1005.0, + 294.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1000.0, + 1408.0, + 1000.0, + 1408.0, + 1034.0, + 320.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1029.0, + 801.0, + 1029.0, + 801.0, + 1059.0, + 322.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1078.0, + 1405.0, + 1078.0, + 1405.0, + 1111.0, + 296.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1106.0, + 1407.0, + 1106.0, + 1407.0, + 1140.0, + 321.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1130.0, + 401.0, + 1130.0, + 401.0, + 1166.0, + 322.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 894.0, + 1404.0, + 894.0, + 1404.0, + 925.0, + 297.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 922.0, + 950.0, + 922.0, + 950.0, + 951.0, + 323.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 730.0, + 1405.0, + 730.0, + 1405.0, + 774.0, + 292.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 764.0, + 994.0, + 764.0, + 994.0, + 796.0, + 321.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 291.0, + 324.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 287.0, + 1136.0, + 287.0, + 1136.0, + 320.0, + 323.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 336.0, + 1404.0, + 336.0, + 1404.0, + 371.0, + 296.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 366.0, + 669.0, + 366.0, + 669.0, + 396.0, + 319.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1763.0, + 1406.0, + 1763.0, + 1406.0, + 1796.0, + 296.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1791.0, + 1406.0, + 1791.0, + 1406.0, + 1825.0, + 320.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1820.0, + 490.0, + 1820.0, + 490.0, + 1850.0, + 323.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1946.0, + 1404.0, + 1946.0, + 1404.0, + 1982.0, + 295.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1978.0, + 1404.0, + 1978.0, + 1404.0, + 2007.0, + 324.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2004.0, + 592.0, + 2004.0, + 592.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 418.0, + 1404.0, + 418.0, + 1404.0, + 448.0, + 297.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 444.0, + 1406.0, + 444.0, + 1406.0, + 477.0, + 323.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 474.0, + 1404.0, + 474.0, + 1404.0, + 507.0, + 322.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 626.0, + 1405.0, + 626.0, + 1405.0, + 666.0, + 294.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 689.0, + 324.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 684.0, + 758.0, + 684.0, + 758.0, + 717.0, + 323.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1863.0, + 1404.0, + 1863.0, + 1404.0, + 1907.0, + 293.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1899.0, + 1201.0, + 1899.0, + 1201.0, + 1930.0, + 322.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 523.0, + 1404.0, + 523.0, + 1404.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 553.0, + 1403.0, + 553.0, + 1403.0, + 582.0, + 325.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 581.0, + 670.0, + 581.0, + 670.0, + 609.0, + 320.0, + 609.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 814, + 75, + 814, + 105, + 299, + 105 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.824 + }, + { + "category_id": 1, + "poly": [ + 292, + 146, + 1409, + 146, + 1409, + 2039, + 292, + 2039 + ], + "score": 0.541 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 233.0, + 1408.0, + 233.0, + 1408.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1296.0, + 260.0, + 1296.0, + 293.0, + 322.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 306.0, + 1408.0, + 306.0, + 1408.0, + 347.0, + 291.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 339.0, + 502.0, + 339.0, + 502.0, + 372.0, + 320.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 382.0, + 1408.0, + 382.0, + 1408.0, + 426.0, + 291.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 414.0, + 1237.0, + 414.0, + 1237.0, + 451.0, + 316.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 463.0, + 1406.0, + 463.0, + 1406.0, + 503.0, + 293.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 495.0, + 1394.0, + 495.0, + 1394.0, + 528.0, + 320.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 544.0, + 1406.0, + 544.0, + 1406.0, + 578.0, + 295.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 576.0, + 751.0, + 576.0, + 751.0, + 603.0, + 322.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 621.0, + 1408.0, + 621.0, + 1408.0, + 654.0, + 293.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 650.0, + 1406.0, + 650.0, + 1406.0, + 684.0, + 322.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 675.0, + 392.0, + 675.0, + 392.0, + 711.0, + 320.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 729.0, + 1406.0, + 729.0, + 1406.0, + 762.0, + 295.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 754.0, + 1148.0, + 754.0, + 1148.0, + 789.0, + 316.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 800.0, + 1406.0, + 800.0, + 1406.0, + 839.0, + 293.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 833.0, + 1191.0, + 833.0, + 1191.0, + 866.0, + 318.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 879.0, + 1410.0, + 879.0, + 1410.0, + 920.0, + 291.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 912.0, + 1368.0, + 912.0, + 1368.0, + 945.0, + 320.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 962.0, + 1404.0, + 962.0, + 1404.0, + 995.0, + 297.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 987.0, + 975.0, + 987.0, + 975.0, + 1024.0, + 316.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1034.0, + 1410.0, + 1034.0, + 1410.0, + 1076.0, + 291.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1068.0, + 1273.0, + 1068.0, + 1273.0, + 1101.0, + 318.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1408.0, + 1113.0, + 1408.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1144.0, + 584.0, + 1144.0, + 584.0, + 1175.0, + 318.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1188.0, + 1408.0, + 1188.0, + 1408.0, + 1230.0, + 293.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1219.0, + 1408.0, + 1219.0, + 1408.0, + 1259.0, + 318.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1248.0, + 673.0, + 1248.0, + 673.0, + 1285.0, + 316.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1300.0, + 1408.0, + 1300.0, + 1408.0, + 1333.0, + 295.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1329.0, + 838.0, + 1329.0, + 838.0, + 1362.0, + 318.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1377.0, + 1408.0, + 1377.0, + 1408.0, + 1410.0, + 293.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1402.0, + 1408.0, + 1402.0, + 1408.0, + 1439.0, + 316.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1431.0, + 1117.0, + 1431.0, + 1117.0, + 1468.0, + 316.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1479.0, + 1406.0, + 1479.0, + 1406.0, + 1520.0, + 291.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1512.0, + 1404.0, + 1512.0, + 1404.0, + 1545.0, + 322.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1537.0, + 1009.0, + 1537.0, + 1009.0, + 1574.0, + 314.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1587.0, + 943.0, + 1587.0, + 943.0, + 1620.0, + 295.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 1404.0, + 1636.0, + 1404.0, + 1670.0, + 295.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1668.0, + 1406.0, + 1668.0, + 1406.0, + 1701.0, + 322.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1695.0, + 688.0, + 1695.0, + 688.0, + 1728.0, + 320.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1744.0, + 1404.0, + 1744.0, + 1404.0, + 1778.0, + 295.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1769.0, + 1406.0, + 1769.0, + 1406.0, + 1807.0, + 316.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1819.0, + 1408.0, + 1819.0, + 1408.0, + 1859.0, + 293.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1850.0, + 884.0, + 1850.0, + 884.0, + 1883.0, + 320.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1900.0, + 1406.0, + 1900.0, + 1406.0, + 1933.0, + 295.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1925.0, + 905.0, + 1925.0, + 905.0, + 1964.0, + 318.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1975.0, + 1408.0, + 1975.0, + 1408.0, + 2014.0, + 295.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 2004.0, + 685.0, + 2004.0, + 685.0, + 2039.0, + 316.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 299, + 231, + 1402, + 231, + 1402, + 316, + 299, + 316 + ], + "score": 0.802 + }, + { + "category_id": 1, + "poly": [ + 296, + 441, + 1403, + 441, + 1403, + 499, + 296, + 499 + ], + "score": 0.799 + }, + { + "category_id": 1, + "poly": [ + 295, + 336, + 1400, + 336, + 1400, + 422, + 295, + 422 + ], + "score": 0.745 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1406.0, + 231.0, + 1406.0, + 263.0, + 295.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 291.0, + 323.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 289.0, + 664.0, + 289.0, + 664.0, + 317.0, + 322.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 439.0, + 1404.0, + 439.0, + 1404.0, + 476.0, + 295.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 470.0, + 738.0, + 470.0, + 738.0, + 501.0, + 323.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 335.0, + 1405.0, + 335.0, + 1405.0, + 371.0, + 295.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 362.0, + 1406.0, + 362.0, + 1406.0, + 401.0, + 319.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 393.0, + 634.0, + 393.0, + 634.0, + 426.0, + 322.0, + 426.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 434, + 1381, + 434, + 1381, + 901, + 301, + 901 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 295, + 928, + 1407, + 928, + 1407, + 1185, + 295, + 1185 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 367, + 1817, + 1333, + 1817, + 1333, + 1971, + 367, + 1971 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 296, + 1589, + 1407, + 1589, + 1407, + 1653, + 296, + 1653 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 571, + 1672, + 1128, + 1672, + 1128, + 1716, + 571, + 1716 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 298, + 220, + 537, + 220, + 537, + 282, + 298, + 282 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 293, + 1759, + 1309, + 1759, + 1309, + 1797, + 293, + 1797 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 296, + 1299, + 1194, + 1299, + 1194, + 1335, + 296, + 1335 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 1227, + 470, + 1227, + 470, + 1262, + 299, + 1262 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 813, + 74, + 813, + 106, + 298, + 106 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 321, + 2003, + 1372, + 2003, + 1372, + 2036, + 321, + 2036 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 296, + 1546, + 1140, + 1546, + 1140, + 1580, + 296, + 1580 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 371, + 1359, + 1294, + 1359, + 1294, + 1526, + 371, + 1526 + ], + "score": 0.834 + }, + { + "category_id": 0, + "poly": [ + 297, + 329, + 1057, + 329, + 1057, + 369, + 297, + 369 + ], + "score": 0.507 + }, + { + "category_id": 1, + "poly": [ + 297, + 329, + 1057, + 329, + 1057, + 369, + 297, + 369 + ], + "score": 0.453 + }, + { + "category_id": 8, + "poly": [ + 371, + 1359, + 1294, + 1359, + 1294, + 1526, + 371, + 1526 + ], + "score": 0.152 + }, + { + "category_id": 13, + "poly": [ + 898, + 1430, + 988, + 1430, + 988, + 1470, + 898, + 1470 + ], + "score": 0.93, + "latex": "Y _ { y } ^ { 0 } = y" + }, + { + "category_id": 13, + "poly": [ + 990, + 1761, + 1177, + 1761, + 1177, + 1796, + 990, + 1796 + ], + "score": 0.93, + "latex": "\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1096, + 1325, + 1096, + 1325, + 1130, + 1100, + 1130 + ], + "score": 0.93, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( \\cdot | \\boldsymbol { y } ) = \\boldsymbol { W } ^ { T } f _ { \\boldsymbol { y } } + \\boldsymbol { b }" + }, + { + "category_id": 13, + "poly": [ + 608, + 1430, + 812, + 1430, + 812, + 1470, + 608, + 1470 + ], + "score": 0.93, + "latex": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )" + }, + { + "category_id": 14, + "poly": [ + 365, + 1814, + 1332, + 1814, + 1332, + 1973, + 365, + 1973 + ], + "score": 0.92, + "latex": "\\begin{array} { l } { \\displaystyle \\mathop { \\operatorname* { m i n } } _ { \\textsc { i r } } \\sum _ { x \\in \\mathcal { X } , \\ y \\in \\mathcal { V } } \\Gamma ( x , y ) d ( x , y ) } \\\\ { \\mathrm { s u b j e c t ~ t o ~ } \\displaystyle \\sum _ { x } \\Gamma ( x , y ) = P y ( y ) \\forall y , \\quad \\displaystyle \\sum _ { y } \\Gamma ( x , y ) = P _ { \\mathcal { X } } ( x ) \\forall x , \\quad \\Gamma ( x , y ) \\geq 0 \\forall x , y } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 479, + 1068, + 644, + 1068, + 644, + 1100, + 479, + 1100 + ], + "score": 0.91, + "latex": "z _ { \\theta } ( y ) = h _ { \\theta } ( f _ { y } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1622, + 413, + 1622, + 413, + 1655, + 297, + 1655 + ], + "score": 0.91, + "latex": "\\mathcal { X } , \\mathcal { Y } \\subset \\mathcal { S }" + }, + { + "category_id": 13, + "poly": [ + 488, + 1433, + 561, + 1433, + 561, + 1465, + 488, + 1465 + ], + "score": 0.9, + "latex": "y \\in Y" + }, + { + "category_id": 13, + "poly": [ + 780, + 1590, + 896, + 1590, + 896, + 1623, + 780, + 1623 + ], + "score": 0.9, + "latex": "d , d ^ { \\prime } \\in \\mathbb { M }" + }, + { + "category_id": 13, + "poly": [ + 937, + 958, + 1218, + 958, + 1218, + 988, + 937, + 988 + ], + "score": 0.9, + "latex": "\\bar { \\Psi } _ { x } : = \\Psi ( x ) , \\Psi _ { y } : = \\bar { \\Psi } ( y )" + }, + { + "category_id": 13, + "poly": [ + 991, + 931, + 1051, + 931, + 1051, + 960, + 991, + 960 + ], + "score": 0.89, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1041, + 376, + 1041, + 376, + 1070, + 298, + 1070 + ], + "score": 0.89, + "latex": "f _ { \\theta } ( \\hat { \\Psi } _ { y } )" + }, + { + "category_id": 13, + "poly": [ + 940, + 987, + 1062, + 987, + 1062, + 1016, + 940, + 1016 + ], + "score": 0.88, + "latex": "x \\Psi ( x ) = x" + }, + { + "category_id": 13, + "poly": [ + 1021, + 1042, + 1050, + 1042, + 1050, + 1068, + 1021, + 1068 + ], + "score": 0.87, + "latex": "h _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1015, + 1105, + 1015, + 1105, + 1042, + 1079, + 1042 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 14, + "poly": [ + 571, + 1672, + 1127, + 1672, + 1127, + 1714, + 571, + 1714 + ], + "score": 0.86, + "latex": "\\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } )" + }, + { + "category_id": 13, + "poly": [ + 558, + 1105, + 587, + 1105, + 587, + 1128, + 558, + 1128 + ], + "score": 0.85, + "latex": "\\pi \\theta" + }, + { + "category_id": 13, + "poly": [ + 298, + 1069, + 464, + 1069, + 464, + 1099, + 298, + 1099 + ], + "score": 0.84, + "latex": "z _ { \\theta } ( x ) = h _ { \\theta } ( f _ { x } )" + }, + { + "category_id": 14, + "poly": [ + 513, + 1357, + 1295, + 1357, + 1295, + 1418, + 513, + 1418 + ], + "score": 0.83, + "latex": "\\begin{array} { r } { { ! } \\mathbb { E } _ { t \\ge 0 } [ \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) ] = \\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1128, + 346, + 1128, + 346, + 1156, + 298, + 1156 + ], + "score": 0.83, + "latex": "W , b" + }, + { + "category_id": 13, + "poly": [ + 1299, + 1047, + 1325, + 1047, + 1325, + 1068, + 1299, + 1068 + ], + "score": 0.83, + "latex": "z _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 536, + 959, + 560, + 959, + 560, + 983, + 536, + 983 + ], + "score": 0.82, + "latex": "\\Psi" + }, + { + "category_id": 13, + "poly": [ + 496, + 987, + 519, + 987, + 519, + 1011, + 496, + 1011 + ], + "score": 0.82, + "latex": "\\Psi" + }, + { + "category_id": 13, + "poly": [ + 1191, + 1013, + 1330, + 1013, + 1330, + 1044, + 1191, + 1044 + ], + "score": 0.81, + "latex": "{ \\bar { f } } _ { x } = f _ { \\theta } ( \\Psi _ { x } )" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1014, + 1404, + 1014, + 1404, + 1043, + 1345, + 1043 + ], + "score": 0.76, + "latex": "f _ { y } =" + }, + { + "category_id": 14, + "poly": [ + 387, + 1481, + 961, + 1481, + 961, + 1525, + 387, + 1525 + ], + "score": 0.74, + "latex": "T V ^ { n } ( Y _ { y } ^ { k } ) = \\mathbb { E } _ { 0 \\leq t < n } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { k + t } ) , \\pi ^ { * } ( Y _ { y } ^ { k + t } ) ) ." + }, + { + "category_id": 13, + "poly": [ + 1230, + 1592, + 1325, + 1592, + 1325, + 1625, + 1230, + 1625 + ], + "score": 0.73, + "latex": "P _ { X } , P _ { y }" + }, + { + "category_id": 13, + "poly": [ + 889, + 715, + 923, + 715, + 923, + 755, + 889, + 755 + ], + "score": 0.64, + "latex": "\\mathsf { h } _ { \\mathsf { \\theta } }" + }, + { + "category_id": 13, + "poly": [ + 519, + 1359, + 1297, + 1359, + 1297, + 1417, + 519, + 1417 + ], + "score": 0.6, + "latex": "\\begin{array} { r } { \\mathbb { E } _ { t \\geq 0 } [ \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) ] = \\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\geq 0 } \\gamma ^ { t } T V ( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) ) \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 771, + 686, + 800, + 686, + 800, + 733, + 771, + 733 + ], + "score": 0.5, + "latex": "\\mathsf { f } _ { \\mathsf { y } }" + }, + { + "category_id": 13, + "poly": [ + 1010, + 692, + 1045, + 692, + 1045, + 732, + 1010, + 732 + ], + "score": 0.49, + "latex": "z _ { y }" + }, + { + "category_id": 13, + "poly": [ + 666, + 715, + 692, + 715, + 692, + 755, + 666, + 755 + ], + "score": 0.41, + "latex": "{ \\mathfrak { f } } _ { \\mathfrak { g } }" + }, + { + "category_id": 13, + "poly": [ + 1230, + 1593, + 1269, + 1593, + 1269, + 1623, + 1230, + 1623 + ], + "score": 0.33, + "latex": "P _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 342, + 565, + 365, + 565, + 365, + 588, + 342, + 588 + ], + "score": 0.32, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 549, + 693, + 591, + 693, + 591, + 733, + 549, + 733 + ], + "score": 0.32, + "latex": "\\Psi _ { \\ y }" + }, + { + "category_id": 13, + "poly": [ + 1013, + 541, + 1048, + 541, + 1048, + 578, + 1013, + 578 + ], + "score": 0.26, + "latex": "Z _ { _ { X } }" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 435.0, + 911.0, + 435.0, + 911.0, + 479.0, + 466.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 433.0, + 1103.0, + 433.0, + 1103.0, + 483.0, + 943.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 483.0, + 401.0, + 483.0, + 401.0, + 534.0, + 308.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 537.0, + 580.0, + 537.0, + 580.0, + 556.0, + 551.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 526.0, + 799.0, + 526.0, + 799.0, + 562.0, + 770.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 556.0, + 341.0, + 556.0, + 341.0, + 598.0, + 335.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 556.0, + 375.0, + 556.0, + 375.0, + 598.0, + 366.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 559.0, + 595.0, + 559.0, + 595.0, + 580.0, + 574.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 559.0, + 804.0, + 559.0, + 804.0, + 580.0, + 782.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 579.0, + 1384.0, + 579.0, + 1384.0, + 628.0, + 1191.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 665.0, + 362.0, + 665.0, + 362.0, + 674.0, + 349.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 610.0, + 1370.0, + 610.0, + 1370.0, + 671.0, + 1222.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 689.0, + 577.0, + 689.0, + 577.0, + 707.0, + 548.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 773.0, + 852.0, + 773.0, + 852.0, + 823.0, + 784.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 785.0, + 1247.0, + 785.0, + 1247.0, + 837.0, + 995.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 822.0, + 1208.0, + 822.0, + 1208.0, + 880.0, + 1066.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 847.0, + 957.0, + 847.0, + 957.0, + 914.0, + 847.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.25, + 512.0, + 908.25, + 512.0, + 908.25, + 540.0, + 902.25, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 497.0, + 712.0, + 497.0, + 712.0, + 552.0, + 648.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.25, + 680.0, + 812.25, + 680.0, + 812.25, + 731.0, + 755.25, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 708.5, + 920.0, + 708.5, + 920.0, + 752.0, + 891.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.75, + 699.5, + 709.75, + 699.5, + 709.75, + 760.5, + 645.75, + 760.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 808.5, + 917.0, + 808.5, + 917.0, + 849.5, + 887.0, + 849.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 929.0, + 990.0, + 929.0, + 990.0, + 964.0, + 294.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 929.0, + 1406.0, + 929.0, + 1406.0, + 964.0, + 1052.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 952.0, + 535.0, + 952.0, + 535.0, + 996.0, + 290.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 952.0, + 936.0, + 952.0, + 936.0, + 996.0, + 561.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 952.0, + 1407.0, + 952.0, + 1407.0, + 996.0, + 1219.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 987.0, + 495.0, + 987.0, + 495.0, + 1018.0, + 296.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 987.0, + 939.0, + 987.0, + 939.0, + 1018.0, + 520.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1018.0, + 1063.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 1078.0, + 1013.0, + 1078.0, + 1048.0, + 293.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1013.0, + 1190.0, + 1013.0, + 1190.0, + 1048.0, + 1106.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1013.0, + 1344.0, + 1013.0, + 1344.0, + 1048.0, + 1331.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1042.0, + 1020.0, + 1042.0, + 1020.0, + 1073.0, + 377.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1042.0, + 1298.0, + 1042.0, + 1298.0, + 1073.0, + 1051.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1042.0, + 1408.0, + 1042.0, + 1408.0, + 1073.0, + 1326.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1068.0, + 297.0, + 1068.0, + 297.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1068.0, + 478.0, + 1068.0, + 478.0, + 1102.0, + 465.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1068.0, + 1406.0, + 1068.0, + 1406.0, + 1102.0, + 645.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1092.0, + 557.0, + 1092.0, + 557.0, + 1136.0, + 290.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1092.0, + 1099.0, + 1092.0, + 1099.0, + 1136.0, + 588.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1092.0, + 1408.0, + 1092.0, + 1408.0, + 1136.0, + 1326.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1126.0, + 1405.0, + 1126.0, + 1405.0, + 1159.0, + 347.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1154.0, + 1155.0, + 1154.0, + 1155.0, + 1187.0, + 294.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 215.0, + 545.0, + 215.0, + 545.0, + 287.0, + 288.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1223.0, + 475.0, + 1223.0, + 475.0, + 1269.0, + 292.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 814.0, + 72.0, + 814.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 328.0, + 1058.0, + 328.0, + 1058.0, + 373.0, + 294.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1586.0, + 779.0, + 1586.0, + 779.0, + 1630.0, + 291.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1586.0, + 1229.0, + 1586.0, + 1229.0, + 1630.0, + 897.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1586.0, + 1407.0, + 1586.0, + 1407.0, + 1630.0, + 1326.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1621.0, + 529.0, + 1621.0, + 529.0, + 1654.0, + 414.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1755.0, + 989.0, + 1755.0, + 989.0, + 1802.0, + 292.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1755.0, + 1310.0, + 1755.0, + 1310.0, + 1802.0, + 1178.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1298.0, + 1196.0, + 1298.0, + 1196.0, + 1338.0, + 295.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1997.0, + 1376.0, + 1997.0, + 1376.0, + 2039.0, + 329.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1544.0, + 1139.0, + 1544.0, + 1139.0, + 1584.0, + 295.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1356.0, + 512.0, + 1356.0, + 512.0, + 1419.0, + 365.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1423.0, + 487.0, + 1423.0, + 487.0, + 1478.0, + 367.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1423.0, + 607.0, + 1423.0, + 607.0, + 1478.0, + 562.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1423.0, + 897.0, + 1423.0, + 897.0, + 1478.0, + 813.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1423.0, + 1005.0, + 1423.0, + 1005.0, + 1478.0, + 989.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1475.0, + 386.0, + 1475.0, + 386.0, + 1531.0, + 368.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1475.0, + 965.0, + 1475.0, + 965.0, + 1531.0, + 962.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 328.0, + 1058.0, + 328.0, + 1058.0, + 373.0, + 294.0, + 373.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 291, + 842, + 1464, + 842, + 1464, + 1382, + 291, + 1382 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 372, + 269, + 1328, + 269, + 1328, + 598, + 372, + 598 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 671, + 1776, + 1029, + 1776, + 1029, + 1844, + 671, + 1844 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 404, + 1632, + 1293, + 1632, + 1293, + 1708, + 404, + 1708 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 297, + 1717, + 1401, + 1717, + 1401, + 1779, + 297, + 1779 + ], + "score": 0.934 + }, + { + "category_id": 8, + "poly": [ + 452, + 1914, + 1244, + 1914, + 1244, + 1991, + 452, + 1991 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 435, + 689, + 1265, + 689, + 1265, + 770, + 435, + 770 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 453, + 1479, + 1248, + 1479, + 1248, + 1562, + 453, + 1562 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 296, + 1593, + 1164, + 1593, + 1164, + 1627, + 296, + 1627 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 295, + 1426, + 1115, + 1426, + 1115, + 1472, + 295, + 1472 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 297, + 2002, + 546, + 2002, + 546, + 2035, + 297, + 2035 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 298, + 226, + 1241, + 226, + 1241, + 264, + 298, + 264 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 295, + 648, + 748, + 648, + 748, + 682, + 295, + 682 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 296, + 1862, + 1289, + 1862, + 1289, + 1904, + 296, + 1904 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 296, + 802, + 371, + 802, + 371, + 834, + 296, + 834 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2114, + 835, + 2114 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 1374, + 2003, + 1404, + 2003, + 1404, + 2033, + 1374, + 2033 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 1373, + 1384, + 1405, + 1384, + 1405, + 1415, + 1373, + 1415 + ], + "score": 0.814 + }, + { + "category_id": 2, + "poly": [ + 1373, + 600, + 1405, + 600, + 1405, + 631, + 1373, + 631 + ], + "score": 0.76 + }, + { + "category_id": 2, + "poly": [ + 1373, + 600, + 1405, + 600, + 1405, + 632, + 1373, + 632 + ], + "score": 0.351 + }, + { + "category_id": 14, + "poly": [ + 372, + 270, + 1333, + 270, + 1333, + 598, + 372, + 598 + ], + "score": 0.95, + "latex": "\\begin{array} { l } { \\displaystyle \\mathcal { W } ^ { 1 } ( d ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| = \\mathcal { W } ^ { 1 } ( d - d ^ { \\prime } + d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) \\leq \\| d - d ^ { \\prime } \\| } \\\\ { \\displaystyle \\quad \\quad \\quad = \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( d ( x , y ) - d ^ { \\prime } ( x , y ) + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad \\leq \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) \\left( \\| d - d ^ { \\prime } \\| + d ^ { \\prime } ( x , y ) \\right) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\operatorname* { m i n } _ { \\displaystyle \\Gamma } \\sum _ { x \\in \\mathcal { X } , y \\in \\mathcal { Y } } \\Gamma ( x , y ) d ^ { \\prime } ( x , y ) } \\\\ { \\displaystyle \\quad \\quad = \\| d - d ^ { \\prime } \\| + \\mathcal { W } ^ { 1 } ( d ^ { \\prime } ) ( P _ { \\mathcal { X } } , P _ { \\mathcal { Y } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 491, + 1862, + 653, + 1862, + 653, + 1907, + 491, + 1907 + ], + "score": 0.93, + "latex": "\\textstyle { \\frac { 1 - \\gamma } { 1 + \\gamma } } T V ^ { n } ( Y _ { . } ^ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 964, + 229, + 1045, + 229, + 1045, + 264, + 964, + 264 + ], + "score": 0.93, + "latex": "\\mathscr { W } ^ { 1 } ( d )" + }, + { + "category_id": 14, + "poly": [ + 669, + 1771, + 1030, + 1771, + 1030, + 1844, + 669, + 1844 + ], + "score": 0.93, + "latex": "T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { \\ast } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } )" + }, + { + "category_id": 14, + "poly": [ + 452, + 1911, + 1248, + 1911, + 1248, + 1996, + 452, + 1996 + ], + "score": 0.92, + "latex": "\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\frac { 1 - \\gamma } { 1 + \\gamma } T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 807, + 1717, + 1164, + 1717, + 1164, + 1756, + 807, + 1756 + ], + "score": 0.92, + "latex": "W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )" + }, + { + "category_id": 14, + "poly": [ + 402, + 1632, + 1297, + 1632, + 1297, + 1710, + 402, + 1710 + ], + "score": 0.91, + "latex": "\\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) = \\sum _ { y _ { 1 } \\in Y } P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) - \\sum _ { x \\in X } P ^ { \\pi ^ { * } } ( x | \\tilde { x } _ { y _ { 0 } } ) \\mathbf { 0 }" + }, + { + "category_id": 13, + "poly": [ + 550, + 651, + 633, + 651, + 633, + 682, + 550, + 682 + ], + "score": 0.91, + "latex": "y _ { 0 } \\in Y" + }, + { + "category_id": 14, + "poly": [ + 434, + 690, + 1263, + 690, + 1263, + 773, + 434, + 773 + ], + "score": 0.91, + "latex": "\\sum _ { y _ { 1 } \\in Y } \\left( P ^ { \\tilde { \\pi } } ( y _ { 1 } | y _ { 0 } ) - P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) \\right) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\leq \\frac { 2 } { 1 - \\gamma } T V ( \\tilde { \\pi } ( y _ { 0 } ) , \\pi ^ { * } ( y _ { 0 } ) )" + }, + { + "category_id": 14, + "poly": [ + 450, + 1479, + 1248, + 1479, + 1248, + 1564, + 450, + 1564 + ], + "score": 0.9, + "latex": "\\sum _ { y _ { 1 } \\in Y } { P ^ { \\pi ^ { * } } ( y _ { 1 } | y _ { 0 } ) T V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) } \\leq { \\frac { 1 + \\gamma } { 1 - \\gamma } } W _ { 1 } ( d ^ { * } ) \\left( P ^ { \\pi ^ { * } } ( \\cdot | { \\tilde { x } _ { y _ { 0 } } } ) , P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 921, + 1862, + 1281, + 1862, + 1281, + 1901, + 921, + 1901 + ], + "score": 0.89, + "latex": "W _ { 1 } ( d ^ { * } ) ( P ^ { \\pi ^ { * } } ( \\cdot | y _ { 0 } ) , P ^ { \\pi ^ { * } } ( \\cdot | \\tilde { x } _ { y _ { 0 } } ) )" + }, + { + "category_id": 14, + "poly": [ + 310, + 839, + 1428, + 839, + 1428, + 1387, + 310, + 1387 + ], + "score": 0.88, + "latex": "\\begin{array} { r l } { \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | | B \\rangle } ) - P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) \\leq | \\displaystyle \\sum _ { i \\in \\mathcal { K } } ( P ^ { \\kappa } ( g _ { 1 | | B \\rangle } ) - P ^ { \\kappa ^ { \\star } } ( g _ { 1 | | B \\rangle } ) ) T V ^ { n } ( Y _ { n } ^ { 0 } ) | } & { } \\\\ & { \\leq \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha | \\hat { \\sigma } ( a ) | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( a | \\mathfrak { g } _ { 1 } ) ) | T V ^ { n } ( Y _ { n } ^ { 0 } ) } \\\\ & { \\leq \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\displaystyle \\sum _ { i \\in \\mathcal { K } } \\sum _ { i \\in \\mathcal { K } } P ^ { \\kappa } ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) | \\overline { { \\tau } } ( \\mathfrak { g } ( | \\mathfrak { g } _ { 1 } \\rangle ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } ) | ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | \\displaystyle \\sum _ { i \\in \\mathcal { K } } P ( g _ { 1 } | \\mathfrak { g } _ { 1 } , \\alpha ) } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & { = \\displaystyle \\frac { 1 } { 1 - \\gamma } \\displaystyle \\sum _ { i \\in \\mathcal { K } } | \\overline { { \\tau } } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) - \\pi ^ { \\star } ( \\mathfrak { g } _ { 1 } | \\mathfrak { g } _ { 1 } ) | } \\\\ & = \\ \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 664, + 1429, + 999, + 1429, + 999, + 1474, + 664, + 1474 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { \\mathbf { \\dot { \\mathbf { \\theta } } T } V ^ { n } ( Y _ { y _ { 1 } } ^ { 0 } ) \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y _ { 1 } } , y _ { 1 } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 550, + 1434, + 634, + 1434, + 634, + 1466, + 550, + 1466 + ], + "score": 0.86, + "latex": "y _ { 0 } \\in Y" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 2008.0, + 1403.0, + 2008.0, + 1403.0, + 2034.0, + 1378.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 1389.0, + 1403.0, + 1389.0, + 1403.0, + 1416.0, + 1378.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 605.0, + 1403.0, + 605.0, + 1403.0, + 631.0, + 1379.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 606.0, + 1403.0, + 606.0, + 1403.0, + 634.0, + 1378.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1710.0, + 806.0, + 1710.0, + 806.0, + 1762.0, + 288.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1710.0, + 1407.0, + 1710.0, + 1407.0, + 1762.0, + 1165.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1749.0, + 408.0, + 1749.0, + 408.0, + 1786.0, + 292.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1591.0, + 1164.0, + 1591.0, + 1164.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1422.0, + 549.0, + 1422.0, + 549.0, + 1479.0, + 288.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1422.0, + 663.0, + 1422.0, + 663.0, + 1479.0, + 635.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1422.0, + 1124.0, + 1422.0, + 1124.0, + 1479.0, + 1000.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 547.0, + 2002.0, + 547.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 963.0, + 226.0, + 963.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 226.0, + 1244.0, + 226.0, + 1244.0, + 268.0, + 1046.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 646.0, + 549.0, + 646.0, + 549.0, + 686.0, + 296.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 646.0, + 750.0, + 646.0, + 750.0, + 686.0, + 634.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1854.0, + 490.0, + 1854.0, + 490.0, + 1913.0, + 289.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1854.0, + 920.0, + 1854.0, + 920.0, + 1913.0, + 654.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1854.0, + 1296.0, + 1854.0, + 1296.0, + 1913.0, + 1282.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 797.0, + 376.0, + 797.0, + 376.0, + 844.0, + 290.0, + 844.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1052, + 1406, + 1052, + 1406, + 1164, + 296, + 1164 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 310, + 1251, + 1538, + 1251, + 1538, + 2036, + 310, + 2036 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 358, + 497, + 1341, + 497, + 1341, + 709, + 358, + 709 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 298, + 391, + 1404, + 391, + 1404, + 487, + 298, + 487 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 531, + 916, + 1162, + 916, + 1162, + 1028, + 531, + 1028 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 295, + 833, + 1397, + 833, + 1397, + 908, + 295, + 908 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 294, + 753, + 1405, + 753, + 1405, + 819, + 294, + 819 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1176, + 1400, + 1176, + 1400, + 1238, + 298, + 1238 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 814, + 73, + 814, + 106, + 297, + 106 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 298, + 326, + 1120, + 326, + 1120, + 361, + 298, + 361 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 790, + 228, + 790, + 263, + 298, + 263 + ], + "score": 0.914 + }, + { + "category_id": 8, + "poly": [ + 474, + 274, + 1226, + 274, + 1226, + 318, + 474, + 318 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2114, + 835, + 2114 + ], + "score": 0.774 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2114, + 835, + 2114 + ], + "score": 0.198 + }, + { + "category_id": 14, + "poly": [ + 312, + 1243, + 1428, + 1243, + 1428, + 2016, + 312, + 2016 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { \\gamma ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\mathcal { K } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) - \\gamma \\varepsilon \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\Bigg [ \\varepsilon \\gamma \\varepsilon ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\varepsilon \\gamma ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\varepsilon ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) \\Bigg ] } \\\\ & { - \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\Delta \\mathbf { x } ^ { \\varepsilon } \\Bigg ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & { \\quad + \\sum _ { s \\in \\mathcal { N } } ^ { \\varepsilon } \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) + \\gamma \\gamma ^ { \\varepsilon } ( \\Phi _ { 0 } ^ { \\varepsilon } ) \\gamma ^ { \\varepsilon } ( \\mathbf { x } ^ { \\varepsilon } ) } \\\\ & \\quad - \\sum _ { s \\in \\mathcal { N } } ^ \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 527, + 1116, + 708, + 1116, + 708, + 1165, + 527, + 1165 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { T V ^ { n } \\leq \\frac { 1 - \\gamma ^ { n + 1 } } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 533, + 915, + 1164, + 915, + 1164, + 1026, + 533, + 1026 + ], + "score": 0.94, + "latex": "\\mathbb { E } _ { Y _ { y } ^ { t } } \\left[ \\sum _ { t \\ge 0 } \\gamma ^ { t } T V \\left( \\tilde { \\pi } ( Y _ { y } ^ { t } ) , \\pi ^ { * } ( Y _ { y } ^ { t } ) \\right) \\right] \\le \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) ." + }, + { + "category_id": 14, + "poly": [ + 357, + 493, + 1340, + 493, + 1340, + 709, + 357, + 709 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathcal { F } ( d ) ( x , y ) - \\mathcal { F } ( d ^ { \\prime } ) ( x , y ) } \\\\ & { \\quad = \\gamma \\left( \\mathcal { W } _ { 1 } ( d ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad \\overset { L e m m a ~ \\ : 1 } { \\leq } \\gamma \\left( \\| d - d ^ { \\prime } \\| + \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) - \\mathcal { W } _ { 1 } ( d ^ { \\prime } ) ( P ^ { \\pi _ { \\mathcal { X } } ^ { * } } ( \\cdot | x ) , P _ { y } ^ { \\pi ^ { * } } ( \\cdot | y ) ) \\right) } \\\\ & { \\quad = \\gamma \\| d - d ^ { \\prime } \\| } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 962, + 834, + 1176, + 834, + 1176, + 874, + 962, + 874 + ], + "score": 0.93, + "latex": "Y _ { y } ^ { t } \\sim P ^ { \\tilde { \\pi } } ( \\cdot | Y _ { y } ^ { t - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 341, + 1178, + 447, + 1178, + 447, + 1212, + 341, + 1212 + ], + "score": 0.93, + "latex": "P _ { t } ^ { \\pi } ( y ^ { \\prime } | y )" + }, + { + "category_id": 13, + "poly": [ + 753, + 870, + 842, + 870, + 842, + 910, + 753, + 910 + ], + "score": 0.92, + "latex": "Y _ { y } ^ { 0 } = y" + }, + { + "category_id": 13, + "poly": [ + 366, + 756, + 700, + 756, + 700, + 790, + 366, + 790 + ], + "score": 0.92, + "latex": "\\| \\mathcal F ( d ) - \\mathcal F ( d ^ { \\prime } ) \\| \\le \\gamma \\| d - d ^ { \\prime } \\|" + }, + { + "category_id": 13, + "poly": [ + 901, + 1178, + 982, + 1178, + 982, + 1211, + 901, + 1211 + ], + "score": 0.92, + "latex": "y ^ { \\prime } \\in Y" + }, + { + "category_id": 13, + "poly": [ + 1283, + 423, + 1397, + 423, + 1397, + 455, + 1283, + 455 + ], + "score": 0.92, + "latex": "d , d ^ { \\prime } \\in \\mathbb { M }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1124, + 404, + 1124, + 404, + 1160, + 298, + 1160 + ], + "score": 0.91, + "latex": "d ^ { * } ( \\tilde { x } _ { y } , y )" + }, + { + "category_id": 13, + "poly": [ + 915, + 1084, + 1406, + 1084, + 1406, + 1119, + 915, + 1119 + ], + "score": 0.91, + "latex": "T V ( \\tilde { \\pi } ( y ) , \\pi ^ { * } ( y ) ) = T V ( \\pi ^ { * } ( \\tilde { x } _ { y } ) , \\pi ^ { * } ( y ) ) \\leq" + }, + { + "category_id": 13, + "poly": [ + 1129, + 759, + 1199, + 759, + 1199, + 789, + 1129, + 789 + ], + "score": 0.91, + "latex": "\\gamma < 1" + }, + { + "category_id": 13, + "poly": [ + 844, + 837, + 916, + 837, + 916, + 869, + 844, + 869 + ], + "score": 0.9, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 401, + 1084, + 488, + 1084, + 488, + 1112, + 401, + 1112 + ], + "score": 0.9, + "latex": "T V ^ { n + \\bar { 1 } }" + }, + { + "category_id": 13, + "poly": [ + 696, + 1086, + 765, + 1086, + 765, + 1112, + 696, + 1112 + ], + "score": 0.89, + "latex": "n = 1" + }, + { + "category_id": 14, + "poly": [ + 474, + 273, + 1226, + 273, + 1226, + 316, + 474, + 316 + ], + "score": 0.88, + "latex": "\\mathcal F ( d ) ( x , y ) = \\mathrm { D I S T } \\big ( \\pi ^ { * } ( x ) , \\pi ^ { * } ( y ) \\big ) + \\gamma \\mathcal W _ { 1 } ( d ) \\big ( P _ { \\mathcal X } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal V } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1056, + 1174, + 1056, + 1174, + 1082, + 1114, + 1082 + ], + "score": 0.87, + "latex": "T V ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 588, + 395, + 613, + 395, + 613, + 421, + 588, + 421 + ], + "score": 0.86, + "latex": "\\mathcal { F }" + }, + { + "category_id": 13, + "poly": [ + 607, + 425, + 633, + 425, + 633, + 451, + 607, + 451 + ], + "score": 0.86, + "latex": "\\mathcal { F }" + }, + { + "category_id": 13, + "poly": [ + 424, + 790, + 452, + 790, + 452, + 814, + 424, + 814 + ], + "score": 0.85, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 789, + 758, + 815, + 758, + 815, + 784, + 789, + 784 + ], + "score": 0.85, + "latex": "\\mathcal { F }" + }, + { + "category_id": 13, + "poly": [ + 487, + 455, + 562, + 455, + 562, + 482, + 487, + 482 + ], + "score": 0.84, + "latex": "x \\in \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 574, + 455, + 646, + 455, + 646, + 486, + 574, + 486 + ], + "score": 0.82, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 653, + 232, + 679, + 232, + 679, + 259, + 653, + 259 + ], + "score": 0.82, + "latex": "\\mathcal { F }" + }, + { + "category_id": 13, + "poly": [ + 817, + 1125, + 863, + 1125, + 863, + 1154, + 817, + 1154 + ], + "score": 0.79, + "latex": "T V" + }, + { + "category_id": 13, + "poly": [ + 553, + 1212, + 570, + 1212, + 570, + 1239, + 553, + 1239 + ], + "score": 0.78, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1044, + 1182, + 1058, + 1182, + 1058, + 1206, + 1044, + 1206 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1185, + 1401, + 1185, + 1401, + 1206, + 1381, + 1206 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1080, + 876, + 1099, + 876, + 1099, + 900, + 1080, + 900 + ], + "score": 0.64, + "latex": "\\tilde { \\pi }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1052.0, + 1113.0, + 1052.0, + 1113.0, + 1085.0, + 296.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1052.0, + 1403.0, + 1052.0, + 1403.0, + 1085.0, + 1175.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1076.0, + 400.0, + 1076.0, + 400.0, + 1123.0, + 291.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1076.0, + 695.0, + 1076.0, + 695.0, + 1123.0, + 489.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1076.0, + 914.0, + 1076.0, + 914.0, + 1123.0, + 766.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1124.0, + 816.0, + 1124.0, + 816.0, + 1160.0, + 709.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1124.0, + 1308.0, + 1124.0, + 1308.0, + 1160.0, + 864.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.75, + 1108.5, + 720.75, + 1108.5, + 720.75, + 1167.0, + 286.75, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 392.0, + 587.0, + 392.0, + 587.0, + 427.0, + 294.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 392.0, + 1405.0, + 392.0, + 1405.0, + 427.0, + 614.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 423.0, + 606.0, + 423.0, + 606.0, + 457.0, + 294.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 423.0, + 1282.0, + 423.0, + 1282.0, + 457.0, + 634.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 423.0, + 1408.0, + 423.0, + 1408.0, + 457.0, + 1398.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 455.0, + 486.0, + 455.0, + 486.0, + 486.0, + 296.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 455.0, + 573.0, + 455.0, + 573.0, + 486.0, + 563.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 455.0, + 656.0, + 455.0, + 656.0, + 486.0, + 647.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 824.0, + 843.0, + 824.0, + 843.0, + 888.0, + 288.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 824.0, + 961.0, + 824.0, + 961.0, + 888.0, + 917.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 824.0, + 1409.0, + 824.0, + 1409.0, + 888.0, + 1177.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 866.0, + 752.0, + 866.0, + 752.0, + 914.0, + 294.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 866.0, + 1079.0, + 866.0, + 1079.0, + 914.0, + 843.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 866.0, + 1222.0, + 866.0, + 1222.0, + 914.0, + 1100.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 750.0, + 365.0, + 750.0, + 365.0, + 793.0, + 292.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 750.0, + 788.0, + 750.0, + 788.0, + 793.0, + 701.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 750.0, + 1128.0, + 750.0, + 1128.0, + 793.0, + 816.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 750.0, + 1406.0, + 750.0, + 1406.0, + 793.0, + 1200.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 423.0, + 783.0, + 423.0, + 821.0, + 296.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 783.0, + 466.0, + 783.0, + 466.0, + 821.0, + 453.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1174.0, + 340.0, + 1174.0, + 340.0, + 1216.0, + 293.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1174.0, + 900.0, + 1174.0, + 900.0, + 1216.0, + 448.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1174.0, + 1043.0, + 1174.0, + 1043.0, + 1216.0, + 983.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1174.0, + 1380.0, + 1174.0, + 1380.0, + 1216.0, + 1059.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1174.0, + 1405.0, + 1174.0, + 1405.0, + 1216.0, + 1402.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 552.0, + 1206.0, + 552.0, + 1241.0, + 294.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1206.0, + 746.0, + 1206.0, + 746.0, + 1241.0, + 571.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 325.0, + 1123.0, + 325.0, + 1123.0, + 364.0, + 294.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 223.0, + 652.0, + 223.0, + 652.0, + 271.0, + 292.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 223.0, + 791.0, + 223.0, + 791.0, + 271.0, + 680.0, + 271.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1459, + 1406, + 1459, + 1406, + 1651, + 296, + 1651 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1663, + 1405, + 1663, + 1405, + 1788, + 298, + 1788 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 879, + 1405, + 879, + 1405, + 1035, + 296, + 1035 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1405, + 1880, + 1405, + 2035, + 299, + 2035 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1308, + 1405, + 1308, + 1405, + 1401, + 297, + 1401 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 535, + 1406, + 535, + 1406, + 629, + 297, + 629 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 295, + 695, + 1406, + 695, + 1406, + 789, + 295, + 789 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 682, + 272, + 1018, + 272, + 1018, + 340, + 682, + 340 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 578, + 1408, + 1117, + 1408, + 1117, + 1450, + 578, + 1450 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 439, + 637, + 1256, + 637, + 1256, + 685, + 439, + 685 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 461, + 1244, + 1237, + 1244, + 1237, + 1288, + 461, + 1288 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 299, + 1200, + 1121, + 1200, + 1121, + 1235, + 299, + 1235 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 346, + 602, + 346, + 602, + 379, + 298, + 379 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 783, + 229, + 783, + 262, + 299, + 262 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 296, + 488, + 902, + 488, + 902, + 521, + 296, + 521 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 386, + 798, + 1311, + 798, + 1311, + 872, + 386, + 872 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 300, + 421, + 707, + 421, + 707, + 457, + 300, + 457 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 1077, + 777, + 1077, + 777, + 1113, + 300, + 1113 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1250, + 1400, + 1250, + 1400, + 1282, + 1339, + 1282 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 298, + 1143, + 590, + 1143, + 590, + 1177, + 298, + 1177 + ], + "score": 0.906 + }, + { + "category_id": 9, + "poly": [ + 1341, + 640, + 1400, + 640, + 1400, + 672, + 1341, + 672 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1412, + 1400, + 1412, + 1400, + 1443, + 1339, + 1443 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 298, + 1822, + 1171, + 1822, + 1171, + 1857, + 298, + 1857 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 863, + 2087, + 863, + 2112, + 835, + 2112 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 1374, + 347, + 1403, + 347, + 1403, + 375, + 1374, + 375 + ], + "score": 0.828 + }, + { + "category_id": 13, + "poly": [ + 914, + 911, + 1041, + 911, + 1041, + 945, + 914, + 945 + ], + "score": 0.94, + "latex": "P _ { \\mathcal { Y } } ( x ) = 0" + }, + { + "category_id": 14, + "poly": [ + 679, + 269, + 1016, + 269, + 1016, + 341, + 679, + 341 + ], + "score": 0.94, + "latex": "T V ^ { n } ( Y _ { y } ^ { 0 } ) \\leq \\frac { 1 + \\gamma } { 1 - \\gamma } d ^ { * } ( \\tilde { x } _ { y } , y ) ," + }, + { + "category_id": 13, + "poly": [ + 731, + 1585, + 888, + 1585, + 888, + 1623, + 731, + 1623 + ], + "score": 0.94, + "latex": "\\tau _ { y } ^ { * } = \\{ y _ { t } \\} _ { t = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 517, + 1585, + 679, + 1585, + 679, + 1622, + 517, + 1622 + ], + "score": 0.94, + "latex": "\\tau _ { \\mathcal { X } } ^ { * } = \\{ x _ { t } \\} _ { t = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 585, + 911, + 713, + 911, + 713, + 945, + 585, + 945 + ], + "score": 0.93, + "latex": "P _ { \\mathcal { X } } ( y ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1583, + 1397, + 1583, + 1397, + 1624, + 1213, + 1624 + ], + "score": 0.93, + "latex": "y _ { t + 1 } = P _ { y } ^ { \\pi ^ { * } } ( y _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 973, + 1584, + 1162, + 1584, + 1162, + 1621, + 973, + 1621 + ], + "score": 0.93, + "latex": "x _ { t + 1 } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 858, + 727, + 935, + 727, + 935, + 760, + 858, + 760 + ], + "score": 0.93, + "latex": "\\mathscr { W } _ { 1 } ( d )" + }, + { + "category_id": 14, + "poly": [ + 385, + 795, + 1312, + 795, + 1312, + 874, + 385, + 874 + ], + "score": 0.92, + "latex": "\\operatorname* { m a x } _ { \\mathbf { u } , \\mathbf { v } } \\sum _ { x \\in \\mathcal { X } } P ( x ) u _ { x } - \\sum _ { y \\in \\mathcal { Y } } P ( y ) v _ { y } \\quad \\mathrm { s u b j e c t ~ t o } \\forall x \\in \\mathcal { X } , y \\in \\mathcal { Y } \\quad u _ { x } - v _ { y } \\leq d ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 369, + 567, + 468, + 567, + 468, + 599, + 369, + 599 + ], + "score": 0.92, + "latex": "d _ { \\sim } ( x , y )" + }, + { + "category_id": 14, + "poly": [ + 580, + 1407, + 1120, + 1407, + 1120, + 1451, + 580, + 1451 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma d ^ { * } ( x ^ { \\prime } , y ^ { \\prime } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 373, + 727, + 545, + 727, + 545, + 761, + 373, + 761 + ], + "score": 0.92, + "latex": "\\mathcal { W } _ { 1 } : \\mathbb { M } \\mathbb { M } _ { p }" + }, + { + "category_id": 13, + "poly": [ + 791, + 912, + 864, + 912, + 864, + 944, + 791, + 944 + ], + "score": 0.91, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 938, + 1204, + 991, + 1204, + 991, + 1235, + 938, + 1235 + ], + "score": 0.91, + "latex": "\\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 14, + "poly": [ + 441, + 636, + 1258, + 636, + 1258, + 687, + 441, + 687 + ], + "score": 0.9, + "latex": "\\mathcal { F } _ { \\sim } ^ { e } ( d ) ( x , y ) = \\operatorname* { m a x } _ { a \\in A } \\vert R ( x , a ) - R ( y , a ) \\vert + \\gamma \\mathcal { W } _ { 1 } ( d ) \\left( P ^ { a } ( \\cdot \\vert x ) , P ^ { a } ( \\cdot \\vert y ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1493, + 1375, + 1493, + 1375, + 1525, + 1322, + 1525 + ], + "score": 0.9, + "latex": "\\mathcal { M } _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 379, + 599, + 539, + 599, + 539, + 629, + 379, + 629 + ], + "score": 0.9, + "latex": "\\mathcal { F } _ { \\sim } ^ { e } : \\mathbb { M } \\mathbb { M }" + }, + { + "category_id": 13, + "poly": [ + 831, + 1204, + 886, + 1204, + 886, + 1233, + 831, + 1233 + ], + "score": 0.9, + "latex": "\\mathcal { M } _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 1118, + 912, + 1195, + 912, + 1195, + 940, + 1118, + 940 + ], + "score": 0.9, + "latex": "x \\in \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1215, + 1496, + 1271, + 1496, + 1271, + 1523, + 1215, + 1523 + ], + "score": 0.89, + "latex": "\\mathcal { M } _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 698, + 1135, + 698, + 1135, + 728, + 1094, + 728 + ], + "score": 0.89, + "latex": "\\mathcal { F } _ { \\sim } ^ { e }" + }, + { + "category_id": 13, + "poly": [ + 1207, + 537, + 1401, + 537, + 1401, + 566, + 1207, + 566 + ], + "score": 0.89, + "latex": "d _ { \\sim } : \\mathcal { S } \\times \\mathcal { S } \\to \\mathbb { R }" + }, + { + "category_id": 14, + "poly": [ + 461, + 1242, + 1238, + 1242, + 1238, + 1288, + 461, + 1288 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { d ^ { * } ( x , y ) = \\mathrm { D I S T } \\big ( \\pi _ { \\mathcal { X } } ^ { * } ( x ) , \\pi _ { \\mathcal { Y } } ^ { * } ( y ) \\big ) + \\gamma \\mathcal { W } _ { 1 } \\big ( d ^ { * } ) \\big ( P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( \\cdot | x ) , P _ { \\mathcal { Y } } ^ { \\pi ^ { * } } ( \\cdot | y ) \\big ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 327, + 942, + 362, + 942, + 362, + 971, + 327, + 971 + ], + "score": 0.88, + "latex": "d _ { \\sim }" + }, + { + "category_id": 13, + "poly": [ + 361, + 698, + 402, + 698, + 402, + 728, + 361, + 728 + ], + "score": 0.88, + "latex": "\\mathcal { F } _ { \\sim } ^ { e }" + }, + { + "category_id": 13, + "poly": [ + 531, + 1458, + 677, + 1458, + 677, + 1498, + 531, + 1498 + ], + "score": 0.87, + "latex": "y ^ { \\prime } = P _ { y } ^ { \\pi ^ { * } } ( y )" + }, + { + "category_id": 13, + "poly": [ + 689, + 697, + 725, + 697, + 725, + 726, + 689, + 726 + ], + "score": 0.87, + "latex": "d _ { \\sim }" + }, + { + "category_id": 13, + "poly": [ + 732, + 1372, + 761, + 1372, + 761, + 1398, + 732, + 1398 + ], + "score": 0.87, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 923, + 973, + 955, + 973, + 955, + 1000, + 923, + 1000 + ], + "score": 0.86, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1458, + 520, + 1458, + 520, + 1496, + 372, + 1496 + ], + "score": 0.83, + "latex": "x ^ { \\prime } = P _ { \\mathcal { X } } ^ { \\pi ^ { * } } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1174, + 882, + 1201, + 882, + 1201, + 908, + 1174, + 908 + ], + "score": 0.83, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 1462, + 1185, + 1462, + 1185, + 1496, + 1109, + 1496 + ], + "score": 0.83, + "latex": "\\pi _ { \\mathcal { X } } ^ { * } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1253, + 882, + 1277, + 882, + 1277, + 911, + 1253, + 911 + ], + "score": 0.82, + "latex": "\\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 728, + 1214, + 728, + 1214, + 754, + 1196, + 754 + ], + "score": 0.81, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 753, + 237, + 772, + 237, + 772, + 258, + 753, + 258 + ], + "score": 0.73, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 297, + 1498, + 351, + 1498, + 351, + 1525, + 297, + 1525 + ], + "score": 0.66, + "latex": "x , y" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1462, + 1270, + 1462, + 1270, + 1496, + 1196, + 1496 + ], + "score": 0.58, + "latex": "\\pi _ { \\mathcal { X } } ^ { * } ( y )" + }, + { + "category_id": 13, + "poly": [ + 1110, + 1462, + 1270, + 1462, + 1270, + 1496, + 1110, + 1496 + ], + "score": 0.29, + "latex": "\\pi _ { \\mathcal { X } } ^ { \\ast } ( x ) , \\pi _ { \\mathcal { X } } ^ { \\ast } ( y )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 417.0, + 710.0, + 417.0, + 710.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1073.0, + 779.0, + 1073.0, + 779.0, + 1118.0, + 294.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1141.0, + 592.0, + 1141.0, + 592.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1822.0, + 1175.0, + 1822.0, + 1175.0, + 1861.0, + 295.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 350.0, + 1403.0, + 350.0, + 1403.0, + 378.0, + 1377.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1456.0, + 371.0, + 1456.0, + 371.0, + 1500.0, + 293.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1456.0, + 530.0, + 1456.0, + 530.0, + 1500.0, + 521.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1456.0, + 1108.0, + 1456.0, + 1108.0, + 1500.0, + 678.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1456.0, + 1407.0, + 1456.0, + 1407.0, + 1500.0, + 1271.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 296.0, + 1490.0, + 296.0, + 1529.0, + 293.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1490.0, + 1214.0, + 1490.0, + 1214.0, + 1529.0, + 352.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1490.0, + 1321.0, + 1490.0, + 1321.0, + 1529.0, + 1272.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1490.0, + 1407.0, + 1490.0, + 1407.0, + 1529.0, + 1376.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1521.0, + 1408.0, + 1521.0, + 1408.0, + 1558.0, + 294.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1551.0, + 1407.0, + 1551.0, + 1407.0, + 1589.0, + 293.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1577.0, + 516.0, + 1577.0, + 516.0, + 1631.0, + 290.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1577.0, + 730.0, + 1577.0, + 730.0, + 1631.0, + 680.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1577.0, + 972.0, + 1577.0, + 972.0, + 1631.0, + 889.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1577.0, + 1212.0, + 1577.0, + 1212.0, + 1631.0, + 1163.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1577.0, + 1413.0, + 1577.0, + 1413.0, + 1631.0, + 1398.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1616.0, + 1409.0, + 1616.0, + 1409.0, + 1653.0, + 294.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1662.0, + 1406.0, + 1662.0, + 1406.0, + 1699.0, + 293.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1695.0, + 1406.0, + 1695.0, + 1406.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1724.0, + 1406.0, + 1724.0, + 1406.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1756.0, + 1333.0, + 1756.0, + 1333.0, + 1789.0, + 295.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 881.0, + 1173.0, + 881.0, + 1173.0, + 914.0, + 296.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 881.0, + 1252.0, + 881.0, + 1252.0, + 914.0, + 1202.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 881.0, + 1405.0, + 881.0, + 1405.0, + 914.0, + 1278.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 910.0, + 584.0, + 910.0, + 584.0, + 948.0, + 295.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 910.0, + 790.0, + 910.0, + 790.0, + 948.0, + 714.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 910.0, + 913.0, + 910.0, + 913.0, + 948.0, + 865.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 910.0, + 1117.0, + 910.0, + 1117.0, + 948.0, + 1042.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 910.0, + 1405.0, + 910.0, + 1405.0, + 948.0, + 1196.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 937.0, + 326.0, + 937.0, + 326.0, + 981.0, + 291.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 937.0, + 1406.0, + 937.0, + 1406.0, + 981.0, + 363.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 970.0, + 922.0, + 970.0, + 922.0, + 1008.0, + 294.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 970.0, + 1406.0, + 970.0, + 1406.0, + 1008.0, + 956.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1004.0, + 981.0, + 1004.0, + 981.0, + 1037.0, + 295.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1915.0, + 298.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1912.0, + 1403.0, + 1912.0, + 1403.0, + 1946.0, + 297.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 886.0, + 2003.0, + 886.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 1406.0, + 1305.0, + 1406.0, + 1346.0, + 294.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1406.0, + 1339.0, + 1406.0, + 1375.0, + 294.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1370.0, + 731.0, + 1370.0, + 731.0, + 1404.0, + 296.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1370.0, + 1390.0, + 1370.0, + 1390.0, + 1404.0, + 762.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 535.0, + 1206.0, + 535.0, + 1206.0, + 569.0, + 295.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 535.0, + 1406.0, + 535.0, + 1406.0, + 569.0, + 1402.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 565.0, + 368.0, + 565.0, + 368.0, + 600.0, + 294.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 565.0, + 1409.0, + 565.0, + 1409.0, + 600.0, + 469.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 592.0, + 378.0, + 592.0, + 378.0, + 633.0, + 293.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 592.0, + 577.0, + 592.0, + 577.0, + 633.0, + 540.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 695.0, + 360.0, + 695.0, + 360.0, + 729.0, + 294.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 695.0, + 688.0, + 695.0, + 688.0, + 729.0, + 403.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 695.0, + 1093.0, + 695.0, + 1093.0, + 729.0, + 726.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 695.0, + 1404.0, + 695.0, + 1404.0, + 729.0, + 1136.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 724.0, + 372.0, + 724.0, + 372.0, + 761.0, + 293.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 724.0, + 857.0, + 724.0, + 857.0, + 761.0, + 546.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 724.0, + 1195.0, + 724.0, + 1195.0, + 761.0, + 936.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 724.0, + 1406.0, + 724.0, + 1406.0, + 761.0, + 1215.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 755.0, + 637.0, + 755.0, + 637.0, + 793.0, + 296.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1196.0, + 830.0, + 1196.0, + 830.0, + 1244.0, + 292.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1196.0, + 937.0, + 1196.0, + 937.0, + 1244.0, + 887.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1196.0, + 1122.0, + 1196.0, + 1122.0, + 1244.0, + 992.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 342.0, + 604.0, + 342.0, + 604.0, + 383.0, + 295.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 752.0, + 229.0, + 752.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 229.0, + 783.0, + 229.0, + 783.0, + 265.0, + 773.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 486.0, + 905.0, + 486.0, + 905.0, + 523.0, + 294.0, + 523.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1028, + 1406, + 1028, + 1406, + 1245, + 296, + 1245 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 295, + 1706, + 1406, + 1706, + 1406, + 1801, + 295, + 1801 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1807, + 1404, + 1807, + 1404, + 1916, + 300, + 1916 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 697, + 1405, + 697, + 1405, + 824, + 296, + 824 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 296, + 1407, + 1404, + 1407, + 1404, + 1500, + 296, + 1500 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 296, + 1544, + 1407, + 1544, + 1407, + 1685, + 296, + 1685 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 390, + 217, + 1295, + 217, + 1295, + 508, + 390, + 508 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 294, + 836, + 1403, + 836, + 1403, + 900, + 294, + 900 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 1248, + 1402, + 1248, + 1402, + 1316, + 295, + 1316 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 417, + 1925, + 1276, + 1925, + 1276, + 2043, + 417, + 2043 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 298, + 601, + 1399, + 601, + 1399, + 680, + 298, + 680 + ], + "score": 0.936 + }, + { + "category_id": 8, + "poly": [ + 467, + 905, + 1230, + 905, + 1230, + 980, + 467, + 980 + ], + "score": 0.916 + }, + { + "category_id": 4, + "poly": [ + 295, + 533, + 1400, + 533, + 1400, + 593, + 295, + 593 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 296, + 1349, + 953, + 1349, + 953, + 1383, + 296, + 1383 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 813, + 75, + 813, + 105, + 298, + 105 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 293, + 982, + 1399, + 982, + 1399, + 1018, + 293, + 1018 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1507, + 1400, + 1507, + 1400, + 1537, + 1339, + 1537 + ], + "score": 0.882 + }, + { + "category_id": 8, + "poly": [ + 417, + 1502, + 1279, + 1502, + 1279, + 1542, + 417, + 1542 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 13, + "poly": [ + 894, + 1574, + 1164, + 1574, + 1164, + 1610, + 894, + 1610 + ], + "score": 0.92, + "latex": "d \\big ( ( x , \\pi _ { 1 } \\bar { ) } , ( y , \\pi _ { 2 } ) \\big ) ^ { \\cdot } \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 330, + 1281, + 465, + 1281, + 465, + 1316, + 330, + 1316 + ], + "score": 0.92, + "latex": "d ^ { * } ( x , \\psi ( x ) )" + }, + { + "category_id": 13, + "poly": [ + 913, + 1804, + 1106, + 1804, + 1106, + 1840, + 913, + 1840 + ], + "score": 0.92, + "latex": "\\hat { d } : \\mathcal { S } \\times \\mathcal { S } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 511, + 730, + 799, + 730, + 799, + 766, + 511, + 766 + ], + "score": 0.92, + "latex": "\\pi _ { \\mathcal { X } } ^ { \\ast } ( x _ { 1 } ) = \\pi _ { \\mathcal { Y } } ^ { \\ast } ( y _ { 1 } ) = a _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1250, + 1100, + 1250, + 1100, + 1284, + 1019, + 1284 + ], + "score": 0.92, + "latex": "M _ { \\psi ( \\chi ) }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 699, + 1406, + 699, + 1406, + 734, + 1283, + 734 + ], + "score": 0.91, + "latex": "\\pi _ { \\mathcal { X } } ^ { * } ( x _ { 0 } ) =" + }, + { + "category_id": 13, + "poly": [ + 498, + 1438, + 825, + 1438, + 825, + 1472, + 498, + 1472 + ], + "score": 0.91, + "latex": "\\dot { d } : ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\times ( \\boldsymbol { S } \\times \\boldsymbol { \\Pi } ) \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 602, + 1255, + 602, + 1255, + 636, + 1103, + 636 + ], + "score": 0.91, + "latex": "\\forall ( x , y ) \\in ~ { \\mathcal { L } }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 982, + 1399, + 982, + 1399, + 1016, + 1116, + 1016 + ], + "score": 0.91, + "latex": "d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1841, + 1408, + 1841, + 1408, + 1881, + 1040, + 1881 + ], + "score": 0.91, + "latex": "\\hat { d } ( x , y ) ~ = ~ \\mathrm { D I S T } \\big ( \\hat { \\pi } ( x ) , \\hat { \\pi } ( y ) \\big ) ~ + ~" + }, + { + "category_id": 13, + "poly": [ + 740, + 634, + 1063, + 634, + 1063, + 680, + 740, + 680 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { d _ { \\sim } ^ { \\ast } ( x , y ) = \\frac { \\left| R _ { \\mathrm { m a x } } - R _ { \\mathrm { m i n } } \\right| } { 1 - \\gamma } - \\epsilon } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 387, + 1607, + 855, + 1607, + 855, + 1645, + 387, + 1645 + ], + "score": 0.9, + "latex": "d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big ) \\ = \\ d \\big ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 365, + 982, + 578, + 982, + 578, + 1017, + 365, + 1017 + ], + "score": 0.9, + "latex": "r _ { y } > \\left( 1 + 1 / \\gamma \\right) r _ { x }" + }, + { + "category_id": 13, + "poly": [ + 1052, + 735, + 1121, + 735, + 1121, + 764, + 1052, + 764 + ], + "score": 0.89, + "latex": "r _ { x } , r _ { y }" + }, + { + "category_id": 13, + "poly": [ + 718, + 982, + 998, + 982, + 998, + 1016, + 718, + 1016 + ], + "score": 0.89, + "latex": "d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) < d _ { \\sim } ( x _ { 0 } , y _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1878, + 630, + 1878, + 630, + 1920, + 297, + 1920 + ], + "score": 0.89, + "latex": "\\gamma \\mathcal { W } _ { 1 } ( \\hat { d } ) \\big ( P ^ { \\hat { \\pi } } ( \\cdot \\vert x ) , P ^ { \\hat { \\pi } } ( \\cdot \\vert y ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 1001, + 1285, + 1077, + 1285, + 1077, + 1311, + 1001, + 1311 + ], + "score": 0.89, + "latex": "x \\in \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 297, + 730, + 454, + 730, + 454, + 764, + 297, + 764 + ], + "score": 0.89, + "latex": "\\pi _ { y } ^ { * } ( y _ { 0 } ) = a _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 870, + 1286, + 870, + 1286, + 900, + 1252, + 900 + ], + "score": 0.88, + "latex": "d _ { \\sim } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 990, + 870, + 1022, + 870, + 1022, + 896, + 990, + 896 + ], + "score": 0.88, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 903, + 870, + 939, + 870, + 939, + 898, + 903, + 898 + ], + "score": 0.88, + "latex": "d _ { \\sim }" + }, + { + "category_id": 13, + "poly": [ + 930, + 603, + 978, + 603, + 978, + 634, + 930, + 634 + ], + "score": 0.88, + "latex": "M _ { \\mathcal { Y } }" + }, + { + "category_id": 13, + "poly": [ + 555, + 568, + 620, + 568, + 620, + 592, + 555, + 592 + ], + "score": 0.88, + "latex": "x _ { 0 } , y _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 869, + 567, + 933, + 567, + 933, + 592, + 869, + 592 + ], + "score": 0.88, + "latex": "x _ { 2 } , y _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 819, + 603, + 869, + 603, + 869, + 633, + 819, + 633 + ], + "score": 0.88, + "latex": "M _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 650, + 1248, + 700, + 1248, + 700, + 1279, + 650, + 1279 + ], + "score": 0.87, + "latex": "M _ { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 883, + 1546, + 925, + 1546, + 925, + 1574, + 883, + 1574 + ], + "score": 0.87, + "latex": "\\mathcal { W } _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1225, + 641, + 1290, + 641, + 1290, + 669, + 1225, + 669 + ], + "score": 0.87, + "latex": "\\epsilon > 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 762, + 509, + 762, + 509, + 794, + 297, + 794 + ], + "score": 0.86, + "latex": "r _ { y } > ( 1 + 1 / \\gamma ) r _ { x }" + }, + { + "category_id": 13, + "poly": [ + 352, + 798, + 380, + 798, + 380, + 823, + 352, + 823 + ], + "score": 0.86, + "latex": "y _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1323, + 844, + 1354, + 844, + 1354, + 869, + 1323, + 869 + ], + "score": 0.86, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 995, + 765, + 1027, + 765, + 1027, + 791, + 995, + 791 + ], + "score": 0.86, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1341, + 602, + 1403, + 602, + 1403, + 634, + 1341, + 634 + ], + "score": 0.85, + "latex": "{ \\mathcal { L } } =" + }, + { + "category_id": 13, + "poly": [ + 297, + 1647, + 989, + 1647, + 989, + 1684, + 297, + 1684 + ], + "score": 0.84, + "latex": "d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) < d \\bigl ( ( x , \\pi _ { 1 } ) , ( z , \\pi _ { 3 } ) \\bigr ) + d \\bigl ( ( z , \\pi _ { 3 } ) , ( y , \\pi _ { 2 } ) \\bigr )" + }, + { + "category_id": 13, + "poly": [ + 1374, + 766, + 1402, + 766, + 1402, + 793, + 1374, + 793 + ], + "score": 0.84, + "latex": "y _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 887, + 944, + 1226, + 944, + 1226, + 979, + 887, + 979 + ], + "score": 0.82, + "latex": "d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 1 } ) = | r _ { y } - r _ { x } | + \\gamma r _ { x }" + }, + { + "category_id": 13, + "poly": [ + 923, + 844, + 942, + 844, + 942, + 869, + 923, + 869 + ], + "score": 0.82, + "latex": "y" + }, + { + "category_id": 14, + "poly": [ + 417, + 1924, + 1276, + 1924, + 1276, + 1968, + 417, + 1968 + ], + "score": 0.79, + "latex": "| d ^ { * } ( x , y ) - { \\hat { d } } ( x , y ) | < { { \\cal A } } \\bigl ( ( x , \\pi ^ { * } ) , ( x , { \\hat { \\pi } } ) \\bigr ) _ { \\bot } + { { \\cal A } } \\bigl ( ( y , { \\hat { \\pi } } ) , ( y , \\pi ^ { * } ) \\bigr ) _ { \\bot } ." + }, + { + "category_id": 13, + "poly": [ + 587, + 1577, + 605, + 1577, + 605, + 1604, + 587, + 1604 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 923, + 1613, + 943, + 1613, + 943, + 1639, + 923, + 1639 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1546, + 1402, + 1546, + 1402, + 1572, + 1383, + 1572 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1577, + 1240, + 1577, + 1240, + 1604, + 1221, + 1604 + ], + "score": 0.76, + "latex": "d" + }, + { + "category_id": 14, + "poly": [ + 415, + 1502, + 1280, + 1502, + 1280, + 1544, + 415, + 1544 + ], + "score": 0.75, + "latex": "d \\bigl ( ( x , \\pi _ { 1 } ) , ( y , \\pi _ { 2 } ) \\bigr ) = \\mathrm { D I S T } \\bigl ( \\pi _ { 1 } ( x ) , \\pi _ { 2 } ( y ) \\bigr ) + \\gamma { \\mathcal W } _ { 1 } ( d ) \\bigl ( P ^ { \\pi _ { 1 } } ( \\cdot | x ) , P ^ { \\pi _ { 2 } } ( \\cdot | y ) \\bigr ) ." + }, + { + "category_id": 13, + "poly": [ + 890, + 1284, + 911, + 1284, + 911, + 1315, + 890, + 1315 + ], + "score": 0.75, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 570, + 637, + 728, + 637, + 728, + 674, + 570, + 674 + ], + "score": 0.72, + "latex": "d ^ { * } ( x , y ) = 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 910, + 1847, + 932, + 1847, + 932, + 1873, + 910, + 1873 + ], + "score": 0.71, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 904, + 1440, + 928, + 1440, + 928, + 1466, + 904, + 1466 + ], + "score": 0.67, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1440, + 1305, + 1440, + 1305, + 1466, + 1287, + 1466 + ], + "score": 0.6, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1253, + 1440, + 1274, + 1440, + 1274, + 1466, + 1253, + 1466 + ], + "score": 0.6, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 481, + 639, + 556, + 639, + 556, + 672, + 481, + 672 + ], + "score": 0.54, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 729, + 1847, + 751, + 1847, + 751, + 1874, + 729, + 1874 + ], + "score": 0.53, + "latex": "\\hat { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1283, + 319, + 1283, + 319, + 1315, + 297, + 1315 + ], + "score": 0.47, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 547, + 985, + 577, + 985, + 577, + 1014, + 547, + 1014 + ], + "score": 0.44, + "latex": "r _ { x }" + }, + { + "category_id": 13, + "poly": [ + 903, + 2016, + 922, + 2016, + 922, + 2036, + 903, + 2036 + ], + "score": 0.39, + "latex": "x" + }, + { + "category_id": 14, + "poly": [ + 468, + 902, + 986, + 902, + 986, + 981, + 468, + 981 + ], + "score": 0.39, + "latex": "\\begin{array} { r l } & { d _ { \\sim } ( x _ { 0 } , y _ { 0 } ) = d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y _ { 0 } ) = ( 1 + \\gamma ) | r _ { y } - r _ { x } | } \\\\ & { d _ { \\sim } ( x _ { 0 } , y _ { 1 } ) = \\operatorname* { m a x } \\left( ( 1 + \\gamma ) r _ { x } , r _ { y } \\right) , \\ d _ { \\sim } ^ { \\ast } ( x _ { 0 } , y } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1207, + 2015, + 1227, + 2015, + 1227, + 2041, + 1207, + 2041 + ], + "score": 0.35, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1255, + 426, + 1283, + 426, + 1283, + 446, + 1255, + 446 + ], + "score": 0.34, + "latex": "a _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 445, + 300, + 473, + 300, + 473, + 319, + 445, + 319 + ], + "score": 0.33, + "latex": "r _ { x }" + }, + { + "category_id": 13, + "poly": [ + 479, + 764, + 508, + 764, + 508, + 792, + 479, + 792 + ], + "score": 0.31, + "latex": "r _ { x }" + }, + { + "category_id": 13, + "poly": [ + 298, + 637, + 467, + 637, + 467, + 674, + 298, + 674 + ], + "score": 0.29, + "latex": "\\{ ( x , y ) ~ | x \\in \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 441, + 412, + 471, + 412, + 471, + 431, + 441, + 431 + ], + "score": 0.25, + "latex": "x _ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 244.0, + 1068.0, + 244.0, + 1068.0, + 280.0, + 1026.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 289.0, + 444.0, + 289.0, + 444.0, + 326.0, + 390.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 289.0, + 482.0, + 289.0, + 482.0, + 326.0, + 474.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 287.0, + 783.0, + 287.0, + 783.0, + 327.0, + 729.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 288.0, + 1255.0, + 288.0, + 1255.0, + 326.0, + 1202.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 310.0, + 486.0, + 310.0, + 486.0, + 355.0, + 385.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 314.0, + 824.0, + 314.0, + 824.0, + 350.0, + 681.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 311.0, + 953.0, + 311.0, + 953.0, + 355.0, + 858.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 313.0, + 1297.0, + 313.0, + 1297.0, + 352.0, + 1154.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 331.0, + 645.0, + 331.0, + 645.0, + 373.0, + 506.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 333.0, + 1117.0, + 333.0, + 1117.0, + 373.0, + 979.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 385.0, + 826.0, + 385.0, + 826.0, + 424.0, + 774.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 401.0, + 949.0, + 401.0, + 949.0, + 437.0, + 907.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 401.0, + 1186.0, + 401.0, + 1186.0, + 437.0, + 1144.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 384.0, + 1299.0, + 384.0, + 1299.0, + 424.0, + 1247.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 422.0, + 657.0, + 422.0, + 657.0, + 447.0, + 632.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 426.0, + 745.0, + 426.0, + 745.0, + 438.0, + 732.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 422.0, + 1127.0, + 422.0, + 1127.0, + 447.0, + 1102.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 424.0, + 1219.0, + 424.0, + 1219.0, + 440.0, + 1203.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 447.0, + 616.0, + 447.0, + 616.0, + 483.0, + 534.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 445.0, + 1087.0, + 445.0, + 1087.0, + 485.0, + 1009.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 472.0, + 646.0, + 472.0, + 646.0, + 506.0, + 504.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 469.0, + 1123.0, + 469.0, + 1123.0, + 509.0, + 973.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 246.5, + 592.0, + 246.5, + 592.0, + 276.5, + 555.0, + 276.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 287.0, + 954.0, + 287.0, + 954.0, + 327.5, + 862.0, + 327.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 359.5, + 479.0, + 359.5, + 479.0, + 377.5, + 459.0, + 377.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.25, + 356.0, + 959.25, + 356.0, + 959.25, + 383.5, + 925.25, + 383.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 400.5, + 475.0, + 400.5, + 475.0, + 434.5, + 436.0, + 434.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 400.5, + 711.0, + 400.5, + 711.0, + 435.5, + 673.0, + 435.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.25, + 417.5, + 814.25, + 417.5, + 814.25, + 449.0, + 779.25, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 418.5, + 1287.0, + 418.5, + 1287.0, + 448.5, + 1251.0, + 448.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 532.0, + 1405.0, + 532.0, + 1405.0, + 568.0, + 294.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 563.0, + 554.0, + 563.0, + 554.0, + 594.0, + 295.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 563.0, + 868.0, + 563.0, + 868.0, + 594.0, + 621.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 563.0, + 1168.0, + 563.0, + 1168.0, + 594.0, + 934.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1348.0, + 953.0, + 1348.0, + 953.0, + 1386.0, + 294.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 814.0, + 73.0, + 814.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2122.0, + 831.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1028.0, + 1406.0, + 1028.0, + 1406.0, + 1063.0, + 295.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1060.0, + 1408.0, + 1060.0, + 1408.0, + 1094.0, + 295.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1088.0, + 1406.0, + 1088.0, + 1406.0, + 1124.0, + 294.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1119.0, + 1406.0, + 1119.0, + 1406.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1186.0, + 294.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1180.0, + 1404.0, + 1180.0, + 1404.0, + 1217.0, + 293.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1213.0, + 821.0, + 1213.0, + 821.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1706.0, + 1406.0, + 1706.0, + 1406.0, + 1744.0, + 295.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1736.0, + 1406.0, + 1736.0, + 1406.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1768.0, + 1324.0, + 1768.0, + 1324.0, + 1806.0, + 295.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 912.0, + 1805.0, + 912.0, + 1846.0, + 295.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1805.0, + 1406.0, + 1805.0, + 1406.0, + 1846.0, + 1107.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1842.0, + 728.0, + 1842.0, + 728.0, + 1882.0, + 295.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1842.0, + 909.0, + 1842.0, + 909.0, + 1882.0, + 752.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1842.0, + 1039.0, + 1842.0, + 1039.0, + 1882.0, + 933.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1880.0, + 754.0, + 1880.0, + 754.0, + 1920.0, + 631.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 692.0, + 1282.0, + 692.0, + 1282.0, + 737.0, + 291.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 727.0, + 510.0, + 727.0, + 510.0, + 768.0, + 455.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 727.0, + 1051.0, + 727.0, + 1051.0, + 768.0, + 800.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 727.0, + 1408.0, + 727.0, + 1408.0, + 768.0, + 1122.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 757.0, + 296.0, + 757.0, + 296.0, + 801.0, + 290.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 757.0, + 994.0, + 757.0, + 994.0, + 801.0, + 510.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 757.0, + 1373.0, + 757.0, + 1373.0, + 801.0, + 1028.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 757.0, + 1407.0, + 757.0, + 1407.0, + 801.0, + 1403.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 787.0, + 351.0, + 787.0, + 351.0, + 829.0, + 293.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 787.0, + 395.0, + 787.0, + 395.0, + 829.0, + 381.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 1407.0, + 1404.0, + 1407.0, + 1444.0, + 294.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1435.0, + 497.0, + 1435.0, + 497.0, + 1474.0, + 293.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1435.0, + 903.0, + 1435.0, + 903.0, + 1474.0, + 826.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1435.0, + 1252.0, + 1435.0, + 1252.0, + 1474.0, + 929.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 1435.0, + 1286.0, + 1435.0, + 1286.0, + 1474.0, + 1275.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1435.0, + 1406.0, + 1435.0, + 1406.0, + 1474.0, + 1306.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 556.0, + 1470.0, + 556.0, + 1504.0, + 296.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1543.0, + 882.0, + 1543.0, + 882.0, + 1578.0, + 296.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1543.0, + 1382.0, + 1543.0, + 1382.0, + 1578.0, + 926.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1543.0, + 1407.0, + 1543.0, + 1407.0, + 1578.0, + 1403.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1576.0, + 586.0, + 1576.0, + 586.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1576.0, + 893.0, + 1576.0, + 893.0, + 1611.0, + 606.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1576.0, + 1220.0, + 1576.0, + 1220.0, + 1611.0, + 1165.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1576.0, + 1409.0, + 1576.0, + 1409.0, + 1611.0, + 1241.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 386.0, + 1607.0, + 386.0, + 1648.0, + 293.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1607.0, + 922.0, + 1607.0, + 922.0, + 1648.0, + 856.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1607.0, + 1409.0, + 1607.0, + 1409.0, + 1648.0, + 944.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1643.0, + 1000.0, + 1643.0, + 1000.0, + 1686.0, + 990.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 833.0, + 922.0, + 833.0, + 922.0, + 873.0, + 294.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 833.0, + 1322.0, + 833.0, + 1322.0, + 873.0, + 943.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 833.0, + 1405.0, + 833.0, + 1405.0, + 873.0, + 1355.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 868.0, + 902.0, + 868.0, + 902.0, + 904.0, + 296.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 868.0, + 989.0, + 868.0, + 989.0, + 904.0, + 940.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 868.0, + 1251.0, + 868.0, + 1251.0, + 904.0, + 1023.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 868.0, + 1298.0, + 868.0, + 1298.0, + 904.0, + 1287.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1243.0, + 649.0, + 1243.0, + 649.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1243.0, + 1018.0, + 1243.0, + 1018.0, + 1287.0, + 701.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1243.0, + 1406.0, + 1243.0, + 1406.0, + 1287.0, + 1101.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1279.0, + 329.0, + 1279.0, + 329.0, + 1317.0, + 320.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1279.0, + 889.0, + 1279.0, + 889.0, + 1317.0, + 466.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1279.0, + 1000.0, + 1279.0, + 1000.0, + 1317.0, + 912.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1279.0, + 1090.0, + 1279.0, + 1090.0, + 1317.0, + 1078.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 599.0, + 818.0, + 599.0, + 818.0, + 638.0, + 293.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 599.0, + 929.0, + 599.0, + 929.0, + 638.0, + 870.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 599.0, + 1102.0, + 599.0, + 1102.0, + 638.0, + 979.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 599.0, + 1340.0, + 599.0, + 1340.0, + 638.0, + 1256.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 622.0, + 297.0, + 622.0, + 297.0, + 688.0, + 291.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 622.0, + 480.0, + 622.0, + 480.0, + 688.0, + 468.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 622.0, + 569.0, + 622.0, + 569.0, + 688.0, + 557.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 622.0, + 739.0, + 622.0, + 739.0, + 688.0, + 729.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 622.0, + 1224.0, + 622.0, + 1224.0, + 688.0, + 1064.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 622.0, + 1306.0, + 622.0, + 1306.0, + 688.0, + 1291.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 974.0, + 364.0, + 974.0, + 364.0, + 1024.0, + 292.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 974.0, + 717.0, + 974.0, + 717.0, + 1024.0, + 579.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 974.0, + 1115.0, + 974.0, + 1115.0, + 1024.0, + 999.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 974.0, + 1403.0, + 974.0, + 1403.0, + 1024.0, + 1400.0, + 1024.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1562, + 1405, + 1562, + 1405, + 1839, + 297, + 1839 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1072, + 1405, + 1072, + 1405, + 1319, + 297, + 1319 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 742, + 1405, + 742, + 1405, + 927, + 297, + 927 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1333, + 1406, + 1333, + 1406, + 1547, + 297, + 1547 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1853, + 1405, + 1853, + 1405, + 2037, + 297, + 2037 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 293, + 226, + 1404, + 226, + 1404, + 291, + 293, + 291 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 307, + 326, + 1394, + 326, + 1394, + 497, + 307, + 497 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 664, + 723, + 664, + 723, + 699, + 299, + 699 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 301, + 993, + 756, + 993, + 756, + 1029, + 301, + 1029 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 2, + "poly": [ + 1376, + 534, + 1403, + 534, + 1403, + 563, + 1376, + 563 + ], + "score": 0.828 + }, + { + "category_id": 14, + "poly": [ + 304, + 326, + 1388, + 326, + 1388, + 497, + 304, + 497 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { d ^ { * } ( x , y ) = d ( ( x , \\pi ^ { * } ) , ( y , \\pi ^ { * } ) ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\\\ & { \\qquad < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) + d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { d ^ { * } ( x , y ) - \\hat { d } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) \\quad : \\ \\hat { d } ( x , y ) = d \\big ( ( x , \\hat { \\pi } ) , ( y , \\hat { \\pi } ) \\big ) } \\\\ & { { \\mathrm { S i m i l a r l y } } , \\ \\hat { d } ( x , y ) - d ^ { * } ( x , y ) < d \\big ( ( x , \\pi ^ { * } ) , ( x , \\hat { \\pi } ) \\big ) + d \\big ( ( y , \\hat { \\pi } ) , ( y , \\pi ^ { * } ) \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 350, + 833, + 931, + 833, + 931, + 870, + 350, + 870 + ], + "score": 0.9, + "latex": "\\mathcal { L } ( \\theta ) = \\mathbb { E } _ { s _ { i } , s _ { j } } [ ( \\| f _ { \\theta } ( s _ { i } ) - f _ { \\theta } ( s _ { j } ) \\| _ { 2 } - d ^ { * } ( s _ { i } , s _ { j } ) ) ^ { \\bar { 2 } } ]" + }, + { + "category_id": 13, + "poly": [ + 1322, + 806, + 1350, + 806, + 1350, + 837, + 1322, + 837 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 595, + 261, + 628, + 261, + 628, + 286, + 595, + 286 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 806, + 1100, + 806, + 1100, + 832, + 1070, + 832 + ], + "score": 0.86, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 488, + 232, + 517, + 232, + 517, + 258, + 488, + 258 + ], + "score": 0.84, + "latex": "d ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 768, + 226, + 787, + 226, + 787, + 259, + 768, + 259 + ], + "score": 0.82, + "latex": "\\hat { d }" + }, + { + "category_id": 13, + "poly": [ + 678, + 260, + 699, + 260, + 699, + 286, + 678, + 286 + ], + "score": 0.8, + "latex": "\\hat { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1375, + 866, + 1402, + 866, + 1402, + 894, + 1375, + 894 + ], + "score": 0.36, + "latex": "l 2" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 661.0, + 727.0, + 661.0, + 727.0, + 705.0, + 293.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 992.0, + 760.0, + 992.0, + 760.0, + 1032.0, + 294.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 534.0, + 1404.0, + 534.0, + 1404.0, + 568.0, + 1378.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1563.0, + 1405.0, + 1563.0, + 1405.0, + 1596.0, + 296.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1626.0, + 295.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1621.0, + 1406.0, + 1621.0, + 1406.0, + 1659.0, + 292.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1651.0, + 1405.0, + 1651.0, + 1405.0, + 1690.0, + 292.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1718.0, + 295.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1714.0, + 1407.0, + 1714.0, + 1407.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1780.0, + 294.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1774.0, + 1406.0, + 1774.0, + 1406.0, + 1812.0, + 292.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 442.0, + 1805.0, + 442.0, + 1839.0, + 295.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1071.0, + 1406.0, + 1071.0, + 1406.0, + 1106.0, + 295.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1105.0, + 1403.0, + 1105.0, + 1403.0, + 1135.0, + 296.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1134.0, + 1407.0, + 1134.0, + 1407.0, + 1169.0, + 291.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 1407.0, + 1163.0, + 1407.0, + 1199.0, + 294.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1226.0, + 1405.0, + 1226.0, + 1405.0, + 1259.0, + 295.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1257.0, + 1406.0, + 1257.0, + 1406.0, + 1287.0, + 295.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1286.0, + 1407.0, + 1286.0, + 1407.0, + 1320.0, + 294.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 777.0, + 294.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 773.0, + 1403.0, + 773.0, + 1403.0, + 808.0, + 294.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 803.0, + 1069.0, + 803.0, + 1069.0, + 840.0, + 292.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 803.0, + 1321.0, + 803.0, + 1321.0, + 840.0, + 1101.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 803.0, + 1405.0, + 803.0, + 1405.0, + 840.0, + 1351.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 832.0, + 349.0, + 832.0, + 349.0, + 873.0, + 291.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 832.0, + 1407.0, + 832.0, + 1407.0, + 873.0, + 932.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 865.0, + 1374.0, + 865.0, + 1374.0, + 901.0, + 294.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 865.0, + 1406.0, + 865.0, + 1406.0, + 901.0, + 1403.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 894.0, + 807.0, + 894.0, + 807.0, + 930.0, + 294.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 1408.0, + 1328.0, + 1408.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1360.0, + 1406.0, + 1360.0, + 1406.0, + 1400.0, + 292.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1393.0, + 1408.0, + 1393.0, + 1408.0, + 1428.0, + 295.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1408.0, + 1423.0, + 1408.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 1407.0, + 1453.0, + 1407.0, + 1488.0, + 295.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1487.0, + 1403.0, + 1487.0, + 1403.0, + 1517.0, + 296.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 1392.0, + 1514.0, + 1392.0, + 1550.0, + 294.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1852.0, + 1406.0, + 1852.0, + 1406.0, + 1888.0, + 295.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1885.0, + 1405.0, + 1885.0, + 1405.0, + 1917.0, + 296.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1408.0, + 1913.0, + 1408.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1970.0, + 1407.0, + 1970.0, + 1407.0, + 2014.0, + 291.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2006.0, + 441.0, + 2006.0, + 441.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 487.0, + 224.0, + 487.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 224.0, + 767.0, + 224.0, + 767.0, + 268.0, + 518.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 224.0, + 1407.0, + 224.0, + 1407.0, + 268.0, + 788.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 256.0, + 594.0, + 256.0, + 594.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 256.0, + 677.0, + 256.0, + 677.0, + 295.0, + 629.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 256.0, + 845.0, + 256.0, + 845.0, + 295.0, + 700.0, + 295.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1024, + 1405, + 1024, + 1405, + 1210, + 296, + 1210 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1223, + 1404, + 1223, + 1404, + 1408, + 297, + 1408 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1423, + 1404, + 1423, + 1404, + 1517, + 297, + 1517 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 318, + 389, + 1381, + 389, + 1381, + 794, + 318, + 794 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 302, + 1652, + 1395, + 1652, + 1395, + 1915, + 302, + 1915 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 297, + 825, + 1407, + 825, + 1407, + 912, + 297, + 912 + ], + "score": 0.959 + }, + { + "category_id": 4, + "poly": [ + 296, + 1941, + 1402, + 1941, + 1402, + 2025, + 296, + 2025 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 1530, + 1401, + 1530, + 1401, + 1595, + 298, + 1595 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 866, + 2088, + 866, + 2113, + 835, + 2113 + ], + "score": 0.863 + }, + { + "category_id": 0, + "poly": [ + 299, + 225, + 769, + 225, + 769, + 262, + 299, + 262 + ], + "score": 0.833 + }, + { + "category_id": 13, + "poly": [ + 921, + 1531, + 1096, + 1531, + 1096, + 1563, + 921, + 1563 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { \\mathrm { I L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 1220, + 1424, + 1304, + 1424, + 1304, + 1452, + 1220, + 1452 + ], + "score": 0.9, + "latex": "k = 6 4" + }, + { + "category_id": 13, + "poly": [ + 656, + 1286, + 721, + 1286, + 721, + 1315, + 656, + 1315 + ], + "score": 0.9, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1532, + 1224, + 1532, + 1224, + 1562, + 1176, + 1562 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } _ { \\mathrm { I L } }" + }, + { + "category_id": 13, + "poly": [ + 542, + 1286, + 608, + 1286, + 608, + 1315, + 542, + 1315 + ], + "score": 0.86, + "latex": "4 \\times 4" + }, + { + "category_id": 13, + "poly": [ + 292, + 1118, + 358, + 1118, + 358, + 1147, + 292, + 1147 + ], + "score": 0.86, + "latex": "+ 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 451, + 1563, + 531, + 1563, + 531, + 1594, + 451, + 1594 + ], + "score": 0.85, + "latex": "\\mathcal { L } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 469, + 1286, + 533, + 1286, + 533, + 1315, + 469, + 1315 + ], + "score": 0.84, + "latex": "8 \\times 8" + }, + { + "category_id": 13, + "poly": [ + 1219, + 1057, + 1258, + 1057, + 1258, + 1085, + 1219, + 1085 + ], + "score": 0.83, + "latex": "+ 1" + }, + { + "category_id": 13, + "poly": [ + 843, + 1458, + 872, + 1458, + 872, + 1485, + 843, + 1485 + ], + "score": 0.83, + "latex": "z _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1568, + 1199, + 1568, + 1199, + 1590, + 1179, + 1590 + ], + "score": 0.77, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 391.0, + 1042.0, + 391.0, + 1042.0, + 418.0, + 579.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 600.0, + 1083.0, + 600.0, + 1083.0, + 627.0, + 621.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1654.0, + 337.0, + 1654.0, + 337.0, + 1865.0, + 304.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1659.0, + 497.0, + 1659.0, + 497.0, + 1685.0, + 443.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1655.0, + 699.0, + 1655.0, + 699.0, + 1865.0, + 667.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1659.0, + 860.0, + 1659.0, + 860.0, + 1685.0, + 807.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1650.0, + 1069.0, + 1650.0, + 1069.0, + 1868.0, + 1029.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1659.0, + 1223.0, + 1659.0, + 1223.0, + 1685.0, + 1170.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1674.0, + 386.0, + 1674.0, + 386.0, + 1700.0, + 338.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 1679.0, + 596.0, + 1679.0, + 596.0, + 1712.0, + 444.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1673.0, + 749.0, + 1673.0, + 749.0, + 1702.0, + 701.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1677.0, + 961.0, + 1677.0, + 961.0, + 1713.0, + 805.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1674.0, + 1112.0, + 1674.0, + 1112.0, + 1700.0, + 1063.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1683.0, + 1322.0, + 1683.0, + 1322.0, + 1710.0, + 1169.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1707.0, + 385.0, + 1707.0, + 385.0, + 1733.0, + 338.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1705.0, + 750.0, + 1705.0, + 750.0, + 1736.0, + 699.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1704.0, + 1112.0, + 1704.0, + 1112.0, + 1736.0, + 1062.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1738.0, + 386.0, + 1738.0, + 386.0, + 1767.0, + 338.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1738.0, + 749.0, + 1738.0, + 749.0, + 1767.0, + 701.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1738.0, + 1112.0, + 1738.0, + 1112.0, + 1767.0, + 1063.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1770.0, + 386.0, + 1770.0, + 386.0, + 1801.0, + 335.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1770.0, + 749.0, + 1770.0, + 749.0, + 1801.0, + 699.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1770.0, + 1113.0, + 1770.0, + 1113.0, + 1801.0, + 1062.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1804.0, + 386.0, + 1804.0, + 386.0, + 1833.0, + 338.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1804.0, + 747.0, + 1804.0, + 747.0, + 1830.0, + 702.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1804.0, + 1111.0, + 1804.0, + 1111.0, + 1831.0, + 1065.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1835.0, + 387.0, + 1835.0, + 387.0, + 1864.0, + 348.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1835.0, + 750.0, + 1835.0, + 750.0, + 1864.0, + 711.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1835.0, + 1113.0, + 1835.0, + 1113.0, + 1864.0, + 1073.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1857.0, + 402.0, + 1857.0, + 402.0, + 1881.0, + 381.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1857.0, + 477.0, + 1857.0, + 477.0, + 1883.0, + 432.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1857.0, + 609.0, + 1857.0, + 609.0, + 1882.0, + 489.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1855.0, + 666.0, + 1855.0, + 666.0, + 1881.0, + 610.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1857.0, + 765.0, + 1857.0, + 765.0, + 1879.0, + 744.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1857.0, + 840.0, + 1857.0, + 840.0, + 1883.0, + 795.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1857.0, + 908.0, + 1857.0, + 908.0, + 1882.0, + 853.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1857.0, + 970.0, + 1857.0, + 970.0, + 1882.0, + 914.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1855.0, + 1030.0, + 1855.0, + 1030.0, + 1882.0, + 976.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1857.0, + 1128.0, + 1857.0, + 1128.0, + 1879.0, + 1107.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1857.0, + 1204.0, + 1857.0, + 1204.0, + 1883.0, + 1158.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1857.0, + 1271.0, + 1857.0, + 1271.0, + 1882.0, + 1216.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1857.0, + 1333.0, + 1857.0, + 1333.0, + 1882.0, + 1277.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1855.0, + 1393.0, + 1855.0, + 1393.0, + 1882.0, + 1337.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1880.0, + 615.0, + 1880.0, + 615.0, + 1916.0, + 420.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1882.0, + 978.0, + 1882.0, + 978.0, + 1915.0, + 784.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1880.0, + 1341.0, + 1880.0, + 1341.0, + 1916.0, + 1146.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 858.0, + 295.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 854.0, + 1408.0, + 854.0, + 1408.0, + 887.0, + 295.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 879.0, + 562.0, + 879.0, + 562.0, + 915.0, + 293.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1971.0, + 295.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2000.0, + 294.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1997.0, + 490.0, + 1997.0, + 490.0, + 2027.0, + 296.0, + 2027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 223.0, + 774.0, + 223.0, + 774.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1023.0, + 1405.0, + 1023.0, + 1405.0, + 1061.0, + 293.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1057.0, + 1218.0, + 1057.0, + 1218.0, + 1089.0, + 295.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1057.0, + 1403.0, + 1057.0, + 1403.0, + 1089.0, + 1259.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1085.0, + 1407.0, + 1085.0, + 1407.0, + 1120.0, + 293.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1116.0, + 1406.0, + 1116.0, + 1406.0, + 1151.0, + 359.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1146.0, + 1406.0, + 1146.0, + 1406.0, + 1181.0, + 293.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1221.0, + 1177.0, + 1221.0, + 1213.0, + 294.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 1406.0, + 1224.0, + 1406.0, + 1260.0, + 294.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1254.0, + 1402.0, + 1254.0, + 1402.0, + 1286.0, + 295.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1286.0, + 468.0, + 1286.0, + 468.0, + 1318.0, + 295.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1286.0, + 541.0, + 1286.0, + 541.0, + 1318.0, + 534.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1286.0, + 655.0, + 1286.0, + 655.0, + 1318.0, + 609.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1318.0, + 722.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1314.0, + 1405.0, + 1314.0, + 1405.0, + 1350.0, + 294.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 1406.0, + 1346.0, + 1406.0, + 1382.0, + 294.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1377.0, + 602.0, + 1377.0, + 602.0, + 1412.0, + 295.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1420.0, + 1219.0, + 1420.0, + 1219.0, + 1459.0, + 296.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1420.0, + 1406.0, + 1420.0, + 1406.0, + 1459.0, + 1305.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1451.0, + 842.0, + 1451.0, + 842.0, + 1492.0, + 292.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1451.0, + 1405.0, + 1451.0, + 1405.0, + 1492.0, + 873.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1485.0, + 1350.0, + 1485.0, + 1350.0, + 1518.0, + 295.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 920.0, + 1528.0, + 920.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1528.0, + 1175.0, + 1528.0, + 1175.0, + 1566.0, + 1097.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1566.0, + 1225.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 450.0, + 1561.0, + 450.0, + 1597.0, + 294.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1561.0, + 1178.0, + 1561.0, + 1178.0, + 1597.0, + 532.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1561.0, + 1212.0, + 1561.0, + 1212.0, + 1597.0, + 1200.0, + 1597.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 351, + 194, + 1348, + 194, + 1348, + 1183, + 351, + 1183 + ], + "score": 0.978 + }, + { + "category_id": 4, + "poly": [ + 295, + 1207, + 1407, + 1207, + 1407, + 1405, + 295, + 1405 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 302, + 1689, + 1396, + 1689, + 1396, + 1947, + 302, + 1947 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 295, + 1972, + 1407, + 1972, + 1407, + 2058, + 295, + 2058 + ], + "score": 0.956 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 812, + 76, + 812, + 104, + 298, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.849 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2114, + 834, + 2114 + ], + "score": 0.119 + }, + { + "category_id": 13, + "poly": [ + 738, + 1262, + 812, + 1262, + 812, + 1295, + 738, + 1295 + ], + "score": 0.92, + "latex": "( \\dot { i } , j ) ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1301, + 1210, + 1398, + 1210, + 1398, + 1238, + 1301, + 1238 + ], + "score": 0.9, + "latex": "\\gamma = 0 . 9 9" + }, + { + "category_id": 13, + "poly": [ + 362, + 1004, + 401, + 1004, + 401, + 1026, + 362, + 1026 + ], + "score": 0.87, + "latex": "y _ { 1 0 }" + }, + { + "category_id": 13, + "poly": [ + 430, + 1295, + 456, + 1295, + 456, + 1322, + 430, + 1322 + ], + "score": 0.86, + "latex": "y _ { j }" + }, + { + "category_id": 13, + "poly": [ + 358, + 1296, + 385, + 1296, + 385, + 1319, + 358, + 1319 + ], + "score": 0.85, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1062, + 401, + 1062, + 401, + 1084, + 373, + 1084 + ], + "score": 0.84, + "latex": "y _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 399, + 1088, + 430, + 1088, + 430, + 1108, + 399, + 1108 + ], + "score": 0.84, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1242, + 323, + 1242, + 323, + 1264, + 298, + 1264 + ], + "score": 0.83, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 496, + 1146, + 525, + 1146, + 525, + 1172, + 496, + 1172 + ], + "score": 0.82, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 369, + 1241, + 393, + 1241, + 393, + 1265, + 369, + 1265 + ], + "score": 0.82, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 362, + 771, + 402, + 771, + 402, + 793, + 362, + 793 + ], + "score": 0.78, + "latex": "y _ { 5 0 }" + }, + { + "category_id": 13, + "poly": [ + 362, + 507, + 402, + 507, + 402, + 529, + 362, + 529 + ], + "score": 0.78, + "latex": "y _ { 1 0 }" + }, + { + "category_id": 13, + "poly": [ + 362, + 946, + 402, + 946, + 402, + 968, + 362, + 968 + ], + "score": 0.74, + "latex": "y _ { 2 0 }" + }, + { + "category_id": 13, + "poly": [ + 451, + 1088, + 494, + 1088, + 494, + 1108, + 451, + 1108 + ], + "score": 0.73, + "latex": "x _ { 1 0 }" + }, + { + "category_id": 13, + "poly": [ + 374, + 567, + 403, + 567, + 403, + 589, + 374, + 589 + ], + "score": 0.72, + "latex": "y _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 401, + 592, + 432, + 592, + 432, + 613, + 401, + 613 + ], + "score": 0.69, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 362, + 829, + 401, + 829, + 401, + 851, + 362, + 851 + ], + "score": 0.66, + "latex": "y _ { 4 0 }" + }, + { + "category_id": 13, + "poly": [ + 980, + 1348, + 1037, + 1348, + 1037, + 1376, + 980, + 1376 + ], + "score": 0.64, + "latex": "( T V )" + }, + { + "category_id": 13, + "poly": [ + 362, + 887, + 401, + 887, + 401, + 910, + 362, + 910 + ], + "score": 0.62, + "latex": "y _ { 3 0 }" + }, + { + "category_id": 13, + "poly": [ + 509, + 1087, + 553, + 1087, + 553, + 1109, + 509, + 1109 + ], + "score": 0.49, + "latex": "x _ { 2 0 }" + }, + { + "category_id": 13, + "poly": [ + 362, + 447, + 402, + 447, + 402, + 469, + 362, + 469 + ], + "score": 0.36, + "latex": "y _ { 2 0 }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1085, + 612, + 1085, + 612, + 1110, + 568, + 1110 + ], + "score": 0.26, + "latex": "x _ { 3 0 }" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 199.0, + 693.0, + 199.0, + 693.0, + 243.0, + 469.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 196.0, + 1128.0, + 196.0, + 1128.0, + 247.0, + 1031.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 253.0, + 414.0, + 253.0, + 414.0, + 302.0, + 352.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 234.0, + 847.0, + 234.0, + 847.0, + 280.0, + 786.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 254.0, + 917.0, + 254.0, + 917.0, + 302.0, + 853.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 236.0, + 1349.0, + 236.0, + 1349.0, + 280.0, + 1278.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 286.0, + 847.0, + 286.0, + 847.0, + 333.0, + 786.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 288.0, + 1349.0, + 288.0, + 1349.0, + 332.0, + 1279.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 311.0, + 416.0, + 311.0, + 416.0, + 363.0, + 350.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 314.0, + 917.0, + 314.0, + 917.0, + 361.0, + 854.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 336.0, + 846.0, + 336.0, + 846.0, + 379.0, + 787.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 339.0, + 1348.0, + 339.0, + 1348.0, + 380.0, + 1280.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 372.0, + 419.0, + 372.0, + 419.0, + 424.0, + 351.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 387.0, + 846.0, + 387.0, + 846.0, + 429.0, + 782.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 374.0, + 917.0, + 374.0, + 917.0, + 421.0, + 854.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 389.0, + 1348.0, + 389.0, + 1348.0, + 429.0, + 1279.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 431.0, + 361.0, + 431.0, + 361.0, + 483.0, + 351.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 431.0, + 416.0, + 431.0, + 416.0, + 483.0, + 403.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 440.0, + 834.0, + 440.0, + 834.0, + 481.0, + 786.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 434.0, + 916.0, + 434.0, + 916.0, + 478.0, + 855.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 439.0, + 1348.0, + 439.0, + 1348.0, + 480.0, + 1280.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 492.0, + 361.0, + 492.0, + 361.0, + 539.0, + 352.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 492.0, + 416.0, + 492.0, + 416.0, + 539.0, + 403.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 488.0, + 835.0, + 488.0, + 835.0, + 534.0, + 788.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 491.0, + 917.0, + 491.0, + 917.0, + 538.0, + 853.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 490.0, + 1348.0, + 490.0, + 1348.0, + 531.0, + 1281.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 552.0, + 373.0, + 552.0, + 373.0, + 598.0, + 364.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 552.0, + 414.0, + 552.0, + 414.0, + 598.0, + 404.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 540.0, + 835.0, + 540.0, + 835.0, + 586.0, + 787.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 551.0, + 916.0, + 551.0, + 916.0, + 595.0, + 865.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 538.0, + 1349.0, + 538.0, + 1349.0, + 582.0, + 1280.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 578.0, + 400.0, + 578.0, + 400.0, + 621.0, + 395.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 578.0, + 742.0, + 578.0, + 742.0, + 621.0, + 433.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 573.0, + 1244.0, + 573.0, + 1244.0, + 619.0, + 895.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 589.0, + 1348.0, + 589.0, + 1348.0, + 628.0, + 1283.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 646.0, + 724.0, + 646.0, + 724.0, + 683.0, + 473.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 643.0, + 1245.0, + 643.0, + 1245.0, + 680.0, + 956.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 701.0, + 724.0, + 701.0, + 724.0, + 749.0, + 427.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 710.0, + 847.0, + 710.0, + 847.0, + 753.0, + 772.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 695.0, + 1230.0, + 695.0, + 1230.0, + 742.0, + 943.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 686.0, + 1345.0, + 686.0, + 1345.0, + 722.0, + 1270.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 757.0, + 361.0, + 757.0, + 361.0, + 804.0, + 351.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 757.0, + 415.0, + 757.0, + 415.0, + 804.0, + 403.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 753.0, + 917.0, + 753.0, + 917.0, + 801.0, + 854.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 766.0, + 1349.0, + 766.0, + 1349.0, + 806.0, + 1287.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 811.0, + 361.0, + 811.0, + 361.0, + 867.0, + 351.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 811.0, + 418.0, + 811.0, + 418.0, + 867.0, + 402.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 802.0, + 849.0, + 802.0, + 849.0, + 850.0, + 771.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 810.0, + 918.0, + 810.0, + 918.0, + 863.0, + 855.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 850.0, + 1348.0, + 850.0, + 1348.0, + 887.0, + 1291.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 873.0, + 361.0, + 873.0, + 361.0, + 920.0, + 352.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 873.0, + 415.0, + 873.0, + 415.0, + 920.0, + 402.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 872.0, + 916.0, + 872.0, + 916.0, + 920.0, + 853.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 899.0, + 847.0, + 899.0, + 847.0, + 939.0, + 772.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 932.0, + 361.0, + 932.0, + 361.0, + 980.0, + 351.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 932.0, + 414.0, + 932.0, + 414.0, + 980.0, + 403.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 935.0, + 917.0, + 935.0, + 917.0, + 982.0, + 854.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 930.0, + 1349.0, + 930.0, + 1349.0, + 971.0, + 1289.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 990.0, + 361.0, + 990.0, + 361.0, + 1038.0, + 352.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 990.0, + 414.0, + 990.0, + 414.0, + 1038.0, + 402.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 994.0, + 846.0, + 994.0, + 846.0, + 1033.0, + 775.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 995.0, + 916.0, + 995.0, + 916.0, + 1041.0, + 854.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1012.0, + 1349.0, + 1012.0, + 1349.0, + 1054.0, + 1292.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1047.0, + 372.0, + 1047.0, + 372.0, + 1116.0, + 364.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1047.0, + 450.0, + 1047.0, + 450.0, + 1116.0, + 431.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1047.0, + 508.0, + 1047.0, + 508.0, + 1116.0, + 495.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1047.0, + 567.0, + 1047.0, + 567.0, + 1116.0, + 554.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1047.0, + 738.0, + 1047.0, + 738.0, + 1116.0, + 613.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1092.0, + 806.0, + 1092.0, + 806.0, + 1120.0, + 774.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1054.0, + 944.0, + 1054.0, + 944.0, + 1121.0, + 867.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1079.0, + 1251.0, + 1079.0, + 1251.0, + 1125.0, + 949.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1088.0, + 1348.0, + 1088.0, + 1348.0, + 1134.0, + 1269.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1142.0, + 495.0, + 1142.0, + 495.0, + 1177.0, + 457.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1142.0, + 740.0, + 1142.0, + 740.0, + 1177.0, + 526.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1146.0, + 1229.0, + 1146.0, + 1229.0, + 1189.0, + 972.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 415.0, + 486.0, + 415.0, + 486.0, + 433.0, + 405.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 414.0, + 991.0, + 414.0, + 991.0, + 430.0, + 902.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 370.0, + 742.0, + 370.0, + 742.0, + 384.0, + 480.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.25, + 591.5, + 816.25, + 591.5, + 816.25, + 628.0, + 784.25, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 912.5, + 997.0, + 912.5, + 997.0, + 927.5, + 909.0, + 927.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1208.0, + 1300.0, + 1208.0, + 1300.0, + 1241.0, + 296.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1208.0, + 1408.0, + 1208.0, + 1408.0, + 1241.0, + 1399.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1234.0, + 297.0, + 1234.0, + 297.0, + 1270.0, + 293.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1234.0, + 368.0, + 1234.0, + 368.0, + 1270.0, + 324.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1234.0, + 1405.0, + 1234.0, + 1405.0, + 1270.0, + 394.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1257.0, + 737.0, + 1257.0, + 737.0, + 1301.0, + 290.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1257.0, + 1409.0, + 1257.0, + 1409.0, + 1301.0, + 813.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1291.0, + 357.0, + 1291.0, + 357.0, + 1325.0, + 294.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1291.0, + 429.0, + 1291.0, + 429.0, + 1325.0, + 386.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1291.0, + 1405.0, + 1291.0, + 1405.0, + 1325.0, + 457.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1319.0, + 1405.0, + 1319.0, + 1405.0, + 1353.0, + 296.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1349.0, + 979.0, + 1349.0, + 979.0, + 1378.0, + 296.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1378.0, + 1038.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1374.0, + 1292.0, + 1374.0, + 1292.0, + 1407.0, + 296.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1687.0, + 337.0, + 1687.0, + 337.0, + 1901.0, + 304.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1689.0, + 398.0, + 1689.0, + 398.0, + 1723.0, + 336.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1690.0, + 1066.0, + 1690.0, + 1066.0, + 1899.0, + 1029.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1692.0, + 699.0, + 1692.0, + 699.0, + 1898.0, + 666.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1692.0, + 760.0, + 1692.0, + 760.0, + 1721.0, + 699.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1692.0, + 1124.0, + 1692.0, + 1124.0, + 1721.0, + 1063.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1696.0, + 966.0, + 1696.0, + 966.0, + 1718.0, + 813.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1729.0, + 394.0, + 1729.0, + 394.0, + 1756.0, + 348.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1730.0, + 756.0, + 1730.0, + 756.0, + 1756.0, + 710.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1720.0, + 906.0, + 1720.0, + 906.0, + 1745.0, + 814.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1730.0, + 1120.0, + 1730.0, + 1120.0, + 1756.0, + 1073.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1763.0, + 394.0, + 1763.0, + 394.0, + 1791.0, + 348.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1765.0, + 756.0, + 1765.0, + 756.0, + 1792.0, + 710.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1765.0, + 1121.0, + 1765.0, + 1121.0, + 1793.0, + 1073.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1799.0, + 394.0, + 1799.0, + 394.0, + 1825.0, + 348.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1799.0, + 756.0, + 1799.0, + 756.0, + 1827.0, + 710.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1801.0, + 1120.0, + 1801.0, + 1120.0, + 1827.0, + 1073.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1834.0, + 394.0, + 1834.0, + 394.0, + 1860.0, + 348.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1834.0, + 634.0, + 1834.0, + 634.0, + 1858.0, + 481.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1835.0, + 756.0, + 1835.0, + 756.0, + 1862.0, + 710.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1834.0, + 1119.0, + 1834.0, + 1119.0, + 1862.0, + 1073.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1833.0, + 1361.0, + 1833.0, + 1361.0, + 1860.0, + 1207.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1857.0, + 573.0, + 1857.0, + 573.0, + 1881.0, + 482.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1854.0, + 1300.0, + 1854.0, + 1300.0, + 1885.0, + 1207.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1869.0, + 395.0, + 1869.0, + 395.0, + 1896.0, + 358.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1869.0, + 757.0, + 1869.0, + 757.0, + 1898.0, + 720.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1869.0, + 1121.0, + 1869.0, + 1121.0, + 1898.0, + 1083.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1890.0, + 410.0, + 1890.0, + 410.0, + 1911.0, + 391.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1889.0, + 483.0, + 1889.0, + 483.0, + 1915.0, + 440.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1888.0, + 667.0, + 1888.0, + 667.0, + 1914.0, + 492.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1890.0, + 772.0, + 1890.0, + 772.0, + 1912.0, + 753.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1891.0, + 847.0, + 1891.0, + 847.0, + 1914.0, + 804.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1888.0, + 1032.0, + 1888.0, + 1032.0, + 1915.0, + 854.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1890.0, + 1136.0, + 1890.0, + 1136.0, + 1912.0, + 1116.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1891.0, + 1207.0, + 1891.0, + 1207.0, + 1914.0, + 1166.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1888.0, + 1394.0, + 1888.0, + 1394.0, + 1915.0, + 1221.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1912.0, + 619.0, + 1912.0, + 619.0, + 1947.0, + 424.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1912.0, + 979.0, + 1912.0, + 979.0, + 1947.0, + 790.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1912.0, + 1342.0, + 1912.0, + 1342.0, + 1947.0, + 1153.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2002.0, + 294.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2030.0, + 296.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2029.0, + 491.0, + 2029.0, + 491.0, + 2059.0, + 296.0, + 2059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1436, + 1405, + 1436, + 1405, + 1560, + 297, + 1560 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 332, + 729, + 1391, + 729, + 1391, + 1141, + 332, + 1141 + ], + "score": 0.974 + }, + { + "category_id": 5, + "poly": [ + 326, + 1686, + 713, + 1686, + 713, + 2028, + 326, + 2028 + ], + "score": 0.972, + "html": "
HyperparameterValue
Learning rate decay Training epochs0.999
Optimizer2000 Adam
Batch size (Imitation)256
Num training tasks I-scale Parameter (β)18
Embedding size (k)0.01
Batch Size (LCME)64
57
|T57
" + }, + { + "category_id": 5, + "poly": [ + 812, + 453, + 1369, + 453, + 1369, + 644, + 812, + 644 + ], + "score": 0.971, + "html": "
MethodRed (%)Green (%)
RandConv6.2 (0.4)99.6 (0.2)
Dropout and l2 reg.19.5(0.2)100.0 (0.0)
RandConv +PSEs29.8 (1.3)99.6 (0.2)
PSEs37.9 (1.9)100.0 (0.0)
" + }, + { + "category_id": 3, + "poly": [ + 804, + 1651, + 1389, + 1651, + 1389, + 1901, + 804, + 1901 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 318, + 328, + 716, + 328, + 716, + 523, + 318, + 523 + ], + "score": 0.964 + }, + { + "category_id": 4, + "poly": [ + 296, + 548, + 746, + 548, + 746, + 633, + 296, + 633 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 297, + 1191, + 1405, + 1191, + 1405, + 1305, + 297, + 1305 + ], + "score": 0.948 + }, + { + "category_id": 4, + "poly": [ + 791, + 1936, + 1403, + 1936, + 1403, + 1993, + 791, + 1993 + ], + "score": 0.942 + }, + { + "category_id": 6, + "poly": [ + 783, + 316, + 1397, + 316, + 1397, + 430, + 783, + 430 + ], + "score": 0.937 + }, + { + "category_id": 6, + "poly": [ + 295, + 1605, + 743, + 1605, + 743, + 1662, + 295, + 1662 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 299, + 1374, + 614, + 1374, + 614, + 1405, + 299, + 1405 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 300, + 231, + 733, + 231, + 733, + 261, + 300, + 261 + ], + "score": 0.784 + }, + { + "category_id": 6, + "poly": [ + 297, + 1604, + 742, + 1604, + 742, + 1663, + 297, + 1663 + ], + "score": 0.22 + }, + { + "category_id": 13, + "poly": [ + 1017, + 318, + 1057, + 318, + 1057, + 345, + 1017, + 345 + ], + "score": 0.72, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 345, + 1896, + 366, + 1896, + 366, + 1923, + 345, + 1923 + ], + "score": 0.42, + "latex": "\\Gamma" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 732.0, + 451.0, + 732.0, + 451.0, + 761.0, + 383.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 732.0, + 675.0, + 732.0, + 675.0, + 763.0, + 518.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 734.0, + 854.0, + 734.0, + 854.0, + 761.0, + 696.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 732.0, + 1388.0, + 732.0, + 1388.0, + 763.0, + 857.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 941.0, + 451.0, + 941.0, + 451.0, + 974.0, + 383.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 945.0, + 676.0, + 945.0, + 676.0, + 972.0, + 517.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 941.0, + 853.0, + 941.0, + 853.0, + 972.0, + 695.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 944.0, + 1225.0, + 944.0, + 1225.0, + 972.0, + 872.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 945.0, + 1387.0, + 945.0, + 1387.0, + 972.0, + 1230.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1658.0, + 847.0, + 1658.0, + 847.0, + 1670.0, + 835.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1658.0, + 916.0, + 1658.0, + 916.0, + 1671.0, + 902.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1657.0, + 942.0, + 1657.0, + 942.0, + 1673.0, + 921.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1656.0, + 962.0, + 1656.0, + 962.0, + 1673.0, + 943.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1656.0, + 1072.0, + 1656.0, + 1072.0, + 1674.0, + 1011.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1653.0, + 1187.0, + 1653.0, + 1187.0, + 1677.0, + 1120.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1657.0, + 1252.0, + 1657.0, + 1252.0, + 1672.0, + 1235.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1655.0, + 1277.0, + 1655.0, + 1277.0, + 1674.0, + 1253.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1656.0, + 1295.0, + 1656.0, + 1295.0, + 1673.0, + 1278.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1660.0, + 1357.0, + 1660.0, + 1357.0, + 1665.0, + 1352.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1659.0, + 1380.0, + 1659.0, + 1380.0, + 1670.0, + 1372.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1679.0, + 826.0, + 1679.0, + 826.0, + 1690.0, + 813.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1675.0, + 942.0, + 1675.0, + 942.0, + 1696.0, + 920.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1676.0, + 1053.0, + 1676.0, + 1053.0, + 1696.0, + 1031.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1675.0, + 1164.0, + 1675.0, + 1164.0, + 1697.0, + 1142.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1675.0, + 1275.0, + 1675.0, + 1275.0, + 1696.0, + 1253.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1682.0, + 1379.0, + 1682.0, + 1379.0, + 1689.0, + 1372.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1745.0, + 826.0, + 1745.0, + 826.0, + 1757.0, + 813.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1744.0, + 939.0, + 1744.0, + 939.0, + 1760.0, + 921.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1743.0, + 1053.0, + 1743.0, + 1053.0, + 1762.0, + 1033.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1744.0, + 1162.0, + 1744.0, + 1162.0, + 1761.0, + 1144.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1744.0, + 1272.0, + 1744.0, + 1272.0, + 1760.0, + 1256.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1746.0, + 1380.0, + 1746.0, + 1380.0, + 1758.0, + 1372.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1767.0, + 847.0, + 1767.0, + 847.0, + 1780.0, + 832.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1766.0, + 917.0, + 1766.0, + 917.0, + 1783.0, + 901.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1770.0, + 935.0, + 1770.0, + 935.0, + 1780.0, + 927.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1765.0, + 1072.0, + 1765.0, + 1072.0, + 1785.0, + 1011.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1765.0, + 1183.0, + 1765.0, + 1183.0, + 1785.0, + 1124.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1768.0, + 1250.0, + 1768.0, + 1250.0, + 1780.0, + 1236.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1767.0, + 1272.0, + 1767.0, + 1272.0, + 1783.0, + 1256.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1767.0, + 1293.0, + 1767.0, + 1293.0, + 1781.0, + 1280.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1767.0, + 1361.0, + 1767.0, + 1361.0, + 1779.0, + 1348.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1790.0, + 826.0, + 1790.0, + 826.0, + 1803.0, + 812.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1788.0, + 939.0, + 1788.0, + 939.0, + 1804.0, + 921.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1787.0, + 1053.0, + 1787.0, + 1053.0, + 1809.0, + 1031.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1787.0, + 1163.0, + 1787.0, + 1163.0, + 1808.0, + 1142.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1788.0, + 1272.0, + 1788.0, + 1272.0, + 1804.0, + 1256.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1792.0, + 1380.0, + 1792.0, + 1380.0, + 1802.0, + 1372.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1855.0, + 939.0, + 1855.0, + 939.0, + 1871.0, + 921.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1854.0, + 1053.0, + 1854.0, + 1053.0, + 1874.0, + 1032.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1855.0, + 1162.0, + 1855.0, + 1162.0, + 1872.0, + 1144.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1855.0, + 1273.0, + 1855.0, + 1273.0, + 1871.0, + 1256.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1857.0, + 1381.0, + 1857.0, + 1381.0, + 1870.0, + 1370.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1877.0, + 917.0, + 1877.0, + 917.0, + 1894.0, + 901.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1880.0, + 936.0, + 1880.0, + 936.0, + 1893.0, + 925.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1877.0, + 963.0, + 1877.0, + 963.0, + 1896.0, + 944.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1876.0, + 1074.0, + 1876.0, + 1074.0, + 1895.0, + 1011.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1873.0, + 1189.0, + 1873.0, + 1189.0, + 1899.0, + 1119.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1876.0, + 1274.0, + 1876.0, + 1274.0, + 1895.0, + 1234.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1877.0, + 1295.0, + 1877.0, + 1295.0, + 1895.0, + 1278.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1880.0, + 1359.0, + 1880.0, + 1359.0, + 1889.0, + 1351.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1882.0, + 1379.0, + 1882.0, + 1379.0, + 1890.0, + 1372.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 360.0, + 331.0, + 360.0, + 331.0, + 369.0, + 321.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 359.0, + 539.0, + 359.0, + 539.0, + 370.0, + 527.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 390.0, + 330.0, + 390.0, + 330.0, + 399.0, + 321.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 389.0, + 539.0, + 389.0, + 539.0, + 400.0, + 527.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 419.0, + 330.0, + 419.0, + 330.0, + 428.0, + 322.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 416.0, + 434.0, + 416.0, + 434.0, + 443.0, + 417.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 417.0, + 539.0, + 417.0, + 539.0, + 429.0, + 527.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 449.0, + 330.0, + 449.0, + 330.0, + 457.0, + 321.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 448.0, + 539.0, + 448.0, + 539.0, + 456.0, + 528.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 479.0, + 330.0, + 479.0, + 330.0, + 487.0, + 321.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 477.0, + 537.0, + 477.0, + 537.0, + 485.0, + 528.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 514.0, + 369.0, + 514.0, + 369.0, + 521.0, + 360.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 514.0, + 398.0, + 514.0, + 398.0, + 521.0, + 390.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 513.0, + 428.0, + 513.0, + 428.0, + 521.0, + 420.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 512.0, + 578.0, + 512.0, + 578.0, + 520.0, + 566.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 511.0, + 608.0, + 511.0, + 608.0, + 521.0, + 593.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 511.0, + 638.0, + 511.0, + 638.0, + 521.0, + 621.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 512.0, + 666.0, + 512.0, + 666.0, + 519.0, + 652.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 512.0, + 694.0, + 512.0, + 694.0, + 519.0, + 680.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 548.0, + 746.0, + 548.0, + 746.0, + 578.0, + 297.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 576.0, + 745.0, + 576.0, + 745.0, + 606.0, + 296.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 604.0, + 675.0, + 604.0, + 675.0, + 634.0, + 298.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1192.0, + 1405.0, + 1192.0, + 1405.0, + 1225.0, + 295.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1220.0, + 1403.0, + 1220.0, + 1403.0, + 1253.0, + 295.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1243.0, + 1405.0, + 1243.0, + 1405.0, + 1284.0, + 292.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1275.0, + 1403.0, + 1275.0, + 1403.0, + 1308.0, + 295.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1934.0, + 1405.0, + 1934.0, + 1405.0, + 1967.0, + 790.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1964.0, + 1090.0, + 1964.0, + 1090.0, + 1995.0, + 791.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 316.0, + 1016.0, + 316.0, + 1016.0, + 345.0, + 783.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 316.0, + 1397.0, + 316.0, + 1397.0, + 345.0, + 1058.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 342.0, + 1398.0, + 342.0, + 1398.0, + 374.0, + 782.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 372.0, + 1397.0, + 372.0, + 1397.0, + 401.0, + 783.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 399.0, + 1169.0, + 399.0, + 1169.0, + 429.0, + 784.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1604.0, + 744.0, + 1604.0, + 744.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1632.0, + 742.0, + 1632.0, + 742.0, + 1664.0, + 296.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1374.0, + 616.0, + 1374.0, + 616.0, + 1407.0, + 297.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 736.0, + 229.0, + 736.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1604.0, + 743.0, + 1604.0, + 743.0, + 1635.0, + 295.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1631.0, + 743.0, + 1631.0, + 743.0, + 1664.0, + 295.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 1403.0, + 1436.0, + 1403.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1464.0, + 1406.0, + 1464.0, + 1406.0, + 1503.0, + 291.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1500.0, + 1403.0, + 1500.0, + 1403.0, + 1533.0, + 295.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1527.0, + 988.0, + 1527.0, + 988.0, + 1565.0, + 292.0, + 1565.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 329, + 1066, + 1362, + 1066, + 1362, + 1276, + 329, + 1276 + ], + "score": 0.982, + "html": "
HyperparameterPSMπ*-bisimulation
CMEsl2-embeddingsCMEsl2-embeddings
Learning Rate4×10-35×10-44.7×10-41×10-4
Contrastive Temperature (1/入)1.015×10-11
Auxiliary loss coefficient (α)5.01×10-11×10-11×10-6
" + }, + { + "category_id": 5, + "poly": [ + 296, + 333, + 1415, + 333, + 1415, + 561, + 296, + 561 + ], + "score": 0.979, + "html": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-33.2×10-37×10-32.6×10-3
l2-reg. coefficient4.3×10-41×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)11.05×10-1
Auxiliary loss coefficient (α)11×10115.0
" + }, + { + "category_id": 5, + "poly": [ + 307, + 701, + 1387, + 701, + 1387, + 929, + 307, + 929 + ], + "score": 0.978, + "html": "
HyperparameterDropout and l2-reg.PSEsRandConv RandConv +PSEs
Learning Rate4×10-36×10-35×10-32.6×10-3
l2-reg. coefficient4.3×10-47×10-511
Dropout coefficient3×10-1111
Contrastive Temperature (1/入)15×10-115×10-1
Auxiliary loss coefficient (α)15.015.0
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1319, + 1407, + 1319, + 1407, + 1567, + 297, + 1567 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1679, + 1405, + 1679, + 1405, + 1774, + 298, + 1774 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1910, + 1406, + 1910, + 1406, + 2036, + 297, + 2036 + ], + "score": 0.975 + }, + { + "category_id": 6, + "poly": [ + 296, + 224, + 1408, + 224, + 1408, + 310, + 296, + 310 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 501, + 1786, + 1196, + 1786, + 1196, + 1898, + 501, + 1898 + ], + "score": 0.954 + }, + { + "category_id": 6, + "poly": [ + 298, + 985, + 1400, + 985, + 1400, + 1043, + 298, + 1043 + ], + "score": 0.938 + }, + { + "category_id": 6, + "poly": [ + 299, + 619, + 1401, + 619, + 1401, + 680, + 299, + 680 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 300, + 1609, + 768, + 1609, + 768, + 1647, + 300, + 1647 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.906 + }, + { + "category_id": 9, + "poly": [ + 1339, + 1827, + 1400, + 1827, + 1400, + 1859, + 1339, + 1859 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2113, + 835, + 2113 + ], + "score": 0.87 + }, + { + "category_id": 14, + "poly": [ + 499, + 1785, + 1194, + 1785, + 1194, + 1899, + 499, + 1899 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathrm { m i n i m i z e } \\quad E _ { s _ { 0 } \\sim \\mathcal { D } } \\left[ \\frac { 1 } { 2 } \\sum _ { t = 0 } ^ { \\infty } s _ { t } ^ { T } Q s _ { t } + a _ { t } ^ { T } R a _ { t } \\right] , } \\\\ & { \\mathrm { s u b j e c t \\ t o } \\quad s _ { t + 1 } = A s _ { t } + B a _ { t } , o _ { t } = \\left[ \\begin{array} { l } { 0 . 1 W _ { c } } \\\\ { W _ { d } } \\end{array} \\right] s _ { t } , a _ { t } = K o _ { t } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1261, + 1912, + 1372, + 1912, + 1372, + 1942, + 1261, + 1942 + ], + "score": 0.91, + "latex": "a _ { t } \\in \\mathbb { R } ^ { n _ { a } }" + }, + { + "category_id": 13, + "poly": [ + 752, + 1912, + 860, + 1912, + 860, + 1942, + 752, + 1942 + ], + "score": 0.91, + "latex": "s _ { t } \\in \\mathbb { R } ^ { n _ { s } }" + }, + { + "category_id": 13, + "poly": [ + 913, + 1973, + 954, + 1973, + 954, + 2003, + 913, + 2003 + ], + "score": 0.89, + "latex": "W _ { d }" + }, + { + "category_id": 13, + "poly": [ + 654, + 1505, + 722, + 1505, + 722, + 1535, + 654, + 1535 + ], + "score": 0.87, + "latex": "( - 4 \\% )" + }, + { + "category_id": 13, + "poly": [ + 296, + 423, + 324, + 423, + 324, + 454, + 296, + 454 + ], + "score": 0.83, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 560, + 1944, + 589, + 1944, + 589, + 1970, + 560, + 1970 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 821, + 1974, + 862, + 1974, + 862, + 2003, + 821, + 2003 + ], + "score": 0.81, + "latex": "W _ { c }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1913, + 399, + 1913, + 399, + 1939, + 373, + 1939 + ], + "score": 0.8, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 309, + 792, + 336, + 792, + 336, + 823, + 309, + 823 + ], + "score": 0.8, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1371, + 1948, + 1396, + 1948, + 1396, + 1973, + 1371, + 1973 + ], + "score": 0.78, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 783, + 1977, + 808, + 1977, + 808, + 2003, + 783, + 2003 + ], + "score": 0.73, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 797, + 714, + 824, + 714, + 824, + 745, + 797, + 745 + ], + "score": 0.41, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 223.0, + 1402.0, + 223.0, + 1402.0, + 258.0, + 293.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 252.0, + 1405.0, + 252.0, + 1405.0, + 284.0, + 295.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 280.0, + 827.0, + 280.0, + 827.0, + 311.0, + 294.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 986.0, + 1402.0, + 986.0, + 1402.0, + 1016.0, + 296.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1012.0, + 1323.0, + 1012.0, + 1323.0, + 1045.0, + 293.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 621.0, + 1403.0, + 621.0, + 1403.0, + 652.0, + 295.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 647.0, + 1323.0, + 647.0, + 1323.0, + 681.0, + 293.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1609.0, + 773.0, + 1609.0, + 773.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1322.0, + 1403.0, + 1322.0, + 1403.0, + 1352.0, + 296.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1351.0, + 1405.0, + 1351.0, + 1405.0, + 1385.0, + 294.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1383.0, + 1405.0, + 1383.0, + 1405.0, + 1416.0, + 295.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1413.0, + 1405.0, + 1413.0, + 1405.0, + 1447.0, + 295.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1442.0, + 1409.0, + 1442.0, + 1409.0, + 1476.0, + 295.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1405.0, + 1473.0, + 1405.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1501.0, + 653.0, + 1501.0, + 653.0, + 1537.0, + 292.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1501.0, + 1409.0, + 1501.0, + 1409.0, + 1537.0, + 723.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1532.0, + 875.0, + 1532.0, + 875.0, + 1567.0, + 294.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1714.0, + 297.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 1408.0, + 1712.0, + 1408.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1739.0, + 1372.0, + 1739.0, + 1372.0, + 1778.0, + 295.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 372.0, + 1909.0, + 372.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1909.0, + 751.0, + 1909.0, + 751.0, + 1946.0, + 400.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1909.0, + 1260.0, + 1909.0, + 1260.0, + 1946.0, + 861.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1946.0, + 1373.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 559.0, + 1939.0, + 559.0, + 1980.0, + 292.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1939.0, + 1370.0, + 1939.0, + 1370.0, + 1980.0, + 590.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1939.0, + 1409.0, + 1939.0, + 1409.0, + 1980.0, + 1397.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 782.0, + 1972.0, + 782.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1972.0, + 820.0, + 1972.0, + 820.0, + 2007.0, + 809.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1972.0, + 912.0, + 1972.0, + 912.0, + 2007.0, + 863.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 955.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 747, + 1405, + 747, + 1405, + 1082, + 296, + 1082 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 545, + 1334, + 1156, + 1334, + 1156, + 1647, + 545, + 1647 + ], + "score": 0.979, + "html": "
ParameterSeting
AOrthogonal matrix, scaled 0.8
BI20×20
nx20
na20
QI20×20
RI20×20
KiAiOrthogonal Initialization, scaled 0.001
WdRandom semi-orthogonal matrix
" + }, + { + "category_id": 5, + "poly": [ + 382, + 333, + 1310, + 333, + 1310, + 571, + 382, + 571 + ], + "score": 0.979, + "html": "
MethodNumber of Distractors (nd)
500100010000
Overparametrization (Song et al., 2019)25.8 (1.5)24.9 (1.1)24.9 (0.4)
IPO (Sonar et al., 202O) (IRM + Policy opt.)32.6 (5.0)27.3 (2.8)24.8 (0.4)
Weight Sparsity (ll-reg.)28.2 (0.0)28.2 (0.0)28.2 (0.0)
PSM (State aggregation)0.03 (0.0)0.03 (0.0)0.02 (0.0)
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1095, + 1404, + 1095, + 1404, + 1250, + 298, + 1250 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 610, + 1405, + 610, + 1405, + 734, + 297, + 734 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1679, + 1403, + 1679, + 1403, + 1773, + 298, + 1773 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 295, + 1971, + 1402, + 1971, + 1402, + 2034, + 295, + 2034 + ], + "score": 0.94 + }, + { + "category_id": 6, + "poly": [ + 297, + 223, + 1405, + 223, + 1405, + 309, + 297, + 309 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 298, + 1811, + 893, + 1811, + 893, + 1844, + 298, + 1844 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.862 + }, + { + "category_id": 6, + "poly": [ + 571, + 1279, + 1123, + 1279, + 1123, + 1312, + 571, + 1312 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 298, + 1870, + 1399, + 1870, + 1399, + 1933, + 298, + 1933 + ], + "score": 0.675 + }, + { + "category_id": 0, + "poly": [ + 298, + 1870, + 1399, + 1870, + 1399, + 1933, + 298, + 1933 + ], + "score": 0.371 + }, + { + "category_id": 6, + "poly": [ + 572, + 1280, + 1125, + 1280, + 1125, + 1312, + 572, + 1312 + ], + "score": 0.204 + }, + { + "category_id": 13, + "poly": [ + 1091, + 750, + 1234, + 750, + 1234, + 780, + 1091, + 780 + ], + "score": 0.93, + "latex": "W _ { c } s _ { t } \\in \\mathbb { R } ^ { n _ { s } }" + }, + { + "category_id": 13, + "poly": [ + 1177, + 1096, + 1251, + 1096, + 1251, + 1129, + 1177, + 1129 + ], + "score": 0.92, + "latex": "K _ { 1 } ( o )" + }, + { + "category_id": 13, + "poly": [ + 863, + 780, + 1007, + 780, + 1007, + 811, + 863, + 811 + ], + "score": 0.92, + "latex": "W _ { d } s _ { t } \\in \\mathbb { R } ^ { n _ { d } }" + }, + { + "category_id": 13, + "poly": [ + 958, + 1097, + 1094, + 1097, + 1094, + 1127, + 958, + 1127 + ], + "score": 0.91, + "latex": "K = K _ { 1 } K _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 510, + 943, + 976, + 943, + 976, + 1024, + 510, + 1024 + ], + "score": 0.91, + "latex": "K _ { \\star } { } ^ { \\prime } = \\left[ \\begin{array} { c c } { 1 0 \\alpha W _ { c } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\\\ { ( 1 - \\alpha ) W _ { d } P _ { \\star } ^ { \\mathrm { ~ \\top ~ } } } \\end{array} \\right] ^ { T } \\quad \\forall \\alpha \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 710, + 611, + 906, + 611, + 906, + 644, + 710, + 644 + ], + "score": 0.91, + "latex": "( A , B , Q , R , W _ { c } )" + }, + { + "category_id": 13, + "poly": [ + 587, + 842, + 826, + 842, + 826, + 919, + 587, + 919 + ], + "score": 0.9, + "latex": "K _ { \\star } = \\bigg [ 1 0 W _ { c } P _ { \\star } ^ { ~ T } \\bigg ] ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 990, + 253, + 1064, + 253, + 1064, + 277, + 990, + 277 + ], + "score": 0.9, + "latex": "N = 2" + }, + { + "category_id": 13, + "poly": [ + 349, + 816, + 423, + 816, + 423, + 842, + 349, + 842 + ], + "score": 0.89, + "latex": "n _ { s } , n _ { d }" + }, + { + "category_id": 13, + "poly": [ + 436, + 703, + 478, + 703, + 478, + 733, + 436, + 733 + ], + "score": 0.89, + "latex": "W _ { d }" + }, + { + "category_id": 13, + "poly": [ + 694, + 813, + 818, + 813, + 818, + 841, + 694, + 841 + ], + "score": 0.89, + "latex": "n _ { s } < < n _ { d }" + }, + { + "category_id": 13, + "poly": [ + 330, + 612, + 372, + 612, + 372, + 642, + 330, + 642 + ], + "score": 0.88, + "latex": "W _ { d }" + }, + { + "category_id": 13, + "poly": [ + 1139, + 914, + 1181, + 914, + 1181, + 944, + 1139, + 944 + ], + "score": 0.88, + "latex": "W _ { d }" + }, + { + "category_id": 13, + "poly": [ + 854, + 1974, + 961, + 1974, + 961, + 2004, + 854, + 2004 + ], + "score": 0.88, + "latex": "( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )" + }, + { + "category_id": 13, + "poly": [ + 913, + 867, + 947, + 867, + 947, + 896, + 913, + 896 + ], + "score": 0.86, + "latex": "P _ { \\star }" + }, + { + "category_id": 13, + "poly": [ + 817, + 255, + 849, + 255, + 849, + 279, + 817, + 279 + ], + "score": 0.85, + "latex": "n _ { d }" + }, + { + "category_id": 13, + "poly": [ + 524, + 920, + 549, + 920, + 549, + 944, + 524, + 944 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 583, + 1978, + 629, + 1978, + 629, + 2004, + 583, + 2004 + ], + "score": 0.83, + "latex": "x , y" + }, + { + "category_id": 13, + "poly": [ + 852, + 1026, + 879, + 1026, + 879, + 1050, + 852, + 1050 + ], + "score": 0.83, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1374, + 1979, + 1400, + 1979, + 1400, + 2003, + 1374, + 2003 + ], + "score": 0.81, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1128, + 817, + 1128, + 817, + 1156, + 788, + 1156 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1173, + 643, + 1203, + 643, + 1203, + 670, + 1173, + 670 + ], + "score": 0.79, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1276, + 642, + 1305, + 642, + 1305, + 670, + 1276, + 670 + ], + "score": 0.76, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 646, + 1134, + 662, + 1134, + 662, + 1155, + 646, + 1155 + ], + "score": 0.63, + "latex": "o" + }, + { + "category_id": 13, + "poly": [ + 377, + 1902, + 406, + 1902, + 406, + 1928, + 377, + 1928 + ], + "score": 0.62, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 579, + 497, + 606, + 497, + 606, + 527, + 579, + 527 + ], + "score": 0.5, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1403.0, + 225.0, + 1403.0, + 254.0, + 295.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 251.0, + 816.0, + 251.0, + 816.0, + 283.0, + 294.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 251.0, + 989.0, + 251.0, + 989.0, + 283.0, + 850.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 251.0, + 1406.0, + 251.0, + 1406.0, + 283.0, + 1065.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 280.0, + 1061.0, + 280.0, + 1061.0, + 309.0, + 296.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1810.0, + 894.0, + 1810.0, + 894.0, + 1847.0, + 294.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1277.0, + 1128.0, + 1277.0, + 1128.0, + 1314.0, + 568.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1867.0, + 1404.0, + 1867.0, + 1404.0, + 1909.0, + 294.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1898.0, + 376.0, + 1898.0, + 376.0, + 1937.0, + 293.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1898.0, + 1126.0, + 1898.0, + 1126.0, + 1937.0, + 407.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 1278.0, + 1128.0, + 1278.0, + 1128.0, + 1315.0, + 569.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 748.0, + 1090.0, + 748.0, + 1090.0, + 782.0, + 295.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 748.0, + 1405.0, + 748.0, + 1405.0, + 782.0, + 1235.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 779.0, + 862.0, + 779.0, + 862.0, + 813.0, + 294.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 779.0, + 1405.0, + 779.0, + 1405.0, + 813.0, + 1008.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 811.0, + 348.0, + 811.0, + 348.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 811.0, + 693.0, + 811.0, + 693.0, + 845.0, + 424.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 811.0, + 1403.0, + 811.0, + 1403.0, + 845.0, + 819.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 837.0, + 586.0, + 837.0, + 586.0, + 901.0, + 294.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 837.0, + 912.0, + 837.0, + 912.0, + 901.0, + 827.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 837.0, + 1402.0, + 837.0, + 1402.0, + 901.0, + 948.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 913.0, + 523.0, + 913.0, + 523.0, + 948.0, + 296.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 913.0, + 1138.0, + 913.0, + 1138.0, + 948.0, + 550.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 913.0, + 1403.0, + 913.0, + 1403.0, + 948.0, + 1182.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 968.0, + 509.0, + 968.0, + 509.0, + 1007.0, + 295.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 970.0, + 1405.0, + 970.0, + 1405.0, + 1005.0, + 977.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1017.0, + 851.0, + 1017.0, + 851.0, + 1053.0, + 295.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1017.0, + 1403.0, + 1017.0, + 1403.0, + 1053.0, + 880.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1049.0, + 1289.0, + 1049.0, + 1289.0, + 1085.0, + 295.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1097.0, + 957.0, + 1097.0, + 957.0, + 1130.0, + 297.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1097.0, + 1176.0, + 1097.0, + 1176.0, + 1130.0, + 1095.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1097.0, + 1404.0, + 1097.0, + 1404.0, + 1130.0, + 1252.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1126.0, + 645.0, + 1126.0, + 645.0, + 1163.0, + 294.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1126.0, + 787.0, + 1126.0, + 787.0, + 1163.0, + 663.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1126.0, + 1405.0, + 1126.0, + 1405.0, + 1163.0, + 818.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1157.0, + 1405.0, + 1157.0, + 1405.0, + 1194.0, + 293.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1187.0, + 1405.0, + 1187.0, + 1405.0, + 1224.0, + 293.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1216.0, + 874.0, + 1216.0, + 874.0, + 1255.0, + 292.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 610.0, + 329.0, + 610.0, + 329.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 610.0, + 709.0, + 610.0, + 709.0, + 646.0, + 373.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 610.0, + 1406.0, + 610.0, + 1406.0, + 646.0, + 907.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 638.0, + 1172.0, + 638.0, + 1172.0, + 678.0, + 292.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 638.0, + 1275.0, + 638.0, + 1275.0, + 678.0, + 1204.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 638.0, + 1405.0, + 638.0, + 1405.0, + 678.0, + 1306.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 673.0, + 1403.0, + 673.0, + 1403.0, + 706.0, + 295.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 700.0, + 435.0, + 700.0, + 435.0, + 736.0, + 293.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 700.0, + 490.0, + 700.0, + 490.0, + 736.0, + 479.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1713.0, + 294.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1744.0, + 293.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1741.0, + 1408.0, + 1741.0, + 1408.0, + 1774.0, + 293.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 582.0, + 1968.0, + 582.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1968.0, + 853.0, + 1968.0, + 853.0, + 2010.0, + 630.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1968.0, + 1373.0, + 1968.0, + 1373.0, + 2010.0, + 962.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2010.0, + 1401.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1403.0, + 2004.0, + 1403.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1867.0, + 1404.0, + 1867.0, + 1404.0, + 1909.0, + 294.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1898.0, + 376.0, + 1898.0, + 376.0, + 1937.0, + 293.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1898.0, + 1126.0, + 1898.0, + 1126.0, + 1937.0, + 407.0, + 1937.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1005, + 1405, + 1005, + 1405, + 1252, + 296, + 1252 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 296, + 1265, + 1406, + 1265, + 1406, + 1454, + 296, + 1454 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 662, + 1405, + 662, + 1405, + 818, + 296, + 818 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 305, + 1406, + 305, + 1406, + 444, + 297, + 444 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 543, + 1538, + 1158, + 1538, + 1158, + 1651, + 543, + 1651 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 580, + 1899, + 1121, + 1899, + 1121, + 1959, + 580, + 1959 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 298, + 929, + 1401, + 929, + 1401, + 992, + 298, + 992 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 292, + 1658, + 1401, + 1658, + 1401, + 1724, + 292, + 1724 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1404, + 1972, + 1404, + 2037, + 299, + 2037 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 301, + 585, + 1402, + 585, + 1402, + 650, + 301, + 650 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 297, + 1465, + 1401, + 1465, + 1401, + 1531, + 297, + 1531 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 294, + 228, + 1404, + 228, + 1404, + 295, + 294, + 295 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 430, + 502, + 1271, + 502, + 1271, + 576, + 430, + 576 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 553, + 1784, + 1145, + 1784, + 1145, + 1846, + 553, + 1846 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 299, + 1855, + 852, + 1855, + 852, + 1889, + 299, + 1889 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 300, + 860, + 790, + 860, + 790, + 897, + 300, + 897 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 298, + 1735, + 1398, + 1735, + 1398, + 1771, + 298, + 1771 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 296, + 455, + 1381, + 455, + 1381, + 491, + 296, + 491 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.915 + }, + { + "category_id": 9, + "poly": [ + 1338, + 522, + 1401, + 522, + 1401, + 553, + 1338, + 553 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 865, + 2087, + 865, + 2113, + 834, + 2113 + ], + "score": 0.879 + }, + { + "category_id": 14, + "poly": [ + 431, + 499, + 1269, + 499, + 1269, + 577, + 431, + 577 + ], + "score": 0.94, + "latex": "K _ { 1 } ( o _ { t } { ^ x } - o _ { t } { ^ y } ) = K _ { 1 } [ W _ { d { ^ x } } - W _ { d { ^ y } } ] s _ { t } = 0 \\implies K _ { d } ( W _ { d { ^ x } } - W _ { d { ^ y } } ) s _ { t } = 0" + }, + { + "category_id": 14, + "poly": [ + 580, + 1898, + 1117, + 1898, + 1117, + 1960, + 580, + 1960 + ], + "score": 0.93, + "latex": "J _ { \\alpha } ( \\mathcal { D } ) = E _ { \\underset { a _ { t } \\sim \\pi _ { \\theta } ( \\cdot | s _ { t } ) } { s _ { t } \\sim \\mathcal { D } } } [ - \\alpha \\log \\pi _ { \\theta } ( a _ { t } | s _ { t } ) - \\alpha \\bar { \\mathcal { H } } ] ," + }, + { + "category_id": 13, + "poly": [ + 910, + 409, + 1057, + 409, + 1057, + 441, + 910, + 441 + ], + "score": 0.93, + "latex": "K _ { d } \\in \\mathbb { R } ^ { h \\times n _ { d } }" + }, + { + "category_id": 13, + "poly": [ + 766, + 1467, + 883, + 1467, + 883, + 1501, + 766, + 1501 + ], + "score": 0.93, + "latex": "Q _ { \\theta } ( s _ { t } , a _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 451, + 2002, + 573, + 2002, + 573, + 2036, + 451, + 2036 + ], + "score": 0.93, + "latex": "\\bar { \\mathcal { H } } = - | \\bar { \\mathcal { A } } |" + }, + { + "category_id": 13, + "poly": [ + 710, + 408, + 856, + 408, + 856, + 441, + 710, + 441 + ], + "score": 0.93, + "latex": "K _ { s } \\in \\mathbb { R } ^ { h \\times n _ { s } }" + }, + { + "category_id": 13, + "poly": [ + 892, + 1737, + 999, + 1737, + 999, + 1771, + 892, + 1771 + ], + "score": 0.93, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 642, + 1389, + 973, + 1389, + 973, + 1423, + 642, + 1423 + ], + "score": 0.92, + "latex": "a _ { t } \\dot { = } \\operatorname { t a n h } ( \\mu _ { \\theta } ( s _ { t } ) + \\sigma _ { \\theta } ( s _ { t } ) \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 652, + 305, + 864, + 305, + 864, + 380, + 652, + 380 + ], + "score": 0.92, + "latex": "o _ { t } { } ^ { y } = \\left[ \\stackrel { 0 . 1 } { W } _ { d ^ { y } } \\right] s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1011, + 376, + 1108, + 376, + 1108, + 409, + 1011, + 409 + ], + "score": 0.92, + "latex": "K _ { 1 } ( o _ { t } ^ { x } )" + }, + { + "category_id": 14, + "poly": [ + 552, + 1781, + 1146, + 1781, + 1146, + 1847, + 552, + 1847 + ], + "score": 0.92, + "latex": "J _ { \\pi } ( \\mathcal { D } ) = E _ { s _ { t } \\sim \\mathcal { D } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\theta } ( \\cdot | s _ { t } ) | | \\exp \\{ \\frac { 1 } { \\alpha } Q _ { \\theta } ( s _ { t } , \\cdot ) \\} ) ] ." + }, + { + "category_id": 13, + "poly": [ + 436, + 616, + 683, + 616, + 683, + 650, + 436, + 650 + ], + "score": 0.92, + "latex": "K _ { d } ( W _ { d ^ { x } } - W _ { d ^ { y } } ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1391, + 1196, + 1391, + 1196, + 1423, + 1056, + 1423 + ], + "score": 0.92, + "latex": "\\epsilon \\sim \\mathcal { N } ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 507, + 787, + 597, + 787, + 597, + 816, + 507, + 816 + ], + "score": 0.92, + "latex": "K _ { d } = 0" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1359, + 1164, + 1359, + 1164, + 1391, + 1056, + 1391 + ], + "score": 0.91, + "latex": "\\pi _ { \\boldsymbol { \\theta } } ( a _ { t } | \\boldsymbol { s } _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 442, + 409, + 631, + 409, + 631, + 444, + 442, + 444 + ], + "score": 0.91, + "latex": "K _ { 1 } = [ K _ { s } \\quad K _ { d } ]" + }, + { + "category_id": 13, + "poly": [ + 391, + 305, + 601, + 305, + 601, + 380, + 391, + 380 + ], + "score": 0.91, + "latex": "o _ { t } { } ^ { x } = \\left[ \\begin{array} { c } { 0 . 1 W _ { c } } \\\\ { W _ { d ^ { x } } } \\end{array} \\right] s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1296, + 328, + 1353, + 328, + 1353, + 358, + 1296, + 358 + ], + "score": 0.91, + "latex": "W _ { d ^ { x } }" + }, + { + "category_id": 13, + "poly": [ + 370, + 1970, + 450, + 1970, + 450, + 2001, + 370, + 2001 + ], + "score": 0.91, + "latex": "\\bar { \\mathcal { H } } \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1298, + 1367, + 1298, + 1367, + 1332, + 1214, + 1332 + ], + "score": 0.91, + "latex": "( S , \\mathcal { A } , p , r , \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1161, + 376, + 1257, + 376, + 1257, + 408, + 1161, + 408 + ], + "score": 0.91, + "latex": "K _ { 1 } ( o _ { t } ^ { y } )" + }, + { + "category_id": 14, + "poly": [ + 542, + 1537, + 1155, + 1537, + 1155, + 1652, + 542, + 1652 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { J _ { Q } ( \\mathcal { D } ) = E _ { ( s _ { t } , a _ { t } , s _ { t } ^ { \\prime } ) \\sim \\mathcal { D } } [ ( Q _ { \\theta } ( s _ { t } , a _ { t } ) - y _ { t } ) ^ { 2 } ] } \\\\ & { \\qquad a _ { t } ^ { \\prime } { \\sim } \\pi ( \\cdot | s _ { t } ^ { \\prime } ) } \\\\ & { y _ { t } = r ( s _ { t } , a _ { t } ) + \\gamma [ Q _ { \\theta ^ { \\prime } } ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) - \\alpha \\log \\pi _ { \\theta } ( a _ { t } ^ { \\prime } | s _ { t } ^ { \\prime } ) ] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1111, + 408, + 1317, + 408, + 1317, + 441, + 1111, + 441 + ], + "score": 0.89, + "latex": "K _ { 1 } \\in \\mathbb { R } ^ { h \\times ( n _ { s } + n _ { d } ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 378, + 353, + 378, + 353, + 408, + 298, + 408 + ], + "score": 0.89, + "latex": "W _ { d } y" + }, + { + "category_id": 13, + "poly": [ + 1040, + 459, + 1148, + 459, + 1148, + 489, + 1040, + 489 + ], + "score": 0.89, + "latex": "( o _ { t } { } ^ { x } , o _ { t } { } ^ { y } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1299, + 334, + 1299, + 334, + 1330, + 298, + 1330 + ], + "score": 0.88, + "latex": "Q _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 550, + 1303, + 581, + 1303, + 581, + 1329, + 550, + 1329 + ], + "score": 0.85, + "latex": "\\pi _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1329, + 1395, + 1360, + 1395, + 1360, + 1420, + 1329, + 1420 + ], + "score": 0.85, + "latex": "\\sigma _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 759, + 1660, + 783, + 1660, + 783, + 1687, + 759, + 1687 + ], + "score": 0.85, + "latex": "\\theta ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 647, + 267, + 673, + 267, + 673, + 292, + 647, + 292 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 517, + 1394, + 543, + 1394, + 543, + 1420, + 517, + 1420 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1394, + 1279, + 1394, + 1279, + 1422, + 1247, + 1422 + ], + "score": 0.84, + "latex": "\\mu _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1661, + 397, + 1661, + 397, + 1687, + 372, + 1687 + ], + "score": 0.82, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 439, + 1334, + 458, + 1334, + 458, + 1361, + 439, + 1361 + ], + "score": 0.82, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 1346, + 465, + 1371, + 465, + 1371, + 488, + 1346, + 488 + ], + "score": 0.82, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 783, + 1304, + 805, + 1304, + 805, + 1326, + 783, + 1326 + ], + "score": 0.8, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 908, + 1330, + 925, + 1330, + 925, + 1356, + 908, + 1356 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 568, + 1863, + 588, + 1863, + 588, + 1884, + 568, + 1884 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 806, + 266, + 819, + 266, + 819, + 289, + 806, + 289 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 856.0, + 794.0, + 856.0, + 794.0, + 902.0, + 292.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 871.0, + 2083.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1007.0, + 1403.0, + 1007.0, + 1403.0, + 1041.0, + 295.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1038.0, + 1406.0, + 1038.0, + 1406.0, + 1072.0, + 295.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1068.0, + 1403.0, + 1068.0, + 1403.0, + 1102.0, + 295.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1098.0, + 1406.0, + 1098.0, + 1406.0, + 1132.0, + 294.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1130.0, + 1405.0, + 1130.0, + 1405.0, + 1163.0, + 295.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1191.0, + 296.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1189.0, + 1405.0, + 1189.0, + 1405.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1221.0, + 470.0, + 1221.0, + 470.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1264.0, + 1406.0, + 1264.0, + 1406.0, + 1301.0, + 293.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1295.0, + 297.0, + 1295.0, + 297.0, + 1336.0, + 294.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1295.0, + 549.0, + 1295.0, + 549.0, + 1336.0, + 335.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1295.0, + 782.0, + 1295.0, + 782.0, + 1336.0, + 582.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1295.0, + 1213.0, + 1295.0, + 1213.0, + 1336.0, + 806.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1295.0, + 1406.0, + 1295.0, + 1406.0, + 1336.0, + 1368.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1329.0, + 438.0, + 1329.0, + 438.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1329.0, + 907.0, + 1329.0, + 907.0, + 1365.0, + 459.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1329.0, + 1406.0, + 1329.0, + 1406.0, + 1365.0, + 926.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1055.0, + 1358.0, + 1055.0, + 1393.0, + 294.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1358.0, + 1408.0, + 1358.0, + 1408.0, + 1393.0, + 1165.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1389.0, + 516.0, + 1389.0, + 516.0, + 1425.0, + 295.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 1389.0, + 641.0, + 1389.0, + 641.0, + 1425.0, + 544.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1389.0, + 1055.0, + 1389.0, + 1055.0, + 1425.0, + 974.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1389.0, + 1246.0, + 1389.0, + 1246.0, + 1425.0, + 1197.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1389.0, + 1328.0, + 1389.0, + 1328.0, + 1425.0, + 1280.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1389.0, + 1406.0, + 1389.0, + 1406.0, + 1425.0, + 1361.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1419.0, + 754.0, + 1419.0, + 754.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 696.0, + 296.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 695.0, + 1403.0, + 695.0, + 1403.0, + 728.0, + 295.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 725.0, + 1403.0, + 725.0, + 1403.0, + 759.0, + 295.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 755.0, + 1407.0, + 755.0, + 1407.0, + 788.0, + 295.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 506.0, + 784.0, + 506.0, + 820.0, + 294.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 784.0, + 1024.0, + 784.0, + 1024.0, + 820.0, + 598.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 787.0, + 1406.0, + 787.0, + 1406.0, + 817.0, + 1373.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 322.0, + 390.0, + 322.0, + 390.0, + 364.0, + 289.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 291.0, + 651.0, + 291.0, + 651.0, + 384.0, + 602.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 291.0, + 1295.0, + 291.0, + 1295.0, + 384.0, + 865.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 291.0, + 1408.0, + 291.0, + 1408.0, + 384.0, + 1354.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 372.0, + 297.0, + 372.0, + 297.0, + 413.0, + 294.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 372.0, + 1010.0, + 372.0, + 1010.0, + 413.0, + 354.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 372.0, + 1160.0, + 372.0, + 1160.0, + 413.0, + 1109.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 372.0, + 1408.0, + 372.0, + 1408.0, + 413.0, + 1258.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 403.0, + 441.0, + 403.0, + 441.0, + 444.0, + 291.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 403.0, + 709.0, + 403.0, + 709.0, + 444.0, + 632.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 403.0, + 909.0, + 403.0, + 909.0, + 444.0, + 857.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 403.0, + 1110.0, + 403.0, + 1110.0, + 444.0, + 1058.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 403.0, + 1330.0, + 403.0, + 1330.0, + 444.0, + 1318.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.75, + 335.5, + 452.75, + 335.5, + 452.75, + 352.5, + 436.75, + 352.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 335.5, + 551.0, + 335.5, + 551.0, + 380.0, + 487.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 929.0, + 1405.0, + 929.0, + 1405.0, + 964.0, + 296.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 961.0, + 1044.0, + 961.0, + 1044.0, + 992.0, + 296.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1659.0, + 371.0, + 1659.0, + 371.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1659.0, + 758.0, + 1659.0, + 758.0, + 1695.0, + 398.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1659.0, + 1403.0, + 1659.0, + 1403.0, + 1695.0, + 784.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1688.0, + 1247.0, + 1688.0, + 1247.0, + 1727.0, + 292.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1967.0, + 369.0, + 1967.0, + 369.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1967.0, + 1406.0, + 1967.0, + 1406.0, + 2008.0, + 451.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1995.0, + 450.0, + 1995.0, + 450.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1995.0, + 586.0, + 1995.0, + 586.0, + 2039.0, + 574.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 584.0, + 1404.0, + 584.0, + 1404.0, + 620.0, + 296.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 615.0, + 435.0, + 615.0, + 435.0, + 651.0, + 296.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 615.0, + 694.0, + 615.0, + 694.0, + 651.0, + 684.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1464.0, + 765.0, + 1464.0, + 765.0, + 1504.0, + 293.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1464.0, + 1406.0, + 1464.0, + 1406.0, + 1504.0, + 884.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1497.0, + 493.0, + 1497.0, + 493.0, + 1529.0, + 296.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 646.0, + 260.0, + 646.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 260.0, + 805.0, + 260.0, + 805.0, + 297.0, + 674.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 260.0, + 831.0, + 260.0, + 831.0, + 297.0, + 820.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1855.0, + 567.0, + 1855.0, + 567.0, + 1892.0, + 296.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1855.0, + 853.0, + 1855.0, + 853.0, + 1892.0, + 589.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1733.0, + 891.0, + 1733.0, + 891.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1733.0, + 1404.0, + 1733.0, + 1404.0, + 1777.0, + 1000.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 450.0, + 1039.0, + 450.0, + 1039.0, + 499.0, + 293.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 450.0, + 1345.0, + 450.0, + 1345.0, + 499.0, + 1149.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 450.0, + 1384.0, + 450.0, + 1384.0, + 499.0, + 1372.0, + 499.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 334, + 238, + 1384, + 238, + 1384, + 1958, + 334, + 1958 + ], + "score": 0.964 + }, + { + "category_id": 4, + "poly": [ + 298, + 1970, + 1407, + 1970, + 1407, + 2055, + 298, + 2055 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.852 + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 298.0, + 1128.0, + 298.0, + 1128.0, + 315.0, + 1111.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 469.0, + 1339.0, + 469.0, + 1339.0, + 537.0, + 1277.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 829.0, + 1154.0, + 829.0, + 1154.0, + 855.0, + 1131.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 977.0, + 751.0, + 977.0, + 751.0, + 1045.0, + 691.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.75, + 1699.5, + 1110.75, + 1699.5, + 1110.75, + 1726.5, + 1061.75, + 1726.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2002.0, + 295.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1997.0, + 1405.0, + 1997.0, + 1405.0, + 2029.0, + 293.0, + 2029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2028.0, + 485.0, + 2028.0, + 485.0, + 2056.0, + 296.0, + 2056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 871.0, + 2083.0, + 871.0, + 2123.0, + 829.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 25, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 518, + 1405, + 518, + 1405, + 824, + 297, + 824 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1417, + 1405, + 1417, + 1405, + 1723, + 297, + 1723 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 912, + 1404, + 912, + 1404, + 1067, + 298, + 1067 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 536, + 1816, + 1155, + 1816, + 1155, + 2038, + 536, + 2038 + ], + "score": 0.978, + "html": "
HyperparameterSetting
Contrastive temperature (1/入)0.1
Auxiliary loss coefficient (α)1.0
I-scale parameter (β)0.1
Batch Size(LcME)128
|1000 // Action Repeat
" + }, + { + "category_id": 1, + "poly": [ + 299, + 290, + 1404, + 290, + 1404, + 412, + 299, + 412 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1081, + 1404, + 1081, + 1404, + 1205, + 297, + 1205 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 301, + 1219, + 1401, + 1219, + 1401, + 1313, + 301, + 1313 + ], + "score": 0.963 + }, + { + "category_id": 6, + "poly": [ + 396, + 1761, + 1296, + 1761, + 1296, + 1794, + 396, + 1794 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 298, + 852, + 864, + 852, + 864, + 884, + 298, + 884 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 299, + 458, + 609, + 458, + 609, + 488, + 299, + 488 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 302, + 1357, + 788, + 1357, + 788, + 1389, + 302, + 1389 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 864, + 2087, + 864, + 2113, + 834, + 2113 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 741, + 229, + 741, + 261, + 300, + 261 + ], + "score": 0.877 + }, + { + "category_id": 13, + "poly": [ + 714, + 1220, + 896, + 1220, + 896, + 1251, + 714, + 1251 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { \\mathrm { R L } } + \\alpha \\mathcal { L } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 609, + 1632, + 702, + 1632, + 702, + 1661, + 609, + 1661 + ], + "score": 0.9, + "latex": "8 4 \\times 8 4" + }, + { + "category_id": 13, + "poly": [ + 716, + 1252, + 796, + 1252, + 796, + 1285, + 716, + 1285 + ], + "score": 0.9, + "latex": "J _ { \\alpha } ( { \\mathcal { D } } ) _ { , }" + }, + { + "category_id": 13, + "poly": [ + 402, + 1005, + 481, + 1005, + 481, + 1037, + 402, + 1037 + ], + "score": 0.9, + "latex": "{ \\mathcal { L } } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 973, + 1221, + 1028, + 1221, + 1028, + 1252, + 973, + 1252 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { \\mathrm { R L } }" + }, + { + "category_id": 13, + "poly": [ + 578, + 1251, + 654, + 1251, + 654, + 1285, + 578, + 1285 + ], + "score": 0.9, + "latex": "J _ { \\pi } ( \\mathcal { D } )" + }, + { + "category_id": 13, + "poly": [ + 489, + 1251, + 565, + 1251, + 565, + 1285, + 489, + 1285 + ], + "score": 0.9, + "latex": "J _ { \\pi } ( \\mathcal { D } )" + }, + { + "category_id": 13, + "poly": [ + 885, + 1252, + 964, + 1252, + 964, + 1283, + 885, + 1283 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } _ { \\mathrm { C M E } }" + }, + { + "category_id": 13, + "poly": [ + 928, + 914, + 1026, + 914, + 1026, + 942, + 928, + 942 + ], + "score": 0.88, + "latex": "k = 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 570, + 550, + 636, + 550, + 636, + 579, + 570, + 579 + ], + "score": 0.88, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 570, + 947, + 600, + 947, + 600, + 975, + 570, + 975 + ], + "score": 0.85, + "latex": "z _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 298, + 322, + 322, + 322, + 322, + 353, + 298, + 353 + ], + "score": 0.83, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 517, + 1288, + 537, + 1288, + 537, + 1309, + 517, + 1309 + ], + "score": 0.8, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1039, + 1510, + 1120, + 1510, + 1120, + 1539, + 1039, + 1539 + ], + "score": 0.67, + "latex": "K = 2" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1510, + 1213, + 1510, + 1213, + 1541, + 1040, + 1541 + ], + "score": 0.52, + "latex": "K = 2 , M = 2" + }, + { + "category_id": 13, + "poly": [ + 594, + 1541, + 647, + 1541, + 647, + 1572, + 594, + 1572 + ], + "score": 0.33, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1759.0, + 1303.0, + 1759.0, + 1303.0, + 1797.0, + 395.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 849.0, + 868.0, + 849.0, + 868.0, + 889.0, + 293.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 457.0, + 615.0, + 457.0, + 615.0, + 492.0, + 294.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1355.0, + 792.0, + 1355.0, + 792.0, + 1393.0, + 294.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2083.0, + 870.0, + 2083.0, + 870.0, + 2126.0, + 830.0, + 2126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 744.0, + 229.0, + 744.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 517.0, + 1403.0, + 517.0, + 1403.0, + 549.0, + 296.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 549.0, + 569.0, + 549.0, + 569.0, + 583.0, + 294.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 549.0, + 1405.0, + 549.0, + 1405.0, + 583.0, + 637.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 580.0, + 1405.0, + 580.0, + 1405.0, + 613.0, + 292.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 609.0, + 1406.0, + 609.0, + 1406.0, + 644.0, + 294.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 641.0, + 1407.0, + 641.0, + 1407.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 672.0, + 1403.0, + 672.0, + 1403.0, + 704.0, + 295.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1407.0, + 699.0, + 1407.0, + 737.0, + 294.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 732.0, + 1405.0, + 732.0, + 1405.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 762.0, + 1405.0, + 762.0, + 1405.0, + 798.0, + 295.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 792.0, + 862.0, + 792.0, + 862.0, + 827.0, + 295.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1419.0, + 1407.0, + 1419.0, + 1407.0, + 1451.0, + 296.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1447.0, + 1407.0, + 1447.0, + 1407.0, + 1486.0, + 294.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1405.0, + 1479.0, + 1405.0, + 1514.0, + 295.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1508.0, + 1038.0, + 1508.0, + 1038.0, + 1545.0, + 291.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1508.0, + 1405.0, + 1508.0, + 1405.0, + 1545.0, + 1214.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1538.0, + 593.0, + 1538.0, + 593.0, + 1575.0, + 294.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1538.0, + 1407.0, + 1538.0, + 1407.0, + 1575.0, + 648.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1566.0, + 1408.0, + 1566.0, + 1408.0, + 1607.0, + 292.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1597.0, + 1406.0, + 1597.0, + 1406.0, + 1637.0, + 292.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 608.0, + 1632.0, + 608.0, + 1664.0, + 295.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1632.0, + 1402.0, + 1632.0, + 1402.0, + 1664.0, + 703.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1659.0, + 1406.0, + 1659.0, + 1406.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1694.0, + 813.0, + 1694.0, + 813.0, + 1726.0, + 295.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 914.0, + 927.0, + 914.0, + 927.0, + 947.0, + 296.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 914.0, + 1405.0, + 914.0, + 1405.0, + 947.0, + 1027.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 942.0, + 569.0, + 942.0, + 569.0, + 979.0, + 294.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 942.0, + 1406.0, + 942.0, + 1406.0, + 979.0, + 601.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 974.0, + 1406.0, + 974.0, + 1406.0, + 1011.0, + 294.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1003.0, + 401.0, + 1003.0, + 401.0, + 1040.0, + 294.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1003.0, + 1409.0, + 1003.0, + 1409.0, + 1040.0, + 482.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1035.0, + 1065.0, + 1035.0, + 1065.0, + 1072.0, + 294.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 287.0, + 1404.0, + 287.0, + 1404.0, + 324.0, + 293.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 353.0, + 323.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 385.0, + 295.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 381.0, + 1406.0, + 381.0, + 1406.0, + 413.0, + 295.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1082.0, + 1407.0, + 1082.0, + 1407.0, + 1119.0, + 294.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1114.0, + 1407.0, + 1114.0, + 1407.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1144.0, + 1406.0, + 1144.0, + 1406.0, + 1180.0, + 294.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1176.0, + 1263.0, + 1176.0, + 1263.0, + 1208.0, + 296.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1215.0, + 713.0, + 1215.0, + 713.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1215.0, + 972.0, + 1215.0, + 972.0, + 1256.0, + 897.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1215.0, + 1405.0, + 1215.0, + 1405.0, + 1256.0, + 1029.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1248.0, + 488.0, + 1248.0, + 488.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 1248.0, + 577.0, + 1248.0, + 577.0, + 1287.0, + 566.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1248.0, + 715.0, + 1248.0, + 715.0, + 1287.0, + 655.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1248.0, + 884.0, + 1248.0, + 884.0, + 1287.0, + 797.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1248.0, + 1405.0, + 1248.0, + 1405.0, + 1287.0, + 965.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1279.0, + 516.0, + 1279.0, + 516.0, + 1315.0, + 296.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1279.0, + 550.0, + 1279.0, + 550.0, + 1315.0, + 538.0, + 1315.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 26, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 564, + 278, + 1129, + 278, + 1129, + 682, + 564, + 682 + ], + "score": 0.979, + "html": "
ParameterSetting
Replay buffer capacity100,000
Seed steps1,000
Batch size (DrQ)512
Discount γ0.99
OptimizerAdam
Learning rate10-3
Critic target update frequency2
Critic Q-function soft-update rate T0.01
Actor update frequency2
Actor log stddev bounds
Init temperature[-10,2] 0.1
" + }, + { + "category_id": 3, + "poly": [ + 296, + 1265, + 1404, + 1265, + 1404, + 1887, + 296, + 1887 + ], + "score": 0.973 + }, + { + "category_id": 5, + "poly": [ + 653, + 834, + 1046, + 834, + 1046, + 1086, + 653, + 1086 + ], + "score": 0.971, + "html": "
Task nameAction repeat
Cartpole Swingup8
Reacher Easy4
Cheetah Run4
Finger Spin2
Ball In Cup Catch4
Walker Walk2
" + }, + { + "category_id": 4, + "poly": [ + 297, + 1913, + 1405, + 1913, + 1405, + 2026, + 297, + 2026 + ], + "score": 0.962 + }, + { + "category_id": 6, + "poly": [ + 297, + 779, + 1390, + 779, + 1390, + 811, + 297, + 811 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 812, + 75, + 812, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 6, + "poly": [ + 296, + 223, + 1392, + 223, + 1392, + 255, + 296, + 255 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.861 + }, + { + "category_id": 0, + "poly": [ + 298, + 1161, + 681, + 1161, + 681, + 1191, + 298, + 1191 + ], + "score": 0.691 + }, + { + "category_id": 2, + "poly": [ + 298, + 1161, + 681, + 1161, + 681, + 1191, + 298, + 1191 + ], + "score": 0.128 + }, + { + "category_id": 13, + "poly": [ + 947, + 1567, + 1005, + 1567, + 1005, + 1595, + 947, + 1595 + ], + "score": 0.83, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 588, + 1858, + 645, + 1858, + 645, + 1884, + 588, + 1884 + ], + "score": 0.83, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 587, + 1567, + 646, + 1567, + 646, + 1595, + 587, + 1595 + ], + "score": 0.82, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 947, + 1858, + 1004, + 1858, + 1004, + 1884, + 947, + 1884 + ], + "score": 0.81, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 1305, + 1567, + 1363, + 1567, + 1363, + 1595, + 1305, + 1595 + ], + "score": 0.8, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 1305, + 1857, + 1362, + 1857, + 1362, + 1884, + 1305, + 1884 + ], + "score": 0.79, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 690, + 432, + 710, + 432, + 710, + 459, + 690, + 459 + ], + "score": 0.34, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 734, + 1942, + 784, + 1942, + 784, + 1970, + 734, + 1970 + ], + "score": 0.26, + "latex": "_ \\mathrm { D r Q }" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1273.0, + 874.0, + 1273.0, + 874.0, + 1301.0, + 748.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1271.0, + 1010.0, + 1271.0, + 1010.0, + 1303.0, + 954.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1318.0, + 629.0, + 1318.0, + 629.0, + 1351.0, + 411.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1315.0, + 986.0, + 1315.0, + 986.0, + 1354.0, + 770.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1320.0, + 1316.0, + 1320.0, + 1316.0, + 1349.0, + 1162.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1340.0, + 735.0, + 1340.0, + 735.0, + 1372.0, + 686.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1355.0, + 378.0, + 1355.0, + 378.0, + 1525.0, + 296.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1375.0, + 1089.0, + 1375.0, + 1089.0, + 1403.0, + 1046.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1395.0, + 732.0, + 1395.0, + 732.0, + 1428.0, + 686.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1421.0, + 1090.0, + 1421.0, + 1090.0, + 1452.0, + 1045.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1439.0, + 371.0, + 1439.0, + 371.0, + 1468.0, + 332.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1451.0, + 732.0, + 1451.0, + 732.0, + 1483.0, + 686.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1479.0, + 371.0, + 1479.0, + 371.0, + 1506.0, + 330.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1468.0, + 1092.0, + 1468.0, + 1092.0, + 1500.0, + 1045.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1522.0, + 372.0, + 1522.0, + 372.0, + 1543.0, + 353.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1506.0, + 733.0, + 1506.0, + 733.0, + 1537.0, + 686.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1518.0, + 1092.0, + 1518.0, + 1092.0, + 1541.0, + 1070.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1537.0, + 397.0, + 1537.0, + 397.0, + 1565.0, + 357.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1538.0, + 455.0, + 1538.0, + 455.0, + 1565.0, + 415.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1538.0, + 513.0, + 1538.0, + 513.0, + 1566.0, + 471.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1538.0, + 569.0, + 1538.0, + 569.0, + 1565.0, + 528.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1538.0, + 629.0, + 1538.0, + 629.0, + 1565.0, + 586.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1536.0, + 684.0, + 1536.0, + 684.0, + 1565.0, + 644.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1534.0, + 758.0, + 1534.0, + 758.0, + 1566.0, + 716.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1538.0, + 815.0, + 1538.0, + 815.0, + 1565.0, + 772.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1536.0, + 873.0, + 1536.0, + 873.0, + 1567.0, + 827.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1538.0, + 929.0, + 1538.0, + 929.0, + 1566.0, + 886.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1537.0, + 987.0, + 1537.0, + 987.0, + 1565.0, + 944.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1537.0, + 1043.0, + 1537.0, + 1043.0, + 1565.0, + 1001.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1537.0, + 1116.0, + 1537.0, + 1116.0, + 1565.0, + 1074.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1538.0, + 1172.0, + 1538.0, + 1172.0, + 1565.0, + 1132.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1538.0, + 1230.0, + 1538.0, + 1230.0, + 1566.0, + 1188.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1538.0, + 1287.0, + 1538.0, + 1287.0, + 1566.0, + 1245.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 1538.0, + 1346.0, + 1538.0, + 1346.0, + 1565.0, + 1303.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1537.0, + 1400.0, + 1537.0, + 1400.0, + 1565.0, + 1361.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1563.0, + 586.0, + 1563.0, + 586.0, + 1595.0, + 388.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1563.0, + 650.0, + 1563.0, + 650.0, + 1595.0, + 647.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1565.0, + 946.0, + 1565.0, + 946.0, + 1596.0, + 747.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1565.0, + 1011.0, + 1565.0, + 1011.0, + 1596.0, + 1006.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1563.0, + 1304.0, + 1563.0, + 1304.0, + 1595.0, + 1104.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 1563.0, + 1367.0, + 1563.0, + 1367.0, + 1595.0, + 1364.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1605.0, + 591.0, + 1605.0, + 591.0, + 1645.0, + 449.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1603.0, + 965.0, + 1603.0, + 965.0, + 1645.0, + 794.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1607.0, + 1316.0, + 1607.0, + 1316.0, + 1640.0, + 1162.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1631.0, + 735.0, + 1631.0, + 735.0, + 1661.0, + 676.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1647.0, + 330.0, + 1647.0, + 330.0, + 1813.0, + 298.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1662.0, + 371.0, + 1662.0, + 371.0, + 1691.0, + 332.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1669.0, + 733.0, + 1669.0, + 733.0, + 1700.0, + 686.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1655.0, + 1092.0, + 1655.0, + 1092.0, + 1686.0, + 1045.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1701.0, + 372.0, + 1701.0, + 372.0, + 1735.0, + 330.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1706.0, + 733.0, + 1706.0, + 733.0, + 1739.0, + 686.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1701.0, + 1090.0, + 1701.0, + 1090.0, + 1733.0, + 1045.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1750.0, + 372.0, + 1750.0, + 372.0, + 1777.0, + 330.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1745.0, + 733.0, + 1745.0, + 733.0, + 1777.0, + 686.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1748.0, + 1092.0, + 1748.0, + 1092.0, + 1779.0, + 1045.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1798.0, + 369.0, + 1798.0, + 369.0, + 1815.0, + 354.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1784.0, + 733.0, + 1784.0, + 733.0, + 1815.0, + 685.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1799.0, + 1090.0, + 1799.0, + 1090.0, + 1823.0, + 1069.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1827.0, + 400.0, + 1827.0, + 400.0, + 1853.0, + 358.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1828.0, + 456.0, + 1828.0, + 456.0, + 1854.0, + 412.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1828.0, + 514.0, + 1828.0, + 514.0, + 1854.0, + 470.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1828.0, + 572.0, + 1828.0, + 572.0, + 1854.0, + 527.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1828.0, + 629.0, + 1828.0, + 629.0, + 1854.0, + 586.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1827.0, + 683.0, + 1827.0, + 683.0, + 1853.0, + 642.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1827.0, + 758.0, + 1827.0, + 758.0, + 1853.0, + 715.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 1828.0, + 815.0, + 1828.0, + 815.0, + 1854.0, + 771.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1828.0, + 873.0, + 1828.0, + 873.0, + 1854.0, + 829.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1828.0, + 929.0, + 1828.0, + 929.0, + 1854.0, + 885.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1828.0, + 986.0, + 1828.0, + 986.0, + 1854.0, + 944.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1827.0, + 1043.0, + 1827.0, + 1043.0, + 1854.0, + 1002.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1827.0, + 1117.0, + 1827.0, + 1117.0, + 1853.0, + 1074.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1828.0, + 1173.0, + 1828.0, + 1173.0, + 1854.0, + 1131.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1828.0, + 1231.0, + 1828.0, + 1231.0, + 1854.0, + 1188.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1828.0, + 1288.0, + 1828.0, + 1288.0, + 1854.0, + 1244.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 1828.0, + 1347.0, + 1828.0, + 1347.0, + 1854.0, + 1303.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1828.0, + 1399.0, + 1828.0, + 1399.0, + 1852.0, + 1362.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1854.0, + 587.0, + 1854.0, + 587.0, + 1886.0, + 388.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1854.0, + 651.0, + 1854.0, + 651.0, + 1886.0, + 646.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1854.0, + 946.0, + 1854.0, + 946.0, + 1886.0, + 747.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1854.0, + 1010.0, + 1854.0, + 1010.0, + 1886.0, + 1005.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1854.0, + 1304.0, + 1854.0, + 1304.0, + 1886.0, + 1105.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1854.0, + 1367.0, + 1854.0, + 1367.0, + 1886.0, + 1363.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1938.0, + 733.0, + 1938.0, + 733.0, + 1978.0, + 292.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 1938.0, + 1405.0, + 1938.0, + 1405.0, + 1978.0, + 785.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1966.0, + 1406.0, + 1966.0, + 1406.0, + 2001.0, + 294.0, + 2001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1996.0, + 755.0, + 1996.0, + 755.0, + 2029.0, + 296.0, + 2029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 779.0, + 1392.0, + 779.0, + 1392.0, + 813.0, + 305.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 223.0, + 1395.0, + 223.0, + 1395.0, + 257.0, + 302.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2122.0, + 830.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 684.0, + 1159.0, + 684.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 684.0, + 1159.0, + 684.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 27, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 296, + 231, + 1403, + 231, + 1403, + 852, + 296, + 852 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 296, + 879, + 1405, + 879, + 1405, + 993, + 296, + 993 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 266, + 1172, + 1369, + 1172, + 1369, + 1868, + 266, + 1868 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 273, + 1951, + 880, + 1951, + 880, + 2034, + 273, + 2034 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 296, + 1120, + 952, + 1120, + 952, + 1151, + 296, + 1151 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 297, + 1896, + 606, + 1896, + 606, + 1926, + 297, + 1926 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 298, + 1051, + 553, + 1051, + 553, + 1085, + 298, + 1085 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2112, + 834, + 2112 + ], + "score": 0.836 + }, + { + "category_id": 13, + "poly": [ + 581, + 1564, + 619, + 1564, + 619, + 1586, + 581, + 1586 + ], + "score": 0.85, + "latex": "+ =" + }, + { + "category_id": 13, + "poly": [ + 588, + 533, + 646, + 533, + 646, + 561, + 588, + 561 + ], + "score": 0.84, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 595, + 1536, + 634, + 1536, + 634, + 1559, + 595, + 1559 + ], + "score": 0.84, + "latex": "+ =" + }, + { + "category_id": 13, + "poly": [ + 947, + 533, + 1005, + 533, + 1005, + 561, + 947, + 561 + ], + "score": 0.83, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 1305, + 533, + 1363, + 533, + 1363, + 561, + 1305, + 561 + ], + "score": 0.83, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 588, + 823, + 646, + 823, + 646, + 850, + 588, + 850 + ], + "score": 0.82, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 580, + 1592, + 619, + 1592, + 619, + 1614, + 580, + 1614 + ], + "score": 0.82, + "latex": "+ =" + }, + { + "category_id": 13, + "poly": [ + 608, + 1510, + 634, + 1510, + 634, + 1530, + 608, + 1530 + ], + "score": 0.81, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 947, + 823, + 1005, + 823, + 1005, + 850, + 947, + 850 + ], + "score": 0.8, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 1306, + 823, + 1363, + 823, + 1363, + 850, + 1306, + 850 + ], + "score": 0.78, + "latex": "( \\times 1 0 ^ { 6 } )" + }, + { + "category_id": 13, + "poly": [ + 443, + 1703, + 470, + 1703, + 470, + 1725, + 443, + 1725 + ], + "score": 0.78, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 911, + 1176, + 943, + 1176, + 943, + 1199, + 911, + 1199 + ], + "score": 0.66, + "latex": "{ } = 0" + }, + { + "category_id": 13, + "poly": [ + 1049, + 1175, + 1137, + 1175, + 1137, + 1200, + 1049, + 1200 + ], + "score": 0.48, + "latex": "s { = } 1 { \\mathrm { e - } } 7" + }, + { + "category_id": 13, + "poly": [ + 355, + 1425, + 379, + 1425, + 379, + 1447, + 355, + 1447 + ], + "score": 0.47, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 730, + 1510, + 751, + 1510, + 751, + 1531, + 730, + 1531 + ], + "score": 0.46, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 415, + 1813, + 439, + 1813, + 439, + 1834, + 415, + 1834 + ], + "score": 0.42, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 787, + 1729, + 814, + 1729, + 814, + 1753, + 787, + 1753 + ], + "score": 0.37, + "latex": "<" + }, + { + "category_id": 13, + "poly": [ + 444, + 1480, + 512, + 1480, + 512, + 1505, + 444, + 1505 + ], + "score": 0.37, + "latex": "\\ c = ~ 1" + }, + { + "category_id": 13, + "poly": [ + 803, + 908, + 852, + 908, + 852, + 935, + 803, + 935 + ], + "score": 0.32, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 1322, + 907, + 1371, + 907, + 1371, + 935, + 1322, + 935 + ], + "score": 0.26, + "latex": "\\mathrm { D r Q }" + }, + { + "category_id": 13, + "poly": [ + 506, + 1482, + 526, + 1482, + 526, + 1503, + 506, + 1503 + ], + "score": 0.25, + "latex": "\\star" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 236.0, + 873.0, + 236.0, + 873.0, + 267.0, + 622.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 236.0, + 1132.0, + 236.0, + 1132.0, + 269.0, + 958.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 284.0, + 628.0, + 284.0, + 628.0, + 316.0, + 413.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 280.0, + 988.0, + 280.0, + 988.0, + 320.0, + 770.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 286.0, + 1316.0, + 286.0, + 1316.0, + 315.0, + 1161.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 324.0, + 370.0, + 324.0, + 370.0, + 490.0, + 296.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 314.0, + 733.0, + 314.0, + 733.0, + 346.0, + 686.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 334.0, + 1091.0, + 334.0, + 1091.0, + 365.0, + 1044.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 389.0, + 372.0, + 389.0, + 372.0, + 416.0, + 330.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 386.0, + 733.0, + 386.0, + 733.0, + 418.0, + 686.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 391.0, + 1091.0, + 391.0, + 1091.0, + 423.0, + 1044.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 450.0, + 372.0, + 450.0, + 372.0, + 478.0, + 329.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 456.0, + 733.0, + 456.0, + 733.0, + 489.0, + 686.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 450.0, + 1092.0, + 450.0, + 1092.0, + 481.0, + 1044.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 461.0, + 1116.0, + 461.0, + 1116.0, + 471.0, + 1106.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 503.0, + 399.0, + 503.0, + 399.0, + 530.0, + 358.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 504.0, + 455.0, + 504.0, + 455.0, + 530.0, + 414.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 504.0, + 513.0, + 504.0, + 513.0, + 531.0, + 471.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 504.0, + 569.0, + 504.0, + 569.0, + 530.0, + 528.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 504.0, + 628.0, + 504.0, + 628.0, + 530.0, + 585.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 501.0, + 684.0, + 501.0, + 684.0, + 530.0, + 643.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 501.0, + 758.0, + 501.0, + 758.0, + 530.0, + 716.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 504.0, + 816.0, + 504.0, + 816.0, + 530.0, + 772.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 504.0, + 871.0, + 504.0, + 871.0, + 531.0, + 829.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 504.0, + 930.0, + 504.0, + 930.0, + 531.0, + 887.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 504.0, + 987.0, + 504.0, + 987.0, + 530.0, + 945.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 501.0, + 1043.0, + 501.0, + 1043.0, + 530.0, + 1002.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 501.0, + 1116.0, + 501.0, + 1116.0, + 530.0, + 1075.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 504.0, + 1173.0, + 504.0, + 1173.0, + 530.0, + 1131.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 504.0, + 1230.0, + 504.0, + 1230.0, + 531.0, + 1189.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 504.0, + 1287.0, + 504.0, + 1287.0, + 531.0, + 1246.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 504.0, + 1346.0, + 504.0, + 1346.0, + 530.0, + 1303.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 503.0, + 1400.0, + 503.0, + 1400.0, + 530.0, + 1360.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 530.0, + 587.0, + 530.0, + 587.0, + 562.0, + 387.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 530.0, + 651.0, + 530.0, + 651.0, + 562.0, + 647.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 530.0, + 946.0, + 530.0, + 946.0, + 562.0, + 746.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 530.0, + 1010.0, + 530.0, + 1010.0, + 562.0, + 1006.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 530.0, + 1304.0, + 530.0, + 1304.0, + 562.0, + 1105.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 530.0, + 1370.0, + 530.0, + 1370.0, + 562.0, + 1364.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 570.0, + 591.0, + 570.0, + 591.0, + 611.0, + 449.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 569.0, + 964.0, + 569.0, + 964.0, + 611.0, + 794.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 573.0, + 1316.0, + 573.0, + 1316.0, + 605.0, + 1161.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 594.0, + 735.0, + 594.0, + 735.0, + 624.0, + 676.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 372.0, + 612.0, + 372.0, + 780.0, + 294.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 639.0, + 733.0, + 639.0, + 733.0, + 670.0, + 686.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 632.0, + 1091.0, + 632.0, + 1091.0, + 663.0, + 1044.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 699.0, + 372.0, + 699.0, + 372.0, + 727.0, + 329.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 684.0, + 733.0, + 684.0, + 733.0, + 715.0, + 686.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 689.0, + 1092.0, + 689.0, + 1092.0, + 720.0, + 1044.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 732.0, + 731.0, + 732.0, + 731.0, + 759.0, + 687.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 756.0, + 372.0, + 756.0, + 372.0, + 789.0, + 330.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 748.0, + 1092.0, + 748.0, + 1092.0, + 779.0, + 1044.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 792.0, + 401.0, + 792.0, + 401.0, + 819.0, + 358.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 794.0, + 456.0, + 794.0, + 456.0, + 821.0, + 413.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 793.0, + 514.0, + 793.0, + 514.0, + 819.0, + 470.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 793.0, + 570.0, + 793.0, + 570.0, + 819.0, + 528.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 793.0, + 628.0, + 793.0, + 628.0, + 819.0, + 585.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 792.0, + 684.0, + 792.0, + 684.0, + 819.0, + 643.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 792.0, + 756.0, + 792.0, + 756.0, + 819.0, + 716.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 796.0, + 813.0, + 796.0, + 813.0, + 818.0, + 773.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 793.0, + 872.0, + 793.0, + 872.0, + 819.0, + 829.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 793.0, + 929.0, + 793.0, + 929.0, + 819.0, + 886.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 793.0, + 987.0, + 793.0, + 987.0, + 819.0, + 945.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 792.0, + 1042.0, + 792.0, + 1042.0, + 819.0, + 1002.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 790.0, + 1117.0, + 790.0, + 1117.0, + 820.0, + 1075.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 796.0, + 1174.0, + 796.0, + 1174.0, + 818.0, + 1131.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 793.0, + 1230.0, + 793.0, + 1230.0, + 819.0, + 1189.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 793.0, + 1287.0, + 793.0, + 1287.0, + 819.0, + 1246.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 793.0, + 1346.0, + 793.0, + 1346.0, + 819.0, + 1303.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 792.0, + 1400.0, + 792.0, + 1400.0, + 819.0, + 1360.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 817.0, + 587.0, + 817.0, + 587.0, + 852.0, + 387.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 817.0, + 651.0, + 817.0, + 651.0, + 852.0, + 647.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 817.0, + 946.0, + 817.0, + 946.0, + 852.0, + 745.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 817.0, + 1009.0, + 817.0, + 1009.0, + 852.0, + 1006.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 817.0, + 1305.0, + 817.0, + 1305.0, + 852.0, + 1105.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 817.0, + 1368.0, + 817.0, + 1368.0, + 852.0, + 1364.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 1406.0, + 877.0, + 1406.0, + 913.0, + 294.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 905.0, + 802.0, + 905.0, + 802.0, + 940.0, + 294.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 905.0, + 1321.0, + 905.0, + 1321.0, + 940.0, + 853.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 905.0, + 1405.0, + 905.0, + 1405.0, + 940.0, + 1372.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 935.0, + 1405.0, + 935.0, + 1405.0, + 967.0, + 294.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 964.0, + 1158.0, + 964.0, + 1158.0, + 994.0, + 296.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1118.0, + 951.0, + 1118.0, + 951.0, + 1154.0, + 295.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1894.0, + 610.0, + 1894.0, + 610.0, + 1929.0, + 294.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1046.0, + 558.0, + 1046.0, + 558.0, + 1092.0, + 291.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2122.0, + 830.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1172.0, + 910.0, + 1172.0, + 910.0, + 1203.0, + 274.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1172.0, + 1048.0, + 1172.0, + 1048.0, + 1203.0, + 944.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1172.0, + 1167.0, + 1172.0, + 1167.0, + 1203.0, + 1138.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1208.0, + 290.0, + 1208.0, + 290.0, + 1225.0, + 274.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1194.0, + 1373.0, + 1194.0, + 1373.0, + 1233.0, + 323.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 1236.0, + 289.0, + 1236.0, + 289.0, + 1255.0, + 275.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1264.0, + 290.0, + 1264.0, + 290.0, + 1281.0, + 274.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1255.0, + 409.0, + 1255.0, + 409.0, + 1287.0, + 324.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 274.0, + 1291.0, + 289.0, + 1291.0, + 289.0, + 1310.0, + 274.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1281.0, + 1360.0, + 1281.0, + 1360.0, + 1316.0, + 353.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1311.0, + 419.0, + 1311.0, + 419.0, + 1346.0, + 348.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 1345.0, + 291.0, + 1345.0, + 291.0, + 1366.0, + 272.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1337.0, + 811.0, + 1337.0, + 811.0, + 1371.0, + 351.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 1374.0, + 289.0, + 1374.0, + 289.0, + 1392.0, + 275.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1366.0, + 1167.0, + 1366.0, + 1167.0, + 1398.0, + 351.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1392.0, + 379.0, + 1392.0, + 379.0, + 1417.0, + 326.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 1401.0, + 291.0, + 1401.0, + 291.0, + 1423.0, + 272.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 1428.0, + 291.0, + 1428.0, + 291.0, + 1449.0, + 272.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1418.0, + 354.0, + 1418.0, + 354.0, + 1454.0, + 321.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1418.0, + 779.0, + 1418.0, + 779.0, + 1454.0, + 380.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 1454.0, + 292.0, + 1454.0, + 292.0, + 1479.0, + 266.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1448.0, + 633.0, + 1448.0, + 633.0, + 1480.0, + 323.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 1484.0, + 290.0, + 1484.0, + 290.0, + 1505.0, + 266.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1479.0, + 443.0, + 1479.0, + 443.0, + 1506.0, + 355.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1479.0, + 707.0, + 1479.0, + 707.0, + 1506.0, + 527.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 1509.0, + 291.0, + 1509.0, + 291.0, + 1535.0, + 266.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1503.0, + 607.0, + 1503.0, + 607.0, + 1535.0, + 354.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1503.0, + 729.0, + 1503.0, + 729.0, + 1535.0, + 635.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1503.0, + 843.0, + 1503.0, + 843.0, + 1535.0, + 752.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1536.0, + 292.0, + 1536.0, + 292.0, + 1564.0, + 265.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1531.0, + 594.0, + 1531.0, + 594.0, + 1562.0, + 354.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1531.0, + 1002.0, + 1531.0, + 1002.0, + 1562.0, + 635.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1565.0, + 292.0, + 1565.0, + 292.0, + 1591.0, + 265.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1561.0, + 580.0, + 1561.0, + 580.0, + 1592.0, + 355.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1561.0, + 993.0, + 1561.0, + 993.0, + 1592.0, + 620.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 1591.0, + 291.0, + 1591.0, + 291.0, + 1618.0, + 266.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1587.0, + 579.0, + 1587.0, + 579.0, + 1618.0, + 355.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1587.0, + 989.0, + 1587.0, + 989.0, + 1618.0, + 620.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1620.0, + 292.0, + 1620.0, + 292.0, + 1646.0, + 265.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1616.0, + 547.0, + 1616.0, + 547.0, + 1647.0, + 354.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1647.0, + 291.0, + 1647.0, + 291.0, + 1673.0, + 265.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1674.0, + 292.0, + 1674.0, + 292.0, + 1700.0, + 265.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1668.0, + 498.0, + 1668.0, + 498.0, + 1702.0, + 323.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 1702.0, + 292.0, + 1702.0, + 292.0, + 1729.0, + 265.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1698.0, + 442.0, + 1698.0, + 442.0, + 1729.0, + 354.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1698.0, + 643.0, + 1698.0, + 643.0, + 1729.0, + 471.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 264.0, + 1730.0, + 292.0, + 1730.0, + 292.0, + 1756.0, + 264.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1723.0, + 786.0, + 1723.0, + 786.0, + 1760.0, + 353.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1723.0, + 885.0, + 1723.0, + 885.0, + 1760.0, + 815.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 261.0, + 1756.0, + 294.0, + 1756.0, + 294.0, + 1785.0, + 261.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1750.0, + 472.0, + 1750.0, + 472.0, + 1787.0, + 381.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 1784.0, + 294.0, + 1784.0, + 294.0, + 1814.0, + 262.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1782.0, + 438.0, + 1782.0, + 438.0, + 1811.0, + 354.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 261.0, + 1811.0, + 294.0, + 1811.0, + 294.0, + 1841.0, + 261.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1809.0, + 414.0, + 1809.0, + 414.0, + 1841.0, + 383.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1809.0, + 569.0, + 1809.0, + 569.0, + 1841.0, + 440.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 1841.0, + 292.0, + 1841.0, + 292.0, + 1866.0, + 262.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1836.0, + 459.0, + 1836.0, + 459.0, + 1868.0, + 325.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 270.0, + 1947.0, + 884.0, + 1947.0, + 884.0, + 1982.0, + 270.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 1984.0, + 291.0, + 1984.0, + 291.0, + 2005.0, + 272.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1977.0, + 823.0, + 1977.0, + 823.0, + 2008.0, + 609.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 2011.0, + 291.0, + 2011.0, + 291.0, + 2032.0, + 272.0, + 2032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 2006.0, + 793.0, + 2006.0, + 793.0, + 2036.0, + 609.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 28, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 262, + 229, + 1392, + 229, + 1392, + 931, + 262, + 931 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 866, + 2088, + 866, + 2113, + 835, + 2113 + ], + "score": 0.83 + }, + { + "category_id": 2, + "poly": [ + 296, + 966, + 315, + 966, + 315, + 987, + 296, + 987 + ], + "score": 0.81 + }, + { + "category_id": 13, + "poly": [ + 505, + 735, + 544, + 735, + 544, + 758, + 505, + 758 + ], + "score": 0.84, + "latex": "+ =" + }, + { + "category_id": 13, + "poly": [ + 504, + 513, + 530, + 513, + 530, + 535, + 504, + 535 + ], + "score": 0.83, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 506, + 790, + 545, + 790, + 545, + 813, + 506, + 813 + ], + "score": 0.81, + "latex": "+ =" + }, + { + "category_id": 13, + "poly": [ + 502, + 596, + 530, + 596, + 530, + 618, + 502, + 618 + ], + "score": 0.81, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 517, + 707, + 545, + 707, + 545, + 728, + 517, + 728 + ], + "score": 0.81, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 608, + 679, + 634, + 679, + 634, + 701, + 608, + 701 + ], + "score": 0.8, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 637, + 374, + 664, + 374, + 664, + 396, + 637, + 396 + ], + "score": 0.8, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 502, + 456, + 529, + 456, + 529, + 480, + 502, + 480 + ], + "score": 0.8, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 578, + 763, + 604, + 763, + 604, + 784, + 578, + 784 + ], + "score": 0.8, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 517, + 541, + 545, + 541, + 545, + 562, + 517, + 562 + ], + "score": 0.79, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1301, + 567, + 1347, + 567, + 1347, + 591, + 1301, + 591 + ], + "score": 0.78, + "latex": "\\mathsf { S } = 1" + }, + { + "category_id": 13, + "poly": [ + 518, + 291, + 545, + 291, + 545, + 313, + 518, + 313 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 578, + 403, + 604, + 403, + 604, + 423, + 578, + 423 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 503, + 874, + 530, + 874, + 530, + 895, + 503, + 895 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1086, + 513, + 1122, + 513, + 1122, + 536, + 1086, + 536 + ], + "score": 0.71, + "latex": "\\ S = 1" + }, + { + "category_id": 13, + "poly": [ + 1165, + 873, + 1197, + 873, + 1197, + 896, + 1165, + 896 + ], + "score": 0.68, + "latex": "^ { = 1 }" + }, + { + "category_id": 13, + "poly": [ + 1240, + 515, + 1256, + 515, + 1256, + 534, + 1240, + 534 + ], + "score": 0.62, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 867, + 570, + 883, + 570, + 883, + 589, + 867, + 589 + ], + "score": 0.57, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 551, + 650, + 573, + 650, + 573, + 674, + 551, + 674 + ], + "score": 0.52, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 741, + 319, + 769, + 319, + 769, + 341, + 741, + 341 + ], + "score": 0.51, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 665, + 235, + 735, + 235, + 735, + 260, + 665, + 260 + ], + "score": 0.5, + "latex": "\\gimel = 1 \\digamma . 0" + }, + { + "category_id": 13, + "poly": [ + 713, + 345, + 769, + 345, + 769, + 370, + 713, + 370 + ], + "score": 0.48, + "latex": "\\_ =" + }, + { + "category_id": 13, + "poly": [ + 612, + 430, + 633, + 430, + 633, + 452, + 612, + 452 + ], + "score": 0.44, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 740, + 347, + 769, + 347, + 769, + 369, + 740, + 369 + ], + "score": 0.41, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1193, + 763, + 1214, + 763, + 1214, + 784, + 1193, + 784 + ], + "score": 0.33, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 965.0, + 316.0, + 965.0, + 316.0, + 988.0, + 297.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 240.0, + 290.0, + 240.0, + 290.0, + 258.0, + 275.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 233.0, + 664.0, + 233.0, + 664.0, + 262.0, + 608.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 233.0, + 765.0, + 233.0, + 765.0, + 262.0, + 736.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 268.0, + 290.0, + 268.0, + 290.0, + 288.0, + 275.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 253.0, + 1068.0, + 253.0, + 1068.0, + 294.0, + 324.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 293.0, + 291.0, + 293.0, + 291.0, + 316.0, + 272.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 288.0, + 517.0, + 288.0, + 517.0, + 316.0, + 325.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 288.0, + 895.0, + 288.0, + 895.0, + 316.0, + 546.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 322.0, + 291.0, + 322.0, + 291.0, + 343.0, + 272.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 316.0, + 740.0, + 316.0, + 740.0, + 347.0, + 325.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 316.0, + 1164.0, + 316.0, + 1164.0, + 347.0, + 770.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 349.0, + 291.0, + 349.0, + 291.0, + 372.0, + 272.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 341.0, + 712.0, + 341.0, + 712.0, + 376.0, + 323.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 341.0, + 1167.0, + 341.0, + 1167.0, + 376.0, + 770.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 376.0, + 291.0, + 376.0, + 291.0, + 398.0, + 272.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 368.0, + 636.0, + 368.0, + 636.0, + 403.0, + 322.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 368.0, + 1305.0, + 368.0, + 1305.0, + 403.0, + 665.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 405.0, + 291.0, + 405.0, + 291.0, + 428.0, + 267.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 394.0, + 577.0, + 394.0, + 577.0, + 432.0, + 322.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 394.0, + 751.0, + 394.0, + 751.0, + 432.0, + 605.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 267.0, + 433.0, + 289.0, + 433.0, + 289.0, + 455.0, + 267.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 431.0, + 611.0, + 431.0, + 611.0, + 455.0, + 327.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 431.0, + 810.0, + 431.0, + 810.0, + 455.0, + 634.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 457.0, + 293.0, + 457.0, + 293.0, + 484.0, + 266.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 456.0, + 501.0, + 456.0, + 501.0, + 481.0, + 323.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 456.0, + 797.0, + 456.0, + 797.0, + 481.0, + 530.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 486.0, + 293.0, + 486.0, + 293.0, + 512.0, + 265.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 514.0, + 293.0, + 514.0, + 293.0, + 540.0, + 265.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 510.0, + 503.0, + 510.0, + 503.0, + 541.0, + 325.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 510.0, + 1085.0, + 510.0, + 1085.0, + 541.0, + 531.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 510.0, + 1239.0, + 510.0, + 1239.0, + 541.0, + 1123.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 510.0, + 1389.0, + 510.0, + 1389.0, + 541.0, + 1257.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 541.0, + 293.0, + 541.0, + 293.0, + 566.0, + 266.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 537.0, + 516.0, + 537.0, + 516.0, + 566.0, + 322.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 537.0, + 694.0, + 537.0, + 694.0, + 566.0, + 546.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 266.0, + 569.0, + 293.0, + 569.0, + 293.0, + 595.0, + 266.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 565.0, + 866.0, + 565.0, + 866.0, + 596.0, + 389.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 565.0, + 1300.0, + 565.0, + 1300.0, + 596.0, + 884.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 565.0, + 1360.0, + 565.0, + 1360.0, + 596.0, + 1348.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 596.0, + 291.0, + 596.0, + 291.0, + 623.0, + 265.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 591.0, + 501.0, + 591.0, + 501.0, + 625.0, + 321.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 591.0, + 1213.0, + 591.0, + 1213.0, + 625.0, + 531.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 625.0, + 293.0, + 625.0, + 293.0, + 650.0, + 265.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 265.0, + 651.0, + 291.0, + 651.0, + 291.0, + 678.0, + 265.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 648.0, + 550.0, + 648.0, + 550.0, + 679.0, + 323.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 648.0, + 649.0, + 648.0, + 649.0, + 679.0, + 574.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 680.0, + 293.0, + 680.0, + 293.0, + 705.0, + 263.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 675.0, + 607.0, + 675.0, + 607.0, + 707.0, + 325.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 675.0, + 973.0, + 675.0, + 973.0, + 707.0, + 635.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 708.0, + 290.0, + 708.0, + 290.0, + 733.0, + 263.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 703.0, + 516.0, + 703.0, + 516.0, + 735.0, + 321.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 703.0, + 1182.0, + 703.0, + 1182.0, + 735.0, + 546.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 734.0, + 291.0, + 734.0, + 291.0, + 760.0, + 263.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 732.0, + 504.0, + 732.0, + 504.0, + 760.0, + 322.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 732.0, + 737.0, + 732.0, + 737.0, + 760.0, + 545.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 763.0, + 291.0, + 763.0, + 291.0, + 789.0, + 263.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 757.0, + 577.0, + 757.0, + 577.0, + 792.0, + 323.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 757.0, + 1192.0, + 757.0, + 1192.0, + 792.0, + 605.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 757.0, + 1301.0, + 757.0, + 1301.0, + 792.0, + 1215.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 791.0, + 293.0, + 791.0, + 293.0, + 816.0, + 263.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 787.0, + 505.0, + 787.0, + 505.0, + 818.0, + 322.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 787.0, + 977.0, + 787.0, + 977.0, + 818.0, + 546.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 816.0, + 294.0, + 816.0, + 294.0, + 846.0, + 262.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 814.0, + 1060.0, + 814.0, + 1060.0, + 846.0, + 385.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 846.0, + 293.0, + 846.0, + 293.0, + 872.0, + 263.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 873.0, + 291.0, + 873.0, + 291.0, + 900.0, + 263.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 869.0, + 502.0, + 869.0, + 502.0, + 900.0, + 323.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 869.0, + 1164.0, + 869.0, + 1164.0, + 900.0, + 531.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 869.0, + 1210.0, + 869.0, + 1210.0, + 900.0, + 1198.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 262.0, + 901.0, + 293.0, + 901.0, + 293.0, + 926.0, + 262.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 897.0, + 1242.0, + 897.0, + 1242.0, + 928.0, + 325.0, + 928.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 29, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/r1WUqIceg/r1WUqIceg.md b/parse/train/r1WUqIceg/r1WUqIceg.md new file mode 100644 index 0000000000000000000000000000000000000000..bfe80cd4d1948d1369e984f30cd9a92d7c700309 --- /dev/null +++ b/parse/train/r1WUqIceg/r1WUqIceg.md @@ -0,0 +1,131 @@ +# IMPROVING STOCHASTIC GRADIENT DESCENT WITH FEEDBACK + +Jayanth Koushik & Hiroaki Hayashi ∗ + +Language Technologies Institute +Carnegie Mellon University +Pittsburgh, PA 15213, USA +{jkoushik,hiroakih}@cs.cmu.edu + +# ABSTRACT + +In this paper we propose a simple and efficient method for improving stochastic gradient descent methods by using feedback from the objective function. The method tracks the relative changes in the objective function with a running average, and uses it to adaptively tune the learning rate in stochastic gradient descent. We specifically apply this idea to modify Adam, a popular algorithm for training deep neural networks. We conduct experiments to compare the resulting algorithm, which we call Eve, with state of the art methods used for training deep learning models. We train CNNs for image classification, and RNNs for language modeling and question answering. Our experiments show that Eve outperforms all other algorithms on these benchmark tasks. We also analyze the behavior of the feedback mechanism during the training process. + +# 1 INTRODUCTION + +Despite several breakthrough results in the last few years, the training of deep learning models remains a challenging problem. This training is a complex, high-dimensional, non-convex, stochastic optimization problem which is not amenable to many standard methods. Currently, the most common approach is to use some variant of stochastic gradient descent. Many extensions have been proposed to the basic gradient descent algorithm - designed to handle specific issues in the training of deep learning models. We review some of these methods in the next section. + +Although variants of simple stochastic gradient descent work quite well in practice, there is still room for improvement. This is easily evidenced by the existence of numerous methods to simplify the optimization problem itself like weight initialization techniques and normalization methods. + +In this work, we seek to improve stochastic gradient descent with a simple method that incorporates feedback from the objective function. The relative changes in the objective function indicate progress of the optimization algorithm. Our main hypothesis is that incorporating information about this change into the optimization algorithm can lead to improved performance - quantified in terms of the progress rate. We keep a running average of the relative changes in the objective function and use it to divide the learning rate. When the average relative change is high, the learning rate is reduced. This can improve the progress if, for example, the algorithm is bouncing around the walls of the objective function. Conversely, when the relative change is low, the learning rate is increased. This can help the algorithm accelerate through flat areas in the loss surface. As we discuss in the next section, such “plateaus” pose a significant challenge for first order methods and can create the illusion of local minima. + +While our method is general i.e. independent of any particular optimization algorithm, in this work we specifically apply the method to modify Adam (Kingma & Ba, 2014), considered to be the state of the art for training deep learning models. We call the resulting algorithm Eve and design experiments to compare it with Adam, as well as other popular methods from the literature. + +The paper is organized as follows. In Section 2, we review recent results related to the optimization of deep neural networks. We also discuss some popular algorithms and their motivations. Our general method, and the specific algorithm Eve are discussed in Section 3. Then in Section 4, we show that Eve consistently outperforms other methods in training convolutional neural networks (CNNs), and recurrent neural networks (RNNs). We also look in some detail, at the behavior of our method in the simple case of convex non-stochastic optimization. Finally we conclude in Section 5. + +# 2 RELATED WORK + +There has been considerable effort to understand the challenges in deep learning optimization. Intuitively, it seems that the non-convex optimization is made difficult by the presence of several poor local optima. However, this geometric intuition proves to be inadequate in reasoning about the high-dimensional case that arises with training deep learning models. Various empirical and theoretical results (Bray & Dean, 2007; Dauphin et al., 2014) have indicated that the problem in high dimensions arises not from local minima, but rather from saddle points. Moreover, a recent paper (Kawaguchi, 2016) proved (for deep linear networks, and under reasonable assumptions, also for deep non-linear networks) that all local minima in fact achieve the same value, and are optimal. The work also showed that all critical points which are not global minima are saddle points. Saddle points can seriously hamper the progress of both first and second order methods. Second order methods like Newton’s method are actually attracted to saddle points and are not suitable for high dimensional non-convex optimization. First order methods can escape from saddle points by following directions of negative curvature. However, such saddle points are usually surrounded by regions of small curvature - plateaus. This makes first order methods very slow near saddle points and can create the illusion of a local minimum. + +To tackle the saddle point problem, Dauphin et al. (2014) propose a second order method that fixes the issue with Newton’s method. Their algorithm builds on considering the behavior of Newton’s method near saddle points. Newton’s method rescales gradients in each eigen-direction with the corresponding inverse eigenvalue. However, near a saddle point, negative eigenvalues can cause the method to move towards the saddle point. Based on this observation, the authors propose using the absolute values of the eigenvalues to rescale the gradients. This saddle-free Newton method is backed by theoretical justifications and empirical results; however due to the computational requirements, second order methods are not very suitable for training large scale models. So we do not compare with such approaches in this work. + +We instead focus on first order methods which only rely on the gradient information. A key issue in training deep learning models is that of sparse gradients. To handle this, Adagrad (Duchi et al., 2011) adaptively changes the learning rate for each parameter, performing larger updates for infrequently updated parameters. However its update rule causes the learning rate to monotonically decrease, which eventually stalls the algorithm. Adadelta (Zeiler, 2012) and RMSProp (Tieleman & Hinton, 2012) are two extensions that try to fix this issue. Finally, a closely related method, and the base for our algorithm Eve (introduced in the next section), is Adam (Kingma & Ba, 2014). Adam incorporates the advantages of both Adagrad and RMSProp - and it has been found to work quite well in practice. Adam uses a running average of the gradient to determine the direction of descent, and scales the learning rate with a running average of the gradient squared. The authors of Adam also propose an extension based on the infinity norm, called Adamax. In our experiments, we compare Eve with both Adam and Adamax. + +# 3 METHOD + +# 3.1 ASSUMPTION + +We do need to make an assumption about the objective function $f$ . We assume that the minimum value of $f$ over its domain is known. While this is true for loss functions encountered in machine learning (like mean squared error or cross entropy), it does not hold if the objective function also includes regularization terms $\boldsymbol { L } _ { 1 }$ , $L _ { 2 }$ etc.). In all our experiments, we used dropout for regularization which is not affected by this assumption. Finally, to simplify notation we assume that the minimum has been subtracted from the objective function i.e. the minimum has been made 0. + +Algorithm 1 Eve: Adam with feedback. Parameters carried over from Adam have the same default values: $\alpha = 1 0 ^ { - 3 }$ , $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 9 9$ , $\epsilon = 1 0 ^ { - 8 }$ . For parameters specific to our method, we recommend default values $\beta _ { 3 } = 0 . 9 9 9$ , $k = 0 . 1$ , $K = 1 0$ . Wherever applicable, products are elementwise. + +
Require: α: learning rate
Require: β1, β2 ∈ [0,1): exponential decay rates for moment estimation in Adam
Require: β3 ∈ [0,1): exponential decay rate for computing relative change
Require: k,K: lower and upper threshold for relative change
Require: ε: fuzz factor
Require: f(θ): objective function
Require: Oo: initial value for parameters
mo = vo =0
do=1
f-1=t=0
while stopping condition is not reached do
t←t+1
gt←Vθf(0t-1)
mt←βimt-1+(1-β1)gt mt
mt←(1-β)
Ut←β2Ut-1+(1-β2)g² Ut
Ut←(1-)
if t>1 then
if f(0t-1) ≥ft-2 then δt←k+1
△t←K+1
else 1
8t←K+
△t←
end if
ct ←min{max{δt, }}
Ft-1←ctft-2
Ft-1-ft-21 rt↑
min{ft-1,ft-2} dt←βgdt-1+(1-β)rt
else
ft-1←f(0t-1)
dt←1
end if
0t←0t-1-Qdv+e mt
+ +# 3.2 ALGORITHM + +The main component of our proposed method is a feedback term that captures the relative change in the objective value. Let $f _ { t - 2 }$ and $f _ { t - 1 }$ denote the values of the objective function at time steps $t - 2$ and t − 1 respectively. Then this change is computed as rt = ft−2−ft−1f $\begin{array} { r } { r _ { t } = { \frac { f _ { t - 2 } - \bar { \cdot } f _ { t - 1 } } { f _ { t - 1 } } } = { \frac { f _ { t - 2 } - f _ { t - 1 } } { f _ { t - 1 } } } } \end{array}$ ft−2−ft−1 if ft−2 ft−1, an d ft−1−ft−2f otherwise. Note that this value is always non-negative but it can be less than or greater than 1 i.e. it captures both relative increase and decrease. We compute a running average using these relative changes to get a smoother estimate. Specifically, we take $d _ { 1 } = 1$ , and for $t > 1$ define $d _ { t }$ as $d _ { t } = \beta d _ { t - 1 } + ( 1 - \beta ) r _ { t }$ . Here $\beta \in [ 0 , 1 )$ is a decay rate - large values correspond to a slowly changing $d _ { t }$ , and vice versa. This simple expression can, however, blow up and lead to instability. To handle this issue, we use a thresholding scheme. A simple thing to do would be to clip $d _ { t }$ as $\operatorname* { m i n } \{ \operatorname* { m a x } \{ k , d _ { t } \} , K \}$ for some suitable $0 < k < K$ . But we found this to not work very well in practice due to the abrupt nature of the clipping. Instead we indirectly clip $d _ { t }$ by smoothly tracking the objective function. Let $\widehat { f } _ { t - 2 }$ be the value of the smoothly tracked objective function at time $t - 2$ with ${ \widehat { f } } _ { 0 } = f _ { 0 }$ . For now, assume $f _ { t - 1 } \leq \widehat { f } _ { t - 2 }$ . We would like to have $k \leq r _ { t } \leq K$ which in this case requires $\begin{array} { r } { \frac { f _ { t - 1 } } { \widehat { f } _ { t - 2 } } \in \left[ \frac { 1 } { K + 1 } , \frac { 1 } { k + 1 } \right] } \end{array}$ . So we compute $\begin{array} { r } { c _ { t } = \operatorname* { m i n } \left\{ \operatorname* { m a x } \left\{ \frac { 1 } { K + 1 } , \frac { f _ { t } } { \widehat { f } _ { t - 1 } } \right\} , \frac { 1 } { k + 1 } \right\} } \end{array}$ and set $\widehat { f } _ { t - 1 } = c _ { t } \widehat { f } _ { t - 2 }$ . Finally $r _ { t }$ is $\frac { \widehat { f } _ { t - 2 } - \widehat { f } _ { t - 1 } } { \widehat { f } _ { t - 1 } }$ . Analogous expressions can also be derived for the case when $f _ { t - 1 } > \widehat { f } _ { t - 2 }$ . This smooth tracking has the additional advantage of making $d _ { t }$ less susceptible to the high variability that comes with training using minibatches. + +![](images/f97087fe23db1dc3a650ad1e44d6c838538276740c378a42457561cffa200ed7.jpg) +Figure 1: Training loss for convolutional neural networks on CIFAR10 and CIFAR100. The vertical axis is shown in log scale. In both cases, our proposed method achieves the best performance. + +Once $d _ { t }$ has been computed, it can be used to modify any gradient descent algorithm by modifying the learning rate $\alpha$ as $\alpha _ { t } = \alpha / d _ { t }$ . Large values of $d _ { t }$ , caused by large changes in the objective function will lead to a smaller effective learning rate. Similarly, small values of $d _ { t }$ will lead to a larger effective learning rate. Since we start with $d _ { 0 } = 1$ , the initial updates will closely follow that of the base algorithm. In the next section, we will look at how $d _ { t }$ evolves during the course of an experiment to get a better understanding of how it affects the training. + +We note again that our method is independent of any particular gradient descent algorithm. However, for this current work, we specifically focus on applying the method to Adam (Kingma & Ba, 2014). This modified algorithm, which we call Eve, is shown in Algorithm 1. We modify the final Adam update by multiplying the denominator $\sqrt { v _ { t } }$ with $d _ { t }$ . In addition to the hyperparameters in Adam, we introduce 3 new hyperparameters $\beta _ { 3 }$ , $k$ , and $K$ . In all our experiments we use the values $\beta _ { 3 } = 0 . 9 9 9$ , $k = 0 . 1$ , and $K = 1 0$ , which we found to work well in practice. + +# 4 EXPERIMENTS + +Now we evaluate our proposed method by comparing Eve with several state of the art algorithms for optimizing deep learning models.1. In all experiments, we used ReLU activations, and initialized weights according to the scheme proposed by Glorot & Bengio (2010). We used minibatches of size 128, and linear decay for the learning rate: $\alpha _ { t } = \alpha / ( 1 + \gamma t )$ $( \gamma$ is the decay rate, picked by searching over a range of values). + +In the figures, SGD refers to vanilla stochastic gradient descent, and SGD Nesterov refers to stochastic gradient descent with Nesterov momentum (Nesterov, 1983) where we set the momentum to 0.9 in all experiments. + +# 4.1 CONVOLUTIONAL NEURAL NETWORKS + +We first trained a 5 layer convolutional neural network for 10-way classification of images from the CIFAR10 dataset (Krizhevsky & Hinton, 2009). The model consisted of 2 blocks of $3 \mathrm { x } 3$ convolutional layers each followed by $2 \mathbf { x } 2$ max pooling and 0.25 dropout (Srivastava et al., 2014). The first block contained 2 layers with 32 filters each, and the second block contained 2 layers with 64 filters each. The convolutional layers were followed by a fully connected layer with 512 units and a 10-way softmax layer. We trained this model for 500 epochs on the training split using various popular methods for training deep learning models, as well as Eve. For each algorithm, we tried learning rates $\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \}$ (for algorithms with suggested default learning rates, we also included them in the search), learning rate decays $\{ 0 , 1 0 ^ { - 2 } , 1 \breve { 0 } ^ { - 3 } , 1 0 ^ { - 4 } \}$ , and picked the pair of values that led to the smallest final training loss. The loss curves are shown in Figure 1a. Eve quickly surpasses all other methods and achieves the lowest final training loss. In the next section we will look at the behavior of the adaptive coefficient $d _ { t }$ to gain some intuition behind this improved performance. + +![](images/f3ff840fdc66fddf215a2fadd32bb8c3e7b6c2e863f57aeacf8c8f8ff6545987.jpg) +Figure 2: Behavior of the tuning coefficient $d _ { t }$ during the experiment shown in Figure 1a. There is an overall trend of acceleration followed by decay, but also more fine-grained behavior as indicated by the bottom-right plot. + +![](images/e2d88fee7e37f8eda5b9a8ded593e0c30bd16264ba4f39e6b5079d4508648bf8.jpg) +Figure 3: Minibatch losses for Eve during the CNN experiment on CIFAR10 (Figure 1a). The variance in the losses increases throughout the training. + +We also trained a larger CNN model using the top-performing algorithms from the previous experiment. This model consisted of 3 blocks of $3 \mathrm { x } 3$ convolutional layers (3 layers per block, and 64, 128, 256 filters per layer in the first, second, and third block respectively) each followed by $2 \mathrm { x } 2 \mathrm { m a x }$ pooling and 0.5 dropout. Then we had 2 fully connected layers with 512 units, each followed by 0.5 dropout, and finally a 100-way softmax. We trained this model on the CIFAR100 (Krizhevsky & Hinton, 2009) dataset for 100 epochs. We again performed a grid search over the same learning rate and decay values as the last experiment. The results are shown in Figure 1b, and once again show that our proposed method improves over state of the art methods for training convolutional neural networks. + +![](images/27f3a9518a4468af14eaa22a22c033deb4cbfe6c54a5d76cf153c1895cc83be7.jpg) +Figure 4: Loss curves and tuning coefficient $d _ { t }$ for batch gradient descent training of a logistic regression model. For Eve, $d _ { t }$ continuously decreases and converges to the lower threshold 0.1. + +![](images/4b9980bf2139c08ee969e7d5bfc125a1ed4d0abbb8ea01e97cfe2b928fe37f6f.jpg) +Figure 5: The left plot shows Adam with different learning rates plotted with Eve (learning rate $1 0 ^ { - 2 }$ ). Although Adam with a larger learning rate can be almost identical to Eve, this is largely dependent on the initial values for Adam as shown in the plot on the right. + +# 4.2 ANALYSIS OF TUNING COEFFICIENT + +Before we consider the next set of experiments on recurrent neural networks, we will first look more closely at the behavior of the tuning coefficient $d _ { t }$ in our algorithm. We will specifically consider the results from the CNN experiment on CIFAR10. Figure 2 shows the progress of $d _ { t }$ throughout the training, and also in two smaller windows. A few things are worth noting here. First is that of the overall trend. There is an initial acceleration followed by a decay. This initial acceleration allows Eve to rapidly overtake other methods, and makes it proceed at a faster pace for about 100 epochs. This acceleration is not equivalent to simply starting with a larger learning rate - in all our experiments we search over a range of learning rate values. The overall trend for $d _ { t }$ can be explained by looking at the minibatch losses at each iteration (as opposed to the loss computed over the entire dataset after each epoch) in Figure 3. Initially, different minibatches achieve similar loss values which leads to $d _ { t }$ decreasing. But as training proceeds, the variance in the minibatch losses increases and $d _ { t }$ eventually increases. However, this overall trend does not capture the complete picture - for example, as shown in the bottom right plot of Figure 2, $d _ { t }$ can actually be decreasing in some regions of the training, adjusting to local structures in the error surface. + +To further study the observed acceleration, and to also motivate the need for clipping, we consider a simpler experiment. We trained a logistic regression model on 1000 images from the MNIST dataset. We used batch gradient descent for training i.e. all 1000 samples were used for computing the gradient at each step. We trained this model using Eve, Adam, Adamax, and SGD Nesterov for 10000 iterations, searching over a large range of values for the learning rate and decay: $\alpha \in$ $\{ 1 0 ^ { - 1 } , 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } \}$ , $\gamma \in \{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } , 1 0 ^ { - 6 } , 0 \}$ . The results are shown in Figure 4. Eve again outperforms all other methods and achieves the lowest training loss. Also, since this is a smooth non-stochastic problem, the tuning coefficient $d _ { t }$ continuously decreases - this makes having a thresholding mechanism important since the learning rate would blow up otherwise. + +![](images/92c46efb82e34367ee757f4c1af8c5474272a094da2588de2dd44fab1ac2d3c6.jpg) +(c) Penn Treebank, Language modeling +Figure 6: Loss curves for experiments with recurrent neural networks. Eve consistently achieves better performance than other methods. + +Although in the previous experiment the effect of our method is to increase the learning rate, it is not equivalent to simply starting with a larger learning rate. We will establish this with a couple simple experiments. First we note that in the previous experiment, the optimal decay rates for both Adam and Eve were 0 - no decay. The optimal learning rate for Eve was $\mathrm { \bar { 1 0 } ^ { - 2 } }$ . Since the tuning coefficient converges to 0.1, we trained Adam using no decay, and learning rates $i \times 1 0 ^ { - 2 }$ where $i$ varies from 1 to 10. The training loss curves are shown in the left plot of Figure 5. While increasing the learning rate with Adam does seem to close the gap with Eve, Eve does remain marginally faster. Moreover, and more importantly, this first plot represents the best-case situation for Adam. With larger learning rates, training becomes increasingly unstable and sensitive to the initial values of the parameters. This is illustrated in the right plot of Figure 5 where we used Eve (with learning rate $1 0 ^ { - 2 }$ ) and Adam (with learning rate $1 0 ^ { - 1 }$ ) 10 times with different random initializations. In some cases, Adam fails to converge whereas Eve always converges - even though Eve eventually reaches a learning rate of $1 0 ^ { - 1 }$ . This is because very early in the training, the model is quite sensitive at higher learning rates due to larger gradients. Depending on the initial values, the algorithm may or may not converge. So it is advantageous to slowly accelerate as the learning stabilizes rather than start with a larger learning rate. + +# 4.3 RECURRENT NEURAL NETWORKS + +Finally, we evaluated our method on recurrent neural networks (RNNs). We first trained a RNN for character-level language modeling on the Penn Treebank dataset (Marcus et al., 1993). Specifically, the model consisted of a 2-layer character-level Gated Recurrent Unit (Chung et al., 2014) with hidden layers of size 256, 0.5 dropout between layers, and sequences of 100 characters. We adopted $1 0 ^ { - 3 }$ as the initial learning rate for Adam, Eve, and RMSProp. For Adamax, we used $2 \times 1 0 ^ { - 3 }$ as the learning rate since it is the suggested value. We used $3 \times 1 0 ^ { - 4 }$ for the learning rate decay. We trained this model for 100 epochs using each of the algorithms. The results, plotted in Figure 6c, clearly show that our method achieves the best results. Eve optimizes the model to a lower final loss than the other models. + +We trained another RNN-based model for the question & answering task. Specifically, we chose two question types among 20 types from the bAbI dataset (Weston et al., 2015), Q19 and Q14. The dataset consists of pairs of supporting story sentences and a question. Different types of pairs are said to require different reasoning schemes. For our test case, Q19 and Q14 correspond to Path Finding and Time Reasoning respectively. We picked Q19 since it is reported to have the lowest baseline score, and we picked Q14 randomly from the remaining questions. The model consisted of two parts, one for encoding story sentences and another for query. Both included an embedding layer with 256 hidden units, and 0.3 dropout. Next query word embeddings were fed into a GRU one token at a time, to compute a sentence representation. Both story and query sequences were truncated to the maximum sequence length of 100. Finally, the sequence of word embeddings from story sentences and the repeated encoded representation of a query were combined together to serve as input for each time step in another GRU, with 0.3 dropout. We searched for the learning rate and decay from a range of values, $\alpha \in \{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \}$ and $\gamma \in \{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 0 \}$ . The results, shown in Figures 6a, and 6b show that Eve again improves over all other methods. + +# 5 CONCLUSION + +We proposed a simple and efficient method for incorporating feedback in to stochastic gradient descent algorithms. We used this method to create Eve, a modified version of the Adam algorithm. Experiments with a variety of models showed that the proposed method can help improve the optimization of deep learning models. + +For future work, we would look to theoretically analyze our method and its effects. While we have tried to evaluate our algorithm Eve on a variety of tasks, additional experiments on larger scale problems would further highlight the strength of our approach. We are making code for our method and the experiments publicly available to encourage more research on this method. + +# REFERENCES + +Alan J Bray and David S Dean. Statistics of critical points of gaussian fields on large-dimensional spaces. Physical review letters, 98(15):150201, 2007. + +Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. + +Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In Advances in neural information processing systems, pp. 2933–2941, 2014. + +John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. + +Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Aistats, volume 9, pp. 249–256, 2010. + +Kenji Kawaguchi. Deep learning without poor local minima. In Advances in Neural Information Processing Systems (NIPS), 2016. to appear. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. + +Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993. + +Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence o (1/k2). In Doklady an SSSR, volume 269, pp. 543–547, 1983. + +Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014. + +Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4(2), 2012. + +Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015. + +Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012. \ No newline at end of file diff --git a/parse/train/r1WUqIceg/r1WUqIceg_content_list.json b/parse/train/r1WUqIceg/r1WUqIceg_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..204e22f0f344dfcfa1c7c9e5b525bbd629f86441 --- /dev/null +++ b/parse/train/r1WUqIceg/r1WUqIceg_content_list.json @@ -0,0 +1,769 @@ +[ + { + "type": "text", + "text": "IMPROVING STOCHASTIC GRADIENT DESCENT WITH FEEDBACK ", + "text_level": 1, + "bbox": [ + 174, + 98, + 735, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jayanth Koushik & Hiroaki Hayashi ∗ ", + "bbox": [ + 186, + 170, + 450, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language Technologies Institute \nCarnegie Mellon University \nPittsburgh, PA 15213, USA \n{jkoushik,hiroakih}@cs.cmu.edu ", + "bbox": [ + 184, + 185, + 480, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 276, + 544, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper we propose a simple and efficient method for improving stochastic gradient descent methods by using feedback from the objective function. The method tracks the relative changes in the objective function with a running average, and uses it to adaptively tune the learning rate in stochastic gradient descent. We specifically apply this idea to modify Adam, a popular algorithm for training deep neural networks. We conduct experiments to compare the resulting algorithm, which we call Eve, with state of the art methods used for training deep learning models. We train CNNs for image classification, and RNNs for language modeling and question answering. Our experiments show that Eve outperforms all other algorithms on these benchmark tasks. We also analyze the behavior of the feedback mechanism during the training process. ", + "bbox": [ + 233, + 305, + 766, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 482, + 336, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite several breakthrough results in the last few years, the training of deep learning models remains a challenging problem. This training is a complex, high-dimensional, non-convex, stochastic optimization problem which is not amenable to many standard methods. Currently, the most common approach is to use some variant of stochastic gradient descent. Many extensions have been proposed to the basic gradient descent algorithm - designed to handle specific issues in the training of deep learning models. We review some of these methods in the next section. ", + "bbox": [ + 174, + 512, + 825, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although variants of simple stochastic gradient descent work quite well in practice, there is still room for improvement. This is easily evidenced by the existence of numerous methods to simplify the optimization problem itself like weight initialization techniques and normalization methods. ", + "bbox": [ + 176, + 603, + 821, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we seek to improve stochastic gradient descent with a simple method that incorporates feedback from the objective function. The relative changes in the objective function indicate progress of the optimization algorithm. Our main hypothesis is that incorporating information about this change into the optimization algorithm can lead to improved performance - quantified in terms of the progress rate. We keep a running average of the relative changes in the objective function and use it to divide the learning rate. When the average relative change is high, the learning rate is reduced. This can improve the progress if, for example, the algorithm is bouncing around the walls of the objective function. Conversely, when the relative change is low, the learning rate is increased. This can help the algorithm accelerate through flat areas in the loss surface. As we discuss in the next section, such “plateaus” pose a significant challenge for first order methods and can create the illusion of local minima. ", + "bbox": [ + 174, + 652, + 825, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While our method is general i.e. independent of any particular optimization algorithm, in this work we specifically apply the method to modify Adam (Kingma & Ba, 2014), considered to be the state of the art for training deep learning models. We call the resulting algorithm Eve and design experiments to compare it with Adam, as well as other popular methods from the literature. ", + "bbox": [ + 174, + 811, + 825, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The paper is organized as follows. In Section 2, we review recent results related to the optimization of deep neural networks. We also discuss some popular algorithms and their motivations. Our general method, and the specific algorithm Eve are discussed in Section 3. Then in Section 4, we show that Eve consistently outperforms other methods in training convolutional neural networks (CNNs), and recurrent neural networks (RNNs). We also look in some detail, at the behavior of our method in the simple case of convex non-stochastic optimization. Finally we conclude in Section 5. ", + "bbox": [ + 174, + 875, + 823, + 904 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 189, + 343, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There has been considerable effort to understand the challenges in deep learning optimization. Intuitively, it seems that the non-convex optimization is made difficult by the presence of several poor local optima. However, this geometric intuition proves to be inadequate in reasoning about the high-dimensional case that arises with training deep learning models. Various empirical and theoretical results (Bray & Dean, 2007; Dauphin et al., 2014) have indicated that the problem in high dimensions arises not from local minima, but rather from saddle points. Moreover, a recent paper (Kawaguchi, 2016) proved (for deep linear networks, and under reasonable assumptions, also for deep non-linear networks) that all local minima in fact achieve the same value, and are optimal. The work also showed that all critical points which are not global minima are saddle points. Saddle points can seriously hamper the progress of both first and second order methods. Second order methods like Newton’s method are actually attracted to saddle points and are not suitable for high dimensional non-convex optimization. First order methods can escape from saddle points by following directions of negative curvature. However, such saddle points are usually surrounded by regions of small curvature - plateaus. This makes first order methods very slow near saddle points and can create the illusion of a local minimum. ", + "bbox": [ + 174, + 228, + 825, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the saddle point problem, Dauphin et al. (2014) propose a second order method that fixes the issue with Newton’s method. Their algorithm builds on considering the behavior of Newton’s method near saddle points. Newton’s method rescales gradients in each eigen-direction with the corresponding inverse eigenvalue. However, near a saddle point, negative eigenvalues can cause the method to move towards the saddle point. Based on this observation, the authors propose using the absolute values of the eigenvalues to rescale the gradients. This saddle-free Newton method is backed by theoretical justifications and empirical results; however due to the computational requirements, second order methods are not very suitable for training large scale models. So we do not compare with such approaches in this work. ", + "bbox": [ + 174, + 443, + 825, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We instead focus on first order methods which only rely on the gradient information. A key issue in training deep learning models is that of sparse gradients. To handle this, Adagrad (Duchi et al., 2011) adaptively changes the learning rate for each parameter, performing larger updates for infrequently updated parameters. However its update rule causes the learning rate to monotonically decrease, which eventually stalls the algorithm. Adadelta (Zeiler, 2012) and RMSProp (Tieleman & Hinton, 2012) are two extensions that try to fix this issue. Finally, a closely related method, and the base for our algorithm Eve (introduced in the next section), is Adam (Kingma & Ba, 2014). Adam incorporates the advantages of both Adagrad and RMSProp - and it has been found to work quite well in practice. Adam uses a running average of the gradient to determine the direction of descent, and scales the learning rate with a running average of the gradient squared. The authors of Adam also propose an extension based on the infinity norm, called Adamax. In our experiments, we compare Eve with both Adam and Adamax. ", + "bbox": [ + 174, + 574, + 825, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 METHOD ", + "text_level": 1, + "bbox": [ + 176, + 772, + 281, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 ASSUMPTION ", + "text_level": 1, + "bbox": [ + 176, + 809, + 308, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We do need to make an assumption about the objective function $f$ . We assume that the minimum value of $f$ over its domain is known. While this is true for loss functions encountered in machine learning (like mean squared error or cross entropy), it does not hold if the objective function also includes regularization terms $\\boldsymbol { L } _ { 1 }$ , $L _ { 2 }$ etc.). In all our experiments, we used dropout for regularization which is not affected by this assumption. Finally, to simplify notation we assume that the minimum has been subtracted from the objective function i.e. the minimum has been made 0. ", + "bbox": [ + 174, + 840, + 825, + 922 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Algorithm 1 Eve: Adam with feedback. Parameters carried over from Adam have the same default values: $\\alpha = 1 0 ^ { - 3 }$ , $\\beta _ { 1 } = 0 . 9$ , $\\beta _ { 2 } = 0 . 9 9 9$ , $\\epsilon = 1 0 ^ { - 8 }$ . For parameters specific to our method, we recommend default values $\\beta _ { 3 } = 0 . 9 9 9$ , $k = 0 . 1$ , $K = 1 0$ . Wherever applicable, products are elementwise. ", + "bbox": [ + 173, + 102, + 825, + 159 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/5e45f59f6a66111f20a95e09b6aafc0193a84735fa47a785faf32d8eae5fb68c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Require: α: learning rate
Require: β1, β2 ∈ [0,1): exponential decay rates for moment estimation in Adam
Require: β3 ∈ [0,1): exponential decay rate for computing relative change
Require: k,K: lower and upper threshold for relative change
Require: ε: fuzz factor
Require: f(θ): objective function
Require: Oo: initial value for parameters
mo = vo =0
do=1
f-1=t=0
while stopping condition is not reached do
t←t+1
gt←Vθf(0t-1)
mt←βimt-1+(1-β1)gt mt
mt←(1-β)
Ut←β2Ut-1+(1-β2)g² Ut
Ut←(1-)
if t>1 then
if f(0t-1) ≥ft-2 then δt←k+1
△t←K+1
else 1
8t←K+
△t←
end if
ct ←min{max{δt, }}
Ft-1←ctft-2
Ft-1-ft-21 rt↑
min{ft-1,ft-2} dt←βgdt-1+(1-β)rt
else
ft-1←f(0t-1)
dt←1
end if
0t←0t-1-Qdv+e mt
", + "bbox": [ + 168, + 156, + 820, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 732, + 303, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main component of our proposed method is a feedback term that captures the relative change in the objective value. Let $f _ { t - 2 }$ and $f _ { t - 1 }$ denote the values of the objective function at time steps $t - 2$ and t − 1 respectively. Then this change is computed as rt = ft−2−ft−1f $\\begin{array} { r } { r _ { t } = { \\frac { f _ { t - 2 } - \\bar { \\cdot } f _ { t - 1 } } { f _ { t - 1 } } } = { \\frac { f _ { t - 2 } - f _ { t - 1 } } { f _ { t - 1 } } } } \\end{array}$ ft−2−ft−1 if ft−2 ft−1, an d ft−1−ft−2f otherwise. Note that this value is always non-negative but it can be less than or greater than 1 i.e. it captures both relative increase and decrease. We compute a running average using these relative changes to get a smoother estimate. Specifically, we take $d _ { 1 } = 1$ , and for $t > 1$ define $d _ { t }$ as $d _ { t } = \\beta d _ { t - 1 } + ( 1 - \\beta ) r _ { t }$ . Here $\\beta \\in [ 0 , 1 )$ is a decay rate - large values correspond to a slowly changing $d _ { t }$ , and vice versa. This simple expression can, however, blow up and lead to instability. To handle this issue, we use a thresholding scheme. A simple thing to do would be to clip $d _ { t }$ as $\\operatorname* { m i n } \\{ \\operatorname* { m a x } \\{ k , d _ { t } \\} , K \\}$ for some suitable $0 < k < K$ . But we found this to not work very well in practice due to the abrupt nature of the clipping. Instead we indirectly clip $d _ { t }$ by smoothly tracking the objective function. Let $\\widehat { f } _ { t - 2 }$ be the value of the smoothly tracked objective function at time $t - 2$ with ${ \\widehat { f } } _ { 0 } = f _ { 0 }$ . For now, assume $f _ { t - 1 } \\leq \\widehat { f } _ { t - 2 }$ . We would like to have $k \\leq r _ { t } \\leq K$ which in this case requires $\\begin{array} { r } { \\frac { f _ { t - 1 } } { \\widehat { f } _ { t - 2 } } \\in \\left[ \\frac { 1 } { K + 1 } , \\frac { 1 } { k + 1 } \\right] } \\end{array}$ . So we compute $\\begin{array} { r } { c _ { t } = \\operatorname* { m i n } \\left\\{ \\operatorname* { m a x } \\left\\{ \\frac { 1 } { K + 1 } , \\frac { f _ { t } } { \\widehat { f } _ { t - 1 } } \\right\\} , \\frac { 1 } { k + 1 } \\right\\} } \\end{array}$ and set $\\widehat { f } _ { t - 1 } = c _ { t } \\widehat { f } _ { t - 2 }$ . Finally $r _ { t }$ is $\\frac { \\widehat { f } _ { t - 2 } - \\widehat { f } _ { t - 1 } } { \\widehat { f } _ { t - 1 } }$ . Analogous expressions can also be derived for the case when $f _ { t - 1 } > \\widehat { f } _ { t - 2 }$ . This smooth tracking has the additional advantage of making $d _ { t }$ less susceptible to the high variability that comes with training using minibatches. ", + "bbox": [ + 173, + 758, + 826, + 924 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f97087fe23db1dc3a650ad1e44d6c838538276740c378a42457561cffa200ed7.jpg", + "image_caption": [ + "Figure 1: Training loss for convolutional neural networks on CIFAR10 and CIFAR100. The vertical axis is shown in log scale. In both cases, our proposed method achieves the best performance. " + ], + "image_footnote": [], + "bbox": [ + 179, + 99, + 818, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 330, + 825, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once $d _ { t }$ has been computed, it can be used to modify any gradient descent algorithm by modifying the learning rate $\\alpha$ as $\\alpha _ { t } = \\alpha / d _ { t }$ . Large values of $d _ { t }$ , caused by large changes in the objective function will lead to a smaller effective learning rate. Similarly, small values of $d _ { t }$ will lead to a larger effective learning rate. Since we start with $d _ { 0 } = 1$ , the initial updates will closely follow that of the base algorithm. In the next section, we will look at how $d _ { t }$ evolves during the course of an experiment to get a better understanding of how it affects the training. ", + "bbox": [ + 173, + 450, + 825, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note again that our method is independent of any particular gradient descent algorithm. However, for this current work, we specifically focus on applying the method to Adam (Kingma & Ba, 2014). This modified algorithm, which we call Eve, is shown in Algorithm 1. We modify the final Adam update by multiplying the denominator $\\sqrt { v _ { t } }$ with $d _ { t }$ . In addition to the hyperparameters in Adam, we introduce 3 new hyperparameters $\\beta _ { 3 }$ , $k$ , and $K$ . In all our experiments we use the values $\\beta _ { 3 } = 0 . 9 9 9$ , $k = 0 . 1$ , and $K = 1 0$ , which we found to work well in practice. ", + "bbox": [ + 173, + 541, + 826, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 647, + 326, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now we evaluate our proposed method by comparing Eve with several state of the art algorithms for optimizing deep learning models.1. In all experiments, we used ReLU activations, and initialized weights according to the scheme proposed by Glorot & Bengio (2010). We used minibatches of size 128, and linear decay for the learning rate: $\\alpha _ { t } = \\alpha / ( 1 + \\gamma t )$ $( \\gamma$ is the decay rate, picked by searching over a range of values). ", + "bbox": [ + 174, + 679, + 825, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the figures, SGD refers to vanilla stochastic gradient descent, and SGD Nesterov refers to stochastic gradient descent with Nesterov momentum (Nesterov, 1983) where we set the momentum to 0.9 in all experiments. ", + "bbox": [ + 174, + 756, + 825, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 CONVOLUTIONAL NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 815, + 488, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first trained a 5 layer convolutional neural network for 10-way classification of images from the CIFAR10 dataset (Krizhevsky & Hinton, 2009). The model consisted of 2 blocks of $3 \\mathrm { x } 3$ convolutional layers each followed by $2 \\mathbf { x } 2$ max pooling and 0.25 dropout (Srivastava et al., 2014). The first block contained 2 layers with 32 filters each, and the second block contained 2 layers with 64 filters each. The convolutional layers were followed by a fully connected layer with 512 units and a 10-way softmax layer. We trained this model for 500 epochs on the training split using various popular methods for training deep learning models, as well as Eve. For each algorithm, we tried learning rates $\\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}$ (for algorithms with suggested default learning rates, we also included them in the search), learning rate decays $\\{ 0 , 1 0 ^ { - 2 } , 1 \\breve { 0 } ^ { - 3 } , 1 0 ^ { - 4 } \\}$ , and picked the pair of values that led to the smallest final training loss. The loss curves are shown in Figure 1a. Eve quickly surpasses all other methods and achieves the lowest final training loss. In the next section we will look at the behavior of the adaptive coefficient $d _ { t }$ to gain some intuition behind this improved performance. ", + "bbox": [ + 174, + 843, + 825, + 885 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3ff840fdc66fddf215a2fadd32bb8c3e7b6c2e863f57aeacf8c8f8ff6545987.jpg", + "image_caption": [ + "Figure 2: Behavior of the tuning coefficient $d _ { t }$ during the experiment shown in Figure 1a. There is an overall trend of acceleration followed by decay, but also more fine-grained behavior as indicated by the bottom-right plot. " + ], + "image_footnote": [], + "bbox": [ + 176, + 102, + 823, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e2d88fee7e37f8eda5b9a8ded593e0c30bd16264ba4f39e6b5079d4508648bf8.jpg", + "image_caption": [ + "Figure 3: Minibatch losses for Eve during the CNN experiment on CIFAR10 (Figure 1a). The variance in the losses increases throughout the training. " + ], + "image_footnote": [], + "bbox": [ + 313, + 397, + 683, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 665, + 825, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also trained a larger CNN model using the top-performing algorithms from the previous experiment. This model consisted of 3 blocks of $3 \\mathrm { x } 3$ convolutional layers (3 layers per block, and 64, 128, 256 filters per layer in the first, second, and third block respectively) each followed by $2 \\mathrm { x } 2 \\mathrm { m a x }$ pooling and 0.5 dropout. Then we had 2 fully connected layers with 512 units, each followed by 0.5 dropout, and finally a 100-way softmax. We trained this model on the CIFAR100 (Krizhevsky & Hinton, 2009) dataset for 100 epochs. We again performed a grid search over the same learning rate and decay values as the last experiment. The results are shown in Figure 1b, and once again show that our proposed method improves over state of the art methods for training convolutional neural networks. ", + "bbox": [ + 173, + 797, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27f3a9518a4468af14eaa22a22c033deb4cbfe6c54a5d76cf153c1895cc83be7.jpg", + "image_caption": [ + "Figure 4: Loss curves and tuning coefficient $d _ { t }$ for batch gradient descent training of a logistic regression model. For Eve, $d _ { t }$ continuously decreases and converges to the lower threshold 0.1. " + ], + "image_footnote": [], + "bbox": [ + 184, + 109, + 815, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b9980bf2139c08ee969e7d5bfc125a1ed4d0abbb8ea01e97cfe2b928fe37f6f.jpg", + "image_caption": [ + "Figure 5: The left plot shows Adam with different learning rates plotted with Eve (learning rate $1 0 ^ { - 2 }$ ). Although Adam with a larger learning rate can be almost identical to Eve, this is largely dependent on the initial values for Adam as shown in the plot on the right. " + ], + "image_footnote": [], + "bbox": [ + 184, + 363, + 810, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 ANALYSIS OF TUNING COEFFICIENT ", + "text_level": 1, + "bbox": [ + 178, + 640, + 465, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before we consider the next set of experiments on recurrent neural networks, we will first look more closely at the behavior of the tuning coefficient $d _ { t }$ in our algorithm. We will specifically consider the results from the CNN experiment on CIFAR10. Figure 2 shows the progress of $d _ { t }$ throughout the training, and also in two smaller windows. A few things are worth noting here. First is that of the overall trend. There is an initial acceleration followed by a decay. This initial acceleration allows Eve to rapidly overtake other methods, and makes it proceed at a faster pace for about 100 epochs. This acceleration is not equivalent to simply starting with a larger learning rate - in all our experiments we search over a range of learning rate values. The overall trend for $d _ { t }$ can be explained by looking at the minibatch losses at each iteration (as opposed to the loss computed over the entire dataset after each epoch) in Figure 3. Initially, different minibatches achieve similar loss values which leads to $d _ { t }$ decreasing. But as training proceeds, the variance in the minibatch losses increases and $d _ { t }$ eventually increases. However, this overall trend does not capture the complete picture - for example, as shown in the bottom right plot of Figure 2, $d _ { t }$ can actually be decreasing in some regions of the training, adjusting to local structures in the error surface. ", + "bbox": [ + 174, + 666, + 825, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further study the observed acceleration, and to also motivate the need for clipping, we consider a simpler experiment. We trained a logistic regression model on 1000 images from the MNIST dataset. We used batch gradient descent for training i.e. all 1000 samples were used for computing the gradient at each step. We trained this model using Eve, Adam, Adamax, and SGD Nesterov for 10000 iterations, searching over a large range of values for the learning rate and decay: $\\alpha \\in$ $\\{ 1 0 ^ { - 1 } , 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } \\}$ , $\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } , 1 0 ^ { - 6 } , 0 \\}$ . The results are shown in Figure 4. Eve again outperforms all other methods and achieves the lowest training loss. Also, since this is a smooth non-stochastic problem, the tuning coefficient $d _ { t }$ continuously decreases - this makes having a thresholding mechanism important since the learning rate would blow up otherwise. ", + "bbox": [ + 174, + 868, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92c46efb82e34367ee757f4c1af8c5474272a094da2588de2dd44fab1ac2d3c6.jpg", + "image_caption": [ + "(c) Penn Treebank, Language modeling ", + "Figure 6: Loss curves for experiments with recurrent neural networks. Eve consistently achieves better performance than other methods. " + ], + "image_footnote": [], + "bbox": [ + 181, + 98, + 818, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 493, + 825, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although in the previous experiment the effect of our method is to increase the learning rate, it is not equivalent to simply starting with a larger learning rate. We will establish this with a couple simple experiments. First we note that in the previous experiment, the optimal decay rates for both Adam and Eve were 0 - no decay. The optimal learning rate for Eve was $\\mathrm { \\bar { 1 0 } ^ { - 2 } }$ . Since the tuning coefficient converges to 0.1, we trained Adam using no decay, and learning rates $i \\times 1 0 ^ { - 2 }$ where $i$ varies from 1 to 10. The training loss curves are shown in the left plot of Figure 5. While increasing the learning rate with Adam does seem to close the gap with Eve, Eve does remain marginally faster. Moreover, and more importantly, this first plot represents the best-case situation for Adam. With larger learning rates, training becomes increasingly unstable and sensitive to the initial values of the parameters. This is illustrated in the right plot of Figure 5 where we used Eve (with learning rate $1 0 ^ { - 2 }$ ) and Adam (with learning rate $1 0 ^ { - 1 }$ ) 10 times with different random initializations. In some cases, Adam fails to converge whereas Eve always converges - even though Eve eventually reaches a learning rate of $1 0 ^ { - 1 }$ . This is because very early in the training, the model is quite sensitive at higher learning rates due to larger gradients. Depending on the initial values, the algorithm may or may not converge. So it is advantageous to slowly accelerate as the learning stabilizes rather than start with a larger learning rate. ", + "bbox": [ + 174, + 570, + 825, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 RECURRENT NEURAL NETWORKS", + "text_level": 1, + "bbox": [ + 178, + 813, + 449, + 827 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we evaluated our method on recurrent neural networks (RNNs). We first trained a RNN for character-level language modeling on the Penn Treebank dataset (Marcus et al., 1993). Specifically, the model consisted of a 2-layer character-level Gated Recurrent Unit (Chung et al., 2014) with hidden layers of size 256, 0.5 dropout between layers, and sequences of 100 characters. We adopted $1 0 ^ { - 3 }$ as the initial learning rate for Adam, Eve, and RMSProp. For Adamax, we used $2 \\times 1 0 ^ { - 3 }$ as the learning rate since it is the suggested value. We used $3 \\times 1 0 ^ { - 4 }$ for the learning rate decay. We trained this model for 100 epochs using each of the algorithms. The results, plotted in Figure 6c, clearly show that our method achieves the best results. Eve optimizes the model to a lower final loss than the other models. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 103, + 823, + 145 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We trained another RNN-based model for the question & answering task. Specifically, we chose two question types among 20 types from the bAbI dataset (Weston et al., 2015), Q19 and Q14. The dataset consists of pairs of supporting story sentences and a question. Different types of pairs are said to require different reasoning schemes. For our test case, Q19 and Q14 correspond to Path Finding and Time Reasoning respectively. We picked Q19 since it is reported to have the lowest baseline score, and we picked Q14 randomly from the remaining questions. The model consisted of two parts, one for encoding story sentences and another for query. Both included an embedding layer with 256 hidden units, and 0.3 dropout. Next query word embeddings were fed into a GRU one token at a time, to compute a sentence representation. Both story and query sequences were truncated to the maximum sequence length of 100. Finally, the sequence of word embeddings from story sentences and the repeated encoded representation of a query were combined together to serve as input for each time step in another GRU, with 0.3 dropout. We searched for the learning rate and decay from a range of values, $\\alpha \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}$ and $\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 0 \\}$ . The results, shown in Figures 6a, and 6b show that Eve again improves over all other methods. ", + "bbox": [ + 174, + 154, + 825, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 367, + 318, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a simple and efficient method for incorporating feedback in to stochastic gradient descent algorithms. We used this method to create Eve, a modified version of the Adam algorithm. Experiments with a variety of models showed that the proposed method can help improve the optimization of deep learning models. ", + "bbox": [ + 174, + 398, + 825, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For future work, we would look to theoretically analyze our method and its effects. While we have tried to evaluate our algorithm Eve on a variety of tasks, additional experiments on larger scale problems would further highlight the strength of our approach. We are making code for our method and the experiments publicly available to encourage more research on this method. ", + "bbox": [ + 176, + 462, + 825, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 537, + 285, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alan J Bray and David S Dean. Statistics of critical points of gaussian fields on large-dimensional spaces. Physical review letters, 98(15):150201, 2007. ", + "bbox": [ + 174, + 560, + 821, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. ", + "bbox": [ + 173, + 597, + 821, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In Advances in neural information processing systems, pp. 2933–2941, 2014. ", + "bbox": [ + 176, + 635, + 823, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. ", + "bbox": [ + 174, + 685, + 823, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Aistats, volume 9, pp. 249–256, 2010. ", + "bbox": [ + 176, + 723, + 823, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kenji Kawaguchi. Deep learning without poor local minima. In Advances in Neural Information Processing Systems (NIPS), 2016. to appear. ", + "bbox": [ + 176, + 761, + 825, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 174, + 797, + 823, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. ", + "bbox": [ + 171, + 834, + 825, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993. ", + "bbox": [ + 174, + 858, + 823, + 887 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence o (1/k2). In Doklady an SSSR, volume 269, pp. 543–547, 1983. ", + "bbox": [ + 176, + 895, + 821, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014. ", + "bbox": [ + 178, + 103, + 823, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4(2), 2012. ", + "bbox": [ + 173, + 155, + 825, + 184 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks. arXiv preprint arXiv:1502.05698, 2015. ", + "bbox": [ + 174, + 193, + 823, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012. ", + "bbox": [ + 174, + 244, + 825, + 272 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/parse/train/r1WUqIceg/r1WUqIceg_middle.json b/parse/train/r1WUqIceg/r1WUqIceg_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..5892849a272c18f5b2193891e475452499b16194 --- /dev/null +++ b/parse/train/r1WUqIceg/r1WUqIceg_middle.json @@ -0,0 +1,19665 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 450, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 453, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 453, + 98 + ], + "score": 1.0, + "content": "IMPROVING STOCHASTIC GRADIENT DESCENT", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 233, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 233, + 118 + ], + "score": 1.0, + "content": "WITH FEEDBACK", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 276, + 146 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 279, + 148 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 279, + 148 + ], + "score": 1.0, + "content": "Jayanth Koushik & Hiroaki Hayashi ∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 113, + 147, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 112, + 147, + 244, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 147, + 244, + 158 + ], + "score": 1.0, + "content": "Language Technologies Institute", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 226, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 226, + 169 + ], + "score": 1.0, + "content": "Carnegie Mellon University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 168, + 225, + 179 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 225, + 179 + ], + "score": 1.0, + "content": "Pittsburgh, PA 15213, USA", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 180, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 114, + 180, + 294, + 190 + ], + "score": 1.0, + "content": "{jkoushik,hiroakih}@cs.cmu.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 242, + 469, + 363 + ], + "lines": [ + { + "bbox": [ + 141, + 242, + 469, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 469, + 255 + ], + "score": 1.0, + "content": "In this paper we propose a simple and efficient method for improving stochastic", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "score": 1.0, + "content": "gradient descent methods by using feedback from the objective function. The", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 263, + 470, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 278 + ], + "score": 1.0, + "content": "method tracks the relative changes in the objective function with a running average,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 276, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 470, + 288 + ], + "score": 1.0, + "content": "and uses it to adaptively tune the learning rate in stochastic gradient descent. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 286, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 469, + 300 + ], + "score": 1.0, + "content": "specifically apply this idea to modify Adam, a popular algorithm for training deep", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 298, + 470, + 310 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 470, + 310 + ], + "score": 1.0, + "content": "neural networks. We conduct experiments to compare the resulting algorithm,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 469, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 321 + ], + "score": 1.0, + "content": "which we call Eve, with state of the art methods used for training deep learning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 318, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 469, + 333 + ], + "score": 1.0, + "content": "models. We train CNNs for image classification, and RNNs for language modeling", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "score": 1.0, + "content": "and question answering. Our experiments show that Eve outperforms all other", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 341, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 341, + 470, + 353 + ], + "score": 1.0, + "content": "algorithms on these benchmark tasks. We also analyze the behavior of the feedback", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 352, + 302, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 302, + 365 + ], + "score": 1.0, + "content": "mechanism during the training process.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 382, + 206, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Despite several breakthrough results in the last few years, the training of deep learning models", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "remains a challenging problem. This training is a complex, high-dimensional, non-convex, stochastic", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "optimization problem which is not amenable to many standard methods. Currently, the most common", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "approach is to use some variant of stochastic gradient descent. Many extensions have been proposed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "score": 1.0, + "content": "to the basic gradient descent algorithm - designed to handle specific issues in the training of deep", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 462, + 391, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 391, + 473 + ], + "score": 1.0, + "content": "learning models. We review some of these methods in the next section.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 108, + 478, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "Although variants of simple stochastic gradient descent work quite well in practice, there is still room", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "for improvement. This is easily evidenced by the existence of numerous methods to simplify the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 500, + 476, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 476, + 513 + ], + "score": 1.0, + "content": "optimization problem itself like weight initialization techniques and normalization methods.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 517, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 504, + 529 + ], + "score": 1.0, + "content": "In this work, we seek to improve stochastic gradient descent with a simple method that incorporates", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "feedback from the objective function. The relative changes in the objective function indicate progress", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "of the optimization algorithm. Our main hypothesis is that incorporating information about this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "change into the optimization algorithm can lead to improved performance - quantified in terms of the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "progress rate. We keep a running average of the relative changes in the objective function and use it", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 507, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 507, + 585 + ], + "score": 1.0, + "content": "to divide the learning rate. When the average relative change is high, the learning rate is reduced.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "This can improve the progress if, for example, the algorithm is bouncing around the walls of the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 504, + 606 + ], + "score": 1.0, + "content": "objective function. Conversely, when the relative change is low, the learning rate is increased. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "can help the algorithm accelerate through flat areas in the loss surface. As we discuss in the next", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "section, such “plateaus” pose a significant challenge for first order methods and can create the illusion", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 626, + 175, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 175, + 639 + ], + "score": 1.0, + "content": "of local minima.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "While our method is general i.e. independent of any particular optimization algorithm, in this work", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "we specifically apply the method to modify Adam (Kingma & Ba, 2014), considered to be the state of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "the art for training deep learning models. We call the resulting algorithm Eve and design experiments", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 423, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 423, + 689 + ], + "score": 1.0, + "content": "to compare it with Adam, as well as other popular methods from the literature.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 693, + 504, + 716 + ], + "lines": [ + { + "bbox": [ + 106, + 693, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 505, + 706 + ], + "score": 1.0, + "content": "The paper is organized as follows. In Section 2, we review recent results related to the optimization", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 704, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 704, + 505, + 717 + ], + "score": 1.0, + "content": "of deep neural networks. We also discuss some popular algorithms and their motivations. Our general", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 722, + 192, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 194, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 194, + 733 + ], + "score": 1.0, + "content": "∗Equal contribution.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 450, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 453, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 453, + 98 + ], + "score": 1.0, + "content": "IMPROVING STOCHASTIC GRADIENT DESCENT", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 233, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 233, + 118 + ], + "score": 1.0, + "content": "WITH FEEDBACK", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 276, + 146 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 279, + 148 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 279, + 148 + ], + "score": 1.0, + "content": "Jayanth Koushik & Hiroaki Hayashi ∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 111, + 134, + 279, + 148 + ] + }, + { + "type": "list", + "bbox": [ + 113, + 147, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 112, + 147, + 244, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 147, + 244, + 158 + ], + "score": 1.0, + "content": "Language Technologies Institute", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 156, + 226, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 226, + 169 + ], + "score": 1.0, + "content": "Carnegie Mellon University", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 168, + 225, + 179 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 225, + 179 + ], + "score": 1.0, + "content": "Pittsburgh, PA 15213, USA", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 180, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 114, + 180, + 294, + 190 + ], + "score": 1.0, + "content": "{jkoushik,hiroakih}@cs.cmu.edu", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + } + ], + "index": 4.5, + "bbox_fs": [ + 111, + 147, + 294, + 190 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 335, + 232 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 242, + 469, + 363 + ], + "lines": [ + { + "bbox": [ + 141, + 242, + 469, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 469, + 255 + ], + "score": 1.0, + "content": "In this paper we propose a simple and efficient method for improving stochastic", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "score": 1.0, + "content": "gradient descent methods by using feedback from the objective function. The", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 263, + 470, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 470, + 278 + ], + "score": 1.0, + "content": "method tracks the relative changes in the objective function with a running average,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 276, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 470, + 288 + ], + "score": 1.0, + "content": "and uses it to adaptively tune the learning rate in stochastic gradient descent. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 286, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 469, + 300 + ], + "score": 1.0, + "content": "specifically apply this idea to modify Adam, a popular algorithm for training deep", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 298, + 470, + 310 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 470, + 310 + ], + "score": 1.0, + "content": "neural networks. We conduct experiments to compare the resulting algorithm,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 469, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 321 + ], + "score": 1.0, + "content": "which we call Eve, with state of the art methods used for training deep learning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 318, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 469, + 333 + ], + "score": 1.0, + "content": "models. We train CNNs for image classification, and RNNs for language modeling", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 470, + 343 + ], + "score": 1.0, + "content": "and question answering. Our experiments show that Eve outperforms all other", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 341, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 341, + 470, + 353 + ], + "score": 1.0, + "content": "algorithms on these benchmark tasks. We also analyze the behavior of the feedback", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 352, + 302, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 302, + 365 + ], + "score": 1.0, + "content": "mechanism during the training process.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13, + "bbox_fs": [ + 141, + 242, + 470, + 365 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 382, + 206, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Despite several breakthrough results in the last few years, the training of deep learning models", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "remains a challenging problem. This training is a complex, high-dimensional, non-convex, stochastic", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "optimization problem which is not amenable to many standard methods. Currently, the most common", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "approach is to use some variant of stochastic gradient descent. Many extensions have been proposed", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "score": 1.0, + "content": "to the basic gradient descent algorithm - designed to handle specific issues in the training of deep", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 462, + 391, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 391, + 473 + ], + "score": 1.0, + "content": "learning models. We review some of these methods in the next section.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 406, + 506, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 478, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "Although variants of simple stochastic gradient descent work quite well in practice, there is still room", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "for improvement. This is easily evidenced by the existence of numerous methods to simplify the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 500, + 476, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 476, + 513 + ], + "score": 1.0, + "content": "optimization problem itself like weight initialization techniques and normalization methods.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 478, + 505, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 517, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 504, + 529 + ], + "score": 1.0, + "content": "In this work, we seek to improve stochastic gradient descent with a simple method that incorporates", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "feedback from the objective function. The relative changes in the objective function indicate progress", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "of the optimization algorithm. Our main hypothesis is that incorporating information about this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "change into the optimization algorithm can lead to improved performance - quantified in terms of the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "progress rate. We keep a running average of the relative changes in the objective function and use it", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 507, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 507, + 585 + ], + "score": 1.0, + "content": "to divide the learning rate. When the average relative change is high, the learning rate is reduced.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "This can improve the progress if, for example, the algorithm is bouncing around the walls of the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 504, + 606 + ], + "score": 1.0, + "content": "objective function. Conversely, when the relative change is low, the learning rate is increased. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "can help the algorithm accelerate through flat areas in the loss surface. As we discuss in the next", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "section, such “plateaus” pose a significant challenge for first order methods and can create the illusion", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 626, + 175, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 175, + 639 + ], + "score": 1.0, + "content": "of local minima.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 517, + 507, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "While our method is general i.e. independent of any particular optimization algorithm, in this work", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "we specifically apply the method to modify Adam (Kingma & Ba, 2014), considered to be the state of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "the art for training deep learning models. We call the resulting algorithm Eve and design experiments", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 423, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 423, + 689 + ], + "score": 1.0, + "content": "to compare it with Adam, as well as other popular methods from the literature.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 642, + 506, + 689 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 693, + 504, + 716 + ], + "lines": [ + { + "bbox": [ + 106, + 693, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 505, + 706 + ], + "score": 1.0, + "content": "The paper is organized as follows. In Section 2, we review recent results related to the optimization", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 704, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 704, + 505, + 717 + ], + "score": 1.0, + "content": "of deep neural networks. We also discuss some popular algorithms and their motivations. Our general", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "method, and the specific algorithm Eve are discussed in Section 3. Then in Section 4, we show that", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "Eve consistently outperforms other methods in training convolutional neural networks (CNNs), and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "recurrent neural networks (RNNs). We also look in some detail, at the behavior of our method in the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 448, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 448, + 127 + ], + "score": 1.0, + "content": "simple case of convex non-stochastic optimization. Finally we conclude in Section 5.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 693, + 505, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "method, and the specific algorithm Eve are discussed in Section 3. Then in Section 4, we show that", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "Eve consistently outperforms other methods in training convolutional neural networks (CNNs), and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "recurrent neural networks (RNNs). We also look in some detail, at the behavior of our method in the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 448, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 448, + 127 + ], + "score": 1.0, + "content": "simple case of convex non-stochastic optimization. Finally we conclude in Section 5.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 108, + 150, + 210, + 163 + ], + "lines": [ + { + "bbox": [ + 104, + 149, + 213, + 166 + ], + "spans": [ + { + "bbox": [ + 104, + 149, + 213, + 166 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 506, + 192 + ], + "score": 1.0, + "content": "There has been considerable effort to understand the challenges in deep learning optimization.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "Intuitively, it seems that the non-convex optimization is made difficult by the presence of several", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "poor local optima. However, this geometric intuition proves to be inadequate in reasoning about", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "the high-dimensional case that arises with training deep learning models. Various empirical and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 238 + ], + "score": 1.0, + "content": "theoretical results (Bray & Dean, 2007; Dauphin et al., 2014) have indicated that the problem in high", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "dimensions arises not from local minima, but rather from saddle points. Moreover, a recent paper", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "(Kawaguchi, 2016) proved (for deep linear networks, and under reasonable assumptions, also for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "deep non-linear networks) that all local minima in fact achieve the same value, and are optimal. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "work also showed that all critical points which are not global minima are saddle points. Saddle points", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "can seriously hamper the progress of both first and second order methods. Second order methods like", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "Newton’s method are actually attracted to saddle points and are not suitable for high dimensional", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "non-convex optimization. First order methods can escape from saddle points by following directions", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 313, + 504, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 504, + 325 + ], + "score": 1.0, + "content": "of negative curvature. However, such saddle points are usually surrounded by regions of small", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "curvature - plateaus. This makes first order methods very slow near saddle points and can create the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 334, + 222, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 222, + 347 + ], + "score": 1.0, + "content": "illusion of a local minimum.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "To tackle the saddle point problem, Dauphin et al. (2014) propose a second order method that fixes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "the issue with Newton’s method. Their algorithm builds on considering the behavior of Newton’s", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 374, + 504, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 504, + 385 + ], + "score": 1.0, + "content": "method near saddle points. Newton’s method rescales gradients in each eigen-direction with the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "corresponding inverse eigenvalue. However, near a saddle point, negative eigenvalues can cause the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "method to move towards the saddle point. Based on this observation, the authors propose using the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 406, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 418 + ], + "score": 1.0, + "content": "absolute values of the eigenvalues to rescale the gradients. This saddle-free Newton method is backed", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "by theoretical justifications and empirical results; however due to the computational requirements,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "second order methods are not very suitable for training large scale models. So we do not compare", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 439, + 246, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 246, + 451 + ], + "score": 1.0, + "content": "with such approaches in this work.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "We instead focus on first order methods which only rely on the gradient information. A key issue in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "score": 1.0, + "content": "training deep learning models is that of sparse gradients. To handle this, Adagrad (Duchi et al., 2011)", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "adaptively changes the learning rate for each parameter, performing larger updates for infrequently", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 489, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 507, + 502 + ], + "score": 1.0, + "content": "updated parameters. However its update rule causes the learning rate to monotonically decrease,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 498, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 513 + ], + "score": 1.0, + "content": "which eventually stalls the algorithm. Adadelta (Zeiler, 2012) and RMSProp (Tieleman & Hinton,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "2012) are two extensions that try to fix this issue. Finally, a closely related method, and the base", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "for our algorithm Eve (introduced in the next section), is Adam (Kingma & Ba, 2014). Adam", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "incorporates the advantages of both Adagrad and RMSProp - and it has been found to work quite", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 543, + 507, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 507, + 557 + ], + "score": 1.0, + "content": "well in practice. Adam uses a running average of the gradient to determine the direction of descent,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "and scales the learning rate with a running average of the gradient squared. The authors of Adam also", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "propose an extension based on the infinity norm, called Adamax. In our experiments, we compare", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 576, + 247, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 247, + 588 + ], + "score": 1.0, + "content": "Eve with both Adam and Adamax.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 612, + 172, + 625 + ], + "lines": [ + { + "bbox": [ + 104, + 610, + 174, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 174, + 628 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "title", + "bbox": [ + 108, + 641, + 189, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 190, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 190, + 654 + ], + "score": 1.0, + "content": "3.1 ASSUMPTION", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 369, + 678 + ], + "score": 1.0, + "content": "We do need to make an assumption about the objective function", + "type": "text" + }, + { + "bbox": [ + 370, + 666, + 376, + 677 + ], + "score": 0.82, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 665, + 505, + 678 + ], + "score": 1.0, + "content": ". We assume that the minimum", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 141, + 689 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 142, + 677, + 149, + 689 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "over its domain is known. While this is true for loss functions encountered in machine", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "learning (like mean squared error or cross entropy), it does not hold if the objective function also", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 225, + 712 + ], + "score": 1.0, + "content": "includes regularization terms", + "type": "text" + }, + { + "bbox": [ + 226, + 699, + 237, + 710 + ], + "score": 0.74, + "content": "\\boldsymbol { L } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 698, + 241, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 241, + 699, + 254, + 710 + ], + "score": 0.73, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "etc.). In all our experiments, we used dropout for regularization", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "which is not affected by this assumption. Finally, to simplify notation we assume that the minimum", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 441, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 441, + 732 + ], + "score": 1.0, + "content": "has been subtracted from the objective function i.e. the minimum has been made 0.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 505, + 127 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 150, + 210, + 163 + ], + "lines": [ + { + "bbox": [ + 104, + 149, + 213, + 166 + ], + "spans": [ + { + "bbox": [ + 104, + 149, + 213, + 166 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 506, + 192 + ], + "score": 1.0, + "content": "There has been considerable effort to understand the challenges in deep learning optimization.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "Intuitively, it seems that the non-convex optimization is made difficult by the presence of several", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "poor local optima. However, this geometric intuition proves to be inadequate in reasoning about", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "the high-dimensional case that arises with training deep learning models. Various empirical and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 238 + ], + "score": 1.0, + "content": "theoretical results (Bray & Dean, 2007; Dauphin et al., 2014) have indicated that the problem in high", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "dimensions arises not from local minima, but rather from saddle points. Moreover, a recent paper", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "(Kawaguchi, 2016) proved (for deep linear networks, and under reasonable assumptions, also for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "deep non-linear networks) that all local minima in fact achieve the same value, and are optimal. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "work also showed that all critical points which are not global minima are saddle points. Saddle points", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "can seriously hamper the progress of both first and second order methods. Second order methods like", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "Newton’s method are actually attracted to saddle points and are not suitable for high dimensional", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "non-convex optimization. First order methods can escape from saddle points by following directions", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 313, + 504, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 504, + 325 + ], + "score": 1.0, + "content": "of negative curvature. However, such saddle points are usually surrounded by regions of small", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "curvature - plateaus. This makes first order methods very slow near saddle points and can create the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 334, + 222, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 222, + 347 + ], + "score": 1.0, + "content": "illusion of a local minimum.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 180, + 506, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "To tackle the saddle point problem, Dauphin et al. (2014) propose a second order method that fixes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "the issue with Newton’s method. Their algorithm builds on considering the behavior of Newton’s", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 374, + 504, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 504, + 385 + ], + "score": 1.0, + "content": "method near saddle points. Newton’s method rescales gradients in each eigen-direction with the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "corresponding inverse eigenvalue. However, near a saddle point, negative eigenvalues can cause the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "method to move towards the saddle point. Based on this observation, the authors propose using the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 406, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 506, + 418 + ], + "score": 1.0, + "content": "absolute values of the eigenvalues to rescale the gradients. This saddle-free Newton method is backed", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 417, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 506, + 430 + ], + "score": 1.0, + "content": "by theoretical justifications and empirical results; however due to the computational requirements,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "second order methods are not very suitable for training large scale models. So we do not compare", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 439, + 246, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 246, + 451 + ], + "score": 1.0, + "content": "with such approaches in this work.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 351, + 506, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "We instead focus on first order methods which only rely on the gradient information. A key issue in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "score": 1.0, + "content": "training deep learning models is that of sparse gradients. To handle this, Adagrad (Duchi et al., 2011)", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "adaptively changes the learning rate for each parameter, performing larger updates for infrequently", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 489, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 507, + 502 + ], + "score": 1.0, + "content": "updated parameters. However its update rule causes the learning rate to monotonically decrease,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 498, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 513 + ], + "score": 1.0, + "content": "which eventually stalls the algorithm. Adadelta (Zeiler, 2012) and RMSProp (Tieleman & Hinton,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "2012) are two extensions that try to fix this issue. Finally, a closely related method, and the base", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "for our algorithm Eve (introduced in the next section), is Adam (Kingma & Ba, 2014). Adam", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "incorporates the advantages of both Adagrad and RMSProp - and it has been found to work quite", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 543, + 507, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 507, + 557 + ], + "score": 1.0, + "content": "well in practice. Adam uses a running average of the gradient to determine the direction of descent,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "and scales the learning rate with a running average of the gradient squared. The authors of Adam also", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "propose an extension based on the infinity norm, called Adamax. In our experiments, we compare", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 576, + 247, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 247, + 588 + ], + "score": 1.0, + "content": "Eve with both Adam and Adamax.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 455, + 507, + 588 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 612, + 172, + 625 + ], + "lines": [ + { + "bbox": [ + 104, + 610, + 174, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 174, + 628 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "title", + "bbox": [ + 108, + 641, + 189, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 190, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 190, + 654 + ], + "score": 1.0, + "content": "3.1 ASSUMPTION", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 369, + 678 + ], + "score": 1.0, + "content": "We do need to make an assumption about the objective function", + "type": "text" + }, + { + "bbox": [ + 370, + 666, + 376, + 677 + ], + "score": 0.82, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 665, + 505, + 678 + ], + "score": 1.0, + "content": ". We assume that the minimum", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 141, + 689 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 142, + 677, + 149, + 689 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "over its domain is known. While this is true for loss functions encountered in machine", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "learning (like mean squared error or cross entropy), it does not hold if the objective function also", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 225, + 712 + ], + "score": 1.0, + "content": "includes regularization terms", + "type": "text" + }, + { + "bbox": [ + 226, + 699, + 237, + 710 + ], + "score": 0.74, + "content": "\\boldsymbol { L } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 698, + 241, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 241, + 699, + 254, + 710 + ], + "score": 0.73, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "etc.). In all our experiments, we used dropout for regularization", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "which is not affected by this assumption. Finally, to simplify notation we assume that the minimum", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 441, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 441, + 732 + ], + "score": 1.0, + "content": "has been subtracted from the objective function i.e. the minimum has been made 0.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 665, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 93 + ], + "score": 1.0, + "content": "Algorithm 1 Eve: Adam with feedback. Parameters carried over from Adam have the same default", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 138, + 105 + ], + "score": 1.0, + "content": "values:", + "type": "text" + }, + { + "bbox": [ + 138, + 93, + 182, + 104 + ], + "score": 0.9, + "content": "\\alpha = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 92, + 186, + 105 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 186, + 93, + 225, + 105 + ], + "score": 0.87, + "content": "\\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 92, + 229, + 105 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 229, + 93, + 278, + 105 + ], + "score": 0.86, + "content": "\\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 92, + 282, + 105 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 282, + 93, + 323, + 104 + ], + "score": 0.88, + "content": "\\epsilon = 1 0 ^ { - 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 92, + 506, + 105 + ], + "score": 1.0, + "content": ". For parameters specific to our method, we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 218, + 116 + ], + "score": 1.0, + "content": "recommend default values", + "type": "text" + }, + { + "bbox": [ + 218, + 105, + 270, + 116 + ], + "score": 0.87, + "content": "\\beta _ { 3 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 104, + 274, + 116 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 274, + 104, + 311, + 115 + ], + "score": 0.84, + "content": "k = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 104, + 316, + 116 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 316, + 104, + 353, + 114 + ], + "score": 0.87, + "content": "K = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 104, + 505, + 116 + ], + "score": 1.0, + "content": ". Wherever applicable, products are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 161, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 161, + 127 + ], + "score": 1.0, + "content": "elementwise.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table", + "bbox": [ + 103, + 124, + 502, + 559 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 103, + 124, + 502, + 559 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 103, + 124, + 502, + 559 + ], + "spans": [ + { + "bbox": [ + 103, + 124, + 502, + 559 + ], + "score": 0.908, + "html": "
Require: α: learning rate
Require: β1, β2 ∈ [0,1): exponential decay rates for moment estimation in Adam
Require: β3 ∈ [0,1): exponential decay rate for computing relative change
Require: k,K: lower and upper threshold for relative change
Require: ε: fuzz factor
Require: f(θ): objective function
Require: Oo: initial value for parameters
mo = vo =0
do=1
f-1=t=0
while stopping condition is not reached do
t←t+1
gt←Vθf(0t-1)
mt←βimt-1+(1-β1)gt mt
mt←(1-β)
Ut←β2Ut-1+(1-β2)g² Ut
Ut←(1-)
if t>1 then
if f(0t-1) ≥ft-2 then δt←k+1
△t←K+1
else 1
8t←K+
△t←
end if
ct ←min{max{δt, }}
Ft-1←ctft-2
Ft-1-ft-21 rt↑
min{ft-1,ft-2} dt←βgdt-1+(1-β)rt
else
ft-1←f(0t-1)
dt←1
end if
0t←0t-1-Qdv+e mt
", + "type": "table", + "image_path": "5e45f59f6a66111f20a95e09b6aafc0193a84735fa47a785faf32d8eae5fb68c.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 103, + 124, + 502, + 269.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 103, + 269.0, + 502, + 414.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 103, + 414.0, + 502, + 559.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 580, + 186, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 187, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 187, + 593 + ], + "score": 1.0, + "content": "3.2 ALGORITHM", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "The main component of our proposed method is a feedback term that captures the relative change in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 612, + 504, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 201, + 626 + ], + "score": 1.0, + "content": "the objective value. Let", + "type": "text" + }, + { + "bbox": [ + 202, + 613, + 221, + 624 + ], + "score": 0.91, + "content": "f _ { t - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 612, + 239, + 626 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 613, + 259, + 624 + ], + "score": 0.91, + "content": "f _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 612, + 482, + 626 + ], + "score": 1.0, + "content": "denote the values of the objective function at time steps", + "type": "text" + }, + { + "bbox": [ + 482, + 613, + 504, + 623 + ], + "score": 0.83, + "content": "t - 2", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 622, + 507, + 641 + ], + "spans": [ + { + "bbox": [ + 103, + 622, + 392, + 640 + ], + "score": 1.0, + "content": "and t − 1 respectively. Then this change is computed as rt = ft−2−ft−1f", + "type": "text" + }, + { + "bbox": [ + 325, + 623, + 441, + 641 + ], + "score": 0.94, + "content": "\\begin{array} { r } { r _ { t } = { \\frac { f _ { t - 2 } - \\bar { \\cdot } f _ { t - 1 } } { f _ { t - 1 } } } = { \\frac { f _ { t - 2 } - f _ { t - 1 } } { f _ { t - 1 } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 623, + 507, + 638 + ], + "score": 1.0, + "content": "ft−2−ft−1 if ft−2 ft−1,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 638, + 507, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 123, + 653 + ], + "score": 1.0, + "content": "an", + "type": "text" + }, + { + "bbox": [ + 119, + 638, + 507, + 657 + ], + "score": 1.0, + "content": "d ft−1−ft−2f otherwise. Note that this value is always non-negative but it can be less than or greater", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "than 1 i.e. it captures both relative increase and decrease. We compute a running average using these", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 375, + 678 + ], + "score": 1.0, + "content": "relative changes to get a smoother estimate. Specifically, we take", + "type": "text" + }, + { + "bbox": [ + 375, + 666, + 405, + 677 + ], + "score": 0.91, + "content": "d _ { 1 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 666, + 441, + 678 + ], + "score": 1.0, + "content": ", and for", + "type": "text" + }, + { + "bbox": [ + 441, + 666, + 465, + 676 + ], + "score": 0.9, + "content": "t > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 666, + 493, + 678 + ], + "score": 1.0, + "content": "define", + "type": "text" + }, + { + "bbox": [ + 494, + 666, + 504, + 677 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 118, + 689 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 118, + 677, + 219, + 689 + ], + "score": 0.91, + "content": "d _ { t } = \\beta d _ { t - 1 } + ( 1 - \\beta ) r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 677, + 247, + 689 + ], + "score": 1.0, + "content": ". Here", + "type": "text" + }, + { + "bbox": [ + 248, + 677, + 289, + 689 + ], + "score": 0.92, + "content": "\\beta \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "is a decay rate - large values correspond to a slowly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 146, + 700 + ], + "score": 1.0, + "content": "changing", + "type": "text" + }, + { + "bbox": [ + 146, + 689, + 156, + 699 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 688, + 506, + 700 + ], + "score": 1.0, + "content": ", and vice versa. This simple expression can, however, blow up and lead to instability.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 482, + 712 + ], + "score": 1.0, + "content": "To handle this issue, we use a thresholding scheme. A simple thing to do would be to clip", + "type": "text" + }, + { + "bbox": [ + 482, + 699, + 492, + 710 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 709, + 195, + 722 + ], + "score": 0.92, + "content": "\\operatorname* { m i n } \\{ \\operatorname* { m a x } \\{ k , d _ { t } \\} , K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 710, + 270, + 722 + ], + "score": 1.0, + "content": "for some suitable", + "type": "text" + }, + { + "bbox": [ + 271, + 710, + 321, + 720 + ], + "score": 0.9, + "content": "0 < k < K", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 710, + 506, + 722 + ], + "score": 1.0, + "content": ". But we found this to not work very well in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 406, + 734 + ], + "score": 1.0, + "content": "practice due to the abrupt nature of the clipping. Instead we indirectly clip", + "type": "text" + }, + { + "bbox": [ + 406, + 721, + 416, + 732 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "by smoothly tracking", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 93 + ], + "score": 1.0, + "content": "Algorithm 1 Eve: Adam with feedback. Parameters carried over from Adam have the same default", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 138, + 105 + ], + "score": 1.0, + "content": "values:", + "type": "text" + }, + { + "bbox": [ + 138, + 93, + 182, + 104 + ], + "score": 0.9, + "content": "\\alpha = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 92, + 186, + 105 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 186, + 93, + 225, + 105 + ], + "score": 0.87, + "content": "\\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 92, + 229, + 105 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 229, + 93, + 278, + 105 + ], + "score": 0.86, + "content": "\\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 92, + 282, + 105 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 282, + 93, + 323, + 104 + ], + "score": 0.88, + "content": "\\epsilon = 1 0 ^ { - 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 92, + 506, + 105 + ], + "score": 1.0, + "content": ". For parameters specific to our method, we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 218, + 116 + ], + "score": 1.0, + "content": "recommend default values", + "type": "text" + }, + { + "bbox": [ + 218, + 105, + 270, + 116 + ], + "score": 0.87, + "content": "\\beta _ { 3 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 104, + 274, + 116 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 274, + 104, + 311, + 115 + ], + "score": 0.84, + "content": "k = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 104, + 316, + 116 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 316, + 104, + 353, + 114 + ], + "score": 0.87, + "content": "K = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 104, + 505, + 116 + ], + "score": 1.0, + "content": ". Wherever applicable, products are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 161, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 161, + 127 + ], + "score": 1.0, + "content": "elementwise.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 127 + ] + }, + { + "type": "table", + "bbox": [ + 103, + 124, + 502, + 559 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 103, + 124, + 502, + 559 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 103, + 124, + 502, + 559 + ], + "spans": [ + { + "bbox": [ + 103, + 124, + 502, + 559 + ], + "score": 0.908, + "html": "
Require: α: learning rate
Require: β1, β2 ∈ [0,1): exponential decay rates for moment estimation in Adam
Require: β3 ∈ [0,1): exponential decay rate for computing relative change
Require: k,K: lower and upper threshold for relative change
Require: ε: fuzz factor
Require: f(θ): objective function
Require: Oo: initial value for parameters
mo = vo =0
do=1
f-1=t=0
while stopping condition is not reached do
t←t+1
gt←Vθf(0t-1)
mt←βimt-1+(1-β1)gt mt
mt←(1-β)
Ut←β2Ut-1+(1-β2)g² Ut
Ut←(1-)
if t>1 then
if f(0t-1) ≥ft-2 then δt←k+1
△t←K+1
else 1
8t←K+
△t←
end if
ct ←min{max{δt, }}
Ft-1←ctft-2
Ft-1-ft-21 rt↑
min{ft-1,ft-2} dt←βgdt-1+(1-β)rt
else
ft-1←f(0t-1)
dt←1
end if
0t←0t-1-Qdv+e mt
", + "type": "table", + "image_path": "5e45f59f6a66111f20a95e09b6aafc0193a84735fa47a785faf32d8eae5fb68c.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 103, + 124, + 502, + 269.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 103, + 269.0, + 502, + 414.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 103, + 414.0, + 502, + 559.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 580, + 186, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 187, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 187, + 593 + ], + "score": 1.0, + "content": "3.2 ALGORITHM", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "The main component of our proposed method is a feedback term that captures the relative change in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 612, + 504, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 201, + 626 + ], + "score": 1.0, + "content": "the objective value. Let", + "type": "text" + }, + { + "bbox": [ + 202, + 613, + 221, + 624 + ], + "score": 0.91, + "content": "f _ { t - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 612, + 239, + 626 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 613, + 259, + 624 + ], + "score": 0.91, + "content": "f _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 612, + 482, + 626 + ], + "score": 1.0, + "content": "denote the values of the objective function at time steps", + "type": "text" + }, + { + "bbox": [ + 482, + 613, + 504, + 623 + ], + "score": 0.83, + "content": "t - 2", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 622, + 507, + 641 + ], + "spans": [ + { + "bbox": [ + 103, + 622, + 392, + 640 + ], + "score": 1.0, + "content": "and t − 1 respectively. Then this change is computed as rt = ft−2−ft−1f", + "type": "text" + }, + { + "bbox": [ + 325, + 623, + 441, + 641 + ], + "score": 0.94, + "content": "\\begin{array} { r } { r _ { t } = { \\frac { f _ { t - 2 } - \\bar { \\cdot } f _ { t - 1 } } { f _ { t - 1 } } } = { \\frac { f _ { t - 2 } - f _ { t - 1 } } { f _ { t - 1 } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 623, + 507, + 638 + ], + "score": 1.0, + "content": "ft−2−ft−1 if ft−2 ft−1,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 638, + 507, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 123, + 653 + ], + "score": 1.0, + "content": "an", + "type": "text" + }, + { + "bbox": [ + 119, + 638, + 507, + 657 + ], + "score": 1.0, + "content": "d ft−1−ft−2f otherwise. Note that this value is always non-negative but it can be less than or greater", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "than 1 i.e. it captures both relative increase and decrease. We compute a running average using these", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 375, + 678 + ], + "score": 1.0, + "content": "relative changes to get a smoother estimate. Specifically, we take", + "type": "text" + }, + { + "bbox": [ + 375, + 666, + 405, + 677 + ], + "score": 0.91, + "content": "d _ { 1 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 666, + 441, + 678 + ], + "score": 1.0, + "content": ", and for", + "type": "text" + }, + { + "bbox": [ + 441, + 666, + 465, + 676 + ], + "score": 0.9, + "content": "t > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 666, + 493, + 678 + ], + "score": 1.0, + "content": "define", + "type": "text" + }, + { + "bbox": [ + 494, + 666, + 504, + 677 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 118, + 689 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 118, + 677, + 219, + 689 + ], + "score": 0.91, + "content": "d _ { t } = \\beta d _ { t - 1 } + ( 1 - \\beta ) r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 677, + 247, + 689 + ], + "score": 1.0, + "content": ". Here", + "type": "text" + }, + { + "bbox": [ + 248, + 677, + 289, + 689 + ], + "score": 0.92, + "content": "\\beta \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "is a decay rate - large values correspond to a slowly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 146, + 700 + ], + "score": 1.0, + "content": "changing", + "type": "text" + }, + { + "bbox": [ + 146, + 689, + 156, + 699 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 688, + 506, + 700 + ], + "score": 1.0, + "content": ", and vice versa. This simple expression can, however, blow up and lead to instability.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 482, + 712 + ], + "score": 1.0, + "content": "To handle this issue, we use a thresholding scheme. A simple thing to do would be to clip", + "type": "text" + }, + { + "bbox": [ + 482, + 699, + 492, + 710 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 709, + 195, + 722 + ], + "score": 0.92, + "content": "\\operatorname* { m i n } \\{ \\operatorname* { m a x } \\{ k , d _ { t } \\} , K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 710, + 270, + 722 + ], + "score": 1.0, + "content": "for some suitable", + "type": "text" + }, + { + "bbox": [ + 271, + 710, + 321, + 720 + ], + "score": 0.9, + "content": "0 < k < K", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 710, + 506, + 722 + ], + "score": 1.0, + "content": ". But we found this to not work very well in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 406, + 734 + ], + "score": 1.0, + "content": "practice due to the abrupt nature of the clipping. Instead we indirectly clip", + "type": "text" + }, + { + "bbox": [ + 406, + 721, + 416, + 732 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "by smoothly tracking", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 262, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 213, + 277 + ], + "score": 1.0, + "content": "the objective function. Let", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 213, + 262, + 233, + 276 + ], + "score": 0.92, + "content": "\\widehat { f } _ { t - 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 233, + 263, + 482, + 277 + ], + "score": 1.0, + "content": "be the value of the smoothly tracked objective function at time", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 482, + 265, + 504, + 275 + ], + "score": 0.7, + "content": "t - 2", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 276, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 127, + 291 + ], + "score": 1.0, + "content": "with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 127, + 276, + 162, + 290 + ], + "score": 0.93, + "content": "{ \\widehat { f } } _ { 0 } = f _ { 0 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 162, + 276, + 240, + 291 + ], + "score": 1.0, + "content": ". For now, assume", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 240, + 276, + 293, + 290 + ], + "score": 0.92, + "content": "f _ { t - 1 } \\leq \\widehat { f } _ { t - 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 293, + 276, + 394, + 291 + ], + "score": 1.0, + "content": ". We would like to have", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 394, + 278, + 448, + 289 + ], + "score": 0.9, + "content": "k \\leq r _ { t } \\leq K", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 448, + 276, + 505, + 291 + ], + "score": 1.0, + "content": "which in this", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 288, + 509, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 163, + 311 + ], + "score": 1.0, + "content": "case requires", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 163, + 289, + 245, + 309 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\frac { f _ { t - 1 } } { \\widehat { f } _ { t - 2 } } \\in \\left[ \\frac { 1 } { K + 1 } , \\frac { 1 } { k + 1 } \\right] } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 246, + 288, + 318, + 314 + ], + "score": 1.0, + "content": ". So we compute", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 318, + 289, + 472, + 309 + ], + "score": 0.94, + "content": "\\begin{array} { r } { c _ { t } = \\operatorname* { m i n } \\left\\{ \\operatorname* { m a x } \\left\\{ \\frac { 1 } { K + 1 } , \\frac { f _ { t } } { \\widehat { f } _ { t - 1 } } \\right\\} , \\frac { 1 } { k + 1 } \\right\\} } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 472, + 288, + 509, + 314 + ], + "score": 1.0, + "content": "and set", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 304, + 509, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 166, + 324 + ], + "score": 0.92, + "content": "\\widehat { f } _ { t - 1 } = c _ { t } \\widehat { f } _ { t - 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 166, + 304, + 200, + 330 + ], + "score": 1.0, + "content": ". Finally", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 200, + 313, + 209, + 323 + ], + "score": 0.78, + "content": "r _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 210, + 304, + 219, + 330 + ], + "score": 1.0, + "content": "is", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 220, + 308, + 261, + 329 + ], + "score": 0.93, + "content": "\\frac { \\widehat { f } _ { t - 2 } - \\widehat { f } _ { t - 1 } } { \\widehat { f } _ { t - 1 } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 261, + 304, + 509, + 330 + ], + "score": 1.0, + "content": ". Analogous expressions can also be derived for the case when", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 326, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 158, + 342 + ], + "score": 0.92, + "content": "f _ { t - 1 } > \\widehat { f } _ { t - 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 158, + 326, + 405, + 343 + ], + "score": 1.0, + "content": ". This smooth tracking has the additional advantage of making", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 406, + 330, + 415, + 341 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 416, + 326, + 506, + 343 + ], + "score": 1.0, + "content": "less susceptible to the", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 341, + 346, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 346, + 353 + ], + "score": 1.0, + "content": "high variability that comes with training using minibatches.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 13, + "bbox_fs": [ + 103, + 601, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 79, + 501, + 210 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 79, + 501, + 210 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 79, + 501, + 210 + ], + "spans": [ + { + "bbox": [ + 110, + 79, + 501, + 210 + ], + "score": 0.971, + "type": "image", + "image_path": "f97087fe23db1dc3a650ad1e44d6c838538276740c378a42457561cffa200ed7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 79, + 501, + 122.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 122.66666666666666, + 501, + 166.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 166.33333333333331, + 501, + 209.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 219, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "Figure 1: Training loss for convolutional neural networks on CIFAR10 and CIFAR100. The vertical", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 231, + 483, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 483, + 244 + ], + "score": 1.0, + "content": "axis is shown in log scale. In both cases, our proposed method achieves the best performance.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 262, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 213, + 277 + ], + "score": 1.0, + "content": "the objective function. Let", + "type": "text" + }, + { + "bbox": [ + 213, + 262, + 233, + 276 + ], + "score": 0.92, + "content": "\\widehat { f } _ { t - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 263, + 482, + 277 + ], + "score": 1.0, + "content": "be the value of the smoothly tracked objective function at time", + "type": "text" + }, + { + "bbox": [ + 482, + 265, + 504, + 275 + ], + "score": 0.7, + "content": "t - 2", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 276, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 127, + 291 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 276, + 162, + 290 + ], + "score": 0.93, + "content": "{ \\widehat { f } } _ { 0 } = f _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 276, + 240, + 291 + ], + "score": 1.0, + "content": ". For now, assume", + "type": "text" + }, + { + "bbox": [ + 240, + 276, + 293, + 290 + ], + "score": 0.92, + "content": "f _ { t - 1 } \\leq \\widehat { f } _ { t - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 276, + 394, + 291 + ], + "score": 1.0, + "content": ". We would like to have", + "type": "text" + }, + { + "bbox": [ + 394, + 278, + 448, + 289 + ], + "score": 0.9, + "content": "k \\leq r _ { t } \\leq K", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 276, + 505, + 291 + ], + "score": 1.0, + "content": "which in this", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 288, + 509, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 163, + 311 + ], + "score": 1.0, + "content": "case requires", + "type": "text" + }, + { + "bbox": [ + 163, + 289, + 245, + 309 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\frac { f _ { t - 1 } } { \\widehat { f } _ { t - 2 } } \\in \\left[ \\frac { 1 } { K + 1 } , \\frac { 1 } { k + 1 } \\right] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 288, + 318, + 314 + ], + "score": 1.0, + "content": ". So we compute", + "type": "text" + }, + { + "bbox": [ + 318, + 289, + 472, + 309 + ], + "score": 0.94, + "content": "\\begin{array} { r } { c _ { t } = \\operatorname* { m i n } \\left\\{ \\operatorname* { m a x } \\left\\{ \\frac { 1 } { K + 1 } , \\frac { f _ { t } } { \\widehat { f } _ { t - 1 } } \\right\\} , \\frac { 1 } { k + 1 } \\right\\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 288, + 509, + 314 + ], + "score": 1.0, + "content": "and set", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 304, + 509, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 166, + 324 + ], + "score": 0.92, + "content": "\\widehat { f } _ { t - 1 } = c _ { t } \\widehat { f } _ { t - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 304, + 200, + 330 + ], + "score": 1.0, + "content": ". Finally", + "type": "text" + }, + { + "bbox": [ + 200, + 313, + 209, + 323 + ], + "score": 0.78, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 304, + 219, + 330 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 220, + 308, + 261, + 329 + ], + "score": 0.93, + "content": "\\frac { \\widehat { f } _ { t - 2 } - \\widehat { f } _ { t - 1 } } { \\widehat { f } _ { t - 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 304, + 509, + 330 + ], + "score": 1.0, + "content": ". Analogous expressions can also be derived for the case when", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 326, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 158, + 342 + ], + "score": 0.92, + "content": "f _ { t - 1 } > \\widehat { f } _ { t - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 326, + 405, + 343 + ], + "score": 1.0, + "content": ". This smooth tracking has the additional advantage of making", + "type": "text" + }, + { + "bbox": [ + 406, + 330, + 415, + 341 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 326, + 506, + 343 + ], + "score": 1.0, + "content": "less susceptible to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 341, + 346, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 346, + 353 + ], + "score": 1.0, + "content": "high variability that comes with training using minibatches.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 129, + 371 + ], + "score": 1.0, + "content": "Once", + "type": "text" + }, + { + "bbox": [ + 130, + 358, + 140, + 369 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 356, + 505, + 371 + ], + "score": 1.0, + "content": "has been computed, it can be used to modify any gradient descent algorithm by modifying", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 177, + 381 + ], + "score": 1.0, + "content": "the learning rate", + "type": "text" + }, + { + "bbox": [ + 177, + 371, + 185, + 379 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 368, + 199, + 381 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 199, + 369, + 246, + 380 + ], + "score": 0.93, + "content": "\\alpha _ { t } = \\alpha / d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 368, + 320, + 381 + ], + "score": 1.0, + "content": ". Large values of", + "type": "text" + }, + { + "bbox": [ + 320, + 369, + 330, + 380 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 368, + 505, + 381 + ], + "score": 1.0, + "content": ", caused by large changes in the objective", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 437, + 393 + ], + "score": 1.0, + "content": "function will lead to a smaller effective learning rate. Similarly, small values of", + "type": "text" + }, + { + "bbox": [ + 437, + 380, + 447, + 391 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "will lead to a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 303, + 403 + ], + "score": 1.0, + "content": "larger effective learning rate. Since we start with", + "type": "text" + }, + { + "bbox": [ + 304, + 391, + 333, + 402 + ], + "score": 0.92, + "content": "d _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 390, + 506, + 403 + ], + "score": 1.0, + "content": ", the initial updates will closely follow that", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 363, + 415 + ], + "score": 1.0, + "content": "of the base algorithm. In the next section, we will look at how", + "type": "text" + }, + { + "bbox": [ + 363, + 402, + 374, + 413 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "evolves during the course of an", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 412, + 387, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 387, + 426 + ], + "score": 1.0, + "content": "experiment to get a better understanding of how it affects the training.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 506, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "score": 1.0, + "content": "We note again that our method is independent of any particular gradient descent algorithm. However,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "for this current work, we specifically focus on applying the method to Adam (Kingma & Ba, 2014).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 452, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 463 + ], + "score": 1.0, + "content": "This modified algorithm, which we call Eve, is shown in Algorithm 1. We modify the final Adam", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 262, + 475 + ], + "score": 1.0, + "content": "update by multiplying the denominator", + "type": "text" + }, + { + "bbox": [ + 262, + 463, + 280, + 475 + ], + "score": 0.91, + "content": "\\sqrt { v _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 462, + 300, + 475 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 301, + 463, + 311, + 474 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 462, + 506, + 475 + ], + "score": 1.0, + "content": ". In addition to the hyperparameters in Adam, we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 473, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 239, + 486 + ], + "score": 1.0, + "content": "introduce 3 new hyperparameters", + "type": "text" + }, + { + "bbox": [ + 239, + 474, + 250, + 485 + ], + "score": 0.59, + "content": "\\beta _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 473, + 254, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 254, + 474, + 261, + 484 + ], + "score": 0.47, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 473, + 281, + 486 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 281, + 474, + 291, + 483 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 473, + 455, + 486 + ], + "score": 1.0, + "content": ". In all our experiments we use the values", + "type": "text" + }, + { + "bbox": [ + 455, + 473, + 502, + 485 + ], + "score": 0.91, + "content": "\\beta _ { 3 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 473, + 506, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 483, + 363, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 139, + 495 + ], + "score": 0.89, + "content": "k = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 483, + 160, + 497 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 160, + 484, + 193, + 495 + ], + "score": 0.9, + "content": "K = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 483, + 363, + 497 + ], + "score": 1.0, + "content": ", which we found to work well in practice.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 108, + 513, + 200, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 201, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 201, + 527 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 538, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "Now we evaluate our proposed method by comparing Eve with several state of the art algorithms for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "optimizing deep learning models.1. In all experiments, we used ReLU activations, and initialized", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "weights according to the scheme proposed by Glorot & Bengio (2010). We used minibatches of size", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 570, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 275, + 586 + ], + "score": 1.0, + "content": "128, and linear decay for the learning rate:", + "type": "text" + }, + { + "bbox": [ + 275, + 571, + 343, + 583 + ], + "score": 0.92, + "content": "\\alpha _ { t } = \\alpha / ( 1 + \\gamma t )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 573, + 356, + 583 + ], + "score": 0.71, + "content": "( \\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 570, + 505, + 586 + ], + "score": 1.0, + "content": "is the decay rate, picked by searching", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 582, + 202, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 202, + 595 + ], + "score": 1.0, + "content": "over a range of values).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "In the figures, SGD refers to vanilla stochastic gradient descent, and SGD Nesterov refers to stochastic", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "gradient descent with Nesterov momentum (Nesterov, 1983) where we set the momentum to 0.9 in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 621, + 173, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 173, + 634 + ], + "score": 1.0, + "content": "all experiments.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 299, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 300, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 300, + 660 + ], + "score": 1.0, + "content": "4.1 CONVOLUTIONAL NEURAL NETWORKS", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "We first trained a 5 layer convolutional neural network for 10-way classification of images from the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 432, + 691 + ], + "score": 1.0, + "content": "CIFAR10 dataset (Krizhevsky & Hinton, 2009). The model consisted of 2 blocks of", + "type": "text" + }, + { + "bbox": [ + 433, + 679, + 449, + 689 + ], + "score": 0.37, + "content": "3 \\mathrm { x } 3", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "convolutional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 690, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 203, + 702 + ], + "score": 1.0, + "content": "layers each followed by", + "type": "text" + }, + { + "bbox": [ + 204, + 690, + 221, + 700 + ], + "score": 0.67, + "content": "2 \\mathbf { x } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 690, + 506, + 702 + ], + "score": 1.0, + "content": "max pooling and 0.25 dropout (Srivastava et al., 2014). The first block", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 712, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "1Full technical details of the experiments and additional results are available at https://github.com/", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 257, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 257, + 732 + ], + "score": 1.0, + "content": "jayanthkoushik/sgd-feedback.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 79, + 501, + 210 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 79, + 501, + 210 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 79, + 501, + 210 + ], + "spans": [ + { + "bbox": [ + 110, + 79, + 501, + 210 + ], + "score": 0.971, + "type": "image", + "image_path": "f97087fe23db1dc3a650ad1e44d6c838538276740c378a42457561cffa200ed7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 79, + 501, + 122.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 122.66666666666666, + 501, + 166.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 166.33333333333331, + 501, + 209.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 219, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "Figure 1: Training loss for convolutional neural networks on CIFAR10 and CIFAR100. The vertical", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 231, + 483, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 483, + 244 + ], + "score": 1.0, + "content": "axis is shown in log scale. In both cases, our proposed method achieves the best performance.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 106, + 262, + 505, + 353 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 106, + 262, + 509, + 353 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 129, + 371 + ], + "score": 1.0, + "content": "Once", + "type": "text" + }, + { + "bbox": [ + 130, + 358, + 140, + 369 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 356, + 505, + 371 + ], + "score": 1.0, + "content": "has been computed, it can be used to modify any gradient descent algorithm by modifying", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 177, + 381 + ], + "score": 1.0, + "content": "the learning rate", + "type": "text" + }, + { + "bbox": [ + 177, + 371, + 185, + 379 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 368, + 199, + 381 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 199, + 369, + 246, + 380 + ], + "score": 0.93, + "content": "\\alpha _ { t } = \\alpha / d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 368, + 320, + 381 + ], + "score": 1.0, + "content": ". Large values of", + "type": "text" + }, + { + "bbox": [ + 320, + 369, + 330, + 380 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 368, + 505, + 381 + ], + "score": 1.0, + "content": ", caused by large changes in the objective", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 437, + 393 + ], + "score": 1.0, + "content": "function will lead to a smaller effective learning rate. Similarly, small values of", + "type": "text" + }, + { + "bbox": [ + 437, + 380, + 447, + 391 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "will lead to a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 303, + 403 + ], + "score": 1.0, + "content": "larger effective learning rate. Since we start with", + "type": "text" + }, + { + "bbox": [ + 304, + 391, + 333, + 402 + ], + "score": 0.92, + "content": "d _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 390, + 506, + 403 + ], + "score": 1.0, + "content": ", the initial updates will closely follow that", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 363, + 415 + ], + "score": 1.0, + "content": "of the base algorithm. In the next section, we will look at how", + "type": "text" + }, + { + "bbox": [ + 363, + 402, + 374, + 413 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "evolves during the course of an", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 412, + 387, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 387, + 426 + ], + "score": 1.0, + "content": "experiment to get a better understanding of how it affects the training.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 356, + 506, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 506, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "score": 1.0, + "content": "We note again that our method is independent of any particular gradient descent algorithm. However,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "for this current work, we specifically focus on applying the method to Adam (Kingma & Ba, 2014).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 452, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 463 + ], + "score": 1.0, + "content": "This modified algorithm, which we call Eve, is shown in Algorithm 1. We modify the final Adam", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 262, + 475 + ], + "score": 1.0, + "content": "update by multiplying the denominator", + "type": "text" + }, + { + "bbox": [ + 262, + 463, + 280, + 475 + ], + "score": 0.91, + "content": "\\sqrt { v _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 462, + 300, + 475 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 301, + 463, + 311, + 474 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 462, + 506, + 475 + ], + "score": 1.0, + "content": ". In addition to the hyperparameters in Adam, we", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 473, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 239, + 486 + ], + "score": 1.0, + "content": "introduce 3 new hyperparameters", + "type": "text" + }, + { + "bbox": [ + 239, + 474, + 250, + 485 + ], + "score": 0.59, + "content": "\\beta _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 473, + 254, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 254, + 474, + 261, + 484 + ], + "score": 0.47, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 473, + 281, + 486 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 281, + 474, + 291, + 483 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 473, + 455, + 486 + ], + "score": 1.0, + "content": ". In all our experiments we use the values", + "type": "text" + }, + { + "bbox": [ + 455, + 473, + 502, + 485 + ], + "score": 0.91, + "content": "\\beta _ { 3 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 473, + 506, + 486 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 483, + 363, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 139, + 495 + ], + "score": 0.89, + "content": "k = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 483, + 160, + 497 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 160, + 484, + 193, + 495 + ], + "score": 0.9, + "content": "K = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 483, + 363, + 497 + ], + "score": 1.0, + "content": ", which we found to work well in practice.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 428, + 507, + 497 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 513, + 200, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 201, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 201, + 527 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 538, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "Now we evaluate our proposed method by comparing Eve with several state of the art algorithms for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "optimizing deep learning models.1. In all experiments, we used ReLU activations, and initialized", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "weights according to the scheme proposed by Glorot & Bengio (2010). We used minibatches of size", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 570, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 275, + 586 + ], + "score": 1.0, + "content": "128, and linear decay for the learning rate:", + "type": "text" + }, + { + "bbox": [ + 275, + 571, + 343, + 583 + ], + "score": 0.92, + "content": "\\alpha _ { t } = \\alpha / ( 1 + \\gamma t )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 573, + 356, + 583 + ], + "score": 0.71, + "content": "( \\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 570, + 505, + 586 + ], + "score": 1.0, + "content": "is the decay rate, picked by searching", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 582, + 202, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 202, + 595 + ], + "score": 1.0, + "content": "over a range of values).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 537, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "In the figures, SGD refers to vanilla stochastic gradient descent, and SGD Nesterov refers to stochastic", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "gradient descent with Nesterov momentum (Nesterov, 1983) where we set the momentum to 0.9 in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 621, + 173, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 173, + 634 + ], + "score": 1.0, + "content": "all experiments.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 599, + 506, + 634 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 299, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 300, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 300, + 660 + ], + "score": 1.0, + "content": "4.1 CONVOLUTIONAL NEURAL NETWORKS", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "We first trained a 5 layer convolutional neural network for 10-way classification of images from the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 432, + 691 + ], + "score": 1.0, + "content": "CIFAR10 dataset (Krizhevsky & Hinton, 2009). The model consisted of 2 blocks of", + "type": "text" + }, + { + "bbox": [ + 433, + 679, + 449, + 689 + ], + "score": 0.37, + "content": "3 \\mathrm { x } 3", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "convolutional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 690, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 203, + 702 + ], + "score": 1.0, + "content": "layers each followed by", + "type": "text" + }, + { + "bbox": [ + 204, + 690, + 221, + 700 + ], + "score": 0.67, + "content": "2 \\mathbf { x } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 690, + 506, + 702 + ], + "score": 1.0, + "content": "max pooling and 0.25 dropout (Srivastava et al., 2014). The first block", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 506, + 540 + ], + "score": 1.0, + "content": "contained 2 layers with 32 filters each, and the second block contained 2 layers with 64 filters each.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "The convolutional layers were followed by a fully connected layer with 512 units and a 10-way", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "softmax layer. We trained this model for 500 epochs on the training split using various popular", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "methods for training deep learning models, as well as Eve. For each algorithm, we tried learning rates", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 571, + 189, + 584 + ], + "score": 0.91, + "content": "\\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 189, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "(for algorithms with suggested default learning rates, we also included them in", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 237, + 596 + ], + "score": 1.0, + "content": "the search), learning rate decays", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 237, + 582, + 328, + 595 + ], + "score": 0.91, + "content": "\\{ 0 , 1 0 ^ { - 2 } , 1 \\breve { 0 } ^ { - 3 } , 1 0 ^ { - 4 } \\}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 328, + 582, + 505, + 596 + ], + "score": 1.0, + "content": ", and picked the pair of values that led to the", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "smallest final training loss. The loss curves are shown in Figure 1a. Eve quickly surpasses all other", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "methods and achieves the lowest final training loss. In the next section we will look at the behavior", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 614, + 456, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 212, + 630 + ], + "score": 1.0, + "content": "of the adaptive coefficient", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 212, + 616, + 222, + 627 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 222, + 614, + 456, + 630 + ], + "score": 1.0, + "content": "to gain some intuition behind this improved performance.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 667, + 506, + 702 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 81, + 504, + 232 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 81, + 504, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 81, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 81, + 504, + 232 + ], + "score": 0.968, + "type": "image", + "image_path": "f3ff840fdc66fddf215a2fadd32bb8c3e7b6c2e863f57aeacf8c8f8ff6545987.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 81, + 504, + 131.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 131.33333333333334, + 504, + 181.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 181.66666666666669, + 504, + 232.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 242, + 505, + 276 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 279, + 255 + ], + "score": 1.0, + "content": "Figure 2: Behavior of the tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 279, + 243, + 289, + 254 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "during the experiment shown in Figure 1a. There is an", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "overall trend of acceleration followed by decay, but also more fine-grained behavior as indicated by", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 264, + 194, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 194, + 277 + ], + "score": 1.0, + "content": "the bottom-right plot.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 192, + 315, + 418, + 447 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 192, + 315, + 418, + 447 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 192, + 315, + 418, + 447 + ], + "spans": [ + { + "bbox": [ + 192, + 315, + 418, + 447 + ], + "score": 0.967, + "type": "image", + "image_path": "e2d88fee7e37f8eda5b9a8ded593e0c30bd16264ba4f39e6b5079d4508648bf8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 192, + 315, + 418, + 329.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 329.6666666666667, + 418, + 344.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 344.33333333333337, + 418, + 359.00000000000006 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 192, + 359.00000000000006, + 418, + 373.66666666666674 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 192, + 373.66666666666674, + 418, + 388.3333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 192, + 388.3333333333334, + 418, + 403.0000000000001 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 192, + 403.0000000000001, + 418, + 417.6666666666668 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 192, + 417.6666666666668, + 418, + 432.3333333333335 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 192, + 432.3333333333335, + 418, + 447.00000000000017 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 461, + 504, + 485 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "Figure 3: Minibatch losses for Eve during the CNN experiment on CIFAR10 (Figure 1a). The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 472, + 329, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 329, + 487 + ], + "score": 1.0, + "content": "variance in the losses increases throughout the training.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + } + ], + "index": 12.75 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 506, + 540 + ], + "score": 1.0, + "content": "contained 2 layers with 32 filters each, and the second block contained 2 layers with 64 filters each.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "The convolutional layers were followed by a fully connected layer with 512 units and a 10-way", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "softmax layer. We trained this model for 500 epochs on the training split using various popular", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "methods for training deep learning models, as well as Eve. For each algorithm, we tried learning rates", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 571, + 189, + 584 + ], + "score": 0.91, + "content": "\\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "(for algorithms with suggested default learning rates, we also included them in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 237, + 596 + ], + "score": 1.0, + "content": "the search), learning rate decays", + "type": "text" + }, + { + "bbox": [ + 237, + 582, + 328, + 595 + ], + "score": 0.91, + "content": "\\{ 0 , 1 0 ^ { - 2 } , 1 \\breve { 0 } ^ { - 3 } , 1 0 ^ { - 4 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 582, + 505, + 596 + ], + "score": 1.0, + "content": ", and picked the pair of values that led to the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "smallest final training loss. The loss curves are shown in Figure 1a. Eve quickly surpasses all other", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "methods and achieves the lowest final training loss. In the next section we will look at the behavior", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 614, + 456, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 212, + 630 + ], + "score": 1.0, + "content": "of the adaptive coefficient", + "type": "text" + }, + { + "bbox": [ + 212, + 616, + 222, + 627 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 614, + 456, + 630 + ], + "score": 1.0, + "content": "to gain some intuition behind this improved performance.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "We also trained a larger CNN model using the top-performing algorithms from the previous exper-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 286, + 656 + ], + "score": 1.0, + "content": "iment. This model consisted of 3 blocks of", + "type": "text" + }, + { + "bbox": [ + 287, + 644, + 304, + 654 + ], + "score": 0.33, + "content": "3 \\mathrm { x } 3", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "convolutional layers (3 layers per block, and 64,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 654, + 504, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 468, + 668 + ], + "score": 1.0, + "content": "128, 256 filters per layer in the first, second, and third block respectively) each followed by", + "type": "text" + }, + { + "bbox": [ + 468, + 655, + 504, + 666 + ], + "score": 0.33, + "content": "2 \\mathrm { x } 2 \\mathrm { m a x }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "pooling and 0.5 dropout. Then we had 2 fully connected layers with 512 units, each followed by 0.5", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "dropout, and finally a 100-way softmax. We trained this model on the CIFAR100 (Krizhevsky &", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "Hinton, 2009) dataset for 100 epochs. We again performed a grid search over the same learning rate", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "and decay values as the last experiment. The results are shown in Figure 1b, and once again show", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "that our proposed method improves over state of the art methods for training convolutional neural", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 721, + 148, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 148, + 733 + ], + "score": 1.0, + "content": "networks.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 81, + 504, + 232 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 81, + 504, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 81, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 81, + 504, + 232 + ], + "score": 0.968, + "type": "image", + "image_path": "f3ff840fdc66fddf215a2fadd32bb8c3e7b6c2e863f57aeacf8c8f8ff6545987.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 81, + 504, + 131.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 131.33333333333334, + 504, + 181.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 181.66666666666669, + 504, + 232.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 242, + 505, + 276 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 279, + 255 + ], + "score": 1.0, + "content": "Figure 2: Behavior of the tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 279, + 243, + 289, + 254 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "during the experiment shown in Figure 1a. There is an", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "overall trend of acceleration followed by decay, but also more fine-grained behavior as indicated by", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 264, + 194, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 194, + 277 + ], + "score": 1.0, + "content": "the bottom-right plot.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 192, + 315, + 418, + 447 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 192, + 315, + 418, + 447 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 192, + 315, + 418, + 447 + ], + "spans": [ + { + "bbox": [ + 192, + 315, + 418, + 447 + ], + "score": 0.967, + "type": "image", + "image_path": "e2d88fee7e37f8eda5b9a8ded593e0c30bd16264ba4f39e6b5079d4508648bf8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 192, + 315, + 418, + 329.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 329.6666666666667, + 418, + 344.33333333333337 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 344.33333333333337, + 418, + 359.00000000000006 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 192, + 359.00000000000006, + 418, + 373.66666666666674 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 192, + 373.66666666666674, + 418, + 388.3333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 192, + 388.3333333333334, + 418, + 403.0000000000001 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 192, + 403.0000000000001, + 418, + 417.6666666666668 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 192, + 417.6666666666668, + 418, + 432.3333333333335 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 192, + 432.3333333333335, + 418, + 447.00000000000017 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 461, + 504, + 485 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "Figure 3: Minibatch losses for Eve during the CNN experiment on CIFAR10 (Figure 1a). The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 472, + 329, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 329, + 487 + ], + "score": 1.0, + "content": "variance in the losses increases throughout the training.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + } + ], + "index": 12.75 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 505, + 628 + ], + "lines": [], + "index": 21, + "bbox_fs": [ + 105, + 528, + 506, + 630 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "We also trained a larger CNN model using the top-performing algorithms from the previous exper-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 286, + 656 + ], + "score": 1.0, + "content": "iment. This model consisted of 3 blocks of", + "type": "text" + }, + { + "bbox": [ + 287, + 644, + 304, + 654 + ], + "score": 0.33, + "content": "3 \\mathrm { x } 3", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "convolutional layers (3 layers per block, and 64,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 654, + 504, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 468, + 668 + ], + "score": 1.0, + "content": "128, 256 filters per layer in the first, second, and third block respectively) each followed by", + "type": "text" + }, + { + "bbox": [ + 468, + 655, + 504, + 666 + ], + "score": 0.33, + "content": "2 \\mathrm { x } 2 \\mathrm { m a x }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "pooling and 0.5 dropout. Then we had 2 fully connected layers with 512 units, each followed by 0.5", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "dropout, and finally a 100-way softmax. We trained this model on the CIFAR100 (Krizhevsky &", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "Hinton, 2009) dataset for 100 epochs. We again performed a grid search over the same learning rate", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "and decay values as the last experiment. The results are shown in Figure 1b, and once again show", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "that our proposed method improves over state of the art methods for training convolutional neural", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 721, + 148, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 148, + 733 + ], + "score": 1.0, + "content": "networks.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 87, + 499, + 233 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 87, + 499, + 233 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 87, + 499, + 233 + ], + "spans": [ + { + "bbox": [ + 113, + 87, + 499, + 233 + ], + "score": 0.965, + "type": "image", + "image_path": "27f3a9518a4468af14eaa22a22c033deb4cbfe6c54a5d76cf153c1895cc83be7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 87, + 499, + 135.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 135.66666666666666, + 499, + 184.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 184.33333333333331, + 499, + 232.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 248, + 505, + 272 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 294, + 261 + ], + "score": 1.0, + "content": "Figure 4: Loss curves and tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 295, + 249, + 305, + 260 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "for batch gradient descent training of a logistic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 489, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 217, + 272 + ], + "score": 1.0, + "content": "regression model. For Eve,", + "type": "text" + }, + { + "bbox": [ + 217, + 260, + 227, + 271 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 260, + 489, + 272 + ], + "score": 1.0, + "content": "continuously decreases and converges to the lower threshold 0.1.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 113, + 288, + 496, + 437 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 288, + 496, + 437 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 288, + 496, + 437 + ], + "spans": [ + { + "bbox": [ + 113, + 288, + 496, + 437 + ], + "score": 0.97, + "type": "image", + "image_path": "4b9980bf2139c08ee969e7d5bfc125a1ed4d0abbb8ea01e97cfe2b928fe37f6f.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 113, + 288, + 496, + 337.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 337.6666666666667, + 496, + 387.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 113, + 387.33333333333337, + 496, + 437.00000000000006 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 452, + 506, + 486 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 450, + 478, + 465 + ], + "score": 1.0, + "content": "Figure 5: The left plot shows Adam with different learning rates plotted with Eve (learning rate", + "type": "text" + }, + { + "bbox": [ + 478, + 451, + 500, + 462 + ], + "score": 0.9, + "content": "1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 450, + 506, + 465 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Although Adam with a larger learning rate can be almost identical to Eve, this is largely dependent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 474, + 360, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 360, + 486 + ], + "score": 1.0, + "content": "on the initial values for Adam as shown in the plot on the right.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 109, + 507, + 285, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 286, + 520 + ], + "score": 1.0, + "content": "4.2 ANALYSIS OF TUNING COEFFICIENT", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Before we consider the next set of experiments on recurrent neural networks, we will first look more", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 294, + 552 + ], + "score": 1.0, + "content": "closely at the behavior of the tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 294, + 540, + 304, + 550 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "in our algorithm. We will specifically consider the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 431, + 563 + ], + "score": 1.0, + "content": "results from the CNN experiment on CIFAR10. Figure 2 shows the progress of", + "type": "text" + }, + { + "bbox": [ + 432, + 551, + 441, + 561 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "throughout the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "training, and also in two smaller windows. A few things are worth noting here. First is that of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "overall trend. There is an initial acceleration followed by a decay. This initial acceleration allows Eve", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "to rapidly overtake other methods, and makes it proceed at a faster pace for about 100 epochs. This", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "acceleration is not equivalent to simply starting with a larger learning rate - in all our experiments we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 368, + 618 + ], + "score": 1.0, + "content": "search over a range of learning rate values. The overall trend for", + "type": "text" + }, + { + "bbox": [ + 368, + 605, + 378, + 616 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "can be explained by looking at", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "the minibatch losses at each iteration (as opposed to the loss computed over the entire dataset after", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 627, + 504, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 494, + 640 + ], + "score": 1.0, + "content": "each epoch) in Figure 3. Initially, different minibatches achieve similar loss values which leads to", + "type": "text" + }, + { + "bbox": [ + 494, + 627, + 504, + 638 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 451, + 651 + ], + "score": 1.0, + "content": "decreasing. But as training proceeds, the variance in the minibatch losses increases and", + "type": "text" + }, + { + "bbox": [ + 451, + 639, + 461, + 649 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "eventually", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 650, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 661 + ], + "score": 1.0, + "content": "increases. However, this overall trend does not capture the complete picture - for example, as shown", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 658, + 507, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 254, + 674 + ], + "score": 1.0, + "content": "in the bottom right plot of Figure 2,", + "type": "text" + }, + { + "bbox": [ + 255, + 660, + 265, + 671 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 658, + 507, + 674 + ], + "score": 1.0, + "content": "can actually be decreasing in some regions of the training,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 671, + 299, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 299, + 683 + ], + "score": 1.0, + "content": "adjusting to local structures in the error surface.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "To further study the observed acceleration, and to also motivate the need for clipping, we consider", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "a simpler experiment. We trained a logistic regression model on 1000 images from the MNIST", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "dataset. We used batch gradient descent for training i.e. all 1000 samples were used for computing", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "the gradient at each step. We trained this model using Eve, Adam, Adamax, and SGD Nesterov", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 87, + 499, + 233 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 87, + 499, + 233 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 87, + 499, + 233 + ], + "spans": [ + { + "bbox": [ + 113, + 87, + 499, + 233 + ], + "score": 0.965, + "type": "image", + "image_path": "27f3a9518a4468af14eaa22a22c033deb4cbfe6c54a5d76cf153c1895cc83be7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 87, + 499, + 135.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 135.66666666666666, + 499, + 184.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 184.33333333333331, + 499, + 232.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 248, + 505, + 272 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 294, + 261 + ], + "score": 1.0, + "content": "Figure 4: Loss curves and tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 295, + 249, + 305, + 260 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "for batch gradient descent training of a logistic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 260, + 489, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 217, + 272 + ], + "score": 1.0, + "content": "regression model. For Eve,", + "type": "text" + }, + { + "bbox": [ + 217, + 260, + 227, + 271 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 260, + 489, + 272 + ], + "score": 1.0, + "content": "continuously decreases and converges to the lower threshold 0.1.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 113, + 288, + 496, + 437 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 288, + 496, + 437 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 288, + 496, + 437 + ], + "spans": [ + { + "bbox": [ + 113, + 288, + 496, + 437 + ], + "score": 0.97, + "type": "image", + "image_path": "4b9980bf2139c08ee969e7d5bfc125a1ed4d0abbb8ea01e97cfe2b928fe37f6f.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 113, + 288, + 496, + 337.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 337.6666666666667, + 496, + 387.33333333333337 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 113, + 387.33333333333337, + 496, + 437.00000000000006 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 452, + 506, + 486 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 450, + 478, + 465 + ], + "score": 1.0, + "content": "Figure 5: The left plot shows Adam with different learning rates plotted with Eve (learning rate", + "type": "text" + }, + { + "bbox": [ + 478, + 451, + 500, + 462 + ], + "score": 0.9, + "content": "1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 450, + 506, + 465 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Although Adam with a larger learning rate can be almost identical to Eve, this is largely dependent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 474, + 360, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 360, + 486 + ], + "score": 1.0, + "content": "on the initial values for Adam as shown in the plot on the right.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 109, + 507, + 285, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 286, + 520 + ], + "score": 1.0, + "content": "4.2 ANALYSIS OF TUNING COEFFICIENT", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 528, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Before we consider the next set of experiments on recurrent neural networks, we will first look more", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 294, + 552 + ], + "score": 1.0, + "content": "closely at the behavior of the tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 294, + 540, + 304, + 550 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "in our algorithm. We will specifically consider the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 431, + 563 + ], + "score": 1.0, + "content": "results from the CNN experiment on CIFAR10. Figure 2 shows the progress of", + "type": "text" + }, + { + "bbox": [ + 432, + 551, + 441, + 561 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "throughout the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "training, and also in two smaller windows. A few things are worth noting here. First is that of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "overall trend. There is an initial acceleration followed by a decay. This initial acceleration allows Eve", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "to rapidly overtake other methods, and makes it proceed at a faster pace for about 100 epochs. This", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "acceleration is not equivalent to simply starting with a larger learning rate - in all our experiments we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 368, + 618 + ], + "score": 1.0, + "content": "search over a range of learning rate values. The overall trend for", + "type": "text" + }, + { + "bbox": [ + 368, + 605, + 378, + 616 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "can be explained by looking at", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "the minibatch losses at each iteration (as opposed to the loss computed over the entire dataset after", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 627, + 504, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 494, + 640 + ], + "score": 1.0, + "content": "each epoch) in Figure 3. Initially, different minibatches achieve similar loss values which leads to", + "type": "text" + }, + { + "bbox": [ + 494, + 627, + 504, + 638 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 451, + 651 + ], + "score": 1.0, + "content": "decreasing. But as training proceeds, the variance in the minibatch losses increases and", + "type": "text" + }, + { + "bbox": [ + 451, + 639, + 461, + 649 + ], + "score": 0.88, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "eventually", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 650, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 661 + ], + "score": 1.0, + "content": "increases. However, this overall trend does not capture the complete picture - for example, as shown", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 658, + 507, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 254, + 674 + ], + "score": 1.0, + "content": "in the bottom right plot of Figure 2,", + "type": "text" + }, + { + "bbox": [ + 255, + 660, + 265, + 671 + ], + "score": 0.86, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 658, + 507, + 674 + ], + "score": 1.0, + "content": "can actually be decreasing in some regions of the training,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 671, + 299, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 299, + 683 + ], + "score": 1.0, + "content": "adjusting to local structures in the error surface.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 527, + 507, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "To further study the observed acceleration, and to also motivate the need for clipping, we consider", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "a simpler experiment. We trained a logistic regression model on 1000 images from the MNIST", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "dataset. We used batch gradient descent for training i.e. all 1000 samples were used for computing", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "the gradient at each step. We trained this model using Eve, Adam, Adamax, and SGD Nesterov", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 486, + 403 + ], + "score": 1.0, + "content": "for 10000 iterations, searching over a large range of values for the learning rate and decay:", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 486, + 392, + 505, + 402 + ], + "score": 0.83, + "content": "\\alpha \\in", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 400, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 239, + 414 + ], + "score": 0.66, + "content": "\\{ 1 0 ^ { - 1 } , 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } \\}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 239, + 400, + 241, + 416 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 242, + 402, + 402, + 414 + ], + "score": 0.83, + "content": "\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } , 1 0 ^ { - 6 } , 0 \\}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 403, + 400, + 506, + 416 + ], + "score": 1.0, + "content": ". The results are shown in", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "Figure 4. Eve again outperforms all other methods and achieves the lowest training loss. Also, since", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 351, + 436 + ], + "score": 1.0, + "content": "this is a smooth non-stochastic problem, the tuning coefficient", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 352, + 425, + 361, + 435 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 362, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "continuously decreases - this makes", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 435, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 480, + 447 + ], + "score": 1.0, + "content": "having a thresholding mechanism important since the learning rate would blow up otherwise.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 78, + 501, + 324 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 78, + 501, + 324 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 78, + 501, + 324 + ], + "spans": [ + { + "bbox": [ + 111, + 78, + 501, + 324 + ], + "score": 0.975, + "type": "image", + "image_path": "92c46efb82e34367ee757f4c1af8c5474272a094da2588de2dd44fab1ac2d3c6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 78, + 501, + 160.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 160.0, + 501, + 242.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 242.0, + 501, + 324.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 233, + 325, + 378, + 336 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 232, + 323, + 378, + 338 + ], + "spans": [ + { + "bbox": [ + 232, + 323, + 378, + 338 + ], + "score": 1.0, + "content": "(c) Penn Treebank, Language modeling", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 344, + 504, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 6: Loss curves for experiments with recurrent neural networks. Eve consistently achieves", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 356, + 265, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 265, + 367 + ], + "score": 1.0, + "content": "better performance than other methods.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 486, + 403 + ], + "score": 1.0, + "content": "for 10000 iterations, searching over a large range of values for the learning rate and decay:", + "type": "text" + }, + { + "bbox": [ + 486, + 392, + 505, + 402 + ], + "score": 0.83, + "content": "\\alpha \\in", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 400, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 239, + 414 + ], + "score": 0.66, + "content": "\\{ 1 0 ^ { - 1 } , 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 400, + 241, + 416 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 242, + 402, + 402, + 414 + ], + "score": 0.83, + "content": "\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } , 1 0 ^ { - 6 } , 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 400, + 506, + 416 + ], + "score": 1.0, + "content": ". The results are shown in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "Figure 4. Eve again outperforms all other methods and achieves the lowest training loss. Also, since", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 351, + 436 + ], + "score": 1.0, + "content": "this is a smooth non-stochastic problem, the tuning coefficient", + "type": "text" + }, + { + "bbox": [ + 352, + 425, + 361, + 435 + ], + "score": 0.87, + "content": "d _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "continuously decreases - this makes", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 435, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 480, + 447 + ], + "score": 1.0, + "content": "having a thresholding mechanism important since the learning rate would blow up otherwise.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "Although in the previous experiment the effect of our method is to increase the learning rate, it is not", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "equivalent to simply starting with a larger learning rate. We will establish this with a couple simple", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "experiments. First we note that in the previous experiment, the optimal decay rates for both Adam", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 368, + 498 + ], + "score": 1.0, + "content": "and Eve were 0 - no decay. The optimal learning rate for Eve was", + "type": "text" + }, + { + "bbox": [ + 369, + 484, + 390, + 495 + ], + "score": 0.89, + "content": "\\mathrm { \\bar { 1 0 } ^ { - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 484, + 506, + 498 + ], + "score": 1.0, + "content": ". Since the tuning coefficient", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 495, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 380, + 508 + ], + "score": 1.0, + "content": "converges to 0.1, we trained Adam using no decay, and learning rates", + "type": "text" + }, + { + "bbox": [ + 380, + 495, + 417, + 506 + ], + "score": 0.92, + "content": "i \\times 1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 495, + 445, + 508 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 445, + 497, + 450, + 506 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 495, + 506, + 508 + ], + "score": 1.0, + "content": "varies from 1", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "to 10. The training loss curves are shown in the left plot of Figure 5. While increasing the learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "rate with Adam does seem to close the gap with Eve, Eve does remain marginally faster. Moreover,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 527, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 506, + 543 + ], + "score": 1.0, + "content": "and more importantly, this first plot represents the best-case situation for Adam. With larger learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "score": 1.0, + "content": "rates, training becomes increasingly unstable and sensitive to the initial values of the parameters.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 436, + 563 + ], + "score": 1.0, + "content": "This is illustrated in the right plot of Figure 5 where we used Eve (with learning rate", + "type": "text" + }, + { + "bbox": [ + 436, + 550, + 458, + 561 + ], + "score": 0.87, + "content": "1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 550, + 506, + 563 + ], + "score": 1.0, + "content": ") and Adam", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 183, + 574 + ], + "score": 1.0, + "content": "(with learning rate", + "type": "text" + }, + { + "bbox": [ + 183, + 561, + 205, + 572 + ], + "score": 0.91, + "content": "1 0 ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 561, + 506, + 574 + ], + "score": 1.0, + "content": ") 10 times with different random initializations. In some cases, Adam fails", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "to converge whereas Eve always converges - even though Eve eventually reaches a learning rate of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 582, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 128, + 594 + ], + "score": 0.89, + "content": "1 0 ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 582, + 506, + 597 + ], + "score": 1.0, + "content": ". This is because very early in the training, the model is quite sensitive at higher learning rates", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 593, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 608 + ], + "score": 1.0, + "content": "due to larger gradients. Depending on the initial values, the algorithm may or may not converge. So it", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "is advantageous to slowly accelerate as the learning stabilizes rather than start with a larger learning", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 617, + 128, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 128, + 629 + ], + "score": 1.0, + "content": "rate.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 109, + 644, + 275, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 277, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 277, + 657 + ], + "score": 1.0, + "content": "4.3 RECURRENT NEURAL NETWORKS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "Finally, we evaluated our method on recurrent neural networks (RNNs). We first trained a RNN for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "character-level language modeling on the Penn Treebank dataset (Marcus et al., 1993). Specifically,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "the model consisted of a 2-layer character-level Gated Recurrent Unit (Chung et al., 2014) with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "hidden layers of size 256, 0.5 dropout between layers, and sequences of 100 characters. We adopted", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 128, + 720 + ], + "score": 0.89, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 708, + 453, + 722 + ], + "score": 1.0, + "content": "as the initial learning rate for Adam, Eve, and RMSProp. For Adamax, we used", + "type": "text" + }, + { + "bbox": [ + 454, + 709, + 492, + 720 + ], + "score": 0.92, + "content": "2 \\times 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 719, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 337, + 733 + ], + "score": 1.0, + "content": "the learning rate since it is the suggested value. We used", + "type": "text" + }, + { + "bbox": [ + 338, + 720, + 376, + 731 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 719, + 505, + 733 + ], + "score": 1.0, + "content": "for the learning rate decay. We", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 78, + 501, + 324 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 78, + 501, + 324 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 78, + 501, + 324 + ], + "spans": [ + { + "bbox": [ + 111, + 78, + 501, + 324 + ], + "score": 0.975, + "type": "image", + "image_path": "92c46efb82e34367ee757f4c1af8c5474272a094da2588de2dd44fab1ac2d3c6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 78, + 501, + 160.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 160.0, + 501, + 242.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 242.0, + 501, + 324.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 233, + 325, + 378, + 336 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 232, + 323, + 378, + 338 + ], + "spans": [ + { + "bbox": [ + 232, + 323, + 378, + 338 + ], + "score": 1.0, + "content": "(c) Penn Treebank, Language modeling", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 344, + 504, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 6: Loss curves for experiments with recurrent neural networks. Eve consistently achieves", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 356, + 265, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 265, + 367 + ], + "score": 1.0, + "content": "better performance than other methods.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 446 + ], + "lines": [], + "index": 8, + "bbox_fs": [ + 105, + 391, + 506, + 447 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "Although in the previous experiment the effect of our method is to increase the learning rate, it is not", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "equivalent to simply starting with a larger learning rate. We will establish this with a couple simple", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "experiments. First we note that in the previous experiment, the optimal decay rates for both Adam", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 368, + 498 + ], + "score": 1.0, + "content": "and Eve were 0 - no decay. The optimal learning rate for Eve was", + "type": "text" + }, + { + "bbox": [ + 369, + 484, + 390, + 495 + ], + "score": 0.89, + "content": "\\mathrm { \\bar { 1 0 } ^ { - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 484, + 506, + 498 + ], + "score": 1.0, + "content": ". Since the tuning coefficient", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 495, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 380, + 508 + ], + "score": 1.0, + "content": "converges to 0.1, we trained Adam using no decay, and learning rates", + "type": "text" + }, + { + "bbox": [ + 380, + 495, + 417, + 506 + ], + "score": 0.92, + "content": "i \\times 1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 495, + 445, + 508 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 445, + 497, + 450, + 506 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 495, + 506, + 508 + ], + "score": 1.0, + "content": "varies from 1", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "to 10. The training loss curves are shown in the left plot of Figure 5. While increasing the learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "rate with Adam does seem to close the gap with Eve, Eve does remain marginally faster. Moreover,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 527, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 506, + 543 + ], + "score": 1.0, + "content": "and more importantly, this first plot represents the best-case situation for Adam. With larger learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "score": 1.0, + "content": "rates, training becomes increasingly unstable and sensitive to the initial values of the parameters.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 436, + 563 + ], + "score": 1.0, + "content": "This is illustrated in the right plot of Figure 5 where we used Eve (with learning rate", + "type": "text" + }, + { + "bbox": [ + 436, + 550, + 458, + 561 + ], + "score": 0.87, + "content": "1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 550, + 506, + 563 + ], + "score": 1.0, + "content": ") and Adam", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 183, + 574 + ], + "score": 1.0, + "content": "(with learning rate", + "type": "text" + }, + { + "bbox": [ + 183, + 561, + 205, + 572 + ], + "score": 0.91, + "content": "1 0 ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 561, + 506, + 574 + ], + "score": 1.0, + "content": ") 10 times with different random initializations. In some cases, Adam fails", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "to converge whereas Eve always converges - even though Eve eventually reaches a learning rate of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 582, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 128, + 594 + ], + "score": 0.89, + "content": "1 0 ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 582, + 506, + 597 + ], + "score": 1.0, + "content": ". This is because very early in the training, the model is quite sensitive at higher learning rates", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 593, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 608 + ], + "score": 1.0, + "content": "due to larger gradients. Depending on the initial values, the algorithm may or may not converge. So it", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "is advantageous to slowly accelerate as the learning stabilizes rather than start with a larger learning", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 617, + 128, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 128, + 629 + ], + "score": 1.0, + "content": "rate.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 452, + 507, + 629 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 644, + 275, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 277, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 277, + 657 + ], + "score": 1.0, + "content": "4.3 RECURRENT NEURAL NETWORKS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "Finally, we evaluated our method on recurrent neural networks (RNNs). We first trained a RNN for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "character-level language modeling on the Penn Treebank dataset (Marcus et al., 1993). Specifically,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "the model consisted of a 2-layer character-level Gated Recurrent Unit (Chung et al., 2014) with", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "hidden layers of size 256, 0.5 dropout between layers, and sequences of 100 characters. We adopted", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 128, + 720 + ], + "score": 0.89, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 708, + 453, + 722 + ], + "score": 1.0, + "content": "as the initial learning rate for Adam, Eve, and RMSProp. For Adamax, we used", + "type": "text" + }, + { + "bbox": [ + 454, + 709, + 492, + 720 + ], + "score": 0.92, + "content": "2 \\times 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 719, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 337, + 733 + ], + "score": 1.0, + "content": "the learning rate since it is the suggested value. We used", + "type": "text" + }, + { + "bbox": [ + 338, + 720, + 376, + 731 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 719, + 505, + 733 + ], + "score": 1.0, + "content": "for the learning rate decay. We", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "trained this model for 100 epochs using each of the algorithms. The results, plotted in Figure 6c,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "clearly show that our method achieves the best results. Eve optimizes the model to a lower final loss", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 198, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 198, + 116 + ], + "score": 1.0, + "content": "than the other models.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 666, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "trained this model for 100 epochs using each of the algorithms. The results, plotted in Figure 6c,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "clearly show that our method achieves the best results. Eve optimizes the model to a lower final loss", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 198, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 198, + 116 + ], + "score": 1.0, + "content": "than the other models.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "We trained another RNN-based model for the question & answering task. Specifically, we chose", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "two question types among 20 types from the bAbI dataset (Weston et al., 2015), Q19 and Q14. The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "dataset consists of pairs of supporting story sentences and a question. Different types of pairs are said", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "to require different reasoning schemes. For our test case, Q19 and Q14 correspond to Path Finding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "and Time Reasoning respectively. We picked Q19 since it is reported to have the lowest baseline", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "score, and we picked Q14 randomly from the remaining questions. The model consisted of two parts,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "one for encoding story sentences and another for query. Both included an embedding layer with 256", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "hidden units, and 0.3 dropout. Next query word embeddings were fed into a GRU one token at a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 223 + ], + "score": 1.0, + "content": "time, to compute a sentence representation. Both story and query sequences were truncated to the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "maximum sequence length of 100. Finally, the sequence of word embeddings from story sentences", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "and the repeated encoded representation of a query were combined together to serve as input for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 504, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 504, + 254 + ], + "score": 1.0, + "content": "each time step in another GRU, with 0.3 dropout. We searched for the learning rate and decay from", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 251, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 181, + 266 + ], + "score": 1.0, + "content": "a range of values,", + "type": "text" + }, + { + "bbox": [ + 181, + 252, + 282, + 265 + ], + "score": 0.91, + "content": "\\alpha \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 251, + 300, + 266 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 301, + 252, + 411, + 265 + ], + "score": 0.92, + "content": "\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 251, + 506, + 266 + ], + "score": 1.0, + "content": ". The results, shown in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 398, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 398, + 276 + ], + "score": 1.0, + "content": "Figures 6a, and 6b show that Eve again improves over all other methods.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 291, + 195, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 197, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 197, + 307 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "score": 1.0, + "content": "We proposed a simple and efficient method for incorporating feedback in to stochastic gradient", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 327, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 506, + 339 + ], + "score": 1.0, + "content": "descent algorithms. We used this method to create Eve, a modified version of the Adam algorithm.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "Experiments with a variety of models showed that the proposed method can help improve the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 349, + 260, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 260, + 361 + ], + "score": 1.0, + "content": "optimization of deep learning models.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 108, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "For future work, we would look to theoretically analyze our method and its effects. While we have", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "tried to evaluate our algorithm Eve on a variety of tasks, additional experiments on larger scale", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "problems would further highlight the strength of our approach. We are making code for our method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 399, + 439, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 439, + 410 + ], + "score": 1.0, + "content": "and the experiments publicly available to encourage more research on this method.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 107, + 426, + 175, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 176, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 176, + 440 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 503, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "Alan J Bray and David S Dean. Statistics of critical points of gaussian fields on large-dimensional", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 456, + 333, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 456, + 333, + 468 + ], + "score": 1.0, + "content": "spaces. Physical review letters, 98(15):150201, 2007.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 503, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 485, + 497, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 485, + 497, + 497 + ], + "score": 1.0, + "content": "gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 504, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 525, + 483, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 525, + 483, + 538 + ], + "score": 1.0, + "content": "optimization. In Advances in neural information processing systems, pp. 2933–2941, 2014.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 555, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 116, + 555, + 482, + 567 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 108, + 573, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 584, + 324, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 584, + 324, + 596 + ], + "score": 1.0, + "content": "networks. In Aistats, volume 9, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 108, + 603, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "Kenji Kawaguchi. Deep learning without poor local minima. In Advances in Neural Information", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 613, + 297, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 613, + 297, + 627 + ], + "score": 1.0, + "content": "Processing Systems (NIPS), 2016. to appear.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 504, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 645 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 643, + 214, + 654 + ], + "spans": [ + { + "bbox": [ + 115, + 643, + 214, + 654 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 105, + 661, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 661, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 506, + 675 + ], + "score": 1.0, + "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 691, + 469, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 469, + 703 + ], + "score": 1.0, + "content": "corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of conver-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 397, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 397, + 732 + ], + "score": 1.0, + "content": "gence o (1/k2). In Doklady an SSSR, volume 269, pp. 543–547, 1983.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 106, + 82, + 506, + 116 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "We trained another RNN-based model for the question & answering task. Specifically, we chose", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "two question types among 20 types from the bAbI dataset (Weston et al., 2015), Q19 and Q14. The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "dataset consists of pairs of supporting story sentences and a question. Different types of pairs are said", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "to require different reasoning schemes. For our test case, Q19 and Q14 correspond to Path Finding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "and Time Reasoning respectively. We picked Q19 since it is reported to have the lowest baseline", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "score, and we picked Q14 randomly from the remaining questions. The model consisted of two parts,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "one for encoding story sentences and another for query. Both included an embedding layer with 256", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "hidden units, and 0.3 dropout. Next query word embeddings were fed into a GRU one token at a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 506, + 223 + ], + "score": 1.0, + "content": "time, to compute a sentence representation. Both story and query sequences were truncated to the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 233 + ], + "score": 1.0, + "content": "maximum sequence length of 100. Finally, the sequence of word embeddings from story sentences", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "and the repeated encoded representation of a query were combined together to serve as input for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 504, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 504, + 254 + ], + "score": 1.0, + "content": "each time step in another GRU, with 0.3 dropout. We searched for the learning rate and decay from", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 251, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 181, + 266 + ], + "score": 1.0, + "content": "a range of values,", + "type": "text" + }, + { + "bbox": [ + 181, + 252, + 282, + 265 + ], + "score": 0.91, + "content": "\\alpha \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 251, + 300, + 266 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 301, + 252, + 411, + 265 + ], + "score": 0.92, + "content": "\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 251, + 506, + 266 + ], + "score": 1.0, + "content": ". The results, shown in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 398, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 398, + 276 + ], + "score": 1.0, + "content": "Figures 6a, and 6b show that Eve again improves over all other methods.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 120, + 506, + 276 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 291, + 195, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 197, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 197, + 307 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "score": 1.0, + "content": "We proposed a simple and efficient method for incorporating feedback in to stochastic gradient", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 327, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 506, + 339 + ], + "score": 1.0, + "content": "descent algorithms. We used this method to create Eve, a modified version of the Adam algorithm.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "Experiments with a variety of models showed that the proposed method can help improve the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 349, + 260, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 260, + 361 + ], + "score": 1.0, + "content": "optimization of deep learning models.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 316, + 506, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "For future work, we would look to theoretically analyze our method and its effects. While we have", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "tried to evaluate our algorithm Eve on a variety of tasks, additional experiments on larger scale", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "problems would further highlight the strength of our approach. We are making code for our method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 399, + 439, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 439, + 410 + ], + "score": 1.0, + "content": "and the experiments publicly available to encourage more research on this method.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 366, + 505, + 410 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 426, + 175, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 176, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 176, + 440 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 503, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "Alan J Bray and David S Dean. Statistics of critical points of gaussian fields on large-dimensional", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 456, + 333, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 456, + 333, + 468 + ], + "score": 1.0, + "content": "spaces. Physical review letters, 98(15):150201, 2007.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 444, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 473, + 503, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 485, + 497, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 485, + 497, + 497 + ], + "score": 1.0, + "content": "gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 474, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 504, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Yann N Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 525, + 483, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 525, + 483, + 538 + ], + "score": 1.0, + "content": "optimization. In Advances in neural information processing systems, pp. 2933–2941, 2014.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 503, + 506, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 555, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 116, + 555, + 482, + 567 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 543, + 505, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 573, + 504, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 584, + 324, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 584, + 324, + 596 + ], + "score": 1.0, + "content": "networks. In Aistats, volume 9, pp. 249–256, 2010.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 573, + 505, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 603, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "Kenji Kawaguchi. Deep learning without poor local minima. In Advances in Neural Information", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 613, + 297, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 613, + 297, + 627 + ], + "score": 1.0, + "content": "Processing Systems (NIPS), 2016. to appear.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 603, + 505, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 504, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 645 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 643, + 214, + 654 + ], + "spans": [ + { + "bbox": [ + 115, + 643, + 214, + 654 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 630, + 506, + 654 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 661, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 661, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 506, + 675 + ], + "score": 1.0, + "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 661, + 506, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 691, + 469, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 469, + 703 + ], + "score": 1.0, + "content": "corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 680, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of conver-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 397, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 397, + 732 + ], + "score": 1.0, + "content": "gence o (1/k2). In Doklady an SSSR, volume 269, pp. 543–547, 1983.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 506, + 108 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 257, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 257, + 116 + ], + "score": 1.0, + "content": "Research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4(2), 2012.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 505, + 178 + ], + "score": 1.0, + "content": "Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 304, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 304, + 187 + ], + "score": 1.0, + "content": "tasks. arXiv preprint arXiv:1502.05698, 2015.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "score": 1.0, + "content": "Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 143, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 143, + 217 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 506, + 108 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 257, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 257, + 116 + ], + "score": 1.0, + "content": "Research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 506, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 147 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4(2), 2012.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 122, + 506, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 505, + 178 + ], + "score": 1.0, + "content": "Joulin, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 304, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 304, + 187 + ], + "score": 1.0, + "content": "tasks. arXiv preprint arXiv:1502.05698, 2015.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 153, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "score": 1.0, + "content": "Matthew D Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 143, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 143, + 217 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 193, + 507, + 217 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/r1WUqIceg/r1WUqIceg_model.json b/parse/train/r1WUqIceg/r1WUqIceg_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5e270e1a3a2a47a2ecd5cc78c3acd44c2b462d72 --- /dev/null +++ b/parse/train/r1WUqIceg/r1WUqIceg_model.json @@ -0,0 +1,13006 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 674, + 1303, + 674, + 1303, + 1010, + 398, + 1010 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1437, + 1404, + 1437, + 1404, + 1772, + 298, + 1772 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1130, + 1403, + 1130, + 1403, + 1312, + 299, + 1312 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1788, + 1403, + 1788, + 1403, + 1911, + 299, + 1911 + ], + "score": 0.969 + }, + { + "category_id": 0, + "poly": [ + 298, + 219, + 1251, + 219, + 1251, + 323, + 298, + 323 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 302, + 1329, + 1399, + 1329, + 1399, + 1422, + 302, + 1422 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 299, + 1927, + 1400, + 1927, + 1400, + 1989, + 299, + 1989 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 302, + 1062, + 573, + 1062, + 573, + 1097, + 302, + 1097 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 334, + 2006, + 535, + 2006, + 535, + 2033, + 334, + 2033 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 610, + 926, + 610, + 926, + 643, + 773, + 643 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.734 + }, + { + "category_id": 1, + "poly": [ + 318, + 375, + 768, + 375, + 768, + 408, + 318, + 408 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 314, + 410, + 817, + 410, + 817, + 529, + 314, + 529 + ], + "score": 0.616 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 216.0, + 1260.0, + 216.0, + 1260.0, + 273.0, + 293.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 272.0, + 649.0, + 272.0, + 649.0, + 329.0, + 295.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1059.0, + 579.0, + 1059.0, + 579.0, + 1106.0, + 294.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 2001.0, + 541.0, + 2001.0, + 541.0, + 2038.0, + 329.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 609.0, + 932.0, + 609.0, + 932.0, + 646.0, + 769.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 674.0, + 1305.0, + 674.0, + 1305.0, + 710.0, + 394.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 704.0, + 1306.0, + 704.0, + 1306.0, + 741.0, + 392.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 733.0, + 1308.0, + 733.0, + 1308.0, + 773.0, + 392.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 767.0, + 1307.0, + 767.0, + 1307.0, + 802.0, + 394.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 796.0, + 1305.0, + 796.0, + 1305.0, + 834.0, + 393.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 829.0, + 1307.0, + 829.0, + 1307.0, + 862.0, + 394.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 856.0, + 1304.0, + 856.0, + 1304.0, + 894.0, + 393.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 886.0, + 1305.0, + 886.0, + 1305.0, + 925.0, + 392.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 919.0, + 1307.0, + 919.0, + 1307.0, + 954.0, + 394.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 949.0, + 1306.0, + 949.0, + 1306.0, + 981.0, + 396.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 979.0, + 839.0, + 979.0, + 839.0, + 1015.0, + 394.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 1402.0, + 1438.0, + 1402.0, + 1472.0, + 293.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1466.0, + 1405.0, + 1466.0, + 1405.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1404.0, + 1529.0, + 1404.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1559.0, + 1407.0, + 1559.0, + 1407.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1588.0, + 1409.0, + 1588.0, + 1409.0, + 1627.0, + 293.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1618.0, + 1406.0, + 1618.0, + 1406.0, + 1655.0, + 293.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1653.0, + 1402.0, + 1653.0, + 1402.0, + 1684.0, + 297.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1681.0, + 1406.0, + 1681.0, + 1406.0, + 1714.0, + 293.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1747.0, + 294.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1739.0, + 487.0, + 1739.0, + 487.0, + 1775.0, + 295.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1129.0, + 1405.0, + 1129.0, + 1405.0, + 1164.0, + 295.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1196.0, + 293.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1191.0, + 1407.0, + 1191.0, + 1407.0, + 1225.0, + 293.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1255.0, + 293.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1249.0, + 1405.0, + 1249.0, + 1405.0, + 1290.0, + 292.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1284.0, + 1088.0, + 1284.0, + 1088.0, + 1315.0, + 295.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1786.0, + 1407.0, + 1786.0, + 1407.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1853.0, + 295.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1176.0, + 1882.0, + 1176.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1328.0, + 1405.0, + 1328.0, + 1405.0, + 1366.0, + 296.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1357.0, + 1405.0, + 1357.0, + 1405.0, + 1396.0, + 293.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1390.0, + 1324.0, + 1390.0, + 1324.0, + 1425.0, + 296.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1926.0, + 1404.0, + 1926.0, + 1404.0, + 1962.0, + 295.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1957.0, + 1404.0, + 1957.0, + 1404.0, + 1992.0, + 295.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 373.0, + 775.0, + 373.0, + 775.0, + 413.0, + 310.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 409.0, + 680.0, + 409.0, + 680.0, + 440.0, + 313.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 435.0, + 630.0, + 435.0, + 630.0, + 472.0, + 311.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 468.0, + 625.0, + 468.0, + 625.0, + 499.0, + 313.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 500.0, + 817.0, + 500.0, + 817.0, + 530.0, + 317.0, + 530.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 977, + 1405, + 977, + 1405, + 1251, + 298, + 1251 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1266, + 1405, + 1266, + 1405, + 1632, + 298, + 1632 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 504, + 1405, + 504, + 1405, + 961, + 298, + 961 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2033, + 298, + 2033 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1403, + 229, + 1403, + 353, + 299, + 353 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 301, + 419, + 586, + 419, + 586, + 455, + 301, + 455 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 300, + 1700, + 480, + 1700, + 480, + 1737, + 300, + 1737 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 1783, + 526, + 1783, + 526, + 1815, + 300, + 1815 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 856, + 76, + 856, + 104, + 298, + 104 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.68 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.197 + }, + { + "category_id": 13, + "poly": [ + 395, + 1882, + 416, + 1882, + 416, + 1914, + 395, + 1914 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1852, + 1046, + 1852, + 1046, + 1883, + 1028, + 1883 + ], + "score": 0.82, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 628, + 1944, + 661, + 1944, + 661, + 1974, + 628, + 1974 + ], + "score": 0.74, + "latex": "\\boldsymbol { L } _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 671, + 1944, + 706, + 1944, + 706, + 1973, + 671, + 1973 + ], + "score": 0.73, + "latex": "L _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 416.0, + 593.0, + 416.0, + 593.0, + 463.0, + 291.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1695.0, + 486.0, + 1695.0, + 486.0, + 1745.0, + 290.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1783.0, + 530.0, + 1783.0, + 530.0, + 1819.0, + 295.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 976.0, + 1403.0, + 976.0, + 1403.0, + 1011.0, + 295.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1005.0, + 1405.0, + 1005.0, + 1405.0, + 1041.0, + 295.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1039.0, + 1402.0, + 1039.0, + 1402.0, + 1071.0, + 296.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1105.0, + 295.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1098.0, + 1405.0, + 1098.0, + 1405.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1128.0, + 1406.0, + 1128.0, + 1406.0, + 1163.0, + 293.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 1408.0, + 1159.0, + 1408.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1191.0, + 1403.0, + 1191.0, + 1403.0, + 1227.0, + 295.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1222.0, + 686.0, + 1222.0, + 686.0, + 1254.0, + 296.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1266.0, + 1405.0, + 1266.0, + 1405.0, + 1301.0, + 293.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1299.0, + 1407.0, + 1299.0, + 1407.0, + 1332.0, + 293.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1363.0, + 295.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1359.0, + 1410.0, + 1359.0, + 1410.0, + 1397.0, + 295.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1386.0, + 1410.0, + 1386.0, + 1410.0, + 1426.0, + 292.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1417.0, + 1406.0, + 1417.0, + 1406.0, + 1454.0, + 293.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1451.0, + 1405.0, + 1451.0, + 1405.0, + 1485.0, + 295.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1515.0, + 293.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1410.0, + 1511.0, + 1410.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 1405.0, + 1542.0, + 1405.0, + 1576.0, + 295.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1573.0, + 1403.0, + 1573.0, + 1403.0, + 1607.0, + 295.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1602.0, + 687.0, + 1602.0, + 687.0, + 1635.0, + 295.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 502.0, + 1407.0, + 502.0, + 1407.0, + 536.0, + 295.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 534.0, + 1403.0, + 534.0, + 1403.0, + 568.0, + 295.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 563.0, + 1406.0, + 563.0, + 1406.0, + 601.0, + 292.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 596.0, + 1405.0, + 596.0, + 1405.0, + 631.0, + 295.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 621.0, + 1406.0, + 621.0, + 1406.0, + 663.0, + 292.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 654.0, + 1406.0, + 654.0, + 1406.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 686.0, + 1405.0, + 686.0, + 1405.0, + 721.0, + 295.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 718.0, + 1403.0, + 718.0, + 1403.0, + 752.0, + 295.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 747.0, + 1403.0, + 747.0, + 1403.0, + 782.0, + 295.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 779.0, + 1403.0, + 779.0, + 1403.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 843.0, + 295.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 838.0, + 1403.0, + 838.0, + 1403.0, + 874.0, + 292.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 872.0, + 1402.0, + 872.0, + 1402.0, + 903.0, + 296.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 900.0, + 1405.0, + 900.0, + 1405.0, + 935.0, + 295.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 930.0, + 618.0, + 930.0, + 618.0, + 964.0, + 295.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1027.0, + 1848.0, + 1027.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1885.0, + 1047.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1882.0, + 394.0, + 1882.0, + 394.0, + 1914.0, + 296.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1914.0, + 417.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 627.0, + 1940.0, + 627.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1940.0, + 670.0, + 1940.0, + 670.0, + 1978.0, + 662.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1978.0, + 707.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1226.0, + 2001.0, + 1226.0, + 2035.0, + 293.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 263.0, + 295.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 294.0, + 295.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 323.0, + 1246.0, + 323.0, + 1246.0, + 355.0, + 295.0, + 355.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1671, + 1406, + 1671, + 1406, + 2035, + 296, + 2035 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 288, + 345, + 1397, + 345, + 1397, + 1554, + 288, + 1554 + ], + "score": 0.908, + "html": "
Require: α: learning rate
Require: β1, β2 ∈ [0,1): exponential decay rates for moment estimation in Adam
Require: β3 ∈ [0,1): exponential decay rate for computing relative change
Require: k,K: lower and upper threshold for relative change
Require: ε: fuzz factor
Require: f(θ): objective function
Require: Oo: initial value for parameters
mo = vo =0
do=1
f-1=t=0
while stopping condition is not reached do
t←t+1
gt←Vθf(0t-1)
mt←βimt-1+(1-β1)gt mt
mt←(1-β)
Ut←β2Ut-1+(1-β2)g² Ut
Ut←(1-)
if t>1 then
if f(0t-1) ≥ft-2 then δt←k+1
△t←K+1
else 1
8t←K+
△t←
end if
ct ←min{max{δt, }}
Ft-1←ctft-2
Ft-1-ft-21 rt↑
min{ft-1,ft-2} dt←βgdt-1+(1-β)rt
else
ft-1←f(0t-1)
dt←1
end if
0t←0t-1-Qdv+e mt
" + }, + { + "category_id": 0, + "poly": [ + 299, + 1613, + 517, + 1613, + 517, + 1644, + 299, + 1644 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 1403, + 227, + 1403, + 352, + 297, + 352 + ], + "score": 0.85 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.607 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 299, + 360, + 594, + 360, + 594, + 389, + 299, + 389 + ], + "score": 0.091 + }, + { + "category_id": 13, + "poly": [ + 903, + 1733, + 1227, + 1733, + 1227, + 1781, + 903, + 1781 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { r _ { t } = { \\frac { f _ { t - 2 } - \\bar { \\cdot } f _ { t - 1 } } { f _ { t - 1 } } } = { \\frac { f _ { t - 2 } - f _ { t - 1 } } { f _ { t - 1 } } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 345, + 1776, + 459, + 1776, + 459, + 1824, + 345, + 1824 + ], + "score": 0.93, + "latex": "\\frac { f _ { t - 1 } - f _ { t - 2 } } { f _ { t - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1972, + 543, + 1972, + 543, + 2006, + 298, + 2006 + ], + "score": 0.92, + "latex": "\\operatorname* { m i n } \\{ \\operatorname* { m a x } \\{ k , d _ { t } \\} , K \\}" + }, + { + "category_id": 13, + "poly": [ + 689, + 1881, + 804, + 1881, + 804, + 1915, + 689, + 1915 + ], + "score": 0.92, + "latex": "\\beta \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1043, + 1852, + 1125, + 1852, + 1125, + 1881, + 1043, + 1881 + ], + "score": 0.91, + "latex": "d _ { 1 } = 1" + }, + { + "category_id": 13, + "poly": [ + 329, + 1881, + 611, + 1881, + 611, + 1914, + 329, + 1914 + ], + "score": 0.91, + "latex": "d _ { t } = \\beta d _ { t - 1 } + ( 1 - \\beta ) r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 667, + 1704, + 721, + 1704, + 721, + 1736, + 667, + 1736 + ], + "score": 0.91, + "latex": "f _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 562, + 1705, + 616, + 1705, + 616, + 1736, + 562, + 1736 + ], + "score": 0.91, + "latex": "f _ { t - 2 }" + }, + { + "category_id": 13, + "poly": [ + 753, + 1973, + 892, + 1973, + 892, + 2001, + 753, + 2001 + ], + "score": 0.9, + "latex": "0 < k < K" + }, + { + "category_id": 13, + "poly": [ + 1226, + 1852, + 1293, + 1852, + 1293, + 1880, + 1226, + 1880 + ], + "score": 0.9, + "latex": "t > 1" + }, + { + "category_id": 13, + "poly": [ + 415, + 389, + 578, + 389, + 578, + 421, + 415, + 421 + ], + "score": 0.9, + "latex": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 415, + 510, + 469, + 510, + 469, + 543, + 415, + 543 + ], + "score": 0.9, + "latex": "f ( \\theta )" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1738, + 1396, + 1738, + 1396, + 1775, + 1255, + 1775 + ], + "score": 0.9, + "latex": "f _ { t - 2 } \\geq f _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 415, + 420, + 537, + 420, + 537, + 451, + 415, + 451 + ], + "score": 0.9, + "latex": "\\beta _ { 3 } \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 386, + 259, + 506, + 259, + 506, + 290, + 386, + 290 + ], + "score": 0.9, + "latex": "\\alpha = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 785, + 259, + 898, + 259, + 898, + 289, + 785, + 289 + ], + "score": 0.88, + "latex": "\\epsilon = 1 0 ^ { - 8 }" + }, + { + "category_id": 13, + "poly": [ + 414, + 451, + 472, + 451, + 472, + 480, + 414, + 480 + ], + "score": 0.88, + "latex": "k , K" + }, + { + "category_id": 13, + "poly": [ + 879, + 291, + 982, + 291, + 982, + 319, + 879, + 319 + ], + "score": 0.87, + "latex": "K = 1 0" + }, + { + "category_id": 13, + "poly": [ + 608, + 292, + 750, + 292, + 750, + 323, + 608, + 323 + ], + "score": 0.87, + "latex": "\\beta _ { 3 } = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 1130, + 2004, + 1157, + 2004, + 1157, + 2034, + 1130, + 2034 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 518, + 260, + 627, + 260, + 627, + 292, + 518, + 292 + ], + "score": 0.87, + "latex": "\\beta _ { 1 } = 0 . 9" + }, + { + "category_id": 13, + "poly": [ + 1341, + 1944, + 1369, + 1944, + 1369, + 1973, + 1341, + 1973 + ], + "score": 0.86, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 433, + 922, + 612, + 922, + 612, + 960, + 433, + 960 + ], + "score": 0.86, + "latex": "f ( \\widehat { \\theta } _ { t - 1 } ) \\geq \\widehat { f } _ { t - 2 }" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1851, + 1400, + 1851, + 1400, + 1881, + 1373, + 1881 + ], + "score": 0.86, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 408, + 1914, + 434, + 1914, + 434, + 1942, + 408, + 1942 + ], + "score": 0.86, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 638, + 260, + 774, + 260, + 774, + 292, + 638, + 292 + ], + "score": 0.86, + "latex": "\\beta _ { 2 } = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 343, + 1740, + 400, + 1740, + 400, + 1769, + 343, + 1769 + ], + "score": 0.86, + "latex": "t - 1" + }, + { + "category_id": 13, + "poly": [ + 763, + 291, + 864, + 291, + 864, + 320, + 763, + 320 + ], + "score": 0.84, + "latex": "k = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 414, + 543, + 443, + 543, + 443, + 571, + 414, + 571 + ], + "score": 0.83, + "latex": "\\theta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1704, + 1402, + 1704, + 1402, + 1733, + 1340, + 1733 + ], + "score": 0.83, + "latex": "t - 2" + }, + { + "category_id": 13, + "poly": [ + 391, + 901, + 454, + 901, + 454, + 925, + 391, + 925 + ], + "score": 0.79, + "latex": "t > 1" + }, + { + "category_id": 14, + "poly": [ + 320, + 572, + 480, + 572, + 480, + 670, + 320, + 670 + ], + "score": 0.75, + "latex": "\\begin{array} { l } { { m _ { 0 } = v _ { 0 } = 0 } } \\\\ { { d _ { 0 } = 1 } } \\\\ { { \\widehat { f } _ { - 1 } = t = 0 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 415, + 364, + 436, + 364, + 436, + 386, + 415, + 386 + ], + "score": 0.7, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 414, + 485, + 433, + 485, + 433, + 507, + 414, + 507 + ], + "score": 0.6, + "latex": "\\epsilon { : }" + }, + { + "category_id": 14, + "poly": [ + 366, + 1436, + 632, + 1436, + 632, + 1484, + 366, + 1484 + ], + "score": 0.46, + "latex": "\\begin{array} { r } { \\widehat { \\theta } _ { t } \\gets \\theta _ { t - 1 } - \\alpha \\frac { \\widehat { m } _ { t } } { d _ { t } \\sqrt { v _ { t } } + \\epsilon } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 442, + 960, + 595, + 960, + 595, + 988, + 442, + 988 + ], + "score": 0.46, + "latex": "\\delta _ { t } \\gets k + 1" + }, + { + "category_id": 13, + "poly": [ + 407, + 1510, + 434, + 1510, + 434, + 1541, + 407, + 1541 + ], + "score": 0.41, + "latex": "\\theta _ { t }" + }, + { + "category_id": 14, + "poly": [ + 358, + 697, + 695, + 697, + 695, + 899, + 358, + 899 + ], + "score": 0.35, + "latex": "\\begin{array} { r l } & { t \\gets t + 1 } \\\\ & { g _ { t } \\gets \\nabla _ { \\theta } f ( \\theta _ { t - 1 } ) } \\\\ & { m _ { t } \\gets \\beta _ { 1 } m _ { t - 1 } + ( 1 - \\beta _ { 1 } ) g _ { t } } \\\\ & { \\widehat { m } _ { t } \\gets \\frac { m _ { t } } { ( 1 - \\beta _ { 1 } ^ { t } ) } } \\\\ & { v _ { t } \\gets \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) g _ { t } ^ { 2 } } \\\\ & { \\widehat { v } _ { t } \\gets \\frac { v _ { t } } { ( 1 - \\beta _ { 2 } ^ { t } ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 366, + 862, + 510, + 862, + 510, + 901, + 366, + 901 + ], + "score": 0.3, + "latex": "\\begin{array} { r } { \\widehat { v } _ { t } \\gets \\frac { v _ { t } } { ( 1 - \\beta _ { 2 } ^ { t } ) } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 401, + 1342, + 665, + 1342, + 665, + 1414, + 401, + 1414 + ], + "score": 0.27, + "latex": "\\mathop { \\frown } _ { \\hat { f } _ { t - 1 } } \\gets f ( \\theta _ { t - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 365, + 1435, + 632, + 1435, + 632, + 1484, + 365, + 1484 + ], + "score": 0.26, + "latex": "\\begin{array} { r } { \\widehat { \\theta _ { t } } \\theta _ { t - 1 } - \\alpha \\frac { \\widehat { m } _ { t } } { d _ { t } \\sqrt { \\widehat { v } _ { t } } + \\epsilon } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1610.0, + 521.0, + 1610.0, + 521.0, + 1649.0, + 294.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 1404.0, + 1671.0, + 1404.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 561.0, + 1702.0, + 561.0, + 1739.0, + 294.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1702.0, + 666.0, + 1702.0, + 666.0, + 1739.0, + 617.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1702.0, + 1339.0, + 1702.0, + 1339.0, + 1739.0, + 722.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1702.0, + 1407.0, + 1702.0, + 1407.0, + 1739.0, + 1403.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1754.0, + 1200.0, + 1754.0, + 1200.0, + 1788.0, + 1141.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1782.0, + 344.0, + 1782.0, + 344.0, + 1814.0, + 292.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1817.0, + 1407.0, + 1817.0, + 1407.0, + 1857.0, + 291.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1042.0, + 1851.0, + 1042.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1851.0, + 1225.0, + 1851.0, + 1225.0, + 1884.0, + 1126.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1851.0, + 1372.0, + 1851.0, + 1372.0, + 1884.0, + 1294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1884.0, + 1401.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 328.0, + 1882.0, + 328.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1882.0, + 688.0, + 1882.0, + 688.0, + 1916.0, + 612.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1916.0, + 805.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 407.0, + 1912.0, + 407.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1912.0, + 1407.0, + 1912.0, + 1407.0, + 1946.0, + 435.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1340.0, + 1939.0, + 1340.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1979.0, + 1370.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 1973.0, + 752.0, + 1973.0, + 752.0, + 2006.0, + 544.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2006.0, + 893.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1129.0, + 2002.0, + 1129.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2039.0, + 1158.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1731.5, + 1409.0, + 1731.5, + 1409.0, + 1774.0, + 1106.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.25, + 1730.0, + 1089.25, + 1730.0, + 1089.25, + 1779.0, + 287.25, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1774.0, + 1411.0, + 1774.0, + 1411.0, + 1826.0, + 331.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 261.0, + 296.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 257.0, + 385.0, + 257.0, + 385.0, + 294.0, + 292.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 257.0, + 517.0, + 257.0, + 517.0, + 294.0, + 507.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 257.0, + 637.0, + 257.0, + 637.0, + 294.0, + 628.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 257.0, + 784.0, + 257.0, + 784.0, + 294.0, + 775.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 257.0, + 1407.0, + 257.0, + 1407.0, + 294.0, + 899.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 291.0, + 607.0, + 291.0, + 607.0, + 323.0, + 293.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 291.0, + 762.0, + 291.0, + 762.0, + 323.0, + 751.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 291.0, + 878.0, + 291.0, + 878.0, + 323.0, + 865.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 291.0, + 1404.0, + 291.0, + 1404.0, + 323.0, + 983.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 322.0, + 448.0, + 322.0, + 448.0, + 353.0, + 296.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 353.0, + 414.0, + 353.0, + 414.0, + 397.0, + 293.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 353.0, + 597.0, + 353.0, + 597.0, + 397.0, + 437.0, + 397.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 730, + 1404, + 730, + 1404, + 981, + 296, + 981 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1193, + 1406, + 1193, + 1406, + 1378, + 297, + 1378 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 994, + 1404, + 994, + 1404, + 1180, + 297, + 1180 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1495, + 1404, + 1495, + 1404, + 1651, + 298, + 1651 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 306, + 220, + 1393, + 220, + 1393, + 586, + 306, + 586 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1856, + 1403, + 1856, + 1403, + 1949, + 299, + 1949 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 1664, + 1403, + 1664, + 1403, + 1758, + 299, + 1758 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 295, + 610, + 1404, + 610, + 1404, + 676, + 295, + 676 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 297, + 1978, + 1402, + 1978, + 1402, + 2033, + 297, + 2033 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 300, + 1425, + 558, + 1425, + 558, + 1460, + 300, + 1460 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 301, + 1797, + 831, + 1797, + 831, + 1830, + 301, + 1830 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2111, + 840, + 2111 + ], + "score": 0.784 + }, + { + "category_id": 13, + "poly": [ + 455, + 804, + 683, + 804, + 683, + 860, + 455, + 860 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\frac { f _ { t - 1 } } { \\widehat { f } _ { t - 2 } } \\in \\left[ \\frac { 1 } { K + 1 } , \\frac { 1 } { k + 1 } \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 885, + 804, + 1312, + 804, + 1312, + 861, + 885, + 861 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { c _ { t } = \\operatorname* { m i n } \\left\\{ \\operatorname* { m a x } \\left\\{ \\frac { 1 } { K + 1 } , \\frac { f _ { t } } { \\widehat { f } _ { t - 1 } } \\right\\} , \\frac { 1 } { k + 1 } \\right\\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 612, + 858, + 726, + 858, + 726, + 915, + 612, + 915 + ], + "score": 0.93, + "latex": "\\frac { \\widehat { f } _ { t - 2 } - \\widehat { f } _ { t - 1 } } { \\widehat { f } _ { t - 1 } }" + }, + { + "category_id": 13, + "poly": [ + 554, + 1026, + 684, + 1026, + 684, + 1058, + 554, + 1058 + ], + "score": 0.93, + "latex": "\\alpha _ { t } = \\alpha / d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 355, + 768, + 450, + 768, + 450, + 806, + 355, + 806 + ], + "score": 0.93, + "latex": "{ \\widehat { f } } _ { 0 } = f _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 861, + 462, + 861, + 462, + 902, + 297, + 902 + ], + "score": 0.92, + "latex": "\\widehat { f } _ { t - 1 } = c _ { t } \\widehat { f } _ { t - 2 }" + }, + { + "category_id": 13, + "poly": [ + 765, + 1588, + 955, + 1588, + 955, + 1622, + 765, + 1622 + ], + "score": 0.92, + "latex": "\\alpha _ { t } = \\alpha / ( 1 + \\gamma t )" + }, + { + "category_id": 13, + "poly": [ + 668, + 767, + 814, + 767, + 814, + 807, + 668, + 807 + ], + "score": 0.92, + "latex": "f _ { t - 1 } \\leq \\widehat { f } _ { t - 2 }" + }, + { + "category_id": 13, + "poly": [ + 593, + 730, + 648, + 730, + 648, + 769, + 593, + 769 + ], + "score": 0.92, + "latex": "\\widehat { f } _ { t - 2 }" + }, + { + "category_id": 13, + "poly": [ + 845, + 1087, + 925, + 1087, + 925, + 1117, + 845, + 1117 + ], + "score": 0.92, + "latex": "d _ { 0 } = 1" + }, + { + "category_id": 13, + "poly": [ + 298, + 910, + 440, + 910, + 440, + 950, + 298, + 950 + ], + "score": 0.92, + "latex": "f _ { t - 1 } > \\widehat { f } _ { t - 2 }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1316, + 1397, + 1316, + 1397, + 1348, + 1266, + 1348 + ], + "score": 0.91, + "latex": "\\beta _ { 3 } = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 729, + 1287, + 779, + 1287, + 779, + 1320, + 729, + 1320 + ], + "score": 0.91, + "latex": "\\sqrt { v _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1096, + 773, + 1245, + 773, + 1245, + 804, + 1096, + 804 + ], + "score": 0.9, + "latex": "k \\leq r _ { t } \\leq K" + }, + { + "category_id": 13, + "poly": [ + 446, + 1347, + 538, + 1347, + 538, + 1375, + 446, + 1375 + ], + "score": 0.9, + "latex": "K = 1 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 1347, + 387, + 1347, + 387, + 1376, + 297, + 1376 + ], + "score": 0.89, + "latex": "k = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1119, + 1039, + 1119, + 1039, + 1148, + 1011, + 1148 + ], + "score": 0.88, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 837, + 1287, + 864, + 1287, + 864, + 1317, + 837, + 1317 + ], + "score": 0.88, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 891, + 1027, + 918, + 1027, + 918, + 1056, + 891, + 1056 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1057, + 1243, + 1057, + 1243, + 1087, + 1216, + 1087 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 919, + 1155, + 919, + 1155, + 948, + 1128, + 948 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 362, + 996, + 390, + 996, + 390, + 1026, + 362, + 1026 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 783, + 1318, + 811, + 1318, + 811, + 1344, + 783, + 1344 + ], + "score": 0.82, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 557, + 871, + 583, + 871, + 583, + 899, + 557, + 899 + ], + "score": 0.78, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 494, + 1031, + 516, + 1031, + 516, + 1054, + 494, + 1054 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 965, + 1592, + 990, + 1592, + 990, + 1621, + 965, + 1621 + ], + "score": 0.71, + "latex": "( \\gamma" + }, + { + "category_id": 13, + "poly": [ + 1340, + 738, + 1402, + 738, + 1402, + 766, + 1340, + 766 + ], + "score": 0.7, + "latex": "t - 2" + }, + { + "category_id": 13, + "poly": [ + 567, + 1918, + 614, + 1918, + 614, + 1946, + 567, + 1946 + ], + "score": 0.67, + "latex": "2 \\mathbf { x } 2" + }, + { + "category_id": 13, + "poly": [ + 666, + 1318, + 696, + 1318, + 696, + 1348, + 666, + 1348 + ], + "score": 0.59, + "latex": "\\beta _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 880, + 293, + 931, + 293, + 931, + 310, + 880, + 310 + ], + "score": 0.53, + "latex": "2 . 6 \\times 1 0 ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 881, + 363, + 930, + 363, + 930, + 376, + 881, + 376 + ], + "score": 0.53, + "latex": "1 . 6 \\times 1 0 ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 327, + 428, + 385, + 428, + 385, + 442, + 327, + 442 + ], + "score": 0.49, + "latex": "5 . 4 \\times 1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1319, + 726, + 1319, + 726, + 1346, + 707, + 1346 + ], + "score": 0.47, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 311, + 359, + 384, + 359, + 384, + 382, + 311, + 382 + ], + "score": 0.38, + "latex": "\\stackrel { \\mathcal { U } } { \\leq } 1 . 6 \\times 1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1888, + 1248, + 1888, + 1248, + 1916, + 1203, + 1916 + ], + "score": 0.37, + "latex": "3 \\mathrm { x } 3" + }, + { + "category_id": 13, + "poly": [ + 873, + 427, + 931, + 427, + 931, + 443, + 873, + 443 + ], + "score": 0.34, + "latex": "9 . 8 \\times 1 0 ^ { - 1 }" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 223.0, + 390.0, + 223.0, + 390.0, + 244.0, + 331.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 223.0, + 936.0, + 223.0, + 936.0, + 244.0, + 875.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 240.0, + 770.0, + 240.0, + 770.0, + 262.0, + 737.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 240.0, + 1316.0, + 240.0, + 1316.0, + 262.0, + 1282.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 257.0, + 782.0, + 257.0, + 782.0, + 281.0, + 737.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 260.0, + 1327.0, + 260.0, + 1327.0, + 280.0, + 1282.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 277.0, + 794.0, + 277.0, + 794.0, + 297.0, + 737.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 278.0, + 1340.0, + 278.0, + 1340.0, + 297.0, + 1285.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 290.0, + 390.0, + 290.0, + 390.0, + 310.0, + 321.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 293.0, + 826.0, + 293.0, + 826.0, + 316.0, + 736.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 290.0, + 936.0, + 290.0, + 936.0, + 311.0, + 932.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 293.0, + 1371.0, + 293.0, + 1371.0, + 316.0, + 1280.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 308.0, + 803.0, + 308.0, + 803.0, + 335.0, + 735.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 330.0, + 775.0, + 330.0, + 775.0, + 351.0, + 736.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 351.0, + 389.0, + 351.0, + 389.0, + 385.0, + 385.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 346.0, + 798.0, + 346.0, + 798.0, + 386.0, + 736.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 354.0, + 880.0, + 354.0, + 880.0, + 385.0, + 856.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 354.0, + 937.0, + 354.0, + 937.0, + 385.0, + 931.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 423.0, + 326.0, + 423.0, + 326.0, + 446.0, + 320.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 423.0, + 391.0, + 423.0, + 391.0, + 446.0, + 386.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 425.0, + 872.0, + 425.0, + 872.0, + 445.0, + 866.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 425.0, + 936.0, + 425.0, + 936.0, + 445.0, + 932.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 492.0, + 936.0, + 492.0, + 936.0, + 512.0, + 866.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 509.0, + 401.0, + 509.0, + 401.0, + 524.0, + 386.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 507.0, + 495.0, + 507.0, + 495.0, + 526.0, + 467.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 507.0, + 582.0, + 507.0, + 582.0, + 526.0, + 555.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 505.0, + 671.0, + 505.0, + 671.0, + 528.0, + 641.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 507.0, + 758.0, + 507.0, + 758.0, + 526.0, + 731.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 507.0, + 846.0, + 507.0, + 846.0, + 526.0, + 819.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 509.0, + 947.0, + 509.0, + 947.0, + 524.0, + 931.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 507.0, + 1037.0, + 507.0, + 1037.0, + 526.0, + 1016.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 506.0, + 1126.0, + 506.0, + 1126.0, + 526.0, + 1103.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 506.0, + 1214.0, + 506.0, + 1214.0, + 526.0, + 1191.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 507.0, + 1301.0, + 507.0, + 1301.0, + 526.0, + 1279.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 507.0, + 1392.0, + 507.0, + 1392.0, + 526.0, + 1364.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 521.0, + 637.0, + 521.0, + 637.0, + 545.0, + 589.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 520.0, + 1182.0, + 520.0, + 1182.0, + 545.0, + 1134.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 556.0, + 647.0, + 556.0, + 647.0, + 584.0, + 505.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 556.0, + 1200.0, + 556.0, + 1200.0, + 585.0, + 1045.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.75, + 489.0, + 392.75, + 489.0, + 392.75, + 514.5, + 320.75, + 514.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 611.0, + 1405.0, + 611.0, + 1405.0, + 647.0, + 295.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 1342.0, + 642.0, + 1342.0, + 678.0, + 295.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2010.0, + 332.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 716.0, + 2004.0, + 716.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1425.0, + 561.0, + 1425.0, + 561.0, + 1465.0, + 293.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1794.0, + 835.0, + 1794.0, + 835.0, + 1834.0, + 293.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 837.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 732.0, + 592.0, + 732.0, + 592.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 732.0, + 1339.0, + 732.0, + 1339.0, + 770.0, + 649.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 732.0, + 1406.0, + 732.0, + 1406.0, + 770.0, + 1403.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 769.0, + 354.0, + 769.0, + 354.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 769.0, + 667.0, + 769.0, + 667.0, + 810.0, + 451.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 769.0, + 1095.0, + 769.0, + 1095.0, + 810.0, + 815.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 769.0, + 1405.0, + 769.0, + 1405.0, + 810.0, + 1246.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 805.0, + 454.0, + 805.0, + 454.0, + 866.0, + 295.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 801.0, + 884.0, + 801.0, + 884.0, + 874.0, + 684.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 801.0, + 1416.0, + 801.0, + 1416.0, + 874.0, + 1313.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 845.0, + 296.0, + 845.0, + 296.0, + 919.0, + 286.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 845.0, + 556.0, + 845.0, + 556.0, + 919.0, + 463.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 845.0, + 611.0, + 845.0, + 611.0, + 919.0, + 584.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 845.0, + 1415.0, + 845.0, + 1415.0, + 919.0, + 727.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 908.0, + 297.0, + 908.0, + 297.0, + 955.0, + 291.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 908.0, + 1127.0, + 908.0, + 1127.0, + 955.0, + 441.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 908.0, + 1407.0, + 908.0, + 1407.0, + 955.0, + 1156.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 963.0, + 948.0, + 963.0, + 982.0, + 295.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1190.0, + 1409.0, + 1190.0, + 1409.0, + 1232.0, + 292.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1408.0, + 1223.0, + 1408.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1256.0, + 1404.0, + 1256.0, + 1404.0, + 1288.0, + 296.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1285.0, + 728.0, + 1285.0, + 728.0, + 1321.0, + 296.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1285.0, + 836.0, + 1285.0, + 836.0, + 1321.0, + 780.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1285.0, + 1406.0, + 1285.0, + 1406.0, + 1321.0, + 865.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1316.0, + 665.0, + 1316.0, + 665.0, + 1352.0, + 295.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1316.0, + 706.0, + 1316.0, + 706.0, + 1352.0, + 697.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1316.0, + 782.0, + 1316.0, + 782.0, + 1352.0, + 727.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1316.0, + 1265.0, + 1316.0, + 1265.0, + 1352.0, + 812.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1316.0, + 1408.0, + 1316.0, + 1408.0, + 1352.0, + 1398.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1344.0, + 296.0, + 1344.0, + 296.0, + 1382.0, + 292.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1344.0, + 445.0, + 1344.0, + 445.0, + 1382.0, + 388.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1344.0, + 1011.0, + 1344.0, + 1011.0, + 1382.0, + 539.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 991.0, + 361.0, + 991.0, + 361.0, + 1032.0, + 295.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 991.0, + 1405.0, + 991.0, + 1405.0, + 1032.0, + 391.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1023.0, + 493.0, + 1023.0, + 493.0, + 1061.0, + 294.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1023.0, + 553.0, + 1023.0, + 553.0, + 1061.0, + 517.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1023.0, + 890.0, + 1023.0, + 890.0, + 1061.0, + 685.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1023.0, + 1405.0, + 1023.0, + 1405.0, + 1061.0, + 919.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1054.0, + 1215.0, + 1054.0, + 1215.0, + 1092.0, + 292.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1054.0, + 1406.0, + 1054.0, + 1406.0, + 1092.0, + 1244.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1086.0, + 844.0, + 1086.0, + 844.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1086.0, + 1406.0, + 1086.0, + 1406.0, + 1121.0, + 926.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1115.0, + 1010.0, + 1115.0, + 1010.0, + 1153.0, + 294.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1115.0, + 1406.0, + 1115.0, + 1406.0, + 1153.0, + 1040.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1077.0, + 1146.0, + 1077.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1494.0, + 1406.0, + 1494.0, + 1406.0, + 1530.0, + 293.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1526.0, + 1405.0, + 1526.0, + 1405.0, + 1561.0, + 293.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1557.0, + 1406.0, + 1557.0, + 1406.0, + 1592.0, + 292.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1584.0, + 764.0, + 1584.0, + 764.0, + 1628.0, + 291.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1584.0, + 964.0, + 1584.0, + 964.0, + 1628.0, + 956.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1584.0, + 1405.0, + 1584.0, + 1405.0, + 1628.0, + 991.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1617.0, + 562.0, + 1617.0, + 562.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1855.0, + 1405.0, + 1855.0, + 1405.0, + 1893.0, + 294.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1888.0, + 1202.0, + 1888.0, + 1202.0, + 1922.0, + 295.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1888.0, + 1405.0, + 1888.0, + 1405.0, + 1922.0, + 1249.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 566.0, + 1918.0, + 566.0, + 1952.0, + 295.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1918.0, + 1407.0, + 1918.0, + 1407.0, + 1952.0, + 615.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1404.0, + 1665.0, + 1404.0, + 1699.0, + 294.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1695.0, + 1407.0, + 1695.0, + 1407.0, + 1730.0, + 293.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1725.0, + 482.0, + 1725.0, + 482.0, + 1763.0, + 295.0, + 1763.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1405, + 1758, + 1405, + 2034, + 297, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1466, + 1405, + 1466, + 1405, + 1745, + 297, + 1745 + ], + "score": 0.981 + }, + { + "category_id": 3, + "poly": [ + 302, + 225, + 1400, + 225, + 1400, + 646, + 302, + 646 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 535, + 875, + 1162, + 875, + 1162, + 1242, + 535, + 1242 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 297, + 673, + 1404, + 673, + 1404, + 768, + 297, + 768 + ], + "score": 0.954 + }, + { + "category_id": 4, + "poly": [ + 295, + 1283, + 1400, + 1283, + 1400, + 1348, + 295, + 1348 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.845 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.734 + }, + { + "category_id": 13, + "poly": [ + 298, + 1588, + 525, + 1588, + 525, + 1624, + 298, + 1624 + ], + "score": 0.91, + "latex": "\\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}" + }, + { + "category_id": 13, + "poly": [ + 660, + 1618, + 912, + 1618, + 912, + 1654, + 660, + 1654 + ], + "score": 0.91, + "latex": "\\{ 0 , 1 0 ^ { - 2 } , 1 \\breve { 0 } ^ { - 3 } , 1 0 ^ { - 4 } \\}" + }, + { + "category_id": 13, + "poly": [ + 591, + 1713, + 618, + 1713, + 618, + 1743, + 591, + 1743 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 776, + 676, + 804, + 676, + 804, + 706, + 776, + 706 + ], + "score": 0.86, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 555, + 1003, + 586, + 1003, + 586, + 1020, + 555, + 1020 + ], + "score": 0.56, + "latex": "1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 555, + 1065, + 586, + 1065, + 586, + 1081, + 555, + 1081 + ], + "score": 0.45, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 1302, + 1821, + 1402, + 1821, + 1402, + 1850, + 1302, + 1850 + ], + "score": 0.33, + "latex": "2 \\mathrm { x } 2 \\mathrm { m a x }" + }, + { + "category_id": 13, + "poly": [ + 798, + 1791, + 845, + 1791, + 845, + 1819, + 798, + 1819 + ], + "score": 0.33, + "latex": "3 \\mathrm { x } 3" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 230.0, + 343.0, + 230.0, + 343.0, + 251.0, + 324.0, + 251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 226.0, + 1056.0, + 226.0, + 1056.0, + 254.0, + 1019.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 300.0, + 1054.0, + 300.0, + 1054.0, + 329.0, + 1018.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 350.0, + 341.0, + 350.0, + 341.0, + 365.0, + 326.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 375.0, + 1057.0, + 375.0, + 1057.0, + 403.0, + 1018.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 398.0, + 1069.0, + 398.0, + 1069.0, + 419.0, + 1048.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 395.0, + 1138.0, + 395.0, + 1138.0, + 422.0, + 1109.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 397.0, + 1202.0, + 397.0, + 1202.0, + 420.0, + 1172.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 397.0, + 1267.0, + 397.0, + 1267.0, + 420.0, + 1237.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 397.0, + 1331.0, + 397.0, + 1331.0, + 420.0, + 1301.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 397.0, + 1394.0, + 397.0, + 1394.0, + 420.0, + 1366.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 410.0, + 323.0, + 410.0, + 323.0, + 425.0, + 307.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 430.0, + 1057.0, + 430.0, + 1057.0, + 459.0, + 1018.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 459.0, + 1054.0, + 459.0, + 1054.0, + 489.0, + 1018.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 489.0, + 1054.0, + 489.0, + 1054.0, + 519.0, + 1018.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 521.0, + 1054.0, + 521.0, + 1054.0, + 548.0, + 1018.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 549.0, + 1053.0, + 549.0, + 1053.0, + 578.0, + 1018.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 585.0, + 341.0, + 585.0, + 341.0, + 602.0, + 326.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 580.0, + 1057.0, + 580.0, + 1057.0, + 607.0, + 1019.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 603.0, + 357.0, + 603.0, + 357.0, + 621.0, + 341.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 601.0, + 495.0, + 601.0, + 495.0, + 625.0, + 460.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 601.0, + 624.0, + 601.0, + 624.0, + 625.0, + 588.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 600.0, + 755.0, + 600.0, + 755.0, + 626.0, + 715.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 601.0, + 881.0, + 601.0, + 881.0, + 625.0, + 846.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 598.0, + 1012.0, + 598.0, + 1012.0, + 626.0, + 973.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 598.0, + 1078.0, + 598.0, + 1078.0, + 626.0, + 1038.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 598.0, + 1143.0, + 598.0, + 1143.0, + 626.0, + 1103.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 598.0, + 1208.0, + 598.0, + 1208.0, + 626.0, + 1167.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 598.0, + 1272.0, + 598.0, + 1272.0, + 626.0, + 1232.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 600.0, + 1336.0, + 600.0, + 1336.0, + 626.0, + 1297.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 601.0, + 1396.0, + 601.0, + 1396.0, + 623.0, + 1363.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 617.0, + 702.0, + 617.0, + 702.0, + 651.0, + 639.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 618.0, + 1252.0, + 618.0, + 1252.0, + 648.0, + 1188.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 372.0, + 672.0, + 372.0, + 672.0, + 405.0, + 581.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.25, + 488.0, + 1126.25, + 488.0, + 1126.25, + 517.0, + 1063.25, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 434.5, + 611.0, + 434.5, + 611.0, + 467.5, + 417.0, + 467.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 875.0, + 591.0, + 875.0, + 591.0, + 901.0, + 560.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 936.0, + 592.0, + 936.0, + 592.0, + 962.0, + 560.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1023.0, + 559.0, + 1023.0, + 559.0, + 1063.0, + 538.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1060.0, + 554.0, + 1060.0, + 554.0, + 1085.0, + 548.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1060.0, + 592.0, + 1060.0, + 592.0, + 1085.0, + 587.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1204.0, + 604.0, + 1204.0, + 604.0, + 1224.0, + 588.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1202.0, + 755.0, + 1202.0, + 755.0, + 1225.0, + 707.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1202.0, + 895.0, + 1202.0, + 895.0, + 1225.0, + 841.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1202.0, + 1030.0, + 1202.0, + 1030.0, + 1225.0, + 978.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1202.0, + 1167.0, + 1202.0, + 1167.0, + 1225.0, + 1112.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1220.0, + 901.0, + 1220.0, + 901.0, + 1247.0, + 835.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 996.0, + 595.0, + 996.0, + 595.0, + 1024.5, + 548.0, + 1024.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.75, + 1120.0, + 595.75, + 1120.0, + 595.75, + 1149.0, + 547.75, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1182.0, + 598.0, + 1182.0, + 598.0, + 1209.5, + 547.0, + 1209.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 775.0, + 672.0, + 775.0, + 711.0, + 294.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 672.0, + 1406.0, + 672.0, + 1406.0, + 711.0, + 805.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 702.0, + 1405.0, + 702.0, + 1405.0, + 742.0, + 292.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 735.0, + 541.0, + 735.0, + 541.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1322.0, + 294.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 915.0, + 1312.0, + 915.0, + 1353.0, + 294.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1759.0, + 1408.0, + 1759.0, + 1408.0, + 1795.0, + 295.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 797.0, + 1788.0, + 797.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1788.0, + 1408.0, + 1788.0, + 1408.0, + 1824.0, + 846.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1819.0, + 1301.0, + 1819.0, + 1301.0, + 1856.0, + 295.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1819.0, + 1407.0, + 1819.0, + 1407.0, + 1856.0, + 1403.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1407.0, + 1848.0, + 1407.0, + 1887.0, + 292.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1883.0, + 1406.0, + 1883.0, + 1406.0, + 1916.0, + 296.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 413.0, + 2003.0, + 413.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1469.0, + 1407.0, + 1469.0, + 1407.0, + 1502.0, + 295.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 1405.0, + 1496.0, + 1405.0, + 1536.0, + 294.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1407.0, + 1528.0, + 1407.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1558.0, + 1405.0, + 1558.0, + 1405.0, + 1596.0, + 292.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1584.0, + 297.0, + 1584.0, + 297.0, + 1629.0, + 294.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1584.0, + 1407.0, + 1584.0, + 1407.0, + 1629.0, + 526.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1619.0, + 659.0, + 1619.0, + 659.0, + 1656.0, + 295.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1619.0, + 1405.0, + 1619.0, + 1405.0, + 1656.0, + 913.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1649.0, + 1407.0, + 1649.0, + 1407.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1681.0, + 1407.0, + 1681.0, + 1407.0, + 1717.0, + 295.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1708.0, + 590.0, + 1708.0, + 590.0, + 1752.0, + 292.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1708.0, + 1268.0, + 1708.0, + 1268.0, + 1752.0, + 619.0, + 1752.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1468, + 1405, + 1468, + 1405, + 1895, + 298, + 1895 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 299, + 1912, + 1402, + 1912, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 315, + 800, + 1380, + 800, + 1380, + 1214, + 315, + 1214 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 314, + 242, + 1387, + 242, + 1387, + 649, + 314, + 649 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 296, + 1256, + 1407, + 1256, + 1407, + 1350, + 296, + 1350 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 297, + 689, + 1405, + 689, + 1405, + 756, + 297, + 756 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 303, + 1410, + 792, + 1410, + 792, + 1442, + 303, + 1442 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.824 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 854, + 76, + 854, + 104, + 300, + 104 + ], + "score": 0.301 + }, + { + "category_id": 13, + "poly": [ + 1329, + 1254, + 1390, + 1254, + 1390, + 1286, + 1329, + 1286 + ], + "score": 0.9, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1683, + 1051, + 1683, + 1051, + 1713, + 1024, + 1713 + ], + "score": 0.88, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1775, + 1282, + 1775, + 1282, + 1804, + 1255, + 1804 + ], + "score": 0.88, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1531, + 1227, + 1531, + 1227, + 1560, + 1200, + 1560 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 820, + 692, + 848, + 692, + 848, + 723, + 820, + 723 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 818, + 1501, + 845, + 1501, + 845, + 1530, + 818, + 1530 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 870, + 1039, + 902, + 1039, + 902, + 1054, + 870, + 1054 + ], + "score": 0.87, + "latex": "1 0 ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 1374, + 1743, + 1400, + 1743, + 1400, + 1773, + 1374, + 1773 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 604, + 724, + 631, + 724, + 631, + 753, + 604, + 753 + ], + "score": 0.86, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 870, + 1001, + 903, + 1001, + 903, + 1016, + 870, + 1016 + ], + "score": 0.86, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 341, + 333, + 414, + 333, + 414, + 350, + 341, + 350 + ], + "score": 0.86, + "latex": "3 . 3 \\times 1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 709, + 1835, + 737, + 1835, + 737, + 1865, + 709, + 1865 + ], + "score": 0.86, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 870, + 1116, + 903, + 1116, + 903, + 1131, + 870, + 1131 + ], + "score": 0.83, + "latex": "1 0 ^ { - 7 }" + }, + { + "category_id": 13, + "poly": [ + 340, + 505, + 414, + 505, + 414, + 522, + 340, + 522 + ], + "score": 0.83, + "latex": "7 . 7 \\times 1 0 ^ { - 6 }" + }, + { + "category_id": 13, + "poly": [ + 870, + 924, + 902, + 924, + 902, + 940, + 870, + 940 + ], + "score": 0.83, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 870, + 963, + 902, + 963, + 902, + 979, + 870, + 979 + ], + "score": 0.82, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 870, + 1078, + 902, + 1078, + 902, + 1092, + 870, + 1092 + ], + "score": 0.81, + "latex": "1 0 ^ { - 6 }" + }, + { + "category_id": 13, + "poly": [ + 340, + 1097, + 373, + 1097, + 373, + 1112, + 340, + 1112 + ], + "score": 0.8, + "latex": "1 0 ^ { - 7 }" + }, + { + "category_id": 13, + "poly": [ + 870, + 886, + 902, + 886, + 902, + 902, + 870, + 902 + ], + "score": 0.79, + "latex": "1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 340, + 591, + 413, + 591, + 413, + 607, + 340, + 607 + ], + "score": 0.77, + "latex": "1 . 2 \\times 1 0 ^ { - 7 }" + }, + { + "category_id": 13, + "poly": [ + 340, + 1039, + 373, + 1039, + 373, + 1054, + 340, + 1054 + ], + "score": 0.72, + "latex": "1 0 ^ { - 6 }" + }, + { + "category_id": 13, + "poly": [ + 340, + 924, + 373, + 924, + 373, + 940, + 340, + 940 + ], + "score": 0.71, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 869, + 1153, + 902, + 1153, + 902, + 1169, + 869, + 1169 + ], + "score": 0.62, + "latex": "1 0 ^ { - 8 }" + }, + { + "category_id": 13, + "poly": [ + 349, + 245, + 413, + 245, + 413, + 263, + 349, + 263 + ], + "score": 0.6, + "latex": "2 . 2 \\times 1 0 ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 878, + 846, + 902, + 846, + 902, + 863, + 878, + 863 + ], + "score": 0.58, + "latex": "1 0 ^ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 340, + 808, + 374, + 808, + 374, + 825, + 340, + 825 + ], + "score": 0.58, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 732, + 935, + 814, + 935, + 814, + 955, + 732, + 955 + ], + "score": 0.52, + "latex": " { \\boldsymbol { \\kappa } } \\ 1 0 \\times 1 0 ^ { - 2 } )" + }, + { + "category_id": 13, + "poly": [ + 340, + 866, + 373, + 866, + 373, + 882, + 340, + 882 + ], + "score": 0.52, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 878, + 808, + 902, + 808, + 902, + 825, + 878, + 825 + ], + "score": 0.51, + "latex": "1 0 ^ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 338, + 418, + 414, + 418, + 414, + 438, + 338, + 438 + ], + "score": 0.5, + "latex": "5 . 1 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 498, + 1031, + 557, + 1031, + 557, + 1046, + 498, + 1046 + ], + "score": 0.49, + "latex": "\\alpha \\mathrm { ~ 1 ~ } \\times \\mathrm { ~ 1 0 ~ }" + }, + { + "category_id": 13, + "poly": [ + 318, + 413, + 335, + 413, + 335, + 444, + 318, + 444 + ], + "score": 0.49, + "latex": "\\underline { { \\hat { \\mathcal { G } } } }" + }, + { + "category_id": 13, + "poly": [ + 511, + 1055, + 576, + 1055, + 576, + 1072, + 511, + 1072 + ], + "score": 0.47, + "latex": "2 \\times 1 0 ^ { - 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1034, + 421, + 1081, + 421, + 1081, + 436, + 1034, + 436 + ], + "score": 0.37, + "latex": "\\approx 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 741, + 858, + 807, + 858, + 807, + 878, + 741, + 878 + ], + "score": 0.36, + "latex": "7 \\times 1 0 ^ { - 2 } )" + }, + { + "category_id": 13, + "poly": [ + 723, + 910, + 807, + 910, + 807, + 929, + 723, + 929 + ], + "score": 0.32, + "latex": "( \\alpha 9 \\times 1 0 ^ { - 2 } )" + }, + { + "category_id": 13, + "poly": [ + 740, + 885, + 807, + 885, + 807, + 903, + 740, + 903 + ], + "score": 0.29, + "latex": "8 \\times 1 0 ^ { - 2 } )" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 801.0, + 339.0, + 801.0, + 339.0, + 830.0, + 332.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 801.0, + 380.0, + 801.0, + 380.0, + 830.0, + 375.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 802.0, + 877.0, + 802.0, + 877.0, + 832.0, + 872.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 802.0, + 908.0, + 802.0, + 908.0, + 832.0, + 903.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 857.0, + 339.0, + 857.0, + 339.0, + 888.0, + 331.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 857.0, + 382.0, + 857.0, + 382.0, + 888.0, + 374.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 824.0, + 740.0, + 824.0, + 740.0, + 883.0, + 663.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 824.0, + 816.0, + 824.0, + 816.0, + 883.0, + 808.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 836.0, + 877.0, + 836.0, + 877.0, + 871.0, + 871.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 836.0, + 912.0, + 836.0, + 912.0, + 871.0, + 903.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 825.0, + 1351.0, + 825.0, + 1351.0, + 875.0, + 1294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 875.0, + 739.0, + 875.0, + 739.0, + 908.0, + 662.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 875.0, + 817.0, + 875.0, + 817.0, + 908.0, + 808.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 877.0, + 869.0, + 877.0, + 869.0, + 909.0, + 860.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 877.0, + 910.0, + 877.0, + 910.0, + 909.0, + 903.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 914.0, + 339.0, + 914.0, + 339.0, + 946.0, + 334.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 914.0, + 383.0, + 914.0, + 383.0, + 946.0, + 374.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 898.0, + 722.0, + 898.0, + 722.0, + 956.0, + 662.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 898.0, + 823.0, + 898.0, + 823.0, + 956.0, + 815.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 915.0, + 869.0, + 915.0, + 869.0, + 946.0, + 859.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 915.0, + 911.0, + 915.0, + 911.0, + 946.0, + 903.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 951.0, + 869.0, + 951.0, + 869.0, + 987.0, + 862.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 951.0, + 912.0, + 951.0, + 912.0, + 987.0, + 903.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 966.0, + 345.0, + 966.0, + 345.0, + 1015.0, + 318.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 971.0, + 381.0, + 971.0, + 381.0, + 1004.0, + 329.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 996.0, + 479.0, + 996.0, + 479.0, + 1028.0, + 438.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 992.0, + 869.0, + 992.0, + 869.0, + 1024.0, + 860.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 992.0, + 910.0, + 992.0, + 910.0, + 1024.0, + 904.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1018.0, + 497.0, + 1018.0, + 497.0, + 1054.0, + 436.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1018.0, + 568.0, + 1018.0, + 568.0, + 1054.0, + 558.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1031.0, + 869.0, + 1031.0, + 869.0, + 1059.0, + 860.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1031.0, + 909.0, + 1031.0, + 909.0, + 1059.0, + 903.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1046.0, + 510.0, + 1046.0, + 510.0, + 1077.0, + 432.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1046.0, + 586.0, + 1046.0, + 586.0, + 1077.0, + 577.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1070.0, + 585.0, + 1070.0, + 585.0, + 1102.0, + 436.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1069.0, + 869.0, + 1069.0, + 869.0, + 1098.0, + 861.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1069.0, + 909.0, + 1069.0, + 909.0, + 1098.0, + 903.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1084.0, + 339.0, + 1084.0, + 339.0, + 1121.0, + 333.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1084.0, + 383.0, + 1084.0, + 383.0, + 1121.0, + 374.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1096.0, + 586.0, + 1096.0, + 586.0, + 1129.0, + 431.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 1107.0, + 869.0, + 1107.0, + 869.0, + 1137.0, + 859.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1107.0, + 910.0, + 1107.0, + 910.0, + 1137.0, + 904.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1119.0, + 587.0, + 1119.0, + 587.0, + 1155.0, + 434.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1144.0, + 385.0, + 1144.0, + 385.0, + 1175.0, + 331.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1146.0, + 868.0, + 1146.0, + 868.0, + 1173.0, + 860.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1146.0, + 911.0, + 1146.0, + 911.0, + 1173.0, + 903.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1166.0, + 396.0, + 1166.0, + 396.0, + 1192.0, + 372.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1164.0, + 521.0, + 1164.0, + 521.0, + 1194.0, + 471.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1164.0, + 635.0, + 1164.0, + 635.0, + 1194.0, + 578.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1164.0, + 747.0, + 1164.0, + 747.0, + 1194.0, + 691.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1164.0, + 857.0, + 1164.0, + 857.0, + 1194.0, + 801.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1168.0, + 925.0, + 1168.0, + 925.0, + 1191.0, + 902.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1165.0, + 1049.0, + 1165.0, + 1049.0, + 1194.0, + 1002.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1164.0, + 1164.0, + 1164.0, + 1164.0, + 1194.0, + 1108.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1164.0, + 1276.0, + 1164.0, + 1276.0, + 1194.0, + 1220.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 1164.0, + 1386.0, + 1164.0, + 1386.0, + 1194.0, + 1330.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1179.0, + 642.0, + 1179.0, + 642.0, + 1221.0, + 573.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1182.0, + 1168.0, + 1182.0, + 1168.0, + 1217.0, + 1105.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1012.0, + 1011.0, + 1012.0, + 1011.0, + 1031.5, + 954.0, + 1031.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.75, + 1027.0, + 386.75, + 1027.0, + 386.75, + 1063.0, + 328.75, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 236.0, + 348.0, + 236.0, + 348.0, + 269.0, + 341.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 236.0, + 422.0, + 236.0, + 422.0, + 269.0, + 414.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 260.0, + 920.0, + 260.0, + 920.0, + 292.0, + 874.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 280.0, + 936.0, + 280.0, + 936.0, + 316.0, + 876.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 305.0, + 951.0, + 305.0, + 951.0, + 335.0, + 875.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 325.0, + 340.0, + 325.0, + 340.0, + 352.0, + 330.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 325.0, + 421.0, + 325.0, + 421.0, + 352.0, + 415.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 326.0, + 990.0, + 326.0, + 990.0, + 360.0, + 872.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 333.0, + 1089.0, + 333.0, + 1089.0, + 361.0, + 1053.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 405.0, + 421.0, + 405.0, + 421.0, + 448.0, + 415.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 410.0, + 1033.0, + 410.0, + 1033.0, + 443.0, + 1025.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 410.0, + 1091.0, + 410.0, + 1091.0, + 443.0, + 1082.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 497.0, + 339.0, + 497.0, + 339.0, + 525.0, + 332.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 497.0, + 420.0, + 497.0, + 420.0, + 525.0, + 415.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 491.0, + 1090.0, + 491.0, + 1090.0, + 519.0, + 1052.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 605.0, + 435.0, + 605.0, + 435.0, + 628.0, + 414.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 603.0, + 563.0, + 603.0, + 563.0, + 631.0, + 516.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 603.0, + 678.0, + 603.0, + 678.0, + 631.0, + 631.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 603.0, + 793.0, + 603.0, + 793.0, + 631.0, + 745.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 603.0, + 908.0, + 603.0, + 908.0, + 631.0, + 860.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 600.0, + 1026.0, + 600.0, + 1026.0, + 631.0, + 970.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 619.0, + 745.0, + 619.0, + 745.0, + 656.0, + 678.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.75, + 579.5, + 423.75, + 579.5, + 423.75, + 612.5, + 331.75, + 612.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1252.0, + 1328.0, + 1252.0, + 1328.0, + 1292.0, + 291.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 1252.0, + 1408.0, + 1252.0, + 1408.0, + 1292.0, + 1391.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1284.0, + 1407.0, + 1284.0, + 1407.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1318.0, + 1000.0, + 1318.0, + 1000.0, + 1352.0, + 295.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 690.0, + 819.0, + 690.0, + 819.0, + 726.0, + 295.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 690.0, + 1405.0, + 690.0, + 1405.0, + 726.0, + 849.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 723.0, + 603.0, + 723.0, + 603.0, + 756.0, + 294.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 723.0, + 1359.0, + 723.0, + 1359.0, + 756.0, + 632.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1409.0, + 797.0, + 1409.0, + 797.0, + 1446.0, + 297.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1466.0, + 1406.0, + 1466.0, + 1406.0, + 1504.0, + 293.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1499.0, + 817.0, + 1499.0, + 817.0, + 1535.0, + 295.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1499.0, + 1405.0, + 1499.0, + 1405.0, + 1535.0, + 846.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1529.0, + 1199.0, + 1529.0, + 1199.0, + 1565.0, + 295.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 1529.0, + 1403.0, + 1529.0, + 1403.0, + 1565.0, + 1228.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1560.0, + 1403.0, + 1560.0, + 1403.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1589.0, + 1406.0, + 1589.0, + 1406.0, + 1626.0, + 293.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1621.0, + 1405.0, + 1621.0, + 1405.0, + 1656.0, + 295.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1651.0, + 1406.0, + 1651.0, + 1406.0, + 1688.0, + 293.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1023.0, + 1680.0, + 1023.0, + 1717.0, + 293.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1680.0, + 1407.0, + 1680.0, + 1407.0, + 1717.0, + 1052.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 1407.0, + 1711.0, + 1407.0, + 1749.0, + 293.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 1373.0, + 1742.0, + 1373.0, + 1778.0, + 295.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1778.0, + 1401.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1254.0, + 1771.0, + 1254.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1809.0, + 1283.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1806.0, + 1403.0, + 1806.0, + 1403.0, + 1837.0, + 296.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1830.0, + 708.0, + 1830.0, + 708.0, + 1873.0, + 292.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1830.0, + 1410.0, + 1830.0, + 1410.0, + 1873.0, + 738.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1866.0, + 833.0, + 1866.0, + 833.0, + 1898.0, + 297.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1976.0, + 290.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1257, + 1405, + 1257, + 1405, + 1745, + 298, + 1745 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1404, + 1850, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 310, + 218, + 1394, + 218, + 1394, + 901, + 310, + 901 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1088, + 1403, + 1088, + 1403, + 1241, + 298, + 1241 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 297, + 958, + 1402, + 958, + 1402, + 1021, + 297, + 1021 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 303, + 1790, + 766, + 1790, + 766, + 1820, + 303, + 1820 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 856, + 76, + 856, + 104, + 297, + 104 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.748 + }, + { + "category_id": 4, + "poly": [ + 649, + 903, + 1050, + 903, + 1050, + 934, + 649, + 934 + ], + "score": 0.726 + }, + { + "category_id": 13, + "poly": [ + 1262, + 1970, + 1369, + 1970, + 1369, + 2002, + 1262, + 2002 + ], + "score": 0.92, + "latex": "2 \\times 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1058, + 1377, + 1161, + 1377, + 1161, + 1407, + 1058, + 1407 + ], + "score": 0.92, + "latex": "i \\times 1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 510, + 1561, + 570, + 1561, + 570, + 1591, + 510, + 1591 + ], + "score": 0.91, + "latex": "1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 939, + 2002, + 1047, + 2002, + 1047, + 2032, + 939, + 2032 + ], + "score": 0.91, + "latex": "3 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1970, + 356, + 1970, + 356, + 2001, + 296, + 2001 + ], + "score": 0.89, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1621, + 356, + 1621, + 356, + 1652, + 296, + 1652 + ], + "score": 0.89, + "latex": "1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1025, + 1347, + 1084, + 1347, + 1084, + 1377, + 1025, + 1377 + ], + "score": 0.89, + "latex": "\\mathrm { \\bar { 1 0 } ^ { - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 978, + 1181, + 1005, + 1181, + 1005, + 1210, + 978, + 1210 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1529, + 1273, + 1529, + 1273, + 1561, + 1213, + 1561 + ], + "score": 0.87, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 673, + 1117, + 1119, + 1117, + 1119, + 1152, + 673, + 1152 + ], + "score": 0.83, + "latex": "\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } , 1 0 ^ { - 6 } , 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1090, + 1404, + 1090, + 1404, + 1119, + 1351, + 1119 + ], + "score": 0.83, + "latex": "\\alpha \\in" + }, + { + "category_id": 13, + "poly": [ + 1238, + 1381, + 1252, + 1381, + 1252, + 1406, + 1238, + 1406 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 327, + 429, + 385, + 429, + 385, + 442, + 327, + 442 + ], + "score": 0.73, + "latex": "3 . 8 \\times 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1116, + 664, + 1116, + 664, + 1152, + 297, + 1152 + ], + "score": 0.66, + "latex": "\\{ 1 0 ^ { - 1 } , 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 1 0 ^ { - 5 } \\}" + }, + { + "category_id": 13, + "poly": [ + 311, + 359, + 385, + 359, + 385, + 383, + 311, + 383 + ], + "score": 0.46, + "latex": "\\underline { { \\hat { \\ell } } } \\ 2 . 9 \\times 1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 328, + 495, + 387, + 495, + 387, + 509, + 328, + 509 + ], + "score": 0.35, + "latex": "4 . 9 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 873, + 428, + 931, + 428, + 931, + 442, + 873, + 442 + ], + "score": 0.34, + "latex": "3 . 6 \\times 1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1117, + 370, + 1117, + 370, + 1151, + 299, + 1151 + ], + "score": 0.32, + "latex": "\\lbrace 1 0 ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 450, + 1117, + 510, + 1117, + 510, + 1150, + 450, + 1150 + ], + "score": 0.25, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 223.0, + 390.0, + 223.0, + 390.0, + 245.0, + 327.0, + 245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 222.0, + 936.0, + 222.0, + 936.0, + 246.0, + 874.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 239.0, + 418.0, + 239.0, + 418.0, + 249.0, + 408.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 240.0, + 795.0, + 240.0, + 795.0, + 262.0, + 762.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 240.0, + 1342.0, + 240.0, + 1342.0, + 262.0, + 1307.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 256.0, + 808.0, + 256.0, + 808.0, + 281.0, + 762.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 256.0, + 1354.0, + 256.0, + 1354.0, + 281.0, + 1308.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 277.0, + 820.0, + 277.0, + 820.0, + 298.0, + 763.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 277.0, + 1365.0, + 277.0, + 1365.0, + 298.0, + 1310.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 289.0, + 389.0, + 289.0, + 389.0, + 311.0, + 320.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 290.0, + 935.0, + 290.0, + 935.0, + 311.0, + 869.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 351.0, + 391.0, + 351.0, + 391.0, + 387.0, + 386.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 345.0, + 942.0, + 345.0, + 942.0, + 392.0, + 848.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 424.0, + 326.0, + 424.0, + 326.0, + 446.0, + 320.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 424.0, + 391.0, + 424.0, + 391.0, + 446.0, + 386.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 424.0, + 872.0, + 424.0, + 872.0, + 445.0, + 866.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 424.0, + 935.0, + 424.0, + 935.0, + 445.0, + 932.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 492.0, + 327.0, + 492.0, + 327.0, + 512.0, + 321.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 492.0, + 392.0, + 492.0, + 392.0, + 512.0, + 388.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 492.0, + 936.0, + 492.0, + 936.0, + 512.0, + 866.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 507.0, + 403.0, + 507.0, + 403.0, + 526.0, + 383.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 507.0, + 492.0, + 507.0, + 492.0, + 526.0, + 470.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 507.0, + 581.0, + 507.0, + 581.0, + 526.0, + 558.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 507.0, + 669.0, + 507.0, + 669.0, + 526.0, + 645.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 506.0, + 756.0, + 506.0, + 756.0, + 526.0, + 732.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 506.0, + 848.0, + 506.0, + 848.0, + 528.0, + 817.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 508.0, + 947.0, + 508.0, + 947.0, + 524.0, + 932.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 507.0, + 1038.0, + 507.0, + 1038.0, + 526.0, + 1016.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 507.0, + 1126.0, + 507.0, + 1126.0, + 526.0, + 1104.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 507.0, + 1213.0, + 507.0, + 1213.0, + 526.0, + 1191.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 507.0, + 1301.0, + 507.0, + 1301.0, + 526.0, + 1279.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 507.0, + 1391.0, + 507.0, + 1391.0, + 525.0, + 1365.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 520.0, + 636.0, + 520.0, + 636.0, + 545.0, + 588.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 520.0, + 1183.0, + 520.0, + 1183.0, + 545.0, + 1134.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 555.0, + 654.0, + 555.0, + 654.0, + 587.0, + 499.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 554.0, + 1199.0, + 554.0, + 1199.0, + 586.0, + 1044.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 590.0, + 623.0, + 590.0, + 623.0, + 613.0, + 592.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 608.0, + 1069.0, + 608.0, + 1069.0, + 632.0, + 1035.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 623.0, + 625.0, + 623.0, + 625.0, + 650.0, + 591.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 626.0, + 1082.0, + 626.0, + 1082.0, + 652.0, + 1034.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 643.0, + 1103.0, + 643.0, + 1103.0, + 673.0, + 1033.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 657.0, + 625.0, + 657.0, + 625.0, + 684.0, + 591.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 694.0, + 624.0, + 694.0, + 624.0, + 717.0, + 589.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 701.0, + 602.0, + 701.0, + 602.0, + 745.0, + 576.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 727.0, + 623.0, + 727.0, + 623.0, + 750.0, + 592.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 763.0, + 623.0, + 763.0, + 623.0, + 786.0, + 592.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 797.0, + 624.0, + 797.0, + 624.0, + 820.0, + 592.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 831.0, + 624.0, + 831.0, + 624.0, + 854.0, + 593.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 849.0, + 634.0, + 849.0, + 634.0, + 867.0, + 618.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 846.0, + 736.0, + 846.0, + 736.0, + 869.0, + 709.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 847.0, + 831.0, + 847.0, + 831.0, + 868.0, + 808.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 846.0, + 930.0, + 846.0, + 930.0, + 869.0, + 901.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 846.0, + 1026.0, + 846.0, + 1026.0, + 869.0, + 999.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 847.0, + 1126.0, + 847.0, + 1126.0, + 869.0, + 1092.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 863.0, + 893.0, + 863.0, + 893.0, + 889.0, + 842.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 293.5, + 825.0, + 293.5, + 825.0, + 316.5, + 762.0, + 316.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 292.5, + 1371.0, + 292.5, + 1371.0, + 316.5, + 1307.0, + 316.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 665.5, + 1095.0, + 665.5, + 1095.0, + 689.5, + 1033.0, + 689.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 958.0, + 1404.0, + 958.0, + 1404.0, + 993.0, + 296.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 990.0, + 738.0, + 990.0, + 738.0, + 1021.0, + 295.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1787.0, + 772.0, + 1787.0, + 772.0, + 1825.0, + 295.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 898.0, + 1051.0, + 898.0, + 1051.0, + 940.0, + 647.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1257.0, + 1405.0, + 1257.0, + 1405.0, + 1292.0, + 295.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1288.0, + 1405.0, + 1288.0, + 1405.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1353.0, + 293.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 1024.0, + 1347.0, + 1024.0, + 1385.0, + 293.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1347.0, + 1406.0, + 1347.0, + 1406.0, + 1385.0, + 1085.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1377.0, + 1057.0, + 1377.0, + 1057.0, + 1413.0, + 292.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1377.0, + 1237.0, + 1377.0, + 1237.0, + 1413.0, + 1162.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1377.0, + 1406.0, + 1377.0, + 1406.0, + 1413.0, + 1253.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1407.0, + 1406.0, + 1407.0, + 1406.0, + 1447.0, + 292.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1439.0, + 1408.0, + 1439.0, + 1408.0, + 1475.0, + 292.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1464.0, + 1407.0, + 1464.0, + 1407.0, + 1511.0, + 291.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1499.0, + 1410.0, + 1499.0, + 1410.0, + 1538.0, + 291.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1529.0, + 1212.0, + 1529.0, + 1212.0, + 1566.0, + 292.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1529.0, + 1406.0, + 1529.0, + 1406.0, + 1566.0, + 1274.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1560.0, + 509.0, + 1560.0, + 509.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1560.0, + 1406.0, + 1560.0, + 1406.0, + 1595.0, + 571.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1590.0, + 1407.0, + 1590.0, + 1407.0, + 1627.0, + 291.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1619.0, + 295.0, + 1619.0, + 295.0, + 1660.0, + 292.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1619.0, + 1406.0, + 1619.0, + 1406.0, + 1660.0, + 357.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1649.0, + 1407.0, + 1649.0, + 1407.0, + 1690.0, + 292.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1722.0, + 291.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1715.0, + 356.0, + 1715.0, + 356.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1882.0, + 296.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1879.0, + 1408.0, + 1879.0, + 1408.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 295.0, + 1969.0, + 295.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1969.0, + 1261.0, + 1969.0, + 1261.0, + 2008.0, + 357.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1969.0, + 1408.0, + 1969.0, + 1408.0, + 2008.0, + 1370.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 938.0, + 1999.0, + 938.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1999.0, + 1405.0, + 1999.0, + 1405.0, + 2038.0, + 1048.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1350.0, + 1087.0, + 1350.0, + 1122.0, + 294.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1112.0, + 296.0, + 1112.0, + 296.0, + 1157.0, + 293.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1112.0, + 672.0, + 1112.0, + 672.0, + 1157.0, + 665.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1112.0, + 1408.0, + 1112.0, + 1408.0, + 1157.0, + 1120.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1147.0, + 1405.0, + 1147.0, + 1405.0, + 1186.0, + 293.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1180.0, + 977.0, + 1180.0, + 977.0, + 1213.0, + 296.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1180.0, + 1404.0, + 1180.0, + 1404.0, + 1213.0, + 1006.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1211.0, + 1334.0, + 1211.0, + 1334.0, + 1244.0, + 296.0, + 1244.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 339, + 1404, + 339, + 1404, + 766, + 298, + 766 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 299, + 879, + 1404, + 879, + 1404, + 1001, + 299, + 1001 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 300, + 1017, + 1403, + 1017, + 1403, + 1140, + 300, + 1140 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 302, + 229, + 1400, + 229, + 1400, + 321, + 302, + 321 + ], + "score": 0.957 + }, + { + "category_id": 0, + "poly": [ + 300, + 809, + 543, + 809, + 543, + 844, + 300, + 844 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 299, + 1186, + 488, + 1186, + 488, + 1220, + 299, + 1220 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 856, + 76, + 856, + 104, + 299, + 104 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 299, + 1236, + 1399, + 1236, + 1399, + 1299, + 299, + 1299 + ], + "score": 0.815 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.776 + }, + { + "category_id": 1, + "poly": [ + 297, + 1316, + 1399, + 1316, + 1399, + 1382, + 297, + 1382 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 300, + 1675, + 1403, + 1675, + 1403, + 1740, + 300, + 1740 + ], + "score": 0.673 + }, + { + "category_id": 1, + "poly": [ + 299, + 1756, + 1402, + 1756, + 1402, + 1820, + 299, + 1820 + ], + "score": 0.625 + }, + { + "category_id": 1, + "poly": [ + 300, + 1593, + 1401, + 1593, + 1401, + 1657, + 300, + 1657 + ], + "score": 0.596 + }, + { + "category_id": 1, + "poly": [ + 302, + 1398, + 1400, + 1398, + 1400, + 1494, + 302, + 1494 + ], + "score": 0.588 + }, + { + "category_id": 1, + "poly": [ + 293, + 1838, + 1403, + 1838, + 1403, + 1872, + 293, + 1872 + ], + "score": 0.586 + }, + { + "category_id": 1, + "poly": [ + 299, + 1511, + 1401, + 1511, + 1401, + 1575, + 299, + 1575 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1399, + 1972, + 1399, + 2034, + 300, + 2034 + ], + "score": 0.533 + }, + { + "category_id": 1, + "poly": [ + 298, + 1889, + 1401, + 1889, + 1401, + 1954, + 298, + 1954 + ], + "score": 0.524 + }, + { + "category_id": 13, + "poly": [ + 837, + 702, + 1142, + 702, + 1142, + 737, + 837, + 737 + ], + "score": 0.92, + "latex": "\\gamma \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } , 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 504, + 702, + 784, + 702, + 784, + 737, + 504, + 737 + ], + "score": 0.91, + "latex": "\\alpha \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 3 } , 1 0 ^ { - 4 } \\}" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 805.0, + 549.0, + 805.0, + 549.0, + 853.0, + 292.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1186.0, + 490.0, + 1186.0, + 490.0, + 1223.0, + 296.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 336.0, + 1405.0, + 336.0, + 1405.0, + 374.0, + 295.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1406.0, + 369.0, + 1406.0, + 404.0, + 294.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 1408.0, + 398.0, + 1408.0, + 437.0, + 293.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 468.0, + 291.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1406.0, + 461.0, + 1406.0, + 497.0, + 294.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 490.0, + 1407.0, + 490.0, + 1407.0, + 526.0, + 293.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 521.0, + 1406.0, + 521.0, + 1406.0, + 556.0, + 293.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1408.0, + 551.0, + 1408.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 582.0, + 1408.0, + 582.0, + 1408.0, + 621.0, + 293.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 613.0, + 1406.0, + 613.0, + 1406.0, + 649.0, + 293.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 641.0, + 1405.0, + 641.0, + 1405.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 1402.0, + 672.0, + 1402.0, + 707.0, + 294.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 698.0, + 503.0, + 698.0, + 503.0, + 741.0, + 289.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 698.0, + 836.0, + 698.0, + 836.0, + 741.0, + 785.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 698.0, + 1408.0, + 698.0, + 1408.0, + 741.0, + 1143.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 734.0, + 1107.0, + 734.0, + 1107.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 880.0, + 1404.0, + 880.0, + 1404.0, + 912.0, + 295.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 910.0, + 1406.0, + 910.0, + 1406.0, + 942.0, + 295.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 941.0, + 1404.0, + 941.0, + 1404.0, + 974.0, + 295.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 972.0, + 723.0, + 972.0, + 723.0, + 1004.0, + 295.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1018.0, + 1403.0, + 1018.0, + 1403.0, + 1051.0, + 295.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1048.0, + 1404.0, + 1048.0, + 1404.0, + 1081.0, + 295.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1080.0, + 1404.0, + 1080.0, + 1404.0, + 1112.0, + 296.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1109.0, + 1220.0, + 1109.0, + 1220.0, + 1141.0, + 295.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 259.0, + 1404.0, + 259.0, + 1404.0, + 294.0, + 296.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 291.0, + 552.0, + 291.0, + 552.0, + 324.0, + 297.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1236.0, + 1404.0, + 1236.0, + 1404.0, + 1271.0, + 296.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1268.0, + 926.0, + 1268.0, + 926.0, + 1301.0, + 323.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1353.0, + 293.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1349.0, + 1383.0, + 1349.0, + 1383.0, + 1383.0, + 321.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1675.0, + 1403.0, + 1675.0, + 1403.0, + 1711.0, + 296.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1705.0, + 825.0, + 1705.0, + 825.0, + 1742.0, + 320.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1751.0, + 1406.0, + 1751.0, + 1406.0, + 1794.0, + 293.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1788.0, + 596.0, + 1788.0, + 596.0, + 1819.0, + 322.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1593.0, + 1403.0, + 1593.0, + 1403.0, + 1629.0, + 296.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1624.0, + 900.0, + 1624.0, + 900.0, + 1657.0, + 321.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1436.0, + 297.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1428.0, + 1406.0, + 1428.0, + 1406.0, + 1468.0, + 321.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1461.0, + 1344.0, + 1461.0, + 1344.0, + 1497.0, + 322.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1837.0, + 1406.0, + 1837.0, + 1406.0, + 1876.0, + 296.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1510.0, + 1405.0, + 1510.0, + 1405.0, + 1547.0, + 294.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1542.0, + 1341.0, + 1542.0, + 1341.0, + 1576.0, + 323.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 1104.0, + 2003.0, + 1104.0, + 2035.0, + 321.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1889.0, + 1405.0, + 1889.0, + 1405.0, + 1925.0, + 294.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1922.0, + 1304.0, + 1922.0, + 1304.0, + 1955.0, + 323.0, + 1955.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 539, + 1404, + 539, + 1404, + 601, + 298, + 601 + ], + "score": 0.783 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 305, + 228, + 1402, + 228, + 1402, + 323, + 305, + 323 + ], + "score": 0.763 + }, + { + "category_id": 1, + "poly": [ + 298, + 425, + 1401, + 425, + 1401, + 519, + 298, + 519 + ], + "score": 0.695 + }, + { + "category_id": 1, + "poly": [ + 295, + 342, + 1403, + 342, + 1403, + 406, + 295, + 406 + ], + "score": 0.653 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 857, + 74, + 857, + 105, + 298, + 105 + ], + "score": 0.623 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 857, + 74, + 857, + 105, + 298, + 105 + ], + "score": 0.226 + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 537.0, + 1409.0, + 537.0, + 1409.0, + 576.0, + 293.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 566.0, + 399.0, + 566.0, + 399.0, + 605.0, + 320.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 227.0, + 1408.0, + 227.0, + 1408.0, + 266.0, + 297.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 257.0, + 1406.0, + 257.0, + 1406.0, + 300.0, + 321.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 293.0, + 715.0, + 293.0, + 715.0, + 323.0, + 324.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 461.0, + 294.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 453.0, + 1405.0, + 453.0, + 1405.0, + 496.0, + 320.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 488.0, + 847.0, + 488.0, + 847.0, + 522.0, + 324.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 339.0, + 1404.0, + 339.0, + 1404.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 374.0, + 1408.0, + 374.0, + 1408.0, + 409.0, + 322.0, + 409.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ryeK6nNFDr/ryeK6nNFDr.md b/parse/train/ryeK6nNFDr/ryeK6nNFDr.md new file mode 100644 index 0000000000000000000000000000000000000000..702b76dad16a8456b532630fe3f37ae92c9f9ec0 --- /dev/null +++ b/parse/train/ryeK6nNFDr/ryeK6nNFDr.md @@ -0,0 +1,394 @@ +# EFFECTIVE AND ROBUST DETECTION OF ADVER-SARIAL EXAMPLES VIA BENFORD-FOURIER COEFFI-CIENTS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Adversarial examples have been well known as a serious threat to deep neural networks (DNNs). To ensure successful and safe operations of DNNs on realworld tasks, it is urgent to equip DNNs with effective defense strategies. In this work, we study the detection of adversarial examples, based on the assumption that the output and internal responses of one DNN model for both adversarial and benign examples follow the generalized Gaussian distribution (GGD), but with different parameters (i.e., shape factor, mean, and variance). GGD is a general distribution family to cover many popular distributions (e.g., Laplacian, Gaussian, or uniform). It is more likely to approximate the intrinsic distributions of internal responses than any specific distribution. Besides, since the shape factor is more robust to different databases rather than the other two parameters, we propose to construct discriminative features via the shape factor for adversarial detection, employing the magnitude of Benford-Fourier coefficients (MBF), which can be easily estimated using responses. Finally, a support vector machine is trained as the adversarial detector through leveraging the MBF features. Through the Kolmogorov-Smirnov (KS) test (Massey Jr (1951)), We empirically verify that: 1) the posterior vectors of both adversarial and benign examples follow GGD; 2) the extracted MBF features of adversarial and benign examples follow different distributions. Extensive experiments in terms of image classification demonstrate that the proposed detector is much more effective and robust on detecting adversarial examples of different crafting methods and different sources, in contrast to state-of-the-art adversarial detection methods. + +# 1 INTRODUCTION + +Deep neural networks (DNNs) have achieved a remarkable success in many important applications, such as image classification, face recognition, object detection, etc. In the meanwhile, DNNs have been shown to be very vulnerable to adversarial examples. However, many real-world scenarios have very restrictive requirements about the robustness of DNNs, such as face verification for login, or semantic segmentation in autonomous driving. Adversarial examples are a serious threat to the applications of DNNs to these important tasks. Since many kinds of adversarial attack methods have been proposed to fool DNNs, it is more urgent to equip effective defensive strategies to ensure the safety of deep models in real-world applications. However, defense seems to be more challenging than attack, as it has to face adversarial examples from unknown crafting methods and unknown data sources. Typical defensive strategies include adversarial training, adversarial de-noising, and adversarial detection. Compared to the former two strategies, adversarial detection is somewhat more cost-effective, as it often does not need to re-train or modify the original DNN model. + +There are two main challenges for adversarial detection. (1) The adversarial examples are designed to camouflage themselves to be close to the corresponding benign examples in the input space. Then, where and how to extract the discriminative information to train the detector? (2) The data sources and the generating methods of adversarial examples are often inaccessible to the detector. In this case, the detector can be stably effective across different data sources and different attack methods? In other words, a good adversarial detector is required to be not only effective to distinguish adversarial and benign examples, but also robust to different data sources and attack methods. + +To satisfy the first requirement of effectiveness, we utilize the other principle of crafting adversarial examples that the outputs between benign and adversarial examples should be large, to encourage the change of the final prediction. It means that the imperceptible difference between benign and adversarial examples in the input space is enlarged along the DNN model, leading to the significant difference in the output space. Inspired by this fact, we assume that the output or the responses of internal layers of the DNN model should include the discriminative information for benign and adversarial examples. A few works have attempted to extract different types of discriminative features from the output or the internal responses, such as kernel density (KD) (Feinman et al. (2017)) and the local intrinsic dimensionality (LID) (Ma et al. (2018)), etc. To achieve the robustness, the extracted discriminative features should model the intrinsic difference between adversarial and benign examples, rather than the difference from the changes of data sources or attack methods. Many existing methods have shown the effectiveness to some extent on detecting adversarial examples of specific data sources and attack methods. However, their robustness, especially across different data sources, has not been well studied and verified. + +In this work, we propose a novel detection method based on the assumption that the internal responses of both adversarial and benign examples follow the generalized Gaussian distribution (GGD) (Varanasi & Aazhang (1989)), but with different parameters, including the shape factor, mean, and variance. The rationale behind this assumption is that GGD covers many popular distributions with varied shape factors (such as Laplacian, Gaussian, or uniform distribution), such that GGD is more likely to approximate the intrinsic response distributions rather than one specific distribution. Moreover, mean and variance of GGD may vary significantly with respect to different classes and data sources, even for benign examples, while the shape factor is more robust. For example, the mean and variance of two Gaussian distributions could be totally different, but their shape factors are the same (i.e., 2). Thus, we propose to use the shape factor as an effective and robust discrimination between adversarial and benign examples. However, it is difficult to exactly estimate the shape factor in practice. We resort to the magnitude of Benford-Fourier coefficients (Pasquini et al. (2014)), which is a function of the shape factor. It can be easily estimated using internal responses, according to the definition of Fourier transform. Then, the magnitudes estimated from internal responses of different convolutional layers are concatenated as a novel representation. Finally, a support vector machine (SVM) (Vapnik (2013)) is trained using the new representations as the adversarial detector. Extensive experiments carried out on several databases verify the effectiveness and robustness of the proposed detection method. To further verify the rationale of our assumption, we present the empirical analysis through the Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)). The KS test verifies that 1) the posterior vectors of both adversarial and benign examples predicted by the CNN model follow the distribution of GGD, but with different parameters, and 2) the MBF features of adversarial and benign examples follow different distributions, and the MBF features of adversarial examples crafted from different attack methods follow the same distribution, as well as that the MBF features of adversarial/benign examples from different data sources follow the same distribution. Moreover, we visualize the statistics (i.e., mean $\pm$ standard deviation) of the extracted MBF features for adversarial and benign examples. The visualization reveals the distinct difference between adversarial and benign examples. These empirical analysis demonstrates the effectiveness and robustness of the proposed MBF detector. + +# 2 RELATED WORK + +In this section we focus on reviewing existing works about adversarial detection, while those of other adversarial defensive strategies are out of scope of this work. + +The general idea of most existing detection methods is learning or constructing a new representation to discriminate adversarial and benign examples, utilizing the outputs or immediate responses of an original classification network. Li & Li (2017) trained a cascading classifier based on the principal component analysis (PCA) (Pearson (1901)) statistics of responses from each convolutional layer of the defended convolutional neural network (CNN) model. An example is recognized as benign if all single classifiers of the cascade predict it as benign, otherwise adversarial. Lu et al. (2017) proposed SafetyNet by adding a RBF-SVM classifier to detect adversarial examples, at the end of the original classification network. Metzen et al. (2017) proposed a detection network along with an original classification network, which takes the internal responses of the original network as inputs. It shows effectiveness on detecting adversarial examples generated by simple attacks (such as FGSM (Goodfellow et al. (2014)) and JSMA (Papernot et al. (2016))), while performs much worse when facing more advanced attacks (such as C&W (Carlini & Wagner (2017a))). It tells that this method is sensitive to attack methods. Grosse et al. (2017) attempted to detect adversarial examples using the statistical test of maximum mean discrepancy (MMD). Although above detection methods show effectiveness on some attack methods and some databases, but a thorough evaluation presented in Carlini & Wagner (2017b) has shown that these methods are sensitive to attack methods or databases, and they can be somewhat easily invaded by new attacks. + +Some recent works proposed to utilize neighboring samples in the same database to construct a better representation of a current sample. Feinman et al. (2017) defined two metrics based on the responses of the final hidden layer of the classification neural network, including kernel density estimation (KDE) and Bayesian neural network uncertainty (BU). If the metric score of KDE/BU is lower/higher than a pre-defined threshold, then the example is predicted as adversarial. Ma et al. (2018) utilized the local intrinsic dimensionality (LID) to measure the characterization of adversarial regions of DNNs. LID describes the distance between one example and its $k$ -nearest neighboring sample in the feature space of immediate responses of the original classification network. The distances computed from different layers are concatenated as the example representation, which is then used to train a shallow classifier to discriminate adversarial and benign examples. Zheng & Hong (2018) defined the intrinsic hidden state distribution (IHSD) of the responses of the original classification network to model different classes. The Gaussian mixture model (GMM) was used to approximate IHSD of each class. Then, the posterior probability of one sample assigned to GMM is computed as the metric. If the probability is lower than a pre-defined threshold, then it is recognized as adversarial. Lee et al. (2018) computed the class-conditional Gaussian distribution of the responses of the original classification network based on the whole training set. Then, the Mahalanobis distance between one sample and its nearest class-conditional Gaussian distribution is used as the metric for detection. If the distance is larger than a pre-defined threshold, then it is detected as adversarial. Compared to some aforementioned single-representation-based detection methods, these joint-representation-based methods showed better performances on some databases. However, the detection cost for each example is much higher, as the responses of its neighboring samples should also be computed. Besides, since the representation is highly dependent on the neighbors or all training examples, the detection performance may be sensitive to data sources, which will be studied in later experiments. + +There are also some other approaches that do not construct representations from the responses of the original classification network. Hendrycks & Gimpel (2016) adopted PCA statistics to discriminate adversarial and benign images, independent of any DNN model. However, the study presented in Carlini & Wagner (2017b) has demonstrated that this method works for MNIST but not for CIFAR-10, and PCA statistics are not robust features to detect adversarial images. Pang et al. (2018) proposed a novel loss called reverse cross-entropy (RCE) to train the classification network, such that the distance measured by kernel density (Feinman et al. (2017)) between adversarial and benign examples could be enlarged. Samangouei et al. (2018) proposed Defense-GAN to model the distribution of benign examples using a generative adversarial network (GAN). If the Wasserstein distance between one example and its corresponding example generated by the GAN is larger than a pre-defined value, then it is detected as adversarial. However, these above three methods are much more costly than other detection methods. + +# 3 PRELIMINARIES + +# 3.1 GENERALIZED GAUSSIAN DISTRIBUTION + +Assume that a random variable $\mathcal { X } ~ \in ~ \mathbb { R } ^ { d }$ follows the generalized Gaussian distribution (GGD) (Varanasi $\&$ Aazhang (1989)). Then, its probability density function (PDF) is formulated with two positive parameters, including the shape factor $c$ and the standard deviation $\sigma$ , as follows + +$$ +\mathcal { P } _ { \mathcal { X } } ( x ) = A \cdot e ^ { - \lvert \beta x \rvert ^ { c } } , +$$ + +where β = 1σ Γ(3/c)Γ(1/c)  and $\begin{array} { r } { A = \frac { \beta c } { 2 \Gamma ( 1 / c ) } } \end{array}$ , with $\Gamma ( \cdot )$ being the Gamma function. Note that the mean parameter $\mu$ is omitted above, as $\mu$ has no relation with the shape of distribution and we set it as 0 without loss of generality. A nice characteristic of GGD is that it covers many popular distributions + +with varied shape factors. For example, when $c = 1$ , then it becomes the Laplacian distribution; when $c = 2$ , then it is the Gaussian distribution with a variance of √ √ $\sigma ^ { 2 }$ ; when $c \to + \infty$ , then it is specified as a uniform distribution on $( - \sqrt { 2 } \sigma , \sqrt { 2 } \sigma )$ . + +# 3.2 BENFORD-FOURIER COEFFICIENTS + +Although generalized Gaussian distribution (GGD) is able to cover a bunch of distributions, it is hard to depict the exact forms of GGD precisely. To this end, we further define a random variable $\mathcal { Z } = \log _ { 1 0 } | \mathcal { X } |$ mod 1 for detecting and distinguishing different form of GGD, of which the PDF is formulated by means of Fourier Series as Perez-Gonz ´ alez et al. (2007), with the fundamental period ´ being fixed as $2 \pi$ , + +$$ +\mathcal { P } _ { Z } ( z ) = 1 + 2 \sum _ { n = 1 } ^ { + \infty } \lfloor A _ { n } \cos { ( 2 \pi n z ) } + B _ { n } \sin { ( 2 \pi n z ) } \rfloor = 1 + 2 \sum _ { n = 1 } ^ { + \infty } \lfloor a _ { n } \rfloor \cos ( 2 \pi n z + \phi _ { n } ) , +$$ + +where $z \in [ 0 , 1 )$ corresponds to the domain of random variable $\mathcal { Z }$ , the phase of Fourier Series is explained as $\begin{array} { r } { \phi _ { n } = \arctan \left( - \frac { B _ { n } } { A _ { n } } \right) } \end{array}$ , and the magnitude denotes $| a _ { n } | = \sqrt { A _ { n } ^ { 2 } + B _ { n } ^ { 2 } }$ . $a _ { n } = | a _ { n } | \cdot e ^ { j \phi _ { n } }$ denotes the $n$ -th Fourier coefficient of $\mathcal { P } _ { \mathcal { Z } } ( z )$ evaluated at $2 \pi n$ , and its definition is + +$$ +a _ { n } = \int _ { - \infty } ^ { + \infty } \mathcal { P } _ { \mathcal { Z } } ( z ) \cdot e ^ { - j 2 \pi n \log _ { 1 0 } z } d z = \frac { 2 A e ^ { \frac { j 2 \pi n \log \beta } { \log 1 0 } } } { \beta c } \cdot \Gamma \Bigg ( \frac { - j 2 \pi n + \log 1 0 } { c \log 1 0 } \Bigg ) . +$$ + +$a _ { n }$ is also called as Benford-Fourier coefficient. Note that $a _ { n }$ is a complex number, and its magnitude can be calculated as follows + +$$ +| a _ { n } | = \bigg ( \prod _ { k = 0 } ^ { + \infty } \bigg [ 1 + \big ( \frac { 2 \pi n } { \log 1 0 ( c k + 1 ) } \big ) ^ { 2 } \bigg ] ^ { - 1 } \bigg ) ^ { \frac { 1 } { 2 } } . +$$ + +Note that $\left| a _ { n } \right|$ gets smaller as $n \in \mathbb N$ increases. And, an interesting property of $\left| a _ { n } \right|$ is that it only depends on the shape factor $c$ , while is independent of the parameter $\sigma$ . Thus, one set of the absolute values of Benford-Fourier coefficients $\{ | a _ { n } | \} _ { n \in \mathbb { N } }$ correspond to one identical $c$ , i.e., one identical special distribution of GGD. In other words, we could use $\{ | a _ { n } | \} _ { n \in \mathbb { N } }$ as features or representations to discriminate different special distributions of GGD. + +However, if it is often difficult to know or even estimate the shape factor $c$ , we cannot compute the value of $\left| a _ { n } \right|$ . But fortunately, recalling that $a _ { n }$ is the $n$ -th Fourier coefficient of $\mathcal { P } _ { \mathcal { Z } } ( z )$ evaluated at $2 \pi n$ , we can derive an easy estimation. Specifically, assume that $\mathbf { x } = \{ x _ { 1 } , \dots , x _ { M } \}$ is a set of $M$ i.i.d. points sampled from GGD with the same shape factor $c$ . Then, the corresponding BenfordFourier coefficients can be estimated as follows (Pasquini et al. (2014)): + +$$ +\hat { a } _ { n } = \frac { \sum _ { m = 1 } ^ { M } e ^ { - j 2 \pi n \log _ { 1 0 } | x _ { m } | } } { M } = \frac { 1 } { M } \sum _ { m = 1 } ^ { M } \left[ \cos \left( 2 \pi n \log _ { 1 0 } | x _ { m } | \right) - j \sin \left( 2 \pi n \log _ { 1 0 } | x _ { m } | \right) \right] . +$$ + +The gap between $\hat { a } _ { n }$ and $a _ { n }$ is analyzed in Theorem 1. It tells that $\hat { a } _ { n }$ gets closer to $a _ { n }$ as $M$ increases. For clarity, we firstly introduce a few notations: $\mathcal { T } = e ^ { - j 2 \pi n \log _ { 1 0 } | \mathcal { X } | }$ is a random variable with $\mathcal { X }$ obeying the generalized Gaussian distribution, and $\hat { a } _ { n }$ is an observation of the random variable $\begin{array} { r } { \mathcal { Y } = \bar { \frac { 1 } { M } } \bar { \sum _ { m = 1 } ^ { M } } \mathcal { T } _ { m } } \end{array}$ . Due to the space limit, the proof will be presented in supplementary material. + +Theorem 1 Assume that the estimation error $\varepsilon _ { n } = { \hat { a } } _ { n } - a _ { n }$ is an observation of the random variable $\mathcal { E } = \mathcal { V } - a _ { n }$ . $| \mathcal { E } |$ follows the Rayleigh distribution (Papoulis & Pillai (2001)), of which the probability density function $( P D F )$ is formulated as + +$$ +\mathcal { P } _ { | \mathcal { E } | } \left( r \right) = 2 M r e ^ { - M r ^ { 2 } } . +$$ + +And the expectation and variance are respectively shown as follows: + +$$ +E { \bigl ( } | { \mathcal { E } } | { \bigr ) } = { \frac { 1 } { 2 } } { \sqrt { \frac { \pi } { M } } } , D ( | { \mathcal { E } } | ) = { \frac { 4 - \pi } { 4 M } } , +$$ + +which implies that the estimation error $\varepsilon _ { n }$ gets closer to $\boldsymbol { \theta }$ as the number of samples M increases. + +# 4 ADVERSARIAL DETECTION VIA BENFORD-FOURIER COEFFICIENTS + +# 4.1 TRAINING PROCEDURE OF ADVERSARIAL DETECTOR + +There are three stages to train the proposed adversarial detector, including: 1) building a training set based on benign images; 2) extracting novel representations of the training set via Benford-Fourier coefficients; 3) training a SVM classifier as the adversarial detector. They will be explained in details sequentially. And, the overall training procedure is briefly summarized in Algorithm 1. + +Build a training set. Firstly, we collect $N$ clean images $\left\{ \mathbf { x } _ { 1 } , \dotsc , \mathbf { x } _ { N } \right\}$ , which can be correctly predicted by $f _ { \theta }$ . Then, we adopt one adversarial attack method (e.g., C&W (Carlini & Wagner (2017a)) or BIM ( Kurakin et al. (2016))) to generate one adversarial image corresponding to each clean image. The crafted $N$ adversarial examples are denoted as $\big \{ \hat { \mathbf { x } } _ { 1 } , \dotsc , \hat { \mathbf { x } } _ { N } \big \}$ . Besides, to avoid that the noisy image (polluted by some kind of non-malicious noises but still can be correctly predicted by $f _ { \theta } )$ is incorrectly detected as adversarial, we also craft one noisy image by adding small random Gaussian noises onto each clean image. These $N$ noisy examples are denoted as $\left\{ bar { \mathbf { x } } _ { 1 } , \ldots , \bar { \mathbf { x } } _ { N } \right\}$ . Note that, hereafter benign examples include both clean and Gaussian noisy examples. Consequently, we obtain one training set with $3 N$ examples, denoted as $\mathcal { D } _ { t r } = \{ ( \mathbf { x } _ { i } , - 1 \bar { ) } , ( \bar { \mathbf { x } } _ { i } , \bar { { ~ - } 1 } ) , ( \hat { \mathbf { x } } _ { i } , + 1 \bar { ) } \} _ { i = 1 , \ldots , N }$ . + +Extract novel representations. We firstly feed the $i$ -th training image from $\mathcal { D } _ { t r }$ into $f _ { \theta }$ . We concatenate all response entries of the $l$ -th layer in $f _ { \theta }$ to obtain one vector $\mathbf { r } _ { i } ^ { l }$ . Then, we estimate the corresponding Benford-Fourier coefficients according to Eq. (5), as follows + +$$ +( \hat { a } _ { n } ) _ { i } ^ { l } = \frac { 1 } { M _ { i } ^ { l } } \sum _ { m = 1 } ^ { M _ { i } ^ { l } } e ^ { - j 2 \pi n \log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | } , +$$ + +where $M _ { i } ^ { l }$ indicates the length of $\mathbf { r } _ { i } ^ { l }$ . The magnitude of $( \hat { a } _ { n } ) _ { i } ^ { l }$ is computed as follows + +$$ +| ( \hat { a } _ { n } ) _ { i } ^ { l } | = \frac { 1 } { M _ { i } ^ { l } } \sum _ { m = 1 } ^ { M _ { i } ^ { l } } \Bigg ( \cos ^ { 2 } \left( 2 \pi n \log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \right) + \sin ^ { 2 } \left( 2 \pi n \log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \right) \Bigg ) ^ { \frac { 1 } { 2 } } . +$$ + +Then, we extract one $T$ -dimensional feature vector $\mathbf { a } _ { i } ^ { l } = [ | ( \hat { a } _ { 1 } ) _ { i } ^ { l } | , \dots , | ( \hat { a } _ { T } ) _ { i } ^ { l } | ] \in \mathbb { R } _ { + } ^ { T }$ for the $i$ -th training image from the $l$ -th layer. We set $T = 1 6$ in experiments, as $| ( \hat { a } _ { n } ) _ { i } ^ { l } |$ of larger $n$ is too small for discrimination. Finally, we concatenate the feature vectors of all layers to form a long vector $\hat { \mathbf { a } } _ { i } = [ \hat { \mathbf { a } } _ { i } ^ { 1 } ; \hdots ; \hat { \mathbf { a } } _ { i } ^ { L } ] \in \mathbb { R } _ { + } ^ { T \check { L } }$ , with $L$ being the number of layers in $f _ { \theta }$ . Consequently, we obtain a novel representation of the training set $\mathcal { D } _ { t r }$ , denoted as $\hat { \mathcal { A } } _ { t r } = \{ ( \hat { \bf { a } } _ { i } , \pm 1 ) \} _ { i = 1 , \dots , 3 N }$ , where the label of ${ \bf a } _ { i }$ (i.e., $+ 1$ or $^ { - 1 }$ ) is directly obtained from $\mathcal { D } _ { t r }$ . + +Train an adversarial detector. Finally, we train a binary SVM classifier based on $\hat { A } _ { t r }$ . The trained SVM classifier will serve as the adversarial detector for the CNN model $f _ { \theta }$ . + +Testing. One novel testing example is firstly predicted as adversarial or not by the trained adversarial detector . If adversarial, then it is rejected; otherwise, it is fed into $f _ { \theta }$ to predict its class label. + +Remark. Note that in the derivation of $a _ { n }$ (see Eq. (3)), the mean parameter of GGD is set to 0. In experiments, we calculate the mean values of internal-layer responses of all networks for every image, and find that mean parameters at most layers are close to 0, while the mean parameters at a few layers could be large. However, the mean value is subtracted from each response entry when we extract the MBF features in our experiments. Thus, the derived $a _ { n }$ is applicable to our task. Besides, according to Theorem 1, the estimation error of $( \hat { a } _ { n } ) _ { i } ^ { l }$ is inversely proportional to $M _ { i } ^ { l }$ . It tells that the coefficient estimated from the larger-sized layer is more accurate. In many neural networks (e.g., AlexNet (Krizhevsky et al. (2012))), the response sizes of high layers get smaller, which means the less accurate estimation. However, we believe that the discrimination between adversarial and benign examples in higher layers is more evident than that in lower layers. There is a trade-off between estimation accuracy and discrimination. This is why we concatenate the estimated magnitudes of all layers together to construct the novel representation. + +# 4.2 EXPERIMENTAL SETTINGS + +Databases and network architectures We conduct experiments on three databases, including CIFAR-10 (Krizhevsky et al. (2014)), SVHN (Netzer et al. (2011)), and a subset of ImageNet (Deng + +Algorithm 1 Training the adversarial detector via the magnitude of Benford-Fourier coefficients. + +Require: The trained CNN model $f _ { \theta }$ with $L$ layers, and the training set $\mathcal { D } _ { t r }$ + +1: for $i = 1$ to $| \mathcal { D } _ { t r } |$ do: +2: for $l = 1$ to $L$ do: +3: Compute $| ( \hat { a } _ { n } ) _ { i } ^ { l } |$ as Eq. (8), with $n = 1 , \dots , T$ ; +4: Concatenate $\{ | ( \hat { a } _ { n } ) _ { i } ^ { l } | \} _ { n = 1 , \dots , T }$ to obtain a vector $\hat { \mathbf { a } } _ { i } ^ { l }$ ; +5: end for +6: Concatenate $\{ \hat { \mathbf { a } } _ { i } ^ { l } \} _ { l = 1 , \dots , L }$ to obtain a long vector $\hat { \mathbf { a } } _ { i }$ ; +7: end for +8: Build a novel representation of the training set, denoted as $\hat { \mathcal { A } } _ { t r } = \{ ( \hat { \mathbf { a } } _ { i } , \pm 1 ) \} _ { i = 1 , \dots , | \mathcal { D } _ { t r } | }$ ; +9: Train a binary SVM classifier based on $\hat { A } _ { t r }$ ; +return The trained binary SVM classifier. + +et al. (2009)). In terms of CIFAR-10 and SVHN, we adopt the same settings as the compared method LID (Ma et al. (2018)). Specifically, a 33-layer network pre-trained on the training set of CIFAR-10 achieves $8 2 . 3 7 \%$ accuracy on the testing set with 10, 000 benign images; a 19-layer network pretrained on the training set of SVHN achieves $9 2 . 6 \%$ accuracy on the testing set with 26, 032 benign images. Then, we add a small noise drawn from ${ \mathcal { N } } ( 0 , \sigma ^ { 2 } )$ on each testing image, with $\sigma$ being the similar level of the $\ell _ { 2 }$ norm of adversarial perturbations on the same database. If both the benign and its noisy image can be correctly predicted by the classification network, then it is picked out for training the detection. We finally collect 8, 175 and 23, 862 benign images from CIFATR-10 and SVHN, respectively. These images are randomly partitioned to the $8 0 \%$ training set and the $2 0 \%$ testing set, used for the training and testing of the detector. We also collect a subset from ImageNet, including 800 benign images of 8 classes (snowbird, spoonbill, bobtail, Leonberg, hamster, proboscis monkey, cypripedium calceolus, and earthstar). The 100 images of each class contain 50 testing images and 50 randomly selected training images. We fine-tune the checkpoints of both AlexNet and VGG pre-trained on ImageNet1 on these 800 images to achieve $1 0 0 \%$ accuracy. Then, 785 benign images are kept for detection, as both their noisy images and themselves can be correctly predicted by both the fine-tuned AlexNet and VGG models. These 785 images are then randomly partitioned to 400 training and 305 testing images used for detection. For each database, as described in Section 4.1, one noisy and one adversarial image are generated for each benign image; then, all of benign, noisy, and adversarial images are used for detection. + +Attack methods We adopt four popular adversarial attack methods to craft adversarial examples, including basic iterative method (BIM (Kurakin et al. (2016))), CarliniWagnerL2Attack (CW-L2 (Carlini & Wagner (2017a))), DeepFool ( (Moosavi-Dezfooli et al. (2016))), and random projected gradient descent (R-PGD (Madry et al. (2017))). They are implemented by Foolbox2. + +Compared detection methods We compare with three state-of-the-art and open-sourced adversarial detection methods, including $\mathrm { K D + B U } ^ { 3 }$ (Feinman et al. (2017)), Mahalanobis distance4 (M-D) (Lee et al. (2018)), and $\mathrm { L I D } ^ { 5 }$ (Ma et al. (2018)). Note that another recent work called I-defender (Zheng & Hong (2018)) is not compared, as its code is not available. To ensure the fair comparison, the SVM classifier is trained with all compared methods, implemented by the fitcsvm6 function in MATLAB. There are two important hyper-parameters in LID, i.e., the size of mini-batch and the number of neighbors. On CIFAR-10 and SVHN, they are respectively set as 100 and 20, as suggested in Ma et al. (2018); on ImageNet, as there are only 400 benign training images, they are respectively set as 50 and 20 in experiments. Moreover, we find that there are some unfair settings in the implementations of compared methods. For example, $\mathrm { K D + B U }$ utilizes the extra 50, 000 images of CIFAR-10 to compute the kernel density of each training and testing image; M-D also uses these extra images to compute the mean and co-variance of GMM. Since extra images of the similar distribution with the training images are often unavailable, we believe that extra images should not be used to ensure the fair comparison. Thus, extra images are not used for $\mathrm { K D + B U }$ and M-D in our experiments. Besides, + +Table 1: Detection results in the non-transfer case. + +
databaseDetectorAUROC(%)Accuracy (%)
BIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
CIFAR-10KD+BU (Feinman et al. (2017))79.082.880.677.774.273.071.772.6
M-D (Lee et al. (2018))51.748.453.752.066.766.766.766.7
LID (Ma et al. (2018))87.987.486.682.772.678.578.070.2
MBF99.699.696.999.498.898.391.898.5
SVHNKD+BU (Feinman et al. (2017)) M-D (Lee et al. (2018))81.585.184.182.578.680.079.378.6
LID (Ma et al. (2018))49.750.049.650.166.766.766.766.7
91.788.792.290.583.380.884.081.5
ImageNet- AlexNetMBF KD+BU (Feinman et al. (2017))99.799.999.399.598.898.391.898.5
M-D (Lee et al. (2018))50.852.6 60.151.151.534.436.834.935.3
LID (Ma et al. (2018))48.1 71.770.946.5 71.954.1 72.366.7 68.766.6 60.566.866.7
ImageNet-MBF99.999.699.999.898.697.865.8 98.868.3
KD+BU (Feinman et al. (2017))57.157.756.258.642.843.641.798.4
M-D (Lee et al. (2018))63.764.845.966.967.165.666.744.8
89.567.4
VGG16LID (Ma et al. (2018)) MBF82.6 99.884.2 100.0100.084.2 100.077.8 99.676.7 99.683.3 100.076.4 100.0
+ +Table 2: Detection results evaluated by AUROC $( \% )$ in the attack-transfer case. + +
databaseTest attack→BIMCW-L2DeepFoolR-PGD
TrainattackKD+BU (Feinman et al. (2017))/M-D(Lee et al.(2018))/LID (Ma et al.(2018)/MBF
CIFAR-10BIM CW-L2 DeepFool R-PGD79.0/51.7787.9799.6 83.4/47.9/90.3/99.6 83.6/53.5 /89.7 /99.3 78.1/51.5/86.1/99.678.1/52.2/85.7799.4 82.8/48.4/87.4/99.6 83.1/51.4/86.1/99.1 77.3 /51.9 /84.5 /99.576.4/52.4/86.4/87.8 80.3 /46.5 /87.0 / 88.6 80.6/53.7/86.6/96.978.6/50.9/87.6/99.4 82.9/48.2/88.9/99.5 83.1/53.6/87.9/98.9 77.7/52.1/82.7/99.4
SVHNBIM CW-L2 DeepFool R-PGD81.5749.7791.7799.7 83.6/50.8/91.6/99.7 85.0/50.0/91.6/99.7 82.5 /49.9 /91.2 /99.683.1/51.0/88.3/99.7 85.1/50.0/88.7/99.9 86.3/ 50.0 /87.9 /99.8 84.0/49.5/87.8/99.675.8/52.4/84.9/88.3 80.1/48.8/91.9/97.5 82.5/49.7/92.0/97.3 84.1/49.6/92.2/99.3 81.2/50.2/91.6/97.481.5750.4/90.7799.5 83.6/49.1/90.8/99.7 85.0/49.4/90.8 /99.6 82.5/50.1/90.5 /99.5
ImageNet- AlexNetBIM CW-L2 DeepFool R-PGD50.8/48.1/71.7/99.9 52.3/50.7/69.7/99.8 51.8/53.6/71.3/99.9 51.6/52.4/72.5 /99.850.8/44.9/70.8/99.6 52.6/60.1/70.9/99.7 51.9/53.5/69.7/99.7 51.3/ 50.4/72.3/99.650.4/49.7772.1/99.8 51.2/52.1/66.0/99.7 51.2/46.5 /71.9 /99.9 50.3 /50.5 /72.2/99.750.7/50.4/70.8/99.8 52.2/55.8/69.8/99.8 51.8/51.2/71.2/99.8
ImageNet- VGG-16BIM CW-L2 DeepFool R-PGD57.1/63.7/82.6/99.8 57.3/64.3/81.9/100.0 58.7/45.5/ 80.2/99.7 58.6/62.3/81.1/100.057.1/57.2/84.3/99.5 57.7/64.8/84.2/100.0 59.4/46.1/79.1/99.3 58.9/60.8/83.7/100.053.8/54.4/90.4/100.0 53.8/56.6/90.2/100.0 56.2/45.9/89.5/100.0 55.3/54.4/90.8/100.051.5 /54.1/72.3 /99.8 57.1/65.4/84.2/100.0 57.5/67.6/84.3/100.0 59.0 /44.0/ 80.9 /99.7 58.6/66.9/84.2/100.0
+ +Table 3: Detection results evaluated by accuracy $( \% )$ in the attack-transfer case. + +
databaseTest attack→BIM CW-L2 DeepFool R-PGDKD+BU(Feinman et al. (2017))/M-D(Lee et al. (2018))/LID (Ma et al. (2018)/MBF
CIFAR-10BIMCW-L2DeepFoolR-PGD74.2/66.7/72.6/98.874.2/66.7/80.3 /98.174.0/66.7 /80.4/95.273.1/66.7/77.7 /98.666.773.0/66.7 /78.5 /98.3 71.7 /66.7/77.9 /89.573.8/66.7/61.3/96.871.7 /66.7/77.9 /89.5 73.8 /66.7 /73.6 /98.2
66.771.9 / 66.7 /75.1/98.5 71.0 /66.7/75.4/89.773.4/66.7/74.8/95.471.0 /66.7/75.4/89.7 72.6/66.7/70.2/98.5
SVHNBIMCW-L2DeepFoolR-PGD78.6/66.77/83.3/98.7 80.5/66.7779.4/98.9 77.7/66.7/83.5/95.8 78.6/66.7/82.9/98.278.2/66.7/83.5 /98.0 80.0/66.7 /80.8/99.2 77.1/66.7/84.0 /95.9 78.1/66.7/80.7/97.379.8/66.7/83.9/98.0 81.8/66.7/79.9/98.4 79.3/66.7/84.0/97.3 79.9/66.7/80.7/98.078.6/66.7 /83.6 /98.6 80.5 / 66.7 /79.9/98.9 77.7 /66.7 /83.8/96.5 78.6/ 66.7 /81.5 /98.8
ImageNet-AlexNetBIMCW-L2DeepFoolR-PGD34.4/66.7/68.7798.6 34.4/66.6/68.6797.336.5/66.2/64.6/98.0 36.8/66.6/60.6/97.8 35.0 /66.1/60.7/98.134.3/66.7/65.0/98.435.0 /66.1/60.7/98.1 36.4/66.1/62.6/98.034.9/66.8/65.8/98.8 35.9 /66.6/68.3 /98.634.1/66.7/67.0/98.3 35.3/66.7/68.3/98.4
36.0 /66.7/69.4/98.835.4/66.7/67.5/98.566.835.2/66.7/67.7/97.6 34.1/66.7/67.0/98.3
ImageNet-VGG-16BIMCW-L2DeepFoolR-PGD42.8/67.1777.8/98.643.1/64.8/75.2/98.045.4/66.7 /76.2/98.844.8 /66.3/76.3/98.566.643.6/65.6/76.7 /97.8 38.4/ 62.5/82.5 /98.142.7/67.9/76.9/98.438.4/ 62.5/82.5 /98.1 43.3/67.2/73.9/98.041.8/66.7 /83.3/98.8 45.8 /66.7/74.4/98.640.6/64.9/82.4/98.3 45.8/67.4/76.4/98.4
66.745.4/66.6/74.1/97.6 40.6/64.9/82.4/98.3
+ +LID utilizes other benign testing images as neighborhoods to extract features for each testing image. It is unfair to utilize the information of benign or adversarial for neighboring testing images. In our experiments, we use benign training images as neighborhoods for each testing image. + +Three comparison cases and evaluation metrics We conduct experiments of three cases, including: 1) non-transfer, both training and testing adversarial examples are crafted by the same attack method; 2) attack-transfer, both training and testing adversarial examples are crafted by different attack methods; 3) data-transfer, both training and testing adversarial examples are crafted by the same attack method, but the data sources of training and testing benign examples are different. Two widely used metrics are used to evaluate the detection performance, including area under the receiver operating characteristics (AUROC), and detection accuracy, which is the diagonal summation of the confusion matrix, using 0.5 as the threshold of the posterior probability. Higher values of both metrics indicate better performance. + +# 4.3 RESULTS + +Detection results in the non-transfer case are shown in Table 1. The proposed MBF method shows the best performance in all cases, and is much superior to all compared methods. LID performs the second-best in most case. $\mathrm { K D + B U }$ gives somewhat good detection performance on CIFAR-10 and SVHN, but performs very poor on ImageNet. It tells that $\mathrm { K D + B U }$ is very sensitive to different databases and networks. M-D gives almost the degenerate results. Note that the results of M-D on + +Table 4: Detection results evaluated by AUROC score $( \% )$ in the data-transfer case. All detectors are trained on the train set, and tested on the out-of-sample set (including 365 images) of the ImageNet database. The best results are highlighted in bold. + +
ImageNet-AlexNetImageNet-VGG-16
DetectorBIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
KD+BU66.369.464.768.884.085.079.985.2
M-D48.749.250.051.054.550.050.456.1
LID69.167.672.971.484.288.489.283.7
MBF99.199.099.699.499.399.599.399.5
+ +Table 5: Attack Failure Rates $( \% )$ of adaptive attacks with different detectors. See context for details. + +
Original CW-L2LIDMBF
CIFAR-100.088.4100.0
SVHN0.085.692.9
+ +CIFAR-10 and SVHN reported in (Lee et al. (2018)) are very high, but their networks (i.e., ResNet (He et al. (2016)) and DenseNet (Huang et al. (2017)) are different from that used in our experiments. It implies that M-D may be suitable for very deep neural networks, but not for shallower networks. + +Detection results in the attack-transfer case evaluated by AUROC and accuracy are presented in Tables 2 and 3, respectively. MBF still shows much better performance in all transfer cases than all compared methods, and the changes among detecting different attacks are very small. It verifies the robustness of MBF to different attacks. + +We also conduct a data-transfer experiment on ImageNet. Specifically, we collect extra 365 images of 8 classes (same with the classes used for the detection training, see Section 4.2) through searching the class names in Baidu and Facebook. These 365 benign and their noisy images can be correctly predicted by both the fine-tuned AlexNet and VGG models. The results on these images are shown in Table 4. MBF still shows the best performance. And, compared to the corresponding results in Table 1, the AUROC/accuracy scores of MBF on detecting different attacks change very gently. It verifies the robustness of MBF to different data sources. + +# 4.4 ADAPTIVE ATTACK AGAINST ADVERSARIAL DETECTION + +Similar to Carlini & Wagner (2017b) and Ma et al. (2018), we also present an adaptive white-box attack, of which the goal is not only to fool the classifier, but also to evade the adversarial detector. It measures how easy to bypass the detector. Specifically, by combining the CW-L2 attack (Carlini & Wagner (2017a)) and the proposed MBF detector, the adaptive attack is formulated as follows: + +$$ +\operatorname * { a r g m i n } _ { \mathbf { x } _ { a d v } \in [ 0 , 1 ] } \| \mathbf { x } - \mathbf { x } _ { a d v } \| _ { 2 } ^ { 2 } + \alpha \cdot \big [ - l ( f _ { \theta } ( \mathbf { x } _ { a d v } ) , y ) + \| D ( f _ { \theta } ( \mathbf { x } _ { a d v } ) ) - D ( f _ { \theta } ( \mathbf { x } ) ) \| _ { 1 } \big ] , +$$ + +where $\mathbf { x }$ denotes the clean input, $y$ indicates the ground-truth label. $l ( \mathbf { x } _ { a d v } , y )$ is the Hinge loss. $D ( f _ { \pmb { \theta } } ( \mathbf { x } _ { a d v } ) )$ represents the extracted features by the the detector $D$ (specified later), from $f _ { \pmb { \theta } } ( \mathbf { x } _ { a d v } )$ . The minimization of the $\ell _ { 1 }$ term $| | D ( f _ { \theta } ( \mathbf { x } _ { a d v } ) ) ^ { * } - D ( f _ { \theta } ( \mathbf { x } ) ) | | _ { 1 }$ encourages the detection features of adversarial and benign examples to be close, such that the detector could be evaded. + +We conduct the above adaptive attack on both CIFAR-10 and SVHN. Specifically, 1000 clean examples are randomly selected from the test set of each database. For each generated adversarial example, if it fails to fool the classifier $f _ { \theta }$ , we record 1. The attack failure rate over all 1000 adversarial examples is computed as the metric. The higher rate indicates that it is more difficult to evade the detector. Since LID performs much better than $\mathrm { K D + B U }$ and M-D in above experiments, here we only compare LID and the proposed MBF detector. For clarity, the detection features are only extracted from the soft-max layer for both LID and MBF. The original CW-L2 attack (i.e., without the $\ell _ { 1 }$ term) is also presented as the baseline. With the same setting in (Ma et al. (2018)), the trade-off parameter $\alpha$ is determined by binary search within the range $[ \bar { 1 0 } ^ { - 3 } , 1 0 ^ { 6 } ]$ . The results are shown in Table 5. The attack failure rates of MBF on both databases are much higher than those of LID. It tells that it is more difficult to evade MBF than LID. + +# 4.5 HYPOTHESIS TEST VIA KOLMOGOROV-SMIRNOV TEST + +Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)) is a non-parametric test method in statistics, to test whether a sample follows a reference probability distribution (one-sample KS test), or whether two samples follow the same distribution (two-sample KS test). Specifically, in one-sample KS test, the distance between the empirical distribution function of one sample and the cumulative distribution function of the reference probability distribution is measured. Then, the $p$ -value corresponding to the obtained distance is computed. If the $p$ -value is larger than the significance level $\alpha$ (here we set $\alpha = 0 . 0 5$ ), then the null hypothesis that the sample follows the reference distribution is accepted; otherwise, rejected. Similarly, in two-sample KS test, the distance between the empirical distribution functions of two samples is computed. If the corresponding $p$ -value is larger than $\alpha$ , then it accepts that two samples follow the same distribution. The KS test conducted below is implemented by the python function scipy.stats.ks 2samp7. + +Table 6: The $p$ -value of KS hypothesis test among clean samples, noisy samples, and adversarial samples crafted by four methods, respectively, on ImageNet-AlexNet. + +
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
+ +Table 7: The $p$ -value of two-sample KS test among MBF coefficients of different types of examples. + +
ImageNet-VGG-16(clean, noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
+ +Table 8: The $p$ -value of KS hypothesis test among MBF coefficients of different data sources. + +
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
(train set,test set)0.8390.8580.3680.8420.2460.221
(train set, out-of-sample set)0.0490.0540.0940.4070.7370.088
+ +Hypothesis test 1. Here we verify that whether the posterior vectors of both adversarial and benign examples follow GGD. We denote the posterior vector of one adversarial example as $\bf p \mathrm { _ { a d v } }$ , and that of one benign example as $\mathbf { p } _ { \mathrm { { b e n } } }$ . The distribution of GGD is denoted as ${ \mathcal { P } } _ { \mathrm { G G D } }$ . Then, we conduct two one-sample KS tests, including: + +• H1.1 The test of adversarial examples: $H _ { 0 }$ : $\mathbf { p } _ { \mathrm { a d v } } \sim \mathcal { P } _ { \mathrm { G G D - a d v } }$ ; $H _ { 1 }$ : padv 6∼ PGGD-adv. +• H1.2 The test of benign examples: $H _ { 0 }$ : $\mathbf { p } _ { \mathrm { b e n } } \sim \mathcal { P } _ { \mathrm { G G D - b e n } }$ ; $H _ { 1 }$ : pben 6∼ PGGD-ben. + +In H1.1, the reference distribution $\mathcal { P } _ { \mathrm { G G D - a d v } }$ is firstly estimated from $\mathbf { p } _ { \mathrm { a d v } }$ , using the estimated method proposed in (Lasmar et al. (2009)). To alleviate the uncertainty of the estimation, we draw 500 samples from the estimated $\mathcal { P } _ { \mathrm { G G D - a d v } }$ . Then, we conduct the two-sample KS test between $\mathbf { p } _ { \mathrm { a d v } }$ and these 500 samples respectively. The average $p$ -value over 1000 tests is recorded. The mean of the average $p$ -values over the whole database is reported. H1.2 is conducted similarly. The results tested on ImageNet-AlexNet are shown in Table 6. Due to the space limit, results on other databases and models will be presented in the supplementary material. In all cases, the $p$ -values are larger than the significance level 0.05. Hence, we can conclude that the posterior vectors of both adversarial and benign examples follow GGD. However, note that the parameters of their corresponding GGD are different, which will be verified in the following test. + +Hypothesis test 2. Here we verify that whether the extracted MBF features of adversarial and benign examples follow the same empirical distribution. We denote the MBF feature vector of one adversarial example as $\mathbf { m _ { \mathrm { a d v } } }$ , and the corresponding empirical distribution is denoted as $\hat { \mathcal { P } } _ { \mathrm { a d v - M B F } }$ . Similarly, we define $\mathbf { m } _ { \mathrm { b e n } }$ and $\hat { \mathcal { P } } _ { \mathrm { b e n - M B F } }$ for benign examples. Then, we conduct the following four two-sample KS tests: + +• H2.1 The test between adversarial and benign examples: $H _ { 0 }$ : $\hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } = \hat { \mathcal { P } } _ { \mathrm { b e n - M B F } }$ ; $H _ { 1 }$ : $\hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } \neq \hat { \mathcal { P } } _ { \mathrm { b e n - M B F } }$ . +• H2.2 The test between adversarial examples crafted from different attack methods: $H _ { 0 }$ : $\hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { a t t a c k - 1 } } = \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { a t t a c k - 2 } }$ ; $H _ { 1 } \colon \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { a t t a c k - 1 } } \neq \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { a t t a c k - 2 } }$ . +H2.3 The test between adversarial examples from different data sources (i.e., train, test and +out-of-sample set): $H _ { 0 }$ : $\begin{array} { r } { \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { s o u r c e - 1 } } = \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { s o u r c e - 2 } } } \end{array}$ ; $H _ { 1 } \colon \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { s o u r c e - 1 } } \neq \hat { \mathcal { P } } _ { \mathrm { a d v - M B F } } ^ { \mathrm { s o u r c e - 2 } }$ . + +![](images/e6e248dabc29baa792b249691b317cc1ab671db8b1d5f02fc3e9a7bc0e6db03a.jpg) +Figure 1: Statistics (mean $\pm$ standard deviation) of MBF coefficients on train (top row), test (median row), and out-of-sample (bottom row) set of ImageNet-VGG-16. + +• H2.4 The test between benign examples from different data sources (i.e., train, test and out-of-sample set): $H _ { 0 }$ : $\hat { \mathcal { P } } _ { \mathrm { b e n - M B F } } ^ { \mathrm { s o u r c e - 1 } } = \hat { \mathcal { P } } _ { \mathrm { b e n - M B F } } ^ { \mathrm { s o u r c e - 2 } }$ ; $H _ { 1 } \colon \hat { \mathcal { P } } _ { \mathrm { b e n - M B F } } ^ { \mathrm { s o u r c e - 1 } } \neq \hat { \mathcal { P } } _ { \mathrm { b e n - M B F } } ^ { \mathrm { s o u r c e - 2 } }$ . + +In above four two-sample KS tests, we test on the 16-dimensional MBF features extracted from the soft-max layer. Since the implementation scipy.stats.ks 2samp cannot compare two vectors, we compare the feature of each dimension separately, then report the average $p$ -value over all dimensions. Specifically, when comparing two sets of samples, we firstly concatenate the feature of each dimension across all samples in the same set, leading to 16 long vectors for each set. Then, each pair of two long vectors corresponding to the same dimension from two sets is compared by KS test. The average $p$ -value over all 16 dimensions is reported. The $p$ -values of ${ \bf H } 2 . 1$ are shown in Table 7 (see the column “(clean, BIM)”). The $p$ -values on both train and test set are 0. Thus, the hypothesis $H _ { 0 }$ is rejected, i.e., the MBF features of adversarial and benign examples follow different distributions. The $p$ -values of ${ \bf H } 2 . 2$ are shown in Table 7. We pick two groups of attack methods, i.e., (BIM, DeepFool) and (CW-L2, R-PGD). The $p$ -values of (BIM, DeepFool) are 0, while the $p$ -values of (CW-L2, R-PGD) are larger than 0.05. It demonstrates that the distributions of adversarial examples crafted from different attack methods are possible to be different. The $p$ -values of ${ \bf H } 2 . 3$ are shown in Table 8. The $p$ -values of all types of adversarial examples exceed 0.05. Thus, the MBF features of adversarial examples from different data sources follow the same distribution. The $p$ -values of H2.4 are shown in Table 8. Only the $p$ -value of “(train set, out-of-sample set)” of clean examples is slightly lower than 0.05, while the values of other cases exceed 0.05. Thus, in most cases, the MBF features of benign examples from different data sources follow the same distribution. + +From above analysis, we obtain the following conclusions: 1) The extracted MBF features of adversarial and benign examples follow different empirical distributions. It explains why MBF features are effective for detecting adversarial and benign examples; 2) The extracted MBF features of adversarial/benign examples from the train, test and out-of-sample sets follow the same empirical distribution. Although the extracted MBF features of adversarial examples crafted from different attack methods may not follow the same empirical distribution, the significant difference between benign and different adversarial distributions can still lead to the good detection performance in the attack-transfer case. It explains why MBF features are robust across different attack methods and different data sources. Moreover, we visualize the statistics of each dimension of MBF features, i.e., mean and standard deviation, as shown in Fig. 1. These visualizations also support above conclusions. Due to space limit, more KS tests and visualizations on different databases and networks will be presented in the supplementary material. + +# 5 CONCLUSION + +This work has proposed a novel adversarial detection method, dubbed MBF. The assumption behind is that the internal responses of the classification network of both adversarial and benign examples follow the generalized Gaussian distribution (GGD), but with different shape factors. The magnitude of Benford-Fourier coefficient is a function w.r.t. the shape factor, and can be easily estimated based on responses. Thus, it can serve as the discriminative features between adversarial and benign examples. The extensive experiments conducted on several databases, as well as the empirical analysis via KS test, demonstrate the superior effectiveness and robustness to different attacks and different data sources of the proposed MBF method, over state-of-the-art detection methods. + +# REFERENCES + +Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pp. 39–57. IEEE, 2017a. + +Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pp. 3–14. ACM, 2017b. + +Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee, 2009. + +Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. + +Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. + +Nathaniel R Goodman. Statistical analysis based on a certain multivariate complex gaussian distribution (an introduction). The Annals of mathematical statistics, 34(1):152–177, 1963. + +Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and Patrick McDaniel. On the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. + +Dan Hendrycks and Kevin Gimpel. Early methods for detecting adversarial images. arXiv preprint arXiv:1608.00530, 2016. + +Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4700–4708, 2017. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. + +Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html, 55, 2014. + +Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533, 2016. + +Nour-Eddine Lasmar, Youssef Stitou, and Yannick Berthoumieu. Multiscale skewed heavy tailed model for texture analysis. In 2009 16th IEEE International Conference on Image Processing (ICIP), pp. 2281–2284. IEEE, 2009. + +Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems, pp. 7167–7177, 2018. + +Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764– 5772, 2017. + +Jiajun Lu, Theerasit Issaranon, and David Forsyth. Safetynet: Detecting and rejecting adversarial examples robustly. In Proceedings of the IEEE International Conference on Computer Vision, pp. 446–454, 2017. + +Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck, Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using local intrinsic dimensionality. In International Conference on Learning Representations, 2018. + +Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. + +Frank J Massey Jr. The kolmogorov-smirnov test for goodness of fit. Journal of the American statistical Association, 46(253):68–78, 1951. + +Jan Hendrik Metzen, Tim Genewein, Volker Fischer, and Bastian Bischoff. On detecting adversarial perturbations. In International Conference on Learning Representations, 2017. + +Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2574–2582, 2016. + +Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. + +Tianyu Pang, Chao Du, Yinpeng Dong, and Jun Zhu. Towards robust detection of adversarial examples. In Advances in Neural Information Processing Systems, pp. 4579–4589, 2018. + +Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy, pp. 372–387. IEEE, 2016. + +A. Papoulis and S.U. Pillai. Probability, Random Variables, and Stochastic Processes. McGrawHill Series in Electrical and Computer Engineering. McGraw-Hill, 2001. ISBN 9780072817256. URL https://books.google.com.hk/books?id $=$ Mal4OgAACAAJ. + +Cecilia Pasquini, Fernando Perez-Gonz ´ alez, and Giulia Boato. A benford-fourier jpeg compression ´ detector. In 2014 IEEE International Conference on Image Processing, pp. 5322–5326. IEEE, 2014. + +Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–572, 1901. + +Fernando Perez-Gonz ´ alez, Greg L Heileman, and Chaouki T Abdallah. Benford’s lawin image ´ processing. In 2007 IEEE International Conference on Image Processing, volume 1, pp. I–405. IEEE, 2007. + +Murray Rosenblatt. A central limit theorem and a strong mixing condition. Proceedings of the National Academy of Sciences of the United States of America, 42(1):43, 1956. + +Pouya Samangouei, Maya Kabkab, and Rama Chellappa. Defense-gan: Protecting classifiers against adversarial attacks using generative models. arXiv preprint arXiv:1805.06605, 2018. + +Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 2013. + +Mahesh K Varanasi and Behnaam Aazhang. Parametric generalized gaussian density estimation. The Journal of the Acoustical Society of America, 86(4):1404–1415, 1989. + +Zhihao Zheng and Pengyu Hong. Robust detection of adversarial attacks by modeling the intrinsic properties of deep neural networks. In Advances in Neural Information Processing Systems, pp. 7913–7922, 2018. + +# A APPENDIX + +A.1 PROOF OF THEOREM 1 + +Proof 1 Assume that all the variables $\{ \mathcal { T } _ { m } \} _ { m = 1 , \dots , M }$ are independent and identically distributed. Applying the central limit theorem Rosenblatt (1956) to the real and imaginary parts of $_ { \mathcal { V } }$ , we can obtain that both parts asymptotically follow the Gaussian distribution + +$$ +\mathcal { V } = \frac { 1 } { M } \sum _ { m = 1 } ^ { M } \mathcal { T } _ { m } \sim \mathcal { N } \big ( E ( \mathcal { T } _ { 1 } ) , \frac { 1 } { M } D ( \mathcal { T } _ { 1 } ) \big ) , +$$ + +where + +$$ +\begin{array} { l } { \displaystyle E ( \mathcal { T } _ { 1 } ) = E \left( e ^ { - j 2 \pi n \log _ { 1 0 } | \mathcal { X } _ { 1 } | } \right) = \int _ { - \infty } ^ { + \infty } \mathcal { P } _ { \mathcal { X } _ { 1 } } ( x ) \cdot e ^ { - j 2 \pi n \log _ { 1 0 } | x | } d x = a _ { n } , } \\ { \displaystyle D ( \mathcal { T } _ { 1 } ) = E \big ( \left| \mathcal { T } _ { 1 } \right| ^ { 2 } \big ) - \left| E \left( \mathcal { T } _ { 1 } \right) \right| ^ { 2 } = E \big ( \left| e ^ { - j 2 \pi n \log _ { 1 0 } | \mathcal { X } _ { 1 } | } \right| ^ { 2 } \big ) - \left| a _ { n } \right| ^ { 2 } = 1 - \left| a _ { n } \right| ^ { 2 } . } \end{array} +$$ + +The PDF of $\mathcal { V }$ can be rewritten as follows: + +$$ +\mathcal { V } \sim \mathcal { N } ( a _ { n } , \frac { 1 - | a _ { n } | ^ { 2 } } { M } ) . +$$ + +Thus, we obtain + +$$ +\mathcal { E } \sim \mathcal { N } ( 0 , \frac { 1 - | a _ { n } | ^ { 2 } } { M } ) . +$$ + +Besides, the pseudo variance Goodman (1963) of $\mathcal { T } _ { 1 }$ is + +$$ +\mathcal { T } _ { \mathcal { T } _ { 1 } , \mathcal { T } _ { 1 } } = E ( \mathcal { T } _ { 1 } ^ { 2 } ) - E ( \mathcal { T } _ { 1 } ) ^ { 2 } = E \big ( e ^ { - j 2 \pi n \log _ { 1 0 } \left| \mathcal { X } _ { 1 } ^ { 2 } \right| } \big ) - a _ { n } ^ { 2 } = a _ { 2 n } - a _ { n } ^ { 2 } . +$$ + +Correspondingly, + +$$ +\mathcal { I } _ { \mathcal { E } , \mathcal { E } } = \frac { 1 } { M } \mathcal { I } _ { \mathcal { T } _ { 1 } , \mathcal { T } _ { 1 } } = \frac { a _ { 2 n } - a _ { n } ^ { 2 } } { M } . +$$ + +Since $\left( a _ { 2 n } - a _ { n } ^ { 2 } \right)$ is bounded, we have $\begin{array} { r } { \operatorname* { l i m } _ { M \to \infty } \mathcal { I } _ { \mathcal { E } , \mathcal { E } } = 0 } \end{array}$ . Thus, we obtain that the random variable $\mathcal { E }$ follows circularly-symmetric complex Gaussian distribution, because the sufficient and necessary condition is that mean value and pseudo variance equal zero Goodman (1963). It implies that both the real and imaginary part of $\mathcal { E }$ follow the same Gaussian distribution and they are independent. Thus, the magnitude of this complex random variable follows the Rayleigh distribution Papoulis & Pillai (2001), and the probability density function of $| \mathcal { E } |$ can be formulated as + +$$ +\mathcal { P } _ { | \varepsilon | } \left( r \right) = \frac { r } { s ^ { 2 } } e ^ { - r ^ { 2 } / 2 s ^ { 2 } } , +$$ + +where s is the scale parameter. Knowing the properties of the Rayleigh distributionPapoulis & Pillai (2001), we have: + +$$ +s ^ { 2 } = { \frac { 1 } { 2 } } D ( { \mathcal { E } } ) = { \frac { 1 - | a _ { n } | ^ { 2 } } { 2 M } } . +$$ + +Utilizing the fact that $\left| a _ { n } \right| ^ { 2 }$ is close to 0 when n is a modest number, we obtain that $\begin{array} { r } { D \left( \mathcal { E } \right) \approx \frac { 1 } { M } } \end{array}$ , leading to $\textstyle s ^ { 2 } = { \frac { 1 } { 2 M } }$ . Then, we obtain + +$$ +E { \bigl ( } | { \mathcal { E } } | { \bigr ) } = { \frac { 1 } { 2 } } { \sqrt { \frac { \pi } { M } } } , D { \bigl ( } | { \mathcal { E } } | { \bigr ) } = { \frac { 4 - \pi } { 4 M } } . +$$ + +It is easy to observe that both $E { \big ( } | { \mathcal { E } } | { \big ) }$ and $D ( \left. \boldsymbol { \mathcal { E } } \right. )$ are close to zero when the number of samples M increases. It implies that the estimation error $\varepsilon _ { n }$ gets closer to $O$ as $M$ increases. + +Table 9: The $p$ -value of KS hypothesis test among clean samples, noisy samples, and adversarial samples crafted by four methods, respectively. + +
cleannoisyBIMCW-L2DeepFoolR-PGD
CIFAR-10train set0.0740.0740.1580.1390.1550.159
test set0.0720.0720.1580.1380.2160.216
SVHNtrain set0.1040.1040.2490.1940.2220.245
test set0.1060.1060.2490.1930.2190.239
ImageNet-AlexNettrain set0.1160.1160.2530.2350.2420.253
test set0.1170.1170.2610.2360.2430.259
ImageNet-VGG-16train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
+ +Table 10: The $p$ -value of two-sample KS test among MBF coefficients of different types of examples. + +
(clean,noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
CIFAR-10train set0.6400.0000.0000.003
test set0.7640.0000.0300.000
SVHNtrain set0.5590.0000.0300.000
test set0.6850.0000.0000.000
ImageNet-AlexNettrain set1.0000.0000.0020.134
test set0.9930.0000.4770.000
ImageNet-VGG-16train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
+ +Table 11: The $p$ -value of KS test among MBF coefficients from different data sources. + +
clean noisyBIMCW-L2DeepFoolR-PGD
CIFAR-10(train set, test set)0.120 0.1880.3330.4560.5840.223
SVHN(train set, test set)0.494 0.7080.5230.1470.4910.645
ImageNet- AlexNet(train set, test set) (train set, out-of-sample)0.153 0.148 0.000 0.0000.370 0.1690.605 0.1010.410 0.2060.386 0.158
ImageNet- VGG-16(train set, test set) (train set, out-of-sample)0.839 0.858 0.049 0.0540.368 0.0940.842 0.4070.246 0.7370.221 0.088
+ +![](images/951766bb96a13d49ac27f0e9767ec43586237ffa3aba2ff5527718e6f9c18193.jpg) +Figure 2: Statistics (mean $\pm$ standard deviation) of MBF coefficients on train (top row) and test (bottom row) set of CIFAR-10. + +![](images/f90ea0187e4a20a3df47ccc875e9571b38f3c0e93fe3a9c53bf08df094171e49.jpg) +Figure 3: Statistics (mean $\pm$ standard deviation) of MBF coefficients on train (top row) and test (bottom row) set of SVHN. + +![](images/a2307b9dd6d8b208fa0ca9565e779759c161a7c1546a2a84a8e287388b4c42f5.jpg) +Figure 4: Statistics (mean $\pm$ standard deviation) of MBF coefficients on train (top row) and test (median row), and out-of-sample (bottom row) set of ImageNet-AlexNet. + +Table 12: Detection results evaluated by AUROC score $( \% )$ on nearly 4000 images from the testset of MNIST. The best results are highlighted in bold. + +
Iteration numberin each GD runNumber of GD runsBIMCW-L2DeepFoolR-PGD
200100.7440.9800.7020.746
100100.6890.9730.6500.690
10050.6640.9690.6280.666
10020.6120.9590.5860.614
MBF1.0000.9911.0001.000
+ +# A.2 ADDITIONAL EMPIRICAL ANALYSIS + +Here we present additional empirical analysis on more databases and networks, as shown in Tables 9, 10 and 11. The $p$ -values in most cases also supports the conclusions obtained in the main manuscript. We also present more visualizations in Figs. 2, 3 and 4. These visualizations also demonstrate the distinct difference of MBF features between adversarial and benign examples. + +# A.3 COMPARISON BETWEEN DEFENSE-GAN + +We adopt official code8 of Defense-GAN (Samangouei et al. (2018)) and train an generative model on train set of MNIST dataset after 130, 000 epochs (As the official code of Defense-GAN currently only supports three datasets including MNIST, Fashion-MNIST, and CelebA). The classification model attains an accuracy of $9 8 . 7 \%$ on the test set of MNIST dataset. We still utilize four attack strategies, including BIM, CW-L2, DeepFool and R-PGD. After removing images that are misclassified and failed to attack, we pick 6000 images to train SVMs based on our MBF features, and leave nearly 4000 images for testing. There are two key hyper-parameters in Defense-GAN, including the iteration number in each GD run, and the total number of GD runs. We have tried multiple settings of these two hyper-parameters. Table 12 show the AUROC score on nearly 4000 images. Our MBF method exceeds Defense-GAN by a large margin in quantitative performance. + +# A.4 DETAILED PARAMETERS OF ADVERSARIAL ATTACK + +Four popular adversarial attack methods are adopted to craft adversarial examples, including basic iterative method (BIM), CarliniWagnerL2Attack (CW-L2), DeepFool, and random projected gradient descent (R-PGD). We emphasize detailed parameters of these attack strategies for reproducibility, which are BIM $\mathrm { ( e p s { = } 0 . 3 }$ , stepsize $= 0 . 0 5$ , iterations ${ \it \Omega } = 1 0$ ), CW-L2 (binary search steps $^ { = 5 }$ , confidence $= 0 . 0$ , learning rate $= 0 . 0 0 5$ , max iterations $\mathord { \left. \begin{array} { r l l } { - 1 0 0 0 } \end{array} \right. }$ ), DeepFool (max step $\mathord { \mathrm { \ s } } = 1 0 0$ ), R-PGD $\mathrm { \ e p s { = } 0 . 3 }$ , stepsize ${ \mathrm { \Omega } } { = } 0 . 0 1$ , iterations $\scriptstyle = 4 0$ ). All of the attack strategies are implemented by Foolbox. \ No newline at end of file diff --git a/parse/train/ryeK6nNFDr/ryeK6nNFDr_content_list.json b/parse/train/ryeK6nNFDr/ryeK6nNFDr_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f703e78b1da9df03514863da1bb507dee0b7a212 --- /dev/null +++ b/parse/train/ryeK6nNFDr/ryeK6nNFDr_content_list.json @@ -0,0 +1,1895 @@ +[ + { + "type": "text", + "text": "EFFECTIVE AND ROBUST DETECTION OF ADVER-SARIAL EXAMPLES VIA BENFORD-FOURIER COEFFI-CIENTS", + "text_level": 1, + "bbox": [ + 174, + 98, + 821, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 195, + 398, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 261, + 544, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial examples have been well known as a serious threat to deep neural networks (DNNs). To ensure successful and safe operations of DNNs on realworld tasks, it is urgent to equip DNNs with effective defense strategies. In this work, we study the detection of adversarial examples, based on the assumption that the output and internal responses of one DNN model for both adversarial and benign examples follow the generalized Gaussian distribution (GGD), but with different parameters (i.e., shape factor, mean, and variance). GGD is a general distribution family to cover many popular distributions (e.g., Laplacian, Gaussian, or uniform). It is more likely to approximate the intrinsic distributions of internal responses than any specific distribution. Besides, since the shape factor is more robust to different databases rather than the other two parameters, we propose to construct discriminative features via the shape factor for adversarial detection, employing the magnitude of Benford-Fourier coefficients (MBF), which can be easily estimated using responses. Finally, a support vector machine is trained as the adversarial detector through leveraging the MBF features. Through the Kolmogorov-Smirnov (KS) test (Massey Jr (1951)), We empirically verify that: 1) the posterior vectors of both adversarial and benign examples follow GGD; 2) the extracted MBF features of adversarial and benign examples follow different distributions. Extensive experiments in terms of image classification demonstrate that the proposed detector is much more effective and robust on detecting adversarial examples of different crafting methods and different sources, in contrast to state-of-the-art adversarial detection methods. ", + "bbox": [ + 233, + 291, + 764, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 621, + 336, + 637 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have achieved a remarkable success in many important applications, such as image classification, face recognition, object detection, etc. In the meanwhile, DNNs have been shown to be very vulnerable to adversarial examples. However, many real-world scenarios have very restrictive requirements about the robustness of DNNs, such as face verification for login, or semantic segmentation in autonomous driving. Adversarial examples are a serious threat to the applications of DNNs to these important tasks. Since many kinds of adversarial attack methods have been proposed to fool DNNs, it is more urgent to equip effective defensive strategies to ensure the safety of deep models in real-world applications. However, defense seems to be more challenging than attack, as it has to face adversarial examples from unknown crafting methods and unknown data sources. Typical defensive strategies include adversarial training, adversarial de-noising, and adversarial detection. Compared to the former two strategies, adversarial detection is somewhat more cost-effective, as it often does not need to re-train or modify the original DNN model. ", + "bbox": [ + 174, + 652, + 825, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are two main challenges for adversarial detection. (1) The adversarial examples are designed to camouflage themselves to be close to the corresponding benign examples in the input space. Then, where and how to extract the discriminative information to train the detector? (2) The data sources and the generating methods of adversarial examples are often inaccessible to the detector. In this case, the detector can be stably effective across different data sources and different attack methods? In other words, a good adversarial detector is required to be not only effective to distinguish adversarial and benign examples, but also robust to different data sources and attack methods. ", + "bbox": [ + 174, + 827, + 823, + 922 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To satisfy the first requirement of effectiveness, we utilize the other principle of crafting adversarial examples that the outputs between benign and adversarial examples should be large, to encourage the change of the final prediction. It means that the imperceptible difference between benign and adversarial examples in the input space is enlarged along the DNN model, leading to the significant difference in the output space. Inspired by this fact, we assume that the output or the responses of internal layers of the DNN model should include the discriminative information for benign and adversarial examples. A few works have attempted to extract different types of discriminative features from the output or the internal responses, such as kernel density (KD) (Feinman et al. (2017)) and the local intrinsic dimensionality (LID) (Ma et al. (2018)), etc. To achieve the robustness, the extracted discriminative features should model the intrinsic difference between adversarial and benign examples, rather than the difference from the changes of data sources or attack methods. Many existing methods have shown the effectiveness to some extent on detecting adversarial examples of specific data sources and attack methods. However, their robustness, especially across different data sources, has not been well studied and verified. ", + "bbox": [ + 174, + 103, + 825, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel detection method based on the assumption that the internal responses of both adversarial and benign examples follow the generalized Gaussian distribution (GGD) (Varanasi & Aazhang (1989)), but with different parameters, including the shape factor, mean, and variance. The rationale behind this assumption is that GGD covers many popular distributions with varied shape factors (such as Laplacian, Gaussian, or uniform distribution), such that GGD is more likely to approximate the intrinsic response distributions rather than one specific distribution. Moreover, mean and variance of GGD may vary significantly with respect to different classes and data sources, even for benign examples, while the shape factor is more robust. For example, the mean and variance of two Gaussian distributions could be totally different, but their shape factors are the same (i.e., 2). Thus, we propose to use the shape factor as an effective and robust discrimination between adversarial and benign examples. However, it is difficult to exactly estimate the shape factor in practice. We resort to the magnitude of Benford-Fourier coefficients (Pasquini et al. (2014)), which is a function of the shape factor. It can be easily estimated using internal responses, according to the definition of Fourier transform. Then, the magnitudes estimated from internal responses of different convolutional layers are concatenated as a novel representation. Finally, a support vector machine (SVM) (Vapnik (2013)) is trained using the new representations as the adversarial detector. Extensive experiments carried out on several databases verify the effectiveness and robustness of the proposed detection method. To further verify the rationale of our assumption, we present the empirical analysis through the Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)). The KS test verifies that 1) the posterior vectors of both adversarial and benign examples predicted by the CNN model follow the distribution of GGD, but with different parameters, and 2) the MBF features of adversarial and benign examples follow different distributions, and the MBF features of adversarial examples crafted from different attack methods follow the same distribution, as well as that the MBF features of adversarial/benign examples from different data sources follow the same distribution. Moreover, we visualize the statistics (i.e., mean $\\pm$ standard deviation) of the extracted MBF features for adversarial and benign examples. The visualization reveals the distinct difference between adversarial and benign examples. These empirical analysis demonstrates the effectiveness and robustness of the proposed MBF detector. ", + "bbox": [ + 174, + 306, + 825, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 715, + 339, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section we focus on reviewing existing works about adversarial detection, while those of other adversarial defensive strategies are out of scope of this work. ", + "bbox": [ + 174, + 750, + 823, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The general idea of most existing detection methods is learning or constructing a new representation to discriminate adversarial and benign examples, utilizing the outputs or immediate responses of an original classification network. Li & Li (2017) trained a cascading classifier based on the principal component analysis (PCA) (Pearson (1901)) statistics of responses from each convolutional layer of the defended convolutional neural network (CNN) model. An example is recognized as benign if all single classifiers of the cascade predict it as benign, otherwise adversarial. Lu et al. (2017) proposed SafetyNet by adding a RBF-SVM classifier to detect adversarial examples, at the end of the original classification network. Metzen et al. (2017) proposed a detection network along with an original classification network, which takes the internal responses of the original network as inputs. It shows effectiveness on detecting adversarial examples generated by simple attacks (such as FGSM (Goodfellow et al. (2014)) and JSMA (Papernot et al. (2016))), while performs much worse when facing more advanced attacks (such as C&W (Carlini & Wagner (2017a))). It tells that this method is sensitive to attack methods. Grosse et al. (2017) attempted to detect adversarial examples using the statistical test of maximum mean discrepancy (MMD). Although above detection methods show effectiveness on some attack methods and some databases, but a thorough evaluation presented in Carlini & Wagner (2017b) has shown that these methods are sensitive to attack methods or databases, and they can be somewhat easily invaded by new attacks. ", + "bbox": [ + 174, + 785, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some recent works proposed to utilize neighboring samples in the same database to construct a better representation of a current sample. Feinman et al. (2017) defined two metrics based on the responses of the final hidden layer of the classification neural network, including kernel density estimation (KDE) and Bayesian neural network uncertainty (BU). If the metric score of KDE/BU is lower/higher than a pre-defined threshold, then the example is predicted as adversarial. Ma et al. (2018) utilized the local intrinsic dimensionality (LID) to measure the characterization of adversarial regions of DNNs. LID describes the distance between one example and its $k$ -nearest neighboring sample in the feature space of immediate responses of the original classification network. The distances computed from different layers are concatenated as the example representation, which is then used to train a shallow classifier to discriminate adversarial and benign examples. Zheng & Hong (2018) defined the intrinsic hidden state distribution (IHSD) of the responses of the original classification network to model different classes. The Gaussian mixture model (GMM) was used to approximate IHSD of each class. Then, the posterior probability of one sample assigned to GMM is computed as the metric. If the probability is lower than a pre-defined threshold, then it is recognized as adversarial. Lee et al. (2018) computed the class-conditional Gaussian distribution of the responses of the original classification network based on the whole training set. Then, the Mahalanobis distance between one sample and its nearest class-conditional Gaussian distribution is used as the metric for detection. If the distance is larger than a pre-defined threshold, then it is detected as adversarial. Compared to some aforementioned single-representation-based detection methods, these joint-representation-based methods showed better performances on some databases. However, the detection cost for each example is much higher, as the responses of its neighboring samples should also be computed. Besides, since the representation is highly dependent on the neighbors or all training examples, the detection performance may be sensitive to data sources, which will be studied in later experiments. ", + "bbox": [ + 174, + 208, + 825, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are also some other approaches that do not construct representations from the responses of the original classification network. Hendrycks & Gimpel (2016) adopted PCA statistics to discriminate adversarial and benign images, independent of any DNN model. However, the study presented in Carlini & Wagner (2017b) has demonstrated that this method works for MNIST but not for CIFAR-10, and PCA statistics are not robust features to detect adversarial images. Pang et al. (2018) proposed a novel loss called reverse cross-entropy (RCE) to train the classification network, such that the distance measured by kernel density (Feinman et al. (2017)) between adversarial and benign examples could be enlarged. Samangouei et al. (2018) proposed Defense-GAN to model the distribution of benign examples using a generative adversarial network (GAN). If the Wasserstein distance between one example and its corresponding example generated by the GAN is larger than a pre-defined value, then it is detected as adversarial. However, these above three methods are much more costly than other detection methods. ", + "bbox": [ + 174, + 547, + 825, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 736, + 339, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 GENERALIZED GAUSSIAN DISTRIBUTION ", + "text_level": 1, + "bbox": [ + 174, + 768, + 501, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume that a random variable $\\mathcal { X } ~ \\in ~ \\mathbb { R } ^ { d }$ follows the generalized Gaussian distribution (GGD) (Varanasi $\\&$ Aazhang (1989)). Then, its probability density function (PDF) is formulated with two positive parameters, including the shape factor $c$ and the standard deviation $\\sigma$ , as follows ", + "bbox": [ + 174, + 794, + 825, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/749d0e34905c0f5b3529b4411d95ccd4fd3b1b23f07643de44cfdf9789915504.jpg", + "text": "$$\n\\mathcal { P } _ { \\mathcal { X } } ( x ) = A \\cdot e ^ { - \\lvert \\beta x \\rvert ^ { c } } ,\n$$", + "text_format": "latex", + "bbox": [ + 426, + 844, + 571, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where β = 1σ \u0000 Γ(3/c)Γ(1/c) \u0001 and $\\begin{array} { r } { A = \\frac { \\beta c } { 2 \\Gamma ( 1 / c ) } } \\end{array}$ , with $\\Gamma ( \\cdot )$ being the Gamma function. Note that the mean parameter $\\mu$ is omitted above, as $\\mu$ has no relation with the shape of distribution and we set it as 0 without loss of generality. A nice characteristic of GGD is that it covers many popular distributions ", + "bbox": [ + 174, + 875, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with varied shape factors. For example, when $c = 1$ , then it becomes the Laplacian distribution; when $c = 2$ , then it is the Gaussian distribution with a variance of √ √ $\\sigma ^ { 2 }$ ; when $c \\to + \\infty$ , then it is specified as a uniform distribution on $( - \\sqrt { 2 } \\sigma , \\sqrt { 2 } \\sigma )$ . ", + "bbox": [ + 174, + 103, + 825, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 BENFORD-FOURIER COEFFICIENTS ", + "text_level": 1, + "bbox": [ + 176, + 162, + 459, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although generalized Gaussian distribution (GGD) is able to cover a bunch of distributions, it is hard to depict the exact forms of GGD precisely. To this end, we further define a random variable $\\mathcal { Z } = \\log _ { 1 0 } | \\mathcal { X } |$ mod 1 for detecting and distinguishing different form of GGD, of which the PDF is formulated by means of Fourier Series as Perez-Gonz ´ alez et al. (2007), with the fundamental period ´ being fixed as $2 \\pi$ , ", + "bbox": [ + 174, + 188, + 825, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/46c52c4fa59c4e2d00b3ed521c2e6873b0ae03399402a4950cbde421f3f6ffb7.jpg", + "text": "$$\n\\mathcal { P } _ { Z } ( z ) = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor A _ { n } \\cos { ( 2 \\pi n z ) } + B _ { n } \\sin { ( 2 \\pi n z ) } \\rfloor = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor a _ { n } \\rfloor \\cos ( 2 \\pi n z + \\phi _ { n } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 194, + 263, + 782, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z \\in [ 0 , 1 )$ corresponds to the domain of random variable $\\mathcal { Z }$ , the phase of Fourier Series is explained as $\\begin{array} { r } { \\phi _ { n } = \\arctan \\left( - \\frac { B _ { n } } { A _ { n } } \\right) } \\end{array}$ , and the magnitude denotes $| a _ { n } | = \\sqrt { A _ { n } ^ { 2 } + B _ { n } ^ { 2 } }$ . $a _ { n } = | a _ { n } | \\cdot e ^ { j \\phi _ { n } }$ denotes the $n$ -th Fourier coefficient of $\\mathcal { P } _ { \\mathcal { Z } } ( z )$ evaluated at $2 \\pi n$ , and its definition is ", + "bbox": [ + 176, + 311, + 823, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/efbb6c783d4614a8299f4216750213fe7c22579c183a5e5cb35dae6b6e055e6b.jpg", + "text": "$$\na _ { n } = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { Z } } ( z ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } z } d z = \\frac { 2 A e ^ { \\frac { j 2 \\pi n \\log \\beta } { \\log 1 0 } } } { \\beta c } \\cdot \\Gamma \\Bigg ( \\frac { - j 2 \\pi n + \\log 1 0 } { c \\log 1 0 } \\Bigg ) .\n$$", + "text_format": "latex", + "bbox": [ + 240, + 364, + 758, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$a _ { n }$ is also called as Benford-Fourier coefficient. Note that $a _ { n }$ is a complex number, and its magnitude can be calculated as follows ", + "bbox": [ + 173, + 409, + 825, + 438 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/0313939955d3e7556cd3ce3a967cf82aad04df7f0b0783863901fa195bb87a35.jpg", + "text": "$$\n| a _ { n } | = \\bigg ( \\prod _ { k = 0 } ^ { + \\infty } \\bigg [ 1 + \\big ( \\frac { 2 \\pi n } { \\log 1 0 ( c k + 1 ) } \\big ) ^ { 2 } \\bigg ] ^ { - 1 } \\bigg ) ^ { \\frac { 1 } { 2 } } .\n$$", + "text_format": "latex", + "bbox": [ + 346, + 440, + 651, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $\\left| a _ { n } \\right|$ gets smaller as $n \\in \\mathbb N$ increases. And, an interesting property of $\\left| a _ { n } \\right|$ is that it only depends on the shape factor $c$ , while is independent of the parameter $\\sigma$ . Thus, one set of the absolute values of Benford-Fourier coefficients $\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }$ correspond to one identical $c$ , i.e., one identical special distribution of GGD. In other words, we could use $\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }$ as features or representations to discriminate different special distributions of GGD. ", + "bbox": [ + 173, + 488, + 825, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, if it is often difficult to know or even estimate the shape factor $c$ , we cannot compute the value of $\\left| a _ { n } \\right|$ . But fortunately, recalling that $a _ { n }$ is the $n$ -th Fourier coefficient of $\\mathcal { P } _ { \\mathcal { Z } } ( z )$ evaluated at $2 \\pi n$ , we can derive an easy estimation. Specifically, assume that $\\mathbf { x } = \\{ x _ { 1 } , \\dots , x _ { M } \\}$ is a set of $M$ i.i.d. points sampled from GGD with the same shape factor $c$ . Then, the corresponding BenfordFourier coefficients can be estimated as follows (Pasquini et al. (2014)): ", + "bbox": [ + 173, + 565, + 825, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/d880ea67f8e77aae0d5ceec51b211a6c2889e4952e2f70f61e0eab97324a542e.jpg", + "text": "$$\n\\hat { a } _ { n } = \\frac { \\sum _ { m = 1 } ^ { M } e ^ { - j 2 \\pi n \\log _ { 1 0 } | x _ { m } | } } { M } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\left[ \\cos \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) - j \\sin \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 187, + 641, + 790, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The gap between $\\hat { a } _ { n }$ and $a _ { n }$ is analyzed in Theorem 1. It tells that $\\hat { a } _ { n }$ gets closer to $a _ { n }$ as $M$ increases. For clarity, we firstly introduce a few notations: $\\mathcal { T } = e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } | }$ is a random variable with $\\mathcal { X }$ obeying the generalized Gaussian distribution, and $\\hat { a } _ { n }$ is an observation of the random variable $\\begin{array} { r } { \\mathcal { Y } = \\bar { \\frac { 1 } { M } } \\bar { \\sum _ { m = 1 } ^ { M } } \\mathcal { T } _ { m } } \\end{array}$ . Due to the space limit, the proof will be presented in supplementary material. ", + "bbox": [ + 173, + 689, + 825, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 Assume that the estimation error $\\varepsilon _ { n } = { \\hat { a } } _ { n } - a _ { n }$ is an observation of the random variable $\\mathcal { E } = \\mathcal { V } - a _ { n }$ . $| \\mathcal { E } |$ follows the Rayleigh distribution (Papoulis & Pillai (2001)), of which the probability density function $( P D F )$ is formulated as ", + "bbox": [ + 173, + 775, + 825, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/068986864f1121d3f477d0f01fd36df05a07e4cfc6a74658b99c256b28b0ee54.jpg", + "text": "$$\n\\mathcal { P } _ { | \\mathcal { E } | } \\left( r \\right) = 2 M r e ^ { - M r ^ { 2 } } .\n$$", + "text_format": "latex", + "bbox": [ + 418, + 823, + 578, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "And the expectation and variance are respectively shown as follows: ", + "bbox": [ + 171, + 849, + 622, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/c85cd3d17b8682c0a24a8ee94c97ad201648c8c0e1bc95a65758264883d030f4.jpg", + "text": "$$\nE { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D ( | { \\mathcal { E } } | ) = { \\frac { 4 - \\pi } { 4 M } } ,\n$$", + "text_format": "latex", + "bbox": [ + 374, + 869, + 622, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which implies that the estimation error $\\varepsilon _ { n }$ gets closer to $\\boldsymbol { \\theta }$ as the number of samples M increases. ", + "bbox": [ + 171, + 909, + 808, + 925 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 ADVERSARIAL DETECTION VIA BENFORD-FOURIER COEFFICIENTS ", + "text_level": 1, + "bbox": [ + 171, + 102, + 764, + 118 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 TRAINING PROCEDURE OF ADVERSARIAL DETECTOR ", + "text_level": 1, + "bbox": [ + 173, + 132, + 584, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are three stages to train the proposed adversarial detector, including: 1) building a training set based on benign images; 2) extracting novel representations of the training set via Benford-Fourier coefficients; 3) training a SVM classifier as the adversarial detector. They will be explained in details sequentially. And, the overall training procedure is briefly summarized in Algorithm 1. ", + "bbox": [ + 174, + 159, + 821, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Build a training set. Firstly, we collect $N$ clean images $\\left\\{ \\mathbf { x } _ { 1 } , \\dotsc , \\mathbf { x } _ { N } \\right\\}$ , which can be correctly predicted by $f _ { \\theta }$ . Then, we adopt one adversarial attack method (e.g., C&W (Carlini & Wagner (2017a)) or BIM ( Kurakin et al. (2016))) to generate one adversarial image corresponding to each clean image. The crafted $N$ adversarial examples are denoted as $\\big \\{ \\hat { \\mathbf { x } } _ { 1 } , \\dotsc , \\hat { \\mathbf { x } } _ { N } \\big \\}$ . Besides, to avoid that the noisy image (polluted by some kind of non-malicious noises but still can be correctly predicted by $f _ { \\theta } )$ is incorrectly detected as adversarial, we also craft one noisy image by adding small random Gaussian noises onto each clean image. These $N$ noisy examples are denoted as $\\left\\{ bar { \\mathbf { x } } _ { 1 } , \\ldots , \\bar { \\mathbf { x } } _ { N } \\right\\}$ . Note that, hereafter benign examples include both clean and Gaussian noisy examples. Consequently, we obtain one training set with $3 N$ examples, denoted as $\\mathcal { D } _ { t r } = \\{ ( \\mathbf { x } _ { i } , - 1 \\bar { ) } , ( \\bar { \\mathbf { x } } _ { i } , \\bar { { ~ - } 1 } ) , ( \\hat { \\mathbf { x } } _ { i } , + 1 \\bar { ) } \\} _ { i = 1 , \\ldots , N }$ . ", + "bbox": [ + 173, + 219, + 825, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extract novel representations. We firstly feed the $i$ -th training image from $\\mathcal { D } _ { t r }$ into $f _ { \\theta }$ . We concatenate all response entries of the $l$ -th layer in $f _ { \\theta }$ to obtain one vector $\\mathbf { r } _ { i } ^ { l }$ . Then, we estimate the corresponding Benford-Fourier coefficients according to Eq. (5), as follows ", + "bbox": [ + 173, + 351, + 825, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/b0c2baba103458e9fc8c223b8296cf5db64662c047e8b26952024979114e50ab.jpg", + "text": "$$\n( \\hat { a } _ { n } ) _ { i } ^ { l } = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } e ^ { - j 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | } ,\n$$", + "text_format": "latex", + "bbox": [ + 375, + 396, + 620, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M _ { i } ^ { l }$ indicates the length of $\\mathbf { r } _ { i } ^ { l }$ . The magnitude of $( \\hat { a } _ { n } ) _ { i } ^ { l }$ is computed as follows ", + "bbox": [ + 174, + 445, + 725, + 462 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/d664b67c28464c0f5bfdd1180d1f04d25ee0d5d13fedc8d6f50866f80533c8cd.jpg", + "text": "$$\n| ( \\hat { a } _ { n } ) _ { i } ^ { l } | = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } \\Bigg ( \\cos ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) + \\sin ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) \\Bigg ) ^ { \\frac { 1 } { 2 } } .\n$$", + "text_format": "latex", + "bbox": [ + 243, + 465, + 754, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we extract one $T$ -dimensional feature vector $\\mathbf { a } _ { i } ^ { l } = [ | ( \\hat { a } _ { 1 } ) _ { i } ^ { l } | , \\dots , | ( \\hat { a } _ { T } ) _ { i } ^ { l } | ] \\in \\mathbb { R } _ { + } ^ { T }$ for the $i$ -th training image from the $l$ -th layer. We set $T = 1 6$ in experiments, as $| ( \\hat { a } _ { n } ) _ { i } ^ { l } |$ of larger $n$ is too small for discrimination. Finally, we concatenate the feature vectors of all layers to form a long vector $\\hat { \\mathbf { a } } _ { i } = [ \\hat { \\mathbf { a } } _ { i } ^ { 1 } ; \\hdots ; \\hat { \\mathbf { a } } _ { i } ^ { L } ] \\in \\mathbb { R } _ { + } ^ { T \\check { L } }$ , with $L$ being the number of layers in $f _ { \\theta }$ . Consequently, we obtain a novel representation of the training set $\\mathcal { D } _ { t r }$ , denoted as $\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\bf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , 3 N }$ , where the label of ${ \\bf a } _ { i }$ (i.e., $+ 1$ or $^ { - 1 }$ ) is directly obtained from $\\mathcal { D } _ { t r }$ . ", + "bbox": [ + 173, + 521, + 825, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Train an adversarial detector. Finally, we train a binary SVM classifier based on $\\hat { A } _ { t r }$ . The trained SVM classifier will serve as the adversarial detector for the CNN model $f _ { \\theta }$ . ", + "bbox": [ + 173, + 619, + 821, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing. One novel testing example is firstly predicted as adversarial or not by the trained adversarial detector . If adversarial, then it is rejected; otherwise, it is fed into $f _ { \\theta }$ to predict its class label. ", + "bbox": [ + 173, + 654, + 821, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark. Note that in the derivation of $a _ { n }$ (see Eq. (3)), the mean parameter of GGD is set to 0. In experiments, we calculate the mean values of internal-layer responses of all networks for every image, and find that mean parameters at most layers are close to 0, while the mean parameters at a few layers could be large. However, the mean value is subtracted from each response entry when we extract the MBF features in our experiments. Thus, the derived $a _ { n }$ is applicable to our task. Besides, according to Theorem 1, the estimation error of $( \\hat { a } _ { n } ) _ { i } ^ { l }$ is inversely proportional to $M _ { i } ^ { l }$ . It tells that the coefficient estimated from the larger-sized layer is more accurate. In many neural networks (e.g., AlexNet (Krizhevsky et al. (2012))), the response sizes of high layers get smaller, which means the less accurate estimation. However, we believe that the discrimination between adversarial and benign examples in higher layers is more evident than that in lower layers. There is a trade-off between estimation accuracy and discrimination. This is why we concatenate the estimated magnitudes of all layers together to construct the novel representation. ", + "bbox": [ + 174, + 689, + 825, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 EXPERIMENTAL SETTINGS ", + "text_level": 1, + "bbox": [ + 176, + 873, + 398, + 887 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Databases and network architectures We conduct experiments on three databases, including CIFAR-10 (Krizhevsky et al. (2014)), SVHN (Netzer et al. (2011)), and a subset of ImageNet (Deng ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Training the adversarial detector via the magnitude of Benford-Fourier coefficients. ", + "bbox": [ + 173, + 103, + 816, + 118 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Require: The trained CNN model $f _ { \\theta }$ with $L$ layers, and the training set $\\mathcal { D } _ { t r }$ ", + "bbox": [ + 174, + 121, + 629, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1: for $i = 1$ to $| \\mathcal { D } _ { t r } |$ do: \n2: for $l = 1$ to $L$ do: \n3: Compute $| ( \\hat { a } _ { n } ) _ { i } ^ { l } |$ as Eq. (8), with $n = 1 , \\dots , T$ ; \n4: Concatenate $\\{ | ( \\hat { a } _ { n } ) _ { i } ^ { l } | \\} _ { n = 1 , \\dots , T }$ to obtain a vector $\\hat { \\mathbf { a } } _ { i } ^ { l }$ ; \n5: end for \n6: Concatenate $\\{ \\hat { \\mathbf { a } } _ { i } ^ { l } \\} _ { l = 1 , \\dots , L }$ to obtain a long vector $\\hat { \\mathbf { a } } _ { i }$ ; \n7: end for \n8: Build a novel representation of the training set, denoted as $\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\mathbf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , | \\mathcal { D } _ { t r } | }$ ; \n9: Train a binary SVM classifier based on $\\hat { A } _ { t r }$ ; \nreturn The trained binary SVM classifier. ", + "bbox": [ + 179, + 132, + 733, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "et al. (2009)). In terms of CIFAR-10 and SVHN, we adopt the same settings as the compared method LID (Ma et al. (2018)). Specifically, a 33-layer network pre-trained on the training set of CIFAR-10 achieves $8 2 . 3 7 \\%$ accuracy on the testing set with 10, 000 benign images; a 19-layer network pretrained on the training set of SVHN achieves $9 2 . 6 \\%$ accuracy on the testing set with 26, 032 benign images. Then, we add a small noise drawn from ${ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } )$ on each testing image, with $\\sigma$ being the similar level of the $\\ell _ { 2 }$ norm of adversarial perturbations on the same database. If both the benign and its noisy image can be correctly predicted by the classification network, then it is picked out for training the detection. We finally collect 8, 175 and 23, 862 benign images from CIFATR-10 and SVHN, respectively. These images are randomly partitioned to the $8 0 \\%$ training set and the $2 0 \\%$ testing set, used for the training and testing of the detector. We also collect a subset from ImageNet, including 800 benign images of 8 classes (snowbird, spoonbill, bobtail, Leonberg, hamster, proboscis monkey, cypripedium calceolus, and earthstar). The 100 images of each class contain 50 testing images and 50 randomly selected training images. We fine-tune the checkpoints of both AlexNet and VGG pre-trained on ImageNet1 on these 800 images to achieve $1 0 0 \\%$ accuracy. Then, 785 benign images are kept for detection, as both their noisy images and themselves can be correctly predicted by both the fine-tuned AlexNet and VGG models. These 785 images are then randomly partitioned to 400 training and 305 testing images used for detection. For each database, as described in Section 4.1, one noisy and one adversarial image are generated for each benign image; then, all of benign, noisy, and adversarial images are used for detection. ", + "bbox": [ + 173, + 299, + 825, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack methods We adopt four popular adversarial attack methods to craft adversarial examples, including basic iterative method (BIM (Kurakin et al. (2016))), CarliniWagnerL2Attack (CW-L2 (Carlini & Wagner (2017a))), DeepFool ( (Moosavi-Dezfooli et al. (2016))), and random projected gradient descent (R-PGD (Madry et al. (2017))). They are implemented by Foolbox2. ", + "bbox": [ + 174, + 569, + 825, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared detection methods We compare with three state-of-the-art and open-sourced adversarial detection methods, including $\\mathrm { K D + B U } ^ { 3 }$ (Feinman et al. (2017)), Mahalanobis distance4 (M-D) (Lee et al. (2018)), and $\\mathrm { L I D } ^ { 5 }$ (Ma et al. (2018)). Note that another recent work called I-defender (Zheng & Hong (2018)) is not compared, as its code is not available. To ensure the fair comparison, the SVM classifier is trained with all compared methods, implemented by the fitcsvm6 function in MATLAB. There are two important hyper-parameters in LID, i.e., the size of mini-batch and the number of neighbors. On CIFAR-10 and SVHN, they are respectively set as 100 and 20, as suggested in Ma et al. (2018); on ImageNet, as there are only 400 benign training images, they are respectively set as 50 and 20 in experiments. Moreover, we find that there are some unfair settings in the implementations of compared methods. For example, $\\mathrm { K D + B U }$ utilizes the extra 50, 000 images of CIFAR-10 to compute the kernel density of each training and testing image; M-D also uses these extra images to compute the mean and co-variance of GMM. Since extra images of the similar distribution with the training images are often unavailable, we believe that extra images should not be used to ensure the fair comparison. Thus, extra images are not used for $\\mathrm { K D + B U }$ and M-D in our experiments. Besides, ", + "bbox": [ + 173, + 630, + 825, + 824 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d024e527f7cda4b02a67ead484dc23c7889e4450d6e1bd0d6490c6229dde8930.jpg", + "table_caption": [ + "Table 1: Detection results in the non-transfer case. " + ], + "table_footnote": [], + "table_body": "
databaseDetectorAUROC(%)Accuracy (%)
BIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
CIFAR-10KD+BU (Feinman et al. (2017))79.082.880.677.774.273.071.772.6
M-D (Lee et al. (2018))51.748.453.752.066.766.766.766.7
LID (Ma et al. (2018))87.987.486.682.772.678.578.070.2
MBF99.699.696.999.498.898.391.898.5
SVHNKD+BU (Feinman et al. (2017)) M-D (Lee et al. (2018))81.585.184.182.578.680.079.378.6
LID (Ma et al. (2018))49.750.049.650.166.766.766.766.7
91.788.792.290.583.380.884.081.5
ImageNet- AlexNetMBF KD+BU (Feinman et al. (2017))99.799.999.399.598.898.391.898.5
M-D (Lee et al. (2018))50.852.6 60.151.151.534.436.834.935.3
LID (Ma et al. (2018))48.1 71.770.946.5 71.954.1 72.366.7 68.766.6 60.566.866.7
ImageNet-MBF99.999.699.999.898.697.865.8 98.868.3
KD+BU (Feinman et al. (2017))57.157.756.258.642.843.641.798.4
M-D (Lee et al. (2018))63.764.845.966.967.165.666.744.8
89.567.4
VGG16LID (Ma et al. (2018)) MBF82.6 99.884.2 100.0100.084.2 100.077.8 99.676.7 99.683.3 100.076.4 100.0
", + "bbox": [ + 196, + 126, + 799, + 284 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb083cac0a3e20d6cea6f26727565045ea368fd4ac7db491d45a921712eb9bee.jpg", + "table_caption": [ + "Table 2: Detection results evaluated by AUROC $( \\% )$ in the attack-transfer case. " + ], + "table_footnote": [], + "table_body": "
databaseTest attack→BIMCW-L2DeepFoolR-PGD
TrainattackKD+BU (Feinman et al. (2017))/M-D(Lee et al.(2018))/LID (Ma et al.(2018)/MBF
CIFAR-10BIM CW-L2 DeepFool R-PGD79.0/51.7787.9799.6 83.4/47.9/90.3/99.6 83.6/53.5 /89.7 /99.3 78.1/51.5/86.1/99.678.1/52.2/85.7799.4 82.8/48.4/87.4/99.6 83.1/51.4/86.1/99.1 77.3 /51.9 /84.5 /99.576.4/52.4/86.4/87.8 80.3 /46.5 /87.0 / 88.6 80.6/53.7/86.6/96.978.6/50.9/87.6/99.4 82.9/48.2/88.9/99.5 83.1/53.6/87.9/98.9 77.7/52.1/82.7/99.4
SVHNBIM CW-L2 DeepFool R-PGD81.5749.7791.7799.7 83.6/50.8/91.6/99.7 85.0/50.0/91.6/99.7 82.5 /49.9 /91.2 /99.683.1/51.0/88.3/99.7 85.1/50.0/88.7/99.9 86.3/ 50.0 /87.9 /99.8 84.0/49.5/87.8/99.675.8/52.4/84.9/88.3 80.1/48.8/91.9/97.5 82.5/49.7/92.0/97.3 84.1/49.6/92.2/99.3 81.2/50.2/91.6/97.481.5750.4/90.7799.5 83.6/49.1/90.8/99.7 85.0/49.4/90.8 /99.6 82.5/50.1/90.5 /99.5
ImageNet- AlexNetBIM CW-L2 DeepFool R-PGD50.8/48.1/71.7/99.9 52.3/50.7/69.7/99.8 51.8/53.6/71.3/99.9 51.6/52.4/72.5 /99.850.8/44.9/70.8/99.6 52.6/60.1/70.9/99.7 51.9/53.5/69.7/99.7 51.3/ 50.4/72.3/99.650.4/49.7772.1/99.8 51.2/52.1/66.0/99.7 51.2/46.5 /71.9 /99.9 50.3 /50.5 /72.2/99.750.7/50.4/70.8/99.8 52.2/55.8/69.8/99.8 51.8/51.2/71.2/99.8
ImageNet- VGG-16BIM CW-L2 DeepFool R-PGD57.1/63.7/82.6/99.8 57.3/64.3/81.9/100.0 58.7/45.5/ 80.2/99.7 58.6/62.3/81.1/100.057.1/57.2/84.3/99.5 57.7/64.8/84.2/100.0 59.4/46.1/79.1/99.3 58.9/60.8/83.7/100.053.8/54.4/90.4/100.0 53.8/56.6/90.2/100.0 56.2/45.9/89.5/100.0 55.3/54.4/90.8/100.051.5 /54.1/72.3 /99.8 57.1/65.4/84.2/100.0 57.5/67.6/84.3/100.0 59.0 /44.0/ 80.9 /99.7 58.6/66.9/84.2/100.0
", + "bbox": [ + 194, + 309, + 803, + 460 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88ec8aa6b35da30e3bb67fda7c18fc761e062fbf4a7d36de115eb99f540d4bd5.jpg", + "table_caption": [ + "Table 3: Detection results evaluated by accuracy $( \\% )$ in the attack-transfer case. " + ], + "table_footnote": [], + "table_body": "
databaseTest attack→BIM CW-L2 DeepFool R-PGDKD+BU(Feinman et al. (2017))/M-D(Lee et al. (2018))/LID (Ma et al. (2018)/MBF
CIFAR-10BIMCW-L2DeepFoolR-PGD74.2/66.7/72.6/98.874.2/66.7/80.3 /98.174.0/66.7 /80.4/95.273.1/66.7/77.7 /98.666.773.0/66.7 /78.5 /98.3 71.7 /66.7/77.9 /89.573.8/66.7/61.3/96.871.7 /66.7/77.9 /89.5 73.8 /66.7 /73.6 /98.2
66.771.9 / 66.7 /75.1/98.5 71.0 /66.7/75.4/89.773.4/66.7/74.8/95.471.0 /66.7/75.4/89.7 72.6/66.7/70.2/98.5
SVHNBIMCW-L2DeepFoolR-PGD78.6/66.77/83.3/98.7 80.5/66.7779.4/98.9 77.7/66.7/83.5/95.8 78.6/66.7/82.9/98.278.2/66.7/83.5 /98.0 80.0/66.7 /80.8/99.2 77.1/66.7/84.0 /95.9 78.1/66.7/80.7/97.379.8/66.7/83.9/98.0 81.8/66.7/79.9/98.4 79.3/66.7/84.0/97.3 79.9/66.7/80.7/98.078.6/66.7 /83.6 /98.6 80.5 / 66.7 /79.9/98.9 77.7 /66.7 /83.8/96.5 78.6/ 66.7 /81.5 /98.8
ImageNet-AlexNetBIMCW-L2DeepFoolR-PGD34.4/66.7/68.7798.6 34.4/66.6/68.6797.336.5/66.2/64.6/98.0 36.8/66.6/60.6/97.8 35.0 /66.1/60.7/98.134.3/66.7/65.0/98.435.0 /66.1/60.7/98.1 36.4/66.1/62.6/98.034.9/66.8/65.8/98.8 35.9 /66.6/68.3 /98.634.1/66.7/67.0/98.3 35.3/66.7/68.3/98.4
36.0 /66.7/69.4/98.835.4/66.7/67.5/98.566.835.2/66.7/67.7/97.6 34.1/66.7/67.0/98.3
ImageNet-VGG-16BIMCW-L2DeepFoolR-PGD42.8/67.1777.8/98.643.1/64.8/75.2/98.045.4/66.7 /76.2/98.844.8 /66.3/76.3/98.566.643.6/65.6/76.7 /97.8 38.4/ 62.5/82.5 /98.142.7/67.9/76.9/98.438.4/ 62.5/82.5 /98.1 43.3/67.2/73.9/98.041.8/66.7 /83.3/98.8 45.8 /66.7/74.4/98.640.6/64.9/82.4/98.3 45.8/67.4/76.4/98.4
66.745.4/66.6/74.1/97.6 40.6/64.9/82.4/98.3
", + "bbox": [ + 194, + 482, + 802, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LID utilizes other benign testing images as neighborhoods to extract features for each testing image. It is unfair to utilize the information of benign or adversarial for neighboring testing images. In our experiments, we use benign training images as neighborhoods for each testing image. ", + "bbox": [ + 174, + 648, + 823, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Three comparison cases and evaluation metrics We conduct experiments of three cases, including: 1) non-transfer, both training and testing adversarial examples are crafted by the same attack method; 2) attack-transfer, both training and testing adversarial examples are crafted by different attack methods; 3) data-transfer, both training and testing adversarial examples are crafted by the same attack method, but the data sources of training and testing benign examples are different. Two widely used metrics are used to evaluate the detection performance, including area under the receiver operating characteristics (AUROC), and detection accuracy, which is the diagonal summation of the confusion matrix, using 0.5 as the threshold of the posterior probability. Higher values of both metrics indicate better performance. ", + "bbox": [ + 173, + 695, + 825, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 832, + 277, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Detection results in the non-transfer case are shown in Table 1. The proposed MBF method shows the best performance in all cases, and is much superior to all compared methods. LID performs the second-best in most case. $\\mathrm { K D + B U }$ gives somewhat good detection performance on CIFAR-10 and SVHN, but performs very poor on ImageNet. It tells that $\\mathrm { K D + B U }$ is very sensitive to different databases and networks. M-D gives almost the degenerate results. Note that the results of M-D on ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69efaee09e76d4a1b56e04fd81944dc90fd9797a28a35b7f953f4304f89b2339.jpg", + "table_caption": [ + "Table 4: Detection results evaluated by AUROC score $( \\% )$ in the data-transfer case. All detectors are trained on the train set, and tested on the out-of-sample set (including 365 images) of the ImageNet database. The best results are highlighted in bold. " + ], + "table_footnote": [], + "table_body": "
ImageNet-AlexNetImageNet-VGG-16
DetectorBIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
KD+BU66.369.464.768.884.085.079.985.2
M-D48.749.250.051.054.550.050.456.1
LID69.167.672.971.484.288.489.283.7
MBF99.199.099.699.499.399.599.399.5
", + "bbox": [ + 250, + 156, + 746, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b975bebdd9d344b5960d73838588cd6994286a09202d3cfa99fe01e1e30d4edd.jpg", + "table_caption": [ + "Table 5: Attack Failure Rates $( \\% )$ of adaptive attacks with different detectors. See context for details. " + ], + "table_footnote": [], + "table_body": "
Original CW-L2LIDMBF
CIFAR-100.088.4100.0
SVHN0.085.692.9
", + "bbox": [ + 369, + 250, + 629, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CIFAR-10 and SVHN reported in (Lee et al. (2018)) are very high, but their networks (i.e., ResNet (He et al. (2016)) and DenseNet (Huang et al. (2017)) are different from that used in our experiments. It implies that M-D may be suitable for very deep neural networks, but not for shallower networks. ", + "bbox": [ + 174, + 292, + 825, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection results in the attack-transfer case evaluated by AUROC and accuracy are presented in Tables 2 and 3, respectively. MBF still shows much better performance in all transfer cases than all compared methods, and the changes among detecting different attacks are very small. It verifies the robustness of MBF to different attacks. ", + "bbox": [ + 174, + 342, + 825, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct a data-transfer experiment on ImageNet. Specifically, we collect extra 365 images of 8 classes (same with the classes used for the detection training, see Section 4.2) through searching the class names in Baidu and Facebook. These 365 benign and their noisy images can be correctly predicted by both the fine-tuned AlexNet and VGG models. The results on these images are shown in Table 4. MBF still shows the best performance. And, compared to the corresponding results in Table 1, the AUROC/accuracy scores of MBF on detecting different attacks change very gently. It verifies the robustness of MBF to different data sources. ", + "bbox": [ + 174, + 405, + 825, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 ADAPTIVE ATTACK AGAINST ADVERSARIAL DETECTION ", + "text_level": 1, + "bbox": [ + 173, + 518, + 611, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to Carlini & Wagner (2017b) and Ma et al. (2018), we also present an adaptive white-box attack, of which the goal is not only to fool the classifier, but also to evade the adversarial detector. It measures how easy to bypass the detector. Specifically, by combining the CW-L2 attack (Carlini & Wagner (2017a)) and the proposed MBF detector, the adaptive attack is formulated as follows: ", + "bbox": [ + 173, + 544, + 825, + 601 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/ef0772f1a645b7f33bb82d6f5e88667e7a460e99dac440dd60000a4074a74b04.jpg", + "text": "$$\n\\operatorname * { a r g m i n } _ { \\mathbf { x } _ { a d v } \\in [ 0 , 1 ] } \\| \\mathbf { x } - \\mathbf { x } _ { a d v } \\| _ { 2 } ^ { 2 } + \\alpha \\cdot \\big [ - l ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) , y ) + \\| D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) - D ( f _ { \\theta } ( \\mathbf { x } ) ) \\| _ { 1 } \\big ] ,\n$$", + "text_format": "latex", + "bbox": [ + 223, + 603, + 769, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf { x }$ denotes the clean input, $y$ indicates the ground-truth label. $l ( \\mathbf { x } _ { a d v } , y )$ is the Hinge loss. $D ( f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } ) )$ represents the extracted features by the the detector $D$ (specified later), from $f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } )$ . The minimization of the $\\ell _ { 1 }$ term $| | D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) ^ { * } - D ( f _ { \\theta } ( \\mathbf { x } ) ) | | _ { 1 }$ encourages the detection features of adversarial and benign examples to be close, such that the detector could be evaded. ", + "bbox": [ + 174, + 637, + 825, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct the above adaptive attack on both CIFAR-10 and SVHN. Specifically, 1000 clean examples are randomly selected from the test set of each database. For each generated adversarial example, if it fails to fool the classifier $f _ { \\theta }$ , we record 1. The attack failure rate over all 1000 adversarial examples is computed as the metric. The higher rate indicates that it is more difficult to evade the detector. Since LID performs much better than $\\mathrm { K D + B U }$ and M-D in above experiments, here we only compare LID and the proposed MBF detector. For clarity, the detection features are only extracted from the soft-max layer for both LID and MBF. The original CW-L2 attack (i.e., without the $\\ell _ { 1 }$ term) is also presented as the baseline. With the same setting in (Ma et al. (2018)), the trade-off parameter $\\alpha$ is determined by binary search within the range $[ \\bar { 1 0 } ^ { - 3 } , 1 0 ^ { 6 } ]$ . The results are shown in Table 5. The attack failure rates of MBF on both databases are much higher than those of LID. It tells that it is more difficult to evade MBF than LID. ", + "bbox": [ + 173, + 699, + 825, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5 HYPOTHESIS TEST VIA KOLMOGOROV-SMIRNOV TEST ", + "text_level": 1, + "bbox": [ + 174, + 869, + 593, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)) is a non-parametric test method in statistics, to test whether a sample follows a reference probability distribution (one-sample KS test), or whether two samples follow the same distribution (two-sample KS test). Specifically, in one-sample KS test, the distance between the empirical distribution function of one sample and the cumulative distribution function of the reference probability distribution is measured. Then, the $p$ -value corresponding to the obtained distance is computed. If the $p$ -value is larger than the significance level $\\alpha$ (here we set $\\alpha = 0 . 0 5$ ), then the null hypothesis that the sample follows the reference distribution is accepted; otherwise, rejected. Similarly, in two-sample KS test, the distance between the empirical distribution functions of two samples is computed. If the corresponding $p$ -value is larger than $\\alpha$ , then it accepts that two samples follow the same distribution. The KS test conducted below is implemented by the python function scipy.stats.ks 2samp7. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f128add81b448d21c04af4561663c3d499b3b4be016b057dcf77c25e5527a83f.jpg", + "table_caption": [ + "Table 6: The $p$ -value of KS hypothesis test among clean samples, noisy samples, and adversarial samples crafted by four methods, respectively, on ImageNet-AlexNet. " + ], + "table_footnote": [], + "table_body": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
", + "bbox": [ + 254, + 142, + 741, + 184 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/adf2d35488b42e10bd5631f045e040f1271834d43043993f162f1e8389239dc3.jpg", + "table_caption": [ + "Table 7: The $p$ -value of two-sample KS test among MBF coefficients of different types of examples. " + ], + "table_footnote": [], + "table_body": "
ImageNet-VGG-16(clean, noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
", + "bbox": [ + 212, + 210, + 782, + 251 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/16d5e9200efba8879ec80096b9d49580ef3e3677a6ac01a4aa00dc1b7da652c6.jpg", + "table_caption": [ + "Table 8: The $p$ -value of KS hypothesis test among MBF coefficients of different data sources. " + ], + "table_footnote": [], + "table_body": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
(train set,test set)0.8390.8580.3680.8420.2460.221
(train set, out-of-sample set)0.0490.0540.0940.4070.7370.088
", + "bbox": [ + 225, + 277, + 766, + 318 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 327, + 825, + 453 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hypothesis test 1. Here we verify that whether the posterior vectors of both adversarial and benign examples follow GGD. We denote the posterior vector of one adversarial example as $\\bf p \\mathrm { _ { a d v } }$ , and that of one benign example as $\\mathbf { p } _ { \\mathrm { { b e n } } }$ . The distribution of GGD is denoted as ${ \\mathcal { P } } _ { \\mathrm { G G D } }$ . Then, we conduct two one-sample KS tests, including: ", + "bbox": [ + 176, + 459, + 823, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "• H1.1 The test of adversarial examples: $H _ { 0 }$ : $\\mathbf { p } _ { \\mathrm { a d v } } \\sim \\mathcal { P } _ { \\mathrm { G G D - a d v } }$ ; $H _ { 1 }$ : padv 6∼ PGGD-adv. \n• H1.2 The test of benign examples: $H _ { 0 }$ : $\\mathbf { p } _ { \\mathrm { b e n } } \\sim \\mathcal { P } _ { \\mathrm { G G D - b e n } }$ ; $H _ { 1 }$ : pben 6∼ PGGD-ben. ", + "bbox": [ + 215, + 525, + 779, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In H1.1, the reference distribution $\\mathcal { P } _ { \\mathrm { G G D - a d v } }$ is firstly estimated from $\\mathbf { p } _ { \\mathrm { a d v } }$ , using the estimated method proposed in (Lasmar et al. (2009)). To alleviate the uncertainty of the estimation, we draw 500 samples from the estimated $\\mathcal { P } _ { \\mathrm { G G D - a d v } }$ . Then, we conduct the two-sample KS test between $\\mathbf { p } _ { \\mathrm { a d v } }$ and these 500 samples respectively. The average $p$ -value over 1000 tests is recorded. The mean of the average $p$ -values over the whole database is reported. H1.2 is conducted similarly. The results tested on ImageNet-AlexNet are shown in Table 6. Due to the space limit, results on other databases and models will be presented in the supplementary material. In all cases, the $p$ -values are larger than the significance level 0.05. Hence, we can conclude that the posterior vectors of both adversarial and benign examples follow GGD. However, note that the parameters of their corresponding GGD are different, which will be verified in the following test. ", + "bbox": [ + 173, + 569, + 825, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hypothesis test 2. Here we verify that whether the extracted MBF features of adversarial and benign examples follow the same empirical distribution. We denote the MBF feature vector of one adversarial example as $\\mathbf { m _ { \\mathrm { a d v } } }$ , and the corresponding empirical distribution is denoted as $\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } }$ . Similarly, we define $\\mathbf { m } _ { \\mathrm { b e n } }$ and $\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }$ for benign examples. Then, we conduct the following four two-sample KS tests: ", + "bbox": [ + 174, + 715, + 825, + 790 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "• H2.1 The test between adversarial and benign examples: $H _ { 0 }$ : $\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }$ ; $H _ { 1 }$ : $\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }$ . \n• H2.2 The test between adversarial examples crafted from different attack methods: $H _ { 0 }$ : $\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }$ ; $H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }$ . \nH2.3 The test between adversarial examples from different data sources (i.e., train, test and \nout-of-sample set): $H _ { 0 }$ : $\\begin{array} { r } { \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } } } \\end{array}$ ; $H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } }$ . ", + "bbox": [ + 215, + 800, + 825, + 904 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e6e248dabc29baa792b249691b317cc1ab671db8b1d5f02fc3e9a7bc0e6db03a.jpg", + "image_caption": [ + "Figure 1: Statistics (mean $\\pm$ standard deviation) of MBF coefficients on train (top row), test (median row), and out-of-sample (bottom row) set of ImageNet-VGG-16. " + ], + "image_footnote": [], + "bbox": [ + 222, + 104, + 772, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "• H2.4 The test between benign examples from different data sources (i.e., train, test and out-of-sample set): $H _ { 0 }$ : $\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }$ ; $H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }$ . ", + "bbox": [ + 217, + 313, + 821, + 347 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In above four two-sample KS tests, we test on the 16-dimensional MBF features extracted from the soft-max layer. Since the implementation scipy.stats.ks 2samp cannot compare two vectors, we compare the feature of each dimension separately, then report the average $p$ -value over all dimensions. Specifically, when comparing two sets of samples, we firstly concatenate the feature of each dimension across all samples in the same set, leading to 16 long vectors for each set. Then, each pair of two long vectors corresponding to the same dimension from two sets is compared by KS test. The average $p$ -value over all 16 dimensions is reported. The $p$ -values of ${ \\bf H } 2 . 1$ are shown in Table 7 (see the column “(clean, BIM)”). The $p$ -values on both train and test set are 0. Thus, the hypothesis $H _ { 0 }$ is rejected, i.e., the MBF features of adversarial and benign examples follow different distributions. The $p$ -values of ${ \\bf H } 2 . 2$ are shown in Table 7. We pick two groups of attack methods, i.e., (BIM, DeepFool) and (CW-L2, R-PGD). The $p$ -values of (BIM, DeepFool) are 0, while the $p$ -values of (CW-L2, R-PGD) are larger than 0.05. It demonstrates that the distributions of adversarial examples crafted from different attack methods are possible to be different. The $p$ -values of ${ \\bf H } 2 . 3$ are shown in Table 8. The $p$ -values of all types of adversarial examples exceed 0.05. Thus, the MBF features of adversarial examples from different data sources follow the same distribution. The $p$ -values of H2.4 are shown in Table 8. Only the $p$ -value of “(train set, out-of-sample set)” of clean examples is slightly lower than 0.05, while the values of other cases exceed 0.05. Thus, in most cases, the MBF features of benign examples from different data sources follow the same distribution. ", + "bbox": [ + 174, + 356, + 825, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "From above analysis, we obtain the following conclusions: 1) The extracted MBF features of adversarial and benign examples follow different empirical distributions. It explains why MBF features are effective for detecting adversarial and benign examples; 2) The extracted MBF features of adversarial/benign examples from the train, test and out-of-sample sets follow the same empirical distribution. Although the extracted MBF features of adversarial examples crafted from different attack methods may not follow the same empirical distribution, the significant difference between benign and different adversarial distributions can still lead to the good detection performance in the attack-transfer case. It explains why MBF features are robust across different attack methods and different data sources. Moreover, we visualize the statistics of each dimension of MBF features, i.e., mean and standard deviation, as shown in Fig. 1. These visualizations also support above conclusions. Due to space limit, more KS tests and visualizations on different databases and networks will be presented in the supplementary material. ", + "bbox": [ + 174, + 612, + 825, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 787, + 318, + 804 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This work has proposed a novel adversarial detection method, dubbed MBF. The assumption behind is that the internal responses of the classification network of both adversarial and benign examples follow the generalized Gaussian distribution (GGD), but with different shape factors. The magnitude of Benford-Fourier coefficient is a function w.r.t. the shape factor, and can be easily estimated based on responses. Thus, it can serve as the discriminative features between adversarial and benign examples. The extensive experiments conducted on several databases, as well as the empirical analysis via KS test, demonstrate the superior effectiveness and robustness to different attacks and different data sources of the proposed MBF method, over state-of-the-art detection methods. ", + "bbox": [ + 174, + 811, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 285, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pp. 39–57. IEEE, 2017a. ", + "bbox": [ + 169, + 126, + 823, + 155 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pp. 3–14. ACM, 2017b. ", + "bbox": [ + 173, + 162, + 823, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee, 2009. ", + "bbox": [ + 173, + 213, + 823, + 256 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. ", + "bbox": [ + 171, + 263, + 825, + 294 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. ", + "bbox": [ + 171, + 301, + 825, + 330 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nathaniel R Goodman. Statistical analysis based on a certain multivariate complex gaussian distribution (an introduction). The Annals of mathematical statistics, 34(1):152–177, 1963. ", + "bbox": [ + 173, + 338, + 823, + 368 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and Patrick McDaniel. On the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017. ", + "bbox": [ + 173, + 376, + 823, + 405 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016. ", + "bbox": [ + 173, + 412, + 823, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dan Hendrycks and Kevin Gimpel. Early methods for detecting adversarial images. arXiv preprint arXiv:1608.00530, 2016. ", + "bbox": [ + 169, + 463, + 825, + 493 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4700–4708, 2017. ", + "bbox": [ + 176, + 501, + 823, + 544 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ", + "bbox": [ + 176, + 551, + 821, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. The cifar-10 dataset. online: http://www.cs.toronto.edu/kriz/cifar.html, 55, 2014. ", + "bbox": [ + 168, + 603, + 825, + 632 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533, 2016. ", + "bbox": [ + 171, + 640, + 823, + 670 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nour-Eddine Lasmar, Youssef Stitou, and Yannick Berthoumieu. Multiscale skewed heavy tailed model for texture analysis. In 2009 16th IEEE International Conference on Image Processing (ICIP), pp. 2281–2284. IEEE, 2009. ", + "bbox": [ + 174, + 676, + 823, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems, pp. 7167–7177, 2018. ", + "bbox": [ + 174, + 728, + 823, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764– 5772, 2017. ", + "bbox": [ + 173, + 779, + 823, + 821 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jiajun Lu, Theerasit Issaranon, and David Forsyth. Safetynet: Detecting and rejecting adversarial examples robustly. In Proceedings of the IEEE International Conference on Computer Vision, pp. 446–454, 2017. ", + "bbox": [ + 173, + 829, + 823, + 872 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck, Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using local intrinsic dimensionality. In International Conference on Learning Representations, 2018. ", + "bbox": [ + 176, + 881, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Frank J Massey Jr. The kolmogorov-smirnov test for goodness of fit. Journal of the American statistical Association, 46(253):68–78, 1951. ", + "bbox": [ + 173, + 155, + 823, + 184 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jan Hendrik Metzen, Tim Genewein, Volker Fischer, and Bastian Bischoff. On detecting adversarial perturbations. In International Conference on Learning Representations, 2017. ", + "bbox": [ + 173, + 193, + 823, + 222 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2574–2582, 2016. ", + "bbox": [ + 178, + 229, + 823, + 273 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. ", + "bbox": [ + 176, + 281, + 821, + 310 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tianyu Pang, Chao Du, Yinpeng Dong, and Jun Zhu. Towards robust detection of adversarial examples. In Advances in Neural Information Processing Systems, pp. 4579–4589, 2018. ", + "bbox": [ + 176, + 319, + 821, + 349 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy, pp. 372–387. IEEE, 2016. ", + "bbox": [ + 173, + 357, + 825, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A. Papoulis and S.U. Pillai. Probability, Random Variables, and Stochastic Processes. McGrawHill Series in Electrical and Computer Engineering. McGraw-Hill, 2001. ISBN 9780072817256. URL https://books.google.com.hk/books?id $=$ Mal4OgAACAAJ. ", + "bbox": [ + 174, + 409, + 825, + 452 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cecilia Pasquini, Fernando Perez-Gonz ´ alez, and Giulia Boato. A benford-fourier jpeg compression ´ detector. In 2014 IEEE International Conference on Image Processing, pp. 5322–5326. IEEE, 2014. ", + "bbox": [ + 173, + 460, + 825, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–572, 1901. ", + "bbox": [ + 174, + 512, + 821, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fernando Perez-Gonz ´ alez, Greg L Heileman, and Chaouki T Abdallah. Benford’s lawin image ´ processing. In 2007 IEEE International Conference on Image Processing, volume 1, pp. I–405. IEEE, 2007. ", + "bbox": [ + 171, + 549, + 823, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Murray Rosenblatt. A central limit theorem and a strong mixing condition. Proceedings of the National Academy of Sciences of the United States of America, 42(1):43, 1956. ", + "bbox": [ + 169, + 601, + 825, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pouya Samangouei, Maya Kabkab, and Rama Chellappa. Defense-gan: Protecting classifiers against adversarial attacks using generative models. arXiv preprint arXiv:1805.06605, 2018. ", + "bbox": [ + 171, + 638, + 823, + 669 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 2013. ", + "bbox": [ + 171, + 676, + 823, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mahesh K Varanasi and Behnaam Aazhang. Parametric generalized gaussian density estimation. The Journal of the Acoustical Society of America, 86(4):1404–1415, 1989. ", + "bbox": [ + 171, + 714, + 821, + 744 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Zhihao Zheng and Pengyu Hong. Robust detection of adversarial attacks by modeling the intrinsic properties of deep neural networks. In Advances in Neural Information Processing Systems, pp. 7913–7922, 2018. ", + "bbox": [ + 173, + 752, + 823, + 795 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 299, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1 PROOF OF THEOREM 1 ", + "bbox": [ + 176, + 133, + 375, + 148 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof 1 Assume that all the variables $\\{ \\mathcal { T } _ { m } \\} _ { m = 1 , \\dots , M }$ are independent and identically distributed. Applying the central limit theorem Rosenblatt (1956) to the real and imaginary parts of $_ { \\mathcal { V } }$ , we can obtain that both parts asymptotically follow the Gaussian distribution ", + "bbox": [ + 173, + 159, + 825, + 204 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/0270ee8c2464771f30bcfc969534922fd98e9655c21d659640419a9685d918e8.jpg", + "text": "$$\n\\mathcal { V } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { T } _ { m } \\sim \\mathcal { N } \\big ( E ( \\mathcal { T } _ { 1 } ) , \\frac { 1 } { M } D ( \\mathcal { T } _ { 1 } ) \\big ) ,\n$$", + "text_format": "latex", + "bbox": [ + 356, + 212, + 640, + 253 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 173, + 268, + 215, + 282 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/9029be1948795385fe01c66baae9a81fcd9762a05d42f2e2145cbbb6f6e9dc44.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle E ( \\mathcal { T } _ { 1 } ) = E \\left( e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right) = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { X } _ { 1 } } ( x ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } | x | } d x = a _ { n } , } \\\\ { \\displaystyle D ( \\mathcal { T } _ { 1 } ) = E \\big ( \\left| \\mathcal { T } _ { 1 } \\right| ^ { 2 } \\big ) - \\left| E \\left( \\mathcal { T } _ { 1 } \\right) \\right| ^ { 2 } = E \\big ( \\left| e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right| ^ { 2 } \\big ) - \\left| a _ { n } \\right| ^ { 2 } = 1 - \\left| a _ { n } \\right| ^ { 2 } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 232, + 287, + 766, + 356 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The PDF of $\\mathcal { V }$ can be rewritten as follows: ", + "bbox": [ + 176, + 367, + 454, + 382 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/6ae7c6afcab816a8de62e4a29f5e9edd698fde894a0135aed7697535dd698b4e.jpg", + "text": "$$\n\\mathcal { V } \\sim \\mathcal { N } ( a _ { n } , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) .\n$$", + "text_format": "latex", + "bbox": [ + 418, + 390, + 580, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Thus, we obtain ", + "bbox": [ + 174, + 428, + 281, + 441 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/1ee7357ccb4d65fbe42ccc1b5b87617528b90b7da44dfb9d109cac24144be28f.jpg", + "text": "$$\n\\mathcal { E } \\sim \\mathcal { N } ( 0 , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) .\n$$", + "text_format": "latex", + "bbox": [ + 423, + 448, + 573, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Besides, the pseudo variance Goodman (1963) of $\\mathcal { T } _ { 1 }$ is ", + "bbox": [ + 174, + 493, + 534, + 508 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/eed2d9e58d94f0dd4558d74684a3fda7e7d3f02d41e8a792433ae66aacdad552.jpg", + "text": "$$\n\\mathcal { T } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = E ( \\mathcal { T } _ { 1 } ^ { 2 } ) - E ( \\mathcal { T } _ { 1 } ) ^ { 2 } = E \\big ( e ^ { - j 2 \\pi n \\log _ { 1 0 } \\left| \\mathcal { X } _ { 1 } ^ { 2 } \\right| } \\big ) - a _ { n } ^ { 2 } = a _ { 2 n } - a _ { n } ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 267, + 516, + 730, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Correspondingly, ", + "bbox": [ + 173, + 544, + 289, + 559 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/c3744e73aa7ab5407eb57adfce801a8d059323fd7875630eb1763d2f2f8f11be.jpg", + "text": "$$\n\\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = \\frac { 1 } { M } \\mathcal { I } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = \\frac { a _ { 2 n } - a _ { n } ^ { 2 } } { M } .\n$$", + "text_format": "latex", + "bbox": [ + 392, + 565, + 606, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since $\\left( a _ { 2 n } - a _ { n } ^ { 2 } \\right)$ is bounded, we have $\\begin{array} { r } { \\operatorname* { l i m } _ { M \\to \\infty } \\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = 0 } \\end{array}$ . Thus, we obtain that the random variable $\\mathcal { E }$ follows circularly-symmetric complex Gaussian distribution, because the sufficient and necessary condition is that mean value and pseudo variance equal zero Goodman (1963). It implies that both the real and imaginary part of $\\mathcal { E }$ follow the same Gaussian distribution and they are independent. Thus, the magnitude of this complex random variable follows the Rayleigh distribution Papoulis & Pillai (2001), and the probability density function of $| \\mathcal { E } |$ can be formulated as ", + "bbox": [ + 173, + 604, + 825, + 689 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/8fcd3d6f2ab95407bc14119a563351950db51c91de02d82b5c69332fca8e2fdf.jpg", + "text": "$$\n\\mathcal { P } _ { | \\varepsilon | } \\left( r \\right) = \\frac { r } { s ^ { 2 } } e ^ { - r ^ { 2 } / 2 s ^ { 2 } } ,\n$$", + "text_format": "latex", + "bbox": [ + 419, + 696, + 576, + 723 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where s is the scale parameter. Knowing the properties of the Rayleigh distributionPapoulis & Pillai (2001), we have: ", + "bbox": [ + 173, + 729, + 825, + 758 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/666b6315498def3bdc6e938b8a4b4c9847345e478a1039a9a2be73c12da1da1c.jpg", + "text": "$$\ns ^ { 2 } = { \\frac { 1 } { 2 } } D ( { \\mathcal { E } } ) = { \\frac { 1 - | a _ { n } | ^ { 2 } } { 2 M } } .\n$$", + "text_format": "latex", + "bbox": [ + 408, + 765, + 589, + 796 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Utilizing the fact that $\\left| a _ { n } \\right| ^ { 2 }$ is close to 0 when n is a modest number, we obtain that $\\begin{array} { r } { D \\left( \\mathcal { E } \\right) \\approx \\frac { 1 } { M } } \\end{array}$ , leading to $\\textstyle s ^ { 2 } = { \\frac { 1 } { 2 M } }$ . Then, we obtain ", + "bbox": [ + 174, + 810, + 825, + 844 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/c8e8e8cb610860d6011ba2f448bfc677e4f4ac66bebe896f1cefb665bde6c066.jpg", + "text": "$$\nE { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 4 - \\pi } { 4 M } } .\n$$", + "text_format": "latex", + "bbox": [ + 370, + 852, + 627, + 887 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is easy to observe that both $E { \\big ( } | { \\mathcal { E } } | { \\big ) }$ and $D ( \\left. \\boldsymbol { \\mathcal { E } } \\right. )$ are close to zero when the number of samples M increases. It implies that the estimation error $\\varepsilon _ { n }$ gets closer to $O$ as $M$ increases. ", + "bbox": [ + 173, + 893, + 823, + 925 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/fa3230b33a62941de39e3eb72b6593e1a48dc25c0980b935c88410a70d320d21.jpg", + "table_caption": [ + "Table 9: The $p$ -value of KS hypothesis test among clean samples, noisy samples, and adversarial samples crafted by four methods, respectively. " + ], + "table_footnote": [], + "table_body": "
cleannoisyBIMCW-L2DeepFoolR-PGD
CIFAR-10train set0.0740.0740.1580.1390.1550.159
test set0.0720.0720.1580.1380.2160.216
SVHNtrain set0.1040.1040.2490.1940.2220.245
test set0.1060.1060.2490.1930.2190.239
ImageNet-AlexNettrain set0.1160.1160.2530.2350.2420.253
test set0.1170.1170.2610.2360.2430.259
ImageNet-VGG-16train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
", + "bbox": [ + 223, + 155, + 774, + 272 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/6da3000751d59c307bab64c1baf09f6e3cb5b7f008aacbef64e8fb64f37a5192.jpg", + "table_caption": [ + "Table 10: The $p$ -value of two-sample KS test among MBF coefficients of different types of examples. " + ], + "table_footnote": [], + "table_body": "
(clean,noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
CIFAR-10train set0.6400.0000.0000.003
test set0.7640.0000.0300.000
SVHNtrain set0.5590.0000.0300.000
test set0.6850.0000.0000.000
ImageNet-AlexNettrain set1.0000.0000.0020.134
test set0.9930.0000.4770.000
ImageNet-VGG-16train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
", + "bbox": [ + 181, + 316, + 815, + 434 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/432208b1f366590fad0e738578fcaad0579c7b3400760126f5ff1bf2e1a454f4.jpg", + "table_caption": [ + "Table 11: The $p$ -value of KS test among MBF coefficients from different data sources. " + ], + "table_footnote": [], + "table_body": "
clean noisyBIMCW-L2DeepFoolR-PGD
CIFAR-10(train set, test set)0.120 0.1880.3330.4560.5840.223
SVHN(train set, test set)0.494 0.7080.5230.1470.4910.645
ImageNet- AlexNet(train set, test set) (train set, out-of-sample)0.153 0.148 0.000 0.0000.370 0.1690.605 0.1010.410 0.2060.386 0.158
ImageNet- VGG-16(train set, test set) (train set, out-of-sample)0.839 0.858 0.049 0.0540.368 0.0940.842 0.4070.246 0.7370.221 0.088
", + "bbox": [ + 173, + 483, + 838, + 585 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/951766bb96a13d49ac27f0e9767ec43586237ffa3aba2ff5527718e6f9c18193.jpg", + "image_caption": [ + "Figure 2: Statistics (mean $\\pm$ standard deviation) of MBF coefficients on train (top row) and test (bottom row) set of CIFAR-10. " + ], + "image_footnote": [], + "bbox": [ + 220, + 609, + 772, + 724 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f90ea0187e4a20a3df47ccc875e9571b38f3c0e93fe3a9c53bf08df094171e49.jpg", + "image_caption": [ + "Figure 3: Statistics (mean $\\pm$ standard deviation) of MBF coefficients on train (top row) and test (bottom row) set of SVHN. " + ], + "image_footnote": [], + "bbox": [ + 222, + 776, + 772, + 891 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a2307b9dd6d8b208fa0ca9565e779759c161a7c1546a2a84a8e287388b4c42f5.jpg", + "image_caption": [ + "Figure 4: Statistics (mean $\\pm$ standard deviation) of MBF coefficients on train (top row) and test (median row), and out-of-sample (bottom row) set of ImageNet-AlexNet. " + ], + "image_footnote": [], + "bbox": [ + 222, + 104, + 772, + 268 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/33ed125ffa3caba424d560cc534dab9906f2037fcc8f12426c7dfde11ab5814d.jpg", + "table_caption": [ + "Table 12: Detection results evaluated by AUROC score $( \\% )$ on nearly 4000 images from the testset of MNIST. The best results are highlighted in bold. " + ], + "table_footnote": [], + "table_body": "
Iteration numberin each GD runNumber of GD runsBIMCW-L2DeepFoolR-PGD
200100.7440.9800.7020.746
100100.6890.9730.6500.690
10050.6640.9690.6280.666
10020.6120.9590.5860.614
MBF1.0000.9911.0001.000
", + "bbox": [ + 232, + 342, + 764, + 411 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.2 ADDITIONAL EMPIRICAL ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 435, + 468, + 450 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Here we present additional empirical analysis on more databases and networks, as shown in Tables 9, 10 and 11. The $p$ -values in most cases also supports the conclusions obtained in the main manuscript. We also present more visualizations in Figs. 2, 3 and 4. These visualizations also demonstrate the distinct difference of MBF features between adversarial and benign examples. ", + "bbox": [ + 174, + 462, + 825, + 517 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.3 COMPARISON BETWEEN DEFENSE-GAN ", + "text_level": 1, + "bbox": [ + 174, + 535, + 498, + 549 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We adopt official code8 of Defense-GAN (Samangouei et al. (2018)) and train an generative model on train set of MNIST dataset after 130, 000 epochs (As the official code of Defense-GAN currently only supports three datasets including MNIST, Fashion-MNIST, and CelebA). The classification model attains an accuracy of $9 8 . 7 \\%$ on the test set of MNIST dataset. We still utilize four attack strategies, including BIM, CW-L2, DeepFool and R-PGD. After removing images that are misclassified and failed to attack, we pick 6000 images to train SVMs based on our MBF features, and leave nearly 4000 images for testing. There are two key hyper-parameters in Defense-GAN, including the iteration number in each GD run, and the total number of GD runs. We have tried multiple settings of these two hyper-parameters. Table 12 show the AUROC score on nearly 4000 images. Our MBF method exceeds Defense-GAN by a large margin in quantitative performance. ", + "bbox": [ + 174, + 560, + 825, + 700 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.4 DETAILED PARAMETERS OF ADVERSARIAL ATTACK ", + "text_level": 1, + "bbox": [ + 174, + 717, + 576, + 731 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Four popular adversarial attack methods are adopted to craft adversarial examples, including basic iterative method (BIM), CarliniWagnerL2Attack (CW-L2), DeepFool, and random projected gradient descent (R-PGD). We emphasize detailed parameters of these attack strategies for reproducibility, which are BIM $\\mathrm { ( e p s { = } 0 . 3 }$ , stepsize $= 0 . 0 5$ , iterations ${ \\it \\Omega } = 1 0$ ), CW-L2 (binary search steps $^ { = 5 }$ , confidence $= 0 . 0$ , learning rate $= 0 . 0 0 5$ , max iterations $\\mathord { \\left. \\begin{array} { r l l } { - 1 0 0 0 } \\end{array} \\right. }$ ), DeepFool (max step $\\mathord { \\mathrm { \\ s } } = 1 0 0$ ), R-PGD $\\mathrm { \\ e p s { = } 0 . 3 }$ , stepsize ${ \\mathrm { \\Omega } } { = } 0 . 0 1$ , iterations $\\scriptstyle = 4 0$ ). All of the attack strategies are implemented by Foolbox. ", + "bbox": [ + 174, + 742, + 825, + 827 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/parse/train/ryeK6nNFDr/ryeK6nNFDr_middle.json b/parse/train/ryeK6nNFDr/ryeK6nNFDr_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..aea4ddb90bc28ca9031643f7016acdaaac98aebb --- /dev/null +++ b/parse/train/ryeK6nNFDr/ryeK6nNFDr_middle.json @@ -0,0 +1,45482 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 503, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "score": 1.0, + "content": "EFFECTIVE AND ROBUST DETECTION OF ADVER-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 98, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 98, + 505, + 117 + ], + "score": 1.0, + "content": "SARIAL EXAMPLES VIA BENFORD-FOURIER COEFFI-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 121, + 163, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 163, + 136 + ], + "score": 1.0, + "content": "CIENTS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 155, + 244, + 177 + ], + "lines": [ + { + "bbox": [ + 113, + 156, + 201, + 167 + ], + "spans": [ + { + "bbox": [ + 113, + 156, + 201, + 167 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 166, + 245, + 178 + ], + "spans": [ + { + "bbox": [ + 112, + 166, + 245, + 178 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 278, + 207, + 333, + 218 + ], + "lines": [ + { + "bbox": [ + 276, + 206, + 335, + 219 + ], + "spans": [ + { + "bbox": [ + 276, + 206, + 335, + 219 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 231, + 468, + 472 + ], + "lines": [ + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "score": 1.0, + "content": "Adversarial examples have been well known as a serious threat to deep neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 243, + 469, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 469, + 254 + ], + "score": 1.0, + "content": "networks (DNNs). To ensure successful and safe operations of DNNs on real-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "score": 1.0, + "content": "world tasks, it is urgent to equip DNNs with effective defense strategies. In this", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "score": 1.0, + "content": "work, we study the detection of adversarial examples, based on the assumption", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 275, + 469, + 286 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 469, + 286 + ], + "score": 1.0, + "content": "that the output and internal responses of one DNN model for both adversarial and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 286, + 469, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 469, + 297 + ], + "score": 1.0, + "content": "benign examples follow the generalized Gaussian distribution (GGD), but with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 296, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 470, + 309 + ], + "score": 1.0, + "content": "different parameters (i.e., shape factor, mean, and variance). GGD is a general", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "distribution family to cover many popular distributions (e.g., Laplacian, Gaussian,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 319, + 469, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 319, + 469, + 330 + ], + "score": 1.0, + "content": "or uniform). It is more likely to approximate the intrinsic distributions of internal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 329, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 470, + 343 + ], + "score": 1.0, + "content": "responses than any specific distribution. Besides, since the shape factor is more", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 340, + 469, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 469, + 353 + ], + "score": 1.0, + "content": "robust to different databases rather than the other two parameters, we propose", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 352, + 469, + 364 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 469, + 364 + ], + "score": 1.0, + "content": "to construct discriminative features via the shape factor for adversarial detection,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 362, + 470, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 362, + 470, + 375 + ], + "score": 1.0, + "content": "employing the magnitude of Benford-Fourier coefficients (MBF), which can be", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 374, + 469, + 385 + ], + "spans": [ + { + "bbox": [ + 141, + 374, + 469, + 385 + ], + "score": 1.0, + "content": "easily estimated using responses. Finally, a support vector machine is trained", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 384, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 470, + 397 + ], + "score": 1.0, + "content": "as the adversarial detector through leveraging the MBF features. Through the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 395, + 470, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 395, + 470, + 408 + ], + "score": 1.0, + "content": "Kolmogorov-Smirnov (KS) test (Massey Jr (1951)), We empirically verify that:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 406, + 470, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 470, + 419 + ], + "score": 1.0, + "content": "1) the posterior vectors of both adversarial and benign examples follow GGD; 2)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 417, + 470, + 430 + ], + "spans": [ + { + "bbox": [ + 141, + 417, + 470, + 430 + ], + "score": 1.0, + "content": "the extracted MBF features of adversarial and benign examples follow different", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 428, + 470, + 441 + ], + "spans": [ + { + "bbox": [ + 141, + 428, + 470, + 441 + ], + "score": 1.0, + "content": "distributions. Extensive experiments in terms of image classification demonstrate", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 439, + 469, + 451 + ], + "spans": [ + { + "bbox": [ + 141, + 439, + 469, + 451 + ], + "score": 1.0, + "content": "that the proposed detector is much more effective and robust on detecting adver-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 450, + 469, + 462 + ], + "spans": [ + { + "bbox": [ + 141, + 450, + 469, + 462 + ], + "score": 1.0, + "content": "sarial examples of different crafting methods and different sources, in contrast to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 462, + 327, + 472 + ], + "spans": [ + { + "bbox": [ + 142, + 462, + 327, + 472 + ], + "score": 1.0, + "content": "state-of-the-art adversarial detection methods.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 492, + 206, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 208, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 208, + 508 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "Deep neural networks (DNNs) have achieved a remarkable success in many important applications,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "such as image classification, face recognition, object detection, etc. In the meanwhile, DNNs have", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "been shown to be very vulnerable to adversarial examples. However, many real-world scenarios", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "have very restrictive requirements about the robustness of DNNs, such as face verification for login,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "or semantic segmentation in autonomous driving. Adversarial examples are a serious threat to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "applications of DNNs to these important tasks. Since many kinds of adversarial attack methods have", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "been proposed to fool DNNs, it is more urgent to equip effective defensive strategies to ensure the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 608 + ], + "score": 1.0, + "content": "safety of deep models in real-world applications. However, defense seems to be more challenging", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "than attack, as it has to face adversarial examples from unknown crafting methods and unknown", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "data sources. Typical defensive strategies include adversarial training, adversarial de-noising, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "adversarial detection. Compared to the former two strategies, adversarial detection is somewhat", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 474, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 474, + 650 + ], + "score": 1.0, + "content": "more cost-effective, as it often does not need to re-train or modify the original DNN model.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "There are two main challenges for adversarial detection. (1) The adversarial examples are designed", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "to camouflage themselves to be close to the corresponding benign examples in the input space. Then,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "where and how to extract the discriminative information to train the detector? (2) The data sources", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "and the generating methods of adversarial examples are often inaccessible to the detector. In this", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 710 + ], + "score": 1.0, + "content": "case, the detector can be stably effective across different data sources and different attack meth-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "ods? In other words, a good adversarial detector is required to be not only effective to distinguish", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 482, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 482, + 732 + ], + "score": 1.0, + "content": "adversarial and benign examples, but also robust to different data sources and attack methods.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 503, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "score": 1.0, + "content": "EFFECTIVE AND ROBUST DETECTION OF ADVER-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 98, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 104, + 98, + 505, + 117 + ], + "score": 1.0, + "content": "SARIAL EXAMPLES VIA BENFORD-FOURIER COEFFI-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 121, + 163, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 163, + 136 + ], + "score": 1.0, + "content": "CIENTS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 155, + 244, + 177 + ], + "lines": [ + { + "bbox": [ + 113, + 156, + 201, + 167 + ], + "spans": [ + { + "bbox": [ + 113, + 156, + 201, + 167 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 166, + 245, + 178 + ], + "spans": [ + { + "bbox": [ + 112, + 166, + 245, + 178 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 112, + 156, + 245, + 178 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 207, + 333, + 218 + ], + "lines": [ + { + "bbox": [ + 276, + 206, + 335, + 219 + ], + "spans": [ + { + "bbox": [ + 276, + 206, + 335, + 219 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 231, + 468, + 472 + ], + "lines": [ + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 231, + 469, + 243 + ], + "score": 1.0, + "content": "Adversarial examples have been well known as a serious threat to deep neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 243, + 469, + 254 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 469, + 254 + ], + "score": 1.0, + "content": "networks (DNNs). To ensure successful and safe operations of DNNs on real-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 264 + ], + "score": 1.0, + "content": "world tasks, it is urgent to equip DNNs with effective defense strategies. In this", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 263, + 469, + 276 + ], + "score": 1.0, + "content": "work, we study the detection of adversarial examples, based on the assumption", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 275, + 469, + 286 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 469, + 286 + ], + "score": 1.0, + "content": "that the output and internal responses of one DNN model for both adversarial and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 286, + 469, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 469, + 297 + ], + "score": 1.0, + "content": "benign examples follow the generalized Gaussian distribution (GGD), but with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 296, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 296, + 470, + 309 + ], + "score": 1.0, + "content": "different parameters (i.e., shape factor, mean, and variance). GGD is a general", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "distribution family to cover many popular distributions (e.g., Laplacian, Gaussian,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 319, + 469, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 319, + 469, + 330 + ], + "score": 1.0, + "content": "or uniform). It is more likely to approximate the intrinsic distributions of internal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 329, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 470, + 343 + ], + "score": 1.0, + "content": "responses than any specific distribution. Besides, since the shape factor is more", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 340, + 469, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 469, + 353 + ], + "score": 1.0, + "content": "robust to different databases rather than the other two parameters, we propose", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 352, + 469, + 364 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 469, + 364 + ], + "score": 1.0, + "content": "to construct discriminative features via the shape factor for adversarial detection,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 362, + 470, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 362, + 470, + 375 + ], + "score": 1.0, + "content": "employing the magnitude of Benford-Fourier coefficients (MBF), which can be", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 374, + 469, + 385 + ], + "spans": [ + { + "bbox": [ + 141, + 374, + 469, + 385 + ], + "score": 1.0, + "content": "easily estimated using responses. Finally, a support vector machine is trained", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 384, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 470, + 397 + ], + "score": 1.0, + "content": "as the adversarial detector through leveraging the MBF features. Through the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 395, + 470, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 395, + 470, + 408 + ], + "score": 1.0, + "content": "Kolmogorov-Smirnov (KS) test (Massey Jr (1951)), We empirically verify that:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 406, + 470, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 470, + 419 + ], + "score": 1.0, + "content": "1) the posterior vectors of both adversarial and benign examples follow GGD; 2)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 417, + 470, + 430 + ], + "spans": [ + { + "bbox": [ + 141, + 417, + 470, + 430 + ], + "score": 1.0, + "content": "the extracted MBF features of adversarial and benign examples follow different", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 428, + 470, + 441 + ], + "spans": [ + { + "bbox": [ + 141, + 428, + 470, + 441 + ], + "score": 1.0, + "content": "distributions. Extensive experiments in terms of image classification demonstrate", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 439, + 469, + 451 + ], + "spans": [ + { + "bbox": [ + 141, + 439, + 469, + 451 + ], + "score": 1.0, + "content": "that the proposed detector is much more effective and robust on detecting adver-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 450, + 469, + 462 + ], + "spans": [ + { + "bbox": [ + 141, + 450, + 469, + 462 + ], + "score": 1.0, + "content": "sarial examples of different crafting methods and different sources, in contrast to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 462, + 327, + 472 + ], + "spans": [ + { + "bbox": [ + 142, + 462, + 327, + 472 + ], + "score": 1.0, + "content": "state-of-the-art adversarial detection methods.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 16.5, + "bbox_fs": [ + 141, + 231, + 470, + 472 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 492, + 206, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 491, + 208, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 208, + 508 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "Deep neural networks (DNNs) have achieved a remarkable success in many important applications,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "such as image classification, face recognition, object detection, etc. In the meanwhile, DNNs have", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "been shown to be very vulnerable to adversarial examples. However, many real-world scenarios", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "have very restrictive requirements about the robustness of DNNs, such as face verification for login,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "or semantic segmentation in autonomous driving. Adversarial examples are a serious threat to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "applications of DNNs to these important tasks. Since many kinds of adversarial attack methods have", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "been proposed to fool DNNs, it is more urgent to equip effective defensive strategies to ensure the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 608 + ], + "score": 1.0, + "content": "safety of deep models in real-world applications. However, defense seems to be more challenging", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "than attack, as it has to face adversarial examples from unknown crafting methods and unknown", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "data sources. Typical defensive strategies include adversarial training, adversarial de-noising, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "adversarial detection. Compared to the former two strategies, adversarial detection is somewhat", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 474, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 474, + 650 + ], + "score": 1.0, + "content": "more cost-effective, as it often does not need to re-train or modify the original DNN model.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 517, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "There are two main challenges for adversarial detection. (1) The adversarial examples are designed", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "to camouflage themselves to be close to the corresponding benign examples in the input space. Then,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "where and how to extract the discriminative information to train the detector? (2) The data sources", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "and the generating methods of adversarial examples are often inaccessible to the detector. In this", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 710 + ], + "score": 1.0, + "content": "case, the detector can be stably effective across different data sources and different attack meth-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "ods? In other words, a good adversarial detector is required to be not only effective to distinguish", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 482, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 482, + 732 + ], + "score": 1.0, + "content": "adversarial and benign examples, but also robust to different data sources and attack methods.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 655, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "To satisfy the first requirement of effectiveness, we utilize the other principle of crafting adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 108 + ], + "score": 1.0, + "content": "examples that the outputs between benign and adversarial examples should be large, to encourage", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "the change of the final prediction. It means that the imperceptible difference between benign and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "adversarial examples in the input space is enlarged along the DNN model, leading to the significant", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "difference in the output space. Inspired by this fact, we assume that the output or the responses", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "of internal layers of the DNN model should include the discriminative information for benign and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "adversarial examples. A few works have attempted to extract different types of discriminative fea-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "tures from the output or the internal responses, such as kernel density (KD) (Feinman et al. (2017))", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "score": 1.0, + "content": "and the local intrinsic dimensionality (LID) (Ma et al. (2018)), etc. To achieve the robustness, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 182, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 193 + ], + "score": 1.0, + "content": "extracted discriminative features should model the intrinsic difference between adversarial and be-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "nign examples, rather than the difference from the changes of data sources or attack methods. Many", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "score": 1.0, + "content": "existing methods have shown the effectiveness to some extent on detecting adversarial examples of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "specific data sources and attack methods. However, their robustness, especially across different data", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 296, + 237 + ], + "score": 1.0, + "content": "sources, has not been well studied and verified.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "In this work, we propose a novel detection method based on the assumption that the internal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "responses of both adversarial and benign examples follow the generalized Gaussian distribution", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "(GGD) (Varanasi & Aazhang (1989)), but with different parameters, including the shape factor,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "mean, and variance. The rationale behind this assumption is that GGD covers many popular distri-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "butions with varied shape factors (such as Laplacian, Gaussian, or uniform distribution), such that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "score": 1.0, + "content": "GGD is more likely to approximate the intrinsic response distributions rather than one specific distri-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "bution. Moreover, mean and variance of GGD may vary significantly with respect to different classes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "and data sources, even for benign examples, while the shape factor is more robust. For example, the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "mean and variance of two Gaussian distributions could be totally different, but their shape factors are", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "the same (i.e., 2). Thus, we propose to use the shape factor as an effective and robust discrimination", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 351, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 504, + 364 + ], + "score": 1.0, + "content": "between adversarial and benign examples. However, it is difficult to exactly estimate the shape fac-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "tor in practice. We resort to the magnitude of Benford-Fourier coefficients (Pasquini et al. (2014)),", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 372, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 387 + ], + "score": 1.0, + "content": "which is a function of the shape factor. It can be easily estimated using internal responses, according", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "to the definition of Fourier transform. Then, the magnitudes estimated from internal responses of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "different convolutional layers are concatenated as a novel representation. Finally, a support vector", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 406, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 504, + 419 + ], + "score": 1.0, + "content": "machine (SVM) (Vapnik (2013)) is trained using the new representations as the adversarial detec-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "score": 1.0, + "content": "tor. Extensive experiments carried out on several databases verify the effectiveness and robustness", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "of the proposed detection method. To further verify the rationale of our assumption, we present", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "the empirical analysis through the Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)). The KS", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "test verifies that 1) the posterior vectors of both adversarial and benign examples predicted by the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "CNN model follow the distribution of GGD, but with different parameters, and 2) the MBF fea-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "score": 1.0, + "content": "tures of adversarial and benign examples follow different distributions, and the MBF features of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "adversarial examples crafted from different attack methods follow the same distribution, as well as", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "that the MBF features of adversarial/benign examples from different data sources follow the same", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 350, + 517 + ], + "score": 1.0, + "content": "distribution. Moreover, we visualize the statistics (i.e., mean", + "type": "text" + }, + { + "bbox": [ + 350, + 505, + 361, + 515 + ], + "score": 0.7, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "standard deviation) of the extracted", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "MBF features for adversarial and benign examples. The visualization reveals the distinct difference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "between adversarial and benign examples. These empirical analysis demonstrates the effectiveness", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 537, + 292, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 292, + 551 + ], + "score": 1.0, + "content": "and robustness of the proposed MBF detector.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 208, + 580 + ], + "lines": [ + { + "bbox": [ + 104, + 566, + 211, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 566, + 211, + 582 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 504, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "In this section we focus on reviewing existing works about adversarial detection, while those of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 604, + 375, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 375, + 617 + ], + "score": 1.0, + "content": "other adversarial defensive strategies are out of scope of this work.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "The general idea of most existing detection methods is learning or constructing a new representation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "to discriminate adversarial and benign examples, utilizing the outputs or immediate responses of an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "original classification network. Li & Li (2017) trained a cascading classifier based on the principal", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "component analysis (PCA) (Pearson (1901)) statistics of responses from each convolutional layer", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "of the defended convolutional neural network (CNN) model. An example is recognized as benign", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "if all single classifiers of the cascade predict it as benign, otherwise adversarial. Lu et al. (2017)", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "proposed SafetyNet by adding a RBF-SVM classifier to detect adversarial examples, at the end of", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the original classification network. Metzen et al. (2017) proposed a detection network along with an", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "original classification network, which takes the internal responses of the original network as inputs.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "It shows effectiveness on detecting adversarial examples generated by simple attacks (such as FGSM", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 49.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 236 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "To satisfy the first requirement of effectiveness, we utilize the other principle of crafting adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 108 + ], + "score": 1.0, + "content": "examples that the outputs between benign and adversarial examples should be large, to encourage", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "the change of the final prediction. It means that the imperceptible difference between benign and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "adversarial examples in the input space is enlarged along the DNN model, leading to the significant", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "difference in the output space. Inspired by this fact, we assume that the output or the responses", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "of internal layers of the DNN model should include the discriminative information for benign and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "adversarial examples. A few works have attempted to extract different types of discriminative fea-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "tures from the output or the internal responses, such as kernel density (KD) (Feinman et al. (2017))", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "score": 1.0, + "content": "and the local intrinsic dimensionality (LID) (Ma et al. (2018)), etc. To achieve the robustness, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 182, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 193 + ], + "score": 1.0, + "content": "extracted discriminative features should model the intrinsic difference between adversarial and be-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "nign examples, rather than the difference from the changes of data sources or attack methods. Many", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "score": 1.0, + "content": "existing methods have shown the effectiveness to some extent on detecting adversarial examples of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "specific data sources and attack methods. However, their robustness, especially across different data", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 225, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 296, + 237 + ], + "score": 1.0, + "content": "sources, has not been well studied and verified.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 82, + 506, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "In this work, we propose a novel detection method based on the assumption that the internal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "responses of both adversarial and benign examples follow the generalized Gaussian distribution", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "(GGD) (Varanasi & Aazhang (1989)), but with different parameters, including the shape factor,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "mean, and variance. The rationale behind this assumption is that GGD covers many popular distri-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "butions with varied shape factors (such as Laplacian, Gaussian, or uniform distribution), such that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "score": 1.0, + "content": "GGD is more likely to approximate the intrinsic response distributions rather than one specific distri-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "bution. Moreover, mean and variance of GGD may vary significantly with respect to different classes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "and data sources, even for benign examples, while the shape factor is more robust. For example, the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "mean and variance of two Gaussian distributions could be totally different, but their shape factors are", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "the same (i.e., 2). Thus, we propose to use the shape factor as an effective and robust discrimination", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 351, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 504, + 364 + ], + "score": 1.0, + "content": "between adversarial and benign examples. However, it is difficult to exactly estimate the shape fac-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 376 + ], + "score": 1.0, + "content": "tor in practice. We resort to the magnitude of Benford-Fourier coefficients (Pasquini et al. (2014)),", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 372, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 387 + ], + "score": 1.0, + "content": "which is a function of the shape factor. It can be easily estimated using internal responses, according", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "to the definition of Fourier transform. Then, the magnitudes estimated from internal responses of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "different convolutional layers are concatenated as a novel representation. Finally, a support vector", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 406, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 504, + 419 + ], + "score": 1.0, + "content": "machine (SVM) (Vapnik (2013)) is trained using the new representations as the adversarial detec-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 429 + ], + "score": 1.0, + "content": "tor. Extensive experiments carried out on several databases verify the effectiveness and robustness", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "of the proposed detection method. To further verify the rationale of our assumption, we present", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "the empirical analysis through the Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)). The KS", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "test verifies that 1) the posterior vectors of both adversarial and benign examples predicted by the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "CNN model follow the distribution of GGD, but with different parameters, and 2) the MBF fea-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "score": 1.0, + "content": "tures of adversarial and benign examples follow different distributions, and the MBF features of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "adversarial examples crafted from different attack methods follow the same distribution, as well as", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "that the MBF features of adversarial/benign examples from different data sources follow the same", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 350, + 517 + ], + "score": 1.0, + "content": "distribution. Moreover, we visualize the statistics (i.e., mean", + "type": "text" + }, + { + "bbox": [ + 350, + 505, + 361, + 515 + ], + "score": 0.7, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "standard deviation) of the extracted", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "MBF features for adversarial and benign examples. The visualization reveals the distinct difference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "between adversarial and benign examples. These empirical analysis demonstrates the effectiveness", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 537, + 292, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 292, + 551 + ], + "score": 1.0, + "content": "and robustness of the proposed MBF detector.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 242, + 506, + 551 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 208, + 580 + ], + "lines": [ + { + "bbox": [ + 104, + 566, + 211, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 566, + 211, + 582 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 504, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "In this section we focus on reviewing existing works about adversarial detection, while those of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 604, + 375, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 375, + 617 + ], + "score": 1.0, + "content": "other adversarial defensive strategies are out of scope of this work.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 594, + 505, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "The general idea of most existing detection methods is learning or constructing a new representation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "to discriminate adversarial and benign examples, utilizing the outputs or immediate responses of an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "original classification network. Li & Li (2017) trained a cascading classifier based on the principal", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "component analysis (PCA) (Pearson (1901)) statistics of responses from each convolutional layer", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "of the defended convolutional neural network (CNN) model. An example is recognized as benign", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "if all single classifiers of the cascade predict it as benign, otherwise adversarial. Lu et al. (2017)", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "proposed SafetyNet by adding a RBF-SVM classifier to detect adversarial examples, at the end of", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "the original classification network. Metzen et al. (2017) proposed a detection network along with an", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "original classification network, which takes the internal responses of the original network as inputs.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "It shows effectiveness on detecting adversarial examples generated by simple attacks (such as FGSM", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(Goodfellow et al. (2014)) and JSMA (Papernot et al. (2016))), while performs much worse when", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "facing more advanced attacks (such as C&W (Carlini & Wagner (2017a))). It tells that this method", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "is sensitive to attack methods. Grosse et al. (2017) attempted to detect adversarial examples using", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "the statistical test of maximum mean discrepancy (MMD). Although above detection methods show", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "effectiveness on some attack methods and some databases, but a thorough evaluation presented in", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "Carlini & Wagner (2017b) has shown that these methods are sensitive to attack methods or databases,", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 336, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 336, + 161 + ], + "score": 1.0, + "content": "and they can be somewhat easily invaded by new attacks.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 49.5, + "bbox_fs": [ + 104, + 621, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(Goodfellow et al. (2014)) and JSMA (Papernot et al. (2016))), while performs much worse when", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "facing more advanced attacks (such as C&W (Carlini & Wagner (2017a))). It tells that this method", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "is sensitive to attack methods. Grosse et al. (2017) attempted to detect adversarial examples using", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "the statistical test of maximum mean discrepancy (MMD). Although above detection methods show", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 138 + ], + "score": 1.0, + "content": "effectiveness on some attack methods and some databases, but a thorough evaluation presented in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "Carlini & Wagner (2017b) has shown that these methods are sensitive to attack methods or databases,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 336, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 336, + 161 + ], + "score": 1.0, + "content": "and they can be somewhat easily invaded by new attacks.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "Some recent works proposed to utilize neighboring samples in the same database to construct a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "better representation of a current sample. Feinman et al. (2017) defined two metrics based on the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "responses of the final hidden layer of the classification neural network, including kernel density", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "estimation (KDE) and Bayesian neural network uncertainty (BU). If the metric score of KDE/BU", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "is lower/higher than a pre-defined threshold, then the example is predicted as adversarial. Ma et al.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "(2018) utilized the local intrinsic dimensionality (LID) to measure the characterization of adversarial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 414, + 244 + ], + "score": 1.0, + "content": "regions of DNNs. LID describes the distance between one example and its", + "type": "text" + }, + { + "bbox": [ + 414, + 231, + 421, + 241 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "-nearest neighboring", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "sample in the feature space of immediate responses of the original classification network. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "distances computed from different layers are concatenated as the example representation, which is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "then used to train a shallow classifier to discriminate adversarial and benign examples. Zheng &", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "Hong (2018) defined the intrinsic hidden state distribution (IHSD) of the responses of the original", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "classification network to model different classes. The Gaussian mixture model (GMM) was used", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "score": 1.0, + "content": "to approximate IHSD of each class. Then, the posterior probability of one sample assigned to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "GMM is computed as the metric. If the probability is lower than a pre-defined threshold, then it is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "recognized as adversarial. Lee et al. (2018) computed the class-conditional Gaussian distribution", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "score": 1.0, + "content": "of the responses of the original classification network based on the whole training set. Then, the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "Mahalanobis distance between one sample and its nearest class-conditional Gaussian distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "score": 1.0, + "content": "is used as the metric for detection. If the distance is larger than a pre-defined threshold, then it is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "detected as adversarial. Compared to some aforementioned single-representation-based detection", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "methods, these joint-representation-based methods showed better performances on some databases.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "score": 1.0, + "content": "However, the detection cost for each example is much higher, as the responses of its neighboring", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 396, + 504, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 504, + 407 + ], + "score": 1.0, + "content": "samples should also be computed. Besides, since the representation is highly dependent on the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 405, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 420 + ], + "score": 1.0, + "content": "neighbors or all training examples, the detection performance may be sensitive to data sources,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 417, + 278, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 278, + 429 + ], + "score": 1.0, + "content": "which will be studied in later experiments.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "There are also some other approaches that do not construct representations from the responses of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "the original classification network. Hendrycks & Gimpel (2016) adopted PCA statistics to dis-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "criminate adversarial and benign images, independent of any DNN model. However, the study", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "presented in Carlini & Wagner (2017b) has demonstrated that this method works for MNIST but not", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "for CIFAR-10, and PCA statistics are not robust features to detect adversarial images. Pang et al.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "(2018) proposed a novel loss called reverse cross-entropy (RCE) to train the classification network,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "such that the distance measured by kernel density (Feinman et al. (2017)) between adversarial and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "benign examples could be enlarged. Samangouei et al. (2018) proposed Defense-GAN to model the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "score": 1.0, + "content": "distribution of benign examples using a generative adversarial network (GAN). If the Wasserstein", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 533, + 504, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 504, + 545 + ], + "score": 1.0, + "content": "distance between one example and its corresponding example generated by the GAN is larger than", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 544, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 555 + ], + "score": 1.0, + "content": "a pre-defined value, then it is detected as adversarial. However, these above three methods are much", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 555, + 276, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 276, + 566 + ], + "score": 1.0, + "content": "more costly than other detection methods.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 108, + 583, + 208, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 209, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 209, + 598 + ], + "score": 1.0, + "content": "3 PRELIMINARIES", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 107, + 609, + 307, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 309, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 309, + 622 + ], + "score": 1.0, + "content": "3.1 GENERALIZED GAUSSIAN DISTRIBUTION", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 241, + 643 + ], + "score": 1.0, + "content": "Assume that a random variable", + "type": "text" + }, + { + "bbox": [ + 241, + 629, + 281, + 640 + ], + "score": 0.92, + "content": "\\mathcal { X } ~ \\in ~ \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 628, + 505, + 643 + ], + "score": 1.0, + "content": "follows the generalized Gaussian distribution (GGD)", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 146, + 653 + ], + "score": 1.0, + "content": "(Varanasi", + "type": "text" + }, + { + "bbox": [ + 147, + 641, + 156, + 651 + ], + "score": 0.27, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "Aazhang (1989)). Then, its probability density function (PDF) is formulated with two", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 653, + 463, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 294, + 664 + ], + "score": 1.0, + "content": "positive parameters, including the shape factor", + "type": "text" + }, + { + "bbox": [ + 295, + 654, + 300, + 662 + ], + "score": 0.81, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 653, + 408, + 664 + ], + "score": 1.0, + "content": "and the standard deviation", + "type": "text" + }, + { + "bbox": [ + 408, + 654, + 415, + 662 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 653, + 463, + 664 + ], + "score": 1.0, + "content": ", as follows", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "interline_equation", + "bbox": [ + 261, + 669, + 350, + 685 + ], + "lines": [ + { + "bbox": [ + 261, + 669, + 350, + 685 + ], + "spans": [ + { + "bbox": [ + 261, + 669, + 350, + 685 + ], + "score": 0.92, + "content": "\\mathcal { P } _ { \\mathcal { X } } ( x ) = A \\cdot e ^ { - \\lvert \\beta x \\rvert ^ { c } } ,", + "type": "interline_equation", + "image_path": "749d0e34905c0f5b3529b4411d95ccd4fd3b1b23f07643de44cfdf9789915504.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 261, + 669, + 350, + 685 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 693, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 692, + 508, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 195, + 712 + ], + "score": 1.0, + "content": "where β = 1σ \u0000 Γ(3/c)Γ(1/c) \u0001", + "type": "text" + }, + { + "bbox": [ + 201, + 692, + 220, + 716 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 221, + 694, + 271, + 711 + ], + "score": 0.94, + "content": "\\begin{array} { r } { A = \\frac { \\beta c } { 2 \\Gamma ( 1 / c ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 692, + 295, + 716 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 295, + 697, + 313, + 709 + ], + "score": 0.89, + "content": "\\Gamma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 692, + 508, + 716 + ], + "score": 1.0, + "content": "being the Gamma function. Note that the mean", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 149, + 722 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 149, + 712, + 156, + 721 + ], + "score": 0.81, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 709, + 240, + 722 + ], + "score": 1.0, + "content": "is omitted above, as", + "type": "text" + }, + { + "bbox": [ + 241, + 712, + 248, + 721 + ], + "score": 0.82, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "has no relation with the shape of distribution and we set it as 0", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "without loss of generality. A nice characteristic of GGD is that it covers many popular distributions", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 105, + 82, + 506, + 161 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "Some recent works proposed to utilize neighboring samples in the same database to construct a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "better representation of a current sample. Feinman et al. (2017) defined two metrics based on the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "responses of the final hidden layer of the classification neural network, including kernel density", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "estimation (KDE) and Bayesian neural network uncertainty (BU). If the metric score of KDE/BU", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "is lower/higher than a pre-defined threshold, then the example is predicted as adversarial. Ma et al.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "(2018) utilized the local intrinsic dimensionality (LID) to measure the characterization of adversarial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 414, + 244 + ], + "score": 1.0, + "content": "regions of DNNs. LID describes the distance between one example and its", + "type": "text" + }, + { + "bbox": [ + 414, + 231, + 421, + 241 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "-nearest neighboring", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "sample in the feature space of immediate responses of the original classification network. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "distances computed from different layers are concatenated as the example representation, which is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "then used to train a shallow classifier to discriminate adversarial and benign examples. Zheng &", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "Hong (2018) defined the intrinsic hidden state distribution (IHSD) of the responses of the original", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "classification network to model different classes. The Gaussian mixture model (GMM) was used", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "score": 1.0, + "content": "to approximate IHSD of each class. Then, the posterior probability of one sample assigned to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "GMM is computed as the metric. If the probability is lower than a pre-defined threshold, then it is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "recognized as adversarial. Lee et al. (2018) computed the class-conditional Gaussian distribution", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 341 + ], + "score": 1.0, + "content": "of the responses of the original classification network based on the whole training set. Then, the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "Mahalanobis distance between one sample and its nearest class-conditional Gaussian distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 364 + ], + "score": 1.0, + "content": "is used as the metric for detection. If the distance is larger than a pre-defined threshold, then it is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "detected as adversarial. Compared to some aforementioned single-representation-based detection", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "methods, these joint-representation-based methods showed better performances on some databases.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 398 + ], + "score": 1.0, + "content": "However, the detection cost for each example is much higher, as the responses of its neighboring", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 396, + 504, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 504, + 407 + ], + "score": 1.0, + "content": "samples should also be computed. Besides, since the representation is highly dependent on the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 405, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 420 + ], + "score": 1.0, + "content": "neighbors or all training examples, the detection performance may be sensitive to data sources,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 417, + 278, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 278, + 429 + ], + "score": 1.0, + "content": "which will be studied in later experiments.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 165, + 506, + 429 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "There are also some other approaches that do not construct representations from the responses of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "the original classification network. Hendrycks & Gimpel (2016) adopted PCA statistics to dis-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "criminate adversarial and benign images, independent of any DNN model. However, the study", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "presented in Carlini & Wagner (2017b) has demonstrated that this method works for MNIST but not", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 490 + ], + "score": 1.0, + "content": "for CIFAR-10, and PCA statistics are not robust features to detect adversarial images. Pang et al.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "(2018) proposed a novel loss called reverse cross-entropy (RCE) to train the classification network,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "such that the distance measured by kernel density (Feinman et al. (2017)) between adversarial and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "benign examples could be enlarged. Samangouei et al. (2018) proposed Defense-GAN to model the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "score": 1.0, + "content": "distribution of benign examples using a generative adversarial network (GAN). If the Wasserstein", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 533, + 504, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 504, + 545 + ], + "score": 1.0, + "content": "distance between one example and its corresponding example generated by the GAN is larger than", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 544, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 555 + ], + "score": 1.0, + "content": "a pre-defined value, then it is detected as adversarial. However, these above three methods are much", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 555, + 276, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 276, + 566 + ], + "score": 1.0, + "content": "more costly than other detection methods.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 433, + 506, + 566 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 583, + 208, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 209, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 209, + 598 + ], + "score": 1.0, + "content": "3 PRELIMINARIES", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 107, + 609, + 307, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 309, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 309, + 622 + ], + "score": 1.0, + "content": "3.1 GENERALIZED GAUSSIAN DISTRIBUTION", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 241, + 643 + ], + "score": 1.0, + "content": "Assume that a random variable", + "type": "text" + }, + { + "bbox": [ + 241, + 629, + 281, + 640 + ], + "score": 0.92, + "content": "\\mathcal { X } ~ \\in ~ \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 628, + 505, + 643 + ], + "score": 1.0, + "content": "follows the generalized Gaussian distribution (GGD)", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 146, + 653 + ], + "score": 1.0, + "content": "(Varanasi", + "type": "text" + }, + { + "bbox": [ + 147, + 641, + 156, + 651 + ], + "score": 0.27, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "Aazhang (1989)). Then, its probability density function (PDF) is formulated with two", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 653, + 463, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 294, + 664 + ], + "score": 1.0, + "content": "positive parameters, including the shape factor", + "type": "text" + }, + { + "bbox": [ + 295, + 654, + 300, + 662 + ], + "score": 0.81, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 653, + 408, + 664 + ], + "score": 1.0, + "content": "and the standard deviation", + "type": "text" + }, + { + "bbox": [ + 408, + 654, + 415, + 662 + ], + "score": 0.76, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 653, + 463, + 664 + ], + "score": 1.0, + "content": ", as follows", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 628, + 505, + 664 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 261, + 669, + 350, + 685 + ], + "lines": [ + { + "bbox": [ + 261, + 669, + 350, + 685 + ], + "spans": [ + { + "bbox": [ + 261, + 669, + 350, + 685 + ], + "score": 0.92, + "content": "\\mathcal { P } _ { \\mathcal { X } } ( x ) = A \\cdot e ^ { - \\lvert \\beta x \\rvert ^ { c } } ,", + "type": "interline_equation", + "image_path": "749d0e34905c0f5b3529b4411d95ccd4fd3b1b23f07643de44cfdf9789915504.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 261, + 669, + 350, + 685 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 693, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 692, + 508, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 195, + 712 + ], + "score": 1.0, + "content": "where β = 1σ \u0000 Γ(3/c)Γ(1/c) \u0001", + "type": "text" + }, + { + "bbox": [ + 201, + 692, + 220, + 716 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 221, + 694, + 271, + 711 + ], + "score": 0.94, + "content": "\\begin{array} { r } { A = \\frac { \\beta c } { 2 \\Gamma ( 1 / c ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 692, + 295, + 716 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 295, + 697, + 313, + 709 + ], + "score": 0.89, + "content": "\\Gamma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 692, + 508, + 716 + ], + "score": 1.0, + "content": "being the Gamma function. Note that the mean", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 149, + 722 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 149, + 712, + 156, + 721 + ], + "score": 0.81, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 709, + 240, + 722 + ], + "score": 1.0, + "content": "is omitted above, as", + "type": "text" + }, + { + "bbox": [ + 241, + 712, + 248, + 721 + ], + "score": 0.82, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "has no relation with the shape of distribution and we set it as 0", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "without loss of generality. A nice characteristic of GGD is that it covers many popular distributions", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 692, + 508, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 297, + 96 + ], + "score": 1.0, + "content": "with varied shape factors. For example, when", + "type": "text" + }, + { + "bbox": [ + 298, + 83, + 324, + 93 + ], + "score": 0.9, + "content": "c = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 81, + 505, + 96 + ], + "score": 1.0, + "content": ", then it becomes the Laplacian distribution;", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 131, + 106 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 94, + 157, + 104 + ], + "score": 0.89, + "content": "c = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 93, + 380, + 106 + ], + "score": 1.0, + "content": ", then it is the Gaussian distribution with a variance of √ √", + "type": "text" + }, + { + "bbox": [ + 380, + 93, + 392, + 104 + ], + "score": 0.88, + "content": "\\sigma ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 93, + 421, + 106 + ], + "score": 1.0, + "content": "; when", + "type": "text" + }, + { + "bbox": [ + 421, + 95, + 462, + 104 + ], + "score": 0.89, + "content": "c \\to + \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 93, + 505, + 106 + ], + "score": 1.0, + "content": ", then it is", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 321, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 258, + 117 + ], + "score": 1.0, + "content": "specified as a uniform distribution on", + "type": "text" + }, + { + "bbox": [ + 258, + 104, + 316, + 117 + ], + "score": 0.91, + "content": "( - \\sqrt { 2 } \\sigma , \\sqrt { 2 } \\sigma )", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 104, + 321, + 117 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 129, + 281, + 141 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 281, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 281, + 141 + ], + "score": 1.0, + "content": "3.2 BENFORD-FOURIER COEFFICIENTS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "Although generalized Gaussian distribution (GGD) is able to cover a bunch of distributions, it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "hard to depict the exact forms of GGD precisely. To this end, we further define a random variable", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 169, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 107, + 172, + 167, + 184 + ], + "score": 0.87, + "content": "\\mathcal { Z } = \\log _ { 1 0 } | \\mathcal { X } |", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 169, + 507, + 186 + ], + "score": 1.0, + "content": "mod 1 for detecting and distinguishing different form of GGD, of which the PDF is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "formulated by means of Fourier Series as Perez-Gonz ´ alez et al. (2007), with the fundamental period ´", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 181, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 164, + 207 + ], + "score": 1.0, + "content": "being fixed as", + "type": "text" + }, + { + "bbox": [ + 164, + 194, + 176, + 204 + ], + "score": 0.82, + "content": "2 \\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 192, + 181, + 207 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 209, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 119, + 209, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 119, + 209, + 479, + 243 + ], + "score": 0.94, + "content": "\\mathcal { P } _ { Z } ( z ) = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor A _ { n } \\cos { ( 2 \\pi n z ) } + B _ { n } \\sin { ( 2 \\pi n z ) } \\rfloor = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor a _ { n } \\rfloor \\cos ( 2 \\pi n z + \\phi _ { n } ) ,", + "type": "interline_equation", + "image_path": "46c52c4fa59c4e2d00b3ed521c2e6873b0ae03399402a4950cbde421f3f6ffb7.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 119, + 209, + 479, + 220.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 119, + 220.33333333333334, + 479, + 231.66666666666669 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 119, + 231.66666666666669, + 479, + 243.00000000000003 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 247, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 134, + 261 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 248, + 176, + 260 + ], + "score": 0.92, + "content": "z \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 246, + 369, + 261 + ], + "score": 1.0, + "content": "corresponds to the domain of random variable", + "type": "text" + }, + { + "bbox": [ + 369, + 248, + 378, + 258 + ], + "score": 0.81, + "content": "\\mathcal { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 246, + 506, + 261 + ], + "score": 1.0, + "content": ", the phase of Fourier Series is", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 254, + 503, + 279 + ], + "spans": [ + { + "bbox": [ + 102, + 254, + 158, + 279 + ], + "score": 1.0, + "content": "explained as", + "type": "text" + }, + { + "bbox": [ + 158, + 259, + 245, + 274 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\phi _ { n } = \\arctan \\left( - \\frac { B _ { n } } { A _ { n } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 254, + 356, + 279 + ], + "score": 1.0, + "content": ", and the magnitude denotes", + "type": "text" + }, + { + "bbox": [ + 357, + 259, + 435, + 273 + ], + "score": 0.91, + "content": "| a _ { n } | = \\sqrt { A _ { n } ^ { 2 } + B _ { n } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 254, + 439, + 279 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 439, + 259, + 503, + 273 + ], + "score": 0.9, + "content": "a _ { n } = | a _ { n } | \\cdot e ^ { j \\phi _ { n } }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 441, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 154, + 285 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 154, + 275, + 161, + 283 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 272, + 261, + 285 + ], + "score": 1.0, + "content": "-th Fourier coefficient of", + "type": "text" + }, + { + "bbox": [ + 261, + 273, + 288, + 285 + ], + "score": 0.93, + "content": "\\mathcal { P } _ { \\mathcal { Z } } ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 272, + 339, + 285 + ], + "score": 1.0, + "content": "evaluated at", + "type": "text" + }, + { + "bbox": [ + 340, + 274, + 357, + 284 + ], + "score": 0.85, + "content": "2 \\pi n", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 272, + 441, + 285 + ], + "score": 1.0, + "content": ", and its definition is", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 289, + 464, + 321 + ], + "lines": [ + { + "bbox": [ + 147, + 289, + 464, + 321 + ], + "spans": [ + { + "bbox": [ + 147, + 289, + 464, + 321 + ], + "score": 0.93, + "content": "a _ { n } = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { Z } } ( z ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } z } d z = \\frac { 2 A e ^ { \\frac { j 2 \\pi n \\log \\beta } { \\log 1 0 } } } { \\beta c } \\cdot \\Gamma \\Bigg ( \\frac { - j 2 \\pi n + \\log 1 0 } { c \\log 1 0 } \\Bigg ) .", + "type": "interline_equation", + "image_path": "efbb6c783d4614a8299f4216750213fe7c22579c183a5e5cb35dae6b6e055e6b.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 147, + 289, + 464, + 299.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 147, + 299.6666666666667, + 464, + 310.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 147, + 310.33333333333337, + 464, + 321.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 107, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 118, + 335 + ], + "score": 0.81, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 323, + 334, + 337 + ], + "score": 1.0, + "content": "is also called as Benford-Fourier coefficient. Note that", + "type": "text" + }, + { + "bbox": [ + 335, + 326, + 346, + 335 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "is a complex number, and its magnitude", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 336, + 221, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 221, + 347 + ], + "score": 1.0, + "content": "can be calculated as follows", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 349, + 399, + 383 + ], + "lines": [ + { + "bbox": [ + 212, + 349, + 399, + 383 + ], + "spans": [ + { + "bbox": [ + 212, + 349, + 399, + 383 + ], + "score": 0.94, + "content": "| a _ { n } | = \\bigg ( \\prod _ { k = 0 } ^ { + \\infty } \\bigg [ 1 + \\big ( \\frac { 2 \\pi n } { \\log 1 0 ( c k + 1 ) } \\big ) ^ { 2 } \\bigg ] ^ { - 1 } \\bigg ) ^ { \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "0313939955d3e7556cd3ce3a967cf82aad04df7f0b0783863901fa195bb87a35.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 349, + 399, + 366.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 212, + 366.0, + 399, + 383.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 146, + 401 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 147, + 388, + 164, + 400 + ], + "score": 0.92, + "content": "\\left| a _ { n } \\right|", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 388, + 228, + 401 + ], + "score": 1.0, + "content": "gets smaller as", + "type": "text" + }, + { + "bbox": [ + 228, + 388, + 257, + 398 + ], + "score": 0.91, + "content": "n \\in \\mathbb N", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 388, + 430, + 401 + ], + "score": 1.0, + "content": "increases. And, an interesting property of", + "type": "text" + }, + { + "bbox": [ + 430, + 388, + 447, + 400 + ], + "score": 0.91, + "content": "\\left| a _ { n } \\right|", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "is that it only", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 218, + 411 + ], + "score": 1.0, + "content": "depends on the shape factor", + "type": "text" + }, + { + "bbox": [ + 218, + 402, + 224, + 409 + ], + "score": 0.72, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 399, + 378, + 411 + ], + "score": 1.0, + "content": ", while is independent of the parameter", + "type": "text" + }, + { + "bbox": [ + 379, + 401, + 385, + 409 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 399, + 505, + 411 + ], + "score": 1.0, + "content": ". Thus, one set of the absolute", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 409, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 264, + 424 + ], + "score": 1.0, + "content": "values of Benford-Fourier coefficients", + "type": "text" + }, + { + "bbox": [ + 264, + 410, + 307, + 422 + ], + "score": 0.92, + "content": "\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 409, + 422, + 424 + ], + "score": 1.0, + "content": "correspond to one identical", + "type": "text" + }, + { + "bbox": [ + 423, + 412, + 428, + 420 + ], + "score": 0.56, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 409, + 506, + 424 + ], + "score": 1.0, + "content": ", i.e., one identical", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 342, + 434 + ], + "score": 1.0, + "content": "special distribution of GGD. In other words, we could use", + "type": "text" + }, + { + "bbox": [ + 342, + 421, + 385, + 433 + ], + "score": 0.93, + "content": "\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "as features or representations", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 432, + 324, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 324, + 444 + ], + "score": 1.0, + "content": "to discriminate different special distributions of GGD.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 400, + 461 + ], + "score": 1.0, + "content": "However, if it is often difficult to know or even estimate the shape factor", + "type": "text" + }, + { + "bbox": [ + 400, + 452, + 406, + 459 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 448, + 505, + 461 + ], + "score": 1.0, + "content": ", we cannot compute the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 142, + 472 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 142, + 460, + 160, + 472 + ], + "score": 0.91, + "content": "\\left| a _ { n } \\right|", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 459, + 288, + 472 + ], + "score": 1.0, + "content": ". But fortunately, recalling that", + "type": "text" + }, + { + "bbox": [ + 288, + 461, + 300, + 471 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 459, + 326, + 472 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 327, + 461, + 334, + 470 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 459, + 435, + 472 + ], + "score": 1.0, + "content": "-th Fourier coefficient of", + "type": "text" + }, + { + "bbox": [ + 435, + 460, + 463, + 471 + ], + "score": 0.9, + "content": "\\mathcal { P } _ { \\mathcal { Z } } ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "evaluated", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 104, + 470, + 116, + 484 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 117, + 471, + 135, + 481 + ], + "score": 0.85, + "content": "2 \\pi n", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 470, + 382, + 484 + ], + "score": 1.0, + "content": ", we can derive an easy estimation. Specifically, assume that", + "type": "text" + }, + { + "bbox": [ + 382, + 471, + 461, + 483 + ], + "score": 0.91, + "content": "\\mathbf { x } = \\{ x _ { 1 } , \\dots , x _ { M } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "is a set of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 481, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 107, + 482, + 118, + 492 + ], + "score": 0.78, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 481, + 357, + 495 + ], + "score": 1.0, + "content": "i.i.d. points sampled from GGD with the same shape factor", + "type": "text" + }, + { + "bbox": [ + 357, + 484, + 363, + 492 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 481, + 505, + 495 + ], + "score": 1.0, + "content": ". Then, the corresponding Benford-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 492, + 395, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 395, + 505 + ], + "score": 1.0, + "content": "Fourier coefficients can be estimated as follows (Pasquini et al. (2014)):", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 508, + 484, + 542 + ], + "lines": [ + { + "bbox": [ + 115, + 508, + 484, + 542 + ], + "spans": [ + { + "bbox": [ + 115, + 508, + 484, + 542 + ], + "score": 0.94, + "content": "\\hat { a } _ { n } = \\frac { \\sum _ { m = 1 } ^ { M } e ^ { - j 2 \\pi n \\log _ { 1 0 } | x _ { m } | } } { M } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\left[ \\cos \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) - j \\sin \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) \\right] .", + "type": "interline_equation", + "image_path": "d880ea67f8e77aae0d5ceec51b211a6c2889e4952e2f70f61e0eab97324a542e.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 115, + 508, + 484, + 519.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 115, + 519.3333333333334, + 484, + 530.6666666666667 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 115, + 530.6666666666667, + 484, + 542.0000000000001 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 178, + 559 + ], + "score": 1.0, + "content": "The gap between", + "type": "text" + }, + { + "bbox": [ + 179, + 547, + 191, + 558 + ], + "score": 0.89, + "content": "\\hat { a } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 546, + 210, + 559 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 210, + 548, + 222, + 558 + ], + "score": 0.87, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 546, + 381, + 559 + ], + "score": 1.0, + "content": "is analyzed in Theorem 1. It tells that", + "type": "text" + }, + { + "bbox": [ + 382, + 547, + 393, + 558 + ], + "score": 0.88, + "content": "\\hat { a } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 546, + 452, + 559 + ], + "score": 1.0, + "content": "gets closer to", + "type": "text" + }, + { + "bbox": [ + 452, + 549, + 464, + 558 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 546, + 478, + 559 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 478, + 547, + 489, + 556 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "in-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 556, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 104, + 556, + 336, + 571 + ], + "score": 1.0, + "content": "creases. For clarity, we firstly introduce a few notations:", + "type": "text" + }, + { + "bbox": [ + 336, + 558, + 420, + 569 + ], + "score": 0.91, + "content": "\\mathcal { T } = e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } | }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 556, + 506, + 571 + ], + "score": 1.0, + "content": "is a random variable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 570, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 127, + 581 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 570, + 137, + 579 + ], + "score": 0.83, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 570, + 342, + 581 + ], + "score": 1.0, + "content": "obeying the generalized Gaussian distribution, and", + "type": "text" + }, + { + "bbox": [ + 342, + 570, + 354, + 581 + ], + "score": 0.88, + "content": "\\hat { a } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 570, + 505, + 581 + ], + "score": 1.0, + "content": "is an observation of the random vari-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 102, + 577, + 508, + 603 + ], + "spans": [ + { + "bbox": [ + 102, + 577, + 127, + 603 + ], + "score": 1.0, + "content": "able", + "type": "text" + }, + { + "bbox": [ + 127, + 580, + 207, + 596 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { Y } = \\bar { \\frac { 1 } { M } } \\bar { \\sum _ { m = 1 } ^ { M } } \\mathcal { T } _ { m } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 577, + 508, + 603 + ], + "score": 1.0, + "content": ". Due to the space limit, the proof will be presented in supplementary", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 147, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 147, + 604 + ], + "score": 1.0, + "content": "material.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 288, + 628 + ], + "score": 1.0, + "content": "Theorem 1 Assume that the estimation error", + "type": "text" + }, + { + "bbox": [ + 288, + 616, + 342, + 626 + ], + "score": 0.92, + "content": "\\varepsilon _ { n } = { \\hat { a } } _ { n } - a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 613, + 506, + 628 + ], + "score": 1.0, + "content": "is an observation of the random variable", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 624, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 153, + 637 + ], + "score": 0.87, + "content": "\\mathcal { E } = \\mathcal { V } - a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 624, + 157, + 639 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 158, + 626, + 170, + 637 + ], + "score": 0.82, + "content": "| \\mathcal { E } |", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 624, + 505, + 639 + ], + "score": 1.0, + "content": "follows the Rayleigh distribution (Papoulis & Pillai (2001)), of which the probability", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 636, + 268, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 172, + 650 + ], + "score": 1.0, + "content": "density function", + "type": "text" + }, + { + "bbox": [ + 173, + 637, + 200, + 648 + ], + "score": 0.68, + "content": "( P D F )", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 636, + 268, + 650 + ], + "score": 1.0, + "content": "is formulated as", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 652, + 354, + 670 + ], + "lines": [ + { + "bbox": [ + 256, + 652, + 354, + 670 + ], + "spans": [ + { + "bbox": [ + 256, + 652, + 354, + 670 + ], + "score": 0.92, + "content": "\\mathcal { P } _ { | \\mathcal { E } | } \\left( r \\right) = 2 M r e ^ { - M r ^ { 2 } } .", + "type": "interline_equation", + "image_path": "068986864f1121d3f477d0f01fd36df05a07e4cfc6a74658b99c256b28b0ee54.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 256, + 652, + 354, + 670 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 673, + 381, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 672, + 382, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 382, + 686 + ], + "score": 1.0, + "content": "And the expectation and variance are respectively shown as follows:", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 689, + 381, + 717 + ], + "lines": [ + { + "bbox": [ + 229, + 689, + 381, + 717 + ], + "spans": [ + { + "bbox": [ + 229, + 689, + 381, + 717 + ], + "score": 0.95, + "content": "E { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D ( | { \\mathcal { E } } | ) = { \\frac { 4 - \\pi } { 4 M } } ,", + "type": "interline_equation", + "image_path": "c85cd3d17b8682c0a24a8ee94c97ad201648c8c0e1bc95a65758264883d030f4.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 229, + 689, + 381, + 717 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 495, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 497, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 264, + 734 + ], + "score": 1.0, + "content": "which implies that the estimation error", + "type": "text" + }, + { + "bbox": [ + 264, + 725, + 275, + 732 + ], + "score": 0.89, + "content": "\\varepsilon _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 720, + 332, + 734 + ], + "score": 1.0, + "content": "gets closer to", + "type": "text" + }, + { + "bbox": [ + 332, + 723, + 338, + 730 + ], + "score": 0.82, + "content": "\\boldsymbol { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 720, + 497, + 734 + ], + "score": 1.0, + "content": "as the number of samples M increases.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "4", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 297, + 96 + ], + "score": 1.0, + "content": "with varied shape factors. For example, when", + "type": "text" + }, + { + "bbox": [ + 298, + 83, + 324, + 93 + ], + "score": 0.9, + "content": "c = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 81, + 505, + 96 + ], + "score": 1.0, + "content": ", then it becomes the Laplacian distribution;", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 131, + 106 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 131, + 94, + 157, + 104 + ], + "score": 0.89, + "content": "c = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 93, + 380, + 106 + ], + "score": 1.0, + "content": ", then it is the Gaussian distribution with a variance of √ √", + "type": "text" + }, + { + "bbox": [ + 380, + 93, + 392, + 104 + ], + "score": 0.88, + "content": "\\sigma ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 93, + 421, + 106 + ], + "score": 1.0, + "content": "; when", + "type": "text" + }, + { + "bbox": [ + 421, + 95, + 462, + 104 + ], + "score": 0.89, + "content": "c \\to + \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 93, + 505, + 106 + ], + "score": 1.0, + "content": ", then it is", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 321, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 258, + 117 + ], + "score": 1.0, + "content": "specified as a uniform distribution on", + "type": "text" + }, + { + "bbox": [ + 258, + 104, + 316, + 117 + ], + "score": 0.91, + "content": "( - \\sqrt { 2 } \\sigma , \\sqrt { 2 } \\sigma )", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 104, + 321, + 117 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 129, + 281, + 141 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 281, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 281, + 141 + ], + "score": 1.0, + "content": "3.2 BENFORD-FOURIER COEFFICIENTS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "Although generalized Gaussian distribution (GGD) is able to cover a bunch of distributions, it is", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "hard to depict the exact forms of GGD precisely. To this end, we further define a random variable", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 169, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 107, + 172, + 167, + 184 + ], + "score": 0.87, + "content": "\\mathcal { Z } = \\log _ { 1 0 } | \\mathcal { X } |", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 169, + 507, + 186 + ], + "score": 1.0, + "content": "mod 1 for detecting and distinguishing different form of GGD, of which the PDF is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "formulated by means of Fourier Series as Perez-Gonz ´ alez et al. (2007), with the fundamental period ´", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 181, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 164, + 207 + ], + "score": 1.0, + "content": "being fixed as", + "type": "text" + }, + { + "bbox": [ + 164, + 194, + 176, + 204 + ], + "score": 0.82, + "content": "2 \\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 192, + 181, + 207 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 150, + 507, + 207 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 209, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 119, + 209, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 119, + 209, + 479, + 243 + ], + "score": 0.94, + "content": "\\mathcal { P } _ { Z } ( z ) = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor A _ { n } \\cos { ( 2 \\pi n z ) } + B _ { n } \\sin { ( 2 \\pi n z ) } \\rfloor = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor a _ { n } \\rfloor \\cos ( 2 \\pi n z + \\phi _ { n } ) ,", + "type": "interline_equation", + "image_path": "46c52c4fa59c4e2d00b3ed521c2e6873b0ae03399402a4950cbde421f3f6ffb7.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 119, + 209, + 479, + 220.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 119, + 220.33333333333334, + 479, + 231.66666666666669 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 119, + 231.66666666666669, + 479, + 243.00000000000003 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 247, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 134, + 261 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 248, + 176, + 260 + ], + "score": 0.92, + "content": "z \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 246, + 369, + 261 + ], + "score": 1.0, + "content": "corresponds to the domain of random variable", + "type": "text" + }, + { + "bbox": [ + 369, + 248, + 378, + 258 + ], + "score": 0.81, + "content": "\\mathcal { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 246, + 506, + 261 + ], + "score": 1.0, + "content": ", the phase of Fourier Series is", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 102, + 254, + 503, + 279 + ], + "spans": [ + { + "bbox": [ + 102, + 254, + 158, + 279 + ], + "score": 1.0, + "content": "explained as", + "type": "text" + }, + { + "bbox": [ + 158, + 259, + 245, + 274 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\phi _ { n } = \\arctan \\left( - \\frac { B _ { n } } { A _ { n } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 254, + 356, + 279 + ], + "score": 1.0, + "content": ", and the magnitude denotes", + "type": "text" + }, + { + "bbox": [ + 357, + 259, + 435, + 273 + ], + "score": 0.91, + "content": "| a _ { n } | = \\sqrt { A _ { n } ^ { 2 } + B _ { n } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 254, + 439, + 279 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 439, + 259, + 503, + 273 + ], + "score": 0.9, + "content": "a _ { n } = | a _ { n } | \\cdot e ^ { j \\phi _ { n } }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 441, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 154, + 285 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 154, + 275, + 161, + 283 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 272, + 261, + 285 + ], + "score": 1.0, + "content": "-th Fourier coefficient of", + "type": "text" + }, + { + "bbox": [ + 261, + 273, + 288, + 285 + ], + "score": 0.93, + "content": "\\mathcal { P } _ { \\mathcal { Z } } ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 272, + 339, + 285 + ], + "score": 1.0, + "content": "evaluated at", + "type": "text" + }, + { + "bbox": [ + 340, + 274, + 357, + 284 + ], + "score": 0.85, + "content": "2 \\pi n", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 272, + 441, + 285 + ], + "score": 1.0, + "content": ", and its definition is", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 102, + 246, + 506, + 285 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 147, + 289, + 464, + 321 + ], + "lines": [ + { + "bbox": [ + 147, + 289, + 464, + 321 + ], + "spans": [ + { + "bbox": [ + 147, + 289, + 464, + 321 + ], + "score": 0.93, + "content": "a _ { n } = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { Z } } ( z ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } z } d z = \\frac { 2 A e ^ { \\frac { j 2 \\pi n \\log \\beta } { \\log 1 0 } } } { \\beta c } \\cdot \\Gamma \\Bigg ( \\frac { - j 2 \\pi n + \\log 1 0 } { c \\log 1 0 } \\Bigg ) .", + "type": "interline_equation", + "image_path": "efbb6c783d4614a8299f4216750213fe7c22579c183a5e5cb35dae6b6e055e6b.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 147, + 289, + 464, + 299.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 147, + 299.6666666666667, + 464, + 310.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 147, + 310.33333333333337, + 464, + 321.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 107, + 323, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 118, + 335 + ], + "score": 0.81, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 323, + 334, + 337 + ], + "score": 1.0, + "content": "is also called as Benford-Fourier coefficient. Note that", + "type": "text" + }, + { + "bbox": [ + 335, + 326, + 346, + 335 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 323, + 505, + 337 + ], + "score": 1.0, + "content": "is a complex number, and its magnitude", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 336, + 221, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 221, + 347 + ], + "score": 1.0, + "content": "can be calculated as follows", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 323, + 505, + 347 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 349, + 399, + 383 + ], + "lines": [ + { + "bbox": [ + 212, + 349, + 399, + 383 + ], + "spans": [ + { + "bbox": [ + 212, + 349, + 399, + 383 + ], + "score": 0.94, + "content": "| a _ { n } | = \\bigg ( \\prod _ { k = 0 } ^ { + \\infty } \\bigg [ 1 + \\big ( \\frac { 2 \\pi n } { \\log 1 0 ( c k + 1 ) } \\big ) ^ { 2 } \\bigg ] ^ { - 1 } \\bigg ) ^ { \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "0313939955d3e7556cd3ce3a967cf82aad04df7f0b0783863901fa195bb87a35.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 349, + 399, + 366.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 212, + 366.0, + 399, + 383.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 146, + 401 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 147, + 388, + 164, + 400 + ], + "score": 0.92, + "content": "\\left| a _ { n } \\right|", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 388, + 228, + 401 + ], + "score": 1.0, + "content": "gets smaller as", + "type": "text" + }, + { + "bbox": [ + 228, + 388, + 257, + 398 + ], + "score": 0.91, + "content": "n \\in \\mathbb N", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 388, + 430, + 401 + ], + "score": 1.0, + "content": "increases. And, an interesting property of", + "type": "text" + }, + { + "bbox": [ + 430, + 388, + 447, + 400 + ], + "score": 0.91, + "content": "\\left| a _ { n } \\right|", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "is that it only", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 218, + 411 + ], + "score": 1.0, + "content": "depends on the shape factor", + "type": "text" + }, + { + "bbox": [ + 218, + 402, + 224, + 409 + ], + "score": 0.72, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 399, + 378, + 411 + ], + "score": 1.0, + "content": ", while is independent of the parameter", + "type": "text" + }, + { + "bbox": [ + 379, + 401, + 385, + 409 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 399, + 505, + 411 + ], + "score": 1.0, + "content": ". Thus, one set of the absolute", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 409, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 264, + 424 + ], + "score": 1.0, + "content": "values of Benford-Fourier coefficients", + "type": "text" + }, + { + "bbox": [ + 264, + 410, + 307, + 422 + ], + "score": 0.92, + "content": "\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 409, + 422, + 424 + ], + "score": 1.0, + "content": "correspond to one identical", + "type": "text" + }, + { + "bbox": [ + 423, + 412, + 428, + 420 + ], + "score": 0.56, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 409, + 506, + 424 + ], + "score": 1.0, + "content": ", i.e., one identical", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 342, + 434 + ], + "score": 1.0, + "content": "special distribution of GGD. In other words, we could use", + "type": "text" + }, + { + "bbox": [ + 342, + 421, + 385, + 433 + ], + "score": 0.93, + "content": "\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "as features or representations", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 432, + 324, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 324, + 444 + ], + "score": 1.0, + "content": "to discriminate different special distributions of GGD.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 388, + 506, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 448, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 400, + 461 + ], + "score": 1.0, + "content": "However, if it is often difficult to know or even estimate the shape factor", + "type": "text" + }, + { + "bbox": [ + 400, + 452, + 406, + 459 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 448, + 505, + 461 + ], + "score": 1.0, + "content": ", we cannot compute the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 142, + 472 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 142, + 460, + 160, + 472 + ], + "score": 0.91, + "content": "\\left| a _ { n } \\right|", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 459, + 288, + 472 + ], + "score": 1.0, + "content": ". But fortunately, recalling that", + "type": "text" + }, + { + "bbox": [ + 288, + 461, + 300, + 471 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 459, + 326, + 472 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 327, + 461, + 334, + 470 + ], + "score": 0.79, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 459, + 435, + 472 + ], + "score": 1.0, + "content": "-th Fourier coefficient of", + "type": "text" + }, + { + "bbox": [ + 435, + 460, + 463, + 471 + ], + "score": 0.9, + "content": "\\mathcal { P } _ { \\mathcal { Z } } ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "evaluated", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 104, + 470, + 116, + 484 + ], + "score": 1.0, + "content": "at", + "type": "text" + }, + { + "bbox": [ + 117, + 471, + 135, + 481 + ], + "score": 0.85, + "content": "2 \\pi n", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 470, + 382, + 484 + ], + "score": 1.0, + "content": ", we can derive an easy estimation. Specifically, assume that", + "type": "text" + }, + { + "bbox": [ + 382, + 471, + 461, + 483 + ], + "score": 0.91, + "content": "\\mathbf { x } = \\{ x _ { 1 } , \\dots , x _ { M } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "is a set of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 481, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 107, + 482, + 118, + 492 + ], + "score": 0.78, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 481, + 357, + 495 + ], + "score": 1.0, + "content": "i.i.d. points sampled from GGD with the same shape factor", + "type": "text" + }, + { + "bbox": [ + 357, + 484, + 363, + 492 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 481, + 505, + 495 + ], + "score": 1.0, + "content": ". Then, the corresponding Benford-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 492, + 395, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 395, + 505 + ], + "score": 1.0, + "content": "Fourier coefficients can be estimated as follows (Pasquini et al. (2014)):", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 104, + 448, + 506, + 505 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 508, + 484, + 542 + ], + "lines": [ + { + "bbox": [ + 115, + 508, + 484, + 542 + ], + "spans": [ + { + "bbox": [ + 115, + 508, + 484, + 542 + ], + "score": 0.94, + "content": "\\hat { a } _ { n } = \\frac { \\sum _ { m = 1 } ^ { M } e ^ { - j 2 \\pi n \\log _ { 1 0 } | x _ { m } | } } { M } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\left[ \\cos \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) - j \\sin \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) \\right] .", + "type": "interline_equation", + "image_path": "d880ea67f8e77aae0d5ceec51b211a6c2889e4952e2f70f61e0eab97324a542e.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 115, + 508, + 484, + 519.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 115, + 519.3333333333334, + 484, + 530.6666666666667 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 115, + 530.6666666666667, + 484, + 542.0000000000001 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 178, + 559 + ], + "score": 1.0, + "content": "The gap between", + "type": "text" + }, + { + "bbox": [ + 179, + 547, + 191, + 558 + ], + "score": 0.89, + "content": "\\hat { a } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 546, + 210, + 559 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 210, + 548, + 222, + 558 + ], + "score": 0.87, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 546, + 381, + 559 + ], + "score": 1.0, + "content": "is analyzed in Theorem 1. It tells that", + "type": "text" + }, + { + "bbox": [ + 382, + 547, + 393, + 558 + ], + "score": 0.88, + "content": "\\hat { a } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 546, + 452, + 559 + ], + "score": 1.0, + "content": "gets closer to", + "type": "text" + }, + { + "bbox": [ + 452, + 549, + 464, + 558 + ], + "score": 0.86, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 546, + 478, + 559 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 478, + 547, + 489, + 556 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "in-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 556, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 104, + 556, + 336, + 571 + ], + "score": 1.0, + "content": "creases. For clarity, we firstly introduce a few notations:", + "type": "text" + }, + { + "bbox": [ + 336, + 558, + 420, + 569 + ], + "score": 0.91, + "content": "\\mathcal { T } = e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } | }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 556, + 506, + 571 + ], + "score": 1.0, + "content": "is a random variable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 570, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 127, + 581 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 570, + 137, + 579 + ], + "score": 0.83, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 570, + 342, + 581 + ], + "score": 1.0, + "content": "obeying the generalized Gaussian distribution, and", + "type": "text" + }, + { + "bbox": [ + 342, + 570, + 354, + 581 + ], + "score": 0.88, + "content": "\\hat { a } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 570, + 505, + 581 + ], + "score": 1.0, + "content": "is an observation of the random vari-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 102, + 577, + 508, + 603 + ], + "spans": [ + { + "bbox": [ + 102, + 577, + 127, + 603 + ], + "score": 1.0, + "content": "able", + "type": "text" + }, + { + "bbox": [ + 127, + 580, + 207, + 596 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { Y } = \\bar { \\frac { 1 } { M } } \\bar { \\sum _ { m = 1 } ^ { M } } \\mathcal { T } _ { m } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 577, + 508, + 603 + ], + "score": 1.0, + "content": ". Due to the space limit, the proof will be presented in supplementary", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 594, + 147, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 147, + 604 + ], + "score": 1.0, + "content": "material.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 102, + 546, + 508, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 288, + 628 + ], + "score": 1.0, + "content": "Theorem 1 Assume that the estimation error", + "type": "text" + }, + { + "bbox": [ + 288, + 616, + 342, + 626 + ], + "score": 0.92, + "content": "\\varepsilon _ { n } = { \\hat { a } } _ { n } - a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 613, + 506, + 628 + ], + "score": 1.0, + "content": "is an observation of the random variable", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 624, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 153, + 637 + ], + "score": 0.87, + "content": "\\mathcal { E } = \\mathcal { V } - a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 624, + 157, + 639 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 158, + 626, + 170, + 637 + ], + "score": 0.82, + "content": "| \\mathcal { E } |", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 624, + 505, + 639 + ], + "score": 1.0, + "content": "follows the Rayleigh distribution (Papoulis & Pillai (2001)), of which the probability", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 636, + 268, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 172, + 650 + ], + "score": 1.0, + "content": "density function", + "type": "text" + }, + { + "bbox": [ + 173, + 637, + 200, + 648 + ], + "score": 0.68, + "content": "( P D F )", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 636, + 268, + 650 + ], + "score": 1.0, + "content": "is formulated as", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 613, + 506, + 650 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 652, + 354, + 670 + ], + "lines": [ + { + "bbox": [ + 256, + 652, + 354, + 670 + ], + "spans": [ + { + "bbox": [ + 256, + 652, + 354, + 670 + ], + "score": 0.92, + "content": "\\mathcal { P } _ { | \\mathcal { E } | } \\left( r \\right) = 2 M r e ^ { - M r ^ { 2 } } .", + "type": "interline_equation", + "image_path": "068986864f1121d3f477d0f01fd36df05a07e4cfc6a74658b99c256b28b0ee54.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 256, + 652, + 354, + 670 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 673, + 381, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 672, + 382, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 382, + 686 + ], + "score": 1.0, + "content": "And the expectation and variance are respectively shown as follows:", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 672, + 382, + 686 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 689, + 381, + 717 + ], + "lines": [ + { + "bbox": [ + 229, + 689, + 381, + 717 + ], + "spans": [ + { + "bbox": [ + 229, + 689, + 381, + 717 + ], + "score": 0.95, + "content": "E { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D ( | { \\mathcal { E } } | ) = { \\frac { 4 - \\pi } { 4 M } } ,", + "type": "interline_equation", + "image_path": "c85cd3d17b8682c0a24a8ee94c97ad201648c8c0e1bc95a65758264883d030f4.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 229, + 689, + 381, + 717 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 495, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 497, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 264, + 734 + ], + "score": 1.0, + "content": "which implies that the estimation error", + "type": "text" + }, + { + "bbox": [ + 264, + 725, + 275, + 732 + ], + "score": 0.89, + "content": "\\varepsilon _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 720, + 332, + 734 + ], + "score": 1.0, + "content": "gets closer to", + "type": "text" + }, + { + "bbox": [ + 332, + 723, + 338, + 730 + ], + "score": 0.82, + "content": "\\boldsymbol { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 720, + 497, + 734 + ], + "score": 1.0, + "content": "as the number of samples M increases.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 720, + 497, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 105, + 81, + 468, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 470, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 470, + 95 + ], + "score": 1.0, + "content": "4 ADVERSARIAL DETECTION VIA BENFORD-FOURIER COEFFICIENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 106, + 105, + 358, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 360, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 360, + 118 + ], + "score": 1.0, + "content": "4.1 TRAINING PROCEDURE OF ADVERSARIAL DETECTOR", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 503, + 171 + ], + "lines": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "score": 1.0, + "content": "There are three stages to train the proposed adversarial detector, including: 1) building a training set", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "based on benign images; 2) extracting novel representations of the training set via Benford-Fourier", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "coefficients; 3) training a SVM classifier as the adversarial detector. They will be explained in details", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 455, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 455, + 171 + ], + "score": 1.0, + "content": "sequentially. And, the overall training procedure is briefly summarized in Algorithm 1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 267, + 188 + ], + "score": 1.0, + "content": "Build a training set. Firstly, we collect", + "type": "text" + }, + { + "bbox": [ + 267, + 175, + 277, + 185 + ], + "score": 0.75, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 174, + 333, + 188 + ], + "score": 1.0, + "content": "clean images", + "type": "text" + }, + { + "bbox": [ + 333, + 175, + 389, + 187 + ], + "score": 0.92, + "content": "\\left\\{ \\mathbf { x } _ { 1 } , \\dotsc , \\mathbf { x } _ { N } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 174, + 505, + 188 + ], + "score": 1.0, + "content": ", which can be correctly pre-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 145, + 199 + ], + "score": 1.0, + "content": "dicted by", + "type": "text" + }, + { + "bbox": [ + 145, + 186, + 156, + 197 + ], + "score": 0.84, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 185, + 505, + 199 + ], + "score": 1.0, + "content": ". Then, we adopt one adversarial attack method (e.g., C&W (Carlini & Wagner (2017a))", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "or BIM ( Kurakin et al. (2016))) to generate one adversarial image corresponding to each clean im-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 207, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 104, + 207, + 176, + 221 + ], + "score": 1.0, + "content": "age. The crafted", + "type": "text" + }, + { + "bbox": [ + 177, + 208, + 187, + 218 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 207, + 337, + 221 + ], + "score": 1.0, + "content": "adversarial examples are denoted as", + "type": "text" + }, + { + "bbox": [ + 337, + 208, + 394, + 220 + ], + "score": 0.93, + "content": "\\big \\{ \\hat { \\mathbf { x } } _ { 1 } , \\dotsc , \\hat { \\mathbf { x } } _ { N } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 207, + 506, + 221 + ], + "score": 1.0, + "content": ". Besides, to avoid that the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "noisy image (polluted by some kind of non-malicious noises but still can be correctly predicted", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 119, + 243 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 230, + 132, + 241 + ], + "score": 0.79, + "content": "f _ { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "is incorrectly detected as adversarial, we also craft one noisy image by adding small random", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 240, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 290, + 254 + ], + "score": 1.0, + "content": "Gaussian noises onto each clean image. These", + "type": "text" + }, + { + "bbox": [ + 290, + 241, + 300, + 250 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 240, + 422, + 254 + ], + "score": 1.0, + "content": "noisy examples are denoted as", + "type": "text" + }, + { + "bbox": [ + 422, + 241, + 479, + 253 + ], + "score": 0.93, + "content": "\\left\\{ bar { \\mathbf { x } } _ { 1 } , \\ldots , \\bar { \\mathbf { x } } _ { N } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 240, + 505, + 254 + ], + "score": 1.0, + "content": ". Note", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "that, hereafter benign examples include both clean and Gaussian noisy examples. Consequently, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 261, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 216, + 277 + ], + "score": 1.0, + "content": "obtain one training set with", + "type": "text" + }, + { + "bbox": [ + 216, + 263, + 231, + 272 + ], + "score": 0.81, + "content": "3 N", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 261, + 318, + 277 + ], + "score": 1.0, + "content": "examples, denoted as", + "type": "text" + }, + { + "bbox": [ + 318, + 262, + 501, + 275 + ], + "score": 0.91, + "content": "\\mathcal { D } _ { t r } = \\{ ( \\mathbf { x } _ { i } , - 1 \\bar { ) } , ( \\bar { \\mathbf { x } } _ { i } , \\bar { { ~ - } 1 } ) , ( \\hat { \\mathbf { x } } _ { i } , + 1 \\bar { ) } \\} _ { i = 1 , \\ldots , N }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 261, + 506, + 277 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 326, + 291 + ], + "score": 1.0, + "content": "Extract novel representations. We firstly feed the", + "type": "text" + }, + { + "bbox": [ + 326, + 280, + 330, + 289 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 279, + 430, + 291 + ], + "score": 1.0, + "content": "-th training image from", + "type": "text" + }, + { + "bbox": [ + 431, + 279, + 447, + 290 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 469, + 279, + 480, + 290 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 279, + 505, + 291 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 259, + 302 + ], + "score": 1.0, + "content": "concatenate all response entries of the", + "type": "text" + }, + { + "bbox": [ + 260, + 291, + 264, + 299 + ], + "score": 0.7, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 289, + 308, + 302 + ], + "score": 1.0, + "content": "-th layer in", + "type": "text" + }, + { + "bbox": [ + 309, + 290, + 319, + 301 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 289, + 402, + 302 + ], + "score": 1.0, + "content": "to obtain one vector", + "type": "text" + }, + { + "bbox": [ + 402, + 289, + 411, + 302 + ], + "score": 0.88, + "content": "\\mathbf { r } _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 289, + 505, + 302 + ], + "score": 1.0, + "content": ". Then, we estimate the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 410, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 410, + 312 + ], + "score": 1.0, + "content": "corresponding Benford-Fourier coefficients according to Eq. (5), as follows", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 314, + 380, + 351 + ], + "lines": [ + { + "bbox": [ + 230, + 314, + 380, + 351 + ], + "spans": [ + { + "bbox": [ + 230, + 314, + 380, + 351 + ], + "score": 0.95, + "content": "( \\hat { a } _ { n } ) _ { i } ^ { l } = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } e ^ { - j 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | } ,", + "type": "interline_equation", + "image_path": "b0c2baba103458e9fc8c223b8296cf5db64662c047e8b26952024979114e50ab.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 314, + 380, + 332.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 230, + 332.5, + 380, + 351.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 444, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 446, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 133, + 368 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 353, + 148, + 366 + ], + "score": 0.91, + "content": "M _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 352, + 240, + 368 + ], + "score": 1.0, + "content": "indicates the length of", + "type": "text" + }, + { + "bbox": [ + 240, + 354, + 249, + 366 + ], + "score": 0.88, + "content": "\\mathbf { r } _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 352, + 328, + 368 + ], + "score": 1.0, + "content": ". The magnitude of", + "type": "text" + }, + { + "bbox": [ + 328, + 353, + 350, + 367 + ], + "score": 0.93, + "content": "( \\hat { a } _ { n } ) _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 352, + 446, + 368 + ], + "score": 1.0, + "content": "is computed as follows", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 369, + 462, + 406 + ], + "lines": [ + { + "bbox": [ + 149, + 369, + 462, + 406 + ], + "spans": [ + { + "bbox": [ + 149, + 369, + 462, + 406 + ], + "score": 0.94, + "content": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } | = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } \\Bigg ( \\cos ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) + \\sin ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) \\Bigg ) ^ { \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "d664b67c28464c0f5bfdd1180d1f04d25ee0d5d13fedc8d6f50866f80533c8cd.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 149, + 369, + 462, + 381.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 149, + 381.3333333333333, + 462, + 393.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 149, + 393.66666666666663, + 462, + 405.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 413, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 195, + 427 + ], + "score": 1.0, + "content": "Then, we extract one", + "type": "text" + }, + { + "bbox": [ + 196, + 415, + 204, + 424 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 413, + 317, + 427 + ], + "score": 1.0, + "content": "-dimensional feature vector", + "type": "text" + }, + { + "bbox": [ + 318, + 413, + 455, + 427 + ], + "score": 0.93, + "content": "\\mathbf { a } _ { i } ^ { l } = [ | ( \\hat { a } _ { 1 } ) _ { i } ^ { l } | , \\dots , | ( \\hat { a } _ { T } ) _ { i } ^ { l } | ] \\in \\mathbb { R } _ { + } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 413, + 488, + 427 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 488, + 415, + 493, + 424 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 413, + 505, + 427 + ], + "score": 1.0, + "content": "-th", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 426, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 203, + 441 + ], + "score": 1.0, + "content": "training image from the", + "type": "text" + }, + { + "bbox": [ + 203, + 428, + 208, + 437 + ], + "score": 0.62, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 426, + 274, + 441 + ], + "score": 1.0, + "content": "-th layer. We set", + "type": "text" + }, + { + "bbox": [ + 274, + 427, + 306, + 437 + ], + "score": 0.89, + "content": "T = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 426, + 382, + 441 + ], + "score": 1.0, + "content": "in experiments, as", + "type": "text" + }, + { + "bbox": [ + 383, + 426, + 411, + 439 + ], + "score": 0.91, + "content": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } |", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 426, + 448, + 441 + ], + "score": 1.0, + "content": "of larger", + "type": "text" + }, + { + "bbox": [ + 448, + 429, + 455, + 437 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 426, + 506, + 441 + ], + "score": 1.0, + "content": "is too small", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "for discrimination. Finally, we concatenate the feature vectors of all layers to form a long vector", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 446, + 508, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 216, + 462 + ], + "score": 0.91, + "content": "\\hat { \\mathbf { a } } _ { i } = [ \\hat { \\mathbf { a } } _ { i } ^ { 1 } ; \\hdots ; \\hat { \\mathbf { a } } _ { i } ^ { L } ] \\in \\mathbb { R } _ { + } ^ { T \\check { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 446, + 242, + 464 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 242, + 450, + 250, + 459 + ], + "score": 0.8, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 446, + 376, + 464 + ], + "score": 1.0, + "content": "being the number of layers in", + "type": "text" + }, + { + "bbox": [ + 377, + 450, + 388, + 460 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 446, + 508, + 464 + ], + "score": 1.0, + "content": ". Consequently, we obtain a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 461, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 262, + 476 + ], + "score": 1.0, + "content": "novel representation of the training set", + "type": "text" + }, + { + "bbox": [ + 262, + 462, + 279, + 474 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 462, + 327, + 476 + ], + "score": 1.0, + "content": ", denoted as", + "type": "text" + }, + { + "bbox": [ + 328, + 461, + 438, + 475 + ], + "score": 0.92, + "content": "\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\bf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , 3 N }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 462, + 506, + 476 + ], + "score": 1.0, + "content": ", where the label", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 315, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 117, + 487 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 475, + 127, + 485 + ], + "score": 0.85, + "content": "{ \\bf a } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 473, + 149, + 487 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 149, + 474, + 163, + 485 + ], + "score": 0.84, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 473, + 175, + 487 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 175, + 474, + 190, + 485 + ], + "score": 0.76, + "content": "^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 473, + 294, + 487 + ], + "score": 1.0, + "content": ") is directly obtained from", + "type": "text" + }, + { + "bbox": [ + 294, + 474, + 310, + 485 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 473, + 315, + 487 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 503, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 504, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 436, + 502 + ], + "score": 1.0, + "content": "Train an adversarial detector. Finally, we train a binary SVM classifier based on", + "type": "text" + }, + { + "bbox": [ + 436, + 489, + 452, + 502 + ], + "score": 0.91, + "content": "\\hat { A } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 489, + 504, + 502 + ], + "score": 1.0, + "content": ". The trained", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 500, + 410, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 396, + 515 + ], + "score": 1.0, + "content": "SVM classifier will serve as the adversarial detector for the CNN model", + "type": "text" + }, + { + "bbox": [ + 396, + 502, + 406, + 514 + ], + "score": 0.84, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 500, + 410, + 515 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 518, + 503, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "Testing. One novel testing example is firstly predicted as adversarial or not by the trained adversarial", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 528, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 372, + 541 + ], + "score": 1.0, + "content": "detector . If adversarial, then it is rejected; otherwise, it is fed into", + "type": "text" + }, + { + "bbox": [ + 372, + 529, + 383, + 541 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 528, + 482, + 541 + ], + "score": 1.0, + "content": "to predict its class label.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 269, + 559 + ], + "score": 1.0, + "content": "Remark. Note that in the derivation of", + "type": "text" + }, + { + "bbox": [ + 270, + 548, + 281, + 557 + ], + "score": 0.85, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "(see Eq. (3)), the mean parameter of GGD is set to 0.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "In experiments, we calculate the mean values of internal-layer responses of all networks for every", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "image, and find that mean parameters at most layers are close to 0, while the mean parameters", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 593 + ], + "score": 1.0, + "content": "at a few layers could be large. However, the mean value is subtracted from each response entry", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 410, + 602 + ], + "score": 1.0, + "content": "when we extract the MBF features in our experiments. Thus, the derived", + "type": "text" + }, + { + "bbox": [ + 410, + 591, + 422, + 601 + ], + "score": 0.85, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "is applicable to our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 369, + 614 + ], + "score": 1.0, + "content": "task. Besides, according to Theorem 1, the estimation error of", + "type": "text" + }, + { + "bbox": [ + 369, + 600, + 392, + 613 + ], + "score": 0.93, + "content": "( \\hat { a } _ { n } ) _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "is inversely proportional to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 121, + 625 + ], + "score": 0.9, + "content": "M _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 612, + 505, + 626 + ], + "score": 1.0, + "content": ". It tells that the coefficient estimated from the larger-sized layer is more accurate. In many", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "neural networks (e.g., AlexNet (Krizhevsky et al. (2012))), the response sizes of high layers get", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 636, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 646 + ], + "score": 1.0, + "content": "smaller, which means the less accurate estimation. However, we believe that the discrimination", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "between adversarial and benign examples in higher layers is more evident than that in lower layers.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "There is a trade-off between estimation accuracy and discrimination. This is why we concatenate", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 668, + 445, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 445, + 680 + ], + "score": 1.0, + "content": "the estimated magnitudes of all layers together to construct the novel representation.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 108, + 692, + 244, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 245, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 245, + 704 + ], + "score": 1.0, + "content": "4.2 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "Databases and network architectures We conduct experiments on three databases, including", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 718, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 506, + 735 + ], + "score": 1.0, + "content": "CIFAR-10 (Krizhevsky et al. (2014)), SVHN (Netzer et al. (2011)), and a subset of ImageNet (Deng", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 105, + 81, + 468, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 470, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 470, + 95 + ], + "score": 1.0, + "content": "4 ADVERSARIAL DETECTION VIA BENFORD-FOURIER COEFFICIENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 106, + 105, + 358, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 360, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 360, + 118 + ], + "score": 1.0, + "content": "4.1 TRAINING PROCEDURE OF ADVERSARIAL DETECTOR", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 503, + 171 + ], + "lines": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "score": 1.0, + "content": "There are three stages to train the proposed adversarial detector, including: 1) building a training set", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "based on benign images; 2) extracting novel representations of the training set via Benford-Fourier", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "coefficients; 3) training a SVM classifier as the adversarial detector. They will be explained in details", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 455, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 455, + 171 + ], + "score": 1.0, + "content": "sequentially. And, the overall training procedure is briefly summarized in Algorithm 1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 125, + 505, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 267, + 188 + ], + "score": 1.0, + "content": "Build a training set. Firstly, we collect", + "type": "text" + }, + { + "bbox": [ + 267, + 175, + 277, + 185 + ], + "score": 0.75, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 174, + 333, + 188 + ], + "score": 1.0, + "content": "clean images", + "type": "text" + }, + { + "bbox": [ + 333, + 175, + 389, + 187 + ], + "score": 0.92, + "content": "\\left\\{ \\mathbf { x } _ { 1 } , \\dotsc , \\mathbf { x } _ { N } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 174, + 505, + 188 + ], + "score": 1.0, + "content": ", which can be correctly pre-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 145, + 199 + ], + "score": 1.0, + "content": "dicted by", + "type": "text" + }, + { + "bbox": [ + 145, + 186, + 156, + 197 + ], + "score": 0.84, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 185, + 505, + 199 + ], + "score": 1.0, + "content": ". Then, we adopt one adversarial attack method (e.g., C&W (Carlini & Wagner (2017a))", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "or BIM ( Kurakin et al. (2016))) to generate one adversarial image corresponding to each clean im-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 207, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 104, + 207, + 176, + 221 + ], + "score": 1.0, + "content": "age. The crafted", + "type": "text" + }, + { + "bbox": [ + 177, + 208, + 187, + 218 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 207, + 337, + 221 + ], + "score": 1.0, + "content": "adversarial examples are denoted as", + "type": "text" + }, + { + "bbox": [ + 337, + 208, + 394, + 220 + ], + "score": 0.93, + "content": "\\big \\{ \\hat { \\mathbf { x } } _ { 1 } , \\dotsc , \\hat { \\mathbf { x } } _ { N } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 207, + 506, + 221 + ], + "score": 1.0, + "content": ". Besides, to avoid that the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "noisy image (polluted by some kind of non-malicious noises but still can be correctly predicted", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 119, + 243 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 230, + 132, + 241 + ], + "score": 0.79, + "content": "f _ { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "is incorrectly detected as adversarial, we also craft one noisy image by adding small random", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 240, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 290, + 254 + ], + "score": 1.0, + "content": "Gaussian noises onto each clean image. These", + "type": "text" + }, + { + "bbox": [ + 290, + 241, + 300, + 250 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 240, + 422, + 254 + ], + "score": 1.0, + "content": "noisy examples are denoted as", + "type": "text" + }, + { + "bbox": [ + 422, + 241, + 479, + 253 + ], + "score": 0.93, + "content": "\\left\\{ bar { \\mathbf { x } } _ { 1 } , \\ldots , \\bar { \\mathbf { x } } _ { N } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 240, + 505, + 254 + ], + "score": 1.0, + "content": ". Note", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "that, hereafter benign examples include both clean and Gaussian noisy examples. Consequently, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 261, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 216, + 277 + ], + "score": 1.0, + "content": "obtain one training set with", + "type": "text" + }, + { + "bbox": [ + 216, + 263, + 231, + 272 + ], + "score": 0.81, + "content": "3 N", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 261, + 318, + 277 + ], + "score": 1.0, + "content": "examples, denoted as", + "type": "text" + }, + { + "bbox": [ + 318, + 262, + 501, + 275 + ], + "score": 0.91, + "content": "\\mathcal { D } _ { t r } = \\{ ( \\mathbf { x } _ { i } , - 1 \\bar { ) } , ( \\bar { \\mathbf { x } } _ { i } , \\bar { { ~ - } 1 } ) , ( \\hat { \\mathbf { x } } _ { i } , + 1 \\bar { ) } \\} _ { i = 1 , \\ldots , N }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 261, + 506, + 277 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 174, + 506, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 326, + 291 + ], + "score": 1.0, + "content": "Extract novel representations. We firstly feed the", + "type": "text" + }, + { + "bbox": [ + 326, + 280, + 330, + 289 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 279, + 430, + 291 + ], + "score": 1.0, + "content": "-th training image from", + "type": "text" + }, + { + "bbox": [ + 431, + 279, + 447, + 290 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 469, + 279, + 480, + 290 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 279, + 505, + 291 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 259, + 302 + ], + "score": 1.0, + "content": "concatenate all response entries of the", + "type": "text" + }, + { + "bbox": [ + 260, + 291, + 264, + 299 + ], + "score": 0.7, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 289, + 308, + 302 + ], + "score": 1.0, + "content": "-th layer in", + "type": "text" + }, + { + "bbox": [ + 309, + 290, + 319, + 301 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 289, + 402, + 302 + ], + "score": 1.0, + "content": "to obtain one vector", + "type": "text" + }, + { + "bbox": [ + 402, + 289, + 411, + 302 + ], + "score": 0.88, + "content": "\\mathbf { r } _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 289, + 505, + 302 + ], + "score": 1.0, + "content": ". Then, we estimate the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 300, + 410, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 410, + 312 + ], + "score": 1.0, + "content": "corresponding Benford-Fourier coefficients according to Eq. (5), as follows", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 279, + 505, + 312 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 314, + 380, + 351 + ], + "lines": [ + { + "bbox": [ + 230, + 314, + 380, + 351 + ], + "spans": [ + { + "bbox": [ + 230, + 314, + 380, + 351 + ], + "score": 0.95, + "content": "( \\hat { a } _ { n } ) _ { i } ^ { l } = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } e ^ { - j 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | } ,", + "type": "interline_equation", + "image_path": "b0c2baba103458e9fc8c223b8296cf5db64662c047e8b26952024979114e50ab.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 314, + 380, + 332.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 230, + 332.5, + 380, + 351.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 444, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 446, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 133, + 368 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 353, + 148, + 366 + ], + "score": 0.91, + "content": "M _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 352, + 240, + 368 + ], + "score": 1.0, + "content": "indicates the length of", + "type": "text" + }, + { + "bbox": [ + 240, + 354, + 249, + 366 + ], + "score": 0.88, + "content": "\\mathbf { r } _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 352, + 328, + 368 + ], + "score": 1.0, + "content": ". The magnitude of", + "type": "text" + }, + { + "bbox": [ + 328, + 353, + 350, + 367 + ], + "score": 0.93, + "content": "( \\hat { a } _ { n } ) _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 352, + 446, + 368 + ], + "score": 1.0, + "content": "is computed as follows", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 352, + 446, + 368 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 369, + 462, + 406 + ], + "lines": [ + { + "bbox": [ + 149, + 369, + 462, + 406 + ], + "spans": [ + { + "bbox": [ + 149, + 369, + 462, + 406 + ], + "score": 0.94, + "content": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } | = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } \\Bigg ( \\cos ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) + \\sin ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) \\Bigg ) ^ { \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "d664b67c28464c0f5bfdd1180d1f04d25ee0d5d13fedc8d6f50866f80533c8cd.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 149, + 369, + 462, + 381.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 149, + 381.3333333333333, + 462, + 393.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 149, + 393.66666666666663, + 462, + 405.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 413, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 195, + 427 + ], + "score": 1.0, + "content": "Then, we extract one", + "type": "text" + }, + { + "bbox": [ + 196, + 415, + 204, + 424 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 413, + 317, + 427 + ], + "score": 1.0, + "content": "-dimensional feature vector", + "type": "text" + }, + { + "bbox": [ + 318, + 413, + 455, + 427 + ], + "score": 0.93, + "content": "\\mathbf { a } _ { i } ^ { l } = [ | ( \\hat { a } _ { 1 } ) _ { i } ^ { l } | , \\dots , | ( \\hat { a } _ { T } ) _ { i } ^ { l } | ] \\in \\mathbb { R } _ { + } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 413, + 488, + 427 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 488, + 415, + 493, + 424 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 413, + 505, + 427 + ], + "score": 1.0, + "content": "-th", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 426, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 203, + 441 + ], + "score": 1.0, + "content": "training image from the", + "type": "text" + }, + { + "bbox": [ + 203, + 428, + 208, + 437 + ], + "score": 0.62, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 426, + 274, + 441 + ], + "score": 1.0, + "content": "-th layer. We set", + "type": "text" + }, + { + "bbox": [ + 274, + 427, + 306, + 437 + ], + "score": 0.89, + "content": "T = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 426, + 382, + 441 + ], + "score": 1.0, + "content": "in experiments, as", + "type": "text" + }, + { + "bbox": [ + 383, + 426, + 411, + 439 + ], + "score": 0.91, + "content": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } |", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 426, + 448, + 441 + ], + "score": 1.0, + "content": "of larger", + "type": "text" + }, + { + "bbox": [ + 448, + 429, + 455, + 437 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 426, + 506, + 441 + ], + "score": 1.0, + "content": "is too small", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "for discrimination. Finally, we concatenate the feature vectors of all layers to form a long vector", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 446, + 508, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 216, + 462 + ], + "score": 0.91, + "content": "\\hat { \\mathbf { a } } _ { i } = [ \\hat { \\mathbf { a } } _ { i } ^ { 1 } ; \\hdots ; \\hat { \\mathbf { a } } _ { i } ^ { L } ] \\in \\mathbb { R } _ { + } ^ { T \\check { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 446, + 242, + 464 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 242, + 450, + 250, + 459 + ], + "score": 0.8, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 446, + 376, + 464 + ], + "score": 1.0, + "content": "being the number of layers in", + "type": "text" + }, + { + "bbox": [ + 377, + 450, + 388, + 460 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 446, + 508, + 464 + ], + "score": 1.0, + "content": ". Consequently, we obtain a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 461, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 262, + 476 + ], + "score": 1.0, + "content": "novel representation of the training set", + "type": "text" + }, + { + "bbox": [ + 262, + 462, + 279, + 474 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 462, + 327, + 476 + ], + "score": 1.0, + "content": ", denoted as", + "type": "text" + }, + { + "bbox": [ + 328, + 461, + 438, + 475 + ], + "score": 0.92, + "content": "\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\bf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , 3 N }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 462, + 506, + 476 + ], + "score": 1.0, + "content": ", where the label", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 315, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 117, + 487 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 475, + 127, + 485 + ], + "score": 0.85, + "content": "{ \\bf a } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 473, + 149, + 487 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 149, + 474, + 163, + 485 + ], + "score": 0.84, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 473, + 175, + 487 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 175, + 474, + 190, + 485 + ], + "score": 0.76, + "content": "^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 473, + 294, + 487 + ], + "score": 1.0, + "content": ") is directly obtained from", + "type": "text" + }, + { + "bbox": [ + 294, + 474, + 310, + 485 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 473, + 315, + 487 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 413, + 508, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 503, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 504, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 436, + 502 + ], + "score": 1.0, + "content": "Train an adversarial detector. Finally, we train a binary SVM classifier based on", + "type": "text" + }, + { + "bbox": [ + 436, + 489, + 452, + 502 + ], + "score": 0.91, + "content": "\\hat { A } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 489, + 504, + 502 + ], + "score": 1.0, + "content": ". The trained", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 500, + 410, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 396, + 515 + ], + "score": 1.0, + "content": "SVM classifier will serve as the adversarial detector for the CNN model", + "type": "text" + }, + { + "bbox": [ + 396, + 502, + 406, + 514 + ], + "score": 0.84, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 500, + 410, + 515 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 489, + 504, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 518, + 503, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "Testing. One novel testing example is firstly predicted as adversarial or not by the trained adversarial", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 528, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 372, + 541 + ], + "score": 1.0, + "content": "detector . If adversarial, then it is rejected; otherwise, it is fed into", + "type": "text" + }, + { + "bbox": [ + 372, + 529, + 383, + 541 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 528, + 482, + 541 + ], + "score": 1.0, + "content": "to predict its class label.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 517, + 505, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 269, + 559 + ], + "score": 1.0, + "content": "Remark. Note that in the derivation of", + "type": "text" + }, + { + "bbox": [ + 270, + 548, + 281, + 557 + ], + "score": 0.85, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "(see Eq. (3)), the mean parameter of GGD is set to 0.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "In experiments, we calculate the mean values of internal-layer responses of all networks for every", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "image, and find that mean parameters at most layers are close to 0, while the mean parameters", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 593 + ], + "score": 1.0, + "content": "at a few layers could be large. However, the mean value is subtracted from each response entry", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 410, + 602 + ], + "score": 1.0, + "content": "when we extract the MBF features in our experiments. Thus, the derived", + "type": "text" + }, + { + "bbox": [ + 410, + 591, + 422, + 601 + ], + "score": 0.85, + "content": "a _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "is applicable to our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 369, + 614 + ], + "score": 1.0, + "content": "task. Besides, according to Theorem 1, the estimation error of", + "type": "text" + }, + { + "bbox": [ + 369, + 600, + 392, + 613 + ], + "score": 0.93, + "content": "( \\hat { a } _ { n } ) _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "is inversely proportional to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 612, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 121, + 625 + ], + "score": 0.9, + "content": "M _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 612, + 505, + 626 + ], + "score": 1.0, + "content": ". It tells that the coefficient estimated from the larger-sized layer is more accurate. In many", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "neural networks (e.g., AlexNet (Krizhevsky et al. (2012))), the response sizes of high layers get", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 636, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 646 + ], + "score": 1.0, + "content": "smaller, which means the less accurate estimation. However, we believe that the discrimination", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "between adversarial and benign examples in higher layers is more evident than that in lower layers.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "There is a trade-off between estimation accuracy and discrimination. This is why we concatenate", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 668, + 445, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 445, + 680 + ], + "score": 1.0, + "content": "the estimated magnitudes of all layers together to construct the novel representation.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 546, + 506, + 680 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 692, + 244, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 245, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 245, + 704 + ], + "score": 1.0, + "content": "4.2 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "Databases and network architectures We conduct experiments on three databases, including", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 718, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 506, + 735 + ], + "score": 1.0, + "content": "CIFAR-10 (Krizhevsky et al. (2014)), SVHN (Netzer et al. (2011)), and a subset of ImageNet (Deng", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 708, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 500, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 500, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 500, + 96 + ], + "score": 1.0, + "content": "Algorithm 1 Training the adversarial detector via the magnitude of Benford-Fourier coefficients.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 385, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 384, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 234, + 109 + ], + "score": 1.0, + "content": "Require: The trained CNN model", + "type": "text" + }, + { + "bbox": [ + 234, + 96, + 244, + 106 + ], + "score": 0.85, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 94, + 263, + 109 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 264, + 96, + 271, + 105 + ], + "score": 0.78, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 94, + 369, + 109 + ], + "score": 1.0, + "content": "layers, and the training set", + "type": "text" + }, + { + "bbox": [ + 369, + 96, + 384, + 106 + ], + "score": 0.85, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 110, + 105, + 449, + 210 + ], + "lines": [ + { + "bbox": [ + 109, + 103, + 203, + 117 + ], + "spans": [ + { + "bbox": [ + 109, + 103, + 135, + 117 + ], + "score": 1.0, + "content": "1: for", + "type": "text" + }, + { + "bbox": [ + 135, + 106, + 156, + 115 + ], + "score": 0.89, + "content": "i = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 103, + 167, + 117 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 167, + 106, + 187, + 116 + ], + "score": 0.89, + "content": "| \\mathcal { D } _ { t r } |", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 103, + 203, + 117 + ], + "score": 1.0, + "content": "do:", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 114, + 203, + 126 + ], + "spans": [ + { + "bbox": [ + 109, + 114, + 120, + 126 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 134, + 115, + 148, + 126 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 149, + 116, + 169, + 125 + ], + "score": 0.88, + "content": "l = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 115, + 180, + 126 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 180, + 117, + 187, + 124 + ], + "score": 0.73, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 115, + 203, + 126 + ], + "score": 1.0, + "content": "do:", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 124, + 327, + 137 + ], + "spans": [ + { + "bbox": [ + 109, + 124, + 120, + 137 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 148, + 124, + 184, + 137 + ], + "score": 1.0, + "content": "Compute", + "type": "text" + }, + { + "bbox": [ + 184, + 125, + 210, + 137 + ], + "score": 0.92, + "content": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } |", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 124, + 269, + 137 + ], + "score": 1.0, + "content": "as Eq. (8), with", + "type": "text" + }, + { + "bbox": [ + 270, + 126, + 321, + 136 + ], + "score": 0.9, + "content": "n = 1 , \\dots , T", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 124, + 327, + 137 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 134, + 344, + 149 + ], + "spans": [ + { + "bbox": [ + 109, + 136, + 120, + 147 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 147, + 134, + 195, + 149 + ], + "score": 1.0, + "content": "Concatenate", + "type": "text" + }, + { + "bbox": [ + 196, + 136, + 262, + 148 + ], + "score": 0.92, + "content": "\\{ | ( \\hat { a } _ { n } ) _ { i } ^ { l } | \\} _ { n = 1 , \\dots , T }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 134, + 327, + 149 + ], + "score": 1.0, + "content": "to obtain a vector", + "type": "text" + }, + { + "bbox": [ + 328, + 136, + 338, + 147 + ], + "score": 0.87, + "content": "\\hat { \\mathbf { a } } _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 134, + 344, + 149 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 146, + 165, + 157 + ], + "spans": [ + { + "bbox": [ + 109, + 146, + 120, + 157 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 135, + 146, + 165, + 157 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 155, + 330, + 170 + ], + "spans": [ + { + "bbox": [ + 109, + 156, + 120, + 168 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 133, + 155, + 182, + 170 + ], + "score": 1.0, + "content": "Concatenate", + "type": "text" + }, + { + "bbox": [ + 182, + 156, + 229, + 168 + ], + "score": 0.92, + "content": "\\{ \\hat { \\mathbf { a } } _ { i } ^ { l } \\} _ { l = 1 , \\dots , L }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 155, + 313, + 170 + ], + "score": 1.0, + "content": "to obtain a long vector", + "type": "text" + }, + { + "bbox": [ + 313, + 157, + 322, + 167 + ], + "score": 0.83, + "content": "\\hat { \\mathbf { a } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 155, + 330, + 170 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 165, + 153, + 178 + ], + "spans": [ + { + "bbox": [ + 108, + 165, + 153, + 178 + ], + "score": 1.0, + "content": "7: end for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 173, + 449, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 173, + 333, + 190 + ], + "score": 1.0, + "content": "8: Build a novel representation of the training set, denoted as", + "type": "text" + }, + { + "bbox": [ + 333, + 175, + 443, + 188 + ], + "score": 0.88, + "content": "\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\mathbf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , | \\mathcal { D } _ { t r } | }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 173, + 449, + 190 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 109, + 187, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 109, + 187, + 264, + 200 + ], + "score": 1.0, + "content": "9: Train a binary SVM classifier based on", + "type": "text" + }, + { + "bbox": [ + 264, + 187, + 280, + 199 + ], + "score": 0.89, + "content": "\\hat { A } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 187, + 286, + 200 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 120, + 199, + 274, + 210 + ], + "spans": [ + { + "bbox": [ + 120, + 199, + 274, + 210 + ], + "score": 1.0, + "content": "return The trained binary SVM classifier.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "score": 1.0, + "content": "et al. (2009)). In terms of CIFAR-10 and SVHN, we adopt the same settings as the compared method", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "LID (Ma et al. (2018)). Specifically, a 33-layer network pre-trained on the training set of CIFAR-10", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 257, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 143, + 273 + ], + "score": 1.0, + "content": "achieves", + "type": "text" + }, + { + "bbox": [ + 144, + 259, + 176, + 270 + ], + "score": 0.89, + "content": "8 2 . 3 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 257, + 505, + 273 + ], + "score": 1.0, + "content": "accuracy on the testing set with 10, 000 benign images; a 19-layer network pre-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 287, + 283 + ], + "score": 1.0, + "content": "trained on the training set of SVHN achieves", + "type": "text" + }, + { + "bbox": [ + 288, + 270, + 315, + 281 + ], + "score": 0.88, + "content": "9 2 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "accuracy on the testing set with 26, 032 benign", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 303, + 293 + ], + "score": 1.0, + "content": "images. Then, we add a small noise drawn from", + "type": "text" + }, + { + "bbox": [ + 303, + 281, + 342, + 293 + ], + "score": 0.93, + "content": "{ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 281, + 456, + 293 + ], + "score": 1.0, + "content": "on each testing image, with", + "type": "text" + }, + { + "bbox": [ + 456, + 282, + 464, + 291 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "being the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 182, + 304 + ], + "score": 1.0, + "content": "similar level of the", + "type": "text" + }, + { + "bbox": [ + 182, + 292, + 192, + 303 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "norm of adversarial perturbations on the same database. If both the benign and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "its noisy image can be correctly predicted by the classification network, then it is picked out for train-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "ing the detection. We finally collect 8, 175 and 23, 862 benign images from CIFATR-10 and SVHN,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 324, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 324, + 340, + 339 + ], + "score": 1.0, + "content": "respectively. These images are randomly partitioned to the", + "type": "text" + }, + { + "bbox": [ + 341, + 325, + 361, + 336 + ], + "score": 0.89, + "content": "8 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 324, + 439, + 339 + ], + "score": 1.0, + "content": "training set and the", + "type": "text" + }, + { + "bbox": [ + 439, + 325, + 459, + 335 + ], + "score": 0.89, + "content": "2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 324, + 506, + 339 + ], + "score": 1.0, + "content": "testing set,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "used for the training and testing of the detector. We also collect a subset from ImageNet, including", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 346, + 504, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 504, + 360 + ], + "score": 1.0, + "content": "800 benign images of 8 classes (snowbird, spoonbill, bobtail, Leonberg, hamster, proboscis monkey,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "score": 1.0, + "content": "cypripedium calceolus, and earthstar). The 100 images of each class contain 50 testing images", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "and 50 randomly selected training images. We fine-tune the checkpoints of both AlexNet and VGG", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 336, + 393 + ], + "score": 1.0, + "content": "pre-trained on ImageNet1 on these 800 images to achieve", + "type": "text" + }, + { + "bbox": [ + 336, + 379, + 361, + 390 + ], + "score": 0.88, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "accuracy. Then, 785 benign images", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "are kept for detection, as both their noisy images and themselves can be correctly predicted by both", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "score": 1.0, + "content": "the fine-tuned AlexNet and VGG models. These 785 images are then randomly partitioned to 400", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "training and 305 testing images used for detection. For each database, as described in Section 4.1,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "one noisy and one adversarial image are generated for each benign image; then, all of benign, noisy,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 435, + 291, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 291, + 446 + ], + "score": 1.0, + "content": "and adversarial images are used for detection.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "Attack methods We adopt four popular adversarial attack methods to craft adversarial examples,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "including basic iterative method (BIM (Kurakin et al. (2016))), CarliniWagnerL2Attack (CW-L2", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "score": 1.0, + "content": "(Carlini & Wagner (2017a))), DeepFool ( (Moosavi-Dezfooli et al. (2016))), and random projected", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 483, + 450, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 450, + 496 + ], + "score": 1.0, + "content": "gradient descent (R-PGD (Madry et al. (2017))). They are implemented by Foolbox2.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "Compared detection methods We compare with three state-of-the-art and open-sourced adversarial", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 509, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 224, + 523 + ], + "score": 1.0, + "content": "detection methods, including", + "type": "text" + }, + { + "bbox": [ + 225, + 509, + 263, + 521 + ], + "score": 0.92, + "content": "\\mathrm { K D + B U } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 509, + 506, + 523 + ], + "score": 1.0, + "content": "(Feinman et al. (2017)), Mahalanobis distance4 (M-D) (Lee", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 519, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 104, + 519, + 180, + 535 + ], + "score": 1.0, + "content": "et al. (2018)), and", + "type": "text" + }, + { + "bbox": [ + 180, + 520, + 203, + 532 + ], + "score": 0.75, + "content": "\\mathrm { L I D } ^ { 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 519, + 506, + 535 + ], + "score": 1.0, + "content": "(Ma et al. (2018)). Note that another recent work called I-defender (Zheng", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "& Hong (2018)) is not compared, as its code is not available. To ensure the fair comparison, the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "SVM classifier is trained with all compared methods, implemented by the fitcsvm6 function in MAT-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "score": 1.0, + "content": "LAB. There are two important hyper-parameters in LID, i.e., the size of mini-batch and the number", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "of neighbors. On CIFAR-10 and SVHN, they are respectively set as 100 and 20, as suggested in Ma", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "et al. (2018); on ImageNet, as there are only 400 benign training images, they are respectively set as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "50 and 20 in experiments. Moreover, we find that there are some unfair settings in the implementa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 274, + 610 + ], + "score": 1.0, + "content": "tions of compared methods. For example,", + "type": "text" + }, + { + "bbox": [ + 274, + 598, + 309, + 608 + ], + "score": 0.69, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "utilizes the extra 50, 000 images of CIFAR-10 to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "compute the kernel density of each training and testing image; M-D also uses these extra images to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "score": 1.0, + "content": "compute the mean and co-variance of GMM. Since extra images of the similar distribution with the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "training images are often unavailable, we believe that extra images should not be used to ensure the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 315, + 654 + ], + "score": 1.0, + "content": "fair comparison. Thus, extra images are not used for", + "type": "text" + }, + { + "bbox": [ + 316, + 642, + 351, + 652 + ], + "score": 0.73, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "and M-D in our experiments. Besides,", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 41.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 666, + 358, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 665, + 325, + 678 + ], + "spans": [ + { + "bbox": [ + 119, + 665, + 325, + 678 + ], + "score": 1.0, + "content": "1https://pytorch.org/docs/robust/torchvision/models.html", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 677, + 267, + 688 + ], + "spans": [ + { + "bbox": [ + 119, + 677, + 267, + 688 + ], + "score": 1.0, + "content": "2https://foolbox.readthedocs.io/en/latest/", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 686, + 339, + 701 + ], + "spans": [ + { + "bbox": [ + 117, + 686, + 339, + 701 + ], + "score": 1.0, + "content": "3 https://github.com/rfeinman/detecting-adversarial-samples/", + "type": "text" + } + ] + }, + { + "bbox": [ + 116, + 697, + 339, + 713 + ], + "spans": [ + { + "bbox": [ + 116, + 697, + 339, + 713 + ], + "score": 1.0, + "content": "4 https://github.com/pokaxpoka/deep Mahalanobis detector/", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 708, + 359, + 723 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 359, + 723 + ], + "score": 1.0, + "content": "5 https://github.com/xingjunm/lid adversarial subspace detection/", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 313, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 313, + 734 + ], + "score": 1.0, + "content": "6https://www.mathworks.com/help/stats/fitcsvm.html", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 500, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 500, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 500, + 96 + ], + "score": 1.0, + "content": "Algorithm 1 Training the adversarial detector via the magnitude of Benford-Fourier coefficients.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 81, + 500, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 385, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 384, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 234, + 109 + ], + "score": 1.0, + "content": "Require: The trained CNN model", + "type": "text" + }, + { + "bbox": [ + 234, + 96, + 244, + 106 + ], + "score": 0.85, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 94, + 263, + 109 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 264, + 96, + 271, + 105 + ], + "score": 0.78, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 94, + 369, + 109 + ], + "score": 1.0, + "content": "layers, and the training set", + "type": "text" + }, + { + "bbox": [ + 369, + 96, + 384, + 106 + ], + "score": 0.85, + "content": "\\mathcal { D } _ { t r }", + "type": "inline_equation" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 94, + 384, + 109 + ] + }, + { + "type": "index", + "bbox": [ + 110, + 105, + 449, + 210 + ], + "lines": [ + { + "bbox": [ + 109, + 103, + 203, + 117 + ], + "spans": [ + { + "bbox": [ + 109, + 103, + 135, + 117 + ], + "score": 1.0, + "content": "1: for", + "type": "text" + }, + { + "bbox": [ + 135, + 106, + 156, + 115 + ], + "score": 0.89, + "content": "i = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 103, + 167, + 117 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 167, + 106, + 187, + 116 + ], + "score": 0.89, + "content": "| \\mathcal { D } _ { t r } |", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 103, + 203, + 117 + ], + "score": 1.0, + "content": "do:", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 114, + 203, + 126 + ], + "spans": [ + { + "bbox": [ + 109, + 114, + 120, + 126 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 134, + 115, + 148, + 126 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 149, + 116, + 169, + 125 + ], + "score": 0.88, + "content": "l = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 115, + 180, + 126 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 180, + 117, + 187, + 124 + ], + "score": 0.73, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 115, + 203, + 126 + ], + "score": 1.0, + "content": "do:", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 124, + 327, + 137 + ], + "spans": [ + { + "bbox": [ + 109, + 124, + 120, + 137 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 148, + 124, + 184, + 137 + ], + "score": 1.0, + "content": "Compute", + "type": "text" + }, + { + "bbox": [ + 184, + 125, + 210, + 137 + ], + "score": 0.92, + "content": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } |", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 124, + 269, + 137 + ], + "score": 1.0, + "content": "as Eq. (8), with", + "type": "text" + }, + { + "bbox": [ + 270, + 126, + 321, + 136 + ], + "score": 0.9, + "content": "n = 1 , \\dots , T", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 124, + 327, + 137 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 134, + 344, + 149 + ], + "spans": [ + { + "bbox": [ + 109, + 136, + 120, + 147 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 147, + 134, + 195, + 149 + ], + "score": 1.0, + "content": "Concatenate", + "type": "text" + }, + { + "bbox": [ + 196, + 136, + 262, + 148 + ], + "score": 0.92, + "content": "\\{ | ( \\hat { a } _ { n } ) _ { i } ^ { l } | \\} _ { n = 1 , \\dots , T }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 134, + 327, + 149 + ], + "score": 1.0, + "content": "to obtain a vector", + "type": "text" + }, + { + "bbox": [ + 328, + 136, + 338, + 147 + ], + "score": 0.87, + "content": "\\hat { \\mathbf { a } } _ { i } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 134, + 344, + 149 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 146, + 165, + 157 + ], + "spans": [ + { + "bbox": [ + 109, + 146, + 120, + 157 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 135, + 146, + 165, + 157 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 155, + 330, + 170 + ], + "spans": [ + { + "bbox": [ + 109, + 156, + 120, + 168 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 133, + 155, + 182, + 170 + ], + "score": 1.0, + "content": "Concatenate", + "type": "text" + }, + { + "bbox": [ + 182, + 156, + 229, + 168 + ], + "score": 0.92, + "content": "\\{ \\hat { \\mathbf { a } } _ { i } ^ { l } \\} _ { l = 1 , \\dots , L }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 155, + 313, + 170 + ], + "score": 1.0, + "content": "to obtain a long vector", + "type": "text" + }, + { + "bbox": [ + 313, + 157, + 322, + 167 + ], + "score": 0.83, + "content": "\\hat { \\mathbf { a } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 155, + 330, + 170 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 108, + 165, + 153, + 178 + ], + "spans": [ + { + "bbox": [ + 108, + 165, + 153, + 178 + ], + "score": 1.0, + "content": "7: end for", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 108, + 173, + 449, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 173, + 333, + 190 + ], + "score": 1.0, + "content": "8: Build a novel representation of the training set, denoted as", + "type": "text" + }, + { + "bbox": [ + 333, + 175, + 443, + 188 + ], + "score": 0.88, + "content": "\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\mathbf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , | \\mathcal { D } _ { t r } | }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 173, + 449, + 190 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 187, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 109, + 187, + 264, + 200 + ], + "score": 1.0, + "content": "9: Train a binary SVM classifier based on", + "type": "text" + }, + { + "bbox": [ + 264, + 187, + 280, + 199 + ], + "score": 0.89, + "content": "\\hat { A } _ { t r }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 187, + 286, + 200 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 120, + 199, + 274, + 210 + ], + "spans": [ + { + "bbox": [ + 120, + 199, + 274, + 210 + ], + "score": 1.0, + "content": "return The trained binary SVM classifier.", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + } + ], + "index": 6.5, + "bbox_fs": [ + 108, + 103, + 449, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "score": 1.0, + "content": "et al. (2009)). In terms of CIFAR-10 and SVHN, we adopt the same settings as the compared method", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "LID (Ma et al. (2018)). Specifically, a 33-layer network pre-trained on the training set of CIFAR-10", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 257, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 143, + 273 + ], + "score": 1.0, + "content": "achieves", + "type": "text" + }, + { + "bbox": [ + 144, + 259, + 176, + 270 + ], + "score": 0.89, + "content": "8 2 . 3 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 257, + 505, + 273 + ], + "score": 1.0, + "content": "accuracy on the testing set with 10, 000 benign images; a 19-layer network pre-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 287, + 283 + ], + "score": 1.0, + "content": "trained on the training set of SVHN achieves", + "type": "text" + }, + { + "bbox": [ + 288, + 270, + 315, + 281 + ], + "score": 0.88, + "content": "9 2 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "accuracy on the testing set with 26, 032 benign", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 303, + 293 + ], + "score": 1.0, + "content": "images. Then, we add a small noise drawn from", + "type": "text" + }, + { + "bbox": [ + 303, + 281, + 342, + 293 + ], + "score": 0.93, + "content": "{ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 281, + 456, + 293 + ], + "score": 1.0, + "content": "on each testing image, with", + "type": "text" + }, + { + "bbox": [ + 456, + 282, + 464, + 291 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "being the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 182, + 304 + ], + "score": 1.0, + "content": "similar level of the", + "type": "text" + }, + { + "bbox": [ + 182, + 292, + 192, + 303 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "norm of adversarial perturbations on the same database. If both the benign and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "its noisy image can be correctly predicted by the classification network, then it is picked out for train-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "ing the detection. We finally collect 8, 175 and 23, 862 benign images from CIFATR-10 and SVHN,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 324, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 324, + 340, + 339 + ], + "score": 1.0, + "content": "respectively. These images are randomly partitioned to the", + "type": "text" + }, + { + "bbox": [ + 341, + 325, + 361, + 336 + ], + "score": 0.89, + "content": "8 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 324, + 439, + 339 + ], + "score": 1.0, + "content": "training set and the", + "type": "text" + }, + { + "bbox": [ + 439, + 325, + 459, + 335 + ], + "score": 0.89, + "content": "2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 324, + 506, + 339 + ], + "score": 1.0, + "content": "testing set,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "used for the training and testing of the detector. We also collect a subset from ImageNet, including", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 346, + 504, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 504, + 360 + ], + "score": 1.0, + "content": "800 benign images of 8 classes (snowbird, spoonbill, bobtail, Leonberg, hamster, proboscis monkey,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "score": 1.0, + "content": "cypripedium calceolus, and earthstar). The 100 images of each class contain 50 testing images", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "and 50 randomly selected training images. We fine-tune the checkpoints of both AlexNet and VGG", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 336, + 393 + ], + "score": 1.0, + "content": "pre-trained on ImageNet1 on these 800 images to achieve", + "type": "text" + }, + { + "bbox": [ + 336, + 379, + 361, + 390 + ], + "score": 0.88, + "content": "1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "accuracy. Then, 785 benign images", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "are kept for detection, as both their noisy images and themselves can be correctly predicted by both", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 413 + ], + "score": 1.0, + "content": "the fine-tuned AlexNet and VGG models. These 785 images are then randomly partitioned to 400", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "training and 305 testing images used for detection. For each database, as described in Section 4.1,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "one noisy and one adversarial image are generated for each benign image; then, all of benign, noisy,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 435, + 291, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 291, + 446 + ], + "score": 1.0, + "content": "and adversarial images are used for detection.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 237, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 451, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 463 + ], + "score": 1.0, + "content": "Attack methods We adopt four popular adversarial attack methods to craft adversarial examples,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "including basic iterative method (BIM (Kurakin et al. (2016))), CarliniWagnerL2Attack (CW-L2", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "score": 1.0, + "content": "(Carlini & Wagner (2017a))), DeepFool ( (Moosavi-Dezfooli et al. (2016))), and random projected", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 483, + 450, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 450, + 496 + ], + "score": 1.0, + "content": "gradient descent (R-PGD (Madry et al. (2017))). They are implemented by Foolbox2.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 450, + 506, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 499, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "Compared detection methods We compare with three state-of-the-art and open-sourced adversarial", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 509, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 224, + 523 + ], + "score": 1.0, + "content": "detection methods, including", + "type": "text" + }, + { + "bbox": [ + 225, + 509, + 263, + 521 + ], + "score": 0.92, + "content": "\\mathrm { K D + B U } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 509, + 506, + 523 + ], + "score": 1.0, + "content": "(Feinman et al. (2017)), Mahalanobis distance4 (M-D) (Lee", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 519, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 104, + 519, + 180, + 535 + ], + "score": 1.0, + "content": "et al. (2018)), and", + "type": "text" + }, + { + "bbox": [ + 180, + 520, + 203, + 532 + ], + "score": 0.75, + "content": "\\mathrm { L I D } ^ { 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 519, + 506, + 535 + ], + "score": 1.0, + "content": "(Ma et al. (2018)). Note that another recent work called I-defender (Zheng", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "& Hong (2018)) is not compared, as its code is not available. To ensure the fair comparison, the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "SVM classifier is trained with all compared methods, implemented by the fitcsvm6 function in MAT-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "score": 1.0, + "content": "LAB. There are two important hyper-parameters in LID, i.e., the size of mini-batch and the number", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "of neighbors. On CIFAR-10 and SVHN, they are respectively set as 100 and 20, as suggested in Ma", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "et al. (2018); on ImageNet, as there are only 400 benign training images, they are respectively set as", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "50 and 20 in experiments. Moreover, we find that there are some unfair settings in the implementa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 274, + 610 + ], + "score": 1.0, + "content": "tions of compared methods. For example,", + "type": "text" + }, + { + "bbox": [ + 274, + 598, + 309, + 608 + ], + "score": 0.69, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "utilizes the extra 50, 000 images of CIFAR-10 to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "compute the kernel density of each training and testing image; M-D also uses these extra images to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 632 + ], + "score": 1.0, + "content": "compute the mean and co-variance of GMM. Since extra images of the similar distribution with the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "training images are often unavailable, we believe that extra images should not be used to ensure the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 315, + 654 + ], + "score": 1.0, + "content": "fair comparison. Thus, extra images are not used for", + "type": "text" + }, + { + "bbox": [ + 316, + 642, + 351, + 652 + ], + "score": 0.73, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "and M-D in our experiments. Besides,", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 499, + 506, + 654 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 100, + 489, + 225 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 203, + 90, + 407, + 100 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 203, + 87, + 408, + 102 + ], + "spans": [ + { + "bbox": [ + 203, + 87, + 408, + 102 + ], + "score": 1.0, + "content": "Table 1: Detection results in the non-transfer case.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 120, + 100, + 489, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 100, + 489, + 225 + ], + "spans": [ + { + "bbox": [ + 120, + 100, + 489, + 225 + ], + "score": 0.965, + "html": "
databaseDetectorAUROC(%)Accuracy (%)
BIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
CIFAR-10KD+BU (Feinman et al. (2017))79.082.880.677.774.273.071.772.6
M-D (Lee et al. (2018))51.748.453.752.066.766.766.766.7
LID (Ma et al. (2018))87.987.486.682.772.678.578.070.2
MBF99.699.696.999.498.898.391.898.5
SVHNKD+BU (Feinman et al. (2017)) M-D (Lee et al. (2018))81.585.184.182.578.680.079.378.6
LID (Ma et al. (2018))49.750.049.650.166.766.766.766.7
91.788.792.290.583.380.884.081.5
ImageNet- AlexNetMBF KD+BU (Feinman et al. (2017))99.799.999.399.598.898.391.898.5
M-D (Lee et al. (2018))50.852.6 60.151.151.534.436.834.935.3
LID (Ma et al. (2018))48.1 71.770.946.5 71.954.1 72.366.7 68.766.6 60.566.866.7
ImageNet-MBF99.999.699.999.898.697.865.8 98.868.3
KD+BU (Feinman et al. (2017))57.157.756.258.642.843.641.798.4
M-D (Lee et al. (2018))63.764.845.966.967.165.666.744.8
89.567.4
VGG16LID (Ma et al. (2018)) MBF82.6 99.884.2 100.0100.084.2 100.077.8 99.676.7 99.683.3 100.076.4 100.0
", + "type": "table", + "image_path": "d024e527f7cda4b02a67ead484dc23c7889e4450d6e1bd0d6490c6229dde8930.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 120, + 100, + 489, + 141.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 141.66666666666666, + 489, + 183.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 120, + 183.33333333333331, + 489, + 224.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 119, + 245, + 492, + 365 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 146, + 231, + 464, + 243 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 229, + 466, + 245 + ], + "spans": [ + { + "bbox": [ + 145, + 229, + 340, + 245 + ], + "score": 1.0, + "content": "Table 2: Detection results evaluated by AUROC", + "type": "text" + }, + { + "bbox": [ + 340, + 231, + 355, + 243 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 229, + 466, + 245 + ], + "score": 1.0, + "content": "in the attack-transfer case.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 119, + 245, + 492, + 365 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 245, + 492, + 365 + ], + "spans": [ + { + "bbox": [ + 119, + 245, + 492, + 365 + ], + "score": 0.845, + "html": "
databaseTest attack→BIMCW-L2DeepFoolR-PGD
TrainattackKD+BU (Feinman et al. (2017))/M-D(Lee et al.(2018))/LID (Ma et al.(2018)/MBF
CIFAR-10BIM CW-L2 DeepFool R-PGD79.0/51.7787.9799.6 83.4/47.9/90.3/99.6 83.6/53.5 /89.7 /99.3 78.1/51.5/86.1/99.678.1/52.2/85.7799.4 82.8/48.4/87.4/99.6 83.1/51.4/86.1/99.1 77.3 /51.9 /84.5 /99.576.4/52.4/86.4/87.8 80.3 /46.5 /87.0 / 88.6 80.6/53.7/86.6/96.978.6/50.9/87.6/99.4 82.9/48.2/88.9/99.5 83.1/53.6/87.9/98.9 77.7/52.1/82.7/99.4
SVHNBIM CW-L2 DeepFool R-PGD81.5749.7791.7799.7 83.6/50.8/91.6/99.7 85.0/50.0/91.6/99.7 82.5 /49.9 /91.2 /99.683.1/51.0/88.3/99.7 85.1/50.0/88.7/99.9 86.3/ 50.0 /87.9 /99.8 84.0/49.5/87.8/99.675.8/52.4/84.9/88.3 80.1/48.8/91.9/97.5 82.5/49.7/92.0/97.3 84.1/49.6/92.2/99.3 81.2/50.2/91.6/97.481.5750.4/90.7799.5 83.6/49.1/90.8/99.7 85.0/49.4/90.8 /99.6 82.5/50.1/90.5 /99.5
ImageNet- AlexNetBIM CW-L2 DeepFool R-PGD50.8/48.1/71.7/99.9 52.3/50.7/69.7/99.8 51.8/53.6/71.3/99.9 51.6/52.4/72.5 /99.850.8/44.9/70.8/99.6 52.6/60.1/70.9/99.7 51.9/53.5/69.7/99.7 51.3/ 50.4/72.3/99.650.4/49.7772.1/99.8 51.2/52.1/66.0/99.7 51.2/46.5 /71.9 /99.9 50.3 /50.5 /72.2/99.750.7/50.4/70.8/99.8 52.2/55.8/69.8/99.8 51.8/51.2/71.2/99.8
ImageNet- VGG-16BIM CW-L2 DeepFool R-PGD57.1/63.7/82.6/99.8 57.3/64.3/81.9/100.0 58.7/45.5/ 80.2/99.7 58.6/62.3/81.1/100.057.1/57.2/84.3/99.5 57.7/64.8/84.2/100.0 59.4/46.1/79.1/99.3 58.9/60.8/83.7/100.053.8/54.4/90.4/100.0 53.8/56.6/90.2/100.0 56.2/45.9/89.5/100.0 55.3/54.4/90.8/100.051.5 /54.1/72.3 /99.8 57.1/65.4/84.2/100.0 57.5/67.6/84.3/100.0 59.0 /44.0/ 80.9 /99.7 58.6/66.9/84.2/100.0
", + "type": "table", + "image_path": "fb083cac0a3e20d6cea6f26727565045ea368fd4ac7db491d45a921712eb9bee.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 119, + 245, + 492, + 285.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 119, + 285.0, + 492, + 325.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 119, + 325.0, + 492, + 365.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "table", + "bbox": [ + 119, + 382, + 491, + 506 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 145, + 369, + 465, + 380 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 144, + 367, + 467, + 382 + ], + "spans": [ + { + "bbox": [ + 144, + 367, + 340, + 382 + ], + "score": 1.0, + "content": "Table 3: Detection results evaluated by accuracy", + "type": "text" + }, + { + "bbox": [ + 341, + 369, + 355, + 380 + ], + "score": 0.71, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 367, + 467, + 382 + ], + "score": 1.0, + "content": "in the attack-transfer case.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 119, + 382, + 491, + 506 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 119, + 382, + 491, + 506 + ], + "spans": [ + { + "bbox": [ + 119, + 382, + 491, + 506 + ], + "score": 0.924, + "html": "
databaseTest attack→BIM CW-L2 DeepFool R-PGDKD+BU(Feinman et al. (2017))/M-D(Lee et al. (2018))/LID (Ma et al. (2018)/MBF
CIFAR-10BIMCW-L2DeepFoolR-PGD74.2/66.7/72.6/98.874.2/66.7/80.3 /98.174.0/66.7 /80.4/95.273.1/66.7/77.7 /98.666.773.0/66.7 /78.5 /98.3 71.7 /66.7/77.9 /89.573.8/66.7/61.3/96.871.7 /66.7/77.9 /89.5 73.8 /66.7 /73.6 /98.2
66.771.9 / 66.7 /75.1/98.5 71.0 /66.7/75.4/89.773.4/66.7/74.8/95.471.0 /66.7/75.4/89.7 72.6/66.7/70.2/98.5
SVHNBIMCW-L2DeepFoolR-PGD78.6/66.77/83.3/98.7 80.5/66.7779.4/98.9 77.7/66.7/83.5/95.8 78.6/66.7/82.9/98.278.2/66.7/83.5 /98.0 80.0/66.7 /80.8/99.2 77.1/66.7/84.0 /95.9 78.1/66.7/80.7/97.379.8/66.7/83.9/98.0 81.8/66.7/79.9/98.4 79.3/66.7/84.0/97.3 79.9/66.7/80.7/98.078.6/66.7 /83.6 /98.6 80.5 / 66.7 /79.9/98.9 77.7 /66.7 /83.8/96.5 78.6/ 66.7 /81.5 /98.8
ImageNet-AlexNetBIMCW-L2DeepFoolR-PGD34.4/66.7/68.7798.6 34.4/66.6/68.6797.336.5/66.2/64.6/98.0 36.8/66.6/60.6/97.8 35.0 /66.1/60.7/98.134.3/66.7/65.0/98.435.0 /66.1/60.7/98.1 36.4/66.1/62.6/98.034.9/66.8/65.8/98.8 35.9 /66.6/68.3 /98.634.1/66.7/67.0/98.3 35.3/66.7/68.3/98.4
36.0 /66.7/69.4/98.835.4/66.7/67.5/98.566.835.2/66.7/67.7/97.6 34.1/66.7/67.0/98.3
ImageNet-VGG-16BIMCW-L2DeepFoolR-PGD42.8/67.1777.8/98.643.1/64.8/75.2/98.045.4/66.7 /76.2/98.844.8 /66.3/76.3/98.566.643.6/65.6/76.7 /97.8 38.4/ 62.5/82.5 /98.142.7/67.9/76.9/98.438.4/ 62.5/82.5 /98.1 43.3/67.2/73.9/98.041.8/66.7 /83.3/98.8 45.8 /66.7/74.4/98.640.6/64.9/82.4/98.3 45.8/67.4/76.4/98.4
66.745.4/66.6/74.1/97.6 40.6/64.9/82.4/98.3
", + "type": "table", + "image_path": "88ec8aa6b35da30e3bb67fda7c18fc761e062fbf4a7d36de115eb99f540d4bd5.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 119, + 382, + 491, + 423.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 119, + 423.3333333333333, + 491, + 464.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 119, + 464.66666666666663, + 491, + 505.99999999999994 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 504, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 504, + 527 + ], + "score": 1.0, + "content": "LID utilizes other benign testing images as neighborhoods to extract features for each testing image.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "It is unfair to utilize the information of benign or adversarial for neighboring testing images. In our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 536, + 450, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 450, + 549 + ], + "score": 1.0, + "content": "experiments, we use benign training images as neighborhoods for each testing image.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "Three comparison cases and evaluation metrics We conduct experiments of three cases, includ-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 563, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 574 + ], + "score": 1.0, + "content": "ing: 1) non-transfer, both training and testing adversarial examples are crafted by the same attack", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "method; 2) attack-transfer, both training and testing adversarial examples are crafted by different", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "attack methods; 3) data-transfer, both training and testing adversarial examples are crafted by the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "same attack method, but the data sources of training and testing benign examples are different. Two", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "widely used metrics are used to evaluate the detection performance, including area under the re-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "ceiver operating characteristics (AUROC), and detection accuracy, which is the diagonal summation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "of the confusion matrix, using 0.5 as the threshold of the posterior probability. Higher values of both", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 639, + 253, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 253, + 653 + ], + "score": 1.0, + "content": "metrics indicate better performance.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 107, + 659, + 170, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 171, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 171, + 671 + ], + "score": 1.0, + "content": "4.3 RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "Detection results in the non-transfer case are shown in Table 1. The proposed MBF method shows", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "the best performance in all cases, and is much superior to all compared methods. LID performs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 228, + 711 + ], + "score": 1.0, + "content": "the second-best in most case.", + "type": "text" + }, + { + "bbox": [ + 229, + 699, + 264, + 709 + ], + "score": 0.67, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "gives somewhat good detection performance on CIFAR-10", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 354, + 722 + ], + "score": 1.0, + "content": "and SVHN, but performs very poor on ImageNet. It tells that", + "type": "text" + }, + { + "bbox": [ + 355, + 710, + 390, + 720 + ], + "score": 0.65, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "is very sensitive to different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "databases and networks. M-D gives almost the degenerate results. Note that the results of M-D on", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 100, + 489, + 225 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 203, + 90, + 407, + 100 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 203, + 87, + 408, + 102 + ], + "spans": [ + { + "bbox": [ + 203, + 87, + 408, + 102 + ], + "score": 1.0, + "content": "Table 1: Detection results in the non-transfer case.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 120, + 100, + 489, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 100, + 489, + 225 + ], + "spans": [ + { + "bbox": [ + 120, + 100, + 489, + 225 + ], + "score": 0.965, + "html": "
databaseDetectorAUROC(%)Accuracy (%)
BIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
CIFAR-10KD+BU (Feinman et al. (2017))79.082.880.677.774.273.071.772.6
M-D (Lee et al. (2018))51.748.453.752.066.766.766.766.7
LID (Ma et al. (2018))87.987.486.682.772.678.578.070.2
MBF99.699.696.999.498.898.391.898.5
SVHNKD+BU (Feinman et al. (2017)) M-D (Lee et al. (2018))81.585.184.182.578.680.079.378.6
LID (Ma et al. (2018))49.750.049.650.166.766.766.766.7
91.788.792.290.583.380.884.081.5
ImageNet- AlexNetMBF KD+BU (Feinman et al. (2017))99.799.999.399.598.898.391.898.5
M-D (Lee et al. (2018))50.852.6 60.151.151.534.436.834.935.3
LID (Ma et al. (2018))48.1 71.770.946.5 71.954.1 72.366.7 68.766.6 60.566.866.7
ImageNet-MBF99.999.699.999.898.697.865.8 98.868.3
KD+BU (Feinman et al. (2017))57.157.756.258.642.843.641.798.4
M-D (Lee et al. (2018))63.764.845.966.967.165.666.744.8
89.567.4
VGG16LID (Ma et al. (2018)) MBF82.6 99.884.2 100.0100.084.2 100.077.8 99.676.7 99.683.3 100.076.4 100.0
", + "type": "table", + "image_path": "d024e527f7cda4b02a67ead484dc23c7889e4450d6e1bd0d6490c6229dde8930.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 120, + 100, + 489, + 141.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 141.66666666666666, + 489, + 183.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 120, + 183.33333333333331, + 489, + 224.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 119, + 245, + 492, + 365 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 146, + 231, + 464, + 243 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 229, + 466, + 245 + ], + "spans": [ + { + "bbox": [ + 145, + 229, + 340, + 245 + ], + "score": 1.0, + "content": "Table 2: Detection results evaluated by AUROC", + "type": "text" + }, + { + "bbox": [ + 340, + 231, + 355, + 243 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 229, + 466, + 245 + ], + "score": 1.0, + "content": "in the attack-transfer case.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 119, + 245, + 492, + 365 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 119, + 245, + 492, + 365 + ], + "spans": [ + { + "bbox": [ + 119, + 245, + 492, + 365 + ], + "score": 0.845, + "html": "
databaseTest attack→BIMCW-L2DeepFoolR-PGD
TrainattackKD+BU (Feinman et al. (2017))/M-D(Lee et al.(2018))/LID (Ma et al.(2018)/MBF
CIFAR-10BIM CW-L2 DeepFool R-PGD79.0/51.7787.9799.6 83.4/47.9/90.3/99.6 83.6/53.5 /89.7 /99.3 78.1/51.5/86.1/99.678.1/52.2/85.7799.4 82.8/48.4/87.4/99.6 83.1/51.4/86.1/99.1 77.3 /51.9 /84.5 /99.576.4/52.4/86.4/87.8 80.3 /46.5 /87.0 / 88.6 80.6/53.7/86.6/96.978.6/50.9/87.6/99.4 82.9/48.2/88.9/99.5 83.1/53.6/87.9/98.9 77.7/52.1/82.7/99.4
SVHNBIM CW-L2 DeepFool R-PGD81.5749.7791.7799.7 83.6/50.8/91.6/99.7 85.0/50.0/91.6/99.7 82.5 /49.9 /91.2 /99.683.1/51.0/88.3/99.7 85.1/50.0/88.7/99.9 86.3/ 50.0 /87.9 /99.8 84.0/49.5/87.8/99.675.8/52.4/84.9/88.3 80.1/48.8/91.9/97.5 82.5/49.7/92.0/97.3 84.1/49.6/92.2/99.3 81.2/50.2/91.6/97.481.5750.4/90.7799.5 83.6/49.1/90.8/99.7 85.0/49.4/90.8 /99.6 82.5/50.1/90.5 /99.5
ImageNet- AlexNetBIM CW-L2 DeepFool R-PGD50.8/48.1/71.7/99.9 52.3/50.7/69.7/99.8 51.8/53.6/71.3/99.9 51.6/52.4/72.5 /99.850.8/44.9/70.8/99.6 52.6/60.1/70.9/99.7 51.9/53.5/69.7/99.7 51.3/ 50.4/72.3/99.650.4/49.7772.1/99.8 51.2/52.1/66.0/99.7 51.2/46.5 /71.9 /99.9 50.3 /50.5 /72.2/99.750.7/50.4/70.8/99.8 52.2/55.8/69.8/99.8 51.8/51.2/71.2/99.8
ImageNet- VGG-16BIM CW-L2 DeepFool R-PGD57.1/63.7/82.6/99.8 57.3/64.3/81.9/100.0 58.7/45.5/ 80.2/99.7 58.6/62.3/81.1/100.057.1/57.2/84.3/99.5 57.7/64.8/84.2/100.0 59.4/46.1/79.1/99.3 58.9/60.8/83.7/100.053.8/54.4/90.4/100.0 53.8/56.6/90.2/100.0 56.2/45.9/89.5/100.0 55.3/54.4/90.8/100.051.5 /54.1/72.3 /99.8 57.1/65.4/84.2/100.0 57.5/67.6/84.3/100.0 59.0 /44.0/ 80.9 /99.7 58.6/66.9/84.2/100.0
", + "type": "table", + "image_path": "fb083cac0a3e20d6cea6f26727565045ea368fd4ac7db491d45a921712eb9bee.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 119, + 245, + 492, + 285.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 119, + 285.0, + 492, + 325.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 119, + 325.0, + 492, + 365.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "table", + "bbox": [ + 119, + 382, + 491, + 506 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 145, + 369, + 465, + 380 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 144, + 367, + 467, + 382 + ], + "spans": [ + { + "bbox": [ + 144, + 367, + 340, + 382 + ], + "score": 1.0, + "content": "Table 3: Detection results evaluated by accuracy", + "type": "text" + }, + { + "bbox": [ + 341, + 369, + 355, + 380 + ], + "score": 0.71, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 367, + 467, + 382 + ], + "score": 1.0, + "content": "in the attack-transfer case.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 119, + 382, + 491, + 506 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 119, + 382, + 491, + 506 + ], + "spans": [ + { + "bbox": [ + 119, + 382, + 491, + 506 + ], + "score": 0.924, + "html": "
databaseTest attack→BIM CW-L2 DeepFool R-PGDKD+BU(Feinman et al. (2017))/M-D(Lee et al. (2018))/LID (Ma et al. (2018)/MBF
CIFAR-10BIMCW-L2DeepFoolR-PGD74.2/66.7/72.6/98.874.2/66.7/80.3 /98.174.0/66.7 /80.4/95.273.1/66.7/77.7 /98.666.773.0/66.7 /78.5 /98.3 71.7 /66.7/77.9 /89.573.8/66.7/61.3/96.871.7 /66.7/77.9 /89.5 73.8 /66.7 /73.6 /98.2
66.771.9 / 66.7 /75.1/98.5 71.0 /66.7/75.4/89.773.4/66.7/74.8/95.471.0 /66.7/75.4/89.7 72.6/66.7/70.2/98.5
SVHNBIMCW-L2DeepFoolR-PGD78.6/66.77/83.3/98.7 80.5/66.7779.4/98.9 77.7/66.7/83.5/95.8 78.6/66.7/82.9/98.278.2/66.7/83.5 /98.0 80.0/66.7 /80.8/99.2 77.1/66.7/84.0 /95.9 78.1/66.7/80.7/97.379.8/66.7/83.9/98.0 81.8/66.7/79.9/98.4 79.3/66.7/84.0/97.3 79.9/66.7/80.7/98.078.6/66.7 /83.6 /98.6 80.5 / 66.7 /79.9/98.9 77.7 /66.7 /83.8/96.5 78.6/ 66.7 /81.5 /98.8
ImageNet-AlexNetBIMCW-L2DeepFoolR-PGD34.4/66.7/68.7798.6 34.4/66.6/68.6797.336.5/66.2/64.6/98.0 36.8/66.6/60.6/97.8 35.0 /66.1/60.7/98.134.3/66.7/65.0/98.435.0 /66.1/60.7/98.1 36.4/66.1/62.6/98.034.9/66.8/65.8/98.8 35.9 /66.6/68.3 /98.634.1/66.7/67.0/98.3 35.3/66.7/68.3/98.4
36.0 /66.7/69.4/98.835.4/66.7/67.5/98.566.835.2/66.7/67.7/97.6 34.1/66.7/67.0/98.3
ImageNet-VGG-16BIMCW-L2DeepFoolR-PGD42.8/67.1777.8/98.643.1/64.8/75.2/98.045.4/66.7 /76.2/98.844.8 /66.3/76.3/98.566.643.6/65.6/76.7 /97.8 38.4/ 62.5/82.5 /98.142.7/67.9/76.9/98.438.4/ 62.5/82.5 /98.1 43.3/67.2/73.9/98.041.8/66.7 /83.3/98.8 45.8 /66.7/74.4/98.640.6/64.9/82.4/98.3 45.8/67.4/76.4/98.4
66.745.4/66.6/74.1/97.6 40.6/64.9/82.4/98.3
", + "type": "table", + "image_path": "88ec8aa6b35da30e3bb67fda7c18fc761e062fbf4a7d36de115eb99f540d4bd5.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 119, + 382, + 491, + 423.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 119, + 423.3333333333333, + 491, + 464.66666666666663 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 119, + 464.66666666666663, + 491, + 505.99999999999994 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 504, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 504, + 527 + ], + "score": 1.0, + "content": "LID utilizes other benign testing images as neighborhoods to extract features for each testing image.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "It is unfair to utilize the information of benign or adversarial for neighboring testing images. In our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 536, + 450, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 450, + 549 + ], + "score": 1.0, + "content": "experiments, we use benign training images as neighborhoods for each testing image.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 512, + 505, + 549 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "Three comparison cases and evaluation metrics We conduct experiments of three cases, includ-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 563, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 574 + ], + "score": 1.0, + "content": "ing: 1) non-transfer, both training and testing adversarial examples are crafted by the same attack", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "method; 2) attack-transfer, both training and testing adversarial examples are crafted by different", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "attack methods; 3) data-transfer, both training and testing adversarial examples are crafted by the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "same attack method, but the data sources of training and testing benign examples are different. Two", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "widely used metrics are used to evaluate the detection performance, including area under the re-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "ceiver operating characteristics (AUROC), and detection accuracy, which is the diagonal summation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "of the confusion matrix, using 0.5 as the threshold of the posterior probability. Higher values of both", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 639, + 253, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 253, + 653 + ], + "score": 1.0, + "content": "metrics indicate better performance.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 551, + 506, + 653 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 659, + 170, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 171, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 171, + 671 + ], + "score": 1.0, + "content": "4.3 RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "Detection results in the non-transfer case are shown in Table 1. The proposed MBF method shows", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "the best performance in all cases, and is much superior to all compared methods. LID performs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 228, + 711 + ], + "score": 1.0, + "content": "the second-best in most case.", + "type": "text" + }, + { + "bbox": [ + 229, + 699, + 264, + 709 + ], + "score": 0.67, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "gives somewhat good detection performance on CIFAR-10", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 354, + 722 + ], + "score": 1.0, + "content": "and SVHN, but performs very poor on ImageNet. It tells that", + "type": "text" + }, + { + "bbox": [ + 355, + 710, + 390, + 720 + ], + "score": 0.65, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "is very sensitive to different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "databases and networks. M-D gives almost the degenerate results. Note that the results of M-D on", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 676, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 124, + 457, + 178 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 89, + 504, + 122 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 320, + 102 + ], + "score": 1.0, + "content": "Table 4: Detection results evaluated by AUROC score", + "type": "text" + }, + { + "bbox": [ + 321, + 90, + 336, + 100 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "in the data-transfer case. All detectors are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "score": 1.0, + "content": "trained on the train set, and tested on the out-of-sample set (including 365 images) of the ImageNet", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 110, + 307, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 307, + 123 + ], + "score": 1.0, + "content": "database. The best results are highlighted in bold.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 153, + 124, + 457, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 124, + 457, + 178 + ], + "spans": [ + { + "bbox": [ + 153, + 124, + 457, + 178 + ], + "score": 0.962, + "html": "
ImageNet-AlexNetImageNet-VGG-16
DetectorBIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
KD+BU66.369.464.768.884.085.079.985.2
M-D48.749.250.051.054.550.050.456.1
LID69.167.672.971.484.288.489.283.7
MBF99.199.099.699.499.399.599.399.5
", + "type": "table", + "image_path": "69efaee09e76d4a1b56e04fd81944dc90fd9797a28a35b7f953f4304f89b2339.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 153, + 124, + 457, + 142.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 153, + 142.0, + 457, + 160.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 153, + 160.0, + 457, + 178.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 226, + 198, + 385, + 226 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 185, + 504, + 196 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 184, + 504, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 224, + 196 + ], + "score": 1.0, + "content": "Table 5: Attack Failure Rates", + "type": "text" + }, + { + "bbox": [ + 224, + 185, + 239, + 196 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 184, + 504, + 196 + ], + "score": 1.0, + "content": "of adaptive attacks with different detectors. See context for details.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 226, + 198, + 385, + 226 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 226, + 198, + 385, + 226 + ], + "spans": [ + { + "bbox": [ + 226, + 198, + 385, + 226 + ], + "score": 0.956, + "html": "
Original CW-L2LIDMBF
CIFAR-100.088.4100.0
SVHN0.085.692.9
", + "type": "table", + "image_path": "b975bebdd9d344b5960d73838588cd6994286a09202d3cfa99fe01e1e30d4edd.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 198, + 385, + 212.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 226, + 212.0, + 385, + 226.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "text", + "bbox": [ + 107, + 232, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "CIFAR-10 and SVHN reported in (Lee et al. (2018)) are very high, but their networks (i.e., ResNet", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "score": 1.0, + "content": "(He et al. (2016)) and DenseNet (Huang et al. (2017)) are different from that used in our experiments.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 255, + 503, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 503, + 267 + ], + "score": 1.0, + "content": "It implies that M-D may be suitable for very deep neural networks, but not for shallower networks.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 271, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 272, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 283 + ], + "score": 1.0, + "content": "Detection results in the attack-transfer case evaluated by AUROC and accuracy are presented in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "Tables 2 and 3, respectively. MBF still shows much better performance in all transfer cases than all", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "compared methods, and the changes among detecting different attacks are very small. It verifies the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 304, + 264, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 264, + 315 + ], + "score": 1.0, + "content": "robustness of MBF to different attacks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "score": 1.0, + "content": "We also conduct a data-transfer experiment on ImageNet. Specifically, we collect extra 365 images", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "of 8 classes (same with the classes used for the detection training, see Section 4.2) through searching", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "the class names in Baidu and Facebook. These 365 benign and their noisy images can be correctly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "predicted by both the fine-tuned AlexNet and VGG models. The results on these images are shown", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "in Table 4. MBF still shows the best performance. And, compared to the corresponding results in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "Table 1, the AUROC/accuracy scores of MBF on detecting different attacks change very gently. It", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 387, + 333, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 333, + 399 + ], + "score": 1.0, + "content": "verifies the robustness of MBF to different data sources.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 106, + 411, + 374, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 375, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 375, + 424 + ], + "score": 1.0, + "content": "4.4 ADAPTIVE ATTACK AGAINST ADVERSARIAL DETECTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 505, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "Similar to Carlini & Wagner (2017b) and Ma et al. (2018), we also present an adaptive white-box", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "attack, of which the goal is not only to fool the classifier, but also to evade the adversarial detector.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "It measures how easy to bypass the detector. Specifically, by combining the CW-L2 attack (Carlini", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 465, + 496, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 496, + 477 + ], + "score": 1.0, + "content": "& Wagner (2017a)) and the proposed MBF detector, the adaptive attack is formulated as follows:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 137, + 478, + 471, + 502 + ], + "lines": [ + { + "bbox": [ + 137, + 478, + 471, + 502 + ], + "spans": [ + { + "bbox": [ + 137, + 478, + 471, + 502 + ], + "score": 0.89, + "content": "\\operatorname * { a r g m i n } _ { \\mathbf { x } _ { a d v } \\in [ 0 , 1 ] } \\| \\mathbf { x } - \\mathbf { x } _ { a d v } \\| _ { 2 } ^ { 2 } + \\alpha \\cdot \\big [ - l ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) , y ) + \\| D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) - D ( f _ { \\theta } ( \\mathbf { x } ) ) \\| _ { 1 } \\big ] ,", + "type": "interline_equation", + "image_path": "ef0772f1a645b7f33bb82d6f5e88667e7a460e99dac440dd60000a4074a74b04.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 137, + 478, + 471, + 502 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 133, + 518 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 507, + 142, + 515 + ], + "score": 0.6, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 505, + 243, + 518 + ], + "score": 1.0, + "content": "denotes the clean input,", + "type": "text" + }, + { + "bbox": [ + 244, + 507, + 251, + 517 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 505, + 388, + 518 + ], + "score": 1.0, + "content": "indicates the ground-truth label.", + "type": "text" + }, + { + "bbox": [ + 388, + 505, + 429, + 517 + ], + "score": 0.93, + "content": "l ( \\mathbf { x } _ { a d v } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "is the Hinge loss.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 515, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 160, + 528 + ], + "score": 0.91, + "content": "D ( f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 515, + 365, + 530 + ], + "score": 1.0, + "content": "represents the extracted features by the the detector", + "type": "text" + }, + { + "bbox": [ + 365, + 517, + 375, + 526 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 515, + 463, + 530 + ], + "score": 1.0, + "content": "(specified later), from", + "type": "text" + }, + { + "bbox": [ + 464, + 516, + 501, + 528 + ], + "score": 0.92, + "content": "f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 515, + 506, + 530 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 526, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 205, + 540 + ], + "score": 1.0, + "content": "The minimization of the", + "type": "text" + }, + { + "bbox": [ + 205, + 527, + 215, + 538 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 526, + 237, + 540 + ], + "score": 1.0, + "content": "term", + "type": "text" + }, + { + "bbox": [ + 238, + 527, + 358, + 539 + ], + "score": 0.93, + "content": "| | D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) ^ { * } - D ( f _ { \\theta } ( \\mathbf { x } ) ) | | _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 526, + 506, + 540 + ], + "score": 1.0, + "content": "encourages the detection features of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 539, + 443, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 443, + 550 + ], + "score": 1.0, + "content": "adversarial and benign examples to be close, such that the detector could be evaded.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "We conduct the above adaptive attack on both CIFAR-10 and SVHN. Specifically, 1000 clean ex-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "amples are randomly selected from the test set of each database. For each generated adversarial", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 577, + 504, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 263, + 589 + ], + "score": 1.0, + "content": "example, if it fails to fool the classifier", + "type": "text" + }, + { + "bbox": [ + 264, + 577, + 275, + 588 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 578, + 504, + 589 + ], + "score": 1.0, + "content": ", we record 1. The attack failure rate over all 1000 adver-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "sarial examples is computed as the metric. The higher rate indicates that it is more difficult to evade", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 308, + 611 + ], + "score": 1.0, + "content": "the detector. Since LID performs much better than", + "type": "text" + }, + { + "bbox": [ + 308, + 599, + 343, + 609 + ], + "score": 0.8, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "and M-D in above experiments, here we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "only compare LID and the proposed MBF detector. For clarity, the detection features are only ex-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "tracted from the soft-max layer for both LID and MBF. The original CW-L2 attack (i.e., without the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 116, + 643 + ], + "score": 0.86, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "term) is also presented as the baseline. With the same setting in (Ma et al. (2018)), the trade-off", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 641, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 149, + 655 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 149, + 645, + 157, + 653 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 641, + 352, + 655 + ], + "score": 1.0, + "content": "is determined by binary search within the range", + "type": "text" + }, + { + "bbox": [ + 352, + 642, + 399, + 656 + ], + "score": 0.92, + "content": "[ \\bar { 1 0 } ^ { - 3 } , 1 0 ^ { 6 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 641, + 506, + 655 + ], + "score": 1.0, + "content": ". The results are shown in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "Table 5. The attack failure rates of MBF on both databases are much higher than those of LID. It", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 665, + 318, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 318, + 676 + ], + "score": 1.0, + "content": "tells that it is more difficult to evade MBF than LID.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 363, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 365, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 365, + 702 + ], + "score": 1.0, + "content": "4.5 HYPOTHESIS TEST VIA KOLMOGOROV-SMIRNOV TEST", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)) is a non-parametric test method in statis-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "tics, to test whether a sample follows a reference probability distribution (one-sample KS test), or", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 124, + 457, + 178 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 89, + 504, + 122 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 320, + 102 + ], + "score": 1.0, + "content": "Table 4: Detection results evaluated by AUROC score", + "type": "text" + }, + { + "bbox": [ + 321, + 90, + 336, + 100 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "in the data-transfer case. All detectors are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "score": 1.0, + "content": "trained on the train set, and tested on the out-of-sample set (including 365 images) of the ImageNet", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 110, + 307, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 307, + 123 + ], + "score": 1.0, + "content": "database. The best results are highlighted in bold.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 153, + 124, + 457, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 124, + 457, + 178 + ], + "spans": [ + { + "bbox": [ + 153, + 124, + 457, + 178 + ], + "score": 0.962, + "html": "
ImageNet-AlexNetImageNet-VGG-16
DetectorBIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
KD+BU66.369.464.768.884.085.079.985.2
M-D48.749.250.051.054.550.050.456.1
LID69.167.672.971.484.288.489.283.7
MBF99.199.099.699.499.399.599.399.5
", + "type": "table", + "image_path": "69efaee09e76d4a1b56e04fd81944dc90fd9797a28a35b7f953f4304f89b2339.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 153, + 124, + 457, + 142.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 153, + 142.0, + 457, + 160.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 153, + 160.0, + 457, + 178.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 226, + 198, + 385, + 226 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 185, + 504, + 196 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 184, + 504, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 224, + 196 + ], + "score": 1.0, + "content": "Table 5: Attack Failure Rates", + "type": "text" + }, + { + "bbox": [ + 224, + 185, + 239, + 196 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 184, + 504, + 196 + ], + "score": 1.0, + "content": "of adaptive attacks with different detectors. See context for details.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 226, + 198, + 385, + 226 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 226, + 198, + 385, + 226 + ], + "spans": [ + { + "bbox": [ + 226, + 198, + 385, + 226 + ], + "score": 0.956, + "html": "
Original CW-L2LIDMBF
CIFAR-100.088.4100.0
SVHN0.085.692.9
", + "type": "table", + "image_path": "b975bebdd9d344b5960d73838588cd6994286a09202d3cfa99fe01e1e30d4edd.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 198, + 385, + 212.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 226, + 212.0, + 385, + 226.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "text", + "bbox": [ + 107, + 232, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "CIFAR-10 and SVHN reported in (Lee et al. (2018)) are very high, but their networks (i.e., ResNet", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "score": 1.0, + "content": "(He et al. (2016)) and DenseNet (Huang et al. (2017)) are different from that used in our experiments.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 255, + 503, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 503, + 267 + ], + "score": 1.0, + "content": "It implies that M-D may be suitable for very deep neural networks, but not for shallower networks.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 232, + 506, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 271, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 272, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 283 + ], + "score": 1.0, + "content": "Detection results in the attack-transfer case evaluated by AUROC and accuracy are presented in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "Tables 2 and 3, respectively. MBF still shows much better performance in all transfer cases than all", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "compared methods, and the changes among detecting different attacks are very small. It verifies the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 304, + 264, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 264, + 315 + ], + "score": 1.0, + "content": "robustness of MBF to different attacks.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 272, + 505, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "score": 1.0, + "content": "We also conduct a data-transfer experiment on ImageNet. Specifically, we collect extra 365 images", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "of 8 classes (same with the classes used for the detection training, see Section 4.2) through searching", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "the class names in Baidu and Facebook. These 365 benign and their noisy images can be correctly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "predicted by both the fine-tuned AlexNet and VGG models. The results on these images are shown", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "in Table 4. MBF still shows the best performance. And, compared to the corresponding results in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "Table 1, the AUROC/accuracy scores of MBF on detecting different attacks change very gently. It", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 387, + 333, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 333, + 399 + ], + "score": 1.0, + "content": "verifies the robustness of MBF to different data sources.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 320, + 506, + 399 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 411, + 374, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 411, + 375, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 375, + 424 + ], + "score": 1.0, + "content": "4.4 ADAPTIVE ATTACK AGAINST ADVERSARIAL DETECTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 505, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "Similar to Carlini & Wagner (2017b) and Ma et al. (2018), we also present an adaptive white-box", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "attack, of which the goal is not only to fool the classifier, but also to evade the adversarial detector.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "It measures how easy to bypass the detector. Specifically, by combining the CW-L2 attack (Carlini", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 465, + 496, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 496, + 477 + ], + "score": 1.0, + "content": "& Wagner (2017a)) and the proposed MBF detector, the adaptive attack is formulated as follows:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 432, + 505, + 477 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 137, + 478, + 471, + 502 + ], + "lines": [ + { + "bbox": [ + 137, + 478, + 471, + 502 + ], + "spans": [ + { + "bbox": [ + 137, + 478, + 471, + 502 + ], + "score": 0.89, + "content": "\\operatorname * { a r g m i n } _ { \\mathbf { x } _ { a d v } \\in [ 0 , 1 ] } \\| \\mathbf { x } - \\mathbf { x } _ { a d v } \\| _ { 2 } ^ { 2 } + \\alpha \\cdot \\big [ - l ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) , y ) + \\| D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) - D ( f _ { \\theta } ( \\mathbf { x } ) ) \\| _ { 1 } \\big ] ,", + "type": "interline_equation", + "image_path": "ef0772f1a645b7f33bb82d6f5e88667e7a460e99dac440dd60000a4074a74b04.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 137, + 478, + 471, + 502 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 133, + 518 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 507, + 142, + 515 + ], + "score": 0.6, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 505, + 243, + 518 + ], + "score": 1.0, + "content": "denotes the clean input,", + "type": "text" + }, + { + "bbox": [ + 244, + 507, + 251, + 517 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 505, + 388, + 518 + ], + "score": 1.0, + "content": "indicates the ground-truth label.", + "type": "text" + }, + { + "bbox": [ + 388, + 505, + 429, + 517 + ], + "score": 0.93, + "content": "l ( \\mathbf { x } _ { a d v } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "is the Hinge loss.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 515, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 160, + 528 + ], + "score": 0.91, + "content": "D ( f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 515, + 365, + 530 + ], + "score": 1.0, + "content": "represents the extracted features by the the detector", + "type": "text" + }, + { + "bbox": [ + 365, + 517, + 375, + 526 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 515, + 463, + 530 + ], + "score": 1.0, + "content": "(specified later), from", + "type": "text" + }, + { + "bbox": [ + 464, + 516, + 501, + 528 + ], + "score": 0.92, + "content": "f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 515, + 506, + 530 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 526, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 205, + 540 + ], + "score": 1.0, + "content": "The minimization of the", + "type": "text" + }, + { + "bbox": [ + 205, + 527, + 215, + 538 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 526, + 237, + 540 + ], + "score": 1.0, + "content": "term", + "type": "text" + }, + { + "bbox": [ + 238, + 527, + 358, + 539 + ], + "score": 0.93, + "content": "| | D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) ^ { * } - D ( f _ { \\theta } ( \\mathbf { x } ) ) | | _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 526, + 506, + 540 + ], + "score": 1.0, + "content": "encourages the detection features of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 539, + 443, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 443, + 550 + ], + "score": 1.0, + "content": "adversarial and benign examples to be close, such that the detector could be evaded.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 505, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "We conduct the above adaptive attack on both CIFAR-10 and SVHN. Specifically, 1000 clean ex-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "amples are randomly selected from the test set of each database. For each generated adversarial", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 577, + 504, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 263, + 589 + ], + "score": 1.0, + "content": "example, if it fails to fool the classifier", + "type": "text" + }, + { + "bbox": [ + 264, + 577, + 275, + 588 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 578, + 504, + 589 + ], + "score": 1.0, + "content": ", we record 1. The attack failure rate over all 1000 adver-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "sarial examples is computed as the metric. The higher rate indicates that it is more difficult to evade", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 308, + 611 + ], + "score": 1.0, + "content": "the detector. Since LID performs much better than", + "type": "text" + }, + { + "bbox": [ + 308, + 599, + 343, + 609 + ], + "score": 0.8, + "content": "\\mathrm { K D + B U }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 598, + 506, + 611 + ], + "score": 1.0, + "content": "and M-D in above experiments, here we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "only compare LID and the proposed MBF detector. For clarity, the detection features are only ex-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "tracted from the soft-max layer for both LID and MBF. The original CW-L2 attack (i.e., without the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 116, + 643 + ], + "score": 0.86, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "term) is also presented as the baseline. With the same setting in (Ma et al. (2018)), the trade-off", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 641, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 149, + 655 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 149, + 645, + 157, + 653 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 641, + 352, + 655 + ], + "score": 1.0, + "content": "is determined by binary search within the range", + "type": "text" + }, + { + "bbox": [ + 352, + 642, + 399, + 656 + ], + "score": 0.92, + "content": "[ \\bar { 1 0 } ^ { - 3 } , 1 0 ^ { 6 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 641, + 506, + 655 + ], + "score": 1.0, + "content": ". The results are shown in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 506, + 666 + ], + "score": 1.0, + "content": "Table 5. The attack failure rates of MBF on both databases are much higher than those of LID. It", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 665, + 318, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 318, + 676 + ], + "score": 1.0, + "content": "tells that it is more difficult to evade MBF than LID.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 555, + 506, + 676 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 363, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 365, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 365, + 702 + ], + "score": 1.0, + "content": "4.5 HYPOTHESIS TEST VIA KOLMOGOROV-SMIRNOV TEST", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Kolmogorov-Smirnov test (KS test) (Massey Jr (1951)) is a non-parametric test method in statis-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "tics, to test whether a sample follows a reference probability distribution (one-sample KS test), or", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 260, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 504, + 271 + ], + "score": 1.0, + "content": "whether two samples follow the same distribution (two-sample KS test). Specifically, in one-sample", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 271, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 282 + ], + "score": 1.0, + "content": "KS test, the distance between the empirical distribution function of one sample and the cumulative", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 446, + 294 + ], + "score": 1.0, + "content": "distribution function of the reference probability distribution is measured. Then, the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 446, + 283, + 453, + 293 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 453, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "-value corre-", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 322, + 305 + ], + "score": 1.0, + "content": "sponding to the obtained distance is computed. If the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 322, + 294, + 329, + 304 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 329, + 292, + 496, + 305 + ], + "score": 1.0, + "content": "-value is larger than the significance level", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 496, + 294, + 504, + 302 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 158, + 315 + ], + "score": 1.0, + "content": "(here we set", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 158, + 304, + 199, + 314 + ], + "score": 0.86, + "content": "\\alpha = 0 . 0 5", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 200, + 304, + 505, + 315 + ], + "score": 1.0, + "content": "), then the null hypothesis that the sample follows the reference distribution", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 328 + ], + "score": 1.0, + "content": "is accepted; otherwise, rejected. Similarly, in two-sample KS test, the distance between the empir-", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 416, + 338 + ], + "score": 1.0, + "content": "ical distribution functions of two samples is computed. If the corresponding", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 417, + 327, + 424, + 337 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 424, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "-value is larger than", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 336, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 107, + 338, + 114, + 346 + ], + "score": 0.72, + "content": "\\alpha", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 114, + 336, + 506, + 348 + ], + "score": 1.0, + "content": ", then it accepts that two samples follow the same distribution. The KS test conducted below is", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 356, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 356, + 360 + ], + "score": 1.0, + "content": "implemented by the python function scipy.stats.ks 2samp7.", + "type": "text", + "cross_page": true + } + ], + "index": 21 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 113, + 454, + 146 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 162, + 102 + ], + "score": 1.0, + "content": "Table 6: The", + "type": "text" + }, + { + "bbox": [ + 163, + 91, + 169, + 101 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "-value of KS hypothesis test among clean samples, noisy samples, and adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 388, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 388, + 113 + ], + "score": 1.0, + "content": "samples crafted by four methods, respectively, on ImageNet-AlexNet.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 156, + 113, + 454, + 146 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 113, + 454, + 146 + ], + "spans": [ + { + "bbox": [ + 156, + 113, + 454, + 146 + ], + "score": 0.963, + "html": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
", + "type": "table", + "image_path": "f128add81b448d21c04af4561663c3d499b3b4be016b057dcf77c25e5527a83f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 156, + 113, + 454, + 124.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 156, + 124.0, + 454, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 156, + 135.0, + 454, + 146.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 130, + 167, + 479, + 199 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 153, + 504, + 165 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 159, + 167 + ], + "score": 1.0, + "content": "Table 7: The", + "type": "text" + }, + { + "bbox": [ + 159, + 155, + 165, + 165 + ], + "score": 0.76, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 151, + 505, + 167 + ], + "score": 1.0, + "content": "-value of two-sample KS test among MBF coefficients of different types of examples.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 130, + 167, + 479, + 199 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 130, + 167, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 479, + 199 + ], + "score": 0.948, + "html": "
ImageNet-VGG-16(clean, noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
", + "type": "table", + "image_path": "adf2d35488b42e10bd5631f045e040f1271834d43043993f162f1e8389239dc3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 130, + 167, + 479, + 177.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 130, + 177.66666666666666, + 479, + 188.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 130, + 188.33333333333331, + 479, + 198.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "table", + "bbox": [ + 138, + 220, + 469, + 252 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 119, + 206, + 492, + 218 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 118, + 205, + 493, + 219 + ], + "spans": [ + { + "bbox": [ + 118, + 205, + 171, + 219 + ], + "score": 1.0, + "content": "Table 8: The", + "type": "text" + }, + { + "bbox": [ + 171, + 208, + 178, + 218 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 205, + 493, + 219 + ], + "score": 1.0, + "content": "-value of KS hypothesis test among MBF coefficients of different data sources.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 138, + 220, + 469, + 252 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 138, + 220, + 469, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 469, + 252 + ], + "score": 0.85, + "html": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
(train set,test set)0.8390.8580.3680.8420.2460.221
(train set, out-of-sample set)0.0490.0540.0940.4070.7370.088
", + "type": "table", + "image_path": "16d5e9200efba8879ec80096b9d49580ef3e3677a6ac01a4aa00dc1b7da652c6.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 138, + 220, + 469, + 230.66666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 138, + 230.66666666666666, + 469, + 241.33333333333331 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 138, + 241.33333333333331, + 469, + 251.99999999999997 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 505, + 359 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 504, + 271 + ], + "score": 1.0, + "content": "whether two samples follow the same distribution (two-sample KS test). Specifically, in one-sample", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 271, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 282 + ], + "score": 1.0, + "content": "KS test, the distance between the empirical distribution function of one sample and the cumulative", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 446, + 294 + ], + "score": 1.0, + "content": "distribution function of the reference probability distribution is measured. Then, the", + "type": "text" + }, + { + "bbox": [ + 446, + 283, + 453, + 293 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "-value corre-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 292, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 322, + 305 + ], + "score": 1.0, + "content": "sponding to the obtained distance is computed. If the", + "type": "text" + }, + { + "bbox": [ + 322, + 294, + 329, + 304 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 292, + 496, + 305 + ], + "score": 1.0, + "content": "-value is larger than the significance level", + "type": "text" + }, + { + "bbox": [ + 496, + 294, + 504, + 302 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 158, + 315 + ], + "score": 1.0, + "content": "(here we set", + "type": "text" + }, + { + "bbox": [ + 158, + 304, + 199, + 314 + ], + "score": 0.86, + "content": "\\alpha = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 304, + 505, + 315 + ], + "score": 1.0, + "content": "), then the null hypothesis that the sample follows the reference distribution", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 328 + ], + "score": 1.0, + "content": "is accepted; otherwise, rejected. Similarly, in two-sample KS test, the distance between the empir-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 416, + 338 + ], + "score": 1.0, + "content": "ical distribution functions of two samples is computed. If the corresponding", + "type": "text" + }, + { + "bbox": [ + 417, + 327, + 424, + 337 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "-value is larger than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 336, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 107, + 338, + 114, + 346 + ], + "score": 0.72, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 336, + 506, + 348 + ], + "score": 1.0, + "content": ", then it accepts that two samples follow the same distribution. The KS test conducted below is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 347, + 356, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 356, + 360 + ], + "score": 1.0, + "content": "implemented by the python function scipy.stats.ks 2samp7.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 108, + 364, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "Hypothesis test 1. Here we verify that whether the posterior vectors of both adversarial and benign", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 448, + 388 + ], + "score": 1.0, + "content": "examples follow GGD. We denote the posterior vector of one adversarial example as", + "type": "text" + }, + { + "bbox": [ + 448, + 376, + 466, + 387 + ], + "score": 0.87, + "content": "\\bf p \\mathrm { _ { a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 375, + 506, + 388 + ], + "score": 1.0, + "content": ", and that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 209, + 398 + ], + "score": 1.0, + "content": "of one benign example as", + "type": "text" + }, + { + "bbox": [ + 210, + 387, + 228, + 398 + ], + "score": 0.83, + "content": "\\mathbf { p } _ { \\mathrm { { b e n } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 387, + 386, + 398 + ], + "score": 1.0, + "content": ". The distribution of GGD is denoted as", + "type": "text" + }, + { + "bbox": [ + 387, + 386, + 410, + 397 + ], + "score": 0.89, + "content": "{ \\mathcal { P } } _ { \\mathrm { G G D } }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 387, + 505, + 398 + ], + "score": 1.0, + "content": ". Then, we conduct two", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 397, + 236, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 236, + 410 + ], + "score": 1.0, + "content": "one-sample KS tests, including:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 132, + 416, + 477, + 443 + ], + "lines": [ + { + "bbox": [ + 130, + 414, + 478, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 299, + 431 + ], + "score": 1.0, + "content": "• H1.1 The test of adversarial examples:", + "type": "text" + }, + { + "bbox": [ + 299, + 417, + 312, + 428 + ], + "score": 0.83, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 414, + 316, + 431 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 317, + 417, + 383, + 429 + ], + "score": 0.78, + "content": "\\mathbf { p } _ { \\mathrm { a d v } } \\sim \\mathcal { P } _ { \\mathrm { G G D - a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 414, + 389, + 431 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 389, + 417, + 402, + 428 + ], + "score": 0.72, + "content": "H _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 414, + 478, + 431 + ], + "score": 1.0, + "content": ": padv 6∼ PGGD-adv.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 131, + 429, + 461, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 429, + 282, + 445 + ], + "score": 1.0, + "content": "• H1.2 The test of benign examples:", + "type": "text" + }, + { + "bbox": [ + 282, + 432, + 296, + 443 + ], + "score": 0.77, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 429, + 300, + 445 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 300, + 432, + 367, + 443 + ], + "score": 0.72, + "content": "\\mathbf { p } _ { \\mathrm { b e n } } \\sim \\mathcal { P } _ { \\mathrm { G G D - b e n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 429, + 372, + 445 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 373, + 432, + 386, + 443 + ], + "score": 0.53, + "content": "H _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 429, + 461, + 445 + ], + "score": 1.0, + "content": ": pben 6∼ PGGD-ben.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 254, + 464 + ], + "score": 1.0, + "content": "In H1.1, the reference distribution", + "type": "text" + }, + { + "bbox": [ + 254, + 451, + 291, + 462 + ], + "score": 0.9, + "content": "\\mathcal { P } _ { \\mathrm { G G D - a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 451, + 397, + 464 + ], + "score": 1.0, + "content": "is firstly estimated from", + "type": "text" + }, + { + "bbox": [ + 398, + 453, + 416, + 462 + ], + "score": 0.87, + "content": "\\mathbf { p } _ { \\mathrm { a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 451, + 505, + 464 + ], + "score": 1.0, + "content": ", using the estimated", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "method proposed in (Lasmar et al. (2009)). To alleviate the uncertainty of the estimation, we draw", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 471, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 237, + 487 + ], + "score": 1.0, + "content": "500 samples from the estimated", + "type": "text" + }, + { + "bbox": [ + 237, + 473, + 273, + 484 + ], + "score": 0.89, + "content": "\\mathcal { P } _ { \\mathrm { G G D - a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 471, + 485, + 487 + ], + "score": 1.0, + "content": ". Then, we conduct the two-sample KS test between", + "type": "text" + }, + { + "bbox": [ + 486, + 475, + 504, + 484 + ], + "score": 0.79, + "content": "\\mathbf { p } _ { \\mathrm { a d v } }", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 304, + 497 + ], + "score": 1.0, + "content": "and these 500 samples respectively. The average", + "type": "text" + }, + { + "bbox": [ + 305, + 486, + 311, + 496 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "-value over 1000 tests is recorded. The mean of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 155, + 507 + ], + "score": 1.0, + "content": "the average", + "type": "text" + }, + { + "bbox": [ + 155, + 497, + 161, + 507 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "-values over the whole database is reported. H1.2 is conducted similarly. The results", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "tested on ImageNet-AlexNet are shown in Table 6. Due to the space limit, results on other databases", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 428, + 530 + ], + "score": 1.0, + "content": "and models will be presented in the supplementary material. In all cases, the", + "type": "text" + }, + { + "bbox": [ + 428, + 519, + 434, + 529 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "-values are larger", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "than the significance level 0.05. Hence, we can conclude that the posterior vectors of both adver-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "sarial and benign examples follow GGD. However, note that the parameters of their corresponding", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 549, + 358, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 358, + 563 + ], + "score": 1.0, + "content": "GGD are different, which will be verified in the following test.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "Hypothesis test 2. Here we verify that whether the extracted MBF features of adversarial and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "benign examples follow the same empirical distribution. We denote the MBF feature vector of one", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 104, + 588, + 201, + 604 + ], + "score": 1.0, + "content": "adversarial example as", + "type": "text" + }, + { + "bbox": [ + 201, + 592, + 222, + 601 + ], + "score": 0.88, + "content": "\\mathbf { m _ { \\mathrm { a d v } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 588, + 465, + 604 + ], + "score": 1.0, + "content": ", and the corresponding empirical distribution is denoted as", + "type": "text" + }, + { + "bbox": [ + 465, + 588, + 501, + 601 + ], + "score": 0.79, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 588, + 506, + 604 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 189, + 615 + ], + "score": 1.0, + "content": "Similarly, we define", + "type": "text" + }, + { + "bbox": [ + 190, + 604, + 211, + 614 + ], + "score": 0.91, + "content": "\\mathbf { m } _ { \\mathrm { b e n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 601, + 230, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 230, + 601, + 266, + 614 + ], + "score": 0.85, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "for benign examples. Then, we conduct the following four", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 614, + 194, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 194, + 626 + ], + "score": 1.0, + "content": "two-sample KS tests:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 132, + 634, + 505, + 716 + ], + "lines": [ + { + "bbox": [ + 131, + 633, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 131, + 634, + 374, + 649 + ], + "score": 1.0, + "content": "• H2.1 The test between adversarial and benign examples:", + "type": "text" + }, + { + "bbox": [ + 374, + 635, + 387, + 646 + ], + "score": 0.73, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 634, + 392, + 649 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 392, + 633, + 477, + 647 + ], + "score": 0.8, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 634, + 487, + 649 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 487, + 635, + 501, + 647 + ], + "score": 0.85, + "content": "H _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 634, + 505, + 649 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 142, + 646, + 231, + 662 + ], + "spans": [ + { + "bbox": [ + 142, + 646, + 226, + 660 + ], + "score": 0.91, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 646, + 231, + 662 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 661, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 132, + 661, + 487, + 675 + ], + "score": 1.0, + "content": "• H2.2 The test between adversarial examples crafted from different attack methods:", + "type": "text" + }, + { + "bbox": [ + 487, + 663, + 501, + 673 + ], + "score": 0.88, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 661, + 505, + 675 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 142, + 668, + 345, + 693 + ], + "spans": [ + { + "bbox": [ + 142, + 673, + 225, + 688 + ], + "score": 0.87, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 668, + 234, + 693 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 234, + 673, + 336, + 688 + ], + "score": 0.73, + "content": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 668, + 345, + 693 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 138, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 138, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "H2.3 The test between adversarial examples from different data sources (i.e., train, test and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 138, + 696, + 435, + 720 + ], + "spans": [ + { + "bbox": [ + 138, + 696, + 220, + 720 + ], + "score": 1.0, + "content": "out-of-sample set):", + "type": "text" + }, + { + "bbox": [ + 221, + 702, + 234, + 713 + ], + "score": 0.83, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 696, + 239, + 720 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 239, + 700, + 321, + 715 + ], + "score": 0.81, + "content": "\\begin{array} { r } { \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 707, + 327, + 715 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 330, + 700, + 432, + 715 + ], + "score": 0.75, + "content": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 707, + 435, + 715 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 119, + 721, + 431, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 719, + 432, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 719, + 432, + 734 + ], + "score": 1.0, + "content": "7https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.stats.ks 2samp.html", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 113, + 454, + 146 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 162, + 102 + ], + "score": 1.0, + "content": "Table 6: The", + "type": "text" + }, + { + "bbox": [ + 163, + 91, + 169, + 101 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "-value of KS hypothesis test among clean samples, noisy samples, and adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 388, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 388, + 113 + ], + "score": 1.0, + "content": "samples crafted by four methods, respectively, on ImageNet-AlexNet.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 156, + 113, + 454, + 146 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 113, + 454, + 146 + ], + "spans": [ + { + "bbox": [ + 156, + 113, + 454, + 146 + ], + "score": 0.963, + "html": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
", + "type": "table", + "image_path": "f128add81b448d21c04af4561663c3d499b3b4be016b057dcf77c25e5527a83f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 156, + 113, + 454, + 124.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 156, + 124.0, + 454, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 156, + 135.0, + 454, + 146.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 130, + 167, + 479, + 199 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 153, + 504, + 165 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 159, + 167 + ], + "score": 1.0, + "content": "Table 7: The", + "type": "text" + }, + { + "bbox": [ + 159, + 155, + 165, + 165 + ], + "score": 0.76, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 151, + 505, + 167 + ], + "score": 1.0, + "content": "-value of two-sample KS test among MBF coefficients of different types of examples.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 130, + 167, + 479, + 199 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 130, + 167, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 479, + 199 + ], + "score": 0.948, + "html": "
ImageNet-VGG-16(clean, noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
", + "type": "table", + "image_path": "adf2d35488b42e10bd5631f045e040f1271834d43043993f162f1e8389239dc3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 130, + 167, + 479, + 177.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 130, + 177.66666666666666, + 479, + 188.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 130, + 188.33333333333331, + 479, + 198.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "table", + "bbox": [ + 138, + 220, + 469, + 252 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 119, + 206, + 492, + 218 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 118, + 205, + 493, + 219 + ], + "spans": [ + { + "bbox": [ + 118, + 205, + 171, + 219 + ], + "score": 1.0, + "content": "Table 8: The", + "type": "text" + }, + { + "bbox": [ + 171, + 208, + 178, + 218 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 205, + 493, + 219 + ], + "score": 1.0, + "content": "-value of KS hypothesis test among MBF coefficients of different data sources.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 138, + 220, + 469, + 252 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 138, + 220, + 469, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 469, + 252 + ], + "score": 0.85, + "html": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
(train set,test set)0.8390.8580.3680.8420.2460.221
(train set, out-of-sample set)0.0490.0540.0940.4070.7370.088
", + "type": "table", + "image_path": "16d5e9200efba8879ec80096b9d49580ef3e3677a6ac01a4aa00dc1b7da652c6.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 138, + 220, + 469, + 230.66666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 138, + 230.66666666666666, + 469, + 241.33333333333331 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 138, + 241.33333333333331, + 469, + 251.99999999999997 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 505, + 359 + ], + "lines": [], + "index": 17, + "bbox_fs": [ + 105, + 260, + 506, + 360 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 364, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "Hypothesis test 1. Here we verify that whether the posterior vectors of both adversarial and benign", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 448, + 388 + ], + "score": 1.0, + "content": "examples follow GGD. We denote the posterior vector of one adversarial example as", + "type": "text" + }, + { + "bbox": [ + 448, + 376, + 466, + 387 + ], + "score": 0.87, + "content": "\\bf p \\mathrm { _ { a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 375, + 506, + 388 + ], + "score": 1.0, + "content": ", and that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 209, + 398 + ], + "score": 1.0, + "content": "of one benign example as", + "type": "text" + }, + { + "bbox": [ + 210, + 387, + 228, + 398 + ], + "score": 0.83, + "content": "\\mathbf { p } _ { \\mathrm { { b e n } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 387, + 386, + 398 + ], + "score": 1.0, + "content": ". The distribution of GGD is denoted as", + "type": "text" + }, + { + "bbox": [ + 387, + 386, + 410, + 397 + ], + "score": 0.89, + "content": "{ \\mathcal { P } } _ { \\mathrm { G G D } }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 387, + 505, + 398 + ], + "score": 1.0, + "content": ". Then, we conduct two", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 397, + 236, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 236, + 410 + ], + "score": 1.0, + "content": "one-sample KS tests, including:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 364, + 506, + 410 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 416, + 477, + 443 + ], + "lines": [ + { + "bbox": [ + 130, + 414, + 478, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 299, + 431 + ], + "score": 1.0, + "content": "• H1.1 The test of adversarial examples:", + "type": "text" + }, + { + "bbox": [ + 299, + 417, + 312, + 428 + ], + "score": 0.83, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 414, + 316, + 431 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 317, + 417, + 383, + 429 + ], + "score": 0.78, + "content": "\\mathbf { p } _ { \\mathrm { a d v } } \\sim \\mathcal { P } _ { \\mathrm { G G D - a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 414, + 389, + 431 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 389, + 417, + 402, + 428 + ], + "score": 0.72, + "content": "H _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 414, + 478, + 431 + ], + "score": 1.0, + "content": ": padv 6∼ PGGD-adv.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 429, + 461, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 429, + 282, + 445 + ], + "score": 1.0, + "content": "• H1.2 The test of benign examples:", + "type": "text" + }, + { + "bbox": [ + 282, + 432, + 296, + 443 + ], + "score": 0.77, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 429, + 300, + 445 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 300, + 432, + 367, + 443 + ], + "score": 0.72, + "content": "\\mathbf { p } _ { \\mathrm { b e n } } \\sim \\mathcal { P } _ { \\mathrm { G G D - b e n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 429, + 372, + 445 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 373, + 432, + 386, + 443 + ], + "score": 0.53, + "content": "H _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 429, + 461, + 445 + ], + "score": 1.0, + "content": ": pben 6∼ PGGD-ben.", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 26.5, + "bbox_fs": [ + 130, + 414, + 478, + 445 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 451, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 254, + 464 + ], + "score": 1.0, + "content": "In H1.1, the reference distribution", + "type": "text" + }, + { + "bbox": [ + 254, + 451, + 291, + 462 + ], + "score": 0.9, + "content": "\\mathcal { P } _ { \\mathrm { G G D - a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 451, + 397, + 464 + ], + "score": 1.0, + "content": "is firstly estimated from", + "type": "text" + }, + { + "bbox": [ + 398, + 453, + 416, + 462 + ], + "score": 0.87, + "content": "\\mathbf { p } _ { \\mathrm { a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 451, + 505, + 464 + ], + "score": 1.0, + "content": ", using the estimated", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "method proposed in (Lasmar et al. (2009)). To alleviate the uncertainty of the estimation, we draw", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 471, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 237, + 487 + ], + "score": 1.0, + "content": "500 samples from the estimated", + "type": "text" + }, + { + "bbox": [ + 237, + 473, + 273, + 484 + ], + "score": 0.89, + "content": "\\mathcal { P } _ { \\mathrm { G G D - a d v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 471, + 485, + 487 + ], + "score": 1.0, + "content": ". Then, we conduct the two-sample KS test between", + "type": "text" + }, + { + "bbox": [ + 486, + 475, + 504, + 484 + ], + "score": 0.79, + "content": "\\mathbf { p } _ { \\mathrm { a d v } }", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 304, + 497 + ], + "score": 1.0, + "content": "and these 500 samples respectively. The average", + "type": "text" + }, + { + "bbox": [ + 305, + 486, + 311, + 496 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "-value over 1000 tests is recorded. The mean of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 155, + 507 + ], + "score": 1.0, + "content": "the average", + "type": "text" + }, + { + "bbox": [ + 155, + 497, + 161, + 507 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "-values over the whole database is reported. H1.2 is conducted similarly. The results", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "tested on ImageNet-AlexNet are shown in Table 6. Due to the space limit, results on other databases", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 428, + 530 + ], + "score": 1.0, + "content": "and models will be presented in the supplementary material. In all cases, the", + "type": "text" + }, + { + "bbox": [ + 428, + 519, + 434, + 529 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "-values are larger", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "than the significance level 0.05. Hence, we can conclude that the posterior vectors of both adver-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "sarial and benign examples follow GGD. However, note that the parameters of their corresponding", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 549, + 358, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 358, + 563 + ], + "score": 1.0, + "content": "GGD are different, which will be verified in the following test.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32.5, + "bbox_fs": [ + 104, + 451, + 506, + 563 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "Hypothesis test 2. Here we verify that whether the extracted MBF features of adversarial and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "benign examples follow the same empirical distribution. We denote the MBF feature vector of one", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 104, + 588, + 201, + 604 + ], + "score": 1.0, + "content": "adversarial example as", + "type": "text" + }, + { + "bbox": [ + 201, + 592, + 222, + 601 + ], + "score": 0.88, + "content": "\\mathbf { m _ { \\mathrm { a d v } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 588, + 465, + 604 + ], + "score": 1.0, + "content": ", and the corresponding empirical distribution is denoted as", + "type": "text" + }, + { + "bbox": [ + 465, + 588, + 501, + 601 + ], + "score": 0.79, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 588, + 506, + 604 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 189, + 615 + ], + "score": 1.0, + "content": "Similarly, we define", + "type": "text" + }, + { + "bbox": [ + 190, + 604, + 211, + 614 + ], + "score": 0.91, + "content": "\\mathbf { m } _ { \\mathrm { b e n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 601, + 230, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 230, + 601, + 266, + 614 + ], + "score": 0.85, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "for benign examples. Then, we conduct the following four", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 614, + 194, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 194, + 626 + ], + "score": 1.0, + "content": "two-sample KS tests:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 567, + 506, + 626 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 634, + 505, + 716 + ], + "lines": [ + { + "bbox": [ + 131, + 633, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 131, + 634, + 374, + 649 + ], + "score": 1.0, + "content": "• H2.1 The test between adversarial and benign examples:", + "type": "text" + }, + { + "bbox": [ + 374, + 635, + 387, + 646 + ], + "score": 0.73, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 634, + 392, + 649 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 392, + 633, + 477, + 647 + ], + "score": 0.8, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 634, + 487, + 649 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 487, + 635, + 501, + 647 + ], + "score": 0.85, + "content": "H _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 634, + 505, + 649 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 646, + 231, + 662 + ], + "spans": [ + { + "bbox": [ + 142, + 646, + 226, + 660 + ], + "score": 0.91, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 646, + 231, + 662 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 661, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 132, + 661, + 487, + 675 + ], + "score": 1.0, + "content": "• H2.2 The test between adversarial examples crafted from different attack methods:", + "type": "text" + }, + { + "bbox": [ + 487, + 663, + 501, + 673 + ], + "score": 0.88, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 661, + 505, + 675 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 668, + 345, + 693 + ], + "spans": [ + { + "bbox": [ + 142, + 673, + 225, + 688 + ], + "score": 0.87, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 668, + 234, + 693 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 234, + 673, + 336, + 688 + ], + "score": 0.73, + "content": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 668, + 345, + 693 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 138, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 138, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "H2.3 The test between adversarial examples from different data sources (i.e., train, test and", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 138, + 696, + 435, + 720 + ], + "spans": [ + { + "bbox": [ + 138, + 696, + 220, + 720 + ], + "score": 1.0, + "content": "out-of-sample set):", + "type": "text" + }, + { + "bbox": [ + 221, + 702, + 234, + 713 + ], + "score": 0.83, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 696, + 239, + 720 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 239, + 700, + 321, + 715 + ], + "score": 0.81, + "content": "\\begin{array} { r } { \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 707, + 327, + 715 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 330, + 700, + 432, + 715 + ], + "score": 0.75, + "content": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 707, + 435, + 715 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 45.5, + "bbox_fs": [ + 131, + 633, + 505, + 720 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "score": 0.974, + "type": "image", + "image_path": "e6e248dabc29baa792b249691b317cc1ab671db8b1d5f02fc3e9a7bc0e6db03a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 83, + 473, + 126.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 126.33333333333334, + 473, + 169.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 169.66666666666669, + 473, + 213.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 217, + 504, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 212, + 230 + ], + "score": 1.0, + "content": "Figure 1: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 212, + 218, + 222, + 228 + ], + "score": 0.58, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 217, + 505, + 230 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row), test (median", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 228, + 367, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 367, + 241 + ], + "score": 1.0, + "content": "row), and out-of-sample (bottom row) set of ImageNet-VGG-16.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 133, + 248, + 503, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "• H2.4 The test between benign examples from different data sources (i.e., train, test and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 254, + 438, + 279 + ], + "spans": [ + { + "bbox": [ + 137, + 254, + 220, + 279 + ], + "score": 1.0, + "content": "out-of-sample set):", + "type": "text" + }, + { + "bbox": [ + 221, + 260, + 234, + 272 + ], + "score": 0.83, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 254, + 239, + 279 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 239, + 259, + 321, + 274 + ], + "score": 0.79, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 254, + 330, + 279 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 330, + 259, + 433, + 274 + ], + "score": 0.65, + "content": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 254, + 438, + 279 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "In above four two-sample KS tests, we test on the 16-dimensional MBF features extracted from the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "score": 1.0, + "content": "soft-max layer. Since the implementation scipy.stats.ks 2samp cannot compare two vectors, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 304, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 408, + 316 + ], + "score": 1.0, + "content": "compare the feature of each dimension separately, then report the average", + "type": "text" + }, + { + "bbox": [ + 408, + 306, + 414, + 316 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 304, + 505, + 316 + ], + "score": 1.0, + "content": "-value over all dimen-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "sions. Specifically, when comparing two sets of samples, we firstly concatenate the feature of each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "dimension across all samples in the same set, leading to 16 long vectors for each set. Then, each pair", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 337, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 505, + 349 + ], + "score": 1.0, + "content": "of two long vectors corresponding to the same dimension from two sets is compared by KS test. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 347, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 139, + 361 + ], + "score": 1.0, + "content": "average", + "type": "text" + }, + { + "bbox": [ + 140, + 349, + 146, + 359 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 347, + 331, + 361 + ], + "score": 1.0, + "content": "-value over all 16 dimensions is reported. The", + "type": "text" + }, + { + "bbox": [ + 331, + 349, + 338, + 359 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 347, + 378, + 361 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + }, + { + "bbox": [ + 378, + 348, + 400, + 358 + ], + "score": 0.28, + "content": "{ \\bf H } 2 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 347, + 506, + 361 + ], + "score": 1.0, + "content": "are shown in Table 7 (see", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 358, + 504, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 241, + 371 + ], + "score": 1.0, + "content": "the column “(clean, BIM)”). The", + "type": "text" + }, + { + "bbox": [ + 241, + 361, + 248, + 370 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 358, + 489, + 371 + ], + "score": 1.0, + "content": "-values on both train and test set are 0. Thus, the hypothesis", + "type": "text" + }, + { + "bbox": [ + 490, + 359, + 504, + 370 + ], + "score": 0.88, + "content": "H _ { 0 }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "is rejected, i.e., the MBF features of adversarial and benign examples follow different distributions.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 125, + 393 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 125, + 382, + 132, + 392 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 381, + 174, + 393 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + }, + { + "bbox": [ + 174, + 380, + 196, + 391 + ], + "score": 0.32, + "content": "{ \\bf H } 2 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "are shown in Table 7. We pick two groups of attack methods, i.e., (BIM,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 267, + 403 + ], + "score": 1.0, + "content": "DeepFool) and (CW-L2, R-PGD). The", + "type": "text" + }, + { + "bbox": [ + 267, + 393, + 273, + 403 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 390, + 457, + 403 + ], + "score": 1.0, + "content": "-values of (BIM, DeepFool) are 0, while the", + "type": "text" + }, + { + "bbox": [ + 457, + 393, + 464, + 403 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "(CW-L2, R-PGD) are larger than 0.05. It demonstrates that the distributions of adversarial examples", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 391, + 426 + ], + "score": 1.0, + "content": "crafted from different attack methods are possible to be different. The", + "type": "text" + }, + { + "bbox": [ + 391, + 415, + 398, + 425 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 413, + 438, + 426 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + }, + { + "bbox": [ + 439, + 414, + 461, + 424 + ], + "score": 0.3, + "content": "{ \\bf H } 2 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "are shown", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 171, + 437 + ], + "score": 1.0, + "content": "in Table 8. The", + "type": "text" + }, + { + "bbox": [ + 171, + 425, + 178, + 436 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "-values of all types of adversarial examples exceed 0.05. Thus, the MBF features", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 457, + 447 + ], + "score": 1.0, + "content": "of adversarial examples from different data sources follow the same distribution. The", + "type": "text" + }, + { + "bbox": [ + 457, + 437, + 464, + 447 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 254, + 459 + ], + "score": 1.0, + "content": "H2.4 are shown in Table 8. Only the", + "type": "text" + }, + { + "bbox": [ + 254, + 448, + 261, + 457 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 445, + 506, + 459 + ], + "score": 1.0, + "content": "-value of “(train set, out-of-sample set)” of clean examples is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "slightly lower than 0.05, while the values of other cases exceed 0.05. Thus, in most cases, the MBF", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 468, + 447, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 447, + 480 + ], + "score": 1.0, + "content": "features of benign examples from different data sources follow the same distribution.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "From above analysis, we obtain the following conclusions: 1) The extracted MBF features of adver-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 496, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 508 + ], + "score": 1.0, + "content": "sarial and benign examples follow different empirical distributions. It explains why MBF features", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "are effective for detecting adversarial and benign examples; 2) The extracted MBF features of ad-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "versarial/benign examples from the train, test and out-of-sample sets follow the same empirical", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "distribution. Although the extracted MBF features of adversarial examples crafted from different", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 541, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 551 + ], + "score": 1.0, + "content": "attack methods may not follow the same empirical distribution, the significant difference between", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "benign and different adversarial distributions can still lead to the good detection performance in the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "attack-transfer case. It explains why MBF features are robust across different attack methods and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 571, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 586 + ], + "score": 1.0, + "content": "different data sources. Moreover, we visualize the statistics of each dimension of MBF features, i.e.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 584, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 595 + ], + "score": 1.0, + "content": "mean and standard deviation, as shown in Fig. 1. These visualizations also support above conclu-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "sions. Due to space limit, more KS tests and visualizations on different databases and networks will", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 606, + 291, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 291, + 618 + ], + "score": 1.0, + "content": "be presented in the supplementary material.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 624, + 195, + 637 + ], + "lines": [ + { + "bbox": [ + 104, + 622, + 198, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 622, + 198, + 641 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "This work has proposed a novel adversarial detection method, dubbed MBF. The assumption behind", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "is that the internal responses of the classification network of both adversarial and benign examples", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "follow the generalized Gaussian distribution (GGD), but with different shape factors. The magni-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "tude of Benford-Fourier coefficient is a function w.r.t. the shape factor, and can be easily estimated", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "based on responses. Thus, it can serve as the discriminative features between adversarial and be-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "nign examples. The extensive experiments conducted on several databases, as well as the empirical", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "analysis via KS test, demonstrate the superior effectiveness and robustness to different attacks and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 477, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 477, + 732 + ], + "score": 1.0, + "content": "different data sources of the proposed MBF method, over state-of-the-art detection methods.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "score": 0.974, + "type": "image", + "image_path": "e6e248dabc29baa792b249691b317cc1ab671db8b1d5f02fc3e9a7bc0e6db03a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 83, + 473, + 126.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 126.33333333333334, + 473, + 169.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 169.66666666666669, + 473, + 213.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 217, + 504, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 212, + 230 + ], + "score": 1.0, + "content": "Figure 1: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 212, + 218, + 222, + 228 + ], + "score": 0.58, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 217, + 505, + 230 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row), test (median", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 228, + 367, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 367, + 241 + ], + "score": 1.0, + "content": "row), and out-of-sample (bottom row) set of ImageNet-VGG-16.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 133, + 248, + 503, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "• H2.4 The test between benign examples from different data sources (i.e., train, test and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 254, + 438, + 279 + ], + "spans": [ + { + "bbox": [ + 137, + 254, + 220, + 279 + ], + "score": 1.0, + "content": "out-of-sample set):", + "type": "text" + }, + { + "bbox": [ + 221, + 260, + 234, + 272 + ], + "score": 0.83, + "content": "H _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 254, + 239, + 279 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 239, + 259, + 321, + 274 + ], + "score": 0.79, + "content": "\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 254, + 330, + 279 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 330, + 259, + 433, + 274 + ], + "score": 0.65, + "content": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 254, + 438, + 279 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 131, + 248, + 505, + 279 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "In above four two-sample KS tests, we test on the 16-dimensional MBF features extracted from the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 306 + ], + "score": 1.0, + "content": "soft-max layer. Since the implementation scipy.stats.ks 2samp cannot compare two vectors, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 304, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 408, + 316 + ], + "score": 1.0, + "content": "compare the feature of each dimension separately, then report the average", + "type": "text" + }, + { + "bbox": [ + 408, + 306, + 414, + 316 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 304, + 505, + 316 + ], + "score": 1.0, + "content": "-value over all dimen-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "sions. Specifically, when comparing two sets of samples, we firstly concatenate the feature of each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "dimension across all samples in the same set, leading to 16 long vectors for each set. Then, each pair", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 337, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 505, + 349 + ], + "score": 1.0, + "content": "of two long vectors corresponding to the same dimension from two sets is compared by KS test. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 347, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 139, + 361 + ], + "score": 1.0, + "content": "average", + "type": "text" + }, + { + "bbox": [ + 140, + 349, + 146, + 359 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 347, + 331, + 361 + ], + "score": 1.0, + "content": "-value over all 16 dimensions is reported. The", + "type": "text" + }, + { + "bbox": [ + 331, + 349, + 338, + 359 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 347, + 378, + 361 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + }, + { + "bbox": [ + 378, + 348, + 400, + 358 + ], + "score": 0.28, + "content": "{ \\bf H } 2 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 347, + 506, + 361 + ], + "score": 1.0, + "content": "are shown in Table 7 (see", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 358, + 504, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 241, + 371 + ], + "score": 1.0, + "content": "the column “(clean, BIM)”). The", + "type": "text" + }, + { + "bbox": [ + 241, + 361, + 248, + 370 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 358, + 489, + 371 + ], + "score": 1.0, + "content": "-values on both train and test set are 0. Thus, the hypothesis", + "type": "text" + }, + { + "bbox": [ + 490, + 359, + 504, + 370 + ], + "score": 0.88, + "content": "H _ { 0 }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "is rejected, i.e., the MBF features of adversarial and benign examples follow different distributions.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 125, + 393 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 125, + 382, + 132, + 392 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 381, + 174, + 393 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + }, + { + "bbox": [ + 174, + 380, + 196, + 391 + ], + "score": 0.32, + "content": "{ \\bf H } 2 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "are shown in Table 7. We pick two groups of attack methods, i.e., (BIM,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 267, + 403 + ], + "score": 1.0, + "content": "DeepFool) and (CW-L2, R-PGD). The", + "type": "text" + }, + { + "bbox": [ + 267, + 393, + 273, + 403 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 390, + 457, + 403 + ], + "score": 1.0, + "content": "-values of (BIM, DeepFool) are 0, while the", + "type": "text" + }, + { + "bbox": [ + 457, + 393, + 464, + 403 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "(CW-L2, R-PGD) are larger than 0.05. It demonstrates that the distributions of adversarial examples", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 391, + 426 + ], + "score": 1.0, + "content": "crafted from different attack methods are possible to be different. The", + "type": "text" + }, + { + "bbox": [ + 391, + 415, + 398, + 425 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 413, + 438, + 426 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + }, + { + "bbox": [ + 439, + 414, + 461, + 424 + ], + "score": 0.3, + "content": "{ \\bf H } 2 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "are shown", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 171, + 437 + ], + "score": 1.0, + "content": "in Table 8. The", + "type": "text" + }, + { + "bbox": [ + 171, + 425, + 178, + 436 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "-values of all types of adversarial examples exceed 0.05. Thus, the MBF features", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 457, + 447 + ], + "score": 1.0, + "content": "of adversarial examples from different data sources follow the same distribution. The", + "type": "text" + }, + { + "bbox": [ + 457, + 437, + 464, + 447 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "-values of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 254, + 459 + ], + "score": 1.0, + "content": "H2.4 are shown in Table 8. Only the", + "type": "text" + }, + { + "bbox": [ + 254, + 448, + 261, + 457 + ], + "score": 0.82, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 445, + 506, + 459 + ], + "score": 1.0, + "content": "-value of “(train set, out-of-sample set)” of clean examples is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 469 + ], + "score": 1.0, + "content": "slightly lower than 0.05, while the values of other cases exceed 0.05. Thus, in most cases, the MBF", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 468, + 447, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 447, + 480 + ], + "score": 1.0, + "content": "features of benign examples from different data sources follow the same distribution.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 282, + 506, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "From above analysis, we obtain the following conclusions: 1) The extracted MBF features of adver-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 496, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 508 + ], + "score": 1.0, + "content": "sarial and benign examples follow different empirical distributions. It explains why MBF features", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "are effective for detecting adversarial and benign examples; 2) The extracted MBF features of ad-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "versarial/benign examples from the train, test and out-of-sample sets follow the same empirical", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "distribution. Although the extracted MBF features of adversarial examples crafted from different", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 541, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 551 + ], + "score": 1.0, + "content": "attack methods may not follow the same empirical distribution, the significant difference between", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "benign and different adversarial distributions can still lead to the good detection performance in the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "attack-transfer case. It explains why MBF features are robust across different attack methods and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 571, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 586 + ], + "score": 1.0, + "content": "different data sources. Moreover, we visualize the statistics of each dimension of MBF features, i.e.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 584, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 595 + ], + "score": 1.0, + "content": "mean and standard deviation, as shown in Fig. 1. These visualizations also support above conclu-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "sions. Due to space limit, more KS tests and visualizations on different databases and networks will", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 606, + 291, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 291, + 618 + ], + "score": 1.0, + "content": "be presented in the supplementary material.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 485, + 506, + 618 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 624, + 195, + 637 + ], + "lines": [ + { + "bbox": [ + 104, + 622, + 198, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 622, + 198, + 641 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "This work has proposed a novel adversarial detection method, dubbed MBF. The assumption behind", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "is that the internal responses of the classification network of both adversarial and benign examples", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "follow the generalized Gaussian distribution (GGD), but with different shape factors. The magni-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "tude of Benford-Fourier coefficient is a function w.r.t. the shape factor, and can be easily estimated", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "based on responses. Thus, it can serve as the discriminative features between adversarial and be-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "nign examples. The extensive experiments conducted on several databases, as well as the empirical", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "analysis via KS test, demonstrate the superior effectiveness and robustness to different attacks and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 477, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 477, + 732 + ], + "score": 1.0, + "content": "different data sources of the proposed MBF method, over state-of-the-art detection methods.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 644, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 100, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 112 + ], + "score": 1.0, + "content": "Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 390, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 390, + 123 + ], + "score": 1.0, + "content": "IEEE Symposium on Security and Privacy, pp. 39–57. IEEE, 2017a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 142 + ], + "score": 1.0, + "content": "Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 116, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 151, + 251, + 163 + ], + "spans": [ + { + "bbox": [ + 116, + 151, + 251, + 163 + ], + "score": 1.0, + "content": "Security, pp. 3–14. ACM, 2017b.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 169, + 504, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 504, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 504, + 182 + ], + "score": 1.0, + "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 191, + 218, + 203 + ], + "spans": [ + { + "bbox": [ + 115, + 191, + 218, + 203 + ], + "score": 1.0, + "content": "pp. 248–255. Ieee, 2009.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 209, + 505, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "score": 1.0, + "content": "Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 221, + 374, + 233 + ], + "spans": [ + { + "bbox": [ + 116, + 221, + 374, + 233 + ], + "score": 1.0, + "content": "samples from artifacts. arXiv preprint arXiv:1703.00410, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 105, + 239, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "score": 1.0, + "content": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 250, + 317, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 317, + 263 + ], + "score": 1.0, + "content": "examples. arXiv preprint arXiv:1412.6572, 2014.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 504, + 292 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "Nathaniel R Goodman. Statistical analysis based on a certain multivariate complex gaussian distri-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 279, + 461, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 279, + 461, + 292 + ], + "score": 1.0, + "content": "bution (an introduction). The Annals of mathematical statistics, 34(1):152–177, 1963.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 504, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and Patrick McDaniel. On", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 309, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 116, + 309, + 480, + 322 + ], + "score": 1.0, + "content": "the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 342 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 337, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 115, + 337, + 506, + 353 + ], + "score": 1.0, + "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 349, + 181, + 361 + ], + "spans": [ + { + "bbox": [ + 117, + 349, + 181, + 361 + ], + "score": 1.0, + "content": "770–778, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 104, + 367, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 382 + ], + "score": 1.0, + "content": "Dan Hendrycks and Kevin Gimpel. Early methods for detecting adversarial images. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 379, + 219, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 219, + 391 + ], + "score": 1.0, + "content": "arXiv:1608.00530, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 397, + 504, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 420, + 257, + 432 + ], + "spans": [ + { + "bbox": [ + 116, + 420, + 257, + 432 + ], + "score": 1.0, + "content": "recognition, pp. 4700–4708, 2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 437, + 503, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 459, + 143, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 143, + 471 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 103, + 478, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 347, + 491 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton.", + "type": "text" + }, + { + "bbox": [ + 362, + 477, + 460, + 491 + ], + "score": 1.0, + "content": "The cifar-10 dataset.", + "type": "text" + }, + { + "bbox": [ + 472, + 477, + 506, + 491 + ], + "score": 1.0, + "content": "online:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 488, + 323, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 323, + 502 + ], + "score": 1.0, + "content": "http://www.cs.toronto.edu/kriz/cifar.html, 55, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 105, + 507, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 504, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 504, + 520 + ], + "score": 1.0, + "content": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 519, + 278, + 530 + ], + "spans": [ + { + "bbox": [ + 116, + 519, + 278, + 530 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1607.02533, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Nour-Eddine Lasmar, Youssef Stitou, and Yannick Berthoumieu. Multiscale skewed heavy tailed", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 546, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 505, + 562 + ], + "score": 1.0, + "content": "model for texture analysis. In 2009 16th IEEE International Conference on Image Processing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 559, + 264, + 571 + ], + "spans": [ + { + "bbox": [ + 117, + 559, + 264, + 571 + ], + "score": 1.0, + "content": "(ICIP), pp. 2281–2284. IEEE, 2009.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 591 + ], + "score": 1.0, + "content": "Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 116, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "out-of-distribution samples and adversarial attacks. In Advances in Neural Information Process-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 599, + 258, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 258, + 612 + ], + "score": 1.0, + "content": "ing Systems, pp. 7167–7177, 2018.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764–", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 638, + 168, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 168, + 652 + ], + "score": 1.0, + "content": "5772, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "Jiajun Lu, Theerasit Issaranon, and David Forsyth. Safetynet: Detecting and rejecting adversarial", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 667, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 114, + 667, + 506, + 684 + ], + "score": 1.0, + "content": "examples robustly. In Proceedings of the IEEE International Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 680, + 182, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 182, + 691 + ], + "score": 1.0, + "content": "446–454, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 720, + 497, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 497, + 733 + ], + "score": 1.0, + "content": "local intrinsic dimensionality. In International Conference on Learning Representations, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 104, + 100, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 112 + ], + "score": 1.0, + "content": "Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 390, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 390, + 123 + ], + "score": 1.0, + "content": "IEEE Symposium on Security and Privacy, pp. 39–57. IEEE, 2017a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 99, + 505, + 123 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 142 + ], + "score": 1.0, + "content": "Nicholas Carlini and David Wagner. Adversarial examples are not easily detected: Bypassing ten", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 116, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 151, + 251, + 163 + ], + "spans": [ + { + "bbox": [ + 116, + 151, + 251, + 163 + ], + "score": 1.0, + "content": "Security, pp. 3–14. ACM, 2017b.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 128, + 505, + 163 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 169, + 504, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 504, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 504, + 182 + ], + "score": 1.0, + "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 191, + 218, + 203 + ], + "spans": [ + { + "bbox": [ + 115, + 191, + 218, + 203 + ], + "score": 1.0, + "content": "pp. 248–255. Ieee, 2009.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 169, + 505, + 203 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 209, + 505, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 223 + ], + "score": 1.0, + "content": "Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 221, + 374, + 233 + ], + "spans": [ + { + "bbox": [ + 116, + 221, + 374, + 233 + ], + "score": 1.0, + "content": "samples from artifacts. arXiv preprint arXiv:1703.00410, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 209, + 505, + 233 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 239, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 253 + ], + "score": 1.0, + "content": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 250, + 317, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 317, + 263 + ], + "score": 1.0, + "content": "examples. arXiv preprint arXiv:1412.6572, 2014.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 238, + 505, + 263 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 504, + 292 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "Nathaniel R Goodman. Statistical analysis based on a certain multivariate complex gaussian distri-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 279, + 461, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 279, + 461, + 292 + ], + "score": 1.0, + "content": "bution (an introduction). The Annals of mathematical statistics, 34(1):152–177, 1963.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 268, + 505, + 292 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 504, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and Patrick McDaniel. On", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 309, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 116, + 309, + 480, + 322 + ], + "score": 1.0, + "content": "the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 297, + 506, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 342 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 337, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 115, + 337, + 506, + 353 + ], + "score": 1.0, + "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 349, + 181, + 361 + ], + "spans": [ + { + "bbox": [ + 117, + 349, + 181, + 361 + ], + "score": 1.0, + "content": "770–778, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 325, + 506, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 367, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 382 + ], + "score": 1.0, + "content": "Dan Hendrycks and Kevin Gimpel. Early methods for detecting adversarial images. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 379, + 219, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 219, + 391 + ], + "score": 1.0, + "content": "arXiv:1608.00530, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 366, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 397, + 504, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 420, + 257, + 432 + ], + "spans": [ + { + "bbox": [ + 116, + 420, + 257, + 432 + ], + "score": 1.0, + "content": "recognition, pp. 4700–4708, 2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 396, + 506, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 437, + 503, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 459, + 143, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 143, + 471 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 437, + 505, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 478, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 347, + 491 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton.", + "type": "text" + }, + { + "bbox": [ + 362, + 477, + 460, + 491 + ], + "score": 1.0, + "content": "The cifar-10 dataset.", + "type": "text" + }, + { + "bbox": [ + 472, + 477, + 506, + 491 + ], + "score": 1.0, + "content": "online:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 488, + 323, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 323, + 502 + ], + "score": 1.0, + "content": "http://www.cs.toronto.edu/kriz/cifar.html, 55, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 477, + 506, + 502 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 507, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 504, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 504, + 520 + ], + "score": 1.0, + "content": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 519, + 278, + 530 + ], + "spans": [ + { + "bbox": [ + 116, + 519, + 278, + 530 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1607.02533, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 507, + 504, + 530 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Nour-Eddine Lasmar, Youssef Stitou, and Yannick Berthoumieu. Multiscale skewed heavy tailed", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 546, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 505, + 562 + ], + "score": 1.0, + "content": "model for texture analysis. In 2009 16th IEEE International Conference on Image Processing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 559, + 264, + 571 + ], + "spans": [ + { + "bbox": [ + 117, + 559, + 264, + 571 + ], + "score": 1.0, + "content": "(ICIP), pp. 2281–2284. IEEE, 2009.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 537, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 591 + ], + "score": 1.0, + "content": "Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 116, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "out-of-distribution samples and adversarial attacks. In Advances in Neural Information Process-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 599, + 258, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 258, + 612 + ], + "score": 1.0, + "content": "ing Systems, pp. 7167–7177, 2018.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 576, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Xin Li and Fuxin Li. Adversarial examples detection in deep networks with convolutional filter", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "statistics. In Proceedings of the IEEE International Conference on Computer Vision, pp. 5764–", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 638, + 168, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 168, + 652 + ], + "score": 1.0, + "content": "5772, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 617, + 505, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "Jiajun Lu, Theerasit Issaranon, and David Forsyth. Safetynet: Detecting and rejecting adversarial", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 667, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 114, + 667, + 506, + 684 + ], + "score": 1.0, + "content": "examples robustly. In Proceedings of the IEEE International Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 680, + 182, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 182, + 691 + ], + "score": 1.0, + "content": "446–454, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 657, + 506, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 720, + 497, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 497, + 733 + ], + "score": 1.0, + "content": "local intrinsic dimensionality. In International Conference on Learning Representations, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 699, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 144, + 116 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 144, + 116 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "Frank J Massey Jr. The kolmogorov-smirnov test for goodness of fit. Journal of the American", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 298, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 298, + 146 + ], + "score": 1.0, + "content": "statistical Association, 46(253):68–78, 1951.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Jan Hendrik Metzen, Tim Genewein, Volker Fischer, and Bastian Bischoff. On detecting adversarial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 435, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 435, + 176 + ], + "score": 1.0, + "content": "perturbations. In International Conference on Learning Representations, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 109, + 182, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "accurate method to fool deep neural networks. In Proceedings of the IEEE conference on com-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 206, + 355, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 206, + 355, + 217 + ], + "score": 1.0, + "content": "puter vision and pattern recognition, pp. 2574–2582, 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 108, + 223, + 503, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 237 + ], + "score": 1.0, + "content": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 382, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 382, + 247 + ], + "score": 1.0, + "content": "digits in natural images with unsupervised feature learning. 2011.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 108, + 253, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "Tianyu Pang, Chao Du, Yinpeng Dong, and Jun Zhu. Towards robust detection of adversarial exam-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 453, + 277 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 453, + 277 + ], + "score": 1.0, + "content": "ples. In Advances in Neural Information Processing Systems, pp. 4579–4589, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 504, + 308 + ], + "score": 1.0, + "content": "Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Sympo-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 306, + 345, + 317 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 345, + 317 + ], + "score": 1.0, + "content": "sium on Security and Privacy, pp. 372–387. IEEE, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "A. Papoulis and S.U. Pillai. Probability, Random Variables, and Stochastic Processes. McGraw-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "Hill Series in Electrical and Computer Engineering. McGraw-Hill, 2001. ISBN 9780072817256.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 345, + 437, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 345, + 354, + 360 + ], + "score": 1.0, + "content": "URL https://books.google.com.hk/books?id", + "type": "text" + }, + { + "bbox": [ + 355, + 348, + 361, + 356 + ], + "score": 0.4, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 345, + 437, + 360 + ], + "score": 1.0, + "content": "Mal4OgAACAAJ.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 379 + ], + "score": 1.0, + "content": "Cecilia Pasquini, Fernando Perez-Gonz ´ alez, and Giulia Boato. A benford-fourier jpeg compression ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 116, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "detector. In 2014 IEEE International Conference on Image Processing, pp. 5322–5326. IEEE,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 385, + 143, + 400 + ], + "spans": [ + { + "bbox": [ + 115, + 385, + 143, + 400 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 503, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 417, + 497, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 497, + 430 + ], + "score": 1.0, + "content": "Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–572, 1901.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 105, + 435, + 504, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 450 + ], + "score": 1.0, + "content": "Fernando Perez-Gonz ´ alez, Greg L Heileman, and Chaouki T Abdallah. Benford’s lawin image ´", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "processing. In 2007 IEEE International Conference on Image Processing, volume 1, pp. I–405.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 458, + 168, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 168, + 469 + ], + "score": 1.0, + "content": "IEEE, 2007.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 104, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "Murray Rosenblatt. A central limit theorem and a strong mixing condition. Proceedings of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 487, + 433, + 500 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 433, + 500 + ], + "score": 1.0, + "content": "National Academy of Sciences of the United States of America, 42(1):43, 1956.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 105, + 506, + 504, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 504, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 504, + 519 + ], + "score": 1.0, + "content": "Pouya Samangouei, Maya Kabkab, and Rama Chellappa. Defense-gan: Protecting classifiers against", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 517, + 457, + 530 + ], + "spans": [ + { + "bbox": [ + 116, + 517, + 457, + 530 + ], + "score": 1.0, + "content": "adversarial attacks using generative models. arXiv preprint arXiv:1805.06605, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 105, + 536, + 504, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 550 + ], + "score": 1.0, + "content": "Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 545, + 144, + 560 + ], + "spans": [ + { + "bbox": [ + 114, + 545, + 144, + 560 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 105, + 566, + 503, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 504, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 504, + 580 + ], + "score": 1.0, + "content": "Mahesh K Varanasi and Behnaam Aazhang. Parametric generalized gaussian density estimation.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 577, + 416, + 590 + ], + "spans": [ + { + "bbox": [ + 116, + 577, + 416, + 590 + ], + "score": 1.0, + "content": "The Journal of the Acoustical Society of America, 86(4):1404–1415, 1989.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "score": 1.0, + "content": "Zhihao Zheng and Pengyu Hong. Robust detection of adversarial attacks by modeling the intrinsic", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 606, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 606, + 506, + 622 + ], + "score": 1.0, + "content": "properties of deep neural networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 618, + 191, + 630 + ], + "spans": [ + { + "bbox": [ + 117, + 618, + 191, + 630 + ], + "score": 1.0, + "content": "7913–7922, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 144, + 116 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 144, + 116 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "Frank J Massey Jr. The kolmogorov-smirnov test for goodness of fit. Journal of the American", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 298, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 298, + 146 + ], + "score": 1.0, + "content": "statistical Association, 46(253):68–78, 1951.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 122, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Jan Hendrik Metzen, Tim Genewein, Volker Fischer, and Bastian Bischoff. On detecting adversarial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 435, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 435, + 176 + ], + "score": 1.0, + "content": "perturbations. In International Conference on Learning Representations, 2017.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 153, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 182, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "accurate method to fool deep neural networks. In Proceedings of the IEEE conference on com-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 206, + 355, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 206, + 355, + 217 + ], + "score": 1.0, + "content": "puter vision and pattern recognition, pp. 2574–2582, 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 182, + 505, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 223, + 503, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 222, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 237 + ], + "score": 1.0, + "content": "Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 382, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 382, + 247 + ], + "score": 1.0, + "content": "digits in natural images with unsupervised feature learning. 2011.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 222, + 505, + 247 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 253, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "Tianyu Pang, Chao Du, Yinpeng Dong, and Jun Zhu. Towards robust detection of adversarial exam-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 453, + 277 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 453, + 277 + ], + "score": 1.0, + "content": "ples. In Advances in Neural Information Processing Systems, pp. 4579–4589, 2018.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 252, + 505, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 504, + 308 + ], + "score": 1.0, + "content": "Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Sympo-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 306, + 345, + 317 + ], + "spans": [ + { + "bbox": [ + 115, + 306, + 345, + 317 + ], + "score": 1.0, + "content": "sium on Security and Privacy, pp. 372–387. IEEE, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 284, + 505, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "A. Papoulis and S.U. Pillai. Probability, Random Variables, and Stochastic Processes. McGraw-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "Hill Series in Electrical and Computer Engineering. McGraw-Hill, 2001. ISBN 9780072817256.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 345, + 437, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 345, + 354, + 360 + ], + "score": 1.0, + "content": "URL https://books.google.com.hk/books?id", + "type": "text" + }, + { + "bbox": [ + 355, + 348, + 361, + 356 + ], + "score": 0.4, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 345, + 437, + 360 + ], + "score": 1.0, + "content": "Mal4OgAACAAJ.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 325, + 505, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 365, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 379 + ], + "score": 1.0, + "content": "Cecilia Pasquini, Fernando Perez-Gonz ´ alez, and Giulia Boato. A benford-fourier jpeg compression ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 116, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "detector. In 2014 IEEE International Conference on Image Processing, pp. 5322–5326. IEEE,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 385, + 143, + 400 + ], + "spans": [ + { + "bbox": [ + 115, + 385, + 143, + 400 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 364, + 505, + 400 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 503, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Karl Pearson. Liii. on lines and planes of closest fit to systems of points in space. The London,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 417, + 497, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 497, + 430 + ], + "score": 1.0, + "content": "Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–572, 1901.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 406, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 435, + 504, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 450 + ], + "score": 1.0, + "content": "Fernando Perez-Gonz ´ alez, Greg L Heileman, and Chaouki T Abdallah. Benford’s lawin image ´", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "processing. In 2007 IEEE International Conference on Image Processing, volume 1, pp. I–405.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 458, + 168, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 168, + 469 + ], + "score": 1.0, + "content": "IEEE, 2007.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 434, + 506, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "Murray Rosenblatt. A central limit theorem and a strong mixing condition. Proceedings of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 487, + 433, + 500 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 433, + 500 + ], + "score": 1.0, + "content": "National Academy of Sciences of the United States of America, 42(1):43, 1956.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 476, + 505, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 506, + 504, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 504, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 504, + 519 + ], + "score": 1.0, + "content": "Pouya Samangouei, Maya Kabkab, and Rama Chellappa. Defense-gan: Protecting classifiers against", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 517, + 457, + 530 + ], + "spans": [ + { + "bbox": [ + 116, + 517, + 457, + 530 + ], + "score": 1.0, + "content": "adversarial attacks using generative models. arXiv preprint arXiv:1805.06605, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 507, + 504, + 530 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 536, + 504, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 550 + ], + "score": 1.0, + "content": "Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 545, + 144, + 560 + ], + "spans": [ + { + "bbox": [ + 114, + 545, + 144, + 560 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 534, + 505, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 566, + 503, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 504, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 504, + 580 + ], + "score": 1.0, + "content": "Mahesh K Varanasi and Behnaam Aazhang. Parametric generalized gaussian density estimation.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 577, + 416, + 590 + ], + "spans": [ + { + "bbox": [ + 116, + 577, + 416, + 590 + ], + "score": 1.0, + "content": "The Journal of the Acoustical Society of America, 86(4):1404–1415, 1989.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 566, + 504, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 596, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 610 + ], + "score": 1.0, + "content": "Zhihao Zheng and Pengyu Hong. Robust detection of adversarial attacks by modeling the intrinsic", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 606, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 606, + 506, + 622 + ], + "score": 1.0, + "content": "properties of deep neural networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 618, + 191, + 630 + ], + "spans": [ + { + "bbox": [ + 117, + 618, + 191, + 630 + ], + "score": 1.0, + "content": "7913–7922, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 595, + 506, + 630 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 183, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 106, + 230, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "score": 1.0, + "content": "A.1 PROOF OF THEOREM 1", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 104, + 126, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 104, + 126, + 263, + 143 + ], + "score": 1.0, + "content": "Proof 1 Assume that all the variables", + "type": "text" + }, + { + "bbox": [ + 264, + 128, + 325, + 140 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { m } \\} _ { m = 1 , \\dots , M }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 126, + 505, + 143 + ], + "score": 1.0, + "content": "are independent and identically distributed.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 462, + 153 + ], + "score": 1.0, + "content": "Applying the central limit theorem Rosenblatt (1956) to the real and imaginary parts of", + "type": "text" + }, + { + "bbox": [ + 462, + 142, + 470, + 150 + ], + "score": 0.78, + "content": "_ { \\mathcal { V } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 138, + 505, + 153 + ], + "score": 1.0, + "content": ", we can", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 151, + 388, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 388, + 162 + ], + "score": 1.0, + "content": "obtain that both parts asymptotically follow the Gaussian distribution", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 168, + 392, + 201 + ], + "lines": [ + { + "bbox": [ + 218, + 168, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 218, + 168, + 392, + 201 + ], + "score": 0.94, + "content": "\\mathcal { V } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { T } _ { m } \\sim \\mathcal { N } \\big ( E ( \\mathcal { T } _ { 1 } ) , \\frac { 1 } { M } D ( \\mathcal { T } _ { 1 } ) \\big ) ,", + "type": "interline_equation", + "image_path": "0270ee8c2464771f30bcfc969534922fd98e9655c21d659640419a9685d918e8.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 168, + 392, + 184.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 218, + 184.5, + 392, + 201.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 132, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 135, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 135, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 228, + 469, + 282 + ], + "lines": [ + { + "bbox": [ + 142, + 228, + 469, + 282 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 469, + 282 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle E ( \\mathcal { T } _ { 1 } ) = E \\left( e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right) = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { X } _ { 1 } } ( x ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } | x | } d x = a _ { n } , } \\\\ { \\displaystyle D ( \\mathcal { T } _ { 1 } ) = E \\big ( \\left| \\mathcal { T } _ { 1 } \\right| ^ { 2 } \\big ) - \\left| E \\left( \\mathcal { T } _ { 1 } \\right) \\right| ^ { 2 } = E \\big ( \\left| e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right| ^ { 2 } \\big ) - \\left| a _ { n } \\right| ^ { 2 } = 1 - \\left| a _ { n } \\right| ^ { 2 } . } \\end{array}", + "type": "interline_equation", + "image_path": "9029be1948795385fe01c66baae9a81fcd9762a05d42f2e2145cbbb6f6e9dc44.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 142, + 228, + 469, + 246.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 142, + 246.0, + 469, + 264.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 142, + 264.0, + 469, + 282.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 291, + 278, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 279, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 156, + 305 + ], + "score": 1.0, + "content": "The PDF of", + "type": "text" + }, + { + "bbox": [ + 156, + 292, + 165, + 303 + ], + "score": 0.66, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 290, + 279, + 305 + ], + "score": 1.0, + "content": "can be rewritten as follows:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 309, + 355, + 334 + ], + "lines": [ + { + "bbox": [ + 256, + 309, + 355, + 334 + ], + "spans": [ + { + "bbox": [ + 256, + 309, + 355, + 334 + ], + "score": 0.94, + "content": "\\mathcal { V } \\sim \\mathcal { N } ( a _ { n } , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) .", + "type": "interline_equation", + "image_path": "6ae7c6afcab816a8de62e4a29f5e9edd698fde894a0135aed7697535dd698b4e.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 256, + 309, + 355, + 334 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 172, + 350 + ], + "lines": [ + { + "bbox": [ + 106, + 337, + 173, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 173, + 351 + ], + "score": 1.0, + "content": "Thus, we obtain", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 355, + 351, + 380 + ], + "lines": [ + { + "bbox": [ + 259, + 355, + 351, + 380 + ], + "spans": [ + { + "bbox": [ + 259, + 355, + 351, + 380 + ], + "score": 0.95, + "content": "\\mathcal { E } \\sim \\mathcal { N } ( 0 , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) .", + "type": "interline_equation", + "image_path": "1ee7357ccb4d65fbe42ccc1b5b87617528b90b7da44dfb9d109cac24144be28f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 259, + 355, + 351, + 380 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 327, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 327, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 306, + 404 + ], + "score": 1.0, + "content": "Besides, the pseudo variance Goodman (1963) of", + "type": "text" + }, + { + "bbox": [ + 307, + 393, + 316, + 403 + ], + "score": 0.85, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 391, + 327, + 404 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 409, + 447, + 426 + ], + "lines": [ + { + "bbox": [ + 164, + 409, + 447, + 426 + ], + "spans": [ + { + "bbox": [ + 164, + 409, + 447, + 426 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = E ( \\mathcal { T } _ { 1 } ^ { 2 } ) - E ( \\mathcal { T } _ { 1 } ) ^ { 2 } = E \\big ( e ^ { - j 2 \\pi n \\log _ { 1 0 } \\left| \\mathcal { X } _ { 1 } ^ { 2 } \\right| } \\big ) - a _ { n } ^ { 2 } = a _ { 2 n } - a _ { n } ^ { 2 } .", + "type": "interline_equation", + "image_path": "eed2d9e58d94f0dd4558d74684a3fda7e7d3f02d41e8a792433ae66aacdad552.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 164, + 409, + 447, + 426 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 177, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 178, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 178, + 446 + ], + "score": 1.0, + "content": "Correspondingly,", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 448, + 371, + 473 + ], + "lines": [ + { + "bbox": [ + 240, + 448, + 371, + 473 + ], + "spans": [ + { + "bbox": [ + 240, + 448, + 371, + 473 + ], + "score": 0.94, + "content": "\\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = \\frac { 1 } { M } \\mathcal { I } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = \\frac { a _ { 2 n } - a _ { n } ^ { 2 } } { M } .", + "type": "interline_equation", + "image_path": "c3744e73aa7ab5407eb57adfce801a8d059323fd7875630eb1763d2f2f8f11be.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 240, + 448, + 371, + 473 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 131, + 493 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 479, + 173, + 492 + ], + "score": 0.91, + "content": "\\left( a _ { 2 n } - a _ { n } ^ { 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 479, + 256, + 493 + ], + "score": 1.0, + "content": "is bounded, we have", + "type": "text" + }, + { + "bbox": [ + 257, + 480, + 335, + 492 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\operatorname* { l i m } _ { M \\to \\infty } \\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 479, + 506, + 493 + ], + "score": 1.0, + "content": ". Thus, we obtain that the random variable", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 489, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 107, + 493, + 113, + 501 + ], + "score": 0.72, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 489, + 506, + 505 + ], + "score": 1.0, + "content": "follows circularly-symmetric complex Gaussian distribution, because the sufficient and necessary", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "condition is that mean value and pseudo variance equal zero Goodman (1963). It implies that both", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 232, + 525 + ], + "score": 1.0, + "content": "the real and imaginary part of", + "type": "text" + }, + { + "bbox": [ + 233, + 513, + 240, + 523 + ], + "score": 0.71, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "follow the same Gaussian distribution and they are independent.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Thus, the magnitude of this complex random variable follows the Rayleigh distribution Papoulis &", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 534, + 418, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 317, + 547 + ], + "score": 1.0, + "content": "Pillai (2001), and the probability density function of", + "type": "text" + }, + { + "bbox": [ + 318, + 535, + 330, + 546 + ], + "score": 0.89, + "content": "| \\mathcal { E } |", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 534, + 418, + 547 + ], + "score": 1.0, + "content": "can be formulated as", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 552, + 353, + 573 + ], + "lines": [ + { + "bbox": [ + 257, + 552, + 353, + 573 + ], + "spans": [ + { + "bbox": [ + 257, + 552, + 353, + 573 + ], + "score": 0.94, + "content": "\\mathcal { P } _ { | \\varepsilon | } \\left( r \\right) = \\frac { r } { s ^ { 2 } } e ^ { - r ^ { 2 } / 2 s ^ { 2 } } ,", + "type": "interline_equation", + "image_path": "8fcd3d6f2ab95407bc14119a563351950db51c91de02d82b5c69332fca8e2fdf.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 257, + 552, + 353, + 573 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "where s is the scale parameter. Knowing the properties of the Rayleigh distributionPapoulis & Pillai", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 588, + 177, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 177, + 602 + ], + "score": 1.0, + "content": "(2001), we have:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 606, + 361, + 631 + ], + "lines": [ + { + "bbox": [ + 250, + 606, + 361, + 631 + ], + "spans": [ + { + "bbox": [ + 250, + 606, + 361, + 631 + ], + "score": 0.94, + "content": "s ^ { 2 } = { \\frac { 1 } { 2 } } D ( { \\mathcal { E } } ) = { \\frac { 1 - | a _ { n } | ^ { 2 } } { 2 M } } .", + "type": "interline_equation", + "image_path": "666b6315498def3bdc6e938b8a4b4c9847345e478a1039a9a2be73c12da1da1c.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 250, + 606, + 361, + 631 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 196, + 660 + ], + "score": 1.0, + "content": "Utilizing the fact that", + "type": "text" + }, + { + "bbox": [ + 196, + 643, + 218, + 656 + ], + "score": 0.92, + "content": "\\left| a _ { n } \\right| ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 639, + 450, + 660 + ], + "score": 1.0, + "content": "is close to 0 when n is a modest number, we obtain that", + "type": "text" + }, + { + "bbox": [ + 450, + 642, + 501, + 657 + ], + "score": 0.93, + "content": "\\begin{array} { r } { D \\left( \\mathcal { E } \\right) \\approx \\frac { 1 } { M } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 639, + 506, + 660 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 653, + 260, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 149, + 673 + ], + "score": 1.0, + "content": "leading to", + "type": "text" + }, + { + "bbox": [ + 150, + 656, + 187, + 670 + ], + "score": 0.93, + "content": "\\textstyle s ^ { 2 } = { \\frac { 1 } { 2 M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 653, + 260, + 673 + ], + "score": 1.0, + "content": ". Then, we obtain", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 675, + 384, + 703 + ], + "lines": [ + { + "bbox": [ + 227, + 675, + 384, + 703 + ], + "spans": [ + { + "bbox": [ + 227, + 675, + 384, + 703 + ], + "score": 0.95, + "content": "E { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 4 - \\pi } { 4 M } } .", + "type": "interline_equation", + "image_path": "c8e8e8cb610860d6011ba2f448bfc677e4f4ac66bebe896f1cefb665bde6c066.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 227, + 675, + 384, + 703 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 504, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 225, + 722 + ], + "score": 1.0, + "content": "It is easy to observe that both", + "type": "text" + }, + { + "bbox": [ + 226, + 710, + 258, + 723 + ], + "score": 0.91, + "content": "E { \\big ( } | { \\mathcal { E } } | { \\big ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 708, + 277, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 277, + 710, + 310, + 723 + ], + "score": 0.93, + "content": "D ( \\left. \\boldsymbol { \\mathcal { E } } \\right. )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "are close to zero when the number of samples M", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 720, + 429, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 289, + 733 + ], + "score": 1.0, + "content": "increases. It implies that the estimation error", + "type": "text" + }, + { + "bbox": [ + 289, + 725, + 300, + 732 + ], + "score": 0.88, + "content": "\\varepsilon _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 720, + 357, + 733 + ], + "score": 1.0, + "content": "gets closer to", + "type": "text" + }, + { + "bbox": [ + 357, + 722, + 363, + 730 + ], + "score": 0.61, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 720, + 375, + 733 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 376, + 721, + 386, + 730 + ], + "score": 0.27, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 720, + 429, + 733 + ], + "score": 1.0, + "content": "increases.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 183, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 106, + 230, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "score": 1.0, + "content": "A.1 PROOF OF THEOREM 1", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 106, + 231, + 119 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 104, + 126, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 104, + 126, + 263, + 143 + ], + "score": 1.0, + "content": "Proof 1 Assume that all the variables", + "type": "text" + }, + { + "bbox": [ + 264, + 128, + 325, + 140 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { m } \\} _ { m = 1 , \\dots , M }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 126, + 505, + 143 + ], + "score": 1.0, + "content": "are independent and identically distributed.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 462, + 153 + ], + "score": 1.0, + "content": "Applying the central limit theorem Rosenblatt (1956) to the real and imaginary parts of", + "type": "text" + }, + { + "bbox": [ + 462, + 142, + 470, + 150 + ], + "score": 0.78, + "content": "_ { \\mathcal { V } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 138, + 505, + 153 + ], + "score": 1.0, + "content": ", we can", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 151, + 388, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 388, + 162 + ], + "score": 1.0, + "content": "obtain that both parts asymptotically follow the Gaussian distribution", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 126, + 505, + 162 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 168, + 392, + 201 + ], + "lines": [ + { + "bbox": [ + 218, + 168, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 218, + 168, + 392, + 201 + ], + "score": 0.94, + "content": "\\mathcal { V } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { T } _ { m } \\sim \\mathcal { N } \\big ( E ( \\mathcal { T } _ { 1 } ) , \\frac { 1 } { M } D ( \\mathcal { T } _ { 1 } ) \\big ) ,", + "type": "interline_equation", + "image_path": "0270ee8c2464771f30bcfc969534922fd98e9655c21d659640419a9685d918e8.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 168, + 392, + 184.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 218, + 184.5, + 392, + 201.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 132, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 135, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 135, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 212, + 135, + 225 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 228, + 469, + 282 + ], + "lines": [ + { + "bbox": [ + 142, + 228, + 469, + 282 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 469, + 282 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle E ( \\mathcal { T } _ { 1 } ) = E \\left( e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right) = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { X } _ { 1 } } ( x ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } | x | } d x = a _ { n } , } \\\\ { \\displaystyle D ( \\mathcal { T } _ { 1 } ) = E \\big ( \\left| \\mathcal { T } _ { 1 } \\right| ^ { 2 } \\big ) - \\left| E \\left( \\mathcal { T } _ { 1 } \\right) \\right| ^ { 2 } = E \\big ( \\left| e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right| ^ { 2 } \\big ) - \\left| a _ { n } \\right| ^ { 2 } = 1 - \\left| a _ { n } \\right| ^ { 2 } . } \\end{array}", + "type": "interline_equation", + "image_path": "9029be1948795385fe01c66baae9a81fcd9762a05d42f2e2145cbbb6f6e9dc44.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 142, + 228, + 469, + 246.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 142, + 246.0, + 469, + 264.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 142, + 264.0, + 469, + 282.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 291, + 278, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 279, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 156, + 305 + ], + "score": 1.0, + "content": "The PDF of", + "type": "text" + }, + { + "bbox": [ + 156, + 292, + 165, + 303 + ], + "score": 0.66, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 290, + 279, + 305 + ], + "score": 1.0, + "content": "can be rewritten as follows:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 290, + 279, + 305 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 309, + 355, + 334 + ], + "lines": [ + { + "bbox": [ + 256, + 309, + 355, + 334 + ], + "spans": [ + { + "bbox": [ + 256, + 309, + 355, + 334 + ], + "score": 0.94, + "content": "\\mathcal { V } \\sim \\mathcal { N } ( a _ { n } , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) .", + "type": "interline_equation", + "image_path": "6ae7c6afcab816a8de62e4a29f5e9edd698fde894a0135aed7697535dd698b4e.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 256, + 309, + 355, + 334 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 172, + 350 + ], + "lines": [ + { + "bbox": [ + 106, + 337, + 173, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 173, + 351 + ], + "score": 1.0, + "content": "Thus, we obtain", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 337, + 173, + 351 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 355, + 351, + 380 + ], + "lines": [ + { + "bbox": [ + 259, + 355, + 351, + 380 + ], + "spans": [ + { + "bbox": [ + 259, + 355, + 351, + 380 + ], + "score": 0.95, + "content": "\\mathcal { E } \\sim \\mathcal { N } ( 0 , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) .", + "type": "interline_equation", + "image_path": "1ee7357ccb4d65fbe42ccc1b5b87617528b90b7da44dfb9d109cac24144be28f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 259, + 355, + 351, + 380 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 327, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 327, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 306, + 404 + ], + "score": 1.0, + "content": "Besides, the pseudo variance Goodman (1963) of", + "type": "text" + }, + { + "bbox": [ + 307, + 393, + 316, + 403 + ], + "score": 0.85, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 391, + 327, + 404 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 391, + 327, + 404 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 409, + 447, + 426 + ], + "lines": [ + { + "bbox": [ + 164, + 409, + 447, + 426 + ], + "spans": [ + { + "bbox": [ + 164, + 409, + 447, + 426 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = E ( \\mathcal { T } _ { 1 } ^ { 2 } ) - E ( \\mathcal { T } _ { 1 } ) ^ { 2 } = E \\big ( e ^ { - j 2 \\pi n \\log _ { 1 0 } \\left| \\mathcal { X } _ { 1 } ^ { 2 } \\right| } \\big ) - a _ { n } ^ { 2 } = a _ { 2 n } - a _ { n } ^ { 2 } .", + "type": "interline_equation", + "image_path": "eed2d9e58d94f0dd4558d74684a3fda7e7d3f02d41e8a792433ae66aacdad552.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 164, + 409, + 447, + 426 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 177, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 178, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 178, + 446 + ], + "score": 1.0, + "content": "Correspondingly,", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 428, + 178, + 446 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 448, + 371, + 473 + ], + "lines": [ + { + "bbox": [ + 240, + 448, + 371, + 473 + ], + "spans": [ + { + "bbox": [ + 240, + 448, + 371, + 473 + ], + "score": 0.94, + "content": "\\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = \\frac { 1 } { M } \\mathcal { I } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = \\frac { a _ { 2 n } - a _ { n } ^ { 2 } } { M } .", + "type": "interline_equation", + "image_path": "c3744e73aa7ab5407eb57adfce801a8d059323fd7875630eb1763d2f2f8f11be.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 240, + 448, + 371, + 473 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 131, + 493 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 479, + 173, + 492 + ], + "score": 0.91, + "content": "\\left( a _ { 2 n } - a _ { n } ^ { 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 479, + 256, + 493 + ], + "score": 1.0, + "content": "is bounded, we have", + "type": "text" + }, + { + "bbox": [ + 257, + 480, + 335, + 492 + ], + "score": 0.86, + "content": "\\begin{array} { r } { \\operatorname* { l i m } _ { M \\to \\infty } \\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 479, + 506, + 493 + ], + "score": 1.0, + "content": ". Thus, we obtain that the random variable", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 489, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 107, + 493, + 113, + 501 + ], + "score": 0.72, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 489, + 506, + 505 + ], + "score": 1.0, + "content": "follows circularly-symmetric complex Gaussian distribution, because the sufficient and necessary", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "condition is that mean value and pseudo variance equal zero Goodman (1963). It implies that both", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 232, + 525 + ], + "score": 1.0, + "content": "the real and imaginary part of", + "type": "text" + }, + { + "bbox": [ + 233, + 513, + 240, + 523 + ], + "score": 0.71, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "follow the same Gaussian distribution and they are independent.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Thus, the magnitude of this complex random variable follows the Rayleigh distribution Papoulis &", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 534, + 418, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 317, + 547 + ], + "score": 1.0, + "content": "Pillai (2001), and the probability density function of", + "type": "text" + }, + { + "bbox": [ + 318, + 535, + 330, + 546 + ], + "score": 0.89, + "content": "| \\mathcal { E } |", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 534, + 418, + 547 + ], + "score": 1.0, + "content": "can be formulated as", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 479, + 506, + 547 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 552, + 353, + 573 + ], + "lines": [ + { + "bbox": [ + 257, + 552, + 353, + 573 + ], + "spans": [ + { + "bbox": [ + 257, + 552, + 353, + 573 + ], + "score": 0.94, + "content": "\\mathcal { P } _ { | \\varepsilon | } \\left( r \\right) = \\frac { r } { s ^ { 2 } } e ^ { - r ^ { 2 } / 2 s ^ { 2 } } ,", + "type": "interline_equation", + "image_path": "8fcd3d6f2ab95407bc14119a563351950db51c91de02d82b5c69332fca8e2fdf.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 257, + 552, + 353, + 573 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "where s is the scale parameter. Knowing the properties of the Rayleigh distributionPapoulis & Pillai", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 588, + 177, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 177, + 602 + ], + "score": 1.0, + "content": "(2001), we have:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 577, + 506, + 602 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 606, + 361, + 631 + ], + "lines": [ + { + "bbox": [ + 250, + 606, + 361, + 631 + ], + "spans": [ + { + "bbox": [ + 250, + 606, + 361, + 631 + ], + "score": 0.94, + "content": "s ^ { 2 } = { \\frac { 1 } { 2 } } D ( { \\mathcal { E } } ) = { \\frac { 1 - | a _ { n } | ^ { 2 } } { 2 M } } .", + "type": "interline_equation", + "image_path": "666b6315498def3bdc6e938b8a4b4c9847345e478a1039a9a2be73c12da1da1c.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 250, + 606, + 361, + 631 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 196, + 660 + ], + "score": 1.0, + "content": "Utilizing the fact that", + "type": "text" + }, + { + "bbox": [ + 196, + 643, + 218, + 656 + ], + "score": 0.92, + "content": "\\left| a _ { n } \\right| ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 639, + 450, + 660 + ], + "score": 1.0, + "content": "is close to 0 when n is a modest number, we obtain that", + "type": "text" + }, + { + "bbox": [ + 450, + 642, + 501, + 657 + ], + "score": 0.93, + "content": "\\begin{array} { r } { D \\left( \\mathcal { E } \\right) \\approx \\frac { 1 } { M } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 639, + 506, + 660 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 653, + 260, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 149, + 673 + ], + "score": 1.0, + "content": "leading to", + "type": "text" + }, + { + "bbox": [ + 150, + 656, + 187, + 670 + ], + "score": 0.93, + "content": "\\textstyle s ^ { 2 } = { \\frac { 1 } { 2 M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 653, + 260, + 673 + ], + "score": 1.0, + "content": ". Then, we obtain", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 639, + 506, + 673 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 675, + 384, + 703 + ], + "lines": [ + { + "bbox": [ + 227, + 675, + 384, + 703 + ], + "spans": [ + { + "bbox": [ + 227, + 675, + 384, + 703 + ], + "score": 0.95, + "content": "E { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 4 - \\pi } { 4 M } } .", + "type": "interline_equation", + "image_path": "c8e8e8cb610860d6011ba2f448bfc677e4f4ac66bebe896f1cefb665bde6c066.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 227, + 675, + 384, + 703 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 504, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 225, + 722 + ], + "score": 1.0, + "content": "It is easy to observe that both", + "type": "text" + }, + { + "bbox": [ + 226, + 710, + 258, + 723 + ], + "score": 0.91, + "content": "E { \\big ( } | { \\mathcal { E } } | { \\big ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 708, + 277, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 277, + 710, + 310, + 723 + ], + "score": 0.93, + "content": "D ( \\left. \\boldsymbol { \\mathcal { E } } \\right. )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "are close to zero when the number of samples M", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 720, + 429, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 289, + 733 + ], + "score": 1.0, + "content": "increases. It implies that the estimation error", + "type": "text" + }, + { + "bbox": [ + 289, + 725, + 300, + 732 + ], + "score": 0.88, + "content": "\\varepsilon _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 720, + 357, + 733 + ], + "score": 1.0, + "content": "gets closer to", + "type": "text" + }, + { + "bbox": [ + 357, + 722, + 363, + 730 + ], + "score": 0.61, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 720, + 375, + 733 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 376, + 721, + 386, + 730 + ], + "score": 0.27, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 720, + 429, + 733 + ], + "score": 1.0, + "content": "increases.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 708, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 123, + 474, + 216 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 98, + 504, + 121 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 97, + 505, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 162, + 111 + ], + "score": 1.0, + "content": "Table 9: The", + "type": "text" + }, + { + "bbox": [ + 162, + 100, + 169, + 110 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 97, + 505, + 111 + ], + "score": 1.0, + "content": "-value of KS hypothesis test among clean samples, noisy samples, and adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 109, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 294, + 123 + ], + "score": 1.0, + "content": "samples crafted by four methods, respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 137, + 123, + 474, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 123, + 474, + 216 + ], + "spans": [ + { + "bbox": [ + 137, + 123, + 474, + 216 + ], + "score": 0.981, + "html": "
cleannoisyBIMCW-L2DeepFoolR-PGD
CIFAR-10train set0.0740.0740.1580.1390.1550.159
test set0.0720.0720.1580.1380.2160.216
SVHNtrain set0.1040.1040.2490.1940.2220.245
test set0.1060.1060.2490.1930.2190.239
ImageNet-AlexNettrain set0.1160.1160.2530.2350.2420.253
test set0.1170.1170.2610.2360.2430.259
ImageNet-VGG-16train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
", + "type": "table", + "image_path": "fa3230b33a62941de39e3eb72b6593e1a48dc25c0980b935c88410a70d320d21.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 137, + 123, + 474, + 154.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 137, + 154.0, + 474, + 185.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 137, + 185.0, + 474, + 216.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 111, + 251, + 499, + 344 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 237, + 502, + 249 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 237, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 163, + 251 + ], + "score": 1.0, + "content": "Table 10: The", + "type": "text" + }, + { + "bbox": [ + 163, + 240, + 169, + 249 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 237, + 504, + 251 + ], + "score": 1.0, + "content": "-value of two-sample KS test among MBF coefficients of different types of examples.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 111, + 251, + 499, + 344 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 251, + 499, + 344 + ], + "spans": [ + { + "bbox": [ + 111, + 251, + 499, + 344 + ], + "score": 0.977, + "html": "
(clean,noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
CIFAR-10train set0.6400.0000.0000.003
test set0.7640.0000.0300.000
SVHNtrain set0.5590.0000.0300.000
test set0.6850.0000.0000.000
ImageNet-AlexNettrain set1.0000.0000.0020.134
test set0.9930.0000.4770.000
ImageNet-VGG-16train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
", + "type": "table", + "image_path": "6da3000751d59c307bab64c1baf09f6e3cb5b7f008aacbef64e8fb64f37a5192.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 111, + 251, + 499, + 282.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 282.0, + 499, + 313.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 313.0, + 499, + 344.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "table", + "bbox": [ + 106, + 383, + 513, + 464 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 132, + 369, + 477, + 381 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 131, + 368, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 190, + 382 + ], + "score": 1.0, + "content": "Table 11: The", + "type": "text" + }, + { + "bbox": [ + 191, + 371, + 197, + 381 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 368, + 479, + 382 + ], + "score": 1.0, + "content": "-value of KS test among MBF coefficients from different data sources.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 106, + 383, + 513, + 464 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 383, + 513, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 513, + 464 + ], + "score": 0.979, + "html": "
clean noisyBIMCW-L2DeepFoolR-PGD
CIFAR-10(train set, test set)0.120 0.1880.3330.4560.5840.223
SVHN(train set, test set)0.494 0.7080.5230.1470.4910.645
ImageNet- AlexNet(train set, test set) (train set, out-of-sample)0.153 0.148 0.000 0.0000.370 0.1690.605 0.1010.410 0.2060.386 0.158
ImageNet- VGG-16(train set, test set) (train set, out-of-sample)0.839 0.858 0.049 0.0540.368 0.0940.842 0.4070.246 0.7370.221 0.088
", + "type": "table", + "image_path": "432208b1f366590fad0e738578fcaad0579c7b3400760126f5ff1bf2e1a454f4.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 106, + 383, + 513, + 410.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 410.0, + 513, + 437.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 437.0, + 513, + 464.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 10.0 + }, + { + "type": "image", + "bbox": [ + 135, + 483, + 473, + 574 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 483, + 473, + 574 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 483, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 135, + 483, + 473, + 574 + ], + "score": 0.968, + "type": "image", + "image_path": "951766bb96a13d49ac27f0e9767ec43586237ffa3aba2ff5527718e6f9c18193.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 135, + 483, + 473, + 513.3333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 135, + 513.3333333333334, + 473, + 543.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 135, + 543.6666666666667, + 473, + 574.0000000000001 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 578, + 504, + 600 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 576, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 576, + 219, + 591 + ], + "score": 1.0, + "content": "Figure 2: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 219, + 578, + 230, + 588 + ], + "score": 0.71, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 576, + 506, + 591 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row) and test", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 588, + 232, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 232, + 600 + ], + "score": 1.0, + "content": "(bottom row) set of CIFAR-10.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "index": 15.25 + }, + { + "type": "image", + "bbox": [ + 136, + 615, + 473, + 706 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 615, + 473, + 706 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 136, + 615, + 473, + 706 + ], + "spans": [ + { + "bbox": [ + 136, + 615, + 473, + 706 + ], + "score": 0.971, + "type": "image", + "image_path": "f90ea0187e4a20a3df47ccc875e9571b38f3c0e93fe3a9c53bf08df094171e49.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 136, + 615, + 473, + 645.3333333333334 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 136, + 645.3333333333334, + 473, + 675.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 136, + 675.6666666666667, + 473, + 706.0000000000001 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 709, + 505, + 733 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 218, + 723 + ], + "score": 1.0, + "content": "Figure 3: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 219, + 711, + 230, + 721 + ], + "score": 0.71, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row) and test", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 721, + 218, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 218, + 733 + ], + "score": 1.0, + "content": "(bottom row) set of SVHN.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 20.25 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 123, + 474, + 216 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 98, + 504, + 121 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 97, + 505, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 162, + 111 + ], + "score": 1.0, + "content": "Table 9: The", + "type": "text" + }, + { + "bbox": [ + 162, + 100, + 169, + 110 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 97, + 505, + 111 + ], + "score": 1.0, + "content": "-value of KS hypothesis test among clean samples, noisy samples, and adversarial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 109, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 294, + 123 + ], + "score": 1.0, + "content": "samples crafted by four methods, respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 137, + 123, + 474, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 123, + 474, + 216 + ], + "spans": [ + { + "bbox": [ + 137, + 123, + 474, + 216 + ], + "score": 0.981, + "html": "
cleannoisyBIMCW-L2DeepFoolR-PGD
CIFAR-10train set0.0740.0740.1580.1390.1550.159
test set0.0720.0720.1580.1380.2160.216
SVHNtrain set0.1040.1040.2490.1940.2220.245
test set0.1060.1060.2490.1930.2190.239
ImageNet-AlexNettrain set0.1160.1160.2530.2350.2420.253
test set0.1170.1170.2610.2360.2430.259
ImageNet-VGG-16train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
", + "type": "table", + "image_path": "fa3230b33a62941de39e3eb72b6593e1a48dc25c0980b935c88410a70d320d21.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 137, + 123, + 474, + 154.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 137, + 154.0, + 474, + 185.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 137, + 185.0, + 474, + 216.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 111, + 251, + 499, + 344 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 237, + 502, + 249 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 237, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 163, + 251 + ], + "score": 1.0, + "content": "Table 10: The", + "type": "text" + }, + { + "bbox": [ + 163, + 240, + 169, + 249 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 237, + 504, + 251 + ], + "score": 1.0, + "content": "-value of two-sample KS test among MBF coefficients of different types of examples.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 111, + 251, + 499, + 344 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 251, + 499, + 344 + ], + "spans": [ + { + "bbox": [ + 111, + 251, + 499, + 344 + ], + "score": 0.977, + "html": "
(clean,noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
CIFAR-10train set0.6400.0000.0000.003
test set0.7640.0000.0300.000
SVHNtrain set0.5590.0000.0300.000
test set0.6850.0000.0000.000
ImageNet-AlexNettrain set1.0000.0000.0020.134
test set0.9930.0000.4770.000
ImageNet-VGG-16train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
", + "type": "table", + "image_path": "6da3000751d59c307bab64c1baf09f6e3cb5b7f008aacbef64e8fb64f37a5192.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 111, + 251, + 499, + 282.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 282.0, + 499, + 313.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 313.0, + 499, + 344.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "table", + "bbox": [ + 106, + 383, + 513, + 464 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 132, + 369, + 477, + 381 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 131, + 368, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 190, + 382 + ], + "score": 1.0, + "content": "Table 11: The", + "type": "text" + }, + { + "bbox": [ + 191, + 371, + 197, + 381 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 368, + 479, + 382 + ], + "score": 1.0, + "content": "-value of KS test among MBF coefficients from different data sources.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 106, + 383, + 513, + 464 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 383, + 513, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 513, + 464 + ], + "score": 0.979, + "html": "
clean noisyBIMCW-L2DeepFoolR-PGD
CIFAR-10(train set, test set)0.120 0.1880.3330.4560.5840.223
SVHN(train set, test set)0.494 0.7080.5230.1470.4910.645
ImageNet- AlexNet(train set, test set) (train set, out-of-sample)0.153 0.148 0.000 0.0000.370 0.1690.605 0.1010.410 0.2060.386 0.158
ImageNet- VGG-16(train set, test set) (train set, out-of-sample)0.839 0.858 0.049 0.0540.368 0.0940.842 0.4070.246 0.7370.221 0.088
", + "type": "table", + "image_path": "432208b1f366590fad0e738578fcaad0579c7b3400760126f5ff1bf2e1a454f4.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 106, + 383, + 513, + 410.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 410.0, + 513, + 437.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 437.0, + 513, + 464.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 10.0 + }, + { + "type": "image", + "bbox": [ + 135, + 483, + 473, + 574 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 483, + 473, + 574 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 483, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 135, + 483, + 473, + 574 + ], + "score": 0.968, + "type": "image", + "image_path": "951766bb96a13d49ac27f0e9767ec43586237ffa3aba2ff5527718e6f9c18193.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 135, + 483, + 473, + 513.3333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 135, + 513.3333333333334, + 473, + 543.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 135, + 543.6666666666667, + 473, + 574.0000000000001 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 578, + 504, + 600 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 576, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 576, + 219, + 591 + ], + "score": 1.0, + "content": "Figure 2: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 219, + 578, + 230, + 588 + ], + "score": 0.71, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 576, + 506, + 591 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row) and test", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 588, + 232, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 232, + 600 + ], + "score": 1.0, + "content": "(bottom row) set of CIFAR-10.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "index": 15.25 + }, + { + "type": "image", + "bbox": [ + 136, + 615, + 473, + 706 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 615, + 473, + 706 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 136, + 615, + 473, + 706 + ], + "spans": [ + { + "bbox": [ + 136, + 615, + 473, + 706 + ], + "score": 0.971, + "type": "image", + "image_path": "f90ea0187e4a20a3df47ccc875e9571b38f3c0e93fe3a9c53bf08df094171e49.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 136, + 615, + 473, + 645.3333333333334 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 136, + 645.3333333333334, + 473, + 675.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 136, + 675.6666666666667, + 473, + 706.0000000000001 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 709, + 505, + 733 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 218, + 723 + ], + "score": 1.0, + "content": "Figure 3: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 219, + 711, + 230, + 721 + ], + "score": 0.71, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row) and test", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 721, + 218, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 218, + 733 + ], + "score": 1.0, + "content": "(bottom row) set of SVHN.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 20.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "score": 0.974, + "type": "image", + "image_path": "a2307b9dd6d8b208fa0ca9565e779759c161a7c1546a2a84a8e287388b4c42f5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 83, + 473, + 126.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 126.33333333333334, + 473, + 169.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 169.66666666666669, + 473, + 213.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 217, + 505, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 219, + 230 + ], + "score": 1.0, + "content": "Figure 4: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 219, + 218, + 230, + 228 + ], + "score": 0.64, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row) and test", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 227, + 401, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 401, + 240 + ], + "score": 1.0, + "content": "(median row), and out-of-sample (bottom row) set of ImageNet-AlexNet.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 142, + 271, + 468, + 326 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 248, + 504, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 330, + 261 + ], + "score": 1.0, + "content": "Table 12: Detection results evaluated by AUROC score", + "type": "text" + }, + { + "bbox": [ + 331, + 248, + 346, + 259 + ], + "score": 0.68, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "on nearly 4000 images from the testset", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 259, + 313, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 313, + 270 + ], + "score": 1.0, + "content": "of MNIST. The best results are highlighted in bold.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 142, + 271, + 468, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 271, + 468, + 326 + ], + "spans": [ + { + "bbox": [ + 142, + 271, + 468, + 326 + ], + "score": 0.976, + "html": "
Iteration numberin each GD runNumber of GD runsBIMCW-L2DeepFoolR-PGD
200100.7440.9800.7020.746
100100.6890.9730.6500.690
10050.6640.9690.6280.666
10020.6120.9590.5860.614
MBF1.0000.9911.0001.000
", + "type": "table", + "image_path": "33ed125ffa3caba424d560cc534dab9906f2037fcc8f12426c7dfde11ab5814d.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 142, + 271, + 468, + 289.3333333333333 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 142, + 289.3333333333333, + 468, + 307.66666666666663 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 142, + 307.66666666666663, + 468, + 325.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 288, + 358 + ], + "score": 1.0, + "content": "A.2 ADDITIONAL EMPIRICAL ANALYSIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Here we present additional empirical analysis on more databases and networks, as shown in Tables 9,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 167, + 390 + ], + "score": 1.0, + "content": "10 and 11. The", + "type": "text" + }, + { + "bbox": [ + 168, + 379, + 174, + 389 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "-values in most cases also supports the conclusions obtained in the main manuscript.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "We also present more visualizations in Figs. 2, 3 and 4. These visualizations also demonstrate the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 399, + 420, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 420, + 412 + ], + "score": 1.0, + "content": "distinct difference of MBF features between adversarial and benign examples.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 424, + 305, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 306, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 306, + 437 + ], + "score": 1.0, + "content": "A.3 COMPARISON BETWEEN DEFENSE-GAN", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "We adopt official code8 of Defense-GAN (Samangouei et al. (2018)) and train an generative model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 456, + 504, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 504, + 467 + ], + "score": 1.0, + "content": "on train set of MNIST dataset after 130, 000 epochs (As the official code of Defense-GAN currently", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "only supports three datasets including MNIST, Fashion-MNIST, and CelebA). The classification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 226, + 489 + ], + "score": 1.0, + "content": "model attains an accuracy of", + "type": "text" + }, + { + "bbox": [ + 226, + 478, + 254, + 488 + ], + "score": 0.87, + "content": "9 8 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "on the test set of MNIST dataset. We still utilize four attack", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "strategies, including BIM, CW-L2, DeepFool and R-PGD. After removing images that are misclas-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "score": 1.0, + "content": "sified and failed to attack, we pick 6000 images to train SVMs based on our MBF features, and leave", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "nearly 4000 images for testing. There are two key hyper-parameters in Defense-GAN, including the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 519, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 535 + ], + "score": 1.0, + "content": "iteration number in each GD run, and the total number of GD runs. We have tried multiple settings", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "of these two hyper-parameters. Table 12 show the AUROC score on nearly 4000 images. Our MBF", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 542, + 419, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 419, + 556 + ], + "score": 1.0, + "content": "method exceeds Defense-GAN by a large margin in quantitative performance.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 107, + 568, + 353, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 355, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 355, + 580 + ], + "score": 1.0, + "content": "A.4 DETAILED PARAMETERS OF ADVERSARIAL ATTACK", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Four popular adversarial attack methods are adopted to craft adversarial examples, including ba-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "sic iterative method (BIM), CarliniWagnerL2Attack (CW-L2), DeepFool, and random projected", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "gradient descent (R-PGD). We emphasize detailed parameters of these attack strategies for repro-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 216, + 635 + ], + "score": 1.0, + "content": "ducibility, which are BIM", + "type": "text" + }, + { + "bbox": [ + 216, + 622, + 249, + 633 + ], + "score": 0.66, + "content": "\\mathrm { ( e p s { = } 0 . 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 622, + 285, + 635 + ], + "score": 1.0, + "content": ", stepsize", + "type": "text" + }, + { + "bbox": [ + 285, + 622, + 309, + 632 + ], + "score": 0.71, + "content": "= 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 622, + 351, + 635 + ], + "score": 1.0, + "content": ", iterations", + "type": "text" + }, + { + "bbox": [ + 351, + 622, + 368, + 632 + ], + "score": 0.62, + "content": "{ \\it \\Omega } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 622, + 489, + 635 + ], + "score": 1.0, + "content": "), CW-L2 (binary search steps", + "type": "text" + }, + { + "bbox": [ + 490, + 622, + 501, + 632 + ], + "score": 0.66, + "content": "^ { = 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 622, + 505, + 635 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 150, + 645 + ], + "score": 1.0, + "content": "confidence", + "type": "text" + }, + { + "bbox": [ + 150, + 633, + 169, + 643 + ], + "score": 0.68, + "content": "= 0 . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 632, + 226, + 645 + ], + "score": 1.0, + "content": ", learning rate", + "type": "text" + }, + { + "bbox": [ + 226, + 633, + 255, + 643 + ], + "score": 0.61, + "content": "= 0 . 0 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 632, + 319, + 645 + ], + "score": 1.0, + "content": ", max iterations", + "type": "text" + }, + { + "bbox": [ + 319, + 633, + 346, + 643 + ], + "score": 0.7, + "content": "\\mathord { \\left. \\begin{array} { r l l } { - 1 0 0 0 } \\end{array} \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 632, + 440, + 645 + ], + "score": 1.0, + "content": "), DeepFool (max step", + "type": "text" + }, + { + "bbox": [ + 441, + 633, + 464, + 643 + ], + "score": 0.66, + "content": "\\mathord { \\mathrm { \\ s } } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "), R-PGD", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 110, + 643, + 494, + 656 + ], + "spans": [ + { + "bbox": [ + 110, + 644, + 143, + 655 + ], + "score": 0.38, + "content": "\\mathrm { \\ e p s { = } 0 . 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 643, + 178, + 656 + ], + "score": 1.0, + "content": ", stepsize", + "type": "text" + }, + { + "bbox": [ + 178, + 644, + 202, + 654 + ], + "score": 0.78, + "content": "{ \\mathrm { \\Omega } } { = } 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 643, + 243, + 656 + ], + "score": 1.0, + "content": ", iterations", + "type": "text" + }, + { + "bbox": [ + 244, + 644, + 260, + 654 + ], + "score": 0.75, + "content": "\\scriptstyle = 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 643, + 494, + 656 + ], + "score": 1.0, + "content": "). All of the attack strategies are implemented by Foolbox.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 272, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 274, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 274, + 734 + ], + "score": 1.0, + "content": "8 https://github.com/kabkabm/defensegan/", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 83, + 473, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 83, + 473, + 213 + ], + "score": 0.974, + "type": "image", + "image_path": "a2307b9dd6d8b208fa0ca9565e779759c161a7c1546a2a84a8e287388b4c42f5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 136, + 83, + 473, + 126.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 136, + 126.33333333333334, + 473, + 169.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 136, + 169.66666666666669, + 473, + 213.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 217, + 505, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 219, + 230 + ], + "score": 1.0, + "content": "Figure 4: Statistics (mean", + "type": "text" + }, + { + "bbox": [ + 219, + 218, + 230, + 228 + ], + "score": 0.64, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "standard deviation) of MBF coefficients on train (top row) and test", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 227, + 401, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 401, + 240 + ], + "score": 1.0, + "content": "(median row), and out-of-sample (bottom row) set of ImageNet-AlexNet.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 142, + 271, + 468, + 326 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 248, + 504, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 330, + 261 + ], + "score": 1.0, + "content": "Table 12: Detection results evaluated by AUROC score", + "type": "text" + }, + { + "bbox": [ + 331, + 248, + 346, + 259 + ], + "score": 0.68, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "on nearly 4000 images from the testset", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 259, + 313, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 313, + 270 + ], + "score": 1.0, + "content": "of MNIST. The best results are highlighted in bold.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 142, + 271, + 468, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 271, + 468, + 326 + ], + "spans": [ + { + "bbox": [ + 142, + 271, + 468, + 326 + ], + "score": 0.976, + "html": "
Iteration numberin each GD runNumber of GD runsBIMCW-L2DeepFoolR-PGD
200100.7440.9800.7020.746
100100.6890.9730.6500.690
10050.6640.9690.6280.666
10020.6120.9590.5860.614
MBF1.0000.9911.0001.000
", + "type": "table", + "image_path": "33ed125ffa3caba424d560cc534dab9906f2037fcc8f12426c7dfde11ab5814d.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 142, + 271, + 468, + 289.3333333333333 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 142, + 289.3333333333333, + 468, + 307.66666666666663 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 142, + 307.66666666666663, + 468, + 325.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 288, + 358 + ], + "score": 1.0, + "content": "A.2 ADDITIONAL EMPIRICAL ANALYSIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Here we present additional empirical analysis on more databases and networks, as shown in Tables 9,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 167, + 390 + ], + "score": 1.0, + "content": "10 and 11. The", + "type": "text" + }, + { + "bbox": [ + 168, + 379, + 174, + 389 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "-values in most cases also supports the conclusions obtained in the main manuscript.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "We also present more visualizations in Figs. 2, 3 and 4. These visualizations also demonstrate the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 399, + 420, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 420, + 412 + ], + "score": 1.0, + "content": "distinct difference of MBF features between adversarial and benign examples.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 366, + 506, + 412 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 424, + 305, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 306, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 306, + 437 + ], + "score": 1.0, + "content": "A.3 COMPARISON BETWEEN DEFENSE-GAN", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 444, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "We adopt official code8 of Defense-GAN (Samangouei et al. (2018)) and train an generative model", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 456, + 504, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 504, + 467 + ], + "score": 1.0, + "content": "on train set of MNIST dataset after 130, 000 epochs (As the official code of Defense-GAN currently", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "only supports three datasets including MNIST, Fashion-MNIST, and CelebA). The classification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 226, + 489 + ], + "score": 1.0, + "content": "model attains an accuracy of", + "type": "text" + }, + { + "bbox": [ + 226, + 478, + 254, + 488 + ], + "score": 0.87, + "content": "9 8 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "on the test set of MNIST dataset. We still utilize four attack", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "strategies, including BIM, CW-L2, DeepFool and R-PGD. After removing images that are misclas-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 504, + 511 + ], + "score": 1.0, + "content": "sified and failed to attack, we pick 6000 images to train SVMs based on our MBF features, and leave", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "nearly 4000 images for testing. There are two key hyper-parameters in Defense-GAN, including the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 519, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 535 + ], + "score": 1.0, + "content": "iteration number in each GD run, and the total number of GD runs. We have tried multiple settings", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "of these two hyper-parameters. Table 12 show the AUROC score on nearly 4000 images. Our MBF", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 542, + 419, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 419, + 556 + ], + "score": 1.0, + "content": "method exceeds Defense-GAN by a large margin in quantitative performance.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 443, + 506, + 556 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 568, + 353, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 355, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 355, + 580 + ], + "score": 1.0, + "content": "A.4 DETAILED PARAMETERS OF ADVERSARIAL ATTACK", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Four popular adversarial attack methods are adopted to craft adversarial examples, including ba-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "sic iterative method (BIM), CarliniWagnerL2Attack (CW-L2), DeepFool, and random projected", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "gradient descent (R-PGD). We emphasize detailed parameters of these attack strategies for repro-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 216, + 635 + ], + "score": 1.0, + "content": "ducibility, which are BIM", + "type": "text" + }, + { + "bbox": [ + 216, + 622, + 249, + 633 + ], + "score": 0.66, + "content": "\\mathrm { ( e p s { = } 0 . 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 622, + 285, + 635 + ], + "score": 1.0, + "content": ", stepsize", + "type": "text" + }, + { + "bbox": [ + 285, + 622, + 309, + 632 + ], + "score": 0.71, + "content": "= 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 622, + 351, + 635 + ], + "score": 1.0, + "content": ", iterations", + "type": "text" + }, + { + "bbox": [ + 351, + 622, + 368, + 632 + ], + "score": 0.62, + "content": "{ \\it \\Omega } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 622, + 489, + 635 + ], + "score": 1.0, + "content": "), CW-L2 (binary search steps", + "type": "text" + }, + { + "bbox": [ + 490, + 622, + 501, + 632 + ], + "score": 0.66, + "content": "^ { = 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 622, + 505, + 635 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 150, + 645 + ], + "score": 1.0, + "content": "confidence", + "type": "text" + }, + { + "bbox": [ + 150, + 633, + 169, + 643 + ], + "score": 0.68, + "content": "= 0 . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 632, + 226, + 645 + ], + "score": 1.0, + "content": ", learning rate", + "type": "text" + }, + { + "bbox": [ + 226, + 633, + 255, + 643 + ], + "score": 0.61, + "content": "= 0 . 0 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 632, + 319, + 645 + ], + "score": 1.0, + "content": ", max iterations", + "type": "text" + }, + { + "bbox": [ + 319, + 633, + 346, + 643 + ], + "score": 0.7, + "content": "\\mathord { \\left. \\begin{array} { r l l } { - 1 0 0 0 } \\end{array} \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 632, + 440, + 645 + ], + "score": 1.0, + "content": "), DeepFool (max step", + "type": "text" + }, + { + "bbox": [ + 441, + 633, + 464, + 643 + ], + "score": 0.66, + "content": "\\mathord { \\mathrm { \\ s } } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "), R-PGD", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 110, + 643, + 494, + 656 + ], + "spans": [ + { + "bbox": [ + 110, + 644, + 143, + 655 + ], + "score": 0.38, + "content": "\\mathrm { \\ e p s { = } 0 . 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 643, + 178, + 656 + ], + "score": 1.0, + "content": ", stepsize", + "type": "text" + }, + { + "bbox": [ + 178, + 644, + 202, + 654 + ], + "score": 0.78, + "content": "{ \\mathrm { \\Omega } } { = } 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 643, + 243, + 656 + ], + "score": 1.0, + "content": ", iterations", + "type": "text" + }, + { + "bbox": [ + 244, + 644, + 260, + 654 + ], + "score": 0.75, + "content": "\\scriptstyle = 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 643, + 494, + 656 + ], + "score": 1.0, + "content": "). All of the attack strategies are implemented by Foolbox.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 588, + 505, + 656 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ryeK6nNFDr/ryeK6nNFDr_model.json b/parse/train/ryeK6nNFDr/ryeK6nNFDr_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0b2a9a0134e5f0ea66178b2a9718e3910f7578b --- /dev/null +++ b/parse/train/ryeK6nNFDr/ryeK6nNFDr_model.json @@ -0,0 +1,21049 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 642, + 1302, + 642, + 1302, + 1312, + 398, + 1312 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1438, + 1404, + 1438, + 1404, + 1803, + 298, + 1803 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1820, + 1402, + 1820, + 1402, + 2033, + 298, + 2033 + ], + "score": 0.979 + }, + { + "category_id": 0, + "poly": [ + 299, + 219, + 1399, + 219, + 1399, + 377, + 299, + 377 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 313, + 433, + 680, + 433, + 680, + 493, + 313, + 493 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 302, + 1369, + 573, + 1369, + 573, + 1404, + 302, + 1404 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 773, + 575, + 926, + 575, + 926, + 608, + 773, + 608 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.612 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 217.0, + 1403.0, + 217.0, + 1403.0, + 271.0, + 294.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 274.0, + 1404.0, + 274.0, + 1404.0, + 325.0, + 291.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 337.0, + 453.0, + 337.0, + 453.0, + 380.0, + 296.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1366.0, + 579.0, + 1366.0, + 579.0, + 1413.0, + 294.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 574.0, + 932.0, + 574.0, + 932.0, + 611.0, + 769.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 643.0, + 1305.0, + 643.0, + 1305.0, + 676.0, + 396.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 675.0, + 1304.0, + 675.0, + 1304.0, + 706.0, + 394.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 704.0, + 1305.0, + 704.0, + 1305.0, + 736.0, + 394.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 733.0, + 1305.0, + 733.0, + 1305.0, + 768.0, + 394.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 765.0, + 1305.0, + 765.0, + 1305.0, + 797.0, + 393.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 795.0, + 1305.0, + 795.0, + 1305.0, + 827.0, + 393.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 824.0, + 1306.0, + 824.0, + 1306.0, + 861.0, + 393.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 856.0, + 1305.0, + 856.0, + 1305.0, + 890.0, + 394.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 888.0, + 1305.0, + 888.0, + 1305.0, + 918.0, + 394.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 915.0, + 1307.0, + 915.0, + 1307.0, + 954.0, + 392.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 946.0, + 1305.0, + 946.0, + 1305.0, + 982.0, + 393.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 978.0, + 1305.0, + 978.0, + 1305.0, + 1012.0, + 393.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1007.0, + 1306.0, + 1007.0, + 1306.0, + 1042.0, + 393.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1040.0, + 1305.0, + 1040.0, + 1305.0, + 1071.0, + 393.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1068.0, + 1306.0, + 1068.0, + 1306.0, + 1103.0, + 392.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1099.0, + 1306.0, + 1099.0, + 1306.0, + 1135.0, + 394.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1129.0, + 1306.0, + 1129.0, + 1306.0, + 1164.0, + 394.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1160.0, + 1306.0, + 1160.0, + 1306.0, + 1195.0, + 393.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1190.0, + 1306.0, + 1190.0, + 1306.0, + 1225.0, + 394.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1221.0, + 1304.0, + 1221.0, + 1304.0, + 1255.0, + 394.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1252.0, + 1305.0, + 1252.0, + 1305.0, + 1285.0, + 394.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1284.0, + 911.0, + 1284.0, + 911.0, + 1313.0, + 395.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1404.0, + 1438.0, + 1404.0, + 1474.0, + 295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1469.0, + 1405.0, + 1469.0, + 1405.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1502.0, + 1401.0, + 1502.0, + 1401.0, + 1532.0, + 296.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1527.0, + 1405.0, + 1527.0, + 1405.0, + 1566.0, + 292.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1594.0, + 293.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1405.0, + 1591.0, + 1405.0, + 1625.0, + 294.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1622.0, + 1405.0, + 1622.0, + 1405.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1691.0, + 292.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1406.0, + 1682.0, + 1406.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1406.0, + 1713.0, + 1406.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1774.0, + 1317.0, + 1774.0, + 1317.0, + 1807.0, + 293.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1820.0, + 1404.0, + 1820.0, + 1404.0, + 1854.0, + 296.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1881.0, + 1406.0, + 1881.0, + 1406.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1914.0, + 1404.0, + 1914.0, + 1404.0, + 1942.0, + 294.0, + 1942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1974.0, + 293.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1339.0, + 2004.0, + 1339.0, + 2034.0, + 294.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 434.0, + 560.0, + 434.0, + 560.0, + 466.0, + 315.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 462.0, + 681.0, + 462.0, + 681.0, + 497.0, + 312.0, + 497.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 230, + 1403, + 230, + 1403, + 657, + 299, + 657 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 675, + 1404, + 675, + 1404, + 1524, + 298, + 1524 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1728, + 1403, + 1728, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1651, + 1400, + 1651, + 1400, + 1712, + 299, + 1712 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 300, + 1577, + 580, + 1577, + 580, + 1612, + 300, + 1612 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.705 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 859, + 2089, + 859, + 2112, + 841, + 2112 + ], + "score": 0.147 + }, + { + "category_id": 13, + "poly": [ + 974, + 1405, + 1003, + 1405, + 1003, + 1433, + 974, + 1433 + ], + "score": 0.7, + "latex": "\\pm" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1573.0, + 587.0, + 1573.0, + 587.0, + 1619.0, + 291.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 838.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 300.0, + 293.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 357.0, + 294.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 350.0, + 1405.0, + 350.0, + 1405.0, + 389.0, + 293.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 384.0, + 1404.0, + 384.0, + 1404.0, + 416.0, + 295.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 1404.0, + 413.0, + 1404.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 1404.0, + 444.0, + 1404.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 473.0, + 1405.0, + 473.0, + 1405.0, + 509.0, + 293.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 506.0, + 1404.0, + 506.0, + 1404.0, + 537.0, + 295.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 533.0, + 1407.0, + 533.0, + 1407.0, + 571.0, + 292.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 565.0, + 1407.0, + 565.0, + 1407.0, + 599.0, + 294.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 594.0, + 1407.0, + 594.0, + 1407.0, + 631.0, + 293.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 823.0, + 627.0, + 823.0, + 659.0, + 294.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 708.0, + 294.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 703.0, + 1406.0, + 703.0, + 1406.0, + 739.0, + 291.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 772.0, + 292.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 765.0, + 1404.0, + 765.0, + 1404.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 796.0, + 1406.0, + 796.0, + 1406.0, + 830.0, + 294.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 823.0, + 1406.0, + 823.0, + 1406.0, + 862.0, + 292.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 856.0, + 1405.0, + 856.0, + 1405.0, + 891.0, + 294.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 886.0, + 1403.0, + 886.0, + 1403.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 915.0, + 1405.0, + 915.0, + 1405.0, + 951.0, + 292.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 948.0, + 1405.0, + 948.0, + 1405.0, + 982.0, + 294.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 1402.0, + 977.0, + 1402.0, + 1012.0, + 294.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1006.0, + 1404.0, + 1006.0, + 1404.0, + 1045.0, + 292.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1036.0, + 1405.0, + 1036.0, + 1405.0, + 1075.0, + 293.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1070.0, + 1406.0, + 1070.0, + 1406.0, + 1104.0, + 293.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1098.0, + 1406.0, + 1098.0, + 1406.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1129.0, + 1402.0, + 1129.0, + 1402.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1163.0, + 1404.0, + 1163.0, + 1404.0, + 1193.0, + 296.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1190.0, + 1406.0, + 1190.0, + 1406.0, + 1226.0, + 293.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1284.0, + 1405.0, + 1284.0, + 1405.0, + 1315.0, + 296.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 1406.0, + 1312.0, + 1406.0, + 1347.0, + 294.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1344.0, + 1404.0, + 1344.0, + 1404.0, + 1375.0, + 296.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1406.0, + 296.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 973.0, + 1403.0, + 973.0, + 1438.0, + 294.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1403.0, + 1405.0, + 1403.0, + 1405.0, + 1438.0, + 1004.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1434.0, + 1405.0, + 1434.0, + 1405.0, + 1469.0, + 293.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1466.0, + 1404.0, + 1466.0, + 1404.0, + 1497.0, + 294.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1494.0, + 813.0, + 1494.0, + 813.0, + 1531.0, + 293.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1726.0, + 1405.0, + 1726.0, + 1405.0, + 1764.0, + 294.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1825.0, + 292.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1404.0, + 1849.0, + 1404.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1407.0, + 1881.0, + 1407.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1910.0, + 1408.0, + 1910.0, + 1408.0, + 1946.0, + 291.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1042.0, + 1679.0, + 1042.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 460, + 1404, + 460, + 1404, + 1191, + 298, + 1191 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1207, + 1404, + 1207, + 1404, + 1573, + 298, + 1573 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 445, + 298, + 445 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1749, + 1404, + 1749, + 1404, + 1844, + 298, + 1844 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1927, + 1403, + 1927, + 1403, + 2036, + 299, + 2036 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 725, + 1862, + 974, + 1862, + 974, + 1905, + 725, + 1905 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 298, + 1692, + 855, + 1692, + 855, + 1725, + 298, + 1725 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 300, + 1621, + 578, + 1621, + 578, + 1656, + 300, + 1656 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1871, + 1399, + 1871, + 1399, + 1900, + 1366, + 1900 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.656 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.458 + }, + { + "category_id": 13, + "poly": [ + 614, + 1930, + 754, + 1930, + 754, + 1977, + 614, + 1977 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { A = \\frac { \\beta c } { 2 \\Gamma ( 1 / c ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 373, + 1925, + 560, + 1925, + 560, + 1977, + 373, + 1977 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\beta = \\frac { 1 } { \\sigma } \\left( \\frac { \\Gamma \\left( 3 / c \\right) } { \\Gamma \\left( 1 / c \\right) } \\right) ^ { \\frac { 1 } { 2 } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 725, + 1861, + 973, + 1861, + 973, + 1904, + 725, + 1904 + ], + "score": 0.92, + "latex": "\\mathcal { P } _ { \\mathcal { X } } ( x ) = A \\cdot e ^ { - \\lvert \\beta x \\rvert ^ { c } } ," + }, + { + "category_id": 13, + "poly": [ + 671, + 1749, + 782, + 1749, + 782, + 1780, + 671, + 1780 + ], + "score": 0.92, + "latex": "\\mathcal { X } ~ \\in ~ \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 822, + 1937, + 871, + 1937, + 871, + 1970, + 822, + 1970 + ], + "score": 0.89, + "latex": "\\Gamma ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1152, + 643, + 1171, + 643, + 1171, + 671, + 1152, + 671 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 670, + 1980, + 690, + 1980, + 690, + 2004, + 670, + 2004 + ], + "score": 0.82, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 820, + 1818, + 836, + 1818, + 836, + 1840, + 820, + 1840 + ], + "score": 0.81, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 416, + 1979, + 436, + 1979, + 436, + 2005, + 416, + 2005 + ], + "score": 0.81, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1819, + 1155, + 1819, + 1155, + 1840, + 1136, + 1840 + ], + "score": 0.76, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 409, + 1783, + 436, + 1783, + 436, + 1811, + 409, + 1811 + ], + "score": 0.27, + "latex": "\\&" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1692.0, + 859.0, + 1692.0, + 859.0, + 1728.0, + 294.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1618.0, + 582.0, + 1618.0, + 582.0, + 1662.0, + 293.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 1408.0, + 461.0, + 1408.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 490.0, + 1406.0, + 490.0, + 1406.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 520.0, + 1404.0, + 520.0, + 1404.0, + 558.0, + 293.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 551.0, + 1404.0, + 551.0, + 1404.0, + 586.0, + 293.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 582.0, + 1404.0, + 582.0, + 1404.0, + 615.0, + 292.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 614.0, + 1405.0, + 614.0, + 1405.0, + 645.0, + 297.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 641.0, + 1151.0, + 641.0, + 1151.0, + 680.0, + 293.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 641.0, + 1405.0, + 641.0, + 1405.0, + 680.0, + 1172.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1406.0, + 673.0, + 1406.0, + 708.0, + 294.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 1405.0, + 703.0, + 1405.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 733.0, + 1405.0, + 733.0, + 1405.0, + 769.0, + 293.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 765.0, + 1406.0, + 765.0, + 1406.0, + 800.0, + 294.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 828.0, + 296.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 824.0, + 1407.0, + 824.0, + 1407.0, + 861.0, + 292.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 855.0, + 1406.0, + 855.0, + 1406.0, + 890.0, + 294.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 886.0, + 1406.0, + 886.0, + 1406.0, + 921.0, + 293.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 918.0, + 1404.0, + 918.0, + 1404.0, + 949.0, + 296.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 947.0, + 1406.0, + 947.0, + 1406.0, + 981.0, + 293.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 974.0, + 1406.0, + 974.0, + 1406.0, + 1013.0, + 292.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1038.0, + 1404.0, + 1038.0, + 1404.0, + 1073.0, + 294.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 1405.0, + 1065.0, + 1405.0, + 1107.0, + 292.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1101.0, + 1402.0, + 1101.0, + 1402.0, + 1132.0, + 296.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1127.0, + 1408.0, + 1127.0, + 1408.0, + 1167.0, + 292.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1161.0, + 774.0, + 1161.0, + 774.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1204.0, + 1406.0, + 1204.0, + 1406.0, + 1243.0, + 295.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1237.0, + 1405.0, + 1237.0, + 1405.0, + 1271.0, + 296.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1269.0, + 1404.0, + 1269.0, + 1404.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1298.0, + 1407.0, + 1298.0, + 1407.0, + 1334.0, + 292.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1329.0, + 1403.0, + 1329.0, + 1403.0, + 1362.0, + 295.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1391.0, + 1405.0, + 1391.0, + 1405.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1456.0, + 292.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1449.0, + 1406.0, + 1449.0, + 1406.0, + 1487.0, + 293.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1483.0, + 1402.0, + 1483.0, + 1402.0, + 1514.0, + 296.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1513.0, + 1405.0, + 1513.0, + 1405.0, + 1544.0, + 293.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1544.0, + 768.0, + 1544.0, + 768.0, + 1574.0, + 294.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 1405.0, + 288.0, + 1405.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 350.0, + 1405.0, + 350.0, + 1405.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 381.0, + 1407.0, + 381.0, + 1407.0, + 417.0, + 294.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 935.0, + 413.0, + 935.0, + 448.0, + 294.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1747.0, + 670.0, + 1747.0, + 670.0, + 1787.0, + 293.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1787.0, + 783.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1782.0, + 408.0, + 1782.0, + 408.0, + 1816.0, + 294.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1782.0, + 1405.0, + 1782.0, + 1405.0, + 1816.0, + 437.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1814.0, + 819.0, + 1814.0, + 819.0, + 1845.0, + 294.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1814.0, + 1135.0, + 1814.0, + 1135.0, + 1845.0, + 837.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1814.0, + 1287.0, + 1814.0, + 1287.0, + 1845.0, + 1156.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1923.0, + 613.0, + 1923.0, + 613.0, + 1989.0, + 561.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1923.0, + 821.0, + 1923.0, + 821.0, + 1989.0, + 755.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1923.0, + 1413.0, + 1923.0, + 1413.0, + 1989.0, + 872.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 415.0, + 1972.0, + 415.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1972.0, + 669.0, + 1972.0, + 669.0, + 2006.0, + 437.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2006.0, + 691.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.25, + 1927.5, + 543.25, + 1927.5, + 543.25, + 1978.5, + 292.25, + 1978.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1077, + 1404, + 1077, + 1404, + 1233, + 297, + 1233 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1247, + 1404, + 1247, + 1404, + 1402, + 297, + 1402 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 414, + 1404, + 414, + 1404, + 571, + 298, + 571 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 300, + 688, + 1402, + 688, + 1402, + 793, + 300, + 793 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1518, + 1404, + 1518, + 1404, + 1682, + 297, + 1682 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 295, + 1708, + 1405, + 1708, + 1405, + 1802, + 295, + 1802 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 588, + 975, + 1106, + 975, + 1106, + 1064, + 588, + 1064 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 409, + 806, + 1286, + 806, + 1286, + 890, + 409, + 890 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 295, + 901, + 1404, + 901, + 1404, + 964, + 295, + 964 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 636, + 1916, + 1063, + 1916, + 1063, + 1991, + 636, + 1991 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 316, + 1416, + 1334, + 1416, + 1334, + 1507, + 316, + 1507 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 713, + 1816, + 982, + 1816, + 982, + 1861, + 713, + 1861 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 332, + 587, + 1331, + 587, + 1331, + 676, + 332, + 676 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 294, + 1871, + 1059, + 1871, + 1059, + 1904, + 294, + 1904 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 293, + 2002, + 1377, + 2002, + 1377, + 2037, + 293, + 2037 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 105, + 297, + 105 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 301, + 359, + 781, + 359, + 781, + 392, + 301, + 392 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1405, + 228, + 1405, + 325, + 299, + 325 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1004, + 1400, + 1004, + 1400, + 1033, + 1366, + 1033 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 837, + 1400, + 837, + 1400, + 868, + 1366, + 868 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1826, + 1401, + 1826, + 1401, + 1856, + 1365, + 1856 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1366, + 616, + 1400, + 616, + 1400, + 646, + 1366, + 646 + ], + "score": 0.871 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1447, + 1400, + 1447, + 1400, + 1477, + 1367, + 1477 + ], + "score": 0.807 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.792 + }, + { + "category_id": 14, + "poly": [ + 638, + 1915, + 1060, + 1915, + 1060, + 1992, + 638, + 1992 + ], + "score": 0.95, + "latex": "E { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D ( | { \\mathcal { E } } | ) = { \\frac { 4 - \\pi } { 4 M } } ," + }, + { + "category_id": 14, + "poly": [ + 332, + 583, + 1331, + 583, + 1331, + 677, + 332, + 677 + ], + "score": 0.94, + "latex": "\\mathcal { P } _ { Z } ( z ) = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor A _ { n } \\cos { ( 2 \\pi n z ) } + B _ { n } \\sin { ( 2 \\pi n z ) } \\rfloor = 1 + 2 \\sum _ { n = 1 } ^ { + \\infty } \\lfloor a _ { n } \\rfloor \\cos ( 2 \\pi n z + \\phi _ { n } ) ," + }, + { + "category_id": 14, + "poly": [ + 320, + 1413, + 1345, + 1413, + 1345, + 1508, + 320, + 1508 + ], + "score": 0.94, + "latex": "\\hat { a } _ { n } = \\frac { \\sum _ { m = 1 } ^ { M } e ^ { - j 2 \\pi n \\log _ { 1 0 } | x _ { m } | } } { M } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\left[ \\cos \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) - j \\sin \\left( 2 \\pi n \\log _ { 1 0 } | x _ { m } | \\right) \\right] ." + }, + { + "category_id": 14, + "poly": [ + 589, + 971, + 1111, + 971, + 1111, + 1066, + 589, + 1066 + ], + "score": 0.94, + "latex": "| a _ { n } | = \\bigg ( \\prod _ { k = 0 } ^ { + \\infty } \\bigg [ 1 + \\big ( \\frac { 2 \\pi n } { \\log 1 0 ( c k + 1 ) } \\big ) ^ { 2 } \\bigg ] ^ { - 1 } \\bigg ) ^ { \\frac { 1 } { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 354, + 1613, + 577, + 1613, + 577, + 1656, + 354, + 1656 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathcal { Y } = \\bar { \\frac { 1 } { M } } \\bar { \\sum _ { m = 1 } ^ { M } } \\mathcal { T } _ { m } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 410, + 804, + 1291, + 804, + 1291, + 892, + 410, + 892 + ], + "score": 0.93, + "latex": "a _ { n } = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { Z } } ( z ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } z } d z = \\frac { 2 A e ^ { \\frac { j 2 \\pi n \\log \\beta } { \\log 1 0 } } } { \\beta c } \\cdot \\Gamma \\Bigg ( \\frac { - j 2 \\pi n + \\log 1 0 } { c \\log 1 0 } \\Bigg ) ." + }, + { + "category_id": 13, + "poly": [ + 726, + 760, + 801, + 760, + 801, + 794, + 726, + 794 + ], + "score": 0.93, + "latex": "\\mathcal { P } _ { \\mathcal { Z } } ( z )" + }, + { + "category_id": 13, + "poly": [ + 951, + 1170, + 1070, + 1170, + 1070, + 1204, + 951, + 1204 + ], + "score": 0.93, + "latex": "\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }" + }, + { + "category_id": 14, + "poly": [ + 713, + 1813, + 986, + 1813, + 986, + 1862, + 713, + 1862 + ], + "score": 0.92, + "latex": "\\mathcal { P } _ { | \\mathcal { E } | } \\left( r \\right) = 2 M r e ^ { - M r ^ { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 735, + 1140, + 854, + 1140, + 854, + 1174, + 735, + 1174 + ], + "score": 0.92, + "latex": "\\{ | a _ { n } | \\} _ { n \\in \\mathbb { N } }" + }, + { + "category_id": 13, + "poly": [ + 409, + 1080, + 458, + 1080, + 458, + 1113, + 409, + 1113 + ], + "score": 0.92, + "latex": "\\left| a _ { n } \\right|" + }, + { + "category_id": 13, + "poly": [ + 801, + 1712, + 951, + 1712, + 951, + 1741, + 801, + 1741 + ], + "score": 0.92, + "latex": "\\varepsilon _ { n } = { \\hat { a } } _ { n } - a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 374, + 689, + 490, + 689, + 490, + 723, + 374, + 723 + ], + "score": 0.92, + "latex": "z \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1063, + 1309, + 1283, + 1309, + 1283, + 1342, + 1063, + 1342 + ], + "score": 0.91, + "latex": "\\mathbf { x } = \\{ x _ { 1 } , \\dots , x _ { M } \\}" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1080, + 1244, + 1080, + 1244, + 1113, + 1196, + 1113 + ], + "score": 0.91, + "latex": "\\left| a _ { n } \\right|" + }, + { + "category_id": 13, + "poly": [ + 992, + 721, + 1210, + 721, + 1210, + 761, + 992, + 761 + ], + "score": 0.91, + "latex": "| a _ { n } | = \\sqrt { A _ { n } ^ { 2 } + B _ { n } ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 935, + 1551, + 1167, + 1551, + 1167, + 1581, + 935, + 1581 + ], + "score": 0.91, + "latex": "\\mathcal { T } = e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } | }" + }, + { + "category_id": 13, + "poly": [ + 397, + 1279, + 445, + 1279, + 445, + 1312, + 397, + 1312 + ], + "score": 0.91, + "latex": "\\left| a _ { n } \\right|" + }, + { + "category_id": 13, + "poly": [ + 719, + 289, + 879, + 289, + 879, + 325, + 719, + 325 + ], + "score": 0.91, + "latex": "( - \\sqrt { 2 } \\sigma , \\sqrt { 2 } \\sigma )" + }, + { + "category_id": 13, + "poly": [ + 636, + 1079, + 715, + 1079, + 715, + 1108, + 636, + 1108 + ], + "score": 0.91, + "latex": "n \\in \\mathbb N" + }, + { + "category_id": 13, + "poly": [ + 828, + 232, + 901, + 232, + 901, + 259, + 828, + 259 + ], + "score": 0.9, + "latex": "c = 1" + }, + { + "category_id": 13, + "poly": [ + 1222, + 722, + 1398, + 722, + 1398, + 759, + 1222, + 759 + ], + "score": 0.9, + "latex": "a _ { n } = | a _ { n } | \\cdot e ^ { j \\phi _ { n } }" + }, + { + "category_id": 13, + "poly": [ + 1211, + 1280, + 1287, + 1280, + 1287, + 1311, + 1211, + 1311 + ], + "score": 0.9, + "latex": "\\mathcal { P } _ { \\mathcal { Z } } ( z )" + }, + { + "category_id": 13, + "poly": [ + 735, + 2016, + 764, + 2016, + 764, + 2034, + 735, + 2034 + ], + "score": 0.89, + "latex": "\\varepsilon _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1172, + 264, + 1286, + 264, + 1286, + 291, + 1172, + 291 + ], + "score": 0.89, + "latex": "c \\to + \\infty" + }, + { + "category_id": 13, + "poly": [ + 366, + 262, + 438, + 262, + 438, + 290, + 366, + 290 + ], + "score": 0.89, + "latex": "c = 2" + }, + { + "category_id": 13, + "poly": [ + 498, + 1521, + 531, + 1521, + 531, + 1550, + 498, + 1550 + ], + "score": 0.89, + "latex": "\\hat { a } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 952, + 1585, + 985, + 1585, + 985, + 1614, + 952, + 1614 + ], + "score": 0.88, + "latex": "\\hat { a } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1521, + 1094, + 1521, + 1094, + 1550, + 1062, + 1550 + ], + "score": 0.88, + "latex": "\\hat { a } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1058, + 260, + 1089, + 260, + 1089, + 289, + 1058, + 289 + ], + "score": 0.88, + "latex": "\\sigma ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 478, + 466, + 478, + 466, + 512, + 298, + 512 + ], + "score": 0.87, + "latex": "\\mathcal { Z } = \\log _ { 1 0 } | \\mathcal { X } |" + }, + { + "category_id": 13, + "poly": [ + 297, + 1740, + 426, + 1740, + 426, + 1771, + 297, + 1771 + ], + "score": 0.87, + "latex": "\\mathcal { E } = \\mathcal { V } - a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 440, + 721, + 681, + 721, + 681, + 763, + 440, + 763 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\phi _ { n } = \\arctan \\left( - \\frac { B _ { n } } { A _ { n } } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 585, + 1524, + 618, + 1524, + 618, + 1550, + 585, + 1550 + ], + "score": 0.87, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 931, + 906, + 962, + 906, + 962, + 933, + 931, + 933 + ], + "score": 0.86, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 801, + 1283, + 834, + 1283, + 834, + 1309, + 801, + 1309 + ], + "score": 0.86, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1258, + 1525, + 1291, + 1525, + 1291, + 1550, + 1258, + 1550 + ], + "score": 0.86, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 945, + 762, + 994, + 762, + 994, + 789, + 945, + 789 + ], + "score": 0.85, + "latex": "2 \\pi n" + }, + { + "category_id": 13, + "poly": [ + 325, + 1310, + 376, + 1310, + 376, + 1337, + 325, + 1337 + ], + "score": 0.85, + "latex": "2 \\pi n" + }, + { + "category_id": 13, + "poly": [ + 354, + 1585, + 381, + 1585, + 381, + 1611, + 354, + 1611 + ], + "score": 0.83, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1329, + 1521, + 1361, + 1521, + 1361, + 1547, + 1329, + 1547 + ], + "score": 0.82, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 924, + 2009, + 940, + 2009, + 940, + 2029, + 924, + 2029 + ], + "score": 0.82, + "latex": "\\boldsymbol { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 439, + 1739, + 474, + 1739, + 474, + 1772, + 439, + 1772 + ], + "score": 0.82, + "latex": "| \\mathcal { E } |" + }, + { + "category_id": 13, + "poly": [ + 458, + 541, + 491, + 541, + 491, + 568, + 458, + 568 + ], + "score": 0.82, + "latex": "2 \\pi" + }, + { + "category_id": 13, + "poly": [ + 299, + 909, + 329, + 909, + 329, + 933, + 299, + 933 + ], + "score": 0.81, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1027, + 691, + 1052, + 691, + 1052, + 717, + 1027, + 717 + ], + "score": 0.81, + "latex": "\\mathcal { Z }" + }, + { + "category_id": 13, + "poly": [ + 909, + 1283, + 928, + 1283, + 928, + 1306, + 909, + 1306 + ], + "score": 0.79, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 430, + 765, + 449, + 765, + 449, + 788, + 430, + 788 + ], + "score": 0.79, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 298, + 1340, + 330, + 1340, + 330, + 1368, + 298, + 1368 + ], + "score": 0.78, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1115, + 1072, + 1115, + 1072, + 1137, + 1053, + 1137 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 608, + 1118, + 623, + 1118, + 623, + 1137, + 608, + 1137 + ], + "score": 0.72, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 994, + 1346, + 1009, + 1346, + 1009, + 1367, + 994, + 1367 + ], + "score": 0.7, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 1113, + 1256, + 1128, + 1256, + 1128, + 1276, + 1113, + 1276 + ], + "score": 0.69, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 481, + 1772, + 556, + 1772, + 556, + 1802, + 481, + 1802 + ], + "score": 0.68, + "latex": "( P D F )" + }, + { + "category_id": 13, + "poly": [ + 1175, + 1147, + 1191, + 1147, + 1191, + 1168, + 1175, + 1168 + ], + "score": 0.56, + "latex": "c" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 358.0, + 783.0, + 358.0, + 783.0, + 394.0, + 294.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 863.0, + 2085.0, + 863.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1078.0, + 408.0, + 1078.0, + 408.0, + 1115.0, + 294.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1078.0, + 635.0, + 1078.0, + 635.0, + 1115.0, + 459.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1078.0, + 1195.0, + 1078.0, + 1195.0, + 1115.0, + 716.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1078.0, + 1404.0, + 1078.0, + 1404.0, + 1115.0, + 1245.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1110.0, + 607.0, + 1110.0, + 607.0, + 1143.0, + 296.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1110.0, + 1052.0, + 1110.0, + 1052.0, + 1143.0, + 624.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1110.0, + 1404.0, + 1110.0, + 1404.0, + 1143.0, + 1073.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 734.0, + 1138.0, + 734.0, + 1179.0, + 294.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1138.0, + 1174.0, + 1138.0, + 1174.0, + 1179.0, + 855.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1138.0, + 1406.0, + 1138.0, + 1406.0, + 1179.0, + 1192.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1167.0, + 950.0, + 1167.0, + 950.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1167.0, + 1407.0, + 1167.0, + 1407.0, + 1208.0, + 1071.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1200.0, + 902.0, + 1200.0, + 902.0, + 1234.0, + 295.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1246.0, + 1112.0, + 1246.0, + 1112.0, + 1282.0, + 295.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1246.0, + 1404.0, + 1246.0, + 1404.0, + 1282.0, + 1129.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1275.0, + 396.0, + 1275.0, + 396.0, + 1313.0, + 295.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1275.0, + 800.0, + 1275.0, + 800.0, + 1313.0, + 446.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1275.0, + 908.0, + 1275.0, + 908.0, + 1313.0, + 835.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1275.0, + 1210.0, + 1275.0, + 1210.0, + 1313.0, + 929.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1275.0, + 1405.0, + 1275.0, + 1405.0, + 1313.0, + 1288.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1306.0, + 324.0, + 1306.0, + 324.0, + 1346.0, + 291.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1306.0, + 1062.0, + 1306.0, + 1062.0, + 1346.0, + 377.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1306.0, + 1407.0, + 1306.0, + 1407.0, + 1346.0, + 1284.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1338.0, + 993.0, + 1338.0, + 993.0, + 1375.0, + 331.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1338.0, + 1405.0, + 1338.0, + 1405.0, + 1375.0, + 1010.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1368.0, + 1098.0, + 1368.0, + 1098.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 418.0, + 1405.0, + 418.0, + 1405.0, + 451.0, + 296.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 1406.0, + 444.0, + 1406.0, + 483.0, + 292.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 471.0, + 297.0, + 471.0, + 297.0, + 518.0, + 291.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 471.0, + 1409.0, + 471.0, + 1409.0, + 518.0, + 467.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 509.0, + 1404.0, + 509.0, + 1404.0, + 542.0, + 296.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 457.0, + 535.0, + 457.0, + 577.0, + 292.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 535.0, + 504.0, + 535.0, + 504.0, + 577.0, + 492.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 685.0, + 373.0, + 685.0, + 373.0, + 725.0, + 295.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 685.0, + 1026.0, + 685.0, + 1026.0, + 725.0, + 491.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 685.0, + 1407.0, + 685.0, + 1407.0, + 725.0, + 1053.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 708.0, + 439.0, + 708.0, + 439.0, + 776.0, + 286.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 708.0, + 991.0, + 708.0, + 991.0, + 776.0, + 682.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 708.0, + 1221.0, + 708.0, + 1221.0, + 776.0, + 1211.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 708.0, + 1409.0, + 708.0, + 1409.0, + 776.0, + 1399.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 757.0, + 429.0, + 757.0, + 429.0, + 794.0, + 295.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 757.0, + 725.0, + 757.0, + 725.0, + 794.0, + 450.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 757.0, + 944.0, + 757.0, + 944.0, + 794.0, + 802.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 757.0, + 1225.0, + 757.0, + 1225.0, + 794.0, + 995.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1517.0, + 497.0, + 1517.0, + 497.0, + 1554.0, + 295.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1517.0, + 584.0, + 1517.0, + 584.0, + 1554.0, + 532.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1517.0, + 1061.0, + 1517.0, + 1061.0, + 1554.0, + 619.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1517.0, + 1257.0, + 1517.0, + 1257.0, + 1554.0, + 1095.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1517.0, + 1328.0, + 1517.0, + 1328.0, + 1554.0, + 1292.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1517.0, + 1404.0, + 1517.0, + 1404.0, + 1554.0, + 1362.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1546.0, + 934.0, + 1546.0, + 934.0, + 1588.0, + 290.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1546.0, + 1407.0, + 1546.0, + 1407.0, + 1588.0, + 1168.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1584.0, + 353.0, + 1584.0, + 353.0, + 1616.0, + 296.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1584.0, + 951.0, + 1584.0, + 951.0, + 1616.0, + 382.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1584.0, + 1404.0, + 1584.0, + 1404.0, + 1616.0, + 986.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1605.0, + 353.0, + 1605.0, + 353.0, + 1676.0, + 284.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1605.0, + 1413.0, + 1605.0, + 1413.0, + 1676.0, + 578.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1651.0, + 411.0, + 1651.0, + 411.0, + 1679.0, + 293.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1705.0, + 800.0, + 1705.0, + 800.0, + 1745.0, + 293.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1705.0, + 1406.0, + 1705.0, + 1406.0, + 1745.0, + 952.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1736.0, + 296.0, + 1736.0, + 296.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1736.0, + 438.0, + 1736.0, + 438.0, + 1777.0, + 427.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1736.0, + 1405.0, + 1736.0, + 1405.0, + 1777.0, + 475.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1767.0, + 480.0, + 1767.0, + 480.0, + 1807.0, + 293.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1767.0, + 747.0, + 1767.0, + 747.0, + 1807.0, + 557.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 298.0, + 899.0, + 298.0, + 938.0, + 292.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 899.0, + 930.0, + 899.0, + 930.0, + 938.0, + 330.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 938.0, + 963.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 934.0, + 615.0, + 934.0, + 615.0, + 966.0, + 295.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1869.0, + 1062.0, + 1869.0, + 1062.0, + 1908.0, + 294.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2000.0, + 734.0, + 2000.0, + 734.0, + 2039.0, + 295.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 2000.0, + 923.0, + 2000.0, + 923.0, + 2039.0, + 765.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 2000.0, + 1383.0, + 2000.0, + 1383.0, + 2039.0, + 941.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 827.0, + 225.0, + 827.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 225.0, + 1405.0, + 225.0, + 1405.0, + 267.0, + 902.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 259.0, + 365.0, + 259.0, + 365.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 259.0, + 1057.0, + 259.0, + 1057.0, + 295.0, + 439.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 259.0, + 1171.0, + 259.0, + 1171.0, + 295.0, + 1090.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 295.0, + 1287.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 718.0, + 289.0, + 718.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 289.0, + 893.0, + 289.0, + 893.0, + 327.0, + 880.0, + 327.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1517, + 1404, + 1517, + 1404, + 1887, + 298, + 1887 + ], + "score": 0.986 + }, + { + "category_id": 1, + "poly": [ + 297, + 1148, + 1403, + 1148, + 1403, + 1351, + 297, + 1351 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 351, + 1399, + 351, + 1399, + 475, + 299, + 475 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 639, + 879, + 1059, + 879, + 1059, + 974, + 639, + 974 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 297, + 1364, + 1398, + 1364, + 1398, + 1428, + 297, + 1428 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 295, + 1439, + 1398, + 1439, + 1398, + 1503, + 295, + 1503 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 416, + 1030, + 1281, + 1030, + 1281, + 1124, + 416, + 1124 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1401, + 1972, + 1401, + 2034, + 300, + 2034 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 299, + 983, + 1235, + 983, + 1235, + 1019, + 299, + 1019 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 293, + 225, + 1302, + 225, + 1302, + 262, + 293, + 262 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 297, + 486, + 1404, + 486, + 1404, + 765, + 297, + 765 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 300, + 1923, + 680, + 1923, + 680, + 1954, + 300, + 1954 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 914, + 1400, + 914, + 1400, + 945, + 1366, + 945 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 297, + 774, + 1403, + 774, + 1403, + 868, + 297, + 868 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1065, + 1400, + 1065, + 1400, + 1094, + 1366, + 1094 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.715 + }, + { + "category_id": 0, + "poly": [ + 297, + 293, + 997, + 293, + 997, + 325, + 297, + 325 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 297, + 293, + 997, + 293, + 997, + 325, + 297, + 325 + ], + "score": 0.257 + }, + { + "category_id": 14, + "poly": [ + 640, + 874, + 1057, + 874, + 1057, + 976, + 640, + 976 + ], + "score": 0.95, + "latex": "( \\hat { a } _ { n } ) _ { i } ^ { l } = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } e ^ { - j 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | } ," + }, + { + "category_id": 14, + "poly": [ + 416, + 1025, + 1284, + 1025, + 1284, + 1128, + 416, + 1128 + ], + "score": 0.94, + "latex": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } | = \\frac { 1 } { M _ { i } ^ { l } } \\sum _ { m = 1 } ^ { M _ { i } ^ { l } } \\Bigg ( \\cos ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) + \\sin ^ { 2 } \\left( 2 \\pi n \\log _ { 1 0 } | ( r _ { m } ) _ { i } ^ { l } | \\right) \\Bigg ) ^ { \\frac { 1 } { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 1027, + 1669, + 1090, + 1669, + 1090, + 1704, + 1027, + 1704 + ], + "score": 0.93, + "latex": "( \\hat { a } _ { n } ) _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1174, + 670, + 1331, + 670, + 1331, + 703, + 1174, + 703 + ], + "score": 0.93, + "latex": "\\left\\{ bar { \\mathbf { x } } _ { 1 } , \\ldots , \\bar { \\mathbf { x } } _ { N } \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 938, + 579, + 1096, + 579, + 1096, + 612, + 938, + 612 + ], + "score": 0.93, + "latex": "\\big \\{ \\hat { \\mathbf { x } } _ { 1 } , \\dotsc , \\hat { \\mathbf { x } } _ { N } \\big \\}" + }, + { + "category_id": 13, + "poly": [ + 884, + 1149, + 1266, + 1149, + 1266, + 1187, + 884, + 1187 + ], + "score": 0.93, + "latex": "\\mathbf { a } _ { i } ^ { l } = [ | ( \\hat { a } _ { 1 } ) _ { i } ^ { l } | , \\dots , | ( \\hat { a } _ { T } ) _ { i } ^ { l } | ] \\in \\mathbb { R } _ { + } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 913, + 983, + 974, + 983, + 974, + 1020, + 913, + 1020 + ], + "score": 0.93, + "latex": "( \\hat { a } _ { n } ) _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 912, + 1282, + 1217, + 1282, + 1217, + 1320, + 912, + 1320 + ], + "score": 0.92, + "latex": "\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\bf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , 3 N }" + }, + { + "category_id": 13, + "poly": [ + 927, + 487, + 1083, + 487, + 1083, + 521, + 927, + 521 + ], + "score": 0.92, + "latex": "\\left\\{ \\mathbf { x } _ { 1 } , \\dotsc , \\mathbf { x } _ { N } \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 1064, + 1186, + 1142, + 1186, + 1142, + 1221, + 1064, + 1221 + ], + "score": 0.91, + "latex": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } |" + }, + { + "category_id": 13, + "poly": [ + 297, + 1245, + 602, + 1245, + 602, + 1284, + 297, + 1284 + ], + "score": 0.91, + "latex": "\\hat { \\mathbf { a } } _ { i } = [ \\hat { \\mathbf { a } } _ { i } ^ { 1 } ; \\hdots ; \\hat { \\mathbf { a } } _ { i } ^ { L } ] \\in \\mathbb { R } _ { + } ^ { T \\check { L } }" + }, + { + "category_id": 13, + "poly": [ + 372, + 983, + 413, + 983, + 413, + 1019, + 372, + 1019 + ], + "score": 0.91, + "latex": "M _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 885, + 730, + 1394, + 730, + 1394, + 765, + 885, + 765 + ], + "score": 0.91, + "latex": "\\mathcal { D } _ { t r } = \\{ ( \\mathbf { x } _ { i } , - 1 \\bar { ) } , ( \\bar { \\mathbf { x } } _ { i } , \\bar { { ~ - } 1 } ) , ( \\hat { \\mathbf { x } } _ { i } , + 1 \\bar { ) } \\} _ { i = 1 , \\ldots , N }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1360, + 1258, + 1360, + 1258, + 1396, + 1213, + 1396 + ], + "score": 0.91, + "latex": "\\hat { A } _ { t r }" + }, + { + "category_id": 13, + "poly": [ + 819, + 1318, + 863, + 1318, + 863, + 1348, + 819, + 1348 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { t r }" + }, + { + "category_id": 13, + "poly": [ + 1198, + 776, + 1243, + 776, + 1243, + 806, + 1198, + 806 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { t r }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1701, + 338, + 1701, + 338, + 1737, + 298, + 1737 + ], + "score": 0.9, + "latex": "M _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 730, + 1286, + 775, + 1286, + 775, + 1317, + 730, + 1317 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { t r }" + }, + { + "category_id": 13, + "poly": [ + 763, + 1188, + 852, + 1188, + 852, + 1216, + 763, + 1216 + ], + "score": 0.89, + "latex": "T = 1 6" + }, + { + "category_id": 13, + "poly": [ + 669, + 984, + 694, + 984, + 694, + 1019, + 669, + 1019 + ], + "score": 0.88, + "latex": "\\mathbf { r } _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1118, + 805, + 1143, + 805, + 1143, + 839, + 1118, + 839 + ], + "score": 0.88, + "latex": "\\mathbf { r } _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 859, + 807, + 888, + 807, + 888, + 838, + 859, + 838 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1472, + 1066, + 1472, + 1066, + 1503, + 1035, + 1503 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1250, + 1078, + 1250, + 1078, + 1280, + 1048, + 1280 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1305, + 776, + 1336, + 776, + 1336, + 808, + 1305, + 808 + ], + "score": 0.86, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 750, + 1523, + 783, + 1523, + 783, + 1549, + 750, + 1549 + ], + "score": 0.85, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 327, + 1320, + 354, + 1320, + 354, + 1348, + 327, + 1348 + ], + "score": 0.85, + "latex": "{ \\bf a } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1140, + 1643, + 1173, + 1643, + 1173, + 1670, + 1140, + 1670 + ], + "score": 0.85, + "latex": "a _ { n }" + }, + { + "category_id": 13, + "poly": [ + 415, + 1319, + 455, + 1319, + 455, + 1348, + 415, + 1348 + ], + "score": 0.84, + "latex": "+ 1" + }, + { + "category_id": 13, + "poly": [ + 1101, + 1397, + 1130, + 1397, + 1130, + 1429, + 1101, + 1429 + ], + "score": 0.84, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 405, + 519, + 435, + 519, + 435, + 549, + 405, + 549 + ], + "score": 0.84, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 545, + 1154, + 569, + 1154, + 569, + 1180, + 545, + 1180 + ], + "score": 0.83, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 601, + 732, + 643, + 732, + 643, + 758, + 601, + 758 + ], + "score": 0.81, + "latex": "3 N" + }, + { + "category_id": 13, + "poly": [ + 492, + 580, + 521, + 580, + 521, + 606, + 492, + 606 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 674, + 1250, + 697, + 1250, + 697, + 1276, + 674, + 1276 + ], + "score": 0.8, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 332, + 640, + 369, + 640, + 369, + 672, + 332, + 672 + ], + "score": 0.79, + "latex": "f _ { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 807, + 672, + 835, + 672, + 835, + 697, + 807, + 697 + ], + "score": 0.78, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 907, + 778, + 919, + 778, + 919, + 803, + 907, + 803 + ], + "score": 0.78, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 488, + 1319, + 528, + 1319, + 528, + 1348, + 488, + 1348 + ], + "score": 0.76, + "latex": "^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1193, + 1266, + 1193, + 1266, + 1215, + 1246, + 1215 + ], + "score": 0.76, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 743, + 488, + 772, + 488, + 772, + 515, + 743, + 515 + ], + "score": 0.75, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1357, + 1155, + 1370, + 1155, + 1370, + 1180, + 1357, + 1180 + ], + "score": 0.71, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 723, + 809, + 734, + 809, + 734, + 833, + 723, + 833 + ], + "score": 0.7, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 566, + 1189, + 579, + 1189, + 579, + 1215, + 566, + 1215 + ], + "score": 0.62, + "latex": "l" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 223.0, + 1306.0, + 223.0, + 1306.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1921.0, + 683.0, + 1921.0, + 683.0, + 1957.0, + 295.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 1001.0, + 293.0, + 1001.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1518.0, + 749.0, + 1518.0, + 749.0, + 1553.0, + 296.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1518.0, + 1405.0, + 1518.0, + 1405.0, + 1553.0, + 784.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1548.0, + 1404.0, + 1548.0, + 1404.0, + 1585.0, + 293.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1580.0, + 1405.0, + 1580.0, + 1405.0, + 1613.0, + 293.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1648.0, + 292.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1638.0, + 1139.0, + 1638.0, + 1139.0, + 1674.0, + 295.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1638.0, + 1406.0, + 1638.0, + 1406.0, + 1674.0, + 1174.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1669.0, + 1026.0, + 1669.0, + 1026.0, + 1706.0, + 293.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1706.0, + 1091.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1740.0, + 339.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1734.0, + 1406.0, + 1734.0, + 1406.0, + 1771.0, + 293.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1767.0, + 1404.0, + 1767.0, + 1404.0, + 1797.0, + 296.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1794.0, + 1404.0, + 1794.0, + 1404.0, + 1831.0, + 293.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1826.0, + 1405.0, + 1826.0, + 1405.0, + 1860.0, + 296.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1856.0, + 1237.0, + 1856.0, + 1237.0, + 1890.0, + 294.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1149.0, + 544.0, + 1149.0, + 544.0, + 1188.0, + 292.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1149.0, + 883.0, + 1149.0, + 883.0, + 1188.0, + 570.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1149.0, + 1356.0, + 1149.0, + 1356.0, + 1188.0, + 1267.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1149.0, + 1405.0, + 1149.0, + 1405.0, + 1188.0, + 1371.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1184.0, + 565.0, + 1184.0, + 565.0, + 1225.0, + 292.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1184.0, + 762.0, + 1184.0, + 762.0, + 1225.0, + 580.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1184.0, + 1063.0, + 1184.0, + 1063.0, + 1225.0, + 853.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1184.0, + 1245.0, + 1184.0, + 1245.0, + 1225.0, + 1143.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1184.0, + 1407.0, + 1184.0, + 1407.0, + 1225.0, + 1267.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1407.0, + 1216.0, + 1407.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1240.0, + 296.0, + 1240.0, + 296.0, + 1289.0, + 290.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1240.0, + 673.0, + 1240.0, + 673.0, + 1289.0, + 603.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1240.0, + 1047.0, + 1240.0, + 1047.0, + 1289.0, + 698.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1240.0, + 1412.0, + 1240.0, + 1412.0, + 1289.0, + 1079.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1284.0, + 729.0, + 1284.0, + 729.0, + 1323.0, + 291.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1284.0, + 911.0, + 1284.0, + 911.0, + 1323.0, + 776.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1284.0, + 1406.0, + 1284.0, + 1406.0, + 1323.0, + 1218.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1315.0, + 326.0, + 1315.0, + 326.0, + 1354.0, + 292.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1315.0, + 414.0, + 1315.0, + 414.0, + 1354.0, + 355.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1315.0, + 487.0, + 1315.0, + 487.0, + 1354.0, + 456.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1315.0, + 818.0, + 1315.0, + 818.0, + 1354.0, + 529.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1315.0, + 876.0, + 1315.0, + 876.0, + 1354.0, + 864.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 349.0, + 1405.0, + 349.0, + 1405.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 384.0, + 1404.0, + 384.0, + 1404.0, + 417.0, + 295.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 414.0, + 1404.0, + 414.0, + 1404.0, + 447.0, + 295.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 445.0, + 1266.0, + 445.0, + 1266.0, + 477.0, + 293.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1361.0, + 1212.0, + 1361.0, + 1212.0, + 1397.0, + 296.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1361.0, + 1402.0, + 1361.0, + 1402.0, + 1397.0, + 1259.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 1100.0, + 1391.0, + 1100.0, + 1433.0, + 293.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1391.0, + 1140.0, + 1391.0, + 1140.0, + 1433.0, + 1131.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1403.0, + 1438.0, + 1403.0, + 1474.0, + 295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 1034.0, + 1468.0, + 1034.0, + 1504.0, + 295.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1468.0, + 1340.0, + 1468.0, + 1340.0, + 1504.0, + 1067.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1997.0, + 1406.0, + 1997.0, + 1406.0, + 2044.0, + 293.0, + 2044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 980.0, + 371.0, + 980.0, + 371.0, + 1023.0, + 295.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 980.0, + 668.0, + 980.0, + 668.0, + 1023.0, + 414.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 980.0, + 912.0, + 980.0, + 912.0, + 1023.0, + 695.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 980.0, + 1241.0, + 980.0, + 1241.0, + 1023.0, + 975.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 486.0, + 742.0, + 486.0, + 742.0, + 523.0, + 294.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 486.0, + 926.0, + 486.0, + 926.0, + 523.0, + 773.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 486.0, + 1404.0, + 486.0, + 1404.0, + 523.0, + 1084.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 516.0, + 404.0, + 516.0, + 404.0, + 553.0, + 295.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 516.0, + 1404.0, + 516.0, + 1404.0, + 553.0, + 436.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 576.0, + 491.0, + 576.0, + 491.0, + 615.0, + 291.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 576.0, + 937.0, + 576.0, + 937.0, + 615.0, + 522.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 576.0, + 1406.0, + 576.0, + 1406.0, + 615.0, + 1097.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 608.0, + 1406.0, + 608.0, + 1406.0, + 644.0, + 292.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 639.0, + 331.0, + 639.0, + 331.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 639.0, + 1405.0, + 639.0, + 1405.0, + 676.0, + 370.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 669.0, + 806.0, + 669.0, + 806.0, + 706.0, + 295.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 669.0, + 1173.0, + 669.0, + 1173.0, + 706.0, + 836.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 669.0, + 1405.0, + 669.0, + 1405.0, + 706.0, + 1332.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 698.0, + 1406.0, + 698.0, + 1406.0, + 736.0, + 292.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 727.0, + 600.0, + 727.0, + 600.0, + 770.0, + 291.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 727.0, + 884.0, + 727.0, + 884.0, + 770.0, + 644.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 727.0, + 1408.0, + 727.0, + 1408.0, + 770.0, + 1395.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 775.0, + 906.0, + 775.0, + 906.0, + 809.0, + 296.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 775.0, + 1197.0, + 775.0, + 1197.0, + 809.0, + 920.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 775.0, + 1304.0, + 775.0, + 1304.0, + 809.0, + 1244.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 775.0, + 1404.0, + 775.0, + 1404.0, + 809.0, + 1337.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 803.0, + 722.0, + 803.0, + 722.0, + 840.0, + 292.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 803.0, + 858.0, + 803.0, + 858.0, + 840.0, + 735.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 803.0, + 1117.0, + 803.0, + 1117.0, + 840.0, + 889.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 803.0, + 1405.0, + 803.0, + 1405.0, + 840.0, + 1144.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 836.0, + 1141.0, + 836.0, + 1141.0, + 869.0, + 295.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 293.0, + 1001.0, + 293.0, + 1001.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 659, + 1405, + 659, + 1405, + 1238, + 297, + 1238 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1387, + 1404, + 1387, + 1404, + 1815, + 297, + 1815 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1253, + 1403, + 1253, + 1403, + 1375, + 298, + 1375 + ], + "score": 0.967 + }, + { + "category_id": 2, + "poly": [ + 330, + 1852, + 996, + 1852, + 996, + 2035, + 330, + 2035 + ], + "score": 0.954 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 306, + 292, + 1249, + 292, + 1249, + 586, + 306, + 586 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.793 + }, + { + "category_id": 1, + "poly": [ + 298, + 268, + 1072, + 268, + 1072, + 297, + 298, + 297 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1391, + 228, + 1391, + 262, + 295, + 262 + ], + "score": 0.221 + }, + { + "category_id": 13, + "poly": [ + 844, + 781, + 951, + 781, + 951, + 815, + 844, + 815 + ], + "score": 0.93, + "latex": "{ \\mathcal { N } } ( 0 , \\sigma ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 545, + 379, + 729, + 379, + 729, + 412, + 545, + 412 + ], + "score": 0.92, + "latex": "\\{ | ( \\hat { a } _ { n } ) _ { i } ^ { l } | \\} _ { n = 1 , \\dots , T }" + }, + { + "category_id": 13, + "poly": [ + 507, + 434, + 637, + 434, + 637, + 467, + 507, + 467 + ], + "score": 0.92, + "latex": "\\{ \\hat { \\mathbf { a } } _ { i } ^ { l } \\} _ { l = 1 , \\dots , L }" + }, + { + "category_id": 13, + "poly": [ + 625, + 1416, + 733, + 1416, + 733, + 1448, + 625, + 1448 + ], + "score": 0.92, + "latex": "\\mathrm { K D + B U } ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 513, + 348, + 586, + 348, + 586, + 381, + 513, + 381 + ], + "score": 0.92, + "latex": "| ( \\hat { a } _ { n } ) _ { i } ^ { l } |" + }, + { + "category_id": 13, + "poly": [ + 750, + 351, + 892, + 351, + 892, + 380, + 750, + 380 + ], + "score": 0.9, + "latex": "n = 1 , \\dots , T" + }, + { + "category_id": 13, + "poly": [ + 736, + 522, + 779, + 522, + 779, + 555, + 736, + 555 + ], + "score": 0.89, + "latex": "\\hat { A } _ { t r }" + }, + { + "category_id": 13, + "poly": [ + 465, + 295, + 521, + 295, + 521, + 324, + 465, + 324 + ], + "score": 0.89, + "latex": "| \\mathcal { D } _ { t r } |" + }, + { + "category_id": 13, + "poly": [ + 400, + 721, + 491, + 721, + 491, + 751, + 400, + 751 + ], + "score": 0.89, + "latex": "8 2 . 3 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1222, + 903, + 1277, + 903, + 1277, + 933, + 1222, + 933 + ], + "score": 0.89, + "latex": "2 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 948, + 903, + 1003, + 903, + 1003, + 934, + 948, + 934 + ], + "score": 0.89, + "latex": "8 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 377, + 296, + 436, + 296, + 436, + 321, + 377, + 321 + ], + "score": 0.89, + "latex": "i = 1" + }, + { + "category_id": 13, + "poly": [ + 936, + 1055, + 1004, + 1055, + 1004, + 1085, + 936, + 1085 + ], + "score": 0.88, + "latex": "1 0 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 927, + 487, + 1233, + 487, + 1233, + 524, + 927, + 524 + ], + "score": 0.88, + "latex": "\\hat { \\mathcal { A } } _ { t r } = \\{ ( \\hat { \\mathbf { a } } _ { i } , \\pm 1 ) \\} _ { i = 1 , \\dots , | \\mathcal { D } _ { t r } | }" + }, + { + "category_id": 13, + "poly": [ + 800, + 751, + 876, + 751, + 876, + 781, + 800, + 781 + ], + "score": 0.88, + "latex": "9 2 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 507, + 813, + 535, + 813, + 535, + 843, + 507, + 843 + ], + "score": 0.88, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 414, + 324, + 472, + 324, + 472, + 348, + 414, + 348 + ], + "score": 0.88, + "latex": "l = 1" + }, + { + "category_id": 13, + "poly": [ + 912, + 378, + 939, + 378, + 939, + 410, + 912, + 410 + ], + "score": 0.87, + "latex": "\\hat { \\mathbf { a } } _ { i } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1026, + 269, + 1067, + 269, + 1067, + 295, + 1026, + 295 + ], + "score": 0.85, + "latex": "\\mathcal { D } _ { t r }" + }, + { + "category_id": 13, + "poly": [ + 652, + 268, + 679, + 268, + 679, + 296, + 652, + 296 + ], + "score": 0.85, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 871, + 438, + 897, + 438, + 897, + 464, + 871, + 464 + ], + "score": 0.83, + "latex": "\\hat { \\mathbf { a } } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 734, + 269, + 754, + 269, + 754, + 293, + 734, + 293 + ], + "score": 0.78, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 1269, + 786, + 1291, + 786, + 1291, + 810, + 1269, + 810 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 502, + 1447, + 564, + 1447, + 564, + 1478, + 502, + 1478 + ], + "score": 0.75, + "latex": "\\mathrm { L I D } ^ { 5 }" + }, + { + "category_id": 13, + "poly": [ + 502, + 325, + 522, + 325, + 522, + 347, + 502, + 347 + ], + "score": 0.73, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 878, + 1784, + 975, + 1784, + 975, + 1813, + 878, + 1813 + ], + "score": 0.73, + "latex": "\\mathrm { K D + B U }" + }, + { + "category_id": 13, + "poly": [ + 763, + 1662, + 861, + 1662, + 861, + 1691, + 763, + 1691 + ], + "score": 0.69, + "latex": "\\mathrm { K D + B U }" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1849.0, + 904.0, + 1849.0, + 904.0, + 1886.0, + 331.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1882.0, + 743.0, + 1882.0, + 743.0, + 1913.0, + 332.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1908.0, + 944.0, + 1908.0, + 944.0, + 1948.0, + 325.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1937.0, + 943.0, + 1937.0, + 943.0, + 1982.0, + 324.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1967.0, + 999.0, + 1967.0, + 999.0, + 2010.0, + 325.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1998.0, + 872.0, + 1998.0, + 872.0, + 2040.0, + 329.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 1406.0, + 659.0, + 1406.0, + 694.0, + 294.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 689.0, + 1403.0, + 689.0, + 1403.0, + 724.0, + 295.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 716.0, + 399.0, + 716.0, + 399.0, + 759.0, + 292.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 716.0, + 1405.0, + 716.0, + 1405.0, + 759.0, + 492.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 799.0, + 750.0, + 799.0, + 788.0, + 294.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 750.0, + 1405.0, + 750.0, + 1405.0, + 788.0, + 877.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 843.0, + 781.0, + 843.0, + 816.0, + 294.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 781.0, + 1268.0, + 781.0, + 1268.0, + 816.0, + 952.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 781.0, + 1405.0, + 781.0, + 1405.0, + 816.0, + 1292.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 506.0, + 812.0, + 506.0, + 847.0, + 295.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 812.0, + 1405.0, + 812.0, + 1405.0, + 847.0, + 536.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 842.0, + 1403.0, + 842.0, + 1403.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 872.0, + 1403.0, + 872.0, + 1403.0, + 907.0, + 292.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 900.0, + 947.0, + 900.0, + 947.0, + 942.0, + 291.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 900.0, + 1221.0, + 900.0, + 1221.0, + 942.0, + 1004.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 900.0, + 1406.0, + 900.0, + 1406.0, + 942.0, + 1278.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 931.0, + 1406.0, + 931.0, + 1406.0, + 971.0, + 294.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 962.0, + 1402.0, + 962.0, + 1402.0, + 1001.0, + 294.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 993.0, + 1406.0, + 993.0, + 1406.0, + 1032.0, + 292.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1026.0, + 1405.0, + 1026.0, + 1405.0, + 1061.0, + 295.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1053.0, + 935.0, + 1053.0, + 935.0, + 1092.0, + 292.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1053.0, + 1405.0, + 1053.0, + 1405.0, + 1092.0, + 1005.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1122.0, + 295.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1118.0, + 1403.0, + 1118.0, + 1403.0, + 1149.0, + 296.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1147.0, + 1405.0, + 1147.0, + 1405.0, + 1182.0, + 295.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1405.0, + 1177.0, + 1405.0, + 1214.0, + 294.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1209.0, + 809.0, + 1209.0, + 809.0, + 1241.0, + 296.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1422.0, + 296.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1414.0, + 624.0, + 1414.0, + 624.0, + 1455.0, + 292.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1414.0, + 1406.0, + 1414.0, + 1406.0, + 1455.0, + 734.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1442.0, + 501.0, + 1442.0, + 501.0, + 1488.0, + 290.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1442.0, + 1407.0, + 1442.0, + 1407.0, + 1488.0, + 565.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1406.0, + 1479.0, + 1406.0, + 1515.0, + 295.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1508.0, + 1405.0, + 1508.0, + 1405.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1536.0, + 1406.0, + 1536.0, + 1406.0, + 1577.0, + 292.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1567.0, + 1405.0, + 1567.0, + 1405.0, + 1607.0, + 292.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1637.0, + 294.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1663.0, + 762.0, + 1663.0, + 762.0, + 1695.0, + 296.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1695.0, + 862.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1691.0, + 1405.0, + 1691.0, + 1405.0, + 1727.0, + 295.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1723.0, + 1406.0, + 1723.0, + 1406.0, + 1756.0, + 293.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1753.0, + 1405.0, + 1753.0, + 1405.0, + 1789.0, + 295.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1782.0, + 877.0, + 1782.0, + 877.0, + 1819.0, + 294.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1782.0, + 1406.0, + 1782.0, + 1406.0, + 1819.0, + 976.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1252.0, + 1404.0, + 1252.0, + 1404.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1280.0, + 1408.0, + 1280.0, + 1408.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1314.0, + 1404.0, + 1314.0, + 1404.0, + 1346.0, + 296.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1344.0, + 1251.0, + 1344.0, + 1251.0, + 1378.0, + 293.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 288.0, + 376.0, + 288.0, + 376.0, + 326.0, + 304.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 288.0, + 464.0, + 288.0, + 464.0, + 326.0, + 437.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 288.0, + 565.0, + 288.0, + 565.0, + 326.0, + 522.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 319.0, + 335.0, + 319.0, + 335.0, + 352.0, + 305.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 321.0, + 413.0, + 321.0, + 413.0, + 351.0, + 374.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 321.0, + 501.0, + 321.0, + 501.0, + 351.0, + 473.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 321.0, + 565.0, + 321.0, + 565.0, + 351.0, + 523.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 347.0, + 335.0, + 347.0, + 335.0, + 381.0, + 305.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 347.0, + 512.0, + 347.0, + 512.0, + 382.0, + 412.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 347.0, + 749.0, + 347.0, + 749.0, + 382.0, + 587.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 347.0, + 909.0, + 347.0, + 909.0, + 382.0, + 893.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 379.0, + 335.0, + 379.0, + 335.0, + 411.0, + 304.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 374.0, + 544.0, + 374.0, + 544.0, + 415.0, + 410.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 374.0, + 911.0, + 374.0, + 911.0, + 415.0, + 730.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 374.0, + 957.0, + 374.0, + 957.0, + 415.0, + 940.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 407.0, + 335.0, + 407.0, + 335.0, + 438.0, + 304.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 408.0, + 460.0, + 408.0, + 460.0, + 437.0, + 377.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 436.0, + 335.0, + 436.0, + 335.0, + 467.0, + 304.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 431.0, + 506.0, + 431.0, + 506.0, + 473.0, + 372.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 431.0, + 870.0, + 431.0, + 870.0, + 473.0, + 638.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 431.0, + 917.0, + 431.0, + 917.0, + 473.0, + 898.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 460.0, + 425.0, + 460.0, + 425.0, + 496.0, + 302.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 483.0, + 926.0, + 483.0, + 926.0, + 530.0, + 300.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 483.0, + 1249.0, + 483.0, + 1249.0, + 530.0, + 1234.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 522.0, + 735.0, + 522.0, + 735.0, + 557.0, + 304.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 522.0, + 795.0, + 522.0, + 795.0, + 557.0, + 780.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 553.0, + 763.0, + 553.0, + 763.0, + 586.0, + 336.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 263.0, + 651.0, + 263.0, + 651.0, + 303.0, + 294.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 263.0, + 733.0, + 263.0, + 733.0, + 303.0, + 680.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 263.0, + 1025.0, + 263.0, + 1025.0, + 303.0, + 755.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1391.0, + 225.0, + 1391.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1532, + 1403, + 1532, + 1403, + 1808, + 297, + 1808 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1402, + 1881, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 334, + 280, + 1361, + 280, + 1361, + 625, + 334, + 625 + ], + "score": 0.965, + "html": "
databaseDetectorAUROC(%)Accuracy (%)
BIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
CIFAR-10KD+BU (Feinman et al. (2017))79.082.880.677.774.273.071.772.6
M-D (Lee et al. (2018))51.748.453.752.066.766.766.766.7
LID (Ma et al. (2018))87.987.486.682.772.678.578.070.2
MBF99.699.696.999.498.898.391.898.5
SVHNKD+BU (Feinman et al. (2017)) M-D (Lee et al. (2018))81.585.184.182.578.680.079.378.6
LID (Ma et al. (2018))49.750.049.650.166.766.766.766.7
91.788.792.290.583.380.884.081.5
ImageNet- AlexNetMBF KD+BU (Feinman et al. (2017))99.799.999.399.598.898.391.898.5
M-D (Lee et al. (2018))50.852.6 60.151.151.534.436.834.935.3
LID (Ma et al. (2018))48.1 71.770.946.5 71.954.1 72.366.7 68.766.6 60.566.866.7
ImageNet-MBF99.999.699.999.898.697.865.8 98.868.3
KD+BU (Feinman et al. (2017))57.157.756.258.642.843.641.798.4
M-D (Lee et al. (2018))63.764.845.966.967.165.666.744.8
89.567.4
VGG16LID (Ma et al. (2018)) MBF82.6 99.884.2 100.0100.084.2 100.077.8 99.676.7 99.683.3 100.076.4 100.0
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1428, + 1402, + 1428, + 1402, + 1520, + 299, + 1520 + ], + "score": 0.962 + }, + { + "category_id": 5, + "poly": [ + 333, + 1062, + 1364, + 1062, + 1364, + 1407, + 333, + 1407 + ], + "score": 0.924, + "html": "
databaseTest attack→BIM CW-L2 DeepFool R-PGDKD+BU(Feinman et al. (2017))/M-D(Lee et al. (2018))/LID (Ma et al. (2018)/MBF
CIFAR-10BIMCW-L2DeepFoolR-PGD74.2/66.7/72.6/98.874.2/66.7/80.3 /98.174.0/66.7 /80.4/95.273.1/66.7/77.7 /98.666.773.0/66.7 /78.5 /98.3 71.7 /66.7/77.9 /89.573.8/66.7/61.3/96.871.7 /66.7/77.9 /89.5 73.8 /66.7 /73.6 /98.2
66.771.9 / 66.7 /75.1/98.5 71.0 /66.7/75.4/89.773.4/66.7/74.8/95.471.0 /66.7/75.4/89.7 72.6/66.7/70.2/98.5
SVHNBIMCW-L2DeepFoolR-PGD78.6/66.77/83.3/98.7 80.5/66.7779.4/98.9 77.7/66.7/83.5/95.8 78.6/66.7/82.9/98.278.2/66.7/83.5 /98.0 80.0/66.7 /80.8/99.2 77.1/66.7/84.0 /95.9 78.1/66.7/80.7/97.379.8/66.7/83.9/98.0 81.8/66.7/79.9/98.4 79.3/66.7/84.0/97.3 79.9/66.7/80.7/98.078.6/66.7 /83.6 /98.6 80.5 / 66.7 /79.9/98.9 77.7 /66.7 /83.8/96.5 78.6/ 66.7 /81.5 /98.8
ImageNet-AlexNetBIMCW-L2DeepFoolR-PGD34.4/66.7/68.7798.6 34.4/66.6/68.6797.336.5/66.2/64.6/98.0 36.8/66.6/60.6/97.8 35.0 /66.1/60.7/98.134.3/66.7/65.0/98.435.0 /66.1/60.7/98.1 36.4/66.1/62.6/98.034.9/66.8/65.8/98.8 35.9 /66.6/68.3 /98.634.1/66.7/67.0/98.3 35.3/66.7/68.3/98.4
36.0 /66.7/69.4/98.835.4/66.7/67.5/98.566.835.2/66.7/67.7/97.6 34.1/66.7/67.0/98.3
ImageNet-VGG-16BIMCW-L2DeepFoolR-PGD42.8/67.1777.8/98.643.1/64.8/75.2/98.045.4/66.7 /76.2/98.844.8 /66.3/76.3/98.566.643.6/65.6/76.7 /97.8 38.4/ 62.5/82.5 /98.142.7/67.9/76.9/98.438.4/ 62.5/82.5 /98.1 43.3/67.2/73.9/98.041.8/66.7 /83.3/98.8 45.8 /66.7/74.4/98.640.6/64.9/82.4/98.3 45.8/67.4/76.4/98.4
66.745.4/66.6/74.1/97.6 40.6/64.9/82.4/98.3
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.91 + }, + { + "category_id": 6, + "poly": [ + 564, + 251, + 1132, + 251, + 1132, + 278, + 564, + 278 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 299, + 1831, + 474, + 1831, + 474, + 1861, + 299, + 1861 + ], + "score": 0.886 + }, + { + "category_id": 5, + "poly": [ + 332, + 681, + 1368, + 681, + 1368, + 1015, + 332, + 1015 + ], + "score": 0.845, + "html": "
databaseTest attack→BIMCW-L2DeepFoolR-PGD
TrainattackKD+BU (Feinman et al. (2017))/M-D(Lee et al.(2018))/LID (Ma et al.(2018)/MBF
CIFAR-10BIM CW-L2 DeepFool R-PGD79.0/51.7787.9799.6 83.4/47.9/90.3/99.6 83.6/53.5 /89.7 /99.3 78.1/51.5/86.1/99.678.1/52.2/85.7799.4 82.8/48.4/87.4/99.6 83.1/51.4/86.1/99.1 77.3 /51.9 /84.5 /99.576.4/52.4/86.4/87.8 80.3 /46.5 /87.0 / 88.6 80.6/53.7/86.6/96.978.6/50.9/87.6/99.4 82.9/48.2/88.9/99.5 83.1/53.6/87.9/98.9 77.7/52.1/82.7/99.4
SVHNBIM CW-L2 DeepFool R-PGD81.5749.7791.7799.7 83.6/50.8/91.6/99.7 85.0/50.0/91.6/99.7 82.5 /49.9 /91.2 /99.683.1/51.0/88.3/99.7 85.1/50.0/88.7/99.9 86.3/ 50.0 /87.9 /99.8 84.0/49.5/87.8/99.675.8/52.4/84.9/88.3 80.1/48.8/91.9/97.5 82.5/49.7/92.0/97.3 84.1/49.6/92.2/99.3 81.2/50.2/91.6/97.481.5750.4/90.7799.5 83.6/49.1/90.8/99.7 85.0/49.4/90.8 /99.6 82.5/50.1/90.5 /99.5
ImageNet- AlexNetBIM CW-L2 DeepFool R-PGD50.8/48.1/71.7/99.9 52.3/50.7/69.7/99.8 51.8/53.6/71.3/99.9 51.6/52.4/72.5 /99.850.8/44.9/70.8/99.6 52.6/60.1/70.9/99.7 51.9/53.5/69.7/99.7 51.3/ 50.4/72.3/99.650.4/49.7772.1/99.8 51.2/52.1/66.0/99.7 51.2/46.5 /71.9 /99.9 50.3 /50.5 /72.2/99.750.7/50.4/70.8/99.8 52.2/55.8/69.8/99.8 51.8/51.2/71.2/99.8
ImageNet- VGG-16BIM CW-L2 DeepFool R-PGD57.1/63.7/82.6/99.8 57.3/64.3/81.9/100.0 58.7/45.5/ 80.2/99.7 58.6/62.3/81.1/100.057.1/57.2/84.3/99.5 57.7/64.8/84.2/100.0 59.4/46.1/79.1/99.3 58.9/60.8/83.7/100.053.8/54.4/90.4/100.0 53.8/56.6/90.2/100.0 56.2/45.9/89.5/100.0 55.3/54.4/90.8/100.051.5 /54.1/72.3 /99.8 57.1/65.4/84.2/100.0 57.5/67.6/84.3/100.0 59.0 /44.0/ 80.9 /99.7 58.6/66.9/84.2/100.0
" + }, + { + "category_id": 6, + "poly": [ + 408, + 643, + 1290, + 643, + 1290, + 675, + 408, + 675 + ], + "score": 0.656 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2110, + 842, + 2110 + ], + "score": 0.637 + }, + { + "category_id": 6, + "poly": [ + 405, + 1026, + 1292, + 1026, + 1292, + 1056, + 405, + 1056 + ], + "score": 0.629 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2110, + 842, + 2110 + ], + "score": 0.236 + }, + { + "category_id": 13, + "poly": [ + 946, + 643, + 988, + 643, + 988, + 675, + 946, + 675 + ], + "score": 0.75, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 948, + 1026, + 988, + 1026, + 988, + 1056, + 948, + 1056 + ], + "score": 0.71, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 637, + 1943, + 735, + 1943, + 735, + 1972, + 637, + 1972 + ], + "score": 0.67, + "latex": "\\mathrm { K D + B U }" + }, + { + "category_id": 13, + "poly": [ + 987, + 1973, + 1085, + 1973, + 1085, + 2002, + 987, + 2002 + ], + "score": 0.65, + "latex": "\\mathrm { K D + B U }" + }, + { + "category_id": 13, + "poly": [ + 533, + 1066, + 554, + 1066, + 554, + 1083, + 533, + 1083 + ], + "score": 0.42, + "latex": "" + }, + { + "category_id": 13, + "poly": [ + 526, + 683, + 547, + 683, + 547, + 699, + 526, + 699 + ], + "score": 0.26, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 244.0, + 1135.0, + 244.0, + 1135.0, + 285.0, + 564.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1828.0, + 476.0, + 1828.0, + 476.0, + 1866.0, + 294.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 638.0, + 945.0, + 638.0, + 945.0, + 681.0, + 403.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 638.0, + 1296.0, + 638.0, + 1296.0, + 681.0, + 989.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1021.0, + 947.0, + 1021.0, + 947.0, + 1062.0, + 400.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1021.0, + 1298.0, + 1021.0, + 1298.0, + 1062.0, + 989.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1533.0, + 1404.0, + 1533.0, + 1404.0, + 1566.0, + 297.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1564.0, + 1405.0, + 1564.0, + 1405.0, + 1597.0, + 296.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1592.0, + 1407.0, + 1592.0, + 1407.0, + 1630.0, + 292.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1622.0, + 1405.0, + 1622.0, + 1405.0, + 1660.0, + 292.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1653.0, + 1407.0, + 1653.0, + 1407.0, + 1691.0, + 292.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1682.0, + 1405.0, + 1682.0, + 1405.0, + 1722.0, + 295.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1717.0, + 1404.0, + 1717.0, + 1404.0, + 1750.0, + 296.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1747.0, + 1404.0, + 1747.0, + 1404.0, + 1780.0, + 296.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1776.0, + 703.0, + 1776.0, + 703.0, + 1814.0, + 292.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1943.0, + 636.0, + 1943.0, + 636.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 736.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 986.0, + 1973.0, + 986.0, + 2006.0, + 297.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 1086.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2002.0, + 1403.0, + 2002.0, + 1403.0, + 2035.0, + 297.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1424.0, + 1402.0, + 1424.0, + 1402.0, + 1466.0, + 293.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1457.0, + 1404.0, + 1457.0, + 1404.0, + 1493.0, + 293.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1490.0, + 1250.0, + 1490.0, + 1250.0, + 1527.0, + 294.0, + 1527.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1541, + 1405, + 1541, + 1405, + 1877, + 297, + 1877 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 892, + 1404, + 892, + 1404, + 1108, + 298, + 1108 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1404, + 1403, + 1404, + 1403, + 1528, + 298, + 1528 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1199, + 1403, + 1199, + 1403, + 1324, + 297, + 1324 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 755, + 1404, + 755, + 1404, + 877, + 298, + 877 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 647, + 1403, + 647, + 1403, + 740, + 298, + 740 + ], + "score": 0.968 + }, + { + "category_id": 5, + "poly": [ + 425, + 347, + 1272, + 347, + 1272, + 496, + 425, + 496 + ], + "score": 0.962, + "html": "
ImageNet-AlexNetImageNet-VGG-16
DetectorBIMCW-L2DeepFoolR-PGDBIMCW-L2DeepFoolR-PGD
KD+BU66.369.464.768.884.085.079.985.2
M-D48.749.250.051.054.550.050.456.1
LID69.167.672.971.484.288.489.283.7
MBF99.199.099.699.499.399.599.399.5
" + }, + { + "category_id": 5, + "poly": [ + 628, + 550, + 1071, + 550, + 1071, + 630, + 628, + 630 + ], + "score": 0.956, + "html": "
Original CW-L2LIDMBF
CIFAR-100.088.4100.0
SVHN0.085.692.9
" + }, + { + "category_id": 6, + "poly": [ + 300, + 249, + 1401, + 249, + 1401, + 341, + 300, + 341 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1398, + 1971, + 1398, + 2034, + 297, + 2034 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 298, + 1915, + 1010, + 1915, + 1010, + 1946, + 298, + 1946 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 297, + 1144, + 1041, + 1144, + 1041, + 1175, + 297, + 1175 + ], + "score": 0.911 + }, + { + "category_id": 8, + "poly": [ + 384, + 1332, + 1311, + 1332, + 1311, + 1392, + 384, + 1392 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1340, + 1399, + 1340, + 1399, + 1369, + 1366, + 1369 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.793 + }, + { + "category_id": 6, + "poly": [ + 301, + 514, + 1401, + 514, + 1401, + 545, + 301, + 545 + ], + "score": 0.659 + }, + { + "category_id": 1, + "poly": [ + 301, + 514, + 1401, + 514, + 1401, + 545, + 301, + 545 + ], + "score": 0.271 + }, + { + "category_id": 13, + "poly": [ + 1080, + 1404, + 1193, + 1404, + 1193, + 1438, + 1080, + 1438 + ], + "score": 0.93, + "latex": "l ( \\mathbf { x } _ { a d v } , y )" + }, + { + "category_id": 13, + "poly": [ + 662, + 1465, + 995, + 1465, + 995, + 1499, + 662, + 1499 + ], + "score": 0.93, + "latex": "| | D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) ^ { * } - D ( f _ { \\theta } ( \\mathbf { x } ) ) | | _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1289, + 1436, + 1393, + 1436, + 1393, + 1469, + 1289, + 1469 + ], + "score": 0.92, + "latex": "f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } )" + }, + { + "category_id": 13, + "poly": [ + 980, + 1785, + 1111, + 1785, + 1111, + 1823, + 980, + 1823 + ], + "score": 0.92, + "latex": "[ \\bar { 1 0 } ^ { - 3 } , 1 0 ^ { 6 } ]" + }, + { + "category_id": 13, + "poly": [ + 298, + 1435, + 446, + 1435, + 446, + 1469, + 298, + 1469 + ], + "score": 0.91, + "latex": "D ( f _ { \\pmb { \\theta } } ( \\mathbf { x } _ { a d v } ) )" + }, + { + "category_id": 14, + "poly": [ + 383, + 1329, + 1311, + 1329, + 1311, + 1397, + 383, + 1397 + ], + "score": 0.89, + "latex": "\\operatorname * { a r g m i n } _ { \\mathbf { x } _ { a d v } \\in [ 0 , 1 ] } \\| \\mathbf { x } - \\mathbf { x } _ { a d v } \\| _ { 2 } ^ { 2 } + \\alpha \\cdot \\big [ - l ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) , y ) + \\| D ( f _ { \\theta } ( \\mathbf { x } _ { a d v } ) ) - D ( f _ { \\theta } ( \\mathbf { x } ) ) \\| _ { 1 } \\big ] ," + }, + { + "category_id": 13, + "poly": [ + 572, + 1466, + 599, + 1466, + 599, + 1496, + 572, + 1496 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1757, + 324, + 1757, + 324, + 1788, + 297, + 1788 + ], + "score": 0.86, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 734, + 1605, + 764, + 1605, + 764, + 1636, + 734, + 1636 + ], + "score": 0.86, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1015, + 1437, + 1042, + 1437, + 1042, + 1463, + 1015, + 1463 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 858, + 1665, + 955, + 1665, + 955, + 1694, + 858, + 1694 + ], + "score": 0.8, + "latex": "\\mathrm { K D + B U }" + }, + { + "category_id": 13, + "poly": [ + 678, + 1411, + 698, + 1411, + 698, + 1437, + 678, + 1437 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 415, + 1792, + 437, + 1792, + 437, + 1814, + 415, + 1814 + ], + "score": 0.77, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 892, + 251, + 935, + 251, + 935, + 280, + 892, + 280 + ], + "score": 0.75, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 624, + 516, + 666, + 516, + 666, + 545, + 624, + 545 + ], + "score": 0.75, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1410, + 396, + 1410, + 396, + 1432, + 373, + 1432 + ], + "score": 0.6, + "latex": "\\mathbf { x }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 247.0, + 891.0, + 247.0, + 891.0, + 284.0, + 294.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 247.0, + 1405.0, + 247.0, + 1405.0, + 284.0, + 936.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 281.0, + 1405.0, + 281.0, + 1405.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 308.0, + 853.0, + 308.0, + 853.0, + 343.0, + 295.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1015.0, + 1911.0, + 1015.0, + 1951.0, + 293.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1143.0, + 1043.0, + 1143.0, + 1043.0, + 1178.0, + 295.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 513.0, + 623.0, + 513.0, + 623.0, + 547.0, + 297.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 513.0, + 1402.0, + 513.0, + 1402.0, + 547.0, + 667.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 1403.0, + 1542.0, + 1403.0, + 1578.0, + 295.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1575.0, + 1403.0, + 1575.0, + 1403.0, + 1606.0, + 296.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1606.0, + 733.0, + 1606.0, + 733.0, + 1637.0, + 295.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1606.0, + 1402.0, + 1606.0, + 1402.0, + 1637.0, + 765.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1403.0, + 1633.0, + 1403.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1662.0, + 857.0, + 1662.0, + 857.0, + 1699.0, + 294.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1662.0, + 1406.0, + 1662.0, + 1406.0, + 1699.0, + 956.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1695.0, + 1405.0, + 1695.0, + 1405.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1725.0, + 1406.0, + 1725.0, + 1406.0, + 1760.0, + 294.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1791.0, + 325.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1783.0, + 414.0, + 1783.0, + 414.0, + 1822.0, + 291.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1783.0, + 979.0, + 1783.0, + 979.0, + 1822.0, + 438.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1783.0, + 1406.0, + 1783.0, + 1406.0, + 1822.0, + 1112.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1851.0, + 295.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1848.0, + 884.0, + 1848.0, + 884.0, + 1879.0, + 295.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 890.0, + 1405.0, + 890.0, + 1405.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 921.0, + 1404.0, + 921.0, + 1404.0, + 961.0, + 293.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 1403.0, + 951.0, + 1403.0, + 990.0, + 293.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 984.0, + 1406.0, + 984.0, + 1406.0, + 1021.0, + 292.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1014.0, + 1405.0, + 1014.0, + 1405.0, + 1049.0, + 292.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1042.0, + 1408.0, + 1042.0, + 1408.0, + 1083.0, + 292.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1075.0, + 925.0, + 1075.0, + 925.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 372.0, + 1403.0, + 372.0, + 1439.0, + 294.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1403.0, + 677.0, + 1403.0, + 677.0, + 1439.0, + 397.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1403.0, + 1079.0, + 1403.0, + 1079.0, + 1439.0, + 699.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1403.0, + 1404.0, + 1403.0, + 1404.0, + 1439.0, + 1194.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1431.0, + 297.0, + 1431.0, + 297.0, + 1473.0, + 292.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1431.0, + 1014.0, + 1431.0, + 1014.0, + 1473.0, + 447.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1431.0, + 1288.0, + 1431.0, + 1288.0, + 1473.0, + 1043.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1431.0, + 1407.0, + 1431.0, + 1407.0, + 1473.0, + 1394.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1462.0, + 571.0, + 1462.0, + 571.0, + 1501.0, + 292.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1462.0, + 661.0, + 1462.0, + 661.0, + 1501.0, + 600.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1462.0, + 1408.0, + 1462.0, + 1408.0, + 1501.0, + 996.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1498.0, + 1231.0, + 1498.0, + 1231.0, + 1528.0, + 293.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1202.0, + 1404.0, + 1202.0, + 1404.0, + 1235.0, + 296.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1232.0, + 1405.0, + 1232.0, + 1405.0, + 1265.0, + 295.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1262.0, + 1405.0, + 1262.0, + 1405.0, + 1297.0, + 291.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1293.0, + 1378.0, + 1293.0, + 1378.0, + 1326.0, + 295.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 756.0, + 1404.0, + 756.0, + 1404.0, + 788.0, + 294.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 786.0, + 1404.0, + 786.0, + 1404.0, + 818.0, + 296.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 816.0, + 1405.0, + 816.0, + 1405.0, + 851.0, + 293.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 847.0, + 734.0, + 847.0, + 734.0, + 876.0, + 296.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 646.0, + 1407.0, + 646.0, + 1407.0, + 684.0, + 294.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 677.0, + 1402.0, + 677.0, + 1402.0, + 710.0, + 296.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 709.0, + 1399.0, + 709.0, + 1399.0, + 743.0, + 294.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2008.0, + 296.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 513.0, + 623.0, + 513.0, + 623.0, + 547.0, + 297.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 513.0, + 1402.0, + 513.0, + 1402.0, + 547.0, + 667.0, + 547.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 720, + 1404, + 720, + 1404, + 998, + 297, + 998 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1253, + 1405, + 1253, + 1405, + 1560, + 297, + 1560 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1575, + 1404, + 1575, + 1404, + 1739, + 298, + 1739 + ], + "score": 0.969 + }, + { + "category_id": 5, + "poly": [ + 435, + 316, + 1262, + 316, + 1262, + 408, + 435, + 408 + ], + "score": 0.963, + "html": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
" + }, + { + "category_id": 1, + "poly": [ + 301, + 1012, + 1401, + 1012, + 1401, + 1136, + 301, + 1136 + ], + "score": 0.962 + }, + { + "category_id": 5, + "poly": [ + 362, + 464, + 1333, + 464, + 1333, + 555, + 362, + 555 + ], + "score": 0.948, + "html": "
ImageNet-VGG-16(clean, noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
" + }, + { + "category_id": 1, + "poly": [ + 368, + 1762, + 1404, + 1762, + 1404, + 1989, + 368, + 1989 + ], + "score": 0.941 + }, + { + "category_id": 6, + "poly": [ + 297, + 248, + 1400, + 248, + 1400, + 311, + 297, + 311 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 369, + 1157, + 1325, + 1157, + 1325, + 1233, + 369, + 1233 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 299, + 427, + 1400, + 427, + 1400, + 459, + 299, + 459 + ], + "score": 0.81 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.804 + }, + { + "category_id": 6, + "poly": [ + 332, + 574, + 1367, + 574, + 1367, + 606, + 332, + 606 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 333, + 2004, + 1198, + 2004, + 1198, + 2035, + 333, + 2035 + ], + "score": 0.442 + }, + { + "category_id": 2, + "poly": [ + 333, + 2004, + 1198, + 2004, + 1198, + 2035, + 333, + 2035 + ], + "score": 0.231 + }, + { + "category_id": 13, + "poly": [ + 528, + 1678, + 588, + 1678, + 588, + 1706, + 528, + 1706 + ], + "score": 0.91, + "latex": "\\mathbf { m } _ { \\mathrm { b e n } }" + }, + { + "category_id": 13, + "poly": [ + 396, + 1795, + 630, + 1795, + 630, + 1835, + 396, + 1835 + ], + "score": 0.91, + "latex": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1255, + 809, + 1255, + 809, + 1285, + 707, + 1285 + ], + "score": 0.9, + "latex": "\\mathcal { P } _ { \\mathrm { G G D - a d v } }" + }, + { + "category_id": 13, + "poly": [ + 1075, + 1074, + 1140, + 1074, + 1140, + 1105, + 1075, + 1105 + ], + "score": 0.89, + "latex": "{ \\mathcal { P } } _ { \\mathrm { G G D } }" + }, + { + "category_id": 13, + "poly": [ + 660, + 1316, + 760, + 1316, + 760, + 1347, + 660, + 1347 + ], + "score": 0.89, + "latex": "\\mathcal { P } _ { \\mathrm { G G D - a d v } }" + }, + { + "category_id": 13, + "poly": [ + 560, + 1645, + 619, + 1645, + 619, + 1671, + 560, + 1671 + ], + "score": 0.88, + "latex": "\\mathbf { m _ { \\mathrm { a d v } } }" + }, + { + "category_id": 13, + "poly": [ + 1355, + 1843, + 1392, + 1843, + 1392, + 1872, + 1355, + 1872 + ], + "score": 0.88, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1046, + 1297, + 1046, + 1297, + 1075, + 1246, + 1075 + ], + "score": 0.87, + "latex": "\\bf p \\mathrm { _ { a d v } }" + }, + { + "category_id": 13, + "poly": [ + 1106, + 1260, + 1156, + 1260, + 1156, + 1286, + 1106, + 1286 + ], + "score": 0.87, + "latex": "\\mathbf { p } _ { \\mathrm { a d v } }" + }, + { + "category_id": 13, + "poly": [ + 396, + 1871, + 627, + 1871, + 627, + 1912, + 396, + 1912 + ], + "score": 0.87, + "latex": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 441, + 845, + 555, + 845, + 555, + 874, + 441, + 874 + ], + "score": 0.86, + "latex": "\\alpha = 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 640, + 1670, + 739, + 1670, + 739, + 1707, + 640, + 1707 + ], + "score": 0.85, + "latex": "\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }" + }, + { + "category_id": 13, + "poly": [ + 1354, + 1765, + 1392, + 1765, + 1392, + 1798, + 1354, + 1798 + ], + "score": 0.85, + "latex": "H _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 832, + 1161, + 868, + 1161, + 868, + 1190, + 832, + 1190 + ], + "score": 0.83, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 584, + 1076, + 634, + 1076, + 634, + 1106, + 584, + 1106 + ], + "score": 0.83, + "latex": "\\mathbf { p } _ { \\mathrm { { b e n } } }" + }, + { + "category_id": 13, + "poly": [ + 614, + 1950, + 650, + 1950, + 650, + 1983, + 614, + 1983 + ], + "score": 0.83, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 666, + 1946, + 893, + 1946, + 893, + 1987, + 666, + 1987 + ], + "score": 0.81, + "latex": "\\begin{array} { r } { \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 896, + 818, + 914, + 818, + 914, + 845, + 896, + 845 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1241, + 788, + 1259, + 788, + 1259, + 815, + 1241, + 815 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1091, + 1761, + 1327, + 1761, + 1327, + 1798, + 1091, + 1798 + ], + "score": 0.8, + "latex": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } }" + }, + { + "category_id": 13, + "poly": [ + 1190, + 1443, + 1207, + 1443, + 1207, + 1470, + 1190, + 1470 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 453, + 255, + 470, + 255, + 470, + 281, + 453, + 281 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1159, + 910, + 1178, + 910, + 1178, + 937, + 1159, + 937 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1350, + 1320, + 1401, + 1320, + 1401, + 1347, + 1350, + 1347 + ], + "score": 0.79, + "latex": "\\mathbf { p } _ { \\mathrm { a d v } }" + }, + { + "category_id": 13, + "poly": [ + 848, + 1351, + 866, + 1351, + 866, + 1378, + 848, + 1378 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1294, + 1635, + 1394, + 1635, + 1394, + 1672, + 1294, + 1672 + ], + "score": 0.79, + "latex": "\\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } }" + }, + { + "category_id": 13, + "poly": [ + 881, + 1160, + 1066, + 1160, + 1066, + 1192, + 881, + 1192 + ], + "score": 0.78, + "latex": "\\mathbf { p } _ { \\mathrm { a d v } } \\sim \\mathcal { P } _ { \\mathrm { G G D - a d v } }" + }, + { + "category_id": 13, + "poly": [ + 432, + 1382, + 449, + 1382, + 449, + 1409, + 432, + 1409 + ], + "score": 0.78, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 476, + 579, + 495, + 579, + 495, + 607, + 476, + 607 + ], + "score": 0.78, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 786, + 1200, + 823, + 1200, + 823, + 1231, + 786, + 1231 + ], + "score": 0.77, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 443, + 433, + 461, + 433, + 461, + 459, + 443, + 459 + ], + "score": 0.76, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 919, + 1946, + 1202, + 1946, + 1202, + 1987, + 919, + 1987 + ], + "score": 0.75, + "latex": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { s o u r c e - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 1379, + 818, + 1402, + 818, + 1402, + 841, + 1379, + 841 + ], + "score": 0.73, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1764, + 1076, + 1764, + 1076, + 1797, + 1040, + 1797 + ], + "score": 0.73, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 651, + 1871, + 934, + 1871, + 934, + 1912, + 651, + 1912 + ], + "score": 0.73, + "latex": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { a d v - M B F } } ^ { \\mathrm { a t t a c k - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 940, + 318, + 940, + 318, + 962, + 298, + 962 + ], + "score": 0.72, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 836, + 1200, + 1020, + 1200, + 1020, + 1232, + 836, + 1232 + ], + "score": 0.72, + "latex": "\\mathbf { p } _ { \\mathrm { b e n } } \\sim \\mathcal { P } _ { \\mathrm { G G D - b e n } }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 1160, + 1119, + 1160, + 1119, + 1191, + 1083, + 1191 + ], + "score": 0.72, + "latex": "H _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 651, + 1873, + 689, + 1873, + 689, + 1909, + 651, + 1909 + ], + "score": 0.58, + "latex": "H _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1200, + 1074, + 1200, + 1074, + 1231, + 1037, + 1231 + ], + "score": 0.53, + "latex": "H _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 919, + 1948, + 955, + 1948, + 955, + 1984, + 919, + 1984 + ], + "score": 0.3, + "latex": "H _ { 1 }" + }, + { + "category_id": 5, + "poly": [ + 385, + 612, + 1305, + 612, + 1305, + 702, + 385, + 702 + ], + "score": 0.85, + "html": "
ImageNet-VGG-16cleannoisyBIMCW-L2DeepFoolR-PGD
(train set,test set)0.8390.8580.3680.8420.2460.221
(train set, out-of-sample set)0.0490.0540.0940.4070.7370.088
" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 245.0, + 452.0, + 245.0, + 452.0, + 286.0, + 293.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 245.0, + 1405.0, + 245.0, + 1405.0, + 286.0, + 471.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 278.0, + 1078.0, + 278.0, + 1078.0, + 315.0, + 292.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 421.0, + 442.0, + 421.0, + 442.0, + 465.0, + 293.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 421.0, + 1404.0, + 421.0, + 1404.0, + 465.0, + 462.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 572.0, + 475.0, + 572.0, + 475.0, + 611.0, + 328.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 572.0, + 1372.0, + 572.0, + 1372.0, + 611.0, + 496.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1998.0, + 1200.0, + 1998.0, + 1200.0, + 2041.0, + 331.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 723.0, + 1402.0, + 723.0, + 1402.0, + 755.0, + 296.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 753.0, + 1405.0, + 753.0, + 1405.0, + 785.0, + 296.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 780.0, + 1240.0, + 780.0, + 1240.0, + 818.0, + 294.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 780.0, + 1406.0, + 780.0, + 1406.0, + 818.0, + 1260.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 813.0, + 895.0, + 813.0, + 895.0, + 848.0, + 294.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 813.0, + 1378.0, + 813.0, + 1378.0, + 848.0, + 915.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 813.0, + 1406.0, + 813.0, + 1406.0, + 848.0, + 1403.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 845.0, + 440.0, + 845.0, + 440.0, + 877.0, + 296.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 845.0, + 1405.0, + 845.0, + 1405.0, + 877.0, + 556.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 873.0, + 1404.0, + 873.0, + 1404.0, + 912.0, + 292.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 902.0, + 1158.0, + 902.0, + 1158.0, + 940.0, + 292.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 902.0, + 1405.0, + 902.0, + 1405.0, + 940.0, + 1179.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 934.0, + 297.0, + 934.0, + 297.0, + 969.0, + 294.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 934.0, + 1406.0, + 934.0, + 1406.0, + 969.0, + 319.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 964.0, + 990.0, + 964.0, + 990.0, + 1001.0, + 293.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1254.0, + 706.0, + 1254.0, + 706.0, + 1289.0, + 294.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1254.0, + 1105.0, + 1254.0, + 1105.0, + 1289.0, + 810.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1254.0, + 1405.0, + 1254.0, + 1405.0, + 1289.0, + 1157.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1286.0, + 1403.0, + 1286.0, + 1403.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1311.0, + 659.0, + 1311.0, + 659.0, + 1354.0, + 291.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1311.0, + 1349.0, + 1311.0, + 1349.0, + 1354.0, + 761.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1311.0, + 1406.0, + 1311.0, + 1406.0, + 1354.0, + 1402.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1345.0, + 847.0, + 1345.0, + 847.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1345.0, + 1407.0, + 1345.0, + 1407.0, + 1381.0, + 867.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1376.0, + 431.0, + 1376.0, + 431.0, + 1410.0, + 292.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1410.0, + 450.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1408.0, + 1403.0, + 1408.0, + 1403.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1437.0, + 1189.0, + 1437.0, + 1189.0, + 1473.0, + 294.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1437.0, + 1405.0, + 1437.0, + 1405.0, + 1473.0, + 1208.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1469.0, + 1403.0, + 1469.0, + 1403.0, + 1501.0, + 295.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1526.0, + 997.0, + 1526.0, + 997.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1607.0, + 296.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1636.0, + 559.0, + 1636.0, + 559.0, + 1678.0, + 291.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1636.0, + 1293.0, + 1636.0, + 1293.0, + 1678.0, + 620.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1636.0, + 1406.0, + 1636.0, + 1406.0, + 1678.0, + 1395.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1672.0, + 527.0, + 1672.0, + 527.0, + 1711.0, + 293.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1672.0, + 639.0, + 1672.0, + 639.0, + 1711.0, + 589.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1672.0, + 1405.0, + 1672.0, + 1405.0, + 1711.0, + 740.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1708.0, + 539.0, + 1708.0, + 539.0, + 1741.0, + 294.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1012.0, + 1404.0, + 1012.0, + 1404.0, + 1048.0, + 296.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1042.0, + 1245.0, + 1042.0, + 1245.0, + 1079.0, + 295.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 1042.0, + 1407.0, + 1042.0, + 1407.0, + 1079.0, + 1298.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1075.0, + 583.0, + 1075.0, + 583.0, + 1108.0, + 296.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1075.0, + 1074.0, + 1075.0, + 1074.0, + 1108.0, + 635.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1075.0, + 1405.0, + 1075.0, + 1405.0, + 1108.0, + 1141.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1104.0, + 656.0, + 1104.0, + 656.0, + 1140.0, + 296.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1762.0, + 1039.0, + 1762.0, + 1039.0, + 1804.0, + 365.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1762.0, + 1090.0, + 1762.0, + 1090.0, + 1804.0, + 1077.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1762.0, + 1353.0, + 1762.0, + 1353.0, + 1804.0, + 1328.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1762.0, + 1405.0, + 1762.0, + 1405.0, + 1804.0, + 1393.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1795.0, + 395.0, + 1795.0, + 395.0, + 1841.0, + 392.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1795.0, + 644.0, + 1795.0, + 644.0, + 1841.0, + 631.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1838.0, + 1354.0, + 1838.0, + 1354.0, + 1876.0, + 367.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1838.0, + 1404.0, + 1838.0, + 1404.0, + 1876.0, + 1393.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1856.0, + 395.0, + 1856.0, + 395.0, + 1926.0, + 385.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1856.0, + 650.0, + 1856.0, + 650.0, + 1926.0, + 628.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1856.0, + 959.0, + 1856.0, + 959.0, + 1926.0, + 935.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1914.0, + 1404.0, + 1914.0, + 1404.0, + 1949.0, + 386.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1935.0, + 613.0, + 1935.0, + 613.0, + 2000.0, + 384.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1935.0, + 665.0, + 1935.0, + 665.0, + 2000.0, + 651.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1964.0, + 910.0, + 1964.0, + 910.0, + 1987.0, + 894.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1964.0, + 1211.0, + 1964.0, + 1211.0, + 1988.0, + 1203.0, + 1988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1871.0, + 492.0, + 1871.0, + 492.0, + 1897.0, + 396.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1941.0, + 915.0, + 1941.0, + 915.0, + 1980.5, + 760.0, + 1980.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1943.5, + 1075.0, + 1943.5, + 1075.0, + 1984.5, + 909.0, + 1984.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1152.0, + 831.0, + 1152.0, + 831.0, + 1198.0, + 363.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1152.0, + 880.0, + 1152.0, + 880.0, + 1198.0, + 869.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1152.0, + 1082.0, + 1152.0, + 1082.0, + 1198.0, + 1067.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1152.0, + 1328.0, + 1152.0, + 1328.0, + 1198.0, + 1120.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1194.0, + 785.0, + 1194.0, + 785.0, + 1238.0, + 365.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1194.0, + 835.0, + 1194.0, + 835.0, + 1238.0, + 824.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1194.0, + 1036.0, + 1194.0, + 1036.0, + 1238.0, + 1021.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1194.0, + 1282.0, + 1194.0, + 1282.0, + 1238.0, + 1075.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1998.0, + 1200.0, + 1998.0, + 1200.0, + 2041.0, + 331.0, + 2041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1348, + 1403, + 1348, + 1403, + 1714, + 298, + 1714 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 785, + 1404, + 785, + 1404, + 1332, + 298, + 1332 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1403, + 1788, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 379, + 232, + 1315, + 232, + 1315, + 592, + 379, + 592 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 370, + 691, + 1399, + 691, + 1399, + 764, + 370, + 764 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 858, + 76, + 858, + 105, + 297, + 105 + ], + "score": 0.904 + }, + { + "category_id": 4, + "poly": [ + 296, + 604, + 1402, + 604, + 1402, + 667, + 296, + 667 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 1736, + 544, + 1736, + 544, + 1772, + 300, + 1772 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.853 + }, + { + "category_id": 13, + "poly": [ + 1362, + 998, + 1401, + 998, + 1401, + 1028, + 1362, + 1028 + ], + "score": 0.88, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 614, + 724, + 650, + 724, + 650, + 758, + 614, + 758 + ], + "score": 0.83, + "latex": "H _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 743, + 1093, + 761, + 1093, + 761, + 1120, + 743, + 1120 + ], + "score": 0.82, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1088, + 1154, + 1106, + 1154, + 1106, + 1181, + 1088, + 1181 + ], + "score": 0.82, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 708, + 1246, + 725, + 1246, + 725, + 1272, + 708, + 1272 + ], + "score": 0.82, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 922, + 971, + 939, + 971, + 939, + 999, + 922, + 999 + ], + "score": 0.82, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 672, + 1003, + 690, + 1003, + 690, + 1029, + 672, + 1029 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1271, + 1093, + 1289, + 1093, + 1289, + 1120, + 1271, + 1120 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 349, + 1063, + 368, + 1063, + 368, + 1091, + 349, + 1091 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 477, + 1183, + 495, + 1183, + 495, + 1212, + 477, + 1212 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1135, + 851, + 1152, + 851, + 1152, + 878, + 1135, + 878 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 389, + 972, + 407, + 972, + 407, + 999, + 389, + 999 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1272, + 1215, + 1290, + 1215, + 1290, + 1242, + 1272, + 1242 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 665, + 721, + 893, + 721, + 893, + 762, + 665, + 762 + ], + "score": 0.79, + "latex": "\\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } = \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 918, + 722, + 1203, + 722, + 1203, + 763, + 918, + 763 + ], + "score": 0.65, + "latex": "H _ { 1 } \\colon \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 1 } } \\neq \\hat { \\mathcal { P } } _ { \\mathrm { b e n - M B F } } ^ { \\mathrm { s o u r c e - 2 } }" + }, + { + "category_id": 13, + "poly": [ + 918, + 723, + 956, + 723, + 956, + 759, + 918, + 759 + ], + "score": 0.62, + "latex": "H _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 591, + 607, + 619, + 607, + 619, + 634, + 591, + 634 + ], + "score": 0.58, + "latex": "\\pm" + }, + { + "category_id": 13, + "poly": [ + 486, + 1058, + 547, + 1058, + 547, + 1087, + 486, + 1087 + ], + "score": 0.32, + "latex": "{ \\bf H } 2 . 2" + }, + { + "category_id": 13, + "poly": [ + 1220, + 1150, + 1281, + 1150, + 1281, + 1178, + 1220, + 1178 + ], + "score": 0.3, + "latex": "{ \\bf H } 2 . 3" + }, + { + "category_id": 13, + "poly": [ + 1052, + 967, + 1112, + 967, + 1112, + 995, + 1052, + 995 + ], + "score": 0.28, + "latex": "{ \\bf H } 2 . 1" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 255.0, + 863.0, + 255.0, + 863.0, + 314.0, + 766.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 268.0, + 755.0, + 268.0, + 755.0, + 300.0, + 725.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 340.0, + 613.0, + 340.0, + 613.0, + 424.0, + 397.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 379.0, + 646.0, + 379.0, + 646.0, + 387.0, + 634.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 342.0, + 986.0, + 342.0, + 986.0, + 447.0, + 695.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 467.0, + 579.0, + 467.0, + 579.0, + 518.0, + 514.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 446.0, + 1294.0, + 446.0, + 1294.0, + 570.0, + 683.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 563.0, + 500.0, + 563.0, + 500.0, + 596.0, + 406.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 558.0, + 659.0, + 558.0, + 659.0, + 601.0, + 563.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 562.0, + 813.0, + 562.0, + 813.0, + 595.0, + 721.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 561.0, + 983.0, + 561.0, + 983.0, + 594.0, + 870.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 561.0, + 1152.0, + 561.0, + 1152.0, + 598.0, + 1015.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 562.0, + 1294.0, + 562.0, + 1294.0, + 594.0, + 1184.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.25, + 271.0, + 1243.25, + 271.0, + 1243.25, + 296.0, + 1197.25, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 604.0, + 590.0, + 604.0, + 590.0, + 639.0, + 295.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 604.0, + 1404.0, + 604.0, + 1404.0, + 639.0, + 620.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 634.0, + 1020.0, + 634.0, + 1020.0, + 670.0, + 292.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1730.0, + 550.0, + 1730.0, + 550.0, + 1781.0, + 290.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 1404.0, + 1349.0, + 1404.0, + 1383.0, + 294.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1379.0, + 1407.0, + 1379.0, + 1407.0, + 1413.0, + 293.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1410.0, + 1404.0, + 1410.0, + 1404.0, + 1444.0, + 294.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1441.0, + 1404.0, + 1441.0, + 1404.0, + 1475.0, + 296.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1469.0, + 1405.0, + 1469.0, + 1405.0, + 1504.0, + 293.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1503.0, + 1404.0, + 1503.0, + 1404.0, + 1533.0, + 296.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1404.0, + 1561.0, + 1404.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1588.0, + 1405.0, + 1588.0, + 1405.0, + 1629.0, + 293.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1624.0, + 1403.0, + 1624.0, + 1403.0, + 1655.0, + 296.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1654.0, + 1404.0, + 1654.0, + 1404.0, + 1686.0, + 293.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1684.0, + 809.0, + 1684.0, + 809.0, + 1717.0, + 293.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 1404.0, + 786.0, + 1404.0, + 817.0, + 294.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 851.0, + 293.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 846.0, + 1134.0, + 846.0, + 1134.0, + 880.0, + 292.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 846.0, + 1404.0, + 846.0, + 1404.0, + 880.0, + 1153.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 909.0, + 293.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 904.0, + 1405.0, + 904.0, + 1405.0, + 940.0, + 294.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 938.0, + 1405.0, + 938.0, + 1405.0, + 970.0, + 296.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 965.0, + 388.0, + 965.0, + 388.0, + 1004.0, + 292.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 965.0, + 921.0, + 965.0, + 921.0, + 1004.0, + 408.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 965.0, + 1051.0, + 965.0, + 1051.0, + 1004.0, + 940.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 965.0, + 1408.0, + 965.0, + 1408.0, + 1004.0, + 1113.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 671.0, + 995.0, + 671.0, + 1033.0, + 293.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 995.0, + 1361.0, + 995.0, + 1361.0, + 1033.0, + 691.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1027.0, + 1406.0, + 1027.0, + 1406.0, + 1062.0, + 293.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1059.0, + 348.0, + 1059.0, + 348.0, + 1094.0, + 294.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1059.0, + 485.0, + 1059.0, + 485.0, + 1094.0, + 369.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1094.0, + 548.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1086.0, + 742.0, + 1086.0, + 742.0, + 1122.0, + 293.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1086.0, + 1270.0, + 1086.0, + 1270.0, + 1122.0, + 762.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1086.0, + 1405.0, + 1086.0, + 1405.0, + 1122.0, + 1290.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1119.0, + 1405.0, + 1119.0, + 1405.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1149.0, + 1087.0, + 1149.0, + 1087.0, + 1185.0, + 293.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1149.0, + 1219.0, + 1149.0, + 1219.0, + 1185.0, + 1107.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1149.0, + 1406.0, + 1149.0, + 1406.0, + 1185.0, + 1282.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1179.0, + 476.0, + 1179.0, + 476.0, + 1214.0, + 293.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1179.0, + 1406.0, + 1179.0, + 1406.0, + 1214.0, + 496.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 1271.0, + 1211.0, + 1271.0, + 1243.0, + 294.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1211.0, + 1405.0, + 1211.0, + 1405.0, + 1243.0, + 1291.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1238.0, + 707.0, + 1238.0, + 707.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1275.0, + 726.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1271.0, + 1406.0, + 1271.0, + 1406.0, + 1305.0, + 293.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1301.0, + 1244.0, + 1301.0, + 1244.0, + 1335.0, + 292.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1791.0, + 1403.0, + 1791.0, + 1403.0, + 1821.0, + 296.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1822.0, + 1403.0, + 1822.0, + 1403.0, + 1852.0, + 294.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1403.0, + 1850.0, + 1403.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1913.0, + 1403.0, + 1913.0, + 1403.0, + 1943.0, + 296.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1976.0, + 292.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1325.0, + 2001.0, + 1325.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 690.0, + 1404.0, + 690.0, + 1404.0, + 724.0, + 366.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 707.0, + 613.0, + 707.0, + 613.0, + 777.0, + 381.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 707.0, + 664.0, + 707.0, + 664.0, + 777.0, + 651.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 707.0, + 917.0, + 707.0, + 917.0, + 777.0, + 894.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 707.0, + 1217.0, + 707.0, + 1217.0, + 777.0, + 1204.0, + 777.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.824 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 488, + 227, + 488, + 261, + 300, + 261 + ], + "score": 0.784 + }, + { + "category_id": 1, + "poly": [ + 290, + 279, + 1402, + 279, + 1402, + 342, + 290, + 342 + ], + "score": 0.778 + }, + { + "category_id": 1, + "poly": [ + 293, + 1409, + 1401, + 1409, + 1401, + 1475, + 293, + 1475 + ], + "score": 0.757 + }, + { + "category_id": 1, + "poly": [ + 287, + 1328, + 1403, + 1328, + 1403, + 1394, + 287, + 1394 + ], + "score": 0.746 + }, + { + "category_id": 1, + "poly": [ + 291, + 1022, + 1404, + 1022, + 1404, + 1087, + 291, + 1087 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 297, + 747, + 1401, + 747, + 1401, + 812, + 297, + 812 + ], + "score": 0.733 + }, + { + "category_id": 1, + "poly": [ + 299, + 1491, + 1401, + 1491, + 1401, + 1586, + 299, + 1586 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 294, + 664, + 1404, + 664, + 1404, + 730, + 294, + 730 + ], + "score": 0.724 + }, + { + "category_id": 1, + "poly": [ + 300, + 1215, + 1399, + 1215, + 1399, + 1310, + 300, + 1310 + ], + "score": 0.72 + }, + { + "category_id": 1, + "poly": [ + 296, + 470, + 1401, + 470, + 1401, + 565, + 296, + 565 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 300, + 1103, + 1401, + 1103, + 1401, + 1199, + 300, + 1199 + ], + "score": 0.712 + }, + { + "category_id": 1, + "poly": [ + 292, + 583, + 1403, + 583, + 1403, + 648, + 292, + 648 + ], + "score": 0.709 + }, + { + "category_id": 1, + "poly": [ + 297, + 1716, + 1401, + 1716, + 1401, + 1811, + 297, + 1811 + ], + "score": 0.709 + }, + { + "category_id": 1, + "poly": [ + 296, + 828, + 1402, + 828, + 1402, + 894, + 296, + 894 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 297, + 909, + 1402, + 909, + 1402, + 1004, + 297, + 1004 + ], + "score": 0.705 + }, + { + "category_id": 1, + "poly": [ + 299, + 1603, + 1402, + 1603, + 1402, + 1699, + 299, + 1699 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 297, + 359, + 1400, + 359, + 1400, + 453, + 297, + 453 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 295, + 1827, + 1401, + 1827, + 1401, + 1922, + 295, + 1922 + ], + "score": 0.654 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1403, + 1941, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.645 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 490.0, + 227.0, + 490.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 277.0, + 1404.0, + 277.0, + 1404.0, + 313.0, + 293.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 308.0, + 1086.0, + 308.0, + 1086.0, + 342.0, + 322.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1410.0, + 1402.0, + 1410.0, + 1402.0, + 1446.0, + 296.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1442.0, + 774.0, + 1442.0, + 774.0, + 1474.0, + 323.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1326.0, + 965.0, + 1326.0, + 965.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1327.0, + 1278.0, + 1327.0, + 1278.0, + 1364.0, + 1006.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1327.0, + 1407.0, + 1327.0, + 1407.0, + 1366.0, + 1312.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1358.0, + 898.0, + 1358.0, + 898.0, + 1395.0, + 322.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1018.0, + 1405.0, + 1018.0, + 1405.0, + 1062.0, + 292.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1054.0, + 609.0, + 1054.0, + 609.0, + 1087.0, + 323.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 747.0, + 1405.0, + 747.0, + 1405.0, + 783.0, + 295.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 777.0, + 1282.0, + 777.0, + 1282.0, + 812.0, + 322.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1492.0, + 1405.0, + 1492.0, + 1405.0, + 1526.0, + 295.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1519.0, + 1405.0, + 1519.0, + 1405.0, + 1562.0, + 320.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1554.0, + 735.0, + 1554.0, + 735.0, + 1588.0, + 325.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 663.0, + 1404.0, + 663.0, + 1404.0, + 703.0, + 294.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 697.0, + 883.0, + 697.0, + 883.0, + 731.0, + 323.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1214.0, + 1404.0, + 1214.0, + 1404.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1283.0, + 321.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1275.0, + 399.0, + 1275.0, + 399.0, + 1311.0, + 320.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 470.0, + 1402.0, + 470.0, + 1402.0, + 508.0, + 294.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 503.0, + 1405.0, + 503.0, + 1405.0, + 539.0, + 321.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 532.0, + 608.0, + 532.0, + 608.0, + 566.0, + 320.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1102.0, + 1406.0, + 1102.0, + 1406.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1136.0, + 1405.0, + 1136.0, + 1405.0, + 1171.0, + 324.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1167.0, + 715.0, + 1167.0, + 715.0, + 1200.0, + 324.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 582.0, + 1404.0, + 582.0, + 1404.0, + 621.0, + 294.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 614.0, + 1039.0, + 614.0, + 1039.0, + 649.0, + 323.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1751.0, + 295.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1749.0, + 1405.0, + 1749.0, + 1405.0, + 1783.0, + 322.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1774.0, + 467.0, + 1774.0, + 467.0, + 1812.0, + 321.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 1406.0, + 827.0, + 1406.0, + 865.0, + 294.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 860.0, + 1335.0, + 860.0, + 1335.0, + 896.0, + 324.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 905.0, + 1403.0, + 905.0, + 1403.0, + 952.0, + 292.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 937.0, + 1406.0, + 937.0, + 1406.0, + 982.0, + 321.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 972.0, + 504.0, + 972.0, + 504.0, + 1004.0, + 325.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 1406.0, + 1601.0, + 1406.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1635.0, + 1404.0, + 1635.0, + 1404.0, + 1671.0, + 323.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1666.0, + 718.0, + 1666.0, + 718.0, + 1700.0, + 324.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 357.0, + 1404.0, + 357.0, + 1404.0, + 395.0, + 295.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 391.0, + 1404.0, + 391.0, + 1404.0, + 425.0, + 323.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 421.0, + 698.0, + 421.0, + 698.0, + 454.0, + 323.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1827.0, + 1405.0, + 1827.0, + 1405.0, + 1865.0, + 294.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1855.0, + 1406.0, + 1855.0, + 1406.0, + 1901.0, + 319.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1890.0, + 506.0, + 1890.0, + 506.0, + 1922.0, + 324.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1975.0, + 296.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1968.0, + 1405.0, + 1968.0, + 1405.0, + 2010.0, + 320.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2000.0, + 1383.0, + 2000.0, + 1383.0, + 2038.0, + 321.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 301, + 229, + 1401, + 229, + 1401, + 323, + 301, + 323 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 296, + 1657, + 1401, + 1657, + 1401, + 1752, + 296, + 1752 + ], + "score": 0.814 + }, + { + "category_id": 1, + "poly": [ + 294, + 1490, + 1401, + 1490, + 1401, + 1555, + 294, + 1555 + ], + "score": 0.756 + }, + { + "category_id": 1, + "poly": [ + 292, + 1407, + 1401, + 1407, + 1401, + 1473, + 292, + 1473 + ], + "score": 0.753 + }, + { + "category_id": 1, + "poly": [ + 293, + 1574, + 1398, + 1574, + 1398, + 1639, + 293, + 1639 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 290, + 1324, + 1403, + 1324, + 1403, + 1389, + 290, + 1389 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 299, + 902, + 1404, + 902, + 1404, + 996, + 299, + 996 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 298, + 1128, + 1399, + 1128, + 1399, + 1194, + 298, + 1194 + ], + "score": 0.709 + }, + { + "category_id": 1, + "poly": [ + 295, + 343, + 1401, + 343, + 1401, + 406, + 295, + 406 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 297, + 1014, + 1405, + 1014, + 1405, + 1108, + 297, + 1108 + ], + "score": 0.703 + }, + { + "category_id": 1, + "poly": [ + 294, + 1209, + 1401, + 1209, + 1401, + 1305, + 294, + 1305 + ], + "score": 0.683 + }, + { + "category_id": 1, + "poly": [ + 300, + 704, + 1398, + 704, + 1398, + 770, + 300, + 770 + ], + "score": 0.664 + }, + { + "category_id": 1, + "poly": [ + 300, + 621, + 1398, + 621, + 1398, + 686, + 300, + 686 + ], + "score": 0.65 + }, + { + "category_id": 1, + "poly": [ + 303, + 508, + 1400, + 508, + 1400, + 603, + 303, + 603 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 296, + 425, + 1401, + 425, + 1401, + 490, + 296, + 490 + ], + "score": 0.628 + }, + { + "category_id": 1, + "poly": [ + 297, + 787, + 1403, + 787, + 1403, + 884, + 297, + 884 + ], + "score": 0.622 + }, + { + "category_id": 13, + "poly": [ + 987, + 969, + 1004, + 969, + 1004, + 989, + 987, + 989 + ], + "score": 0.4, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 1402.0, + 230.0, + 1402.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 295.0, + 322.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 288.0, + 400.0, + 288.0, + 400.0, + 324.0, + 318.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1695.0, + 293.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1685.0, + 1406.0, + 1685.0, + 1406.0, + 1729.0, + 319.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1719.0, + 532.0, + 1719.0, + 532.0, + 1751.0, + 325.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1486.0, + 1405.0, + 1486.0, + 1405.0, + 1529.0, + 293.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1516.0, + 400.0, + 1516.0, + 400.0, + 1558.0, + 319.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1410.0, + 1402.0, + 1410.0, + 1402.0, + 1443.0, + 296.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1437.0, + 1272.0, + 1437.0, + 1272.0, + 1474.0, + 324.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 1402.0, + 1573.0, + 1402.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1604.0, + 1156.0, + 1604.0, + 1156.0, + 1639.0, + 323.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1323.0, + 1404.0, + 1323.0, + 1404.0, + 1362.0, + 293.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1355.0, + 1205.0, + 1355.0, + 1205.0, + 1390.0, + 321.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 903.0, + 1403.0, + 903.0, + 1403.0, + 937.0, + 295.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 932.0, + 1405.0, + 932.0, + 1405.0, + 970.0, + 322.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 960.0, + 986.0, + 960.0, + 986.0, + 1000.0, + 322.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 960.0, + 1216.0, + 960.0, + 1216.0, + 1000.0, + 1005.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1129.0, + 1403.0, + 1129.0, + 1403.0, + 1165.0, + 296.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1160.0, + 1381.0, + 1160.0, + 1381.0, + 1196.0, + 324.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 341.0, + 1405.0, + 341.0, + 1405.0, + 380.0, + 293.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 375.0, + 830.0, + 375.0, + 830.0, + 406.0, + 322.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1053.0, + 294.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1046.0, + 1405.0, + 1046.0, + 1405.0, + 1080.0, + 323.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1071.0, + 399.0, + 1071.0, + 399.0, + 1112.0, + 320.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1207.0, + 1406.0, + 1207.0, + 1406.0, + 1250.0, + 293.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1242.0, + 1404.0, + 1242.0, + 1404.0, + 1279.0, + 319.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1273.0, + 468.0, + 1273.0, + 468.0, + 1305.0, + 322.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1403.0, + 702.0, + 1403.0, + 744.0, + 294.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 737.0, + 1261.0, + 737.0, + 1261.0, + 771.0, + 321.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 617.0, + 1404.0, + 617.0, + 1404.0, + 661.0, + 295.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 652.0, + 1062.0, + 652.0, + 1062.0, + 687.0, + 321.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 508.0, + 1405.0, + 508.0, + 1405.0, + 546.0, + 297.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 538.0, + 1405.0, + 538.0, + 1405.0, + 577.0, + 320.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 573.0, + 987.0, + 573.0, + 987.0, + 604.0, + 322.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 427.0, + 1403.0, + 427.0, + 1403.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 458.0, + 1209.0, + 458.0, + 1209.0, + 491.0, + 321.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 789.0, + 1403.0, + 789.0, + 1403.0, + 823.0, + 295.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 817.0, + 1402.0, + 817.0, + 1402.0, + 858.0, + 322.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 850.0, + 959.0, + 850.0, + 959.0, + 883.0, + 322.0, + 883.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1331, + 1405, + 1331, + 1405, + 1519, + 296, + 1519 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 351, + 1404, + 351, + 1404, + 452, + 296, + 452 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 716, + 1533, + 981, + 1533, + 981, + 1593, + 716, + 1593 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 695, + 1685, + 1002, + 1685, + 1002, + 1758, + 695, + 1758 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 669, + 1248, + 1031, + 1248, + 1031, + 1317, + 669, + 1317 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 607, + 470, + 1091, + 470, + 1091, + 560, + 607, + 560 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 631, + 1877, + 1066, + 1877, + 1066, + 1953, + 631, + 1953 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 712, + 860, + 984, + 860, + 984, + 929, + 712, + 929 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 296, + 1969, + 1402, + 1969, + 1402, + 2037, + 296, + 2037 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 1786, + 1405, + 1786, + 1405, + 1860, + 298, + 1860 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 296, + 1606, + 1405, + 1606, + 1405, + 1670, + 296, + 1670 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 722, + 988, + 976, + 988, + 976, + 1059, + 722, + 1059 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 452, + 1136, + 1244, + 1136, + 1244, + 1186, + 452, + 1186 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 299, + 942, + 478, + 942, + 478, + 974, + 299, + 974 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 1198, + 492, + 1198, + 492, + 1231, + 297, + 1231 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 298, + 1087, + 910, + 1087, + 910, + 1121, + 298, + 1121 + ], + "score": 0.927 + }, + { + "category_id": 8, + "poly": [ + 401, + 636, + 1225, + 636, + 1225, + 711, + 401, + 711 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 300, + 810, + 774, + 810, + 774, + 844, + 300, + 844 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 297, + 593, + 369, + 593, + 369, + 623, + 297, + 623 + ], + "score": 0.924 + }, + { + "category_id": 8, + "poly": [ + 397, + 717, + 1301, + 717, + 1301, + 779, + 397, + 779 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1546, + 1400, + 1546, + 1400, + 1578, + 1351, + 1578 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1268, + 1400, + 1268, + 1400, + 1300, + 1351, + 1300 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1707, + 1400, + 1707, + 1400, + 1739, + 1351, + 1739 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1010, + 1401, + 1010, + 1401, + 1042, + 1351, + 1042 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1147, + 1400, + 1147, + 1400, + 1179, + 1352, + 1179 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1351, + 881, + 1401, + 881, + 1401, + 913, + 1351, + 913 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1900, + 1400, + 1900, + 1400, + 1931, + 1351, + 1931 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1351, + 498, + 1401, + 498, + 1401, + 530, + 1351, + 530 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1351, + 658, + 1400, + 658, + 1400, + 690, + 1351, + 690 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1352, + 735, + 1401, + 735, + 1401, + 767, + 1352, + 767 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.87 + }, + { + "category_id": 1, + "poly": [ + 300, + 295, + 640, + 295, + 640, + 329, + 300, + 329 + ], + "score": 0.868 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 509, + 226, + 509, + 262, + 300, + 262 + ], + "score": 0.839 + }, + { + "category_id": 14, + "poly": [ + 722, + 988, + 977, + 988, + 977, + 1057, + 722, + 1057 + ], + "score": 0.95, + "latex": "\\mathcal { E } \\sim \\mathcal { N } ( 0 , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) ." + }, + { + "category_id": 14, + "poly": [ + 632, + 1877, + 1068, + 1877, + 1068, + 1953, + 632, + 1953 + ], + "score": 0.95, + "latex": "E { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 1 } { 2 } } { \\sqrt { \\frac { \\pi } { M } } } , D { \\bigl ( } | { \\mathcal { E } } | { \\bigr ) } = { \\frac { 4 - \\pi } { 4 M } } ." + }, + { + "category_id": 14, + "poly": [ + 696, + 1684, + 1004, + 1684, + 1004, + 1755, + 696, + 1755 + ], + "score": 0.94, + "latex": "s ^ { 2 } = { \\frac { 1 } { 2 } } D ( { \\mathcal { E } } ) = { \\frac { 1 - | a _ { n } | ^ { 2 } } { 2 M } } ." + }, + { + "category_id": 14, + "poly": [ + 712, + 859, + 987, + 859, + 987, + 928, + 712, + 928 + ], + "score": 0.94, + "latex": "\\mathcal { V } \\sim \\mathcal { N } ( a _ { n } , \\frac { 1 - | a _ { n } | ^ { 2 } } { M } ) ." + }, + { + "category_id": 14, + "poly": [ + 608, + 467, + 1090, + 467, + 1090, + 561, + 608, + 561 + ], + "score": 0.94, + "latex": "\\mathcal { V } = \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { T } _ { m } \\sim \\mathcal { N } \\big ( E ( \\mathcal { T } _ { 1 } ) , \\frac { 1 } { M } D ( \\mathcal { T } _ { 1 } ) \\big ) ," + }, + { + "category_id": 14, + "poly": [ + 716, + 1534, + 982, + 1534, + 982, + 1593, + 716, + 1593 + ], + "score": 0.94, + "latex": "\\mathcal { P } _ { | \\varepsilon | } \\left( r \\right) = \\frac { r } { s ^ { 2 } } e ^ { - r ^ { 2 } / 2 s ^ { 2 } } ," + }, + { + "category_id": 14, + "poly": [ + 669, + 1247, + 1032, + 1247, + 1032, + 1315, + 669, + 1315 + ], + "score": 0.94, + "latex": "\\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = \\frac { 1 } { M } \\mathcal { I } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = \\frac { a _ { 2 n } - a _ { n } ^ { 2 } } { M } ." + }, + { + "category_id": 13, + "poly": [ + 417, + 1824, + 520, + 1824, + 520, + 1862, + 417, + 1862 + ], + "score": 0.93, + "latex": "\\textstyle s ^ { 2 } = { \\frac { 1 } { 2 M } }" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1786, + 1392, + 1786, + 1392, + 1827, + 1251, + 1827 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { D \\left( \\mathcal { E } \\right) \\approx \\frac { 1 } { M } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 397, + 634, + 1303, + 634, + 1303, + 784, + 397, + 784 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { \\displaystyle E ( \\mathcal { T } _ { 1 } ) = E \\left( e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right) = \\int _ { - \\infty } ^ { + \\infty } \\mathcal { P } _ { \\mathcal { X } _ { 1 } } ( x ) \\cdot e ^ { - j 2 \\pi n \\log _ { 1 0 } | x | } d x = a _ { n } , } \\\\ { \\displaystyle D ( \\mathcal { T } _ { 1 } ) = E \\big ( \\left| \\mathcal { T } _ { 1 } \\right| ^ { 2 } \\big ) - \\left| E \\left( \\mathcal { T } _ { 1 } \\right) \\right| ^ { 2 } = E \\big ( \\left| e ^ { - j 2 \\pi n \\log _ { 1 0 } | \\mathcal { X } _ { 1 } | } \\right| ^ { 2 } \\big ) - \\left| a _ { n } \\right| ^ { 2 } = 1 - \\left| a _ { n } \\right| ^ { 2 } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 772, + 1974, + 862, + 1974, + 862, + 2009, + 772, + 2009 + ], + "score": 0.93, + "latex": "D ( \\left. \\boldsymbol { \\mathcal { E } } \\right. )" + }, + { + "category_id": 13, + "poly": [ + 734, + 358, + 903, + 358, + 903, + 391, + 734, + 391 + ], + "score": 0.92, + "latex": "\\{ \\mathcal { T } _ { m } \\} _ { m = 1 , \\dots , M }" + }, + { + "category_id": 13, + "poly": [ + 547, + 1788, + 607, + 1788, + 607, + 1823, + 547, + 1823 + ], + "score": 0.92, + "latex": "\\left| a _ { n } \\right| ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 628, + 1974, + 717, + 1974, + 717, + 2009, + 628, + 2009 + ], + "score": 0.91, + "latex": "E { \\big ( } | { \\mathcal { E } } | { \\big ) }" + }, + { + "category_id": 14, + "poly": [ + 456, + 1137, + 1243, + 1137, + 1243, + 1185, + 456, + 1185 + ], + "score": 0.91, + "latex": "\\mathcal { T } _ { \\mathcal { T } _ { 1 } , \\mathcal { T } _ { 1 } } = E ( \\mathcal { T } _ { 1 } ^ { 2 } ) - E ( \\mathcal { T } _ { 1 } ) ^ { 2 } = E \\big ( e ^ { - j 2 \\pi n \\log _ { 1 0 } \\left| \\mathcal { X } _ { 1 } ^ { 2 } \\right| } \\big ) - a _ { n } ^ { 2 } = a _ { 2 n } - a _ { n } ^ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 365, + 1332, + 483, + 1332, + 483, + 1367, + 365, + 1367 + ], + "score": 0.91, + "latex": "\\left( a _ { 2 n } - a _ { n } ^ { 2 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 884, + 1488, + 918, + 1488, + 918, + 1519, + 884, + 1519 + ], + "score": 0.89, + "latex": "| \\mathcal { E } |" + }, + { + "category_id": 13, + "poly": [ + 805, + 2015, + 834, + 2015, + 834, + 2035, + 805, + 2035 + ], + "score": 0.88, + "latex": "\\varepsilon _ { n }" + }, + { + "category_id": 13, + "poly": [ + 714, + 1334, + 932, + 1334, + 932, + 1368, + 714, + 1368 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { \\operatorname* { l i m } _ { M \\to \\infty } \\mathcal { I } _ { \\mathcal { E } , \\mathcal { E } } = 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 853, + 1092, + 880, + 1092, + 880, + 1120, + 853, + 1120 + ], + "score": 0.85, + "latex": "\\mathcal { T } _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1285, + 395, + 1306, + 395, + 1306, + 418, + 1285, + 418 + ], + "score": 0.78, + "latex": "_ { \\mathcal { V } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1370, + 316, + 1370, + 316, + 1392, + 298, + 1392 + ], + "score": 0.72, + "latex": "\\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 648, + 1427, + 667, + 1427, + 667, + 1453, + 648, + 1453 + ], + "score": 0.71, + "latex": "\\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 435, + 813, + 460, + 813, + 460, + 842, + 435, + 842 + ], + "score": 0.66, + "latex": "\\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 994, + 2007, + 1010, + 2007, + 1010, + 2029, + 994, + 2029 + ], + "score": 0.61, + "latex": "O" + }, + { + "category_id": 13, + "poly": [ + 1045, + 2004, + 1073, + 2004, + 1073, + 2030, + 1045, + 2030 + ], + "score": 0.27, + "latex": "M" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 220.0, + 514.0, + 220.0, + 514.0, + 271.0, + 294.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1331.0, + 364.0, + 1331.0, + 364.0, + 1371.0, + 293.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1331.0, + 713.0, + 1331.0, + 713.0, + 1371.0, + 484.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1331.0, + 1406.0, + 1331.0, + 1406.0, + 1371.0, + 933.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 297.0, + 1359.0, + 297.0, + 1403.0, + 293.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1359.0, + 1407.0, + 1359.0, + 1407.0, + 1403.0, + 317.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1395.0, + 1403.0, + 1395.0, + 1403.0, + 1428.0, + 295.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1422.0, + 647.0, + 1422.0, + 647.0, + 1460.0, + 293.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1422.0, + 1403.0, + 1422.0, + 1403.0, + 1460.0, + 668.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1491.0, + 294.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1485.0, + 883.0, + 1485.0, + 883.0, + 1521.0, + 294.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1485.0, + 1163.0, + 1485.0, + 1163.0, + 1521.0, + 919.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 350.0, + 733.0, + 350.0, + 733.0, + 398.0, + 290.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 350.0, + 1405.0, + 350.0, + 1405.0, + 398.0, + 904.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 385.0, + 1284.0, + 385.0, + 1284.0, + 427.0, + 291.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 385.0, + 1405.0, + 385.0, + 1405.0, + 427.0, + 1307.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 420.0, + 1078.0, + 420.0, + 1078.0, + 452.0, + 295.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1969.0, + 627.0, + 1969.0, + 627.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1969.0, + 771.0, + 1969.0, + 771.0, + 2006.0, + 718.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1969.0, + 1403.0, + 1969.0, + 1403.0, + 2006.0, + 863.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 804.0, + 2002.0, + 804.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 2002.0, + 993.0, + 2002.0, + 993.0, + 2038.0, + 835.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 2002.0, + 1044.0, + 2002.0, + 1044.0, + 2038.0, + 1011.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 2002.0, + 1192.0, + 2002.0, + 1192.0, + 2038.0, + 1074.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1775.0, + 546.0, + 1775.0, + 546.0, + 1834.0, + 292.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1775.0, + 1250.0, + 1775.0, + 1250.0, + 1834.0, + 608.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1775.0, + 1406.0, + 1775.0, + 1406.0, + 1834.0, + 1393.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1816.0, + 416.0, + 1816.0, + 416.0, + 1871.0, + 291.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1816.0, + 723.0, + 1816.0, + 723.0, + 1871.0, + 521.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1406.0, + 1605.0, + 1406.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1634.0, + 492.0, + 1634.0, + 492.0, + 1674.0, + 293.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 938.0, + 482.0, + 938.0, + 482.0, + 977.0, + 295.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1191.0, + 496.0, + 1191.0, + 496.0, + 1239.0, + 295.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1087.0, + 852.0, + 1087.0, + 852.0, + 1124.0, + 296.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1087.0, + 911.0, + 1087.0, + 911.0, + 1124.0, + 881.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 807.0, + 434.0, + 807.0, + 434.0, + 848.0, + 295.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 807.0, + 776.0, + 807.0, + 776.0, + 848.0, + 461.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 590.0, + 375.0, + 590.0, + 375.0, + 627.0, + 293.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 295.0, + 643.0, + 295.0, + 643.0, + 332.0, + 296.0, + 332.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 381, + 342, + 1318, + 342, + 1318, + 602, + 381, + 602 + ], + "score": 0.981, + "html": "
cleannoisyBIMCW-L2DeepFoolR-PGD
CIFAR-10train set0.0740.0740.1580.1390.1550.159
test set0.0720.0720.1580.1380.2160.216
SVHNtrain set0.1040.1040.2490.1940.2220.245
test set0.1060.1060.2490.1930.2190.239
ImageNet-AlexNettrain set0.1160.1160.2530.2350.2420.253
test set0.1170.1170.2610.2360.2430.259
ImageNet-VGG-16train set0.1140.1140.2360.2280.2400.235
test set0.1140.1140.2310.2280.2390.232
" + }, + { + "category_id": 5, + "poly": [ + 297, + 1064, + 1426, + 1064, + 1426, + 1290, + 297, + 1290 + ], + "score": 0.979, + "html": "
clean noisyBIMCW-L2DeepFoolR-PGD
CIFAR-10(train set, test set)0.120 0.1880.3330.4560.5840.223
SVHN(train set, test set)0.494 0.7080.5230.1470.4910.645
ImageNet- AlexNet(train set, test set) (train set, out-of-sample)0.153 0.148 0.000 0.0000.370 0.1690.605 0.1010.410 0.2060.386 0.158
ImageNet- VGG-16(train set, test set) (train set, out-of-sample)0.839 0.858 0.049 0.0540.368 0.0940.842 0.4070.246 0.7370.221 0.088
" + }, + { + "category_id": 5, + "poly": [ + 310, + 699, + 1388, + 699, + 1388, + 958, + 310, + 958 + ], + "score": 0.977, + "html": "
(clean,noisy)(clean,BIM)(BIM,DeepFool)(CW-L2,R-PGD)
CIFAR-10train set0.6400.0000.0000.003
test set0.7640.0000.0300.000
SVHNtrain set0.5590.0000.0300.000
test set0.6850.0000.0000.000
ImageNet-AlexNettrain set1.0000.0000.0020.134
test set0.9930.0000.4770.000
ImageNet-VGG-16train set0.9980.0000.0000.236
test set1.0000.0000.0000.225
" + }, + { + "category_id": 3, + "poly": [ + 378, + 1711, + 1314, + 1711, + 1314, + 1963, + 378, + 1963 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 377, + 1344, + 1314, + 1344, + 1314, + 1596, + 377, + 1596 + ], + "score": 0.968 + }, + { + "category_id": 6, + "poly": [ + 298, + 274, + 1400, + 274, + 1400, + 338, + 298, + 338 + ], + "score": 0.933 + }, + { + "category_id": 4, + "poly": [ + 292, + 1972, + 1404, + 1972, + 1404, + 2037, + 292, + 2037 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 291, + 1606, + 1402, + 1606, + 1402, + 1669, + 291, + 1669 + ], + "score": 0.911 + }, + { + "category_id": 6, + "poly": [ + 369, + 1025, + 1326, + 1025, + 1326, + 1059, + 369, + 1059 + ], + "score": 0.91 + }, + { + "category_id": 6, + "poly": [ + 299, + 660, + 1397, + 660, + 1397, + 694, + 299, + 694 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.868 + }, + { + "category_id": 13, + "poly": [ + 452, + 279, + 471, + 279, + 471, + 306, + 452, + 306 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 531, + 1032, + 549, + 1032, + 549, + 1059, + 531, + 1059 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 454, + 667, + 472, + 667, + 472, + 694, + 454, + 694 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 609, + 1976, + 639, + 1976, + 639, + 2003, + 609, + 2003 + ], + "score": 0.71, + "latex": "\\pm" + }, + { + "category_id": 13, + "poly": [ + 610, + 1608, + 639, + 1608, + 639, + 1636, + 610, + 1636 + ], + "score": 0.71, + "latex": "\\pm" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1701.0, + 1321.0, + 1701.0, + 1321.0, + 1828.0, + 688.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1817.0, + 1313.0, + 1817.0, + 1313.0, + 1937.0, + 687.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1933.0, + 500.0, + 1933.0, + 500.0, + 1966.0, + 408.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1932.0, + 658.0, + 1932.0, + 658.0, + 1968.0, + 562.0, + 1968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1933.0, + 812.0, + 1933.0, + 812.0, + 1962.0, + 722.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1931.0, + 984.0, + 1931.0, + 984.0, + 1966.0, + 868.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1932.0, + 1151.0, + 1932.0, + 1151.0, + 1967.0, + 1014.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1932.0, + 1294.0, + 1932.0, + 1294.0, + 1963.0, + 1184.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1337.0, + 1151.0, + 1337.0, + 1151.0, + 1465.0, + 688.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1454.0, + 1119.0, + 1454.0, + 1119.0, + 1564.0, + 696.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1566.0, + 501.0, + 1566.0, + 501.0, + 1599.0, + 408.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1565.0, + 659.0, + 1565.0, + 659.0, + 1601.0, + 562.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1566.0, + 811.0, + 1566.0, + 811.0, + 1595.0, + 722.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1566.0, + 983.0, + 1566.0, + 983.0, + 1595.0, + 868.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1565.0, + 1150.0, + 1565.0, + 1150.0, + 1600.0, + 1014.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1565.0, + 1294.0, + 1565.0, + 1294.0, + 1596.0, + 1185.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.75, + 1487.0, + 1282.75, + 1487.0, + 1282.75, + 1532.0, + 1194.75, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 271.0, + 451.0, + 271.0, + 451.0, + 311.0, + 294.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 271.0, + 1405.0, + 271.0, + 1405.0, + 311.0, + 472.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 303.0, + 818.0, + 303.0, + 818.0, + 343.0, + 294.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 608.0, + 1968.0, + 608.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2010.0, + 640.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 608.0, + 2003.0, + 608.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1602.0, + 609.0, + 1602.0, + 609.0, + 1643.0, + 291.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1602.0, + 1407.0, + 1602.0, + 1407.0, + 1643.0, + 640.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 647.0, + 1636.0, + 647.0, + 1668.0, + 295.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1024.0, + 530.0, + 1024.0, + 530.0, + 1062.0, + 366.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1024.0, + 1332.0, + 1024.0, + 1332.0, + 1062.0, + 550.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 659.0, + 453.0, + 659.0, + 453.0, + 698.0, + 295.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 659.0, + 1402.0, + 659.0, + 1402.0, + 698.0, + 473.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1235, + 1404, + 1235, + 1404, + 1542, + 298, + 1542 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1636, + 1405, + 1636, + 1405, + 1821, + 298, + 1821 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1018, + 1403, + 1018, + 1403, + 1141, + 299, + 1141 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 397, + 755, + 1301, + 755, + 1301, + 906, + 397, + 906 + ], + "score": 0.976, + "html": "
Iteration numberin each GD runNumber of GD runsBIMCW-L2DeepFoolR-PGD
200100.7440.9800.7020.746
100100.6890.9730.6500.690
10050.6640.9690.6280.666
10020.6120.9590.5860.614
MBF1.0000.9911.0001.000
" + }, + { + "category_id": 3, + "poly": [ + 379, + 232, + 1315, + 232, + 1315, + 592, + 379, + 592 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 296, + 604, + 1403, + 604, + 1403, + 668, + 296, + 668 + ], + "score": 0.928 + }, + { + "category_id": 6, + "poly": [ + 294, + 689, + 1400, + 689, + 1400, + 751, + 294, + 751 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 758, + 2006, + 758, + 2033, + 333, + 2033 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 301, + 961, + 798, + 961, + 798, + 993, + 301, + 993 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 298, + 1178, + 849, + 1178, + 849, + 1210, + 298, + 1210 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.854 + }, + { + "category_id": 0, + "poly": [ + 299, + 1579, + 981, + 1579, + 981, + 1609, + 299, + 1609 + ], + "score": 0.842 + }, + { + "category_id": 13, + "poly": [ + 630, + 1328, + 706, + 1328, + 706, + 1357, + 630, + 1357 + ], + "score": 0.87, + "latex": "9 8 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 467, + 1053, + 484, + 1053, + 484, + 1081, + 467, + 1081 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 496, + 1790, + 563, + 1790, + 563, + 1819, + 496, + 1819 + ], + "score": 0.78, + "latex": "{ \\mathrm { \\Omega } } { = } 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 678, + 1791, + 724, + 1791, + 724, + 1818, + 678, + 1818 + ], + "score": 0.75, + "latex": "\\scriptstyle = 4 0" + }, + { + "category_id": 13, + "poly": [ + 794, + 1729, + 861, + 1729, + 861, + 1758, + 794, + 1758 + ], + "score": 0.71, + "latex": "= 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 888, + 1760, + 963, + 1760, + 963, + 1788, + 888, + 1788 + ], + "score": 0.7, + "latex": "\\mathord { \\left. \\begin{array} { r l l } { - 1 0 0 0 } \\end{array} \\right. }" + }, + { + "category_id": 13, + "poly": [ + 920, + 691, + 962, + 691, + 962, + 720, + 920, + 720 + ], + "score": 0.68, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 418, + 1760, + 471, + 1760, + 471, + 1788, + 418, + 1788 + ], + "score": 0.68, + "latex": "= 0 . 0" + }, + { + "category_id": 13, + "poly": [ + 1225, + 1760, + 1290, + 1760, + 1290, + 1788, + 1225, + 1788 + ], + "score": 0.66, + "latex": "\\mathord { \\mathrm { \\ s } } = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 1362, + 1730, + 1393, + 1730, + 1393, + 1757, + 1362, + 1757 + ], + "score": 0.66, + "latex": "^ { = 5 }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1730, + 694, + 1730, + 694, + 1759, + 601, + 1759 + ], + "score": 0.66, + "latex": "\\mathrm { ( e p s { = } 0 . 3 }" + }, + { + "category_id": 13, + "poly": [ + 610, + 607, + 639, + 607, + 639, + 634, + 610, + 634 + ], + "score": 0.64, + "latex": "\\pm" + }, + { + "category_id": 13, + "poly": [ + 977, + 1730, + 1023, + 1730, + 1023, + 1757, + 977, + 1757 + ], + "score": 0.62, + "latex": "{ \\it \\Omega } = 1 0" + }, + { + "category_id": 13, + "poly": [ + 630, + 1760, + 711, + 1760, + 711, + 1788, + 630, + 1788 + ], + "score": 0.61, + "latex": "= 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 306, + 1790, + 398, + 1790, + 398, + 1822, + 306, + 1822 + ], + "score": 0.38, + "latex": "\\mathrm { \\ e p s { = } 0 . 3 }" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 229.0, + 620.0, + 229.0, + 620.0, + 316.0, + 456.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 233.0, + 1312.0, + 233.0, + 1312.0, + 338.0, + 683.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 347.0, + 605.0, + 347.0, + 605.0, + 425.0, + 477.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 331.0, + 1324.0, + 331.0, + 1324.0, + 463.0, + 667.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 445.0, + 999.0, + 445.0, + 999.0, + 559.0, + 383.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 452.0, + 1313.0, + 452.0, + 1313.0, + 565.0, + 1000.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 563.0, + 500.0, + 563.0, + 500.0, + 596.0, + 405.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 560.0, + 658.0, + 560.0, + 658.0, + 601.0, + 562.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 563.0, + 812.0, + 563.0, + 812.0, + 595.0, + 721.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 562.0, + 983.0, + 562.0, + 983.0, + 596.0, + 870.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 561.0, + 1152.0, + 561.0, + 1152.0, + 598.0, + 1015.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 562.0, + 1294.0, + 562.0, + 1294.0, + 596.0, + 1182.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 600.0, + 609.0, + 600.0, + 609.0, + 641.0, + 294.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 600.0, + 1405.0, + 600.0, + 1405.0, + 641.0, + 640.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 633.0, + 1115.0, + 633.0, + 1115.0, + 669.0, + 296.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 685.0, + 919.0, + 685.0, + 919.0, + 726.0, + 292.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 685.0, + 1405.0, + 685.0, + 1405.0, + 726.0, + 963.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 720.0, + 870.0, + 720.0, + 870.0, + 752.0, + 296.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1998.0, + 762.0, + 1998.0, + 762.0, + 2041.0, + 328.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 959.0, + 800.0, + 959.0, + 800.0, + 996.0, + 296.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1178.0, + 851.0, + 1178.0, + 851.0, + 1214.0, + 296.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1579.0, + 987.0, + 1579.0, + 987.0, + 1613.0, + 296.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1233.0, + 1405.0, + 1233.0, + 1405.0, + 1272.0, + 295.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1402.0, + 1268.0, + 1402.0, + 1299.0, + 294.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1331.0, + 293.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1326.0, + 629.0, + 1326.0, + 629.0, + 1361.0, + 293.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1326.0, + 1406.0, + 1326.0, + 1406.0, + 1361.0, + 707.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1394.0, + 293.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1390.0, + 1402.0, + 1390.0, + 1402.0, + 1422.0, + 296.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1444.0, + 1405.0, + 1444.0, + 1405.0, + 1488.0, + 292.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1508.0, + 1166.0, + 1508.0, + 1166.0, + 1546.0, + 293.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 1403.0, + 1636.0, + 1403.0, + 1671.0, + 295.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1667.0, + 1405.0, + 1667.0, + 1405.0, + 1703.0, + 293.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1697.0, + 1405.0, + 1697.0, + 1405.0, + 1735.0, + 292.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1728.0, + 600.0, + 1728.0, + 600.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1728.0, + 793.0, + 1728.0, + 793.0, + 1764.0, + 695.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1728.0, + 976.0, + 1728.0, + 976.0, + 1764.0, + 862.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1728.0, + 1361.0, + 1728.0, + 1361.0, + 1764.0, + 1024.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1728.0, + 1403.0, + 1728.0, + 1403.0, + 1764.0, + 1394.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1758.0, + 417.0, + 1758.0, + 417.0, + 1792.0, + 293.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1758.0, + 629.0, + 1758.0, + 629.0, + 1792.0, + 472.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1758.0, + 887.0, + 1758.0, + 887.0, + 1792.0, + 712.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1758.0, + 1224.0, + 1758.0, + 1224.0, + 1792.0, + 964.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1758.0, + 1405.0, + 1758.0, + 1405.0, + 1792.0, + 1291.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1788.0, + 305.0, + 1788.0, + 305.0, + 1824.0, + 296.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1788.0, + 495.0, + 1788.0, + 495.0, + 1824.0, + 399.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1788.0, + 677.0, + 1788.0, + 677.0, + 1824.0, + 564.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1788.0, + 1373.0, + 1788.0, + 1373.0, + 1824.0, + 725.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1019.0, + 1404.0, + 1019.0, + 1404.0, + 1052.0, + 295.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1047.0, + 466.0, + 1047.0, + 466.0, + 1084.0, + 293.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1047.0, + 1404.0, + 1047.0, + 1404.0, + 1084.0, + 485.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1077.0, + 1407.0, + 1077.0, + 1407.0, + 1114.0, + 293.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1109.0, + 1167.0, + 1109.0, + 1167.0, + 1145.0, + 293.0, + 1145.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ryuxYmvel/ryuxYmvel.md b/parse/train/ryuxYmvel/ryuxYmvel.md new file mode 100644 index 0000000000000000000000000000000000000000..65e8b552e43d471772d22fcb97f616a44e1ebdc3 --- /dev/null +++ b/parse/train/ryuxYmvel/ryuxYmvel.md @@ -0,0 +1,262 @@ +# HOLSTEP: A MACHINE LEARNING DATASET FOR HIGHER-ORDER LOGIC THEOREM PROVING + +Cezary Kaliszyk +University of Innsbruck +cezary.kaliszyk@uibk.ac.at + +François Chollet, Christian Szegedy Google Research {fchollet,szegedy}@google.com + +# ABSTRACT + +Large computer-understandable proofs consist of millions of intermediate logical steps. The vast majority of such steps originate from manually selected and manually guided heuristics applied to intermediate goals. So far, machine learning has generally not been used to filter or generate these steps. In this paper, we introduce a new dataset based on Higher-Order Logic (HOL) proofs, for the purpose of developing new machine learning-based theorem-proving strategies. We make this dataset publicly available under the BSD license. We propose various machine learning tasks that can be performed on this dataset, and discuss their significance for theorem proving. We also benchmark a set of simple baseline machine learning models suited for the tasks (including logistic regression, convolutional neural networks and recurrent neural networks). The results of our baseline models show the promise of applying machine learning to HOL theorem proving. + +# 1 INTRODUCTION + +As the usability of interactive theorem proving (ITP) systems (Harrison et al., 2014) grows, its use becomes a more common way of establishing the correctness of software as well as mathematical proofs. Today, ITPs are used for software certification projects ranging from compilers (Leroy, 2009) and operating system components (Chen et al., 2016; Klein et al., 2014), to establishing the absolute correctness of large proofs in mathematics such as the Kepler conjecture (Hales et al., 2015) and the Feit-Thomson Theorem (Gonthier et al., 2013). + +For results of such significance to be possible, the theorem libraries of these ITPs must contain all necessary basic mathematical properties, accompanied with formal proofs. This means that the size of many ITP libraries can be measured in dozens of thousands of theorems (Grabowski et al., 2010; Blanchette et al., 2015) and billions of individual proof steps. While the general direction of the proofs is specified by humans (by providing the goal to prove, specifying intermediate steps, or applying certain automated tactics), the majority of such proof steps are actually found by automated reasoning-based proof search (Kaliszyk & Urban, 2015b), with very little application of machine learning techniques so far. + +At the same time, fast progress has been unfolding in machine learning applied to tasks that involve logical inference, such as natural language question answering (Sukhbaatar et al., 2015), knowledge base completion (Socher et al., 2013a), automated translation (Wu et al., 2016), and premise selection in the context of theorem proving (Alemi et al., 2016). Deep learning in particular has proven to be a powerful tool for embedding semantic meaning and logical relationships into geometric spaces, specifically via models such as convolutional neural networks, recurrent neural networks, and tree-recursive neural networks. These advances strongly suggest that deep learning may have become mature enough to yield significant advances in automated theorem proving. Remarkably, it has recently become possible to build a system, AlphaGo (Silver et al., 2016), blending classical AI techniques such as Monte-Carlo tree search and modern deep learning techniques, capable of playing the game of Go at super-human levels. We should note that theorem proving and Go playing are conceptually related, since both consist in searching for specific nodes in trees of states with extremely large arity and relatively large depth, which involves node evaluation decision (how valuable is this state?) and policy decisions (which node should be expanded next?). The success of AlphaGo can thus serve as encouragement on the road to building deep learning-augmented theorem provers that would blend classical techniques developed over the past few decades with the latest machine learning advances. + +Fast progress in specific machine learning verticals has occasionally been achieved thanks to the release of specialized datasets (often with associated competitions, e.g. the ImageNet dataset for large-scale image classification (Deng et al., 2009)) serving as an experimental testbed and public benchmark of current progress, thus focusing the efforts of the research community. We hope that releasing a theorem proving dataset suited for specific machine learning tasks can serve the same purpose in the vertical of applying machine learning to theorem proving. + +# 1.1 CONTRIBUTION AND OVERVIEW + +First, we develop a dataset for machine learning based on the proof steps used in a large interactive proof section 2. We focus on the HOL Light (Harrison, 2009) ITP, its multivariate analysis library (Harrison, 2013), as well as the formal proof of the Kepler conjecture (Hales et al., 2010). These formalizations constitute a diverse proof dataset containing basic mathematics, analysis, trigonometry, as well as reasoning about data structures such as graphs. Furthermore these formal proof developments have been used as benchmarks for automated reasoning techniques (Kaliszyk & Urban, 2014). + +The dataset consists of 2,013,046 training examples and 196,030 testing examples that originate from 11,400 proofs. Precisely half of the examples are statements that were useful in the currently proven conjectures and half are steps that have been derived either manually or as part of the automated proof search but were not necessary in the final proofs. The dataset contains only proofs of non-trivial theorems, that also do not focus on computation but rather on actual theorem proving. For each proof, the conjecture that is being proven as well as its dependencies (axioms) and may be exploited in machine learning tasks. Furthermore, for each statement both its human-readable (pretty-printed) statement and a tokenization designed to make machine learning tasks more manageable are included. + +Next, in section 3 we discuss the proof step classification tasks that can be attempted using the dataset, and we discuss the usefulness of these tasks in interactive and automated theorem proving. These tasks include unconditioned classification (without access to conjectures and dependencies) and conjecture-conditioned classification (with access to the conjecture) of proof steps as being useful or not in a proof. We outline the use of such classification capabilities for search space pruning and internal guidance, as well as for generation of intermediate steps or possible new lemma statements. + +Finally, in section 4 we propose three baseline models for the proof step classification tasks, and we experimentally evaluate the models on the data in section 5. The models considered include both a relatively simple regression model, as well as deep learning models based on convolutional and recurrent neural networks. + +# 1.2 RELATED WORK + +The use of machine learning in interactive and automated theorem proving has so far focused on three tasks: premise selection, strategy selection, and internal guidance. We shortly explain these. + +Given a large library of proven facts and a user given conjecture, the multi-label classification problem of selecting the facts that are most likely to lead to a successful proof of the conjecture has been usually called relevance filtering or premise selection (Alama et al., 2014). This is crucial for the efficiency of modern automation techniques for ITPs (Blanchette et al., 2016), which today can usually solve $4 0 { - } 5 0 \%$ of the conjectures in theorem proving libraries. Similarly most competitive ATPs today (Sutcliffe, 2016) implement the SInE classifier (Hoder & Voronkov, 2011). + +A second theorem proving task where machine learning has been of importance is strategy selection. With the development of automated theorem provers came many parameters that control their execution. In fact, modern ATPs, such as E (Schulz, 2013) and Vampire (Kovács & Voronkov, 2013), include complete strategy description languages that allow a user to specify the orderings, weighting functions, literal selection strategies, etc. Rather than optimizing the search strategy globally, one can choose the strategy based on the currently considered problem. For this some frameworks use machine learning (Bridge et al., 2014; Kühlwein & Urban, 2015). + +Finally, an automated theorem prover may use machine learning for choosing the actual inference steps. It has been shown to significantly reduce the proof search in first-order tableaux by the selection of extension steps to use (Urban et al., 2011), and has been also successfully applied in monomorphic higher-order logic proving (Färber & Brown, 2016). Data/proof mining has also been applied on the level of interactive theorem proving tactics (Duncan, 2007) to extract and reuse repeating patterns. + +# 2 DATASET EXTRACTION + +We focus on the HOL Light theorem prover for two reasons. First, it follows the LCF approach1). This means that complicated inferences are reduced to the most primitive ones and the data extraction related modifications can be restricted the primitive inferences and it is relatively easy to extract proof steps at an arbitrary selected level of granularity. Second, HOL Light implements higher-order logic (Church, 1940) as its foundation, which on the one hand is powerful enough to encode most of today’s formal proofs, and on the other hand allows for an easy integration of many powerful automation mechanisms (Baader & Nipkow, 1998; Paulson, 1999). + +When selecting the theorems to record, we choose an intermediate approach between HOL Light ProofRecording (Obua & Skalberg, 2006) and the HOL/Import one (Kaliszyk & Krauss, 2013). The theorems that are derived by most common proof functions are extracted by patching these functions like in the former approach, and the remaining theorems are extracted from the underlying OCaml programming language interpreter. In certain cases decision procedures derive theorems to be reused in subsequent invocations. We detect such values by looking at theorems used across proof blocks and avoid extracting such reused unrelated subproofs. + +All kernel-level inferences are recorded together with their respective arguments in a trace file. The trace is processed offline to extract the dependencies of the facts, detect used proof boundaries, mark the used and unused steps, and mark the training and testing examples. Only proofs that have sufficiently many used and unused steps are considered useful for the dataset. The annotated proof trace is processed again by a HOL kernel saving the actual training and testing examples originating from non-trivial reasoning steps. Training and testing examples are grouped by proof: for each proof the conjecture (statement that is finally proved), the dependencies of the theorem are constant, and a list of used and not used intermediate statements is provided. This means that the conjectures used in the training and testing sets are normally disjoint. + +For each statement, whether it is the conjecture, a proof dependency, or an intermediate statement, both a fully parenthesised HOL Light human-like printout is provided, as well as a predefined tokenization. The standard HOL Light printer uses parentheses and operator priorities to make its notations somewhat similar to textbook-style mathematics, while at the same time preserving the complete unambiguity of the order of applications (this is particularly visible for associative operators). The tokenization that we propose attempts to reduce the number of parentheses. To do this we compute the maximum number of arguments that each symbol needs to be applied to, and only mark partial application. This means that fully applied functions (more than $90 \%$ of the applications) do not require neither application operators nor parentheses. Top-level universal quantifications are eliminated, bound variables are represented by their de Bruijn indices (the distance from the corresponding abstraction in the parse tree of the term) and free variables are renamed canonically. Since the Hindley-Milner type inference Hindley (1969) mechanisms will be sufficient to reconstruct the most-general types of the expressions well enough for automated-reasoning techniques Kaliszyk et al. (2015) we erase all type information. Table 1 presents some dataset statistics. The dataset, the description of the used format, the scripts used to generate it and baseline models code are available: + +http://cl-informatik.uibk.ac.at/cek/holstep/ + +Table 1: HolStep dataset statistics + +
TrainTestPositiveNegative
Examples201304619603011045381104538
Avg. length503.18440.20535.52459.66
Avg. tokens87.0180.6295.4877.40
Conjectures99991411--
Avg. dependencies29.5822.8211
+ +# 3 MACHINE LEARNING TASKS + +# 3.1 TASKS DESCRIPTION + +This dataset makes possible several tasks well-suited for machine learning most of which are highly relevant for theorem proving: + +• Predicting whether a statement is useful in the proof of a given conjecture; +• Predicting the dependencies of a proof statement (premise selection); +• Predicting whether a statement is an important one (human named); +• Predicting which conjecture a particular intermediate statement originates from; +• Predicting the name given to a statement; +• Generating intermediate statements useful in the proof of a given conjecture; +• Generating the conjecture the current proof will lead to. + +In what follows we focus on the first task: classifying proof step statements as being useful or not in the context of a given proof. This task may be further specialized into two different tasks: + +• Unconditioned classification of proof steps: determining how likely a given proof is to be useful for the proof it occurred in, based solely on the content of statement (i.e. by only providing the model with the step statement itself, absent any context). Conditioned classification of proof steps: determining how likely a given proof is to be useful for the proof it occurred in, with “conditioning” on the conjecture statement that the proof was aiming to attain, i.e. by providing the model with both the step statement and the conjecture statement). + +In the dataset, for every proof we provide the same number of useful and non-useful steps. As such, the proof step classification problem is a balanced two-class classification problem, where a random baseline would yield an accuracy of 0.5. + +# 3.2 RELEVANCE TO INTERACTIVE AND AUTOMATED THEOREM PROVING + +In the interaction with an interactive theorem prover, the tasks that require most human time are: the search for good intermediate steps; the search for automation techniques able to justify the individual steps, and searching theorem proving libraries for the necessary simpler facts. These three problems directly correspond to the machine learning tasks proposed in the previous subsection. Being able to predict the usefulness of a statement will significantly improve many automation techniques. The generation of good intermediate lemmas or intermediate steps can improve level of granularity of the proof steps. Understanding the correspondence between statements and their names can allow users to search for statements in the libraries more efficiently (Aspinall & Kaliszyk, 2016). Premise selection and filtering are already used in many theorem proving systems, and generation of succeeding steps corresponds to conjecturing and theory exploration. + +![](images/821894620258089e462f597721fc7f2f41b4042c203e976e2ffdd198b87566c3.jpg) +Figure 1: Unconditioned classification model architectures. + +# 4 BASELINE MODELS + +For each task (conditioned and unconditioned classification), we propose three different deep learning architectures, meant to provide a baseline for the classification performance that can be achieved on this dataset. Our models cover a range of architecture features (from convolutional networks to recurrent networks), aiming at probing what characteristics of the data are the most helpful for usefulness classification. + +Our models are implemented in TensorFlow (Abadi et al., 2015) using the Keras framework (Chollet, 2015). Each model was trained on a single Nvidia K80 GPU. Training only takes a few hours per model, which makes running these experiments accessible to most people (they could even be run on a laptop CPU). We are releasing all of our benchmark code as open-source software 2 so as to allow others to reproduce our results and improve upon our models. + +# 4.1 UNCONDITIONED CLASSIFICATION MODELS + +Our three models for this task are as follow: + +• Logistic regression on top of learned token embeddings. This minimal model aims to determine to which extent simple differences between token distribution between useful and non-useful statements can be used to distinguish them. It provides an absolute floor on the performance achievable on this task. • 2-layer 1D convolutional neural network (CNN) with global maxpooling for sequence reduction. This model aims to determine the importance of local patterns of tokens. • 2-layer 1D CNN with LSTM (Hochreiter & Schmidhuber, 1997) sequence reduction. This model aims to determine the importance of order in the features sequences. + +See figure 1 for a layer-by-layer description of these models. + +# 4.2 CONDITIONED CLASSIFICATION MODELS + +For this task, we use versions of the above models that have two siamese branches (identical branches with shared weights), with one branch processing the proof step statement being considered, and the + +![](images/a90e1dc148c2f2fa15c0165b2b7665040e764c96be2465d5ee50f0ad13c726d4.jpg) +Figure 2: Conditioned classification model architectures. + +other branch processing the conjecture. Each branch outputs an embedding; these two embeddings (step embedding and conjecture embedding) are then concatenated and the classified by a fullyconnected network. See figure 2 for a layer-by-layer description of these models. + +# 4.3 INPUT STATEMENTS ENCODING + +It should be noted that all of our models start with an Embedding layer, mapping tokens or characters in the statements to dense vectors in a low-dimensional space. We consider two possible encodings for presenting the input statements (proof steps and conjectures) to the Embedding layers of our models: + +• Character-level encoding of the human-readable versions of the statements, where each character (out of a set of 86 unique characters) in the pretty-printed statements is mapped to a 256-dimensional dense vector. This encoding yields longer statements (training statements are 308 character long on average). +• Token-level encoding of the versions of the statements rendered with our proposed highlevel tokenization scheme. This encoding yields shorter statements (training statements are 60 token long on average), while considerably increasing the size of set of unique tokens (1993 total tokens in the training set). + +Table 2: HolStep proof step classification accuracy without conditioning + +
Logistic regression1D CNN1D CNN-LSTM
Accuracy with char input0.710.820.83
Accuracy with token input0.710.830.77
+ +Table 3: HolStep proof step classification accuracy with conditioning + +
Logistic regressionSiamese 1D CNNSiamese 1D CNN-LSTM
Accuracy with char input0.710.810.83
Accuracy with token input0.710.820.77
+ +# 5 RESULTS + +Experimental results are presented in tables 2 and 3, as well as figs. 3 to 6. + +# 5.1 INFLUENCE OF MODEL ARCHITECTURE + +Our unconditioned logistic regression model yields an accuracy of $71 \%$ , both with character encoding and token encoding (tables 2 and 3). This demonstrates that differences in token or character distributions between useful and non-useful steps alone, absent any context, is sufficient for discriminating between useful and non-useful statements to a reasonable extent. This also demonstrates that the token encoding is not fundamentally more informative than raw character-level statements. + +Additionally, our unconditioned 1D CNN model yields an accuracy of $82 \%$ to $83 \%$ , both with character encoding and token encoding (tables 2 and 3). This demonstrates that patterns of characters or patterns of tokens are considerably more informative than single tokens for the purpose of usefulness classification. + +Finally, our unconditioned convolutional-recurrent model does not improve upon the results of the 1D CNN, which indicates that our models are not able to meaningfully leverage order in the feature sequences into which the statements are encoded. + +# 5.2 INFLUENCE OF INPUT ENCODING + +For the logistic regression model and the 2-layer 1D CNN model, the choice of input encoding seems to have little impact. For the convolutional-recurrent model, the use of the high-level tokenization seems to cause a large decrease in model performance (figs. 4 and 6). This may be due to the fact that token encoding yields shorter sequences, making the use of a LSTM less relevant. + +# 5.3 INFLUENCE OF CONDITIONING ON THE CONJECTURE + +None of our conditioned models appear to be able to improve upon the unconditioned models, which indicates that our architectures are not able to leverage the information provided by the conjecture. The presence of the conditioning does however impact the training profile of our models, in particular by making the 1D CNN model converge faster and overfit significantly quicker (figs. 5 and 6). + +# 6 CONCLUSIONS + +Our baseline deep learning models, albeit fairly weak, are still able to predict statement usefulness with a remarkably high accuracy. Such methods already help first-order automated provers (Kaliszyk & Urban, 2015a) and as the branching factor is higher in HOL the predictions are valuable for a number of practical proving applications. This includes making tableaux-based (Paulson, 1999) and superposition-based (Hurd, 2003) internal ITP proof search significantly more efficient in turn making formalization easier. However, our models do not appear to be able to leverage order in the input sequences, nor conditioning on the conjectures. This is due to the fact that these models are not doing any form of logical reasoning on their input statements; rather they are doing simple pattern matching at the level of n-grams of characters or tokens. This shows the need to focus future efforts on different models that can do reasoning, or alternatively, on systems that blend explicit reasoning (e.g. graph search) with deep learning-based feature learning. A potential new direction would be to leverage the graph structure of HOL statements using e.g. Recursive Neural Tensor Networks (Socher et al., 2013a;b) or other graph-based recursive architectures. + +![](images/a005d1a27b1ed2c048eb84bbd78e94445bdeed1aa42302235e5adc5fb5947e63.jpg) +Figure 3: Training profile of the three unconditioned baseline models with character input. + +![](images/b47242ae74d391eb24307af379669fb0a56e2025f35168a81905c69493fd7b4d.jpg) +Figure 4: Training profile of the three unconditioned baseline models with token input. + +![](images/17c7f989c97c9e8b86b4fe265a96619c2f7ea88a5e06b21d6c64ceddf2338aa8.jpg) +Figure 5: Training profile of the three conditioned baseline models with character input. + +![](images/87d44006190fe92d90fd54144f8b0605b42092602e67eff4e886c823b46d4a17.jpg) +Figure 6: Training profile of the three conditioned baseline models with token input. + +# 6.1 FUTURE WORK + +The dataset focuses on one interactive theorem prover. It would be interesting if the proposed techniques generalize, primarily across ITPs that use the same foundational logic, for example using OpenTheory (Hurd, 2011), and secondarily across fundamentally different ITPs or even ATPs. A significant part of the unused steps originates from trying to fulfill the conditions for rewriting and from calls to intuitionistic tableaux. The main focus is however on the human found proofs so the trained predictions may to an extent mimic the bias on the usefulness in the human proofs. As ATPs are at the moment very week in comparison with human intuition improving this even for the many proofs humans do not find difficult would be an important gain. + +Finally, two of the proposed task for the dataset have been premise selection and intermediate sentence generation. It would be interesting to define more ATP-based ways to evaluate the selected premises, as well as to evaluate generated sentences (Kaliszyk et al., 2015). The set is a relatively large one when it comes to proof step classification, however the number of available premises makes the set a medium-sized set for premise selection in comparison with those of the Mizar Mathematical Library or the seL4 development. + +# ACKNOWLEDGEMENTS + +The first author was partly supported by the ERC starting grant 714034. + +# REFERENCES + +Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. + +Jesse Alama, Tom Heskes, Daniel Kühlwein, Evgeni Tsivtsivadze, and Josef Urban. Premise selection for mathematics by corpus analysis and kernel methods. J. Autom. Reasoning, 52(2): 191–213, 2014. doi: 10.1007/s10817-013-9286-5. + +Alex A. Alemi, François Chollet, Geoffrey Irving, Christian Szegedy, and Josef Urban. DeepMath – Deep sequence models for premise selection. In Daniel D. Lee, Masashi Sugiyama, Ulrike V. Luxburg, Isabelle Guyon, and Roman Garnett (eds.), Advances in Neural Information Processing Systems (NIPS 2016), pp. 2235–2243, 2016. URL https://arxiv.org/abs/1606.04442. + +David Aspinall and Cezary Kaliszyk. What’s in a theorem name? In Jasmin Christian Blanchette and Stephan Merz (eds.), Interactive Theorem Proving (ITP 2016), volume 9807 of LNCS, pp. 459–465. Springer, 2016. doi: 10.1007/978-3-319-43144-4. + +Franz Baader and Tobias Nipkow. Term rewriting and all that. Cambridge University Press, 1998. ISBN 978-0-521-45520-6. + +Jasmin C. Blanchette, Cezary Kaliszyk, Lawrence C. Paulson, and Josef Urban. Hammering towards QED. J. Formalized Reasoning, 9(1):101–148, 2016. ISSN 1972-5787. doi: 10.6092/issn. 1972-5787/4593. + +Jasmin Christian Blanchette, Maximilian P. L. Haslbeck, Daniel Matichuk, and Tobias Nipkow. Mining the Archive of Formal Proofs. In Manfred Kerber, Jacques Carette, Cezary Kaliszyk, Florian Rabe, and Volker Sorge (eds.), Intelligent Computer Mathematics (CICM 2015), volume 9150 of LNCS, pp. 3–17. Springer, 2015. + +James P. Bridge, Sean B. Holden, and Lawrence C. Paulson. Machine learning for first-order theorem proving - learning to select a good heuristic. J. Autom. Reasoning, 53(2):141–172, 2014. doi: 10.1007/s10817-014-9301-5. + +Haogang Chen, Daniel Ziegler, Tej Chajed, Adam Chlipala, M. Frans Kaashoek, and Nickolai Zeldovich. Using crash Hoare logic for certifying the FSCQ file system. In Ajay Gulati and Hakim Weatherspoon (eds.), USENIX 2016. USENIX Association, 2016. + +François Chollet. Keras. https://github.com/fchollet/keras, 2015. + +Alonzo Church. A formulation of the simple theory of types. J. Symb. Log., 5(2):56–68, 1940. doi: 10.2307/2266170. URL http://dx.doi.org/10.2307/2266170. + +J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. + +Hazel Duncan. The Use of Data-Mining for the Automatic Formation of Tactics. PhD thesis, University of Edinburgh, 2007. + +Michael Färber and Chad E. Brown. Internal guidance for Satallax. In Nicola Olivetti and Ashish Tiwari (eds.), International Joint Conference on Automated Reasoning (IJCAR 2016), volume 9706 of LNCS, pp. 349–361. Springer, 2016. doi: 10.1007/978-3-319-40229-1. + +Georges Gonthier, Andrea Asperti, Jeremy Avigad, Yves Bertot, Cyril Cohen, François Garillot, Stéphane Le Roux, Assia Mahboubi, Russell O’Connor, Sidi Ould Biha, Ioana Pasca, Laurence Rideau, Alexey Solovyev, Enrico Tassi, and Laurent Théry. A machine-checked proof of the odd order theorem. In Sandrine Blazy, Christine Paulin-Mohring, and David Pichardie (eds.), Interactive Theorem Proving (ITP 2013), volume 7998 of LNCS, pp. 163–179. Springer, 2013. + +Michael J. C. Gordon, Robin Milner, and Christopher P. Wadsworth. Edinburgh LCF, volume 78 of Lecture Notes in Computer Science. Springer, 1979. ISBN 3-540-09724-4. doi: 10.1007/ 3-540-09724-4. URL http://dx.doi.org/10.1007/3-540-09724-4. + +Adam Grabowski, Artur Kornilowicz, and Adam Naumowicz. Mizar in a nutshell. J. Formalized Reasoning, 3(2):153–245, 2010. doi: 10.6092/issn.1972-5787/1980. + +Thomas Hales, John Harrison, Sean McLaughlin, Tobias Nipkow, Steven Obua, and Roland Zumkeller. A revision of the proof of the Kepler Conjecture. Discrete & Computational Geometry, 44(1):1–34, 2010. + +Thomas C. Hales, Mark Adams, Gertrud Bauer, Dat Tat Dang, John Harrison, Truong Le Hoang, Cezary Kaliszyk, Victor Magron, Sean McLaughlin, Thang Tat Nguyen, Truong Quang Nguyen, Tobias Nipkow, Steven Obua, Joseph Pleso, Jason Rute, Alexey Solovyev, An Hoai Thi Ta, Trung Nam Tran, Diep Thi Trieu, Josef Urban, Ky Khac Vu, and Roland Zumkeller. A formal proof of the Kepler conjecture. CoRR, abs/1501.02155, 2015. + +John Harrison. HOL Light: An overview. In Stefan Berghofer, Tobias Nipkow, Christian Urban, and Makarius Wenzel (eds.), Theorem Proving in Higher Order Logics (TPHOLs 2009), volume 5674 of LNCS, pp. 60–66. Springer, 2009. + +John Harrison. The HOL Light theory of Euclidean space. J. Autom. Reasoning, 50(2):173–190, 2013. doi: 10.1007/s10817-012-9250-9. + +John Harrison, Josef Urban, and Freek Wiedijk. History of interactive theorem proving. In Jörg Siekmann (ed.), Handbook of the History of Logic vol. 9 (Computational Logic), pp. 135–214. Elsevier, 2014. + +R. Hindley. The principal type-scheme of an object in combinatory logic. Transactions of the american mathematical society, 146:29–60, 1969. ISSN 0002-9947. + +Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8): 1735–1780, 1997. + +Kryštof Hoder and Andrei Voronkov. Sine qua non for large theory reasoning. In Nikolaj Bjørner and Viorica Sofronie-Stokkermans (eds.), CADE-23, volume 6803 of LNAI, pp. 299–314. Springer, 2011. + +Joe Hurd. First-order proof tactics in higher-order logic theorem provers. In Myla Archer, Ben Di Vito, and César Muñoz (eds.), Design and Application of Strategies/Tactics in Higher Order Logics (STRATA 2003), number NASA/CP-2003-212448 in NASA Technical Reports, pp. 56–68, September 2003. URL http://www.gilith.com/research/papers. + +Joe Hurd. The OpenTheory standard theory library. In Mihaela Gheorghiu Bobaru, Klaus Havelund, Gerard J. Holzmann, and Rajeev Joshi (eds.), NASA Formal Methods (NFM 2011), volume 6617 of LNCS, pp. 177–191. Springer, 2011. + +Cezary Kaliszyk and Alexander Krauss. Scalable LCF-style proof translation. In Sandrine Blazy, Christine Paulin-Mohring, and David Pichardie (eds.), Interactive Theorem Proving (ITP 2013), volume 7998 of LNCS, pp. 51–66. Springer, 2013. + +Cezary Kaliszyk and Josef Urban. Learning-assisted automated reasoning with Flyspeck. J. Autom. Reasoning, 53(2):173–213, 2014. doi: 10.1007/s10817-014-9303-3. + +Cezary Kaliszyk and Josef Urban. FEMaLeCoP: Fairly efficient machine learning connection prover. In Martin Davis, Ansgar Fehnker, Annabelle McIver, and Andrei Voronkov (eds.), 20th International Conference on Logic for Programming, Artificial Intelligence, and Reasoning (LPAR 2015), volume 9450 of LNCS, pp. 88–96. Springer, 2015a. doi: 10.1007/978-3-662-48899-7. + +Cezary Kaliszyk and Josef Urban. Learning-assisted theorem proving with millions of lemmas. J. Symbolic Computation, 69:109–128, 2015b. doi: 10.1016/j.jsc.2014.09.032. + +Cezary Kaliszyk, Josef Urban, and Jiˇrí Vyskocil. Learning to parse on aligned corpora. In Christian ˇ Urban and Xingyuan Zhang (eds.), Proc. 6h Conference on Interactive Theorem Proving (ITP’15), volume 9236 of LNCS, pp. 227–233. Springer-Verlag, 2015. doi: 10.1007/978-3-319-22102-1_ 15. + +Gerwin Klein, June Andronick, Kevin Elphinstone, Toby C. Murray, Thomas Sewell, Rafal Kolanski, and Gernot Heiser. Comprehensive formal verification of an OS microkernel. ACM Trans. Comput. Syst., 32(1):2, 2014. + +Laura Kovács and Andrei Voronkov. First-order theorem proving and Vampire. In Natasha Sharygina and Helmut Veith (eds.), Computer-Aided Verification (CAV 2013), volume 8044 of LNCS, pp. 1–35. Springer, 2013. + +Daniel Kühlwein and Josef Urban. MaLeS: A framework for automatic tuning of automated theorem provers. J. Autom. Reasoning, 55(2):91–116, 2015. doi: 10.1007/s10817-015-9329-1. + +Xavier Leroy. Formal verification of a realistic compiler. Commun. ACM, 52(7):107–115, 2009. + +Steven Obua and Sebastian Skalberg. Importing HOL into Isabelle/HOL. In Ulrich Furbach and Natarajan Shankar (eds.), International Joint Conference on Automated Reasoning (IJCAR 2006), volume 4130 of LNCS, pp. 298–302. Springer, 2006. + +Lawrence C. Paulson. A generic tableau prover and its integration with Isabelle. J. Universal Computer Science, 5(3):73–87, 1999. + +Stephan Schulz. System description: E 1.8. In Kenneth L. McMillan, Aart Middeldorp, and Andrei Voronkov (eds.), Logic for Programming, Artificial Intelligence (LPAR 2013), volume 8312 of LNCS, pp. 735–743. Springer, 2013. + +David Silver, Aja Huang, Christopher J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of go with deep neural networks and tree search. Nature, 529:484–503, 2016. URL http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html. + +Richard Socher, Danqi Chen, Christopher D. Manning, and Andrew Y. Ng. Reasoning with neural tensor networks for knowledge base completion. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings., pp. 926–934, 2013a. URL http://papers.nips.cc/paper/ 5028-reasoning-with-neural-tensor-networks-for-knowledge-base-completion. + +Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y. $\mathrm { N g }$ , and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1631–1642, Stroudsburg, PA, October 2013b. Association for Computational Linguistics. + +Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. In Advances in Neural Information Processing Systems, pp. 2431–2439, 2015. + +Geoff Sutcliffe. The CADE ATP system competition - CASC. AI Magazine, 37(2):99–101, 2016. + +Josef Urban, Jiˇrí Vyskocil, and Petr Št ˇ epánek. MaLeCoP: Machine learning connection prover. In ˇ Kai Brünnler and George Metcalfe (eds.), TABLEAUX 2011, volume 6793 of LNCS. Springer, 2011. + +Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Lukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural machine translation system: Bridging the gap between human and machine translation. CoRR, abs/1609.08144, 2016. URL http://arxiv.org/abs/1609.08144. \ No newline at end of file diff --git a/parse/train/ryuxYmvel/ryuxYmvel_content_list.json b/parse/train/ryuxYmvel/ryuxYmvel_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc2602da0d38e2014aa4b557b519286e3ff9425d --- /dev/null +++ b/parse/train/ryuxYmvel/ryuxYmvel_content_list.json @@ -0,0 +1,1426 @@ +[ + { + "type": "text", + "text": "HOLSTEP: A MACHINE LEARNING DATASET FOR HIGHER-ORDER LOGIC THEOREM PROVING ", + "text_level": 1, + "bbox": [ + 176, + 98, + 821, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cezary Kaliszyk \nUniversity of Innsbruck \ncezary.kaliszyk@uibk.ac.at ", + "bbox": [ + 184, + 170, + 439, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "François Chollet, Christian Szegedy Google Research {fchollet,szegedy}@google.com ", + "bbox": [ + 493, + 170, + 779, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 250, + 544, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large computer-understandable proofs consist of millions of intermediate logical steps. The vast majority of such steps originate from manually selected and manually guided heuristics applied to intermediate goals. So far, machine learning has generally not been used to filter or generate these steps. In this paper, we introduce a new dataset based on Higher-Order Logic (HOL) proofs, for the purpose of developing new machine learning-based theorem-proving strategies. We make this dataset publicly available under the BSD license. We propose various machine learning tasks that can be performed on this dataset, and discuss their significance for theorem proving. We also benchmark a set of simple baseline machine learning models suited for the tasks (including logistic regression, convolutional neural networks and recurrent neural networks). The results of our baseline models show the promise of applying machine learning to HOL theorem proving. ", + "bbox": [ + 233, + 280, + 764, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 474, + 336, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the usability of interactive theorem proving (ITP) systems (Harrison et al., 2014) grows, its use becomes a more common way of establishing the correctness of software as well as mathematical proofs. Today, ITPs are used for software certification projects ranging from compilers (Leroy, 2009) and operating system components (Chen et al., 2016; Klein et al., 2014), to establishing the absolute correctness of large proofs in mathematics such as the Kepler conjecture (Hales et al., 2015) and the Feit-Thomson Theorem (Gonthier et al., 2013). ", + "bbox": [ + 174, + 506, + 825, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For results of such significance to be possible, the theorem libraries of these ITPs must contain all necessary basic mathematical properties, accompanied with formal proofs. This means that the size of many ITP libraries can be measured in dozens of thousands of theorems (Grabowski et al., 2010; Blanchette et al., 2015) and billions of individual proof steps. While the general direction of the proofs is specified by humans (by providing the goal to prove, specifying intermediate steps, or applying certain automated tactics), the majority of such proof steps are actually found by automated reasoning-based proof search (Kaliszyk & Urban, 2015b), with very little application of machine learning techniques so far. ", + "bbox": [ + 174, + 597, + 825, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At the same time, fast progress has been unfolding in machine learning applied to tasks that involve logical inference, such as natural language question answering (Sukhbaatar et al., 2015), knowledge base completion (Socher et al., 2013a), automated translation (Wu et al., 2016), and premise selection in the context of theorem proving (Alemi et al., 2016). Deep learning in particular has proven to be a powerful tool for embedding semantic meaning and logical relationships into geometric spaces, specifically via models such as convolutional neural networks, recurrent neural networks, and tree-recursive neural networks. These advances strongly suggest that deep learning may have become mature enough to yield significant advances in automated theorem proving. Remarkably, it has recently become possible to build a system, AlphaGo (Silver et al., 2016), blending classical AI techniques such as Monte-Carlo tree search and modern deep learning techniques, capable of playing the game of Go at super-human levels. We should note that theorem proving and Go playing are conceptually related, since both consist in searching for specific nodes in trees of states with extremely large arity and relatively large depth, which involves node evaluation decision (how valuable is this state?) and policy decisions (which node should be expanded next?). The success of AlphaGo can thus serve as encouragement on the road to building deep learning-augmented theorem provers that would blend classical techniques developed over the past few decades with the latest machine learning advances. ", + "bbox": [ + 174, + 715, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fast progress in specific machine learning verticals has occasionally been achieved thanks to the release of specialized datasets (often with associated competitions, e.g. the ImageNet dataset for large-scale image classification (Deng et al., 2009)) serving as an experimental testbed and public benchmark of current progress, thus focusing the efforts of the research community. We hope that releasing a theorem proving dataset suited for specific machine learning tasks can serve the same purpose in the vertical of applying machine learning to theorem proving. ", + "bbox": [ + 174, + 138, + 825, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1 CONTRIBUTION AND OVERVIEW ", + "text_level": 1, + "bbox": [ + 178, + 247, + 437, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we develop a dataset for machine learning based on the proof steps used in a large interactive proof section 2. We focus on the HOL Light (Harrison, 2009) ITP, its multivariate analysis library (Harrison, 2013), as well as the formal proof of the Kepler conjecture (Hales et al., 2010). These formalizations constitute a diverse proof dataset containing basic mathematics, analysis, trigonometry, as well as reasoning about data structures such as graphs. Furthermore these formal proof developments have been used as benchmarks for automated reasoning techniques (Kaliszyk & Urban, 2014). ", + "bbox": [ + 174, + 276, + 825, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The dataset consists of 2,013,046 training examples and 196,030 testing examples that originate from 11,400 proofs. Precisely half of the examples are statements that were useful in the currently proven conjectures and half are steps that have been derived either manually or as part of the automated proof search but were not necessary in the final proofs. The dataset contains only proofs of non-trivial theorems, that also do not focus on computation but rather on actual theorem proving. For each proof, the conjecture that is being proven as well as its dependencies (axioms) and may be exploited in machine learning tasks. Furthermore, for each statement both its human-readable (pretty-printed) statement and a tokenization designed to make machine learning tasks more manageable are included. ", + "bbox": [ + 174, + 381, + 825, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, in section 3 we discuss the proof step classification tasks that can be attempted using the dataset, and we discuss the usefulness of these tasks in interactive and automated theorem proving. These tasks include unconditioned classification (without access to conjectures and dependencies) and conjecture-conditioned classification (with access to the conjecture) of proof steps as being useful or not in a proof. We outline the use of such classification capabilities for search space pruning and internal guidance, as well as for generation of intermediate steps or possible new lemma statements. ", + "bbox": [ + 174, + 513, + 825, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, in section 4 we propose three baseline models for the proof step classification tasks, and we experimentally evaluate the models on the data in section 5. The models considered include both a relatively simple regression model, as well as deep learning models based on convolutional and recurrent neural networks. ", + "bbox": [ + 174, + 617, + 825, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 699, + 331, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The use of machine learning in interactive and automated theorem proving has so far focused on three tasks: premise selection, strategy selection, and internal guidance. We shortly explain these. ", + "bbox": [ + 174, + 728, + 823, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a large library of proven facts and a user given conjecture, the multi-label classification problem of selecting the facts that are most likely to lead to a successful proof of the conjecture has been usually called relevance filtering or premise selection (Alama et al., 2014). This is crucial for the efficiency of modern automation techniques for ITPs (Blanchette et al., 2016), which today can usually solve $4 0 { - } 5 0 \\%$ of the conjectures in theorem proving libraries. Similarly most competitive ATPs today (Sutcliffe, 2016) implement the SInE classifier (Hoder & Voronkov, 2011). ", + "bbox": [ + 174, + 763, + 825, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A second theorem proving task where machine learning has been of importance is strategy selection. With the development of automated theorem provers came many parameters that control their execution. In fact, modern ATPs, such as E (Schulz, 2013) and Vampire (Kovács & Voronkov, 2013), include complete strategy description languages that allow a user to specify the orderings, weighting functions, literal selection strategies, etc. Rather than optimizing the search strategy globally, one can choose the strategy based on the currently considered problem. For this some frameworks use machine learning (Bridge et al., 2014; Kühlwein & Urban, 2015). ", + "bbox": [ + 176, + 854, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, an automated theorem prover may use machine learning for choosing the actual inference steps. It has been shown to significantly reduce the proof search in first-order tableaux by the selection of extension steps to use (Urban et al., 2011), and has been also successfully applied in monomorphic higher-order logic proving (Färber & Brown, 2016). Data/proof mining has also been applied on the level of interactive theorem proving tactics (Duncan, 2007) to extract and reuse repeating patterns. ", + "bbox": [ + 174, + 140, + 825, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 DATASET EXTRACTION ", + "text_level": 1, + "bbox": [ + 176, + 250, + 398, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on the HOL Light theorem prover for two reasons. First, it follows the LCF approach1). This means that complicated inferences are reduced to the most primitive ones and the data extraction related modifications can be restricted the primitive inferences and it is relatively easy to extract proof steps at an arbitrary selected level of granularity. Second, HOL Light implements higher-order logic (Church, 1940) as its foundation, which on the one hand is powerful enough to encode most of today’s formal proofs, and on the other hand allows for an easy integration of many powerful automation mechanisms (Baader & Nipkow, 1998; Paulson, 1999). ", + "bbox": [ + 174, + 285, + 825, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When selecting the theorems to record, we choose an intermediate approach between HOL Light ProofRecording (Obua & Skalberg, 2006) and the HOL/Import one (Kaliszyk & Krauss, 2013). The theorems that are derived by most common proof functions are extracted by patching these functions like in the former approach, and the remaining theorems are extracted from the underlying OCaml programming language interpreter. In certain cases decision procedures derive theorems to be reused in subsequent invocations. We detect such values by looking at theorems used across proof blocks and avoid extracting such reused unrelated subproofs. ", + "bbox": [ + 174, + 388, + 825, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All kernel-level inferences are recorded together with their respective arguments in a trace file. The trace is processed offline to extract the dependencies of the facts, detect used proof boundaries, mark the used and unused steps, and mark the training and testing examples. Only proofs that have sufficiently many used and unused steps are considered useful for the dataset. The annotated proof trace is processed again by a HOL kernel saving the actual training and testing examples originating from non-trivial reasoning steps. Training and testing examples are grouped by proof: for each proof the conjecture (statement that is finally proved), the dependencies of the theorem are constant, and a list of used and not used intermediate statements is provided. This means that the conjectures used in the training and testing sets are normally disjoint. ", + "bbox": [ + 174, + 493, + 825, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each statement, whether it is the conjecture, a proof dependency, or an intermediate statement, both a fully parenthesised HOL Light human-like printout is provided, as well as a predefined tokenization. The standard HOL Light printer uses parentheses and operator priorities to make its notations somewhat similar to textbook-style mathematics, while at the same time preserving the complete unambiguity of the order of applications (this is particularly visible for associative operators). The tokenization that we propose attempts to reduce the number of parentheses. To do this we compute the maximum number of arguments that each symbol needs to be applied to, and only mark partial application. This means that fully applied functions (more than $90 \\%$ of the applications) do not require neither application operators nor parentheses. Top-level universal quantifications are eliminated, bound variables are represented by their de Bruijn indices (the distance from the corresponding abstraction in the parse tree of the term) and free variables are renamed canonically. Since the Hindley-Milner type inference Hindley (1969) mechanisms will be sufficient to reconstruct the most-general types of the expressions well enough for automated-reasoning techniques Kaliszyk et al. (2015) we erase all type information. Table 1 presents some dataset statistics. The dataset, the description of the used format, the scripts used to generate it and baseline models code are available: ", + "bbox": [ + 174, + 626, + 825, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "http://cl-informatik.uibk.ac.at/cek/holstep/ ", + "bbox": [ + 361, + 837, + 638, + 851 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6073529efc3cc681e06d1137becf13db8721e1be8599e3757798b48405083e31.jpg", + "table_caption": [ + "Table 1: HolStep dataset statistics " + ], + "table_footnote": [], + "table_body": "
TrainTestPositiveNegative
Examples201304619603011045381104538
Avg. length503.18440.20535.52459.66
Avg. tokens87.0180.6295.4877.40
Conjectures99991411--
Avg. dependencies29.5822.8211
", + "bbox": [ + 254, + 101, + 741, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 MACHINE LEARNING TASKS ", + "text_level": 1, + "bbox": [ + 176, + 253, + 444, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 TASKS DESCRIPTION ", + "text_level": 1, + "bbox": [ + 176, + 287, + 359, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This dataset makes possible several tasks well-suited for machine learning most of which are highly relevant for theorem proving: ", + "bbox": [ + 173, + 315, + 823, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• Predicting whether a statement is useful in the proof of a given conjecture; \n• Predicting the dependencies of a proof statement (premise selection); \n• Predicting whether a statement is an important one (human named); \n• Predicting which conjecture a particular intermediate statement originates from; \n• Predicting the name given to a statement; \n• Generating intermediate statements useful in the proof of a given conjecture; \n• Generating the conjecture the current proof will lead to. ", + "bbox": [ + 217, + 359, + 758, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In what follows we focus on the first task: classifying proof step statements as being useful or not in the context of a given proof. This task may be further specialized into two different tasks: ", + "bbox": [ + 173, + 526, + 823, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• Unconditioned classification of proof steps: determining how likely a given proof is to be useful for the proof it occurred in, based solely on the content of statement (i.e. by only providing the model with the step statement itself, absent any context). Conditioned classification of proof steps: determining how likely a given proof is to be useful for the proof it occurred in, with “conditioning” on the conjecture statement that the proof was aiming to attain, i.e. by providing the model with both the step statement and the conjecture statement). ", + "bbox": [ + 215, + 570, + 825, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the dataset, for every proof we provide the same number of useful and non-useful steps. As such, the proof step classification problem is a balanced two-class classification problem, where a random baseline would yield an accuracy of 0.5. ", + "bbox": [ + 176, + 694, + 825, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 RELEVANCE TO INTERACTIVE AND AUTOMATED THEOREM PROVING", + "text_level": 1, + "bbox": [ + 176, + 758, + 687, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the interaction with an interactive theorem prover, the tasks that require most human time are: the search for good intermediate steps; the search for automation techniques able to justify the individual steps, and searching theorem proving libraries for the necessary simpler facts. These three problems directly correspond to the machine learning tasks proposed in the previous subsection. Being able to predict the usefulness of a statement will significantly improve many automation techniques. The generation of good intermediate lemmas or intermediate steps can improve level of granularity of the proof steps. Understanding the correspondence between statements and their names can allow users to search for statements in the libraries more efficiently (Aspinall & Kaliszyk, 2016). Premise selection and filtering are already used in many theorem proving systems, and generation of succeeding steps corresponds to conjecturing and theory exploration. ", + "bbox": [ + 173, + 785, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/821894620258089e462f597721fc7f2f41b4042c203e976e2ffdd198b87566c3.jpg", + "image_caption": [ + "Figure 1: Unconditioned classification model architectures. " + ], + "image_footnote": [], + "bbox": [ + 320, + 131, + 683, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 BASELINE MODELS", + "text_level": 1, + "bbox": [ + 176, + 435, + 369, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each task (conditioned and unconditioned classification), we propose three different deep learning architectures, meant to provide a baseline for the classification performance that can be achieved on this dataset. Our models cover a range of architecture features (from convolutional networks to recurrent networks), aiming at probing what characteristics of the data are the most helpful for usefulness classification. ", + "bbox": [ + 174, + 467, + 825, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our models are implemented in TensorFlow (Abadi et al., 2015) using the Keras framework (Chollet, 2015). Each model was trained on a single Nvidia K80 GPU. Training only takes a few hours per model, which makes running these experiments accessible to most people (they could even be run on a laptop CPU). We are releasing all of our benchmark code as open-source software 2 so as to allow others to reproduce our results and improve upon our models. ", + "bbox": [ + 174, + 544, + 825, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 UNCONDITIONED CLASSIFICATION MODELS ", + "text_level": 1, + "bbox": [ + 173, + 631, + 517, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our three models for this task are as follow: ", + "bbox": [ + 176, + 656, + 462, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "• Logistic regression on top of learned token embeddings. This minimal model aims to determine to which extent simple differences between token distribution between useful and non-useful statements can be used to distinguish them. It provides an absolute floor on the performance achievable on this task. • 2-layer 1D convolutional neural network (CNN) with global maxpooling for sequence reduction. This model aims to determine the importance of local patterns of tokens. • 2-layer 1D CNN with LSTM (Hochreiter & Schmidhuber, 1997) sequence reduction. This model aims to determine the importance of order in the features sequences. ", + "bbox": [ + 215, + 683, + 825, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "See figure 1 for a layer-by-layer description of these models. ", + "bbox": [ + 173, + 815, + 570, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 CONDITIONED CLASSIFICATION MODELS ", + "text_level": 1, + "bbox": [ + 174, + 847, + 496, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For this task, we use versions of the above models that have two siamese branches (identical branches with shared weights), with one branch processing the proof step statement being considered, and the ", + "bbox": [ + 178, + 872, + 825, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a90e1dc148c2f2fa15c0165b2b7665040e764c96be2465d5ee50f0ad13c726d4.jpg", + "image_caption": [ + "Figure 2: Conditioned classification model architectures. " + ], + "image_footnote": [], + "bbox": [ + 250, + 136, + 751, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "other branch processing the conjecture. Each branch outputs an embedding; these two embeddings (step embedding and conjecture embedding) are then concatenated and the classified by a fullyconnected network. See figure 2 for a layer-by-layer description of these models. ", + "bbox": [ + 174, + 621, + 825, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 INPUT STATEMENTS ENCODING ", + "text_level": 1, + "bbox": [ + 176, + 690, + 431, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It should be noted that all of our models start with an Embedding layer, mapping tokens or characters in the statements to dense vectors in a low-dimensional space. We consider two possible encodings for presenting the input statements (proof steps and conjectures) to the Embedding layers of our models: ", + "bbox": [ + 176, + 720, + 825, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "• Character-level encoding of the human-readable versions of the statements, where each character (out of a set of 86 unique characters) in the pretty-printed statements is mapped to a 256-dimensional dense vector. This encoding yields longer statements (training statements are 308 character long on average). \n• Token-level encoding of the versions of the statements rendered with our proposed highlevel tokenization scheme. This encoding yields shorter statements (training statements are 60 token long on average), while considerably increasing the size of set of unique tokens (1993 total tokens in the training set). ", + "bbox": [ + 214, + 797, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1cb801e38b574428eabc8f684adacd5e5fc3dfbf2886e132fcdca14fa9e5fd1a.jpg", + "table_caption": [ + "Table 2: HolStep proof step classification accuracy without conditioning " + ], + "table_footnote": [], + "table_body": "
Logistic regression1D CNN1D CNN-LSTM
Accuracy with char input0.710.820.83
Accuracy with token input0.710.830.77
", + "bbox": [ + 228, + 127, + 769, + 199 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4f4c0de5892ddae27d24b3b9fd52e205969fc05f3bcb7d6ca6be57bfb4e9995.jpg", + "table_caption": [ + "Table 3: HolStep proof step classification accuracy with conditioning " + ], + "table_footnote": [], + "table_body": "
Logistic regressionSiamese 1D CNNSiamese 1D CNN-LSTM
Accuracy with char input0.710.810.83
Accuracy with token input0.710.820.77
", + "bbox": [ + 228, + 238, + 769, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 333, + 282, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental results are presented in tables 2 and 3, as well as figs. 3 to 6. ", + "bbox": [ + 176, + 364, + 660, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 INFLUENCE OF MODEL ARCHITECTURE ", + "text_level": 1, + "bbox": [ + 176, + 397, + 483, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our unconditioned logistic regression model yields an accuracy of $71 \\%$ , both with character encoding and token encoding (tables 2 and 3). This demonstrates that differences in token or character distributions between useful and non-useful steps alone, absent any context, is sufficient for discriminating between useful and non-useful statements to a reasonable extent. This also demonstrates that the token encoding is not fundamentally more informative than raw character-level statements. ", + "bbox": [ + 174, + 421, + 825, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, our unconditioned 1D CNN model yields an accuracy of $82 \\%$ to $83 \\%$ , both with character encoding and token encoding (tables 2 and 3). This demonstrates that patterns of characters or patterns of tokens are considerably more informative than single tokens for the purpose of usefulness classification. ", + "bbox": [ + 174, + 498, + 825, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, our unconditioned convolutional-recurrent model does not improve upon the results of the 1D CNN, which indicates that our models are not able to meaningfully leverage order in the feature sequences into which the statements are encoded. ", + "bbox": [ + 176, + 563, + 825, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 INFLUENCE OF INPUT ENCODING ", + "text_level": 1, + "bbox": [ + 176, + 622, + 442, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the logistic regression model and the 2-layer 1D CNN model, the choice of input encoding seems to have little impact. For the convolutional-recurrent model, the use of the high-level tokenization seems to cause a large decrease in model performance (figs. 4 and 6). This may be due to the fact that token encoding yields shorter sequences, making the use of a LSTM less relevant. ", + "bbox": [ + 174, + 647, + 825, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 INFLUENCE OF CONDITIONING ON THE CONJECTURE ", + "text_level": 1, + "bbox": [ + 174, + 722, + 578, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "None of our conditioned models appear to be able to improve upon the unconditioned models, which indicates that our architectures are not able to leverage the information provided by the conjecture. The presence of the conditioning does however impact the training profile of our models, in particular by making the 1D CNN model converge faster and overfit significantly quicker (figs. 5 and 6). ", + "bbox": [ + 174, + 746, + 825, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 821, + 328, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our baseline deep learning models, albeit fairly weak, are still able to predict statement usefulness with a remarkably high accuracy. Such methods already help first-order automated provers (Kaliszyk & Urban, 2015a) and as the branching factor is higher in HOL the predictions are valuable for a number of practical proving applications. This includes making tableaux-based (Paulson, 1999) and superposition-based (Hurd, 2003) internal ITP proof search significantly more efficient in turn making formalization easier. However, our models do not appear to be able to leverage order in the input sequences, nor conditioning on the conjectures. This is due to the fact that these models are not doing any form of logical reasoning on their input statements; rather they are doing simple pattern matching at the level of n-grams of characters or tokens. This shows the need to focus future efforts on different models that can do reasoning, or alternatively, on systems that blend explicit reasoning (e.g. graph search) with deep learning-based feature learning. A potential new direction would be to leverage the graph structure of HOL statements using e.g. Recursive Neural Tensor Networks (Socher et al., 2013a;b) or other graph-based recursive architectures. ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a005d1a27b1ed2c048eb84bbd78e94445bdeed1aa42302235e5adc5fb5947e63.jpg", + "image_caption": [ + "Figure 3: Training profile of the three unconditioned baseline models with character input. " + ], + "image_footnote": [], + "bbox": [ + 209, + 102, + 454, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b47242ae74d391eb24307af379669fb0a56e2025f35168a81905c69493fd7b4d.jpg", + "image_caption": [ + "Figure 4: Training profile of the three unconditioned baseline models with token input. " + ], + "image_footnote": [], + "bbox": [ + 529, + 102, + 772, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17c7f989c97c9e8b86b4fe265a96619c2f7ea88a5e06b21d6c64ceddf2338aa8.jpg", + "image_caption": [ + "Figure 5: Training profile of the three conditioned baseline models with character input. " + ], + "image_footnote": [], + "bbox": [ + 210, + 321, + 454, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87d44006190fe92d90fd54144f8b0605b42092602e67eff4e886c823b46d4a17.jpg", + "image_caption": [ + "Figure 6: Training profile of the three conditioned baseline models with token input. " + ], + "image_footnote": [], + "bbox": [ + 529, + 320, + 771, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 553, + 825, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 FUTURE WORK ", + "text_level": 1, + "bbox": [ + 174, + 691, + 318, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The dataset focuses on one interactive theorem prover. It would be interesting if the proposed techniques generalize, primarily across ITPs that use the same foundational logic, for example using OpenTheory (Hurd, 2011), and secondarily across fundamentally different ITPs or even ATPs. A significant part of the unused steps originates from trying to fulfill the conditions for rewriting and from calls to intuitionistic tableaux. The main focus is however on the human found proofs so the trained predictions may to an extent mimic the bias on the usefulness in the human proofs. As ATPs are at the moment very week in comparison with human intuition improving this even for the many proofs humans do not find difficult would be an important gain. ", + "bbox": [ + 174, + 722, + 825, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, two of the proposed task for the dataset have been premise selection and intermediate sentence generation. It would be interesting to define more ATP-based ways to evaluate the selected premises, as well as to evaluate generated sentences (Kaliszyk et al., 2015). The set is a relatively large one when it comes to proof step classification, however the number of available premises makes the set a medium-sized set for premise selection in comparison with those of the Mizar Mathematical Library or the seL4 development. ", + "bbox": [ + 174, + 840, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENTS ", + "text_level": 1, + "bbox": [ + 176, + 104, + 334, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The first author was partly supported by the ERC starting grant 714034. ", + "bbox": [ + 176, + 130, + 642, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 165, + 285, + 179 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org. ", + "bbox": [ + 176, + 188, + 825, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jesse Alama, Tom Heskes, Daniel Kühlwein, Evgeni Tsivtsivadze, and Josef Urban. Premise selection for mathematics by corpus analysis and kernel methods. J. Autom. Reasoning, 52(2): 191–213, 2014. doi: 10.1007/s10817-013-9286-5. ", + "bbox": [ + 174, + 321, + 825, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex A. Alemi, François Chollet, Geoffrey Irving, Christian Szegedy, and Josef Urban. DeepMath – Deep sequence models for premise selection. In Daniel D. Lee, Masashi Sugiyama, Ulrike V. Luxburg, Isabelle Guyon, and Roman Garnett (eds.), Advances in Neural Information Processing Systems (NIPS 2016), pp. 2235–2243, 2016. URL https://arxiv.org/abs/1606.04442. ", + "bbox": [ + 173, + 373, + 825, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "David Aspinall and Cezary Kaliszyk. What’s in a theorem name? In Jasmin Christian Blanchette and Stephan Merz (eds.), Interactive Theorem Proving (ITP 2016), volume 9807 of LNCS, pp. 459–465. Springer, 2016. doi: 10.1007/978-3-319-43144-4. ", + "bbox": [ + 174, + 438, + 823, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Franz Baader and Tobias Nipkow. Term rewriting and all that. Cambridge University Press, 1998. ISBN 978-0-521-45520-6. ", + "bbox": [ + 169, + 489, + 823, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jasmin C. Blanchette, Cezary Kaliszyk, Lawrence C. Paulson, and Josef Urban. Hammering towards QED. J. Formalized Reasoning, 9(1):101–148, 2016. ISSN 1972-5787. doi: 10.6092/issn. 1972-5787/4593. ", + "bbox": [ + 173, + 526, + 821, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jasmin Christian Blanchette, Maximilian P. L. Haslbeck, Daniel Matichuk, and Tobias Nipkow. Mining the Archive of Formal Proofs. In Manfred Kerber, Jacques Carette, Cezary Kaliszyk, Florian Rabe, and Volker Sorge (eds.), Intelligent Computer Mathematics (CICM 2015), volume 9150 of LNCS, pp. 3–17. Springer, 2015. ", + "bbox": [ + 173, + 577, + 825, + 635 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "James P. Bridge, Sean B. Holden, and Lawrence C. Paulson. Machine learning for first-order theorem proving - learning to select a good heuristic. J. Autom. Reasoning, 53(2):141–172, 2014. doi: 10.1007/s10817-014-9301-5. ", + "bbox": [ + 173, + 643, + 823, + 685 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Haogang Chen, Daniel Ziegler, Tej Chajed, Adam Chlipala, M. Frans Kaashoek, and Nickolai Zeldovich. Using crash Hoare logic for certifying the FSCQ file system. In Ajay Gulati and Hakim Weatherspoon (eds.), USENIX 2016. USENIX Association, 2016. ", + "bbox": [ + 173, + 694, + 825, + 737 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "François Chollet. Keras. https://github.com/fchollet/keras, 2015. ", + "bbox": [ + 173, + 746, + 598, + 761 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alonzo Church. A formulation of the simple theory of types. J. Symb. Log., 5(2):56–68, 1940. doi: 10.2307/2266170. URL http://dx.doi.org/10.2307/2266170. ", + "bbox": [ + 171, + 768, + 823, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. ", + "bbox": [ + 171, + 806, + 823, + 835 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hazel Duncan. The Use of Data-Mining for the Automatic Formation of Tactics. PhD thesis, University of Edinburgh, 2007. ", + "bbox": [ + 169, + 843, + 823, + 872 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michael Färber and Chad E. Brown. Internal guidance for Satallax. In Nicola Olivetti and Ashish Tiwari (eds.), International Joint Conference on Automated Reasoning (IJCAR 2016), volume 9706 of LNCS, pp. 349–361. Springer, 2016. doi: 10.1007/978-3-319-40229-1. ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Georges Gonthier, Andrea Asperti, Jeremy Avigad, Yves Bertot, Cyril Cohen, François Garillot, Stéphane Le Roux, Assia Mahboubi, Russell O’Connor, Sidi Ould Biha, Ioana Pasca, Laurence Rideau, Alexey Solovyev, Enrico Tassi, and Laurent Théry. A machine-checked proof of the odd order theorem. In Sandrine Blazy, Christine Paulin-Mohring, and David Pichardie (eds.), Interactive Theorem Proving (ITP 2013), volume 7998 of LNCS, pp. 163–179. Springer, 2013. ", + "bbox": [ + 173, + 103, + 826, + 174 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Michael J. C. Gordon, Robin Milner, and Christopher P. Wadsworth. Edinburgh LCF, volume 78 of Lecture Notes in Computer Science. Springer, 1979. ISBN 3-540-09724-4. doi: 10.1007/ 3-540-09724-4. URL http://dx.doi.org/10.1007/3-540-09724-4. ", + "bbox": [ + 173, + 181, + 820, + 224 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Adam Grabowski, Artur Kornilowicz, and Adam Naumowicz. Mizar in a nutshell. J. Formalized Reasoning, 3(2):153–245, 2010. doi: 10.6092/issn.1972-5787/1980. ", + "bbox": [ + 173, + 233, + 823, + 262 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Thomas Hales, John Harrison, Sean McLaughlin, Tobias Nipkow, Steven Obua, and Roland Zumkeller. A revision of the proof of the Kepler Conjecture. Discrete & Computational Geometry, 44(1):1–34, 2010. ", + "bbox": [ + 173, + 270, + 825, + 313 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Thomas C. Hales, Mark Adams, Gertrud Bauer, Dat Tat Dang, John Harrison, Truong Le Hoang, Cezary Kaliszyk, Victor Magron, Sean McLaughlin, Thang Tat Nguyen, Truong Quang Nguyen, Tobias Nipkow, Steven Obua, Joseph Pleso, Jason Rute, Alexey Solovyev, An Hoai Thi Ta, Trung Nam Tran, Diep Thi Trieu, Josef Urban, Ky Khac Vu, and Roland Zumkeller. A formal proof of the Kepler conjecture. CoRR, abs/1501.02155, 2015. ", + "bbox": [ + 173, + 320, + 825, + 391 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Harrison. HOL Light: An overview. In Stefan Berghofer, Tobias Nipkow, Christian Urban, and Makarius Wenzel (eds.), Theorem Proving in Higher Order Logics (TPHOLs 2009), volume 5674 of LNCS, pp. 60–66. Springer, 2009. ", + "bbox": [ + 174, + 398, + 821, + 443 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Harrison. The HOL Light theory of Euclidean space. J. Autom. Reasoning, 50(2):173–190, 2013. doi: 10.1007/s10817-012-9250-9. ", + "bbox": [ + 173, + 450, + 823, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Harrison, Josef Urban, and Freek Wiedijk. History of interactive theorem proving. In Jörg Siekmann (ed.), Handbook of the History of Logic vol. 9 (Computational Logic), pp. 135–214. Elsevier, 2014. ", + "bbox": [ + 173, + 487, + 825, + 530 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "R. Hindley. The principal type-scheme of an object in combinatory logic. Transactions of the american mathematical society, 146:29–60, 1969. ISSN 0002-9947. ", + "bbox": [ + 173, + 537, + 823, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 173, + 575, + 823, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kryštof Hoder and Andrei Voronkov. Sine qua non for large theory reasoning. In Nikolaj Bjørner and Viorica Sofronie-Stokkermans (eds.), CADE-23, volume 6803 of LNAI, pp. 299–314. Springer, 2011. ", + "bbox": [ + 174, + 613, + 823, + 655 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Joe Hurd. First-order proof tactics in higher-order logic theorem provers. In Myla Archer, Ben Di Vito, and César Muñoz (eds.), Design and Application of Strategies/Tactics in Higher Order Logics (STRATA 2003), number NASA/CP-2003-212448 in NASA Technical Reports, pp. 56–68, September 2003. URL http://www.gilith.com/research/papers. ", + "bbox": [ + 173, + 664, + 825, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Joe Hurd. The OpenTheory standard theory library. In Mihaela Gheorghiu Bobaru, Klaus Havelund, Gerard J. Holzmann, and Rajeev Joshi (eds.), NASA Formal Methods (NFM 2011), volume 6617 of LNCS, pp. 177–191. Springer, 2011. ", + "bbox": [ + 176, + 728, + 823, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cezary Kaliszyk and Alexander Krauss. Scalable LCF-style proof translation. In Sandrine Blazy, Christine Paulin-Mohring, and David Pichardie (eds.), Interactive Theorem Proving (ITP 2013), volume 7998 of LNCS, pp. 51–66. Springer, 2013. ", + "bbox": [ + 174, + 779, + 823, + 823 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cezary Kaliszyk and Josef Urban. Learning-assisted automated reasoning with Flyspeck. J. Autom. Reasoning, 53(2):173–213, 2014. doi: 10.1007/s10817-014-9303-3. ", + "bbox": [ + 169, + 830, + 821, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cezary Kaliszyk and Josef Urban. FEMaLeCoP: Fairly efficient machine learning connection prover. In Martin Davis, Ansgar Fehnker, Annabelle McIver, and Andrei Voronkov (eds.), 20th International Conference on Logic for Programming, Artificial Intelligence, and Reasoning (LPAR 2015), volume 9450 of LNCS, pp. 88–96. Springer, 2015a. doi: 10.1007/978-3-662-48899-7. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cezary Kaliszyk and Josef Urban. Learning-assisted theorem proving with millions of lemmas. J. Symbolic Computation, 69:109–128, 2015b. doi: 10.1016/j.jsc.2014.09.032. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cezary Kaliszyk, Josef Urban, and Jiˇrí Vyskocil. Learning to parse on aligned corpora. In Christian ˇ Urban and Xingyuan Zhang (eds.), Proc. 6h Conference on Interactive Theorem Proving (ITP’15), volume 9236 of LNCS, pp. 227–233. Springer-Verlag, 2015. doi: 10.1007/978-3-319-22102-1_ 15. ", + "bbox": [ + 173, + 145, + 825, + 200 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gerwin Klein, June Andronick, Kevin Elphinstone, Toby C. Murray, Thomas Sewell, Rafal Kolanski, and Gernot Heiser. Comprehensive formal verification of an OS microkernel. ACM Trans. Comput. Syst., 32(1):2, 2014. ", + "bbox": [ + 176, + 214, + 823, + 257 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Laura Kovács and Andrei Voronkov. First-order theorem proving and Vampire. In Natasha Sharygina and Helmut Veith (eds.), Computer-Aided Verification (CAV 2013), volume 8044 of LNCS, pp. 1–35. Springer, 2013. ", + "bbox": [ + 173, + 270, + 825, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Daniel Kühlwein and Josef Urban. MaLeS: A framework for automatic tuning of automated theorem provers. J. Autom. Reasoning, 55(2):91–116, 2015. doi: 10.1007/s10817-015-9329-1. ", + "bbox": [ + 174, + 325, + 821, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xavier Leroy. Formal verification of a realistic compiler. Commun. ACM, 52(7):107–115, 2009. ", + "bbox": [ + 171, + 366, + 803, + 382 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Steven Obua and Sebastian Skalberg. Importing HOL into Isabelle/HOL. In Ulrich Furbach and Natarajan Shankar (eds.), International Joint Conference on Automated Reasoning (IJCAR 2006), volume 4130 of LNCS, pp. 298–302. Springer, 2006. ", + "bbox": [ + 174, + 393, + 825, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lawrence C. Paulson. A generic tableau prover and its integration with Isabelle. J. Universal Computer Science, 5(3):73–87, 1999. ", + "bbox": [ + 173, + 450, + 823, + 478 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Stephan Schulz. System description: E 1.8. In Kenneth L. McMillan, Aart Middeldorp, and Andrei Voronkov (eds.), Logic for Programming, Artificial Intelligence (LPAR 2013), volume 8312 of LNCS, pp. 735–743. Springer, 2013. ", + "bbox": [ + 173, + 491, + 825, + 535 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Christopher J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of go with deep neural networks and tree search. Nature, 529:484–503, 2016. URL http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html. ", + "bbox": [ + 173, + 546, + 825, + 631 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard Socher, Danqi Chen, Christopher D. Manning, and Andrew Y. Ng. Reasoning with neural tensor networks for knowledge base completion. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Proceedings., pp. 926–934, 2013a. URL http://papers.nips.cc/paper/ 5028-reasoning-with-neural-tensor-networks-for-knowledge-base-completion. ", + "bbox": [ + 173, + 643, + 825, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y. $\\mathrm { N g }$ , and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1631–1642, Stroudsburg, PA, October 2013b. Association for Computational Linguistics. ", + "bbox": [ + 173, + 727, + 825, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. In Advances in Neural Information Processing Systems, pp. 2431–2439, 2015. ", + "bbox": [ + 173, + 810, + 823, + 839 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Geoff Sutcliffe. The CADE ATP system competition - CASC. AI Magazine, 37(2):99–101, 2016. ", + "bbox": [ + 171, + 852, + 812, + 867 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Josef Urban, Jiˇrí Vyskocil, and Petr Št ˇ epánek. MaLeCoP: Machine learning connection prover. In ˇ Kai Brünnler and George Metcalfe (eds.), TABLEAUX 2011, volume 6793 of LNCS. Springer, 2011. ", + "bbox": [ + 174, + 881, + 823, + 922 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Lukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural machine translation system: Bridging the gap between human and machine translation. CoRR, abs/1609.08144, 2016. URL http://arxiv.org/abs/1609.08144. ", + "bbox": [ + 176, + 103, + 825, + 200 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/ryuxYmvel/ryuxYmvel_middle.json b/parse/train/ryuxYmvel/ryuxYmvel_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..927115d130add7730b04952cb04f4559eca199fc --- /dev/null +++ b/parse/train/ryuxYmvel/ryuxYmvel_middle.json @@ -0,0 +1,25799 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "score": 1.0, + "content": "HOLSTEP: A MACHINE LEARNING DATASET FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 430, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 430, + 118 + ], + "score": 1.0, + "content": "HIGHER-ORDER LOGIC THEOREM PROVING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 269, + 168 + ], + "lines": [ + { + "bbox": [ + 112, + 135, + 185, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 135, + 185, + 147 + ], + "score": 1.0, + "content": "Cezary Kaliszyk", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 210, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 210, + 158 + ], + "score": 1.0, + "content": "University of Innsbruck", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 158, + 270, + 170 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 270, + 170 + ], + "score": 1.0, + "content": "cezary.kaliszyk@uibk.ac.at", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 302, + 135, + 477, + 168 + ], + "lines": [ + { + "bbox": [ + 301, + 133, + 456, + 148 + ], + "spans": [ + { + "bbox": [ + 301, + 133, + 456, + 148 + ], + "score": 1.0, + "content": "François Chollet, Christian Szegedy", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 147, + 372, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 147, + 372, + 157 + ], + "score": 1.0, + "content": "Google Research", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 157, + 477, + 169 + ], + "spans": [ + { + "bbox": [ + 302, + 157, + 477, + 169 + ], + "score": 1.0, + "content": "{fchollet,szegedy}@google.com", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 222, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 470, + 235 + ], + "score": 1.0, + "content": "Large computer-understandable proofs consist of millions of intermediate logical", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "steps. The vast majority of such steps originate from manually selected and man-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "ually guided heuristics applied to intermediate goals. So far, machine learning has", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "generally not been used to filter or generate these steps. In this paper, we introduce", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 266, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 140, + 266, + 470, + 280 + ], + "score": 1.0, + "content": "a new dataset based on Higher-Order Logic (HOL) proofs, for the purpose of de-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 277, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 291 + ], + "score": 1.0, + "content": "veloping new machine learning-based theorem-proving strategies. We make this", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "score": 1.0, + "content": "dataset publicly available under the BSD license. We propose various machine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "score": 1.0, + "content": "learning tasks that can be performed on this dataset, and discuss their significance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "score": 1.0, + "content": "for theorem proving. We also benchmark a set of simple baseline machine learn-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "ing models suited for the tasks (including logistic regression, convolutional neural", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "networks and recurrent neural networks). The results of our baseline models show", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 342, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 416, + 357 + ], + "score": 1.0, + "content": "the promise of applying machine learning to HOL theorem proving.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 376, + 206, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "As the usability of interactive theorem proving (ITP) systems (Harrison et al., 2014) grows, its use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "score": 1.0, + "content": "becomes a more common way of establishing the correctness of software as well as mathematical", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "proofs. Today, ITPs are used for software certification projects ranging from compilers (Leroy,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "2009) and operating system components (Chen et al., 2016; Klein et al., 2014), to establishing the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "absolute correctness of large proofs in mathematics such as the Kepler conjecture (Hales et al., 2015)", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 457, + 329, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 329, + 468 + ], + "score": 1.0, + "content": "and the Feit-Thomson Theorem (Gonthier et al., 2013).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "For results of such significance to be possible, the theorem libraries of these ITPs must contain", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "all necessary basic mathematical properties, accompanied with formal proofs. This means that the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 496, + 504, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 504, + 507 + ], + "score": 1.0, + "content": "size of many ITP libraries can be measured in dozens of thousands of theorems (Grabowski et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "score": 1.0, + "content": "2010; Blanchette et al., 2015) and billions of individual proof steps. While the general direction of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "the proofs is specified by humans (by providing the goal to prove, specifying intermediate steps, or", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "applying certain automated tactics), the majority of such proof steps are actually found by automated", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "reasoning-based proof search (Kaliszyk & Urban, 2015b), with very little application of machine", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 214, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 214, + 563 + ], + "score": 1.0, + "content": "learning techniques so far.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "score": 1.0, + "content": "At the same time, fast progress has been unfolding in machine learning applied to tasks that involve", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "logical inference, such as natural language question answering (Sukhbaatar et al., 2015), knowledge", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 589, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 504, + 601 + ], + "score": 1.0, + "content": "base completion (Socher et al., 2013a), automated translation (Wu et al., 2016), and premise selec-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "tion in the context of theorem proving (Alemi et al., 2016). Deep learning in particular has proven", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "to be a powerful tool for embedding semantic meaning and logical relationships into geometric", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "spaces, specifically via models such as convolutional neural networks, recurrent neural networks,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "and tree-recursive neural networks. These advances strongly suggest that deep learning may have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "score": 1.0, + "content": "become mature enough to yield significant advances in automated theorem proving. Remarkably,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "it has recently become possible to build a system, AlphaGo (Silver et al., 2016), blending classi-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "cal AI techniques such as Monte-Carlo tree search and modern deep learning techniques, capable", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "of playing the game of Go at super-human levels. We should note that theorem proving and Go", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "playing are conceptually related, since both consist in searching for specific nodes in trees of states", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "with extremely large arity and relatively large depth, which involves node evaluation decision (how", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "valuable is this state?) and policy decisions (which node should be expanded next?). The success of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "AlphaGo can thus serve as encouragement on the road to building deep learning-augmented theorem", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 43 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 97 + ], + "score": 1.0, + "content": "HOLSTEP: A MACHINE LEARNING DATASET FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 430, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 430, + 118 + ], + "score": 1.0, + "content": "HIGHER-ORDER LOGIC THEOREM PROVING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 113, + 135, + 269, + 168 + ], + "lines": [ + { + "bbox": [ + 112, + 135, + 185, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 135, + 185, + 147 + ], + "score": 1.0, + "content": "Cezary Kaliszyk", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 146, + 210, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 210, + 158 + ], + "score": 1.0, + "content": "University of Innsbruck", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 158, + 270, + 170 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 270, + 170 + ], + "score": 1.0, + "content": "cezary.kaliszyk@uibk.ac.at", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 112, + 135, + 270, + 170 + ] + }, + { + "type": "text", + "bbox": [ + 302, + 135, + 477, + 168 + ], + "lines": [ + { + "bbox": [ + 301, + 133, + 456, + 148 + ], + "spans": [ + { + "bbox": [ + 301, + 133, + 456, + 148 + ], + "score": 1.0, + "content": "François Chollet, Christian Szegedy", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 147, + 372, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 147, + 372, + 157 + ], + "score": 1.0, + "content": "Google Research", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 157, + 477, + 169 + ], + "spans": [ + { + "bbox": [ + 302, + 157, + 477, + 169 + ], + "score": 1.0, + "content": "{fchollet,szegedy}@google.com", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 301, + 133, + 477, + 169 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 222, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 470, + 235 + ], + "score": 1.0, + "content": "Large computer-understandable proofs consist of millions of intermediate logical", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 246 + ], + "score": 1.0, + "content": "steps. The vast majority of such steps originate from manually selected and man-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "ually guided heuristics applied to intermediate goals. So far, machine learning has", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "generally not been used to filter or generate these steps. In this paper, we introduce", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 266, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 140, + 266, + 470, + 280 + ], + "score": 1.0, + "content": "a new dataset based on Higher-Order Logic (HOL) proofs, for the purpose of de-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 277, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 291 + ], + "score": 1.0, + "content": "veloping new machine learning-based theorem-proving strategies. We make this", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 288, + 470, + 301 + ], + "score": 1.0, + "content": "dataset publicly available under the BSD license. We propose various machine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "score": 1.0, + "content": "learning tasks that can be performed on this dataset, and discuss their significance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 323 + ], + "score": 1.0, + "content": "for theorem proving. We also benchmark a set of simple baseline machine learn-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 470, + 334 + ], + "score": 1.0, + "content": "ing models suited for the tasks (including logistic regression, convolutional neural", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 333, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 469, + 344 + ], + "score": 1.0, + "content": "networks and recurrent neural networks). The results of our baseline models show", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 342, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 416, + 357 + ], + "score": 1.0, + "content": "the promise of applying machine learning to HOL theorem proving.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 14.5, + "bbox_fs": [ + 140, + 222, + 470, + 357 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 376, + 206, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "As the usability of interactive theorem proving (ITP) systems (Harrison et al., 2014) grows, its use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "score": 1.0, + "content": "becomes a more common way of establishing the correctness of software as well as mathematical", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "proofs. Today, ITPs are used for software certification projects ranging from compilers (Leroy,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "2009) and operating system components (Chen et al., 2016; Klein et al., 2014), to establishing the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "absolute correctness of large proofs in mathematics such as the Kepler conjecture (Hales et al., 2015)", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 457, + 329, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 329, + 468 + ], + "score": 1.0, + "content": "and the Feit-Thomson Theorem (Gonthier et al., 2013).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 401, + 506, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "For results of such significance to be possible, the theorem libraries of these ITPs must contain", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "all necessary basic mathematical properties, accompanied with formal proofs. This means that the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 496, + 504, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 504, + 507 + ], + "score": 1.0, + "content": "size of many ITP libraries can be measured in dozens of thousands of theorems (Grabowski et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 518 + ], + "score": 1.0, + "content": "2010; Blanchette et al., 2015) and billions of individual proof steps. While the general direction of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "the proofs is specified by humans (by providing the goal to prove, specifying intermediate steps, or", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "applying certain automated tactics), the majority of such proof steps are actually found by automated", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "reasoning-based proof search (Kaliszyk & Urban, 2015b), with very little application of machine", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 214, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 214, + 563 + ], + "score": 1.0, + "content": "learning techniques so far.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 474, + 506, + 563 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 580 + ], + "score": 1.0, + "content": "At the same time, fast progress has been unfolding in machine learning applied to tasks that involve", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "logical inference, such as natural language question answering (Sukhbaatar et al., 2015), knowledge", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 589, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 504, + 601 + ], + "score": 1.0, + "content": "base completion (Socher et al., 2013a), automated translation (Wu et al., 2016), and premise selec-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "tion in the context of theorem proving (Alemi et al., 2016). Deep learning in particular has proven", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "to be a powerful tool for embedding semantic meaning and logical relationships into geometric", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "spaces, specifically via models such as convolutional neural networks, recurrent neural networks,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "and tree-recursive neural networks. These advances strongly suggest that deep learning may have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "score": 1.0, + "content": "become mature enough to yield significant advances in automated theorem proving. Remarkably,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "it has recently become possible to build a system, AlphaGo (Silver et al., 2016), blending classi-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "cal AI techniques such as Monte-Carlo tree search and modern deep learning techniques, capable", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "of playing the game of Go at super-human levels. We should note that theorem proving and Go", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "playing are conceptually related, since both consist in searching for specific nodes in trees of states", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "with extremely large arity and relatively large depth, which involves node evaluation decision (how", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "valuable is this state?) and policy decisions (which node should be expanded next?). The success of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "AlphaGo can thus serve as encouragement on the road to building deep learning-augmented theorem", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "provers that would blend classical techniques developed over the past few decades with the latest", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 219, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 219, + 106 + ], + "score": 1.0, + "content": "machine learning advances.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 566, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "provers that would blend classical techniques developed over the past few decades with the latest", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 219, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 219, + 106 + ], + "score": 1.0, + "content": "machine learning advances.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "Fast progress in specific machine learning verticals has occasionally been achieved thanks to the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "release of specialized datasets (often with associated competitions, e.g. the ImageNet dataset for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "large-scale image classification (Deng et al., 2009)) serving as an experimental testbed and public", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "benchmark of current progress, thus focusing the efforts of the research community. We hope that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "score": 1.0, + "content": "releasing a theorem proving dataset suited for specific machine learning tasks can serve the same", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 399, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 399, + 180 + ], + "score": 1.0, + "content": "purpose in the vertical of applying machine learning to theorem proving.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 109, + 196, + 268, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 270, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 270, + 209 + ], + "score": 1.0, + "content": "1.1 CONTRIBUTION AND OVERVIEW", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 219, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "First, we develop a dataset for machine learning based on the proof steps used in a large interactive", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "proof section 2. We focus on the HOL Light (Harrison, 2009) ITP, its multivariate analysis library", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "(Harrison, 2013), as well as the formal proof of the Kepler conjecture (Hales et al., 2010). These for-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 252, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 267 + ], + "score": 1.0, + "content": "malizations constitute a diverse proof dataset containing basic mathematics, analysis, trigonometry,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "as well as reasoning about data structures such as graphs. Furthermore these formal proof devel-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 274, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 274, + 506, + 288 + ], + "score": 1.0, + "content": "opments have been used as benchmarks for automated reasoning techniques (Kaliszyk & Urban,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 136, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 136, + 298 + ], + "score": 1.0, + "content": "2014).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 401 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "The dataset consists of 2,013,046 training examples and 196,030 testing examples that originate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "from 11,400 proofs. Precisely half of the examples are statements that were useful in the currently", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 336 + ], + "score": 1.0, + "content": "proven conjectures and half are steps that have been derived either manually or as part of the auto-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "mated proof search but were not necessary in the final proofs. The dataset contains only proofs of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "non-trivial theorems, that also do not focus on computation but rather on actual theorem proving.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "For each proof, the conjecture that is being proven as well as its dependencies (axioms) and may", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "be exploited in machine learning tasks. Furthermore, for each statement both its human-readable", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "(pretty-printed) statement and a tokenization designed to make machine learning tasks more man-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 390, + 194, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 194, + 402 + ], + "score": 1.0, + "content": "ageable are included.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Next, in section 3 we discuss the proof step classification tasks that can be attempted using the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "dataset, and we discuss the usefulness of these tasks in interactive and automated theorem proving.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "These tasks include unconditioned classification (without access to conjectures and dependencies)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 439, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 454 + ], + "score": 1.0, + "content": "and conjecture-conditioned classification (with access to the conjecture) of proof steps as being", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 450, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 464 + ], + "score": 1.0, + "content": "useful or not in a proof. We outline the use of such classification capabilities for search space", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "pruning and internal guidance, as well as for generation of intermediate steps or possible new lemma", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 474, + 154, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 154, + 484 + ], + "score": 1.0, + "content": "statements.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 501 + ], + "score": 1.0, + "content": "Finally, in section 4 we propose three baseline models for the proof step classification tasks, and we", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "experimentally evaluate the models on the data in section 5. The models considered include both", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "a relatively simple regression model, as well as deep learning models based on convolutional and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 524, + 213, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 213, + 534 + ], + "score": 1.0, + "content": "recurrent neural networks.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 108, + 554, + 203, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 205, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 205, + 567 + ], + "score": 1.0, + "content": "1.2 RELATED WORK", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "The use of machine learning in interactive and automated theorem proving has so far focused on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 588, + 497, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 497, + 601 + ], + "score": 1.0, + "content": "three tasks: premise selection, strategy selection, and internal guidance. We shortly explain these.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "Given a large library of proven facts and a user given conjecture, the multi-label classification prob-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "lem of selecting the facts that are most likely to lead to a successful proof of the conjecture has", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "been usually called relevance filtering or premise selection (Alama et al., 2014). This is crucial for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "score": 1.0, + "content": "the efficiency of modern automation techniques for ITPs (Blanchette et al., 2016), which today can", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 162, + 661 + ], + "score": 1.0, + "content": "usually solve", + "type": "text" + }, + { + "bbox": [ + 162, + 649, + 197, + 659 + ], + "score": 0.88, + "content": "4 0 { - } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "of the conjectures in theorem proving libraries. Similarly most competitive", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 456, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 456, + 672 + ], + "score": 1.0, + "content": "ATPs today (Sutcliffe, 2016) implement the SInE classifier (Hoder & Voronkov, 2011).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "A second theorem proving task where machine learning has been of importance is strategy selection.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "With the development of automated theorem provers came many parameters that control their exe-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "cution. In fact, modern ATPs, such as E (Schulz, 2013) and Vampire (Kovács & Voronkov, 2013),", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 707, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 707, + 506, + 724 + ], + "score": 1.0, + "content": "include complete strategy description languages that allow a user to specify the orderings, weighting", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "functions, literal selection strategies, etc. Rather than optimizing the search strategy globally, one", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 506, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "Fast progress in specific machine learning verticals has occasionally been achieved thanks to the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "release of specialized datasets (often with associated competitions, e.g. the ImageNet dataset for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "large-scale image classification (Deng et al., 2009)) serving as an experimental testbed and public", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "benchmark of current progress, thus focusing the efforts of the research community. We hope that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "score": 1.0, + "content": "releasing a theorem proving dataset suited for specific machine learning tasks can serve the same", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 399, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 399, + 180 + ], + "score": 1.0, + "content": "purpose in the vertical of applying machine learning to theorem proving.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 110, + 505, + 180 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 196, + 268, + 208 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 270, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 270, + 209 + ], + "score": 1.0, + "content": "1.1 CONTRIBUTION AND OVERVIEW", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 219, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "First, we develop a dataset for machine learning based on the proof steps used in a large interactive", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "proof section 2. We focus on the HOL Light (Harrison, 2009) ITP, its multivariate analysis library", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "(Harrison, 2013), as well as the formal proof of the Kepler conjecture (Hales et al., 2010). These for-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 252, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 267 + ], + "score": 1.0, + "content": "malizations constitute a diverse proof dataset containing basic mathematics, analysis, trigonometry,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "as well as reasoning about data structures such as graphs. Furthermore these formal proof devel-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 274, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 274, + 506, + 288 + ], + "score": 1.0, + "content": "opments have been used as benchmarks for automated reasoning techniques (Kaliszyk & Urban,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 136, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 136, + 298 + ], + "score": 1.0, + "content": "2014).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 104, + 220, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 401 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "The dataset consists of 2,013,046 training examples and 196,030 testing examples that originate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "from 11,400 proofs. Precisely half of the examples are statements that were useful in the currently", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 336 + ], + "score": 1.0, + "content": "proven conjectures and half are steps that have been derived either manually or as part of the auto-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "mated proof search but were not necessary in the final proofs. The dataset contains only proofs of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "non-trivial theorems, that also do not focus on computation but rather on actual theorem proving.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "For each proof, the conjecture that is being proven as well as its dependencies (axioms) and may", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "be exploited in machine learning tasks. Furthermore, for each statement both its human-readable", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "(pretty-printed) statement and a tokenization designed to make machine learning tasks more man-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 390, + 194, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 194, + 402 + ], + "score": 1.0, + "content": "ageable are included.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 302, + 506, + 402 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Next, in section 3 we discuss the proof step classification tasks that can be attempted using the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "dataset, and we discuss the usefulness of these tasks in interactive and automated theorem proving.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "These tasks include unconditioned classification (without access to conjectures and dependencies)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 439, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 454 + ], + "score": 1.0, + "content": "and conjecture-conditioned classification (with access to the conjecture) of proof steps as being", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 450, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 464 + ], + "score": 1.0, + "content": "useful or not in a proof. We outline the use of such classification capabilities for search space", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "pruning and internal guidance, as well as for generation of intermediate steps or possible new lemma", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 474, + 154, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 154, + 484 + ], + "score": 1.0, + "content": "statements.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 406, + 506, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 501 + ], + "score": 1.0, + "content": "Finally, in section 4 we propose three baseline models for the proof step classification tasks, and we", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "experimentally evaluate the models on the data in section 5. The models considered include both", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "a relatively simple regression model, as well as deep learning models based on convolutional and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 524, + 213, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 213, + 534 + ], + "score": 1.0, + "content": "recurrent neural networks.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 490, + 506, + 534 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 554, + 203, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 205, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 205, + 567 + ], + "score": 1.0, + "content": "1.2 RELATED WORK", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "The use of machine learning in interactive and automated theorem proving has so far focused on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 588, + 497, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 497, + 601 + ], + "score": 1.0, + "content": "three tasks: premise selection, strategy selection, and internal guidance. We shortly explain these.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 576, + 505, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "Given a large library of proven facts and a user given conjecture, the multi-label classification prob-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "lem of selecting the facts that are most likely to lead to a successful proof of the conjecture has", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "been usually called relevance filtering or premise selection (Alama et al., 2014). This is crucial for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 651 + ], + "score": 1.0, + "content": "the efficiency of modern automation techniques for ITPs (Blanchette et al., 2016), which today can", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 162, + 661 + ], + "score": 1.0, + "content": "usually solve", + "type": "text" + }, + { + "bbox": [ + 162, + 649, + 197, + 659 + ], + "score": 0.88, + "content": "4 0 { - } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "of the conjectures in theorem proving libraries. Similarly most competitive", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 456, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 456, + 672 + ], + "score": 1.0, + "content": "ATPs today (Sutcliffe, 2016) implement the SInE classifier (Hoder & Voronkov, 2011).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 604, + 506, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "A second theorem proving task where machine learning has been of importance is strategy selection.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "With the development of automated theorem provers came many parameters that control their exe-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "cution. In fact, modern ATPs, such as E (Schulz, 2013) and Vampire (Kovács & Voronkov, 2013),", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 707, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 104, + 707, + 506, + 724 + ], + "score": 1.0, + "content": "include complete strategy description languages that allow a user to specify the orderings, weighting", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "functions, literal selection strategies, etc. Rather than optimizing the search strategy globally, one", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "can choose the strategy based on the currently considered problem. For this some frameworks use", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 370, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 370, + 106 + ], + "score": 1.0, + "content": "machine learning (Bridge et al., 2014; Kühlwein & Urban, 2015).", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 47, + "bbox_fs": [ + 104, + 677, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "can choose the strategy based on the currently considered problem. For this some frameworks use", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 370, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 370, + 106 + ], + "score": 1.0, + "content": "machine learning (Bridge et al., 2014; Kühlwein & Urban, 2015).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "Finally, an automated theorem prover may use machine learning for choosing the actual inference", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "steps. It has been shown to significantly reduce the proof search in first-order tableaux by the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "selection of extension steps to use (Urban et al., 2011), and has been also successfully applied", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "in monomorphic higher-order logic proving (Färber & Brown, 2016). Data/proof mining has also", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "been applied on the level of interactive theorem proving tactics (Duncan, 2007) to extract and reuse", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 167, + 182, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 182, + 178 + ], + "score": 1.0, + "content": "repeating patterns.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 108, + 198, + 244, + 210 + ], + "lines": [ + { + "bbox": [ + 104, + 196, + 246, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 246, + 213 + ], + "score": 1.0, + "content": "2 DATASET EXTRACTION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "We focus on the HOL Light theorem prover for two reasons. First, it follows the LCF approach1).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "This means that complicated inferences are reduced to the most primitive ones and the data extrac-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "tion related modifications can be restricted the primitive inferences and it is relatively easy to extract", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "proof steps at an arbitrary selected level of granularity. Second, HOL Light implements higher-order", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "logic (Church, 1940) as its foundation, which on the one hand is powerful enough to encode most", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "of today’s formal proofs, and on the other hand allows for an easy integration of many powerful", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 376, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 376, + 303 + ], + "score": 1.0, + "content": "automation mechanisms (Baader & Nipkow, 1998; Paulson, 1999).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "score": 1.0, + "content": "When selecting the theorems to record, we choose an intermediate approach between HOL Light", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "ProofRecording (Obua & Skalberg, 2006) and the HOL/Import one (Kaliszyk & Krauss, 2013). The", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "theorems that are derived by most common proof functions are extracted by patching these functions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "like in the former approach, and the remaining theorems are extracted from the underlying OCaml", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "score": 1.0, + "content": "programming language interpreter. In certain cases decision procedures derive theorems to be reused", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "in subsequent invocations. We detect such values by looking at theorems used across proof blocks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 322, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 322, + 387 + ], + "score": 1.0, + "content": "and avoid extracting such reused unrelated subproofs.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "All kernel-level inferences are recorded together with their respective arguments in a trace file. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "score": 1.0, + "content": "trace is processed offline to extract the dependencies of the facts, detect used proof boundaries,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "mark the used and unused steps, and mark the training and testing examples. Only proofs that have", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "sufficiently many used and unused steps are considered useful for the dataset. The annotated proof", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "trace is processed again by a HOL kernel saving the actual training and testing examples originating", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "from non-trivial reasoning steps. Training and testing examples are grouped by proof: for each proof", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "the conjecture (statement that is finally proved), the dependencies of the theorem are constant, and a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "list of used and not used intermediate statements is provided. This means that the conjectures used", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 316, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 316, + 492 + ], + "score": 1.0, + "content": "in the training and testing sets are normally disjoint.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "For each statement, whether it is the conjecture, a proof dependency, or an intermediate statement,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "both a fully parenthesised HOL Light human-like printout is provided, as well as a predefined to-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "kenization. The standard HOL Light printer uses parentheses and operator priorities to make its", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "notations somewhat similar to textbook-style mathematics, while at the same time preserving the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 538, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 553 + ], + "score": 1.0, + "content": "complete unambiguity of the order of applications (this is particularly visible for associative opera-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "tors). The tokenization that we propose attempts to reduce the number of parentheses. To do this we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "compute the maximum number of arguments that each symbol needs to be applied to, and only mark", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 391, + 586 + ], + "score": 1.0, + "content": "partial application. This means that fully applied functions (more than", + "type": "text" + }, + { + "bbox": [ + 391, + 573, + 411, + 583 + ], + "score": 0.86, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "of the applications) do", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 584, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 595 + ], + "score": 1.0, + "content": "not require neither application operators nor parentheses. Top-level universal quantifications are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "eliminated, bound variables are represented by their de Bruijn indices (the distance from the corre-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "sponding abstraction in the parse tree of the term) and free variables are renamed canonically. Since", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "score": 1.0, + "content": "the Hindley-Milner type inference Hindley (1969) mechanisms will be sufficient to reconstruct the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "most-general types of the expressions well enough for automated-reasoning techniques Kaliszyk", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "et al. (2015) we erase all type information. Table 1 presents some dataset statistics. The dataset, the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "description of the used format, the scripts used to generate it and baseline models code are available:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 221, + 663, + 391, + 674 + ], + "lines": [ + { + "bbox": [ + 218, + 661, + 392, + 676 + ], + "spans": [ + { + "bbox": [ + 218, + 661, + 392, + 676 + ], + "score": 1.0, + "content": "http://cl-informatik.uibk.ac.at/cek/holstep/", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 691, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 689, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 119, + 689, + 505, + 704 + ], + "score": 1.0, + "content": "1LCF approach is a software architecture for implementing theorem provers which uses a strongly typed", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "programming language with abstract datatypes (such as OCaml in the case of HOL Light) to separate the small", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 712, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 722 + ], + "score": 1.0, + "content": "trusted core, called the kernel, which verifies the primitive inferences from user code which allows the user to", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 429, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 429, + 733 + ], + "score": 1.0, + "content": "arbitrarily extend the system in a safe manner. For more details see (Gordon et al., 1979).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "Finally, an automated theorem prover may use machine learning for choosing the actual inference", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "steps. It has been shown to significantly reduce the proof search in first-order tableaux by the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "selection of extension steps to use (Urban et al., 2011), and has been also successfully applied", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "in monomorphic higher-order logic proving (Färber & Brown, 2016). Data/proof mining has also", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "been applied on the level of interactive theorem proving tactics (Duncan, 2007) to extract and reuse", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 167, + 182, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 182, + 178 + ], + "score": 1.0, + "content": "repeating patterns.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 110, + 505, + 178 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 198, + 244, + 210 + ], + "lines": [ + { + "bbox": [ + 104, + 196, + 246, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 246, + 213 + ], + "score": 1.0, + "content": "2 DATASET EXTRACTION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "We focus on the HOL Light theorem prover for two reasons. First, it follows the LCF approach1).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "This means that complicated inferences are reduced to the most primitive ones and the data extrac-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "tion related modifications can be restricted the primitive inferences and it is relatively easy to extract", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "proof steps at an arbitrary selected level of granularity. Second, HOL Light implements higher-order", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "logic (Church, 1940) as its foundation, which on the one hand is powerful enough to encode most", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "of today’s formal proofs, and on the other hand allows for an easy integration of many powerful", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 376, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 376, + 303 + ], + "score": 1.0, + "content": "automation mechanisms (Baader & Nipkow, 1998; Paulson, 1999).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 226, + 505, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "score": 1.0, + "content": "When selecting the theorems to record, we choose an intermediate approach between HOL Light", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "ProofRecording (Obua & Skalberg, 2006) and the HOL/Import one (Kaliszyk & Krauss, 2013). The", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "theorems that are derived by most common proof functions are extracted by patching these functions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "like in the former approach, and the remaining theorems are extracted from the underlying OCaml", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "score": 1.0, + "content": "programming language interpreter. In certain cases decision procedures derive theorems to be reused", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "in subsequent invocations. We detect such values by looking at theorems used across proof blocks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 322, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 322, + 387 + ], + "score": 1.0, + "content": "and avoid extracting such reused unrelated subproofs.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 308, + 506, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "All kernel-level inferences are recorded together with their respective arguments in a trace file. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "score": 1.0, + "content": "trace is processed offline to extract the dependencies of the facts, detect used proof boundaries,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "mark the used and unused steps, and mark the training and testing examples. Only proofs that have", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "sufficiently many used and unused steps are considered useful for the dataset. The annotated proof", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "trace is processed again by a HOL kernel saving the actual training and testing examples originating", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "from non-trivial reasoning steps. Training and testing examples are grouped by proof: for each proof", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "the conjecture (statement that is finally proved), the dependencies of the theorem are constant, and a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "list of used and not used intermediate statements is provided. This means that the conjectures used", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 316, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 316, + 492 + ], + "score": 1.0, + "content": "in the training and testing sets are normally disjoint.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 391, + 506, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "For each statement, whether it is the conjecture, a proof dependency, or an intermediate statement,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "both a fully parenthesised HOL Light human-like printout is provided, as well as a predefined to-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "kenization. The standard HOL Light printer uses parentheses and operator priorities to make its", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "notations somewhat similar to textbook-style mathematics, while at the same time preserving the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 538, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 553 + ], + "score": 1.0, + "content": "complete unambiguity of the order of applications (this is particularly visible for associative opera-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "tors). The tokenization that we propose attempts to reduce the number of parentheses. To do this we", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "compute the maximum number of arguments that each symbol needs to be applied to, and only mark", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 391, + 586 + ], + "score": 1.0, + "content": "partial application. This means that fully applied functions (more than", + "type": "text" + }, + { + "bbox": [ + 391, + 573, + 411, + 583 + ], + "score": 0.86, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "of the applications) do", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 584, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 595 + ], + "score": 1.0, + "content": "not require neither application operators nor parentheses. Top-level universal quantifications are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "eliminated, bound variables are represented by their de Bruijn indices (the distance from the corre-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "sponding abstraction in the parse tree of the term) and free variables are renamed canonically. Since", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "score": 1.0, + "content": "the Hindley-Milner type inference Hindley (1969) mechanisms will be sufficient to reconstruct the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "most-general types of the expressions well enough for automated-reasoning techniques Kaliszyk", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "et al. (2015) we erase all type information. Table 1 presents some dataset statistics. The dataset, the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "description of the used format, the scripts used to generate it and baseline models code are available:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 495, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 221, + 663, + 391, + 674 + ], + "lines": [ + { + "bbox": [ + 218, + 661, + 392, + 676 + ], + "spans": [ + { + "bbox": [ + 218, + 661, + 392, + 676 + ], + "score": 1.0, + "content": "http://cl-informatik.uibk.ac.at/cek/holstep/", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47, + "bbox_fs": [ + 218, + 661, + 392, + 676 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 80, + 454, + 159 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 156, + 80, + 454, + 159 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 80, + 454, + 159 + ], + "spans": [ + { + "bbox": [ + 156, + 80, + 454, + 159 + ], + "score": 0.981, + "html": "
TrainTestPositiveNegative
Examples201304619603011045381104538
Avg. length503.18440.20535.52459.66
Avg. tokens87.0180.6295.4877.40
Conjectures99991411--
Avg. dependencies29.5822.8211
", + "type": "table", + "image_path": "6073529efc3cc681e06d1137becf13db8721e1be8599e3757798b48405083e31.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 156, + 80, + 454, + 106.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 156, + 106.33333333333333, + 454, + 132.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 156, + 132.66666666666666, + 454, + 159.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 238, + 167, + 374, + 179 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 236, + 166, + 374, + 180 + ], + "spans": [ + { + "bbox": [ + 236, + 166, + 374, + 180 + ], + "score": 1.0, + "content": "Table 1: HolStep dataset statistics", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 108, + 201, + 272, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 273, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 273, + 216 + ], + "score": 1.0, + "content": "3 MACHINE LEARNING TASKS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 228, + 220, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 221, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 221, + 241 + ], + "score": 1.0, + "content": "3.1 TASKS DESCRIPTION", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 249, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 264 + ], + "score": 1.0, + "content": "This dataset makes possible several tasks well-suited for machine learning most of which are highly", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 261, + 225, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 225, + 275 + ], + "score": 1.0, + "content": "relevant for theorem proving:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 133, + 285, + 464, + 406 + ], + "lines": [ + { + "bbox": [ + 131, + 284, + 443, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 443, + 300 + ], + "score": 1.0, + "content": "• Predicting whether a statement is useful in the proof of a given conjecture;", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 303, + 420, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 420, + 317 + ], + "score": 1.0, + "content": "• Predicting the dependencies of a proof statement (premise selection);", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 322, + 415, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 415, + 335 + ], + "score": 1.0, + "content": "• Predicting whether a statement is an important one (human named);", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 339, + 464, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 464, + 354 + ], + "score": 1.0, + "content": "• Predicting which conjecture a particular intermediate statement originates from;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 357, + 309, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 309, + 371 + ], + "score": 1.0, + "content": "• Predicting the name given to a statement;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 375, + 451, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 451, + 389 + ], + "score": 1.0, + "content": "• Generating intermediate statements useful in the proof of a given conjecture;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 394, + 367, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 367, + 407 + ], + "score": 1.0, + "content": "• Generating the conjecture the current proof will lead to.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 504, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "In what follows we focus on the first task: classifying proof step statements as being useful or not in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 428, + 465, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 465, + 441 + ], + "score": 1.0, + "content": "the context of a given proof. This task may be further specialized into two different tasks:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 132, + 452, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 132, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "• Unconditioned classification of proof steps: determining how likely a given proof is to be", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 141, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "useful for the proof it occurred in, based solely on the content of statement (i.e. by only", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 475, + 426, + 487 + ], + "spans": [ + { + "bbox": [ + 142, + 475, + 426, + 487 + ], + "score": 1.0, + "content": "providing the model with the step statement itself, absent any context).", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 137, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "Conditioned classification of proof steps: determining how likely a given proof is to be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 141, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "useful for the proof it occurred in, with “conditioning” on the conjecture statement that the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 515, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 141, + 515, + 506, + 527 + ], + "score": 1.0, + "content": "proof was aiming to attain, i.e. by providing the model with both the step statement and the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 527, + 232, + 538 + ], + "spans": [ + { + "bbox": [ + 142, + 527, + 232, + 538 + ], + "score": 1.0, + "content": "conjecture statement).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 550, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "In the dataset, for every proof we provide the same number of useful and non-useful steps. As such,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "the proof step classification problem is a balanced two-class classification problem, where a random", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 572, + 269, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 269, + 584 + ], + "score": 1.0, + "content": "baseline would yield an accuracy of 0.5.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 108, + 601, + 421, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 422, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 422, + 613 + ], + "score": 1.0, + "content": "3.2 RELEVANCE TO INTERACTIVE AND AUTOMATED THEOREM PROVING", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "In the interaction with an interactive theorem prover, the tasks that require most human time are: the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "search for good intermediate steps; the search for automation techniques able to justify the individual", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "steps, and searching theorem proving libraries for the necessary simpler facts. These three problems", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "directly correspond to the machine learning tasks proposed in the previous subsection. Being able", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "to predict the usefulness of a statement will significantly improve many automation techniques.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "The generation of good intermediate lemmas or intermediate steps can improve level of granularity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "of the proof steps. Understanding the correspondence between statements and their names can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "allow users to search for statements in the libraries more efficiently (Aspinall & Kaliszyk, 2016).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Premise selection and filtering are already used in many theorem proving systems, and generation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 721, + 394, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 394, + 733 + ], + "score": 1.0, + "content": "of succeeding steps corresponds to conjecturing and theory exploration.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 80, + 454, + 159 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 156, + 80, + 454, + 159 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 80, + 454, + 159 + ], + "spans": [ + { + "bbox": [ + 156, + 80, + 454, + 159 + ], + "score": 0.981, + "html": "
TrainTestPositiveNegative
Examples201304619603011045381104538
Avg. length503.18440.20535.52459.66
Avg. tokens87.0180.6295.4877.40
Conjectures99991411--
Avg. dependencies29.5822.8211
", + "type": "table", + "image_path": "6073529efc3cc681e06d1137becf13db8721e1be8599e3757798b48405083e31.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 156, + 80, + 454, + 106.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 156, + 106.33333333333333, + 454, + 132.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 156, + 132.66666666666666, + 454, + 159.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 238, + 167, + 374, + 179 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 236, + 166, + 374, + 180 + ], + "spans": [ + { + "bbox": [ + 236, + 166, + 374, + 180 + ], + "score": 1.0, + "content": "Table 1: HolStep dataset statistics", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 108, + 201, + 272, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 273, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 273, + 216 + ], + "score": 1.0, + "content": "3 MACHINE LEARNING TASKS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 228, + 220, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 228, + 221, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 221, + 241 + ], + "score": 1.0, + "content": "3.1 TASKS DESCRIPTION", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 249, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 264 + ], + "score": 1.0, + "content": "This dataset makes possible several tasks well-suited for machine learning most of which are highly", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 261, + 225, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 225, + 275 + ], + "score": 1.0, + "content": "relevant for theorem proving:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 249, + 505, + 275 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 285, + 464, + 406 + ], + "lines": [ + { + "bbox": [ + 131, + 284, + 443, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 443, + 300 + ], + "score": 1.0, + "content": "• Predicting whether a statement is useful in the proof of a given conjecture;", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 303, + 420, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 420, + 317 + ], + "score": 1.0, + "content": "• Predicting the dependencies of a proof statement (premise selection);", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 322, + 415, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 415, + 335 + ], + "score": 1.0, + "content": "• Predicting whether a statement is an important one (human named);", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 339, + 464, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 464, + 354 + ], + "score": 1.0, + "content": "• Predicting which conjecture a particular intermediate statement originates from;", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 357, + 309, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 309, + 371 + ], + "score": 1.0, + "content": "• Predicting the name given to a statement;", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 375, + 451, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 451, + 389 + ], + "score": 1.0, + "content": "• Generating intermediate statements useful in the proof of a given conjecture;", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 394, + 367, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 367, + 407 + ], + "score": 1.0, + "content": "• Generating the conjecture the current proof will lead to.", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 11, + "bbox_fs": [ + 131, + 284, + 464, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 504, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "In what follows we focus on the first task: classifying proof step statements as being useful or not in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 428, + 465, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 465, + 441 + ], + "score": 1.0, + "content": "the context of a given proof. This task may be further specialized into two different tasks:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 417, + 505, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 452, + 505, + 537 + ], + "lines": [ + { + "bbox": [ + 132, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "• Unconditioned classification of proof steps: determining how likely a given proof is to be", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 463, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 141, + 463, + 505, + 477 + ], + "score": 1.0, + "content": "useful for the proof it occurred in, based solely on the content of statement (i.e. by only", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 475, + 426, + 487 + ], + "spans": [ + { + "bbox": [ + 142, + 475, + 426, + 487 + ], + "score": 1.0, + "content": "providing the model with the step statement itself, absent any context).", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 137, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "Conditioned classification of proof steps: determining how likely a given proof is to be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 141, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "useful for the proof it occurred in, with “conditioning” on the conjecture statement that the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 515, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 141, + 515, + 506, + 527 + ], + "score": 1.0, + "content": "proof was aiming to attain, i.e. by providing the model with both the step statement and the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 527, + 232, + 538 + ], + "spans": [ + { + "bbox": [ + 142, + 527, + 232, + 538 + ], + "score": 1.0, + "content": "conjecture statement).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20, + "bbox_fs": [ + 132, + 452, + 506, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 550, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "In the dataset, for every proof we provide the same number of useful and non-useful steps. As such,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "the proof step classification problem is a balanced two-class classification problem, where a random", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 572, + 269, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 269, + 584 + ], + "score": 1.0, + "content": "baseline would yield an accuracy of 0.5.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 549, + 505, + 584 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 601, + 421, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 422, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 422, + 613 + ], + "score": 1.0, + "content": "3.2 RELEVANCE TO INTERACTIVE AND AUTOMATED THEOREM PROVING", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "In the interaction with an interactive theorem prover, the tasks that require most human time are: the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "search for good intermediate steps; the search for automation techniques able to justify the individual", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "steps, and searching theorem proving libraries for the necessary simpler facts. These three problems", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "directly correspond to the machine learning tasks proposed in the previous subsection. Being able", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "to predict the usefulness of a statement will significantly improve many automation techniques.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "The generation of good intermediate lemmas or intermediate steps can improve level of granularity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "of the proof steps. Understanding the correspondence between statements and their names can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "allow users to search for statements in the libraries more efficiently (Aspinall & Kaliszyk, 2016).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Premise selection and filtering are already used in many theorem proving systems, and generation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 721, + 394, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 394, + 733 + ], + "score": 1.0, + "content": "of succeeding steps corresponds to conjecturing and theory exploration.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 621, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 104, + 418, + 322 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 186, + 80, + 424, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 186, + 79, + 425, + 93 + ], + "spans": [ + { + "bbox": [ + 186, + 79, + 425, + 93 + ], + "score": 1.0, + "content": "Figure 1: Unconditioned classification model architectures.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 196, + 104, + 418, + 322 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 104, + 418, + 322 + ], + "spans": [ + { + "bbox": [ + 196, + 104, + 418, + 322 + ], + "score": 0.965, + "type": "image", + "image_path": "821894620258089e462f597721fc7f2f41b4042c203e976e2ffdd198b87566c3.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 104, + 418, + 117.625 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 196, + 117.625, + 418, + 131.25 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 196, + 131.25, + 418, + 144.875 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 196, + 144.875, + 418, + 158.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 196, + 158.5, + 418, + 172.125 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 196, + 172.125, + 418, + 185.75 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 196, + 185.75, + 418, + 199.375 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 196, + 199.375, + 418, + 213.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 196, + 213.0, + 418, + 226.625 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 196, + 226.625, + 418, + 240.25 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 196, + 240.25, + 418, + 253.875 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 196, + 253.875, + 418, + 267.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 196, + 267.5, + 418, + 281.125 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 196, + 281.125, + 418, + 294.75 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 196, + 294.75, + 418, + 308.375 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 196, + 308.375, + 418, + 322.0 + ], + "spans": [], + "index": 16 + } + ] + } + ], + "index": 4.25 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 226, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 228, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 228, + 360 + ], + "score": 1.0, + "content": "4 BASELINE MODELS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "For each task (conditioned and unconditioned classification), we propose three different deep learn-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "ing architectures, meant to provide a baseline for the classification performance that can be achieved", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "on this dataset. Our models cover a range of architecture features (from convolutional networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "to recurrent networks), aiming at probing what characteristics of the data are the most helpful for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 415, + 207, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 207, + 426 + ], + "score": 1.0, + "content": "usefulness classification.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "Our models are implemented in TensorFlow (Abadi et al., 2015) using the Keras framework (Chollet,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 441, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 456 + ], + "score": 1.0, + "content": "2015). Each model was trained on a single Nvidia K80 GPU. Training only takes a few hours per", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "model, which makes running these experiments accessible to most people (they could even be run", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "score": 1.0, + "content": "on a laptop CPU). We are releasing all of our benchmark code as open-source software 2 so as to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 475, + 379, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 379, + 487 + ], + "score": 1.0, + "content": "allow others to reproduce our results and improve upon our models.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 106, + 500, + 317, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 318, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 318, + 512 + ], + "score": 1.0, + "content": "4.1 UNCONDITIONED CLASSIFICATION MODELS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 520, + 283, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 284, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 284, + 532 + ], + "score": 1.0, + "content": "Our three models for this task are as follow:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 132, + 541, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 133, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 133, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "• Logistic regression on top of learned token embeddings. This minimal model aims to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 142, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 142, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "determine to which extent simple differences between token distribution between useful", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 564, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 142, + 564, + 505, + 575 + ], + "score": 1.0, + "content": "and non-useful statements can be used to distinguish them. It provides an absolute floor on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 142, + 574, + 304, + 586 + ], + "spans": [ + { + "bbox": [ + 142, + 574, + 304, + 586 + ], + "score": 1.0, + "content": "the performance achievable on this task.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "• 2-layer 1D convolutional neural network (CNN) with global maxpooling for sequence re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 141, + 600, + 469, + 612 + ], + "spans": [ + { + "bbox": [ + 141, + 600, + 469, + 612 + ], + "score": 1.0, + "content": "duction. This model aims to determine the importance of local patterns of tokens.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "• 2-layer 1D CNN with LSTM (Hochreiter & Schmidhuber, 1997) sequence reduction. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 624, + 444, + 638 + ], + "spans": [ + { + "bbox": [ + 141, + 624, + 444, + 638 + ], + "score": 1.0, + "content": "model aims to determine the importance of order in the features sequences.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 349, + 658 + ], + "lines": [ + { + "bbox": [ + 106, + 646, + 350, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 350, + 659 + ], + "score": 1.0, + "content": "See figure 1 for a layer-by-layer description of these models.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 304, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 305, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 305, + 683 + ], + "score": 1.0, + "content": "4.2 CONDITIONED CLASSIFICATION MODELS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 109, + 691, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "For this task, we use versions of the above models that have two siamese branches (identical branches", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 702, + 505, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 715 + ], + "score": 1.0, + "content": "with shared weights), with one branch processing the proof step statement being considered, and the", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 721, + 376, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 376, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 376, + 733 + ], + "score": 1.0, + "content": "2https://github.com/tensorflow/deepmath/tree/master/holstep_baselines", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 104, + 418, + 322 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 186, + 80, + 424, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 186, + 79, + 425, + 93 + ], + "spans": [ + { + "bbox": [ + 186, + 79, + 425, + 93 + ], + "score": 1.0, + "content": "Figure 1: Unconditioned classification model architectures.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 196, + 104, + 418, + 322 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 196, + 104, + 418, + 322 + ], + "spans": [ + { + "bbox": [ + 196, + 104, + 418, + 322 + ], + "score": 0.965, + "type": "image", + "image_path": "821894620258089e462f597721fc7f2f41b4042c203e976e2ffdd198b87566c3.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 196, + 104, + 418, + 117.625 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 196, + 117.625, + 418, + 131.25 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 196, + 131.25, + 418, + 144.875 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 196, + 144.875, + 418, + 158.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 196, + 158.5, + 418, + 172.125 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 196, + 172.125, + 418, + 185.75 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 196, + 185.75, + 418, + 199.375 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 196, + 199.375, + 418, + 213.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 196, + 213.0, + 418, + 226.625 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 196, + 226.625, + 418, + 240.25 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 196, + 240.25, + 418, + 253.875 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 196, + 253.875, + 418, + 267.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 196, + 267.5, + 418, + 281.125 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 196, + 281.125, + 418, + 294.75 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 196, + 294.75, + 418, + 308.375 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 196, + 308.375, + 418, + 322.0 + ], + "spans": [], + "index": 16 + } + ] + } + ], + "index": 4.25 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 226, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 228, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 228, + 360 + ], + "score": 1.0, + "content": "4 BASELINE MODELS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "For each task (conditioned and unconditioned classification), we propose three different deep learn-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "ing architectures, meant to provide a baseline for the classification performance that can be achieved", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "on this dataset. Our models cover a range of architecture features (from convolutional networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "to recurrent networks), aiming at probing what characteristics of the data are the most helpful for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 415, + 207, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 207, + 426 + ], + "score": 1.0, + "content": "usefulness classification.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 370, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "Our models are implemented in TensorFlow (Abadi et al., 2015) using the Keras framework (Chollet,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 441, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 456 + ], + "score": 1.0, + "content": "2015). Each model was trained on a single Nvidia K80 GPU. Training only takes a few hours per", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "model, which makes running these experiments accessible to most people (they could even be run", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 477 + ], + "score": 1.0, + "content": "on a laptop CPU). We are releasing all of our benchmark code as open-source software 2 so as to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 475, + 379, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 379, + 487 + ], + "score": 1.0, + "content": "allow others to reproduce our results and improve upon our models.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 430, + 506, + 487 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 500, + 317, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 318, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 318, + 512 + ], + "score": 1.0, + "content": "4.1 UNCONDITIONED CLASSIFICATION MODELS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 520, + 283, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 284, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 284, + 532 + ], + "score": 1.0, + "content": "Our three models for this task are as follow:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 520, + 284, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 541, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 133, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 133, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "• Logistic regression on top of learned token embeddings. This minimal model aims to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 142, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 142, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "determine to which extent simple differences between token distribution between useful", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 564, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 142, + 564, + 505, + 575 + ], + "score": 1.0, + "content": "and non-useful statements can be used to distinguish them. It provides an absolute floor on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 142, + 574, + 304, + 586 + ], + "spans": [ + { + "bbox": [ + 142, + 574, + 304, + 586 + ], + "score": 1.0, + "content": "the performance achievable on this task.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 132, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "• 2-layer 1D convolutional neural network (CNN) with global maxpooling for sequence re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 141, + 600, + 469, + 612 + ], + "spans": [ + { + "bbox": [ + 141, + 600, + 469, + 612 + ], + "score": 1.0, + "content": "duction. This model aims to determine the importance of local patterns of tokens.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 132, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "• 2-layer 1D CNN with LSTM (Hochreiter & Schmidhuber, 1997) sequence reduction. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 624, + 444, + 638 + ], + "spans": [ + { + "bbox": [ + 141, + 624, + 444, + 638 + ], + "score": 1.0, + "content": "model aims to determine the importance of order in the features sequences.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 132, + 542, + 505, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 349, + 658 + ], + "lines": [ + { + "bbox": [ + 106, + 646, + 350, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 350, + 659 + ], + "score": 1.0, + "content": "See figure 1 for a layer-by-layer description of these models.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 646, + 350, + 659 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 304, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 305, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 305, + 683 + ], + "score": 1.0, + "content": "4.2 CONDITIONED CLASSIFICATION MODELS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 109, + 691, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "For this task, we use versions of the above models that have two siamese branches (identical branches", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 702, + 505, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 715 + ], + "score": 1.0, + "content": "with shared weights), with one branch processing the proof step statement being considered, and the", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 690, + 505, + 715 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 108, + 460, + 459 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 191, + 80, + 419, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 191, + 79, + 420, + 93 + ], + "spans": [ + { + "bbox": [ + 191, + 79, + 420, + 93 + ], + "score": 1.0, + "content": "Figure 2: Conditioned classification model architectures.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 153, + 108, + 460, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 108, + 460, + 459 + ], + "spans": [ + { + "bbox": [ + 153, + 108, + 460, + 459 + ], + "score": 0.973, + "type": "image", + "image_path": "a90e1dc148c2f2fa15c0165b2b7665040e764c96be2465d5ee50f0ad13c726d4.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 153, + 108, + 460, + 225.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 153, + 225.0, + 460, + 342.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 153, + 342.0, + 460, + 459.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "other branch processing the conjecture. Each branch outputs an embedding; these two embeddings", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "(step embedding and conjecture embedding) are then concatenated and the classified by a fully-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 514, + 432, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 432, + 527 + ], + "score": 1.0, + "content": "connected network. See figure 2 for a layer-by-layer description of these models.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 547, + 264, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 265, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 265, + 560 + ], + "score": 1.0, + "content": "4.3 INPUT STATEMENTS ENCODING", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 108, + 571, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 104, + 570, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 505, + 585 + ], + "score": 1.0, + "content": "It should be noted that all of our models start with an Embedding layer, mapping tokens or characters", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "in the statements to dense vectors in a low-dimensional space. We consider two possible encodings", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "for presenting the input statements (proof steps and conjectures) to the Embedding layers of our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 604, + 141, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 141, + 616 + ], + "score": 1.0, + "content": "models:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 131, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 132, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "• Character-level encoding of the human-readable versions of the statements, where each", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 142, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "character (out of a set of 86 unique characters) in the pretty-printed statements is mapped", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 653, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 505, + 667 + ], + "score": 1.0, + "content": "to a 256-dimensional dense vector. This encoding yields longer statements (training state-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 665, + 311, + 678 + ], + "spans": [ + { + "bbox": [ + 141, + 665, + 311, + 678 + ], + "score": 1.0, + "content": "ments are 308 character long on average).", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 133, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "• Token-level encoding of the versions of the statements rendered with our proposed high-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 141, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "level tokenization scheme. This encoding yields shorter statements (training statements are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 141, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "60 token long on average), while considerably increasing the size of set of unique tokens", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 720, + 294, + 732 + ], + "spans": [ + { + "bbox": [ + 141, + 720, + 294, + 732 + ], + "score": 1.0, + "content": "(1993 total tokens in the training set).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 108, + 460, + 459 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 191, + 80, + 419, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 191, + 79, + 420, + 93 + ], + "spans": [ + { + "bbox": [ + 191, + 79, + 420, + 93 + ], + "score": 1.0, + "content": "Figure 2: Conditioned classification model architectures.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image_body", + "bbox": [ + 153, + 108, + 460, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 108, + 460, + 459 + ], + "spans": [ + { + "bbox": [ + 153, + 108, + 460, + 459 + ], + "score": 0.973, + "type": "image", + "image_path": "a90e1dc148c2f2fa15c0165b2b7665040e764c96be2465d5ee50f0ad13c726d4.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 153, + 108, + 460, + 225.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 153, + 225.0, + 460, + 342.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 153, + 342.0, + 460, + 459.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "other branch processing the conjecture. Each branch outputs an embedding; these two embeddings", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "(step embedding and conjecture embedding) are then concatenated and the classified by a fully-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 514, + 432, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 432, + 527 + ], + "score": 1.0, + "content": "connected network. See figure 2 for a layer-by-layer description of these models.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 492, + 505, + 527 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 547, + 264, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 265, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 265, + 560 + ], + "score": 1.0, + "content": "4.3 INPUT STATEMENTS ENCODING", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 108, + 571, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 104, + 570, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 505, + 585 + ], + "score": 1.0, + "content": "It should be noted that all of our models start with an Embedding layer, mapping tokens or characters", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "in the statements to dense vectors in a low-dimensional space. We consider two possible encodings", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 606 + ], + "score": 1.0, + "content": "for presenting the input statements (proof steps and conjectures) to the Embedding layers of our", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 604, + 141, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 141, + 616 + ], + "score": 1.0, + "content": "models:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 570, + 505, + 616 + ] + }, + { + "type": "list", + "bbox": [ + 131, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 132, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "• Character-level encoding of the human-readable versions of the statements, where each", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 142, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "character (out of a set of 86 unique characters) in the pretty-printed statements is mapped", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 653, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 505, + 667 + ], + "score": 1.0, + "content": "to a 256-dimensional dense vector. This encoding yields longer statements (training state-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 665, + 311, + 678 + ], + "spans": [ + { + "bbox": [ + 141, + 665, + 311, + 678 + ], + "score": 1.0, + "content": "ments are 308 character long on average).", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 133, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "• Token-level encoding of the versions of the statements rendered with our proposed high-", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 141, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "level tokenization scheme. This encoding yields shorter statements (training statements are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 141, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "60 token long on average), while considerably increasing the size of set of unique tokens", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 720, + 294, + 732 + ], + "spans": [ + { + "bbox": [ + 141, + 720, + 294, + 732 + ], + "score": 1.0, + "content": "(1993 total tokens in the training set).", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + } + ], + "index": 15.5, + "bbox_fs": [ + 132, + 632, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 101, + 471, + 158 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 160, + 80, + 451, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 78, + 452, + 95 + ], + "spans": [ + { + "bbox": [ + 159, + 78, + 452, + 95 + ], + "score": 1.0, + "content": "Table 2: HolStep proof step classification accuracy without conditioning", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 140, + 101, + 471, + 158 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 101, + 471, + 158 + ], + "spans": [ + { + "bbox": [ + 140, + 101, + 471, + 158 + ], + "score": 0.97, + "html": "
Logistic regression1D CNN1D CNN-LSTM
Accuracy with char input0.710.820.83
Accuracy with token input0.710.830.77
", + "type": "table", + "image_path": "1cb801e38b574428eabc8f684adacd5e5fc3dfbf2886e132fcdca14fa9e5fd1a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 140, + 101, + 471, + 120.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 140, + 120.0, + 471, + 139.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 140, + 139.0, + 471, + 158.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 140, + 189, + 471, + 246 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 165, + 168, + 445, + 180 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 166, + 445, + 182 + ], + "spans": [ + { + "bbox": [ + 165, + 166, + 445, + 182 + ], + "score": 1.0, + "content": "Table 3: HolStep proof step classification accuracy with conditioning", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 140, + 189, + 471, + 246 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 140, + 189, + 471, + 246 + ], + "spans": [ + { + "bbox": [ + 140, + 189, + 471, + 246 + ], + "score": 0.981, + "html": "
Logistic regressionSiamese 1D CNNSiamese 1D CNN-LSTM
Accuracy with char input0.710.810.83
Accuracy with token input0.710.820.77
", + "type": "table", + "image_path": "d4f4c0de5892ddae27d24b3b9fd52e205969fc05f3bcb7d6ca6be57bfb4e9995.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 140, + 189, + 471, + 208.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 140, + 208.0, + 471, + 227.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 140, + 227.0, + 471, + 246.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 107, + 264, + 173, + 277 + ], + "lines": [ + { + "bbox": [ + 104, + 263, + 174, + 280 + ], + "spans": [ + { + "bbox": [ + 104, + 263, + 174, + 280 + ], + "score": 1.0, + "content": "5 RESULTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 108, + 289, + 404, + 301 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 405, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 405, + 303 + ], + "score": 1.0, + "content": "Experimental results are presented in tables 2 and 3, as well as figs. 3 to 6.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 315, + 296, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 298, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 298, + 326 + ], + "score": 1.0, + "content": "5.1 INFLUENCE OF MODEL ARCHITECTURE", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 372, + 348 + ], + "score": 1.0, + "content": "Our unconditioned logistic regression model yields an accuracy of", + "type": "text" + }, + { + "bbox": [ + 373, + 335, + 392, + 345 + ], + "score": 0.88, + "content": "71 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 334, + 505, + 348 + ], + "score": 1.0, + "content": ", both with character encod-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "ing and token encoding (tables 2 and 3). This demonstrates that differences in token or character", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "distributions between useful and non-useful steps alone, absent any context, is sufficient for discrim-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "inating between useful and non-useful statements to a reasonable extent. This also demonstrates that", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 379, + 486, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 486, + 391 + ], + "score": 1.0, + "content": "the token encoding is not fundamentally more informative than raw character-level statements.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 396, + 504, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 388, + 407 + ], + "score": 1.0, + "content": "Additionally, our unconditioned 1D CNN model yields an accuracy of", + "type": "text" + }, + { + "bbox": [ + 388, + 396, + 407, + 406 + ], + "score": 0.88, + "content": "82 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 396, + 419, + 407 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 419, + 396, + 438, + 406 + ], + "score": 0.88, + "content": "83 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 396, + 504, + 407 + ], + "score": 1.0, + "content": ", both with char-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "acter encoding and token encoding (tables 2 and 3). This demonstrates that patterns of characters or", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "patterns of tokens are considerably more informative than single tokens for the purpose of usefulness", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 429, + 163, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 163, + 441 + ], + "score": 1.0, + "content": "classification.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 108, + 446, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "score": 1.0, + "content": "Finally, our unconditioned convolutional-recurrent model does not improve upon the results of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "1D CNN, which indicates that our models are not able to meaningfully leverage order in the feature", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 468, + 307, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 307, + 479 + ], + "score": 1.0, + "content": "sequences into which the statements are encoded.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 493, + 271, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 272, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 272, + 505 + ], + "score": 1.0, + "content": "5.2 INFLUENCE OF INPUT ENCODING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 557 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "For the logistic regression model and the 2-layer 1D CNN model, the choice of input encoding seems", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "to have little impact. For the convolutional-recurrent model, the use of the high-level tokenization", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "seems to cause a large decrease in model performance (figs. 4 and 6). This may be due to the fact", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 546, + 453, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 453, + 558 + ], + "score": 1.0, + "content": "that token encoding yields shorter sequences, making the use of a LSTM less relevant.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 107, + 572, + 354, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 355, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 355, + 584 + ], + "score": 1.0, + "content": "5.3 INFLUENCE OF CONDITIONING ON THE CONJECTURE", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 636 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "None of our conditioned models appear to be able to improve upon the unconditioned models, which", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "indicates that our architectures are not able to leverage the information provided by the conjecture.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "The presence of the conditioning does however impact the training profile of our models, in partic-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 624, + 503, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 503, + 637 + ], + "score": 1.0, + "content": "ular by making the 1D CNN model converge faster and overfit significantly quicker (figs. 5 and 6).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 651, + 201, + 664 + ], + "lines": [ + { + "bbox": [ + 104, + 649, + 203, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 649, + 203, + 667 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "Our baseline deep learning models, albeit fairly weak, are still able to predict statement usefulness", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "with a remarkably high accuracy. Such methods already help first-order automated provers (Kaliszyk", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "& Urban, 2015a) and as the branching factor is higher in HOL the predictions are valuable for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "number of practical proving applications. This includes making tableaux-based (Paulson, 1999)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "and superposition-based (Hurd, 2003) internal ITP proof search significantly more efficient in turn", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 101, + 471, + 158 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 160, + 80, + 451, + 92 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 78, + 452, + 95 + ], + "spans": [ + { + "bbox": [ + 159, + 78, + 452, + 95 + ], + "score": 1.0, + "content": "Table 2: HolStep proof step classification accuracy without conditioning", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 140, + 101, + 471, + 158 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 101, + 471, + 158 + ], + "spans": [ + { + "bbox": [ + 140, + 101, + 471, + 158 + ], + "score": 0.97, + "html": "
Logistic regression1D CNN1D CNN-LSTM
Accuracy with char input0.710.820.83
Accuracy with token input0.710.830.77
", + "type": "table", + "image_path": "1cb801e38b574428eabc8f684adacd5e5fc3dfbf2886e132fcdca14fa9e5fd1a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 140, + 101, + 471, + 120.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 140, + 120.0, + 471, + 139.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 140, + 139.0, + 471, + 158.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 140, + 189, + 471, + 246 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 165, + 168, + 445, + 180 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 166, + 445, + 182 + ], + "spans": [ + { + "bbox": [ + 165, + 166, + 445, + 182 + ], + "score": 1.0, + "content": "Table 3: HolStep proof step classification accuracy with conditioning", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 140, + 189, + 471, + 246 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 140, + 189, + 471, + 246 + ], + "spans": [ + { + "bbox": [ + 140, + 189, + 471, + 246 + ], + "score": 0.981, + "html": "
Logistic regressionSiamese 1D CNNSiamese 1D CNN-LSTM
Accuracy with char input0.710.810.83
Accuracy with token input0.710.820.77
", + "type": "table", + "image_path": "d4f4c0de5892ddae27d24b3b9fd52e205969fc05f3bcb7d6ca6be57bfb4e9995.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 140, + 189, + 471, + 208.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 140, + 208.0, + 471, + 227.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 140, + 227.0, + 471, + 246.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 107, + 264, + 173, + 277 + ], + "lines": [ + { + "bbox": [ + 104, + 263, + 174, + 280 + ], + "spans": [ + { + "bbox": [ + 104, + 263, + 174, + 280 + ], + "score": 1.0, + "content": "5 RESULTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 108, + 289, + 404, + 301 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 405, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 405, + 303 + ], + "score": 1.0, + "content": "Experimental results are presented in tables 2 and 3, as well as figs. 3 to 6.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 288, + 405, + 303 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 315, + 296, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 298, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 298, + 326 + ], + "score": 1.0, + "content": "5.1 INFLUENCE OF MODEL ARCHITECTURE", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 505, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 372, + 348 + ], + "score": 1.0, + "content": "Our unconditioned logistic regression model yields an accuracy of", + "type": "text" + }, + { + "bbox": [ + 373, + 335, + 392, + 345 + ], + "score": 0.88, + "content": "71 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 334, + 505, + 348 + ], + "score": 1.0, + "content": ", both with character encod-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "ing and token encoding (tables 2 and 3). This demonstrates that differences in token or character", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "distributions between useful and non-useful steps alone, absent any context, is sufficient for discrim-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "inating between useful and non-useful statements to a reasonable extent. This also demonstrates that", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 379, + 486, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 486, + 391 + ], + "score": 1.0, + "content": "the token encoding is not fundamentally more informative than raw character-level statements.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 334, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 440 + ], + "lines": [ + { + "bbox": [ + 106, + 396, + 504, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 388, + 407 + ], + "score": 1.0, + "content": "Additionally, our unconditioned 1D CNN model yields an accuracy of", + "type": "text" + }, + { + "bbox": [ + 388, + 396, + 407, + 406 + ], + "score": 0.88, + "content": "82 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 396, + 419, + 407 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 419, + 396, + 438, + 406 + ], + "score": 0.88, + "content": "83 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 396, + 504, + 407 + ], + "score": 1.0, + "content": ", both with char-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "acter encoding and token encoding (tables 2 and 3). This demonstrates that patterns of characters or", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "patterns of tokens are considerably more informative than single tokens for the purpose of usefulness", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 429, + 163, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 163, + 441 + ], + "score": 1.0, + "content": "classification.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 396, + 505, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 446, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "score": 1.0, + "content": "Finally, our unconditioned convolutional-recurrent model does not improve upon the results of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "1D CNN, which indicates that our models are not able to meaningfully leverage order in the feature", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 468, + 307, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 307, + 479 + ], + "score": 1.0, + "content": "sequences into which the statements are encoded.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 445, + 505, + 479 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 493, + 271, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 272, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 272, + 505 + ], + "score": 1.0, + "content": "5.2 INFLUENCE OF INPUT ENCODING", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 557 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "For the logistic regression model and the 2-layer 1D CNN model, the choice of input encoding seems", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "to have little impact. For the convolutional-recurrent model, the use of the high-level tokenization", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "seems to cause a large decrease in model performance (figs. 4 and 6). This may be due to the fact", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 546, + 453, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 453, + 558 + ], + "score": 1.0, + "content": "that token encoding yields shorter sequences, making the use of a LSTM less relevant.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 513, + 505, + 558 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 572, + 354, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 355, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 355, + 584 + ], + "score": 1.0, + "content": "5.3 INFLUENCE OF CONDITIONING ON THE CONJECTURE", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 636 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "None of our conditioned models appear to be able to improve upon the unconditioned models, which", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "indicates that our architectures are not able to leverage the information provided by the conjecture.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "The presence of the conditioning does however impact the training profile of our models, in partic-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 624, + 503, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 503, + 637 + ], + "score": 1.0, + "content": "ular by making the 1D CNN model converge faster and overfit significantly quicker (figs. 5 and 6).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 591, + 506, + 637 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 651, + 201, + 664 + ], + "lines": [ + { + "bbox": [ + 104, + 649, + 203, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 649, + 203, + 667 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "Our baseline deep learning models, albeit fairly weak, are still able to predict statement usefulness", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "with a remarkably high accuracy. Such methods already help first-order automated provers (Kaliszyk", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "& Urban, 2015a) and as the branching factor is higher in HOL the predictions are valuable for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "number of practical proving applications. This includes making tableaux-based (Paulson, 1999)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "and superposition-based (Hurd, 2003) internal ITP proof search significantly more efficient in turn", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "making formalization easier. However, our models do not appear to be able to leverage order in the", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "input sequences, nor conditioning on the conjectures. This is due to the fact that these models are not", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "doing any form of logical reasoning on their input statements; rather they are doing simple pattern", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "matching at the level of n-grams of characters or tokens. This shows the need to focus future efforts", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "score": 1.0, + "content": "on different models that can do reasoning, or alternatively, on systems that blend explicit reasoning", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "(e.g. graph search) with deep learning-based feature learning. A potential new direction would be", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "to leverage the graph structure of HOL statements using e.g. Recursive Neural Tensor Networks", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 515, + 381, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 381, + 527 + ], + "score": 1.0, + "content": "(Socher et al., 2013a;b) or other graph-based recursive architectures.", + "type": "text", + "cross_page": true + } + ], + "index": 23 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 81, + 278, + 203 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 81, + 278, + 203 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 81, + 278, + 203 + ], + "spans": [ + { + "bbox": [ + 128, + 81, + 278, + 203 + ], + "score": 0.97, + "type": "image", + "image_path": "a005d1a27b1ed2c048eb84bbd78e94445bdeed1aa42302235e5adc5fb5947e63.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 128, + 81, + 278, + 142.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 142.0, + 278, + 203.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 213, + 299, + 235 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 212, + 299, + 225 + ], + "spans": [ + { + "bbox": [ + 107, + 212, + 299, + 225 + ], + "score": 1.0, + "content": "Figure 3: Training profile of the three uncondi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 222, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 107, + 222, + 285, + 236 + ], + "score": 1.0, + "content": "tioned baseline models with character input.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 324, + 81, + 473, + 203 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 324, + 81, + 473, + 203 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 324, + 81, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 324, + 81, + 473, + 203 + ], + "score": 0.969, + "type": "image", + "image_path": "b47242ae74d391eb24307af379669fb0a56e2025f35168a81905c69493fd7b4d.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 324, + 81, + 473, + 142.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 324, + 142.0, + 473, + 203.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 311, + 213, + 504, + 235 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 311, + 213, + 503, + 225 + ], + "spans": [ + { + "bbox": [ + 311, + 213, + 503, + 225 + ], + "score": 1.0, + "content": "Figure 4: Training profile of the three uncondi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 222, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 311, + 222, + 476, + 236 + ], + "score": 1.0, + "content": "tioned baseline models with token input.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + } + ], + "index": 4.25 + }, + { + "type": "image", + "bbox": [ + 129, + 255, + 278, + 377 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 255, + 278, + 377 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 129, + 255, + 278, + 377 + ], + "spans": [ + { + "bbox": [ + 129, + 255, + 278, + 377 + ], + "score": 0.968, + "type": "image", + "image_path": "17c7f989c97c9e8b86b4fe265a96619c2f7ea88a5e06b21d6c64ceddf2338aa8.jpg" + } + ] + } + ], + "index": 9.0, + "virtual_lines": [ + { + "bbox": [ + 129, + 255, + 278, + 316.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 129, + 316.0, + 278, + 377.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 386, + 298, + 408 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 386, + 299, + 398 + ], + "spans": [ + { + "bbox": [ + 107, + 386, + 299, + 398 + ], + "score": 1.0, + "content": "Figure 5: Training profile of the three condi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 396, + 285, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 285, + 410 + ], + "score": 1.0, + "content": "tioned baseline models with character input.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + } + ], + "index": 10.75 + }, + { + "type": "image", + "bbox": [ + 324, + 254, + 472, + 377 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 324, + 254, + 472, + 377 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 324, + 254, + 472, + 377 + ], + "spans": [ + { + "bbox": [ + 324, + 254, + 472, + 377 + ], + "score": 0.969, + "type": "image", + "image_path": "87d44006190fe92d90fd54144f8b0605b42092602e67eff4e886c823b46d4a17.jpg" + } + ] + } + ], + "index": 10.0, + "virtual_lines": [ + { + "bbox": [ + 324, + 254, + 472, + 315.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 324, + 315.5, + 472, + 377.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 311, + 387, + 503, + 408 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 311, + 386, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 311, + 386, + 504, + 398 + ], + "score": 1.0, + "content": "Figure 6: Training profile of the three condi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 396, + 475, + 410 + ], + "spans": [ + { + "bbox": [ + 311, + 396, + 475, + 410 + ], + "score": 1.0, + "content": "tioned baseline models with token input.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + } + ], + "index": 12.25 + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "making formalization easier. However, our models do not appear to be able to leverage order in the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "input sequences, nor conditioning on the conjectures. This is due to the fact that these models are not", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "doing any form of logical reasoning on their input statements; rather they are doing simple pattern", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "matching at the level of n-grams of characters or tokens. This shows the need to focus future efforts", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "score": 1.0, + "content": "on different models that can do reasoning, or alternatively, on systems that blend explicit reasoning", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "(e.g. graph search) with deep learning-based feature learning. A potential new direction would be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "to leverage the graph structure of HOL statements using e.g. Recursive Neural Tensor Networks", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 515, + 381, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 381, + 527 + ], + "score": 1.0, + "content": "(Socher et al., 2013a;b) or other graph-based recursive architectures.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 548, + 195, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 197, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 197, + 561 + ], + "score": 1.0, + "content": "6.1 FUTURE WORK", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 504, + 584 + ], + "score": 1.0, + "content": "The dataset focuses on one interactive theorem prover. It would be interesting if the proposed tech-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "niques generalize, primarily across ITPs that use the same foundational logic, for example using", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "OpenTheory (Hurd, 2011), and secondarily across fundamentally different ITPs or even ATPs. A", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "significant part of the unused steps originates from trying to fulfill the conditions for rewriting and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "from calls to intuitionistic tableaux. The main focus is however on the human found proofs so the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "trained predictions may to an extent mimic the bias on the usefulness in the human proofs. As ATPs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "score": 1.0, + "content": "are at the moment very week in comparison with human intuition improving this even for the many", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 649, + 361, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 361, + 662 + ], + "score": 1.0, + "content": "proofs humans do not find difficult would be an important gain.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Finally, two of the proposed task for the dataset have been premise selection and intermediate sen-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "tence generation. It would be interesting to define more ATP-based ways to evaluate the selected", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "premises, as well as to evaluate generated sentences (Kaliszyk et al., 2015). The set is a relatively", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "large one when it comes to proof step classification, however the number of available premises", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "makes the set a medium-sized set for premise selection in comparison with those of the Mizar Math-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 278, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 278, + 733 + ], + "score": 1.0, + "content": "ematical Library or the seL4 development.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 81, + 278, + 203 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 81, + 278, + 203 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 81, + 278, + 203 + ], + "spans": [ + { + "bbox": [ + 128, + 81, + 278, + 203 + ], + "score": 0.97, + "type": "image", + "image_path": "a005d1a27b1ed2c048eb84bbd78e94445bdeed1aa42302235e5adc5fb5947e63.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 128, + 81, + 278, + 142.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 142.0, + 278, + 203.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 213, + 299, + 235 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 212, + 299, + 225 + ], + "spans": [ + { + "bbox": [ + 107, + 212, + 299, + 225 + ], + "score": 1.0, + "content": "Figure 3: Training profile of the three uncondi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 222, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 107, + 222, + 285, + 236 + ], + "score": 1.0, + "content": "tioned baseline models with character input.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 324, + 81, + 473, + 203 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 324, + 81, + 473, + 203 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 324, + 81, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 324, + 81, + 473, + 203 + ], + "score": 0.969, + "type": "image", + "image_path": "b47242ae74d391eb24307af379669fb0a56e2025f35168a81905c69493fd7b4d.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 324, + 81, + 473, + 142.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 324, + 142.0, + 473, + 203.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 311, + 213, + 504, + 235 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 311, + 213, + 503, + 225 + ], + "spans": [ + { + "bbox": [ + 311, + 213, + 503, + 225 + ], + "score": 1.0, + "content": "Figure 4: Training profile of the three uncondi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 222, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 311, + 222, + 476, + 236 + ], + "score": 1.0, + "content": "tioned baseline models with token input.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + } + ], + "index": 4.25 + }, + { + "type": "image", + "bbox": [ + 129, + 255, + 278, + 377 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 255, + 278, + 377 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 129, + 255, + 278, + 377 + ], + "spans": [ + { + "bbox": [ + 129, + 255, + 278, + 377 + ], + "score": 0.968, + "type": "image", + "image_path": "17c7f989c97c9e8b86b4fe265a96619c2f7ea88a5e06b21d6c64ceddf2338aa8.jpg" + } + ] + } + ], + "index": 9.0, + "virtual_lines": [ + { + "bbox": [ + 129, + 255, + 278, + 316.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 129, + 316.0, + 278, + 377.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 386, + 298, + 408 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 386, + 299, + 398 + ], + "spans": [ + { + "bbox": [ + 107, + 386, + 299, + 398 + ], + "score": 1.0, + "content": "Figure 5: Training profile of the three condi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 396, + 285, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 285, + 410 + ], + "score": 1.0, + "content": "tioned baseline models with character input.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + } + ], + "index": 10.75 + }, + { + "type": "image", + "bbox": [ + 324, + 254, + 472, + 377 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 324, + 254, + 472, + 377 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 324, + 254, + 472, + 377 + ], + "spans": [ + { + "bbox": [ + 324, + 254, + 472, + 377 + ], + "score": 0.969, + "type": "image", + "image_path": "87d44006190fe92d90fd54144f8b0605b42092602e67eff4e886c823b46d4a17.jpg" + } + ] + } + ], + "index": 10.0, + "virtual_lines": [ + { + "bbox": [ + 324, + 254, + 472, + 315.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 324, + 315.5, + 472, + 377.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 311, + 387, + 503, + 408 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 311, + 386, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 311, + 386, + 504, + 398 + ], + "score": 1.0, + "content": "Figure 6: Training profile of the three condi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 396, + 475, + 410 + ], + "spans": [ + { + "bbox": [ + 311, + 396, + 475, + 410 + ], + "score": 1.0, + "content": "tioned baseline models with token input.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + } + ], + "index": 12.25 + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 526 + ], + "lines": [], + "index": 19.5, + "bbox_fs": [ + 105, + 438, + 506, + 527 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 548, + 195, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 197, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 197, + 561 + ], + "score": 1.0, + "content": "6.1 FUTURE WORK", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 504, + 584 + ], + "score": 1.0, + "content": "The dataset focuses on one interactive theorem prover. It would be interesting if the proposed tech-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "niques generalize, primarily across ITPs that use the same foundational logic, for example using", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "OpenTheory (Hurd, 2011), and secondarily across fundamentally different ITPs or even ATPs. A", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "significant part of the unused steps originates from trying to fulfill the conditions for rewriting and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "from calls to intuitionistic tableaux. The main focus is however on the human found proofs so the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "trained predictions may to an extent mimic the bias on the usefulness in the human proofs. As ATPs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "score": 1.0, + "content": "are at the moment very week in comparison with human intuition improving this even for the many", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 649, + 361, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 361, + 662 + ], + "score": 1.0, + "content": "proofs humans do not find difficult would be an important gain.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 572, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Finally, two of the proposed task for the dataset have been premise selection and intermediate sen-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "tence generation. It would be interesting to define more ATP-based ways to evaluate the selected", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "premises, as well as to evaluate generated sentences (Kaliszyk et al., 2015). The set is a relatively", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "large one when it comes to proof step classification, however the number of available premises", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "makes the set a medium-sized set for premise selection in comparison with those of the Mizar Math-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 278, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 278, + 733 + ], + "score": 1.0, + "content": "ematical Library or the seL4 development.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 205, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 83, + 206, + 94 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 206, + 94 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 103, + 393, + 114 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 395, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 395, + 116 + ], + "score": 1.0, + "content": "The first author was partly supported by the ERC starting grant 714034.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 131, + 175, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 176, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 176, + 144 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 108, + 149, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vin-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "cent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Watten-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensor-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 236, + 154, + 251 + ], + "spans": [ + { + "bbox": [ + 115, + 236, + 154, + 251 + ], + "score": 1.0, + "content": "flow.org.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 504, + 268 + ], + "score": 1.0, + "content": "Jesse Alama, Tom Heskes, Daniel Kühlwein, Evgeni Tsivtsivadze, and Josef Urban. Premise se-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "lection for mathematics by corpus analysis and kernel methods. J. Autom. Reasoning, 52(2):", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 278, + 321, + 289 + ], + "spans": [ + { + "bbox": [ + 117, + 278, + 321, + 289 + ], + "score": 1.0, + "content": "191–213, 2014. doi: 10.1007/s10817-013-9286-5.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Alex A. Alemi, François Chollet, Geoffrey Irving, Christian Szegedy, and Josef Urban. DeepMath", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "– Deep sequence models for premise selection. In Daniel D. Lee, Masashi Sugiyama, Ulrike V.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "Luxburg, Isabelle Guyon, and Roman Garnett (eds.), Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 329, + 452, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 452, + 342 + ], + "score": 1.0, + "content": "Systems (NIPS 2016), pp. 2235–2243, 2016. URL https://arxiv.org/abs/1606.04442.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 347, + 504, + 381 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "David Aspinall and Cezary Kaliszyk. What’s in a theorem name? In Jasmin Christian Blanchette", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 357, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 357, + 505, + 372 + ], + "score": 1.0, + "content": "and Stephan Merz (eds.), Interactive Theorem Proving (ITP 2016), volume 9807 of LNCS, pp.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 370, + 360, + 381 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 360, + 381 + ], + "score": 1.0, + "content": "459–465. Springer, 2016. doi: 10.1007/978-3-319-43144-4.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 104, + 388, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 504, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 504, + 401 + ], + "score": 1.0, + "content": "Franz Baader and Tobias Nipkow. Term rewriting and all that. Cambridge University Press, 1998.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 399, + 225, + 411 + ], + "spans": [ + { + "bbox": [ + 116, + 399, + 225, + 411 + ], + "score": 1.0, + "content": "ISBN 978-0-521-45520-6.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 503, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "Jasmin C. Blanchette, Cezary Kaliszyk, Lawrence C. Paulson, and Josef Urban. Hammering towards", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "QED. J. Formalized Reasoning, 9(1):101–148, 2016. ISSN 1972-5787. doi: 10.6092/issn.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 440, + 188, + 451 + ], + "spans": [ + { + "bbox": [ + 117, + 440, + 188, + 451 + ], + "score": 1.0, + "content": "1972-5787/4593.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 104, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 104, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "Jasmin Christian Blanchette, Maximilian P. L. Haslbeck, Daniel Matichuk, and Tobias Nipkow.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "Mining the Archive of Formal Proofs. In Manfred Kerber, Jacques Carette, Cezary Kaliszyk,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "Florian Rabe, and Volker Sorge (eds.), Intelligent Computer Mathematics (CICM 2015), volume", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 492, + 281, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 281, + 504 + ], + "score": 1.0, + "content": "9150 of LNCS, pp. 3–17. Springer, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "James P. Bridge, Sean B. Holden, and Lawrence C. Paulson. Machine learning for first-order theo-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "rem proving - learning to select a good heuristic. J. Autom. Reasoning, 53(2):141–172, 2014. doi:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 532, + 237, + 543 + ], + "spans": [ + { + "bbox": [ + 117, + 532, + 237, + 543 + ], + "score": 1.0, + "content": "10.1007/s10817-014-9301-5.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 584 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 504, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 504, + 563 + ], + "score": 1.0, + "content": "Haogang Chen, Daniel Ziegler, Tej Chajed, Adam Chlipala, M. Frans Kaashoek, and Nickolai Zel-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "dovich. Using crash Hoare logic for certifying the FSCQ file system. In Ajay Gulati and Hakim", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 572, + 381, + 585 + ], + "spans": [ + { + "bbox": [ + 116, + 572, + 381, + 585 + ], + "score": 1.0, + "content": "Weatherspoon (eds.), USENIX 2016. USENIX Association, 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 366, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 367, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 367, + 604 + ], + "score": 1.0, + "content": "François Chollet. Keras. https://github.com/fchollet/keras, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 609, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 623 + ], + "score": 1.0, + "content": "Alonzo Church. A formulation of the simple theory of types. J. Symb. Log., 5(2):56–68, 1940. doi:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 620, + 357, + 633 + ], + "spans": [ + { + "bbox": [ + 117, + 620, + 357, + 633 + ], + "score": 1.0, + "content": "10.2307/2266170. URL http://dx.doi.org/10.2307/2266170.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 639, + 504, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "score": 1.0, + "content": "J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 650, + 262, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 262, + 662 + ], + "score": 1.0, + "content": "Image Database. In CVPR09, 2009.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 104, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Hazel Duncan. The Use of Data-Mining for the Automatic Formation of Tactics. PhD thesis, Uni-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 680, + 229, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 229, + 692 + ], + "score": 1.0, + "content": "versity of Edinburgh, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Michael Färber and Chad E. Brown. Internal guidance for Satallax. In Nicola Olivetti and Ashish", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Tiwari (eds.), International Joint Conference on Automated Reasoning (IJCAR 2016), volume", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 720, + 438, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 438, + 732 + ], + "score": 1.0, + "content": "9706 of LNCS, pp. 349–361. Springer, 2016. doi: 10.1007/978-3-319-40229-1.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 205, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 83, + 206, + 94 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 206, + 94 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 103, + 393, + 114 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 395, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 395, + 116 + ], + "score": 1.0, + "content": "The first author was partly supported by the ERC starting grant 714034.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 102, + 395, + 116 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 131, + 175, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 176, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 176, + 144 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 108, + 149, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vin-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "cent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Watten-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensor-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 236, + 154, + 251 + ], + "spans": [ + { + "bbox": [ + 115, + 236, + 154, + 251 + ], + "score": 1.0, + "content": "flow.org.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 149, + 505, + 251 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 504, + 268 + ], + "score": 1.0, + "content": "Jesse Alama, Tom Heskes, Daniel Kühlwein, Evgeni Tsivtsivadze, and Josef Urban. Premise se-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "lection for mathematics by corpus analysis and kernel methods. J. Autom. Reasoning, 52(2):", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 278, + 321, + 289 + ], + "spans": [ + { + "bbox": [ + 117, + 278, + 321, + 289 + ], + "score": 1.0, + "content": "191–213, 2014. doi: 10.1007/s10817-013-9286-5.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 255, + 505, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "Alex A. Alemi, François Chollet, Geoffrey Irving, Christian Szegedy, and Josef Urban. DeepMath", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "– Deep sequence models for premise selection. In Daniel D. Lee, Masashi Sugiyama, Ulrike V.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "Luxburg, Isabelle Guyon, and Roman Garnett (eds.), Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 329, + 452, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 452, + 342 + ], + "score": 1.0, + "content": "Systems (NIPS 2016), pp. 2235–2243, 2016. URL https://arxiv.org/abs/1606.04442.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 296, + 506, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 347, + 504, + 381 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "David Aspinall and Cezary Kaliszyk. What’s in a theorem name? In Jasmin Christian Blanchette", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 357, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 115, + 357, + 505, + 372 + ], + "score": 1.0, + "content": "and Stephan Merz (eds.), Interactive Theorem Proving (ITP 2016), volume 9807 of LNCS, pp.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 370, + 360, + 381 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 360, + 381 + ], + "score": 1.0, + "content": "459–465. Springer, 2016. doi: 10.1007/978-3-319-43144-4.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 348, + 505, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 388, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 504, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 504, + 401 + ], + "score": 1.0, + "content": "Franz Baader and Tobias Nipkow. Term rewriting and all that. Cambridge University Press, 1998.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 399, + 225, + 411 + ], + "spans": [ + { + "bbox": [ + 116, + 399, + 225, + 411 + ], + "score": 1.0, + "content": "ISBN 978-0-521-45520-6.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 387, + 504, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 503, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "Jasmin C. Blanchette, Cezary Kaliszyk, Lawrence C. Paulson, and Josef Urban. Hammering towards", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "QED. J. Formalized Reasoning, 9(1):101–148, 2016. ISSN 1972-5787. doi: 10.6092/issn.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 440, + 188, + 451 + ], + "spans": [ + { + "bbox": [ + 117, + 440, + 188, + 451 + ], + "score": 1.0, + "content": "1972-5787/4593.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 417, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 457, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 104, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 104, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "Jasmin Christian Blanchette, Maximilian P. L. Haslbeck, Daniel Matichuk, and Tobias Nipkow.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "Mining the Archive of Formal Proofs. In Manfred Kerber, Jacques Carette, Cezary Kaliszyk,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "Florian Rabe, and Volker Sorge (eds.), Intelligent Computer Mathematics (CICM 2015), volume", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 492, + 281, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 281, + 504 + ], + "score": 1.0, + "content": "9150 of LNCS, pp. 3–17. Springer, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 457, + 505, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "James P. Bridge, Sean B. Holden, and Lawrence C. Paulson. Machine learning for first-order theo-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "rem proving - learning to select a good heuristic. J. Autom. Reasoning, 53(2):141–172, 2014. doi:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 532, + 237, + 543 + ], + "spans": [ + { + "bbox": [ + 117, + 532, + 237, + 543 + ], + "score": 1.0, + "content": "10.1007/s10817-014-9301-5.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 509, + 505, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 584 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 504, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 504, + 563 + ], + "score": 1.0, + "content": "Haogang Chen, Daniel Ziegler, Tej Chajed, Adam Chlipala, M. Frans Kaashoek, and Nickolai Zel-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "dovich. Using crash Hoare logic for certifying the FSCQ file system. In Ajay Gulati and Hakim", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 572, + 381, + 585 + ], + "spans": [ + { + "bbox": [ + 116, + 572, + 381, + 585 + ], + "score": 1.0, + "content": "Weatherspoon (eds.), USENIX 2016. USENIX Association, 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 550, + 505, + 585 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 366, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 367, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 367, + 604 + ], + "score": 1.0, + "content": "François Chollet. Keras. https://github.com/fchollet/keras, 2015.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 590, + 367, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 609, + 504, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 623 + ], + "score": 1.0, + "content": "Alonzo Church. A formulation of the simple theory of types. J. Symb. Log., 5(2):56–68, 1940. doi:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 620, + 357, + 633 + ], + "spans": [ + { + "bbox": [ + 117, + 620, + 357, + 633 + ], + "score": 1.0, + "content": "10.2307/2266170. URL http://dx.doi.org/10.2307/2266170.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 609, + 505, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 639, + 504, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 652 + ], + "score": 1.0, + "content": "J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 650, + 262, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 650, + 262, + 662 + ], + "score": 1.0, + "content": "Image Database. In CVPR09, 2009.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 638, + 505, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Hazel Duncan. The Use of Data-Mining for the Automatic Formation of Tactics. PhD thesis, Uni-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 680, + 229, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 229, + 692 + ], + "score": 1.0, + "content": "versity of Edinburgh, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 668, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Michael Färber and Chad E. Brown. Internal guidance for Satallax. In Nicola Olivetti and Ashish", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Tiwari (eds.), International Joint Conference on Automated Reasoning (IJCAR 2016), volume", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 720, + 438, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 438, + 732 + ], + "score": 1.0, + "content": "9706 of LNCS, pp. 349–361. Springer, 2016. doi: 10.1007/978-3-319-40229-1.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 699, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Georges Gonthier, Andrea Asperti, Jeremy Avigad, Yves Bertot, Cyril Cohen, François Garillot,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Stéphane Le Roux, Assia Mahboubi, Russell O’Connor, Sidi Ould Biha, Ioana Pasca, Laurence", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 506, + 117 + ], + "score": 1.0, + "content": "Rideau, Alexey Solovyev, Enrico Tassi, and Laurent Théry. A machine-checked proof of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 506, + 129 + ], + "score": 1.0, + "content": "odd order theorem. In Sandrine Blazy, Christine Paulin-Mohring, and David Pichardie (eds.),", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 127, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 116, + 127, + 496, + 138 + ], + "score": 1.0, + "content": "Interactive Theorem Proving (ITP 2013), volume 7998 of LNCS, pp. 163–179. Springer, 2013.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 144, + 502, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 504, + 157 + ], + "score": 1.0, + "content": "Michael J. C. Gordon, Robin Milner, and Christopher P. Wadsworth. Edinburgh LCF, volume 78", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 504, + 168 + ], + "score": 1.0, + "content": "of Lecture Notes in Computer Science. Springer, 1979. ISBN 3-540-09724-4. doi: 10.1007/", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 167, + 375, + 179 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 375, + 179 + ], + "score": 1.0, + "content": "3-540-09724-4. URL http://dx.doi.org/10.1007/3-540-09724-4.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "Adam Grabowski, Artur Kornilowicz, and Adam Naumowicz. Mizar in a nutshell. J. Formalized", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 196, + 391, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 196, + 391, + 208 + ], + "score": 1.0, + "content": "Reasoning, 3(2):153–245, 2010. doi: 10.6092/issn.1972-5787/1980.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 227 + ], + "score": 1.0, + "content": "Thomas Hales, John Harrison, Sean McLaughlin, Tobias Nipkow, Steven Obua, and Roland", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "Zumkeller. A revision of the proof of the Kepler Conjecture. Discrete & Computational Ge-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 223, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 223, + 248 + ], + "score": 1.0, + "content": "ometry, 44(1):1–34, 2010.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 505, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "score": 1.0, + "content": "Thomas C. Hales, Mark Adams, Gertrud Bauer, Dat Tat Dang, John Harrison, Truong Le Hoang,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 264, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 264, + 505, + 280 + ], + "score": 1.0, + "content": "Cezary Kaliszyk, Victor Magron, Sean McLaughlin, Thang Tat Nguyen, Truong Quang Nguyen,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "Tobias Nipkow, Steven Obua, Joseph Pleso, Jason Rute, Alexey Solovyev, An Hoai Thi Ta,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "Trung Nam Tran, Diep Thi Trieu, Josef Urban, Ky Khac Vu, and Roland Zumkeller. A formal", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 299, + 365, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 299, + 365, + 311 + ], + "score": 1.0, + "content": "proof of the Kepler conjecture. CoRR, abs/1501.02155, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 503, + 351 + ], + "lines": [ + { + "bbox": [ + 104, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "John Harrison. HOL Light: An overview. In Stefan Berghofer, Tobias Nipkow, Christian Urban,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 116, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "and Makarius Wenzel (eds.), Theorem Proving in Higher Order Logics (TPHOLs 2009), volume", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 338, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 288, + 352 + ], + "score": 1.0, + "content": "5674 of LNCS, pp. 60–66. Springer, 2009.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "John Harrison. The HOL Light theory of Euclidean space. J. Autom. Reasoning, 50(2):173–190,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 369, + 281, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 369, + 281, + 380 + ], + "score": 1.0, + "content": "2013. doi: 10.1007/s10817-012-9250-9.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 104, + 385, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 104, + 385, + 505, + 400 + ], + "score": 1.0, + "content": "John Harrison, Josef Urban, and Freek Wiedijk. History of interactive theorem proving. In Jörg", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "Siekmann (ed.), Handbook of the History of Logic vol. 9 (Computational Logic), pp. 135–214.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 408, + 178, + 420 + ], + "spans": [ + { + "bbox": [ + 116, + 408, + 178, + 420 + ], + "score": 1.0, + "content": "Elsevier, 2014.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 504, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "R. Hindley. The principal type-scheme of an object in combinatory logic. Transactions of the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 438, + 392, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 392, + 450 + ], + "score": 1.0, + "content": "american mathematical society, 146:29–60, 1969. ISSN 0002-9947.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 466, + 191, + 479 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 191, + 479 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 504, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "score": 1.0, + "content": "Kryštof Hoder and Andrei Voronkov. Sine qua non for large theory reasoning. In Nikolaj Bjørner and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 495, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 115, + 495, + 505, + 510 + ], + "score": 1.0, + "content": "Viorica Sofronie-Stokkermans (eds.), CADE-23, volume 6803 of LNAI, pp. 299–314. Springer,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 506, + 144, + 520 + ], + "spans": [ + { + "bbox": [ + 114, + 506, + 144, + 520 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 526, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Joe Hurd. First-order proof tactics in higher-order logic theorem provers. In Myla Archer, Ben Di", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 116, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "Vito, and César Muñoz (eds.), Design and Application of Strategies/Tactics in Higher Order Log-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "ics (STRATA 2003), number NASA/CP-2003-212448 in NASA Technical Reports, pp. 56–68,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 558, + 367, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 367, + 573 + ], + "score": 1.0, + "content": "September 2003. URL http://www.gilith.com/research/papers.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 504, + 612 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "Joe Hurd. The OpenTheory standard theory library. In Mihaela Gheorghiu Bobaru, Klaus Havelund,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 589, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 589, + 505, + 600 + ], + "score": 1.0, + "content": "Gerard J. Holzmann, and Rajeev Joshi (eds.), NASA Formal Methods (NFM 2011), volume 6617", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 599, + 275, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 275, + 612 + ], + "score": 1.0, + "content": "of LNCS, pp. 177–191. Springer, 2011.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 504, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Alexander Krauss. Scalable LCF-style proof translation. In Sandrine Blazy,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "Christine Paulin-Mohring, and David Pichardie (eds.), Interactive Theorem Proving (ITP 2013),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 640, + 319, + 652 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 319, + 652 + ], + "score": 1.0, + "content": "volume 7998 of LNCS, pp. 51–66. Springer, 2013.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 104, + 658, + 503, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 504, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 504, + 672 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Josef Urban. Learning-assisted automated reasoning with Flyspeck. J. Autom.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 669, + 393, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 393, + 681 + ], + "score": 1.0, + "content": "Reasoning, 53(2):173–213, 2014. doi: 10.1007/s10817-014-9303-3.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Josef Urban. FEMaLeCoP: Fairly efficient machine learning connection", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "prover. In Martin Davis, Ansgar Fehnker, Annabelle McIver, and Andrei Voronkov (eds.), 20th In-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "ternational Conference on Logic for Programming, Artificial Intelligence, and Reasoning (LPAR", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 720, + 492, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 492, + 733 + ], + "score": 1.0, + "content": "2015), volume 9450 of LNCS, pp. 88–96. Springer, 2015a. doi: 10.1007/978-3-662-48899-7.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Georges Gonthier, Andrea Asperti, Jeremy Avigad, Yves Bertot, Cyril Cohen, François Garillot,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Stéphane Le Roux, Assia Mahboubi, Russell O’Connor, Sidi Ould Biha, Ioana Pasca, Laurence", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 506, + 117 + ], + "score": 1.0, + "content": "Rideau, Alexey Solovyev, Enrico Tassi, and Laurent Théry. A machine-checked proof of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 506, + 129 + ], + "score": 1.0, + "content": "odd order theorem. In Sandrine Blazy, Christine Paulin-Mohring, and David Pichardie (eds.),", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 127, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 116, + 127, + 496, + 138 + ], + "score": 1.0, + "content": "Interactive Theorem Proving (ITP 2013), volume 7998 of LNCS, pp. 163–179. Springer, 2013.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 82, + 506, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 144, + 502, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 504, + 157 + ], + "score": 1.0, + "content": "Michael J. C. Gordon, Robin Milner, and Christopher P. Wadsworth. Edinburgh LCF, volume 78", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 504, + 168 + ], + "score": 1.0, + "content": "of Lecture Notes in Computer Science. Springer, 1979. ISBN 3-540-09724-4. doi: 10.1007/", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 167, + 375, + 179 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 375, + 179 + ], + "score": 1.0, + "content": "3-540-09724-4. URL http://dx.doi.org/10.1007/3-540-09724-4.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 145, + 504, + 179 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 185, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 506, + 198 + ], + "score": 1.0, + "content": "Adam Grabowski, Artur Kornilowicz, and Adam Naumowicz. Mizar in a nutshell. J. Formalized", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 196, + 391, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 196, + 391, + 208 + ], + "score": 1.0, + "content": "Reasoning, 3(2):153–245, 2010. doi: 10.6092/issn.1972-5787/1980.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 185, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 227 + ], + "score": 1.0, + "content": "Thomas Hales, John Harrison, Sean McLaughlin, Tobias Nipkow, Steven Obua, and Roland", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "Zumkeller. A revision of the proof of the Kepler Conjecture. Discrete & Computational Ge-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 223, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 223, + 248 + ], + "score": 1.0, + "content": "ometry, 44(1):1–34, 2010.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 213, + 505, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 505, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 268 + ], + "score": 1.0, + "content": "Thomas C. Hales, Mark Adams, Gertrud Bauer, Dat Tat Dang, John Harrison, Truong Le Hoang,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 264, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 264, + 505, + 280 + ], + "score": 1.0, + "content": "Cezary Kaliszyk, Victor Magron, Sean McLaughlin, Thang Tat Nguyen, Truong Quang Nguyen,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "Tobias Nipkow, Steven Obua, Joseph Pleso, Jason Rute, Alexey Solovyev, An Hoai Thi Ta,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "Trung Nam Tran, Diep Thi Trieu, Josef Urban, Ky Khac Vu, and Roland Zumkeller. A formal", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 299, + 365, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 299, + 365, + 311 + ], + "score": 1.0, + "content": "proof of the Kepler conjecture. CoRR, abs/1501.02155, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 253, + 505, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 503, + 351 + ], + "lines": [ + { + "bbox": [ + 104, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "John Harrison. HOL Light: An overview. In Stefan Berghofer, Tobias Nipkow, Christian Urban,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 116, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "and Makarius Wenzel (eds.), Theorem Proving in Higher Order Logics (TPHOLs 2009), volume", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 338, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 288, + 352 + ], + "score": 1.0, + "content": "5674 of LNCS, pp. 60–66. Springer, 2009.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 316, + 505, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "John Harrison. The HOL Light theory of Euclidean space. J. Autom. Reasoning, 50(2):173–190,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 369, + 281, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 369, + 281, + 380 + ], + "score": 1.0, + "content": "2013. doi: 10.1007/s10817-012-9250-9.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 356, + 505, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 104, + 385, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 104, + 385, + 505, + 400 + ], + "score": 1.0, + "content": "John Harrison, Josef Urban, and Freek Wiedijk. History of interactive theorem proving. In Jörg", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "Siekmann (ed.), Handbook of the History of Logic vol. 9 (Computational Logic), pp. 135–214.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 408, + 178, + 420 + ], + "spans": [ + { + "bbox": [ + 116, + 408, + 178, + 420 + ], + "score": 1.0, + "content": "Elsevier, 2014.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 385, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 504, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "R. Hindley. The principal type-scheme of an object in combinatory logic. Transactions of the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 438, + 392, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 392, + 450 + ], + "score": 1.0, + "content": "american mathematical society, 146:29–60, 1969. ISSN 0002-9947.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 426, + 505, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 466, + 191, + 479 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 191, + 479 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 455, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 504, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "score": 1.0, + "content": "Kryštof Hoder and Andrei Voronkov. Sine qua non for large theory reasoning. In Nikolaj Bjørner and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 495, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 115, + 495, + 505, + 510 + ], + "score": 1.0, + "content": "Viorica Sofronie-Stokkermans (eds.), CADE-23, volume 6803 of LNAI, pp. 299–314. Springer,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 506, + 144, + 520 + ], + "spans": [ + { + "bbox": [ + 114, + 506, + 144, + 520 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 485, + 505, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 526, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Joe Hurd. First-order proof tactics in higher-order logic theorem provers. In Myla Archer, Ben Di", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 116, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "Vito, and César Muñoz (eds.), Design and Application of Strategies/Tactics in Higher Order Log-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "ics (STRATA 2003), number NASA/CP-2003-212448 in NASA Technical Reports, pp. 56–68,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 558, + 367, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 367, + 573 + ], + "score": 1.0, + "content": "September 2003. URL http://www.gilith.com/research/papers.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 526, + 505, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 504, + 612 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "Joe Hurd. The OpenTheory standard theory library. In Mihaela Gheorghiu Bobaru, Klaus Havelund,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 589, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 589, + 505, + 600 + ], + "score": 1.0, + "content": "Gerard J. Holzmann, and Rajeev Joshi (eds.), NASA Formal Methods (NFM 2011), volume 6617", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 599, + 275, + 612 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 275, + 612 + ], + "score": 1.0, + "content": "of LNCS, pp. 177–191. Springer, 2011.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 577, + 505, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 504, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 631 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Alexander Krauss. Scalable LCF-style proof translation. In Sandrine Blazy,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "Christine Paulin-Mohring, and David Pichardie (eds.), Interactive Theorem Proving (ITP 2013),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 640, + 319, + 652 + ], + "spans": [ + { + "bbox": [ + 116, + 640, + 319, + 652 + ], + "score": 1.0, + "content": "volume 7998 of LNCS, pp. 51–66. Springer, 2013.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 616, + 505, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 658, + 503, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 504, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 504, + 672 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Josef Urban. Learning-assisted automated reasoning with Flyspeck. J. Autom.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 669, + 393, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 393, + 681 + ], + "score": 1.0, + "content": "Reasoning, 53(2):173–213, 2014. doi: 10.1007/s10817-014-9303-3.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 656, + 504, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Josef Urban. FEMaLeCoP: Fairly efficient machine learning connection", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "prover. In Martin Davis, Ansgar Fehnker, Annabelle McIver, and Andrei Voronkov (eds.), 20th In-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "ternational Conference on Logic for Programming, Artificial Intelligence, and Reasoning (LPAR", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 720, + 492, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 492, + 733 + ], + "score": 1.0, + "content": "2015), volume 9450 of LNCS, pp. 88–96. Springer, 2015a. doi: 10.1007/978-3-662-48899-7.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Josef Urban. Learning-assisted theorem proving with millions of lemmas.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 432, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 432, + 106 + ], + "score": 1.0, + "content": "J. Symbolic Computation, 69:109–128, 2015b. doi: 10.1016/j.jsc.2014.09.032.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 115, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "Cezary Kaliszyk, Josef Urban, and Jiˇrí Vyskocil. Learning to parse on aligned corpora. In Christian ˇ", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "Urban and Xingyuan Zhang (eds.), Proc. 6h Conference on Interactive Theorem Proving (ITP’15),", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "volume 9236 of LNCS, pp. 227–233. Springer-Verlag, 2015. doi: 10.1007/978-3-319-22102-1_", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 148, + 133, + 160 + ], + "spans": [ + { + "bbox": [ + 115, + 148, + 133, + 160 + ], + "score": 1.0, + "content": "15.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 108, + 170, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "Gerwin Klein, June Andronick, Kevin Elphinstone, Toby C. Murray, Thomas Sewell, Rafal Kolan-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 504, + 193 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 504, + 193 + ], + "score": 1.0, + "content": "ski, and Gernot Heiser. Comprehensive formal verification of an OS microkernel. ACM Trans.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 192, + 236, + 205 + ], + "spans": [ + { + "bbox": [ + 117, + 192, + 236, + 205 + ], + "score": 1.0, + "content": "Comput. Syst., 32(1):2, 2014.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 104, + 212, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 505, + 228 + ], + "score": 1.0, + "content": "Laura Kovács and Andrei Voronkov. First-order theorem proving and Vampire. In Natasha Shary-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "gina and Helmut Veith (eds.), Computer-Aided Verification (CAV 2013), volume 8044 of LNCS,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 236, + 221, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 236, + 221, + 249 + ], + "score": 1.0, + "content": "pp. 1–35. Springer, 2013.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 503, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "score": 1.0, + "content": "Daniel Kühlwein and Josef Urban. MaLeS: A framework for automatic tuning of automated theorem", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 270, + 464, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 464, + 281 + ], + "score": 1.0, + "content": "provers. J. Autom. Reasoning, 55(2):91–116, 2015. doi: 10.1007/s10817-015-9329-1.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 290, + 492, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 492, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 492, + 304 + ], + "score": 1.0, + "content": "Xavier Leroy. Formal verification of a realistic compiler. Commun. ACM, 52(7):107–115, 2009.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "Steven Obua and Sebastian Skalberg. Importing HOL into Isabelle/HOL. In Ulrich Furbach and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 115, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "Natarajan Shankar (eds.), International Joint Conference on Automated Reasoning (IJCAR 2006),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 335, + 329, + 347 + ], + "spans": [ + { + "bbox": [ + 116, + 335, + 329, + 347 + ], + "score": 1.0, + "content": "volume 4130 of LNCS, pp. 298–302. Springer, 2006.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "Lawrence C. Paulson. A generic tableau prover and its integration with Isabelle. J. Universal", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 367, + 268, + 380 + ], + "spans": [ + { + "bbox": [ + 117, + 367, + 268, + 380 + ], + "score": 1.0, + "content": "Computer Science, 5(3):73–87, 1999.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "Stephan Schulz. System description: E 1.8. In Kenneth L. McMillan, Aart Middeldorp, and Andrei", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 117, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 117, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "Voronkov (eds.), Logic for Programming, Artificial Intelligence (LPAR 2013), volume 8312 of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 411, + 264, + 425 + ], + "spans": [ + { + "bbox": [ + 115, + 411, + 264, + 425 + ], + "score": 1.0, + "content": "LNCS, pp. 735–743. Springer, 2013.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 504, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 504, + 446 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Christopher J. Maddison, Arthur Guez, Laurent Sifre, George van den", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lil-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "licrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 115, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529:484–503, 2016. URL", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 489, + 412, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 489, + 412, + 501 + ], + "score": 1.0, + "content": "http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 524 + ], + "score": 1.0, + "content": "Richard Socher, Danqi Chen, Christopher D. Manning, and Andrew Y. Ng. Reasoning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 521, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 115, + 521, + 504, + 534 + ], + "score": 1.0, + "content": "with neural tensor networks for knowledge base completion. In Advances in Neural In-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "formation Processing Systems 26: 27th Annual Conference on Neural Information Pro-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 542, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 114, + 542, + 506, + 558 + ], + "score": 1.0, + "content": "cessing Systems 2013. Proceedings., pp. 926–934, 2013a. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 555, + 434, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 434, + 568 + ], + "score": 1.0, + "content": "5028-reasoning-with-neural-tensor-networks-for-knowledge-base-completion.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 487, + 590 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y.", + "type": "text" + }, + { + "bbox": [ + 487, + 577, + 502, + 588 + ], + "score": 0.27, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 576, + 505, + 590 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 596, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 115, + 596, + 506, + 613 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "Processing, pp. 1631–1642, Stroudsburg, PA, October 2013b. Association for Computational Lin-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 620, + 152, + 633 + ], + "spans": [ + { + "bbox": [ + 114, + 620, + 152, + 633 + ], + "score": 1.0, + "content": "guistics.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 642, + 504, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 656 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. In Advances", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 653, + 379, + 666 + ], + "spans": [ + { + "bbox": [ + 115, + 653, + 379, + 666 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 2431–2439, 2015.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 105, + 675, + 497, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 498, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 498, + 689 + ], + "score": 1.0, + "content": "Geoff Sutcliffe. The CADE ATP system competition - CASC. AI Magazine, 37(2):99–101, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Josef Urban, Jiˇrí Vyskocil, and Petr Št ˇ epánek. MaLeCoP: Machine learning connection prover. In ˇ", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "Kai Brünnler and George Metcalfe (eds.), TABLEAUX 2011, volume 6793 of LNCS. Springer,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 718, + 144, + 733 + ], + "spans": [ + { + "bbox": [ + 114, + 718, + 144, + 733 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "Cezary Kaliszyk and Josef Urban. Learning-assisted theorem proving with millions of lemmas.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 432, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 432, + 106 + ], + "score": 1.0, + "content": "J. Symbolic Computation, 69:109–128, 2015b. doi: 10.1016/j.jsc.2014.09.032.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 504, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 115, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "Cezary Kaliszyk, Josef Urban, and Jiˇrí Vyskocil. Learning to parse on aligned corpora. In Christian ˇ", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "Urban and Xingyuan Zhang (eds.), Proc. 6h Conference on Interactive Theorem Proving (ITP’15),", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "volume 9236 of LNCS, pp. 227–233. Springer-Verlag, 2015. doi: 10.1007/978-3-319-22102-1_", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 148, + 133, + 160 + ], + "spans": [ + { + "bbox": [ + 115, + 148, + 133, + 160 + ], + "score": 1.0, + "content": "15.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 115, + 506, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 170, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 182 + ], + "score": 1.0, + "content": "Gerwin Klein, June Andronick, Kevin Elphinstone, Toby C. Murray, Thomas Sewell, Rafal Kolan-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 182, + 504, + 193 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 504, + 193 + ], + "score": 1.0, + "content": "ski, and Gernot Heiser. Comprehensive formal verification of an OS microkernel. ACM Trans.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 192, + 236, + 205 + ], + "spans": [ + { + "bbox": [ + 117, + 192, + 236, + 205 + ], + "score": 1.0, + "content": "Comput. Syst., 32(1):2, 2014.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 170, + 505, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 214, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 104, + 212, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 505, + 228 + ], + "score": 1.0, + "content": "Laura Kovács and Andrei Voronkov. First-order theorem proving and Vampire. In Natasha Shary-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "gina and Helmut Veith (eds.), Computer-Aided Verification (CAV 2013), volume 8044 of LNCS,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 236, + 221, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 236, + 221, + 249 + ], + "score": 1.0, + "content": "pp. 1–35. Springer, 2013.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 212, + 505, + 249 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 503, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 271 + ], + "score": 1.0, + "content": "Daniel Kühlwein and Josef Urban. MaLeS: A framework for automatic tuning of automated theorem", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 270, + 464, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 464, + 281 + ], + "score": 1.0, + "content": "provers. J. Autom. Reasoning, 55(2):91–116, 2015. doi: 10.1007/s10817-015-9329-1.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 257, + 505, + 281 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 290, + 492, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 492, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 492, + 304 + ], + "score": 1.0, + "content": "Xavier Leroy. Formal verification of a realistic compiler. Commun. ACM, 52(7):107–115, 2009.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 290, + 492, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "Steven Obua and Sebastian Skalberg. Importing HOL into Isabelle/HOL. In Ulrich Furbach and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 324, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 115, + 324, + 505, + 337 + ], + "score": 1.0, + "content": "Natarajan Shankar (eds.), International Joint Conference on Automated Reasoning (IJCAR 2006),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 335, + 329, + 347 + ], + "spans": [ + { + "bbox": [ + 116, + 335, + 329, + 347 + ], + "score": 1.0, + "content": "volume 4130 of LNCS, pp. 298–302. Springer, 2006.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 313, + 505, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "Lawrence C. Paulson. A generic tableau prover and its integration with Isabelle. J. Universal", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 367, + 268, + 380 + ], + "spans": [ + { + "bbox": [ + 117, + 367, + 268, + 380 + ], + "score": 1.0, + "content": "Computer Science, 5(3):73–87, 1999.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 356, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 389, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "Stephan Schulz. System description: E 1.8. In Kenneth L. McMillan, Aart Middeldorp, and Andrei", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 117, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 117, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "Voronkov (eds.), Logic for Programming, Artificial Intelligence (LPAR 2013), volume 8312 of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 411, + 264, + 425 + ], + "spans": [ + { + "bbox": [ + 115, + 411, + 264, + 425 + ], + "score": 1.0, + "content": "LNCS, pp. 735–743. Springer, 2013.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 389, + 505, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 504, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 504, + 446 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Christopher J. Maddison, Arthur Guez, Laurent Sifre, George van den", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lil-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "licrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 115, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529:484–503, 2016. URL", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 489, + 412, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 489, + 412, + 501 + ], + "score": 1.0, + "content": "http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 434, + 506, + 501 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 524 + ], + "score": 1.0, + "content": "Richard Socher, Danqi Chen, Christopher D. Manning, and Andrew Y. Ng. Reasoning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 521, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 115, + 521, + 504, + 534 + ], + "score": 1.0, + "content": "with neural tensor networks for knowledge base completion. In Advances in Neural In-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "formation Processing Systems 26: 27th Annual Conference on Neural Information Pro-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 542, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 114, + 542, + 506, + 558 + ], + "score": 1.0, + "content": "cessing Systems 2013. Proceedings., pp. 926–934, 2013a. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 555, + 434, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 434, + 568 + ], + "score": 1.0, + "content": "5028-reasoning-with-neural-tensor-networks-for-knowledge-base-completion.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 509, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 487, + 590 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y.", + "type": "text" + }, + { + "bbox": [ + 487, + 577, + 502, + 588 + ], + "score": 0.27, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 576, + 505, + 590 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 596, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 115, + 596, + 506, + 613 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "Processing, pp. 1631–1642, Stroudsburg, PA, October 2013b. Association for Computational Lin-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 620, + 152, + 633 + ], + "spans": [ + { + "bbox": [ + 114, + 620, + 152, + 633 + ], + "score": 1.0, + "content": "guistics.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 576, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 642, + 504, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 656 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. In Advances", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 653, + 379, + 666 + ], + "spans": [ + { + "bbox": [ + 115, + 653, + 379, + 666 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 2431–2439, 2015.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 642, + 505, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 675, + 497, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 498, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 498, + 689 + ], + "score": 1.0, + "content": "Geoff Sutcliffe. The CADE ATP system competition - CASC. AI Magazine, 37(2):99–101, 2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 674, + 498, + 689 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Josef Urban, Jiˇrí Vyskocil, and Petr Št ˇ epánek. MaLeCoP: Machine learning connection prover. In ˇ", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "Kai Brünnler and George Metcalfe (eds.), TABLEAUX 2011, volume 6793 of LNCS. Springer,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 718, + 144, + 733 + ], + "spans": [ + { + "bbox": [ + 114, + 718, + 144, + 733 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin John-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "son, Xiaobing Liu, Lukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "machine translation system: Bridging the gap between human and machine translation. CoRR,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 149, + 363, + 161 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 363, + 161 + ], + "score": 1.0, + "content": "abs/1609.08144, 2016. URL http://arxiv.org/abs/1609.08144.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin John-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "son, Xiaobing Liu, Lukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. Google’s neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "machine translation system: Bridging the gap between human and machine translation. CoRR,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 149, + 363, + 161 + ], + "spans": [ + { + "bbox": [ + 115, + 149, + 363, + 161 + ], + "score": 1.0, + "content": "abs/1609.08144, 2016. URL http://arxiv.org/abs/1609.08144.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 81, + 506, + 161 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ryuxYmvel/ryuxYmvel_model.json b/parse/train/ryuxYmvel/ryuxYmvel_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b00a02a62f3cc4d9f1b3a34f605d93783d6ff33c --- /dev/null +++ b/parse/train/ryuxYmvel/ryuxYmvel_model.json @@ -0,0 +1,12730 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1576, + 1404, + 1576, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 398, + 619, + 1302, + 619, + 1302, + 986, + 398, + 986 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1315, + 1404, + 1315, + 1404, + 1560, + 298, + 1560 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1116, + 1403, + 1116, + 1403, + 1300, + 299, + 1300 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1399, + 219, + 1399, + 324, + 301, + 324 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 314, + 376, + 748, + 376, + 748, + 469, + 314, + 469 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 839, + 375, + 1326, + 375, + 1326, + 469, + 839, + 469 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 302, + 1046, + 573, + 1046, + 573, + 1081, + 302, + 1081 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.879 + }, + { + "category_id": 0, + "poly": [ + 773, + 550, + 927, + 550, + 927, + 583, + 773, + 583 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.66 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 216.0, + 1404.0, + 216.0, + 1404.0, + 272.0, + 293.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 275.0, + 1196.0, + 275.0, + 1196.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1043.0, + 579.0, + 1043.0, + 579.0, + 1090.0, + 294.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 546.0, + 934.0, + 546.0, + 934.0, + 589.0, + 768.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1573.0, + 1406.0, + 1573.0, + 1406.0, + 1613.0, + 292.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1608.0, + 1404.0, + 1608.0, + 1404.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 1402.0, + 1637.0, + 1402.0, + 1672.0, + 294.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1667.0, + 1405.0, + 1667.0, + 1405.0, + 1704.0, + 293.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1699.0, + 1405.0, + 1699.0, + 1405.0, + 1734.0, + 294.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1728.0, + 1406.0, + 1728.0, + 1406.0, + 1765.0, + 292.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1404.0, + 1760.0, + 1404.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1786.0, + 1405.0, + 1786.0, + 1405.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1820.0, + 1404.0, + 1820.0, + 1404.0, + 1855.0, + 292.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1948.0, + 292.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 619.0, + 1306.0, + 619.0, + 1306.0, + 654.0, + 393.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 650.0, + 1305.0, + 650.0, + 1305.0, + 686.0, + 394.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 681.0, + 1304.0, + 681.0, + 1304.0, + 716.0, + 395.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 712.0, + 1306.0, + 712.0, + 1306.0, + 746.0, + 393.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 740.0, + 1307.0, + 740.0, + 1307.0, + 779.0, + 391.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 771.0, + 1307.0, + 771.0, + 1307.0, + 810.0, + 393.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 802.0, + 1306.0, + 802.0, + 1306.0, + 837.0, + 393.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 834.0, + 1306.0, + 834.0, + 1306.0, + 869.0, + 393.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 863.0, + 1306.0, + 863.0, + 1306.0, + 898.0, + 393.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 894.0, + 1306.0, + 894.0, + 1306.0, + 928.0, + 393.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 926.0, + 1303.0, + 926.0, + 1303.0, + 958.0, + 395.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 952.0, + 1156.0, + 952.0, + 1156.0, + 993.0, + 392.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1350.0, + 294.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1348.0, + 1405.0, + 1348.0, + 1405.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1380.0, + 1402.0, + 1380.0, + 1402.0, + 1410.0, + 296.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 1406.0, + 1408.0, + 1406.0, + 1441.0, + 294.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1436.0, + 1406.0, + 1436.0, + 1406.0, + 1476.0, + 292.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1469.0, + 1406.0, + 1469.0, + 1406.0, + 1503.0, + 292.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 595.0, + 1530.0, + 595.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1115.0, + 1405.0, + 1115.0, + 1405.0, + 1152.0, + 293.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1404.0, + 1148.0, + 1404.0, + 1180.0, + 295.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1407.0, + 1175.0, + 1407.0, + 1217.0, + 292.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1207.0, + 1407.0, + 1207.0, + 1407.0, + 1246.0, + 294.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1241.0, + 1404.0, + 1241.0, + 1404.0, + 1273.0, + 295.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1270.0, + 914.0, + 1270.0, + 914.0, + 1302.0, + 297.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 376.0, + 515.0, + 376.0, + 515.0, + 410.0, + 312.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 408.0, + 585.0, + 408.0, + 585.0, + 441.0, + 312.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 439.0, + 751.0, + 439.0, + 751.0, + 473.0, + 313.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 372.0, + 1268.0, + 372.0, + 1268.0, + 412.0, + 837.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 409.0, + 1036.0, + 409.0, + 1036.0, + 438.0, + 840.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 438.0, + 1327.0, + 438.0, + 1327.0, + 472.0, + 840.0, + 472.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1131, + 1403, + 1131, + 1403, + 1346, + 298, + 1346 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 492, + 298, + 492 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 841, + 1404, + 841, + 1404, + 1115, + 298, + 1115 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1681, + 1403, + 1681, + 1403, + 1865, + 299, + 1865 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 610, + 1403, + 610, + 1403, + 825, + 298, + 825 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1881, + 1402, + 1881, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1361, + 1403, + 1361, + 1403, + 1483, + 298, + 1483 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1604, + 1401, + 1604, + 1401, + 1665, + 299, + 1665 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1400, + 229, + 1400, + 292, + 297, + 292 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 302, + 1539, + 564, + 1539, + 564, + 1571, + 302, + 1571 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 303, + 547, + 746, + 547, + 746, + 578, + 303, + 578 + ], + "score": 0.864 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.682 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 859, + 2089, + 859, + 2112, + 841, + 2112 + ], + "score": 0.204 + }, + { + "category_id": 13, + "poly": [ + 451, + 1804, + 548, + 1804, + 548, + 1833, + 451, + 1833 + ], + "score": 0.88, + "latex": "4 0 { - } 5 0 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1536.0, + 570.0, + 1536.0, + 570.0, + 1576.0, + 295.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 544.0, + 752.0, + 544.0, + 752.0, + 581.0, + 294.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 838.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1129.0, + 1403.0, + 1129.0, + 1403.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1160.0, + 1404.0, + 1160.0, + 1404.0, + 1199.0, + 293.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1190.0, + 1404.0, + 1190.0, + 1404.0, + 1227.0, + 293.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1221.0, + 1407.0, + 1221.0, + 1407.0, + 1262.0, + 292.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1250.0, + 1407.0, + 1250.0, + 1407.0, + 1290.0, + 292.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1320.0, + 294.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1318.0, + 429.0, + 1318.0, + 429.0, + 1347.0, + 294.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 307.0, + 1405.0, + 307.0, + 1405.0, + 343.0, + 294.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 338.0, + 1405.0, + 338.0, + 1405.0, + 373.0, + 293.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 1404.0, + 367.0, + 1404.0, + 403.0, + 293.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 434.0, + 293.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 432.0, + 1404.0, + 432.0, + 1404.0, + 463.0, + 296.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 458.0, + 1109.0, + 458.0, + 1109.0, + 500.0, + 292.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 839.0, + 1406.0, + 839.0, + 1406.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 870.0, + 1404.0, + 870.0, + 1404.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 903.0, + 1406.0, + 903.0, + 1406.0, + 936.0, + 292.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 933.0, + 1406.0, + 933.0, + 1406.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 961.0, + 1406.0, + 961.0, + 1406.0, + 1001.0, + 291.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 991.0, + 1404.0, + 991.0, + 1404.0, + 1030.0, + 292.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1026.0, + 1404.0, + 1026.0, + 1404.0, + 1059.0, + 294.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1053.0, + 1407.0, + 1053.0, + 1407.0, + 1092.0, + 292.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1085.0, + 539.0, + 1085.0, + 539.0, + 1117.0, + 293.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1680.0, + 1403.0, + 1680.0, + 1403.0, + 1716.0, + 295.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1404.0, + 1713.0, + 1404.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1740.0, + 1407.0, + 1740.0, + 1407.0, + 1779.0, + 292.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1769.0, + 1407.0, + 1769.0, + 1407.0, + 1811.0, + 292.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 450.0, + 1802.0, + 450.0, + 1838.0, + 294.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1838.0, + 549.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1269.0, + 1834.0, + 1269.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 1403.0, + 640.0, + 1403.0, + 679.0, + 293.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1403.0, + 673.0, + 1403.0, + 708.0, + 294.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 700.0, + 1407.0, + 700.0, + 1407.0, + 742.0, + 292.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 734.0, + 1403.0, + 734.0, + 1403.0, + 767.0, + 292.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 762.0, + 1407.0, + 762.0, + 1407.0, + 801.0, + 291.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 793.0, + 379.0, + 793.0, + 379.0, + 829.0, + 292.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1882.0, + 1403.0, + 1882.0, + 1403.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1965.0, + 1406.0, + 1965.0, + 1406.0, + 2013.0, + 291.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1403.0, + 2001.0, + 1403.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1394.0, + 294.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1392.0, + 1406.0, + 1392.0, + 1406.0, + 1425.0, + 292.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1422.0, + 1405.0, + 1422.0, + 1405.0, + 1458.0, + 291.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1456.0, + 594.0, + 1456.0, + 594.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1602.0, + 1405.0, + 1602.0, + 1405.0, + 1638.0, + 295.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1634.0, + 1383.0, + 1634.0, + 1383.0, + 1670.0, + 295.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 263.0, + 610.0, + 263.0, + 610.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1087, + 1404, + 1087, + 1404, + 1363, + 298, + 1363 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1378, + 1403, + 1378, + 1403, + 1835, + 298, + 1835 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 857, + 1403, + 857, + 1403, + 1071, + 299, + 1071 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 628, + 1403, + 628, + 1403, + 841, + 298, + 841 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 309, + 1403, + 309, + 1403, + 491, + 298, + 491 + ], + "score": 0.979 + }, + { + "category_id": 2, + "poly": [ + 297, + 1921, + 1403, + 1921, + 1403, + 2033, + 297, + 2033 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 293, + 230, + 1400, + 230, + 1400, + 292, + 293, + 292 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 301, + 550, + 680, + 550, + 680, + 585, + 301, + 585 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 614, + 1843, + 1088, + 1843, + 1088, + 1873, + 614, + 1873 + ], + "score": 0.813 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.658 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.281 + }, + { + "category_id": 13, + "poly": [ + 1088, + 1592, + 1143, + 1592, + 1143, + 1621, + 1088, + 1621 + ], + "score": 0.86, + "latex": "90 \\%" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1916.0, + 1405.0, + 1916.0, + 1405.0, + 1957.0, + 331.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1948.0, + 1405.0, + 1948.0, + 1405.0, + 1982.0, + 293.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1978.0, + 1404.0, + 1978.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1193.0, + 2003.0, + 1193.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 546.0, + 685.0, + 546.0, + 685.0, + 592.0, + 291.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1122.0, + 295.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1114.0, + 1408.0, + 1114.0, + 1408.0, + 1157.0, + 292.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1148.0, + 1406.0, + 1148.0, + 1406.0, + 1185.0, + 293.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 1405.0, + 1178.0, + 1405.0, + 1215.0, + 294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1211.0, + 1405.0, + 1211.0, + 1405.0, + 1247.0, + 293.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1239.0, + 1406.0, + 1239.0, + 1406.0, + 1276.0, + 293.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1268.0, + 1408.0, + 1268.0, + 1408.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1298.0, + 1406.0, + 1298.0, + 1406.0, + 1336.0, + 292.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1332.0, + 879.0, + 1332.0, + 879.0, + 1369.0, + 294.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1376.0, + 1404.0, + 1376.0, + 1404.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1409.0, + 1403.0, + 1409.0, + 1403.0, + 1444.0, + 293.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1437.0, + 1405.0, + 1437.0, + 1405.0, + 1475.0, + 292.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 1405.0, + 1467.0, + 1405.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1497.0, + 1407.0, + 1497.0, + 1407.0, + 1538.0, + 292.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1529.0, + 1407.0, + 1529.0, + 1407.0, + 1568.0, + 293.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1595.0, + 293.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1590.0, + 1087.0, + 1590.0, + 1087.0, + 1628.0, + 291.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1628.0, + 1144.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1624.0, + 1404.0, + 1624.0, + 1404.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1651.0, + 1405.0, + 1651.0, + 1405.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1682.0, + 1407.0, + 1682.0, + 1407.0, + 1718.0, + 292.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 1407.0, + 1712.0, + 1407.0, + 1751.0, + 293.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1810.0, + 292.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 1404.0, + 1804.0, + 1404.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 856.0, + 1407.0, + 856.0, + 1407.0, + 894.0, + 294.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 889.0, + 1404.0, + 889.0, + 1404.0, + 924.0, + 294.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 917.0, + 1404.0, + 917.0, + 1404.0, + 954.0, + 293.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 949.0, + 1405.0, + 949.0, + 1405.0, + 984.0, + 293.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 981.0, + 1405.0, + 981.0, + 1405.0, + 1013.0, + 293.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1046.0, + 294.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1041.0, + 896.0, + 1041.0, + 896.0, + 1076.0, + 294.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 628.0, + 1403.0, + 628.0, + 1403.0, + 662.0, + 296.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 657.0, + 1405.0, + 657.0, + 1405.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 725.0, + 293.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 720.0, + 1404.0, + 720.0, + 1404.0, + 754.0, + 293.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 1405.0, + 750.0, + 1405.0, + 785.0, + 294.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1405.0, + 780.0, + 1405.0, + 816.0, + 293.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 813.0, + 1046.0, + 813.0, + 1046.0, + 843.0, + 296.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 1405.0, + 306.0, + 1405.0, + 343.0, + 293.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 338.0, + 1405.0, + 338.0, + 1405.0, + 372.0, + 293.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1405.0, + 369.0, + 1405.0, + 404.0, + 294.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 435.0, + 294.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 428.0, + 1405.0, + 428.0, + 1405.0, + 465.0, + 293.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 464.0, + 508.0, + 464.0, + 508.0, + 497.0, + 296.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1403.0, + 229.0, + 1403.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 260.0, + 1029.0, + 260.0, + 1029.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1837.0, + 1089.0, + 1837.0, + 1089.0, + 1880.0, + 608.0, + 1880.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 434, + 224, + 1263, + 224, + 1263, + 444, + 434, + 444 + ], + "score": 0.981, + "html": "
TrainTestPositiveNegative
Examples201304619603011045381104538
Avg. length503.18440.20535.52459.66
Avg. tokens87.0180.6295.4877.40
Conjectures99991411--
Avg. dependencies29.5822.8211
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1728, + 1404, + 1728, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 370, + 793, + 1289, + 793, + 1289, + 1128, + 370, + 1128 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 302, + 1528, + 1404, + 1528, + 1404, + 1621, + 302, + 1621 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 369, + 1258, + 1403, + 1258, + 1403, + 1493, + 369, + 1493 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 296, + 696, + 1400, + 696, + 1400, + 761, + 296, + 761 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 295, + 1161, + 1401, + 1161, + 1401, + 1224, + 295, + 1224 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 300, + 560, + 756, + 560, + 756, + 597, + 300, + 597 + ], + "score": 0.912 + }, + { + "category_id": 6, + "poly": [ + 662, + 465, + 1040, + 465, + 1040, + 498, + 662, + 498 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 815, + 75, + 815, + 105, + 298, + 105 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 300, + 636, + 612, + 636, + 612, + 668, + 300, + 668 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.764 + }, + { + "category_id": 0, + "poly": [ + 301, + 1670, + 1170, + 1670, + 1170, + 1699, + 301, + 1699 + ], + "score": 0.753 + }, + { + "category_id": 1, + "poly": [ + 301, + 1670, + 1170, + 1670, + 1170, + 1699, + 301, + 1699 + ], + "score": 0.161 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 558.0, + 759.0, + 558.0, + 759.0, + 602.0, + 292.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 463.0, + 1041.0, + 463.0, + 1041.0, + 501.0, + 658.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 636.0, + 615.0, + 636.0, + 615.0, + 672.0, + 295.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1666.0, + 1174.0, + 1666.0, + 1174.0, + 1704.0, + 293.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1726.0, + 1405.0, + 1726.0, + 1405.0, + 1763.0, + 292.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1793.0, + 292.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1404.0, + 1790.0, + 1404.0, + 1825.0, + 294.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 1404.0, + 1820.0, + 1404.0, + 1856.0, + 295.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1889.0, + 292.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1918.0, + 294.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1977.0, + 292.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1097.0, + 2003.0, + 1097.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 791.0, + 1231.0, + 791.0, + 1231.0, + 836.0, + 366.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 843.0, + 1168.0, + 843.0, + 1168.0, + 882.0, + 368.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 896.0, + 1154.0, + 896.0, + 1154.0, + 931.0, + 369.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 942.0, + 1290.0, + 942.0, + 1290.0, + 985.0, + 366.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 994.0, + 861.0, + 994.0, + 861.0, + 1032.0, + 367.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1043.0, + 1253.0, + 1043.0, + 1253.0, + 1082.0, + 366.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1095.0, + 1021.0, + 1095.0, + 1021.0, + 1132.0, + 367.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1526.0, + 1404.0, + 1526.0, + 1404.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1559.0, + 1404.0, + 1559.0, + 1404.0, + 1593.0, + 297.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1589.0, + 749.0, + 1589.0, + 749.0, + 1623.0, + 297.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1257.0, + 1405.0, + 1257.0, + 1405.0, + 1294.0, + 367.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1287.0, + 1403.0, + 1287.0, + 1403.0, + 1326.0, + 393.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1322.0, + 1185.0, + 1322.0, + 1185.0, + 1355.0, + 395.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1369.0, + 1405.0, + 1369.0, + 1405.0, + 1403.0, + 383.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1399.0, + 1406.0, + 1399.0, + 1406.0, + 1436.0, + 393.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1431.0, + 1406.0, + 1431.0, + 1406.0, + 1466.0, + 393.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1464.0, + 647.0, + 1464.0, + 647.0, + 1496.0, + 396.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 692.0, + 1404.0, + 692.0, + 1404.0, + 735.0, + 294.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 726.0, + 627.0, + 726.0, + 627.0, + 765.0, + 295.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1191.0, + 1294.0, + 1191.0, + 1294.0, + 1226.0, + 295.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1666.0, + 1174.0, + 1666.0, + 1174.0, + 1704.0, + 293.0, + 1704.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1030, + 1404, + 1030, + 1404, + 1183, + 298, + 1183 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1199, + 1404, + 1199, + 1404, + 1352, + 298, + 1352 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 546, + 289, + 1162, + 289, + 1162, + 897, + 546, + 897 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 369, + 1504, + 1403, + 1504, + 1403, + 1770, + 369, + 1770 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 300, + 1446, + 787, + 1446, + 787, + 1477, + 300, + 1477 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 300, + 961, + 630, + 961, + 630, + 996, + 300, + 996 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 298, + 1865, + 847, + 1865, + 847, + 1895, + 298, + 1895 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 329, + 2005, + 1045, + 2005, + 1045, + 2034, + 329, + 2034 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 297, + 1796, + 972, + 1796, + 972, + 1828, + 297, + 1828 + ], + "score": 0.894 + }, + { + "category_id": 4, + "poly": [ + 517, + 224, + 1180, + 224, + 1180, + 256, + 517, + 256 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 297, + 1390, + 882, + 1390, + 882, + 1420, + 297, + 1420 + ], + "score": 0.875 + }, + { + "category_id": 1, + "poly": [ + 303, + 1921, + 1404, + 1921, + 1404, + 1985, + 303, + 1985 + ], + "score": 0.761 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.748 + }, + { + "category_id": 15, + "poly": [ + 605.0, + 291.0, + 687.0, + 291.0, + 687.0, + 318.0, + 605.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 292.0, + 917.0, + 292.0, + 917.0, + 319.0, + 769.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 290.0, + 1166.0, + 290.0, + 1166.0, + 322.0, + 935.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 312.0, + 940.0, + 312.0, + 940.0, + 346.0, + 549.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 315.0, + 1154.0, + 315.0, + 1154.0, + 344.0, + 960.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 344.0, + 694.0, + 344.0, + 694.0, + 366.0, + 595.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 343.0, + 891.0, + 343.0, + 891.0, + 365.0, + 791.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 342.0, + 1097.0, + 342.0, + 1097.0, + 365.0, + 997.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 395.0, + 697.0, + 395.0, + 697.0, + 438.0, + 598.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 393.0, + 893.0, + 393.0, + 893.0, + 434.0, + 794.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 389.0, + 1104.0, + 389.0, + 1104.0, + 431.0, + 1004.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 448.0, + 681.0, + 448.0, + 681.0, + 471.0, + 612.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 444.0, + 877.0, + 444.0, + 877.0, + 467.0, + 809.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 447.0, + 1091.0, + 447.0, + 1091.0, + 470.0, + 1015.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 467.0, + 672.0, + 467.0, + 672.0, + 484.0, + 618.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 461.0, + 869.0, + 461.0, + 869.0, + 481.0, + 816.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 462.0, + 1090.0, + 462.0, + 1090.0, + 485.0, + 1019.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 477.0, + 676.0, + 477.0, + 676.0, + 500.0, + 617.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 474.0, + 872.0, + 474.0, + 872.0, + 494.0, + 814.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 493.0, + 1127.0, + 493.0, + 1127.0, + 523.0, + 981.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 514.0, + 712.0, + 514.0, + 712.0, + 540.0, + 581.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 509.0, + 909.0, + 509.0, + 909.0, + 536.0, + 776.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 534.0, + 673.0, + 534.0, + 673.0, + 551.0, + 620.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 528.0, + 870.0, + 528.0, + 870.0, + 548.0, + 816.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 545.0, + 679.0, + 545.0, + 679.0, + 565.0, + 614.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 542.0, + 875.0, + 542.0, + 875.0, + 562.0, + 810.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 580.0, + 682.0, + 580.0, + 682.0, + 618.0, + 612.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 576.0, + 876.0, + 576.0, + 876.0, + 613.0, + 808.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 612.0, + 675.0, + 612.0, + 675.0, + 632.0, + 618.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 608.0, + 872.0, + 608.0, + 872.0, + 628.0, + 814.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 646.0, + 729.0, + 646.0, + 729.0, + 675.0, + 562.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 642.0, + 908.0, + 642.0, + 908.0, + 681.0, + 779.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 674.0, + 875.0, + 674.0, + 875.0, + 695.0, + 811.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 684.0, + 676.0, + 684.0, + 676.0, + 711.0, + 617.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 704.0, + 675.0, + 704.0, + 675.0, + 724.0, + 618.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 710.0, + 868.0, + 710.0, + 868.0, + 734.0, + 817.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 728.0, + 871.0, + 728.0, + 871.0, + 748.0, + 813.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 741.0, + 684.0, + 741.0, + 684.0, + 761.0, + 609.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 756.0, + 681.0, + 756.0, + 681.0, + 777.0, + 611.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 762.0, + 873.0, + 762.0, + 873.0, + 787.0, + 814.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 779.0, + 870.0, + 779.0, + 870.0, + 799.0, + 813.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 789.0, + 720.0, + 789.0, + 720.0, + 818.0, + 573.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 816.0, + 881.0, + 816.0, + 881.0, + 839.0, + 805.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 833.0, + 877.0, + 833.0, + 877.0, + 853.0, + 809.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 861.0, + 916.0, + 861.0, + 916.0, + 891.0, + 769.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 958.0, + 634.0, + 958.0, + 634.0, + 1001.0, + 293.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1864.0, + 849.0, + 1864.0, + 849.0, + 1899.0, + 295.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 2002.0, + 1045.0, + 2002.0, + 1045.0, + 2038.0, + 332.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 221.0, + 1183.0, + 221.0, + 1183.0, + 261.0, + 517.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1389.0, + 885.0, + 1389.0, + 885.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 1404.0, + 1029.0, + 1404.0, + 1064.0, + 293.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1059.0, + 1404.0, + 1059.0, + 1404.0, + 1095.0, + 294.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1092.0, + 1404.0, + 1092.0, + 1404.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1122.0, + 1404.0, + 1122.0, + 1404.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1153.0, + 577.0, + 1153.0, + 577.0, + 1186.0, + 294.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1196.0, + 1404.0, + 1196.0, + 1404.0, + 1235.0, + 295.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1225.0, + 1404.0, + 1225.0, + 1404.0, + 1268.0, + 293.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1258.0, + 1405.0, + 1258.0, + 1405.0, + 1297.0, + 292.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1287.0, + 1407.0, + 1287.0, + 1407.0, + 1327.0, + 292.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1322.0, + 1053.0, + 1322.0, + 1053.0, + 1355.0, + 296.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1539.0, + 370.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1536.0, + 1405.0, + 1536.0, + 1405.0, + 1568.0, + 395.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1567.0, + 1405.0, + 1567.0, + 1405.0, + 1599.0, + 395.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1597.0, + 847.0, + 1597.0, + 847.0, + 1629.0, + 396.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1634.0, + 1405.0, + 1634.0, + 1405.0, + 1672.0, + 367.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1667.0, + 1304.0, + 1667.0, + 1304.0, + 1702.0, + 392.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1706.0, + 1405.0, + 1706.0, + 1405.0, + 1743.0, + 367.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1736.0, + 1236.0, + 1736.0, + 1236.0, + 1774.0, + 393.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1445.0, + 791.0, + 1445.0, + 791.0, + 1480.0, + 297.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1796.0, + 974.0, + 1796.0, + 974.0, + 1832.0, + 297.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1918.0, + 1404.0, + 1918.0, + 1404.0, + 1957.0, + 296.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1952.0, + 1403.0, + 1952.0, + 1403.0, + 1988.0, + 297.0, + 1988.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 425, + 302, + 1280, + 302, + 1280, + 1277, + 425, + 1277 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1368, + 1403, + 1368, + 1403, + 1461, + 299, + 1461 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1588, + 1404, + 1588, + 1404, + 1711, + 300, + 1711 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 366, + 1756, + 1404, + 1756, + 1404, + 2035, + 366, + 2035 + ], + "score": 0.914 + }, + { + "category_id": 4, + "poly": [ + 531, + 224, + 1166, + 224, + 1166, + 256, + 531, + 256 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 301, + 1522, + 734, + 1522, + 734, + 1553, + 301, + 1553 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 814, + 76, + 814, + 104, + 298, + 104 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.775 + }, + { + "category_id": 15, + "poly": [ + 504.0, + 306.0, + 761.0, + 306.0, + 761.0, + 338.0, + 504.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 305.0, + 1227.0, + 305.0, + 1227.0, + 336.0, + 970.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 333.0, + 720.0, + 333.0, + 720.0, + 362.0, + 545.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 332.0, + 1185.0, + 332.0, + 1185.0, + 361.0, + 1009.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 364.0, + 608.0, + 364.0, + 608.0, + 389.0, + 501.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 364.0, + 761.0, + 364.0, + 761.0, + 390.0, + 644.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 362.0, + 1048.0, + 362.0, + 1048.0, + 387.0, + 942.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 364.0, + 1263.0, + 364.0, + 1263.0, + 390.0, + 1144.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 422.0, + 607.0, + 422.0, + 607.0, + 456.0, + 502.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 423.0, + 757.0, + 423.0, + 757.0, + 457.0, + 649.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 421.0, + 1046.0, + 421.0, + 1046.0, + 456.0, + 941.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 421.0, + 1254.0, + 421.0, + 1254.0, + 456.0, + 1150.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 446.0, + 586.0, + 446.0, + 586.0, + 468.0, + 522.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 448.0, + 733.0, + 448.0, + 733.0, + 467.0, + 671.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 447.0, + 1024.0, + 447.0, + 1024.0, + 466.0, + 963.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 445.0, + 1233.0, + 445.0, + 1233.0, + 467.0, + 1171.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 470.0, + 698.0, + 470.0, + 698.0, + 487.0, + 679.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 477.0, + 1028.0, + 477.0, + 1028.0, + 503.0, + 956.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 477.0, + 1237.0, + 477.0, + 1237.0, + 503.0, + 1164.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 489.0, + 665.0, + 489.0, + 665.0, + 515.0, + 592.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 496.0, + 1021.0, + 496.0, + 1021.0, + 517.0, + 963.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 496.0, + 1229.0, + 496.0, + 1229.0, + 517.0, + 1172.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 510.0, + 1025.0, + 510.0, + 1025.0, + 532.0, + 962.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 510.0, + 1232.0, + 510.0, + 1232.0, + 531.0, + 1169.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 529.0, + 669.0, + 529.0, + 669.0, + 558.0, + 587.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 549.0, + 663.0, + 549.0, + 663.0, + 571.0, + 593.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 548.0, + 1062.0, + 548.0, + 1062.0, + 574.0, + 923.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 547.0, + 1270.0, + 547.0, + 1270.0, + 576.0, + 1134.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 567.0, + 1022.0, + 567.0, + 1022.0, + 589.0, + 964.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 569.0, + 1228.0, + 569.0, + 1228.0, + 588.0, + 1175.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 583.0, + 707.0, + 583.0, + 707.0, + 615.0, + 548.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 580.0, + 1026.0, + 580.0, + 1026.0, + 602.0, + 958.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 581.0, + 1236.0, + 581.0, + 1236.0, + 604.0, + 1168.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 618.0, + 1029.0, + 618.0, + 1029.0, + 644.0, + 956.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 617.0, + 1239.0, + 617.0, + 1239.0, + 646.0, + 1164.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 639.0, + 1019.0, + 639.0, + 1019.0, + 658.0, + 965.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 639.0, + 1228.0, + 639.0, + 1228.0, + 658.0, + 1174.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 654.0, + 729.0, + 654.0, + 729.0, + 682.0, + 537.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 653.0, + 1025.0, + 653.0, + 1025.0, + 675.0, + 962.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 653.0, + 1234.0, + 653.0, + 1234.0, + 675.0, + 1171.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 680.0, + 720.0, + 680.0, + 720.0, + 708.0, + 543.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 690.0, + 1061.0, + 690.0, + 1061.0, + 719.0, + 924.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 690.0, + 1270.0, + 690.0, + 1270.0, + 719.0, + 1133.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 708.0, + 571.0, + 708.0, + 571.0, + 732.0, + 465.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 707.0, + 806.0, + 707.0, + 806.0, + 732.0, + 689.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 710.0, + 1021.0, + 710.0, + 1021.0, + 730.0, + 963.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 710.0, + 1229.0, + 710.0, + 1229.0, + 730.0, + 1173.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 725.0, + 1027.0, + 725.0, + 1027.0, + 747.0, + 959.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 725.0, + 1237.0, + 725.0, + 1237.0, + 747.0, + 1168.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 764.0, + 573.0, + 764.0, + 573.0, + 811.0, + 467.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 764.0, + 799.0, + 764.0, + 799.0, + 811.0, + 693.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 762.0, + 1024.0, + 762.0, + 1024.0, + 802.0, + 963.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 762.0, + 1232.0, + 762.0, + 1232.0, + 802.0, + 1172.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 821.0, + 556.0, + 821.0, + 556.0, + 847.0, + 483.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 821.0, + 783.0, + 821.0, + 783.0, + 847.0, + 709.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 821.0, + 1132.0, + 821.0, + 1132.0, + 850.0, + 1060.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 841.0, + 546.0, + 841.0, + 546.0, + 860.0, + 492.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 841.0, + 772.0, + 841.0, + 772.0, + 860.0, + 718.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 856.0, + 550.0, + 856.0, + 550.0, + 875.0, + 490.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 856.0, + 775.0, + 856.0, + 775.0, + 875.0, + 715.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 863.0, + 1127.0, + 863.0, + 1127.0, + 892.0, + 1064.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 881.0, + 1126.0, + 881.0, + 1126.0, + 904.0, + 1064.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 893.0, + 587.0, + 893.0, + 587.0, + 918.0, + 451.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 893.0, + 813.0, + 893.0, + 813.0, + 918.0, + 677.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 913.0, + 549.0, + 913.0, + 549.0, + 932.0, + 492.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 913.0, + 773.0, + 913.0, + 773.0, + 932.0, + 718.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 926.0, + 553.0, + 926.0, + 553.0, + 946.0, + 484.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 926.0, + 779.0, + 926.0, + 779.0, + 946.0, + 710.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 921.0, + 1137.0, + 921.0, + 1137.0, + 946.0, + 1056.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 940.0, + 1132.0, + 940.0, + 1132.0, + 961.0, + 1060.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 964.0, + 556.0, + 964.0, + 556.0, + 1003.0, + 484.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 964.0, + 781.0, + 964.0, + 781.0, + 1003.0, + 709.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 973.0, + 1174.0, + 973.0, + 1174.0, + 1006.0, + 1017.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 997.0, + 551.0, + 997.0, + 551.0, + 1019.0, + 489.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 998.0, + 775.0, + 998.0, + 775.0, + 1017.0, + 715.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1030.0, + 610.0, + 1030.0, + 610.0, + 1064.0, + 429.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1031.0, + 835.0, + 1031.0, + 835.0, + 1064.0, + 655.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1082.0, + 671.0, + 1082.0, + 671.0, + 1111.0, + 596.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1124.0, + 665.0, + 1124.0, + 665.0, + 1153.0, + 601.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1143.0, + 665.0, + 1143.0, + 665.0, + 1166.0, + 603.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1180.0, + 675.0, + 1180.0, + 675.0, + 1209.0, + 591.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1199.0, + 670.0, + 1199.0, + 670.0, + 1222.0, + 597.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1234.0, + 712.0, + 1234.0, + 712.0, + 1267.0, + 555.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 221.0, + 1169.0, + 221.0, + 1169.0, + 261.0, + 531.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1520.0, + 737.0, + 1520.0, + 737.0, + 1557.0, + 294.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1367.0, + 1404.0, + 1367.0, + 1404.0, + 1405.0, + 294.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1398.0, + 1403.0, + 1398.0, + 1403.0, + 1435.0, + 294.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1430.0, + 1200.0, + 1430.0, + 1200.0, + 1464.0, + 295.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1586.0, + 1405.0, + 1586.0, + 1405.0, + 1625.0, + 291.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1616.0, + 1405.0, + 1616.0, + 1405.0, + 1655.0, + 291.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1648.0, + 1405.0, + 1648.0, + 1405.0, + 1686.0, + 293.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1679.0, + 394.0, + 1679.0, + 394.0, + 1713.0, + 293.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1757.0, + 1404.0, + 1757.0, + 1404.0, + 1789.0, + 369.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1789.0, + 1405.0, + 1789.0, + 1405.0, + 1822.0, + 397.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1816.0, + 1404.0, + 1816.0, + 1404.0, + 1855.0, + 393.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1848.0, + 865.0, + 1848.0, + 865.0, + 1884.0, + 394.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1911.0, + 1403.0, + 1911.0, + 1403.0, + 1948.0, + 371.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1977.0, + 392.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2009.0, + 393.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 2002.0, + 817.0, + 2002.0, + 817.0, + 2036.0, + 394.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 391, + 526, + 1309, + 526, + 1309, + 684, + 391, + 684 + ], + "score": 0.981, + "html": "
Logistic regressionSiamese 1D CNNSiamese 1D CNN-LSTM
Accuracy with char input0.710.810.83
Accuracy with token input0.710.820.77
" + }, + { + "category_id": 1, + "poly": [ + 299, + 930, + 1403, + 930, + 1403, + 1084, + 299, + 1084 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1425, + 1403, + 1425, + 1403, + 1548, + 299, + 1548 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1644, + 1403, + 1644, + 1403, + 1767, + 299, + 1767 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1099, + 1403, + 1099, + 1403, + 1223, + 298, + 1223 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1239, + 1403, + 1239, + 1403, + 1330, + 300, + 1330 + ], + "score": 0.972 + }, + { + "category_id": 5, + "poly": [ + 391, + 281, + 1310, + 281, + 1310, + 441, + 391, + 441 + ], + "score": 0.97, + "html": "
Logistic regression1D CNN1D CNN-LSTM
Accuracy with char input0.710.820.83
Accuracy with token input0.710.830.77
" + }, + { + "category_id": 1, + "poly": [ + 300, + 805, + 1123, + 805, + 1123, + 837, + 300, + 837 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 299, + 735, + 481, + 735, + 481, + 770, + 299, + 770 + ], + "score": 0.917 + }, + { + "category_id": 6, + "poly": [ + 460, + 467, + 1237, + 467, + 1237, + 501, + 460, + 501 + ], + "score": 0.914 + }, + { + "category_id": 6, + "poly": [ + 445, + 224, + 1255, + 224, + 1255, + 258, + 445, + 258 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 298, + 1589, + 984, + 1589, + 984, + 1618, + 298, + 1618 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 300, + 1811, + 561, + 1811, + 561, + 1847, + 300, + 1847 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 815, + 75, + 815, + 104, + 298, + 104 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 302, + 1370, + 754, + 1370, + 754, + 1399, + 302, + 1399 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 302, + 875, + 824, + 875, + 824, + 905, + 302, + 905 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.637 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.281 + }, + { + "category_id": 13, + "poly": [ + 1037, + 932, + 1091, + 932, + 1091, + 961, + 1037, + 961 + ], + "score": 0.88, + "latex": "71 \\%" + }, + { + "category_id": 13, + "poly": [ + 1165, + 1101, + 1218, + 1101, + 1218, + 1130, + 1165, + 1130 + ], + "score": 0.88, + "latex": "83 \\%" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1101, + 1133, + 1101, + 1133, + 1130, + 1079, + 1130 + ], + "score": 0.88, + "latex": "82 \\%" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 731.0, + 484.0, + 731.0, + 484.0, + 778.0, + 291.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 462.0, + 1238.0, + 462.0, + 1238.0, + 507.0, + 460.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 219.0, + 1256.0, + 219.0, + 1256.0, + 264.0, + 443.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1585.0, + 988.0, + 1585.0, + 988.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1804.0, + 566.0, + 1804.0, + 566.0, + 1855.0, + 291.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1368.0, + 756.0, + 1368.0, + 756.0, + 1404.0, + 295.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 874.0, + 830.0, + 874.0, + 830.0, + 908.0, + 295.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 930.0, + 1036.0, + 930.0, + 1036.0, + 967.0, + 295.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 930.0, + 1403.0, + 930.0, + 1403.0, + 967.0, + 1092.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 959.0, + 1405.0, + 959.0, + 1405.0, + 996.0, + 295.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 991.0, + 1404.0, + 991.0, + 1404.0, + 1027.0, + 295.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1024.0, + 1404.0, + 1024.0, + 1404.0, + 1056.0, + 294.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1055.0, + 1350.0, + 1055.0, + 1350.0, + 1088.0, + 295.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1460.0, + 295.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1405.0, + 1457.0, + 1405.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1489.0, + 1405.0, + 1489.0, + 1405.0, + 1521.0, + 293.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1519.0, + 1259.0, + 1519.0, + 1259.0, + 1552.0, + 295.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1680.0, + 293.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1672.0, + 1407.0, + 1672.0, + 1407.0, + 1710.0, + 293.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1703.0, + 1404.0, + 1703.0, + 1404.0, + 1742.0, + 294.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1736.0, + 1398.0, + 1736.0, + 1398.0, + 1772.0, + 294.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1879.0, + 1406.0, + 1879.0, + 1406.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1945.0, + 297.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1407.0, + 1941.0, + 1407.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1100.0, + 1078.0, + 1100.0, + 1078.0, + 1133.0, + 296.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1100.0, + 1164.0, + 1100.0, + 1164.0, + 1133.0, + 1134.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1100.0, + 1402.0, + 1100.0, + 1402.0, + 1133.0, + 1219.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1132.0, + 1405.0, + 1132.0, + 1405.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 1404.0, + 1163.0, + 1404.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1192.0, + 455.0, + 1192.0, + 455.0, + 1225.0, + 295.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1237.0, + 1404.0, + 1237.0, + 1404.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1304.0, + 296.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1300.0, + 853.0, + 1300.0, + 853.0, + 1331.0, + 295.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 802.0, + 1125.0, + 802.0, + 1125.0, + 842.0, + 296.0, + 842.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1589, + 1404, + 1589, + 1404, + 1835, + 298, + 1835 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1402, + 1850, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1218, + 1404, + 1218, + 1404, + 1463, + 298, + 1463 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 358, + 225, + 774, + 225, + 774, + 566, + 358, + 566 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 902, + 708, + 1313, + 708, + 1313, + 1049, + 902, + 1049 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 902, + 225, + 1314, + 225, + 1314, + 566, + 902, + 566 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 361, + 709, + 773, + 709, + 773, + 1048, + 361, + 1048 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 865, + 1075, + 1398, + 1075, + 1398, + 1136, + 865, + 1136 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 298, + 1074, + 829, + 1074, + 829, + 1135, + 298, + 1135 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 864, + 593, + 1400, + 593, + 1400, + 655, + 864, + 655 + ], + "score": 0.947 + }, + { + "category_id": 4, + "poly": [ + 297, + 592, + 831, + 592, + 831, + 655, + 297, + 655 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 299, + 1523, + 543, + 1523, + 543, + 1554, + 299, + 1554 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.792 + }, + { + "category_id": 15, + "poly": [ + 390.0, + 227.0, + 418.0, + 227.0, + 418.0, + 246.0, + 390.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 259.0, + 643.0, + 259.0, + 643.0, + 279.0, + 521.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 274.0, + 624.0, + 274.0, + 624.0, + 297.0, + 521.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 302.0, + 418.0, + 302.0, + 418.0, + 321.0, + 389.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 309.0, + 386.0, + 309.0, + 386.0, + 455.0, + 361.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 316.0, + 678.0, + 316.0, + 678.0, + 337.0, + 516.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 333.0, + 617.0, + 333.0, + 617.0, + 354.0, + 516.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 377.0, + 418.0, + 377.0, + 418.0, + 395.0, + 389.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 381.0, + 766.0, + 381.0, + 766.0, + 403.0, + 585.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 397.0, + 688.0, + 397.0, + 688.0, + 420.0, + 586.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 451.0, + 420.0, + 451.0, + 420.0, + 471.0, + 388.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 523.0, + 425.0, + 523.0, + 425.0, + 555.0, + 387.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 533.0, + 481.0, + 533.0, + 481.0, + 553.0, + 457.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 532.0, + 537.0, + 532.0, + 537.0, + 552.0, + 513.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 533.0, + 591.0, + 533.0, + 591.0, + 552.0, + 567.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 532.0, + 645.0, + 532.0, + 645.0, + 552.0, + 621.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 532.0, + 699.0, + 532.0, + 699.0, + 553.0, + 676.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 533.0, + 754.0, + 533.0, + 754.0, + 553.0, + 731.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 545.0, + 694.0, + 545.0, + 694.0, + 567.0, + 486.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 709.0, + 958.0, + 709.0, + 958.0, + 731.0, + 927.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 745.0, + 1308.0, + 745.0, + 1308.0, + 767.0, + 1192.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 763.0, + 1277.0, + 763.0, + 1277.0, + 783.0, + 1193.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 785.0, + 957.0, + 785.0, + 957.0, + 804.0, + 928.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 797.0, + 928.0, + 797.0, + 928.0, + 942.0, + 905.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 818.0, + 1308.0, + 818.0, + 1308.0, + 837.0, + 1151.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 833.0, + 1238.0, + 833.0, + 1238.0, + 854.0, + 1151.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 860.0, + 956.0, + 860.0, + 956.0, + 878.0, + 928.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 863.0, + 1312.0, + 863.0, + 1312.0, + 885.0, + 1192.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 875.0, + 1279.0, + 875.0, + 1279.0, + 903.0, + 1191.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 935.0, + 957.0, + 935.0, + 957.0, + 953.0, + 928.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1006.0, + 964.0, + 1006.0, + 964.0, + 1038.0, + 926.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1016.0, + 1020.0, + 1016.0, + 1020.0, + 1036.0, + 996.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1015.0, + 1075.0, + 1015.0, + 1075.0, + 1035.0, + 1052.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1015.0, + 1128.0, + 1015.0, + 1128.0, + 1035.0, + 1106.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1015.0, + 1183.0, + 1015.0, + 1183.0, + 1035.0, + 1160.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1014.0, + 1239.0, + 1014.0, + 1239.0, + 1038.0, + 1215.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1016.0, + 1293.0, + 1016.0, + 1293.0, + 1036.0, + 1269.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1027.0, + 1231.0, + 1027.0, + 1231.0, + 1049.0, + 1025.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 226.0, + 957.0, + 226.0, + 957.0, + 248.0, + 927.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 252.0, + 1270.0, + 252.0, + 1270.0, + 275.0, + 1211.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 270.0, + 1311.0, + 270.0, + 1311.0, + 292.0, + 1211.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 302.0, + 957.0, + 302.0, + 957.0, + 321.0, + 928.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 297.0, + 1280.0, + 297.0, + 1280.0, + 320.0, + 1180.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 314.0, + 928.0, + 314.0, + 928.0, + 459.0, + 905.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 315.0, + 1280.0, + 315.0, + 1280.0, + 337.0, + 1181.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 341.0, + 1286.0, + 341.0, + 1286.0, + 362.0, + 1167.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 357.0, + 1265.0, + 357.0, + 1265.0, + 379.0, + 1168.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 377.0, + 955.0, + 377.0, + 955.0, + 395.0, + 928.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 452.0, + 957.0, + 452.0, + 957.0, + 470.0, + 929.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 523.0, + 964.0, + 523.0, + 964.0, + 555.0, + 926.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 533.0, + 1020.0, + 533.0, + 1020.0, + 553.0, + 996.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 532.0, + 1075.0, + 532.0, + 1075.0, + 552.0, + 1052.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 532.0, + 1129.0, + 532.0, + 1129.0, + 552.0, + 1106.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 532.0, + 1184.0, + 532.0, + 1184.0, + 552.0, + 1161.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 532.0, + 1238.0, + 532.0, + 1238.0, + 553.0, + 1215.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 533.0, + 1293.0, + 533.0, + 1293.0, + 553.0, + 1270.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 545.0, + 1233.0, + 545.0, + 1233.0, + 567.0, + 1026.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 709.0, + 419.0, + 709.0, + 419.0, + 731.0, + 388.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 722.0, + 751.0, + 722.0, + 751.0, + 743.0, + 629.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 737.0, + 726.0, + 737.0, + 726.0, + 759.0, + 629.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 754.0, + 713.0, + 754.0, + 713.0, + 777.0, + 630.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 784.0, + 417.0, + 784.0, + 417.0, + 803.0, + 389.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 794.0, + 772.0, + 794.0, + 772.0, + 815.0, + 597.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 795.0, + 388.0, + 795.0, + 388.0, + 940.0, + 365.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 811.0, + 682.0, + 811.0, + 682.0, + 830.0, + 599.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 859.0, + 416.0, + 859.0, + 416.0, + 878.0, + 389.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 865.0, + 759.0, + 865.0, + 759.0, + 886.0, + 580.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 882.0, + 665.0, + 882.0, + 665.0, + 901.0, + 581.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 933.0, + 418.0, + 933.0, + 418.0, + 952.0, + 390.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1006.0, + 425.0, + 1006.0, + 425.0, + 1037.0, + 387.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1014.0, + 482.0, + 1014.0, + 482.0, + 1036.0, + 456.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1015.0, + 536.0, + 1015.0, + 536.0, + 1034.0, + 512.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1015.0, + 590.0, + 1015.0, + 590.0, + 1033.0, + 567.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1015.0, + 644.0, + 1015.0, + 644.0, + 1034.0, + 621.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1015.0, + 698.0, + 1015.0, + 698.0, + 1035.0, + 675.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1016.0, + 753.0, + 1016.0, + 753.0, + 1035.0, + 730.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1026.0, + 691.0, + 1026.0, + 691.0, + 1048.0, + 485.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1073.0, + 1401.0, + 1073.0, + 1401.0, + 1108.0, + 865.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1100.0, + 1322.0, + 1100.0, + 1322.0, + 1140.0, + 865.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1073.0, + 833.0, + 1073.0, + 833.0, + 1108.0, + 298.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1102.0, + 794.0, + 1102.0, + 794.0, + 1139.0, + 297.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 592.0, + 1399.0, + 592.0, + 1399.0, + 626.0, + 866.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 619.0, + 1323.0, + 619.0, + 1323.0, + 657.0, + 865.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 591.0, + 832.0, + 591.0, + 832.0, + 626.0, + 298.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 619.0, + 794.0, + 619.0, + 794.0, + 657.0, + 298.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1519.0, + 549.0, + 1519.0, + 549.0, + 1560.0, + 293.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1402.0, + 1591.0, + 1402.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1618.0, + 1406.0, + 1618.0, + 1406.0, + 1658.0, + 291.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1686.0, + 296.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1406.0, + 1682.0, + 1406.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1747.0, + 294.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1406.0, + 1743.0, + 1406.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1772.0, + 1404.0, + 1772.0, + 1404.0, + 1812.0, + 292.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1803.0, + 1004.0, + 1803.0, + 1004.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1406.0, + 1848.0, + 1406.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1914.0, + 293.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2003.0, + 773.0, + 2003.0, + 773.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1218.0, + 1405.0, + 1218.0, + 1405.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1249.0, + 1406.0, + 1249.0, + 1406.0, + 1282.0, + 293.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1279.0, + 1404.0, + 1279.0, + 1404.0, + 1313.0, + 294.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1311.0, + 1405.0, + 1311.0, + 1405.0, + 1343.0, + 293.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1336.0, + 1405.0, + 1336.0, + 1405.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1369.0, + 1406.0, + 1369.0, + 1406.0, + 1406.0, + 292.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1435.0, + 292.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1432.0, + 1061.0, + 1432.0, + 1061.0, + 1465.0, + 296.0, + 1465.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 301, + 416, + 1404, + 416, + 1404, + 691, + 301, + 691 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 300, + 288, + 1094, + 288, + 1094, + 319, + 300, + 319 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 296, + 1417, + 1400, + 1417, + 1400, + 1510, + 296, + 1510 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 298, + 711, + 1405, + 711, + 1405, + 803, + 298, + 803 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 297, + 1272, + 1403, + 1272, + 1403, + 1399, + 297, + 1399 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 297, + 823, + 1404, + 823, + 1404, + 949, + 297, + 949 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 297, + 1161, + 1399, + 1161, + 1399, + 1254, + 297, + 1254 + ], + "score": 0.868 + }, + { + "category_id": 1, + "poly": [ + 296, + 1529, + 1405, + 1529, + 1405, + 1623, + 296, + 1623 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 296, + 1642, + 1019, + 1642, + 1019, + 1675, + 296, + 1675 + ], + "score": 0.86 + }, + { + "category_id": 1, + "poly": [ + 294, + 1694, + 1401, + 1694, + 1401, + 1757, + 294, + 1757 + ], + "score": 0.854 + }, + { + "category_id": 1, + "poly": [ + 289, + 1079, + 1400, + 1079, + 1400, + 1142, + 289, + 1142 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 298, + 966, + 1401, + 966, + 1401, + 1060, + 298, + 1060 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 294, + 1775, + 1401, + 1775, + 1401, + 1839, + 294, + 1839 + ], + "score": 0.82 + }, + { + "category_id": 1, + "poly": [ + 290, + 1858, + 1400, + 1858, + 1400, + 1922, + 290, + 1922 + ], + "score": 0.82 + }, + { + "category_id": 0, + "poly": [ + 300, + 366, + 488, + 366, + 488, + 397, + 300, + 397 + ], + "score": 0.818 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1403, + 1942, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.797 + }, + { + "category_id": 0, + "poly": [ + 301, + 233, + 571, + 233, + 571, + 260, + 301, + 260 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.768 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 363.0, + 491.0, + 363.0, + 491.0, + 402.0, + 295.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 233.0, + 574.0, + 233.0, + 574.0, + 263.0, + 298.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 414.0, + 1403.0, + 414.0, + 1403.0, + 451.0, + 295.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 445.0, + 1404.0, + 445.0, + 1404.0, + 481.0, + 321.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 477.0, + 1405.0, + 477.0, + 1405.0, + 513.0, + 321.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 543.0, + 322.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 538.0, + 1405.0, + 538.0, + 1405.0, + 574.0, + 322.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 568.0, + 1404.0, + 568.0, + 1404.0, + 604.0, + 322.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 596.0, + 1404.0, + 596.0, + 1404.0, + 636.0, + 321.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 629.0, + 1404.0, + 629.0, + 1404.0, + 665.0, + 321.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 657.0, + 430.0, + 657.0, + 430.0, + 699.0, + 321.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 285.0, + 1098.0, + 285.0, + 1098.0, + 324.0, + 296.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1415.0, + 1404.0, + 1415.0, + 1404.0, + 1454.0, + 292.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1449.0, + 1404.0, + 1449.0, + 1404.0, + 1483.0, + 321.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1480.0, + 659.0, + 1480.0, + 659.0, + 1511.0, + 325.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 709.0, + 1402.0, + 709.0, + 1402.0, + 746.0, + 293.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 741.0, + 1403.0, + 741.0, + 1403.0, + 777.0, + 322.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 774.0, + 894.0, + 774.0, + 894.0, + 804.0, + 326.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1272.0, + 1404.0, + 1272.0, + 1404.0, + 1310.0, + 291.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1302.0, + 1405.0, + 1302.0, + 1405.0, + 1341.0, + 320.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1334.0, + 1404.0, + 1334.0, + 1404.0, + 1370.0, + 320.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1367.0, + 783.0, + 1367.0, + 783.0, + 1401.0, + 322.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 824.0, + 1405.0, + 824.0, + 1405.0, + 860.0, + 295.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 854.0, + 1406.0, + 854.0, + 1406.0, + 891.0, + 321.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 884.0, + 1405.0, + 884.0, + 1405.0, + 924.0, + 320.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 914.0, + 1258.0, + 914.0, + 1258.0, + 951.0, + 322.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1198.0, + 292.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1228.0, + 322.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1223.0, + 524.0, + 1223.0, + 524.0, + 1255.0, + 325.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1530.0, + 1402.0, + 1530.0, + 1402.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1595.0, + 323.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1589.0, + 1061.0, + 1589.0, + 1061.0, + 1627.0, + 324.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1641.0, + 1021.0, + 1641.0, + 1021.0, + 1680.0, + 294.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1692.0, + 1404.0, + 1692.0, + 1404.0, + 1732.0, + 294.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1724.0, + 993.0, + 1724.0, + 993.0, + 1760.0, + 326.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 1402.0, + 1076.0, + 1402.0, + 1116.0, + 293.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1111.0, + 625.0, + 1111.0, + 625.0, + 1142.0, + 324.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 967.0, + 1405.0, + 967.0, + 1405.0, + 1001.0, + 296.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 994.0, + 1405.0, + 994.0, + 1405.0, + 1035.0, + 321.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1028.0, + 1002.0, + 1028.0, + 1002.0, + 1061.0, + 323.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1813.0, + 292.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1807.0, + 729.0, + 1807.0, + 729.0, + 1839.0, + 322.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1858.0, + 1403.0, + 1858.0, + 1403.0, + 1894.0, + 294.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1890.0, + 637.0, + 1890.0, + 637.0, + 1923.0, + 322.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 322.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 1218.0, + 2002.0, + 1218.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 297, + 1268, + 1401, + 1268, + 1401, + 1332, + 297, + 1332 + ], + "score": 0.729 + }, + { + "category_id": 1, + "poly": [ + 296, + 1186, + 1402, + 1186, + 1402, + 1250, + 296, + 1250 + ], + "score": 0.727 + }, + { + "category_id": 1, + "poly": [ + 298, + 1350, + 1401, + 1350, + 1401, + 1443, + 298, + 1443 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 296, + 993, + 1402, + 993, + 1402, + 1056, + 296, + 1056 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 297, + 1462, + 1405, + 1462, + 1405, + 1586, + 297, + 1586 + ], + "score": 0.654 + }, + { + "category_id": 1, + "poly": [ + 296, + 1073, + 1403, + 1073, + 1403, + 1168, + 296, + 1168 + ], + "score": 0.654 + }, + { + "category_id": 1, + "poly": [ + 302, + 1603, + 1402, + 1603, + 1402, + 1700, + 302, + 1700 + ], + "score": 0.654 + }, + { + "category_id": 1, + "poly": [ + 298, + 880, + 1399, + 880, + 1399, + 976, + 298, + 976 + ], + "score": 0.624 + }, + { + "category_id": 1, + "poly": [ + 291, + 1828, + 1399, + 1828, + 1399, + 1892, + 291, + 1892 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 298, + 1716, + 1401, + 1716, + 1401, + 1812, + 298, + 1812 + ], + "score": 0.608 + }, + { + "category_id": 1, + "poly": [ + 295, + 596, + 1403, + 596, + 1403, + 690, + 295, + 690 + ], + "score": 0.546 + }, + { + "category_id": 1, + "poly": [ + 297, + 708, + 1404, + 708, + 1404, + 863, + 297, + 863 + ], + "score": 0.537 + }, + { + "category_id": 1, + "poly": [ + 296, + 514, + 1401, + 514, + 1401, + 578, + 296, + 578 + ], + "score": 0.523 + }, + { + "category_id": 1, + "poly": [ + 297, + 402, + 1397, + 402, + 1397, + 496, + 297, + 496 + ], + "score": 0.519 + }, + { + "category_id": 1, + "poly": [ + 298, + 1910, + 1404, + 1910, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.493 + }, + { + "category_id": 1, + "poly": [ + 297, + 230, + 1406, + 230, + 1406, + 384, + 297, + 384 + ], + "score": 0.477 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1265.0, + 1405.0, + 1265.0, + 1405.0, + 1306.0, + 292.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1297.0, + 533.0, + 1297.0, + 533.0, + 1333.0, + 323.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1186.0, + 1404.0, + 1186.0, + 1404.0, + 1222.0, + 294.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1218.0, + 1090.0, + 1218.0, + 1090.0, + 1251.0, + 320.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 1405.0, + 1348.0, + 1405.0, + 1387.0, + 293.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1418.0, + 322.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1407.0, + 401.0, + 1407.0, + 401.0, + 1446.0, + 319.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 990.0, + 1404.0, + 990.0, + 1404.0, + 1030.0, + 292.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1025.0, + 781.0, + 1025.0, + 781.0, + 1056.0, + 324.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1463.0, + 1403.0, + 1463.0, + 1403.0, + 1496.0, + 295.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1491.0, + 1403.0, + 1491.0, + 1403.0, + 1530.0, + 323.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1523.0, + 1405.0, + 1523.0, + 1405.0, + 1560.0, + 321.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1552.0, + 1022.0, + 1552.0, + 1022.0, + 1592.0, + 321.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1070.0, + 1404.0, + 1070.0, + 1404.0, + 1113.0, + 291.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1105.0, + 1404.0, + 1105.0, + 1404.0, + 1141.0, + 323.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1136.0, + 497.0, + 1136.0, + 497.0, + 1168.0, + 323.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1605.0, + 1403.0, + 1605.0, + 1403.0, + 1639.0, + 296.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1638.0, + 1403.0, + 1638.0, + 1403.0, + 1669.0, + 324.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1666.0, + 765.0, + 1666.0, + 765.0, + 1701.0, + 323.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 878.0, + 1404.0, + 878.0, + 1404.0, + 919.0, + 291.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 913.0, + 1403.0, + 913.0, + 1403.0, + 948.0, + 323.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 941.0, + 800.0, + 941.0, + 800.0, + 979.0, + 322.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1824.0, + 1402.0, + 1824.0, + 1402.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1859.0, + 1093.0, + 1859.0, + 1093.0, + 1893.0, + 322.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1754.0, + 293.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1784.0, + 323.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1779.0, + 887.0, + 1779.0, + 887.0, + 1813.0, + 324.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 593.0, + 1405.0, + 593.0, + 1405.0, + 632.0, + 295.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 626.0, + 1405.0, + 626.0, + 1405.0, + 662.0, + 321.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 660.0, + 621.0, + 660.0, + 621.0, + 690.0, + 322.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 1405.0, + 704.0, + 1405.0, + 747.0, + 294.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 736.0, + 1404.0, + 736.0, + 1404.0, + 778.0, + 321.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 770.0, + 1404.0, + 770.0, + 1404.0, + 807.0, + 322.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 835.0, + 321.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 832.0, + 1016.0, + 832.0, + 1016.0, + 865.0, + 321.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 514.0, + 1406.0, + 514.0, + 1406.0, + 550.0, + 296.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 546.0, + 1088.0, + 546.0, + 1088.0, + 578.0, + 323.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 403.0, + 1402.0, + 403.0, + 1402.0, + 437.0, + 294.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 432.0, + 1402.0, + 432.0, + 1402.0, + 468.0, + 322.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 465.0, + 1043.0, + 465.0, + 1043.0, + 499.0, + 323.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1977.0, + 320.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1973.0, + 1408.0, + 1973.0, + 1408.0, + 2009.0, + 320.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2001.0, + 1367.0, + 2001.0, + 1367.0, + 2037.0, + 321.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1406.0, + 259.0, + 1406.0, + 298.0, + 321.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 288.0, + 1406.0, + 288.0, + 1406.0, + 327.0, + 321.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 316.0, + 1407.0, + 316.0, + 1407.0, + 360.0, + 320.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 353.0, + 1378.0, + 353.0, + 1378.0, + 386.0, + 323.0, + 386.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1205, + 1404, + 1205, + 1404, + 1390, + 297, + 1390 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 297, + 1419, + 1404, + 1419, + 1404, + 1574, + 297, + 1574 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 297, + 1082, + 1403, + 1082, + 1403, + 1178, + 297, + 1178 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 296, + 595, + 1405, + 595, + 1405, + 689, + 296, + 689 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 297, + 1602, + 1404, + 1602, + 1404, + 1757, + 297, + 1757 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 297, + 321, + 1403, + 321, + 1403, + 444, + 297, + 444 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 298, + 869, + 1404, + 869, + 1404, + 965, + 298, + 965 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 296, + 992, + 1402, + 992, + 1402, + 1055, + 296, + 1055 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 298, + 1940, + 1402, + 1940, + 1402, + 2033, + 298, + 2033 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1400, + 229, + 1400, + 294, + 294, + 294 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 302, + 473, + 1400, + 473, + 1400, + 567, + 302, + 567 + ], + "score": 0.896 + }, + { + "category_id": 1, + "poly": [ + 297, + 1786, + 1400, + 1786, + 1400, + 1849, + 297, + 1849 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 298, + 717, + 1398, + 717, + 1398, + 781, + 298, + 781 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 294, + 808, + 1368, + 808, + 1368, + 843, + 294, + 843 + ], + "score": 0.864 + }, + { + "category_id": 1, + "poly": [ + 294, + 1875, + 1383, + 1875, + 1383, + 1910, + 294, + 1910 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.817 + }, + { + "category_id": 13, + "poly": [ + 1355, + 1604, + 1395, + 1604, + 1395, + 1635, + 1355, + 1635 + ], + "score": 0.27, + "latex": "\\mathrm { N g }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1207.0, + 1402.0, + 1207.0, + 1402.0, + 1239.0, + 296.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1235.0, + 1405.0, + 1235.0, + 1405.0, + 1272.0, + 322.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1300.0, + 322.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1296.0, + 1405.0, + 1296.0, + 1405.0, + 1334.0, + 321.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1328.0, + 1407.0, + 1328.0, + 1407.0, + 1364.0, + 322.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1359.0, + 1146.0, + 1359.0, + 1146.0, + 1393.0, + 321.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1416.0, + 1404.0, + 1416.0, + 1404.0, + 1457.0, + 295.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1449.0, + 1402.0, + 1449.0, + 1402.0, + 1485.0, + 322.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1515.0, + 320.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1508.0, + 1406.0, + 1508.0, + 1406.0, + 1550.0, + 319.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1542.0, + 1208.0, + 1542.0, + 1208.0, + 1579.0, + 322.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1083.0, + 1404.0, + 1083.0, + 1404.0, + 1118.0, + 296.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1114.0, + 1405.0, + 1114.0, + 1405.0, + 1148.0, + 325.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1142.0, + 734.0, + 1142.0, + 734.0, + 1181.0, + 322.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 591.0, + 1403.0, + 591.0, + 1403.0, + 634.0, + 291.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 626.0, + 1405.0, + 626.0, + 1405.0, + 663.0, + 320.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 658.0, + 615.0, + 658.0, + 615.0, + 692.0, + 320.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 1354.0, + 1600.0, + 1354.0, + 1640.0, + 294.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1600.0, + 1405.0, + 1600.0, + 1405.0, + 1640.0, + 1396.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1634.0, + 1404.0, + 1634.0, + 1404.0, + 1668.0, + 324.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1658.0, + 1406.0, + 1658.0, + 1406.0, + 1703.0, + 321.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1693.0, + 1404.0, + 1693.0, + 1404.0, + 1730.0, + 321.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1724.0, + 423.0, + 1724.0, + 423.0, + 1759.0, + 318.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 358.0, + 295.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 352.0, + 1407.0, + 352.0, + 1407.0, + 388.0, + 321.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 383.0, + 1407.0, + 383.0, + 1407.0, + 419.0, + 322.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 412.0, + 370.0, + 412.0, + 370.0, + 446.0, + 321.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 870.0, + 1405.0, + 870.0, + 1405.0, + 905.0, + 296.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 937.0, + 322.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 932.0, + 915.0, + 932.0, + 915.0, + 966.0, + 323.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 990.0, + 1407.0, + 990.0, + 1407.0, + 1030.0, + 292.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1022.0, + 746.0, + 1022.0, + 746.0, + 1057.0, + 325.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1404.0, + 1939.0, + 1404.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2011.0, + 321.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1997.0, + 402.0, + 1997.0, + 402.0, + 2038.0, + 319.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1402.0, + 229.0, + 1402.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1201.0, + 259.0, + 1201.0, + 295.0, + 321.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 474.0, + 1403.0, + 474.0, + 1403.0, + 508.0, + 297.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 507.0, + 1401.0, + 507.0, + 1401.0, + 537.0, + 324.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 535.0, + 658.0, + 535.0, + 658.0, + 570.0, + 326.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1784.0, + 1404.0, + 1784.0, + 1404.0, + 1823.0, + 295.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1814.0, + 1055.0, + 1814.0, + 1055.0, + 1852.0, + 322.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 716.0, + 1403.0, + 716.0, + 1403.0, + 753.0, + 294.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 750.0, + 1289.0, + 750.0, + 1289.0, + 781.0, + 320.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 807.0, + 1368.0, + 807.0, + 1368.0, + 847.0, + 294.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1873.0, + 1386.0, + 1873.0, + 1386.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 301, + 229, + 1403, + 229, + 1403, + 444, + 301, + 444 + ], + "score": 0.97 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 104, + 300, + 104 + ], + "score": 0.833 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.833 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 268.0, + 296.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1403.0, + 260.0, + 1403.0, + 295.0, + 322.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 288.0, + 1407.0, + 288.0, + 1407.0, + 329.0, + 320.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 357.0, + 321.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 352.0, + 1404.0, + 352.0, + 1404.0, + 387.0, + 322.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 383.0, + 1405.0, + 383.0, + 1405.0, + 418.0, + 321.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 414.0, + 1011.0, + 414.0, + 1011.0, + 448.0, + 322.0, + 448.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/xzqLpqRzxLq/xzqLpqRzxLq.md b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq.md new file mode 100644 index 0000000000000000000000000000000000000000..a6987e56c8be5772a735eb10349ada719c984c99 --- /dev/null +++ b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq.md @@ -0,0 +1,406 @@ +# IEPT: INSTANCE-LEVEL AND EPISODE-LEVEL PRETEXT TASKS FOR FEW-SHOT LEARNING + +Manli Zhang, Jianhong Zhang & Zhiwu Lu ∗ +Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China {manlizhang,jianhong,luzhiwu}@ruc.edu.cn + +Tao Xiang University of Surrey, Guildford, Surrey, UK t.xiang@surrey.ac.uk + +Mingyu Ding +The University of Hong Kong, Hong Kong +mingyuding@hku.hk + +# Songfang Huang + +Alibaba DAMO Academy, Hangzhou, China songfang.hsf@alibaba-inc.com + +# ABSTRACT + +The need of collecting large quantities of labeled training data for each new task has limited the usefulness of deep neural networks. Given data from a set of source tasks, this limitation can be overcome using two transfer learning approaches: few-shot learning (FSL) and self-supervised learning (SSL). The former aims to learn ‘how to learn’ by designing learning episodes using source tasks to simulate the challenge of solving the target new task with few labeled samples. In contrast, the latter exploits an annotation-free pretext task across all source tasks in order to learn generalizable feature representations. In this work, we propose a novel Instance-level and Episode-level Pretext Task (IEPT) framework that seamlessly integrates SSL into FSL. Specifically, given an FSL episode, we first apply geometric transformations to each instance to generate extended episodes. At the instancelevel, transformation recognition is performed as per standard SSL. Importantly, at the episode-level, two SSL-FSL hybrid learning objectives are devised: (1) The consistency across the predictions of an FSL classifier from different extended episodes is maximized as an episode-level pretext task. (2) The features extracted from each instance across different episodes are integrated to construct a single FSL classifier for meta-learning. Extensive experiments show that our proposed model (i.e., FSL with IEPT) achieves the new state-of-the-art. + +# 1 INTRODUCTION + +Deep convolutional neural networks (CNNs) (Krizhevsky et al., 2012; He et al., 2016b; Huang et al., 2017) have seen tremendous successes in a wide range of application fields, especially in visual recognition. However, the powerful learning ability of CNNs depends on a large amount of manually labeled training data. In practice, for many visual recognition tasks, sufficient manual annotation is either too costly to collect or not feasible (e.g., for rare object classes). This has severely limited the usefulness of CNNs for real-world application scenarios. Attempts have been made recently to mitigate such a limitation from two distinct perspectives, resulting in two popular research lines, both of which aim to transfer knowledge learned from the data of a set of source tasks to a new target one: few-shot learning (FSL) and self-supervised learning (SSL). + +FSL (Fei-Fei et al., 2006; Vinyals et al., 2016; Finn et al., 2017; Snell et al., 2017; Sung et al., 2018) typically takes a ‘learning to learn’ or meta-learning paradigm. That is, it aims to learn an algorithm for learning from few labeled samples, which generalizes well across any tasks. To that end, it adopts an episodic training strategy – the source tasks are arranged into learning episodes, each of which contains $n$ classes and $k$ labeled samples per class to simulate the setting for the target task. Part of the CNN model (e.g., feature extraction subnet, classification layers, or parameter initialization) is then meta-learned for rapid adaptation to new tasks. + +![](images/ac5ad2ee0448bf35ac7fc284a145ae6ff0d67b950b288048f1c82a5d8bf0456d.jpg) +Figure 1: Schematic of our approach to FSL. Given a training episode, we apply 2D rotations by 0, 90, 180, and 270 degrees to each instance to generate four extended episodes. After going through a feature extraction CNN, four losses over three branches are designed: (1) In the top branch, we employ a self-supervised rotation classifier with the instance-level SSL loss $\mathcal { L } _ { i n s t }$ . (2) In the middle branch, an FSL classifier is exploited to predict the FSL classification probabilities for each episode. We maximize the classification consistency among the extended episodes by forcing the four probability distributions to be consistent using $\mathcal { L } _ { e p i s }$ . The average supervised FSL loss $\mathcal { L } _ { a u x }$ is also computed. (3) In the bottom branch, we utilize an integration transformer module to fuse the features extracted from each instance with different rotation transformations; they are then used to compute an integrated FSL loss $\mathcal { L } _ { i n t e g }$ . Among the four losses, $\mathcal { L } _ { i n s t }$ and $\mathcal { L } _ { e p i s }$ are the self-supervised losses, and $\mathcal { L } _ { a u x }$ and $\mathcal { L } _ { i n t e g }$ are the supervised losses. + +In contrast, SSL (Doersch et al., 2015; Noroozi & Favaro, 2016; Iizuka et al., 2016; Doersch & Zisserman, 2017; Noroozi et al., 2018) does not require the source data to be annotated. Instead, it exploits an annotation-free pretext task on the source task data in the hope that a task-generalizable feature representation can be learned from the source tasks for easy adoption or adaptation in a target task. Such a pretext task gets its self-supervised signal at the per-instance level. Examples include rotation and context prediction (Gidaris et al., 2018; Doersch et al., 2015), jigsaw solving (Noroozi & Favaro, 2016), and colorization (Iizuka et al., 2016; Larsson et al., 2016). Since these pretext tasks are class-agnostic, solving them leads to the learning of transferable knowledge. + +Since both FSL and SSL aim to reduce the need of collecting a large amount of labeled training data for a target task by transferring knowledge from a set of source tasks, it is natural to consider combining them in a single framework. Indeed, two recent works (Gidaris et al., 2019; Su et al., 2020) proposed to integrate SSL into FSL by adding an auxiliary SSL pretext task in an FSL model. It showed that the SSL learning objective is complementary to that of FSL and combining them leads to improved FSL performance. However, in (Gidaris et al., 2019; Su et al., 2020), SSL is combined with FSL in a superficial way: it is only taken as a separate auxiliary task for each single training instance and has no effect on the episodic training pipeline of the FSL model. Importantly, by ignoring the class labels of samples, the instance-level SSL learning objective is weak on its own. Since meta-learning across episodes is the essence of most contemporary FSL models, we argue that adding instance-level SSL pretext tasks alone fails to exploit fully the complementarity of the aforementioned FSL and SSL, for which a closer and deeper integration is needed. + +To that end, in this paper we propose a novel Instance-level and Episode-level Pretext Task (IEPT) framework for few-shot recognition. Apart from adding an instance-level pretext SSL task as in (Gidaris et al., 2019; Su et al., 2020), we introduce two episode-level SSL-FSL hybrid learning objectives for seamless SSL-FSL integration. Concretely, as illustrated in Figure 1, our full model has three additional learning objectives (besides the standard FSL one): (1) Different rotation transformations are applied to each original few-shot episode to generate a set of extended episodes, where each image has a rotation label for the instance-level pretext task (i.e., to predict the rotation label). (2) The consistency across the predictions of an FSL classifier from different extended episodes is maximized as an episode-level pretext task. For each training image, the rotation transformation does not change its semantic content and hence its class label; the FSL classifier predictions across different extended episodes thus should be consistent, hence the consistency regularization objective. (3) The correlation of features across instances from these extended episodes is modeled by a transformer-based attention module, optimizing the fusion of the features of each instance/image and its various rotation-transformed versions mainly for task adaptation during meta-testing. Importantly, with these three new learning objectives introduced in IEPT, any meta-learning based FSL model can now benefit more from SSL by fully exploiting their complementarity. + +Our main contributions are: (1) For the first time, we propose both instance-level and episode-level pretext tasks (IEPT) for integrating SSL into FSL. The episode-level pretext task enables episodic training of SSL and hence closer integration of SSL with FSL. (2) In addition to these pretext tasks, FSL further benefits from SSL by integrating features extracted from various rotation-transformed versions of the original training instances. The optimal way of feature integration is learned by a transformer-based attention module, which is mainly designed for task adaptation during meta-testing. (3) Extensive experiments show that the proposed model achieves the new state-of-the-art. + +# 2 RELATED WORK + +Few-Shot Learning. The recent FSL studies are dominated by meta-learning based methods. They can be divided into three groups: (1) Metric-based methods (Vinyals et al., 2016; Snell et al., 2017; Sung et al., 2018; Allen et al., 2019; Xing et al., 2019; Li et al., 2019a;b; Wu et al., 2019; Ye et al., 2020; Afrasiyabi et al., 2020; Liu et al., 2020; Zhang et al., 2020) aim to learn the distance metric between feature embeddings. The focus of these methods is often on meta-learning of a feature-extraction CNN, whilst the classifiers used are of simple form such as a nearest-neighbor classifier. (2) Optimization-based methods (Finn et al., 2017; Ravi & Larochelle, 2017; Rusu et al., 2019; Lee et al., 2019) learn to optimize the model rapidly given a few labeled samples per class in the new task. (3) Model-based methods (Santoro et al., 2016; Munkhdalai & Yu, 2017; Mishra et al., 2018) focus on designing either specific model structures or parameters capable of rapid updating. Apart from these three groups of methods, other FSL methods have attempted feature hallucination (Schwartz et al., 2018; Hariharan & Girshick, 2017; Gao et al., 2018; Wang et al., 2018; Zhang et al., 2019; Tsutsui et al., 2019) which generates additional samples from the given few shots for network finetuning, and parameter predicting (Qiao et al., 2018; Qi et al., 2018; Gidaris & Komodakis, 2019; 2018) which learns to predict part of the parameters of a network given few samples of new classes for quick adaptation. In this work, we adopt the metric-based Prototypical Network (ProtoNet) (Snell et al., 2017) as the basic FSL classifier for the main instantiation of our IEPT framework due to its simplicity and popularity. However, we show that any meta-learning based FSL method can be combined with our IEPT (see results in Figure 2(c)). + +Self-Supervised Learning. In SSL, it is assumed that the source task data is label-free and a pretext task is designed to provide self-supervision signals at the instance-level. Existing SSL approaches differ mainly in the pretext task design. These include predicting the rotation angle (Gidaris et al., 2018) and the context of image patch (Doersch et al., 2015; Nathan Mundhenk et al., 2018), jigsaw solving (Noroozi & Favaro, 2016; Noroozi et al., 2018) (i.e. shuffling and then reordering image patch), and performing images reversion (Iizuka et al., 2016; Pathak et al., 2016; Larsson et al., 2016). SSL has been shown to be beneficial to various down-steam tasks such as semantic object matching (Novotny et al., 2018), object segmentation (Ji et al., 2019) and object detection (Doersch & Zisserman, 2017) by learning transferable feature presentations for these tasks. + +Integrating Self-Supervised Learning into Few-Shot Learning. To the best of our knowledge, only two recent works (Gidaris et al., 2019; Su et al., 2020) have attempted combining SSL with FSL. However, the integration of SSL into FSL is often shallow: the original FSL training pipeline is intact; in the meantime, an additional loss on each image w.r.t. a self-supervised signal like the rotation angle or relative patch location is introduced. With pretext tasks solely at the instance level, combining the two approaches (i.e., SSL and FSL) can only be superficial without fully exploiting the episodic training pipeline unique to FSL. Different from (Gidaris et al., 2019; Su et al., 2020), we introduce an episode-level pretext task to integrate SSL into the episodic training in FSL fully. Specifically, the consistency across the predictions of an FSL classifier from different extended episodes is maximized to reflect the fact that various rotation transformations should not alter the class-label prediction. Moreover, features of each instance and its various rotation-transformed versions are now fused for FSL classification, to integrate SSL with FSL for the supervised classification task. Our experimental results show that thanks to the closer integration of SSL and FSL, our IEPT clearly outperforms (Gidaris et al., 2019; Su et al., 2020) (see Table 1). + +# 3 METHODOLOGY + +# 3.1 PRELIMINARY + +Problem Setting. Given an $n$ -way $k$ -shot FSL task sampled from a test set $\mathcal { D } _ { t }$ , to imitate the test setting, an FSL model is typically trained in an episodic way. That is, $n$ -way $k$ -shot episodes are randomly sampled from a training set $\mathcal { D } _ { s }$ , where the class label space of $\mathcal { D } _ { s }$ has no overlap with that of $\mathcal { D } _ { t }$ . Each episode $E _ { e }$ contains a support set $S _ { e }$ and a query set $\mathcal { Q } _ { e }$ . Concretely, we first randomly sample a set of $n$ classes $\mathcal { C } _ { e }$ from the training set, and then generate $S _ { e }$ and $\mathcal { Q } _ { e }$ by sampling $k$ support samples and $q$ query samples from each class in $\mathcal { C } _ { e }$ , respectively. Formally, we have $\mathcal S _ { e } \stackrel { - } { = } \{ ( x _ { i } , y _ { i } ) \bar { | } y _ { i } \in \mathcal C _ { e } , i \stackrel { - } { = } 1 , . . . , n \stackrel { - } { \times } k \}$ and $\mathcal { Q } _ { e } = \{ ( x _ { i } , y _ { i } ) | y _ { i } \in \mathcal { C } _ { e } , i = 1 , . . . , n \times q \}$ , where $S _ { e } \cap \mathcal { Q } _ { e } = \emptyset$ . For simplicity, we denote $l _ { k } = n \times k$ and $l _ { q } = n \times q$ . In the meta-training stage, the training process has an inner and an outer loop in each episode: in the inner loop, the model is updated using $S _ { e }$ ; its performance is then evaluated on the query set $\mathcal { Q } _ { e }$ in the outer loop to update the model parameters or algorithm that one wants to meta-learn. + +Basic FSL Classifier. We employ ProtoNet (Snell et al., 2017) as the basic FSL model. This model has a feature-extraction CNN and a simple non-parametric classifier. The parameter of the feature extractor is to be meta-learned. Concretely, in the inner loop of an episode, ProtoNet fixes the feature extractor and computes the mean feature embedding for each class as follows: + +$$ +h _ { c } = \frac { 1 } { k } \cdot \sum _ { ( x _ { i } , y _ { i } ) \in S _ { e } } f _ { \phi } ( x _ { i } ) \cdot \mathbb { I } ( y _ { i } = c ) , +$$ + +where class $c \in \mathcal { C } _ { e }$ , $f _ { \phi }$ is a feature extractor with learnable parameters $\phi$ , and $\mathbb { I }$ is the indicator function. By computing the distance between the feature embedding of each query sample and that of the corresponding class, the loss function used to meta-learn $\phi$ in the outer loop is defined as: + +$$ +\mathcal { L } _ { f s l } ( S _ { e } , \mathcal { Q } _ { e } ) = \frac { 1 } { | \mathcal { Q } _ { e } | } \sum _ { ( x _ { i } , y _ { i } ) \in \mathcal { Q } _ { e } } - \log \frac { \exp ( - d ( f _ { \phi } ( x _ { i } ) , h _ { y _ { i } } ) ) } { \sum _ { c \in \mathcal { C } _ { e } } \exp ( - d ( f _ { \phi } ( x _ { i } ) , h _ { c } ) ) } , +$$ + +where $d ( \cdot , \cdot )$ denotes a distance function (e.g., the $l _ { 2 }$ distance). + +# 3.2 PRETEXT TASKS IN IEPT + +The schematic of our IEPT is illustrated in Figure 1. We first define a set of 2D-rotation operators $\mathcal { G } = \{ g _ { r } | r = 0 , . . . , R - 1 \}$ , where $g _ { r }$ means the operator of rotating the image by $r ^ { * 9 0 }$ degrees and $R$ is the total number of rotations ( $R = 4$ in our implementation). Given an original episode $E _ { e } = \{ S _ { e } , \mathcal { Q } _ { e } \}$ as described in Sec. 3.1, we utilize the 2D-rotation operators from $\mathcal { G }$ in turn to transform each image in $E _ { e }$ . This results in a set of $R$ extended episodes (including the original one) $E = \{ \{ \mathcal { S } _ { e } ^ { r } , \bar { \mathcal { Q } } _ { e } ^ { r } \} | r = 0 , . . . , R - 1 \}$ , where $S _ { e } ^ { r } \ = \ \{ ( x _ { i } , y _ { i } , \bar { r } ) | y _ { i } \in C _ { e } , i \ = \ \bar { 1 } , . . . , l _ { k } \}$ and $\mathcal { Q } _ { e } ^ { r } = \{ ( x _ { i } , y _ { i } , r ) | y _ { i } \in C _ { e } , i = 1 , . . . , l _ { q } \} ,$ . Now each episode is denoted as $E _ { e } ^ { r } = \{ ( x _ { i } , y _ { i } , r ) | y _ { i } \in$ $C _ { e } , i = 1 , . . . , l _ { k } , l _ { k } + 1 , . . . , l _ { k } + l _ { q } \}$ , where the first $l _ { k }$ samples are from $S _ { e } ^ { r }$ and the rest from $\mathcal { Q } _ { e } ^ { r }$ Note that $\{ \cal { S } _ { e } ^ { 0 } , \cal { Q } _ { e } ^ { 0 } \}$ is the original episode $\{ \boldsymbol { S } _ { e } , \boldsymbol { \mathcal { Q } } _ { e } \}$ . With the rotation transformations, each sample $( x _ { i } , y _ { i } , r _ { i } )$ in $E$ carries a class label $y _ { i }$ for supervised learning (from the inherent class) and a label $r _ { i }$ from the rotation operator for self-supervised learning. After generating the set of extended episodes $E$ , the feature extractor $f _ { \phi }$ is applied to each image $x _ { i }$ in $E$ . On these episodes, we design two self-supervised pretext tasks, one at the instance-level and the other episode-level. + +Instance-Level Pretext Task. The instance-level task is to recognize different rotation transformations. The idea is that if the model to be meta-learned here (i.e., $f _ { \phi } )$ can be used to distinguish different transformations, it must understand the canonical poses of objects (e.g., animals have legs touching the ground and trees have leaves on top), a vital part of class-agnostic and thus transferable knowledge. With the self-supervised rotation label $r _ { i }$ , we consider the mapping: $f _ { \theta _ { r o t } } : x _ { i } \mapsto r _ { i }$ for each instance $( x _ { i } , y _ { i } , r _ { i } ) \in E$ , where $f _ { \theta _ { r o t } }$ is a rotation classifier with learnable parameters $\theta _ { r o t }$ . Given the input pair $( x _ { i } , r _ { i } )$ , the total instance-level rotation loss is a cross-entropy loss: + +$$ +\mathcal { L } _ { i n s t } = \frac { 1 } { R ( l _ { k } + l _ { q } ) } \sum _ { r = 0 } ^ { R - 1 } \sum _ { ( x _ { i } , y _ { i } , r _ { i } ) \in E _ { e } ^ { r } } - \log \frac { \exp ( [ f _ { \theta _ { r o t } } ( f _ { \phi } ( x _ { i } ) ) ] _ { r _ { i } } ) } { \sum _ { r ^ { \prime } = 0 } ^ { R - 1 } \exp ( [ f _ { \theta _ { r o t } } ( f _ { \phi } ( x _ { i } ) ) ] _ { r ^ { \prime } } ) } , +$$ + +where $[ f _ { \theta _ { r o t } } ( f _ { \phi } ( x _ { i } ) ) ] \in \mathbb { R } ^ { R }$ is the rotation scoring vector and $[ \cdot ] _ { r }$ means taking the $r$ -th element. + +Episode-Level Pretext Task. We design the episode-level task based on a simple principle: although different extended episodes contain images with different rotation transformations, these transformations do not change their class labels. Consequently, the FSL classifier should produce consistent probability distributions for each instance across different extended episodes. Such consistency can be measured using the Kullback–Leibler (KL) divergence. Formally, for each extended episode $\{ S _ { e } ^ { r } , \mathcal { Q } _ { e } ^ { r } \}$ in $E$ , we first define the probability distribution of FSL classification over the query set $\mathcal { Q } _ { e } ^ { r }$ as $\mathcal { P } _ { e } ^ { r } = [ p _ { 1 } ^ { r } ; \cdot \cdot \cdot ; p _ { l _ { q } } ^ { r } ] \in \mathbb { R } ^ { l _ { q } \times n }$ , where $p _ { i } ^ { r } \in \mathbb { R } ^ { n }$ is the probability distribution for $x _ { i }$ in $\mathcal { Q } _ { e } ^ { r }$ with its $c$ -th element $[ p _ { i } ^ { r } ] _ { c }$ $\mathbf { \bar { \psi } } c = 1 , . . . , n )$ being: + +$$ +[ p _ { i } ^ { r } ] _ { c } = \frac { \exp ( - d ( f _ { \phi } ( x _ { i } ) , h _ { c } ^ { r } ) ) } { \sum _ { c ^ { \prime } } \exp ( - d ( f _ { \phi } ( x _ { i } ) , h _ { c ^ { \prime } } ^ { r } ) ) } . +$$ + +The above probability is computed as in Sec. 3.1 and the class embedding $h _ { c } ^ { r }$ is obtained from $S _ { e } ^ { r }$ The mean probability distribution of the $R$ extended episodes is thus given by: + +$$ +\hat { p } _ { i } = \frac { 1 } { R } \cdot \sum _ { r = 0 } ^ { R - 1 } p _ { i } ^ { r } . +$$ + +The total episode-level consistency regularization loss is computed with the KL divergence loss: + +$$ +\mathcal { L } _ { e p i s } = \frac { 1 } { R l _ { q } } \cdot \sum _ { r = 0 } ^ { R - 1 } \sum _ { i = 1 } ^ { l _ { q } } \mathrm { m e a n } ( p _ { i } ^ { r } ( \log p _ { i } ^ { r } - \log \hat { p } _ { i } ) ) . +$$ + +where $\mathrm { m e a n } ( \cdot )$ is an element-wise averaging function. + +# 3.3 INTEGRATED FSL TASK + +The two tasks introduced so far are self-supervised tasks without using the class labels in the query set. Now we describe how in the supervised classification task, the extended episodes can be used. + +Given the set of extended episodes $E$ , we denote the feature set of $E$ as $E _ { e m b }$ , where $E _ { e m b } =$ $\{ f _ { \phi } ( x _ { i } ) | ( x _ { i } , y _ { i } , r ) \in E _ { e } ^ { r } , r = 0 , \cdots , R - 1 , i = 1 , . . . , l _ { k } + l _ { q } \}$ . Note that each extended episode in $E$ corresponds to one specific rotation transformation of the same set of images from the original episode $E _ { e }$ . Therefore, in order to capture the correlation among instances with different transformations and learn how best combine them to form the class mean for meta-learning, an instance attention module is deployed w.r.t. each image in $E _ { e }$ (i.e., all images are assumed to be independent). Specifically, based on $E _ { e m b }$ , we construct the feature tensor $\bar { F } \in \mathbb { R } ^ { ( l _ { k } + l _ { q } ) \times R \times d }$ , where $d$ is the feature dimension. We then adopt a transformer to obtain the integrated representation for FSL classification. The transformer architecture is based on a self-attention mechanism, as in (Vaswani et al., 2017). It receives the triplet input $( F , F , F )$ as $( Q , K , V )$ (Query, Key, and Value, respectively). With $\dot { F } ^ { ( i ) }$ being the $i$ -th row of $F$ (w.r.t. the $i$ -th image in $E _ { e }$ ), the attentive module is defined as: + +$$ +\begin{array} { r l } & { ( \boldsymbol { F } _ { Q } ^ { ( i ) } , \boldsymbol { F } _ { K } ^ { ( i ) } , \boldsymbol { F } _ { V } ^ { ( i ) } ) = ( \boldsymbol { F } ^ { ( i ) } \boldsymbol { W } _ { Q } , \boldsymbol { F } ^ { ( i ) } \boldsymbol { W } _ { K } , \boldsymbol { F } ^ { ( i ) } \boldsymbol { W } _ { V } ) , } \\ & { \boldsymbol { F } _ { a t t } ^ { ( i ) } = \boldsymbol { F } ^ { ( i ) } + \mathrm { s o f t m a x } ( \frac { \boldsymbol { F } _ { Q } ^ { ( i ) } ( \boldsymbol { F } _ { K } ^ { ( i ) } ) ^ { T } } { \sqrt { d _ { K } } } ) \boldsymbol { F } _ { V } ^ { ( i ) } , } \end{array} +$$ + +where $d _ { K } = d$ , and $W _ { Q } , W _ { K } , W _ { V }$ represent the parameters of three fully-connected layers respectively (the parameters of the integration transformer are collected as $\theta _ { i n t }$ ). Note that the key and value are computed from each image and its augmented versions, i.e., they are computed independently without using inter-image correlation. With the attentive feature $F _ { a t t } \in \mathbb { R } ^ { ( l _ { k } + l _ { q } ) \times R \times d }$ , the integrated representation $F _ { i n t e g } = [ F ^ { S } ; F ^ { Q } ] \in \mathbb { R } ^ { ( l _ { k } + l _ { q } ) \times R d }$ ( $F ^ { S }$ and $F ^ { Q }$ are respectively for the support set and query set) is given by: + +$$ +F _ { i n t e g } = \mathrm { f l a t t e n } ( F _ { a t t } ) , +$$ + +where flatten $( \cdot )$ denotes flattening $F _ { a t t }$ along the last two dimensions, i.e., concatenating the attentive features from different extended episodes for the corresponding images. The integrated representation is then inputted to the FSL classifier to define the FSL classification loss: + +$$ +\mathcal { L } _ { i n t e g } = \frac { 1 } { l _ { q } } \cdot \sum _ { i = 1 } ^ { l _ { q } } - \log \frac { \exp ( - d ( F _ { i } ^ { Q } , h _ { y _ { i } } ^ { f } ) ) } { \sum _ { c \in \mathcal { C } _ { e } } \exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) } +$$ + +where the class embedding the integrated FSL task act $\begin{array} { r } { h _ { c } ^ { f } = \frac { 1 } { k } \cdot \sum _ { i = 1 } ^ { l _ { k } } F _ { i } ^ { S } \cdot \mathbb { I } ( y _ { i } = c ) } \end{array}$ is computed on the support set. Note thatrediction averaging. + +# 3.4 TOTAL LOSS + +The total training loss for our full model consists of the self-supervised losses from the pretext tasks and the supervised losses from the FSL tasks. In this work, in addition to $\mathcal { L } _ { i n t e g }$ in Eq. (10), another supervised FSL loss $\mathcal { L } _ { a u x }$ is also used (see Figure 1). $\mathcal { L } _ { a u x }$ is the average FSL classification loss over the extended episodes. Formally, it can be written as: + +$$ +\mathcal { L } _ { a u x } = \frac { 1 } { R } \cdot \sum _ { r = 0 } ^ { R - 1 } \mathcal { L } _ { f s l } ( S _ { e } ^ { r } , \mathcal { Q } _ { e } ^ { r } ) +$$ + +Therefore, the total loss $\mathcal { L } _ { t o t a l }$ for training our full model is given as follows: + +$$ +\begin{array} { r } { \mathcal { L } _ { t o t a l } = \underbrace { \overbrace { w _ { 1 } * \mathcal { L } _ { i n s t } } ^ { i n s t a n c e { - } l e v e l } + \overbrace { w _ { 2 } * \mathcal { L } _ { e p i s } } ^ { e p i s o d e { - } l e v e l } } _ { \mathrm { s e l f - s u p e r v i s e d l o s s } } + \underbrace { w _ { 3 } * \mathcal { L } _ { a u x } + \mathcal { L } _ { i n t e g } } _ { \mathrm { s u p e r v i s e d l o s s } } , } \end{array} +$$ + +where $w _ { 1 } , w _ { 2 } , w _ { 3 }$ are the loss weight hyperparameters. + +# 3.5 INFERENCE + +During the test stage, we only exploit the integrated representation $F _ { i n t e g }$ for the final FSL prediction. The predicted class label for $x _ { i } \in \mathcal { Q } _ { e }$ can be computed with Eq. (10) as: + +$$ +y _ { i } ^ { p r e d } = \underset { y \in \mathcal { C } _ { e } } { \operatorname { a r g m a x } } \frac { \exp ( - d ( F _ { i } ^ { Q } , h _ { y } ^ { f } ) ) } { \sum _ { c \in \mathcal { C } _ { e } } \exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) } . +$$ + +# 3.6 FULL IEPT ALGORITHM + +For easy reproduction, we present the full algorithm for FSL with IEPT in Algorithm 1. Once learned, with the learned $\psi$ , we can perform the inference over the test episodes with Eq. (13). + +# Algorithm 1 FSL with IEPT + +Input: The training set $\mathcal { D } _ { s }$ , the rotation operator set $\mathcal { G }$ +The loss weight hyperparameters $w _ { 1 } , w _ { 2 } , w _ { 3 }$ +Output: The learned $\psi$ +1: Randomly initialize all learnable parameters $\psi = \{ \phi , \theta _ { r o t } , \theta _ { i n t } \}$ +2: for iteration $= 1$ , ..., MaxIteration do +3: Randomly sample episode $E _ { e }$ from $\mathcal { D } _ { s }$ +4: Generate the set of extended episodes $E$ from $E _ { e }$ using $\mathcal { G }$ +5: Compute the SSL loss $\mathcal { L } _ { i n s t }$ for the instance-level pretext task with Eq. (3) +6: Compute the SSL loss $\mathcal { L } _ { e p i s }$ for the episode-level pretext task with Eq. (6) +7: Compute the supervised FSL loss $\mathcal { L } _ { a u x }$ over the extended episodes with Eq. (11) +8: Compute the supervised FSL loss $\mathcal { L } _ { i n t e g }$ for the integrated episode with Eq. (10) +9: $\mathcal { L } _ { t o t a l } = w _ { 1 } * \mathcal { L } _ { i n s t } + w _ { 2 } * \mathcal { L } _ { e p i s } + w _ { 3 } * \mathcal { L } _ { a u x } + \mathcal { L } _ { i n t }$ teg +10: Update $\psi$ based on $\nabla _ { \psi }$ Ltotal +11: end for +12: return $\psi$ . + +# 4 EXPERIMENTS + +# 4.1 EXPERIMENTAL SETUP + +Datasets. Two widely-used FSL datasets are selected: miniImageNet (Vinyals et al., 2016) and tieredImageNet (Ren et al., 2018). The first dataset consists of a total number of 100 classes (600 images per class) and the train/validation/test split is set to 64/16/20 classes as in (Ravi & Larochelle, 2017). The second dataset is a larger dataset including 608 classes totally (nearly 1,200 images per class), which is split into 351/97/160 classes for train/validation/test. Both datasets are subsets sampled from ImageNet (Russakovsky et al., 2015). + +Table 1: Comparative results for 5-way 1/5-shot FSL. The mean classification accuracies (top-1, $\%$ with the $9 5 \%$ confidence intervals are reported. $\dagger$ indicates the result is reproduced by ourselves. + +
MethodminiImageNettieredImageNet
Backbone1-shot5-shot1-shot5-shot
MatchingNet (Vinyals etal.,2016)Conv4-6443.56± 0.8455.31± 0.73
ProtoNet† (Snell et al., 2017)Conv4-6452.61 ± 0.5271.33 ± 0.4153.33 ± 0.5072.10 ± 0.41
MAML (Finn et al., 2017)Conv4-6448.70 ± 1.8463.10 ±0.9251.67 ± 1.8170.30 ±0.08
Relation Net (Sung et al.,2018)Conv4-6450.40 ± 0.8065.30 ±0.7054.48 ± 0.9371.32 ± 0.78
IMP† (Allen et al., 2019)Conv4-6452.91 ± 0.4971.57 ± 0.4253.63 ± 0.5171.89 ± 0.44
DN4 (Li et al., 2019b)Conv4-6451.24 ± 0.7471.02 ± 0.641
DN PARN (Wu et al., 2019)Conv4-6455.22 ± 0.8471.55 ± 0.66
PN+rot (Gidaris et al., 2019)Conv4-6453.63 ± 0.4371.70 ± 0.36
CC+rot (Gidaris et al., 2019)Conv4-6454.83 ± 0.4371.86 ± 0.33
DSN-MR (Simon et al., 2020)Conv4-6455.88 ± 0.9070.50±0.68
Centroid (Afrasiyabi et al., 2020)Conv4-6453.14 ± 1.0671.45 ± 0.72
Neg-Cosine (Liu et al., 2020)Conv4-6452.84 ±0.7670.41 ± 0.66一
IEPT (ours)Conv4-6456.26± 0.4573.91±0.34 58.25±0.48 75.63±0.46
ProtoNet† (Snell et al.,2017)Conv4-51253.25± 0.4473.15 ± 0.3557.88 ±0.5076.82±0.40
MAML (Finn et al., 2017)Conv4-51249.33 ±0.6065.17 ± 0.4952.84± 0.5670.91 ± 0.46
Relation Net (Sung et al.,2018)Conv4-51250.86 ± 0.5767.32 ± 0.4454.69 ±0.5972.71 ±0.43
PN+rot (Gidaris et al.,2019)Conv4-51256.02 ±0.4674.00 ± 0.35一
CC+rot (Gidaris et al., 2019)Conv4-51256.27 ± 0.4374.30 ± 0.33一
IEPT (ours)Conv4-51258.43±0.4675.07±0.3360.91±0.5979.61±0.45
ProtoNet† (Snell et al.,2017)ResNet-1262.39 ±0.5180.53±0.4268.23±0.5084.03 ± 0.41
TADAM (Oreshkin et al., 2018)ResNet-1258.50 ± 0.3076.70 ±0.38一
MetaOptNet (Lee et al.,2019)ResNet-1262.64± 0.6178.63 ± 0.4665.99 ± 0.7281.56 ± 0.63
MTL (Sun et al., 2019)ResNet-1261.20 ± 1.8075.50 ± 0.8065.62 ± 1.8080.61 ±0.90
CAN (Hou et al., 2019)ResNet-1263.85 ±0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
AM3 (Xing et al., 2019)ResNet-1265.21 ± 0.4975.20 ±0.3667.23 ± 0.3478.95 ± 0.22
Shot-Free (Ravichandran et al.,2019)ResNet-1259.04 ± 0.4377.64 ± 0.3966.87 ± 0.4382.64 ±0.43
Neg-Cosine (Liu et al., 2020)ResNet-1263.85 ± 0.8181.57 ± 0.56
Distill (Tian et al., 2020)ResNet-1264.82 ±0.6082.14 ± 0.4371.52 ± 0.6986.03 ±0.49
DSN-MR (Simon et al.,2020)ResNet-1264.60 ±0.7279.51± 0.5067.39 ± 0.8282.85 ±0.56
DeepEMD (Zhang et al.,2020)ResNet-1265.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
FEAT (Ye et al.,2020)ResNet-1266.78±0.2082.05 ± 0.1470.80 ±0.2384.79 ± 0.16
ProtoNet+Rotation (Su et al.,2020)ResNet-1876.00±0.6078.90±0.70
IEPT (ours)ResNet-1267.05±0.44882.90±0.30)72.24±0.50 86.73±0.34
+ +Feature Extractors. For fair comparison with published results, our IEPT adopts three widely-used feature extractors: Conv4-64 (Vinyals et al., 2016), Conv4-512, and ResNet-12 (He et al., 2016a). Particularly, Conv4-512 is almost the same as Conv4-64 except having a different channel size of the last convolution layer. To speed up the training process, as in many previous works (Ye et al., 2020; Zhang et al., 2020; Simon et al., 2020), we pretrain all the feature extractors on the training split of each dataset for our IEPT. Following (He et al., 2016a), we use the temperature scaling skill during the training phase. On both datasets, the input image size is $8 4 \times 8 4$ . The output feature dimensions of Conv4-64, Conv4-512, and ResNet-12 are 64, 512, and 640, respectively. + +Evaluation Metrics. We take the 5-way 5-shot (or 1-shot) FSL evaluation setting, as in previous works. We randomly sample 2,000 episodes from the test split and report the mean classification accuracy (top-1, $\%$ ) as well as the $9 5 \%$ confidence interval. Since the integration transformer copes with each sample independently, we take a strict non-transductive setting during evaluation. + +Implementation Details. PyTorch is used for our implementation. We utilize the Adam optimizer (Kingma & Ba, 2015) for Conv4-64 & Conv4-512 and the SGD optimizer for ResNet-12 to train our IEPT model. The hyperparameters of our IEPT model are selected according to the performance on the validation split.We will release the code soon. + +# 4.2 MAIN RESULTS + +Comparison to State-of-the-Arts. We compare our IEPT with two groups of baselines: (1) Recent SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020); (2) Representative/latest FSL methods (w/o SSL) (Snell et al., 2017; Finn et al., 2017; Lee et al., 2019; Ravichandran et al., 2019; Simon et al., 2020; Zhang et al., 2020; Ye et al., 2020; Liu et al., 2020). The comparative results for 5-way + +Table 2: Ablation study results for our full IEPT model over miniImageNet and tieredImageNet. Our full model includes two self-supervised losses (i.e. $\mathcal { L } _ { e p i s }$ and $\mathcal { L } _ { i n s t . }$ ) and two supervised losses (i.e. $\mathcal { L } _ { a u x }$ and $L _ { i n t e g . }$ ). Conv4-64 is used as the feature extractor. + +
miniImageNettieredImageNet
LintegLinstLepisLaux1-shot5-shot1-shot5-shot
√55.04 ± 0.5272.01 ± 0.4156.98 ± 0.4774.15 ± 0.51
√√55.49 ± 0.5672.54± 0.4657.41 ± 0.5174.65 ± 0.50
√55.88± 0.4372.97 ± 0.4057.76± 0.4575.06 ± 0.40
√55.97 ± 0.5773.28 ± 0.3957.83 ± 0.5575.22 ± 0.48
√?>>>√56.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
+ +![](images/af3c8b4f8135fe31235a01a5a12b5a61f1267d2c4e0e0affdba8e3fca164fe3e.jpg) +Figure 2: (a) Comparison among different combination methods over episodes for FSL with selfsupervision. $\mathbf { ( b ) }$ Illustration of the effect of different choices of $R$ on the performance of our model $R$ denotes the number of extended episodes used for SSL). (c) Comparative results obtained by our IEPT using different basic FSL classifiers (i.e. ProtoNet, FEAT, and IMP). It can be seen clearly that integrated episode-based fusion leads to more separation between classes. All figures present 5-way 1-shot/5-shot results on miniImageNet, using Conv4-64 as the feature extractor. + +1/5-shot FSL are shown in Table 1. We have the following observations: (1) When compared with the representative/latest FSL methods (w/o SSL), our IEPT achieves the best performance on all datasets and under all settings, validating the effectiveness of SSL with IEPT for FSL. (2) Our IEPT also clearly outperforms the two SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020) which only use instance-level pretext tasks, demonstrating the importance of closer/episode-level integration of SSL into FSL. (3) The improvements achieved by our IEPT over ProtoNet range from $2 \%$ to $5 \%$ . Since our IEPT takes ProtoNet as the baseline, the obtained margins provide direct evidence that SSL brings significant benefits to FSL. Note that our IEPT is also shown to be effective under both the fine-grained FSL and cross-domain FSL settings in Sec. 4.3 (see Table 3). + +Ablation Study. Our full IEPT model is trained with four losses (see Eq. (12)), including two self-supervised losses and two supervised losses: the episode-level SSL loss $\mathcal { L } _ { e p i s }$ , the instance-level SSL loss $\mathcal { L } _ { i n s t }$ , the auxiliary FSL loss $\mathcal { L } _ { a u x }$ and the integrated FSL loss $\bar { L } _ { i n t e g }$ . To demonstrate the contribution of each loss, we present the ablation study results for our full IEPT model in Table 2, where Conv4-64 is used as the backbone. We start with $L _ { i n t e g }$ and then add the additional three losses successively. It can be observed that the performance of our model continuously increases when more losses are used, indicating that each loss contributes to the final performance. + +# 4.3 FURTHER EVALUATIONS + +Different Combination Methods over Episodes. We have introduced a transformer-based attention module to fuse the features of each instance from all extended episodes (and an integrated episode can be obtained) for the supervised classification task (see Sec. 3.3). In this experiment, we compare it with two alternative ways of across-episode integration: (1) Averaging extended episodes: the extended episodes are directly fused for FSL classification; (2) Averaging all episodes: the extended episodes as well as the integrated episode are fused for FSL classification. We present the comparative results on miniImageNet in Figure 2(a). For comprehensive comparison, the results of FSL with each single extended episode are also reported. We can observe that: (1) The performance of ‘Episode $0 ^ { \circ }$ ’ is the highest among the four baselines (i.e., FSL with single extended episode), perhaps because the feature extractor is pretrained on the original images without rotation transformations. (2) FSL by averaging extended episodes (i.e., ‘Averaging extended episodes’) indeed improves each of the four baselines. (3) FSL with integrated episode (i.e., ‘Integrated episode’) is superior to FSL by simply averaging extended episodes. (4) Comparing ‘Integrated episode’ with ‘Averaging all episodes’, the performance of FSL with integrated episode is more stable across different settings, furthering validating the usefulness of our across-episode integration. Overall, the episode-integration module is indeed effective in FSL with self-supervision. This is also supported by the visualization results presented in Appendices A.3 & A.4. + +Table 3: Comparative results for the fine-grained FSL on CUB (Wah et al., 2011) and the cross-domain FSL on miniImageNet CUB. + +
MethodBackboneCUBminiImageNet→CUB
1-shot5-shot1-shot5-shot
MatchingNet (Vinyals et al.,2016)Conv4-6461.16±0.8972.86±0.7042.62 ± 0.5556.53± 0.44
ProtoNet (Snell et al.,2017)Conv4-6463.72 ±0.2281.50 ± 0.1550.51 ± 0.5669.28±0.40
MAML (Finn et al.,2017)Conv4-6455.92 ± 0.9572.09 ±0.7643.59 ± 0.5454.18 ± 0.41
Relation Net (Sung et al., 2018)Conv4-6462.45 ± 0.9876.11 ± 0.6949.84 ±0.5468.98 ±0.42
FEAT(Ye et al.,2020)Conv4-6468.87 ±0.2282.90±0.1551.52 ± 0.5470.16 ± 0.40
IEPT(ours)Conv4-6469.97±0.4984.33± 0.3352.68±0.5672.98±0.40
+ +Different Number of Extended Episodes. In all the above experiments, the number of the extended episodes $R$ is set to 4 (rotation by $0 ^ { \circ }$ , $9 0 °$ , $1 8 0 ^ { \circ }$ , $2 7 0 ^ { \circ }$ ). Figure 2(b) shows the impact of the value of $R$ . Note that when $R = 1$ , our IEPT model is equivalent to ProtoNet which is without self-supervision. It can be seen that the performance of our model consistently grows when $R$ increases from 1 to 4. Additionally, the study on exploiting other pretext tasks for our IEPT is presented in Appendix A.1. + +Different Basic FSL Classifiers. As mentioned in Sec. 3.1, we adopt ProtoNet as the basic FSL classifier due to its scalability and simplicity. To further show the effectiveness of our IEPT when other basic FSL classifiers are used, we provide the results obtained by our IEPT using ProtoNet, FEAT, and IMP for FSL in Figure 2(c). It can be clearly observed that our IEPT leads to an improvement of about $1 \%$ over each basic FSL method (ProtoNet, FEAT, or IMP), indicating that our IEPT can be applied to improve a variety of popular FSL methods. + +Comparative Results for Fine-Grained FSL and Cross-Domain FSL. To evaluate our IEPT algorithm under the fine-grained FSL and cross-domain FSL settings, we conduct experiments on CUB (Wah et al., 2011) and miniImageNet $ \mathrm { C U B }$ , respectively. For fine-grained FSL on CUB, following (Ye et al., 2020), we randomly split the dataset into 100 training classes, 50 validation classes, and 50 test classes. For cross-domain FSL on miniImageNet $ \mathrm { C U B }$ , the 100 training classes are from miniImageNet; the 50 validation classes and 50 test classes (using the aforementioned split for fine-grained FSL) are from CUB. Under both settings, we use Conv4-64 as the feature extractor. The 5-way 1/5-shot FSL results are shown in Table 3. Our IEPT clearly achieves the best results, yielding $1 - 3 \%$ improvements over the second-best FEAT. This shows the effectiveness of our IEPT under both fine-grained and cross-domain settings. + +# 5 CONCLUSION + +We have proposed a novel Instance-level and Episode-level Pretext Task (IEPT) framework for integrating SSL into FSL. For the first time, we have introduced an episode-level pretext task for FSL with self-supervision, in addition to the conventional instance-level pretext task. Moreover, we have also developed an episode extension-integration framework by introducing an integration transformer module to fully exploit the extended episodes for FSL. Extensive experiments on two benchmarks demonstrate that the proposed model (i.e., FSL with IEPT) achieves the new state-of-theart. Our ongoing research directions include: exploring other episode-level pretext tasks for FSL with self-supervision, and applying FSL with self-supervision to other vision problems. + +# ACKNOWLEDGMENTS + +This work was supported in part by National Natural Science Foundation of China (61976220 and 61832017), Beijing Outstanding Young Scientist Program (BJJWZYJH012019100020098), Open Project Program Foundation of Key Laboratory of Opto-Electronics Information Processing, Chinese Academy of Sciences (OEIP-O-202006), and Alibaba Innovative Research (AIR) Program. + +# REFERENCES + +Arman Afrasiyabi, Jean-Franc¸ois Lalonde, and Christian Gagne. Associative alignment for few-shot image ´ classification. In European Conference on Computer Vision (ECCV), 2020. + +Kelsey R. Allen, Evan Shelhamer, Hanul Shin, and Joshua B. Tenenbaum. Infinite mixture prototypes for few-shot learning. In International Conference on Machine Learning (ICML), pp. 232–241, 2019. + +Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. International Conference on Learning Representations (ICLR), 2020. + +Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In International Conference on Computer Vision (ICCV), pp. 2051–2060, 2017. + +Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In International Conference on Computer Vision (ICCV), pp. 1422–1430, 2015. + +Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pp. 594–611, 2006. + +Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning (ICML), pp. 1126–1135, 2017. + +Hang Gao, Zheng Shou, Alireza Zareian, Hanwang Zhang, and Shih-Fu Chang. Low-shot learning via covariancepreserving adversarial augmentation networks. In Advances in Neural Information Processing Systems, pp. 975–985, 2018. + +Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4367–4375, 2018. + +Spyros Gidaris and Nikos Komodakis. Generating classification weights with gnn denoising autoencoders for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21–30, 2019. + +Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. + +Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Perez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In International Conference on Computer Vision (ICCV), pp. 8059–8068, 2019. + +Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In International Conference on Computer Vision (ICCV), pp. 3037–3046, 2017. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016a. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016b. + +Ruibing Hou, Hong Chang, MA Bingpeng, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot classification. In Advances in Neural Information Processing Systems, pp. 4003–4014, 2019. + +Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4700–4708, 2017. + +Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Let there be color! joint end-to-end learning of global and local image priors for automatic image colorization with simultaneous classification. ACM Transactions on Graphics (ToG), pp. 1–11, 2016. + +Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image clas- ˜ sification and segmentation. In International Conference on Computer Vision (ICCV), pp. 9865–9874, 2019. + +Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pp. 1097–1105, 2012. + +Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In European Conference on Computer Vision (ECCV), pp. 577–593, 2016. + +Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10657– 10665, 2019. + +Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features for few-shot learning by category traversal. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–10, 2019a. + +Wenbin Li, Lei Wang, Jinglin Xu, Jing Huo, Yang Gao, and Jiebo Luo. Revisiting local descriptor based imageto-class measure for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7260–7268, 2019b. + +Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification. In European Conference on Computer Vision (ECCV), 2020. + +Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. + +Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations (ICLR), 2018. + +Tsendsuren Munkhdalai and Hong Yu. Meta networks. In International Conference on Machine Learning (ICML), pp. 2554–2563, 2017. + +T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9339–9348, 2018. + +Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European Conference on Computer Vision (ECCV), pp. 69–84, 2016. + +Mehdi Noroozi, Ananth Vinjimoor, Paolo Favaro, and Hamed Pirsiavash. Boosting self-supervised learning via knowledge transfer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9359–9367, 2018. + +David Novotny, Samuel Albanie, Diane Larlus, and Andrea Vedaldi. Self-supervised learning of geometrically stable features through probabilistic introspection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3637–3645, 2018. + +Boris Oreshkin, Pau Rodr´ıguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for ´ improved few-shot learning. In Advances in Neural Information Processing Systems, pp. 721–731, 2018. + +Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2536–2544, 2016. + +Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5822–5830, 2018. + +Siyuan Qiao, Chenxi Liu, Wei Shen, and Alan L Yuille. Few-shot image recognition by predicting parameters from activations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7229–7238, 2018. + +Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. + +Avinash Ravichandran, Rahul Bhotika, and Stefano Soatto. Few-shot learning with embedded class models and shot-free meta training. In International Conference on Computer Vision (ICCV), pp. 331–339, 2019. + +Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), 2018. + +Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. In International Journal of Computer Vision (IJCV), pp. 211–252, 2015. + +Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations (ICLR), 2019. + +Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In International Conference on Machine Learning (ICML), pp. 1842–1850, 2016. + +Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris, Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object recognition. In Advances in Neural Information Processing Systems, pp. 2850–2860, 2018. + +Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. Adaptive subspaces for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4136–4145, 2020. + +Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, pp. 4077–4087, 2017. + +Jong-Chyi Su, Subhransu Maji, and Bharath Hariharan. When does self-supervision improve few-shot learning? In European Conference on Computer Vision (ECCV), 2020. + +Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019. + +Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1199–1208, 2018. + +Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision (ECCV), 2020. + +Satoshi Tsutsui, Yanwei Fu, and David Crandall. Meta-reinforced synthetic data for one-shot fine-grained visual recognition. In Advances in Neural Information Processing Systems, pp. 3057–3066, 2019. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pp. 5998–6008, 2017. + +Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pp. 3630–3638, 2016. + +C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. + +Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7278–7286, 2018. + +Ziyang Wu, Yuwei Li, Lihua Guo, and Kui Jia. Parn: Position-aware relation networks for few-shot learning. In International Conference on Computer Vision (ICCV), pp. 6659–6667, 2019. + +Chen Xing, Negar Rostamzadeh, Boris Oreshkin, and Pedro O O. Pinheiro. Adaptive cross-modal few-shot learning. In Advances in Neural Information Processing Systems, pp. 4847–4857, 2019. + +Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-toset functions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8808–8817, 2020. + +Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured classifiers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12203–12213, 2020. + +Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of samples. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2770–2779, 2019. + +# A APPENDIX + +# A.1 COMPARISON AMONG DIFFERENT SSL STRATEGIES + +To generate the extended episodes in IEPT, we apply four rotation transformations (i.e. rotation by $0 ^ { \circ }$ , $9 0 °$ , $1 8 0 ^ { \circ }$ , $2 7 0 ^ { \circ }$ ) to each image. It makes sense to explore whether other self-supervised strategies are also effective for our IEPT. To this end, we exploit shuffling image patches (see Figure 3) for self-supervised learning (SSL). Specifically, we divide each image into $2 ^ { \ast } 2$ patches and reorganize the patch orders to obtain a shuffling label. To compare with the rotation strategy fairly, we choose only four shuffling orders: (1, 2, 3, 4), (2, 3, 4, 1), (3, 4, 1, 2) and (4, 1, 2, 3). Note that the (1, 2, 3, 4) shuffling order equals to the original image. Similar to the rotation strategy, a fully-connected layer is utilized to recognize the shuffling order. The comparative results are shown in Table 4. We can see that both IEPT with shuffling and IEPT with rotation achieve better performance than the original ProtoNet. Particularly, IEPT with shuffling yields $1 - 3 \%$ and $3- 4 \%$ improvements under 5-shot and 1-shot, respectively. This clearly shows the effectiveness of our IEPT for FSL even when different SSL strategies are used to define the pretext tasks. + +Table 4: FSL results obtained by our IEPT using two SSL strategies (i.e. rotation and shuffling image patches). Conv4-64 is used as the feature extractor. + +
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
ProtoNetConv4-6452.61 ±0.5271.33 ± 0.4153.33 ± 0.5072.10± 0.41
IEPT (rotation)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ±0.4875.63 ± 0.46
IEPT (shuffling)Conv4-6455.57 ± 0.6072.84 ± 0.5457.81 ± 0.4874.92 ± 0.50
+ +![](images/2d68786b89daf9ea45fcd13f2c425d6edbd6a37859f608778de12061ff067209.jpg) +Figure 3: Illustration of the self-supervised strategy by shuffling image patches. + +A.2 COMPARISON AMONG DIFFERENT INTEGRATION APPROACHES + +We employ the integration transformer to find the intrinsic correlation of various rotation-transformed instances. The transformer architecture is based on a self-attention mechanism. Concretely, it receives the feature sets of extended episodes as input $Q$ , $K$ and $V$ . Further, it matches each query in $Q$ with a list of keys in $K$ and returns the weighted sum of corresponding values. To show the importance of the transformer module, we compare it with two other integration approaches (i.e. concatenating and averaging) to integrate the features of extended episodes. The comparative results in Table 5 demonstrate that the integration transformer consistently performs better than the simply concatenating/averaging approaches. This suggests that the attention-based integration transformer is a better choice for designing the integration module. + +Table 5: Comparative results obtained by three different approaches to integrating the features from the extended episodes, with Conv4-64 being the feature extractor. + +
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
concatenatingConv4-6451.52±0.6073.36± 0.4950.78±0.6872.79± 0.57
averagingConv4-6451.58 ± 0.6270.97 ± 0.5453.91 ± 0.6972.52 ± 0.59
transformer (ours)Conv4-6456.26± 0.4573.91 ± 0.3458.25 ± 0.4875.63± 0.46
+ +![](images/afa66040ad4c1918609b3c5a1e793531019fcc73f4e6f207f4681c14646f499f.jpg) +Figure 4: Feature visualizations of a number of test episodes using the UMAP algorithm (McInnes et al., 2018). Each row indicates a group of test extended episodes (the first four columns, rotation by $0 ^ { \circ }$ , $9 0 °$ , $1 8 0 ^ { \circ }$ , $2 7 0 ^ { \circ }$ ) and their integrated episode (the last column). The 5-way 5-shot FSL (with Conv4-64) is adopted on miniImageNet. + +# A.3 FEATURE VISUALIZATIONS OF TEST EPISODES + +We provide the feature visualizations of test episodes in Figure 4. It can be seen that an integrated episode (the last one in each row) clearly has a better cluster data structure than the corresponding four extended episodes (the first four ones in each row). This indicates that our transformer-based across-episode integration is indeed effective for few-shot classification with self-supervision. + +# A.4 ATTENTION VISUALIZATION OF TEST EPISODES + +We present attention map visualization of two test episodes (left and right) in Figure 5. Each average attention map is computed by averaging the attention map of all instances of a certain class. We can observe that: (1) The average attention maps from different classes vary significantly, showing that the diverse semantics of different classes can be reflected by our attention-based integration transformer. (2) When the classes of two episodes overlap (e.g., ‘trifle’ and ‘dalmatian’), the average attention maps of an overlapped class from two episodes are similar, illustrating that our attention-based integration transformer can well capture the semantics of classes across episodes. + +![](images/d5c83307373ad302b67daff2626249571668bdb71167a2ce318b0ac9aa79845d.jpg) +Figure 5: Feature and attention visualization of two test episodes (left and right). Both figures present 5-way 5-shot results on miniImageNet, using Conv4-64 as the feature extractor. + +A.5 COMPARISON WITH SIMPLE BASELINE FOR SSL+FSL + +We compare our IEPT with a simple baseline that trains the model with $\mathcal { L } _ { a u x } + \mathcal { L } _ { i n s t }$ and then makes inference by just averaging the outputs of different extended episodes. The results are shown in Table 6. It can be observed that the performance of our IEPT is much more effective than that of simple integration, due to the extra use of $\mathcal { L } _ { i n t e g } + \mathcal { L } _ { e p i s }$ for FSL. + +Table 6: Comparison with the simple baseline that trains the model with $\mathcal { L } _ { a u x } + \mathcal { L } _ { i n s t }$ and then makes inference by just averaging the outputs of different extended episodes. + +
miniImageNettieredImageNet
MethodBackbone1-shot5-shot1-shot5-shot
Laux+LinstConv4-6453.25 ± 0.4671.50 ± 0.4255.06 ± 0.4472.87 ± 0.42
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
+ +# A.6 DIFFERENT ALTERNATIVES OF SELF-SUPERVISED LOSSES + +In Table 7, we provide further ablation study regarding different alternatives of $\mathcal { L } _ { e p i s }$ and $\mathcal { L } _ { i n s t }$ . For the episode-level self-supervised loss $\mathcal { L } _ { e p i s }$ , we compare our implementation (using the KL loss between each distribution and the mean distribution) with that using a pairwise KL loss (i.e., the KL loss between each pair of distributions). For the instance-level self-supervised loss $\mathcal { L } _ { i n s t }$ , we compare our implementation (using the rotation prediction loss) with the recent self-supervised learning technique (Chen et al., 2020). We observe that our implementation achieves slight performance improvements over those using the pairwise KL loss or the contrastive learning loss. + +Table 7: Ablation study results regarding different alternatives of $\mathcal { L } _ { e p i s }$ and $\mathcal { L } _ { i n s t }$ + +
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
IEPT (Linst-SimCLR)Conv4-6456.04 ± 0.4473.67± 0.4158.24± 0.4375.59± 0.41
IEPT (Lepis-Pairwise)Conv4-6455.95 ± 0.4773.72 ±0.4057.91 ± 0.4575.28±0.40
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
+ +# A.7 APPLICATION OF IEPT TO OPTIMIZATION-BASED METHOD MAML + +In Table 8, we show the results obtained by applying our IEPT to the optimization-based model MAML (Finn et al., 2017). We use Conv4-64 as the feature extractor. We can see that our IEPT brings $0 . 7 \% - 2 . 1 \%$ improvements to MAML. This further shows the flexibility (as well as effectiveness) of our IEPT for FSL. + +Table 8: Comparative results by applying IEPT to the optimization-based method MAML. + +
BackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
Method MAMLConv4-6448.70±1.8463.10±0.9251.67 ± 1.8170.30±0.80
MAML+IEPTConv4-6449.68 ± 0.5065.22 ± 0.4852.85 ± 0.5271.04 ± 0.49
+ +![](images/19a2a9e135d017944a544b6d105ac0bab71b8611d3f98f7456e4da154a02510e.jpg) +Figure 6: Visualization of our hyper-parameter analysis under 5-way 1-shot (left) and 5-shot (right) on miniImageNet. Conv4-64 is used as the feature extractor. + +# A.8 HYPER-PARAMETER SENSITIVITY TEST + +We select the hyper-parameters $w _ { 1 }$ , $w _ { 2 }$ and $w _ { 3 }$ from the candidate set $\{ 0 . 1 , 0 . 5 , 1 . 0 , 5 . 0 , 1 0 . 0 \}$ and show the hyper-parameter analysis results in Figure 6. We find that the performance of our IEPT is relatively stable. Concretely, the performance of our IEPT is not sensitive to $w _ { 1 }$ and $w _ { 2 }$ with proper values, but too large $w _ { 1 }$ (i.e. $w _ { 1 } = 1 0 . 0 $ ) tends to cause obvious degradation, perhaps because the FSL task is biased by the rotation prediction loss. \ No newline at end of file diff --git a/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_content_list.json b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..09b442988af2c4dba445b69ddf14571333771687 --- /dev/null +++ b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_content_list.json @@ -0,0 +1,2042 @@ +[ + { + "type": "text", + "text": "IEPT: INSTANCE-LEVEL AND EPISODE-LEVEL PRETEXT TASKS FOR FEW-SHOT LEARNING ", + "text_level": 1, + "bbox": [ + 174, + 98, + 826, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Manli Zhang, Jianhong Zhang & Zhiwu Lu ∗ \nGaoling School of Artificial Intelligence, Renmin University of China, Beijing, China Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China {manlizhang,jianhong,luzhiwu}@ruc.edu.cn ", + "bbox": [ + 184, + 170, + 764, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Xiang University of Surrey, Guildford, Surrey, UK t.xiang@surrey.ac.uk ", + "bbox": [ + 183, + 247, + 473, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingyu Ding \nThe University of Hong Kong, Hong Kong \nmingyuding@hku.hk ", + "bbox": [ + 532, + 247, + 813, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Songfang Huang ", + "text_level": 1, + "bbox": [ + 184, + 311, + 302, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alibaba DAMO Academy, Hangzhou, China songfang.hsf@alibaba-inc.com ", + "bbox": [ + 184, + 325, + 478, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 388, + 544, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The need of collecting large quantities of labeled training data for each new task has limited the usefulness of deep neural networks. Given data from a set of source tasks, this limitation can be overcome using two transfer learning approaches: few-shot learning (FSL) and self-supervised learning (SSL). The former aims to learn ‘how to learn’ by designing learning episodes using source tasks to simulate the challenge of solving the target new task with few labeled samples. In contrast, the latter exploits an annotation-free pretext task across all source tasks in order to learn generalizable feature representations. In this work, we propose a novel Instance-level and Episode-level Pretext Task (IEPT) framework that seamlessly integrates SSL into FSL. Specifically, given an FSL episode, we first apply geometric transformations to each instance to generate extended episodes. At the instancelevel, transformation recognition is performed as per standard SSL. Importantly, at the episode-level, two SSL-FSL hybrid learning objectives are devised: (1) The consistency across the predictions of an FSL classifier from different extended episodes is maximized as an episode-level pretext task. (2) The features extracted from each instance across different episodes are integrated to construct a single FSL classifier for meta-learning. Extensive experiments show that our proposed model (i.e., FSL with IEPT) achieves the new state-of-the-art. ", + "bbox": [ + 233, + 420, + 766, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 695, + 336, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep convolutional neural networks (CNNs) (Krizhevsky et al., 2012; He et al., 2016b; Huang et al., 2017) have seen tremendous successes in a wide range of application fields, especially in visual recognition. However, the powerful learning ability of CNNs depends on a large amount of manually labeled training data. In practice, for many visual recognition tasks, sufficient manual annotation is either too costly to collect or not feasible (e.g., for rare object classes). This has severely limited the usefulness of CNNs for real-world application scenarios. Attempts have been made recently to mitigate such a limitation from two distinct perspectives, resulting in two popular research lines, both of which aim to transfer knowledge learned from the data of a set of source tasks to a new target one: few-shot learning (FSL) and self-supervised learning (SSL). ", + "bbox": [ + 174, + 727, + 825, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "FSL (Fei-Fei et al., 2006; Vinyals et al., 2016; Finn et al., 2017; Snell et al., 2017; Sung et al., 2018) typically takes a ‘learning to learn’ or meta-learning paradigm. That is, it aims to learn an algorithm for learning from few labeled samples, which generalizes well across any tasks. To that end, it adopts an episodic training strategy – the source tasks are arranged into learning episodes, each of which contains $n$ classes and $k$ labeled samples per class to simulate the setting for the target task. Part of the CNN model (e.g., feature extraction subnet, classification layers, or parameter initialization) is then meta-learned for rapid adaptation to new tasks. ", + "bbox": [ + 176, + 859, + 825, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac5ad2ee0448bf35ac7fc284a145ae6ff0d67b950b288048f1c82a5d8bf0456d.jpg", + "image_caption": [ + "Figure 1: Schematic of our approach to FSL. Given a training episode, we apply 2D rotations by 0, 90, 180, and 270 degrees to each instance to generate four extended episodes. After going through a feature extraction CNN, four losses over three branches are designed: (1) In the top branch, we employ a self-supervised rotation classifier with the instance-level SSL loss $\\mathcal { L } _ { i n s t }$ . (2) In the middle branch, an FSL classifier is exploited to predict the FSL classification probabilities for each episode. We maximize the classification consistency among the extended episodes by forcing the four probability distributions to be consistent using $\\mathcal { L } _ { e p i s }$ . The average supervised FSL loss $\\mathcal { L } _ { a u x }$ is also computed. (3) In the bottom branch, we utilize an integration transformer module to fuse the features extracted from each instance with different rotation transformations; they are then used to compute an integrated FSL loss $\\mathcal { L } _ { i n t e g }$ . Among the four losses, $\\mathcal { L } _ { i n s t }$ and $\\mathcal { L } _ { e p i s }$ are the self-supervised losses, and $\\mathcal { L } _ { a u x }$ and $\\mathcal { L } _ { i n t e g }$ are the supervised losses. " + ], + "image_footnote": [], + "bbox": [ + 181, + 111, + 813, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 402, + 825, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, SSL (Doersch et al., 2015; Noroozi & Favaro, 2016; Iizuka et al., 2016; Doersch & Zisserman, 2017; Noroozi et al., 2018) does not require the source data to be annotated. Instead, it exploits an annotation-free pretext task on the source task data in the hope that a task-generalizable feature representation can be learned from the source tasks for easy adoption or adaptation in a target task. Such a pretext task gets its self-supervised signal at the per-instance level. Examples include rotation and context prediction (Gidaris et al., 2018; Doersch et al., 2015), jigsaw solving (Noroozi & Favaro, 2016), and colorization (Iizuka et al., 2016; Larsson et al., 2016). Since these pretext tasks are class-agnostic, solving them leads to the learning of transferable knowledge. ", + "bbox": [ + 173, + 465, + 825, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since both FSL and SSL aim to reduce the need of collecting a large amount of labeled training data for a target task by transferring knowledge from a set of source tasks, it is natural to consider combining them in a single framework. Indeed, two recent works (Gidaris et al., 2019; Su et al., 2020) proposed to integrate SSL into FSL by adding an auxiliary SSL pretext task in an FSL model. It showed that the SSL learning objective is complementary to that of FSL and combining them leads to improved FSL performance. However, in (Gidaris et al., 2019; Su et al., 2020), SSL is combined with FSL in a superficial way: it is only taken as a separate auxiliary task for each single training instance and has no effect on the episodic training pipeline of the FSL model. Importantly, by ignoring the class labels of samples, the instance-level SSL learning objective is weak on its own. Since meta-learning across episodes is the essence of most contemporary FSL models, we argue that adding instance-level SSL pretext tasks alone fails to exploit fully the complementarity of the aforementioned FSL and SSL, for which a closer and deeper integration is needed. ", + "bbox": [ + 174, + 583, + 825, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To that end, in this paper we propose a novel Instance-level and Episode-level Pretext Task (IEPT) framework for few-shot recognition. Apart from adding an instance-level pretext SSL task as in (Gidaris et al., 2019; Su et al., 2020), we introduce two episode-level SSL-FSL hybrid learning objectives for seamless SSL-FSL integration. Concretely, as illustrated in Figure 1, our full model has three additional learning objectives (besides the standard FSL one): (1) Different rotation transformations are applied to each original few-shot episode to generate a set of extended episodes, where each image has a rotation label for the instance-level pretext task (i.e., to predict the rotation label). (2) The consistency across the predictions of an FSL classifier from different extended episodes is maximized as an episode-level pretext task. For each training image, the rotation transformation does not change its semantic content and hence its class label; the FSL classifier predictions across different extended episodes thus should be consistent, hence the consistency regularization objective. (3) The correlation of features across instances from these extended episodes is modeled by a transformer-based attention module, optimizing the fusion of the features of each instance/image and its various rotation-transformed versions mainly for task adaptation during meta-testing. Importantly, with these three new learning objectives introduced in IEPT, any meta-learning based FSL model can now benefit more from SSL by fully exploiting their complementarity. ", + "bbox": [ + 174, + 757, + 825, + 922 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions are: (1) For the first time, we propose both instance-level and episode-level pretext tasks (IEPT) for integrating SSL into FSL. The episode-level pretext task enables episodic training of SSL and hence closer integration of SSL with FSL. (2) In addition to these pretext tasks, FSL further benefits from SSL by integrating features extracted from various rotation-transformed versions of the original training instances. The optimal way of feature integration is learned by a transformer-based attention module, which is mainly designed for task adaptation during meta-testing. (3) Extensive experiments show that the proposed model achieves the new state-of-the-art. ", + "bbox": [ + 174, + 166, + 825, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 290, + 343, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-Shot Learning. The recent FSL studies are dominated by meta-learning based methods. They can be divided into three groups: (1) Metric-based methods (Vinyals et al., 2016; Snell et al., 2017; Sung et al., 2018; Allen et al., 2019; Xing et al., 2019; Li et al., 2019a;b; Wu et al., 2019; Ye et al., 2020; Afrasiyabi et al., 2020; Liu et al., 2020; Zhang et al., 2020) aim to learn the distance metric between feature embeddings. The focus of these methods is often on meta-learning of a feature-extraction CNN, whilst the classifiers used are of simple form such as a nearest-neighbor classifier. (2) Optimization-based methods (Finn et al., 2017; Ravi & Larochelle, 2017; Rusu et al., 2019; Lee et al., 2019) learn to optimize the model rapidly given a few labeled samples per class in the new task. (3) Model-based methods (Santoro et al., 2016; Munkhdalai & Yu, 2017; Mishra et al., 2018) focus on designing either specific model structures or parameters capable of rapid updating. Apart from these three groups of methods, other FSL methods have attempted feature hallucination (Schwartz et al., 2018; Hariharan & Girshick, 2017; Gao et al., 2018; Wang et al., 2018; Zhang et al., 2019; Tsutsui et al., 2019) which generates additional samples from the given few shots for network finetuning, and parameter predicting (Qiao et al., 2018; Qi et al., 2018; Gidaris & Komodakis, 2019; 2018) which learns to predict part of the parameters of a network given few samples of new classes for quick adaptation. In this work, we adopt the metric-based Prototypical Network (ProtoNet) (Snell et al., 2017) as the basic FSL classifier for the main instantiation of our IEPT framework due to its simplicity and popularity. However, we show that any meta-learning based FSL method can be combined with our IEPT (see results in Figure 2(c)). ", + "bbox": [ + 174, + 327, + 826, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning. In SSL, it is assumed that the source task data is label-free and a pretext task is designed to provide self-supervision signals at the instance-level. Existing SSL approaches differ mainly in the pretext task design. These include predicting the rotation angle (Gidaris et al., 2018) and the context of image patch (Doersch et al., 2015; Nathan Mundhenk et al., 2018), jigsaw solving (Noroozi & Favaro, 2016; Noroozi et al., 2018) (i.e. shuffling and then reordering image patch), and performing images reversion (Iizuka et al., 2016; Pathak et al., 2016; Larsson et al., 2016). SSL has been shown to be beneficial to various down-steam tasks such as semantic object matching (Novotny et al., 2018), object segmentation (Ji et al., 2019) and object detection (Doersch & Zisserman, 2017) by learning transferable feature presentations for these tasks. ", + "bbox": [ + 174, + 597, + 825, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Integrating Self-Supervised Learning into Few-Shot Learning. To the best of our knowledge, only two recent works (Gidaris et al., 2019; Su et al., 2020) have attempted combining SSL with FSL. However, the integration of SSL into FSL is often shallow: the original FSL training pipeline is intact; in the meantime, an additional loss on each image w.r.t. a self-supervised signal like the rotation angle or relative patch location is introduced. With pretext tasks solely at the instance level, combining the two approaches (i.e., SSL and FSL) can only be superficial without fully exploiting the episodic training pipeline unique to FSL. Different from (Gidaris et al., 2019; Su et al., 2020), we introduce an episode-level pretext task to integrate SSL into the episodic training in FSL fully. Specifically, the consistency across the predictions of an FSL classifier from different extended episodes is maximized to reflect the fact that various rotation transformations should not alter the class-label prediction. Moreover, features of each instance and its various rotation-transformed versions are now fused for FSL classification, to integrate SSL with FSL for the supervised classification task. Our experimental results show that thanks to the closer integration of SSL and FSL, our IEPT clearly outperforms (Gidaris et al., 2019; Su et al., 2020) (see Table 1). ", + "bbox": [ + 174, + 729, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 METHODOLOGY ", + "text_level": 1, + "bbox": [ + 176, + 102, + 339, + 117 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 PRELIMINARY ", + "text_level": 1, + "bbox": [ + 174, + 125, + 313, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Setting. Given an $n$ -way $k$ -shot FSL task sampled from a test set $\\mathcal { D } _ { t }$ , to imitate the test setting, an FSL model is typically trained in an episodic way. That is, $n$ -way $k$ -shot episodes are randomly sampled from a training set $\\mathcal { D } _ { s }$ , where the class label space of $\\mathcal { D } _ { s }$ has no overlap with that of $\\mathcal { D } _ { t }$ . Each episode $E _ { e }$ contains a support set $S _ { e }$ and a query set $\\mathcal { Q } _ { e }$ . Concretely, we first randomly sample a set of $n$ classes $\\mathcal { C } _ { e }$ from the training set, and then generate $S _ { e }$ and $\\mathcal { Q } _ { e }$ by sampling $k$ support samples and $q$ query samples from each class in $\\mathcal { C } _ { e }$ , respectively. Formally, we have $\\mathcal S _ { e } \\stackrel { - } { = } \\{ ( x _ { i } , y _ { i } ) \\bar { | } y _ { i } \\in \\mathcal C _ { e } , i \\stackrel { - } { = } 1 , . . . , n \\stackrel { - } { \\times } k \\}$ and $\\mathcal { Q } _ { e } = \\{ ( x _ { i } , y _ { i } ) | y _ { i } \\in \\mathcal { C } _ { e } , i = 1 , . . . , n \\times q \\}$ , where $S _ { e } \\cap \\mathcal { Q } _ { e } = \\emptyset$ . For simplicity, we denote $l _ { k } = n \\times k$ and $l _ { q } = n \\times q$ . In the meta-training stage, the training process has an inner and an outer loop in each episode: in the inner loop, the model is updated using $S _ { e }$ ; its performance is then evaluated on the query set $\\mathcal { Q } _ { e }$ in the outer loop to update the model parameters or algorithm that one wants to meta-learn. ", + "bbox": [ + 173, + 148, + 825, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Basic FSL Classifier. We employ ProtoNet (Snell et al., 2017) as the basic FSL model. This model has a feature-extraction CNN and a simple non-parametric classifier. The parameter of the feature extractor is to be meta-learned. Concretely, in the inner loop of an episode, ProtoNet fixes the feature extractor and computes the mean feature embedding for each class as follows: ", + "bbox": [ + 174, + 309, + 825, + 366 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/af53dfdfebf92b14e994742c9d44f9cec3ef0b37d7e37b409e069334abdc5140.jpg", + "text": "$$\nh _ { c } = \\frac { 1 } { k } \\cdot \\sum _ { ( x _ { i } , y _ { i } ) \\in S _ { e } } f _ { \\phi } ( x _ { i } ) \\cdot \\mathbb { I } ( y _ { i } = c ) ,\n$$", + "text_format": "latex", + "bbox": [ + 372, + 371, + 624, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where class $c \\in \\mathcal { C } _ { e }$ , $f _ { \\phi }$ is a feature extractor with learnable parameters $\\phi$ , and $\\mathbb { I }$ is the indicator function. By computing the distance between the feature embedding of each query sample and that of the corresponding class, the loss function used to meta-learn $\\phi$ in the outer loop is defined as: ", + "bbox": [ + 174, + 417, + 823, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f5b0a49ba9ad956ade834f7dac81d4842c7687c92752b1ebb7c1e41414e7ca0f.jpg", + "text": "$$\n\\mathcal { L } _ { f s l } ( S _ { e } , \\mathcal { Q } _ { e } ) = \\frac { 1 } { | \\mathcal { Q } _ { e } | } \\sum _ { ( x _ { i } , y _ { i } ) \\in \\mathcal { Q } _ { e } } - \\log \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { y _ { i } } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ) ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 272, + 465, + 723, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d ( \\cdot , \\cdot )$ denotes a distance function (e.g., the $l _ { 2 }$ distance). ", + "bbox": [ + 173, + 513, + 581, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 PRETEXT TASKS IN IEPT ", + "text_level": 1, + "bbox": [ + 174, + 536, + 392, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The schematic of our IEPT is illustrated in Figure 1. We first define a set of 2D-rotation operators $\\mathcal { G } = \\{ g _ { r } | r = 0 , . . . , R - 1 \\}$ , where $g _ { r }$ means the operator of rotating the image by $r ^ { * 9 0 }$ degrees and $R$ is the total number of rotations ( $R = 4$ in our implementation). Given an original episode $E _ { e } = \\{ S _ { e } , \\mathcal { Q } _ { e } \\}$ as described in Sec. 3.1, we utilize the 2D-rotation operators from $\\mathcal { G }$ in turn to transform each image in $E _ { e }$ . This results in a set of $R$ extended episodes (including the original one) $E = \\{ \\{ \\mathcal { S } _ { e } ^ { r } , \\bar { \\mathcal { Q } } _ { e } ^ { r } \\} | r = 0 , . . . , R - 1 \\}$ , where $S _ { e } ^ { r } \\ = \\ \\{ ( x _ { i } , y _ { i } , \\bar { r } ) | y _ { i } \\in C _ { e } , i \\ = \\ \\bar { 1 } , . . . , l _ { k } \\}$ and $\\mathcal { Q } _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in C _ { e } , i = 1 , . . . , l _ { q } \\} ,$ . Now each episode is denoted as $E _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in$ $C _ { e } , i = 1 , . . . , l _ { k } , l _ { k } + 1 , . . . , l _ { k } + l _ { q } \\}$ , where the first $l _ { k }$ samples are from $S _ { e } ^ { r }$ and the rest from $\\mathcal { Q } _ { e } ^ { r }$ Note that $\\{ \\cal { S } _ { e } ^ { 0 } , \\cal { Q } _ { e } ^ { 0 } \\}$ is the original episode $\\{ \\boldsymbol { S } _ { e } , \\boldsymbol { \\mathcal { Q } } _ { e } \\}$ . With the rotation transformations, each sample $( x _ { i } , y _ { i } , r _ { i } )$ in $E$ carries a class label $y _ { i }$ for supervised learning (from the inherent class) and a label $r _ { i }$ from the rotation operator for self-supervised learning. After generating the set of extended episodes $E$ , the feature extractor $f _ { \\phi }$ is applied to each image $x _ { i }$ in $E$ . On these episodes, we design two self-supervised pretext tasks, one at the instance-level and the other episode-level. ", + "bbox": [ + 173, + 561, + 825, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-Level Pretext Task. The instance-level task is to recognize different rotation transformations. The idea is that if the model to be meta-learned here (i.e., $f _ { \\phi } )$ can be used to distinguish different transformations, it must understand the canonical poses of objects (e.g., animals have legs touching the ground and trees have leaves on top), a vital part of class-agnostic and thus transferable knowledge. With the self-supervised rotation label $r _ { i }$ , we consider the mapping: $f _ { \\theta _ { r o t } } : x _ { i } \\mapsto r _ { i }$ for each instance $( x _ { i } , y _ { i } , r _ { i } ) \\in E$ , where $f _ { \\theta _ { r o t } }$ is a rotation classifier with learnable parameters $\\theta _ { r o t }$ . Given the input pair $( x _ { i } , r _ { i } )$ , the total instance-level rotation loss is a cross-entropy loss: ", + "bbox": [ + 173, + 751, + 826, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/efabe317b4928532f5cf8bbec84ba16e3a6148c3e31c52bd61720019bd8d7f06.jpg", + "text": "$$\n\\mathcal { L } _ { i n s t } = \\frac { 1 } { R ( l _ { k } + l _ { q } ) } \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { ( x _ { i } , y _ { i } , r _ { i } ) \\in E _ { e } ^ { r } } - \\log \\frac { \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r _ { i } } ) } { \\sum _ { r ^ { \\prime } = 0 } ^ { R - 1 } \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r ^ { \\prime } } ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 251, + 854, + 743, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] \\in \\mathbb { R } ^ { R }$ is the rotation scoring vector and $[ \\cdot ] _ { r }$ means taking the $r$ -th element. ", + "bbox": [ + 169, + 909, + 803, + 925 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Episode-Level Pretext Task. We design the episode-level task based on a simple principle: although different extended episodes contain images with different rotation transformations, these transformations do not change their class labels. Consequently, the FSL classifier should produce consistent probability distributions for each instance across different extended episodes. Such consistency can be measured using the Kullback–Leibler (KL) divergence. Formally, for each extended episode $\\{ S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } \\}$ in $E$ , we first define the probability distribution of FSL classification over the query set $\\mathcal { Q } _ { e } ^ { r }$ as $\\mathcal { P } _ { e } ^ { r } = [ p _ { 1 } ^ { r } ; \\cdot \\cdot \\cdot ; p _ { l _ { q } } ^ { r } ] \\in \\mathbb { R } ^ { l _ { q } \\times n }$ , where $p _ { i } ^ { r } \\in \\mathbb { R } ^ { n }$ is the probability distribution for $x _ { i }$ in $\\mathcal { Q } _ { e } ^ { r }$ with its $c$ -th element $[ p _ { i } ^ { r } ] _ { c }$ $\\mathbf { \\bar { \\psi } } c = 1 , . . . , n )$ being: ", + "bbox": [ + 173, + 102, + 826, + 219 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2c3050dc8781955ba9dbc5fa45a1818fd14c0aa77604f3b6b6b1e8ea3d5cb85c.jpg", + "text": "$$\n[ p _ { i } ^ { r } ] _ { c } = \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ^ { r } ) ) } { \\sum _ { c ^ { \\prime } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c ^ { \\prime } } ^ { r } ) ) } .\n$$", + "text_format": "latex", + "bbox": [ + 379, + 220, + 619, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above probability is computed as in Sec. 3.1 and the class embedding $h _ { c } ^ { r }$ is obtained from $S _ { e } ^ { r }$ The mean probability distribution of the $R$ extended episodes is thus given by: ", + "bbox": [ + 173, + 257, + 823, + 285 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/3f129c945a79dc4d05d8d1ab808e4ca1159962847a396d4818341c01b3452008.jpg", + "text": "$$\n\\hat { p } _ { i } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } p _ { i } ^ { r } .\n$$", + "text_format": "latex", + "bbox": [ + 441, + 287, + 558, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total episode-level consistency regularization loss is computed with the KL divergence loss: ", + "bbox": [ + 173, + 332, + 802, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/9daa6c2ade77f619cc6c155e3739cd646215dd62cd89b59454ee5609fbdb7a19.jpg", + "text": "$$\n\\mathcal { L } _ { e p i s } = \\frac { 1 } { R l _ { q } } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { i = 1 } ^ { l _ { q } } \\mathrm { m e a n } ( p _ { i } ^ { r } ( \\log p _ { i } ^ { r } - \\log \\hat { p } _ { i } ) ) .\n$$", + "text_format": "latex", + "bbox": [ + 330, + 348, + 666, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm { m e a n } ( \\cdot )$ is an element-wise averaging function. ", + "bbox": [ + 174, + 395, + 531, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 INTEGRATED FSL TASK ", + "text_level": 1, + "bbox": [ + 174, + 416, + 382, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The two tasks introduced so far are self-supervised tasks without using the class labels in the query set. Now we describe how in the supervised classification task, the extended episodes can be used. ", + "bbox": [ + 173, + 443, + 825, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the set of extended episodes $E$ , we denote the feature set of $E$ as $E _ { e m b }$ , where $E _ { e m b } =$ $\\{ f _ { \\phi } ( x _ { i } ) | ( x _ { i } , y _ { i } , r ) \\in E _ { e } ^ { r } , r = 0 , \\cdots , R - 1 , i = 1 , . . . , l _ { k } + l _ { q } \\}$ . Note that each extended episode in $E$ corresponds to one specific rotation transformation of the same set of images from the original episode $E _ { e }$ . Therefore, in order to capture the correlation among instances with different transformations and learn how best combine them to form the class mean for meta-learning, an instance attention module is deployed w.r.t. each image in $E _ { e }$ (i.e., all images are assumed to be independent). Specifically, based on $E _ { e m b }$ , we construct the feature tensor $\\bar { F } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }$ , where $d$ is the feature dimension. We then adopt a transformer to obtain the integrated representation for FSL classification. The transformer architecture is based on a self-attention mechanism, as in (Vaswani et al., 2017). It receives the triplet input $( F , F , F )$ as $( Q , K , V )$ (Query, Key, and Value, respectively). With $\\dot { F } ^ { ( i ) }$ being the $i$ -th row of $F$ (w.r.t. the $i$ -th image in $E _ { e }$ ), the attentive module is defined as: ", + "bbox": [ + 173, + 477, + 826, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/42b72ef524a028212b41d64b4214ac497599dc9918626fb22811be7f56b4fea4.jpg", + "text": "$$\n\\begin{array} { r l } & { ( \\boldsymbol { F } _ { Q } ^ { ( i ) } , \\boldsymbol { F } _ { K } ^ { ( i ) } , \\boldsymbol { F } _ { V } ^ { ( i ) } ) = ( \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { Q } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { K } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { V } ) , } \\\\ & { \\boldsymbol { F } _ { a t t } ^ { ( i ) } = \\boldsymbol { F } ^ { ( i ) } + \\mathrm { s o f t m a x } ( \\frac { \\boldsymbol { F } _ { Q } ^ { ( i ) } ( \\boldsymbol { F } _ { K } ^ { ( i ) } ) ^ { T } } { \\sqrt { d _ { K } } } ) \\boldsymbol { F } _ { V } ^ { ( i ) } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 325, + 635, + 669, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d _ { K } = d$ , and $W _ { Q } , W _ { K } , W _ { V }$ represent the parameters of three fully-connected layers respectively (the parameters of the integration transformer are collected as $\\theta _ { i n t }$ ). Note that the key and value are computed from each image and its augmented versions, i.e., they are computed independently without using inter-image correlation. With the attentive feature $F _ { a t t } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }$ , the integrated representation $F _ { i n t e g } = [ F ^ { S } ; F ^ { Q } ] \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R d }$ ( $F ^ { S }$ and $F ^ { Q }$ are respectively for the support set and query set) is given by: ", + "bbox": [ + 173, + 699, + 825, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/8780c8be02682008dc4bfb993ddfe86f70bdce00077a0123f7718852eb6f7393.jpg", + "text": "$$\nF _ { i n t e g } = \\mathrm { f l a t t e n } ( F _ { a t t } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 418, + 784, + 576, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where flatten $( \\cdot )$ denotes flattening $F _ { a t t }$ along the last two dimensions, i.e., concatenating the attentive features from different extended episodes for the corresponding images. The integrated representation is then inputted to the FSL classifier to define the FSL classification loss: ", + "bbox": [ + 176, + 801, + 825, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/6fe4d30f749bff32263c8788315d1793d18a141b81fbea68098739d2758d49b4.jpg", + "text": "$$\n\\mathcal { L } _ { i n t e g } = \\frac { 1 } { l _ { q } } \\cdot \\sum _ { i = 1 } ^ { l _ { q } } - \\log \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y _ { i } } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) }\n$$", + "text_format": "latex", + "bbox": [ + 331, + 845, + 666, + 890 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the class embedding the integrated FSL task act $\\begin{array} { r } { h _ { c } ^ { f } = \\frac { 1 } { k } \\cdot \\sum _ { i = 1 } ^ { l _ { k } } F _ { i } ^ { S } \\cdot \\mathbb { I } ( y _ { i } = c ) } \\end{array}$ is computed on the support set. Note thatrediction averaging. ", + "bbox": [ + 173, + 893, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 TOTAL LOSS ", + "text_level": 1, + "bbox": [ + 174, + 103, + 303, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The total training loss for our full model consists of the self-supervised losses from the pretext tasks and the supervised losses from the FSL tasks. In this work, in addition to $\\mathcal { L } _ { i n t e g }$ in Eq. (10), another supervised FSL loss $\\mathcal { L } _ { a u x }$ is also used (see Figure 1). $\\mathcal { L } _ { a u x }$ is the average FSL classification loss over the extended episodes. Formally, it can be written as: ", + "bbox": [ + 174, + 130, + 825, + 185 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/2f4093022e30a93dbeec9c5c4b2ce4413b4beb43881363ab8e48d87496b558e5.jpg", + "text": "$$\n\\mathcal { L } _ { a u x } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\mathcal { L } _ { f s l } ( S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } )\n$$", + "text_format": "latex", + "bbox": [ + 395, + 193, + 602, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the total loss $\\mathcal { L } _ { t o t a l }$ for training our full model is given as follows: ", + "bbox": [ + 173, + 243, + 678, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/7428972404be365c998fd96b04e5ae4729a396ac2be95422a01502c2b04b28b3.jpg", + "text": "$$\n\\begin{array} { r } { \\mathcal { L } _ { t o t a l } = \\underbrace { \\overbrace { w _ { 1 } * \\mathcal { L } _ { i n s t } } ^ { i n s t a n c e { - } l e v e l } + \\overbrace { w _ { 2 } * \\mathcal { L } _ { e p i s } } ^ { e p i s o d e { - } l e v e l } } _ { \\mathrm { s e l f - s u p e r v i s e d l o s s } } + \\underbrace { w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t e g } } _ { \\mathrm { s u p e r v i s e d l o s s } } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 305, + 265, + 692, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w _ { 1 } , w _ { 2 } , w _ { 3 }$ are the loss weight hyperparameters. ", + "bbox": [ + 173, + 325, + 532, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5 INFERENCE ", + "text_level": 1, + "bbox": [ + 174, + 358, + 295, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the test stage, we only exploit the integrated representation $F _ { i n t e g }$ for the final FSL prediction. The predicted class label for $x _ { i } \\in \\mathcal { Q } _ { e }$ can be computed with Eq. (10) as: ", + "bbox": [ + 174, + 383, + 825, + 414 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/2e94de396b91c16c863f7b31f20db819e61061ae5d0a9119814439b597279b54.jpg", + "text": "$$\ny _ { i } ^ { p r e d } = \\underset { y \\in \\mathcal { C } _ { e } } { \\operatorname { a r g m a x } } \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) } .\n$$", + "text_format": "latex", + "bbox": [ + 352, + 420, + 647, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6 FULL IEPT ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 478, + 387, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For easy reproduction, we present the full algorithm for FSL with IEPT in Algorithm 1. Once learned, with the learned $\\psi$ , we can perform the inference over the test episodes with Eq. (13). ", + "bbox": [ + 171, + 505, + 825, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 FSL with IEPT ", + "text_level": 1, + "bbox": [ + 174, + 550, + 362, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: The training set $\\mathcal { D } _ { s }$ , the rotation operator set $\\mathcal { G }$ \nThe loss weight hyperparameters $w _ { 1 } , w _ { 2 } , w _ { 3 }$ \nOutput: The learned $\\psi$ \n1: Randomly initialize all learnable parameters $\\psi = \\{ \\phi , \\theta _ { r o t } , \\theta _ { i n t } \\}$ \n2: for iteration $= 1$ , ..., MaxIteration do \n3: Randomly sample episode $E _ { e }$ from $\\mathcal { D } _ { s }$ \n4: Generate the set of extended episodes $E$ from $E _ { e }$ using $\\mathcal { G }$ \n5: Compute the SSL loss $\\mathcal { L } _ { i n s t }$ for the instance-level pretext task with Eq. (3) \n6: Compute the SSL loss $\\mathcal { L } _ { e p i s }$ for the episode-level pretext task with Eq. (6) \n7: Compute the supervised FSL loss $\\mathcal { L } _ { a u x }$ over the extended episodes with Eq. (11) \n8: Compute the supervised FSL loss $\\mathcal { L } _ { i n t e g }$ for the integrated episode with Eq. (10) \n9: $\\mathcal { L } _ { t o t a l } = w _ { 1 } * \\mathcal { L } _ { i n s t } + w _ { 2 } * \\mathcal { L } _ { e p i s } + w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t }$ teg \n10: Update $\\psi$ based on $\\nabla _ { \\psi }$ Ltotal \n11: end for \n12: return $\\psi$ . ", + "bbox": [ + 174, + 566, + 696, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 781, + 326, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 813, + 375, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Two widely-used FSL datasets are selected: miniImageNet (Vinyals et al., 2016) and tieredImageNet (Ren et al., 2018). The first dataset consists of a total number of 100 classes (600 images per class) and the train/validation/test split is set to 64/16/20 classes as in (Ravi & Larochelle, 2017). The second dataset is a larger dataset including 608 classes totally (nearly 1,200 images per class), which is split into 351/97/160 classes for train/validation/test. Both datasets are subsets sampled from ImageNet (Russakovsky et al., 2015). ", + "bbox": [ + 173, + 839, + 826, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1ab2ff5c5ab953a64fe5c51c547bda001032e318953029864dd4b1643d753419.jpg", + "table_caption": [ + "Table 1: Comparative results for 5-way 1/5-shot FSL. The mean classification accuracies (top-1, $\\%$ with the $9 5 \\%$ confidence intervals are reported. $\\dagger$ indicates the result is reproduced by ourselves. " + ], + "table_footnote": [], + "table_body": "
MethodminiImageNettieredImageNet
Backbone1-shot5-shot1-shot5-shot
MatchingNet (Vinyals etal.,2016)Conv4-6443.56± 0.8455.31± 0.73
ProtoNet† (Snell et al., 2017)Conv4-6452.61 ± 0.5271.33 ± 0.4153.33 ± 0.5072.10 ± 0.41
MAML (Finn et al., 2017)Conv4-6448.70 ± 1.8463.10 ±0.9251.67 ± 1.8170.30 ±0.08
Relation Net (Sung et al.,2018)Conv4-6450.40 ± 0.8065.30 ±0.7054.48 ± 0.9371.32 ± 0.78
IMP† (Allen et al., 2019)Conv4-6452.91 ± 0.4971.57 ± 0.4253.63 ± 0.5171.89 ± 0.44
DN4 (Li et al., 2019b)Conv4-6451.24 ± 0.7471.02 ± 0.641
DN PARN (Wu et al., 2019)Conv4-6455.22 ± 0.8471.55 ± 0.66
PN+rot (Gidaris et al., 2019)Conv4-6453.63 ± 0.4371.70 ± 0.36
CC+rot (Gidaris et al., 2019)Conv4-6454.83 ± 0.4371.86 ± 0.33
DSN-MR (Simon et al., 2020)Conv4-6455.88 ± 0.9070.50±0.68
Centroid (Afrasiyabi et al., 2020)Conv4-6453.14 ± 1.0671.45 ± 0.72
Neg-Cosine (Liu et al., 2020)Conv4-6452.84 ±0.7670.41 ± 0.66一
IEPT (ours)Conv4-6456.26± 0.4573.91±0.34 58.25±0.48 75.63±0.46
ProtoNet† (Snell et al.,2017)Conv4-51253.25± 0.4473.15 ± 0.3557.88 ±0.5076.82±0.40
MAML (Finn et al., 2017)Conv4-51249.33 ±0.6065.17 ± 0.4952.84± 0.5670.91 ± 0.46
Relation Net (Sung et al.,2018)Conv4-51250.86 ± 0.5767.32 ± 0.4454.69 ±0.5972.71 ±0.43
PN+rot (Gidaris et al.,2019)Conv4-51256.02 ±0.4674.00 ± 0.35一
CC+rot (Gidaris et al., 2019)Conv4-51256.27 ± 0.4374.30 ± 0.33一
IEPT (ours)Conv4-51258.43±0.4675.07±0.3360.91±0.5979.61±0.45
ProtoNet† (Snell et al.,2017)ResNet-1262.39 ±0.5180.53±0.4268.23±0.5084.03 ± 0.41
TADAM (Oreshkin et al., 2018)ResNet-1258.50 ± 0.3076.70 ±0.38一
MetaOptNet (Lee et al.,2019)ResNet-1262.64± 0.6178.63 ± 0.4665.99 ± 0.7281.56 ± 0.63
MTL (Sun et al., 2019)ResNet-1261.20 ± 1.8075.50 ± 0.8065.62 ± 1.8080.61 ±0.90
CAN (Hou et al., 2019)ResNet-1263.85 ±0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
AM3 (Xing et al., 2019)ResNet-1265.21 ± 0.4975.20 ±0.3667.23 ± 0.3478.95 ± 0.22
Shot-Free (Ravichandran et al.,2019)ResNet-1259.04 ± 0.4377.64 ± 0.3966.87 ± 0.4382.64 ±0.43
Neg-Cosine (Liu et al., 2020)ResNet-1263.85 ± 0.8181.57 ± 0.56
Distill (Tian et al., 2020)ResNet-1264.82 ±0.6082.14 ± 0.4371.52 ± 0.6986.03 ±0.49
DSN-MR (Simon et al.,2020)ResNet-1264.60 ±0.7279.51± 0.5067.39 ± 0.8282.85 ±0.56
DeepEMD (Zhang et al.,2020)ResNet-1265.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
FEAT (Ye et al.,2020)ResNet-1266.78±0.2082.05 ± 0.1470.80 ±0.2384.79 ± 0.16
ProtoNet+Rotation (Su et al.,2020)ResNet-1876.00±0.6078.90±0.70
IEPT (ours)ResNet-1267.05±0.44882.90±0.30)72.24±0.50 86.73±0.34
", + "bbox": [ + 178, + 141, + 823, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Extractors. For fair comparison with published results, our IEPT adopts three widely-used feature extractors: Conv4-64 (Vinyals et al., 2016), Conv4-512, and ResNet-12 (He et al., 2016a). Particularly, Conv4-512 is almost the same as Conv4-64 except having a different channel size of the last convolution layer. To speed up the training process, as in many previous works (Ye et al., 2020; Zhang et al., 2020; Simon et al., 2020), we pretrain all the feature extractors on the training split of each dataset for our IEPT. Following (He et al., 2016a), we use the temperature scaling skill during the training phase. On both datasets, the input image size is $8 4 \\times 8 4$ . The output feature dimensions of Conv4-64, Conv4-512, and ResNet-12 are 64, 512, and 640, respectively. ", + "bbox": [ + 174, + 583, + 825, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We take the 5-way 5-shot (or 1-shot) FSL evaluation setting, as in previous works. We randomly sample 2,000 episodes from the test split and report the mean classification accuracy (top-1, $\\%$ ) as well as the $9 5 \\%$ confidence interval. Since the integration transformer copes with each sample independently, we take a strict non-transductive setting during evaluation. ", + "bbox": [ + 174, + 702, + 825, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. PyTorch is used for our implementation. We utilize the Adam optimizer (Kingma & Ba, 2015) for Conv4-64 & Conv4-512 and the SGD optimizer for ResNet-12 to train our IEPT model. The hyperparameters of our IEPT model are selected according to the performance on the validation split.We will release the code soon. ", + "bbox": [ + 174, + 763, + 825, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 MAIN RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 840, + 323, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to State-of-the-Arts. We compare our IEPT with two groups of baselines: (1) Recent SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020); (2) Representative/latest FSL methods (w/o SSL) (Snell et al., 2017; Finn et al., 2017; Lee et al., 2019; Ravichandran et al., 2019; Simon et al., 2020; Zhang et al., 2020; Ye et al., 2020; Liu et al., 2020). The comparative results for 5-way ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/782095c5043ea0f9182610faf183276e533078eb62759726e4d5284f243230e1.jpg", + "table_caption": [ + "Table 2: Ablation study results for our full IEPT model over miniImageNet and tieredImageNet. Our full model includes two self-supervised losses (i.e. $\\mathcal { L } _ { e p i s }$ and $\\mathcal { L } _ { i n s t . }$ ) and two supervised losses (i.e. $\\mathcal { L } _ { a u x }$ and $L _ { i n t e g . }$ ). Conv4-64 is used as the feature extractor. " + ], + "table_footnote": [], + "table_body": "
miniImageNettieredImageNet
LintegLinstLepisLaux1-shot5-shot1-shot5-shot
√55.04 ± 0.5272.01 ± 0.4156.98 ± 0.4774.15 ± 0.51
√√55.49 ± 0.5672.54± 0.4657.41 ± 0.5174.65 ± 0.50
√55.88± 0.4372.97 ± 0.4057.76± 0.4575.06 ± 0.40
√55.97 ± 0.5773.28 ± 0.3957.83 ± 0.5575.22 ± 0.48
√?>>>√56.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "bbox": [ + 178, + 156, + 820, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af3c8b4f8135fe31235a01a5a12b5a61f1267d2c4e0e0affdba8e3fca164fe3e.jpg", + "image_caption": [ + "Figure 2: (a) Comparison among different combination methods over episodes for FSL with selfsupervision. $\\mathbf { ( b ) }$ Illustration of the effect of different choices of $R$ on the performance of our model $R$ denotes the number of extended episodes used for SSL). (c) Comparative results obtained by our IEPT using different basic FSL classifiers (i.e. ProtoNet, FEAT, and IMP). It can be seen clearly that integrated episode-based fusion leads to more separation between classes. All figures present 5-way 1-shot/5-shot results on miniImageNet, using Conv4-64 as the feature extractor. " + ], + "image_footnote": [], + "bbox": [ + 187, + 261, + 805, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1/5-shot FSL are shown in Table 1. We have the following observations: (1) When compared with the representative/latest FSL methods (w/o SSL), our IEPT achieves the best performance on all datasets and under all settings, validating the effectiveness of SSL with IEPT for FSL. (2) Our IEPT also clearly outperforms the two SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020) which only use instance-level pretext tasks, demonstrating the importance of closer/episode-level integration of SSL into FSL. (3) The improvements achieved by our IEPT over ProtoNet range from $2 \\%$ to $5 \\%$ . Since our IEPT takes ProtoNet as the baseline, the obtained margins provide direct evidence that SSL brings significant benefits to FSL. Note that our IEPT is also shown to be effective under both the fine-grained FSL and cross-domain FSL settings in Sec. 4.3 (see Table 3). ", + "bbox": [ + 173, + 496, + 825, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. Our full IEPT model is trained with four losses (see Eq. (12)), including two self-supervised losses and two supervised losses: the episode-level SSL loss $\\mathcal { L } _ { e p i s }$ , the instance-level SSL loss $\\mathcal { L } _ { i n s t }$ , the auxiliary FSL loss $\\mathcal { L } _ { a u x }$ and the integrated FSL loss $\\bar { L } _ { i n t e g }$ . To demonstrate the contribution of each loss, we present the ablation study results for our full IEPT model in Table 2, where Conv4-64 is used as the backbone. We start with $L _ { i n t e g }$ and then add the additional three losses successively. It can be observed that the performance of our model continuously increases when more losses are used, indicating that each loss contributes to the final performance. ", + "bbox": [ + 174, + 627, + 825, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 FURTHER EVALUATIONS ", + "text_level": 1, + "bbox": [ + 176, + 744, + 383, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Combination Methods over Episodes. We have introduced a transformer-based attention module to fuse the features of each instance from all extended episodes (and an integrated episode can be obtained) for the supervised classification task (see Sec. 3.3). In this experiment, we compare it with two alternative ways of across-episode integration: (1) Averaging extended episodes: the extended episodes are directly fused for FSL classification; (2) Averaging all episodes: the extended episodes as well as the integrated episode are fused for FSL classification. We present the comparative results on miniImageNet in Figure 2(a). For comprehensive comparison, the results of FSL with each single extended episode are also reported. We can observe that: (1) The performance of ‘Episode $0 ^ { \\circ }$ ’ is the highest among the four baselines (i.e., FSL with single extended episode), perhaps because the feature extractor is pretrained on the original images without rotation transformations. (2) FSL by averaging extended episodes (i.e., ‘Averaging extended episodes’) indeed improves each of the four baselines. (3) FSL with integrated episode (i.e., ‘Integrated episode’) is superior to FSL by simply averaging extended episodes. (4) Comparing ‘Integrated episode’ with ‘Averaging all episodes’, the performance of FSL with integrated episode is more stable across different settings, furthering validating the usefulness of our across-episode integration. Overall, the episode-integration module is indeed effective in FSL with self-supervision. This is also supported by the visualization results presented in Appendices A.3 & A.4. ", + "bbox": [ + 173, + 770, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b9c462107ffd68f09bfe11e7079ebc8a088483e96df432ec01a7e10e9eb18d94.jpg", + "table_caption": [ + "Table 3: Comparative results for the fine-grained FSL on CUB (Wah et al., 2011) and the cross-domain FSL on miniImageNet CUB. " + ], + "table_footnote": [], + "table_body": "
MethodBackboneCUBminiImageNet→CUB
1-shot5-shot1-shot5-shot
MatchingNet (Vinyals et al.,2016)Conv4-6461.16±0.8972.86±0.7042.62 ± 0.5556.53± 0.44
ProtoNet (Snell et al.,2017)Conv4-6463.72 ±0.2281.50 ± 0.1550.51 ± 0.5669.28±0.40
MAML (Finn et al.,2017)Conv4-6455.92 ± 0.9572.09 ±0.7643.59 ± 0.5454.18 ± 0.41
Relation Net (Sung et al., 2018)Conv4-6462.45 ± 0.9876.11 ± 0.6949.84 ±0.5468.98 ±0.42
FEAT(Ye et al.,2020)Conv4-6468.87 ±0.2282.90±0.1551.52 ± 0.5470.16 ± 0.40
IEPT(ours)Conv4-6469.97±0.4984.33± 0.3352.68±0.5672.98±0.40
", + "bbox": [ + 176, + 142, + 816, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 257, + 825, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Different Number of Extended Episodes. In all the above experiments, the number of the extended episodes $R$ is set to 4 (rotation by $0 ^ { \\circ }$ , $9 0 °$ , $1 8 0 ^ { \\circ }$ , $2 7 0 ^ { \\circ }$ ). Figure 2(b) shows the impact of the value of $R$ . Note that when $R = 1$ , our IEPT model is equivalent to ProtoNet which is without self-supervision. It can be seen that the performance of our model consistently grows when $R$ increases from 1 to 4. Additionally, the study on exploiting other pretext tasks for our IEPT is presented in Appendix A.1. ", + "bbox": [ + 174, + 348, + 825, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Different Basic FSL Classifiers. As mentioned in Sec. 3.1, we adopt ProtoNet as the basic FSL classifier due to its scalability and simplicity. To further show the effectiveness of our IEPT when other basic FSL classifiers are used, we provide the results obtained by our IEPT using ProtoNet, FEAT, and IMP for FSL in Figure 2(c). It can be clearly observed that our IEPT leads to an improvement of about $1 \\%$ over each basic FSL method (ProtoNet, FEAT, or IMP), indicating that our IEPT can be applied to improve a variety of popular FSL methods. ", + "bbox": [ + 173, + 424, + 825, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparative Results for Fine-Grained FSL and Cross-Domain FSL. To evaluate our IEPT algorithm under the fine-grained FSL and cross-domain FSL settings, we conduct experiments on CUB (Wah et al., 2011) and miniImageNet $ \\mathrm { C U B }$ , respectively. For fine-grained FSL on CUB, following (Ye et al., 2020), we randomly split the dataset into 100 training classes, 50 validation classes, and 50 test classes. For cross-domain FSL on miniImageNet $ \\mathrm { C U B }$ , the 100 training classes are from miniImageNet; the 50 validation classes and 50 test classes (using the aforementioned split for fine-grained FSL) are from CUB. Under both settings, we use Conv4-64 as the feature extractor. The 5-way 1/5-shot FSL results are shown in Table 3. Our IEPT clearly achieves the best results, yielding $1 - 3 \\%$ improvements over the second-best FEAT. This shows the effectiveness of our IEPT under both fine-grained and cross-domain settings. ", + "bbox": [ + 173, + 515, + 825, + 654 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 678, + 318, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have proposed a novel Instance-level and Episode-level Pretext Task (IEPT) framework for integrating SSL into FSL. For the first time, we have introduced an episode-level pretext task for FSL with self-supervision, in addition to the conventional instance-level pretext task. Moreover, we have also developed an episode extension-integration framework by introducing an integration transformer module to fully exploit the extended episodes for FSL. Extensive experiments on two benchmarks demonstrate that the proposed model (i.e., FSL with IEPT) achieves the new state-of-theart. Our ongoing research directions include: exploring other episode-level pretext tasks for FSL with self-supervision, and applying FSL with self-supervision to other vision problems. ", + "bbox": [ + 174, + 712, + 825, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 843, + 326, + 856 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (61976220 and 61832017), Beijing Outstanding Young Scientist Program (BJJWZYJH012019100020098), Open Project Program Foundation of Key Laboratory of Opto-Electronics Information Processing, Chinese Academy of Sciences (OEIP-O-202006), and Alibaba Innovative Research (AIR) Program. ", + "bbox": [ + 174, + 867, + 825, + 922 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Arman Afrasiyabi, Jean-Franc¸ois Lalonde, and Christian Gagne. Associative alignment for few-shot image ´ classification. In European Conference on Computer Vision (ECCV), 2020. ", + "bbox": [ + 174, + 125, + 823, + 151 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kelsey R. Allen, Evan Shelhamer, Hanul Shin, and Joshua B. Tenenbaum. Infinite mixture prototypes for few-shot learning. In International Conference on Machine Learning (ICML), pp. 232–241, 2019. ", + "bbox": [ + 173, + 160, + 821, + 186 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. International Conference on Learning Representations (ICLR), 2020. ", + "bbox": [ + 176, + 195, + 820, + 223 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In International Conference on Computer Vision (ICCV), pp. 2051–2060, 2017. ", + "bbox": [ + 176, + 231, + 823, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In International Conference on Computer Vision (ICCV), pp. 1422–1430, 2015. ", + "bbox": [ + 174, + 266, + 823, + 294 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pp. 594–611, 2006. ", + "bbox": [ + 174, + 301, + 823, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning (ICML), pp. 1126–1135, 2017. ", + "bbox": [ + 174, + 337, + 823, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hang Gao, Zheng Shou, Alireza Zareian, Hanwang Zhang, and Shih-Fu Chang. Low-shot learning via covariancepreserving adversarial augmentation networks. In Advances in Neural Information Processing Systems, pp. 975–985, 2018. ", + "bbox": [ + 173, + 372, + 823, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4367–4375, 2018. ", + "bbox": [ + 173, + 421, + 823, + 448 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Spyros Gidaris and Nikos Komodakis. Generating classification weights with gnn denoising autoencoders for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21–30, 2019. ", + "bbox": [ + 176, + 455, + 823, + 496 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. ", + "bbox": [ + 173, + 503, + 823, + 531 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Perez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In International Conference on Computer Vision (ICCV), pp. 8059–8068, 2019. ", + "bbox": [ + 176, + 540, + 823, + 579 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In International Conference on Computer Vision (ICCV), pp. 3037–3046, 2017. ", + "bbox": [ + 176, + 588, + 823, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016a. ", + "bbox": [ + 174, + 623, + 823, + 650 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016b. ", + "bbox": [ + 174, + 659, + 823, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ruibing Hou, Hong Chang, MA Bingpeng, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot classification. In Advances in Neural Information Processing Systems, pp. 4003–4014, 2019. ", + "bbox": [ + 173, + 694, + 823, + 722 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4700–4708, 2017. ", + "bbox": [ + 173, + 729, + 823, + 757 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Let there be color! joint end-to-end learning of global and local image priors for automatic image colorization with simultaneous classification. ACM Transactions on Graphics (ToG), pp. 1–11, 2016. ", + "bbox": [ + 176, + 765, + 825, + 805 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image clas- ˜ sification and segmentation. In International Conference on Computer Vision (ICCV), pp. 9865–9874, 2019. ", + "bbox": [ + 173, + 814, + 825, + 852 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. ", + "bbox": [ + 173, + 862, + 821, + 888 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pp. 1097–1105, 2012. ", + "bbox": [ + 171, + 897, + 823, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization. In European Conference on Computer Vision (ECCV), pp. 577–593, 2016. ", + "bbox": [ + 171, + 104, + 825, + 131 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10657– 10665, 2019. ", + "bbox": [ + 174, + 140, + 823, + 179 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features for few-shot learning by category traversal. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–10, 2019a. ", + "bbox": [ + 171, + 188, + 823, + 228 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Wenbin Li, Lei Wang, Jinglin Xu, Jing Huo, Yang Gao, and Jiebo Luo. Revisiting local descriptor based imageto-class measure for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7260–7268, 2019b. ", + "bbox": [ + 171, + 236, + 825, + 276 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters: Understanding margin in few-shot classification. In European Conference on Computer Vision (ECCV), 2020. ", + "bbox": [ + 173, + 285, + 823, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. ", + "bbox": [ + 169, + 320, + 823, + 348 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations (ICLR), 2018. ", + "bbox": [ + 173, + 356, + 823, + 383 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tsendsuren Munkhdalai and Hong Yu. Meta networks. In International Conference on Machine Learning (ICML), pp. 2554–2563, 2017. ", + "bbox": [ + 173, + 392, + 823, + 419 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9339–9348, 2018. ", + "bbox": [ + 171, + 428, + 823, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European Conference on Computer Vision (ECCV), pp. 69–84, 2016. ", + "bbox": [ + 171, + 464, + 823, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mehdi Noroozi, Ananth Vinjimoor, Paolo Favaro, and Hamed Pirsiavash. Boosting self-supervised learning via knowledge transfer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9359–9367, 2018. ", + "bbox": [ + 173, + 500, + 825, + 539 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Novotny, Samuel Albanie, Diane Larlus, and Andrea Vedaldi. Self-supervised learning of geometrically stable features through probabilistic introspection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3637–3645, 2018. ", + "bbox": [ + 173, + 547, + 823, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Boris Oreshkin, Pau Rodr´ıguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for ´ improved few-shot learning. In Advances in Neural Information Processing Systems, pp. 721–731, 2018. ", + "bbox": [ + 173, + 595, + 821, + 623 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2536–2544, 2016. ", + "bbox": [ + 171, + 632, + 823, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5822–5830, 2018. ", + "bbox": [ + 169, + 680, + 825, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Siyuan Qiao, Chenxi Liu, Wei Shen, and Alan L Yuille. Few-shot image recognition by predicting parameters from activations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7229–7238, 2018. ", + "bbox": [ + 174, + 715, + 823, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. ", + "bbox": [ + 173, + 765, + 823, + 791 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Avinash Ravichandran, Rahul Bhotika, and Stefano Soatto. Few-shot learning with embedded class models and shot-free meta training. In International Conference on Computer Vision (ICCV), pp. 331–339, 2019. ", + "bbox": [ + 173, + 800, + 823, + 828 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), 2018. ", + "bbox": [ + 178, + 837, + 823, + 876 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. In International Journal of Computer Vision (IJCV), pp. 211–252, 2015. ", + "bbox": [ + 176, + 885, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations (ICLR), 2019. ", + "bbox": [ + 173, + 104, + 823, + 143 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In International Conference on Machine Learning (ICML), pp. 1842–1850, 2016. ", + "bbox": [ + 174, + 151, + 821, + 191 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris, Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object recognition. In Advances in Neural Information Processing Systems, pp. 2850–2860, 2018. ", + "bbox": [ + 174, + 199, + 823, + 239 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. Adaptive subspaces for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4136–4145, 2020. ", + "bbox": [ + 173, + 247, + 821, + 275 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, pp. 4077–4087, 2017. ", + "bbox": [ + 173, + 282, + 823, + 310 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jong-Chyi Su, Subhransu Maji, and Bharath Hariharan. When does self-supervision improve few-shot learning? In European Conference on Computer Vision (ECCV), 2020. ", + "bbox": [ + 173, + 318, + 823, + 344 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019. ", + "bbox": [ + 171, + 353, + 823, + 381 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1199–1208, 2018. ", + "bbox": [ + 176, + 388, + 823, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision (ECCV), 2020. ", + "bbox": [ + 173, + 436, + 825, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Satoshi Tsutsui, Yanwei Fu, and David Crandall. Meta-reinforced synthetic data for one-shot fine-grained visual recognition. In Advances in Neural Information Processing Systems, pp. 3057–3066, 2019. ", + "bbox": [ + 173, + 484, + 821, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pp. 5998–6008, 2017. ", + "bbox": [ + 171, + 518, + 825, + 559 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems, pp. 3630–3638, 2016. ", + "bbox": [ + 171, + 566, + 823, + 594 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. ", + "bbox": [ + 173, + 602, + 823, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7278–7286, 2018. ", + "bbox": [ + 174, + 637, + 825, + 665 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ziyang Wu, Yuwei Li, Lihua Guo, and Kui Jia. Parn: Position-aware relation networks for few-shot learning. In International Conference on Computer Vision (ICCV), pp. 6659–6667, 2019. ", + "bbox": [ + 174, + 672, + 823, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chen Xing, Negar Rostamzadeh, Boris Oreshkin, and Pedro O O. Pinheiro. Adaptive cross-modal few-shot learning. In Advances in Neural Information Processing Systems, pp. 4847–4857, 2019. ", + "bbox": [ + 173, + 707, + 823, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-toset functions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8808–8817, 2020. ", + "bbox": [ + 174, + 742, + 825, + 782 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured classifiers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12203–12213, 2020. ", + "bbox": [ + 174, + 791, + 823, + 830 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of samples. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2770–2779, 2019. ", + "bbox": [ + 171, + 838, + 825, + 866 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 297, + 117 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1 COMPARISON AMONG DIFFERENT SSL STRATEGIES", + "text_level": 1, + "bbox": [ + 174, + 137, + 578, + 151 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To generate the extended episodes in IEPT, we apply four rotation transformations (i.e. rotation by $0 ^ { \\circ }$ , $9 0 °$ , $1 8 0 ^ { \\circ }$ , $2 7 0 ^ { \\circ }$ ) to each image. It makes sense to explore whether other self-supervised strategies are also effective for our IEPT. To this end, we exploit shuffling image patches (see Figure 3) for self-supervised learning (SSL). Specifically, we divide each image into $2 ^ { \\ast } 2$ patches and reorganize the patch orders to obtain a shuffling label. To compare with the rotation strategy fairly, we choose only four shuffling orders: (1, 2, 3, 4), (2, 3, 4, 1), (3, 4, 1, 2) and (4, 1, 2, 3). Note that the (1, 2, 3, 4) shuffling order equals to the original image. Similar to the rotation strategy, a fully-connected layer is utilized to recognize the shuffling order. The comparative results are shown in Table 4. We can see that both IEPT with shuffling and IEPT with rotation achieve better performance than the original ProtoNet. Particularly, IEPT with shuffling yields $1 - 3 \\%$ and $3- 4 \\%$ improvements under 5-shot and 1-shot, respectively. This clearly shows the effectiveness of our IEPT for FSL even when different SSL strategies are used to define the pretext tasks. ", + "bbox": [ + 173, + 165, + 825, + 332 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/826169c133d349e3d44c759fcc93bff28025782503f1373f27ec4f7ad4152385.jpg", + "table_caption": [ + "Table 4: FSL results obtained by our IEPT using two SSL strategies (i.e. rotation and shuffling image patches). Conv4-64 is used as the feature extractor. " + ], + "table_footnote": [], + "table_body": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
ProtoNetConv4-6452.61 ±0.5271.33 ± 0.4153.33 ± 0.5072.10± 0.41
IEPT (rotation)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ±0.4875.63 ± 0.46
IEPT (shuffling)Conv4-6455.57 ± 0.6072.84 ± 0.5457.81 ± 0.4874.92 ± 0.50
", + "bbox": [ + 186, + 381, + 810, + 445 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2d68786b89daf9ea45fcd13f2c425d6edbd6a37859f608778de12061ff067209.jpg", + "image_caption": [ + "Figure 3: Illustration of the self-supervised strategy by shuffling image patches. " + ], + "image_footnote": [], + "bbox": [ + 178, + 468, + 816, + 584 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.2 COMPARISON AMONG DIFFERENT INTEGRATION APPROACHES", + "bbox": [ + 174, + 641, + 651, + 656 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We employ the integration transformer to find the intrinsic correlation of various rotation-transformed instances. The transformer architecture is based on a self-attention mechanism. Concretely, it receives the feature sets of extended episodes as input $Q$ , $K$ and $V$ . Further, it matches each query in $Q$ with a list of keys in $K$ and returns the weighted sum of corresponding values. To show the importance of the transformer module, we compare it with two other integration approaches (i.e. concatenating and averaging) to integrate the features of extended episodes. The comparative results in Table 5 demonstrate that the integration transformer consistently performs better than the simply concatenating/averaging approaches. This suggests that the attention-based integration transformer is a better choice for designing the integration module. ", + "bbox": [ + 173, + 670, + 825, + 795 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/46c04c223884d28fcbcd433602647d05ad9231a973280b9deba6d14a9738132f.jpg", + "table_caption": [ + "Table 5: Comparative results obtained by three different approaches to integrating the features from the extended episodes, with Conv4-64 being the feature extractor. " + ], + "table_footnote": [], + "table_body": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
concatenatingConv4-6451.52±0.6073.36± 0.4950.78±0.6872.79± 0.57
averagingConv4-6451.58 ± 0.6270.97 ± 0.5453.91 ± 0.6972.52 ± 0.59
transformer (ours)Conv4-6456.26± 0.4573.91 ± 0.3458.25 ± 0.4875.63± 0.46
", + "bbox": [ + 184, + 844, + 813, + 909 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/afa66040ad4c1918609b3c5a1e793531019fcc73f4e6f207f4681c14646f499f.jpg", + "image_caption": [ + "Figure 4: Feature visualizations of a number of test episodes using the UMAP algorithm (McInnes et al., 2018). Each row indicates a group of test extended episodes (the first four columns, rotation by $0 ^ { \\circ }$ , $9 0 °$ , $1 8 0 ^ { \\circ }$ , $2 7 0 ^ { \\circ }$ ) and their integrated episode (the last column). The 5-way 5-shot FSL (with Conv4-64) is adopted on miniImageNet. " + ], + "image_footnote": [], + "bbox": [ + 184, + 102, + 797, + 597 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.3 FEATURE VISUALIZATIONS OF TEST EPISODES ", + "text_level": 1, + "bbox": [ + 174, + 683, + 542, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We provide the feature visualizations of test episodes in Figure 4. It can be seen that an integrated episode (the last one in each row) clearly has a better cluster data structure than the corresponding four extended episodes (the first four ones in each row). This indicates that our transformer-based across-episode integration is indeed effective for few-shot classification with self-supervision. ", + "bbox": [ + 174, + 713, + 825, + 768 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.4 ATTENTION VISUALIZATION OF TEST EPISODES ", + "text_level": 1, + "bbox": [ + 174, + 796, + 549, + 810 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present attention map visualization of two test episodes (left and right) in Figure 5. Each average attention map is computed by averaging the attention map of all instances of a certain class. We can observe that: (1) The average attention maps from different classes vary significantly, showing that the diverse semantics of different classes can be reflected by our attention-based integration transformer. (2) When the classes of two episodes overlap (e.g., ‘trifle’ and ‘dalmatian’), the average attention maps of an overlapped class from two episodes are similar, illustrating that our attention-based integration transformer can well capture the semantics of classes across episodes. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d5c83307373ad302b67daff2626249571668bdb71167a2ce318b0ac9aa79845d.jpg", + "image_caption": [ + "Figure 5: Feature and attention visualization of two test episodes (left and right). Both figures present 5-way 5-shot results on miniImageNet, using Conv4-64 as the feature extractor. " + ], + "image_footnote": [], + "bbox": [ + 191, + 108, + 808, + 265 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.5 COMPARISON WITH SIMPLE BASELINE FOR SSL+FSL", + "bbox": [ + 174, + 335, + 593, + 351 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We compare our IEPT with a simple baseline that trains the model with $\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }$ and then makes inference by just averaging the outputs of different extended episodes. The results are shown in Table 6. It can be observed that the performance of our IEPT is much more effective than that of simple integration, due to the extra use of $\\mathcal { L } _ { i n t e g } + \\mathcal { L } _ { e p i s }$ for FSL. ", + "bbox": [ + 174, + 363, + 825, + 421 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/e2110b7904a72361e913dc8b2042d85b99fd274b6cad602dc6dcbf6ef7b2dea2.jpg", + "table_caption": [ + "Table 6: Comparison with the simple baseline that trains the model with $\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }$ and then makes inference by just averaging the outputs of different extended episodes. " + ], + "table_footnote": [], + "table_body": "
miniImageNettieredImageNet
MethodBackbone1-shot5-shot1-shot5-shot
Laux+LinstConv4-6453.25 ± 0.4671.50 ± 0.4255.06 ± 0.4472.87 ± 0.42
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "bbox": [ + 181, + 479, + 815, + 534 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.6 DIFFERENT ALTERNATIVES OF SELF-SUPERVISED LOSSES ", + "text_level": 1, + "bbox": [ + 174, + 568, + 622, + 582 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In Table 7, we provide further ablation study regarding different alternatives of $\\mathcal { L } _ { e p i s }$ and $\\mathcal { L } _ { i n s t }$ . For the episode-level self-supervised loss $\\mathcal { L } _ { e p i s }$ , we compare our implementation (using the KL loss between each distribution and the mean distribution) with that using a pairwise KL loss (i.e., the KL loss between each pair of distributions). For the instance-level self-supervised loss $\\mathcal { L } _ { i n s t }$ , we compare our implementation (using the rotation prediction loss) with the recent self-supervised learning technique (Chen et al., 2020). We observe that our implementation achieves slight performance improvements over those using the pairwise KL loss or the contrastive learning loss. ", + "bbox": [ + 174, + 594, + 825, + 694 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/2d0670d600723f1311e4dff16c14fd4b30f96753bca322cdab0c9ab09060b90a.jpg", + "table_caption": [ + "Table 7: Ablation study results regarding different alternatives of $\\mathcal { L } _ { e p i s }$ and $\\mathcal { L } _ { i n s t }$ " + ], + "table_footnote": [], + "table_body": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
IEPT (Linst-SimCLR)Conv4-6456.04 ± 0.4473.67± 0.4158.24± 0.4375.59± 0.41
IEPT (Lepis-Pairwise)Conv4-6455.95 ± 0.4773.72 ±0.4057.91 ± 0.4575.28±0.40
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "bbox": [ + 191, + 738, + 805, + 806 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.7 APPLICATION OF IEPT TO OPTIMIZATION-BASED METHOD MAML ", + "text_level": 1, + "bbox": [ + 176, + 838, + 687, + 854 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In Table 8, we show the results obtained by applying our IEPT to the optimization-based model MAML (Finn et al., 2017). We use Conv4-64 as the feature extractor. We can see that our IEPT brings $0 . 7 \\% - 2 . 1 \\%$ improvements to MAML. This further shows the flexibility (as well as effectiveness) of our IEPT for FSL. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/84b9c3b35da25fb4560691fc1312be2bd5050c34561bd624d14f308488569d31.jpg", + "table_caption": [ + "Table 8: Comparative results by applying IEPT to the optimization-based method MAML. " + ], + "table_footnote": [], + "table_body": "
BackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
Method MAMLConv4-6448.70±1.8463.10±0.9251.67 ± 1.8170.30±0.80
MAML+IEPTConv4-6449.68 ± 0.5065.22 ± 0.4852.85 ± 0.5271.04 ± 0.49
", + "bbox": [ + 199, + 127, + 797, + 181 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/19a2a9e135d017944a544b6d105ac0bab71b8611d3f98f7456e4da154a02510e.jpg", + "image_caption": [ + "Figure 6: Visualization of our hyper-parameter analysis under 5-way 1-shot (left) and 5-shot (right) on miniImageNet. Conv4-64 is used as the feature extractor. " + ], + "image_footnote": [], + "bbox": [ + 186, + 200, + 799, + 342 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "A.8 HYPER-PARAMETER SENSITIVITY TEST ", + "text_level": 1, + "bbox": [ + 174, + 397, + 493, + 411 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We select the hyper-parameters $w _ { 1 }$ , $w _ { 2 }$ and $w _ { 3 }$ from the candidate set $\\{ 0 . 1 , 0 . 5 , 1 . 0 , 5 . 0 , 1 0 . 0 \\}$ and show the hyper-parameter analysis results in Figure 6. We find that the performance of our IEPT is relatively stable. Concretely, the performance of our IEPT is not sensitive to $w _ { 1 }$ and $w _ { 2 }$ with proper values, but too large $w _ { 1 }$ (i.e. $w _ { 1 } = 1 0 . 0 $ ) tends to cause obvious degradation, perhaps because the FSL task is biased by the rotation prediction loss. ", + "bbox": [ + 174, + 422, + 825, + 493 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_middle.json b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..1887dad8bfcae35079757b1afa67bcd498b548df --- /dev/null +++ b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_middle.json @@ -0,0 +1,44315 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 508, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 508, + 97 + ], + "score": 1.0, + "content": "IEPT: INSTANCE-LEVEL AND EPISODE-LEVEL PRE-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 401, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 401, + 117 + ], + "score": 1.0, + "content": "TEXT TASKS FOR FEW-SHOT LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 468, + 179 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 308, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 308, + 147 + ], + "score": 1.0, + "content": "Manli Zhang, Jianhong Zhang & Zhiwu Lu ∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 457, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 457, + 159 + ], + "score": 1.0, + "content": "Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 468, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 468, + 170 + ], + "score": 1.0, + "content": "Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 168, + 352, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 352, + 181 + ], + "score": 1.0, + "content": "{manlizhang,jianhong,luzhiwu}@ruc.edu.cn", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 112, + 196, + 290, + 229 + ], + "lines": [ + { + "bbox": [ + 111, + 194, + 160, + 210 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 160, + 210 + ], + "score": 1.0, + "content": "Tao Xiang", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 207, + 290, + 219 + ], + "spans": [ + { + "bbox": [ + 112, + 207, + 290, + 219 + ], + "score": 1.0, + "content": "University of Surrey, Guildford, Surrey, UK", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 218, + 234, + 230 + ], + "spans": [ + { + "bbox": [ + 113, + 218, + 234, + 230 + ], + "score": 1.0, + "content": "t.xiang@surrey.ac.uk", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 326, + 196, + 498, + 229 + ], + "lines": [ + { + "bbox": [ + 324, + 194, + 386, + 211 + ], + "spans": [ + { + "bbox": [ + 324, + 194, + 386, + 211 + ], + "score": 1.0, + "content": "Mingyu Ding", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 205, + 499, + 221 + ], + "spans": [ + { + "bbox": [ + 325, + 205, + 499, + 221 + ], + "score": 1.0, + "content": "The University of Hong Kong, Hong Kong", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 218, + 430, + 230 + ], + "spans": [ + { + "bbox": [ + 325, + 218, + 430, + 230 + ], + "score": 1.0, + "content": "mingyuding@hku.hk", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 113, + 247, + 185, + 258 + ], + "lines": [ + { + "bbox": [ + 111, + 244, + 187, + 261 + ], + "spans": [ + { + "bbox": [ + 111, + 244, + 187, + 261 + ], + "score": 1.0, + "content": "Songfang Huang", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 113, + 258, + 293, + 279 + ], + "lines": [ + { + "bbox": [ + 112, + 257, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 112, + 257, + 294, + 269 + ], + "score": 1.0, + "content": "Alibaba DAMO Academy, Hangzhou, China", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 112, + 269, + 282, + 280 + ], + "spans": [ + { + "bbox": [ + 112, + 269, + 282, + 280 + ], + "score": 1.0, + "content": "songfang.hsf@alibaba-inc.com", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 278, + 308, + 333, + 320 + ], + "lines": [ + { + "bbox": [ + 276, + 308, + 335, + 321 + ], + "spans": [ + { + "bbox": [ + 276, + 308, + 335, + 321 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 143, + 333, + 469, + 530 + ], + "lines": [ + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "The need of collecting large quantities of labeled training data for each new task", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "score": 1.0, + "content": "has limited the usefulness of deep neural networks. Given data from a set of source", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 355, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 470, + 367 + ], + "score": 1.0, + "content": "tasks, this limitation can be overcome using two transfer learning approaches:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "score": 1.0, + "content": "few-shot learning (FSL) and self-supervised learning (SSL). The former aims to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "score": 1.0, + "content": "learn ‘how to learn’ by designing learning episodes using source tasks to simulate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "score": 1.0, + "content": "the challenge of solving the target new task with few labeled samples. In contrast,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "score": 1.0, + "content": "the latter exploits an annotation-free pretext task across all source tasks in order", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 411, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 411, + 469, + 423 + ], + "score": 1.0, + "content": "to learn generalizable feature representations. In this work, we propose a novel", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 421, + 470, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 433 + ], + "score": 1.0, + "content": "Instance-level and Episode-level Pretext Task (IEPT) framework that seamlessly in-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "score": 1.0, + "content": "tegrates SSL into FSL. Specifically, given an FSL episode, we first apply geometric", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 443, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 470, + 456 + ], + "score": 1.0, + "content": "transformations to each instance to generate extended episodes. At the instance-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 453, + 470, + 466 + ], + "spans": [ + { + "bbox": [ + 141, + 453, + 470, + 466 + ], + "score": 1.0, + "content": "level, transformation recognition is performed as per standard SSL. Importantly, at", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 465, + 469, + 477 + ], + "spans": [ + { + "bbox": [ + 141, + 465, + 469, + 477 + ], + "score": 1.0, + "content": "the episode-level, two SSL-FSL hybrid learning objectives are devised: (1) The", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 476, + 470, + 487 + ], + "spans": [ + { + "bbox": [ + 141, + 476, + 470, + 487 + ], + "score": 1.0, + "content": "consistency across the predictions of an FSL classifier from different extended", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 487, + 470, + 498 + ], + "spans": [ + { + "bbox": [ + 142, + 487, + 470, + 498 + ], + "score": 1.0, + "content": "episodes is maximized as an episode-level pretext task. (2) The features extracted", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 497, + 469, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 497, + 469, + 510 + ], + "score": 1.0, + "content": "from each instance across different episodes are integrated to construct a single", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 508, + 469, + 521 + ], + "spans": [ + { + "bbox": [ + 141, + 508, + 469, + 521 + ], + "score": 1.0, + "content": "FSL classifier for meta-learning. Extensive experiments show that our proposed", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 142, + 520, + 391, + 532 + ], + "spans": [ + { + "bbox": [ + 142, + 520, + 391, + 532 + ], + "score": 1.0, + "content": "model (i.e., FSL with IEPT) achieves the new state-of-the-art.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 206, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 208, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 208, + 567 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 588 + ], + "score": 1.0, + "content": "Deep convolutional neural networks (CNNs) (Krizhevsky et al., 2012; He et al., 2016b; Huang et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 601 + ], + "score": 1.0, + "content": "2017) have seen tremendous successes in a wide range of application fields, especially in visual", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "recognition. However, the powerful learning ability of CNNs depends on a large amount of manually", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "labeled training data. In practice, for many visual recognition tasks, sufficient manual annotation is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 619, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 634 + ], + "score": 1.0, + "content": "either too costly to collect or not feasible (e.g., for rare object classes). This has severely limited", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "score": 1.0, + "content": "the usefulness of CNNs for real-world application scenarios. Attempts have been made recently to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "score": 1.0, + "content": "mitigate such a limitation from two distinct perspectives, resulting in two popular research lines, both", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 651, + 507, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 507, + 666 + ], + "score": 1.0, + "content": "of which aim to transfer knowledge learned from the data of a set of source tasks to a new target one:", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 665, + 348, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 348, + 677 + ], + "score": 1.0, + "content": "few-shot learning (FSL) and self-supervised learning (SSL).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 681, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "FSL (Fei-Fei et al., 2006; Vinyals et al., 2016; Finn et al., 2017; Snell et al., 2017; Sung et al., 2018)", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "typically takes a ‘learning to learn’ or meta-learning paradigm. That is, it aims to learn an algorithm", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 702, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 505, + 716 + ], + "score": 1.0, + "content": "for learning from few labeled samples, which generalizes well across any tasks. To that end, it adopts", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 204, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 205, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 205, + 734 + ], + "score": 1.0, + "content": "∗Corresponding author.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 508, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 508, + 97 + ], + "score": 1.0, + "content": "IEPT: INSTANCE-LEVEL AND EPISODE-LEVEL PRE-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 401, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 401, + 117 + ], + "score": 1.0, + "content": "TEXT TASKS FOR FEW-SHOT LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 113, + 135, + 468, + 179 + ], + "lines": [ + { + "bbox": [ + 111, + 134, + 308, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 308, + 147 + ], + "score": 1.0, + "content": "Manli Zhang, Jianhong Zhang & Zhiwu Lu ∗", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 145, + 457, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 457, + 159 + ], + "score": 1.0, + "content": "Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 156, + 468, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 468, + 170 + ], + "score": 1.0, + "content": "Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 168, + 352, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 352, + 181 + ], + "score": 1.0, + "content": "{manlizhang,jianhong,luzhiwu}@ruc.edu.cn", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + } + ], + "index": 3.5, + "bbox_fs": [ + 111, + 134, + 468, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 196, + 290, + 229 + ], + "lines": [ + { + "bbox": [ + 111, + 194, + 160, + 210 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 160, + 210 + ], + "score": 1.0, + "content": "Tao Xiang", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 207, + 290, + 219 + ], + "spans": [ + { + "bbox": [ + 112, + 207, + 290, + 219 + ], + "score": 1.0, + "content": "University of Surrey, Guildford, Surrey, UK", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 218, + 234, + 230 + ], + "spans": [ + { + "bbox": [ + 113, + 218, + 234, + 230 + ], + "score": 1.0, + "content": "t.xiang@surrey.ac.uk", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 111, + 194, + 290, + 230 + ] + }, + { + "type": "list", + "bbox": [ + 326, + 196, + 498, + 229 + ], + "lines": [ + { + "bbox": [ + 324, + 194, + 386, + 211 + ], + "spans": [ + { + "bbox": [ + 324, + 194, + 386, + 211 + ], + "score": 1.0, + "content": "Mingyu Ding", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 325, + 205, + 499, + 221 + ], + "spans": [ + { + "bbox": [ + 325, + 205, + 499, + 221 + ], + "score": 1.0, + "content": "The University of Hong Kong, Hong Kong", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 325, + 218, + 430, + 230 + ], + "spans": [ + { + "bbox": [ + 325, + 218, + 430, + 230 + ], + "score": 1.0, + "content": "mingyuding@hku.hk", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + } + ], + "index": 10, + "bbox_fs": [ + 324, + 194, + 499, + 230 + ] + }, + { + "type": "title", + "bbox": [ + 113, + 247, + 185, + 258 + ], + "lines": [ + { + "bbox": [ + 111, + 244, + 187, + 261 + ], + "spans": [ + { + "bbox": [ + 111, + 244, + 187, + 261 + ], + "score": 1.0, + "content": "Songfang Huang", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 113, + 258, + 293, + 279 + ], + "lines": [ + { + "bbox": [ + 112, + 257, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 112, + 257, + 294, + 269 + ], + "score": 1.0, + "content": "Alibaba DAMO Academy, Hangzhou, China", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 112, + 269, + 282, + 280 + ], + "spans": [ + { + "bbox": [ + 112, + 269, + 282, + 280 + ], + "score": 1.0, + "content": "songfang.hsf@alibaba-inc.com", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 112, + 257, + 294, + 280 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 308, + 333, + 320 + ], + "lines": [ + { + "bbox": [ + 276, + 308, + 335, + 321 + ], + "spans": [ + { + "bbox": [ + 276, + 308, + 335, + 321 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 143, + 333, + 469, + 530 + ], + "lines": [ + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "The need of collecting large quantities of labeled training data for each new task", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "score": 1.0, + "content": "has limited the usefulness of deep neural networks. Given data from a set of source", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 355, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 470, + 367 + ], + "score": 1.0, + "content": "tasks, this limitation can be overcome using two transfer learning approaches:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "score": 1.0, + "content": "few-shot learning (FSL) and self-supervised learning (SSL). The former aims to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "score": 1.0, + "content": "learn ‘how to learn’ by designing learning episodes using source tasks to simulate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "score": 1.0, + "content": "the challenge of solving the target new task with few labeled samples. In contrast,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "score": 1.0, + "content": "the latter exploits an annotation-free pretext task across all source tasks in order", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 411, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 411, + 469, + 423 + ], + "score": 1.0, + "content": "to learn generalizable feature representations. In this work, we propose a novel", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 421, + 470, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 433 + ], + "score": 1.0, + "content": "Instance-level and Episode-level Pretext Task (IEPT) framework that seamlessly in-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "score": 1.0, + "content": "tegrates SSL into FSL. Specifically, given an FSL episode, we first apply geometric", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 443, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 470, + 456 + ], + "score": 1.0, + "content": "transformations to each instance to generate extended episodes. At the instance-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 453, + 470, + 466 + ], + "spans": [ + { + "bbox": [ + 141, + 453, + 470, + 466 + ], + "score": 1.0, + "content": "level, transformation recognition is performed as per standard SSL. Importantly, at", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 465, + 469, + 477 + ], + "spans": [ + { + "bbox": [ + 141, + 465, + 469, + 477 + ], + "score": 1.0, + "content": "the episode-level, two SSL-FSL hybrid learning objectives are devised: (1) The", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 476, + 470, + 487 + ], + "spans": [ + { + "bbox": [ + 141, + 476, + 470, + 487 + ], + "score": 1.0, + "content": "consistency across the predictions of an FSL classifier from different extended", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 487, + 470, + 498 + ], + "spans": [ + { + "bbox": [ + 142, + 487, + 470, + 498 + ], + "score": 1.0, + "content": "episodes is maximized as an episode-level pretext task. (2) The features extracted", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 497, + 469, + 510 + ], + "spans": [ + { + "bbox": [ + 141, + 497, + 469, + 510 + ], + "score": 1.0, + "content": "from each instance across different episodes are integrated to construct a single", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 508, + 469, + 521 + ], + "spans": [ + { + "bbox": [ + 141, + 508, + 469, + 521 + ], + "score": 1.0, + "content": "FSL classifier for meta-learning. Extensive experiments show that our proposed", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 142, + 520, + 391, + 532 + ], + "spans": [ + { + "bbox": [ + 142, + 520, + 391, + 532 + ], + "score": 1.0, + "content": "model (i.e., FSL with IEPT) achieves the new state-of-the-art.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 24.5, + "bbox_fs": [ + 141, + 333, + 470, + 532 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 206, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 208, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 208, + 567 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 588 + ], + "score": 1.0, + "content": "Deep convolutional neural networks (CNNs) (Krizhevsky et al., 2012; He et al., 2016b; Huang et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 601 + ], + "score": 1.0, + "content": "2017) have seen tremendous successes in a wide range of application fields, especially in visual", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "recognition. However, the powerful learning ability of CNNs depends on a large amount of manually", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "labeled training data. In practice, for many visual recognition tasks, sufficient manual annotation is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 619, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 634 + ], + "score": 1.0, + "content": "either too costly to collect or not feasible (e.g., for rare object classes). This has severely limited", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "score": 1.0, + "content": "the usefulness of CNNs for real-world application scenarios. Attempts have been made recently to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 654 + ], + "score": 1.0, + "content": "mitigate such a limitation from two distinct perspectives, resulting in two popular research lines, both", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 651, + 507, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 507, + 666 + ], + "score": 1.0, + "content": "of which aim to transfer knowledge learned from the data of a set of source tasks to a new target one:", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 665, + 348, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 348, + 677 + ], + "score": 1.0, + "content": "few-shot learning (FSL) and self-supervised learning (SSL).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 577, + 507, + 677 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 681, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "FSL (Fei-Fei et al., 2006; Vinyals et al., 2016; Finn et al., 2017; Snell et al., 2017; Sung et al., 2018)", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "typically takes a ‘learning to learn’ or meta-learning paradigm. That is, it aims to learn an algorithm", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 702, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 505, + 716 + ], + "score": 1.0, + "content": "for learning from few labeled samples, which generalizes well across any tasks. To that end, it adopts", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 320, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 330 + ], + "score": 1.0, + "content": "an episodic training strategy – the source tasks are arranged into learning episodes, each of which", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 330, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 142, + 342 + ], + "score": 1.0, + "content": "contains", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 142, + 332, + 150, + 340 + ], + "score": 0.7, + "content": "n", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 150, + 330, + 198, + 342 + ], + "score": 1.0, + "content": "classes and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 198, + 331, + 205, + 340 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 205, + 330, + 506, + 342 + ], + "score": 1.0, + "content": "labeled samples per class to simulate the setting for the target task. Part of", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "the CNN model (e.g., feature extraction subnet, classification layers, or parameter initialization) is", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 352, + 315, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 315, + 364 + ], + "score": 1.0, + "content": "then meta-learned for rapid adaptation to new tasks.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 680, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 88, + 498, + 181 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 88, + 498, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 88, + 498, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 88, + 498, + 181 + ], + "score": 0.96, + "type": "image", + "image_path": "ac5ad2ee0448bf35ac7fc284a145ae6ff0d67b950b288048f1c82a5d8bf0456d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 88, + 498, + 119.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 119.0, + 498, + 150.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 150.0, + 498, + 181.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 187, + 505, + 308 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "Figure 1: Schematic of our approach to FSL. Given a training episode, we apply 2D rotations", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "score": 1.0, + "content": "by 0, 90, 180, and 270 degrees to each instance to generate four extended episodes. After going", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "through a feature extraction CNN, four losses over three branches are designed: (1) In the top branch,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 434, + 234 + ], + "score": 1.0, + "content": "we employ a self-supervised rotation classifier with the instance-level SSL loss", + "type": "text" + }, + { + "bbox": [ + 435, + 220, + 457, + 231 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 219, + 506, + 234 + ], + "score": 1.0, + "content": ". (2) In the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "middle branch, an FSL classifier is exploited to predict the FSL classification probabilities for each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "episode. We maximize the classification consistency among the extended episodes by forcing the four", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 293, + 266 + ], + "score": 1.0, + "content": "probability distributions to be consistent using", + "type": "text" + }, + { + "bbox": [ + 293, + 253, + 316, + 265 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 253, + 454, + 266 + ], + "score": 1.0, + "content": ". The average supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 454, + 253, + 476, + 264 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "is also", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "computed. (3) In the bottom branch, we utilize an integration transformer module to fuse the features", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "extracted from each instance with different rotation transformations; they are then used to compute", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 197, + 299 + ], + "score": 1.0, + "content": "an integrated FSL loss", + "type": "text" + }, + { + "bbox": [ + 197, + 286, + 224, + 298 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 286, + 321, + 299 + ], + "score": 1.0, + "content": ". Among the four losses,", + "type": "text" + }, + { + "bbox": [ + 322, + 286, + 345, + 297 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 286, + 362, + 299 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 363, + 286, + 386, + 298 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 286, + 507, + 299 + ], + "score": 1.0, + "content": "are the self-supervised losses,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 295, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 123, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 297, + 145, + 308 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 295, + 164, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 297, + 191, + 309 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 295, + 296, + 311 + ], + "score": 1.0, + "content": "are the supervised losses.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 319, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 330 + ], + "score": 1.0, + "content": "an episodic training strategy – the source tasks are arranged into learning episodes, each of which", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 330, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 142, + 342 + ], + "score": 1.0, + "content": "contains", + "type": "text" + }, + { + "bbox": [ + 142, + 332, + 150, + 340 + ], + "score": 0.7, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 330, + 198, + 342 + ], + "score": 1.0, + "content": "classes and", + "type": "text" + }, + { + "bbox": [ + 198, + 331, + 205, + 340 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 330, + 506, + 342 + ], + "score": 1.0, + "content": "labeled samples per class to simulate the setting for the target task. Part of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "the CNN model (e.g., feature extraction subnet, classification layers, or parameter initialization) is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 352, + 315, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 315, + 364 + ], + "score": 1.0, + "content": "then meta-learned for rapid adaptation to new tasks.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "In contrast, SSL (Doersch et al., 2015; Noroozi & Favaro, 2016; Iizuka et al., 2016; Doersch &", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "Zisserman, 2017; Noroozi et al., 2018) does not require the source data to be annotated. Instead, it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 391, + 504, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 504, + 403 + ], + "score": 1.0, + "content": "exploits an annotation-free pretext task on the source task data in the hope that a task-generalizable", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "feature representation can be learned from the source tasks for easy adoption or adaptation in a target", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "task. Such a pretext task gets its self-supervised signal at the per-instance level. Examples include", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "rotation and context prediction (Gidaris et al., 2018; Doersch et al., 2015), jigsaw solving (Noroozi &", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "Favaro, 2016), and colorization (Iizuka et al., 2016; Larsson et al., 2016). Since these pretext tasks", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 445, + 428, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 428, + 459 + ], + "score": 1.0, + "content": "are class-agnostic, solving them leads to the learning of transferable knowledge.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 476 + ], + "score": 1.0, + "content": "Since both FSL and SSL aim to reduce the need of collecting a large amount of labeled training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "score": 1.0, + "content": "data for a target task by transferring knowledge from a set of source tasks, it is natural to consider", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "combining them in a single framework. Indeed, two recent works (Gidaris et al., 2019; Su et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 508 + ], + "score": 1.0, + "content": "2020) proposed to integrate SSL into FSL by adding an auxiliary SSL pretext task in an FSL model.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "It showed that the SSL learning objective is complementary to that of FSL and combining them", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "leads to improved FSL performance. However, in (Gidaris et al., 2019; Su et al., 2020), SSL is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "combined with FSL in a superficial way: it is only taken as a separate auxiliary task for each single", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "training instance and has no effect on the episodic training pipeline of the FSL model. Importantly,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 507, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 507, + 563 + ], + "score": 1.0, + "content": "by ignoring the class labels of samples, the instance-level SSL learning objective is weak on its own.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "Since meta-learning across episodes is the essence of most contemporary FSL models, we argue", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "that adding instance-level SSL pretext tasks alone fails to exploit fully the complementarity of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 583, + 438, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 438, + 595 + ], + "score": 1.0, + "content": "aforementioned FSL and SSL, for which a closer and deeper integration is needed.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "To that end, in this paper we propose a novel Instance-level and Episode-level Pretext Task (IEPT)", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "framework for few-shot recognition. Apart from adding an instance-level pretext SSL task as in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 636 + ], + "score": 1.0, + "content": "(Gidaris et al., 2019; Su et al., 2020), we introduce two episode-level SSL-FSL hybrid learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "objectives for seamless SSL-FSL integration. Concretely, as illustrated in Figure 1, our full model", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "has three additional learning objectives (besides the standard FSL one): (1) Different rotation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "transformations are applied to each original few-shot episode to generate a set of extended episodes,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "where each image has a rotation label for the instance-level pretext task (i.e., to predict the rotation", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "label). (2) The consistency across the predictions of an FSL classifier from different extended episodes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "is maximized as an episode-level pretext task. For each training image, the rotation transformation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "does not change its semantic content and hence its class label; the FSL classifier predictions across", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "different extended episodes thus should be consistent, hence the consistency regularization objective.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "(3) The correlation of features across instances from these extended episodes is modeled by a", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 88, + 498, + 181 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 88, + 498, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 88, + 498, + 181 + ], + "spans": [ + { + "bbox": [ + 111, + 88, + 498, + 181 + ], + "score": 0.96, + "type": "image", + "image_path": "ac5ad2ee0448bf35ac7fc284a145ae6ff0d67b950b288048f1c82a5d8bf0456d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 88, + 498, + 119.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 119.0, + 498, + 150.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 150.0, + 498, + 181.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 187, + 505, + 308 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 200 + ], + "score": 1.0, + "content": "Figure 1: Schematic of our approach to FSL. Given a training episode, we apply 2D rotations", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "score": 1.0, + "content": "by 0, 90, 180, and 270 degrees to each instance to generate four extended episodes. After going", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "through a feature extraction CNN, four losses over three branches are designed: (1) In the top branch,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 434, + 234 + ], + "score": 1.0, + "content": "we employ a self-supervised rotation classifier with the instance-level SSL loss", + "type": "text" + }, + { + "bbox": [ + 435, + 220, + 457, + 231 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 219, + 506, + 234 + ], + "score": 1.0, + "content": ". (2) In the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "middle branch, an FSL classifier is exploited to predict the FSL classification probabilities for each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 506, + 254 + ], + "score": 1.0, + "content": "episode. We maximize the classification consistency among the extended episodes by forcing the four", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 293, + 266 + ], + "score": 1.0, + "content": "probability distributions to be consistent using", + "type": "text" + }, + { + "bbox": [ + 293, + 253, + 316, + 265 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 253, + 454, + 266 + ], + "score": 1.0, + "content": ". The average supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 454, + 253, + 476, + 264 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "is also", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "computed. (3) In the bottom branch, we utilize an integration transformer module to fuse the features", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "extracted from each instance with different rotation transformations; they are then used to compute", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 197, + 299 + ], + "score": 1.0, + "content": "an integrated FSL loss", + "type": "text" + }, + { + "bbox": [ + 197, + 286, + 224, + 298 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 286, + 321, + 299 + ], + "score": 1.0, + "content": ". Among the four losses,", + "type": "text" + }, + { + "bbox": [ + 322, + 286, + 345, + 297 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 286, + 362, + 299 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 363, + 286, + 386, + 298 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 286, + 507, + 299 + ], + "score": 1.0, + "content": "are the self-supervised losses,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 295, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 123, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 297, + 145, + 308 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 295, + 164, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 297, + 191, + 309 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 295, + 296, + 311 + ], + "score": 1.0, + "content": "are the supervised losses.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 319, + 505, + 363 + ], + "lines": [], + "index": 15.5, + "bbox_fs": [ + 105, + 320, + 506, + 364 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "In contrast, SSL (Doersch et al., 2015; Noroozi & Favaro, 2016; Iizuka et al., 2016; Doersch &", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "Zisserman, 2017; Noroozi et al., 2018) does not require the source data to be annotated. Instead, it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 391, + 504, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 504, + 403 + ], + "score": 1.0, + "content": "exploits an annotation-free pretext task on the source task data in the hope that a task-generalizable", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "feature representation can be learned from the source tasks for easy adoption or adaptation in a target", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "task. Such a pretext task gets its self-supervised signal at the per-instance level. Examples include", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "rotation and context prediction (Gidaris et al., 2018; Doersch et al., 2015), jigsaw solving (Noroozi &", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "Favaro, 2016), and colorization (Iizuka et al., 2016; Larsson et al., 2016). Since these pretext tasks", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 445, + 428, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 428, + 459 + ], + "score": 1.0, + "content": "are class-agnostic, solving them leads to the learning of transferable knowledge.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 369, + 506, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 476 + ], + "score": 1.0, + "content": "Since both FSL and SSL aim to reduce the need of collecting a large amount of labeled training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 486 + ], + "score": 1.0, + "content": "data for a target task by transferring knowledge from a set of source tasks, it is natural to consider", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "combining them in a single framework. Indeed, two recent works (Gidaris et al., 2019; Su et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 508 + ], + "score": 1.0, + "content": "2020) proposed to integrate SSL into FSL by adding an auxiliary SSL pretext task in an FSL model.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "It showed that the SSL learning objective is complementary to that of FSL and combining them", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "leads to improved FSL performance. However, in (Gidaris et al., 2019; Su et al., 2020), SSL is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "combined with FSL in a superficial way: it is only taken as a separate auxiliary task for each single", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "training instance and has no effect on the episodic training pipeline of the FSL model. Importantly,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 507, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 507, + 563 + ], + "score": 1.0, + "content": "by ignoring the class labels of samples, the instance-level SSL learning objective is weak on its own.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "Since meta-learning across episodes is the essence of most contemporary FSL models, we argue", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "that adding instance-level SSL pretext tasks alone fails to exploit fully the complementarity of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 583, + 438, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 438, + 595 + ], + "score": 1.0, + "content": "aforementioned FSL and SSL, for which a closer and deeper integration is needed.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 461, + 507, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "To that end, in this paper we propose a novel Instance-level and Episode-level Pretext Task (IEPT)", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "framework for few-shot recognition. Apart from adding an instance-level pretext SSL task as in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 636 + ], + "score": 1.0, + "content": "(Gidaris et al., 2019; Su et al., 2020), we introduce two episode-level SSL-FSL hybrid learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "objectives for seamless SSL-FSL integration. Concretely, as illustrated in Figure 1, our full model", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "has three additional learning objectives (besides the standard FSL one): (1) Different rotation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "transformations are applied to each original few-shot episode to generate a set of extended episodes,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "where each image has a rotation label for the instance-level pretext task (i.e., to predict the rotation", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "label). (2) The consistency across the predictions of an FSL classifier from different extended episodes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "is maximized as an episode-level pretext task. For each training image, the rotation transformation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "does not change its semantic content and hence its class label; the FSL classifier predictions across", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "different extended episodes thus should be consistent, hence the consistency regularization objective.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "(3) The correlation of features across instances from these extended episodes is modeled by a", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "transformer-based attention module, optimizing the fusion of the features of each instance/image and", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "its various rotation-transformed versions mainly for task adaptation during meta-testing. Importantly,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "with these three new learning objectives introduced in IEPT, any meta-learning based FSL model can", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 389, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 389, + 128 + ], + "score": 1.0, + "content": "now benefit more from SSL by fully exploiting their complementarity.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 599, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "transformer-based attention module, optimizing the fusion of the features of each instance/image and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "its various rotation-transformed versions mainly for task adaptation during meta-testing. Importantly,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "with these three new learning objectives introduced in IEPT, any meta-learning based FSL model can", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 389, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 389, + 128 + ], + "score": 1.0, + "content": "now benefit more from SSL by fully exploiting their complementarity.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "Our main contributions are: (1) For the first time, we propose both instance-level and episode-level", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "pretext tasks (IEPT) for integrating SSL into FSL. The episode-level pretext task enables episodic", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "training of SSL and hence closer integration of SSL with FSL. (2) In addition to these pretext tasks,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "FSL further benefits from SSL by integrating features extracted from various rotation-transformed", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "versions of the original training instances. The optimal way of feature integration is learned by a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 507, + 201 + ], + "score": 1.0, + "content": "transformer-based attention module, which is mainly designed for task adaptation during meta-testing.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 469, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 469, + 211 + ], + "score": 1.0, + "content": "(3) Extensive experiments show that the proposed model achieves the new state-of-the-art.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 108, + 230, + 210, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 213, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 213, + 245 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 506, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "Few-Shot Learning. The recent FSL studies are dominated by meta-learning based methods. They", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "can be divided into three groups: (1) Metric-based methods (Vinyals et al., 2016; Snell et al., 2017;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "Sung et al., 2018; Allen et al., 2019; Xing et al., 2019; Li et al., 2019a;b; Wu et al., 2019; Ye", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 305 + ], + "score": 1.0, + "content": "et al., 2020; Afrasiyabi et al., 2020; Liu et al., 2020; Zhang et al., 2020) aim to learn the distance", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "metric between feature embeddings. The focus of these methods is often on meta-learning of a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "feature-extraction CNN, whilst the classifiers used are of simple form such as a nearest-neighbor", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "classifier. (2) Optimization-based methods (Finn et al., 2017; Ravi & Larochelle, 2017; Rusu et al.,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "2019; Lee et al., 2019) learn to optimize the model rapidly given a few labeled samples per class in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "the new task. (3) Model-based methods (Santoro et al., 2016; Munkhdalai & Yu, 2017; Mishra et al.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 507, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 507, + 371 + ], + "score": 1.0, + "content": "2018) focus on designing either specific model structures or parameters capable of rapid updating.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "Apart from these three groups of methods, other FSL methods have attempted feature hallucination", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "(Schwartz et al., 2018; Hariharan & Girshick, 2017; Gao et al., 2018; Wang et al., 2018; Zhang et al.,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "2019; Tsutsui et al., 2019) which generates additional samples from the given few shots for network", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "finetuning, and parameter predicting (Qiao et al., 2018; Qi et al., 2018; Gidaris & Komodakis, 2019;", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "2018) which learns to predict part of the parameters of a network given few samples of new classes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "score": 1.0, + "content": "for quick adaptation. In this work, we adopt the metric-based Prototypical Network (ProtoNet) (Snell", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "et al., 2017) as the basic FSL classifier for the main instantiation of our IEPT framework due to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "its simplicity and popularity. However, we show that any meta-learning based FSL method can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 457, + 318, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 318, + 470 + ], + "score": 1.0, + "content": "combined with our IEPT (see results in Figure 2(c)).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "Self-Supervised Learning. In SSL, it is assumed that the source task data is label-free and a pretext", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 483, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 498 + ], + "score": 1.0, + "content": "task is designed to provide self-supervision signals at the instance-level. Existing SSL approaches", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 495, + 507, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 507, + 508 + ], + "score": 1.0, + "content": "differ mainly in the pretext task design. These include predicting the rotation angle (Gidaris et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 504, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 520 + ], + "score": 1.0, + "content": "2018) and the context of image patch (Doersch et al., 2015; Nathan Mundhenk et al., 2018), jigsaw", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "solving (Noroozi & Favaro, 2016; Noroozi et al., 2018) (i.e. shuffling and then reordering image", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "patch), and performing images reversion (Iizuka et al., 2016; Pathak et al., 2016; Larsson et al.,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "2016). SSL has been shown to be beneficial to various down-steam tasks such as semantic object", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "matching (Novotny et al., 2018), object segmentation (Ji et al., 2019) and object detection (Doersch", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 561, + 433, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 433, + 574 + ], + "score": 1.0, + "content": "& Zisserman, 2017) by learning transferable feature presentations for these tasks.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "Integrating Self-Supervised Learning into Few-Shot Learning. To the best of our knowledge,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 589, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 506, + 601 + ], + "score": 1.0, + "content": "only two recent works (Gidaris et al., 2019; Su et al., 2020) have attempted combining SSL with FSL.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "However, the integration of SSL into FSL is often shallow: the original FSL training pipeline is intact;", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "in the meantime, an additional loss on each image w.r.t. a self-supervised signal like the rotation angle", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "score": 1.0, + "content": "or relative patch location is introduced. With pretext tasks solely at the instance level, combining", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 632, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 647 + ], + "score": 1.0, + "content": "the two approaches (i.e., SSL and FSL) can only be superficial without fully exploiting the episodic", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "training pipeline unique to FSL. Different from (Gidaris et al., 2019; Su et al., 2020), we introduce an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "episode-level pretext task to integrate SSL into the episodic training in FSL fully. Specifically, the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "consistency across the predictions of an FSL classifier from different extended episodes is maximized", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "to reflect the fact that various rotation transformations should not alter the class-label prediction.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "Moreover, features of each instance and its various rotation-transformed versions are now fused for", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "FSL classification, to integrate SSL with FSL for the supervised classification task. Our experimental", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "results show that thanks to the closer integration of SSL and FSL, our IEPT clearly outperforms", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 721, + 309, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 309, + 732 + ], + "score": 1.0, + "content": "(Gidaris et al., 2019; Su et al., 2020) (see Table 1).", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 46.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 128 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "Our main contributions are: (1) For the first time, we propose both instance-level and episode-level", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "pretext tasks (IEPT) for integrating SSL into FSL. The episode-level pretext task enables episodic", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "training of SSL and hence closer integration of SSL with FSL. (2) In addition to these pretext tasks,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "FSL further benefits from SSL by integrating features extracted from various rotation-transformed", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "versions of the original training instances. The optimal way of feature integration is learned by a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 507, + 201 + ], + "score": 1.0, + "content": "transformer-based attention module, which is mainly designed for task adaptation during meta-testing.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 199, + 469, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 469, + 211 + ], + "score": 1.0, + "content": "(3) Extensive experiments show that the proposed model achieves the new state-of-the-art.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 132, + 507, + 211 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 230, + 210, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 213, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 213, + 245 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 506, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "Few-Shot Learning. The recent FSL studies are dominated by meta-learning based methods. They", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "can be divided into three groups: (1) Metric-based methods (Vinyals et al., 2016; Snell et al., 2017;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "Sung et al., 2018; Allen et al., 2019; Xing et al., 2019; Li et al., 2019a;b; Wu et al., 2019; Ye", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 305 + ], + "score": 1.0, + "content": "et al., 2020; Afrasiyabi et al., 2020; Liu et al., 2020; Zhang et al., 2020) aim to learn the distance", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "metric between feature embeddings. The focus of these methods is often on meta-learning of a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "feature-extraction CNN, whilst the classifiers used are of simple form such as a nearest-neighbor", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "classifier. (2) Optimization-based methods (Finn et al., 2017; Ravi & Larochelle, 2017; Rusu et al.,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "2019; Lee et al., 2019) learn to optimize the model rapidly given a few labeled samples per class in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "the new task. (3) Model-based methods (Santoro et al., 2016; Munkhdalai & Yu, 2017; Mishra et al.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 507, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 507, + 371 + ], + "score": 1.0, + "content": "2018) focus on designing either specific model structures or parameters capable of rapid updating.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "Apart from these three groups of methods, other FSL methods have attempted feature hallucination", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "(Schwartz et al., 2018; Hariharan & Girshick, 2017; Gao et al., 2018; Wang et al., 2018; Zhang et al.,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "2019; Tsutsui et al., 2019) which generates additional samples from the given few shots for network", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "finetuning, and parameter predicting (Qiao et al., 2018; Qi et al., 2018; Gidaris & Komodakis, 2019;", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 425 + ], + "score": 1.0, + "content": "2018) which learns to predict part of the parameters of a network given few samples of new classes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "score": 1.0, + "content": "for quick adaptation. In this work, we adopt the metric-based Prototypical Network (ProtoNet) (Snell", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "et al., 2017) as the basic FSL classifier for the main instantiation of our IEPT framework due to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 458 + ], + "score": 1.0, + "content": "its simplicity and popularity. However, we show that any meta-learning based FSL method can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 457, + 318, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 318, + 470 + ], + "score": 1.0, + "content": "combined with our IEPT (see results in Figure 2(c)).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 258, + 507, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "Self-Supervised Learning. In SSL, it is assumed that the source task data is label-free and a pretext", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 483, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 498 + ], + "score": 1.0, + "content": "task is designed to provide self-supervision signals at the instance-level. Existing SSL approaches", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 495, + 507, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 507, + 508 + ], + "score": 1.0, + "content": "differ mainly in the pretext task design. These include predicting the rotation angle (Gidaris et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 504, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 520 + ], + "score": 1.0, + "content": "2018) and the context of image patch (Doersch et al., 2015; Nathan Mundhenk et al., 2018), jigsaw", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "solving (Noroozi & Favaro, 2016; Noroozi et al., 2018) (i.e. shuffling and then reordering image", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "patch), and performing images reversion (Iizuka et al., 2016; Pathak et al., 2016; Larsson et al.,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "2016). SSL has been shown to be beneficial to various down-steam tasks such as semantic object", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "matching (Novotny et al., 2018), object segmentation (Ji et al., 2019) and object detection (Doersch", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 561, + 433, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 433, + 574 + ], + "score": 1.0, + "content": "& Zisserman, 2017) by learning transferable feature presentations for these tasks.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 474, + 507, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "Integrating Self-Supervised Learning into Few-Shot Learning. To the best of our knowledge,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 589, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 506, + 601 + ], + "score": 1.0, + "content": "only two recent works (Gidaris et al., 2019; Su et al., 2020) have attempted combining SSL with FSL.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "However, the integration of SSL into FSL is often shallow: the original FSL training pipeline is intact;", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "in the meantime, an additional loss on each image w.r.t. a self-supervised signal like the rotation angle", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "score": 1.0, + "content": "or relative patch location is introduced. With pretext tasks solely at the instance level, combining", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 632, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 647 + ], + "score": 1.0, + "content": "the two approaches (i.e., SSL and FSL) can only be superficial without fully exploiting the episodic", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "training pipeline unique to FSL. Different from (Gidaris et al., 2019; Su et al., 2020), we introduce an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "episode-level pretext task to integrate SSL into the episodic training in FSL fully. Specifically, the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "consistency across the predictions of an FSL classifier from different extended episodes is maximized", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "to reflect the fact that various rotation transformations should not alter the class-label prediction.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "Moreover, features of each instance and its various rotation-transformed versions are now fused for", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "FSL classification, to integrate SSL with FSL for the supervised classification task. Our experimental", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "results show that thanks to the closer integration of SSL and FSL, our IEPT clearly outperforms", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 721, + 309, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 309, + 732 + ], + "score": 1.0, + "content": "(Gidaris et al., 2019; Su et al., 2020) (see Table 1).", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 577, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 208, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 211, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 211, + 97 + ], + "score": 1.0, + "content": "3 METHODOLOGY", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 99, + 192, + 110 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 193, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 193, + 111 + ], + "score": 1.0, + "content": "3.1 PRELIMINARY", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 118, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 222, + 132 + ], + "score": 1.0, + "content": "Problem Setting. Given an", + "type": "text" + }, + { + "bbox": [ + 222, + 122, + 229, + 130 + ], + "score": 0.69, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 119, + 251, + 132 + ], + "score": 1.0, + "content": "-way", + "type": "text" + }, + { + "bbox": [ + 251, + 120, + 258, + 130 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 119, + 415, + 132 + ], + "score": 1.0, + "content": "-shot FSL task sampled from a test set", + "type": "text" + }, + { + "bbox": [ + 416, + 120, + 428, + 131 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 119, + 506, + 132 + ], + "score": 1.0, + "content": ", to imitate the test", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 394, + 143 + ], + "score": 1.0, + "content": "setting, an FSL model is typically trained in an episodic way. That is,", + "type": "text" + }, + { + "bbox": [ + 395, + 133, + 402, + 140 + ], + "score": 0.78, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 131, + 424, + 143 + ], + "score": 1.0, + "content": "-way", + "type": "text" + }, + { + "bbox": [ + 424, + 131, + 430, + 140 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "-shot episodes are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 141, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 264, + 155 + ], + "score": 1.0, + "content": "randomly sampled from a training set", + "type": "text" + }, + { + "bbox": [ + 264, + 142, + 277, + 153 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 141, + 407, + 155 + ], + "score": 1.0, + "content": ", where the class label space of", + "type": "text" + }, + { + "bbox": [ + 407, + 142, + 420, + 153 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 141, + 506, + 155 + ], + "score": 1.0, + "content": "has no overlap with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 136, + 165 + ], + "score": 1.0, + "content": "that of", + "type": "text" + }, + { + "bbox": [ + 137, + 153, + 149, + 163 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 151, + 213, + 165 + ], + "score": 1.0, + "content": ". Each episode", + "type": "text" + }, + { + "bbox": [ + 213, + 153, + 226, + 163 + ], + "score": 0.88, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 151, + 321, + 165 + ], + "score": 1.0, + "content": "contains a support set", + "type": "text" + }, + { + "bbox": [ + 321, + 153, + 333, + 163 + ], + "score": 0.88, + "content": "S _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 151, + 402, + 165 + ], + "score": 1.0, + "content": "and a query set", + "type": "text" + }, + { + "bbox": [ + 402, + 153, + 415, + 164 + ], + "score": 0.88, + "content": "\\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 151, + 506, + 165 + ], + "score": 1.0, + "content": ". Concretely, we first", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 162, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 206, + 177 + ], + "score": 1.0, + "content": "randomly sample a set of", + "type": "text" + }, + { + "bbox": [ + 207, + 165, + 214, + 173 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 162, + 244, + 177 + ], + "score": 1.0, + "content": "classes", + "type": "text" + }, + { + "bbox": [ + 244, + 164, + 255, + 174 + ], + "score": 0.88, + "content": "\\mathcal { C } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 162, + 410, + 177 + ], + "score": 1.0, + "content": "from the training set, and then generate", + "type": "text" + }, + { + "bbox": [ + 411, + 164, + 422, + 174 + ], + "score": 0.89, + "content": "S _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 162, + 440, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 440, + 163, + 453, + 174 + ], + "score": 0.89, + "content": "\\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 162, + 506, + 177 + ], + "score": 1.0, + "content": "by sampling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 113, + 184 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 174, + 203, + 187 + ], + "score": 1.0, + "content": "support samples and", + "type": "text" + }, + { + "bbox": [ + 203, + 176, + 210, + 186 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 174, + 353, + 187 + ], + "score": 1.0, + "content": "query samples from each class in", + "type": "text" + }, + { + "bbox": [ + 354, + 174, + 364, + 185 + ], + "score": 0.86, + "content": "\\mathcal { C } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 174, + 506, + 187 + ], + "score": 1.0, + "content": ", respectively. Formally, we have", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 279, + 197 + ], + "score": 0.91, + "content": "\\mathcal S _ { e } \\stackrel { - } { = } \\{ ( x _ { i } , y _ { i } ) \\bar { | } y _ { i } \\in \\mathcal C _ { e } , i \\stackrel { - } { = } 1 , . . . , n \\stackrel { - } { \\times } k \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 183, + 298, + 199 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 185, + 473, + 197 + ], + "score": 0.92, + "content": "\\mathcal { Q } _ { e } = \\{ ( x _ { i } , y _ { i } ) | y _ { i } \\in \\mathcal { C } _ { e } , i = 1 , . . . , n \\times q \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 183, + 506, + 199 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 195, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 162, + 208 + ], + "score": 0.91, + "content": "S _ { e } \\cap \\mathcal { Q } _ { e } = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 195, + 275, + 210 + ], + "score": 1.0, + "content": ". For simplicity, we denote", + "type": "text" + }, + { + "bbox": [ + 275, + 197, + 324, + 208 + ], + "score": 0.91, + "content": "l _ { k } = n \\times k", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 195, + 343, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 343, + 197, + 390, + 208 + ], + "score": 0.91, + "content": "l _ { q } = n \\times q", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 195, + 507, + 210 + ], + "score": 1.0, + "content": ". In the meta-training stage,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 206, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 506, + 220 + ], + "score": 1.0, + "content": "the training process has an inner and an outer loop in each episode: in the inner loop, the model is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 164, + 231 + ], + "score": 1.0, + "content": "updated using", + "type": "text" + }, + { + "bbox": [ + 165, + 218, + 176, + 230 + ], + "score": 0.88, + "content": "S _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 218, + 382, + 231 + ], + "score": 1.0, + "content": "; its performance is then evaluated on the query set", + "type": "text" + }, + { + "bbox": [ + 382, + 218, + 396, + 230 + ], + "score": 0.89, + "content": "\\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "in the outer loop to update", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 230, + 365, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 365, + 241 + ], + "score": 1.0, + "content": "the model parameters or algorithm that one wants to meta-learn.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 258 + ], + "score": 1.0, + "content": "Basic FSL Classifier. We employ ProtoNet (Snell et al., 2017) as the basic FSL model. This model", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "has a feature-extraction CNN and a simple non-parametric classifier. The parameter of the feature", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 269, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 280 + ], + "score": 1.0, + "content": "extractor is to be meta-learned. Concretely, in the inner loop of an episode, ProtoNet fixes the feature", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 420, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 420, + 291 + ], + "score": 1.0, + "content": "extractor and computes the mean feature embedding for each class as follows:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 294, + 382, + 326 + ], + "lines": [ + { + "bbox": [ + 228, + 294, + 382, + 326 + ], + "spans": [ + { + "bbox": [ + 228, + 294, + 382, + 326 + ], + "score": 0.94, + "content": "h _ { c } = \\frac { 1 } { k } \\cdot \\sum _ { ( x _ { i } , y _ { i } ) \\in S _ { e } } f _ { \\phi } ( x _ { i } ) \\cdot \\mathbb { I } ( y _ { i } = c ) ,", + "type": "interline_equation", + "image_path": "af53dfdfebf92b14e994742c9d44f9cec3ef0b37d7e37b409e069334abdc5140.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 294, + 382, + 310.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 228, + 310.0, + 382, + 326.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 331, + 504, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 157, + 344 + ], + "score": 1.0, + "content": "where class", + "type": "text" + }, + { + "bbox": [ + 157, + 331, + 187, + 343 + ], + "score": 0.69, + "content": "c \\in \\mathcal { C } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 331, + 192, + 344 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 192, + 331, + 203, + 344 + ], + "score": 0.72, + "content": "f _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 331, + 403, + 344 + ], + "score": 1.0, + "content": "is a feature extractor with learnable parameters", + "type": "text" + }, + { + "bbox": [ + 404, + 332, + 411, + 343 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 331, + 433, + 344 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 433, + 332, + 439, + 341 + ], + "score": 0.57, + "content": "\\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "is the indicator", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "score": 1.0, + "content": "function. By computing the distance between the feature embedding of each query sample and that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 352, + 492, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 360, + 366 + ], + "score": 1.0, + "content": "of the corresponding class, the loss function used to meta-learn", + "type": "text" + }, + { + "bbox": [ + 360, + 354, + 368, + 365 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 352, + 492, + 366 + ], + "score": 1.0, + "content": "in the outer loop is defined as:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 369, + 443, + 402 + ], + "lines": [ + { + "bbox": [ + 167, + 369, + 443, + 402 + ], + "spans": [ + { + "bbox": [ + 167, + 369, + 443, + 402 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { f s l } ( S _ { e } , \\mathcal { Q } _ { e } ) = \\frac { 1 } { | \\mathcal { Q } _ { e } | } \\sum _ { ( x _ { i } , y _ { i } ) \\in \\mathcal { Q } _ { e } } - \\log \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { y _ { i } } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ) ) } ,", + "type": "interline_equation", + "image_path": "f5b0a49ba9ad956ade834f7dac81d4842c7687c92752b1ebb7c1e41414e7ca0f.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 167, + 369, + 443, + 380.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 167, + 380.0, + 443, + 391.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 167, + 391.0, + 443, + 402.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 407, + 356, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 357, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 133, + 421 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 407, + 157, + 420 + ], + "score": 0.92, + "content": "d ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 405, + 306, + 421 + ], + "score": 1.0, + "content": "denotes a distance function (e.g., the", + "type": "text" + }, + { + "bbox": [ + 306, + 408, + 315, + 419 + ], + "score": 0.87, + "content": "l _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 405, + 357, + 421 + ], + "score": 1.0, + "content": "distance).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 425, + 240, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 240, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 240, + 438 + ], + "score": 1.0, + "content": "3.2 PRETEXT TASKS IN IEPT", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 445, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "The schematic of our IEPT is illustrated in Figure 1. We first define a set of 2D-rotation operators", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 219, + 468 + ], + "score": 0.91, + "content": "\\mathcal { G } = \\{ g _ { r } | r = 0 , . . . , R - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 456, + 252, + 470 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 252, + 458, + 262, + 468 + ], + "score": 0.85, + "content": "g _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 456, + 448, + 470 + ], + "score": 1.0, + "content": "means the operator of rotating the image by", + "type": "text" + }, + { + "bbox": [ + 448, + 457, + 470, + 467 + ], + "score": 0.87, + "content": "r ^ { * 9 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "degrees", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 124, + 480 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 469, + 133, + 478 + ], + "score": 0.84, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 467, + 267, + 480 + ], + "score": 1.0, + "content": "is the total number of rotations (", + "type": "text" + }, + { + "bbox": [ + 267, + 468, + 295, + 478 + ], + "score": 0.88, + "content": "R = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "in our implementation). Given an original episode", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 107, + 478, + 173, + 491 + ], + "score": 0.93, + "content": "E _ { e } = \\{ S _ { e } , \\mathcal { Q } _ { e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 478, + 453, + 491 + ], + "score": 1.0, + "content": "as described in Sec. 3.1, we utilize the 2D-rotation operators from", + "type": "text" + }, + { + "bbox": [ + 454, + 479, + 462, + 489 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "in turn to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 209, + 502 + ], + "score": 1.0, + "content": "transform each image in", + "type": "text" + }, + { + "bbox": [ + 209, + 490, + 222, + 501 + ], + "score": 0.87, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 490, + 322, + 502 + ], + "score": 1.0, + "content": ". This results in a set of", + "type": "text" + }, + { + "bbox": [ + 323, + 490, + 332, + 500 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "extended episodes (including the original", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 499, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 128, + 514 + ], + "score": 1.0, + "content": "one)", + "type": "text" + }, + { + "bbox": [ + 129, + 500, + 276, + 512 + ], + "score": 0.9, + "content": "E = \\{ \\{ \\mathcal { S } _ { e } ^ { r } , \\bar { \\mathcal { Q } } _ { e } ^ { r } \\} | r = 0 , . . . , R - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 499, + 312, + 514 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 312, + 500, + 486, + 513 + ], + "score": 0.92, + "content": "S _ { e } ^ { r } \\ = \\ \\{ ( x _ { i } , y _ { i } , \\bar { r } ) | y _ { i } \\in C _ { e } , i \\ = \\ \\bar { 1 } , . . . , l _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 499, + 506, + 514 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 107, + 513, + 272, + 523 + ], + "score": 0.86, + "content": "\\mathcal { Q } _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in C _ { e } , i = 1 , . . . , l _ { q } \\} ,", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 511, + 411, + 525 + ], + "score": 1.0, + "content": ". Now each episode is denoted as", + "type": "text" + }, + { + "bbox": [ + 411, + 512, + 505, + 524 + ], + "score": 0.89, + "content": "E _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 520, + 502, + 537 + ], + "spans": [ + { + "bbox": [ + 107, + 524, + 251, + 534 + ], + "score": 0.81, + "content": "C _ { e } , i = 1 , . . . , l _ { k } , l _ { k } + 1 , . . . , l _ { k } + l _ { q } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 520, + 317, + 537 + ], + "score": 1.0, + "content": ", where the first", + "type": "text" + }, + { + "bbox": [ + 318, + 523, + 327, + 533 + ], + "score": 0.87, + "content": "l _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 520, + 402, + 537 + ], + "score": 1.0, + "content": "samples are from", + "type": "text" + }, + { + "bbox": [ + 402, + 523, + 414, + 534 + ], + "score": 0.89, + "content": "S _ { e } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 520, + 489, + 537 + ], + "score": 1.0, + "content": "and the rest from", + "type": "text" + }, + { + "bbox": [ + 489, + 523, + 502, + 534 + ], + "score": 0.88, + "content": "\\mathcal { Q } _ { e } ^ { r }", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 532, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 145, + 548 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 146, + 534, + 185, + 546 + ], + "score": 0.91, + "content": "\\{ \\cal { S } _ { e } ^ { 0 } , \\cal { Q } _ { e } ^ { 0 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 532, + 275, + 548 + ], + "score": 1.0, + "content": "is the original episode", + "type": "text" + }, + { + "bbox": [ + 276, + 534, + 313, + 546 + ], + "score": 0.92, + "content": "\\{ \\boldsymbol { S } _ { e } , \\boldsymbol { \\mathcal { Q } } _ { e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 532, + 506, + 548 + ], + "score": 1.0, + "content": ". With the rotation transformations, each sample", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 545, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 148, + 558 + ], + "score": 0.91, + "content": "( x _ { i } , y _ { i } , r _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 545, + 159, + 559 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 160, + 546, + 169, + 556 + ], + "score": 0.81, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 545, + 248, + 559 + ], + "score": 1.0, + "content": "carries a class label", + "type": "text" + }, + { + "bbox": [ + 248, + 547, + 258, + 557 + ], + "score": 0.85, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 545, + 494, + 559 + ], + "score": 1.0, + "content": "for supervised learning (from the inherent class) and a label", + "type": "text" + }, + { + "bbox": [ + 495, + 547, + 504, + 556 + ], + "score": 0.82, + "content": "r _ { i }", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "from the rotation operator for self-supervised learning. After generating the set of extended episodes", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 107, + 568, + 116, + 577 + ], + "score": 0.79, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 567, + 206, + 580 + ], + "score": 1.0, + "content": ", the feature extractor", + "type": "text" + }, + { + "bbox": [ + 207, + 568, + 218, + 579 + ], + "score": 0.88, + "content": "f _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 567, + 323, + 580 + ], + "score": 1.0, + "content": "is applied to each image", + "type": "text" + }, + { + "bbox": [ + 323, + 569, + 334, + 578 + ], + "score": 0.85, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 567, + 347, + 580 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 347, + 568, + 356, + 577 + ], + "score": 0.8, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 567, + 505, + 580 + ], + "score": 1.0, + "content": ". On these episodes, we design two", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 578, + 435, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 435, + 591 + ], + "score": 1.0, + "content": "self-supervised pretext tasks, one at the instance-level and the other episode-level.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 506, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "Instance-Level Pretext Task. The instance-level task is to recognize different rotation transforma-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 378, + 619 + ], + "score": 1.0, + "content": "tions. The idea is that if the model to be meta-learned here (i.e.,", + "type": "text" + }, + { + "bbox": [ + 378, + 606, + 392, + 618 + ], + "score": 0.84, + "content": "f _ { \\phi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "can be used to distinguish", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "different transformations, it must understand the canonical poses of objects (e.g., animals have legs", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "touching the ground and trees have leaves on top), a vital part of class-agnostic and thus transferable", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 637, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 316, + 653 + ], + "score": 1.0, + "content": "knowledge. With the self-supervised rotation label", + "type": "text" + }, + { + "bbox": [ + 317, + 640, + 326, + 650 + ], + "score": 0.83, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 637, + 440, + 653 + ], + "score": 1.0, + "content": ", we consider the mapping:", + "type": "text" + }, + { + "bbox": [ + 440, + 640, + 504, + 651 + ], + "score": 0.89, + "content": "f _ { \\theta _ { r o t } } : x _ { i } \\mapsto r _ { i }", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 648, + 507, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 177, + 664 + ], + "score": 1.0, + "content": "for each instance", + "type": "text" + }, + { + "bbox": [ + 177, + 650, + 240, + 662 + ], + "score": 0.91, + "content": "( x _ { i } , y _ { i } , r _ { i } ) \\in E", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 648, + 270, + 664 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 271, + 650, + 290, + 662 + ], + "score": 0.92, + "content": "f _ { \\theta _ { r o t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 648, + 486, + 664 + ], + "score": 1.0, + "content": "is a rotation classifier with learnable parameters", + "type": "text" + }, + { + "bbox": [ + 486, + 650, + 502, + 661 + ], + "score": 0.89, + "content": "\\theta _ { r o t }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 648, + 507, + 664 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 660, + 460, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 189, + 674 + ], + "score": 1.0, + "content": "Given the input pair", + "type": "text" + }, + { + "bbox": [ + 189, + 662, + 218, + 673 + ], + "score": 0.92, + "content": "( x _ { i } , r _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 660, + 460, + 674 + ], + "score": 1.0, + "content": ", the total instance-level rotation loss is a cross-entropy loss:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 677, + 455, + 714 + ], + "lines": [ + { + "bbox": [ + 154, + 677, + 455, + 714 + ], + "spans": [ + { + "bbox": [ + 154, + 677, + 455, + 714 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { i n s t } = \\frac { 1 } { R ( l _ { k } + l _ { q } ) } \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { ( x _ { i } , y _ { i } , r _ { i } ) \\in E _ { e } ^ { r } } - \\log \\frac { \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r _ { i } } ) } { \\sum _ { r ^ { \\prime } = 0 } ^ { R - 1 } \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r ^ { \\prime } } ) } ,", + "type": "interline_equation", + "image_path": "efabe317b4928532f5cf8bbec84ba16e3a6148c3e31c52bd61720019bd8d7f06.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 154, + 677, + 455, + 689.3333333333334 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 154, + 689.3333333333334, + 455, + 701.6666666666667 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 154, + 701.6666666666667, + 455, + 714.0000000000001 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 720, + 492, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 494, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 133, + 734 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 720, + 219, + 733 + ], + "score": 0.93, + "content": "[ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] \\in \\mathbb { R } ^ { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 719, + 354, + 734 + ], + "score": 1.0, + "content": "is the rotation scoring vector and", + "type": "text" + }, + { + "bbox": [ + 355, + 720, + 368, + 732 + ], + "score": 0.9, + "content": "[ \\cdot ] _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 719, + 439, + 734 + ], + "score": 1.0, + "content": "means taking the", + "type": "text" + }, + { + "bbox": [ + 440, + 722, + 446, + 730 + ], + "score": 0.77, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 719, + 494, + 734 + ], + "score": 1.0, + "content": "-th element.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 208, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 211, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 211, + 97 + ], + "score": 1.0, + "content": "3 METHODOLOGY", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 99, + 192, + 110 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 193, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 193, + 111 + ], + "score": 1.0, + "content": "3.1 PRELIMINARY", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 118, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 119, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 222, + 132 + ], + "score": 1.0, + "content": "Problem Setting. Given an", + "type": "text" + }, + { + "bbox": [ + 222, + 122, + 229, + 130 + ], + "score": 0.69, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 119, + 251, + 132 + ], + "score": 1.0, + "content": "-way", + "type": "text" + }, + { + "bbox": [ + 251, + 120, + 258, + 130 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 119, + 415, + 132 + ], + "score": 1.0, + "content": "-shot FSL task sampled from a test set", + "type": "text" + }, + { + "bbox": [ + 416, + 120, + 428, + 131 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 119, + 506, + 132 + ], + "score": 1.0, + "content": ", to imitate the test", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 394, + 143 + ], + "score": 1.0, + "content": "setting, an FSL model is typically trained in an episodic way. That is,", + "type": "text" + }, + { + "bbox": [ + 395, + 133, + 402, + 140 + ], + "score": 0.78, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 131, + 424, + 143 + ], + "score": 1.0, + "content": "-way", + "type": "text" + }, + { + "bbox": [ + 424, + 131, + 430, + 140 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "-shot episodes are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 141, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 264, + 155 + ], + "score": 1.0, + "content": "randomly sampled from a training set", + "type": "text" + }, + { + "bbox": [ + 264, + 142, + 277, + 153 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 141, + 407, + 155 + ], + "score": 1.0, + "content": ", where the class label space of", + "type": "text" + }, + { + "bbox": [ + 407, + 142, + 420, + 153 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 141, + 506, + 155 + ], + "score": 1.0, + "content": "has no overlap with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 136, + 165 + ], + "score": 1.0, + "content": "that of", + "type": "text" + }, + { + "bbox": [ + 137, + 153, + 149, + 163 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 151, + 213, + 165 + ], + "score": 1.0, + "content": ". Each episode", + "type": "text" + }, + { + "bbox": [ + 213, + 153, + 226, + 163 + ], + "score": 0.88, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 151, + 321, + 165 + ], + "score": 1.0, + "content": "contains a support set", + "type": "text" + }, + { + "bbox": [ + 321, + 153, + 333, + 163 + ], + "score": 0.88, + "content": "S _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 151, + 402, + 165 + ], + "score": 1.0, + "content": "and a query set", + "type": "text" + }, + { + "bbox": [ + 402, + 153, + 415, + 164 + ], + "score": 0.88, + "content": "\\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 151, + 506, + 165 + ], + "score": 1.0, + "content": ". Concretely, we first", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 162, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 206, + 177 + ], + "score": 1.0, + "content": "randomly sample a set of", + "type": "text" + }, + { + "bbox": [ + 207, + 165, + 214, + 173 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 162, + 244, + 177 + ], + "score": 1.0, + "content": "classes", + "type": "text" + }, + { + "bbox": [ + 244, + 164, + 255, + 174 + ], + "score": 0.88, + "content": "\\mathcal { C } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 162, + 410, + 177 + ], + "score": 1.0, + "content": "from the training set, and then generate", + "type": "text" + }, + { + "bbox": [ + 411, + 164, + 422, + 174 + ], + "score": 0.89, + "content": "S _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 162, + 440, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 440, + 163, + 453, + 174 + ], + "score": 0.89, + "content": "\\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 162, + 506, + 177 + ], + "score": 1.0, + "content": "by sampling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 113, + 184 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 174, + 203, + 187 + ], + "score": 1.0, + "content": "support samples and", + "type": "text" + }, + { + "bbox": [ + 203, + 176, + 210, + 186 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 174, + 353, + 187 + ], + "score": 1.0, + "content": "query samples from each class in", + "type": "text" + }, + { + "bbox": [ + 354, + 174, + 364, + 185 + ], + "score": 0.86, + "content": "\\mathcal { C } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 174, + 506, + 187 + ], + "score": 1.0, + "content": ", respectively. Formally, we have", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 279, + 197 + ], + "score": 0.91, + "content": "\\mathcal S _ { e } \\stackrel { - } { = } \\{ ( x _ { i } , y _ { i } ) \\bar { | } y _ { i } \\in \\mathcal C _ { e } , i \\stackrel { - } { = } 1 , . . . , n \\stackrel { - } { \\times } k \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 183, + 298, + 199 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 185, + 473, + 197 + ], + "score": 0.92, + "content": "\\mathcal { Q } _ { e } = \\{ ( x _ { i } , y _ { i } ) | y _ { i } \\in \\mathcal { C } _ { e } , i = 1 , . . . , n \\times q \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 183, + 506, + 199 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 195, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 162, + 208 + ], + "score": 0.91, + "content": "S _ { e } \\cap \\mathcal { Q } _ { e } = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 195, + 275, + 210 + ], + "score": 1.0, + "content": ". For simplicity, we denote", + "type": "text" + }, + { + "bbox": [ + 275, + 197, + 324, + 208 + ], + "score": 0.91, + "content": "l _ { k } = n \\times k", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 195, + 343, + 210 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 343, + 197, + 390, + 208 + ], + "score": 0.91, + "content": "l _ { q } = n \\times q", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 195, + 507, + 210 + ], + "score": 1.0, + "content": ". In the meta-training stage,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 206, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 506, + 220 + ], + "score": 1.0, + "content": "the training process has an inner and an outer loop in each episode: in the inner loop, the model is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 164, + 231 + ], + "score": 1.0, + "content": "updated using", + "type": "text" + }, + { + "bbox": [ + 165, + 218, + 176, + 230 + ], + "score": 0.88, + "content": "S _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 218, + 382, + 231 + ], + "score": 1.0, + "content": "; its performance is then evaluated on the query set", + "type": "text" + }, + { + "bbox": [ + 382, + 218, + 396, + 230 + ], + "score": 0.89, + "content": "\\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "in the outer loop to update", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 230, + 365, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 365, + 241 + ], + "score": 1.0, + "content": "the model parameters or algorithm that one wants to meta-learn.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7, + "bbox_fs": [ + 104, + 119, + 507, + 241 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 258 + ], + "score": 1.0, + "content": "Basic FSL Classifier. We employ ProtoNet (Snell et al., 2017) as the basic FSL model. This model", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "has a feature-extraction CNN and a simple non-parametric classifier. The parameter of the feature", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 269, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 280 + ], + "score": 1.0, + "content": "extractor is to be meta-learned. Concretely, in the inner loop of an episode, ProtoNet fixes the feature", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 420, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 420, + 291 + ], + "score": 1.0, + "content": "extractor and computes the mean feature embedding for each class as follows:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 246, + 505, + 291 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 294, + 382, + 326 + ], + "lines": [ + { + "bbox": [ + 228, + 294, + 382, + 326 + ], + "spans": [ + { + "bbox": [ + 228, + 294, + 382, + 326 + ], + "score": 0.94, + "content": "h _ { c } = \\frac { 1 } { k } \\cdot \\sum _ { ( x _ { i } , y _ { i } ) \\in S _ { e } } f _ { \\phi } ( x _ { i } ) \\cdot \\mathbb { I } ( y _ { i } = c ) ,", + "type": "interline_equation", + "image_path": "af53dfdfebf92b14e994742c9d44f9cec3ef0b37d7e37b409e069334abdc5140.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 294, + 382, + 310.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 228, + 310.0, + 382, + 326.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 331, + 504, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 157, + 344 + ], + "score": 1.0, + "content": "where class", + "type": "text" + }, + { + "bbox": [ + 157, + 331, + 187, + 343 + ], + "score": 0.69, + "content": "c \\in \\mathcal { C } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 331, + 192, + 344 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 192, + 331, + 203, + 344 + ], + "score": 0.72, + "content": "f _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 331, + 403, + 344 + ], + "score": 1.0, + "content": "is a feature extractor with learnable parameters", + "type": "text" + }, + { + "bbox": [ + 404, + 332, + 411, + 343 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 331, + 433, + 344 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 433, + 332, + 439, + 341 + ], + "score": 0.57, + "content": "\\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "is the indicator", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "score": 1.0, + "content": "function. By computing the distance between the feature embedding of each query sample and that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 352, + 492, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 360, + 366 + ], + "score": 1.0, + "content": "of the corresponding class, the loss function used to meta-learn", + "type": "text" + }, + { + "bbox": [ + 360, + 354, + 368, + 365 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 352, + 492, + 366 + ], + "score": 1.0, + "content": "in the outer loop is defined as:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 331, + 506, + 366 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 369, + 443, + 402 + ], + "lines": [ + { + "bbox": [ + 167, + 369, + 443, + 402 + ], + "spans": [ + { + "bbox": [ + 167, + 369, + 443, + 402 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { f s l } ( S _ { e } , \\mathcal { Q } _ { e } ) = \\frac { 1 } { | \\mathcal { Q } _ { e } | } \\sum _ { ( x _ { i } , y _ { i } ) \\in \\mathcal { Q } _ { e } } - \\log \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { y _ { i } } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ) ) } ,", + "type": "interline_equation", + "image_path": "f5b0a49ba9ad956ade834f7dac81d4842c7687c92752b1ebb7c1e41414e7ca0f.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 167, + 369, + 443, + 380.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 167, + 380.0, + 443, + 391.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 167, + 391.0, + 443, + 402.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 407, + 356, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 357, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 133, + 421 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 407, + 157, + 420 + ], + "score": 0.92, + "content": "d ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 405, + 306, + 421 + ], + "score": 1.0, + "content": "denotes a distance function (e.g., the", + "type": "text" + }, + { + "bbox": [ + 306, + 408, + 315, + 419 + ], + "score": 0.87, + "content": "l _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 405, + 357, + 421 + ], + "score": 1.0, + "content": "distance).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 405, + 357, + 421 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 425, + 240, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 240, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 240, + 438 + ], + "score": 1.0, + "content": "3.2 PRETEXT TASKS IN IEPT", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 445, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "The schematic of our IEPT is illustrated in Figure 1. We first define a set of 2D-rotation operators", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 219, + 468 + ], + "score": 0.91, + "content": "\\mathcal { G } = \\{ g _ { r } | r = 0 , . . . , R - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 456, + 252, + 470 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 252, + 458, + 262, + 468 + ], + "score": 0.85, + "content": "g _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 456, + 448, + 470 + ], + "score": 1.0, + "content": "means the operator of rotating the image by", + "type": "text" + }, + { + "bbox": [ + 448, + 457, + 470, + 467 + ], + "score": 0.87, + "content": "r ^ { * 9 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "degrees", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 124, + 480 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 469, + 133, + 478 + ], + "score": 0.84, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 467, + 267, + 480 + ], + "score": 1.0, + "content": "is the total number of rotations (", + "type": "text" + }, + { + "bbox": [ + 267, + 468, + 295, + 478 + ], + "score": 0.88, + "content": "R = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "in our implementation). Given an original episode", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 107, + 478, + 173, + 491 + ], + "score": 0.93, + "content": "E _ { e } = \\{ S _ { e } , \\mathcal { Q } _ { e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 478, + 453, + 491 + ], + "score": 1.0, + "content": "as described in Sec. 3.1, we utilize the 2D-rotation operators from", + "type": "text" + }, + { + "bbox": [ + 454, + 479, + 462, + 489 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "in turn to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 209, + 502 + ], + "score": 1.0, + "content": "transform each image in", + "type": "text" + }, + { + "bbox": [ + 209, + 490, + 222, + 501 + ], + "score": 0.87, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 490, + 322, + 502 + ], + "score": 1.0, + "content": ". This results in a set of", + "type": "text" + }, + { + "bbox": [ + 323, + 490, + 332, + 500 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "extended episodes (including the original", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 499, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 128, + 514 + ], + "score": 1.0, + "content": "one)", + "type": "text" + }, + { + "bbox": [ + 129, + 500, + 276, + 512 + ], + "score": 0.9, + "content": "E = \\{ \\{ \\mathcal { S } _ { e } ^ { r } , \\bar { \\mathcal { Q } } _ { e } ^ { r } \\} | r = 0 , . . . , R - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 499, + 312, + 514 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 312, + 500, + 486, + 513 + ], + "score": 0.92, + "content": "S _ { e } ^ { r } \\ = \\ \\{ ( x _ { i } , y _ { i } , \\bar { r } ) | y _ { i } \\in C _ { e } , i \\ = \\ \\bar { 1 } , . . . , l _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 499, + 506, + 514 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 107, + 513, + 272, + 523 + ], + "score": 0.86, + "content": "\\mathcal { Q } _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in C _ { e } , i = 1 , . . . , l _ { q } \\} ,", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 511, + 411, + 525 + ], + "score": 1.0, + "content": ". Now each episode is denoted as", + "type": "text" + }, + { + "bbox": [ + 411, + 512, + 505, + 524 + ], + "score": 0.89, + "content": "E _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 520, + 502, + 537 + ], + "spans": [ + { + "bbox": [ + 107, + 524, + 251, + 534 + ], + "score": 0.81, + "content": "C _ { e } , i = 1 , . . . , l _ { k } , l _ { k } + 1 , . . . , l _ { k } + l _ { q } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 520, + 317, + 537 + ], + "score": 1.0, + "content": ", where the first", + "type": "text" + }, + { + "bbox": [ + 318, + 523, + 327, + 533 + ], + "score": 0.87, + "content": "l _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 520, + 402, + 537 + ], + "score": 1.0, + "content": "samples are from", + "type": "text" + }, + { + "bbox": [ + 402, + 523, + 414, + 534 + ], + "score": 0.89, + "content": "S _ { e } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 520, + 489, + 537 + ], + "score": 1.0, + "content": "and the rest from", + "type": "text" + }, + { + "bbox": [ + 489, + 523, + 502, + 534 + ], + "score": 0.88, + "content": "\\mathcal { Q } _ { e } ^ { r }", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 532, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 145, + 548 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 146, + 534, + 185, + 546 + ], + "score": 0.91, + "content": "\\{ \\cal { S } _ { e } ^ { 0 } , \\cal { Q } _ { e } ^ { 0 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 532, + 275, + 548 + ], + "score": 1.0, + "content": "is the original episode", + "type": "text" + }, + { + "bbox": [ + 276, + 534, + 313, + 546 + ], + "score": 0.92, + "content": "\\{ \\boldsymbol { S } _ { e } , \\boldsymbol { \\mathcal { Q } } _ { e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 532, + 506, + 548 + ], + "score": 1.0, + "content": ". With the rotation transformations, each sample", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 545, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 148, + 558 + ], + "score": 0.91, + "content": "( x _ { i } , y _ { i } , r _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 545, + 159, + 559 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 160, + 546, + 169, + 556 + ], + "score": 0.81, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 545, + 248, + 559 + ], + "score": 1.0, + "content": "carries a class label", + "type": "text" + }, + { + "bbox": [ + 248, + 547, + 258, + 557 + ], + "score": 0.85, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 545, + 494, + 559 + ], + "score": 1.0, + "content": "for supervised learning (from the inherent class) and a label", + "type": "text" + }, + { + "bbox": [ + 495, + 547, + 504, + 556 + ], + "score": 0.82, + "content": "r _ { i }", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "from the rotation operator for self-supervised learning. After generating the set of extended episodes", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 107, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 107, + 568, + 116, + 577 + ], + "score": 0.79, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 567, + 206, + 580 + ], + "score": 1.0, + "content": ", the feature extractor", + "type": "text" + }, + { + "bbox": [ + 207, + 568, + 218, + 579 + ], + "score": 0.88, + "content": "f _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 567, + 323, + 580 + ], + "score": 1.0, + "content": "is applied to each image", + "type": "text" + }, + { + "bbox": [ + 323, + 569, + 334, + 578 + ], + "score": 0.85, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 567, + 347, + 580 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 347, + 568, + 356, + 577 + ], + "score": 0.8, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 567, + 505, + 580 + ], + "score": 1.0, + "content": ". On these episodes, we design two", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 578, + 435, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 435, + 591 + ], + "score": 1.0, + "content": "self-supervised pretext tasks, one at the instance-level and the other episode-level.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 444, + 506, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 506, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "Instance-Level Pretext Task. The instance-level task is to recognize different rotation transforma-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 378, + 619 + ], + "score": 1.0, + "content": "tions. The idea is that if the model to be meta-learned here (i.e.,", + "type": "text" + }, + { + "bbox": [ + 378, + 606, + 392, + 618 + ], + "score": 0.84, + "content": "f _ { \\phi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "can be used to distinguish", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "different transformations, it must understand the canonical poses of objects (e.g., animals have legs", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "touching the ground and trees have leaves on top), a vital part of class-agnostic and thus transferable", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 637, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 316, + 653 + ], + "score": 1.0, + "content": "knowledge. With the self-supervised rotation label", + "type": "text" + }, + { + "bbox": [ + 317, + 640, + 326, + 650 + ], + "score": 0.83, + "content": "r _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 637, + 440, + 653 + ], + "score": 1.0, + "content": ", we consider the mapping:", + "type": "text" + }, + { + "bbox": [ + 440, + 640, + 504, + 651 + ], + "score": 0.89, + "content": "f _ { \\theta _ { r o t } } : x _ { i } \\mapsto r _ { i }", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 648, + 507, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 177, + 664 + ], + "score": 1.0, + "content": "for each instance", + "type": "text" + }, + { + "bbox": [ + 177, + 650, + 240, + 662 + ], + "score": 0.91, + "content": "( x _ { i } , y _ { i } , r _ { i } ) \\in E", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 648, + 270, + 664 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 271, + 650, + 290, + 662 + ], + "score": 0.92, + "content": "f _ { \\theta _ { r o t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 648, + 486, + 664 + ], + "score": 1.0, + "content": "is a rotation classifier with learnable parameters", + "type": "text" + }, + { + "bbox": [ + 486, + 650, + 502, + 661 + ], + "score": 0.89, + "content": "\\theta _ { r o t }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 648, + 507, + 664 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 660, + 460, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 189, + 674 + ], + "score": 1.0, + "content": "Given the input pair", + "type": "text" + }, + { + "bbox": [ + 189, + 662, + 218, + 673 + ], + "score": 0.92, + "content": "( x _ { i } , r _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 660, + 460, + 674 + ], + "score": 1.0, + "content": ", the total instance-level rotation loss is a cross-entropy loss:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 594, + 507, + 674 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 677, + 455, + 714 + ], + "lines": [ + { + "bbox": [ + 154, + 677, + 455, + 714 + ], + "spans": [ + { + "bbox": [ + 154, + 677, + 455, + 714 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { i n s t } = \\frac { 1 } { R ( l _ { k } + l _ { q } ) } \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { ( x _ { i } , y _ { i } , r _ { i } ) \\in E _ { e } ^ { r } } - \\log \\frac { \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r _ { i } } ) } { \\sum _ { r ^ { \\prime } = 0 } ^ { R - 1 } \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r ^ { \\prime } } ) } ,", + "type": "interline_equation", + "image_path": "efabe317b4928532f5cf8bbec84ba16e3a6148c3e31c52bd61720019bd8d7f06.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 154, + 677, + 455, + 689.3333333333334 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 154, + 689.3333333333334, + 455, + 701.6666666666667 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 154, + 701.6666666666667, + 455, + 714.0000000000001 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 720, + 492, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 494, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 133, + 734 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 720, + 219, + 733 + ], + "score": 0.93, + "content": "[ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] \\in \\mathbb { R } ^ { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 719, + 354, + 734 + ], + "score": 1.0, + "content": "is the rotation scoring vector and", + "type": "text" + }, + { + "bbox": [ + 355, + 720, + 368, + 732 + ], + "score": 0.9, + "content": "[ \\cdot ] _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 719, + 439, + 734 + ], + "score": 1.0, + "content": "means taking the", + "type": "text" + }, + { + "bbox": [ + 440, + 722, + 446, + 730 + ], + "score": 0.77, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 719, + 494, + 734 + ], + "score": 1.0, + "content": "-th element.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 719, + 494, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Episode-Level Pretext Task. We design the episode-level task based on a simple principle: although", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "different extended episodes contain images with different rotation transformations, these transforma-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "tions do not change their class labels. Consequently, the FSL classifier should produce consistent", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "probability distributions for each instance across different extended episodes. Such consistency can", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "be measured using the Kullback–Leibler (KL) divergence. Formally, for each extended episode", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 136, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 107, + 137, + 146, + 149 + ], + "score": 0.92, + "content": "\\{ S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 136, + 157, + 151 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 157, + 138, + 166, + 147 + ], + "score": 0.8, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 136, + 490, + 151 + ], + "score": 1.0, + "content": ", we first define the probability distribution of FSL classification over the query set", + "type": "text" + }, + { + "bbox": [ + 491, + 137, + 504, + 149 + ], + "score": 0.87, + "content": "\\mathcal { Q } _ { e } ^ { r }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 147, + 117, + 162 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 118, + 148, + 232, + 162 + ], + "score": 0.88, + "content": "\\mathcal { P } _ { e } ^ { r } = [ p _ { 1 } ^ { r } ; \\cdot \\cdot \\cdot ; p _ { l _ { q } } ^ { r } ] \\in \\mathbb { R } ^ { l _ { q } \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 147, + 263, + 162 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 149, + 299, + 160 + ], + "score": 0.93, + "content": "p _ { i } ^ { r } \\in \\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 147, + 435, + 162 + ], + "score": 1.0, + "content": "is the probability distribution for", + "type": "text" + }, + { + "bbox": [ + 435, + 150, + 446, + 159 + ], + "score": 0.84, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 147, + 457, + 162 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 458, + 149, + 471, + 160 + ], + "score": 0.9, + "content": "\\mathcal { Q } _ { e } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "with its", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 160, + 262, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 164, + 112, + 172 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 160, + 158, + 175 + ], + "score": 1.0, + "content": "-th element", + "type": "text" + }, + { + "bbox": [ + 159, + 162, + 178, + 174 + ], + "score": 0.8, + "content": "[ p _ { i } ^ { r } ] _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 163, + 233, + 173 + ], + "score": 0.72, + "content": "\\mathbf { \\bar { \\psi } } c = 1 , . . . , n )", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 160, + 262, + 175 + ], + "score": 1.0, + "content": "being:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 175, + 379, + 203 + ], + "lines": [ + { + "bbox": [ + 232, + 175, + 379, + 203 + ], + "spans": [ + { + "bbox": [ + 232, + 175, + 379, + 203 + ], + "score": 0.95, + "content": "[ p _ { i } ^ { r } ] _ { c } = \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ^ { r } ) ) } { \\sum _ { c ^ { \\prime } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c ^ { \\prime } } ^ { r } ) ) } .", + "type": "interline_equation", + "image_path": "2c3050dc8781955ba9dbc5fa45a1818fd14c0aa77604f3b6b6b1e8ea3d5cb85c.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 232, + 175, + 379, + 203 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 104, + 201, + 502, + 218 + ], + "spans": [ + { + "bbox": [ + 104, + 201, + 408, + 218 + ], + "score": 1.0, + "content": "The above probability is computed as in Sec. 3.1 and the class embedding", + "type": "text" + }, + { + "bbox": [ + 408, + 204, + 420, + 216 + ], + "score": 0.89, + "content": "h _ { c } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 201, + 491, + 218 + ], + "score": 1.0, + "content": "is obtained from", + "type": "text" + }, + { + "bbox": [ + 491, + 206, + 502, + 216 + ], + "score": 0.86, + "content": "S _ { e } ^ { r }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 421, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 268, + 228 + ], + "score": 1.0, + "content": "The mean probability distribution of the", + "type": "text" + }, + { + "bbox": [ + 268, + 215, + 277, + 225 + ], + "score": 0.85, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 213, + 421, + 228 + ], + "score": 1.0, + "content": "extended episodes is thus given by:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 270, + 228, + 342, + 262 + ], + "lines": [ + { + "bbox": [ + 270, + 228, + 342, + 262 + ], + "spans": [ + { + "bbox": [ + 270, + 228, + 342, + 262 + ], + "score": 0.94, + "content": "\\hat { p } _ { i } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } p _ { i } ^ { r } .", + "type": "interline_equation", + "image_path": "3f129c945a79dc4d05d8d1ab808e4ca1159962847a396d4818341c01b3452008.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 270, + 228, + 342, + 245.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 270, + 245.0, + 342, + 262.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 491, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 492, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 492, + 277 + ], + "score": 1.0, + "content": "The total episode-level consistency regularization loss is computed with the KL divergence loss:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 202, + 276, + 408, + 312 + ], + "lines": [ + { + "bbox": [ + 202, + 276, + 408, + 312 + ], + "spans": [ + { + "bbox": [ + 202, + 276, + 408, + 312 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { e p i s } = \\frac { 1 } { R l _ { q } } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { i = 1 } ^ { l _ { q } } \\mathrm { m e a n } ( p _ { i } ^ { r } ( \\log p _ { i } ^ { r } - \\log \\hat { p } _ { i } ) ) .", + "type": "interline_equation", + "image_path": "9daa6c2ade77f619cc6c155e3739cd646215dd62cd89b59454ee5609fbdb7a19.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 276, + 408, + 294.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 202, + 294.0, + 408, + 312.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 325, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 325, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 133, + 327 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 314, + 168, + 326 + ], + "score": 0.62, + "content": "\\mathrm { m e a n } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 313, + 325, + 327 + ], + "score": 1.0, + "content": "is an element-wise averaging function.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 330, + 234, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 234, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 234, + 343 + ], + "score": 1.0, + "content": "3.3 INTEGRATED FSL TASK", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 365 + ], + "score": 1.0, + "content": "The two tasks introduced so far are self-supervised tasks without using the class labels in the query", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 500, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 500, + 374 + ], + "score": 1.0, + "content": "set. Now we describe how in the supervised classification task, the extended episodes can be used.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 378, + 506, + 501 + ], + "lines": [ + { + "bbox": [ + 104, + 376, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 255, + 394 + ], + "score": 1.0, + "content": "Given the set of extended episodes", + "type": "text" + }, + { + "bbox": [ + 255, + 380, + 264, + 389 + ], + "score": 0.81, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 376, + 388, + 394 + ], + "score": 1.0, + "content": ", we denote the feature set of", + "type": "text" + }, + { + "bbox": [ + 389, + 380, + 398, + 389 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 376, + 411, + 394 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 412, + 379, + 435, + 390 + ], + "score": 0.91, + "content": "E _ { e m b }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 376, + 468, + 394 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 468, + 379, + 505, + 390 + ], + "score": 0.89, + "content": "E _ { e m b } =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 390, + 504, + 404 + ], + "spans": [ + { + "bbox": [ + 107, + 390, + 353, + 402 + ], + "score": 0.78, + "content": "\\{ f _ { \\phi } ( x _ { i } ) | ( x _ { i } , y _ { i } , r ) \\in E _ { e } ^ { r } , r = 0 , \\cdots , R - 1 , i = 1 , . . . , l _ { k } + l _ { q } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 390, + 495, + 404 + ], + "score": 1.0, + "content": ". Note that each extended episode in", + "type": "text" + }, + { + "bbox": [ + 495, + 391, + 504, + 400 + ], + "score": 0.81, + "content": "E", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "corresponds to one specific rotation transformation of the same set of images from the original episode", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 412, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 107, + 412, + 119, + 423 + ], + "score": 0.86, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 412, + 506, + 424 + ], + "score": 1.0, + "content": ". Therefore, in order to capture the correlation among instances with different transformations and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "score": 1.0, + "content": "learn how best combine them to form the class mean for meta-learning, an instance attention module", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 433, + 239, + 447 + ], + "score": 1.0, + "content": "is deployed w.r.t. each image in", + "type": "text" + }, + { + "bbox": [ + 239, + 434, + 252, + 446 + ], + "score": 0.88, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "(i.e., all images are assumed to be independent). Specifically,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 443, + 508, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 143, + 460 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 144, + 446, + 167, + 457 + ], + "score": 0.9, + "content": "E _ { e m b }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 443, + 293, + 460 + ], + "score": 1.0, + "content": ", we construct the feature tensor", + "type": "text" + }, + { + "bbox": [ + 293, + 445, + 370, + 456 + ], + "score": 0.92, + "content": "\\bar { F } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 443, + 400, + 460 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 401, + 446, + 407, + 456 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 443, + 508, + 460 + ], + "score": 1.0, + "content": "is the feature dimension.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "We then adopt a transformer to obtain the integrated representation for FSL classification. The", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "transformer architecture is based on a self-attention mechanism, as in (Vaswani et al., 2017). It", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 477, + 504, + 492 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 209, + 492 + ], + "score": 1.0, + "content": "receives the triplet input", + "type": "text" + }, + { + "bbox": [ + 209, + 479, + 247, + 491 + ], + "score": 0.93, + "content": "( F , F , F )", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 477, + 260, + 492 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 261, + 479, + 302, + 491 + ], + "score": 0.92, + "content": "( Q , K , V )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 477, + 486, + 492 + ], + "score": 1.0, + "content": "(Query, Key, and Value, respectively). With", + "type": "text" + }, + { + "bbox": [ + 486, + 478, + 504, + 489 + ], + "score": 0.9, + "content": "\\dot { F } ^ { ( i ) }", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 490, + 454, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 146, + 503 + ], + "score": 1.0, + "content": "being the", + "type": "text" + }, + { + "bbox": [ + 146, + 491, + 150, + 500 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 490, + 191, + 503 + ], + "score": 1.0, + "content": "-th row of", + "type": "text" + }, + { + "bbox": [ + 192, + 491, + 200, + 500 + ], + "score": 0.83, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 490, + 242, + 503 + ], + "score": 1.0, + "content": "(w.r.t. the", + "type": "text" + }, + { + "bbox": [ + 242, + 492, + 246, + 500 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 490, + 296, + 503 + ], + "score": 1.0, + "content": "-th image in", + "type": "text" + }, + { + "bbox": [ + 297, + 491, + 309, + 501 + ], + "score": 0.88, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 490, + 454, + 503 + ], + "score": 1.0, + "content": "), the attentive module is defined as:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 503, + 410, + 553 + ], + "lines": [ + { + "bbox": [ + 199, + 503, + 410, + 553 + ], + "spans": [ + { + "bbox": [ + 199, + 503, + 410, + 553 + ], + "score": 0.79, + "content": "\\begin{array} { r l } & { ( \\boldsymbol { F } _ { Q } ^ { ( i ) } , \\boldsymbol { F } _ { K } ^ { ( i ) } , \\boldsymbol { F } _ { V } ^ { ( i ) } ) = ( \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { Q } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { K } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { V } ) , } \\\\ & { \\boldsymbol { F } _ { a t t } ^ { ( i ) } = \\boldsymbol { F } ^ { ( i ) } + \\mathrm { s o f t m a x } ( \\frac { \\boldsymbol { F } _ { Q } ^ { ( i ) } ( \\boldsymbol { F } _ { K } ^ { ( i ) } ) ^ { T } } { \\sqrt { d _ { K } } } ) \\boldsymbol { F } _ { V } ^ { ( i ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "42b72ef524a028212b41d64b4214ac497599dc9918626fb22811be7f56b4fea4.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 199, + 503, + 410, + 519.6666666666666 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 199, + 519.6666666666666, + 410, + 536.3333333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 199, + 536.3333333333333, + 410, + 552.9999999999999 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 133, + 567 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 554, + 166, + 565 + ], + "score": 0.91, + "content": "d _ { K } = d", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 554, + 187, + 567 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 187, + 554, + 246, + 567 + ], + "score": 0.9, + "content": "W _ { Q } , W _ { K } , W _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "represent the parameters of three fully-connected layers respec-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 372, + 578 + ], + "score": 1.0, + "content": "tively (the parameters of the integration transformer are collected as", + "type": "text" + }, + { + "bbox": [ + 372, + 565, + 389, + 576 + ], + "score": 0.91, + "content": "\\theta _ { i n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "). Note that the key and value", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "are computed from each image and its augmented versions, i.e., they are computed independently", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 585, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 104, + 585, + 359, + 603 + ], + "score": 1.0, + "content": "without using inter-image correlation. With the attentive feature", + "type": "text" + }, + { + "bbox": [ + 359, + 587, + 445, + 599 + ], + "score": 0.93, + "content": "F _ { a t t } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 585, + 506, + 603 + ], + "score": 1.0, + "content": ", the integrated", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 596, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 167, + 615 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 167, + 599, + 307, + 612 + ], + "score": 0.91, + "content": "F _ { i n t e g } = [ F ^ { S } ; F ^ { Q } ] \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R d }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 596, + 312, + 615 + ], + "score": 1.0, + "content": "(", + "type": "text" + }, + { + "bbox": [ + 312, + 599, + 327, + 610 + ], + "score": 0.89, + "content": "F ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 596, + 345, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 345, + 600, + 361, + 610 + ], + "score": 0.9, + "content": "F ^ { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 596, + 507, + 615 + ], + "score": 1.0, + "content": "are respectively for the support set", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 610, + 214, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 214, + 624 + ], + "score": 1.0, + "content": "and query set) is given by:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 621, + 353, + 635 + ], + "lines": [ + { + "bbox": [ + 256, + 621, + 353, + 635 + ], + "spans": [ + { + "bbox": [ + 256, + 621, + 353, + 635 + ], + "score": 0.92, + "content": "F _ { i n t e g } = \\mathrm { f l a t t e n } ( F _ { a t t } ) ,", + "type": "interline_equation", + "image_path": "8780c8be02682008dc4bfb993ddfe86f70bdce00077a0123f7718852eb6f7393.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 256, + 621, + 353, + 635 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 635, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 161, + 648 + ], + "score": 1.0, + "content": "where flatten", + "type": "text" + }, + { + "bbox": [ + 161, + 635, + 172, + 647 + ], + "score": 0.48, + "content": "( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 634, + 244, + 648 + ], + "score": 1.0, + "content": "denotes flattening", + "type": "text" + }, + { + "bbox": [ + 244, + 636, + 262, + 646 + ], + "score": 0.9, + "content": "F _ { a t t }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 634, + 505, + 648 + ], + "score": 1.0, + "content": "along the last two dimensions, i.e., concatenating the attentive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "features from different extended episodes for the corresponding images. The integrated representation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 657, + 400, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 400, + 669 + ], + "score": 1.0, + "content": "is then inputted to the FSL classifier to define the FSL classification loss:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 670, + 408, + 705 + ], + "lines": [ + { + "bbox": [ + 203, + 670, + 408, + 705 + ], + "spans": [ + { + "bbox": [ + 203, + 670, + 408, + 705 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { i n t e g } = \\frac { 1 } { l _ { q } } \\cdot \\sum _ { i = 1 } ^ { l _ { q } } - \\log \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y _ { i } } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) }", + "type": "interline_equation", + "image_path": "6fe4d30f749bff32263c8788315d1793d18a141b81fbea68098739d2758d49b4.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 670, + 408, + 687.5 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 203, + 687.5, + 408, + 705.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 102, + 702, + 510, + 734 + ], + "spans": [ + { + "bbox": [ + 102, + 702, + 215, + 734 + ], + "score": 1.0, + "content": "where the class embedding the integrated FSL task act", + "type": "text" + }, + { + "bbox": [ + 216, + 708, + 338, + 722 + ], + "score": 0.91, + "content": "\\begin{array} { r } { h _ { c } ^ { f } = \\frac { 1 } { k } \\cdot \\sum _ { i = 1 } ^ { l _ { k } } F _ { i } ^ { S } \\cdot \\mathbb { I } ( y _ { i } = c ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 702, + 510, + 734 + ], + "score": 1.0, + "content": "is computed on the support set. Note thatrediction averaging.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Episode-Level Pretext Task. We design the episode-level task based on a simple principle: although", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "different extended episodes contain images with different rotation transformations, these transforma-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "tions do not change their class labels. Consequently, the FSL classifier should produce consistent", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "probability distributions for each instance across different extended episodes. Such consistency can", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "be measured using the Kullback–Leibler (KL) divergence. Formally, for each extended episode", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 136, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 107, + 137, + 146, + 149 + ], + "score": 0.92, + "content": "\\{ S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 136, + 157, + 151 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 157, + 138, + 166, + 147 + ], + "score": 0.8, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 136, + 490, + 151 + ], + "score": 1.0, + "content": ", we first define the probability distribution of FSL classification over the query set", + "type": "text" + }, + { + "bbox": [ + 491, + 137, + 504, + 149 + ], + "score": 0.87, + "content": "\\mathcal { Q } _ { e } ^ { r }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 147, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 104, + 147, + 117, + 162 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 118, + 148, + 232, + 162 + ], + "score": 0.88, + "content": "\\mathcal { P } _ { e } ^ { r } = [ p _ { 1 } ^ { r } ; \\cdot \\cdot \\cdot ; p _ { l _ { q } } ^ { r } ] \\in \\mathbb { R } ^ { l _ { q } \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 147, + 263, + 162 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 149, + 299, + 160 + ], + "score": 0.93, + "content": "p _ { i } ^ { r } \\in \\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 147, + 435, + 162 + ], + "score": 1.0, + "content": "is the probability distribution for", + "type": "text" + }, + { + "bbox": [ + 435, + 150, + 446, + 159 + ], + "score": 0.84, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 147, + 457, + 162 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 458, + 149, + 471, + 160 + ], + "score": 0.9, + "content": "\\mathcal { Q } _ { e } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 147, + 506, + 162 + ], + "score": 1.0, + "content": "with its", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 160, + 262, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 164, + 112, + 172 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 160, + 158, + 175 + ], + "score": 1.0, + "content": "-th element", + "type": "text" + }, + { + "bbox": [ + 159, + 162, + 178, + 174 + ], + "score": 0.8, + "content": "[ p _ { i } ^ { r } ] _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 163, + 233, + 173 + ], + "score": 0.72, + "content": "\\mathbf { \\bar { \\psi } } c = 1 , . . . , n )", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 160, + 262, + 175 + ], + "score": 1.0, + "content": "being:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 82, + 506, + 175 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 175, + 379, + 203 + ], + "lines": [ + { + "bbox": [ + 232, + 175, + 379, + 203 + ], + "spans": [ + { + "bbox": [ + 232, + 175, + 379, + 203 + ], + "score": 0.95, + "content": "[ p _ { i } ^ { r } ] _ { c } = \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ^ { r } ) ) } { \\sum _ { c ^ { \\prime } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c ^ { \\prime } } ^ { r } ) ) } .", + "type": "interline_equation", + "image_path": "2c3050dc8781955ba9dbc5fa45a1818fd14c0aa77604f3b6b6b1e8ea3d5cb85c.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 232, + 175, + 379, + 203 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 104, + 201, + 502, + 218 + ], + "spans": [ + { + "bbox": [ + 104, + 201, + 408, + 218 + ], + "score": 1.0, + "content": "The above probability is computed as in Sec. 3.1 and the class embedding", + "type": "text" + }, + { + "bbox": [ + 408, + 204, + 420, + 216 + ], + "score": 0.89, + "content": "h _ { c } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 201, + 491, + 218 + ], + "score": 1.0, + "content": "is obtained from", + "type": "text" + }, + { + "bbox": [ + 491, + 206, + 502, + 216 + ], + "score": 0.86, + "content": "S _ { e } ^ { r }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 421, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 268, + 228 + ], + "score": 1.0, + "content": "The mean probability distribution of the", + "type": "text" + }, + { + "bbox": [ + 268, + 215, + 277, + 225 + ], + "score": 0.85, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 213, + 421, + 228 + ], + "score": 1.0, + "content": "extended episodes is thus given by:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 201, + 502, + 228 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 270, + 228, + 342, + 262 + ], + "lines": [ + { + "bbox": [ + 270, + 228, + 342, + 262 + ], + "spans": [ + { + "bbox": [ + 270, + 228, + 342, + 262 + ], + "score": 0.94, + "content": "\\hat { p } _ { i } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } p _ { i } ^ { r } .", + "type": "interline_equation", + "image_path": "3f129c945a79dc4d05d8d1ab808e4ca1159962847a396d4818341c01b3452008.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 270, + 228, + 342, + 245.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 270, + 245.0, + 342, + 262.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 491, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 492, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 492, + 277 + ], + "score": 1.0, + "content": "The total episode-level consistency regularization loss is computed with the KL divergence loss:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 262, + 492, + 277 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 202, + 276, + 408, + 312 + ], + "lines": [ + { + "bbox": [ + 202, + 276, + 408, + 312 + ], + "spans": [ + { + "bbox": [ + 202, + 276, + 408, + 312 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { e p i s } = \\frac { 1 } { R l _ { q } } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { i = 1 } ^ { l _ { q } } \\mathrm { m e a n } ( p _ { i } ^ { r } ( \\log p _ { i } ^ { r } - \\log \\hat { p } _ { i } ) ) .", + "type": "interline_equation", + "image_path": "9daa6c2ade77f619cc6c155e3739cd646215dd62cd89b59454ee5609fbdb7a19.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 276, + 408, + 294.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 202, + 294.0, + 408, + 312.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 325, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 325, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 133, + 327 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 314, + 168, + 326 + ], + "score": 0.62, + "content": "\\mathrm { m e a n } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 313, + 325, + 327 + ], + "score": 1.0, + "content": "is an element-wise averaging function.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 313, + 325, + 327 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 330, + 234, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 234, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 234, + 343 + ], + "score": 1.0, + "content": "3.3 INTEGRATED FSL TASK", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 365 + ], + "score": 1.0, + "content": "The two tasks introduced so far are self-supervised tasks without using the class labels in the query", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 500, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 500, + 374 + ], + "score": 1.0, + "content": "set. Now we describe how in the supervised classification task, the extended episodes can be used.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 349, + 505, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 378, + 506, + 501 + ], + "lines": [ + { + "bbox": [ + 104, + 376, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 255, + 394 + ], + "score": 1.0, + "content": "Given the set of extended episodes", + "type": "text" + }, + { + "bbox": [ + 255, + 380, + 264, + 389 + ], + "score": 0.81, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 376, + 388, + 394 + ], + "score": 1.0, + "content": ", we denote the feature set of", + "type": "text" + }, + { + "bbox": [ + 389, + 380, + 398, + 389 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 376, + 411, + 394 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 412, + 379, + 435, + 390 + ], + "score": 0.91, + "content": "E _ { e m b }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 376, + 468, + 394 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 468, + 379, + 505, + 390 + ], + "score": 0.89, + "content": "E _ { e m b } =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 390, + 504, + 404 + ], + "spans": [ + { + "bbox": [ + 107, + 390, + 353, + 402 + ], + "score": 0.78, + "content": "\\{ f _ { \\phi } ( x _ { i } ) | ( x _ { i } , y _ { i } , r ) \\in E _ { e } ^ { r } , r = 0 , \\cdots , R - 1 , i = 1 , . . . , l _ { k } + l _ { q } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 390, + 495, + 404 + ], + "score": 1.0, + "content": ". Note that each extended episode in", + "type": "text" + }, + { + "bbox": [ + 495, + 391, + 504, + 400 + ], + "score": 0.81, + "content": "E", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "corresponds to one specific rotation transformation of the same set of images from the original episode", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 412, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 107, + 412, + 119, + 423 + ], + "score": 0.86, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 412, + 506, + 424 + ], + "score": 1.0, + "content": ". Therefore, in order to capture the correlation among instances with different transformations and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "score": 1.0, + "content": "learn how best combine them to form the class mean for meta-learning, an instance attention module", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 433, + 239, + 447 + ], + "score": 1.0, + "content": "is deployed w.r.t. each image in", + "type": "text" + }, + { + "bbox": [ + 239, + 434, + 252, + 446 + ], + "score": 0.88, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "(i.e., all images are assumed to be independent). Specifically,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 443, + 508, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 143, + 460 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 144, + 446, + 167, + 457 + ], + "score": 0.9, + "content": "E _ { e m b }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 443, + 293, + 460 + ], + "score": 1.0, + "content": ", we construct the feature tensor", + "type": "text" + }, + { + "bbox": [ + 293, + 445, + 370, + 456 + ], + "score": 0.92, + "content": "\\bar { F } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 443, + 400, + 460 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 401, + 446, + 407, + 456 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 443, + 508, + 460 + ], + "score": 1.0, + "content": "is the feature dimension.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "We then adopt a transformer to obtain the integrated representation for FSL classification. The", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "transformer architecture is based on a self-attention mechanism, as in (Vaswani et al., 2017). It", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 477, + 504, + 492 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 209, + 492 + ], + "score": 1.0, + "content": "receives the triplet input", + "type": "text" + }, + { + "bbox": [ + 209, + 479, + 247, + 491 + ], + "score": 0.93, + "content": "( F , F , F )", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 477, + 260, + 492 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 261, + 479, + 302, + 491 + ], + "score": 0.92, + "content": "( Q , K , V )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 477, + 486, + 492 + ], + "score": 1.0, + "content": "(Query, Key, and Value, respectively). With", + "type": "text" + }, + { + "bbox": [ + 486, + 478, + 504, + 489 + ], + "score": 0.9, + "content": "\\dot { F } ^ { ( i ) }", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 490, + 454, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 146, + 503 + ], + "score": 1.0, + "content": "being the", + "type": "text" + }, + { + "bbox": [ + 146, + 491, + 150, + 500 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 490, + 191, + 503 + ], + "score": 1.0, + "content": "-th row of", + "type": "text" + }, + { + "bbox": [ + 192, + 491, + 200, + 500 + ], + "score": 0.83, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 490, + 242, + 503 + ], + "score": 1.0, + "content": "(w.r.t. the", + "type": "text" + }, + { + "bbox": [ + 242, + 492, + 246, + 500 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 490, + 296, + 503 + ], + "score": 1.0, + "content": "-th image in", + "type": "text" + }, + { + "bbox": [ + 297, + 491, + 309, + 501 + ], + "score": 0.88, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 490, + 454, + 503 + ], + "score": 1.0, + "content": "), the attentive module is defined as:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 376, + 508, + 503 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 503, + 410, + 553 + ], + "lines": [ + { + "bbox": [ + 199, + 503, + 410, + 553 + ], + "spans": [ + { + "bbox": [ + 199, + 503, + 410, + 553 + ], + "score": 0.79, + "content": "\\begin{array} { r l } & { ( \\boldsymbol { F } _ { Q } ^ { ( i ) } , \\boldsymbol { F } _ { K } ^ { ( i ) } , \\boldsymbol { F } _ { V } ^ { ( i ) } ) = ( \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { Q } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { K } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { V } ) , } \\\\ & { \\boldsymbol { F } _ { a t t } ^ { ( i ) } = \\boldsymbol { F } ^ { ( i ) } + \\mathrm { s o f t m a x } ( \\frac { \\boldsymbol { F } _ { Q } ^ { ( i ) } ( \\boldsymbol { F } _ { K } ^ { ( i ) } ) ^ { T } } { \\sqrt { d _ { K } } } ) \\boldsymbol { F } _ { V } ^ { ( i ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "42b72ef524a028212b41d64b4214ac497599dc9918626fb22811be7f56b4fea4.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 199, + 503, + 410, + 519.6666666666666 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 199, + 519.6666666666666, + 410, + 536.3333333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 199, + 536.3333333333333, + 410, + 552.9999999999999 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 133, + 567 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 554, + 166, + 565 + ], + "score": 0.91, + "content": "d _ { K } = d", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 554, + 187, + 567 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 187, + 554, + 246, + 567 + ], + "score": 0.9, + "content": "W _ { Q } , W _ { K } , W _ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "represent the parameters of three fully-connected layers respec-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 372, + 578 + ], + "score": 1.0, + "content": "tively (the parameters of the integration transformer are collected as", + "type": "text" + }, + { + "bbox": [ + 372, + 565, + 389, + 576 + ], + "score": 0.91, + "content": "\\theta _ { i n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "). Note that the key and value", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "are computed from each image and its augmented versions, i.e., they are computed independently", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 585, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 104, + 585, + 359, + 603 + ], + "score": 1.0, + "content": "without using inter-image correlation. With the attentive feature", + "type": "text" + }, + { + "bbox": [ + 359, + 587, + 445, + 599 + ], + "score": 0.93, + "content": "F _ { a t t } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 585, + 506, + 603 + ], + "score": 1.0, + "content": ", the integrated", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 596, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 167, + 615 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 167, + 599, + 307, + 612 + ], + "score": 0.91, + "content": "F _ { i n t e g } = [ F ^ { S } ; F ^ { Q } ] \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R d }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 596, + 312, + 615 + ], + "score": 1.0, + "content": "(", + "type": "text" + }, + { + "bbox": [ + 312, + 599, + 327, + 610 + ], + "score": 0.89, + "content": "F ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 596, + 345, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 345, + 600, + 361, + 610 + ], + "score": 0.9, + "content": "F ^ { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 596, + 507, + 615 + ], + "score": 1.0, + "content": "are respectively for the support set", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 610, + 214, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 214, + 624 + ], + "score": 1.0, + "content": "and query set) is given by:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 554, + 507, + 624 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 621, + 353, + 635 + ], + "lines": [ + { + "bbox": [ + 256, + 621, + 353, + 635 + ], + "spans": [ + { + "bbox": [ + 256, + 621, + 353, + 635 + ], + "score": 0.92, + "content": "F _ { i n t e g } = \\mathrm { f l a t t e n } ( F _ { a t t } ) ,", + "type": "interline_equation", + "image_path": "8780c8be02682008dc4bfb993ddfe86f70bdce00077a0123f7718852eb6f7393.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 256, + 621, + 353, + 635 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 635, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 161, + 648 + ], + "score": 1.0, + "content": "where flatten", + "type": "text" + }, + { + "bbox": [ + 161, + 635, + 172, + 647 + ], + "score": 0.48, + "content": "( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 634, + 244, + 648 + ], + "score": 1.0, + "content": "denotes flattening", + "type": "text" + }, + { + "bbox": [ + 244, + 636, + 262, + 646 + ], + "score": 0.9, + "content": "F _ { a t t }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 634, + 505, + 648 + ], + "score": 1.0, + "content": "along the last two dimensions, i.e., concatenating the attentive", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "features from different extended episodes for the corresponding images. The integrated representation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 657, + 400, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 400, + 669 + ], + "score": 1.0, + "content": "is then inputted to the FSL classifier to define the FSL classification loss:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 634, + 505, + 669 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 670, + 408, + 705 + ], + "lines": [ + { + "bbox": [ + 203, + 670, + 408, + 705 + ], + "spans": [ + { + "bbox": [ + 203, + 670, + 408, + 705 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { i n t e g } = \\frac { 1 } { l _ { q } } \\cdot \\sum _ { i = 1 } ^ { l _ { q } } - \\log \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y _ { i } } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) }", + "type": "interline_equation", + "image_path": "6fe4d30f749bff32263c8788315d1793d18a141b81fbea68098739d2758d49b4.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 670, + 408, + 687.5 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 203, + 687.5, + 408, + 705.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 102, + 702, + 510, + 734 + ], + "spans": [ + { + "bbox": [ + 102, + 702, + 215, + 734 + ], + "score": 1.0, + "content": "where the class embedding the integrated FSL task act", + "type": "text" + }, + { + "bbox": [ + 216, + 708, + 338, + 722 + ], + "score": 0.91, + "content": "\\begin{array} { r } { h _ { c } ^ { f } = \\frac { 1 } { k } \\cdot \\sum _ { i = 1 } ^ { l _ { k } } F _ { i } ^ { S } \\cdot \\mathbb { I } ( y _ { i } = c ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 702, + 510, + 734 + ], + "score": 1.0, + "content": "is computed on the support set. Note thatrediction averaging.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46, + "bbox_fs": [ + 102, + 702, + 510, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 186, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 187, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 187, + 96 + ], + "score": 1.0, + "content": "3.4 TOTAL LOSS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 504, + 115 + ], + "score": 1.0, + "content": "The total training loss for our full model consists of the self-supervised losses from the pretext tasks", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 397, + 127 + ], + "score": 1.0, + "content": "and the supervised losses from the FSL tasks. In this work, in addition to", + "type": "text" + }, + { + "bbox": [ + 397, + 114, + 424, + 126 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 113, + 506, + 127 + ], + "score": 1.0, + "content": "in Eq. (10), another", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 191, + 138 + ], + "score": 1.0, + "content": "supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 192, + 126, + 213, + 136 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 125, + 328, + 138 + ], + "score": 1.0, + "content": "is also used (see Figure 1).", + "type": "text" + }, + { + "bbox": [ + 329, + 126, + 351, + 137 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 125, + 505, + 138 + ], + "score": 1.0, + "content": "is the average FSL classification loss", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 341, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 341, + 148 + ], + "score": 1.0, + "content": "over the extended episodes. Formally, it can be written as:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 153, + 369, + 188 + ], + "lines": [ + { + "bbox": [ + 242, + 153, + 369, + 188 + ], + "spans": [ + { + "bbox": [ + 242, + 153, + 369, + 188 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { a u x } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\mathcal { L } _ { f s l } ( S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } )", + "type": "interline_equation", + "image_path": "2f4093022e30a93dbeec9c5c4b2ce4413b4beb43881363ab8e48d87496b558e5.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 153, + 369, + 170.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 242, + 170.5, + 369, + 188.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 415, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 416, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 203, + 207 + ], + "score": 1.0, + "content": "Therefore, the total loss", + "type": "text" + }, + { + "bbox": [ + 203, + 194, + 228, + 204 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { t o t a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 192, + 416, + 207 + ], + "score": 1.0, + "content": "for training our full model is given as follows:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 210, + 424, + 253 + ], + "lines": [ + { + "bbox": [ + 187, + 210, + 424, + 253 + ], + "spans": [ + { + "bbox": [ + 187, + 210, + 424, + 253 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\mathcal { L } _ { t o t a l } = \\underbrace { \\overbrace { w _ { 1 } * \\mathcal { L } _ { i n s t } } ^ { i n s t a n c e { - } l e v e l } + \\overbrace { w _ { 2 } * \\mathcal { L } _ { e p i s } } ^ { e p i s o d e { - } l e v e l } } _ { \\mathrm { s e l f - s u p e r v i s e d l o s s } } + \\underbrace { w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t e g } } _ { \\mathrm { s u p e r v i s e d l o s s } } , } \\end{array}", + "type": "interline_equation", + "image_path": "7428972404be365c998fd96b04e5ae4729a396ac2be95422a01502c2b04b28b3.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 187, + 210, + 424, + 224.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 224.33333333333334, + 424, + 238.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 238.66666666666669, + 424, + 253.00000000000003 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 258, + 326, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 256, + 328, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 133, + 274 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 261, + 178, + 270 + ], + "score": 0.9, + "content": "w _ { 1 } , w _ { 2 } , w _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 256, + 328, + 274 + ], + "score": 1.0, + "content": "are the loss weight hyperparameters.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 284, + 181, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 182, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 182, + 297 + ], + "score": 1.0, + "content": "3.5 INFERENCE", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 367, + 318 + ], + "score": 1.0, + "content": "During the test stage, we only exploit the integrated representation", + "type": "text" + }, + { + "bbox": [ + 368, + 306, + 394, + 317 + ], + "score": 0.92, + "content": "F _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "for the final FSL prediction.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 315, + 397, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 221, + 330 + ], + "score": 1.0, + "content": "The predicted class label for", + "type": "text" + }, + { + "bbox": [ + 222, + 317, + 256, + 327 + ], + "score": 0.92, + "content": "x _ { i } \\in \\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 315, + 397, + 330 + ], + "score": 1.0, + "content": "can be computed with Eq. (10) as:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 216, + 333, + 396, + 367 + ], + "lines": [ + { + "bbox": [ + 216, + 333, + 396, + 367 + ], + "spans": [ + { + "bbox": [ + 216, + 333, + 396, + 367 + ], + "score": 0.94, + "content": "y _ { i } ^ { p r e d } = \\underset { y \\in \\mathcal { C } _ { e } } { \\operatorname { a r g m a x } } \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) } .", + "type": "interline_equation", + "image_path": "2e94de396b91c16c863f7b31f20db819e61061ae5d0a9119814439b597279b54.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 216, + 333, + 396, + 350.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 216, + 350.0, + 396, + 367.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 379, + 237, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 237, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 237, + 391 + ], + "score": 1.0, + "content": "3.6 FULL IEPT ALGORITHM", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 105, + 400, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "For easy reproduction, we present the full algorithm for FSL with IEPT in Algorithm 1. Once learned,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 412, + 449, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 172, + 424 + ], + "score": 1.0, + "content": "with the learned", + "type": "text" + }, + { + "bbox": [ + 173, + 412, + 181, + 423 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 412, + 449, + 424 + ], + "score": 1.0, + "content": ", we can perform the inference over the test episodes with Eq. (13).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 107, + 436, + 222, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 224, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 224, + 448 + ], + "score": 1.0, + "content": "Algorithm 1 FSL with IEPT", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 449, + 426, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 304, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 194, + 461 + ], + "score": 1.0, + "content": "Input: The training set", + "type": "text" + }, + { + "bbox": [ + 194, + 449, + 206, + 459 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 448, + 297, + 461 + ], + "score": 1.0, + "content": ", the rotation operator set", + "type": "text" + }, + { + "bbox": [ + 297, + 450, + 304, + 459 + ], + "score": 0.67, + "content": "\\mathcal { G }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 458, + 298, + 472 + ], + "spans": [ + { + "bbox": [ + 134, + 458, + 256, + 472 + ], + "score": 1.0, + "content": "The loss weight hyperparameters", + "type": "text" + }, + { + "bbox": [ + 256, + 461, + 298, + 469 + ], + "score": 0.77, + "content": "w _ { 1 } , w _ { 2 } , w _ { 3 }", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 468, + 194, + 480 + ], + "spans": [ + { + "bbox": [ + 107, + 468, + 186, + 480 + ], + "score": 1.0, + "content": "Output: The learned", + "type": "text" + }, + { + "bbox": [ + 187, + 470, + 194, + 480 + ], + "score": 0.8, + "content": "\\psi", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 108, + 478, + 356, + 491 + ], + "spans": [ + { + "bbox": [ + 108, + 478, + 282, + 491 + ], + "score": 1.0, + "content": "1: Randomly initialize all learnable parameters", + "type": "text" + }, + { + "bbox": [ + 283, + 479, + 356, + 490 + ], + "score": 0.88, + "content": "\\psi = \\{ \\phi , \\theta _ { r o t } , \\theta _ { i n t } \\}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 110, + 489, + 256, + 499 + ], + "spans": [ + { + "bbox": [ + 110, + 489, + 166, + 499 + ], + "score": 1.0, + "content": "2: for iteration", + "type": "text" + }, + { + "bbox": [ + 167, + 489, + 180, + 498 + ], + "score": 0.81, + "content": "= 1", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 489, + 256, + 499 + ], + "score": 1.0, + "content": ", ..., MaxIteration do", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 110, + 498, + 273, + 510 + ], + "spans": [ + { + "bbox": [ + 110, + 498, + 120, + 510 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 131, + 498, + 228, + 510 + ], + "score": 1.0, + "content": "Randomly sample episode", + "type": "text" + }, + { + "bbox": [ + 229, + 499, + 240, + 509 + ], + "score": 0.87, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 498, + 261, + 510 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 262, + 500, + 273, + 509 + ], + "score": 0.86, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 110, + 508, + 340, + 521 + ], + "spans": [ + { + "bbox": [ + 110, + 509, + 120, + 520 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 131, + 508, + 269, + 521 + ], + "score": 1.0, + "content": "Generate the set of extended episodes", + "type": "text" + }, + { + "bbox": [ + 269, + 510, + 278, + 518 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 508, + 298, + 521 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 298, + 509, + 310, + 519 + ], + "score": 0.85, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 508, + 333, + 521 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 333, + 509, + 340, + 519 + ], + "score": 0.73, + "content": "\\mathcal { G }", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 110, + 518, + 403, + 531 + ], + "spans": [ + { + "bbox": [ + 110, + 519, + 120, + 530 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 131, + 518, + 214, + 531 + ], + "score": 1.0, + "content": "Compute the SSL loss", + "type": "text" + }, + { + "bbox": [ + 214, + 519, + 235, + 529 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 518, + 403, + 531 + ], + "score": 1.0, + "content": "for the instance-level pretext task with Eq. (3)", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 528, + 402, + 541 + ], + "spans": [ + { + "bbox": [ + 110, + 529, + 119, + 539 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 131, + 528, + 214, + 541 + ], + "score": 1.0, + "content": "Compute the SSL loss", + "type": "text" + }, + { + "bbox": [ + 214, + 530, + 235, + 540 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 528, + 402, + 541 + ], + "score": 1.0, + "content": "for the episode-level pretext task with Eq. (6)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 110, + 538, + 425, + 550 + ], + "spans": [ + { + "bbox": [ + 110, + 538, + 120, + 550 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 131, + 539, + 254, + 550 + ], + "score": 1.0, + "content": "Compute the supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 255, + 539, + 275, + 549 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 539, + 425, + 550 + ], + "score": 1.0, + "content": "over the extended episodes with Eq. (11)", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 109, + 548, + 425, + 560 + ], + "spans": [ + { + "bbox": [ + 109, + 548, + 120, + 559 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 131, + 548, + 254, + 560 + ], + "score": 1.0, + "content": "Compute the supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 255, + 549, + 280, + 559 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 548, + 425, + 560 + ], + "score": 1.0, + "content": "for the integrated episode with Eq. (10)", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 109, + 558, + 346, + 572 + ], + "spans": [ + { + "bbox": [ + 109, + 559, + 120, + 569 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 135, + 559, + 335, + 570 + ], + "score": 0.27, + "content": "\\mathcal { L } _ { t o t a l } = w _ { 1 } * \\mathcal { L } _ { i n s t } + w _ { 2 } * \\mathcal { L } _ { e p i s } + w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 558, + 346, + 572 + ], + "score": 1.0, + "content": "teg", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 568, + 244, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 120, + 579 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 131, + 568, + 160, + 581 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 161, + 570, + 168, + 579 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 568, + 203, + 581 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 203, + 569, + 217, + 579 + ], + "score": 0.79, + "content": "\\nabla _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 568, + 244, + 581 + ], + "score": 1.0, + "content": "Ltotal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 578, + 153, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 153, + 589 + ], + "score": 1.0, + "content": "11: end for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 587, + 161, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 149, + 600 + ], + "score": 1.0, + "content": "12: return", + "type": "text" + }, + { + "bbox": [ + 149, + 589, + 157, + 599 + ], + "score": 0.78, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 587, + 161, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 108, + 619, + 200, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 201, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 201, + 634 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 644, + 230, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 231, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 231, + 657 + ], + "score": 1.0, + "content": "4.1 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Datasets. Two widely-used FSL datasets are selected: miniImageNet (Vinyals et al., 2016) and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 677, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 688 + ], + "score": 1.0, + "content": "tieredImageNet (Ren et al., 2018). The first dataset consists of a total number of 100 classes (600", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "images per class) and the train/validation/test split is set to 64/16/20 classes as in (Ravi & Larochelle,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "2017). The second dataset is a larger dataset including 608 classes totally (nearly 1,200 images", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "per class), which is split into 351/97/160 classes for train/validation/test. Both datasets are subsets", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 721, + 316, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 316, + 733 + ], + "score": 1.0, + "content": "sampled from ImageNet (Russakovsky et al., 2015).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 186, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 187, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 187, + 96 + ], + "score": 1.0, + "content": "3.4 TOTAL LOSS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 504, + 115 + ], + "score": 1.0, + "content": "The total training loss for our full model consists of the self-supervised losses from the pretext tasks", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 397, + 127 + ], + "score": 1.0, + "content": "and the supervised losses from the FSL tasks. In this work, in addition to", + "type": "text" + }, + { + "bbox": [ + 397, + 114, + 424, + 126 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 113, + 506, + 127 + ], + "score": 1.0, + "content": "in Eq. (10), another", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 191, + 138 + ], + "score": 1.0, + "content": "supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 192, + 126, + 213, + 136 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 125, + 328, + 138 + ], + "score": 1.0, + "content": "is also used (see Figure 1).", + "type": "text" + }, + { + "bbox": [ + 329, + 126, + 351, + 137 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 125, + 505, + 138 + ], + "score": 1.0, + "content": "is the average FSL classification loss", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 341, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 341, + 148 + ], + "score": 1.0, + "content": "over the extended episodes. Formally, it can be written as:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 103, + 506, + 148 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 153, + 369, + 188 + ], + "lines": [ + { + "bbox": [ + 242, + 153, + 369, + 188 + ], + "spans": [ + { + "bbox": [ + 242, + 153, + 369, + 188 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { a u x } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\mathcal { L } _ { f s l } ( S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } )", + "type": "interline_equation", + "image_path": "2f4093022e30a93dbeec9c5c4b2ce4413b4beb43881363ab8e48d87496b558e5.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 153, + 369, + 170.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 242, + 170.5, + 369, + 188.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 415, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 416, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 203, + 207 + ], + "score": 1.0, + "content": "Therefore, the total loss", + "type": "text" + }, + { + "bbox": [ + 203, + 194, + 228, + 204 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { t o t a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 192, + 416, + 207 + ], + "score": 1.0, + "content": "for training our full model is given as follows:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 192, + 416, + 207 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 210, + 424, + 253 + ], + "lines": [ + { + "bbox": [ + 187, + 210, + 424, + 253 + ], + "spans": [ + { + "bbox": [ + 187, + 210, + 424, + 253 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\mathcal { L } _ { t o t a l } = \\underbrace { \\overbrace { w _ { 1 } * \\mathcal { L } _ { i n s t } } ^ { i n s t a n c e { - } l e v e l } + \\overbrace { w _ { 2 } * \\mathcal { L } _ { e p i s } } ^ { e p i s o d e { - } l e v e l } } _ { \\mathrm { s e l f - s u p e r v i s e d l o s s } } + \\underbrace { w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t e g } } _ { \\mathrm { s u p e r v i s e d l o s s } } , } \\end{array}", + "type": "interline_equation", + "image_path": "7428972404be365c998fd96b04e5ae4729a396ac2be95422a01502c2b04b28b3.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 187, + 210, + 424, + 224.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 224.33333333333334, + 424, + 238.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 238.66666666666669, + 424, + 253.00000000000003 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 258, + 326, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 256, + 328, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 133, + 274 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 261, + 178, + 270 + ], + "score": 0.9, + "content": "w _ { 1 } , w _ { 2 } , w _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 256, + 328, + 274 + ], + "score": 1.0, + "content": "are the loss weight hyperparameters.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 256, + 328, + 274 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 284, + 181, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 182, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 182, + 297 + ], + "score": 1.0, + "content": "3.5 INFERENCE", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 367, + 318 + ], + "score": 1.0, + "content": "During the test stage, we only exploit the integrated representation", + "type": "text" + }, + { + "bbox": [ + 368, + 306, + 394, + 317 + ], + "score": 0.92, + "content": "F _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "for the final FSL prediction.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 315, + 397, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 221, + 330 + ], + "score": 1.0, + "content": "The predicted class label for", + "type": "text" + }, + { + "bbox": [ + 222, + 317, + 256, + 327 + ], + "score": 0.92, + "content": "x _ { i } \\in \\mathcal { Q } _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 315, + 397, + 330 + ], + "score": 1.0, + "content": "can be computed with Eq. (10) as:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 304, + 506, + 330 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 216, + 333, + 396, + 367 + ], + "lines": [ + { + "bbox": [ + 216, + 333, + 396, + 367 + ], + "spans": [ + { + "bbox": [ + 216, + 333, + 396, + 367 + ], + "score": 0.94, + "content": "y _ { i } ^ { p r e d } = \\underset { y \\in \\mathcal { C } _ { e } } { \\operatorname { a r g m a x } } \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) } .", + "type": "interline_equation", + "image_path": "2e94de396b91c16c863f7b31f20db819e61061ae5d0a9119814439b597279b54.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 216, + 333, + 396, + 350.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 216, + 350.0, + 396, + 367.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 379, + 237, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 237, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 237, + 391 + ], + "score": 1.0, + "content": "3.6 FULL IEPT ALGORITHM", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 105, + 400, + 505, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "For easy reproduction, we present the full algorithm for FSL with IEPT in Algorithm 1. Once learned,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 412, + 449, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 172, + 424 + ], + "score": 1.0, + "content": "with the learned", + "type": "text" + }, + { + "bbox": [ + 173, + 412, + 181, + 423 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 412, + 449, + 424 + ], + "score": 1.0, + "content": ", we can perform the inference over the test episodes with Eq. (13).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 400, + 506, + 424 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 436, + 222, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 224, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 224, + 448 + ], + "score": 1.0, + "content": "Algorithm 1 FSL with IEPT", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "index", + "bbox": [ + 107, + 449, + 426, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 304, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 194, + 461 + ], + "score": 1.0, + "content": "Input: The training set", + "type": "text" + }, + { + "bbox": [ + 194, + 449, + 206, + 459 + ], + "score": 0.87, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 448, + 297, + 461 + ], + "score": 1.0, + "content": ", the rotation operator set", + "type": "text" + }, + { + "bbox": [ + 297, + 450, + 304, + 459 + ], + "score": 0.67, + "content": "\\mathcal { G }", + "type": "inline_equation" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 134, + 458, + 298, + 472 + ], + "spans": [ + { + "bbox": [ + 134, + 458, + 256, + 472 + ], + "score": 1.0, + "content": "The loss weight hyperparameters", + "type": "text" + }, + { + "bbox": [ + 256, + 461, + 298, + 469 + ], + "score": 0.77, + "content": "w _ { 1 } , w _ { 2 } , w _ { 3 }", + "type": "inline_equation" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 107, + 468, + 194, + 480 + ], + "spans": [ + { + "bbox": [ + 107, + 468, + 186, + 480 + ], + "score": 1.0, + "content": "Output: The learned", + "type": "text" + }, + { + "bbox": [ + 187, + 470, + 194, + 480 + ], + "score": 0.8, + "content": "\\psi", + "type": "inline_equation" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 108, + 478, + 356, + 491 + ], + "spans": [ + { + "bbox": [ + 108, + 478, + 282, + 491 + ], + "score": 1.0, + "content": "1: Randomly initialize all learnable parameters", + "type": "text" + }, + { + "bbox": [ + 283, + 479, + 356, + 490 + ], + "score": 0.88, + "content": "\\psi = \\{ \\phi , \\theta _ { r o t } , \\theta _ { i n t } \\}", + "type": "inline_equation" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 489, + 256, + 499 + ], + "spans": [ + { + "bbox": [ + 110, + 489, + 166, + 499 + ], + "score": 1.0, + "content": "2: for iteration", + "type": "text" + }, + { + "bbox": [ + 167, + 489, + 180, + 498 + ], + "score": 0.81, + "content": "= 1", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 489, + 256, + 499 + ], + "score": 1.0, + "content": ", ..., MaxIteration do", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 498, + 273, + 510 + ], + "spans": [ + { + "bbox": [ + 110, + 498, + 120, + 510 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 131, + 498, + 228, + 510 + ], + "score": 1.0, + "content": "Randomly sample episode", + "type": "text" + }, + { + "bbox": [ + 229, + 499, + 240, + 509 + ], + "score": 0.87, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 498, + 261, + 510 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 262, + 500, + 273, + 509 + ], + "score": 0.86, + "content": "\\mathcal { D } _ { s }", + "type": "inline_equation" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 508, + 340, + 521 + ], + "spans": [ + { + "bbox": [ + 110, + 509, + 120, + 520 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 131, + 508, + 269, + 521 + ], + "score": 1.0, + "content": "Generate the set of extended episodes", + "type": "text" + }, + { + "bbox": [ + 269, + 510, + 278, + 518 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 508, + 298, + 521 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 298, + 509, + 310, + 519 + ], + "score": 0.85, + "content": "E _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 508, + 333, + 521 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 333, + 509, + 340, + 519 + ], + "score": 0.73, + "content": "\\mathcal { G }", + "type": "inline_equation" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 518, + 403, + 531 + ], + "spans": [ + { + "bbox": [ + 110, + 519, + 120, + 530 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 131, + 518, + 214, + 531 + ], + "score": 1.0, + "content": "Compute the SSL loss", + "type": "text" + }, + { + "bbox": [ + 214, + 519, + 235, + 529 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 518, + 403, + 531 + ], + "score": 1.0, + "content": "for the instance-level pretext task with Eq. (3)", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 528, + 402, + 541 + ], + "spans": [ + { + "bbox": [ + 110, + 529, + 119, + 539 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 131, + 528, + 214, + 541 + ], + "score": 1.0, + "content": "Compute the SSL loss", + "type": "text" + }, + { + "bbox": [ + 214, + 530, + 235, + 540 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 528, + 402, + 541 + ], + "score": 1.0, + "content": "for the episode-level pretext task with Eq. (6)", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 538, + 425, + 550 + ], + "spans": [ + { + "bbox": [ + 110, + 538, + 120, + 550 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 131, + 539, + 254, + 550 + ], + "score": 1.0, + "content": "Compute the supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 255, + 539, + 275, + 549 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 539, + 425, + 550 + ], + "score": 1.0, + "content": "over the extended episodes with Eq. (11)", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 548, + 425, + 560 + ], + "spans": [ + { + "bbox": [ + 109, + 548, + 120, + 559 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 131, + 548, + 254, + 560 + ], + "score": 1.0, + "content": "Compute the supervised FSL loss", + "type": "text" + }, + { + "bbox": [ + 255, + 549, + 280, + 559 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 548, + 425, + 560 + ], + "score": 1.0, + "content": "for the integrated episode with Eq. (10)", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 558, + 346, + 572 + ], + "spans": [ + { + "bbox": [ + 109, + 559, + 120, + 569 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 135, + 559, + 335, + 570 + ], + "score": 0.27, + "content": "\\mathcal { L } _ { t o t a l } = w _ { 1 } * \\mathcal { L } _ { i n s t } + w _ { 2 } * \\mathcal { L } _ { e p i s } + w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 558, + 346, + 572 + ], + "score": 1.0, + "content": "teg", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 568, + 244, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 120, + 579 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 131, + 568, + 160, + 581 + ], + "score": 1.0, + "content": "Update", + "type": "text" + }, + { + "bbox": [ + 161, + 570, + 168, + 579 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 568, + 203, + 581 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 203, + 569, + 217, + 579 + ], + "score": 0.79, + "content": "\\nabla _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 568, + 244, + 581 + ], + "score": 1.0, + "content": "Ltotal", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 578, + 153, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 153, + 589 + ], + "score": 1.0, + "content": "11: end for", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 587, + 161, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 149, + 600 + ], + "score": 1.0, + "content": "12: return", + "type": "text" + }, + { + "bbox": [ + 149, + 589, + 157, + 599 + ], + "score": 0.78, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 587, + 161, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + } + ], + "index": 28, + "bbox_fs": [ + 105, + 448, + 425, + 600 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 619, + 200, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 201, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 201, + 634 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 644, + 230, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 231, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 231, + 657 + ], + "score": 1.0, + "content": "4.1 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Datasets. Two widely-used FSL datasets are selected: miniImageNet (Vinyals et al., 2016) and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 677, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 688 + ], + "score": 1.0, + "content": "tieredImageNet (Ren et al., 2018). The first dataset consists of a total number of 100 classes (600", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "images per class) and the train/validation/test split is set to 64/16/20 classes as in (Ravi & Larochelle,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "2017). The second dataset is a larger dataset including 608 classes totally (nearly 1,200 images", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "per class), which is split into 351/97/160 classes for train/validation/test. Both datasets are subsets", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 721, + 316, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 316, + 733 + ], + "score": 1.0, + "content": "sampled from ImageNet (Russakovsky et al., 2015).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 666, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 112, + 504, + 449 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 89, + 501, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 501, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 491, + 102 + ], + "score": 1.0, + "content": "Table 1: Comparative results for 5-way 1/5-shot FSL. The mean classification accuracies (top-1,", + "type": "text" + }, + { + "bbox": [ + 491, + 90, + 501, + 100 + ], + "score": 0.42, + "content": "\\%", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 493, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 140, + 113 + ], + "score": 1.0, + "content": "with the", + "type": "text" + }, + { + "bbox": [ + 141, + 100, + 160, + 110 + ], + "score": 0.84, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 99, + 297, + 113 + ], + "score": 1.0, + "content": "confidence intervals are reported.", + "type": "text" + }, + { + "bbox": [ + 297, + 101, + 303, + 111 + ], + "score": 0.72, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 99, + 493, + 113 + ], + "score": 1.0, + "content": "indicates the result is reproduced by ourselves.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 112, + 504, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 112, + 504, + 449 + ], + "spans": [ + { + "bbox": [ + 109, + 112, + 504, + 449 + ], + "score": 0.983, + "html": "
MethodminiImageNettieredImageNet
Backbone1-shot5-shot1-shot5-shot
MatchingNet (Vinyals etal.,2016)Conv4-6443.56± 0.8455.31± 0.73
ProtoNet† (Snell et al., 2017)Conv4-6452.61 ± 0.5271.33 ± 0.4153.33 ± 0.5072.10 ± 0.41
MAML (Finn et al., 2017)Conv4-6448.70 ± 1.8463.10 ±0.9251.67 ± 1.8170.30 ±0.08
Relation Net (Sung et al.,2018)Conv4-6450.40 ± 0.8065.30 ±0.7054.48 ± 0.9371.32 ± 0.78
IMP† (Allen et al., 2019)Conv4-6452.91 ± 0.4971.57 ± 0.4253.63 ± 0.5171.89 ± 0.44
DN4 (Li et al., 2019b)Conv4-6451.24 ± 0.7471.02 ± 0.641
DN PARN (Wu et al., 2019)Conv4-6455.22 ± 0.8471.55 ± 0.66
PN+rot (Gidaris et al., 2019)Conv4-6453.63 ± 0.4371.70 ± 0.36
CC+rot (Gidaris et al., 2019)Conv4-6454.83 ± 0.4371.86 ± 0.33
DSN-MR (Simon et al., 2020)Conv4-6455.88 ± 0.9070.50±0.68
Centroid (Afrasiyabi et al., 2020)Conv4-6453.14 ± 1.0671.45 ± 0.72
Neg-Cosine (Liu et al., 2020)Conv4-6452.84 ±0.7670.41 ± 0.66一
IEPT (ours)Conv4-6456.26± 0.4573.91±0.34 58.25±0.48 75.63±0.46
ProtoNet† (Snell et al.,2017)Conv4-51253.25± 0.4473.15 ± 0.3557.88 ±0.5076.82±0.40
MAML (Finn et al., 2017)Conv4-51249.33 ±0.6065.17 ± 0.4952.84± 0.5670.91 ± 0.46
Relation Net (Sung et al.,2018)Conv4-51250.86 ± 0.5767.32 ± 0.4454.69 ±0.5972.71 ±0.43
PN+rot (Gidaris et al.,2019)Conv4-51256.02 ±0.4674.00 ± 0.35一
CC+rot (Gidaris et al., 2019)Conv4-51256.27 ± 0.4374.30 ± 0.33一
IEPT (ours)Conv4-51258.43±0.4675.07±0.3360.91±0.5979.61±0.45
ProtoNet† (Snell et al.,2017)ResNet-1262.39 ±0.5180.53±0.4268.23±0.5084.03 ± 0.41
TADAM (Oreshkin et al., 2018)ResNet-1258.50 ± 0.3076.70 ±0.38一
MetaOptNet (Lee et al.,2019)ResNet-1262.64± 0.6178.63 ± 0.4665.99 ± 0.7281.56 ± 0.63
MTL (Sun et al., 2019)ResNet-1261.20 ± 1.8075.50 ± 0.8065.62 ± 1.8080.61 ±0.90
CAN (Hou et al., 2019)ResNet-1263.85 ±0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
AM3 (Xing et al., 2019)ResNet-1265.21 ± 0.4975.20 ±0.3667.23 ± 0.3478.95 ± 0.22
Shot-Free (Ravichandran et al.,2019)ResNet-1259.04 ± 0.4377.64 ± 0.3966.87 ± 0.4382.64 ±0.43
Neg-Cosine (Liu et al., 2020)ResNet-1263.85 ± 0.8181.57 ± 0.56
Distill (Tian et al., 2020)ResNet-1264.82 ±0.6082.14 ± 0.4371.52 ± 0.6986.03 ±0.49
DSN-MR (Simon et al.,2020)ResNet-1264.60 ±0.7279.51± 0.5067.39 ± 0.8282.85 ±0.56
DeepEMD (Zhang et al.,2020)ResNet-1265.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
FEAT (Ye et al.,2020)ResNet-1266.78±0.2082.05 ± 0.1470.80 ±0.2384.79 ± 0.16
ProtoNet+Rotation (Su et al.,2020)ResNet-1876.00±0.6078.90±0.70
IEPT (ours)ResNet-1267.05±0.44882.90±0.30)72.24±0.50 86.73±0.34
", + "type": "table", + "image_path": "1ab2ff5c5ab953a64fe5c51c547bda001032e318953029864dd4b1643d753419.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 109, + 112, + 504, + 224.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 224.33333333333331, + 504, + 336.66666666666663 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 336.66666666666663, + 504, + 448.99999999999994 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "Feature Extractors. For fair comparison with published results, our IEPT adopts three widely-used", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 474, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 486 + ], + "score": 1.0, + "content": "feature extractors: Conv4-64 (Vinyals et al., 2016), Conv4-512, and ResNet-12 (He et al., 2016a).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "score": 1.0, + "content": "Particularly, Conv4-512 is almost the same as Conv4-64 except having a different channel size of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "last convolution layer. To speed up the training process, as in many previous works (Ye et al., 2020;", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "Zhang et al., 2020; Simon et al., 2020), we pretrain all the feature extractors on the training split of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 515, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 532 + ], + "score": 1.0, + "content": "each dataset for our IEPT. Following (He et al., 2016a), we use the temperature scaling skill during", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 344, + 541 + ], + "score": 1.0, + "content": "the training phase. On both datasets, the input image size is", + "type": "text" + }, + { + "bbox": [ + 344, + 528, + 378, + 539 + ], + "score": 0.9, + "content": "8 4 \\times 8 4", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 528, + 505, + 541 + ], + "score": 1.0, + "content": ". The output feature dimensions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 538, + 411, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 411, + 552 + ], + "score": 1.0, + "content": "of Conv4-64, Conv4-512, and ResNet-12 are 64, 512, and 640, respectively.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "Evaluation Metrics. We take the 5-way 5-shot (or 1-shot) FSL evaluation setting, as in previous", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "works. We randomly sample 2,000 episodes from the test split and report the mean classification", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 173, + 591 + ], + "score": 1.0, + "content": "accuracy (top-1,", + "type": "text" + }, + { + "bbox": [ + 174, + 579, + 183, + 588 + ], + "score": 0.66, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 578, + 243, + 591 + ], + "score": 1.0, + "content": ") as well as the", + "type": "text" + }, + { + "bbox": [ + 244, + 578, + 263, + 589 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "confidence interval. Since the integration transformer copes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 590, + 473, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 473, + 601 + ], + "score": 1.0, + "content": "with each sample independently, we take a strict non-transductive setting during evaluation.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "Implementation Details. PyTorch is used for our implementation. We utilize the Adam optimizer", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "score": 1.0, + "content": "(Kingma & Ba, 2015) for Conv4-64 & Conv4-512 and the SGD optimizer for ResNet-12 to train our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "IEPT model. The hyperparameters of our IEPT model are selected according to the performance on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 639, + 307, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 307, + 651 + ], + "score": 1.0, + "content": "the validation split.We will release the code soon.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 107, + 666, + 198, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 199, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 199, + 679 + ], + "score": 1.0, + "content": "4.2 MAIN RESULTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Comparison to State-of-the-Arts. We compare our IEPT with two groups of baselines: (1) Recent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020); (2) Representative/latest FSL methods", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "(w/o SSL) (Snell et al., 2017; Finn et al., 2017; Lee et al., 2019; Ravichandran et al., 2019; Simon", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "et al., 2020; Zhang et al., 2020; Ye et al., 2020; Liu et al., 2020). The comparative results for 5-way", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 112, + 504, + 449 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 89, + 501, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 501, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 491, + 102 + ], + "score": 1.0, + "content": "Table 1: Comparative results for 5-way 1/5-shot FSL. The mean classification accuracies (top-1,", + "type": "text" + }, + { + "bbox": [ + 491, + 90, + 501, + 100 + ], + "score": 0.42, + "content": "\\%", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 493, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 140, + 113 + ], + "score": 1.0, + "content": "with the", + "type": "text" + }, + { + "bbox": [ + 141, + 100, + 160, + 110 + ], + "score": 0.84, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 99, + 297, + 113 + ], + "score": 1.0, + "content": "confidence intervals are reported.", + "type": "text" + }, + { + "bbox": [ + 297, + 101, + 303, + 111 + ], + "score": 0.72, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 99, + 493, + 113 + ], + "score": 1.0, + "content": "indicates the result is reproduced by ourselves.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 112, + 504, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 112, + 504, + 449 + ], + "spans": [ + { + "bbox": [ + 109, + 112, + 504, + 449 + ], + "score": 0.983, + "html": "
MethodminiImageNettieredImageNet
Backbone1-shot5-shot1-shot5-shot
MatchingNet (Vinyals etal.,2016)Conv4-6443.56± 0.8455.31± 0.73
ProtoNet† (Snell et al., 2017)Conv4-6452.61 ± 0.5271.33 ± 0.4153.33 ± 0.5072.10 ± 0.41
MAML (Finn et al., 2017)Conv4-6448.70 ± 1.8463.10 ±0.9251.67 ± 1.8170.30 ±0.08
Relation Net (Sung et al.,2018)Conv4-6450.40 ± 0.8065.30 ±0.7054.48 ± 0.9371.32 ± 0.78
IMP† (Allen et al., 2019)Conv4-6452.91 ± 0.4971.57 ± 0.4253.63 ± 0.5171.89 ± 0.44
DN4 (Li et al., 2019b)Conv4-6451.24 ± 0.7471.02 ± 0.641
DN PARN (Wu et al., 2019)Conv4-6455.22 ± 0.8471.55 ± 0.66
PN+rot (Gidaris et al., 2019)Conv4-6453.63 ± 0.4371.70 ± 0.36
CC+rot (Gidaris et al., 2019)Conv4-6454.83 ± 0.4371.86 ± 0.33
DSN-MR (Simon et al., 2020)Conv4-6455.88 ± 0.9070.50±0.68
Centroid (Afrasiyabi et al., 2020)Conv4-6453.14 ± 1.0671.45 ± 0.72
Neg-Cosine (Liu et al., 2020)Conv4-6452.84 ±0.7670.41 ± 0.66一
IEPT (ours)Conv4-6456.26± 0.4573.91±0.34 58.25±0.48 75.63±0.46
ProtoNet† (Snell et al.,2017)Conv4-51253.25± 0.4473.15 ± 0.3557.88 ±0.5076.82±0.40
MAML (Finn et al., 2017)Conv4-51249.33 ±0.6065.17 ± 0.4952.84± 0.5670.91 ± 0.46
Relation Net (Sung et al.,2018)Conv4-51250.86 ± 0.5767.32 ± 0.4454.69 ±0.5972.71 ±0.43
PN+rot (Gidaris et al.,2019)Conv4-51256.02 ±0.4674.00 ± 0.35一
CC+rot (Gidaris et al., 2019)Conv4-51256.27 ± 0.4374.30 ± 0.33一
IEPT (ours)Conv4-51258.43±0.4675.07±0.3360.91±0.5979.61±0.45
ProtoNet† (Snell et al.,2017)ResNet-1262.39 ±0.5180.53±0.4268.23±0.5084.03 ± 0.41
TADAM (Oreshkin et al., 2018)ResNet-1258.50 ± 0.3076.70 ±0.38一
MetaOptNet (Lee et al.,2019)ResNet-1262.64± 0.6178.63 ± 0.4665.99 ± 0.7281.56 ± 0.63
MTL (Sun et al., 2019)ResNet-1261.20 ± 1.8075.50 ± 0.8065.62 ± 1.8080.61 ±0.90
CAN (Hou et al., 2019)ResNet-1263.85 ±0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
AM3 (Xing et al., 2019)ResNet-1265.21 ± 0.4975.20 ±0.3667.23 ± 0.3478.95 ± 0.22
Shot-Free (Ravichandran et al.,2019)ResNet-1259.04 ± 0.4377.64 ± 0.3966.87 ± 0.4382.64 ±0.43
Neg-Cosine (Liu et al., 2020)ResNet-1263.85 ± 0.8181.57 ± 0.56
Distill (Tian et al., 2020)ResNet-1264.82 ±0.6082.14 ± 0.4371.52 ± 0.6986.03 ±0.49
DSN-MR (Simon et al.,2020)ResNet-1264.60 ±0.7279.51± 0.5067.39 ± 0.8282.85 ±0.56
DeepEMD (Zhang et al.,2020)ResNet-1265.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
FEAT (Ye et al.,2020)ResNet-1266.78±0.2082.05 ± 0.1470.80 ±0.2384.79 ± 0.16
ProtoNet+Rotation (Su et al.,2020)ResNet-1876.00±0.6078.90±0.70
IEPT (ours)ResNet-1267.05±0.44882.90±0.30)72.24±0.50 86.73±0.34
", + "type": "table", + "image_path": "1ab2ff5c5ab953a64fe5c51c547bda001032e318953029864dd4b1643d753419.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 109, + 112, + 504, + 224.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 224.33333333333331, + 504, + 336.66666666666663 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 336.66666666666663, + 504, + 448.99999999999994 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "Feature Extractors. For fair comparison with published results, our IEPT adopts three widely-used", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 474, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 486 + ], + "score": 1.0, + "content": "feature extractors: Conv4-64 (Vinyals et al., 2016), Conv4-512, and ResNet-12 (He et al., 2016a).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "score": 1.0, + "content": "Particularly, Conv4-512 is almost the same as Conv4-64 except having a different channel size of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "last convolution layer. To speed up the training process, as in many previous works (Ye et al., 2020;", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "Zhang et al., 2020; Simon et al., 2020), we pretrain all the feature extractors on the training split of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 515, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 532 + ], + "score": 1.0, + "content": "each dataset for our IEPT. Following (He et al., 2016a), we use the temperature scaling skill during", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 344, + 541 + ], + "score": 1.0, + "content": "the training phase. On both datasets, the input image size is", + "type": "text" + }, + { + "bbox": [ + 344, + 528, + 378, + 539 + ], + "score": 0.9, + "content": "8 4 \\times 8 4", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 528, + 505, + 541 + ], + "score": 1.0, + "content": ". The output feature dimensions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 538, + 411, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 411, + 552 + ], + "score": 1.0, + "content": "of Conv4-64, Conv4-512, and ResNet-12 are 64, 512, and 640, respectively.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 463, + 507, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "Evaluation Metrics. We take the 5-way 5-shot (or 1-shot) FSL evaluation setting, as in previous", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "works. We randomly sample 2,000 episodes from the test split and report the mean classification", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 173, + 591 + ], + "score": 1.0, + "content": "accuracy (top-1,", + "type": "text" + }, + { + "bbox": [ + 174, + 579, + 183, + 588 + ], + "score": 0.66, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 578, + 243, + 591 + ], + "score": 1.0, + "content": ") as well as the", + "type": "text" + }, + { + "bbox": [ + 244, + 578, + 263, + 589 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "confidence interval. Since the integration transformer copes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 590, + 473, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 473, + 601 + ], + "score": 1.0, + "content": "with each sample independently, we take a strict non-transductive setting during evaluation.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 556, + 505, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "Implementation Details. PyTorch is used for our implementation. We utilize the Adam optimizer", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 630 + ], + "score": 1.0, + "content": "(Kingma & Ba, 2015) for Conv4-64 & Conv4-512 and the SGD optimizer for ResNet-12 to train our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "IEPT model. The hyperparameters of our IEPT model are selected according to the performance on", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 639, + 307, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 307, + 651 + ], + "score": 1.0, + "content": "the validation split.We will release the code soon.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 605, + 506, + 651 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 666, + 198, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 199, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 199, + 679 + ], + "score": 1.0, + "content": "4.2 MAIN RESULTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Comparison to State-of-the-Arts. We compare our IEPT with two groups of baselines: (1) Recent", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020); (2) Representative/latest FSL methods", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "(w/o SSL) (Snell et al., 2017; Finn et al., 2017; Lee et al., 2019; Ravichandran et al., 2019; Simon", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "et al., 2020; Zhang et al., 2020; Ye et al., 2020; Liu et al., 2020). The comparative results for 5-way", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 687, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 124, + 502, + 196 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 123 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 506, + 102 + ], + "score": 1.0, + "content": "Table 2: Ablation study results for our full IEPT model over miniImageNet and tieredImageNet. Our", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 507, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 312, + 113 + ], + "score": 1.0, + "content": "full model includes two self-supervised losses (i.e.", + "type": "text" + }, + { + "bbox": [ + 313, + 100, + 336, + 113 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 99, + 354, + 113 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 354, + 100, + 378, + 111 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { i n s t . }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 99, + 507, + 113 + ], + "score": 1.0, + "content": ") and two supervised losses (i.e.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 111, + 349, + 124 + ], + "spans": [ + { + "bbox": [ + 107, + 111, + 128, + 123 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 111, + 146, + 124 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 147, + 111, + 174, + 124 + ], + "score": 0.87, + "content": "L _ { i n t e g . }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 111, + 349, + 124 + ], + "score": 1.0, + "content": "). Conv4-64 is used as the feature extractor.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 109, + 124, + 502, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 124, + 502, + 196 + ], + "spans": [ + { + "bbox": [ + 109, + 124, + 502, + 196 + ], + "score": 0.732, + "html": "
miniImageNettieredImageNet
LintegLinstLepisLaux1-shot5-shot1-shot5-shot
√55.04 ± 0.5272.01 ± 0.4156.98 ± 0.4774.15 ± 0.51
√√55.49 ± 0.5672.54± 0.4657.41 ± 0.5174.65 ± 0.50
√55.88± 0.4372.97 ± 0.4057.76± 0.4575.06 ± 0.40
√55.97 ± 0.5773.28 ± 0.3957.83 ± 0.5575.22 ± 0.48
√?>>>√56.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "type": "table", + "image_path": "782095c5043ea0f9182610faf183276e533078eb62759726e4d5284f243230e1.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 109, + 124, + 502, + 148.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 148.0, + 502, + 172.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 172.0, + 502, + 196.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 115, + 207, + 493, + 313 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 207, + 493, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 207, + 493, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 493, + 313 + ], + "score": 0.798, + "type": "image", + "image_path": "af3c8b4f8135fe31235a01a5a12b5a61f1267d2c4e0e0affdba8e3fca164fe3e.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 115, + 207, + 493, + 242.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 242.33333333333334, + 493, + 277.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 277.6666666666667, + 493, + 313.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 316, + 506, + 381 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 315, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 327 + ], + "score": 1.0, + "content": "Figure 2: (a) Comparison among different combination methods over episodes for FSL with self-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 327, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 158, + 339 + ], + "score": 1.0, + "content": "supervision.", + "type": "text" + }, + { + "bbox": [ + 159, + 327, + 172, + 338 + ], + "score": 0.33, + "content": "\\mathbf { ( b ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 327, + 361, + 339 + ], + "score": 1.0, + "content": "Illustration of the effect of different choices of", + "type": "text" + }, + { + "bbox": [ + 361, + 327, + 370, + 337 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 327, + 506, + 339 + ], + "score": 1.0, + "content": "on the performance of our model", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 110, + 338, + 119, + 348 + ], + "score": 0.63, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "denotes the number of extended episodes used for SSL). (c) Comparative results obtained by our", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 349, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 360 + ], + "score": 1.0, + "content": "IEPT using different basic FSL classifiers (i.e. ProtoNet, FEAT, and IMP). It can be seen clearly that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 359, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 373 + ], + "score": 1.0, + "content": "integrated episode-based fusion leads to more separation between classes. All figures present 5-way", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 370, + 425, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 425, + 383 + ], + "score": 1.0, + "content": "1-shot/5-shot results on miniImageNet, using Conv4-64 as the feature extractor.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "1/5-shot FSL are shown in Table 1. We have the following observations: (1) When compared with the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 404, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 416 + ], + "score": 1.0, + "content": "representative/latest FSL methods (w/o SSL), our IEPT achieves the best performance on all datasets", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "and under all settings, validating the effectiveness of SSL with IEPT for FSL. (2) Our IEPT also", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "clearly outperforms the two SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020) which only", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "use instance-level pretext tasks, demonstrating the importance of closer/episode-level integration of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 461, + 460 + ], + "score": 1.0, + "content": "SSL into FSL. (3) The improvements achieved by our IEPT over ProtoNet range from", + "type": "text" + }, + { + "bbox": [ + 461, + 448, + 476, + 458 + ], + "score": 0.86, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 447, + 487, + 460 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 488, + 448, + 503, + 458 + ], + "score": 0.87, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 447, + 506, + 460 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Since our IEPT takes ProtoNet as the baseline, the obtained margins provide direct evidence that SSL", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "brings significant benefits to FSL. Note that our IEPT is also shown to be effective under both the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 480, + 403, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 403, + 493 + ], + "score": 1.0, + "content": "fine-grained FSL and cross-domain FSL settings in Sec. 4.3 (see Table 3).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 511 + ], + "score": 1.0, + "content": "Ablation Study. Our full IEPT model is trained with four losses (see Eq. (12)), including two", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 508, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 407, + 522 + ], + "score": 1.0, + "content": "self-supervised losses and two supervised losses: the episode-level SSL loss", + "type": "text" + }, + { + "bbox": [ + 407, + 509, + 430, + 520 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 508, + 506, + 522 + ], + "score": 1.0, + "content": ", the instance-level", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 519, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 144, + 534 + ], + "score": 1.0, + "content": "SSL loss", + "type": "text" + }, + { + "bbox": [ + 144, + 520, + 167, + 531 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 519, + 261, + 534 + ], + "score": 1.0, + "content": ", the auxiliary FSL loss", + "type": "text" + }, + { + "bbox": [ + 261, + 520, + 283, + 531 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 519, + 395, + 534 + ], + "score": 1.0, + "content": "and the integrated FSL loss", + "type": "text" + }, + { + "bbox": [ + 395, + 520, + 422, + 532 + ], + "score": 0.9, + "content": "\\bar { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 519, + 506, + 534 + ], + "score": 1.0, + "content": ". To demonstrate the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "contribution of each loss, we present the ablation study results for our full IEPT model in Table 2,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 338, + 554 + ], + "score": 1.0, + "content": "where Conv4-64 is used as the backbone. We start with", + "type": "text" + }, + { + "bbox": [ + 339, + 542, + 366, + 554 + ], + "score": 0.91, + "content": "L _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "and then add the additional three", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "losses successively. It can be observed that the performance of our model continuously increases", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 563, + 463, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 463, + 576 + ], + "score": 1.0, + "content": "when more losses are used, indicating that each loss contributes to the final performance.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 590, + 235, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 236, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 236, + 602 + ], + "score": 1.0, + "content": "4.3 FURTHER EVALUATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "Different Combination Methods over Episodes. We have introduced a transformer-based attention", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "module to fuse the features of each instance from all extended episodes (and an integrated episode", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "can be obtained) for the supervised classification task (see Sec. 3.3). In this experiment, we compare", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "it with two alternative ways of across-episode integration: (1) Averaging extended episodes: the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "extended episodes are directly fused for FSL classification; (2) Averaging all episodes: the extended", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "episodes as well as the integrated episode are fused for FSL classification. We present the comparative", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "results on miniImageNet in Figure 2(a). For comprehensive comparison, the results of FSL with each", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 491, + 700 + ], + "score": 1.0, + "content": "single extended episode are also reported. We can observe that: (1) The performance of ‘Episode", + "type": "text" + }, + { + "bbox": [ + 491, + 688, + 502, + 698 + ], + "score": 0.78, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "’", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "is the highest among the four baselines (i.e., FSL with single extended episode), perhaps because the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "feature extractor is pretrained on the original images without rotation transformations. (2) FSL by", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "averaging extended episodes (i.e., ‘Averaging extended episodes’) indeed improves each of the four", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 124, + 502, + 196 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 123 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 506, + 102 + ], + "score": 1.0, + "content": "Table 2: Ablation study results for our full IEPT model over miniImageNet and tieredImageNet. Our", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 507, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 312, + 113 + ], + "score": 1.0, + "content": "full model includes two self-supervised losses (i.e.", + "type": "text" + }, + { + "bbox": [ + 313, + 100, + 336, + 113 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 99, + 354, + 113 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 354, + 100, + 378, + 111 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { i n s t . }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 99, + 507, + 113 + ], + "score": 1.0, + "content": ") and two supervised losses (i.e.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 111, + 349, + 124 + ], + "spans": [ + { + "bbox": [ + 107, + 111, + 128, + 123 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 111, + 146, + 124 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 147, + 111, + 174, + 124 + ], + "score": 0.87, + "content": "L _ { i n t e g . }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 111, + 349, + 124 + ], + "score": 1.0, + "content": "). Conv4-64 is used as the feature extractor.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 109, + 124, + 502, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 124, + 502, + 196 + ], + "spans": [ + { + "bbox": [ + 109, + 124, + 502, + 196 + ], + "score": 0.732, + "html": "
miniImageNettieredImageNet
LintegLinstLepisLaux1-shot5-shot1-shot5-shot
√55.04 ± 0.5272.01 ± 0.4156.98 ± 0.4774.15 ± 0.51
√√55.49 ± 0.5672.54± 0.4657.41 ± 0.5174.65 ± 0.50
√55.88± 0.4372.97 ± 0.4057.76± 0.4575.06 ± 0.40
√55.97 ± 0.5773.28 ± 0.3957.83 ± 0.5575.22 ± 0.48
√?>>>√56.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "type": "table", + "image_path": "782095c5043ea0f9182610faf183276e533078eb62759726e4d5284f243230e1.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 109, + 124, + 502, + 148.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 148.0, + 502, + 172.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 172.0, + 502, + 196.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 115, + 207, + 493, + 313 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 207, + 493, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 207, + 493, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 493, + 313 + ], + "score": 0.798, + "type": "image", + "image_path": "af3c8b4f8135fe31235a01a5a12b5a61f1267d2c4e0e0affdba8e3fca164fe3e.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 115, + 207, + 493, + 242.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 242.33333333333334, + 493, + 277.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 277.6666666666667, + 493, + 313.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 316, + 506, + 381 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 315, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 327 + ], + "score": 1.0, + "content": "Figure 2: (a) Comparison among different combination methods over episodes for FSL with self-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 327, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 158, + 339 + ], + "score": 1.0, + "content": "supervision.", + "type": "text" + }, + { + "bbox": [ + 159, + 327, + 172, + 338 + ], + "score": 0.33, + "content": "\\mathbf { ( b ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 327, + 361, + 339 + ], + "score": 1.0, + "content": "Illustration of the effect of different choices of", + "type": "text" + }, + { + "bbox": [ + 361, + 327, + 370, + 337 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 327, + 506, + 339 + ], + "score": 1.0, + "content": "on the performance of our model", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 110, + 338, + 119, + 348 + ], + "score": 0.63, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "denotes the number of extended episodes used for SSL). (c) Comparative results obtained by our", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 349, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 360 + ], + "score": 1.0, + "content": "IEPT using different basic FSL classifiers (i.e. ProtoNet, FEAT, and IMP). It can be seen clearly that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 359, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 373 + ], + "score": 1.0, + "content": "integrated episode-based fusion leads to more separation between classes. All figures present 5-way", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 370, + 425, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 425, + 383 + ], + "score": 1.0, + "content": "1-shot/5-shot results on miniImageNet, using Conv4-64 as the feature extractor.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "1/5-shot FSL are shown in Table 1. We have the following observations: (1) When compared with the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 404, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 416 + ], + "score": 1.0, + "content": "representative/latest FSL methods (w/o SSL), our IEPT achieves the best performance on all datasets", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "and under all settings, validating the effectiveness of SSL with IEPT for FSL. (2) Our IEPT also", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "clearly outperforms the two SSL-based FSL methods (Gidaris et al., 2019; Su et al., 2020) which only", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "use instance-level pretext tasks, demonstrating the importance of closer/episode-level integration of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 461, + 460 + ], + "score": 1.0, + "content": "SSL into FSL. (3) The improvements achieved by our IEPT over ProtoNet range from", + "type": "text" + }, + { + "bbox": [ + 461, + 448, + 476, + 458 + ], + "score": 0.86, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 447, + 487, + 460 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 488, + 448, + 503, + 458 + ], + "score": 0.87, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 447, + 506, + 460 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "Since our IEPT takes ProtoNet as the baseline, the obtained margins provide direct evidence that SSL", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "brings significant benefits to FSL. Note that our IEPT is also shown to be effective under both the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 480, + 403, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 403, + 493 + ], + "score": 1.0, + "content": "fine-grained FSL and cross-domain FSL settings in Sec. 4.3 (see Table 3).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 393, + 506, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 511 + ], + "score": 1.0, + "content": "Ablation Study. Our full IEPT model is trained with four losses (see Eq. (12)), including two", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 508, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 407, + 522 + ], + "score": 1.0, + "content": "self-supervised losses and two supervised losses: the episode-level SSL loss", + "type": "text" + }, + { + "bbox": [ + 407, + 509, + 430, + 520 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 508, + 506, + 522 + ], + "score": 1.0, + "content": ", the instance-level", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 519, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 144, + 534 + ], + "score": 1.0, + "content": "SSL loss", + "type": "text" + }, + { + "bbox": [ + 144, + 520, + 167, + 531 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 519, + 261, + 534 + ], + "score": 1.0, + "content": ", the auxiliary FSL loss", + "type": "text" + }, + { + "bbox": [ + 261, + 520, + 283, + 531 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 519, + 395, + 534 + ], + "score": 1.0, + "content": "and the integrated FSL loss", + "type": "text" + }, + { + "bbox": [ + 395, + 520, + 422, + 532 + ], + "score": 0.9, + "content": "\\bar { L } _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 519, + 506, + 534 + ], + "score": 1.0, + "content": ". To demonstrate the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "contribution of each loss, we present the ablation study results for our full IEPT model in Table 2,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 338, + 554 + ], + "score": 1.0, + "content": "where Conv4-64 is used as the backbone. We start with", + "type": "text" + }, + { + "bbox": [ + 339, + 542, + 366, + 554 + ], + "score": 0.91, + "content": "L _ { i n t e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "and then add the additional three", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "losses successively. It can be observed that the performance of our model continuously increases", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 563, + 463, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 463, + 576 + ], + "score": 1.0, + "content": "when more losses are used, indicating that each loss contributes to the final performance.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 496, + 506, + 576 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 590, + 235, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 236, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 236, + 602 + ], + "score": 1.0, + "content": "4.3 FURTHER EVALUATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "Different Combination Methods over Episodes. We have introduced a transformer-based attention", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "module to fuse the features of each instance from all extended episodes (and an integrated episode", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "can be obtained) for the supervised classification task (see Sec. 3.3). In this experiment, we compare", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "it with two alternative ways of across-episode integration: (1) Averaging extended episodes: the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "extended episodes are directly fused for FSL classification; (2) Averaging all episodes: the extended", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "episodes as well as the integrated episode are fused for FSL classification. We present the comparative", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "results on miniImageNet in Figure 2(a). For comprehensive comparison, the results of FSL with each", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 491, + 700 + ], + "score": 1.0, + "content": "single extended episode are also reported. We can observe that: (1) The performance of ‘Episode", + "type": "text" + }, + { + "bbox": [ + 491, + 688, + 502, + 698 + ], + "score": 0.78, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "’", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "is the highest among the four baselines (i.e., FSL with single extended episode), perhaps because the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "feature extractor is pretrained on the original images without rotation transformations. (2) FSL by", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "averaging extended episodes (i.e., ‘Averaging extended episodes’) indeed improves each of the four", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "baselines. (3) FSL with integrated episode (i.e., ‘Integrated episode’) is superior to FSL by simply", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "averaging extended episodes. (4) Comparing ‘Integrated episode’ with ‘Averaging all episodes’,", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "score": 1.0, + "content": "the performance of FSL with integrated episode is more stable across different settings, furthering", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "validating the usefulness of our across-episode integration. Overall, the episode-integration module", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "is indeed effective in FSL with self-supervision. This is also supported by the visualization results", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 259, + 254, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 254, + 271 + ], + "score": 1.0, + "content": "presented in Appendices A.3 & A.4.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 37, + "bbox_fs": [ + 104, + 610, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 113, + 500, + 190 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 89, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 505, + 102 + ], + "score": 1.0, + "content": "Table 3: Comparative results for the fine-grained FSL on CUB (Wah et al., 2011) and the cross-domain", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 236, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 197, + 113 + ], + "score": 1.0, + "content": "FSL on miniImageNet", + "type": "text" + }, + { + "bbox": [ + 198, + 101, + 210, + 110 + ], + "score": 0.64, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 100, + 236, + 113 + ], + "score": 1.0, + "content": "CUB.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 113, + 500, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 113, + 500, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 113, + 500, + 190 + ], + "score": 0.979, + "html": "
MethodBackboneCUBminiImageNet→CUB
1-shot5-shot1-shot5-shot
MatchingNet (Vinyals et al.,2016)Conv4-6461.16±0.8972.86±0.7042.62 ± 0.5556.53± 0.44
ProtoNet (Snell et al.,2017)Conv4-6463.72 ±0.2281.50 ± 0.1550.51 ± 0.5669.28±0.40
MAML (Finn et al.,2017)Conv4-6455.92 ± 0.9572.09 ±0.7643.59 ± 0.5454.18 ± 0.41
Relation Net (Sung et al., 2018)Conv4-6462.45 ± 0.9876.11 ± 0.6949.84 ±0.5468.98 ±0.42
FEAT(Ye et al.,2020)Conv4-6468.87 ±0.2282.90±0.1551.52 ± 0.5470.16 ± 0.40
IEPT(ours)Conv4-6469.97±0.4984.33± 0.3352.68±0.5672.98±0.40
", + "type": "table", + "image_path": "b9c462107ffd68f09bfe11e7079ebc8a088483e96df432ec01a7e10e9eb18d94.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 113, + 500, + 138.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 138.66666666666666, + 500, + 164.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 164.33333333333331, + 500, + 189.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "baselines. (3) FSL with integrated episode (i.e., ‘Integrated episode’) is superior to FSL by simply", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "averaging extended episodes. (4) Comparing ‘Integrated episode’ with ‘Averaging all episodes’,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 240 + ], + "score": 1.0, + "content": "the performance of FSL with integrated episode is more stable across different settings, furthering", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "validating the usefulness of our across-episode integration. Overall, the episode-integration module", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "is indeed effective in FSL with self-supervision. This is also supported by the visualization results", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 259, + 254, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 254, + 271 + ], + "score": 1.0, + "content": "presented in Appendices A.3 & A.4.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "Different Number of Extended Episodes. In all the above experiments, the number of the extended", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 142, + 299 + ], + "score": 1.0, + "content": "episodes", + "type": "text" + }, + { + "bbox": [ + 143, + 287, + 152, + 297 + ], + "score": 0.76, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 286, + 241, + 299 + ], + "score": 1.0, + "content": "is set to 4 (rotation by", + "type": "text" + }, + { + "bbox": [ + 241, + 287, + 252, + 297 + ], + "score": 0.81, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 286, + 255, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 256, + 287, + 271, + 297 + ], + "score": 0.83, + "content": "9 0 °", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 286, + 275, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 276, + 287, + 296, + 298 + ], + "score": 0.81, + "content": "1 8 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 286, + 299, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 300, + 287, + 321, + 298 + ], + "score": 0.75, + "content": "2 7 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "). Figure 2(b) shows the impact of the value of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 297, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 107, + 298, + 116, + 308 + ], + "score": 0.61, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 297, + 180, + 310 + ], + "score": 1.0, + "content": ". Note that when", + "type": "text" + }, + { + "bbox": [ + 181, + 298, + 208, + 308 + ], + "score": 0.91, + "content": "R = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 297, + 506, + 310 + ], + "score": 1.0, + "content": ", our IEPT model is equivalent to ProtoNet which is without self-supervision.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 308, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 407, + 321 + ], + "score": 1.0, + "content": "It can be seen that the performance of our model consistently grows when", + "type": "text" + }, + { + "bbox": [ + 407, + 309, + 416, + 318 + ], + "score": 0.75, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 308, + 507, + 321 + ], + "score": 1.0, + "content": "increases from 1 to 4.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "Additionally, the study on exploiting other pretext tasks for our IEPT is presented in Appendix A.1.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 348 + ], + "score": 1.0, + "content": "Different Basic FSL Classifiers. As mentioned in Sec. 3.1, we adopt ProtoNet as the basic FSL", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 348, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 506, + 360 + ], + "score": 1.0, + "content": "classifier due to its scalability and simplicity. To further show the effectiveness of our IEPT when other", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "score": 1.0, + "content": "basic FSL classifiers are used, we provide the results obtained by our IEPT using ProtoNet, FEAT,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "and IMP for FSL in Figure 2(c). It can be clearly observed that our IEPT leads to an improvement of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 381, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 131, + 392 + ], + "score": 1.0, + "content": "about", + "type": "text" + }, + { + "bbox": [ + 131, + 381, + 154, + 391 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 381, + 505, + 392 + ], + "score": 1.0, + "content": "over each basic FSL method (ProtoNet, FEAT, or IMP), indicating that our IEPT can be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 392, + 322, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 322, + 404 + ], + "score": 1.0, + "content": "applied to improve a variety of popular FSL methods.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 408, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "Comparative Results for Fine-Grained FSL and Cross-Domain FSL. To evaluate our IEPT", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "score": 1.0, + "content": "algorithm under the fine-grained FSL and cross-domain FSL settings, we conduct experiments on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 284, + 443 + ], + "score": 1.0, + "content": "CUB (Wah et al., 2011) and miniImageNet", + "type": "text" + }, + { + "bbox": [ + 284, + 430, + 319, + 441 + ], + "score": 0.53, + "content": " \\mathrm { C U B }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 429, + 506, + 443 + ], + "score": 1.0, + "content": ", respectively. For fine-grained FSL on CUB,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "following (Ye et al., 2020), we randomly split the dataset into 100 training classes, 50 validation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 374, + 464 + ], + "score": 1.0, + "content": "classes, and 50 test classes. For cross-domain FSL on miniImageNet", + "type": "text" + }, + { + "bbox": [ + 374, + 452, + 408, + 463 + ], + "score": 0.42, + "content": " \\mathrm { C U B }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 453, + 505, + 464 + ], + "score": 1.0, + "content": ", the 100 training classes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "score": 1.0, + "content": "are from miniImageNet; the 50 validation classes and 50 test classes (using the aforementioned split", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "for fine-grained FSL) are from CUB. Under both settings, we use Conv4-64 as the feature extractor.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "The 5-way 1/5-shot FSL results are shown in Table 3. Our IEPT clearly achieves the best results,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 141, + 509 + ], + "score": 1.0, + "content": "yielding", + "type": "text" + }, + { + "bbox": [ + 142, + 496, + 166, + 507 + ], + "score": 0.88, + "content": "1 - 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "improvements over the second-best FEAT. This shows the effectiveness of our IEPT", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 507, + 311, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 311, + 519 + ], + "score": 1.0, + "content": "under both fine-grained and cross-domain settings.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 537, + 195, + 550 + ], + "lines": [ + { + "bbox": [ + 104, + 536, + 197, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 197, + 553 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "We have proposed a novel Instance-level and Episode-level Pretext Task (IEPT) framework for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "score": 1.0, + "content": "integrating SSL into FSL. For the first time, we have introduced an episode-level pretext task for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "FSL with self-supervision, in addition to the conventional instance-level pretext task. Moreover,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "we have also developed an episode extension-integration framework by introducing an integration", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "transformer module to fully exploit the extended episodes for FSL. Extensive experiments on two", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "benchmarks demonstrate that the proposed model (i.e., FSL with IEPT) achieves the new state-of-the-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "art. Our ongoing research directions include: exploring other episode-level pretext tasks for FSL with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 438, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 438, + 654 + ], + "score": 1.0, + "content": "self-supervision, and applying FSL with self-supervision to other vision problems.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 108, + 668, + 200, + 678 + ], + "lines": [ + { + "bbox": [ + 107, + 669, + 200, + 679 + ], + "spans": [ + { + "bbox": [ + 107, + 669, + 200, + 679 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "This work was supported in part by National Natural Science Foundation of China (61976220 and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "61832017), Beijing Outstanding Young Scientist Program (BJJWZYJH012019100020098), Open", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Project Program Foundation of Key Laboratory of Opto-Electronics Information Processing, Chinese", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "score": 1.0, + "content": "Academy of Sciences (OEIP-O-202006), and Alibaba Innovative Research (AIR) Program.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 113, + 500, + 190 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 89, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 505, + 102 + ], + "score": 1.0, + "content": "Table 3: Comparative results for the fine-grained FSL on CUB (Wah et al., 2011) and the cross-domain", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 236, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 197, + 113 + ], + "score": 1.0, + "content": "FSL on miniImageNet", + "type": "text" + }, + { + "bbox": [ + 198, + 101, + 210, + 110 + ], + "score": 0.64, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 100, + 236, + 113 + ], + "score": 1.0, + "content": "CUB.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 113, + 500, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 113, + 500, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 113, + 500, + 190 + ], + "score": 0.979, + "html": "
MethodBackboneCUBminiImageNet→CUB
1-shot5-shot1-shot5-shot
MatchingNet (Vinyals et al.,2016)Conv4-6461.16±0.8972.86±0.7042.62 ± 0.5556.53± 0.44
ProtoNet (Snell et al.,2017)Conv4-6463.72 ±0.2281.50 ± 0.1550.51 ± 0.5669.28±0.40
MAML (Finn et al.,2017)Conv4-6455.92 ± 0.9572.09 ±0.7643.59 ± 0.5454.18 ± 0.41
Relation Net (Sung et al., 2018)Conv4-6462.45 ± 0.9876.11 ± 0.6949.84 ±0.5468.98 ±0.42
FEAT(Ye et al.,2020)Conv4-6468.87 ±0.2282.90±0.1551.52 ± 0.5470.16 ± 0.40
IEPT(ours)Conv4-6469.97±0.4984.33± 0.3352.68±0.5672.98±0.40
", + "type": "table", + "image_path": "b9c462107ffd68f09bfe11e7079ebc8a088483e96df432ec01a7e10e9eb18d94.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 113, + 500, + 138.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 138.66666666666666, + 500, + 164.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 164.33333333333331, + 500, + 189.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 270 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 204, + 506, + 271 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "Different Number of Extended Episodes. In all the above experiments, the number of the extended", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 142, + 299 + ], + "score": 1.0, + "content": "episodes", + "type": "text" + }, + { + "bbox": [ + 143, + 287, + 152, + 297 + ], + "score": 0.76, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 286, + 241, + 299 + ], + "score": 1.0, + "content": "is set to 4 (rotation by", + "type": "text" + }, + { + "bbox": [ + 241, + 287, + 252, + 297 + ], + "score": 0.81, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 286, + 255, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 256, + 287, + 271, + 297 + ], + "score": 0.83, + "content": "9 0 °", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 286, + 275, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 276, + 287, + 296, + 298 + ], + "score": 0.81, + "content": "1 8 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 286, + 299, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 300, + 287, + 321, + 298 + ], + "score": 0.75, + "content": "2 7 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "). Figure 2(b) shows the impact of the value of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 297, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 107, + 298, + 116, + 308 + ], + "score": 0.61, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 297, + 180, + 310 + ], + "score": 1.0, + "content": ". Note that when", + "type": "text" + }, + { + "bbox": [ + 181, + 298, + 208, + 308 + ], + "score": 0.91, + "content": "R = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 297, + 506, + 310 + ], + "score": 1.0, + "content": ", our IEPT model is equivalent to ProtoNet which is without self-supervision.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 308, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 407, + 321 + ], + "score": 1.0, + "content": "It can be seen that the performance of our model consistently grows when", + "type": "text" + }, + { + "bbox": [ + 407, + 309, + 416, + 318 + ], + "score": 0.75, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 308, + 507, + 321 + ], + "score": 1.0, + "content": "increases from 1 to 4.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "Additionally, the study on exploiting other pretext tasks for our IEPT is presented in Appendix A.1.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 276, + 507, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 348 + ], + "score": 1.0, + "content": "Different Basic FSL Classifiers. As mentioned in Sec. 3.1, we adopt ProtoNet as the basic FSL", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 348, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 506, + 360 + ], + "score": 1.0, + "content": "classifier due to its scalability and simplicity. To further show the effectiveness of our IEPT when other", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "score": 1.0, + "content": "basic FSL classifiers are used, we provide the results obtained by our IEPT using ProtoNet, FEAT,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "and IMP for FSL in Figure 2(c). It can be clearly observed that our IEPT leads to an improvement of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 381, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 131, + 392 + ], + "score": 1.0, + "content": "about", + "type": "text" + }, + { + "bbox": [ + 131, + 381, + 154, + 391 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 381, + 505, + 392 + ], + "score": 1.0, + "content": "over each basic FSL method (ProtoNet, FEAT, or IMP), indicating that our IEPT can be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 392, + 322, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 322, + 404 + ], + "score": 1.0, + "content": "applied to improve a variety of popular FSL methods.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 336, + 506, + 404 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 408, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 506, + 420 + ], + "score": 1.0, + "content": "Comparative Results for Fine-Grained FSL and Cross-Domain FSL. To evaluate our IEPT", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 431 + ], + "score": 1.0, + "content": "algorithm under the fine-grained FSL and cross-domain FSL settings, we conduct experiments on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 284, + 443 + ], + "score": 1.0, + "content": "CUB (Wah et al., 2011) and miniImageNet", + "type": "text" + }, + { + "bbox": [ + 284, + 430, + 319, + 441 + ], + "score": 0.53, + "content": " \\mathrm { C U B }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 429, + 506, + 443 + ], + "score": 1.0, + "content": ", respectively. For fine-grained FSL on CUB,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "following (Ye et al., 2020), we randomly split the dataset into 100 training classes, 50 validation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 374, + 464 + ], + "score": 1.0, + "content": "classes, and 50 test classes. For cross-domain FSL on miniImageNet", + "type": "text" + }, + { + "bbox": [ + 374, + 452, + 408, + 463 + ], + "score": 0.42, + "content": " \\mathrm { C U B }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 453, + 505, + 464 + ], + "score": 1.0, + "content": ", the 100 training classes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 504, + 475 + ], + "score": 1.0, + "content": "are from miniImageNet; the 50 validation classes and 50 test classes (using the aforementioned split", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "for fine-grained FSL) are from CUB. Under both settings, we use Conv4-64 as the feature extractor.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "The 5-way 1/5-shot FSL results are shown in Table 3. Our IEPT clearly achieves the best results,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 141, + 509 + ], + "score": 1.0, + "content": "yielding", + "type": "text" + }, + { + "bbox": [ + 142, + 496, + 166, + 507 + ], + "score": 0.88, + "content": "1 - 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "improvements over the second-best FEAT. This shows the effectiveness of our IEPT", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 507, + 311, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 311, + 519 + ], + "score": 1.0, + "content": "under both fine-grained and cross-domain settings.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 408, + 506, + 519 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 537, + 195, + 550 + ], + "lines": [ + { + "bbox": [ + 104, + 536, + 197, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 197, + 553 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "We have proposed a novel Instance-level and Episode-level Pretext Task (IEPT) framework for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 587 + ], + "score": 1.0, + "content": "integrating SSL into FSL. For the first time, we have introduced an episode-level pretext task for", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "FSL with self-supervision, in addition to the conventional instance-level pretext task. Moreover,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "we have also developed an episode extension-integration framework by introducing an integration", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "transformer module to fully exploit the extended episodes for FSL. Extensive experiments on two", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "benchmarks demonstrate that the proposed model (i.e., FSL with IEPT) achieves the new state-of-the-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "art. Our ongoing research directions include: exploring other episode-level pretext tasks for FSL with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 640, + 438, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 438, + 654 + ], + "score": 1.0, + "content": "self-supervision, and applying FSL with self-supervision to other vision problems.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 564, + 506, + 654 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 668, + 200, + 678 + ], + "lines": [ + { + "bbox": [ + 107, + 669, + 200, + 679 + ], + "spans": [ + { + "bbox": [ + 107, + 669, + 200, + 679 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "This work was supported in part by National Natural Science Foundation of China (61976220 and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "61832017), Beijing Outstanding Young Scientist Program (BJJWZYJH012019100020098), Open", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Project Program Foundation of Key Laboratory of Opto-Electronics Information Processing, Chinese", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 471, + 734 + ], + "score": 1.0, + "content": "Academy of Sciences (OEIP-O-202006), and Alibaba Innovative Research (AIR) Program.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 687, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 99, + 504, + 120 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 505, + 112 + ], + "score": 1.0, + "content": "Arman Afrasiyabi, Jean-Franc¸ois Lalonde, and Christian Gagne. Associative alignment for few-shot image ´", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 109, + 389, + 120 + ], + "spans": [ + { + "bbox": [ + 115, + 109, + 389, + 120 + ], + "score": 1.0, + "content": "classification. In European Conference on Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 503, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "Kelsey R. Allen, Evan Shelhamer, Hanul Shin, and Joshua B. Tenenbaum. Infinite mixture prototypes for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 137, + 471, + 149 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 471, + 149 + ], + "score": 1.0, + "content": "few-shot learning. In International Conference on Machine Learning (ICML), pp. 232–241, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 108, + 155, + 502, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 504, + 167 + ], + "score": 1.0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 165, + 492, + 177 + ], + "spans": [ + { + "bbox": [ + 116, + 165, + 492, + 177 + ], + "score": 1.0, + "content": "learning of visual representations. International Conference on Learning Representations (ICLR), 2020.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In International Conference on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 193, + 290, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 290, + 205 + ], + "score": 1.0, + "content": "Computer Vision (ICCV), pp. 2051–2060, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 211, + 504, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 447, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 447, + 234 + ], + "score": 1.0, + "content": "prediction. In International Conference on Computer Vision (ICCV), pp. 1422–1430, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 251 + ], + "score": 1.0, + "content": "Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 250, + 350, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 350, + 261 + ], + "score": 1.0, + "content": "Analysis and Machine Intelligence (TPAMI), pp. 594–611, 2006.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 277, + 451, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 277, + 451, + 289 + ], + "score": 1.0, + "content": "networks. In International Conference on Machine Learning (ICML), pp. 1126–1135, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 504, + 327 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "score": 1.0, + "content": "Hang Gao, Zheng Shou, Alireza Zareian, Hanwang Zhang, and Shih-Fu Chang. Low-shot learning via covariance-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 305, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 506, + 320 + ], + "score": 1.0, + "content": "preserving adversarial augmentation networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 315, + 175, + 327 + ], + "spans": [ + { + "bbox": [ + 116, + 315, + 175, + 327 + ], + "score": 1.0, + "content": "975–985, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In IEEE Conference", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 344, + 394, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 344, + 394, + 356 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pp. 4367–4375, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 108, + 361, + 504, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "Spyros Gidaris and Nikos Komodakis. Generating classification weights with gnn denoising autoencoders for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 115, + 371, + 506, + 384 + ], + "score": 1.0, + "content": "few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21–30,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 381, + 140, + 394 + ], + "spans": [ + { + "bbox": [ + 115, + 381, + 140, + 394 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 504, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 414 + ], + "score": 1.0, + "content": "Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 410, + 414, + 422 + ], + "spans": [ + { + "bbox": [ + 116, + 410, + 414, + 422 + ], + "score": 1.0, + "content": "rotations. In International Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 108, + 428, + 504, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "score": 1.0, + "content": "Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Perez, and Matthieu Cord. Boosting few-shot visual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 438, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 506, + 450 + ], + "score": 1.0, + "content": "learning with self-supervision. In International Conference on Computer Vision (ICCV), pp. 8059–8068,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 448, + 140, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 140, + 459 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 466, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 476, + 395, + 487 + ], + "spans": [ + { + "bbox": [ + 116, + 476, + 395, + 487 + ], + "score": 1.0, + "content": "International Conference on Computer Vision (ICCV), pp. 3037–3046, 2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 504, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 504, + 453, + 516 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 453, + 516 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016a.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 532, + 455, + 544 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 455, + 544 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016b.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 504, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "score": 1.0, + "content": "Ruibing Hou, Hong Chang, MA Bingpeng, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 561, + 451, + 572 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 451, + 572 + ], + "score": 1.0, + "content": "classification. In Advances in Neural Information Processing Systems, pp. 4003–4014, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 107, + 579, + 504, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 579, + 504, + 590 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 589, + 507, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 589, + 507, + 601 + ], + "score": 1.0, + "content": "networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4700–4708, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 108, + 606, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Let there be color! joint end-to-end learning of global", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "and local image priors for automatic image colorization with simultaneous classification. ACM Transactions", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 627, + 247, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 247, + 638 + ], + "score": 1.0, + "content": "on Graphics (ToG), pp. 1–11, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image clas- ˜", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 115, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "sification and segmentation. In International Conference on Computer Vision (ICCV), pp. 9865–9874,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 664, + 140, + 676 + ], + "spans": [ + { + "bbox": [ + 115, + 664, + 140, + 676 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 683, + 503, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 682, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 505, + 696 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 693, + 275, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 693, + 275, + 704 + ], + "score": 1.0, + "content": "on Learning Representations (ICLR), 2015.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 105, + 711, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 438, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 438, + 732 + ], + "score": 1.0, + "content": "networks. In Advances in Neural Information Processing Systems, pp. 1097–1105, 2012.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 99, + 504, + 120 + ], + "lines": [ + { + "bbox": [ + 105, + 98, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 505, + 112 + ], + "score": 1.0, + "content": "Arman Afrasiyabi, Jean-Franc¸ois Lalonde, and Christian Gagne. Associative alignment for few-shot image ´", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 109, + 389, + 120 + ], + "spans": [ + { + "bbox": [ + 115, + 109, + 389, + 120 + ], + "score": 1.0, + "content": "classification. In European Conference on Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 98, + 505, + 120 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 503, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "Kelsey R. Allen, Evan Shelhamer, Hanul Shin, and Joshua B. Tenenbaum. Infinite mixture prototypes for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 137, + 471, + 149 + ], + "spans": [ + { + "bbox": [ + 115, + 137, + 471, + 149 + ], + "score": 1.0, + "content": "few-shot learning. In International Conference on Machine Learning (ICML), pp. 232–241, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 127, + 505, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 155, + 502, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 504, + 167 + ], + "score": 1.0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 165, + 492, + 177 + ], + "spans": [ + { + "bbox": [ + 116, + 165, + 492, + 177 + ], + "score": 1.0, + "content": "learning of visual representations. International Conference on Learning Representations (ICLR), 2020.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 155, + 504, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In International Conference on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 193, + 290, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 290, + 205 + ], + "score": 1.0, + "content": "Computer Vision (ICCV), pp. 2051–2060, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 182, + 505, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 211, + 504, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 222, + 447, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 222, + 447, + 234 + ], + "score": 1.0, + "content": "prediction. In International Conference on Computer Vision (ICCV), pp. 1422–1430, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 211, + 505, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 251 + ], + "score": 1.0, + "content": "Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 250, + 350, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 350, + 261 + ], + "score": 1.0, + "content": "Analysis and Machine Intelligence (TPAMI), pp. 594–611, 2006.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 240, + 505, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 504, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 277, + 451, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 277, + 451, + 289 + ], + "score": 1.0, + "content": "networks. In International Conference on Machine Learning (ICML), pp. 1126–1135, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 267, + 505, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 504, + 327 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "score": 1.0, + "content": "Hang Gao, Zheng Shou, Alireza Zareian, Hanwang Zhang, and Shih-Fu Chang. Low-shot learning via covariance-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 305, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 506, + 320 + ], + "score": 1.0, + "content": "preserving adversarial augmentation networks. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 315, + 175, + 327 + ], + "spans": [ + { + "bbox": [ + 116, + 315, + 175, + 327 + ], + "score": 1.0, + "content": "975–985, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 296, + 506, + 327 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 504, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In IEEE Conference", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 344, + 394, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 344, + 394, + 356 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pp. 4367–4375, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 334, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 361, + 504, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "Spyros Gidaris and Nikos Komodakis. Generating classification weights with gnn denoising autoencoders for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 371, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 115, + 371, + 506, + 384 + ], + "score": 1.0, + "content": "few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 21–30,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 381, + 140, + 394 + ], + "spans": [ + { + "bbox": [ + 115, + 381, + 140, + 394 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 361, + 506, + 394 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 504, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 414 + ], + "score": 1.0, + "content": "Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 410, + 414, + 422 + ], + "spans": [ + { + "bbox": [ + 116, + 410, + 414, + 422 + ], + "score": 1.0, + "content": "rotations. In International Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 399, + 505, + 422 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 428, + 504, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "score": 1.0, + "content": "Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Perez, and Matthieu Cord. Boosting few-shot visual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 438, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 506, + 450 + ], + "score": 1.0, + "content": "learning with self-supervision. In International Conference on Computer Vision (ICCV), pp. 8059–8068,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 448, + 140, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 140, + 459 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 428, + 506, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 466, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 476, + 395, + 487 + ], + "spans": [ + { + "bbox": [ + 116, + 476, + 395, + 487 + ], + "score": 1.0, + "content": "International Conference on Computer Vision (ICCV), pp. 3037–3046, 2017.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 465, + 505, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 504, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 504, + 453, + 516 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 453, + 516 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016a.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 495, + 505, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 532, + 455, + 544 + ], + "spans": [ + { + "bbox": [ + 115, + 532, + 455, + 544 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016b.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 523, + 505, + 544 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 504, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "score": 1.0, + "content": "Ruibing Hou, Hong Chang, MA Bingpeng, Shiguang Shan, and Xilin Chen. Cross attention network for few-shot", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 561, + 451, + 572 + ], + "spans": [ + { + "bbox": [ + 116, + 561, + 451, + 572 + ], + "score": 1.0, + "content": "classification. In Advances in Neural Information Processing Systems, pp. 4003–4014, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 550, + 506, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 107, + 579, + 504, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 579, + 504, + 590 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 589, + 507, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 589, + 507, + 601 + ], + "score": 1.0, + "content": "networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4700–4708, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 107, + 579, + 507, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 606, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Let there be color! joint end-to-end learning of global", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "and local image priors for automatic image colorization with simultaneous classification. ACM Transactions", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 627, + 247, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 247, + 638 + ], + "score": 1.0, + "content": "on Graphics (ToG), pp. 1–11, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 606, + 505, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image clas- ˜", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 115, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "sification and segmentation. In International Conference on Computer Vision (ICCV), pp. 9865–9874,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 664, + 140, + 676 + ], + "spans": [ + { + "bbox": [ + 115, + 664, + 140, + 676 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 645, + 506, + 676 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 683, + 503, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 682, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 505, + 696 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 693, + 275, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 693, + 275, + 704 + ], + "score": 1.0, + "content": "on Learning Representations (ICLR), 2015.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 682, + 505, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 711, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 438, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 438, + 732 + ], + "score": 1.0, + "content": "networks. In Advances in Neural Information Processing Systems, pp. 1097–1105, 2012.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 712, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 83, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 107, + 83, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 506, + 94 + ], + "score": 1.0, + "content": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 385, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 385, + 105 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision (ECCV), pp. 577–593, 2016.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 121, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 114, + 121, + 506, + 133 + ], + "score": 1.0, + "content": "convex optimization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10657–", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 131, + 166, + 142 + ], + "spans": [ + { + "bbox": [ + 116, + 131, + 166, + 142 + ], + "score": 1.0, + "content": "10665, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 105, + 149, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "for few-shot learning by category traversal. In IEEE Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 170, + 212, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 170, + 212, + 182 + ], + "score": 1.0, + "content": "(CVPR), pp. 1–10, 2019a.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 105, + 187, + 505, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "Wenbin Li, Lei Wang, Jinglin Xu, Jing Huo, Yang Gao, and Jiebo Luo. Revisiting local descriptor based image-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "to-class measure for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 208, + 234, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 234, + 219 + ], + "score": 1.0, + "content": "(CVPR), pp. 7260–7268, 2019b.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 504, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 117, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "Understanding margin in few-shot classification. In European Conference on Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 104, + 254, + 504, + 276 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "score": 1.0, + "content": "Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 264, + 342, + 276 + ], + "spans": [ + { + "bbox": [ + 116, + 264, + 342, + 276 + ], + "score": 1.0, + "content": "dimension reduction. arXiv preprint arXiv:1802.03426, 2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 282, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 294 + ], + "score": 1.0, + "content": "Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 293, + 367, + 304 + ], + "spans": [ + { + "bbox": [ + 116, + 293, + 367, + 304 + ], + "score": 1.0, + "content": "International Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 504, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "score": 1.0, + "content": "Tsendsuren Munkhdalai and Hong Yu. Meta networks. In International Conference on Machine Learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 321, + 229, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 229, + 333 + ], + "score": 1.0, + "content": "(ICML), pp. 2554–2563, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 105, + 339, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 349, + 459, + 361 + ], + "spans": [ + { + "bbox": [ + 116, + 349, + 459, + 361 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9339–9348, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 368, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 378, + 366, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 366, + 389 + ], + "score": 1.0, + "content": "European Conference on Computer Vision (ECCV), pp. 69–84, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "score": 1.0, + "content": "Mehdi Noroozi, Ananth Vinjimoor, Paolo Favaro, and Hamed Pirsiavash. Boosting self-supervised learning", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 405, + 507, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 405, + 507, + 420 + ], + "score": 1.0, + "content": "via knowledge transfer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 416, + 184, + 428 + ], + "spans": [ + { + "bbox": [ + 116, + 416, + 184, + 428 + ], + "score": 1.0, + "content": "9359–9367, 2018.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "David Novotny, Samuel Albanie, Diane Larlus, and Andrea Vedaldi. Self-supervised learning of geometrically", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 444, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 116, + 444, + 506, + 456 + ], + "score": 1.0, + "content": "stable features through probabilistic introspection. In IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 455, + 276, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 276, + 466 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 3637–3645, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 503, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "score": 1.0, + "content": "Boris Oreshkin, Pau Rodr´ıguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for ´", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 483, + 495, + 495 + ], + "spans": [ + { + "bbox": [ + 116, + 483, + 495, + 495 + ], + "score": 1.0, + "content": "improved few-shot learning. In Advances in Neural Information Processing Systems, pp. 721–731, 2018.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 501, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "Feature learning by inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 522, + 184, + 532 + ], + "spans": [ + { + "bbox": [ + 116, + 522, + 184, + 532 + ], + "score": 1.0, + "content": "2536–2544, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 104, + 539, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE Conference", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 550, + 394, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 550, + 394, + 561 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pp. 5822–5830, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 504, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "Siyuan Qiao, Chenxi Liu, Wei Shen, and Alan L Yuille. Few-shot image recognition by predicting parameters", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "from activations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7229–7238,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 617, + 275, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 275, + 627 + ], + "score": 1.0, + "content": "on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 504, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "Avinash Ravichandran, Rahul Bhotika, and Stefano Soatto. Few-shot learning with embedded class models and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 645, + 483, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 645, + 483, + 657 + ], + "score": 1.0, + "content": "shot-free meta training. In International Conference on Computer Vision (ICCV), pp. 331–339, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 109, + 663, + 504, + 694 + ], + "lines": [ + { + "bbox": [ + 108, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 108, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 673, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 116, + 673, + 505, + 684 + ], + "score": 1.0, + "content": "Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 683, + 318, + 695 + ], + "spans": [ + { + "bbox": [ + 116, + 683, + 318, + 695 + ], + "score": 1.0, + "content": "Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 108, + 701, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 116, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 114, + 721, + 461, + 733 + ], + "spans": [ + { + "bbox": [ + 114, + 721, + 461, + 733 + ], + "score": 1.0, + "content": "recognition challenge. In International Journal of Computer Vision (IJCV), pp. 211–252, 2015.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 83, + 505, + 104 + ], + "lines": [ + { + "bbox": [ + 107, + 83, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 506, + 94 + ], + "score": 1.0, + "content": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Learning representations for automatic colorization.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 385, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 385, + 105 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision (ECCV), pp. 577–593, 2016.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 107, + 83, + 506, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 121, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 114, + 121, + 506, + 133 + ], + "score": 1.0, + "content": "convex optimization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10657–", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 131, + 166, + 142 + ], + "spans": [ + { + "bbox": [ + 116, + 131, + 166, + 142 + ], + "score": 1.0, + "content": "10665, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 111, + 506, + 142 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 149, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang. Finding task-relevant features", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "for few-shot learning by category traversal. In IEEE Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 170, + 212, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 170, + 212, + 182 + ], + "score": 1.0, + "content": "(CVPR), pp. 1–10, 2019a.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 149, + 505, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 187, + 505, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "Wenbin Li, Lei Wang, Jinglin Xu, Jing Huo, Yang Gao, and Jiebo Luo. Revisiting local descriptor based image-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "to-class measure for few-shot learning. In IEEE Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 208, + 234, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 234, + 219 + ], + "score": 1.0, + "content": "(CVPR), pp. 7260–7268, 2019b.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 187, + 506, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 504, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 239 + ], + "score": 1.0, + "content": "Bin Liu, Yue Cao, Yutong Lin, Qi Li, Zheng Zhang, Mingsheng Long, and Han Hu. Negative margin matters:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 117, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "Understanding margin in few-shot classification. In European Conference on Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 225, + 506, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 254, + 504, + 276 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "score": 1.0, + "content": "Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 264, + 342, + 276 + ], + "spans": [ + { + "bbox": [ + 116, + 264, + 342, + 276 + ], + "score": 1.0, + "content": "dimension reduction. arXiv preprint arXiv:1802.03426, 2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 255, + 505, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 282, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 294 + ], + "score": 1.0, + "content": "Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 293, + 367, + 304 + ], + "spans": [ + { + "bbox": [ + 116, + 293, + 367, + 304 + ], + "score": 1.0, + "content": "International Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 283, + 505, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 504, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "score": 1.0, + "content": "Tsendsuren Munkhdalai and Hong Yu. Meta networks. In International Conference on Machine Learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 321, + 229, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 229, + 333 + ], + "score": 1.0, + "content": "(ICML), pp. 2554–2563, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 309, + 505, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 339, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "T Nathan Mundhenk, Daniel Ho, and Barry Y Chen. Improvements to context based self-supervised learning. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 349, + 459, + 361 + ], + "spans": [ + { + "bbox": [ + 116, + 349, + 459, + 361 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9339–9348, 2018.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 339, + 505, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 368, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 378, + 366, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 366, + 389 + ], + "score": 1.0, + "content": "European Conference on Computer Vision (ECCV), pp. 69–84, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 367, + 505, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 409 + ], + "score": 1.0, + "content": "Mehdi Noroozi, Ananth Vinjimoor, Paolo Favaro, and Hamed Pirsiavash. Boosting self-supervised learning", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 405, + 507, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 405, + 507, + 420 + ], + "score": 1.0, + "content": "via knowledge transfer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 416, + 184, + 428 + ], + "spans": [ + { + "bbox": [ + 116, + 416, + 184, + 428 + ], + "score": 1.0, + "content": "9359–9367, 2018.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 394, + 507, + 428 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "David Novotny, Samuel Albanie, Diane Larlus, and Andrea Vedaldi. Self-supervised learning of geometrically", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 444, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 116, + 444, + 506, + 456 + ], + "score": 1.0, + "content": "stable features through probabilistic introspection. In IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 455, + 276, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 455, + 276, + 466 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 3637–3645, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 434, + 506, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 503, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 484 + ], + "score": 1.0, + "content": "Boris Oreshkin, Pau Rodr´ıguez Lopez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for ´", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 483, + 495, + 495 + ], + "spans": [ + { + "bbox": [ + 116, + 483, + 495, + 495 + ], + "score": 1.0, + "content": "improved few-shot learning. In Advances in Neural Information Processing Systems, pp. 721–731, 2018.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 473, + 505, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 501, + 504, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 506, + 513 + ], + "score": 1.0, + "content": "Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "Feature learning by inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 522, + 184, + 532 + ], + "spans": [ + { + "bbox": [ + 116, + 522, + 184, + 532 + ], + "score": 1.0, + "content": "2536–2544, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 501, + 506, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 539, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE Conference", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 550, + 394, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 550, + 394, + 561 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pp. 5822–5830, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 538, + 505, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 504, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "Siyuan Qiao, Chenxi Liu, Wei Shen, and Alan L Yuille. Few-shot image recognition by predicting parameters", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "from activations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7229–7238,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 140, + 599 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 567, + 506, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 617, + 275, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 275, + 627 + ], + "score": 1.0, + "content": "on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 605, + 505, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 504, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "Avinash Ravichandran, Rahul Bhotika, and Stefano Soatto. Few-shot learning with embedded class models and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 645, + 483, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 645, + 483, + 657 + ], + "score": 1.0, + "content": "shot-free meta training. In International Conference on Computer Vision (ICCV), pp. 331–339, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 635, + 505, + 657 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 663, + 504, + 694 + ], + "lines": [ + { + "bbox": [ + 108, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 108, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 673, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 116, + 673, + 505, + 684 + ], + "score": 1.0, + "content": "Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 683, + 318, + 695 + ], + "spans": [ + { + "bbox": [ + 116, + 683, + 318, + 695 + ], + "score": 1.0, + "content": "Conference on Learning Representations (ICLR), 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 108, + 663, + 505, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 701, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 116, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 114, + 721, + 461, + 733 + ], + "spans": [ + { + "bbox": [ + 114, + 721, + 461, + 733 + ], + "score": 1.0, + "content": "recognition challenge. In International Journal of Computer Vision (IJCV), pp. 211–252, 2015.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 700, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 83, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 106 + ], + "score": 1.0, + "content": "Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 229, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 229, + 114 + ], + "score": 1.0, + "content": "Representations (ICLR), 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 120, + 503, + 152 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 129, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 114, + 129, + 505, + 145 + ], + "score": 1.0, + "content": "with memory-augmented neural networks. In International Conference on Machine Learning (ICML), pp.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 141, + 183, + 152 + ], + "spans": [ + { + "bbox": [ + 116, + 141, + 183, + 152 + ], + "score": 1.0, + "content": "1842–1850, 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 158, + 504, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 169, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 116, + 169, + 505, + 181 + ], + "score": 1.0, + "content": "Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 179, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 179, + 446, + 191 + ], + "score": 1.0, + "content": "recognition. In Advances in Neural Information Processing Systems, pp. 2850–2860, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 503, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. Adaptive subspaces for few-shot learning.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 468, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 468, + 219 + ], + "score": 1.0, + "content": "In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4136–4145, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 504, + 246 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "score": 1.0, + "content": "Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 235, + 344, + 246 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 344, + 246 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 4077–4087, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "score": 1.0, + "content": "Jong-Chyi Su, Subhransu Maji, and Bharath Hariharan. When does self-supervision improve few-shot learning?", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 263, + 336, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 263, + 336, + 274 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 105, + 280, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 107, + 281, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 107, + 281, + 505, + 292 + ], + "score": 1.0, + "content": "Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 290, + 450, + 302 + ], + "spans": [ + { + "bbox": [ + 116, + 290, + 450, + 302 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 308, + 504, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "compare: Relation network for few-shot learning. In IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 328, + 275, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 328, + 275, + 340 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 1199–1208, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 359 + ], + "score": 1.0, + "content": "Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "classification: a good embedding is all you need? In European Conference on Computer Vision (ECCV),", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 366, + 140, + 377 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 140, + 377 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 503, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Satoshi Tsutsui, Yanwei Fu, and David Crandall. Meta-reinforced synthetic data for one-shot fine-grained visual", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 394, + 446, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 446, + 406 + ], + "score": 1.0, + "content": "recognition. In Advances in Neural Information Processing Systems, pp. 3057–3066, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 105, + 411, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 420, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 420, + 507, + 435 + ], + "score": 1.0, + "content": "and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 431, + 185, + 443 + ], + "spans": [ + { + "bbox": [ + 115, + 431, + 185, + 443 + ], + "score": 1.0, + "content": "5998–6008, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 105, + 449, + 504, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "score": 1.0, + "content": "Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 460, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 116, + 460, + 480, + 471 + ], + "score": 1.0, + "content": "for one shot learning. In Advances in Neural Information Processing Systems, pp. 3630–3638, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 504, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 504, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 504, + 489 + ], + "score": 1.0, + "content": "C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 487, + 366, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 366, + 499 + ], + "score": 1.0, + "content": "Report CNS-TR-2011-001, California Institute of Technology, 2011.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 107, + 506, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 506, + 517 + ], + "score": 1.0, + "content": "Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 515, + 468, + 527 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 468, + 527 + ], + "score": 1.0, + "content": "In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7278–7286, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Ziyang Wu, Yuwei Li, Lihua Guo, and Kui Jia. Parn: Position-aware relation networks for few-shot learning. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 543, + 395, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 543, + 395, + 555 + ], + "score": 1.0, + "content": "International Conference on Computer Vision (ICCV), pp. 6659–6667, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 107, + 561, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 107, + 561, + 505, + 572 + ], + "score": 1.0, + "content": "Chen Xing, Negar Rostamzadeh, Boris Oreshkin, and Pedro O O. Pinheiro. Adaptive cross-modal few-shot", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 570, + 435, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 570, + 435, + 583 + ], + "score": 1.0, + "content": "learning. In Advances in Neural Information Processing Systems, pp. 4847–4857, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 599, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 506, + 611 + ], + "score": 1.0, + "content": "set functions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8808–8817,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 609, + 140, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 140, + 620 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 504, + 658 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 637, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 506, + 649 + ], + "score": 1.0, + "content": "differentiable earth mover’s distance and structured classifiers. In IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 647, + 313, + 658 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 313, + 658 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), pp. 12203–12213, 2020.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 105, + 664, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 676 + ], + "score": 1.0, + "content": "Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 675, + 503, + 687 + ], + "spans": [ + { + "bbox": [ + 115, + 675, + 503, + 687 + ], + "score": 1.0, + "content": "samples. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2770–2779, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 83, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 106 + ], + "score": 1.0, + "content": "Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 229, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 229, + 114 + ], + "score": 1.0, + "content": "Representations (ICLR), 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 83, + 505, + 114 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 120, + 503, + 152 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 129, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 114, + 129, + 505, + 145 + ], + "score": 1.0, + "content": "with memory-augmented neural networks. In International Conference on Machine Learning (ICML), pp.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 141, + 183, + 152 + ], + "spans": [ + { + "bbox": [ + 116, + 141, + 183, + 152 + ], + "score": 1.0, + "content": "1842–1850, 2016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 120, + 505, + 152 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 158, + 504, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 169, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 116, + 169, + 505, + 181 + ], + "score": 1.0, + "content": "Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 179, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 179, + 446, + 191 + ], + "score": 1.0, + "content": "recognition. In Advances in Neural Information Processing Systems, pp. 2850–2860, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 158, + 505, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 503, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "Christian Simon, Piotr Koniusz, Richard Nock, and Mehrtash Harandi. Adaptive subspaces for few-shot learning.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 468, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 468, + 219 + ], + "score": 1.0, + "content": "In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4136–4145, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 196, + 505, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 224, + 504, + 246 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 237 + ], + "score": 1.0, + "content": "Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 235, + 344, + 246 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 344, + 246 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 4077–4087, 2017.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 223, + 505, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 266 + ], + "score": 1.0, + "content": "Jong-Chyi Su, Subhransu Maji, and Bharath Hariharan. When does self-supervision improve few-shot learning?", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 263, + 336, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 263, + 336, + 274 + ], + "score": 1.0, + "content": "In European Conference on Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 251, + 506, + 274 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 280, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 107, + 281, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 107, + 281, + 505, + 292 + ], + "score": 1.0, + "content": "Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 290, + 450, + 302 + ], + "spans": [ + { + "bbox": [ + 116, + 290, + 450, + 302 + ], + "score": 1.0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 107, + 281, + 505, + 302 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 308, + 504, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "compare: Relation network for few-shot learning. In IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 328, + 275, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 328, + 275, + 340 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 1199–1208, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 307, + 506, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 359 + ], + "score": 1.0, + "content": "Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "classification: a good embedding is all you need? In European Conference on Computer Vision (ECCV),", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 366, + 140, + 377 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 140, + 377 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 344, + 506, + 377 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 503, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Satoshi Tsutsui, Yanwei Fu, and David Crandall. Meta-reinforced synthetic data for one-shot fine-grained visual", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 394, + 446, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 446, + 406 + ], + "score": 1.0, + "content": "recognition. In Advances in Neural Information Processing Systems, pp. 3057–3066, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 384, + 505, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 411, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 420, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 420, + 507, + 435 + ], + "score": 1.0, + "content": "and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 431, + 185, + 443 + ], + "spans": [ + { + "bbox": [ + 115, + 431, + 185, + 443 + ], + "score": 1.0, + "content": "5998–6008, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 411, + 507, + 443 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 449, + 504, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "score": 1.0, + "content": "Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 460, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 116, + 460, + 480, + 471 + ], + "score": 1.0, + "content": "for one shot learning. In Advances in Neural Information Processing Systems, pp. 3630–3638, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 450, + 505, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 504, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 504, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 504, + 489 + ], + "score": 1.0, + "content": "C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 487, + 366, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 366, + 499 + ], + "score": 1.0, + "content": "Report CNS-TR-2011-001, California Institute of Technology, 2011.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 478, + 504, + 499 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 107, + 506, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 506, + 517 + ], + "score": 1.0, + "content": "Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 515, + 468, + 527 + ], + "spans": [ + { + "bbox": [ + 116, + 515, + 468, + 527 + ], + "score": 1.0, + "content": "In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7278–7286, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 107, + 506, + 506, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Ziyang Wu, Yuwei Li, Lihua Guo, and Kui Jia. Parn: Position-aware relation networks for few-shot learning. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 543, + 395, + 555 + ], + "spans": [ + { + "bbox": [ + 115, + 543, + 395, + 555 + ], + "score": 1.0, + "content": "International Conference on Computer Vision (ICCV), pp. 6659–6667, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 533, + 505, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 107, + 561, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 107, + 561, + 505, + 572 + ], + "score": 1.0, + "content": "Chen Xing, Negar Rostamzadeh, Boris Oreshkin, and Pedro O O. Pinheiro. Adaptive cross-modal few-shot", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 570, + 435, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 570, + 435, + 583 + ], + "score": 1.0, + "content": "learning. In Advances in Neural Information Processing Systems, pp. 4847–4857, 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 107, + 561, + 505, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 599, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 506, + 611 + ], + "score": 1.0, + "content": "set functions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8808–8817,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 609, + 140, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 140, + 620 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 588, + 506, + 620 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 504, + 658 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 637, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 506, + 649 + ], + "score": 1.0, + "content": "differentiable earth mover’s distance and structured classifiers. In IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 647, + 313, + 658 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 313, + 658 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), pp. 12203–12213, 2020.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 627, + 506, + 658 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 664, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 676 + ], + "score": 1.0, + "content": "Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 675, + 503, + 687 + ], + "spans": [ + { + "bbox": [ + 115, + 675, + 503, + 687 + ], + "score": 1.0, + "content": "samples. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2770–2779, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 665, + 505, + 687 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 182, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 184, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 184, + 96 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 109, + 354, + 120 + ], + "lines": [ + { + "bbox": [ + 106, + 109, + 356, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 356, + 122 + ], + "score": 1.0, + "content": "A.1 COMPARISON AMONG DIFFERENT SSL STRATEGIES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 491, + 145 + ], + "score": 1.0, + "content": "To generate the extended episodes in IEPT, we apply four rotation transformations (i.e. rotation by", + "type": "text" + }, + { + "bbox": [ + 492, + 132, + 502, + 142 + ], + "score": 0.82, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 131, + 506, + 145 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 122, + 153 + ], + "score": 0.81, + "content": "9 0 °", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 142, + 126, + 155 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 126, + 142, + 146, + 153 + ], + "score": 0.75, + "content": "1 8 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 142, + 150, + 155 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 151, + 142, + 172, + 153 + ], + "score": 0.73, + "content": "2 7 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 142, + 505, + 155 + ], + "score": 1.0, + "content": ") to each image. It makes sense to explore whether other self-supervised strategies", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "are also effective for our IEPT. To this end, we exploit shuffling image patches (see Figure 3) for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 393, + 177 + ], + "score": 1.0, + "content": "self-supervised learning (SSL). Specifically, we divide each image into", + "type": "text" + }, + { + "bbox": [ + 393, + 165, + 410, + 175 + ], + "score": 0.85, + "content": "2 ^ { \\ast } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "patches and reorganize", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "the patch orders to obtain a shuffling label. To compare with the rotation strategy fairly, we choose", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "only four shuffling orders: (1, 2, 3, 4), (2, 3, 4, 1), (3, 4, 1, 2) and (4, 1, 2, 3). Note that the (1, 2, 3, 4)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "shuffling order equals to the original image. Similar to the rotation strategy, a fully-connected layer is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 506, + 221 + ], + "score": 1.0, + "content": "utilized to recognize the shuffling order. The comparative results are shown in Table 4. We can see", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "that both IEPT with shuffling and IEPT with rotation achieve better performance than the original", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 309, + 243 + ], + "score": 1.0, + "content": "ProtoNet. Particularly, IEPT with shuffling yields", + "type": "text" + }, + { + "bbox": [ + 309, + 230, + 332, + 241 + ], + "score": 0.86, + "content": "1 - 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 230, + 350, + 243 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 350, + 230, + 374, + 241 + ], + "score": 0.87, + "content": "3- 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "improvements under 5-shot and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "score": 1.0, + "content": "1-shot, respectively. This clearly shows the effectiveness of our IEPT for FSL even when different", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 252, + 309, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 309, + 265 + ], + "score": 1.0, + "content": "SSL strategies are used to define the pretext tasks.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 7.5 + }, + { + "type": "table", + "bbox": [ + 114, + 302, + 496, + 353 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 280, + 503, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 277, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 294 + ], + "score": 1.0, + "content": "Table 4: FSL results obtained by our IEPT using two SSL strategies (i.e. rotation and shuffling image", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 312, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 312, + 303 + ], + "score": 1.0, + "content": "patches). Conv4-64 is used as the feature extractor.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "table_body", + "bbox": [ + 114, + 302, + 496, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 302, + 496, + 353 + ], + "spans": [ + { + "bbox": [ + 114, + 302, + 496, + 353 + ], + "score": 0.962, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
ProtoNetConv4-6452.61 ±0.5271.33 ± 0.4153.33 ± 0.5072.10± 0.41
IEPT (rotation)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ±0.4875.63 ± 0.46
IEPT (shuffling)Conv4-6455.57 ± 0.6072.84 ± 0.5457.81 ± 0.4874.92 ± 0.50
", + "type": "table", + "image_path": "826169c133d349e3d44c759fcc93bff28025782503f1373f27ec4f7ad4152385.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 302, + 496, + 319.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 114, + 319.0, + 496, + 336.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 114, + 336.0, + 496, + 353.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 15.75 + }, + { + "type": "image", + "bbox": [ + 109, + 371, + 500, + 463 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 371, + 500, + 463 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 371, + 500, + 463 + ], + "spans": [ + { + "bbox": [ + 109, + 371, + 500, + 463 + ], + "score": 0.962, + "type": "image", + "image_path": "2d68786b89daf9ea45fcd13f2c425d6edbd6a37859f608778de12061ff067209.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 109, + 371, + 500, + 401.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 109, + 401.6666666666667, + 500, + 432.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 109, + 432.33333333333337, + 500, + 463.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 144, + 469, + 463, + 480 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 466, + 464, + 483 + ], + "spans": [ + { + "bbox": [ + 143, + 466, + 464, + 483 + ], + "score": 1.0, + "content": "Figure 3: Illustration of the self-supervised strategy by shuffling image patches.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 21.0 + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 399, + 520 + ], + "lines": [ + { + "bbox": [ + 110, + 507, + 399, + 521 + ], + "spans": [ + { + "bbox": [ + 110, + 507, + 399, + 521 + ], + "score": 1.0, + "content": "A.2 COMPARISON AMONG DIFFERENT INTEGRATION APPROACHES", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "We employ the integration transformer to find the intrinsic correlation of various rotation-transformed", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "score": 1.0, + "content": "instances. The transformer architecture is based on a self-attention mechanism. Concretely, it", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 551, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 325, + 567 + ], + "score": 1.0, + "content": "receives the feature sets of extended episodes as input", + "type": "text" + }, + { + "bbox": [ + 326, + 553, + 335, + 564 + ], + "score": 0.77, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 551, + 338, + 567 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 338, + 553, + 349, + 563 + ], + "score": 0.73, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 551, + 367, + 567 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 553, + 376, + 563 + ], + "score": 0.64, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 551, + 506, + 567 + ], + "score": 1.0, + "content": ". Further, it matches each query", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 117, + 577 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 564, + 127, + 576 + ], + "score": 0.82, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 563, + 214, + 577 + ], + "score": 1.0, + "content": "with a list of keys in", + "type": "text" + }, + { + "bbox": [ + 214, + 564, + 225, + 574 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "and returns the weighted sum of corresponding values. To show the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "importance of the transformer module, we compare it with two other integration approaches (i.e.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "concatenating and averaging) to integrate the features of extended episodes. The comparative results", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "in Table 5 demonstrate that the integration transformer consistently performs better than the simply", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "concatenating/averaging approaches. This suggests that the attention-based integration transformer is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 619, + 317, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 317, + 631 + ], + "score": 1.0, + "content": "a better choice for designing the integration module.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 113, + 669, + 498, + 720 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 646, + 503, + 668 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 646, + 504, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 504, + 659 + ], + "score": 1.0, + "content": "Table 5: Comparative results obtained by three different approaches to integrating the features from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 657, + 370, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 370, + 670 + ], + "score": 1.0, + "content": "the extended episodes, with Conv4-64 being the feature extractor.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "table_body", + "bbox": [ + 113, + 669, + 498, + 720 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 669, + 498, + 720 + ], + "spans": [ + { + "bbox": [ + 113, + 669, + 498, + 720 + ], + "score": 0.976, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
concatenatingConv4-6451.52±0.6073.36± 0.4950.78±0.6872.79± 0.57
averagingConv4-6451.58 ± 0.6270.97 ± 0.5453.91 ± 0.6972.52 ± 0.59
transformer (ours)Conv4-6456.26± 0.4573.91 ± 0.3458.25 ± 0.4875.63± 0.46
", + "type": "table", + "image_path": "46c04c223884d28fcbcd433602647d05ad9231a973280b9deba6d14a9738132f.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 113, + 669, + 498, + 686.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 113, + 686.0, + 498, + 703.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 113, + 703.0, + 498, + 720.0 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 34.75 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 182, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 184, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 184, + 96 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 109, + 354, + 120 + ], + "lines": [ + { + "bbox": [ + 106, + 109, + 356, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 356, + 122 + ], + "score": 1.0, + "content": "A.1 COMPARISON AMONG DIFFERENT SSL STRATEGIES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 491, + 145 + ], + "score": 1.0, + "content": "To generate the extended episodes in IEPT, we apply four rotation transformations (i.e. rotation by", + "type": "text" + }, + { + "bbox": [ + 492, + 132, + 502, + 142 + ], + "score": 0.82, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 131, + 506, + 145 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 122, + 153 + ], + "score": 0.81, + "content": "9 0 °", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 142, + 126, + 155 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 126, + 142, + 146, + 153 + ], + "score": 0.75, + "content": "1 8 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 142, + 150, + 155 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 151, + 142, + 172, + 153 + ], + "score": 0.73, + "content": "2 7 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 142, + 505, + 155 + ], + "score": 1.0, + "content": ") to each image. It makes sense to explore whether other self-supervised strategies", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "are also effective for our IEPT. To this end, we exploit shuffling image patches (see Figure 3) for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 393, + 177 + ], + "score": 1.0, + "content": "self-supervised learning (SSL). Specifically, we divide each image into", + "type": "text" + }, + { + "bbox": [ + 393, + 165, + 410, + 175 + ], + "score": 0.85, + "content": "2 ^ { \\ast } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "patches and reorganize", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "the patch orders to obtain a shuffling label. To compare with the rotation strategy fairly, we choose", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "only four shuffling orders: (1, 2, 3, 4), (2, 3, 4, 1), (3, 4, 1, 2) and (4, 1, 2, 3). Note that the (1, 2, 3, 4)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "shuffling order equals to the original image. Similar to the rotation strategy, a fully-connected layer is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 506, + 221 + ], + "score": 1.0, + "content": "utilized to recognize the shuffling order. The comparative results are shown in Table 4. We can see", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "that both IEPT with shuffling and IEPT with rotation achieve better performance than the original", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 309, + 243 + ], + "score": 1.0, + "content": "ProtoNet. Particularly, IEPT with shuffling yields", + "type": "text" + }, + { + "bbox": [ + 309, + 230, + 332, + 241 + ], + "score": 0.86, + "content": "1 - 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 230, + 350, + 243 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 350, + 230, + 374, + 241 + ], + "score": 0.87, + "content": "3- 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "improvements under 5-shot and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "score": 1.0, + "content": "1-shot, respectively. This clearly shows the effectiveness of our IEPT for FSL even when different", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 252, + 309, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 309, + 265 + ], + "score": 1.0, + "content": "SSL strategies are used to define the pretext tasks.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 131, + 506, + 265 + ] + }, + { + "type": "table", + "bbox": [ + 114, + 302, + 496, + 353 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 280, + 503, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 277, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 294 + ], + "score": 1.0, + "content": "Table 4: FSL results obtained by our IEPT using two SSL strategies (i.e. rotation and shuffling image", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 312, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 312, + 303 + ], + "score": 1.0, + "content": "patches). Conv4-64 is used as the feature extractor.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "table_body", + "bbox": [ + 114, + 302, + 496, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 302, + 496, + 353 + ], + "spans": [ + { + "bbox": [ + 114, + 302, + 496, + 353 + ], + "score": 0.962, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
ProtoNetConv4-6452.61 ±0.5271.33 ± 0.4153.33 ± 0.5072.10± 0.41
IEPT (rotation)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ±0.4875.63 ± 0.46
IEPT (shuffling)Conv4-6455.57 ± 0.6072.84 ± 0.5457.81 ± 0.4874.92 ± 0.50
", + "type": "table", + "image_path": "826169c133d349e3d44c759fcc93bff28025782503f1373f27ec4f7ad4152385.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 302, + 496, + 319.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 114, + 319.0, + 496, + 336.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 114, + 336.0, + 496, + 353.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 15.75 + }, + { + "type": "image", + "bbox": [ + 109, + 371, + 500, + 463 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 371, + 500, + 463 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 371, + 500, + 463 + ], + "spans": [ + { + "bbox": [ + 109, + 371, + 500, + 463 + ], + "score": 0.962, + "type": "image", + "image_path": "2d68786b89daf9ea45fcd13f2c425d6edbd6a37859f608778de12061ff067209.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 109, + 371, + 500, + 401.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 109, + 401.6666666666667, + 500, + 432.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 109, + 432.33333333333337, + 500, + 463.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 144, + 469, + 463, + 480 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 466, + 464, + 483 + ], + "spans": [ + { + "bbox": [ + 143, + 466, + 464, + 483 + ], + "score": 1.0, + "content": "Figure 3: Illustration of the self-supervised strategy by shuffling image patches.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 21.0 + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 399, + 520 + ], + "lines": [ + { + "bbox": [ + 110, + 507, + 399, + 521 + ], + "spans": [ + { + "bbox": [ + 110, + 507, + 399, + 521 + ], + "score": 1.0, + "content": "A.2 COMPARISON AMONG DIFFERENT INTEGRATION APPROACHES", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 110, + 507, + 399, + 521 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "We employ the integration transformer to find the intrinsic correlation of various rotation-transformed", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "score": 1.0, + "content": "instances. The transformer architecture is based on a self-attention mechanism. Concretely, it", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 551, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 325, + 567 + ], + "score": 1.0, + "content": "receives the feature sets of extended episodes as input", + "type": "text" + }, + { + "bbox": [ + 326, + 553, + 335, + 564 + ], + "score": 0.77, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 551, + 338, + 567 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 338, + 553, + 349, + 563 + ], + "score": 0.73, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 551, + 367, + 567 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 553, + 376, + 563 + ], + "score": 0.64, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 551, + 506, + 567 + ], + "score": 1.0, + "content": ". Further, it matches each query", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 117, + 577 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 564, + 127, + 576 + ], + "score": 0.82, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 563, + 214, + 577 + ], + "score": 1.0, + "content": "with a list of keys in", + "type": "text" + }, + { + "bbox": [ + 214, + 564, + 225, + 574 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "and returns the weighted sum of corresponding values. To show the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "importance of the transformer module, we compare it with two other integration approaches (i.e.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "concatenating and averaging) to integrate the features of extended episodes. The comparative results", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "in Table 5 demonstrate that the integration transformer consistently performs better than the simply", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "concatenating/averaging approaches. This suggests that the attention-based integration transformer is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 619, + 317, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 317, + 631 + ], + "score": 1.0, + "content": "a better choice for designing the integration module.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 531, + 506, + 631 + ] + }, + { + "type": "table", + "bbox": [ + 113, + 669, + 498, + 720 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 646, + 503, + 668 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 646, + 504, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 504, + 659 + ], + "score": 1.0, + "content": "Table 5: Comparative results obtained by three different approaches to integrating the features from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 657, + 370, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 370, + 670 + ], + "score": 1.0, + "content": "the extended episodes, with Conv4-64 being the feature extractor.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "table_body", + "bbox": [ + 113, + 669, + 498, + 720 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 669, + 498, + 720 + ], + "spans": [ + { + "bbox": [ + 113, + 669, + 498, + 720 + ], + "score": 0.976, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
concatenatingConv4-6451.52±0.6073.36± 0.4950.78±0.6872.79± 0.57
averagingConv4-6451.58 ± 0.6270.97 ± 0.5453.91 ± 0.6972.52 ± 0.59
transformer (ours)Conv4-6456.26± 0.4573.91 ± 0.3458.25 ± 0.4875.63± 0.46
", + "type": "table", + "image_path": "46c04c223884d28fcbcd433602647d05ad9231a973280b9deba6d14a9738132f.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 113, + 669, + 498, + 686.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 113, + 686.0, + 498, + 703.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 113, + 703.0, + 498, + 720.0 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 34.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 81, + 488, + 473 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 81, + 488, + 473 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 81, + 488, + 473 + ], + "spans": [ + { + "bbox": [ + 113, + 81, + 488, + 473 + ], + "score": 0.946, + "type": "image", + "image_path": "afa66040ad4c1918609b3c5a1e793531019fcc73f4e6f207f4681c14646f499f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 81, + 488, + 211.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 211.66666666666666, + 488, + 342.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 342.3333333333333, + 488, + 473.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 479, + 505, + 524 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "Figure 4: Feature visualizations of a number of test episodes using the UMAP algorithm (McInnes", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "et al., 2018). Each row indicates a group of test extended episodes (the first four columns, rotation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 118, + 515 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 502, + 129, + 512 + ], + "score": 0.81, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 502, + 133, + 515 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 133, + 502, + 149, + 512 + ], + "score": 0.85, + "content": "9 0 °", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 502, + 153, + 515 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 153, + 502, + 173, + 512 + ], + "score": 0.84, + "content": "1 8 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 502, + 177, + 515 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 177, + 502, + 199, + 513 + ], + "score": 0.82, + "content": "2 7 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 502, + 505, + 515 + ], + "score": 1.0, + "content": ") and their integrated episode (the last column). The 5-way 5-shot FSL (with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 513, + 269, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 269, + 524 + ], + "score": 1.0, + "content": "Conv4-64) is adopted on miniImageNet.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "title", + "bbox": [ + 107, + 541, + 332, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 333, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 333, + 554 + ], + "score": 1.0, + "content": "A.3 FEATURE VISUALIZATIONS OF TEST EPISODES", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "We provide the feature visualizations of test episodes in Figure 4. It can be seen that an integrated", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "episode (the last one in each row) clearly has a better cluster data structure than the corresponding", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "four extended episodes (the first four ones in each row). This indicates that our transformer-based", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 597, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 482, + 610 + ], + "score": 1.0, + "content": "across-episode integration is indeed effective for few-shot classification with self-supervision.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 631, + 336, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 631, + 337, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 337, + 644 + ], + "score": 1.0, + "content": "A.4 ATTENTION VISUALIZATION OF TEST EPISODES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "score": 1.0, + "content": "We present attention map visualization of two test episodes (left and right) in Figure 5. Each average", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "attention map is computed by averaging the attention map of all instances of a certain class. We can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "observe that: (1) The average attention maps from different classes vary significantly, showing that the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "diverse semantics of different classes can be reflected by our attention-based integration transformer.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "(2) When the classes of two episodes overlap (e.g., ‘trifle’ and ‘dalmatian’), the average attention", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "maps of an overlapped class from two episodes are similar, illustrating that our attention-based", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 721, + 433, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 433, + 733 + ], + "score": 1.0, + "content": "integration transformer can well capture the semantics of classes across episodes.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 81, + 488, + 473 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 81, + 488, + 473 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 81, + 488, + 473 + ], + "spans": [ + { + "bbox": [ + 113, + 81, + 488, + 473 + ], + "score": 0.946, + "type": "image", + "image_path": "afa66040ad4c1918609b3c5a1e793531019fcc73f4e6f207f4681c14646f499f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 81, + 488, + 211.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 211.66666666666666, + 488, + 342.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 342.3333333333333, + 488, + 473.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 479, + 505, + 524 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "Figure 4: Feature visualizations of a number of test episodes using the UMAP algorithm (McInnes", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "et al., 2018). Each row indicates a group of test extended episodes (the first four columns, rotation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 118, + 515 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 502, + 129, + 512 + ], + "score": 0.81, + "content": "0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 502, + 133, + 515 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 133, + 502, + 149, + 512 + ], + "score": 0.85, + "content": "9 0 °", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 502, + 153, + 515 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 153, + 502, + 173, + 512 + ], + "score": 0.84, + "content": "1 8 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 502, + 177, + 515 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 177, + 502, + 199, + 513 + ], + "score": 0.82, + "content": "2 7 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 502, + 505, + 515 + ], + "score": 1.0, + "content": ") and their integrated episode (the last column). The 5-way 5-shot FSL (with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 513, + 269, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 269, + 524 + ], + "score": 1.0, + "content": "Conv4-64) is adopted on miniImageNet.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "title", + "bbox": [ + 107, + 541, + 332, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 333, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 333, + 554 + ], + "score": 1.0, + "content": "A.3 FEATURE VISUALIZATIONS OF TEST EPISODES", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "We provide the feature visualizations of test episodes in Figure 4. It can be seen that an integrated", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "episode (the last one in each row) clearly has a better cluster data structure than the corresponding", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "four extended episodes (the first four ones in each row). This indicates that our transformer-based", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 597, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 482, + 610 + ], + "score": 1.0, + "content": "across-episode integration is indeed effective for few-shot classification with self-supervision.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 565, + 505, + 610 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 631, + 336, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 631, + 337, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 337, + 644 + ], + "score": 1.0, + "content": "A.4 ATTENTION VISUALIZATION OF TEST EPISODES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "score": 1.0, + "content": "We present attention map visualization of two test episodes (left and right) in Figure 5. Each average", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "attention map is computed by averaging the attention map of all instances of a certain class. We can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "observe that: (1) The average attention maps from different classes vary significantly, showing that the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "diverse semantics of different classes can be reflected by our attention-based integration transformer.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "(2) When the classes of two episodes overlap (e.g., ‘trifle’ and ‘dalmatian’), the average attention", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "maps of an overlapped class from two episodes are similar, illustrating that our attention-based", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 721, + 433, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 433, + 733 + ], + "score": 1.0, + "content": "integration transformer can well capture the semantics of classes across episodes.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 86, + 495, + 210 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 86, + 495, + 210 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 86, + 495, + 210 + ], + "spans": [ + { + "bbox": [ + 117, + 86, + 495, + 210 + ], + "score": 0.968, + "type": "image", + "image_path": "d5c83307373ad302b67daff2626249571668bdb71167a2ce318b0ac9aa79845d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 86, + 495, + 127.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 127.33333333333334, + 495, + 168.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 168.66666666666669, + 495, + 210.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 219, + 504, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "Figure 5: Feature and attention visualization of two test episodes (left and right). Both figures present", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 231, + 425, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 425, + 244 + ], + "score": 1.0, + "content": "5-way 5-shot results on miniImageNet, using Conv4-64 as the feature extractor.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 363, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 364, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 364, + 279 + ], + "score": 1.0, + "content": "A.5 COMPARISON WITH SIMPLE BASELINE FOR SSL+FSL", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 288, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 386, + 301 + ], + "score": 1.0, + "content": "We compare our IEPT with a simple baseline that trains the model with", + "type": "text" + }, + { + "bbox": [ + 386, + 289, + 441, + 300 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "and then makes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "inference by just averaging the outputs of different extended episodes. The results are shown in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "Table 6. It can be observed that the performance of our IEPT is much more effective than that of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 322, + 373, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 274, + 334 + ], + "score": 1.0, + "content": "simple integration, due to the extra use of", + "type": "text" + }, + { + "bbox": [ + 275, + 322, + 335, + 334 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { i n t e g } + \\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 322, + 373, + 334 + ], + "score": 1.0, + "content": "for FSL.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "table", + "bbox": [ + 111, + 380, + 499, + 423 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 357, + 503, + 380 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 408, + 370 + ], + "score": 1.0, + "content": "Table 6: Comparison with the simple baseline that trains the model with", + "type": "text" + }, + { + "bbox": [ + 408, + 357, + 465, + 368 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "and then", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 369, + 415, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 415, + 380 + ], + "score": 1.0, + "content": "makes inference by just averaging the outputs of different extended episodes.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 111, + 380, + 499, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 380, + 499, + 423 + ], + "spans": [ + { + "bbox": [ + 111, + 380, + 499, + 423 + ], + "score": 0.976, + "html": "
miniImageNettieredImageNet
MethodBackbone1-shot5-shot1-shot5-shot
Laux+LinstConv4-6453.25 ± 0.4671.50 ± 0.4255.06 ± 0.4472.87 ± 0.42
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "type": "table", + "image_path": "e2110b7904a72361e913dc8b2042d85b99fd274b6cad602dc6dcbf6ef7b2dea2.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 111, + 380, + 499, + 394.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 111, + 394.3333333333333, + 499, + 408.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 111, + 408.66666666666663, + 499, + 422.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.75 + }, + { + "type": "title", + "bbox": [ + 107, + 450, + 381, + 461 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 382, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 382, + 462 + ], + "score": 1.0, + "content": "A.6 DIFFERENT ALTERNATIVES OF SELF-SUPERVISED LOSSES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 420, + 486 + ], + "score": 1.0, + "content": "In Table 7, we provide further ablation study regarding different alternatives of", + "type": "text" + }, + { + "bbox": [ + 421, + 472, + 443, + 484 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 470, + 461, + 486 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 462, + 472, + 484, + 483 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 470, + 506, + 486 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 262, + 497 + ], + "score": 1.0, + "content": "the episode-level self-supervised loss", + "type": "text" + }, + { + "bbox": [ + 262, + 483, + 285, + 495 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 483, + 506, + 497 + ], + "score": 1.0, + "content": ", we compare our implementation (using the KL loss", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "between each distribution and the mean distribution) with that using a pairwise KL loss (i.e., the KL", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 504, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 428, + 519 + ], + "score": 1.0, + "content": "loss between each pair of distributions). For the instance-level self-supervised loss", + "type": "text" + }, + { + "bbox": [ + 429, + 505, + 451, + 516 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 504, + 506, + 519 + ], + "score": 1.0, + "content": ", we compare", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 514, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 530 + ], + "score": 1.0, + "content": "our implementation (using the rotation prediction loss) with the recent self-supervised learning", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 528, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 539 + ], + "score": 1.0, + "content": "technique (Chen et al., 2020). We observe that our implementation achieves slight performance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 538, + 444, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 444, + 551 + ], + "score": 1.0, + "content": "improvements over those using the pairwise KL loss or the contrastive learning loss.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 117, + 585, + 493, + 639 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 137, + 572, + 466, + 584 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 138, + 568, + 465, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 401, + 588 + ], + "score": 1.0, + "content": "Table 7: Ablation study results regarding different alternatives of", + "type": "text" + }, + { + "bbox": [ + 401, + 573, + 424, + 585 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 568, + 442, + 588 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 442, + 573, + 465, + 583 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "table_body", + "bbox": [ + 117, + 585, + 493, + 639 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 585, + 493, + 639 + ], + "spans": [ + { + "bbox": [ + 117, + 585, + 493, + 639 + ], + "score": 0.978, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
IEPT (Linst-SimCLR)Conv4-6456.04 ± 0.4473.67± 0.4158.24± 0.4375.59± 0.41
IEPT (Lepis-Pairwise)Conv4-6455.95 ± 0.4773.72 ±0.4057.91 ± 0.4575.28±0.40
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "type": "table", + "image_path": "2d0670d600723f1311e4dff16c14fd4b30f96753bca322cdab0c9ab09060b90a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 117, + 585, + 493, + 603.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 117, + 603.0, + 493, + 621.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 117, + 621.0, + 493, + 639.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 108, + 664, + 421, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 423, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 423, + 678 + ], + "score": 1.0, + "content": "A.7 APPLICATION OF IEPT TO OPTIMIZATION-BASED METHOD MAML", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "In Table 8, we show the results obtained by applying our IEPT to the optimization-based model", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "MAML (Finn et al., 2017). We use Conv4-64 as the feature extractor. We can see that our IEPT brings", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 153, + 720 + ], + "score": 0.87, + "content": "0 . 7 \\% - 2 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "improvements to MAML. This further shows the flexibility (as well as effectiveness) of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 182, + 731 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 182, + 731 + ], + "score": 1.0, + "content": "our IEPT for FSL.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 86, + 495, + 210 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 86, + 495, + 210 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 86, + 495, + 210 + ], + "spans": [ + { + "bbox": [ + 117, + 86, + 495, + 210 + ], + "score": 0.968, + "type": "image", + "image_path": "d5c83307373ad302b67daff2626249571668bdb71167a2ce318b0ac9aa79845d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 86, + 495, + 127.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 127.33333333333334, + 495, + 168.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 168.66666666666669, + 495, + 210.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 219, + 504, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "Figure 5: Feature and attention visualization of two test episodes (left and right). Both figures present", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 231, + 425, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 425, + 244 + ], + "score": 1.0, + "content": "5-way 5-shot results on miniImageNet, using Conv4-64 as the feature extractor.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 363, + 278 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 364, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 364, + 279 + ], + "score": 1.0, + "content": "A.5 COMPARISON WITH SIMPLE BASELINE FOR SSL+FSL", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 266, + 364, + 279 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 288, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 386, + 301 + ], + "score": 1.0, + "content": "We compare our IEPT with a simple baseline that trains the model with", + "type": "text" + }, + { + "bbox": [ + 386, + 289, + 441, + 300 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "and then makes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "inference by just averaging the outputs of different extended episodes. The results are shown in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "Table 6. It can be observed that the performance of our IEPT is much more effective than that of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 322, + 373, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 274, + 334 + ], + "score": 1.0, + "content": "simple integration, due to the extra use of", + "type": "text" + }, + { + "bbox": [ + 275, + 322, + 335, + 334 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { i n t e g } + \\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 322, + 373, + 334 + ], + "score": 1.0, + "content": "for FSL.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 288, + 506, + 334 + ] + }, + { + "type": "table", + "bbox": [ + 111, + 380, + 499, + 423 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 357, + 503, + 380 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 408, + 370 + ], + "score": 1.0, + "content": "Table 6: Comparison with the simple baseline that trains the model with", + "type": "text" + }, + { + "bbox": [ + 408, + 357, + 465, + 368 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "and then", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 369, + 415, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 415, + 380 + ], + "score": 1.0, + "content": "makes inference by just averaging the outputs of different extended episodes.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 111, + 380, + 499, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 380, + 499, + 423 + ], + "spans": [ + { + "bbox": [ + 111, + 380, + 499, + 423 + ], + "score": 0.976, + "html": "
miniImageNettieredImageNet
MethodBackbone1-shot5-shot1-shot5-shot
Laux+LinstConv4-6453.25 ± 0.4671.50 ± 0.4255.06 ± 0.4472.87 ± 0.42
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "type": "table", + "image_path": "e2110b7904a72361e913dc8b2042d85b99fd274b6cad602dc6dcbf6ef7b2dea2.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 111, + 380, + 499, + 394.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 111, + 394.3333333333333, + 499, + 408.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 111, + 408.66666666666663, + 499, + 422.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.75 + }, + { + "type": "title", + "bbox": [ + 107, + 450, + 381, + 461 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 382, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 382, + 462 + ], + "score": 1.0, + "content": "A.6 DIFFERENT ALTERNATIVES OF SELF-SUPERVISED LOSSES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 420, + 486 + ], + "score": 1.0, + "content": "In Table 7, we provide further ablation study regarding different alternatives of", + "type": "text" + }, + { + "bbox": [ + 421, + 472, + 443, + 484 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 470, + 461, + 486 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 462, + 472, + 484, + 483 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 470, + 506, + 486 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 262, + 497 + ], + "score": 1.0, + "content": "the episode-level self-supervised loss", + "type": "text" + }, + { + "bbox": [ + 262, + 483, + 285, + 495 + ], + "score": 0.92, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 483, + 506, + 497 + ], + "score": 1.0, + "content": ", we compare our implementation (using the KL loss", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "between each distribution and the mean distribution) with that using a pairwise KL loss (i.e., the KL", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 504, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 428, + 519 + ], + "score": 1.0, + "content": "loss between each pair of distributions). For the instance-level self-supervised loss", + "type": "text" + }, + { + "bbox": [ + 429, + 505, + 451, + 516 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 504, + 506, + 519 + ], + "score": 1.0, + "content": ", we compare", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 514, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 530 + ], + "score": 1.0, + "content": "our implementation (using the rotation prediction loss) with the recent self-supervised learning", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 528, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 539 + ], + "score": 1.0, + "content": "technique (Chen et al., 2020). We observe that our implementation achieves slight performance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 538, + 444, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 444, + 551 + ], + "score": 1.0, + "content": "improvements over those using the pairwise KL loss or the contrastive learning loss.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 470, + 506, + 551 + ] + }, + { + "type": "table", + "bbox": [ + 117, + 585, + 493, + 639 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 137, + 572, + 466, + 584 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 138, + 568, + 465, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 401, + 588 + ], + "score": 1.0, + "content": "Table 7: Ablation study results regarding different alternatives of", + "type": "text" + }, + { + "bbox": [ + 401, + 573, + 424, + 585 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { e p i s }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 568, + 442, + 588 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 442, + 573, + 465, + 583 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { i n s t }", + "type": "inline_equation" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "table_body", + "bbox": [ + 117, + 585, + 493, + 639 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 585, + 493, + 639 + ], + "spans": [ + { + "bbox": [ + 117, + 585, + 493, + 639 + ], + "score": 0.978, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
IEPT (Linst-SimCLR)Conv4-6456.04 ± 0.4473.67± 0.4158.24± 0.4375.59± 0.41
IEPT (Lepis-Pairwise)Conv4-6455.95 ± 0.4773.72 ±0.4057.91 ± 0.4575.28±0.40
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
", + "type": "table", + "image_path": "2d0670d600723f1311e4dff16c14fd4b30f96753bca322cdab0c9ab09060b90a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 117, + 585, + 493, + 603.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 117, + 603.0, + 493, + 621.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 117, + 621.0, + 493, + 639.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 108, + 664, + 421, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 664, + 423, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 423, + 678 + ], + "score": 1.0, + "content": "A.7 APPLICATION OF IEPT TO OPTIMIZATION-BASED METHOD MAML", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "In Table 8, we show the results obtained by applying our IEPT to the optimization-based model", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "MAML (Finn et al., 2017). We use Conv4-64 as the feature extractor. We can see that our IEPT brings", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 153, + 720 + ], + "score": 0.87, + "content": "0 . 7 \\% - 2 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "improvements to MAML. This further shows the flexibility (as well as effectiveness) of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 182, + 731 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 182, + 731 + ], + "score": 1.0, + "content": "our IEPT for FSL.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 687, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 101, + 488, + 144 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 123, + 89, + 486, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 88, + 487, + 103 + ], + "spans": [ + { + "bbox": [ + 123, + 88, + 487, + 103 + ], + "score": 1.0, + "content": "Table 8: Comparative results by applying IEPT to the optimization-based method MAML.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 122, + 101, + 488, + 144 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 101, + 488, + 144 + ], + "spans": [ + { + "bbox": [ + 122, + 101, + 488, + 144 + ], + "score": 0.971, + "html": "
BackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
Method MAMLConv4-6448.70±1.8463.10±0.9251.67 ± 1.8170.30±0.80
MAML+IEPTConv4-6449.68 ± 0.5065.22 ± 0.4852.85 ± 0.5271.04 ± 0.49
", + "type": "table", + "image_path": "84b9c3b35da25fb4560691fc1312be2bd5050c34561bd624d14f308488569d31.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 122, + 101, + 488, + 115.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 115.33333333333333, + 488, + 129.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 129.66666666666666, + 488, + 144.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 114, + 159, + 489, + 271 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 159, + 489, + 271 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 159, + 489, + 271 + ], + "spans": [ + { + "bbox": [ + 114, + 159, + 489, + 271 + ], + "score": 0.966, + "type": "image", + "image_path": "19a2a9e135d017944a544b6d105ac0bab71b8611d3f98f7456e4da154a02510e.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 114, + 159, + 489, + 196.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 114, + 196.33333333333334, + 489, + 233.66666666666669 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 233.66666666666669, + 489, + 271.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 272, + 504, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 6: Visualization of our hyper-parameter analysis under 5-way 1-shot (left) and 5-shot (right)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 283, + 349, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 349, + 295 + ], + "score": 1.0, + "content": "on miniImageNet. Conv4-64 is used as the feature extractor.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "title", + "bbox": [ + 107, + 315, + 302, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 304, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 304, + 327 + ], + "score": 1.0, + "content": "A.8 HYPER-PARAMETER SENSITIVITY TEST", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 235, + 349 + ], + "score": 1.0, + "content": "We select the hyper-parameters", + "type": "text" + }, + { + "bbox": [ + 235, + 338, + 248, + 347 + ], + "score": 0.82, + "content": "w _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 335, + 252, + 349 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 252, + 338, + 264, + 347 + ], + "score": 0.77, + "content": "w _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 335, + 282, + 349 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 283, + 337, + 296, + 347 + ], + "score": 0.86, + "content": "w _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 335, + 389, + 349 + ], + "score": 1.0, + "content": "from the candidate set", + "type": "text" + }, + { + "bbox": [ + 389, + 335, + 487, + 347 + ], + "score": 0.9, + "content": "\\{ 0 . 1 , 0 . 5 , 1 . 0 , 5 . 0 , 1 0 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "show the hyper-parameter analysis results in Figure 6. We find that the performance of our IEPT is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 411, + 371 + ], + "score": 1.0, + "content": "relatively stable. Concretely, the performance of our IEPT is not sensitive to", + "type": "text" + }, + { + "bbox": [ + 412, + 359, + 424, + 369 + ], + "score": 0.86, + "content": "w _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 356, + 442, + 371 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 442, + 359, + 456, + 369 + ], + "score": 0.87, + "content": "w _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 356, + 506, + 371 + ], + "score": 1.0, + "content": "with proper", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 187, + 381 + ], + "score": 1.0, + "content": "values, but too large", + "type": "text" + }, + { + "bbox": [ + 187, + 370, + 199, + 380 + ], + "score": 0.83, + "content": "w _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 368, + 217, + 381 + ], + "score": 1.0, + "content": "(i.e.", + "type": "text" + }, + { + "bbox": [ + 217, + 369, + 262, + 380 + ], + "score": 0.87, + "content": "w _ { 1 } = 1 0 . 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 368, + 506, + 381 + ], + "score": 1.0, + "content": ") tends to cause obvious degradation, perhaps because the FSL", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 379, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 286, + 393 + ], + "score": 1.0, + "content": "task is biased by the rotation prediction loss.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 101, + 488, + 144 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 123, + 89, + 486, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 88, + 487, + 103 + ], + "spans": [ + { + "bbox": [ + 123, + 88, + 487, + 103 + ], + "score": 1.0, + "content": "Table 8: Comparative results by applying IEPT to the optimization-based method MAML.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 122, + 101, + 488, + 144 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 101, + 488, + 144 + ], + "spans": [ + { + "bbox": [ + 122, + 101, + 488, + 144 + ], + "score": 0.971, + "html": "
BackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
Method MAMLConv4-6448.70±1.8463.10±0.9251.67 ± 1.8170.30±0.80
MAML+IEPTConv4-6449.68 ± 0.5065.22 ± 0.4852.85 ± 0.5271.04 ± 0.49
", + "type": "table", + "image_path": "84b9c3b35da25fb4560691fc1312be2bd5050c34561bd624d14f308488569d31.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 122, + 101, + 488, + 115.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 115.33333333333333, + 488, + 129.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 122, + 129.66666666666666, + 488, + 144.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 114, + 159, + 489, + 271 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 159, + 489, + 271 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 159, + 489, + 271 + ], + "spans": [ + { + "bbox": [ + 114, + 159, + 489, + 271 + ], + "score": 0.966, + "type": "image", + "image_path": "19a2a9e135d017944a544b6d105ac0bab71b8611d3f98f7456e4da154a02510e.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 114, + 159, + 489, + 196.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 114, + 196.33333333333334, + 489, + 233.66666666666669 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 114, + 233.66666666666669, + 489, + 271.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 272, + 504, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 6: Visualization of our hyper-parameter analysis under 5-way 1-shot (left) and 5-shot (right)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 283, + 349, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 349, + 295 + ], + "score": 1.0, + "content": "on miniImageNet. Conv4-64 is used as the feature extractor.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "title", + "bbox": [ + 107, + 315, + 302, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 304, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 304, + 327 + ], + "score": 1.0, + "content": "A.8 HYPER-PARAMETER SENSITIVITY TEST", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 235, + 349 + ], + "score": 1.0, + "content": "We select the hyper-parameters", + "type": "text" + }, + { + "bbox": [ + 235, + 338, + 248, + 347 + ], + "score": 0.82, + "content": "w _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 335, + 252, + 349 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 252, + 338, + 264, + 347 + ], + "score": 0.77, + "content": "w _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 335, + 282, + 349 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 283, + 337, + 296, + 347 + ], + "score": 0.86, + "content": "w _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 335, + 389, + 349 + ], + "score": 1.0, + "content": "from the candidate set", + "type": "text" + }, + { + "bbox": [ + 389, + 335, + 487, + 347 + ], + "score": 0.9, + "content": "\\{ 0 . 1 , 0 . 5 , 1 . 0 , 5 . 0 , 1 0 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "show the hyper-parameter analysis results in Figure 6. We find that the performance of our IEPT is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 411, + 371 + ], + "score": 1.0, + "content": "relatively stable. Concretely, the performance of our IEPT is not sensitive to", + "type": "text" + }, + { + "bbox": [ + 412, + 359, + 424, + 369 + ], + "score": 0.86, + "content": "w _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 356, + 442, + 371 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 442, + 359, + 456, + 369 + ], + "score": 0.87, + "content": "w _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 356, + 506, + 371 + ], + "score": 1.0, + "content": "with proper", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 187, + 381 + ], + "score": 1.0, + "content": "values, but too large", + "type": "text" + }, + { + "bbox": [ + 187, + 370, + 199, + 380 + ], + "score": 0.83, + "content": "w _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 368, + 217, + 381 + ], + "score": 1.0, + "content": "(i.e.", + "type": "text" + }, + { + "bbox": [ + 217, + 369, + 262, + 380 + ], + "score": 0.87, + "content": "w _ { 1 } = 1 0 . 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 368, + 506, + 381 + ], + "score": 1.0, + "content": ") tends to cause obvious degradation, perhaps because the FSL", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 379, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 286, + 393 + ], + "score": 1.0, + "content": "task is biased by the rotation prediction loss.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 335, + 506, + 393 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_model.json b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d4f2122615151f71d527014aa2e6f93b1e9f86f --- /dev/null +++ b/parse/train/xzqLpqRzxLq/xzqLpqRzxLq_model.json @@ -0,0 +1,23462 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 925, + 1304, + 925, + 1304, + 1474, + 398, + 1474 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1601, + 1404, + 1601, + 1404, + 1878, + 298, + 1878 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 315, + 375, + 1300, + 375, + 1300, + 499, + 315, + 499 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 300, + 1893, + 1403, + 1893, + 1403, + 1986, + 300, + 1986 + ], + "score": 0.964 + }, + { + "category_id": 0, + "poly": [ + 299, + 219, + 1407, + 219, + 1407, + 323, + 299, + 323 + ], + "score": 0.962 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 567, + 2006, + 567, + 2033, + 333, + 2033 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 302, + 1533, + 573, + 1533, + 573, + 1568, + 302, + 1568 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 313, + 546, + 806, + 546, + 806, + 637, + 313, + 637 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 104, + 300, + 104 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 773, + 858, + 926, + 858, + 926, + 891, + 773, + 891 + ], + "score": 0.853 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 856, + 2088, + 856, + 2112, + 840, + 2112 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 906, + 547, + 1386, + 547, + 1386, + 637, + 906, + 637 + ], + "score": 0.689 + }, + { + "category_id": 1, + "poly": [ + 315, + 718, + 815, + 718, + 815, + 776, + 315, + 776 + ], + "score": 0.638 + }, + { + "category_id": 0, + "poly": [ + 315, + 687, + 516, + 687, + 516, + 717, + 315, + 717 + ], + "score": 0.567 + }, + { + "category_id": 0, + "poly": [ + 907, + 548, + 1067, + 548, + 1067, + 577, + 907, + 577 + ], + "score": 0.352 + }, + { + "category_id": 0, + "poly": [ + 315, + 547, + 440, + 547, + 440, + 577, + 315, + 577 + ], + "score": 0.113 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 216.0, + 1412.0, + 216.0, + 1412.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 274.0, + 1116.0, + 274.0, + 1116.0, + 327.0, + 292.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1998.0, + 572.0, + 1998.0, + 572.0, + 2041.0, + 329.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 579.0, + 1530.0, + 579.0, + 1577.0, + 294.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 857.0, + 932.0, + 857.0, + 932.0, + 894.0, + 769.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 680.0, + 521.0, + 680.0, + 521.0, + 725.0, + 311.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 541.0, + 1073.0, + 541.0, + 1073.0, + 587.0, + 902.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 540.0, + 445.0, + 540.0, + 445.0, + 586.0, + 310.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 927.0, + 1305.0, + 927.0, + 1305.0, + 960.0, + 395.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 956.0, + 1305.0, + 956.0, + 1305.0, + 991.0, + 393.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 988.0, + 1306.0, + 988.0, + 1306.0, + 1020.0, + 394.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1018.0, + 1305.0, + 1018.0, + 1305.0, + 1053.0, + 393.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1049.0, + 1306.0, + 1049.0, + 1306.0, + 1085.0, + 393.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1079.0, + 1307.0, + 1079.0, + 1307.0, + 1114.0, + 393.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1109.0, + 1307.0, + 1109.0, + 1307.0, + 1144.0, + 393.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1142.0, + 1305.0, + 1142.0, + 1305.0, + 1175.0, + 394.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1170.0, + 1307.0, + 1170.0, + 1307.0, + 1205.0, + 393.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1201.0, + 1305.0, + 1201.0, + 1305.0, + 1236.0, + 393.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1232.0, + 1308.0, + 1232.0, + 1308.0, + 1268.0, + 393.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1260.0, + 1306.0, + 1260.0, + 1306.0, + 1296.0, + 393.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1292.0, + 1305.0, + 1292.0, + 1305.0, + 1325.0, + 394.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1323.0, + 1306.0, + 1323.0, + 1306.0, + 1355.0, + 393.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1354.0, + 1306.0, + 1354.0, + 1306.0, + 1386.0, + 395.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1382.0, + 1305.0, + 1382.0, + 1305.0, + 1418.0, + 393.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1413.0, + 1305.0, + 1413.0, + 1305.0, + 1449.0, + 393.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1445.0, + 1088.0, + 1445.0, + 1088.0, + 1478.0, + 395.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1603.0, + 1406.0, + 1603.0, + 1406.0, + 1636.0, + 296.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1630.0, + 1406.0, + 1630.0, + 1406.0, + 1670.0, + 293.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1404.0, + 1665.0, + 1404.0, + 1698.0, + 294.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1694.0, + 1406.0, + 1694.0, + 1406.0, + 1730.0, + 293.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1722.0, + 1408.0, + 1722.0, + 1408.0, + 1763.0, + 292.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1753.0, + 1406.0, + 1753.0, + 1406.0, + 1791.0, + 293.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1819.0, + 296.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1811.0, + 1409.0, + 1811.0, + 1409.0, + 1852.0, + 293.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1848.0, + 967.0, + 1848.0, + 967.0, + 1881.0, + 296.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 374.0, + 856.0, + 374.0, + 856.0, + 409.0, + 310.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 404.0, + 1272.0, + 404.0, + 1272.0, + 442.0, + 310.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 436.0, + 1301.0, + 436.0, + 1301.0, + 473.0, + 309.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 467.0, + 979.0, + 467.0, + 979.0, + 503.0, + 312.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1890.0, + 1408.0, + 1890.0, + 1408.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1923.0, + 1405.0, + 1923.0, + 1405.0, + 1958.0, + 294.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1952.0, + 1404.0, + 1952.0, + 1404.0, + 1991.0, + 293.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 541.0, + 445.0, + 541.0, + 445.0, + 583.0, + 312.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 576.0, + 806.0, + 576.0, + 806.0, + 609.0, + 313.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 608.0, + 651.0, + 608.0, + 651.0, + 639.0, + 314.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 543.0, + 1072.0, + 543.0, + 1072.0, + 584.0, + 903.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 571.0, + 1388.0, + 571.0, + 1388.0, + 615.0, + 903.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 608.0, + 1195.0, + 608.0, + 1195.0, + 640.0, + 903.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 714.0, + 817.0, + 714.0, + 817.0, + 748.0, + 313.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 749.0, + 785.0, + 749.0, + 785.0, + 780.0, + 313.0, + 780.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1286, + 1405, + 1286, + 1405, + 1651, + 298, + 1651 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1668, + 1404, + 1668, + 1404, + 2033, + 298, + 2033 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1025, + 1404, + 1025, + 1404, + 1271, + 297, + 1271 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 887, + 1403, + 887, + 1403, + 1009, + 299, + 1009 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 297, + 520, + 1405, + 520, + 1405, + 856, + 297, + 856 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 309, + 247, + 1385, + 247, + 1385, + 505, + 309, + 505 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.731 + }, + { + "category_id": 13, + "poly": [ + 457, + 826, + 532, + 826, + 532, + 859, + 457, + 859 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 549, + 795, + 623, + 795, + 623, + 829, + 549, + 829 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 816, + 704, + 879, + 704, + 879, + 737, + 816, + 737 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 1009, + 795, + 1073, + 795, + 1073, + 830, + 1009, + 830 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 345, + 826, + 405, + 826, + 405, + 856, + 345, + 856 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 1209, + 613, + 1272, + 613, + 1272, + 643, + 1209, + 643 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 705, + 1324, + 705, + 1324, + 734, + 1263, + 734 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 895, + 795, + 959, + 795, + 959, + 826, + 895, + 826 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 551, + 920, + 571, + 920, + 571, + 946, + 551, + 946 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 397, + 924, + 417, + 924, + 417, + 946, + 397, + 946 + ], + "score": 0.7, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 518.0, + 1406.0, + 518.0, + 1406.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 548.0, + 1406.0, + 548.0, + 1406.0, + 590.0, + 292.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 581.0, + 1407.0, + 581.0, + 1407.0, + 616.0, + 295.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 611.0, + 1208.0, + 611.0, + 1208.0, + 650.0, + 294.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 650.0, + 1273.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 677.0, + 295.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 1406.0, + 673.0, + 1406.0, + 708.0, + 294.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 815.0, + 703.0, + 815.0, + 741.0, + 294.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 703.0, + 1262.0, + 703.0, + 1262.0, + 741.0, + 880.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 703.0, + 1406.0, + 703.0, + 1406.0, + 741.0, + 1325.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 734.0, + 1406.0, + 734.0, + 1406.0, + 767.0, + 294.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 764.0, + 1405.0, + 764.0, + 1405.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 795.0, + 548.0, + 795.0, + 548.0, + 833.0, + 294.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 795.0, + 894.0, + 795.0, + 894.0, + 833.0, + 624.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 795.0, + 1008.0, + 795.0, + 1008.0, + 833.0, + 960.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 795.0, + 1410.0, + 795.0, + 1410.0, + 833.0, + 1074.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 822.0, + 344.0, + 822.0, + 344.0, + 864.0, + 294.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 822.0, + 456.0, + 822.0, + 456.0, + 864.0, + 406.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 822.0, + 823.0, + 822.0, + 823.0, + 864.0, + 533.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 242.0, + 427.0, + 242.0, + 427.0, + 258.0, + 312.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 275.0, + 1021.0, + 275.0, + 1021.0, + 295.0, + 908.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 269.0, + 1207.0, + 269.0, + 1207.0, + 302.0, + 1055.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 330.0, + 841.0, + 330.0, + 841.0, + 341.0, + 832.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 346.0, + 843.0, + 346.0, + 843.0, + 375.0, + 831.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 357.0, + 1004.0, + 357.0, + 1004.0, + 381.0, + 919.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 356.0, + 1359.0, + 356.0, + 1359.0, + 372.0, + 1259.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 372.0, + 765.0, + 372.0, + 765.0, + 391.0, + 734.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 378.0, + 845.0, + 378.0, + 845.0, + 411.0, + 829.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 378.0, + 994.0, + 378.0, + 994.0, + 400.0, + 932.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 374.0, + 1375.0, + 374.0, + 1375.0, + 407.0, + 1240.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 397.0, + 1353.0, + 397.0, + 1353.0, + 417.0, + 1262.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 411.0, + 845.0, + 411.0, + 845.0, + 443.0, + 829.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 446.0, + 825.0, + 446.0, + 825.0, + 472.0, + 783.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 446.0, + 864.0, + 446.0, + 864.0, + 473.0, + 827.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 449.0, + 1002.0, + 449.0, + 1002.0, + 473.0, + 922.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 453.0, + 1086.0, + 453.0, + 1086.0, + 472.0, + 1058.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 439.0, + 1167.0, + 439.0, + 1167.0, + 456.0, + 1119.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 468.0, + 1023.0, + 468.0, + 1023.0, + 492.0, + 907.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 468.0, + 1103.0, + 468.0, + 1103.0, + 489.0, + 1042.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 470.0, + 1374.0, + 470.0, + 1374.0, + 490.0, + 1248.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 334.0, + 1382.0, + 334.0, + 1382.0, + 362.0, + 1235.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 448.5, + 1380.0, + 448.5, + 1380.0, + 477.5, + 1239.0, + 477.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 1406.0, + 1283.0, + 1406.0, + 1323.0, + 293.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1317.0, + 1406.0, + 1317.0, + 1406.0, + 1351.0, + 295.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1347.0, + 1407.0, + 1347.0, + 1407.0, + 1381.0, + 295.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1378.0, + 1408.0, + 1378.0, + 1408.0, + 1412.0, + 295.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1404.0, + 1406.0, + 1404.0, + 1406.0, + 1444.0, + 291.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1437.0, + 1406.0, + 1437.0, + 1406.0, + 1473.0, + 291.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1470.0, + 1405.0, + 1470.0, + 1405.0, + 1504.0, + 295.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1498.0, + 1408.0, + 1498.0, + 1408.0, + 1536.0, + 292.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1529.0, + 1410.0, + 1529.0, + 1410.0, + 1565.0, + 292.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1557.0, + 1406.0, + 1557.0, + 1406.0, + 1597.0, + 292.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1591.0, + 1403.0, + 1591.0, + 1403.0, + 1625.0, + 295.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1621.0, + 1217.0, + 1621.0, + 1217.0, + 1655.0, + 295.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1665.0, + 1408.0, + 1665.0, + 1408.0, + 1705.0, + 292.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1697.0, + 1406.0, + 1697.0, + 1406.0, + 1733.0, + 293.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1723.0, + 1405.0, + 1723.0, + 1405.0, + 1767.0, + 292.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 1402.0, + 1760.0, + 1402.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1404.0, + 1790.0, + 1404.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1408.0, + 1821.0, + 1408.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 1405.0, + 1852.0, + 1405.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1408.0, + 1973.0, + 1408.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1025.0, + 1407.0, + 1025.0, + 1407.0, + 1059.0, + 295.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1053.0, + 1406.0, + 1053.0, + 1406.0, + 1092.0, + 294.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1087.0, + 1402.0, + 1087.0, + 1402.0, + 1120.0, + 296.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1116.0, + 1407.0, + 1116.0, + 1407.0, + 1154.0, + 292.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1404.0, + 1148.0, + 1404.0, + 1182.0, + 295.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 1405.0, + 1178.0, + 1405.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1209.0, + 1405.0, + 1209.0, + 1405.0, + 1243.0, + 295.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1238.0, + 1190.0, + 1238.0, + 1190.0, + 1277.0, + 292.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 889.0, + 1404.0, + 889.0, + 1404.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 919.0, + 396.0, + 919.0, + 396.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 919.0, + 550.0, + 919.0, + 550.0, + 951.0, + 418.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 919.0, + 1407.0, + 919.0, + 1407.0, + 951.0, + 572.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 947.0, + 1407.0, + 947.0, + 1407.0, + 984.0, + 293.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 978.0, + 877.0, + 978.0, + 877.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 721, + 1406, + 721, + 1406, + 1302, + 298, + 1302 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1607, + 1405, + 1607, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1315, + 1404, + 1315, + 1404, + 1592, + 298, + 1592 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1403, + 229, + 1403, + 353, + 299, + 353 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 367, + 1405, + 367, + 1405, + 583, + 298, + 583 + ], + "score": 0.969 + }, + { + "category_id": 0, + "poly": [ + 301, + 641, + 586, + 641, + 586, + 678, + 301, + 678 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.583 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2112, + 842, + 2112 + ], + "score": 0.437 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 639.0, + 593.0, + 639.0, + 593.0, + 683.0, + 292.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 717.0, + 1406.0, + 717.0, + 1406.0, + 759.0, + 292.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 752.0, + 1407.0, + 752.0, + 1407.0, + 787.0, + 295.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 1406.0, + 781.0, + 1406.0, + 816.0, + 293.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 1404.0, + 812.0, + 1404.0, + 848.0, + 295.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 842.0, + 1407.0, + 842.0, + 1407.0, + 879.0, + 293.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 874.0, + 1406.0, + 874.0, + 1406.0, + 909.0, + 293.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 904.0, + 1408.0, + 904.0, + 1408.0, + 939.0, + 295.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 933.0, + 1406.0, + 933.0, + 1406.0, + 970.0, + 293.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 965.0, + 1408.0, + 965.0, + 1408.0, + 1000.0, + 295.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 991.0, + 1409.0, + 991.0, + 1409.0, + 1033.0, + 292.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1027.0, + 1406.0, + 1027.0, + 1406.0, + 1062.0, + 295.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 1407.0, + 1054.0, + 1407.0, + 1092.0, + 293.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1084.0, + 1406.0, + 1084.0, + 1406.0, + 1120.0, + 293.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1114.0, + 1408.0, + 1114.0, + 1408.0, + 1154.0, + 291.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1145.0, + 1406.0, + 1145.0, + 1406.0, + 1183.0, + 293.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1178.0, + 1406.0, + 1178.0, + 1406.0, + 1213.0, + 295.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1208.0, + 1406.0, + 1208.0, + 1406.0, + 1243.0, + 293.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1239.0, + 1406.0, + 1239.0, + 1406.0, + 1274.0, + 293.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1271.0, + 885.0, + 1271.0, + 885.0, + 1306.0, + 295.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1605.0, + 1408.0, + 1605.0, + 1408.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1637.0, + 1408.0, + 1637.0, + 1408.0, + 1672.0, + 295.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1667.0, + 1408.0, + 1667.0, + 1408.0, + 1704.0, + 292.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1696.0, + 1405.0, + 1696.0, + 1405.0, + 1734.0, + 293.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1725.0, + 1406.0, + 1725.0, + 1406.0, + 1767.0, + 292.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1757.0, + 1406.0, + 1757.0, + 1406.0, + 1798.0, + 292.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1827.0, + 293.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1819.0, + 1407.0, + 1819.0, + 1407.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1852.0, + 1405.0, + 1852.0, + 1405.0, + 1884.0, + 296.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1408.0, + 1880.0, + 1408.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 1406.0, + 1913.0, + 1406.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 860.0, + 2003.0, + 860.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1350.0, + 296.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1344.0, + 1408.0, + 1344.0, + 1408.0, + 1385.0, + 292.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1375.0, + 1409.0, + 1375.0, + 1409.0, + 1413.0, + 293.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1402.0, + 1408.0, + 1402.0, + 1408.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1436.0, + 1405.0, + 1436.0, + 1405.0, + 1476.0, + 293.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1468.0, + 1407.0, + 1468.0, + 1407.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1533.0, + 295.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1404.0, + 1529.0, + 1404.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1560.0, + 1204.0, + 1560.0, + 1204.0, + 1596.0, + 292.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1407.0, + 261.0, + 1407.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1407.0, + 290.0, + 1407.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1081.0, + 321.0, + 1081.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 369.0, + 1405.0, + 369.0, + 1405.0, + 403.0, + 296.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 400.0, + 1403.0, + 400.0, + 1403.0, + 434.0, + 293.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 429.0, + 1407.0, + 429.0, + 1407.0, + 464.0, + 295.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1406.0, + 459.0, + 1406.0, + 495.0, + 293.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 491.0, + 1407.0, + 491.0, + 1407.0, + 526.0, + 295.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 1410.0, + 519.0, + 1410.0, + 559.0, + 293.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 553.0, + 1303.0, + 553.0, + 1303.0, + 587.0, + 295.0, + 587.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1237, + 1405, + 1237, + 1405, + 1639, + 295, + 1639 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 296, + 330, + 1405, + 330, + 1405, + 669, + 296, + 669 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 1653, + 1406, + 1653, + 1406, + 1871, + 296, + 1871 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 683, + 1403, + 683, + 1403, + 807, + 298, + 807 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 920, + 1402, + 920, + 1402, + 1014, + 298, + 1014 + ], + "score": 0.978 + }, + { + "category_id": 8, + "poly": [ + 634, + 821, + 1063, + 821, + 1063, + 907, + 634, + 907 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 466, + 1028, + 1230, + 1028, + 1230, + 1117, + 466, + 1117 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 429, + 1887, + 1268, + 1887, + 1268, + 1985, + 429, + 1985 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 295, + 1132, + 991, + 1132, + 991, + 1166, + 295, + 1166 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 291, + 2000, + 1367, + 2000, + 1367, + 2038, + 291, + 2038 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 298, + 1181, + 667, + 1181, + 667, + 1213, + 298, + 1213 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 580, + 226, + 580, + 261, + 300, + 261 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 275, + 535, + 275, + 535, + 307, + 299, + 307 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 840, + 1400, + 840, + 1400, + 870, + 1366, + 870 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1051, + 1400, + 1051, + 1400, + 1080, + 1366, + 1080 + ], + "score": 0.871 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1917, + 1400, + 1917, + 1400, + 1946, + 1366, + 1946 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 857, + 2088, + 857, + 2111, + 840, + 2111 + ], + "score": 0.795 + }, + { + "category_id": 14, + "poly": [ + 634, + 819, + 1062, + 819, + 1062, + 908, + 634, + 908 + ], + "score": 0.94, + "latex": "h _ { c } = \\frac { 1 } { k } \\cdot \\sum _ { ( x _ { i } , y _ { i } ) \\in S _ { e } } f _ { \\phi } ( x _ { i } ) \\cdot \\mathbb { I } ( y _ { i } = c ) ," + }, + { + "category_id": 14, + "poly": [ + 430, + 1883, + 1265, + 1883, + 1265, + 1984, + 430, + 1984 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { i n s t } = \\frac { 1 } { R ( l _ { k } + l _ { q } ) } \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { ( x _ { i } , y _ { i } , r _ { i } ) \\in E _ { e } ^ { r } } - \\log \\frac { \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r _ { i } } ) } { \\sum _ { r ^ { \\prime } = 0 } ^ { R - 1 } \\exp ( [ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] _ { r ^ { \\prime } } ) } ," + }, + { + "category_id": 14, + "poly": [ + 466, + 1027, + 1231, + 1027, + 1231, + 1118, + 466, + 1118 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { f s l } ( S _ { e } , \\mathcal { Q } _ { e } ) = \\frac { 1 } { | \\mathcal { Q } _ { e } | } \\sum _ { ( x _ { i } , y _ { i } ) \\in \\mathcal { Q } _ { e } } - \\log \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { y _ { i } } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ) ) } ," + }, + { + "category_id": 13, + "poly": [ + 372, + 2000, + 611, + 2000, + 611, + 2038, + 372, + 2038 + ], + "score": 0.93, + "latex": "[ f _ { \\theta _ { r o t } } ( f _ { \\phi } ( x _ { i } ) ) ] \\in \\mathbb { R } ^ { R }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1330, + 481, + 1330, + 481, + 1364, + 298, + 1364 + ], + "score": 0.93, + "latex": "E _ { e } = \\{ S _ { e } , \\mathcal { Q } _ { e } \\}" + }, + { + "category_id": 13, + "poly": [ + 767, + 1485, + 872, + 1485, + 872, + 1519, + 767, + 1519 + ], + "score": 0.92, + "latex": "\\{ \\boldsymbol { S } _ { e } , \\boldsymbol { \\mathcal { Q } } _ { e } \\}" + }, + { + "category_id": 13, + "poly": [ + 371, + 1133, + 437, + 1133, + 437, + 1167, + 371, + 1167 + ], + "score": 0.92, + "latex": "d ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 527, + 1840, + 608, + 1840, + 608, + 1871, + 527, + 1871 + ], + "score": 0.92, + "latex": "( x _ { i } , r _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 753, + 1807, + 808, + 1807, + 808, + 1840, + 753, + 1840 + ], + "score": 0.92, + "latex": "f _ { \\theta _ { r o t } }" + }, + { + "category_id": 13, + "poly": [ + 831, + 516, + 1315, + 516, + 1315, + 549, + 831, + 549 + ], + "score": 0.92, + "latex": "\\mathcal { Q } _ { e } = \\{ ( x _ { i } , y _ { i } ) | y _ { i } \\in \\mathcal { C } _ { e } , i = 1 , . . . , n \\times q \\}" + }, + { + "category_id": 13, + "poly": [ + 868, + 1391, + 1351, + 1391, + 1351, + 1425, + 868, + 1425 + ], + "score": 0.92, + "latex": "S _ { e } ^ { r } \\ = \\ \\{ ( x _ { i } , y _ { i } , \\bar { r } ) | y _ { i } \\in C _ { e } , i \\ = \\ \\bar { 1 } , . . . , l _ { k } \\}" + }, + { + "category_id": 13, + "poly": [ + 766, + 549, + 900, + 549, + 900, + 578, + 766, + 578 + ], + "score": 0.91, + "latex": "l _ { k } = n \\times k" + }, + { + "category_id": 13, + "poly": [ + 406, + 1485, + 515, + 1485, + 515, + 1519, + 406, + 1519 + ], + "score": 0.91, + "latex": "\\{ \\cal { S } _ { e } ^ { 0 } , \\cal { Q } _ { e } ^ { 0 } \\}" + }, + { + "category_id": 13, + "poly": [ + 296, + 1517, + 412, + 1517, + 412, + 1550, + 296, + 1550 + ], + "score": 0.91, + "latex": "( x _ { i } , y _ { i } , r _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 296, + 515, + 776, + 515, + 776, + 548, + 296, + 548 + ], + "score": 0.91, + "latex": "\\mathcal S _ { e } \\stackrel { - } { = } \\{ ( x _ { i } , y _ { i } ) \\bar { | } y _ { i } \\in \\mathcal C _ { e } , i \\stackrel { - } { = } 1 , . . . , n \\stackrel { - } { \\times } k \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 548, + 451, + 548, + 451, + 579, + 297, + 579 + ], + "score": 0.91, + "latex": "S _ { e } \\cap \\mathcal { Q } _ { e } = \\emptyset" + }, + { + "category_id": 13, + "poly": [ + 954, + 549, + 1085, + 549, + 1085, + 580, + 954, + 580 + ], + "score": 0.91, + "latex": "l _ { q } = n \\times q" + }, + { + "category_id": 13, + "poly": [ + 297, + 1269, + 610, + 1269, + 610, + 1302, + 297, + 1302 + ], + "score": 0.91, + "latex": "\\mathcal { G } = \\{ g _ { r } | r = 0 , . . . , R - 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 494, + 1807, + 667, + 1807, + 667, + 1839, + 494, + 1839 + ], + "score": 0.91, + "latex": "( x _ { i } , y _ { i } , r _ { i } ) \\in E" + }, + { + "category_id": 13, + "poly": [ + 359, + 1391, + 769, + 1391, + 769, + 1424, + 359, + 1424 + ], + "score": 0.9, + "latex": "E = \\{ \\{ \\mathcal { S } _ { e } ^ { r } , \\bar { \\mathcal { Q } } _ { e } ^ { r } \\} | r = 0 , . . . , R - 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 987, + 2002, + 1023, + 2002, + 1023, + 2036, + 987, + 2036 + ], + "score": 0.9, + "latex": "[ \\cdot ] _ { r }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 395, + 1168, + 395, + 1168, + 425, + 1132, + 425 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1778, + 1400, + 1778, + 1400, + 1810, + 1224, + 1810 + ], + "score": 0.89, + "latex": "f _ { \\theta _ { r o t } } : x _ { i } \\mapsto r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1063, + 608, + 1100, + 608, + 1100, + 639, + 1063, + 639 + ], + "score": 0.89, + "latex": "\\mathcal { Q } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1119, + 1454, + 1152, + 1454, + 1152, + 1486, + 1119, + 1486 + ], + "score": 0.89, + "latex": "S _ { e } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 1224, + 455, + 1261, + 455, + 1261, + 486, + 1224, + 486 + ], + "score": 0.89, + "latex": "\\mathcal { Q } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1142, + 457, + 1174, + 457, + 1174, + 486, + 1142, + 486 + ], + "score": 0.89, + "latex": "S _ { e }" + }, + { + "category_id": 13, + "poly": [ + 735, + 395, + 770, + 395, + 770, + 425, + 735, + 425 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { s }" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1808, + 1397, + 1808, + 1397, + 1838, + 1351, + 1838 + ], + "score": 0.89, + "latex": "\\theta _ { r o t }" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1423, + 1404, + 1423, + 1404, + 1456, + 1144, + 1456 + ], + "score": 0.89, + "latex": "E _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in" + }, + { + "category_id": 13, + "poly": [ + 1361, + 1455, + 1397, + 1455, + 1397, + 1486, + 1361, + 1486 + ], + "score": 0.88, + "latex": "\\mathcal { Q } _ { e } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 594, + 425, + 629, + 425, + 629, + 455, + 594, + 455 + ], + "score": 0.88, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 575, + 1578, + 606, + 1578, + 606, + 1611, + 575, + 1611 + ], + "score": 0.88, + "latex": "f _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 1118, + 426, + 1153, + 426, + 1153, + 456, + 1118, + 456 + ], + "score": 0.88, + "latex": "\\mathcal { Q } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 459, + 608, + 491, + 608, + 491, + 639, + 459, + 639 + ], + "score": 0.88, + "latex": "S _ { e }" + }, + { + "category_id": 13, + "poly": [ + 679, + 456, + 709, + 456, + 709, + 486, + 679, + 486 + ], + "score": 0.88, + "latex": "\\mathcal { C } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 894, + 425, + 926, + 425, + 926, + 455, + 894, + 455 + ], + "score": 0.88, + "latex": "S _ { e }" + }, + { + "category_id": 13, + "poly": [ + 381, + 426, + 415, + 426, + 415, + 455, + 381, + 455 + ], + "score": 0.88, + "latex": "\\mathcal { D } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 743, + 1302, + 822, + 1302, + 822, + 1329, + 743, + 1329 + ], + "score": 0.88, + "latex": "R = 4" + }, + { + "category_id": 13, + "poly": [ + 1156, + 335, + 1189, + 335, + 1189, + 364, + 1156, + 364 + ], + "score": 0.87, + "latex": "\\mathcal { D } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 583, + 1362, + 618, + 1362, + 618, + 1392, + 583, + 1392 + ], + "score": 0.87, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 852, + 1135, + 876, + 1135, + 876, + 1164, + 852, + 1164 + ], + "score": 0.87, + "latex": "l _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 884, + 1454, + 909, + 1454, + 909, + 1483, + 884, + 1483 + ], + "score": 0.87, + "latex": "l _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1270, + 1307, + 1270, + 1307, + 1299, + 1247, + 1299 + ], + "score": 0.87, + "latex": "r ^ { * 9 0 }" + }, + { + "category_id": 13, + "poly": [ + 1002, + 984, + 1023, + 984, + 1023, + 1014, + 1002, + 1014 + ], + "score": 0.86, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 298, + 1425, + 757, + 1425, + 757, + 1455, + 298, + 1455 + ], + "score": 0.86, + "latex": "\\mathcal { Q } _ { e } ^ { r } = \\{ ( x _ { i } , y _ { i } , r ) | y _ { i } \\in C _ { e } , i = 1 , . . . , l _ { q } \\} ," + }, + { + "category_id": 13, + "poly": [ + 984, + 486, + 1013, + 486, + 1013, + 515, + 984, + 515 + ], + "score": 0.86, + "latex": "\\mathcal { C } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 899, + 1582, + 928, + 1582, + 928, + 1608, + 899, + 1608 + ], + "score": 0.85, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1123, + 923, + 1143, + 923, + 1143, + 953, + 1123, + 953 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 690, + 1520, + 717, + 1520, + 717, + 1548, + 690, + 1548 + ], + "score": 0.85, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 701, + 1274, + 729, + 1274, + 729, + 1302, + 701, + 1302 + ], + "score": 0.85, + "latex": "g _ { r }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1303, + 370, + 1303, + 370, + 1328, + 346, + 1328 + ], + "score": 0.84, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1686, + 1090, + 1686, + 1090, + 1719, + 1051, + 1719 + ], + "score": 0.84, + "latex": "f _ { \\phi } )" + }, + { + "category_id": 13, + "poly": [ + 881, + 1780, + 907, + 1780, + 907, + 1807, + 881, + 1807 + ], + "score": 0.83, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1375, + 1521, + 1400, + 1521, + 1400, + 1547, + 1375, + 1547 + ], + "score": 0.82, + "latex": "r _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 366, + 1197, + 366, + 1197, + 391, + 1179, + 391 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1262, + 1332, + 1284, + 1332, + 1284, + 1360, + 1262, + 1360 + ], + "score": 0.81, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 699, + 335, + 718, + 335, + 718, + 362, + 699, + 362 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 445, + 1519, + 470, + 1519, + 470, + 1545, + 445, + 1545 + ], + "score": 0.81, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 299, + 1456, + 698, + 1456, + 698, + 1486, + 299, + 1486 + ], + "score": 0.81, + "latex": "C _ { e } , i = 1 , . . . , l _ { k } , l _ { k } + 1 , . . . , l _ { k } + l _ { q } \\}" + }, + { + "category_id": 13, + "poly": [ + 965, + 1579, + 990, + 1579, + 990, + 1605, + 965, + 1605 + ], + "score": 0.8, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 298, + 1579, + 323, + 1579, + 323, + 1605, + 298, + 1605 + ], + "score": 0.79, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 297, + 486, + 316, + 486, + 316, + 513, + 297, + 513 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 566, + 491, + 584, + 491, + 584, + 517, + 566, + 517 + ], + "score": 0.78, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1098, + 370, + 1117, + 370, + 1117, + 391, + 1098, + 391 + ], + "score": 0.78, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 898, + 1362, + 923, + 1362, + 923, + 1389, + 898, + 1389 + ], + "score": 0.78, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1223, + 2007, + 1239, + 2007, + 1239, + 2030, + 1223, + 2030 + ], + "score": 0.77, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 575, + 460, + 596, + 460, + 596, + 482, + 575, + 482 + ], + "score": 0.77, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 536, + 922, + 566, + 922, + 566, + 957, + 536, + 957 + ], + "score": 0.72, + "latex": "f _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 619, + 340, + 638, + 340, + 638, + 362, + 619, + 362 + ], + "score": 0.69, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 438, + 922, + 521, + 922, + 521, + 954, + 438, + 954 + ], + "score": 0.69, + "latex": "c \\in \\mathcal { C } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1205, + 923, + 1222, + 923, + 1222, + 949, + 1205, + 949 + ], + "score": 0.57, + "latex": "\\mathbb { I }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 669.0, + 1178.0, + 669.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 588.0, + 221.0, + 588.0, + 270.0, + 291.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 274.0, + 537.0, + 274.0, + 537.0, + 310.0, + 294.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1235.0, + 1405.0, + 1235.0, + 1405.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1267.0, + 700.0, + 1267.0, + 700.0, + 1306.0, + 611.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1267.0, + 1246.0, + 1267.0, + 1246.0, + 1306.0, + 730.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1267.0, + 1405.0, + 1267.0, + 1405.0, + 1306.0, + 1308.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 345.0, + 1298.0, + 345.0, + 1336.0, + 294.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1298.0, + 742.0, + 1298.0, + 742.0, + 1336.0, + 371.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1298.0, + 1403.0, + 1298.0, + 1403.0, + 1336.0, + 823.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1329.0, + 1261.0, + 1329.0, + 1261.0, + 1366.0, + 482.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1366.0, + 1285.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 582.0, + 1362.0, + 582.0, + 1395.0, + 295.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1362.0, + 897.0, + 1362.0, + 897.0, + 1395.0, + 619.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1395.0, + 924.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1388.0, + 358.0, + 1388.0, + 358.0, + 1429.0, + 292.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1388.0, + 867.0, + 1388.0, + 867.0, + 1429.0, + 770.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1388.0, + 1407.0, + 1388.0, + 1407.0, + 1429.0, + 1352.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1420.0, + 1143.0, + 1420.0, + 1143.0, + 1460.0, + 758.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1446.0, + 298.0, + 1446.0, + 298.0, + 1494.0, + 292.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1446.0, + 883.0, + 1446.0, + 883.0, + 1494.0, + 699.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1446.0, + 1118.0, + 1446.0, + 1118.0, + 1494.0, + 910.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1446.0, + 1360.0, + 1446.0, + 1360.0, + 1494.0, + 1153.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1446.0, + 1411.0, + 1446.0, + 1411.0, + 1494.0, + 1398.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1480.0, + 405.0, + 1480.0, + 405.0, + 1524.0, + 290.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1480.0, + 766.0, + 1480.0, + 766.0, + 1524.0, + 516.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1480.0, + 1407.0, + 1480.0, + 1407.0, + 1524.0, + 873.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1514.0, + 444.0, + 1514.0, + 444.0, + 1554.0, + 413.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1514.0, + 689.0, + 1514.0, + 689.0, + 1554.0, + 471.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1514.0, + 1374.0, + 1514.0, + 1374.0, + 1554.0, + 718.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1514.0, + 1406.0, + 1514.0, + 1406.0, + 1554.0, + 1401.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1544.0, + 1405.0, + 1544.0, + 1405.0, + 1582.0, + 293.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1575.0, + 574.0, + 1575.0, + 574.0, + 1612.0, + 324.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1575.0, + 898.0, + 1575.0, + 898.0, + 1612.0, + 607.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1575.0, + 964.0, + 1575.0, + 964.0, + 1612.0, + 929.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1575.0, + 1403.0, + 1575.0, + 1403.0, + 1612.0, + 991.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1607.0, + 1210.0, + 1607.0, + 1210.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 332.0, + 618.0, + 332.0, + 618.0, + 367.0, + 294.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 332.0, + 698.0, + 332.0, + 698.0, + 367.0, + 639.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 332.0, + 1155.0, + 332.0, + 1155.0, + 367.0, + 719.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 332.0, + 1406.0, + 332.0, + 1406.0, + 367.0, + 1190.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 364.0, + 1097.0, + 364.0, + 1097.0, + 398.0, + 294.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 364.0, + 1178.0, + 364.0, + 1178.0, + 398.0, + 1118.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 398.0, + 1198.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 392.0, + 734.0, + 392.0, + 734.0, + 431.0, + 293.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 392.0, + 1131.0, + 392.0, + 1131.0, + 431.0, + 771.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 392.0, + 1406.0, + 392.0, + 1406.0, + 431.0, + 1169.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 421.0, + 380.0, + 421.0, + 380.0, + 461.0, + 293.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 421.0, + 593.0, + 421.0, + 593.0, + 461.0, + 416.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 421.0, + 893.0, + 421.0, + 893.0, + 461.0, + 630.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 421.0, + 1117.0, + 421.0, + 1117.0, + 461.0, + 927.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 421.0, + 1407.0, + 421.0, + 1407.0, + 461.0, + 1154.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 451.0, + 574.0, + 451.0, + 574.0, + 492.0, + 291.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 451.0, + 678.0, + 451.0, + 678.0, + 492.0, + 597.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 451.0, + 1141.0, + 451.0, + 1141.0, + 492.0, + 710.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 451.0, + 1223.0, + 451.0, + 1223.0, + 492.0, + 1175.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 451.0, + 1406.0, + 451.0, + 1406.0, + 492.0, + 1262.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 486.0, + 565.0, + 486.0, + 565.0, + 521.0, + 317.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 486.0, + 983.0, + 486.0, + 983.0, + 521.0, + 585.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 486.0, + 1406.0, + 486.0, + 1406.0, + 521.0, + 1014.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 509.0, + 295.0, + 509.0, + 295.0, + 554.0, + 291.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 509.0, + 830.0, + 509.0, + 830.0, + 554.0, + 777.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 509.0, + 1406.0, + 509.0, + 1406.0, + 554.0, + 1316.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 543.0, + 296.0, + 543.0, + 296.0, + 584.0, + 293.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 543.0, + 765.0, + 543.0, + 765.0, + 584.0, + 452.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 543.0, + 953.0, + 543.0, + 953.0, + 584.0, + 901.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 543.0, + 1410.0, + 543.0, + 1410.0, + 584.0, + 1086.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 573.0, + 1407.0, + 573.0, + 1407.0, + 613.0, + 291.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 607.0, + 458.0, + 607.0, + 458.0, + 642.0, + 295.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 607.0, + 1062.0, + 607.0, + 1062.0, + 642.0, + 492.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 607.0, + 1403.0, + 607.0, + 1403.0, + 642.0, + 1101.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 639.0, + 1014.0, + 639.0, + 1014.0, + 670.0, + 296.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1408.0, + 1651.0, + 1408.0, + 1689.0, + 294.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1685.0, + 1050.0, + 1685.0, + 1050.0, + 1720.0, + 294.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1685.0, + 1406.0, + 1685.0, + 1406.0, + 1720.0, + 1091.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1714.0, + 1404.0, + 1714.0, + 1404.0, + 1751.0, + 294.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1747.0, + 1403.0, + 1747.0, + 1403.0, + 1778.0, + 296.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 880.0, + 1771.0, + 880.0, + 1815.0, + 292.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1771.0, + 1223.0, + 1771.0, + 1223.0, + 1815.0, + 908.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1815.0, + 1401.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 493.0, + 1802.0, + 493.0, + 1845.0, + 293.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1802.0, + 752.0, + 1802.0, + 752.0, + 1845.0, + 668.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1802.0, + 1350.0, + 1802.0, + 1350.0, + 1845.0, + 809.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1802.0, + 1411.0, + 1802.0, + 1411.0, + 1845.0, + 1398.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1835.0, + 526.0, + 1835.0, + 526.0, + 1873.0, + 295.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1835.0, + 1278.0, + 1835.0, + 1278.0, + 1873.0, + 609.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 684.0, + 1405.0, + 684.0, + 1405.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 716.0, + 1405.0, + 716.0, + 1405.0, + 748.0, + 294.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 749.0, + 1404.0, + 749.0, + 1404.0, + 778.0, + 296.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 775.0, + 1168.0, + 775.0, + 1168.0, + 810.0, + 292.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 920.0, + 437.0, + 920.0, + 437.0, + 957.0, + 294.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 920.0, + 535.0, + 920.0, + 535.0, + 957.0, + 522.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 920.0, + 1122.0, + 920.0, + 1122.0, + 957.0, + 567.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 920.0, + 1204.0, + 920.0, + 1204.0, + 957.0, + 1144.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 920.0, + 1406.0, + 920.0, + 1406.0, + 957.0, + 1223.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 1406.0, + 952.0, + 1406.0, + 986.0, + 294.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 980.0, + 1001.0, + 980.0, + 1001.0, + 1019.0, + 293.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 980.0, + 1368.0, + 980.0, + 1368.0, + 1019.0, + 1024.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1127.0, + 370.0, + 1127.0, + 370.0, + 1172.0, + 293.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1127.0, + 851.0, + 1127.0, + 851.0, + 1172.0, + 438.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1127.0, + 993.0, + 1127.0, + 993.0, + 1172.0, + 877.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 371.0, + 1999.0, + 371.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1999.0, + 986.0, + 1999.0, + 986.0, + 2039.0, + 612.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1999.0, + 1222.0, + 1999.0, + 1222.0, + 2039.0, + 1024.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1999.0, + 1373.0, + 1999.0, + 1373.0, + 2039.0, + 1240.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 227, + 1406, + 227, + 1406, + 484, + 296, + 484 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 296, + 1052, + 1406, + 1052, + 1406, + 1394, + 296, + 1394 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 1539, + 1405, + 1539, + 1405, + 1730, + 297, + 1730 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 976, + 1404, + 976, + 1404, + 1040, + 296, + 1040 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 300, + 1765, + 1405, + 1765, + 1405, + 1857, + 300, + 1857 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 565, + 1868, + 1134, + 1868, + 1134, + 1961, + 565, + 1961 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 748, + 639, + 946, + 639, + 946, + 726, + 748, + 726 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 646, + 489, + 1051, + 489, + 1051, + 564, + 646, + 564 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 1968, + 1402, + 1968, + 1402, + 2036, + 297, + 2036 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 568, + 1400, + 568, + 1400, + 629, + 297, + 629 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 562, + 774, + 1133, + 774, + 1133, + 866, + 562, + 866 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 558, + 1453, + 1052, + 1453, + 1052, + 1535, + 558, + 1535 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 714, + 1728, + 979, + 1728, + 979, + 1763, + 714, + 1763 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 557, + 1400, + 1143, + 1400, + 1143, + 1446, + 557, + 1446 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 299, + 872, + 903, + 872, + 903, + 905, + 299, + 905 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 295, + 732, + 1366, + 732, + 1366, + 765, + 295, + 765 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 299, + 919, + 652, + 919, + 652, + 952, + 299, + 952 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1897, + 1400, + 1897, + 1400, + 1928, + 1352, + 1928 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 805, + 1400, + 805, + 1400, + 835, + 1366, + 835 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 668, + 1400, + 668, + 1400, + 697, + 1366, + 697 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1731, + 1399, + 1731, + 1399, + 1760, + 1366, + 1760 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 510, + 1400, + 510, + 1400, + 539, + 1366, + 539 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1485, + 1400, + 1485, + 1400, + 1514, + 1366, + 1514 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1410, + 1400, + 1410, + 1400, + 1440, + 1365, + 1440 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.762 + }, + { + "category_id": 14, + "poly": [ + 646, + 487, + 1053, + 487, + 1053, + 565, + 646, + 565 + ], + "score": 0.95, + "latex": "[ p _ { i } ^ { r } ] _ { c } = \\frac { \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c } ^ { r } ) ) } { \\sum _ { c ^ { \\prime } } \\exp ( - d ( f _ { \\phi } ( x _ { i } ) , h _ { c ^ { \\prime } } ^ { r } ) ) } ." + }, + { + "category_id": 14, + "poly": [ + 562, + 769, + 1136, + 769, + 1136, + 868, + 562, + 868 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { e p i s } = \\frac { 1 } { R l _ { q } } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\sum _ { i = 1 } ^ { l _ { q } } \\mathrm { m e a n } ( p _ { i } ^ { r } ( \\log p _ { i } ^ { r } - \\log \\hat { p } _ { i } ) ) ." + }, + { + "category_id": 14, + "poly": [ + 750, + 634, + 950, + 634, + 950, + 729, + 750, + 729 + ], + "score": 0.94, + "latex": "\\hat { p } _ { i } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } p _ { i } ^ { r } ." + }, + { + "category_id": 14, + "poly": [ + 564, + 1862, + 1136, + 1862, + 1136, + 1961, + 564, + 1961 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { i n t e g } = \\frac { 1 } { l _ { q } } \\cdot \\sum _ { i = 1 } ^ { l _ { q } } - \\log \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y _ { i } } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) }" + }, + { + "category_id": 13, + "poly": [ + 582, + 1332, + 687, + 1332, + 687, + 1365, + 582, + 1365 + ], + "score": 0.93, + "latex": "( F , F , F )" + }, + { + "category_id": 13, + "poly": [ + 999, + 1632, + 1238, + 1632, + 1238, + 1666, + 999, + 1666 + ], + "score": 0.93, + "latex": "F _ { a t t } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }" + }, + { + "category_id": 13, + "poly": [ + 734, + 414, + 832, + 414, + 832, + 447, + 734, + 447 + ], + "score": 0.93, + "latex": "p _ { i } ^ { r } \\in \\mathbb { R } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 298, + 382, + 406, + 382, + 406, + 416, + 298, + 416 + ], + "score": 0.92, + "latex": "\\{ S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } \\}" + }, + { + "category_id": 14, + "poly": [ + 713, + 1727, + 981, + 1727, + 981, + 1764, + 713, + 1764 + ], + "score": 0.92, + "latex": "F _ { i n t e g } = \\mathrm { f l a t t e n } ( F _ { a t t } ) ," + }, + { + "category_id": 13, + "poly": [ + 725, + 1332, + 840, + 1332, + 840, + 1365, + 725, + 1365 + ], + "score": 0.92, + "latex": "( Q , K , V )" + }, + { + "category_id": 13, + "poly": [ + 816, + 1237, + 1029, + 1237, + 1029, + 1269, + 816, + 1269 + ], + "score": 0.92, + "latex": "\\bar { F } \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1572, + 1083, + 1572, + 1083, + 1602, + 1035, + 1602 + ], + "score": 0.91, + "latex": "\\theta _ { i n t }" + }, + { + "category_id": 13, + "poly": [ + 1145, + 1055, + 1209, + 1055, + 1209, + 1086, + 1145, + 1086 + ], + "score": 0.91, + "latex": "E _ { e m b }" + }, + { + "category_id": 13, + "poly": [ + 600, + 1968, + 940, + 1968, + 940, + 2008, + 600, + 2008 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { h _ { c } ^ { f } = \\frac { 1 } { k } \\cdot \\sum _ { i = 1 } ^ { l _ { k } } F _ { i } ^ { S } \\cdot \\mathbb { I } ( y _ { i } = c ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 372, + 1541, + 463, + 1541, + 463, + 1572, + 372, + 1572 + ], + "score": 0.91, + "latex": "d _ { K } = d" + }, + { + "category_id": 13, + "poly": [ + 465, + 1665, + 854, + 1665, + 854, + 1702, + 465, + 1702 + ], + "score": 0.91, + "latex": "F _ { i n t e g } = [ F ^ { S } ; F ^ { Q } ] \\in \\mathbb { R } ^ { ( l _ { k } + l _ { q } ) \\times R d }" + }, + { + "category_id": 13, + "poly": [ + 400, + 1241, + 464, + 1241, + 464, + 1271, + 400, + 1271 + ], + "score": 0.9, + "latex": "E _ { e m b }" + }, + { + "category_id": 13, + "poly": [ + 961, + 1667, + 1003, + 1667, + 1003, + 1696, + 961, + 1696 + ], + "score": 0.9, + "latex": "F ^ { Q }" + }, + { + "category_id": 13, + "poly": [ + 679, + 1767, + 729, + 1767, + 729, + 1796, + 679, + 1796 + ], + "score": 0.9, + "latex": "F _ { a t t }" + }, + { + "category_id": 13, + "poly": [ + 1273, + 415, + 1309, + 415, + 1309, + 447, + 1273, + 447 + ], + "score": 0.9, + "latex": "\\mathcal { Q } _ { e } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 1352, + 1330, + 1402, + 1330, + 1402, + 1359, + 1352, + 1359 + ], + "score": 0.9, + "latex": "\\dot { F } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 521, + 1541, + 685, + 1541, + 685, + 1576, + 521, + 1576 + ], + "score": 0.9, + "latex": "W _ { Q } , W _ { K } , W _ { V }" + }, + { + "category_id": 13, + "poly": [ + 1302, + 1055, + 1404, + 1055, + 1404, + 1086, + 1302, + 1086 + ], + "score": 0.89, + "latex": "E _ { e m b } =" + }, + { + "category_id": 13, + "poly": [ + 1136, + 569, + 1168, + 569, + 1168, + 601, + 1136, + 601 + ], + "score": 0.89, + "latex": "h _ { c } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 869, + 1666, + 909, + 1666, + 909, + 1697, + 869, + 1697 + ], + "score": 0.89, + "latex": "F ^ { S }" + }, + { + "category_id": 13, + "poly": [ + 666, + 1208, + 701, + 1208, + 701, + 1239, + 666, + 1239 + ], + "score": 0.88, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 825, + 1365, + 860, + 1365, + 860, + 1394, + 825, + 1394 + ], + "score": 0.88, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 328, + 412, + 647, + 412, + 647, + 452, + 328, + 452 + ], + "score": 0.88, + "latex": "\\mathcal { P } _ { e } ^ { r } = [ p _ { 1 } ^ { r } ; \\cdot \\cdot \\cdot ; p _ { l _ { q } } ^ { r } ] \\in \\mathbb { R } ^ { l _ { q } \\times n }" + }, + { + "category_id": 13, + "poly": [ + 1364, + 383, + 1400, + 383, + 1400, + 416, + 1364, + 416 + ], + "score": 0.87, + "latex": "\\mathcal { Q } _ { e } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1147, + 332, + 1147, + 332, + 1177, + 298, + 1177 + ], + "score": 0.86, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1365, + 573, + 1397, + 573, + 1397, + 601, + 1365, + 601 + ], + "score": 0.86, + "latex": "S _ { e } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 747, + 599, + 772, + 599, + 772, + 625, + 747, + 625 + ], + "score": 0.85, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1211, + 417, + 1239, + 417, + 1239, + 444, + 1211, + 444 + ], + "score": 0.84, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 534, + 1365, + 558, + 1365, + 558, + 1390, + 534, + 1390 + ], + "score": 0.83, + "latex": "F" + }, + { + "category_id": 13, + "poly": [ + 1081, + 1056, + 1106, + 1056, + 1106, + 1082, + 1081, + 1082 + ], + "score": 0.82, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1241, + 1133, + 1241, + 1133, + 1268, + 1114, + 1268 + ], + "score": 0.82, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1376, + 1088, + 1401, + 1088, + 1401, + 1113, + 1376, + 1113 + ], + "score": 0.81, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 710, + 1056, + 734, + 1056, + 734, + 1082, + 710, + 1082 + ], + "score": 0.81, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 442, + 451, + 497, + 451, + 497, + 484, + 442, + 484 + ], + "score": 0.8, + "latex": "[ p _ { i } ^ { r } ] _ { c }" + }, + { + "category_id": 13, + "poly": [ + 438, + 384, + 463, + 384, + 463, + 410, + 438, + 410 + ], + "score": 0.8, + "latex": "E" + }, + { + "category_id": 14, + "poly": [ + 555, + 1398, + 1141, + 1398, + 1141, + 1538, + 555, + 1538 + ], + "score": 0.79, + "latex": "\\begin{array} { r l } & { ( \\boldsymbol { F } _ { Q } ^ { ( i ) } , \\boldsymbol { F } _ { K } ^ { ( i ) } , \\boldsymbol { F } _ { V } ^ { ( i ) } ) = ( \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { Q } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { K } , \\boldsymbol { F } ^ { ( i ) } \\boldsymbol { W } _ { V } ) , } \\\\ & { \\boldsymbol { F } _ { a t t } ^ { ( i ) } = \\boldsymbol { F } ^ { ( i ) } + \\mathrm { s o f t m a x } ( \\frac { \\boldsymbol { F } _ { Q } ^ { ( i ) } ( \\boldsymbol { F } _ { K } ^ { ( i ) } ) ^ { T } } { \\sqrt { d _ { K } } } ) \\boldsymbol { F } _ { V } ^ { ( i ) } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1085, + 983, + 1085, + 983, + 1119, + 298, + 1119 + ], + "score": 0.78, + "latex": "\\{ f _ { \\phi } ( x _ { i } ) | ( x _ { i } , y _ { i } , r ) \\in E _ { e } ^ { r } , r = 0 , \\cdots , R - 1 , i = 1 , . . . , l _ { k } + l _ { q } \\}" + }, + { + "category_id": 13, + "poly": [ + 674, + 1367, + 686, + 1367, + 686, + 1390, + 674, + 1390 + ], + "score": 0.77, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 298, + 456, + 313, + 456, + 313, + 478, + 298, + 478 + ], + "score": 0.77, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 407, + 1365, + 419, + 1365, + 419, + 1390, + 407, + 1390 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 509, + 453, + 648, + 453, + 648, + 483, + 509, + 483 + ], + "score": 0.72, + "latex": "\\mathbf { \\bar { \\psi } } c = 1 , . . . , n )" + }, + { + "category_id": 13, + "poly": [ + 371, + 873, + 467, + 873, + 467, + 906, + 371, + 906 + ], + "score": 0.62, + "latex": "\\mathrm { m e a n } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 449, + 1765, + 479, + 1765, + 479, + 1798, + 449, + 1798 + ], + "score": 0.48, + "latex": "( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 438, + 873, + 467, + 873, + 467, + 905, + 438, + 905 + ], + "score": 0.29, + "latex": "( \\cdot )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 920.0, + 652.0, + 920.0, + 652.0, + 953.0, + 296.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1408.0, + 261.0, + 1408.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 359.0, + 293.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 379.0, + 437.0, + 379.0, + 437.0, + 421.0, + 407.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 379.0, + 1363.0, + 379.0, + 1363.0, + 421.0, + 464.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 379.0, + 1406.0, + 379.0, + 1406.0, + 421.0, + 1401.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 409.0, + 327.0, + 409.0, + 327.0, + 452.0, + 289.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 409.0, + 733.0, + 409.0, + 733.0, + 452.0, + 648.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 409.0, + 1210.0, + 409.0, + 1210.0, + 452.0, + 833.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 409.0, + 1272.0, + 409.0, + 1272.0, + 452.0, + 1240.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 409.0, + 1408.0, + 409.0, + 1408.0, + 452.0, + 1310.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 446.0, + 297.0, + 446.0, + 297.0, + 488.0, + 293.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 446.0, + 441.0, + 446.0, + 441.0, + 488.0, + 314.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 446.0, + 508.0, + 446.0, + 508.0, + 488.0, + 498.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 446.0, + 730.0, + 446.0, + 730.0, + 488.0, + 649.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1047.0, + 709.0, + 1047.0, + 709.0, + 1095.0, + 291.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1047.0, + 1080.0, + 1047.0, + 1080.0, + 1095.0, + 735.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1047.0, + 1144.0, + 1047.0, + 1144.0, + 1095.0, + 1107.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1047.0, + 1301.0, + 1047.0, + 1301.0, + 1095.0, + 1210.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1084.0, + 1375.0, + 1084.0, + 1375.0, + 1123.0, + 984.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1084.0, + 1407.0, + 1084.0, + 1407.0, + 1123.0, + 1402.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1115.0, + 1404.0, + 1115.0, + 1404.0, + 1150.0, + 294.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1145.0, + 297.0, + 1145.0, + 297.0, + 1180.0, + 294.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1145.0, + 1406.0, + 1145.0, + 1406.0, + 1180.0, + 333.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1406.0, + 1177.0, + 1406.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1204.0, + 665.0, + 1204.0, + 665.0, + 1244.0, + 291.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1204.0, + 1407.0, + 1204.0, + 1407.0, + 1244.0, + 702.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1232.0, + 399.0, + 1232.0, + 399.0, + 1278.0, + 289.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1232.0, + 815.0, + 1232.0, + 815.0, + 1278.0, + 465.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1232.0, + 1113.0, + 1232.0, + 1113.0, + 1278.0, + 1030.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1232.0, + 1413.0, + 1232.0, + 1413.0, + 1278.0, + 1134.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1269.0, + 1406.0, + 1269.0, + 1406.0, + 1305.0, + 295.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1301.0, + 1404.0, + 1301.0, + 1404.0, + 1336.0, + 294.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1327.0, + 581.0, + 1327.0, + 581.0, + 1369.0, + 289.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1327.0, + 724.0, + 1327.0, + 724.0, + 1369.0, + 688.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1327.0, + 1351.0, + 1327.0, + 1351.0, + 1369.0, + 841.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1327.0, + 1410.0, + 1327.0, + 1410.0, + 1369.0, + 1403.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 406.0, + 1362.0, + 406.0, + 1398.0, + 295.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1362.0, + 533.0, + 1362.0, + 533.0, + 1398.0, + 420.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1362.0, + 673.0, + 1362.0, + 673.0, + 1398.0, + 559.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1362.0, + 824.0, + 1362.0, + 824.0, + 1398.0, + 687.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1362.0, + 1263.0, + 1362.0, + 1263.0, + 1398.0, + 861.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1540.0, + 371.0, + 1540.0, + 371.0, + 1576.0, + 295.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1540.0, + 520.0, + 1540.0, + 520.0, + 1576.0, + 464.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1540.0, + 1408.0, + 1540.0, + 1408.0, + 1576.0, + 686.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1570.0, + 1034.0, + 1570.0, + 1034.0, + 1606.0, + 295.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1570.0, + 1403.0, + 1570.0, + 1403.0, + 1606.0, + 1084.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1637.0, + 294.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1627.0, + 998.0, + 1627.0, + 998.0, + 1675.0, + 291.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1627.0, + 1408.0, + 1627.0, + 1408.0, + 1675.0, + 1239.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1658.0, + 464.0, + 1658.0, + 464.0, + 1710.0, + 289.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1658.0, + 868.0, + 1658.0, + 868.0, + 1710.0, + 855.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1658.0, + 960.0, + 1658.0, + 960.0, + 1710.0, + 910.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1658.0, + 1411.0, + 1658.0, + 1411.0, + 1710.0, + 1004.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1695.0, + 596.0, + 1695.0, + 596.0, + 1736.0, + 295.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 972.0, + 1404.0, + 972.0, + 1404.0, + 1016.0, + 294.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 1391.0, + 1007.0, + 1391.0, + 1041.0, + 294.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 448.0, + 1763.0, + 448.0, + 1800.0, + 294.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1763.0, + 678.0, + 1763.0, + 678.0, + 1800.0, + 480.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1763.0, + 1405.0, + 1763.0, + 1405.0, + 1800.0, + 730.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1794.0, + 1404.0, + 1794.0, + 1404.0, + 1830.0, + 294.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1825.0, + 1112.0, + 1825.0, + 1112.0, + 1859.0, + 294.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1952.0, + 599.0, + 1952.0, + 599.0, + 2040.0, + 285.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1952.0, + 1417.0, + 1952.0, + 1417.0, + 2040.0, + 941.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1984.0, + 738.0, + 1984.0, + 738.0, + 1993.5, + 735.0, + 1993.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1981.0, + 764.0, + 1981.0, + 764.0, + 1995.5, + 738.0, + 1995.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 561.0, + 1135.0, + 561.0, + 1135.0, + 606.0, + 291.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 561.0, + 1364.0, + 561.0, + 1364.0, + 606.0, + 1169.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 561.0, + 1404.0, + 561.0, + 1404.0, + 606.0, + 1398.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 593.0, + 746.0, + 593.0, + 746.0, + 635.0, + 292.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 593.0, + 1172.0, + 593.0, + 1172.0, + 635.0, + 773.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 870.0, + 370.0, + 870.0, + 370.0, + 909.0, + 295.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 870.0, + 904.0, + 870.0, + 904.0, + 909.0, + 468.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 1367.0, + 728.0, + 1367.0, + 771.0, + 294.0, + 771.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1849, + 1406, + 1849, + 1406, + 2034, + 297, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 287, + 1404, + 287, + 1404, + 411, + 298, + 411 + ], + "score": 0.979 + }, + { + "category_id": 8, + "poly": [ + 598, + 929, + 1098, + 929, + 1098, + 1021, + 598, + 1021 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 522, + 588, + 1175, + 588, + 1175, + 703, + 522, + 703 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 674, + 431, + 1026, + 431, + 1026, + 522, + 674, + 522 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 293, + 1112, + 1405, + 1112, + 1405, + 1177, + 293, + 1177 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 847, + 1405, + 847, + 1405, + 912, + 298, + 912 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 295, + 537, + 1155, + 537, + 1155, + 572, + 295, + 572 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 297, + 719, + 906, + 719, + 906, + 753, + 297, + 753 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 299, + 1054, + 660, + 1054, + 660, + 1087, + 299, + 1087 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 517, + 229, + 517, + 261, + 299, + 261 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 298, + 789, + 504, + 789, + 504, + 821, + 298, + 821 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 300, + 74, + 813, + 74, + 813, + 105, + 300, + 105 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1351, + 460, + 1400, + 460, + 1400, + 492, + 1351, + 492 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1352, + 957, + 1400, + 957, + 1400, + 989, + 1352, + 989 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 300, + 1721, + 557, + 1721, + 557, + 1756, + 300, + 1756 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1352, + 625, + 1401, + 625, + 1401, + 657, + 1352, + 657 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 300, + 1791, + 641, + 1791, + 641, + 1823, + 300, + 1823 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 298, + 1212, + 619, + 1212, + 619, + 1244, + 298, + 1244 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 299, + 1248, + 1186, + 1248, + 1186, + 1672, + 299, + 1672 + ], + "score": 0.673 + }, + { + "category_id": 14, + "poly": [ + 674, + 427, + 1026, + 427, + 1026, + 524, + 674, + 524 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { a u x } = \\frac { 1 } { R } \\cdot \\sum _ { r = 0 } ^ { R - 1 } \\mathcal { L } _ { f s l } ( S _ { e } ^ { r } , \\mathcal { Q } _ { e } ^ { r } )" + }, + { + "category_id": 14, + "poly": [ + 600, + 927, + 1101, + 927, + 1101, + 1020, + 600, + 1020 + ], + "score": 0.94, + "latex": "y _ { i } ^ { p r e d } = \\underset { y \\in \\mathcal { C } _ { e } } { \\operatorname { a r g m a x } } \\frac { \\exp ( - d ( F _ { i } ^ { Q } , h _ { y } ^ { f } ) ) } { \\sum _ { c \\in \\mathcal { C } _ { e } } \\exp ( - d ( F _ { i } ^ { Q } , h _ { c } ^ { f } ) ) } ." + }, + { + "category_id": 13, + "poly": [ + 1023, + 850, + 1096, + 850, + 1096, + 882, + 1023, + 882 + ], + "score": 0.92, + "latex": "F _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 617, + 881, + 713, + 881, + 713, + 910, + 617, + 910 + ], + "score": 0.92, + "latex": "x _ { i } \\in \\mathcal { Q } _ { e }" + }, + { + "category_id": 13, + "poly": [ + 566, + 540, + 636, + 540, + 636, + 569, + 566, + 569 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { t o t a l }" + }, + { + "category_id": 13, + "poly": [ + 534, + 350, + 594, + 350, + 594, + 380, + 534, + 380 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 1105, + 319, + 1180, + 319, + 1180, + 352, + 1105, + 352 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 371, + 726, + 495, + 726, + 495, + 751, + 371, + 751 + ], + "score": 0.9, + "latex": "w _ { 1 } , w _ { 2 } , w _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 914, + 350, + 975, + 350, + 975, + 381, + 914, + 381 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 709, + 1499, + 765, + 1499, + 765, + 1525, + 709, + 1525 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 596, + 1444, + 655, + 1444, + 655, + 1470, + 596, + 1470 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 14, + "poly": [ + 520, + 586, + 1178, + 586, + 1178, + 703, + 520, + 703 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\mathcal { L } _ { t o t a l } = \\underbrace { \\overbrace { w _ { 1 } * \\mathcal { L } _ { i n s t } } ^ { i n s t a n c e { - } l e v e l } + \\overbrace { w _ { 2 } * \\mathcal { L } _ { e p i s } } ^ { e p i s o d e { - } l e v e l } } _ { \\mathrm { s e l f - s u p e r v i s e d l o s s } } + \\underbrace { w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t e g } } _ { \\mathrm { s u p e r v i s e d l o s s } } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 596, + 1473, + 655, + 1473, + 655, + 1500, + 596, + 1500 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 709, + 1527, + 779, + 1527, + 779, + 1555, + 709, + 1555 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 787, + 1331, + 989, + 1331, + 989, + 1362, + 787, + 1362 + ], + "score": 0.88, + "latex": "\\psi = \\{ \\phi , \\theta _ { r o t } , \\theta _ { i n t } \\}" + }, + { + "category_id": 13, + "poly": [ + 637, + 1388, + 669, + 1388, + 669, + 1415, + 637, + 1415 + ], + "score": 0.87, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 540, + 1249, + 573, + 1249, + 573, + 1276, + 540, + 1276 + ], + "score": 0.87, + "latex": "\\mathcal { D } _ { s }" + }, + { + "category_id": 13, + "poly": [ + 728, + 1389, + 761, + 1389, + 761, + 1415, + 728, + 1415 + ], + "score": 0.86, + "latex": "\\mathcal { D } _ { s }" + }, + { + "category_id": 13, + "poly": [ + 830, + 1416, + 862, + 1416, + 862, + 1442, + 830, + 1442 + ], + "score": 0.85, + "latex": "E _ { e }" + }, + { + "category_id": 13, + "poly": [ + 481, + 1146, + 503, + 1146, + 503, + 1176, + 481, + 1176 + ], + "score": 0.84, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 448, + 1584, + 468, + 1584, + 468, + 1610, + 448, + 1610 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 749, + 1417, + 773, + 1417, + 773, + 1440, + 749, + 1440 + ], + "score": 0.82, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 464, + 1361, + 502, + 1361, + 502, + 1386, + 464, + 1386 + ], + "score": 0.81, + "latex": "= 1" + }, + { + "category_id": 13, + "poly": [ + 520, + 1306, + 541, + 1306, + 541, + 1334, + 520, + 1334 + ], + "score": 0.8, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 566, + 1583, + 605, + 1583, + 605, + 1611, + 566, + 1611 + ], + "score": 0.79, + "latex": "\\nabla _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 416, + 1638, + 437, + 1638, + 437, + 1666, + 416, + 1666 + ], + "score": 0.78, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 713, + 1282, + 828, + 1282, + 828, + 1305, + 713, + 1305 + ], + "score": 0.77, + "latex": "w _ { 1 } , w _ { 2 } , w _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 926, + 1416, + 946, + 1416, + 946, + 1442, + 926, + 1442 + ], + "score": 0.73, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 827, + 1251, + 847, + 1251, + 847, + 1276, + 827, + 1276 + ], + "score": 0.67, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 375, + 1554, + 932, + 1554, + 932, + 1584, + 375, + 1584 + ], + "score": 0.27, + "latex": "\\mathcal { L } _ { t o t a l } = w _ { 1 } * \\mathcal { L } _ { i n s t } + w _ { 2 } * \\mathcal { L } _ { e p i s } + w _ { 3 } * \\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n t }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1055.0, + 660.0, + 1055.0, + 660.0, + 1088.0, + 295.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 521.0, + 224.0, + 521.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 506.0, + 786.0, + 506.0, + 825.0, + 294.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1718.0, + 561.0, + 1718.0, + 561.0, + 1762.0, + 293.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1789.0, + 642.0, + 1789.0, + 642.0, + 1826.0, + 294.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 623.0, + 1211.0, + 623.0, + 1246.0, + 294.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1913.0, + 295.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1408.0, + 1911.0, + 1408.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1407.0, + 1971.0, + 1407.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 879.0, + 2003.0, + 879.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 288.0, + 1402.0, + 288.0, + 1402.0, + 321.0, + 296.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 315.0, + 1104.0, + 315.0, + 1104.0, + 355.0, + 292.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 315.0, + 1408.0, + 315.0, + 1408.0, + 355.0, + 1181.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 349.0, + 533.0, + 349.0, + 533.0, + 384.0, + 292.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 349.0, + 913.0, + 349.0, + 913.0, + 384.0, + 595.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 349.0, + 1405.0, + 349.0, + 1405.0, + 384.0, + 976.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 381.0, + 948.0, + 381.0, + 948.0, + 413.0, + 294.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1112.0, + 1407.0, + 1112.0, + 1407.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1146.0, + 480.0, + 1146.0, + 480.0, + 1178.0, + 296.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1146.0, + 1248.0, + 1146.0, + 1248.0, + 1178.0, + 504.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 846.0, + 1022.0, + 846.0, + 1022.0, + 886.0, + 295.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 846.0, + 1408.0, + 846.0, + 1408.0, + 886.0, + 1097.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 876.0, + 616.0, + 876.0, + 616.0, + 917.0, + 295.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 876.0, + 1103.0, + 876.0, + 1103.0, + 917.0, + 714.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 536.0, + 565.0, + 536.0, + 565.0, + 576.0, + 295.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 536.0, + 1158.0, + 536.0, + 1158.0, + 576.0, + 637.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 712.0, + 370.0, + 712.0, + 370.0, + 762.0, + 293.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 712.0, + 912.0, + 712.0, + 912.0, + 762.0, + 496.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1247.0, + 539.0, + 1247.0, + 539.0, + 1281.0, + 295.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1247.0, + 826.0, + 1247.0, + 826.0, + 1281.0, + 574.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1274.0, + 712.0, + 1274.0, + 712.0, + 1312.0, + 374.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1274.0, + 832.0, + 1274.0, + 832.0, + 1312.0, + 829.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1300.0, + 519.0, + 1300.0, + 519.0, + 1336.0, + 298.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1300.0, + 545.0, + 1300.0, + 545.0, + 1336.0, + 542.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1328.0, + 786.0, + 1328.0, + 786.0, + 1365.0, + 301.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1359.0, + 463.0, + 1359.0, + 463.0, + 1387.0, + 308.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1359.0, + 712.0, + 1359.0, + 712.0, + 1387.0, + 503.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1386.0, + 334.0, + 1386.0, + 334.0, + 1417.0, + 306.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1386.0, + 636.0, + 1386.0, + 636.0, + 1417.0, + 365.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1386.0, + 727.0, + 1386.0, + 727.0, + 1417.0, + 670.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1414.0, + 334.0, + 1414.0, + 334.0, + 1445.0, + 306.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1413.0, + 748.0, + 1413.0, + 748.0, + 1448.0, + 365.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1413.0, + 829.0, + 1413.0, + 829.0, + 1448.0, + 774.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1413.0, + 925.0, + 1413.0, + 925.0, + 1448.0, + 863.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1413.0, + 950.0, + 1413.0, + 950.0, + 1448.0, + 947.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1442.0, + 334.0, + 1442.0, + 334.0, + 1473.0, + 306.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1440.0, + 595.0, + 1440.0, + 595.0, + 1475.0, + 365.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1440.0, + 1122.0, + 1440.0, + 1122.0, + 1475.0, + 656.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1471.0, + 332.0, + 1471.0, + 332.0, + 1499.0, + 307.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1467.0, + 595.0, + 1467.0, + 595.0, + 1504.0, + 364.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1467.0, + 1118.0, + 1467.0, + 1118.0, + 1504.0, + 656.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1496.0, + 334.0, + 1496.0, + 334.0, + 1528.0, + 306.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1498.0, + 708.0, + 1498.0, + 708.0, + 1530.0, + 365.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1498.0, + 1183.0, + 1498.0, + 1183.0, + 1530.0, + 766.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1524.0, + 334.0, + 1524.0, + 334.0, + 1554.0, + 305.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1523.0, + 708.0, + 1523.0, + 708.0, + 1558.0, + 364.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1523.0, + 1182.0, + 1523.0, + 1182.0, + 1558.0, + 780.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1553.0, + 334.0, + 1553.0, + 334.0, + 1582.0, + 303.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1550.0, + 374.0, + 1550.0, + 374.0, + 1590.0, + 364.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1550.0, + 962.0, + 1550.0, + 962.0, + 1590.0, + 933.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1581.0, + 336.0, + 1581.0, + 336.0, + 1611.0, + 296.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1579.0, + 447.0, + 1579.0, + 447.0, + 1615.0, + 364.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1579.0, + 565.0, + 1579.0, + 565.0, + 1615.0, + 469.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1579.0, + 679.0, + 1579.0, + 679.0, + 1615.0, + 606.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1607.0, + 426.0, + 1607.0, + 426.0, + 1637.0, + 295.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 415.0, + 1633.0, + 415.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1633.0, + 449.0, + 1633.0, + 449.0, + 1669.0, + 438.0, + 1669.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 304, + 312, + 1402, + 312, + 1402, + 1249, + 304, + 1249 + ], + "score": 0.983, + "html": "
MethodminiImageNettieredImageNet
Backbone1-shot5-shot1-shot5-shot
MatchingNet (Vinyals etal.,2016)Conv4-6443.56± 0.8455.31± 0.73
ProtoNet† (Snell et al., 2017)Conv4-6452.61 ± 0.5271.33 ± 0.4153.33 ± 0.5072.10 ± 0.41
MAML (Finn et al., 2017)Conv4-6448.70 ± 1.8463.10 ±0.9251.67 ± 1.8170.30 ±0.08
Relation Net (Sung et al.,2018)Conv4-6450.40 ± 0.8065.30 ±0.7054.48 ± 0.9371.32 ± 0.78
IMP† (Allen et al., 2019)Conv4-6452.91 ± 0.4971.57 ± 0.4253.63 ± 0.5171.89 ± 0.44
DN4 (Li et al., 2019b)Conv4-6451.24 ± 0.7471.02 ± 0.641
DN PARN (Wu et al., 2019)Conv4-6455.22 ± 0.8471.55 ± 0.66
PN+rot (Gidaris et al., 2019)Conv4-6453.63 ± 0.4371.70 ± 0.36
CC+rot (Gidaris et al., 2019)Conv4-6454.83 ± 0.4371.86 ± 0.33
DSN-MR (Simon et al., 2020)Conv4-6455.88 ± 0.9070.50±0.68
Centroid (Afrasiyabi et al., 2020)Conv4-6453.14 ± 1.0671.45 ± 0.72
Neg-Cosine (Liu et al., 2020)Conv4-6452.84 ±0.7670.41 ± 0.66一
IEPT (ours)Conv4-6456.26± 0.4573.91±0.34 58.25±0.48 75.63±0.46
ProtoNet† (Snell et al.,2017)Conv4-51253.25± 0.4473.15 ± 0.3557.88 ±0.5076.82±0.40
MAML (Finn et al., 2017)Conv4-51249.33 ±0.6065.17 ± 0.4952.84± 0.5670.91 ± 0.46
Relation Net (Sung et al.,2018)Conv4-51250.86 ± 0.5767.32 ± 0.4454.69 ±0.5972.71 ±0.43
PN+rot (Gidaris et al.,2019)Conv4-51256.02 ±0.4674.00 ± 0.35一
CC+rot (Gidaris et al., 2019)Conv4-51256.27 ± 0.4374.30 ± 0.33一
IEPT (ours)Conv4-51258.43±0.4675.07±0.3360.91±0.5979.61±0.45
ProtoNet† (Snell et al.,2017)ResNet-1262.39 ±0.5180.53±0.4268.23±0.5084.03 ± 0.41
TADAM (Oreshkin et al., 2018)ResNet-1258.50 ± 0.3076.70 ±0.38一
MetaOptNet (Lee et al.,2019)ResNet-1262.64± 0.6178.63 ± 0.4665.99 ± 0.7281.56 ± 0.63
MTL (Sun et al., 2019)ResNet-1261.20 ± 1.8075.50 ± 0.8065.62 ± 1.8080.61 ±0.90
CAN (Hou et al., 2019)ResNet-1263.85 ±0.4879.44 ± 0.3469.89 ± 0.5184.23 ± 0.37
AM3 (Xing et al., 2019)ResNet-1265.21 ± 0.4975.20 ±0.3667.23 ± 0.3478.95 ± 0.22
Shot-Free (Ravichandran et al.,2019)ResNet-1259.04 ± 0.4377.64 ± 0.3966.87 ± 0.4382.64 ±0.43
Neg-Cosine (Liu et al., 2020)ResNet-1263.85 ± 0.8181.57 ± 0.56
Distill (Tian et al., 2020)ResNet-1264.82 ±0.6082.14 ± 0.4371.52 ± 0.6986.03 ±0.49
DSN-MR (Simon et al.,2020)ResNet-1264.60 ±0.7279.51± 0.5067.39 ± 0.8282.85 ±0.56
DeepEMD (Zhang et al.,2020)ResNet-1265.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
FEAT (Ye et al.,2020)ResNet-1266.78±0.2082.05 ± 0.1470.80 ±0.2384.79 ± 0.16
ProtoNet+Rotation (Su et al.,2020)ResNet-1876.00±0.6078.90±0.70
IEPT (ours)ResNet-1267.05±0.44882.90±0.30)72.24±0.50 86.73±0.34
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1683, + 1403, + 1683, + 1403, + 1807, + 298, + 1807 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1545, + 1403, + 1545, + 1403, + 1668, + 298, + 1668 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1403, + 1911, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1285, + 1404, + 1285, + 1404, + 1531, + 298, + 1531 + ], + "score": 0.969 + }, + { + "category_id": 6, + "poly": [ + 298, + 249, + 1392, + 249, + 1392, + 310, + 298, + 310 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 299, + 1850, + 550, + 1850, + 550, + 1882, + 299, + 1882 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.745 + }, + { + "category_id": 13, + "poly": [ + 958, + 1469, + 1051, + 1469, + 1051, + 1498, + 958, + 1498 + ], + "score": 0.9, + "latex": "8 4 \\times 8 4" + }, + { + "category_id": 13, + "poly": [ + 678, + 1608, + 733, + 1608, + 733, + 1637, + 678, + 1637 + ], + "score": 0.87, + "latex": "9 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 392, + 280, + 447, + 280, + 447, + 308, + 392, + 308 + ], + "score": 0.84, + "latex": "9 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 827, + 281, + 844, + 281, + 844, + 311, + 827, + 311 + ], + "score": 0.72, + "latex": "\\dagger" + }, + { + "category_id": 13, + "poly": [ + 484, + 1609, + 510, + 1609, + 510, + 1636, + 484, + 1636 + ], + "score": 0.66, + "latex": "\\%" + }, + { + "category_id": 13, + "poly": [ + 1366, + 251, + 1394, + 251, + 1394, + 278, + 1366, + 278 + ], + "score": 0.42, + "latex": "\\%" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 247.0, + 1365.0, + 247.0, + 1365.0, + 285.0, + 293.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 277.0, + 391.0, + 277.0, + 391.0, + 315.0, + 294.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 277.0, + 826.0, + 277.0, + 826.0, + 315.0, + 448.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 277.0, + 1370.0, + 277.0, + 1370.0, + 315.0, + 845.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 553.0, + 1847.0, + 553.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1682.0, + 1407.0, + 1682.0, + 1407.0, + 1719.0, + 293.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1713.0, + 1407.0, + 1713.0, + 1407.0, + 1751.0, + 292.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1782.0, + 292.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1775.0, + 854.0, + 1775.0, + 854.0, + 1811.0, + 294.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1545.0, + 1404.0, + 1545.0, + 1404.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1405.0, + 1577.0, + 1405.0, + 1610.0, + 294.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1608.0, + 483.0, + 1608.0, + 483.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1608.0, + 677.0, + 1608.0, + 677.0, + 1643.0, + 511.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1608.0, + 1405.0, + 1608.0, + 1405.0, + 1643.0, + 734.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1639.0, + 1315.0, + 1639.0, + 1315.0, + 1672.0, + 296.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1999.0, + 1405.0, + 1999.0, + 1405.0, + 2041.0, + 290.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1287.0, + 1405.0, + 1287.0, + 1405.0, + 1320.0, + 294.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1317.0, + 1408.0, + 1317.0, + 1408.0, + 1350.0, + 294.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 1406.0, + 1348.0, + 1406.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1377.0, + 1409.0, + 1377.0, + 1409.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1433.0, + 1405.0, + 1433.0, + 1405.0, + 1478.0, + 292.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1469.0, + 957.0, + 1469.0, + 957.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1469.0, + 1405.0, + 1469.0, + 1405.0, + 1503.0, + 1052.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1497.0, + 1144.0, + 1497.0, + 1144.0, + 1535.0, + 293.0, + 1535.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1697, + 1405, + 1697, + 1405, + 2035, + 297, + 2035 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1092, + 1405, + 1092, + 1405, + 1368, + 297, + 1368 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1382, + 1404, + 1382, + 1404, + 1598, + 298, + 1598 + ], + "score": 0.981 + }, + { + "category_id": 4, + "poly": [ + 296, + 878, + 1406, + 878, + 1406, + 1061, + 296, + 1061 + ], + "score": 0.964 + }, + { + "category_id": 6, + "poly": [ + 297, + 249, + 1405, + 249, + 1405, + 342, + 297, + 342 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 300, + 1639, + 655, + 1639, + 655, + 1669, + 300, + 1669 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.897 + }, + { + "category_id": 3, + "poly": [ + 321, + 577, + 1372, + 577, + 1372, + 871, + 321, + 871 + ], + "score": 0.798 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.795 + }, + { + "category_id": 5, + "poly": [ + 303, + 347, + 1396, + 347, + 1396, + 546, + 303, + 546 + ], + "score": 0.732, + "html": "
miniImageNettieredImageNet
LintegLinstLepisLaux1-shot5-shot1-shot5-shot
√55.04 ± 0.5272.01 ± 0.4156.98 ± 0.4774.15 ± 0.51
√√55.49 ± 0.5672.54± 0.4657.41 ± 0.5174.65 ± 0.50
√55.88± 0.4372.97 ± 0.4057.76± 0.4575.06 ± 0.40
√55.97 ± 0.5773.28 ± 0.3957.83 ± 0.5575.22 ± 0.48
√?>>>√56.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1416, + 1195, + 1416, + 1195, + 1447, + 1133, + 1447 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 942, + 1506, + 1018, + 1506, + 1018, + 1540, + 942, + 1540 + ], + "score": 0.91, + "latex": "L _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 726, + 1445, + 787, + 1445, + 787, + 1475, + 726, + 1475 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 870, + 280, + 934, + 280, + 934, + 314, + 870, + 314 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 401, + 1445, + 465, + 1445, + 465, + 1476, + 401, + 1476 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1446, + 1174, + 1446, + 1174, + 1478, + 1099, + 1478 + ], + "score": 0.9, + "latex": "\\bar { L } _ { i n t e g }" + }, + { + "category_id": 13, + "poly": [ + 298, + 311, + 358, + 311, + 358, + 342, + 298, + 342 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { a u x }" + }, + { + "category_id": 13, + "poly": [ + 985, + 280, + 1051, + 280, + 1051, + 311, + 985, + 311 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { i n s t . }" + }, + { + "category_id": 13, + "poly": [ + 1356, + 1246, + 1398, + 1246, + 1398, + 1274, + 1356, + 1274 + ], + "score": 0.87, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 409, + 311, + 486, + 311, + 486, + 345, + 409, + 345 + ], + "score": 0.87, + "latex": "L _ { i n t e g . }" + }, + { + "category_id": 13, + "poly": [ + 1282, + 1246, + 1324, + 1246, + 1324, + 1274, + 1282, + 1274 + ], + "score": 0.86, + "latex": "2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1366, + 1912, + 1397, + 1912, + 1397, + 1940, + 1366, + 1940 + ], + "score": 0.78, + "latex": "0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 1004, + 911, + 1028, + 911, + 1028, + 938, + 1004, + 938 + ], + "score": 0.78, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 306, + 941, + 331, + 941, + 331, + 968, + 306, + 968 + ], + "score": 0.63, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 442, + 911, + 478, + 911, + 478, + 941, + 442, + 941 + ], + "score": 0.33, + "latex": "\\mathbf { ( b ) }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 877.0, + 1408.0, + 877.0, + 1408.0, + 911.0, + 293.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 909.0, + 441.0, + 909.0, + 441.0, + 944.0, + 293.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 909.0, + 1003.0, + 909.0, + 1003.0, + 944.0, + 479.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 909.0, + 1406.0, + 909.0, + 1406.0, + 944.0, + 1029.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 938.0, + 305.0, + 938.0, + 305.0, + 975.0, + 293.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 938.0, + 1406.0, + 938.0, + 1406.0, + 975.0, + 332.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 971.0, + 1404.0, + 971.0, + 1404.0, + 1002.0, + 295.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 999.0, + 1404.0, + 999.0, + 1404.0, + 1038.0, + 293.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 1181.0, + 1030.0, + 1181.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 247.0, + 1406.0, + 247.0, + 1406.0, + 285.0, + 292.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 276.0, + 869.0, + 276.0, + 869.0, + 316.0, + 292.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 276.0, + 984.0, + 276.0, + 984.0, + 316.0, + 935.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 276.0, + 1410.0, + 276.0, + 1410.0, + 316.0, + 1052.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 310.0, + 408.0, + 310.0, + 408.0, + 347.0, + 359.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 310.0, + 972.0, + 310.0, + 972.0, + 347.0, + 487.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 658.0, + 1637.0, + 658.0, + 1673.0, + 294.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 582.0, + 369.0, + 582.0, + 369.0, + 642.0, + 341.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 590.0, + 528.0, + 590.0, + 528.0, + 654.0, + 402.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 591.0, + 718.0, + 591.0, + 718.0, + 649.0, + 694.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 591.0, + 785.0, + 591.0, + 785.0, + 643.0, + 751.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 591.0, + 1067.0, + 591.0, + 1067.0, + 650.0, + 1046.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 586.0, + 1179.0, + 586.0, + 1179.0, + 658.0, + 1095.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 650.0, + 365.0, + 650.0, + 365.0, + 751.0, + 325.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 646.0, + 501.0, + 646.0, + 501.0, + 680.0, + 400.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 659.0, + 717.0, + 659.0, + 717.0, + 750.0, + 677.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 662.0, + 1067.0, + 662.0, + 1067.0, + 747.0, + 1028.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 681.0, + 365.0, + 681.0, + 365.0, + 701.0, + 342.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 709.0, + 366.0, + 709.0, + 366.0, + 734.0, + 341.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 708.0, + 717.0, + 708.0, + 717.0, + 729.0, + 694.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 708.0, + 1066.0, + 708.0, + 1066.0, + 729.0, + 1046.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 740.0, + 367.0, + 740.0, + 367.0, + 765.0, + 342.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 747.0, + 717.0, + 747.0, + 717.0, + 768.0, + 694.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 750.0, + 1067.0, + 750.0, + 1067.0, + 768.0, + 1046.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 772.0, + 367.0, + 772.0, + 367.0, + 795.0, + 341.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 787.0, + 718.0, + 787.0, + 718.0, + 806.0, + 694.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 789.0, + 1067.0, + 789.0, + 1067.0, + 808.0, + 1044.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 800.0, + 367.0, + 800.0, + 367.0, + 824.0, + 341.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 815.0, + 478.0, + 815.0, + 478.0, + 837.0, + 389.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 814.0, + 622.0, + 814.0, + 622.0, + 839.0, + 535.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 811.0, + 831.0, + 811.0, + 831.0, + 839.0, + 742.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 811.0, + 975.0, + 811.0, + 975.0, + 839.0, + 885.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 815.0, + 1177.0, + 815.0, + 1177.0, + 839.0, + 1097.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 815.0, + 1324.0, + 815.0, + 1324.0, + 839.0, + 1243.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 842.0, + 510.0, + 842.0, + 510.0, + 875.0, + 469.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 840.0, + 864.0, + 840.0, + 864.0, + 877.0, + 820.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 842.0, + 1216.0, + 842.0, + 1216.0, + 877.0, + 1175.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 1405.0, + 1695.0, + 1405.0, + 1733.0, + 294.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1764.0, + 294.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1796.0, + 292.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1825.0, + 294.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 292.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1911.0, + 1365.0, + 1911.0, + 1365.0, + 1946.0, + 291.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1911.0, + 1408.0, + 1911.0, + 1408.0, + 1946.0, + 1398.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1978.0, + 292.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1093.0, + 1403.0, + 1093.0, + 1403.0, + 1126.0, + 295.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1123.0, + 1407.0, + 1123.0, + 1407.0, + 1158.0, + 292.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 1406.0, + 1152.0, + 1406.0, + 1188.0, + 295.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 1406.0, + 1182.0, + 1406.0, + 1221.0, + 294.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1407.0, + 1216.0, + 1407.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1243.0, + 1281.0, + 1243.0, + 1281.0, + 1280.0, + 295.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1243.0, + 1355.0, + 1243.0, + 1355.0, + 1280.0, + 1325.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1243.0, + 1407.0, + 1243.0, + 1407.0, + 1280.0, + 1399.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1274.0, + 1407.0, + 1274.0, + 1407.0, + 1311.0, + 295.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1307.0, + 1405.0, + 1307.0, + 1405.0, + 1340.0, + 296.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1336.0, + 1122.0, + 1336.0, + 1122.0, + 1372.0, + 295.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1380.0, + 1405.0, + 1380.0, + 1405.0, + 1420.0, + 292.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1413.0, + 1132.0, + 1413.0, + 1132.0, + 1452.0, + 293.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1413.0, + 1406.0, + 1413.0, + 1406.0, + 1452.0, + 1196.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1442.0, + 400.0, + 1442.0, + 400.0, + 1484.0, + 293.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1442.0, + 725.0, + 1442.0, + 725.0, + 1484.0, + 466.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1442.0, + 1098.0, + 1442.0, + 1098.0, + 1484.0, + 788.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1442.0, + 1408.0, + 1442.0, + 1408.0, + 1484.0, + 1175.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1475.0, + 1407.0, + 1475.0, + 1407.0, + 1509.0, + 293.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1504.0, + 941.0, + 1504.0, + 941.0, + 1541.0, + 293.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1504.0, + 1406.0, + 1504.0, + 1406.0, + 1541.0, + 1019.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1536.0, + 1405.0, + 1536.0, + 1405.0, + 1571.0, + 293.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1565.0, + 1287.0, + 1565.0, + 1287.0, + 1601.0, + 293.0, + 1601.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1135, + 1405, + 1135, + 1405, + 1441, + 297, + 1441 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1567, + 1405, + 1567, + 1405, + 1814, + 298, + 1814 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 302, + 316, + 1390, + 316, + 1390, + 529, + 302, + 529 + ], + "score": 0.979, + "html": "
MethodBackboneCUBminiImageNet→CUB
1-shot5-shot1-shot5-shot
MatchingNet (Vinyals et al.,2016)Conv4-6461.16±0.8972.86±0.7042.62 ± 0.5556.53± 0.44
ProtoNet (Snell et al.,2017)Conv4-6463.72 ±0.2281.50 ± 0.1550.51 ± 0.5669.28±0.40
MAML (Finn et al.,2017)Conv4-6455.92 ± 0.9572.09 ±0.7643.59 ± 0.5454.18 ± 0.41
Relation Net (Sung et al., 2018)Conv4-6462.45 ± 0.9876.11 ± 0.6949.84 ±0.5468.98 ±0.42
FEAT(Ye et al.,2020)Conv4-6468.87 ±0.2282.90±0.1551.52 ± 0.5470.16 ± 0.40
IEPT(ours)Conv4-6469.97±0.4984.33± 0.3352.68±0.5672.98±0.40
" + }, + { + "category_id": 1, + "poly": [ + 298, + 767, + 1405, + 767, + 1405, + 921, + 298, + 921 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 935, + 1404, + 935, + 1404, + 1120, + 297, + 1120 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 567, + 1404, + 567, + 1404, + 751, + 298, + 751 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1403, + 1911, + 1403, + 2033, + 299, + 2033 + ], + "score": 0.973 + }, + { + "category_id": 6, + "poly": [ + 296, + 248, + 1400, + 248, + 1400, + 310, + 296, + 310 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 300, + 1492, + 543, + 1492, + 543, + 1528, + 300, + 1528 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 301, + 1858, + 557, + 1858, + 557, + 1886, + 301, + 1886 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.778 + }, + { + "category_id": 13, + "poly": [ + 503, + 829, + 578, + 829, + 578, + 857, + 503, + 857 + ], + "score": 0.91, + "latex": "R = 1" + }, + { + "category_id": 13, + "poly": [ + 395, + 1379, + 463, + 1379, + 463, + 1410, + 395, + 1410 + ], + "score": 0.88, + "latex": "1 - 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 366, + 1059, + 429, + 1059, + 429, + 1087, + 366, + 1087 + ], + "score": 0.86, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 712, + 798, + 755, + 798, + 755, + 827, + 712, + 827 + ], + "score": 0.83, + "latex": "9 0 °" + }, + { + "category_id": 13, + "poly": [ + 672, + 799, + 701, + 799, + 701, + 827, + 672, + 827 + ], + "score": 0.81, + "latex": "0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 767, + 798, + 823, + 798, + 823, + 828, + 767, + 828 + ], + "score": 0.81, + "latex": "1 8 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 398, + 799, + 423, + 799, + 423, + 825, + 398, + 825 + ], + "score": 0.76, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 834, + 798, + 893, + 798, + 893, + 828, + 834, + 828 + ], + "score": 0.75, + "latex": "2 7 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 860, + 1157, + 860, + 1157, + 886, + 1132, + 886 + ], + "score": 0.75, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1042, + 1260, + 1076, + 1260, + 1076, + 1286, + 1042, + 1286 + ], + "score": 0.7, + "latex": "" + }, + { + "category_id": 13, + "poly": [ + 550, + 282, + 586, + 282, + 586, + 308, + 550, + 308 + ], + "score": 0.64, + "latex": "" + }, + { + "category_id": 13, + "poly": [ + 298, + 829, + 323, + 829, + 323, + 856, + 298, + 856 + ], + "score": 0.61, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 790, + 1197, + 887, + 1197, + 887, + 1226, + 790, + 1226 + ], + "score": 0.53, + "latex": " \\mathrm { C U B }" + }, + { + "category_id": 13, + "poly": [ + 1041, + 1258, + 1135, + 1258, + 1135, + 1287, + 1041, + 1287 + ], + "score": 0.42, + "latex": " \\mathrm { C U B }" + }, + { + "category_id": 13, + "poly": [ + 791, + 1199, + 827, + 1199, + 827, + 1225, + 791, + 1225 + ], + "score": 0.38, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 249.0, + 1404.0, + 249.0, + 1404.0, + 284.0, + 294.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 279.0, + 549.0, + 279.0, + 549.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 279.0, + 656.0, + 279.0, + 656.0, + 314.0, + 587.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1489.0, + 549.0, + 1489.0, + 549.0, + 1537.0, + 291.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1859.0, + 558.0, + 1859.0, + 558.0, + 1888.0, + 299.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1134.0, + 1407.0, + 1134.0, + 1407.0, + 1169.0, + 295.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1167.0, + 1405.0, + 1167.0, + 1405.0, + 1199.0, + 296.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1194.0, + 789.0, + 1194.0, + 789.0, + 1231.0, + 294.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1194.0, + 1408.0, + 1194.0, + 1408.0, + 1231.0, + 888.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1226.0, + 1405.0, + 1226.0, + 1405.0, + 1261.0, + 295.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1259.0, + 1040.0, + 1259.0, + 1040.0, + 1291.0, + 296.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1259.0, + 1403.0, + 1259.0, + 1403.0, + 1291.0, + 1136.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1288.0, + 1402.0, + 1288.0, + 1402.0, + 1320.0, + 296.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1317.0, + 1408.0, + 1317.0, + 1408.0, + 1354.0, + 292.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1346.0, + 1408.0, + 1346.0, + 1408.0, + 1384.0, + 292.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1378.0, + 394.0, + 1378.0, + 394.0, + 1414.0, + 291.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1378.0, + 1408.0, + 1378.0, + 1408.0, + 1414.0, + 464.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1409.0, + 864.0, + 1409.0, + 864.0, + 1444.0, + 295.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1569.0, + 1403.0, + 1569.0, + 1403.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 1406.0, + 1599.0, + 1406.0, + 1633.0, + 293.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1629.0, + 1407.0, + 1629.0, + 1407.0, + 1665.0, + 293.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1692.0, + 1405.0, + 1692.0, + 1405.0, + 1725.0, + 295.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1721.0, + 1407.0, + 1721.0, + 1407.0, + 1754.0, + 295.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1750.0, + 1405.0, + 1750.0, + 1405.0, + 1786.0, + 293.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1780.0, + 1217.0, + 1780.0, + 1217.0, + 1818.0, + 293.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 767.0, + 1405.0, + 767.0, + 1405.0, + 800.0, + 296.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 796.0, + 397.0, + 796.0, + 397.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 796.0, + 671.0, + 796.0, + 671.0, + 832.0, + 424.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 796.0, + 711.0, + 796.0, + 711.0, + 832.0, + 702.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 796.0, + 766.0, + 796.0, + 766.0, + 832.0, + 756.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 796.0, + 833.0, + 796.0, + 833.0, + 832.0, + 824.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 796.0, + 1406.0, + 796.0, + 1406.0, + 832.0, + 894.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 827.0, + 502.0, + 827.0, + 502.0, + 863.0, + 324.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 827.0, + 1407.0, + 827.0, + 1407.0, + 863.0, + 579.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 857.0, + 1131.0, + 857.0, + 1131.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 857.0, + 1410.0, + 857.0, + 1410.0, + 894.0, + 1158.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 889.0, + 1406.0, + 889.0, + 1406.0, + 926.0, + 293.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 935.0, + 1406.0, + 935.0, + 1406.0, + 969.0, + 294.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 968.0, + 1406.0, + 968.0, + 1406.0, + 1000.0, + 296.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 997.0, + 1407.0, + 997.0, + 1407.0, + 1033.0, + 295.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1061.0, + 296.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1059.0, + 365.0, + 1059.0, + 365.0, + 1091.0, + 296.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1059.0, + 1404.0, + 1059.0, + 1404.0, + 1091.0, + 430.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 897.0, + 1089.0, + 897.0, + 1123.0, + 295.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 568.0, + 1404.0, + 568.0, + 1404.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 598.0, + 1408.0, + 598.0, + 1408.0, + 634.0, + 294.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 626.0, + 1406.0, + 626.0, + 1406.0, + 668.0, + 293.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 1404.0, + 659.0, + 1404.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 689.0, + 1404.0, + 689.0, + 1404.0, + 725.0, + 293.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 721.0, + 708.0, + 721.0, + 708.0, + 754.0, + 293.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1311.0, + 2001.0, + 1311.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 299, + 1373, + 1402, + 1373, + 1402, + 1433, + 299, + 1433 + ], + "score": 0.608 + }, + { + "category_id": 1, + "poly": [ + 299, + 1451, + 1400, + 1451, + 1400, + 1512, + 299, + 1512 + ], + "score": 0.582 + }, + { + "category_id": 1, + "poly": [ + 296, + 1529, + 1401, + 1529, + 1401, + 1590, + 296, + 1590 + ], + "score": 0.556 + }, + { + "category_id": 1, + "poly": [ + 298, + 277, + 1400, + 277, + 1400, + 334, + 298, + 334 + ], + "score": 0.554 + }, + { + "category_id": 1, + "poly": [ + 299, + 743, + 1401, + 743, + 1401, + 804, + 299, + 804 + ], + "score": 0.544 + }, + { + "category_id": 1, + "poly": [ + 297, + 821, + 1402, + 821, + 1402, + 909, + 297, + 909 + ], + "score": 0.543 + }, + { + "category_id": 1, + "poly": [ + 296, + 928, + 1401, + 928, + 1401, + 988, + 296, + 988 + ], + "score": 0.533 + }, + { + "category_id": 1, + "poly": [ + 300, + 1005, + 1401, + 1005, + 1401, + 1092, + 300, + 1092 + ], + "score": 0.521 + }, + { + "category_id": 1, + "poly": [ + 300, + 1189, + 1401, + 1189, + 1401, + 1275, + 300, + 1275 + ], + "score": 0.515 + }, + { + "category_id": 1, + "poly": [ + 300, + 1296, + 1401, + 1296, + 1401, + 1355, + 300, + 1355 + ], + "score": 0.512 + }, + { + "category_id": 1, + "poly": [ + 300, + 509, + 1401, + 509, + 1401, + 570, + 300, + 570 + ], + "score": 0.51 + }, + { + "category_id": 1, + "poly": [ + 299, + 587, + 1401, + 587, + 1401, + 649, + 299, + 649 + ], + "score": 0.505 + }, + { + "category_id": 1, + "poly": [ + 296, + 353, + 1398, + 353, + 1398, + 413, + 296, + 413 + ], + "score": 0.504 + }, + { + "category_id": 1, + "poly": [ + 296, + 1607, + 1402, + 1607, + 1402, + 1669, + 296, + 1669 + ], + "score": 0.499 + }, + { + "category_id": 1, + "poly": [ + 299, + 666, + 1402, + 666, + 1402, + 727, + 299, + 727 + ], + "score": 0.496 + }, + { + "category_id": 1, + "poly": [ + 294, + 1976, + 1402, + 1976, + 1402, + 2034, + 294, + 2034 + ], + "score": 0.485 + }, + { + "category_id": 1, + "poly": [ + 296, + 1111, + 1400, + 1111, + 1400, + 1172, + 296, + 1172 + ], + "score": 0.484 + }, + { + "category_id": 1, + "poly": [ + 300, + 431, + 1396, + 431, + 1396, + 492, + 300, + 492 + ], + "score": 0.483 + }, + { + "category_id": 1, + "poly": [ + 300, + 1686, + 1403, + 1686, + 1403, + 1773, + 300, + 1773 + ], + "score": 0.48 + }, + { + "category_id": 1, + "poly": [ + 296, + 1898, + 1398, + 1898, + 1398, + 1957, + 296, + 1957 + ], + "score": 0.449 + }, + { + "category_id": 1, + "poly": [ + 296, + 1792, + 1403, + 1792, + 1403, + 1877, + 296, + 1877 + ], + "score": 0.423 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1375.0, + 1403.0, + 1375.0, + 1403.0, + 1407.0, + 297.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1401.0, + 1261.0, + 1401.0, + 1261.0, + 1434.0, + 320.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1454.0, + 1404.0, + 1454.0, + 1404.0, + 1485.0, + 297.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1480.0, + 1264.0, + 1480.0, + 1264.0, + 1513.0, + 320.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1406.0, + 1529.0, + 1406.0, + 1563.0, + 294.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1559.0, + 1255.0, + 1559.0, + 1255.0, + 1591.0, + 324.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 274.0, + 1404.0, + 274.0, + 1404.0, + 312.0, + 294.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 304.0, + 1082.0, + 304.0, + 1082.0, + 336.0, + 322.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 742.0, + 1405.0, + 742.0, + 1405.0, + 780.0, + 294.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 771.0, + 1254.0, + 771.0, + 1254.0, + 805.0, + 320.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 823.0, + 1406.0, + 823.0, + 1406.0, + 856.0, + 295.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 849.0, + 1408.0, + 849.0, + 1408.0, + 889.0, + 320.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 877.0, + 487.0, + 877.0, + 487.0, + 910.0, + 323.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 928.0, + 1403.0, + 928.0, + 1403.0, + 963.0, + 296.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 957.0, + 1096.0, + 957.0, + 1096.0, + 989.0, + 321.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1005.0, + 1406.0, + 1005.0, + 1406.0, + 1040.0, + 294.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1033.0, + 1406.0, + 1033.0, + 1406.0, + 1067.0, + 320.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1059.0, + 391.0, + 1059.0, + 391.0, + 1095.0, + 320.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1189.0, + 1406.0, + 1189.0, + 1406.0, + 1223.0, + 294.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1219.0, + 1407.0, + 1219.0, + 1407.0, + 1252.0, + 321.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1245.0, + 390.0, + 1245.0, + 390.0, + 1277.0, + 320.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1294.0, + 1405.0, + 1294.0, + 1405.0, + 1330.0, + 294.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1324.0, + 1099.0, + 1324.0, + 1099.0, + 1355.0, + 323.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 546.0, + 295.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 538.0, + 808.0, + 538.0, + 808.0, + 570.0, + 324.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 588.0, + 1403.0, + 588.0, + 1403.0, + 622.0, + 295.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 617.0, + 1243.0, + 617.0, + 1243.0, + 650.0, + 322.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 353.0, + 1404.0, + 353.0, + 1404.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 383.0, + 1310.0, + 383.0, + 1310.0, + 414.0, + 322.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1611.0, + 1402.0, + 1611.0, + 1402.0, + 1639.0, + 299.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1637.0, + 1409.0, + 1637.0, + 1409.0, + 1670.0, + 321.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 668.0, + 1404.0, + 668.0, + 1404.0, + 699.0, + 295.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 695.0, + 973.0, + 695.0, + 973.0, + 727.0, + 322.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1978.0, + 1404.0, + 1978.0, + 1404.0, + 2009.0, + 297.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 1218.0, + 2003.0, + 1218.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1110.0, + 1405.0, + 1110.0, + 1405.0, + 1150.0, + 294.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1141.0, + 1152.0, + 1141.0, + 1152.0, + 1173.0, + 323.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 433.0, + 1401.0, + 433.0, + 1401.0, + 465.0, + 296.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 461.0, + 1368.0, + 461.0, + 1368.0, + 493.0, + 324.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1685.0, + 1404.0, + 1685.0, + 1404.0, + 1719.0, + 295.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1747.0, + 321.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1742.0, + 687.0, + 1742.0, + 687.0, + 1774.0, + 323.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1895.0, + 1403.0, + 1895.0, + 1403.0, + 1934.0, + 294.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1926.0, + 766.0, + 1926.0, + 766.0, + 1957.0, + 322.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1825.0, + 294.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1821.0, + 1408.0, + 1821.0, + 1408.0, + 1853.0, + 321.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1847.0, + 391.0, + 1847.0, + 391.0, + 1878.0, + 320.0, + 1878.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 296, + 864, + 1400, + 864, + 1400, + 924, + 296, + 924 + ], + "score": 0.628 + }, + { + "category_id": 1, + "poly": [ + 293, + 943, + 1401, + 943, + 1401, + 1004, + 293, + 1004 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 297, + 1763, + 1401, + 1763, + 1401, + 1823, + 297, + 1823 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 294, + 1023, + 1401, + 1023, + 1401, + 1083, + 294, + 1083 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 292, + 231, + 1404, + 231, + 1404, + 291, + 292, + 291 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 297, + 1684, + 1402, + 1684, + 1402, + 1744, + 297, + 1744 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 296, + 785, + 1402, + 785, + 1402, + 846, + 296, + 846 + ], + "score": 0.593 + }, + { + "category_id": 1, + "poly": [ + 297, + 1100, + 1403, + 1100, + 1403, + 1188, + 297, + 1188 + ], + "score": 0.592 + }, + { + "category_id": 1, + "poly": [ + 294, + 522, + 1403, + 522, + 1403, + 609, + 294, + 609 + ], + "score": 0.592 + }, + { + "category_id": 1, + "poly": [ + 295, + 1208, + 1401, + 1208, + 1401, + 1295, + 295, + 1295 + ], + "score": 0.586 + }, + { + "category_id": 1, + "poly": [ + 295, + 1313, + 1399, + 1313, + 1399, + 1374, + 295, + 1374 + ], + "score": 0.585 + }, + { + "category_id": 1, + "poly": [ + 299, + 1577, + 1402, + 1577, + 1402, + 1663, + 299, + 1663 + ], + "score": 0.581 + }, + { + "category_id": 1, + "poly": [ + 305, + 1842, + 1400, + 1842, + 1400, + 1929, + 305, + 1929 + ], + "score": 0.57 + }, + { + "category_id": 1, + "poly": [ + 299, + 310, + 1400, + 310, + 1400, + 395, + 299, + 395 + ], + "score": 0.566 + }, + { + "category_id": 1, + "poly": [ + 294, + 415, + 1402, + 415, + 1402, + 503, + 294, + 503 + ], + "score": 0.56 + }, + { + "category_id": 1, + "poly": [ + 290, + 1498, + 1403, + 1498, + 1403, + 1560, + 290, + 1560 + ], + "score": 0.558 + }, + { + "category_id": 1, + "poly": [ + 294, + 1392, + 1402, + 1392, + 1402, + 1479, + 294, + 1479 + ], + "score": 0.551 + }, + { + "category_id": 1, + "poly": [ + 297, + 628, + 1401, + 628, + 1401, + 689, + 297, + 689 + ], + "score": 0.535 + }, + { + "category_id": 1, + "poly": [ + 300, + 1948, + 1403, + 1948, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.431 + }, + { + "category_id": 1, + "poly": [ + 291, + 707, + 1402, + 707, + 1402, + 767, + 291, + 767 + ], + "score": 0.413 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 859.0, + 1405.0, + 859.0, + 1405.0, + 903.0, + 294.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 893.0, + 638.0, + 893.0, + 638.0, + 925.0, + 321.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 942.0, + 1404.0, + 942.0, + 1404.0, + 979.0, + 295.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 972.0, + 1275.0, + 972.0, + 1275.0, + 1004.0, + 324.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1764.0, + 1403.0, + 1764.0, + 1403.0, + 1795.0, + 297.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1792.0, + 1343.0, + 1792.0, + 1343.0, + 1825.0, + 322.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1020.0, + 1405.0, + 1020.0, + 1405.0, + 1058.0, + 294.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1051.0, + 1018.0, + 1051.0, + 1018.0, + 1082.0, + 323.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 232.0, + 1407.0, + 232.0, + 1407.0, + 263.0, + 298.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 261.0, + 1071.0, + 261.0, + 1071.0, + 292.0, + 324.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1681.0, + 1404.0, + 1681.0, + 1404.0, + 1719.0, + 293.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1714.0, + 765.0, + 1714.0, + 765.0, + 1744.0, + 322.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 787.0, + 1404.0, + 787.0, + 1404.0, + 818.0, + 296.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 815.0, + 1022.0, + 815.0, + 1022.0, + 847.0, + 324.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1097.0, + 1405.0, + 1097.0, + 1405.0, + 1138.0, + 292.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1125.0, + 1409.0, + 1125.0, + 1409.0, + 1168.0, + 320.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1156.0, + 513.0, + 1156.0, + 513.0, + 1189.0, + 323.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1406.0, + 520.0, + 1406.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 551.0, + 1403.0, + 551.0, + 1403.0, + 581.0, + 323.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 578.0, + 651.0, + 578.0, + 651.0, + 611.0, + 322.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1206.0, + 1403.0, + 1206.0, + 1403.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1236.0, + 1406.0, + 1236.0, + 1406.0, + 1269.0, + 323.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1264.0, + 768.0, + 1264.0, + 768.0, + 1297.0, + 323.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1315.0, + 1404.0, + 1315.0, + 1404.0, + 1346.0, + 296.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1343.0, + 1375.0, + 1343.0, + 1375.0, + 1375.0, + 324.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1404.0, + 1577.0, + 1404.0, + 1612.0, + 294.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1607.0, + 1407.0, + 1607.0, + 1407.0, + 1640.0, + 322.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1633.0, + 389.0, + 1633.0, + 389.0, + 1665.0, + 320.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1842.0, + 1404.0, + 1842.0, + 1404.0, + 1875.0, + 300.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1871.0, + 1404.0, + 1871.0, + 1404.0, + 1901.0, + 323.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1898.0, + 884.0, + 1898.0, + 884.0, + 1931.0, + 323.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 310.0, + 1404.0, + 310.0, + 1404.0, + 342.0, + 295.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 337.0, + 1406.0, + 337.0, + 1406.0, + 371.0, + 319.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 366.0, + 462.0, + 366.0, + 462.0, + 396.0, + 323.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 416.0, + 1404.0, + 416.0, + 1404.0, + 450.0, + 294.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 443.0, + 1405.0, + 443.0, + 1405.0, + 479.0, + 321.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 473.0, + 589.0, + 473.0, + 589.0, + 506.0, + 322.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1528.0, + 1095.0, + 1528.0, + 1095.0, + 1559.0, + 322.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1392.0, + 1407.0, + 1392.0, + 1407.0, + 1425.0, + 296.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1417.0, + 1408.0, + 1417.0, + 1408.0, + 1458.0, + 320.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1450.0, + 513.0, + 1450.0, + 513.0, + 1479.0, + 323.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 627.0, + 1408.0, + 627.0, + 1408.0, + 665.0, + 293.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 658.0, + 1405.0, + 658.0, + 1405.0, + 690.0, + 325.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1984.0, + 294.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1977.0, + 1405.0, + 1977.0, + 1405.0, + 2010.0, + 323.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 2004.0, + 1283.0, + 2004.0, + 1283.0, + 2037.0, + 319.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 709.0, + 1403.0, + 709.0, + 1403.0, + 741.0, + 295.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 735.0, + 950.0, + 735.0, + 950.0, + 768.0, + 323.0, + 768.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 299, + 1403, + 1403, + 1403, + 1403, + 1464, + 299, + 1464 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1401, + 1481, + 1401, + 1542, + 298, + 1542 + ], + "score": 0.657 + }, + { + "category_id": 1, + "poly": [ + 293, + 1847, + 1404, + 1847, + 1404, + 1907, + 293, + 1907 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 297, + 231, + 1400, + 231, + 1400, + 317, + 297, + 317 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 297, + 701, + 1400, + 701, + 1400, + 761, + 297, + 761 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 296, + 1327, + 1400, + 1327, + 1400, + 1387, + 296, + 1387 + ], + "score": 0.622 + }, + { + "category_id": 1, + "poly": [ + 297, + 1558, + 1401, + 1558, + 1401, + 1619, + 297, + 1619 + ], + "score": 0.618 + }, + { + "category_id": 1, + "poly": [ + 296, + 623, + 1401, + 623, + 1401, + 684, + 296, + 684 + ], + "score": 0.613 + }, + { + "category_id": 1, + "poly": [ + 292, + 1248, + 1402, + 1248, + 1402, + 1310, + 292, + 1310 + ], + "score": 0.604 + }, + { + "category_id": 1, + "poly": [ + 297, + 962, + 1403, + 962, + 1403, + 1047, + 297, + 1047 + ], + "score": 0.603 + }, + { + "category_id": 1, + "poly": [ + 298, + 1742, + 1402, + 1742, + 1402, + 1828, + 298, + 1828 + ], + "score": 0.588 + }, + { + "category_id": 1, + "poly": [ + 298, + 1636, + 1403, + 1636, + 1403, + 1724, + 298, + 1724 + ], + "score": 0.585 + }, + { + "category_id": 1, + "poly": [ + 294, + 778, + 1400, + 778, + 1400, + 839, + 294, + 839 + ], + "score": 0.584 + }, + { + "category_id": 1, + "poly": [ + 294, + 1143, + 1403, + 1143, + 1403, + 1231, + 294, + 1231 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 300, + 857, + 1400, + 857, + 1400, + 943, + 300, + 943 + ], + "score": 0.579 + }, + { + "category_id": 1, + "poly": [ + 298, + 335, + 1399, + 335, + 1399, + 423, + 298, + 423 + ], + "score": 0.567 + }, + { + "category_id": 1, + "poly": [ + 295, + 1067, + 1399, + 1067, + 1399, + 1126, + 295, + 1126 + ], + "score": 0.566 + }, + { + "category_id": 1, + "poly": [ + 297, + 546, + 1399, + 546, + 1399, + 606, + 297, + 606 + ], + "score": 0.557 + }, + { + "category_id": 1, + "poly": [ + 299, + 440, + 1400, + 440, + 1400, + 529, + 299, + 529 + ], + "score": 0.541 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1406.0, + 1407.0, + 1406.0, + 1407.0, + 1437.0, + 298.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1433.0, + 1301.0, + 1433.0, + 1301.0, + 1465.0, + 323.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1517.0, + 294.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1511.0, + 1098.0, + 1511.0, + 1098.0, + 1543.0, + 322.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1848.0, + 1404.0, + 1848.0, + 1404.0, + 1879.0, + 296.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1876.0, + 1398.0, + 1876.0, + 1398.0, + 1909.0, + 320.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 257.0, + 1405.0, + 257.0, + 1405.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 287.0, + 637.0, + 287.0, + 637.0, + 319.0, + 322.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 698.0, + 1406.0, + 698.0, + 1406.0, + 739.0, + 292.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 731.0, + 935.0, + 731.0, + 935.0, + 762.0, + 323.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1328.0, + 1402.0, + 1328.0, + 1402.0, + 1359.0, + 296.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1355.0, + 1017.0, + 1355.0, + 1017.0, + 1388.0, + 321.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1591.0, + 298.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1586.0, + 1209.0, + 1586.0, + 1209.0, + 1620.0, + 321.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 622.0, + 1405.0, + 622.0, + 1405.0, + 660.0, + 294.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 653.0, + 956.0, + 653.0, + 956.0, + 685.0, + 323.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1250.0, + 1404.0, + 1250.0, + 1404.0, + 1282.0, + 297.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1279.0, + 1334.0, + 1279.0, + 1334.0, + 1311.0, + 324.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 958.0, + 1407.0, + 958.0, + 1407.0, + 998.0, + 296.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 991.0, + 1408.0, + 991.0, + 1408.0, + 1023.0, + 322.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1017.0, + 389.0, + 1017.0, + 389.0, + 1048.0, + 321.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1742.0, + 1404.0, + 1742.0, + 1404.0, + 1775.0, + 296.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1770.0, + 1406.0, + 1770.0, + 1406.0, + 1804.0, + 322.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1798.0, + 870.0, + 1798.0, + 870.0, + 1829.0, + 321.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 1407.0, + 1635.0, + 1407.0, + 1672.0, + 293.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1665.0, + 1407.0, + 1665.0, + 1407.0, + 1699.0, + 321.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1692.0, + 389.0, + 1692.0, + 389.0, + 1723.0, + 321.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 781.0, + 1404.0, + 781.0, + 1404.0, + 812.0, + 298.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 808.0, + 1250.0, + 808.0, + 1250.0, + 840.0, + 323.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1142.0, + 1408.0, + 1142.0, + 1408.0, + 1178.0, + 294.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1169.0, + 1409.0, + 1169.0, + 1409.0, + 1209.0, + 320.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1199.0, + 514.0, + 1199.0, + 514.0, + 1232.0, + 322.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 853.0, + 1406.0, + 853.0, + 1406.0, + 893.0, + 293.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 886.0, + 1405.0, + 886.0, + 1405.0, + 917.0, + 321.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 913.0, + 766.0, + 913.0, + 766.0, + 945.0, + 322.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 334.0, + 1403.0, + 334.0, + 1403.0, + 371.0, + 294.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 361.0, + 1404.0, + 361.0, + 1404.0, + 403.0, + 319.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 394.0, + 511.0, + 394.0, + 511.0, + 424.0, + 324.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1067.0, + 1403.0, + 1067.0, + 1403.0, + 1100.0, + 295.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1095.0, + 1241.0, + 1095.0, + 1241.0, + 1128.0, + 320.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 545.0, + 1403.0, + 545.0, + 1403.0, + 584.0, + 295.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 575.0, + 1301.0, + 575.0, + 1301.0, + 610.0, + 322.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 440.0, + 1405.0, + 440.0, + 1405.0, + 475.0, + 294.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 470.0, + 1405.0, + 470.0, + 1405.0, + 503.0, + 323.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 499.0, + 1240.0, + 499.0, + 1240.0, + 532.0, + 322.0, + 532.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 365, + 1405, + 365, + 1405, + 733, + 297, + 733 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1475, + 1405, + 1475, + 1405, + 1751, + 297, + 1751 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 314, + 1859, + 1385, + 1859, + 1385, + 2001, + 314, + 2001 + ], + "score": 0.976, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
concatenatingConv4-6451.52±0.6073.36± 0.4950.78±0.6872.79± 0.57
averagingConv4-6451.58 ± 0.6270.97 ± 0.5453.91 ± 0.6972.52 ± 0.59
transformer (ours)Conv4-6456.26± 0.4573.91 ± 0.3458.25 ± 0.4875.63± 0.46
" + }, + { + "category_id": 3, + "poly": [ + 304, + 1033, + 1390, + 1033, + 1390, + 1288, + 304, + 1288 + ], + "score": 0.962 + }, + { + "category_id": 5, + "poly": [ + 319, + 840, + 1379, + 840, + 1379, + 981, + 319, + 981 + ], + "score": 0.962, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
ProtoNetConv4-6452.61 ±0.5271.33 ± 0.4153.33 ± 0.5072.10± 0.41
IEPT (rotation)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ±0.4875.63 ± 0.46
IEPT (shuffling)Conv4-6455.57 ± 0.6072.84 ± 0.5457.81 ± 0.4874.92 ± 0.50
" + }, + { + "category_id": 6, + "poly": [ + 298, + 1797, + 1398, + 1797, + 1398, + 1858, + 298, + 1858 + ], + "score": 0.927 + }, + { + "category_id": 6, + "poly": [ + 294, + 778, + 1398, + 778, + 1398, + 839, + 294, + 839 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 812, + 75, + 812, + 104, + 300, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 301, + 227, + 507, + 227, + 507, + 261, + 301, + 261 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2113, + 835, + 2113 + ], + "score": 0.861 + }, + { + "category_id": 4, + "poly": [ + 400, + 1303, + 1288, + 1303, + 1288, + 1336, + 400, + 1336 + ], + "score": 0.855 + }, + { + "category_id": 0, + "poly": [ + 299, + 303, + 985, + 303, + 985, + 336, + 299, + 336 + ], + "score": 0.817 + }, + { + "category_id": 1, + "poly": [ + 299, + 1412, + 1109, + 1412, + 1109, + 1445, + 299, + 1445 + ], + "score": 0.487 + }, + { + "category_id": 0, + "poly": [ + 313, + 1413, + 1104, + 1413, + 1104, + 1444, + 313, + 1444 + ], + "score": 0.235 + }, + { + "category_id": 13, + "poly": [ + 974, + 641, + 1039, + 641, + 1039, + 670, + 974, + 670 + ], + "score": 0.87, + "latex": "3- 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 860, + 641, + 924, + 641, + 924, + 670, + 860, + 670 + ], + "score": 0.86, + "latex": "1 - 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1093, + 459, + 1140, + 459, + 1140, + 487, + 1093, + 487 + ], + "score": 0.85, + "latex": "2 ^ { \\ast } 2" + }, + { + "category_id": 13, + "poly": [ + 1367, + 367, + 1396, + 367, + 1396, + 396, + 1367, + 396 + ], + "score": 0.82, + "latex": "0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 326, + 1568, + 353, + 1568, + 353, + 1600, + 326, + 1600 + ], + "score": 0.82, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 297, + 397, + 340, + 397, + 340, + 427, + 297, + 427 + ], + "score": 0.81, + "latex": "9 0 °" + }, + { + "category_id": 13, + "poly": [ + 596, + 1569, + 625, + 1569, + 625, + 1595, + 596, + 1595 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 906, + 1538, + 931, + 1538, + 931, + 1569, + 906, + 1569 + ], + "score": 0.77, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 352, + 397, + 408, + 397, + 408, + 427, + 352, + 427 + ], + "score": 0.75, + "latex": "1 8 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 941, + 1538, + 971, + 1538, + 971, + 1566, + 941, + 1566 + ], + "score": 0.73, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 420, + 397, + 480, + 397, + 480, + 427, + 420, + 427 + ], + "score": 0.73, + "latex": "2 7 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 1022, + 1538, + 1047, + 1538, + 1047, + 1566, + 1022, + 1566 + ], + "score": 0.64, + "latex": "V" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1795.0, + 1402.0, + 1795.0, + 1402.0, + 1833.0, + 294.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1826.0, + 1029.0, + 1826.0, + 1029.0, + 1863.0, + 294.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 772.0, + 1403.0, + 772.0, + 1403.0, + 817.0, + 292.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 810.0, + 868.0, + 810.0, + 868.0, + 842.0, + 294.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 816.0, + 72.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 222.0, + 513.0, + 222.0, + 513.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1297.0, + 1291.0, + 1297.0, + 1291.0, + 1342.0, + 398.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 303.0, + 989.0, + 303.0, + 989.0, + 339.0, + 296.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1411.0, + 1111.0, + 1411.0, + 1111.0, + 1449.0, + 306.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 365.0, + 1366.0, + 365.0, + 1366.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 365.0, + 1406.0, + 365.0, + 1406.0, + 403.0, + 1397.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 395.0, + 351.0, + 395.0, + 351.0, + 432.0, + 341.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 395.0, + 419.0, + 395.0, + 419.0, + 432.0, + 409.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 395.0, + 1405.0, + 395.0, + 1405.0, + 432.0, + 481.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 426.0, + 1405.0, + 426.0, + 1405.0, + 462.0, + 294.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 456.0, + 1092.0, + 456.0, + 1092.0, + 494.0, + 294.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 456.0, + 1406.0, + 456.0, + 1406.0, + 494.0, + 1141.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 489.0, + 1405.0, + 489.0, + 1405.0, + 523.0, + 295.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 517.0, + 1407.0, + 517.0, + 1407.0, + 555.0, + 294.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 550.0, + 1406.0, + 550.0, + 1406.0, + 584.0, + 295.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 579.0, + 1406.0, + 579.0, + 1406.0, + 614.0, + 295.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 609.0, + 1405.0, + 609.0, + 1405.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 640.0, + 859.0, + 640.0, + 859.0, + 675.0, + 294.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 640.0, + 973.0, + 640.0, + 973.0, + 675.0, + 925.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 640.0, + 1406.0, + 640.0, + 1406.0, + 675.0, + 1040.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 705.0, + 294.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 700.0, + 859.0, + 700.0, + 859.0, + 737.0, + 294.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 1405.0, + 1477.0, + 1405.0, + 1510.0, + 296.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1503.0, + 1407.0, + 1503.0, + 1407.0, + 1543.0, + 292.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1533.0, + 905.0, + 1533.0, + 905.0, + 1576.0, + 291.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1533.0, + 940.0, + 1533.0, + 940.0, + 1576.0, + 932.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1533.0, + 1021.0, + 1533.0, + 1021.0, + 1576.0, + 972.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1533.0, + 1407.0, + 1533.0, + 1407.0, + 1576.0, + 1048.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1565.0, + 325.0, + 1565.0, + 325.0, + 1604.0, + 294.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1565.0, + 595.0, + 1565.0, + 595.0, + 1604.0, + 354.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1565.0, + 1406.0, + 1565.0, + 1406.0, + 1604.0, + 626.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1596.0, + 1407.0, + 1596.0, + 1407.0, + 1633.0, + 294.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1663.0, + 295.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1656.0, + 1405.0, + 1656.0, + 1405.0, + 1695.0, + 292.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1689.0, + 1405.0, + 1689.0, + 1405.0, + 1725.0, + 294.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1721.0, + 882.0, + 1721.0, + 882.0, + 1754.0, + 295.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1412.0, + 1112.0, + 1412.0, + 1112.0, + 1447.0, + 296.0, + 1447.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1404, + 1819, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1570, + 1405, + 1570, + 1405, + 1692, + 298, + 1692 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 297, + 1333, + 1405, + 1333, + 1405, + 1456, + 297, + 1456 + ], + "score": 0.96 + }, + { + "category_id": 3, + "poly": [ + 314, + 227, + 1357, + 227, + 1357, + 1314, + 314, + 1314 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 299, + 1503, + 924, + 1503, + 924, + 1535, + 299, + 1535 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 299, + 1754, + 934, + 1754, + 934, + 1785, + 299, + 1785 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 863, + 2089, + 863, + 2112, + 836, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 13, + "poly": [ + 372, + 1395, + 415, + 1395, + 415, + 1424, + 372, + 1424 + ], + "score": 0.85, + "latex": "9 0 °" + }, + { + "category_id": 13, + "poly": [ + 427, + 1395, + 483, + 1395, + 483, + 1424, + 427, + 1424 + ], + "score": 0.84, + "latex": "1 8 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 494, + 1395, + 554, + 1395, + 554, + 1425, + 494, + 1425 + ], + "score": 0.82, + "latex": "2 7 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 331, + 1395, + 361, + 1395, + 361, + 1424, + 331, + 1424 + ], + "score": 0.81, + "latex": "0 ^ { \\circ }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1334.0, + 1403.0, + 1334.0, + 1403.0, + 1367.0, + 295.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1363.0, + 1406.0, + 1363.0, + 1406.0, + 1399.0, + 292.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1395.0, + 330.0, + 1395.0, + 330.0, + 1431.0, + 294.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1395.0, + 371.0, + 1395.0, + 371.0, + 1431.0, + 362.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1395.0, + 426.0, + 1395.0, + 426.0, + 1431.0, + 416.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1395.0, + 493.0, + 1395.0, + 493.0, + 1431.0, + 484.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1395.0, + 1405.0, + 1395.0, + 1405.0, + 1431.0, + 555.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1426.0, + 748.0, + 1426.0, + 748.0, + 1458.0, + 296.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 265.0, + 424.0, + 265.0, + 424.0, + 275.0, + 414.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 270.0, + 435.0, + 270.0, + 435.0, + 282.0, + 427.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 262.0, + 670.0, + 262.0, + 670.0, + 309.0, + 641.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 271.0, + 1045.0, + 271.0, + 1045.0, + 282.0, + 1036.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 273.0, + 1062.0, + 273.0, + 1062.0, + 280.0, + 1056.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 261.0, + 1317.0, + 261.0, + 1317.0, + 283.0, + 1301.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 282.0, + 443.0, + 282.0, + 443.0, + 292.0, + 434.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 278.0, + 473.0, + 278.0, + 473.0, + 288.0, + 464.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 280.0, + 676.0, + 280.0, + 676.0, + 298.0, + 662.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 290.0, + 456.0, + 290.0, + 456.0, + 306.0, + 419.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 290.0, + 688.0, + 290.0, + 688.0, + 296.0, + 682.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 286.0, + 1045.0, + 286.0, + 1045.0, + 297.0, + 1035.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 299.0, + 685.0, + 299.0, + 685.0, + 309.0, + 674.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 298.0, + 1046.0, + 298.0, + 1046.0, + 314.0, + 1031.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 298.0, + 1203.0, + 298.0, + 1203.0, + 307.0, + 1193.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 303.0, + 428.0, + 303.0, + 428.0, + 319.0, + 411.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 306.0, + 1328.0, + 306.0, + 1328.0, + 317.0, + 1316.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 313.0, + 906.0, + 313.0, + 906.0, + 323.0, + 893.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 323.0, + 423.0, + 323.0, + 423.0, + 333.0, + 414.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 318.0, + 1315.0, + 318.0, + 1315.0, + 333.0, + 1296.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 327.0, + 379.0, + 327.0, + 379.0, + 344.0, + 365.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 338.0, + 434.0, + 338.0, + 434.0, + 349.0, + 422.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 338.0, + 1020.0, + 338.0, + 1020.0, + 366.0, + 985.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 405.0, + 388.0, + 405.0, + 388.0, + 443.0, + 344.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 408.0, + 1221.0, + 408.0, + 1221.0, + 461.0, + 1177.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 428.0, + 390.0, + 428.0, + 390.0, + 480.0, + 339.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 439.0, + 811.0, + 439.0, + 811.0, + 450.0, + 801.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 448.0, + 823.0, + 448.0, + 823.0, + 458.0, + 815.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 449.0, + 832.0, + 449.0, + 832.0, + 455.0, + 826.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 435.0, + 1198.0, + 435.0, + 1198.0, + 453.0, + 1183.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 440.0, + 1279.0, + 440.0, + 1279.0, + 451.0, + 1267.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 440.0, + 1295.0, + 440.0, + 1295.0, + 455.0, + 1280.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 439.0, + 1339.0, + 439.0, + 1339.0, + 515.0, + 1300.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 453.0, + 1306.0, + 453.0, + 1306.0, + 464.0, + 1296.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 459.0, + 584.0, + 459.0, + 584.0, + 469.0, + 573.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 469.0, + 365.0, + 469.0, + 365.0, + 481.0, + 355.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 468.0, + 775.0, + 468.0, + 775.0, + 474.0, + 769.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 468.0, + 796.0, + 468.0, + 796.0, + 474.0, + 790.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 477.0, + 383.0, + 477.0, + 383.0, + 490.0, + 372.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 475.0, + 697.0, + 475.0, + 697.0, + 498.0, + 666.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 482.0, + 774.0, + 482.0, + 774.0, + 496.0, + 765.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 491.0, + 1098.0, + 491.0, + 1098.0, + 501.0, + 1086.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 506.0, + 779.0, + 506.0, + 779.0, + 516.0, + 769.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 580.0, + 593.0, + 580.0, + 593.0, + 591.0, + 581.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 572.0, + 903.0, + 572.0, + 903.0, + 611.0, + 866.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 581.0, + 1270.0, + 581.0, + 1270.0, + 591.0, + 1260.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 591.0, + 589.0, + 591.0, + 589.0, + 613.0, + 557.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 590.0, + 1246.0, + 590.0, + 1246.0, + 606.0, + 1231.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 625.0, + 871.0, + 625.0, + 871.0, + 658.0, + 843.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 615.0, + 1094.0, + 615.0, + 1094.0, + 623.0, + 1084.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 628.0, + 1114.0, + 628.0, + 1114.0, + 639.0, + 1102.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 616.0, + 1210.0, + 616.0, + 1210.0, + 642.0, + 1184.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 606.0, + 1239.0, + 606.0, + 1239.0, + 647.0, + 1215.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 636.0, + 1206.0, + 636.0, + 1206.0, + 656.0, + 1191.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 660.0, + 385.0, + 660.0, + 385.0, + 673.0, + 374.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 654.0, + 1061.0, + 654.0, + 1061.0, + 670.0, + 1044.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 651.0, + 1225.0, + 651.0, + 1225.0, + 665.0, + 1215.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 717.0, + 412.0, + 717.0, + 412.0, + 754.0, + 375.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 727.0, + 371.0, + 727.0, + 371.0, + 745.0, + 356.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 743.0, + 492.0, + 743.0, + 492.0, + 753.0, + 483.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 751.0, + 374.0, + 751.0, + 374.0, + 771.0, + 353.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 759.0, + 1119.0, + 759.0, + 1119.0, + 773.0, + 1109.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 758.0, + 1209.0, + 758.0, + 1209.0, + 773.0, + 1189.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 772.0, + 381.0, + 772.0, + 381.0, + 837.0, + 348.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 771.0, + 1201.0, + 771.0, + 1201.0, + 784.0, + 1184.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 783.0, + 1081.0, + 783.0, + 1081.0, + 792.0, + 1070.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 806.0, + 1102.0, + 806.0, + 1102.0, + 825.0, + 1077.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 787.0, + 1331.0, + 787.0, + 1331.0, + 826.0, + 1288.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 887.0, + 434.0, + 887.0, + 434.0, + 908.0, + 414.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 883.0, + 445.0, + 883.0, + 445.0, + 898.0, + 435.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 873.0, + 673.0, + 873.0, + 673.0, + 911.0, + 641.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 871.0, + 806.0, + 871.0, + 806.0, + 913.0, + 764.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 878.0, + 1084.0, + 878.0, + 1084.0, + 887.0, + 1073.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 890.0, + 1098.0, + 890.0, + 1098.0, + 899.0, + 1088.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 897.0, + 1316.0, + 897.0, + 1316.0, + 908.0, + 1304.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 909.0, + 800.0, + 909.0, + 800.0, + 918.0, + 790.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 907.0, + 1278.0, + 907.0, + 1278.0, + 928.0, + 1259.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 939.0, + 380.0, + 939.0, + 380.0, + 949.0, + 370.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 951.0, + 435.0, + 951.0, + 435.0, + 965.0, + 427.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 950.0, + 1207.0, + 950.0, + 1207.0, + 960.0, + 1198.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 958.0, + 476.0, + 958.0, + 476.0, + 978.0, + 456.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 961.0, + 827.0, + 961.0, + 827.0, + 977.0, + 812.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 957.0, + 1204.0, + 957.0, + 1204.0, + 972.0, + 1184.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 971.0, + 700.0, + 971.0, + 700.0, + 981.0, + 686.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1033.0, + 571.0, + 1033.0, + 571.0, + 1045.0, + 560.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1022.0, + 916.0, + 1022.0, + 916.0, + 1055.0, + 860.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1054.0, + 837.0, + 1054.0, + 837.0, + 1076.0, + 806.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1047.0, + 876.0, + 1047.0, + 876.0, + 1067.0, + 849.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1054.0, + 902.0, + 1054.0, + 902.0, + 1065.0, + 892.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1055.0, + 1093.0, + 1055.0, + 1093.0, + 1065.0, + 1083.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1056.0, + 1118.0, + 1056.0, + 1118.0, + 1068.0, + 1106.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1065.0, + 380.0, + 1065.0, + 380.0, + 1075.0, + 369.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1070.0, + 438.0, + 1070.0, + 438.0, + 1080.0, + 429.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1070.0, + 661.0, + 1070.0, + 661.0, + 1086.0, + 645.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1072.0, + 823.0, + 1072.0, + 823.0, + 1081.0, + 813.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1061.0, + 893.0, + 1061.0, + 893.0, + 1076.0, + 877.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1065.0, + 1114.0, + 1065.0, + 1114.0, + 1081.0, + 1098.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1083.0, + 455.0, + 1083.0, + 455.0, + 1094.0, + 445.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1081.0, + 488.0, + 1081.0, + 488.0, + 1091.0, + 473.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1083.0, + 853.0, + 1083.0, + 853.0, + 1093.0, + 844.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 1097.0, + 454.0, + 1097.0, + 454.0, + 1107.0, + 444.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1092.0, + 1121.0, + 1092.0, + 1121.0, + 1124.0, + 1094.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1090.0, + 1210.0, + 1090.0, + 1210.0, + 1101.0, + 1199.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1109.0, + 451.0, + 1109.0, + 451.0, + 1141.0, + 404.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1109.0, + 659.0, + 1109.0, + 659.0, + 1137.0, + 615.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1111.0, + 705.0, + 1111.0, + 705.0, + 1135.0, + 677.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1111.0, + 1095.0, + 1111.0, + 1095.0, + 1135.0, + 1078.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1124.0, + 1113.0, + 1124.0, + 1113.0, + 1135.0, + 1102.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1118.0, + 1232.0, + 1118.0, + 1232.0, + 1134.0, + 1216.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1198.0, + 1039.0, + 1198.0, + 1039.0, + 1209.0, + 1029.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1216.0, + 422.0, + 1216.0, + 422.0, + 1222.0, + 415.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1218.0, + 616.0, + 1218.0, + 616.0, + 1229.0, + 606.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1213.0, + 871.0, + 1213.0, + 871.0, + 1222.0, + 860.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1208.0, + 892.0, + 1208.0, + 892.0, + 1223.0, + 882.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1214.0, + 997.0, + 1214.0, + 997.0, + 1227.0, + 986.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1232.0, + 370.0, + 1232.0, + 370.0, + 1248.0, + 354.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1233.0, + 590.0, + 1233.0, + 590.0, + 1273.0, + 555.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1228.0, + 864.0, + 1228.0, + 864.0, + 1239.0, + 854.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1226.0, + 904.0, + 1226.0, + 904.0, + 1237.0, + 895.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1243.0, + 1025.0, + 1243.0, + 1025.0, + 1255.0, + 1015.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1224.0, + 1198.0, + 1224.0, + 1198.0, + 1234.0, + 1187.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1227.0, + 1327.0, + 1227.0, + 1327.0, + 1248.0, + 1315.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1249.0, + 443.0, + 1249.0, + 443.0, + 1269.0, + 395.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1248.0, + 613.0, + 1248.0, + 613.0, + 1264.0, + 598.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1255.0, + 907.0, + 1255.0, + 907.0, + 1261.0, + 901.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1256.0, + 1005.0, + 1256.0, + 1005.0, + 1266.0, + 997.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1261.0, + 1106.0, + 1261.0, + 1106.0, + 1271.0, + 1098.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1270.0, + 385.0, + 1270.0, + 385.0, + 1292.0, + 345.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1266.0, + 398.0, + 1266.0, + 398.0, + 1276.0, + 388.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1271.0, + 420.0, + 1271.0, + 420.0, + 1286.0, + 406.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1266.0, + 626.0, + 1266.0, + 626.0, + 1277.0, + 603.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 275.5, + 1303.0, + 275.5, + 1303.0, + 290.0, + 1296.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 301.5, + 397.0, + 301.5, + 397.0, + 314.5, + 393.0, + 314.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 323.0, + 1000.0, + 323.0, + 1000.0, + 342.0, + 976.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 441.0, + 1323.0, + 441.0, + 1323.0, + 452.0, + 1295.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 481.0, + 486.0, + 481.0, + 486.0, + 518.0, + 447.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 465.5, + 988.0, + 465.5, + 988.0, + 473.0, + 976.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.25, + 474.0, + 611.25, + 474.0, + 611.25, + 486.0, + 579.25, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 593.5, + 600.0, + 593.5, + 600.0, + 613.5, + 583.0, + 613.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 588.0, + 368.0, + 588.0, + 368.0, + 596.5, + 347.0, + 596.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.75, + 602.5, + 389.75, + 602.5, + 389.75, + 613.5, + 373.75, + 613.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.75, + 596.0, + 915.75, + 596.0, + 915.75, + 619.0, + 889.75, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 615.5, + 415.0, + 615.5, + 415.0, + 637.5, + 385.0, + 637.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 640.0, + 1076.0, + 640.0, + 1076.0, + 652.0, + 1056.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 720.0, + 700.0, + 720.0, + 700.0, + 758.0, + 654.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.75, + 722.0, + 1017.75, + 722.0, + 1017.75, + 754.0, + 985.75, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 786.5, + 1069.0, + 786.5, + 1069.0, + 804.0, + 1054.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 790.5, + 1088.0, + 790.5, + 1088.0, + 811.5, + 1063.0, + 811.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 807.5, + 610.0, + 807.5, + 610.0, + 826.5, + 588.0, + 826.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 888.5, + 422.0, + 888.5, + 422.0, + 900.0, + 385.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 967.5, + 425.0, + 967.5, + 425.0, + 980.5, + 412.0, + 980.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1036.5, + 590.0, + 1036.5, + 590.0, + 1063.5, + 578.0, + 1063.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1035.0, + 382.0, + 1035.0, + 382.0, + 1070.0, + 340.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.25, + 1073.0, + 453.25, + 1073.0, + 453.25, + 1084.0, + 436.25, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.75, + 1080.0, + 471.75, + 1080.0, + 471.75, + 1095.0, + 458.75, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.25, + 1088.0, + 382.25, + 1088.0, + 382.25, + 1105.5, + 366.25, + 1105.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1192.0, + 997.0, + 1192.0, + 997.0, + 1212.0, + 991.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1503.0, + 927.0, + 1503.0, + 927.0, + 1539.0, + 296.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1753.0, + 938.0, + 1753.0, + 938.0, + 1789.0, + 296.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1817.0, + 1405.0, + 1817.0, + 1405.0, + 1859.0, + 293.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1408.0, + 1910.0, + 1408.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1203.0, + 2004.0, + 1203.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1571.0, + 1405.0, + 1571.0, + 1405.0, + 1603.0, + 296.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1637.0, + 295.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 1405.0, + 1632.0, + 1405.0, + 1665.0, + 293.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1661.0, + 1341.0, + 1661.0, + 1341.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1311, + 1405, + 1311, + 1405, + 1528, + 298, + 1528 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 326, + 1625, + 1372, + 1625, + 1372, + 1775, + 326, + 1775 + ], + "score": 0.978, + "html": "
MethodBackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
IEPT (Linst-SimCLR)Conv4-6456.04 ± 0.4473.67± 0.4158.24± 0.4375.59± 0.41
IEPT (Lepis-Pairwise)Conv4-6455.95 ± 0.4773.72 ±0.4057.91 ± 0.4575.28±0.40
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
" + }, + { + "category_id": 1, + "poly": [ + 298, + 802, + 1404, + 802, + 1404, + 929, + 298, + 929 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 311, + 1057, + 1387, + 1057, + 1387, + 1177, + 311, + 1177 + ], + "score": 0.976, + "html": "
miniImageNettieredImageNet
MethodBackbone1-shot5-shot1-shot5-shot
Laux+LinstConv4-6453.25 ± 0.4671.50 ± 0.4255.06 ± 0.4472.87 ± 0.42
IEPT (ours)Conv4-6456.26 ± 0.4573.91 ± 0.3458.25 ± 0.4875.63 ± 0.46
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1910, + 1405, + 1910, + 1405, + 2034, + 299, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 327, + 239, + 1376, + 239, + 1376, + 584, + 327, + 584 + ], + "score": 0.968 + }, + { + "category_id": 6, + "poly": [ + 297, + 993, + 1398, + 993, + 1398, + 1056, + 297, + 1056 + ], + "score": 0.938 + }, + { + "category_id": 4, + "poly": [ + 294, + 611, + 1401, + 611, + 1401, + 675, + 294, + 675 + ], + "score": 0.934 + }, + { + "category_id": 6, + "poly": [ + 381, + 1591, + 1297, + 1591, + 1297, + 1624, + 381, + 1624 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 298, + 1250, + 1060, + 1250, + 1060, + 1282, + 298, + 1282 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2113, + 835, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 299, + 741, + 1009, + 741, + 1009, + 774, + 299, + 774 + ], + "score": 0.78 + }, + { + "category_id": 0, + "poly": [ + 300, + 1847, + 1170, + 1847, + 1170, + 1881, + 300, + 1881 + ], + "score": 0.76 + }, + { + "category_id": 0, + "poly": [ + 299, + 741, + 1009, + 741, + 1009, + 774, + 299, + 774 + ], + "score": 0.173 + }, + { + "category_id": 1, + "poly": [ + 300, + 1847, + 1170, + 1847, + 1170, + 1881, + 300, + 1881 + ], + "score": 0.137 + }, + { + "category_id": 13, + "poly": [ + 764, + 896, + 933, + 896, + 933, + 930, + 764, + 930 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { i n t e g } + \\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 1074, + 804, + 1226, + 804, + 1226, + 835, + 1074, + 835 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 730, + 1344, + 793, + 1344, + 793, + 1377, + 730, + 1377 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 1170, + 1313, + 1233, + 1313, + 1233, + 1346, + 1170, + 1346 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 1136, + 994, + 1294, + 994, + 1294, + 1024, + 1136, + 1024 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { a u x } + \\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1405, + 1255, + 1405, + 1255, + 1435, + 1192, + 1435 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 1284, + 1313, + 1347, + 1313, + 1347, + 1343, + 1284, + 1343 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1593, + 1178, + 1593, + 1178, + 1625, + 1115, + 1625 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { e p i s }" + }, + { + "category_id": 13, + "poly": [ + 1229, + 1592, + 1292, + 1592, + 1292, + 1622, + 1229, + 1622 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { i n s t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1972, + 425, + 1972, + 425, + 2002, + 297, + 2002 + ], + "score": 0.87, + "latex": "0 . 7 \\% - 2 . 1 \\%" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 246.0, + 390.0, + 246.0, + 390.0, + 267.0, + 352.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 244.0, + 804.0, + 244.0, + 804.0, + 267.0, + 732.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 247.0, + 953.0, + 247.0, + 953.0, + 267.0, + 891.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 247.0, + 1355.0, + 247.0, + 1355.0, + 267.0, + 1286.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 263.0, + 405.0, + 263.0, + 405.0, + 283.0, + 354.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 263.0, + 977.0, + 263.0, + 977.0, + 282.0, + 891.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 276.0, + 422.0, + 276.0, + 422.0, + 302.0, + 352.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 277.0, + 925.0, + 277.0, + 925.0, + 298.0, + 890.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 297.0, + 422.0, + 297.0, + 422.0, + 315.0, + 354.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 293.0, + 927.0, + 293.0, + 927.0, + 313.0, + 891.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 313.0, + 421.0, + 313.0, + 421.0, + 332.0, + 354.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 314.0, + 770.0, + 314.0, + 770.0, + 334.0, + 733.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 310.0, + 919.0, + 310.0, + 919.0, + 327.0, + 891.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 314.0, + 1351.0, + 314.0, + 1351.0, + 333.0, + 1286.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 344.0, + 773.0, + 344.0, + 773.0, + 365.0, + 743.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 379.0, + 781.0, + 379.0, + 781.0, + 398.0, + 732.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 378.0, + 1375.0, + 378.0, + 1375.0, + 396.0, + 1285.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 440.0, + 796.0, + 440.0, + 796.0, + 463.0, + 732.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 444.0, + 1319.0, + 444.0, + 1319.0, + 460.0, + 1286.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 505.0, + 794.0, + 505.0, + 794.0, + 520.0, + 733.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 507.0, + 1315.0, + 507.0, + 1315.0, + 518.0, + 1291.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 564.0, + 796.0, + 564.0, + 796.0, + 587.0, + 730.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 570.0, + 1308.0, + 570.0, + 1308.0, + 583.0, + 1286.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 993.0, + 1135.0, + 993.0, + 1135.0, + 1028.0, + 295.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 993.0, + 1404.0, + 993.0, + 1404.0, + 1028.0, + 1295.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1026.0, + 1155.0, + 1026.0, + 1155.0, + 1058.0, + 297.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 610.0, + 1404.0, + 610.0, + 1404.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 642.0, + 1183.0, + 642.0, + 1183.0, + 678.0, + 296.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1580.0, + 1114.0, + 1580.0, + 1114.0, + 1634.0, + 384.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1580.0, + 1228.0, + 1580.0, + 1228.0, + 1634.0, + 1179.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1580.0, + 1306.0, + 1580.0, + 1306.0, + 1634.0, + 1293.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1247.0, + 1062.0, + 1247.0, + 1062.0, + 1285.0, + 295.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1846.0, + 1177.0, + 1846.0, + 1177.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 741.0, + 1013.0, + 741.0, + 1013.0, + 776.0, + 296.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1308.0, + 1169.0, + 1308.0, + 1169.0, + 1350.0, + 292.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1308.0, + 1283.0, + 1308.0, + 1283.0, + 1350.0, + 1234.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1308.0, + 1407.0, + 1308.0, + 1407.0, + 1350.0, + 1348.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1343.0, + 729.0, + 1343.0, + 729.0, + 1381.0, + 293.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1343.0, + 1406.0, + 1343.0, + 1406.0, + 1381.0, + 794.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1408.0, + 295.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1401.0, + 1191.0, + 1401.0, + 1191.0, + 1442.0, + 292.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1401.0, + 1406.0, + 1401.0, + 1406.0, + 1442.0, + 1256.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1474.0, + 292.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1467.0, + 1403.0, + 1467.0, + 1403.0, + 1498.0, + 296.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1496.0, + 1236.0, + 1496.0, + 1236.0, + 1531.0, + 296.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 802.0, + 1073.0, + 802.0, + 1073.0, + 838.0, + 293.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 802.0, + 1405.0, + 802.0, + 1405.0, + 838.0, + 1227.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 834.0, + 1406.0, + 834.0, + 1406.0, + 870.0, + 293.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 897.0, + 294.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 895.0, + 763.0, + 895.0, + 763.0, + 930.0, + 293.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 895.0, + 1037.0, + 895.0, + 1037.0, + 930.0, + 934.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1978.0, + 292.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2006.0, + 426.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 506.0, + 2002.0, + 506.0, + 2033.0, + 294.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 741.0, + 1013.0, + 741.0, + 1013.0, + 776.0, + 296.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1846.0, + 1177.0, + 1846.0, + 1177.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 933, + 1405, + 933, + 1405, + 1087, + 298, + 1087 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 340, + 281, + 1358, + 281, + 1358, + 402, + 340, + 402 + ], + "score": 0.971, + "html": "
BackboneminiImageNettieredImageNet
1-shot5-shot1-shot5-shot
Method MAMLConv4-6448.70±1.8463.10±0.9251.67 ± 1.8170.30±0.80
MAML+IEPTConv4-6449.68 ± 0.5065.22 ± 0.4852.85 ± 0.5271.04 ± 0.49
" + }, + { + "category_id": 3, + "poly": [ + 317, + 444, + 1361, + 444, + 1361, + 755, + 317, + 755 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 294, + 758, + 1402, + 758, + 1402, + 819, + 294, + 819 + ], + "score": 0.919 + }, + { + "category_id": 6, + "poly": [ + 344, + 249, + 1351, + 249, + 1351, + 281, + 344, + 281 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 864, + 2089, + 864, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 0, + "poly": [ + 299, + 876, + 841, + 876, + 841, + 906, + 299, + 906 + ], + "score": 0.736 + }, + { + "category_id": 13, + "poly": [ + 1083, + 933, + 1354, + 933, + 1354, + 966, + 1083, + 966 + ], + "score": 0.9, + "latex": "\\{ 0 . 1 , 0 . 5 , 1 . 0 , 5 . 0 , 1 0 . 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 604, + 1027, + 729, + 1027, + 729, + 1056, + 604, + 1056 + ], + "score": 0.87, + "latex": "w _ { 1 } = 1 0 . 0 " + }, + { + "category_id": 13, + "poly": [ + 1230, + 998, + 1267, + 998, + 1267, + 1025, + 1230, + 1025 + ], + "score": 0.87, + "latex": "w _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 787, + 938, + 823, + 938, + 823, + 965, + 787, + 965 + ], + "score": 0.86, + "latex": "w _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1145, + 999, + 1180, + 999, + 1180, + 1025, + 1145, + 1025 + ], + "score": 0.86, + "latex": "w _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 521, + 1028, + 555, + 1028, + 555, + 1056, + 521, + 1056 + ], + "score": 0.83, + "latex": "w _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 655, + 940, + 689, + 940, + 689, + 965, + 655, + 965 + ], + "score": 0.82, + "latex": "w _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 701, + 940, + 736, + 940, + 736, + 965, + 701, + 965 + ], + "score": 0.77, + "latex": "w _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 442.0, + 367.0, + 442.0, + 367.0, + 469.0, + 335.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 442.0, + 912.0, + 442.0, + 912.0, + 469.0, + 880.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 507.0, + 365.0, + 507.0, + 365.0, + 531.0, + 336.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 507.0, + 909.0, + 507.0, + 909.0, + 531.0, + 881.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 518.0, + 1024.0, + 518.0, + 1024.0, + 527.0, + 1015.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 534.0, + 364.0, + 534.0, + 364.0, + 646.0, + 311.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 534.0, + 908.0, + 534.0, + 908.0, + 646.0, + 856.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 615.0, + 464.0, + 615.0, + 464.0, + 637.0, + 435.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 615.0, + 1009.0, + 615.0, + 1009.0, + 637.0, + 980.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 633.0, + 364.0, + 633.0, + 364.0, + 658.0, + 337.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 644.0, + 465.0, + 644.0, + 465.0, + 666.0, + 434.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 635.0, + 908.0, + 635.0, + 908.0, + 658.0, + 882.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 644.0, + 1011.0, + 644.0, + 1011.0, + 666.0, + 980.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 670.0, + 467.0, + 670.0, + 467.0, + 695.0, + 433.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 670.0, + 1013.0, + 670.0, + 1013.0, + 695.0, + 978.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 696.0, + 367.0, + 696.0, + 367.0, + 724.0, + 335.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 696.0, + 913.0, + 696.0, + 913.0, + 723.0, + 880.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 712.0, + 381.0, + 712.0, + 381.0, + 738.0, + 347.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 712.0, + 491.0, + 712.0, + 491.0, + 740.0, + 456.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 712.0, + 601.0, + 712.0, + 601.0, + 738.0, + 565.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 712.0, + 708.0, + 712.0, + 708.0, + 740.0, + 672.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 710.0, + 821.0, + 710.0, + 821.0, + 740.0, + 777.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 712.0, + 926.0, + 712.0, + 926.0, + 738.0, + 891.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 712.0, + 1036.0, + 712.0, + 1036.0, + 740.0, + 1000.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 712.0, + 1145.0, + 712.0, + 1145.0, + 738.0, + 1108.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 712.0, + 1254.0, + 712.0, + 1254.0, + 740.0, + 1218.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 710.0, + 1365.0, + 710.0, + 1365.0, + 740.0, + 1322.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 731.0, + 609.0, + 731.0, + 609.0, + 761.0, + 553.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 731.0, + 1154.0, + 731.0, + 1154.0, + 761.0, + 1098.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 756.0, + 1403.0, + 756.0, + 1403.0, + 791.0, + 296.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 787.0, + 970.0, + 787.0, + 970.0, + 821.0, + 293.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 247.0, + 1355.0, + 247.0, + 1355.0, + 287.0, + 344.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 876.0, + 845.0, + 876.0, + 845.0, + 910.0, + 295.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 931.0, + 654.0, + 931.0, + 654.0, + 972.0, + 295.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 931.0, + 700.0, + 931.0, + 700.0, + 972.0, + 690.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 931.0, + 786.0, + 931.0, + 786.0, + 972.0, + 737.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 931.0, + 1082.0, + 931.0, + 1082.0, + 972.0, + 824.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 931.0, + 1406.0, + 931.0, + 1406.0, + 972.0, + 1355.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 965.0, + 1405.0, + 965.0, + 1405.0, + 998.0, + 296.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 991.0, + 1144.0, + 991.0, + 1144.0, + 1031.0, + 292.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 991.0, + 1229.0, + 991.0, + 1229.0, + 1031.0, + 1181.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 991.0, + 1407.0, + 991.0, + 1407.0, + 1031.0, + 1268.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1023.0, + 520.0, + 1023.0, + 520.0, + 1060.0, + 295.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1023.0, + 603.0, + 1023.0, + 603.0, + 1060.0, + 556.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1023.0, + 1406.0, + 1023.0, + 1406.0, + 1060.0, + 730.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 796.0, + 1054.0, + 796.0, + 1092.0, + 293.0, + 1092.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file